content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{sec:introduction}
\texttt{CountSketch}~\cite{CharikarCFC:2002} and its variant {\em feature hashing}~\cite{MoodyD:NIPS1989,WeinbergerDLSA:ICML2009} are
immensely popular dimensionality reduction methods that map
input vectors in $\mathbb{R}^n$ to their sketches in $\mathbb{R}^d$ (where $d\ll n$).
The methods have many applications in machine learning and data analysis and often are used as components in large models or pipelines~\cite{WeinbergerDLSA:ICML2009,pmlr-v5-shi09a,PhamPagh:KDD2013,ChenWTWC:ICML2015,KDD-2016-ChenWTWC,AmiraliSLDSB:ICML2018,SpringKMA:ICML2019,AhleKKPVWZ:SODA2020,CPW:NEURIPS2020}.
The mapping is specified by
{\em internal randomness} $\rho\sim \mathcal{D}$ that determines a set of
$d=\ell\cdot b$ linear measurements vectors $(\boldsymbol{\mu}^{(j,k)})_{j\in [\ell],k\in[b]}$ in $\mathbb{R}^n$.
The sketch $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ of a vector $\boldsymbol{v}\in \mathbb{R}^n$ is the matrix of $d$ linear measurements
\begin{equation} \label{sketchdef:eq}
\text{\texttt{Sketch}}_{\boldsymbol{\rho}}(\boldsymbol{v}) := \left(\langle \boldsymbol{\mu}^{(j,k)},\boldsymbol{v} \rangle\right)_{j\in[\ell], k\in [b]} .
\end{equation}
The salient properties of \texttt{CountSketch} are that (when setting $b=O(1/\epsilon)$ and $\ell = O(\log n)$) the $\ell_2$-heavy hitters of an input $\boldsymbol{v}$, that is, keys $i$ with $v_i^2 > \epsilon \|\boldsymbol{v}\|_2^2$, can be recovered from $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ and that the inner product of two vectors $\boldsymbol{v}$, $\boldsymbol{u}$ can be approximated from their respective sketches $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$, $\text{\texttt{Sketch}}_\rho(\boldsymbol{u})$. This recovery is performed by applying an appropriate {\em estimator} to the sketch, for example, the median estimator~\cite{CharikarCFC:2002} provides estimates on values of keys and supports heavy hitters recovery. But recovery can also be implicit, for example, when the sketch is used as a compression module in a Neural Network~\cite{ChenWTWC:ICML2015}, the recovery of features is learned.
Randomized data structures and algorithms, \texttt{CountSketch} included, are typically analysed under an assumption that the input sequence is generated in a way that does not depend on prior outputs and on the sketch randomness $\rho$. This assumption, however, does not always hold, for example, when there is an intention to construct an adversarial input or when the system has a feedback between inputs and outputs~\cite{SpringKMA:ICML2019,FetchSGD:ICML2020}.
The adaptive setting, where inputs may depend on prior outputs, is more challenging to analyse and
there is
growing interest in quantifying performance and in designing methods that are robust to adaptive inputs. Works in this vein span machine learning~\cite{szegedy2013intriguing,goodfellow2014explaining,athalye2018synthesizing,papernot2017practical}, adaptive data analysis~\cite{Freedman:1983,Ioannidis:2005,FreedmanParadox:2009,HardtUllman:FOCS2014,DworkFHPRR15}, dynamic graph algorithms~\cite{ShiloachEven:JACM1981,AhnGM:SODA2012,gawrychowskiMW:ICALP2020,GutenbergPW:SODA2020,Wajc:STOC2020, BKMNSS22}, and sketching and streaming algorithms~\cite{MironovNS:STOC2008,AhnGM:SODA2012,HardtW:STOC2013,BenEliezerJWY21,HassidimKMMS20,WoodruffZ21,AttiasCSS21,BEO21,GuptaJNRSW:NeurIPS2021,CLNSSS:ICML2022}.
Robustness to adaptive inputs can trivially be achieved by using a fresh data structure for each query, or more finely, for each time the output changes. Hence, $\ell$ independent replicas of a non-robust data structures suffice for supporting $\ell$ adaptive queries. A powerful connection between adaptive robustness and differential privacy \cite{DworkFHPRR15} and utilizing the workhorse of DP composition, yielded essentially a wrapper around $\ell$ independent replicas of a non-robust data structure that supports a quadratic number $\tilde{O}(\ell^2)$ of adaptive queries (or changes to the output) \cite{HassidimKMMS20,GuptaJNRSW:NeurIPS2021,BKMNSS22}.
For the problem of recovering heavy-hitters from $\text{\texttt{CountSketch}}$, the "wrapper" method supports $\approx \epsilon \ell^2$ adaptive queries. The current state of the art~\cite{CLNSSS:ICML2022} is a robust estimator that works with a variant of $\text{\texttt{CountSketch}}$ and supports $\tilde{O}(\ell^2)$ adaptive queries.
Lower bounds on the performance of algorithms to adaptive inputs are obtained by designing an {\em attack}, a sequence of input vectors, that yields a constructed input that is adversarial to the internal randomness $\rho$. Tight lower bounds on the robustness of statistical queries were established by \cite{HardtUllman:FOCS2014,SteinkeUllman:COLT2015}, who designed an attack with a number of queries that is quadratic in the sample size, which matches the known upper bounds \cite{DworkFHPRR15}. Their construction was based on fingerprinting codes \cite{BonehShaw:Crypto1995}. A downside of these constructions is that the inputs used in the attack are not "natural" and hence unlikely to shed some understanding on practical vulnerability in the presence of feedback. Hardt and Woodruff \cite{HardtW:STOC2013} provided an impossibility result for
the task of estimating the norm of the input within a constant factor from (general) linear sketches.
Their construction works with arbitrary correct estimators and produces an adversarial distribution over inputs where the sketch measurements are "far" from their expectations. The attack size, however, has a large polynomial dependence on the sketch size and is far from the respective upper bound.
Ben-Eliezer et al \cite{BenEliezerJWY21} present an attack on the AMS sketch~\cite{ams99} for the task of approximating the $\ell_2$-norm of the input vector. The attack is tailored to a simplified estimator that is linear in the set of linear measurements (whereas the "classic" estimator uses a median of measurements and is not linear). Their attack is efficient in that the number of queries is of the order of the sketch size, rendering the estimator non-robust. It also has an advantage of using "natural" inputs.
More recently, \cite{CLNSSS:ICML2022} presented attacks that are tailored to specific estimators for $\text{\texttt{CountSketch}}$, including an attack of size $O(\ell)$ on the classic median estimator and an attack of size $O(\ell^2)$ on their proposed robust estimator.
\subsection*{Contribution}
Existing works proposed attacks of size that is far from the corresponding known upper bounds or are tailored to a particular estimator. Specifically for $\text{\texttt{CountSketch}}$, there is an upper bound of $O(\ell^2)$ but it is not even known whether there exist estimators that support a super-quadratic number of adaptive inputs. This question is of particular importance because $\text{\texttt{CountSketch}}$ and its variants are the only known efficient sketching method that allow recovery of $\ell_2$-heavy hitters and approximating $\ell_2$ norms and inner products. Moreover, their form as linear measurements is particularly suitable for efficient implementations and integration as components in larger pipelines. Finally, a recent lower bound precludes hope for an efficient deterministic (and hence fully robust) sketch \cite{KamathPW:CCC2021}, so it is likely that the vulnerabilities of $\text{\texttt{CountSketch}}$ are inherent to $\ell_2$-heavy hitter (and approximate $\ell_2$-norm and inner product) recovery from a small sketch.
We construct a {\em universal} attack on \texttt{CountSketch}, that applies against any unknown, potentially non-linear, possibly state maintaining, estimator. We only require that the estimator is correct.
Our attack uses $O(\ell^2)$ queries, matching the $\tilde{O}(\ell^2)$ robust estimator upper bound \cite{CLNSSS:ICML2022}. Moreover, it suffices for the purpose of the attack that the estimator only reports a set of candidate heavy keys without their approximate values (we only require that heavy hitters are reported with very high probability and $0$ value keys are reported with very small probability). Our attack also applies against a correct inner-product estimator (that distinguishes between $\langle \boldsymbol{v},\boldsymbol{u}\rangle=0$ (reported with very small probability) and $\langle \boldsymbol{v},\boldsymbol{u}\rangle^2 \geq \epsilon \|\boldsymbol{v}\|_2^2 \|\boldsymbol{u}\|_2^2$ (reported with high probability.)) Additionally, we apply our attack to correct $\ell_2$-norm estimators applied to an AMS sketch~\cite{ams99} and obtain that an attack of size $O(\ell^2)$ suffices to construct an adversarial input. The AMS sketch can be viewed as a $\text{\texttt{CountSketch}}$ with $b=1$ and is similar to the Johnson Lindenstrauss transform \cite{JLlemma:1984}.
The product of our attack (with high probability) is an {\em adversarial input} $\boldsymbol{v}$ on which the
measurement values of $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ are very biased with respect to their distribution when $\rho\sim\mathcal{D}$. Specifically, the design of $\text{\texttt{CountSketch}}$ results in
linear measurements that are unbiased for any input $\boldsymbol{v}$ under the sketch distribution $\rho\sim \cal{D}$: For each key $i$ and measurement vectors $\boldsymbol{\mu}$ with $\mu_i \not = 0$ it holds that
$\textsf{E}_{\rho}[\langle \boldsymbol{v},\boldsymbol{\mu }\rangle/ \mu_i- v_i]=0$ but the corresponding expected values for our adversarial $\boldsymbol{v}$ in $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ are large
($\geq B \epsilon \|\boldsymbol{v}\|_2^2$ for a desired $B>1$).
This "bias" means that the known standard (and robust) estimators for heavy hitters and inner products would fail on this adversarial input. And generally estimators that satisfy the usual design goal of being correct on any input with high probability over the distribution of $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ may not be correct on the adversarial inputs.
We note however that our result does not preclude the existence of specialized estimators that are correct on our adversarial inputs. We only show that a construction of an input that is adversarial to $\rho$ is possible with any correct estimator.
Finally, our attacks uses "natural" inputs that constitute of a heavy key and random noise. The final adversarial input is a linear combination of the noise components according to the heavy hitter reports and is the only one that depends on prior outputs. The simplicity of this attack suggests "practical" vulnerability of this fundamental sketching technique.
\medskip
{\bf Technique}
Our attacks construct an adversarial input with respect to key $h$. The high level structure is to generate "random tails," $(\boldsymbol{z}^{(t)})_{t\in [r]}$, which are vectors with small random entries. Ideally, we would like to determine for each $\boldsymbol{z}^{(t)}$ whether it is biased up or down with respect to key $h$. Roughly, considering the set $T_h$ of $\ell$ measurement vectors with $\mu_h \in \{-1,1\}$, determine the sign $s^{*(t)}$ of $\frac{1}{\ell} \sum_{\boldsymbol{\mu} \in T_h} \langle \boldsymbol{\mu},\boldsymbol{z}^{(t)} \rangle \cdot \mu_h$.
If we had that, the linear combination
$\boldsymbol{z}^{*(A)}= \sum_{t\in[r]} s^{*(t)} \boldsymbol{z}^{(t)}$ (with large enough $r$) is
an adversarial input. The intuition why this helps is that the bias accumulates linearly with the number of tails $r$ whereas the standard deviation (essentially the $\ell_2$ norm), considering randomness of the selection of tails, increases proportionally to $\sqrt{r}$.
The attack strategy is then to design query vectors of the form
$v^{(t)}_h \boldsymbol{e}_h + \boldsymbol{z}^{(t)}$
so
that from
whether or not $h$ is reported as a heavy hitter candidate
we obtain $s^{(t)}$ that correlates with $s^{*(t)}$. A
higher correlation $\textsf{E}[s^{*(t)} s^{(t)}]$ yields more effective attacks:
With $\textsf{E}[s^{*(t)} s^{(t)}] = \Omega(1)$ we get attacks of size $r=O(\ell)$ and with $\textsf{E}[s^{*(t)} s^{(t)}]=\Omega(1/\sqrt{\ell})$ we get attack of size $r=O(\ell^2)$. We show that we can obtain $s^{(t)}$ with $\textsf{E}[s^{*(t)} s^{(t)}]=\Omega(1/\sqrt{\ell})$ (thus matching the upper bound) against any arbitrary and adaptable estimator as long as it is correct. The difficulty is that such an estimator can be non monotone in the value $v_h$ of the heavy key and change between queries.
Our approach is to select the value $v_h$ of the heavy key in the input vectors uniformly at random from an interval that covers the "uncertainty region" between values that must be reported as heavy to values that are hard to distinguish from $0$ and hence should not be reported. We then observe that {\em in expectation} over the random choice, any correct estimator must have a slight reporting advantage with larger $v_h$. Finally, we show that any particular sketch content is equally likely with symmetrically larger $v_h$ with a tail that is biased down or a smaller $v_h$ with a tail that is biased up. This translates to a slight reporting advantage for $h$ as a candidate heavy hitter when the tail is "biased up." Therefore by taking $s^{(t)}=1$ when $h$ is reported and $s^{(t)}=-1$ otherwise we have the desired correlation.
\medskip
{\bf Remark:} We note that our use of "noise vectors" that have sparsity larger than sketch size is necessary to break robustness. This because with high probability, from a $\text{\texttt{CountSketch}}$ with parameters $b=O(1/\epsilon)$ and $\ell = O(\log n)$, we can fully recover {\em all} vectors $\boldsymbol{v}\in\mathbb{R}^n$ with at most $O(1/\epsilon)$ non-zeros. Therefore, $\text{\texttt{CountSketch}}$ is fully robust to sufficiently sparse inputs, but turns out to be less robust on inputs where a set of
$O(1/\epsilon)$ heavy entries is augmented with noise.
\medskip
{\bf Related work:} In terms of techniques, our work is most related to \cite{CLNSSS:ICML2022} in that the structure of our attack vectors is similar to those used in \cite{CLNSSS:ICML2022} to construct a tailored attack on the classic estimator. The generalization however to a "universal" attack that is effective against arbitrary and unknown estimators was delicate and required multiple new ideas. Our contribution is also related and in a sense complementary to \cite{HardtW:STOC2013} that designed attack on linear sketches that applies with any correct estimator for (approximate) norms. Their attack is much less efficient in that its size is a higher degree polynomial and it uses dependent (adaptive) inputs (whereas with our attack only the final adversarial input depends on prior outputs). The product of their attack are constructed vectors that are in the (approximate) null space of the sketching matrix. These "noise" vectors can have large norms but are "invisible" in the sketch. When such "noise" is added to an input with a signal (say a heavy hitter), the "signal" is suppressed (entry no longer heavy) but can still be recovered from the sketch. Our attack fails the sketch matrix in a complementary way: we construct "noise" vectors that do not involve a signal (a heavy entry) but the sketch mimics a presence of that particular signal.
\medskip
{\bf Overview:}
Our attack is described in Section~\ref{sec:attack_structure} and is
analysed in subsequent sections.
The query vectors used in the attack have a simple sketch structure (described in Section~\ref{sketchdist:sec}) and simplify the estimation task (as described in Section~\ref{sec:estimators}). In Section~\ref{simplifiedminest:sec} we cast the attack on heavy hitters estimator as one on a {\em mean estimator} and carry out the remaining analysis in this simplified context. Section~\ref{minestproofs:sec} includes detailed proofs. In Section~\ref{AMS:sec} we apply our attack technique to $\ell_2$ norm estimators for the AMS sketch~\cite{ams99}.
\section{Preliminaries}
\label{sec:preliminaries}
We use boldface notation for vectors $\boldsymbol{v}$, non boldface
for scalars $v$, $\langle \boldsymbol{v},\boldsymbol{u} \rangle =\sum_i v_i u_i$ for
inner product, and $v\cdot u$ for scalar product. For a vector $\boldsymbol{v}\in \mathbb{R}^n$ we refer to $i\in[n]$ as a {\em key} and $v_i$ as the value of the $i$th key (entry) and denote by $\overline{v}= \frac{1}{n} \sum_{i=1}^n v_i$ the mean value.
For exposition clarity, we use $\approx$ to mean "within a small relative error."
We denote by $\mathcal{N}(v,\sigma^2)$ the normal distribution with mean $v$ and variance $\sigma^2$ and by $\boldsymbol{u} \sim \mathcal{N}_\ell(v,\sigma^2)$ a vector in $\mathbb{R}^\ell$ with entries that are i.i.d.\ $\mathcal{N}(v,\sigma^2)$. The probability density function of $\mathcal{N}_\ell(v,\sigma^2)$ is
\begin{equation} \label{pdfnormal:eq}
f_v(\boldsymbol{u}) =
\prod_{i\in[\ell]} \frac{1}{\sigma\sqrt{2\pi}} e^{-\frac{1}{2} \left(\frac{u_i-v}{\sigma}\right)^2}\ .
\end{equation}
\begin{definition} (heavy hitter) \label{HH:def}
For $\epsilon>0$, and a vector $\boldsymbol{v}\in \mathbb{R}^n$, key $i\in [n]$ is an $\ell_2$-$\epsilon$-{\em heavy hitter} if $v_i^2 \geq \epsilon \|\boldsymbol{v} \|_2^2$.
\end{definition}
Clearly, there can be at most $1/\epsilon$ $\ell_2$-$\epsilon$ heavy hitters.
\begin{definition} (Heavy Hitters Estimator) \label{HHproblem:def}
A $\ell_2$-$\epsilon$-heavy hitters estimator is
applied to a sketch of an input vector $\boldsymbol{v}\in\mathbb{R}^n$ and returns a set of entries $K\subset [n]$. The output is {\em correct} if $K$ includes all the $\ell_2$-$\epsilon$-heavy hitters keys and does not include keys with $v_i=0$.
\end{definition}
\begin{remark} \label{classic:rem}
This correctness definition is weaker than what the classic
estimator applied to $\text{\texttt{CountSketch}}$ provides~\cite{CharikarCFC:2002}:
A sketch with $\ell = O(\log (n/\delta))$ provides with probability $1-\delta$, approximate values for all $n$ keys that are within $\epsilon \|\boldsymbol{v}_{\mathsf{tail}(1/\epsilon)} \|_2^2$, noting that the tail (vector $\boldsymbol{v}_{\mathsf{tail}(1/\epsilon)}$ that is our input $\boldsymbol{v}$ with $1/\epsilon$ heaviest
entries nullified) has smaller norm than
$\boldsymbol{v}$, yielding tighter estimates. In particular, this yields a correct solution to the heavy hitters problem with probability $1-\delta$.
Since our focus here is on designing an attack, our design is stronger against weaker requirements since less information on the randomness is being revealed.
\end{remark}
\begin{definition} (Inner Product Estimator) \label{IPproblem:def}
An inner-product estimator is
applied to sketches of two input vectors $\boldsymbol{v},\boldsymbol{u}\in\mathbb{R}^n$ and returns $s\in \{-1,1\}$. The output is {\em correct}
if $s=-1$ when $\langle \boldsymbol{v},\boldsymbol{u}\rangle=0$ and is $s=1$ when
$\langle \boldsymbol{v},\boldsymbol{u}\rangle^2 \geq \epsilon \|\boldsymbol{v}\|_2^2 \|\boldsymbol{u}\|_2^2$.
\end{definition}
\subsection{\texttt{\text{\texttt{CountSketch}}}}
The sketch~\cite{CharikarCFC:2002} is specified by parameters $(n,\ell,b)$,
where $n$ is the dimension of input vectors, and $d= \ell\cdot b$.
The internal randomness $\rho$ specifies a set of random hash functions
$h_r:[n]\rightarrow [b]$ ($r\in [\ell]$) with the marginals that $\forall k\in [b]$, $i\in [n]$, $\textsf{Pr}[h_j(i) = k]=1/b$, and
$s_{j}:[n]\rightarrow \{-1,1\}$ ($ j\in [\ell]$) so that $\textsf{Pr}[s_{j}(i)=1]=1/2$.
These hash functions define $d = \ell\cdot b$ measurement vectors that are organized as $\ell$ sets of $b$ vectors each
$\boldsymbol{\mu}^{(j,k)}$ ($j\in [\ell]$, $k\in [b]$) where:
\[ \mu^{(j,k)}_i := \mathbbm{1}_{\{h_j(i) = k\}} s_{j}(i) .\]
For an input vector $\boldsymbol{v}\in\mathbb{R}^n$,
$\text{\texttt{Sketch}}_\rho(\boldsymbol{v}) := (\langle \boldsymbol{\mu}^{(j,k)},\boldsymbol{v}\rangle)_{j,k}$
is the set of the respective
measurement values.
Note that for each key $i\in[n]$ there are exactly $\ell$ measurement
vectors with a nonzero $i$th entry: $(\boldsymbol{\mu}^{(j,h_j(i))})_{j\in[\ell]}$ and these measurement vectors are independent (as the only dependency is between measurement in the same
set of $b$, and there is exactly one from each set).
The respective set of $\ell$ adjusted
measurements:
\begin{equation} \label{adjmeasure:eq}
(\langle \boldsymbol{\mu}^{(j,h_j(i))}, \boldsymbol{v} \rangle
\mu^{(j,h_j(i))}_i)_{j\in [\ell]}
\end{equation}
are unbiased estimates of $v_i$:
$\textsf{E}_\rho [\langle \boldsymbol{\mu}^{(j,h_j(i))}, \boldsymbol{v} \rangle
\mu^{(j,h_j(i))}_i ] = v_i$.
The median estimator~\cite{CharikarCFC:2002} uses the median adjusted measurement to estimate the value $v_i$ of each key $i$. The $O(1/\epsilon)$ keys with highest magnitude estimates are then reported as heavy hitters.
When the same randomness $\rho$ is used for $r$ non-adaptive inputs
(inputs selected independently of $\rho$ and prior outputs),
sketch parameter settings of $\ell=\log (r \cdot n/\delta)$ and $b =
O(\epsilon^{-1})$ guarantee that with probability $1-\delta$, all outputs are correct (in the sense of Remark~\ref{classic:rem}).
$\text{\texttt{CountSketch}}$ also supports estimation of inner products.
For two vectors $\boldsymbol{v}$, $\boldsymbol{u}$, we obtain an
unbiased estimate of their inner product from the respective
inner product of the $j\in [\ell]$th row of measurements:
\begin{equation} \label{innerproduct:eq}
\sum_{k\in[b]} \langle \boldsymbol{\mu}^{(j,k)}, \boldsymbol{v} \rangle \cdot
\langle \boldsymbol{\mu}^{(j,k)}, \boldsymbol{u} \rangle\ .
\end{equation}
The median of these $\ell$ estimates is within relative error $\sqrt{\epsilon}$ with probability $1-\exp(\Omega(-\ell))$.
We note that pairwise independent
hash functions $h_j$ and $s_{j}$ suffice for obtaining the guarantees of
Remark~\ref{classic:rem}
\cite{CharikarCFC:2002} whereas 4-wise independence is needed for approximate
inner products. The analysis of the attack we present here, however,
holds even under full randomness.
\subsection{Adversarial Input for \texttt{\text{\texttt{CountSketch}}}}
\begin{definition} [Adversarial input] \label{adversarialinput:def}
We say that an attack $\mathcal{A}$ that is applied to a sketch with randomness $\rho$ and outputs
$i\in [n]$ and $\boldsymbol{z}^{(A)}\in \{-1,0,1\}^n$ (with $z^{(A)}_i=0$) is
$(B,\beta)$-{\em adversarial} (for $B>1$) if, with probability at least $1-\beta$ over the randomness of $\rho,\mathcal{A}$, the adjusted measurements \eqref{adjmeasure:eq} satisfy:
\begin{equation} \label{adverse:def}
\textsf{Pr}_{\rho\sim\mathcal{D},\mathcal{A}}\left[ \frac{1}{\ell} \sum_{j\in [\ell]}
\langle \boldsymbol{\mu}^{(j,h_j(i))}, \boldsymbol{z}^{(A)} \rangle
\mu^{(j,h_j(i))}_i \geq \sqrt{\frac{B}{b}}\| \boldsymbol{z}^{(A)}\|_2 \right] \geq 1-\beta .
\end{equation}
\end{definition}
The adversarial input $\boldsymbol{z}^{(A)}$ is a noise vector (with no heavy hitters) but $\text{\texttt{Sketch}}_\rho(\boldsymbol{z}^{(A)})$ "looks like" (in terms of the average adjusted measurement of key $i$) a sketch of a vector with a heavy key $i$.
It follows from the standard analysis of $\text{\texttt{CountSketch}}$ that the event
\begin{equation} \label{iheavy:eq}
\frac{1}{\ell} \sum_{j\in [\ell]}
\langle \boldsymbol{\mu}^{(j,h_j(i))}, \boldsymbol{v} \rangle
\mu^{(j,h_j(i))}_i \geq \sqrt{\frac{B}{b}}\| \boldsymbol{v}\|_2\ .
\end{equation}
(that correspondence to~\eqref{adverse:def}) is very likely for vectors $\boldsymbol{v}$ such that $h$ is a heavy hitter (Definition~\ref{HH:def}) and extremely unlikely when $h$ is not a heavy hitter, and in particular, when $v_h=0$.
Similarly for inner products,
considering inner product
$\boldsymbol{v}$
with the standard basis vector
$\boldsymbol{e}_i$, the
sketch-based estimates \eqref{innerproduct:eq} of each $j\in \ell$ computed from $\text{\texttt{Sketch}}_\rho(\boldsymbol{e}_i)$ and
$\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$
are equal to the respective
adjusted measurement \eqref{adjmeasure:eq} of $i$ from
$\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$.
The event \eqref{iheavy:eq} is very likely when
$\langle \boldsymbol{e}_i, \boldsymbol{v}\rangle \geq
\epsilon \| \boldsymbol{v}\|_2^2$ and very unlikely when
$\langle \boldsymbol{e}_i, \boldsymbol{v}\rangle=0$, noting that
for our adversarial input $\boldsymbol{z}^{(A)}$ it holds that
$\langle \boldsymbol{e}_i, \boldsymbol{z}^{(A)}\rangle =0$.
While this follows from the standard analysis, the simple structure of our noise vectors allows for a particularly simple argument for bounding the probability of \eqref{iheavy:eq} for $\boldsymbol{v}$ with the structure of $\boldsymbol{z}^{(A)}$ : The distribution of an adjusted measurement of a $\boldsymbol{v} \in \{-1,0,1\}^n$ and $v_i=0$ with support size
$m=|\texttt{supp}{(\boldsymbol{v}}| = \|\boldsymbol{v}\|^2_2$ approaches $\mathcal{N}(0,\frac{m}{b})$ (for large $m/b$), and thus
the average approaches
$\mathcal{N}(0,\frac{m}{\ell\cdot b})$. Therefore, the probability of \eqref{adverse:def} on a random sketch of $\boldsymbol{z}^{(A)}$ is $\leq \exp(-\ell B/2)$ (applying tail bounds on the probability of value exceeding $\sqrt{\ell B}$ standard deviations).
\section{Attack Description}
\label{sec:attack_structure}
We describe our attack against heavy hitters estimators. The modifications needed for it to apply with inner product estimator are described in Section~\ref{IP:sec}.
The attack is an interaction between the following components:
\begin{itemize}
\item
Internal randomness $\rho\sim\mathcal{D}$ that specifies linear measurement vectors $(\boldsymbol{\mu}^{(j,k)})_{j\in[\ell],k\in[b]}$. The sketch $\text{\texttt{Sketch}}_\rho(\boldsymbol{v})$ of a vector
$\boldmath{v}\in \mathbb{R}^n$ is the set of measurements
$(\langle \boldsymbol{\mu}^{(j,k)},\boldsymbol{v}\rangle)_{j\in[\ell],k\in[b]}$.
\item
A {\em query-response algorithm} that at each step $t$ chooses a heavy hitters estimator (see Definition~\ref{HHproblem:def}).
The choice may depend on the randomness $\rho$ and prior queries and responses $(\text{\texttt{Sketch}}_\rho(\boldsymbol{v}^{(t')}),K^{(t')})_{t'<t}$.
The algorithm receives
$\text{\texttt{Sketch}}_\rho(\boldsymbol{v}^{(t)})$, applies the estimator to the sketch,
and outputs $K^{(t)}$.
\item
An {\em adversary} that issues a sequence of input queries $(\boldsymbol{v}^{(t)})_t$ and collects the responses $(K^{(t)})_t$.
The randomness $\rho$ and the sketches of the query vectors $(\text{\texttt{Sketch}}_\rho(\boldsymbol{v}^{(t)}))_t$ are not known to the adversary.
The goal is to construct an
{\em adversarial input vector} $\boldsymbol{z}^{(A)}$ (see Definition~\ref{adversarialinput:def}).
\end{itemize}
Our adversary generates the
{\em query vectors}
$(\boldsymbol{v}^{(t)})_{t\in [r]}$ non-adaptively
as described in Section~\ref{queryvec:sec}. The attack interaction
and its properties are stated in Section~\ref{sec:universal_attack}.
\subsection{Query Vectors} \label{queryvec:sec}
Our attack query vectors $(\boldsymbol{v}^{(t)})_{t\in [r]}$ of
the form:
\begin{equation} \label{input:eq}
\boldsymbol{v}^{(t)} := v^{(t)}_h \boldsymbol{e}_h +
\boldsymbol{z}^{(t)}\in \mathbb{R}^n,
\end{equation}
where $h$ is a special {\em heavy} key, that is selected uniformly
$h\sim \mathcal{U}[n]$ and remains fixed,
$\boldsymbol{e}_h$ is the standard basis vector (axis-aligned unit vector along $h$), and
the vectors $\boldsymbol{z}^{(t)}$ are
{\em tails}. The (randomized) construction of tails is described in
Algorithm~\ref{algo:tails}.
The tail vectors $(\boldsymbol{z}^{(t)})_{t\in[r]}$ have
support of size $|\texttt{supp}(\boldsymbol{z}^{(t)})| = m$ that does not
include key $h$
($h\not\in \texttt{supp}(\boldsymbol{z}^{(t)})$) and so that the supports of different tails are disjoint:
\[
t_1\not= t_2 \implies \texttt{supp}(\boldsymbol{z}^{(t_1)}) \cap \texttt{supp}(\boldsymbol{z}^{(t_2)}) =
\emptyset\ .
\]
For query $t$ and key $i\in\texttt{supp}(\boldsymbol{z}^{(t)})$, the values are selected
i.i.d.\ Rademacher $z^{(t)}_i \sim \mathcal{U}[\{-1,1\}]$.
Note that
$\|\boldsymbol{v}^{(t)}\|_2^2 = (v_h^{(t)})^2+ \|\boldsymbol{z}^{(t)}\|_2^2 = (v_h^{(t)})^2+ m$.
Note that the tails, and (as we shall see) the selection of
$v^{(t)}_h$, and hence the input vectors are constructed
non-adaptively. Only the
final adversarial input vector depends on the output of the estimator on prior queries. The parameter $m$ is set to a value that is polynomial in the sketch size and large enough so that certain approximations hold (see Section~\ref{sketchdist:sec}).
\begin{algorithm2e
{\small
\caption{\texttt{AttackTails}}
\label{algo:tails}
\DontPrintSemicolon
\KwIn{Input dimension $n$, support size $m$, number of tails $r$}
$h \gets \mathcal{U}[n]$ \tcp*{Special heavy hitter key}
$S\gets \{h\}$ \tcp*{Keys used in support}
\For{$t\in [r]$}{
$S' \gets $ random subset of size $m$ from $[n]\setminus S$\;
$\boldsymbol{z}^{(t)} \gets \boldsymbol{0}$\;
\ForEach{$i\in S'$}
{
$\boldsymbol{z}^{(t)}_i \sim \mathcal{U}[\{-1,1\}]$\;
}
$S \gets S\cup S'$
}
\Return{$h, (\boldsymbol{z}^{(t)})_{t\in [r]}$}
}
\end{algorithm2e}
\begin{remark}
The only piece of information needed from the output of the estimator is whether the particular key $h$ is reported as a candidate heavy hitter of $\boldsymbol{v}^{(t)}$, that is, whether $h\in K^{(t)}$.
Note that disclosing additional information
can only make the estimator {\em more} vulnerable to attacks.
\end{remark}
\subsection{Universal Attack}
\label{sec:universal_attack}
\begin{algorithm2e
{\small
\caption{\texttt{Universal Attack on $\text{\texttt{CountSketch}}$} Heavy Hitters Estimators}
\label{algo:attack}
\DontPrintSemicolon
Use $a = \Theta(\sqrt{\frac{\ln(1/\delta_2)}{\ell}})$
and $c = \Theta(1)$\tcp*{With universal constants as in Lemma~\ref{correctest:lemma}}
\KwIn{Initialized $\text{\texttt{CountSketch}}_\rho$ with parameters $(n,\ell,b)$, Query-response algorithm, number of queries $r$, tail support size $m$}
$(h,(\boldsymbol{z}^{(t)})_{t\in [r]}) \gets$
\texttt{AttackTails}$(n,m,r)$\tcp*{Algorithm~\ref{algo:tails}}
\For(\tcp*[h]{Compute Query Vectors}){$t\in [r]$}{
$v^{(t)}_h\sim \mathcal{U}[a\cdot \sigma, (c+2a)\cdot \sigma]$\tcp*{$\sigma := \sqrt{m/b}$}
$\boldsymbol{v}^{(t)} \gets v^{(t)}_h \boldsymbol{e}_h +
\boldsymbol{z}^{(t)}$\tcp*{Query vectors}
}
\For(\tcp*[h]{Apply Query Response}){$t\in [r]$}
{
Choose a correct HH estimator $M^{(t')} $\tcp*{With correct reporting function (Definition~\ref{correctmean:def}), may depend on $(v^{(t')}_h,K^{(t')},M^{(t')})_{t'<t}$ and $\rho$ }
$K^{(t)} \gets M^{(t)}(\text{\texttt{CountSketch}}_\rho(\boldsymbol{v}^{(t)}))$\tcp*{Apply estimator to sketch}
\leIf{$h\in K^{(t)}$}{$s^{(t)}\gets 1$}{$s^{(t)}\gets -1$}
}
\Return{$\boldsymbol{z}^{(A)} \gets \sum_{t\in[r]} s^{(t)} \boldsymbol{z}^{(t)} $}
\tcp*{Adversarial input}
}
\end{algorithm2e}
Our attack interaction is described in Algorithm~\ref{algo:attack}. We generate $r$ attack tails
using Algorithm~\ref{algo:tails}. We then construct $r$ queries of the form \eqref{input:eq} with
i.i.d.\ $v^{(t)}_h\sim \mathcal{U}[a\cdot \sigma, (c+2a)\cdot \sigma]$. At each step $t\in[r]$, we feed the sketch of $\boldsymbol{v}^{(t)}$ to the HH estimator selected by the query response algorithm and collect the output $K^{(t)}$ of the estimator. We then set $s^{(t)} \gets 1$ if $h\in K^{(t)}$ and $s^{(t)} \gets -1$ if $h\not\in K^{(t)}$.
The final step computes the
{\em adversarial input}:
\begin{equation}\label{adversetail:eq}
\boldsymbol{z}^{(A)} := \sum_{t\in[r]} s^{(t)} \boldsymbol{z}^{(t)} \ .
\end{equation}
The statements below only use measurement vectors with nonzero value for key $h$. To simplify the notation, we
use $\boldsymbol{\mu}^{(j)} := \boldsymbol{\mu}^{(j,h_j(h))}$ for $j\in[\ell]$.
For randomness $\rho$, we use the notation $\boldsymbol{\mu}^{(j)}(\rho)$ for the respective measurement
vectors and $\mathcal{A}(\rho)$ for the output distribution of Algorithm~\ref{algo:attack} applied with randomness $\rho$.
The adversarial input has $z^{(A)}_h = 0$ and norm $\|\boldsymbol{z}^{(A)} \|^2_2 = r\cdot m$ (it has support of size $r\cdot m$ with values in the support i.i.d\ Rademacher $U[\{-1,1\}]$).
When an adversarial input $\boldsymbol{z}^{(A)} \sim \mathcal{A}(\rho_0)$ is sketched using
a {\em random} $\rho\sim\mathcal{D}$ it holds that:
\begin{align*}
\forall j\in[\ell],\ \textsf{E}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0),\rho\sim\mathcal{D}}\left[\langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho)\rangle\right] &=0\ \\
\forall j\in[\ell],\ \textsf{Var}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0),\rho\sim\mathcal{D}}\left[\langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho) \rangle\right] &\approx \frac{r\cdot m}{b} = r\sigma^2\ .
\end{align*}
and since the $\ell$ measurements are independent we get:
\begin{align*}
\textsf{E}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0),\rho\sim\mathcal{D}} \left[\frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho)\rangle \cdot \mu^{(j)}_h(\rho)\right] &= 0\\
\textsf{Var}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0),\rho\sim\mathcal{D}} \left[\frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho)\rangle \cdot \mu^{(j)}_h(\rho)\right] &\approx \frac{r}{\ell} \cdot \sigma^2
\end{align*}
The adversarial input $\boldsymbol{z}^{(A)} \sim \mathcal{A}(\rho_0)$ behaves differently with respect to the particular randomness $\rho_0$ it was constructed for. We will establish the following:
\begin{lemma} [Properties of the adversarial input] \label{tailprop:lemma}
\begin{align*}
\textsf{E}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0)}\left[\frac{1}{\ell}\sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho_0)\rangle \cdot \mu^{(j)}_h(\rho_0)\right] &\approx \frac{r}{\ell} \cdot \frac{2\sigma}{c+a}\\
\textsf{Var}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0)}\left[\frac{1}{\ell}\sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho_0)\rangle \cdot \mu^{(j)}_h(\rho_0)\right] &\approx \frac{r}{\ell} \cdot \sigma^2
\end{align*}
\end{lemma}
The proof is provided in Section~\ref{simplifiedminest:sec} (following preparation in Sections~\ref{sketchdist:sec}-\ref{sec:estimators} and details in Section~\ref{minestproofs:sec}). The high level idea, as hinted in the introduction, is that we establish that $h\in K^{(t)}$ and thus $s^{(t)}=1$ is correlated with "positive bias", that is, with the event
$\frac{1}{\ell} \sum_{j\in [\ell]} \langle s^{(t)}\boldsymbol{z}^{(t)}, \boldsymbol{\mu}^{(j)}(\rho_0)\rangle \cdot \mu^{(j)}_h(\rho_0) > 0$. In the sum
$\sum_{t\in[r]} s^{(t)} \boldsymbol{z}^{(t)}$ the bias (which is ``forced'' error on the estimates) increases linearly with $r$ while the $\ell_2$ norm, which corresponds to the standard deviation of the error, increases
proportionally to $\sqrt{r}$.
With Lemma~\ref{tailprop:lemma} in hand, we are ready to show that $\boldsymbol{z}^{(A)}$ is an adversarial input (see Definition~\ref{adversarialinput:def}).
\begin{theorem}[Adversarial input]
If for $B>1$ we use attack of size $r = B\cdot \ell^2 $ then
\begin{align*}
\textsf{E}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0)}\left[\frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)(\rho_0)}\rangle \cdot \mu^{(j)}_h(\rho_0)\right] &\approx \frac{2}{c+a} \sqrt{\frac{B}{b}} \| \boldsymbol{z}^{(A)} \|_2\\
\textsf{Var}_{\rho_0\sim \mathcal{D},\boldsymbol{z}^{(A)}\sim \mathcal{A}(\rho_0)}\left[\frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}(\rho_0)\rangle \cdot \mu^{(j)}_h(\rho_0)\right] &\approx \frac{1}{\ell\cdot b} \|\boldsymbol{z}^{(A)} \|^2_2\ .
\end{align*}
\end{theorem}
\begin{proof}
Using Lemma~\ref{tailprop:lemma},
the expectation of the average with attack size $r=B\ell^2$ is
\begin{align*}
\frac{r}{\ell} \cdot \frac{2\sigma}{c+a} &= \frac{r}{\ell} \sqrt{\frac{m}{b}} \cdot \frac{2}{c+a} \\
&= \frac{2}{c+a} \frac{\sqrt{r}}{\ell} \frac{1}{\sqrt{b}} \sqrt{r\cdot m} \\
&= \frac{2}{c+a} \sqrt{\frac{B}{b}} \|\boldsymbol{z}^{(A)} \|_2 & \text{since $\|\boldsymbol{z}^{(A)} \|_2 = \sqrt{r\cdot m}$ }
\end{align*}
The variance of the average is
\begin{align*}
\frac{r}{\ell} \sigma^2 = \frac{r\cdot m}{\ell\cdot b}= \frac{1}{\ell\cdot b} \|\boldsymbol{z}^{(A)} \|^2_2
\end{align*}
\end{proof}
\subsection{Attack with an inner-product estimator} \label{IP:sec}
We describe the modifications to Algorithm~\ref{algo:attack} needed for the attack
to apply with an inner-product estimator.
We compute the same query vectors
$\boldsymbol{v}^{(t)}_{t\in [r]}$.
At each step $t$, the query response algorithm chooses a correct inner-product estimator $M^{(t)}$
(see Definition~\ref{IPproblem:def}).
The query is issued for the inner
product of
$\boldsymbol{v}^{(t)}$
with the standard basis vector
$\boldsymbol{e}_h$. Note that the value of the inner product is exactly $v^{(t)}_h$ and the requirement of correct reporting of the inner product (Definition~\ref{IPproblem:def}) on these query vectors matches the requirement of a correct heavy hitters reporting of the key $h$ (Definition~\ref{HHproblem:def}).
The input to the estimator are the sketches
$\text{\texttt{Sketch}}_\rho(\boldsymbol{e}_h)$ and
$\text{\texttt{Sketch}}_\rho(\boldsymbol{v}^{(t)})$.
Note that the information available to the estimator from the provided sketches on $v^{(t)}_h$ is the same as with heavy-hitter queries: $\text{\texttt{Sketch}}_\rho(\boldsymbol{e}_h)$ is simply the vector with entries $\mu_h^{(j)}$, which does not add information as $\rho$ and $h$ are assumed to be known to the estimator.
The same analysis therefore applies.
\section{Sketch distribution} \label{sketchdist:sec}
In this section we show that with our particular query inputs \eqref{input:eq}, for large enough $m$, the sketch content that is relevant to determining whether $h$ is a candidate heavy hitter is approximately
$\boldsymbol{u}^{(t)} \sim \mathcal{N}_\ell(v^{(t)}_h,\sigma^2)$ where $\sigma = \sqrt{\frac{m}{b}}$. The random variables
$\boldsymbol{u}^{*(t)} = \boldsymbol{u}^{(t)}- v^{(t)}_h \boldsymbol{1}_\ell$ for $t\in[r]$ are approximately i.i.d.\ from $\mathcal{N}_\ell(0,\sigma^2)$.
$\text{\texttt{CountSketch}}_\rho(\boldsymbol{v}^{(t)})$ includes $\ell\cdot b$
measurements but for inputs known to be of a form \eqref{input:eq} we
can restrict the estimator to only consider the $\ell$ adjusted
measurements \eqref{adjmeasure:eq} of key $h$. We argue below that this
restriction does not limit the power of the estimator.
To simplify our notation going forward, we re-index the set of
relevant measurement vectors and use
$(\boldsymbol{\mu}^{(j)})_{j\in[\ell]}$ for
$(\boldsymbol{\mu}^{(j,h_j(h))})_{j\in[\ell]}$.
We denote by
$\boldsymbol{u}^{(t)}\in\mathbb{R}^\ell$
the random variable vector of the $\ell$ adjusted measurements (see Definition~\ref{adjmeasure:eq})
the sketch provides for $v^{(t)}_h$ and a random tail and by
$\boldsymbol{u}^{*(t)}\in\mathbb{R}^\ell$ the random variable of the respective contributions of the tail component $\boldsymbol{z}^{(t)}$ to these values:
\begin{eqnarray}
u^{*(t)}_j :=& \langle \boldsymbol{z}^{(t)}, \boldsymbol{\mu}^{(j)}(\rho_0) \rangle \cdot \mu^{(j)}_h(\rho_0) \label{ustardef:eq}\\
u^{(t)}_j :=& \langle \boldsymbol{v}^{(t)}, \boldsymbol{\mu}^{(j)}(\rho_0) \rangle \cdot \mu^{(j)}_h(\rho_0) = v^{(t)}_h + u^{*(t)}_j \ . \label{onebucket:eq}
\end{eqnarray}
\begin{lemma}
The distribution of $(\boldsymbol{u}^{*(t)})_{t\in [r]}$
(over $\rho\in \mathcal{D}$ and random tails selection) is such that
$u^{*(t)}_j$ are i.i.d.\
for $t\in [r]$, $j\in [\ell]$) with distribution
\begin{align}
M^{(t)}_j &\sim \mathcal{B}(m,1/b) \nonumber\\
u^{*(t)}_j &\sim 2\cdot \mathcal{B}(M^{(t)}_j,\frac{1}{2})-M^{(t)}_j\ . \label{ustardist:eq}
\end{align}
\end{lemma}
\begin{proof}
Note that the distribution of
$(\boldsymbol{u}^{*(t)})_{t\in [r]}$
does not depend on the choices of
$(v^{(t)}_h)_{t\in [r]}$.
Consider the distribution of
$u^{*(t)}_j$ (for $t\in [r]$, $j\in [\ell]$). The respective
measurement vector is $\boldsymbol{\mu}^{(j)}$.
We have $|\texttt{supp}(\boldsymbol{z})|=m$ and for each $i\in \texttt{supp}(\boldsymbol{z})$, $\mu^{(j)}_i\not=0$ with probability $1/b$. Therefore, the
number of keys in the support that contribute to the
measurement
\[
M^{(t)}_j := \left| \{i \mid z^{(t)}_i \cdot \mu^{(j)}_i \not= 0\}\right|
\]
is a Binomial random variable $M^{(t)}_j \sim \mathcal{B}(m,1/b)$, hence, has expectation $\frac{m}{b}$ and variance $\frac{m}{b}$. The contributions of the $M^{(t)}_j$ keys are i.i.d.\ Rademacher $U\{-1,1\}$ (multiplying by $\mu^{(j)}_h$ which does not change that).
Therefore, the contribution to the measurement, conditioned on $M^{(t)}_j$, has distribution
\begin{equation*}
u^{*(t)}_j \sim 2\cdot \mathcal{B}(M^{(t)}_j,\frac{1}{2})-M^{(t)}_j\ .
\end{equation*}
Note that the random variables
$u^{*(t)}_j$ are symmetric and
$\textsf{E}[u^{*(t)}_j]=0$.
Recall that the values $\mu^{(j)}_i$ for all keys $i\not=h$ and $j\in [\ell]$ are i.i.d.\ ($0$ with probability $(b-1)/b)$ and Rademacher otherwise).
Recall that the random tails are independent of the sketch randomness and the supports $\texttt{supp}(\boldsymbol{z}^{(t)})$ for $t\in [r]$ are disjoint. Even when conditioned on fixing the supports for all $t\in [r]$, all values of keys within the support, across all $t\in [r]$, are i.i.d.\ Rademacher.
We now argue that
$\boldsymbol{u}^{*(t)}$ are i.i.d.\ for different $t\in [r]$. This follows from the disjoint supports of
$\texttt{supp}(\boldsymbol{z}^{(t)})$ for different $t\in [r]$ and therefore
\eqref{ustardef:eq} for different $t$ only depend on disjoint parts of each set of measurement vectors
$(\boldsymbol{\boldsymbol{\mu}^{(j)}})_{j\in [\ell]}$, with only values for keys $i\in \texttt{supp}(\boldsymbol{z}^{(t)})$.
We next observe that for any $t\in [r]$, the values
$u^{*(t)}_j$ are i.i.d.\ for $j\in [\ell]$.
In this case the applicable support $\texttt{supp}(\boldsymbol{z}^{(t)})$
is fixed. But the applicable parts of different measurement vectors $j\in[\ell]$ are independent.
Even when there is overlap in the sets of keys other than $h$ in $\texttt{supp}(\boldsymbol{z}^{(t)})$ that map to different measurements, the measurement vectors randomness is independent, so the contributions of the same key to different measurements it maps to are independent $U\{-1,1\}$.
\end{proof}
We are now ready justify considering only the measurement with $\mu_h\not=0$ in each set $j\in [\ell]$:
The additional measurements have $\boldsymbol{\mu}_h= 0$ and thus do not
depend on $v_h$. They do provide information on
the tail support size parameter $m$ but we can assume $m$ is known to
the estimator. These measurements also provide information on $M^{(t)}_j$, the number of keys in the support that hash to our selected measurement from the each set of $b$ measurements.
This because their sum has distribution
$\mathcal{B}(\mathcal{B}(m-M^{(t)}_j,1/b),1/2)$. But in our analysis we assume that
$m$ is chosen to be large enough so that $M^{(t)}_j$ is very close to its expectation $m/b$ and we may also assume that the exact values $M^{(t)}_j$ are available to the estimator (Essentially, the variations in the value only impacts the "noise" magnitude but not its form).
\begin{lemma}
The distribution of the random variable $\sum_{j\in [\ell]}u^{*(t)}_j$ conditioned on $\rho$ and $\boldsymbol{M}^{(t)}$ is
\[
2 \mathcal{B}\left( \sum_{j\in[\ell]} M^{(t)}_j,1/2\right) - \sum_{j\in[\ell]} M^{(t)}_j\ .
\]
\end{lemma}
\begin{proof}
Taking the sum over $j\in[\ell]$ of the distributions of $u^{*(t)}_j$
\begin{align*}
&\sum_{j\in[\ell]} \left( 2 \mathcal{B}(M^{(t)}_j,1/2)- M^{(t)}_j\right) \\
=& 2 \sum_{j\in[\ell]} \mathcal{B}(M^{(t)}_j,1/2)- \sum_{j\in[\ell]} M^{(t)}_j \\
=& 2\mathcal{B}( \sum_{j\in[\ell]} M^{(t)}_j,1/2) - \sum_{j\in[\ell]} M^{(t)}_j\ ,
\end{align*}
where the last equality follows from properties of a sum of independent Binomial random variables.
\end{proof}
\begin{remark} \label{mchoice:rem}
We choose $m$ to be large enough so that
the following approximations hold:
\begin{itemize}
\item [(i)]
The independent random variables
$M^{(t)}_j \sim \mathcal{B}(m, 1/b)$ are close to their mean $m/b$ for all $t\in[r]$ and $j\in [\ell]$: More precisely, for some $\alpha$ and $\delta'$, with probability $(1-\delta')$, for all $t,j$, $(M^{(t)}_j \in 1\pm \alpha) m/b$. We bound this using Chernoff bounds. We solve
$2\exp(- \frac{m}{b} \alpha^2/3 \leq \delta'/(r \ell )$ to obtain
$m \geq 3 b \alpha^{-2} \ln(2r\ell/\delta')$.
\item [(ii)]
With probability $(1-\delta')$, $r \ell$ random variables
$\mathcal{B}(n,1/2)$ where $n$ is in $(1\pm \alpha) m/b$, are all within
$O(n^{2/3} \alpha)$ from their expected value $n/2$.
When this holds, the normal approximation to the Binomial distribution has relative error within $1\pm \alpha$.
We again apply Chernoff bounds solving for
$2\exp(- (n/2) (n^{-1/3} \alpha)^2/3 \leq \delta'/(r \ell ) $
solving we get
$n^{1/3} \geq 6\alpha^{-2} \ln(2r\ell/\delta')$ and therefore
$m = O(b \alpha^{-6} \ln(2r\ell/\delta')^3)$.
\end{itemize}
\end{remark}
When $m/b$ is large in the sense of Remark~\ref{mchoice:rem}, the distribution of $u^{*(t)}_j$ is approximated well by $\mathcal{N}(0, m/b)$ and
$\boldsymbol{u}^{*(t)}$ is approximated by $\mathcal{N}_\ell(0, m/b)$
and therefore
\begin{eqnarray}
\boldsymbol{u}^{(t)} &\sim& \mathcal{N}_\ell(v^{(t)}_h,\frac{m}{b}) \label{udist:eq}\\
\overline{u} &\sim& \mathcal{N}(v^{(t)}_h, \frac{m}{b} \cdot\frac{1}{\ell})\ .
\end{eqnarray}
The approximation is up to a
relative error of
$\alpha=\alpha'/\ell$ and discretization. We use $\alpha = \alpha'/\ell$ so that a relative error approximation of $1\pm \alpha'$ holds also with respect to a product distribution of $\ell$ such random variables.
For the purpose of cleaner presentation, we will use this approximation going forward. We use
$\sigma := \sqrt{\frac{m}{b}}$.
The analysis carries over by carefully dragging the approximation error term that vanishes for large $m$.
\section{Estimators}
\label{sec:estimators}
In this section we
provide a framework of estimators and
establish properties common to any correct $\ell_2$-$\epsilon$ heavy hitters estimator that is applied to our query vectors.
In its most general form, a query response algorithm fixes before each query $t$ an estimator $M^{(t)}$.
The estimator is applied to the content of the sketch, which on our inputs are i.i.d.\ vectors
$(\boldsymbol{u}^{(t)}\sim \mathcal{N}_\ell(v^{(t)}_h,\sigma^2))_{t\in[r]}$. The estimator is
specified by a {\em reporting function}
$p^{(t)}: \mathbb{R}^\ell \to [0,1]$ so that
$p^{(t)}(\boldsymbol{u}^{(t)}) := \textsf{Pr}[h\in M^{(t)}(\boldsymbol{u}^{(t)})]$
specifies the probability that the returned $K^{(t)}$ includes key $h$ when the sketch content is $\boldsymbol{u}^{(t)}$. We allow the query response algorithm to modify the estimator
arbitrarily between queries and in a way that depends on sketches of prior inputs, prior outputs, and on a maintained state from past queries $(\boldsymbol{u}^{(t')},K^{(t')})_{t'<t}$.
The only constraint that we impose is that (at each step $t$) the output is correct with high probability:
$\ell_2$-$\epsilon$-heavy hitters are reported and $0$ value keys are not reported (see Definition~\ref{HHproblem:def}).
We show that a correct estimator on our query inputs must satisfy the following:
\begin{lemma} [Correct HH estimator basic property] \label{correctest:lemma}
For $\delta_1,\delta_2 \ll 1$,
there are
$a = \Theta(\sqrt{\frac{\ln(1/\delta_2)}{\ell}})$
and $c = \Theta(1)$ so that the following holds.
If the
estimator satisfies (i) if $h$ is a heavy hitter then $\textsf{Pr}[h\in K] \geq 1-\delta_1$ and (ii) if
$v_h=0$ then $\textsf{Pr}[h\not\in K] \geq 1-\delta_2$.
Then
\begin{itemize}
\item $|v_h| \geq c\cdot\sigma$ $\implies$ $\textsf{Pr}[h\in K] \geq 1-\delta_1$
\item $|v_h| \leq a\cdot\sigma$ $\implies$ $\textsf{Pr}[h\not\in K] \geq 1- \frac{1}{\delta_2^{\Omega(1)}}$
\item Otherwise, unrestricted
\end{itemize}
\end{lemma}
\begin{proof}
The tail vectors have $\ell_2^2$ mass $\|\boldsymbol{z} \|^2_2=m$ and
$\|\boldsymbol{v}^{(t)} \|^2_2=m + (v^{(t)}_h)^2$. Key $h$ is $\ell_2$ $\epsilon$-heavy hitter (see Definition~\ref{HH:def}) when
\begin{align*}
v_h^2 \geq \epsilon\cdot \|\boldsymbol{v}\|_2^2 = \epsilon \cdot (v_h^2 + \|\boldsymbol{z}\|_2^2) \implies \\
v_h^2 \geq \frac{\epsilon}{1-\epsilon} \|\boldsymbol{z}\|_2^2 = \frac{\epsilon}{1-\epsilon} m = \frac{b \epsilon}{1-\epsilon}\cdot \sigma \ .
\end{align*}
Recall that the sketch uses $b= O(1/\varepsilon)$, so when $c\approx \frac{b \epsilon}{1-\epsilon} = \Theta(1)$ we get that $|v_h| \geq c\cdot\sigma$ implies that $h$ is $\ell_2$ $\epsilon$-heavy hitter. Therefore an estimator that reports a heavy key with probability at least $1-\delta_1$ must have $\textsf{Pr}[h\in K]\geq 1-\delta_1$.
We now establish the second part. Here we show that if the probability of reporting when $v_h=0$ is at most $\delta_2$ then the probability of reporting
when $|v_h|\leq a\cdot \sigma$ is at most $\delta_2^{1/4}$.
The probability density of a sketch $\boldsymbol{u}$ when $v_h=v$ is:
\[
f_v(\boldsymbol{u}) =
\prod_{i\in[\ell]} \frac{1}{\sigma\sqrt{2\pi}} e^{-\frac{1}{2} \left(\frac{u_i-v}{\sigma}\right)^2}\ ,
\]
We seek to bound the maximum the maximum reporting probability when $v_h \leq a$ subject to reporting probability of $0$ that is at most $\delta_2$. Consider the ratio
\[
\frac{f_v(\boldsymbol{u})}{f_{0}(\boldsymbol{u})} = e^{-\frac{1}{2 \sigma^2}\left( \ell\cdot v^2 - 2\cdot v\cdot \sum_i u_i \right)} = e^{-\frac{1}{2 \sigma^2}\left( \ell\cdot (v^2 - 2 v \overline{u}) \right)}
\]
Note that the ratio depends only on $\overline{u}$ and $v$ and is larger when
$\overline{u}$ is closer to $v$ than to $0$. Therefore, for the goal of bounding the maximum reporting probability with $v$ subject on a bound on the probability for $0$, we place
the reporting probability when $0$ on the largest values of $\overline{u}$. The distribution of $\overline{u}$ with $0$ mean is $\mathcal{N}(0,\sigma^2/\ell)$.
For $\overline{u} \geq \frac{\sigma}{\sqrt{\ell}} \sqrt{\ln(1/\delta_2)}$ provides reporting probability $\Theta(1/\delta_2)$ for $0$. For $v = \frac{1}{2} \frac{\sigma}{\sqrt{\ell}} \sqrt{\ln(1/\delta_2)}$ the reporting probability is $\frac{1}{\delta_2^{\Omega(1)}}$.
\end{proof}
\section{Simplified Mean Estimation}\label{simplifiedminest:sec}
In this section we formulate and state properties of an interaction, described as
Algorithm~\ref{algo:meanestattack},
with a mean estimator of i.i.d\ random variables.
We then show a correspondence between
Algorithm~\ref{algo:attack}
and
Algorithm~\ref{algo:meanestattack}
establish properties of Algorithm~\ref{algo:attack} through this correspondence. In Section~\ref{AMS:sec} we apply this formulation to analyse an attack on norm estimators applied to the AMS sketch.
We use the following definition:
\begin{definition} [correct reporting function] \label{correctmean:def}
A reporting function
$p: \mathbb{R}^\ell \to [0,1]$
is {\em correct} with respect to
parameters: $(\delta,a,c,\ell,\sigma)$ if
\begin{align*}
\forall v\geq c\cdot\sigma,\
\textsf{E}_{\boldsymbol{u} \sim \mathcal{N}_\ell(v,\sigma^2)} [ p^{(t)}(\boldsymbol{u})] \geq 1-\delta\\
\forall v\leq a\cdot\sigma,\
\textsf{E}_{\boldsymbol{u} \sim \mathcal{N}_\ell(v,\sigma^2)} [ p^{(t)}(\boldsymbol{u})] \leq \delta\ .
\end{align*}
\end{definition}
A correct reporting function can be viewed as a simple mean estimator applied to $\ell$ i.i.d.\ samples from $\mathcal{N}(v,\sigma^2)$:
With probability $1-\delta$, the output is $1$ when
$|v|> c\cdot\sigma$ and $-1$ when $|v| < a\cdot\sigma$.
We will establish the following (proof provided in Section~\ref{minestproofs:sec}):
\begin{lemma} [\texttt{MeanEstAttack} Properties]
\label{biasminest:lemma}
Consider Algorithm~\ref{algo:meanestattack} where
$b$ is such that
\[
\sqrt{\frac{\ell}{2\pi}} e^{-\ell b^2/2} \ll \frac{1}{c-a+2b} .
\]
Then the output $\boldsymbol{u}^{(A)} \in \mathbb{R}^\ell$ satisfies
\begin{align}
\textsf{E}_{\mathcal{A}}\left[\overline{u^{(A)}}\right]
&\approx \frac{r}{\ell} \cdot \frac{2 \sigma}{c-a+2b} \label{biasB:eq} \\
\frac{r}{\ell}\sigma^2 \left(1 - \frac{2b}{c-a+2b} \right)^2 &\lessapprox \textsf{Var}_{\mathcal{A}}\left[\overline{u^{(A)}}\right] \lessapprox \frac{r}{\ell}\sigma^2 \label{biasvarB:eq}
\end{align}
\end{lemma}
\begin{algorithm2e
{\small
\caption{\texttt{MeanEstAttack}}
\label{algo:meanestattack}
\DontPrintSemicolon
\KwIn{Parameters $(a,c,\sigma,\delta)$, number of queries $r$, $b\in (0,a]$, A query response algorithm $\mathcal{A}$ that chooses $(a,c,\sigma,\delta)$-correct reporting functions}
\For(\tcp*[h]{Generate queries and reporting functions}){$t\in [r]$}
{
$\mathcal{A}$ chooses a $(a,c,\sigma,\delta)$-correct reporting function $p^{(t)}:\mathbb{R}^\ell\to [0,1]$ \tcp*{Definition~\ref{correctmean:def}. Choice may depend on $(\boldsymbol{u}^{(t')},s^{(t')})_{t'<t}$}
$v^{(t)}\sim \mathcal{U}[a\cdot \sigma, (c+2b)\cdot \sigma]$\;
$\boldsymbol{u}^{*(t)} \sim \mathcal{N}_\ell(0,\sigma^2)$\;
\lFor(\tcp*[h]{Compute $\boldsymbol{u}^{(t)}\in\mathbb{R}^\ell$, note that $\boldsymbol{u}^{(t)} \sim \mathcal{N}_\ell(v^{(t)},\sigma^2)$}){$j\in [\ell]$}{$u^{(t)}_j \gets v^{(t)} + u^{*(t)}_j$}
$s^{(t)} \gets 1$ w.p.\ $p^{(t)}(\boldsymbol{u}^{(t)})$ and $s^{(t)} \gets -1$ otherwise
}
\Return{$\boldsymbol{u}^{(A)} \gets \sum_{t\in[r]} s^{(t)} \boldsymbol{u}^{*(t)} $}
}
\end{algorithm2e}
\subsection{Correspondence between the attack on \texttt{CountSketch} and \texttt{MeanEstAttack}}
We state the attack on $\text{\texttt{CountSketch}}$ in
Algorithm~\ref{algo:attack} in terms of
Algorithm~\ref{algo:meanestattack}.
We first relate the
correctness of an estimator to correctness of its reporting function (proof is immediate):
\begin{lemma} \label{correctp:coro}
If for parameters
$\delta,a,c,\ell,\sigma$ an estimator satisfies (i)
$|v_h| \geq c\cdot\sigma$ $\implies$ $\textsf{Pr}[h\in K] \geq 1-\delta$ and
(ii) $|v_h| \leq a\cdot\sigma$ $\implies$ $\textsf{Pr}[h\not\in K] \geq 1- \delta$, then its reporting function $p$ is correct
with respect to parameters
$(\delta,a,c,\ell,\sigma)$.
\end{lemma}
It follows that an estimator that satisfies the conditions of Lemma~\ref{correctest:lemma} has a reporting function that is correct as in Definition~\ref{correctmean:def} for parameters $(\max\{\delta_1,\delta_2^{\Omega(1)}\},a,c,\ell,\sigma)$.
In the correspondence between
Algorithm~\ref{algo:attack} and
Algorithm~\ref{algo:meanestattack},
the reporting function $M^{(t)}$ corresponds to
$p^{(t)}$, the values $v^{(t)}_h$ to $v^{(t)}$,
the adjusted measurements of key $h$ from the sketch of $\boldsymbol{v}^{(t)}$
correspond to $\boldsymbol{u}^{(t)}$, the adjusted measurements of
key $h$ from the sketch of $\boldsymbol{z}^{(t)}$ corresponds to $\boldsymbol{u}^{*(t)}$, and $s^{(t)}$ plays the same role.
The adjusted measurements of the output of Algorithm~\ref{algo:attack}, adversarial vector
$\boldsymbol{z}^{(A)}$, correspond to
the output $\boldsymbol{u}^{(A)}$ of Algorithm~\ref{algo:meanestattack}.
The parameter values $a,c$ are the same and the parameter $b$ in
Algorithm~\ref{algo:meanestattack}
(unrelated to the sketch parameter) is set to $b\gets a$.
Using this correspondence, we can now provide a proof of
Lemma~\ref{tailprop:lemma}
(Properties of the adversarial tail).
The Lemma follows as a Corollary of Lemma~\ref{biasminest:lemma} using the established relation between Algorithm~\ref{algo:attack} and Algorithm~\ref{algo:meanestattack} and $b=a =\Theta(\sqrt{\ln(1/\delta_2)/\ell}$.
When $\delta_2$ as used to determine $a$ in Lemma~\ref{correctest:lemma} is small enough, we have
$\sigma\cdot \sqrt{\frac{2}{\pi \ell}} e^{- \ell a^2/2} \approx \sigma \cdot \sqrt{\frac{2\cdot \delta_2}{\pi \ell}}$ and the condition of Lemma~\ref{biasminest:lemma} holds.
\begin{proof} [Proof of Lemma~\ref{tailprop:lemma}]
The contributions of the adversarial tail to the adjusted estimates in the sketch are:
\begin{align*}
\langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}\rangle \cdot \mu^{(j)}_h =
\sum_{t\in [r]} s^{(t)} \langle \boldsymbol{z}^{(t)}, \boldsymbol{\mu}^{(j)}\rangle \cdot \mu^{(j)}_h
=
\sum_{t\in [r]} s^{(t)} (u^{(t)}_j-v^{(t)}_h)
=
\sum_{t\in [r]} s^{(t)} u^{*(t)}_j
\ .
\end{align*}
Therefore
\begin{align}
\frac{1}{\ell}\sum_{j\in[\ell]}\langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}\rangle \cdot \mu^{(j)}_h &=\frac{1}{\ell}\sum_{j\in[\ell]}
\sum_{t\in [r]} s^{(t)} u^{*(t)}_j \label{ave:eq}\\
&= \sum_{t\in [r]} s^{(t)} \frac{1}{\ell}\sum_{j\in[\ell]} u^{*(t)}_j \nonumber\\
&= \sum_{t\in [r]} s^{(t)} \overline{u^{*(t)}} = \overline{u^{(A)}} \nonumber\ .
\end{align}
Using Lemma~\ref{biasminest:lemma} and noting that our settings of $a,c$ (see Lemma~\ref{correctest:lemma}) have $c-a \gg a$ we obtain the claim:
\begin{align*}
\textsf{E}_{\rho,\mathcal{A}}\left[\frac{1}{\ell}\sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}\rangle \cdot \mu^{(j)}_h\right] &\approx \frac{r}{\ell} \cdot \frac{2\sigma}{c+a}\\
\textsf{Var}_{\rho,\mathcal{A}}\left[\frac{1}{\ell}\sum_{j\in [\ell]} \langle \boldsymbol{z}^{(A)}, \boldsymbol{\mu}^{(j)}\rangle \cdot \mu^{(j)}_h\right] &\approx \frac{r}{\ell} \cdot \sigma^2
\end{align*}
\end{proof}
\section{Proofs for \texttt{MeanEstAttack} Properties} \label{minestproofs:sec}
We provide the proof details for Lemma~\ref{biasminest:lemma}.
At a high level, we show that for each $t\in [r]$, the value $s^{(t)}$ is correlated with the sign of the normalized bias of the sample:
\begin{equation} \label{normalizedbias:eq}
\alpha^{(t)} := \frac{\overline{u^{*(t)}}}{\sigma} = \frac{\overline{u^{(t)}}-v^{(t)}}{\sigma}\ .
\end{equation}
For $t\in [r]$ with reporting function $p$ we define (superscript $(t)$ omitted for readability from $p$, $\pi$)
\begin{align}
\pi(v) &:= \textsf{Pr}[s^{(t)} = 1 \mid v^{(t)} = v] = \textsf{E}_{\boldsymbol{u}\sim \mathcal{N}_\ell(v,\sigma^2)} [p(\boldsymbol{u})] \label{pidef:eq}\\
\pi(v\mid \alpha) &:= \textsf{Pr}\left[s^{(t)}=1 \mid v^{(t)} = v \textbf{ and } \frac{\overline{u}-v}{\sigma}=\alpha \right]=\textsf{E}_{\boldsymbol{u}\sim \mathcal{N}_\ell(v,\sigma^2) \mid \frac{\overline{u}-v}{\sigma}=\alpha} [p(\boldsymbol{u})] \label{picond:eq}
\end{align}
The probability of $s^{(t)}=1$
when $v^{(t)} = v$ and
the probability of $s^{(t)}=1$ when $v^{(t)}=v$ {\em and } $\alpha^{(t)} = \alpha$.
Note that since $\overline{u} \sim \mathcal{N}(v,\sigma^2/\ell)$ we have $\alpha^{(t)} \sim \mathcal{N}(0,1/\ell)$.
We will establish the following:
\begin{lemma} [Gap Lemma] \label{gap:coro}
For $0< \alpha \leq b\leq a$ (with probability $1-O(\delta)$ over the distribution of $\alpha$) and for any correct reporting function $p$ (and respective $\pi$)
\[
\textsf{E}_{v\sim \mathcal{U}[(a ,(c+2b)\cdot\sigma ]} [\pi(v \mid \alpha) - \pi(v \mid -\alpha) ] \approx \frac{2\alpha}{c-a+2b} .
\]
\end{lemma}
This Lemma implies that when
$v^{(t)} \sim \sigma\cdot [a,c+2b]$ then any correct estimator would exhibit a gap in the probability of reporting $s^{(t)}=1$, depending on the sign of the normalized bias $\alpha^{(t)}$. The gap increases with the magnitude $|\alpha^{(t)}|$. Note that the statement is on the expected {\em difference} between the reporting probabilities for positive or negative bias. The expected probabilities
$\textsf{E}_{v\sim \mathcal{U}[(a ,(c+2b)\cdot\sigma ]} [\pi(v \mid \alpha)$ and
$\textsf{E}_{v\sim \mathcal{U}[(a ,(c+2b)\cdot\sigma ]} [\pi(v \mid -\alpha)$
can vary wildly, be both close to $0$ or close to $1$, for different correct reporting functions.
\subsection{Proof of Lemma~\ref{gap:coro}}
We will use two Lemmas. The first uses the fact that the reporting probability as a function of
$v^{(t)}$ has total increase of almost $1$ between $v=a\cdot \sigma$ to $v = c\cdot \sigma$ to establish its average increases over uniformly random sub-intervals of length $2\alpha\sigma$:
\begin{lemma} [Average increase property] \label{correctestavg:lemma}
For $0< \alpha \leq b \leq a$ (with probability $1-O(\delta)$ over the distribution)
\[
\textsf{E}_{v\sim \mathcal{U}[(a ,(c+2b)\cdot\sigma ]} [\pi(v \mid \alpha) - \pi(v-2\cdot \alpha\cdot\sigma \mid \alpha) ] \approx \frac{2\alpha}{c-a+2b} .
\]
\end{lemma}
\begin{proof}
From Corollary~\ref{correctp:coro}, for almost all values of $\alpha$ (probability $1-O(\delta)$ over the distribution of $\alpha\sim \mathcal{N}(0,1/\ell)$) we must have that
the function $\pi( v \mid \alpha)$ is very close (within a small constant of our choice) to $0$ when
$|v|\leq a\cdot \sigma$ and very close to $1$ when $|v| \geq c\cdot \sigma$.
\begin{align*}
& \textsf{E}_{v\sim \mathcal{U}[a\cdot\sigma ,(c+2b)\cdot\sigma ]} [\pi(v \mid \alpha) - \pi(v -2\alpha\cdot\sigma \mid \alpha) ]\\
&= \frac{1}{c-a+2b} \int_{a\cdot\sigma}^{(c+2b)\cdot\sigma } \pi(v \mid \alpha) - \pi(v -2\alpha\cdot\sigma \mid \alpha) dv \\
&= \frac{1}{c-a+2b} \left(\int_{a\cdot\sigma}^{(c+2b )\cdot\sigma } \pi(v \mid \alpha) dv - \int_{(a-2\alpha)\cdot\sigma}^{(c+2b-2\alpha)\cdot\sigma } \pi(v \mid \alpha) dv \right)\\
&= \frac{1}{c-a+2b} \left(- \int_{(a-2\alpha)\cdot\sigma}^{a \cdot\sigma} \pi(v \mid \alpha) dv + \int_{(c+2b-2\alpha)\cdot\sigma}^{(c+2b)\cdot\sigma} \pi(v \mid \alpha) dv \right) \approx \frac{2\alpha}{c-a+2b} \end{align*}
\end{proof}
\ignore{
\begin{lemma} [Average increase property] \label{correctestavg:lemma}
For any $0<\gamma, \alpha \leq a$ (with v.h.p over the distribution of $\alpha$)
\[
\textsf{E}_{v\sim \mathcal{U}[(0 ,(c+a)\cdot\sigma ]} [\pi(v+\gamma\cdot \sigma \mid \alpha) - \pi(v-\gamma\cdot \sigma \mid \alpha) ] = \Theta(\gamma/c) .
\]
(in particular, this holds when not conditioning on $\alpha$).
\end{lemma}
\begin{proof}
From Lemma~\ref{correctest:lemma},
the function $\pi( v \mid \alpha)$ is very close to $0$ when
$|v|\in [0,a\cdot \sigma]$ and very close to $1$ when $v\geq c\cdot \sigma$.
Its mean growth over an sub-interval of length $2\gamma$ is therefore $\propto \frac{2\gamma}{c+a}$:
\begin{align*}
& \textsf{E}_{v\sim \mathcal{U}[0,(c+a)\cdot\sigma ]} [\pi(v+\gamma\cdot \sigma \mid \alpha) - \pi(v-\gamma\cdot \sigma \mid \alpha) ]\\
&= \frac{1}{c+a} \int_{0}^{(c+a)\cdot\sigma } \pi(v +\gamma\cdot\sigma \mid \alpha) - \pi(v -\gamma\cdot\sigma \mid \alpha) dv \\
&= \frac{1}{c+a} \left(\int_{(\gamma)\cdot\sigma}^{(c+a +\gamma)\cdot\sigma } \pi(v \mid \alpha) dv - \int_{(-\gamma)\cdot\sigma}^{(c+a-\gamma)\cdot\sigma } \pi(v \mid \alpha) dv \right)\\
&= \frac{1}{c+a} \left(- \int_{(-\gamma)\cdot\sigma}^{\gamma \cdot\sigma} \pi(v \mid \alpha) dv + \int_{(c+a-\gamma)\cdot\sigma}^{(c+a+\gamma)\cdot\sigma} \pi(v \mid \alpha) dv \right) \approx \frac{2\gamma}{c+a}
\end{align*}
\end{proof}
}
The following Lemma shows that the probability of
reporting $s^{(t)}=1$, is equally likely when
$v^{(t)} = v + \alpha\cdot\sigma$ and $\overline{u} = v$ and when $v^{(t)} = v - \alpha\cdot\sigma$ and $\overline{u} = v$:
\begin{lemma}\label{sym:lemma}
For any $\alpha>0$, reporting function $p()$, and $v$,
$\pi(v +\alpha\cdot\sigma \mid -\alpha) = \pi(v -\alpha\cdot\sigma \mid +\alpha)$.
\end{lemma}
\begin{proof}
From \eqref{picond:eq}:
\begin{align*}
\pi(v +\alpha\cdot\sigma \mid -\alpha) &= \frac{1}{\int_{\boldsymbol{u} \mid \overline{u}=v} f_{v +\alpha\cdot\sigma}(\boldsymbol{u}) d \boldsymbol{u} } \cdot
\int_{\boldsymbol{u} \mid \overline{u}=v} f_{v +\alpha\cdot\sigma}(\boldsymbol{u}) \cdot p(\boldsymbol{u}) d \boldsymbol{u}\\
\pi(v -\alpha\cdot\sigma \mid \alpha) &=
\frac{1}{\int_{\boldsymbol{u} \mid \overline{u}=v} f_{v +\alpha\cdot\sigma}(\boldsymbol{u}) d \boldsymbol{u} } \cdot
\int_{\boldsymbol{u} \mid \overline{u}=v} f_{v -\alpha\cdot\sigma}(\boldsymbol{u}) \cdot p(\boldsymbol{u}) d \boldsymbol{u}
\end{align*}
The probability $\pi(v +\alpha\cdot\sigma \mid -\alpha)$ is the mean over $\boldsymbol{u}$ with $\overline{u}=v$ of $f_{v +\alpha\cdot\sigma}(\boldsymbol{u}) \cdot p(\boldsymbol{u})$. Similarly, $\pi(v -\alpha\cdot\sigma \mid \alpha)$ is the mean over $\boldsymbol{u}$ with $\overline{u}=v$ of $f_{v -\alpha\cdot\sigma}(\boldsymbol{u}) \cdot p(\boldsymbol{u})$.
It therefore suffices to establish that
$f_{v -\alpha\cdot\sigma}(\boldsymbol{u})= f_{v +\alpha\cdot\sigma}(\boldsymbol{u})$ for $\boldsymbol{u}$ when $\overline{u}=v$ .
This is a
property of a product of normal distributions: for any $\boldsymbol{u} \in \mathbb{R}^\ell$ and $\alpha$, the outcome
$\boldsymbol{u}$
is equally likely with $\mathcal{N}_\ell(\overline{u}+\alpha\cdot\sigma,\sigma^2 )$ and $\mathcal{N}_\ell(\overline{u}-\alpha\cdot\sigma,\sigma^2)$:
\[
f_{\overline{u}+\alpha\cdot\sigma}(\boldsymbol{u}) =f_{\overline{u}-\alpha\cdot\sigma}(\boldsymbol{u}).
\]
We express $\boldsymbol{u}$ in terms of the empirical mean
$\overline{u} := \frac{1}{\ell} \sum_j u_j$
and
the "deviations" from the empirical mean
$(\Delta_j = u_j - \overline{u})_{j=1}^{\ell}$.
Note that by definition $\sum_{j\in [\ell]} \Delta_j = 0$.
Let $\alpha >0$,
using~\eqref{pdfnormal:eq} and $\sum_i \Delta_i=0$ we obtain:
\begin{align*}
f_{\overline{u}+\alpha\cdot\sigma}(\boldsymbol{u})
=& \prod_{i\in[\ell]} \frac{1}{\sigma\sqrt{2\pi}} \exp{-\left(\frac{\overline{u}+\Delta_i -(\overline{u}+\alpha\cdot\sigma)}{\sigma}\right)^2}\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(\frac{-\sum_{i=1}^\ell (\Delta_i^2 + \alpha^2\cdot\sigma^2 -2 \alpha\cdot\sigma\Delta_i)}{\sigma^2} \right)\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(\frac{-\ell \alpha^2\cdot\sigma^2 - \sum_{i=1}^\ell \Delta_i^2 +2 \alpha\cdot\sigma\cdot \sum_{i=1}^\ell \Delta_i)}{\sigma^2} \right)\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(\frac{-\ell \alpha^2\cdot\sigma^2 - \sum_{i=1}^\ell \Delta_i^2}{\sigma^2} \right)=\frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(-\ell \alpha^2 - \frac{\sum_{i=1}^\ell \Delta_i^2}{\sigma^2} \right)
\end{align*}
The calculation for $f_{\overline{u}-\alpha\cdot\sigma}(\boldsymbol{u})$ is similar and yields the same expression:
\begin{align*}
f_{\overline{u}-\alpha\cdot\sigma}(\boldsymbol{u})
=& \prod_{i\in[\ell]} \frac{1}{\sigma\sqrt{2\pi}} \exp{-\left(\frac{\overline{u}+\Delta_i -(\overline{u}-\alpha\cdot\sigma)}{\sigma}\right)^2}\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(\frac{-\sum_{i=1}^\ell (\Delta_i^2 + \alpha^2\cdot\sigma^2 +2 \alpha\cdot\sigma\Delta_i)}{\sigma^2} \right)\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(\frac{-\ell \alpha^2\cdot\sigma^2 - \sum_{i=1}^\ell \Delta_i^2 -2 \alpha\cdot\sigma\cdot \sum_{i=1}^\ell \Delta_i)}{\sigma^2} \right)\\
=& \frac{1}{(\sigma\sqrt{2\pi})^\ell}\exp\left(-\ell \alpha^2 - \frac{\sum_{i=1}^\ell \Delta_i^2}{\sigma^2} \right)
\end{align*}
\end{proof}
Lemma~\ref{gap:coro} follows as a corollary of
Lemma~\ref{correctestavg:lemma} and Lemma~\ref{sym:lemma}: From Lemma~\ref{correctestavg:lemma} we get that
$\pi(v-2\cdot \alpha\cdot\sigma \mid \alpha) = \pi(v \mid -\alpha)$. By making this substitution in the statement of Lemma~\ref{sym:lemma} we obtain Lemma~\ref{gap:coro}.
\subsection{Output properties}
In this section we provide a proof of Lemma~\ref{biasminest:lemma} that states properties of the expectation and variance of the average entry value of the output vector. Additionally, we bound from below the $\ell_2$ norm (Lemma~\ref{squaredbound:lemma}). We use the bound on the $\ell_2$ norm in Section~\ref{AMS:sec}.
\begin{proof} [Proof of Lemma~\ref{biasminest:lemma}]
From linearity of expectation,
\begin{equation} \label{claimE:eq}
\textsf{E}_{\mathcal{A}}\left[\overline{u^{(A)}}\right] =
\textsf{E}\left[\sum_{t\in [r]} s^{(t)}(\overline{u}^{(t)}-v^{(t)}))\right]=
\sigma \textsf{E}\left[\sum_{t\in[r]} s^{(t)} \alpha^{(t)}\right] = \sigma \sum_{t\in [r]} \textsf{E}[s^{(t)} \alpha^{(t)}] ,
\end{equation}
where $\alpha^{(t)}$ is the normalized bias \eqref{normalizedbias:eq}.
We now consider the contribution
of a single query $t \in [r]$:
\begin{equation} \label{onet:eq}
\textsf{E}[s^{(t)}\alpha^{(t)}]\ .
\end{equation}
Recall that the distribution of
$\alpha^{(t)}$
is
$\mathcal{N}(0,1/\ell)$.
We condition on
$|\alpha^{(t)}| = \alpha$ for $\alpha \geq 0$ which has the density function
\begin{equation} \label{evennormal:eq}
2\sqrt{\frac{\ell}{2\pi}}e^{-\frac{1}{2}\ell \alpha^2} .
\end{equation}
We have $s^{(t)}\alpha^{(t)}= \alpha$ if
$\alpha^{(t)} = \alpha$
and $s^{(t)}=1$ or when
$\alpha^{(t)} = \alpha = -\alpha$
and $s^{(t)}=-1$. Similarly, $s^{(t)}\alpha^{(t)} = -\alpha$ if
$\alpha^{(t)} = \alpha$
and $s^{(t)}=-1$ or
when
$\alpha^{(t)} = -\alpha$
and $s^{(t)}=1$.
We now express the expected contribution under this conditioning, for fixed $v^{(t)}$ and fixed correct reporting function $p$. We use that since the distribution of $\alpha^{(t)}$ (for fixed $v^{(t)}$) is symmetrical, there is equal probability $1/2$ for each of
$\alpha^{(t)} = \pm \alpha$. We use
$\pi(v\mid\alpha)$ as defined in
\eqref{picond:eq}.
\begin{align*}
\lefteqn{\textsf{E}_{p}[s^{(t)}\alpha^{(t)} \mid \alpha^{(t)} =\alpha ] =\textsf{E}_{\mathcal{A}}[s^{(t)}(\frac{\overline{u}^{(t)}-v^{(t)}}{\sigma}) \mid \frac{|\overline{u}^{(t)}-v^{(t)}|}{\sigma}=\alpha ] =}\\
&= \frac{1}{2} \alpha \left(\pi(v^{(t)} \mid \alpha)-\pi(v^{(t)} \mid -\alpha) -(1-\pi(v^{(t)} \mid \alpha)) + (1-\pi(v^{(t)}\mid -\alpha))\right) \\
&= \frac{1}{2} \alpha \left(2\pi(v^{(t)}\mid\alpha) - 2\pi(v^{(t)} \mid -\alpha)\right) = \alpha \left(\pi(v^{(t)}\mid\alpha) - \pi(v^{(t)} \mid -\alpha)\right)
\end{align*}
We take the expectation over $v^{(t)} \sim \mathcal{U}[a\cdot\sigma,(c+2b)\cdot\sigma]$ and apply Lemma~\ref{gap:coro}. We obtain that for $\alpha\in [0,b]$:
\begin{align*}
\lefteqn{\textsf{E}_{v^{(t)},p}[s^{(t)} \alpha^{(t)} \mid \alpha^{(t)} =\alpha ] =}\\
&= \alpha \textsf{E}_{v\sim \mathcal{U}[a\cdot\sigma,(c+2b)}[\pi(v\mid\alpha) - \pi(v \mid -\alpha)] \approx \alpha \frac{2\alpha}{c-a+2b} = \frac{2\alpha^2}{c-a+2b}
\end{align*}
Finally, we now take the expectation over $\alpha \geq 0$ using the density
\eqref{evennormal:eq}:
\begin{align}
\textsf{E}_{v^{(t)},p}[s^{(t)} \alpha^{(t)}] &=
\int_0^\infty (2\cdot\sqrt{\frac{\ell}{2\pi}}e^{-\frac{1}{2}\ell \alpha^2})
\textsf{E}_{v,p}[s^{(t)}\alpha^{(t)} \mid |\alpha^{(t)}|=\alpha ]
d\alpha \nonumber \\
&\approx \frac{2}{c-a+2b} 2\cdot\sqrt{\frac{\ell}{2\pi}} \int_0^\infty \alpha^2 e^{-\frac{1}{2}\ell \alpha^2} d\alpha \nonumber \\
&= \frac{4}{c-a+2b} \cdot\sqrt{\frac{\ell}{2\pi}} \ell^{-3/2} \sqrt{\frac{\pi}{2}} = \frac{1}{\ell} \cdot \frac{2}{c-a+2b} \label{pertE:eq}
\end{align}
Noting that the approximation holds as the
contribution of $\alpha\in [b,\infty)$ is negligible when the condition on $b$ holds:
\begin{align*}
& \int_b^\infty (2\cdot\sqrt{\frac{\ell}{2\pi}}e^{-\frac{1}{2}\ell \alpha^2}) \left| \textsf{E}_{v^{(t)},p}[s^{(t)} \alpha^{(t)} \mid |\alpha^{(t)}| =\alpha ] \right | d\alpha \\
&\leq \int_b^\infty (2\cdot\sqrt{\frac{\ell}{2\pi}}e^{-\frac{1}{2}\ell \alpha^2}) \alpha d\alpha \leq \sqrt{\frac{2}{\pi \ell}} e^{- \ell b^2/2}
\end{align*}
Substituting \eqref{pertE:eq} in \eqref{claimE:eq} we establish claim \eqref{biasB:eq} of the Lemma:
$\textsf{E}_{\mathcal{A}}[\overline{u^{(A)}}] \approx \frac{r}{\ell} \sigma \cdot \frac{2}{c-a+2b}$.
We now bound (using independence of $s^{(t)}\alpha^{(t)}$ for different $t$ and thus linearity of the variance)
\begin{equation} \label{claimV:eq}
\textsf{Var}_{\mathcal{A}}\left[\overline{u^{(A)}}\right] =
\textsf{Var}\left[\sum_{t\in [r]} s^{(t)}(\overline{u}^{(t)}-v^{(t)}))\right]=
\sigma^2 \textsf{Var}\left[\sum_{t\in[r]} s^{(t)} \alpha^{(t)}\right] = \sigma^2 \sum_{t\in [r]} \textsf{Var}[s^{(t)} \alpha^{(t)}] .
\end{equation}
We bound
\begin{equation} \label{vart:eq}
\textsf{Var}[s^{(t)}\alpha^{(t)}] .
\end{equation}
We condition on
$|\alpha^{(t)}| = \alpha$ and take expectation over $v$, applying Lemma~\ref{gap:coro}:
We have $s^{(t)}\alpha^{(t)} = \alpha$ with probability
\[
\frac{1}{2}\textsf{E}_v[\pi(v \mid \alpha) - (1-\pi(v \mid -\alpha))]= \frac{1}{2}+ \frac{\alpha}{c-a+2b}
\]
and (similarly)
$s^{(t)}\alpha^{(t)} = - \alpha$ with probability $\frac{1}{2}- \frac{\alpha}{c-a+2b}$.
We obtain
\begin{align}
\lefteqn{\textsf{Var}[s^{(t)}\alpha^{(t)} \mid |\alpha^{(t)} | = \alpha] =}\nonumber\\
&= \textsf{E}[(s^{(t)}\alpha^{(t)})^2\mid |\alpha^{(t)} | = \alpha ] -
\textsf{E}[s^{(t)}\alpha^{(t)}\mid |\alpha^{(t)} | = \alpha ]^2 = \alpha^2 \left(1 - \left(\frac{2\alpha}{c-a+2b}\right)^2 \right)\ .\label{varalpha:eq}
\end{align}
Since $\alpha \in [0,b]$
we have
\[
1 - \frac{2b}{c-a+2b} \leq 1 - \frac{2\alpha}{c-a+2b} \leq 1\ .
\]
Take the expectation of \eqref{varalpha:eq} over $\alpha \geq 0$ using these bounds and density
\eqref{evennormal:eq} we obtain
\begin{align}
\frac{1}{\ell} \left(1 - \frac{2b}{c-a+2b} \right)^2 \leq \textsf{Var}_{v^{(t)},p}[s^{(t)}\alpha^{(t)}] \leq \frac{1}{\ell}\ .\label{bound:eq}
\end{align}
Claim \eqref{biasvarB:eq} in the statement of the Lemma follows by substituting \eqref{bound:eq} in \eqref{claimV:eq}.
In the regime where
$c-a \gg b$, $\frac{2b}{c-a+2b}$ is small and we have
\begin{align*}
\textsf{Var}_{v^{(t)},p}[s^{(t)}\alpha^{(t)}] \approx \frac{1}{\ell}\ .
\end{align*}
\end{proof}
We will use the following in our analysis of the attack on the AMS sketch:
\begin{lemma} \label{squaredbound:lemma}
Under the conditions of Lemma~\ref{biasminest:lemma}, the following holds:
\begin{align}
\textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell}\|\boldsymbol{u}^{(A)}\|^2_2 \right]
&\geq
r\sigma^2\left(1-O(\frac{1}{\sqrt{\ell}}) + \Omega\left(\frac{r}{\ell^2} \left(\frac{2}{c-a+2b}\right)^2 \right) \right) \label{biasSqB:eq}
\end{align}
\end{lemma}
\begin{proof}
We now establish Claim~\eqref{biasSqB:eq}
\begin{align}
\textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell}\|\boldsymbol{u}^{(A)}\|^2_2 \right] &= \frac{1}{\ell} \sum_{j\in [\ell]} \textsf{E}_{\mathcal{A}}\left[(u^{(A)}_j)^2\right] \nonumber\\
&= \frac{1}{\ell} \sum_{j\in [\ell]} \left( \textsf{Var}_{\mathcal{A}} \left[ u^{(A)}_j \right] + \textsf{E}_{\mathcal{A}}\left[ u^{(A)}_j \right]^2 \right)\nonumber\\
&= \frac{1}{\ell} \sum_{j\in [\ell]} \textsf{Var}_{\mathcal{A}} \left[ u^{(A)}_j \right] + \frac{1}{\ell} \sum_{j\in [\ell]} \textsf{E}_{\mathcal{A}}\left[ u^{(A)}_j \right]^2 \label{sqsubbound:eq} \ .
\end{align}
We apply Claim~\eqref{biasB:eq} to bound the second term:
\begin{align}
\frac{1}{\ell} \sum_{j\in [\ell]} \textsf{E}_{\mathcal{A}}\left[ u^{(A)}_j \right]^2 &\geq \textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell} \sum_{j\in [\ell]} u^{(A)}_j \right] & \, \text{applying the inequality $E[X^2]\geq E[X]^2$}\nonumber \\
&= \textsf{E}_{\mathcal{A}}[\overline{u^{(A)}}]^2 \approx \left(\frac{r}{\ell} \sigma^2 \cdot \frac{2}{c-a+2b}\right)^2 & \, \text{Claim~\eqref{biasB:eq}} \label{Esqbound:eq}
\end{align}
We now bound from below
\begin{align}
\sum_{j\in [\ell]} \textsf{Var}_{\mathcal{A}} \left[ u^{(A)}_j \right] &=
\sum_{j\in [\ell]} \textsf{Var}_{\mathcal{A}}\left[ \sum_{t\in[r]} s^{(t)} u^{*(t)}_j \right] \geq r\cdot\ell\cdot\sigma^2 \left(1 -O(\frac{1}{\sqrt{\ell}}) \right)\ . \label{Vsqbound:eq}
\end{align}
We recall that $u^{(t)}_j$ are i.i.d.\ $\mathcal{N}(0,\sigma)$.
We consider a bound where $s^{(t)}\in\{\pm 1\}$ is selected in a worst case way (as to minimize variance) according to the actual $\boldsymbol{u}^{*(t)}\sim \mathcal{N}_{\ell}(0,\sigma)$.
With $\ell=1$ the variance is minimized when the correlation are maximized which is done by having $s^{(t)}u^{*{t)}}$ all with the same sign, that is, behaving like i.i.d.\ $|\mathcal{N}(0,\sigma)|$ that have variance $(1- 2/\pi)\sigma^2$ per term with respective variance
$r(1-2/\pi)\sigma^2$ for the sum. Since the same $s^{(t)}$ multiplies all entries in $\boldsymbol{u}^{*(t)}$, we can in expectation have at most $1/2+ \sqrt{\frac{2}{\pi \ell}}$ of these values with the same sign. Over the total of $r\cdot \ell$ terms we get
expected contribution of at least
$ 2 \sqrt{\frac{2}{\pi \ell}}\cdot (1- 2/\pi)\sigma^2 + (1- 2 \sqrt{\frac{2}{\pi \ell}})\cdot \sigma^2$.
Finally, we substitute \eqref{Esqbound:eq} and \eqref{Vsqbound:eq} in \eqref{sqsubbound:eq} and obtain the claim
\begin{align*}
\frac{1}{\ell} \sum_{j\in [\ell]} \textsf{E}_{\mathcal{A}}\left[ u^{(A)}_j \right]^2 &\geq r\cdot\sigma^2 \left(1 -O(\frac{1}{\sqrt{\ell}}) \right) + \left(\frac{r}{\ell} \sigma^2 \cdot \frac{2}{c-a+2b}\right)^2 \\
&= r\sigma^2 \left(1 - O(\frac{1}{\sqrt{\ell}}) + \frac{r}{\ell^2} (\frac{2}{c-a+2b})^2 \right)
\end{align*}
\end{proof}
\section{Extension: Universal Attack on the AMS Sketch} \label{AMS:sec}
\begin{algorithm2e
{\small
\caption{\texttt{Universal Attack on AMS norm estimation}}
\label{algo:AMSattack}
\DontPrintSemicolon
\KwIn{$\tau$, $\epsilon$, $\delta$ \tcp*{Estimator parameters (Definition~\ref{correctnormest:def})}
Initialized AMS $\text{\texttt{Sketch}}_\rho$ with parameters $(n,\ell)$, number of queries $r$, tail support size $m$}
$\sigma \gets \tau \sqrt{1/2}$; $c \gets \sqrt{1+2\epsilon} $; $a \gets 1$; $b\gets \Theta(\frac{1}{\sqrt{\ell}} \ln(\sqrt\frac{\ell}{\epsilon})$\;
$(h,(\boldsymbol{z}^{(t)})_{t\in [r]}) \gets$
\texttt{AttackTails}$(n,m,r)$\tcp*{Choose tails (Algorithm~\ref{algo:tails})}
$(\boldsymbol{z}^{(t)} \gets \boldsymbol{z}^{(t)} \cdot \frac{\sigma}{\sqrt{m}})_{t\in [r]} $\tcp*{Rescale tails to have $\ell_2$ norm $\sigma$}
\For(\tcp*[h]{Generate Query Vectors}){$t\in [r]$}{
$v^{(t)}_h\sim \mathcal{U}[a\cdot \sigma, (c+2b)\cdot \sigma]$\;
$\boldsymbol{v}^{(t)} \gets v^{(t)}_h \boldsymbol{e}_h +
\boldsymbol{z}^{(t)}$\tcp*{Query vectors}
}
\For(\tcp*[h]{Apply Query Response}){$t\in [r]$}
{
Choose an $(\epsilon,\delta,\tau)$-correct estimator $M^{(t)}$ \tcp*{Definition~\ref{correctnormest:def}, may depend on $(v^{(t')}_h,s^{(t')},M^{(t')})_{t'<t}$ and $\rho$ }
$s^{(t)} \gets M^{(t)}(\text{\texttt{Sketch}}_\rho(\boldsymbol{v}^{(t)})$\tcp*{Apply estimator to sketch}
}
\Return{$\boldsymbol{z}^{(A)} \gets \sum_{t\in[r]} s^{(t)} \boldsymbol{z}^{(t)} $}\tcp*{Adversarial input}
}
\end{algorithm2e}
The AMS sketch~\cite{ams99} and the related Johnson Lindenstrauss transform~\cite{JLlemma:1984} are linear maps of input vectors $\boldsymbol{v} \in \mathbb{R}^n$ to their
sketches in $\mathbb{R}^\ell$. The sketches support recovery of $\ell_2$ norms and distances.
In this section we describe an attack on the AMS sketch that applies with any correct norm estimator.
The sketch is specified by parameters $(n,\ell)$,
where $n$ is the dimension of input vectors and $\ell$ is the number of measurements.
The internal randomness $\rho$ specifies a set of random hash functions
$s_{j}:[n]\rightarrow \{-1,1\}$ ($ j\in [\ell]$) so that $\textsf{Pr}[s_{j}(i)=1]=1/2$.
These hash functions define $\ell$ measurement vectors
$\boldsymbol{\mu}^{(j)}$ ($j\in [\ell]$):
\[ \mu^{(j)}_i := s_{j}(i) .\]
The AMS sketch analysis holds with 4-wise independent functions but our attack analysis holds with full randomness.
Norm estimators $M$ applied to the sketch guarantee that
when $\ell = O(\epsilon^{-2} \log(1/\delta))$
\[
\textsf{Pr}_\rho[| \|\boldsymbol{v}\|_2^2 - M(\text{\texttt{Sketch}}_\rho(\boldsymbol{v}))^2 | \geq \epsilon \|\boldsymbol{v}\|_2^2] < \delta\ .
\]
In particular, for any
$\boldsymbol{v}\in R^n$ and $j\in[\ell]$ it holds that,
$\textsf{E}_{\rho}[\langle \boldsymbol{\mu},\boldsymbol{v}\rangle ^2] = \|\boldsymbol{v}\|^2_2$ and
$\textsf{Var}_{\rho}[\langle \boldsymbol{\mu},\boldsymbol{v}\rangle ^2] = O(\|\boldsymbol{v}\|^2_2)$.
\begin{definition} [Adversarial input for AMS] \label{adversarialinput:def}
We say that an attack $\mathcal{A}$ that is applied to a sketch with randomness $\rho$ and outputs
$\boldsymbol{z}^{(A)}\in \{-1,0,1\}^n$ is
$(\xi,\beta)$-{\em adversarial} (for $\xi>0$) if with probability at least $1-\beta$ over the randomness of $\rho,\mathcal{A}$ it holds that:
\begin{equation}
\textsf{Pr}_{\rho,\mathcal{A}}\left[\frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{\mu}^{(j)}, \boldsymbol{z}^{(A)} \rangle^2
\geq (1+\xi) \cdot \| \boldsymbol{z}^{(A)}\|^2_2\right] \geq 1-\beta\ .
\end{equation}
\end{definition}
Note that when $\ell \gg \frac{1}{ \epsilon^2}\ln(1/\delta)$, for any $\boldsymbol{v}$, we have
\begin{equation}
\textsf{Pr}_{\rho\sim\mathcal{D}} \left[ \frac{1}{\ell} \sum_{j\in [\ell]} \langle \boldsymbol{\mu}^{(j)}(\rho), \boldsymbol{v} \rangle^2
\geq (1+\epsilon) \cdot \| \boldsymbol{v}\|^2_2 \right] \leq \delta\ .
\end{equation}
Algorithm~\ref{algo:AMSattack} describes an
attack that constructs an adversarial input for a sketch with randomness $\rho$. Here we consider both the accuracy
$\epsilon$ (relative error) of the estimator, which is at least $\epsilon \geq 1/\sqrt{\ell}$ (otherwise correct estimators do not exist) and the bias $\xi$ in the product of the attack. The bias also must have $\xi = \Omega(1/\sqrt{\ell})$ (otherwise $\xi$ is within the variability of the output) and naturally we may want $\xi \geq \epsilon$, that is, bias is larger than the accuracy of the estimators we attack.
We establish the following (the proof is presented after a brief discussion that interprets the result):
\begin{lemma}. \label{adverseAMS:lemma}
For $\epsilon, \xi = \Omega(1/\sqrt{\ell})$, any constant $\beta>0$,
and attack size $r = O(\xi \ell^2 \epsilon^2)$,
the output of
Algorithm~\ref{algo:AMSattack} is $(\xi,\beta)$-adversarial.
\end{lemma}
Intuitively, more accurate estimators (that is, smaller $\epsilon$) leak more information on the randomness and hence are easier to attack and we expect attack size to increase with $\epsilon$. A larger bias $\xi$ is harder to accrue and would require a larger attack and hence attack size also increases with $\xi$.
Interestingly, for the AMS sketch, the sketch size parameter $\ell$ controls both
robustness and accuracy, where with less accurate responses we can support more queries. This is in contrast to heavy hitters via $\text{\texttt{CountSketch}}$, where the parameter $b$ controls the accuracy and we use $\ell$ to control confidence and robustness. (In principle, a larger $\ell$ in $\text{\texttt{CountSketch}}$ can also be used to increase accuracy but much less effectively for a given sketch size than increasing $b$). We therefore study the attack size on the AMS sketch in terms of both $\epsilon$ and $\xi$.
The Lemma implies that to generate an $O(1)$-adversarial input for a sketch of size $\ell$ when estimators have accuracy $\epsilon = \Theta(1)$ (what we can get in the non-adaptive setting from a sketch of size $\ell=O(\log(n/\delta))$) it suffices to use $O(\ell^2)$ queries. In this case there is a matching upper bound (up to logarithmic factors) using the "wrapper" method of~\cite{HassidimKMMS20}.
The required attack size when $\xi = O(\epsilon)$ is $r= O(\ell^2 \epsilon^3)$.
For the smallest applicable accuracy $\epsilon = O(1/\sqrt{\ell})$ we obtain that $r= O(\sqrt{\ell})$ queries suffice. In this case, there is a gap with the best known upper bound which is the trivial bound $r=\Omega(1)$ of answering a single query with the sketch.
This gap leaves an interesting open question.
Our attack has a simple structure where the only adaptive input is the final adversarial one. We conjecture that there is more robustness in this case (thus a better upper bound) whereas a more complex attack with multiple adaptive rounds may yield a tight lower bound of $r=\tilde{O}(1)$ for the general case.
\subsection{Proof of Lemma~\ref{adverseAMS:lemma}}
The algorithm and analysis of the attack on the AMS sketch are similar, but simpler, than the attack on $\text{\texttt{CountSketch}}{}$.
We use query vectors of a similar form with varying $v^{(t)}_h$ and fixed tail norm scaled to $\sigma$.
The distribution of
the measurements of the tails (multiplied by $\mu^{(j)}_h$) $u^{*(t)}_j$ is $m^{-1/2} \mathcal{B}(m,1/2)-m/2$, which approaches
$\mathcal{N}(0,\sigma^2)$ for "large enough" $m$. These values are i.i.d.\ for $j\in \ell$ and $t\in [r]$.
The sketch content that is "visible" to the estimator has respective
values $\boldsymbol{u}^{(t)} := v_h^{(t)}\boldsymbol{1}_\ell + \boldsymbol{u}^{*(t)}$
with respective distribution
$\mathcal{N}_\ell(v_h^{(t)},\sigma^2)$.
In our attack we choose
$v^{(t)}_h \sim \mathcal{U}[a,c+2b]\cdot \sigma$, where $a,b,c$ are specified below.
We assume that at each step $t\in [r]$, the query response algorithm chooses an estimator that is correct with respect to $(\epsilon,\delta)$ and $\tau$. The estimator returns only one bit, essentially comparing the norm to $\tau$ with accuracy $\epsilon$ and confidence $1-\delta$:
\begin{definition} [Correct $\ell_2$-norm estimator]\label{correctnormest:def}
A norm estimator is correct for parameters $(\epsilon,\delta)$ and $\tau$
if
$\|\boldsymbol{v}\|^2_2 \geq (1+\epsilon)\tau^2$, the output is $1$ with probability $\geq 1-\delta$ and if
$\|\boldsymbol{v}\|^2_2 \leq \tau^2$ then the output is $-1$ with probability $\geq 1-\delta$.
\end{definition}
\begin{lemma}
A norm estimator applied to sketches of our query inputs is correct for parameters $(\epsilon,\delta)$ and $\tau$ if and only if
its reporting function (see Definition~\ref{correctmean:def}) is correct for parameters
$(\delta,a,c,\ell,\sigma)$ with
\begin{align*}
a &= \sqrt{(\tau/\sigma)^2 -1} \\
c &= \sqrt{(1+\epsilon)(\tau/\sigma)^2 -1}\ .
\end{align*}
\end{lemma}
\begin{proof}
We have
\begin{align*}
\|\boldsymbol{v}^{(t)}\|_2 \geq \tau & \iff
\sqrt{(v^{(t)}_h)^2 +\sigma^2} \geq \tau \implies
((v^{(t)}_h)^2 \geq \tau^2-\sigma^2 = \sigma^2((\tau/\sigma)^2-1) \\ & \implies |v^{(t)}_h|\geq \sigma\sqrt{((\tau/\sigma)^2-1)} = a\cdot \sigma\ .\\
\|\boldsymbol{v}^{(t)}\|^2_2 \geq (1+\epsilon)\tau^2 &\iff |v^{(t)}_h|\geq c \cdot \sigma\ .
\end{align*}
\end{proof}
We analyze
Algorithm~\ref{algo:AMSattack} using a correspondence to Algorithm~\ref{algo:meanestattack}.
We set the query vectors parameters according to the estimator parameters: We choose $\sigma^2 = \tau^2/2$ and thus
\begin{align*}
a &= 1 \\
c &= \sqrt{1+2\epsilon}
\end{align*}
Note that $c-a = \Theta(\epsilon)$ for $\epsilon<1$ and
$c-a = 1+\Theta(\sqrt{\epsilon})$ for $\epsilon > 1$.
We choose
$b = \Theta(\frac{1}{\sqrt{\ell}} \ln(\sqrt\frac{\ell}{\epsilon})$ which satisfies the condition in Lemma~\ref{biasminest:lemma}.
Since $\epsilon \geq \sqrt{\ln(1/\delta)/\ell}$ (otherwise correct estimators don't exist), we have $b = O(\epsilon)$ and thus
$c-a+2b = \Theta(\epsilon)$.
The adversarial input has
$\| \boldsymbol{z}^{(A)} \|^2_2 = r \cdot \sigma^2$ and to assess the bias we need to consider $\textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell}\|\boldsymbol{u}^{(A)}\|^2_2 \right]$.
From Lemma~\ref{biasminest:lemma} using a direct argument
\eqref{Esqbound:eq} we obtain
\[\textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell}\|\boldsymbol{u}^{(A)}\|^2_2 \right]
= r\sigma^2 \frac{r}{\ell^2} \Omega\left(\left(\frac{2}{c-a+2b}\right)^2 \right) =
\| \boldsymbol{z}^{(A)} \|^2_2 \frac{r}{\ell^2} \Omega(\epsilon^{-2}) .\]
This establishes the claim of the Lemma for the
regime of high bias $\xi = \Omega(1)$.
In order to derive tighter bounds for the regime $\xi <1$
we use a refined bound
by applying Lemma~\ref{squaredbound:lemma}:
\begin{align*}
\textsf{E}_{\mathcal{A}}\left[\frac{1}{\ell}\|\boldsymbol{u}^{(A)}\|^2_2 \right]
&=
r\sigma^2\left(1-O(\frac{1}{\sqrt{\ell}}) + \frac{r}{\ell^2} \Omega\left(\left(\frac{2}{c-a+2b}\right)^2 \right) \right) \\
&= \| \boldsymbol{z}^{(A)} \|^2_2 \left(1-O(\frac{1}{\sqrt{\ell}})+\frac{r}{\ell^2} \Omega(\epsilon^{-2}) \right)
\end{align*}
Here for $\xi = \Omega(1/\sqrt{\ell})$, we solve for $\xi = \Omega(r\epsilon^{-2}/\ell^2)$ and obtain the bound
$r = O(\xi \ell^2 \epsilon^2)$.
The confidence bounds of any fixed $\beta$ follows from the bound on the expectation using Markov inequality.
\section*{Conclusion}
Our results suggest interesting directions for followup work. We suspect that our attack technique can be generalized so that it applies with any randomized linear sketch that supports recovery of heavy hitters keys or even other families of features of the input vectors. Our attack techniques might be applicable to trained neural networks as blackbox attacks that construct "random noise" that mimics the presence of a signal. As discussed in the introduction, such attacks complement \cite{HardtW:STOC2013,JayaramWZ:ICLR2020} that constructed "invisible" noise that obfuscates the input but not the output.
\newpage
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:intro}
In this article, we are concerned with numerical symplectic methods for the following charged-particle dynamics (CPD)
(see {\cite{Benettin94,Hairer2018,Hairer2017,Lubich2017,Hairer2020,Lubich2020,WZ}})
\begin{equation}\label{charged-particle sts-cons}
\begin{array}[c]{ll}
{\ddot{x}(t)=\dot{x}(t) \times \frac{B(x(t))}{\epsilon} +F(x(t))}, \quad
x(0)=x_0,\quad \dot{x}(0)=\dot{x}_0,
\end{array}
\end{equation}
where $x(t)\in \RR^3$ represents the position of a particle
moving in an electro-magnetic field, the
magnetic field is described by $B(x) = \nabla_x \times A(x)$ with the vector potential
$A(x) \in \RR^3 $, $\epsilon \in(0,1]$ is a dimensionless parameter which is related to the strength of the magnetic field, and $F(x)$ is the nonlinear function of the form $F(x)=-\nabla U(x)$ with scalar potential $U(x)$. The initial data $x_0$ and $\dot{x}_0$ are
two given real-valued vectors independent of $\eps$. In this paper, we denote the {magnetic field} $B(x)$ by
$B(x)=\big(B_1(x),B_2(x),B_3(x)\big)^{\intercal}$, where $B_i(x) \in \RR $ for $i=1,2,3.$ Then it follows from the definition of the cross
product that
$ \dot{x} \times B(x) = \tilde{B}(x) \dot{x},$ where the skew symmetric
matrix $\tilde{B}(x)$ is given by
$\tilde{B}(x)=\left(
\begin{array}{ccc}
0 & B_3(x) & -B_2(x) \\
-B_3(x) & 0 & B_1(x) \\
B_2(x) & -B_1(x) & 0 \\
\end{array}
\right).
$
According to the conjugate momentum $ p = v + \frac{A(x)}{\epsilon}=v
-\frac{\tilde{B}(x)}{2\epsilon}x$, the CPD
\eqref{charged-particle sts-cons} can be converted into a
Hamiltonian system with the non-separable Hamiltonian
\begin{equation}\label{H}H(x,p)=\frac{1}{2} \norm{p+\frac{\tilde{B}(x)}{2\epsilon}x}^2 + U(x),\end{equation}
where $\norm{ \cdot}$ denotes the Euclidean norm.
For the non-separable
Hamiltonian \eqref{H}, it is well known that its flow is
symplectic (\cite{Hairer2002}), i.e., it preserves the differential 2-form
$\sum\limits_{i=1}^{3} dx_i\wedge dp_i.$ In this work, we are devoted to the formulation and analysis of a novel kind of symplectic methods
for CPD.
{The CPD \eqref{charged-particle sts-cons} arises in many applications such as plasma physics, astrophysics and magnetic fusion devices} (see, e.g.
{\cite{Arnold97,Benettin94,Cary2009,Northrop63}}) {and a large variety of methods has
been constructed and developed for numerical solutions of the system}. For the normal magnetic field $\epsilon \approx 1$,
the Boris method proposed in
\cite{Boris} is popular due to its simplicity and good properties (cf. \cite{Hairer2017, Qin2013}). Other classical schemes have also been proposed for CPD such as multistep methods \cite{Lubich2017,wangwu2020}, splitting methods \cite{Knapp} and so on.
In the recent few {decades}, structure-preserving methods for differential equations have received more and more attention (\cite{bao-zhao,Feng1985,E. Hairer, Hairer2002,S. MacNamara,Sanz-Serna1988,Suris}). There are various structure-preserving methods which have been applied to CPD, including volume-preserving methods
\cite{He2015}, energy-preserving integrators \cite{Brugnano2020, Brugnano2019, li2022,Chacon}, symmetric methods \cite{Lubich2017, wangwu2020} and symplectic methods \cite{ S. Blanes2012, He2017,Y. Shi, Tao2016, Xiao,Xiao21,Zhang2016}. It is noted that these structure-preserving methods are all designed for CPD with normal magnetic field {$ \epsilon \approx 1$} and no analysis has been given for the strong regime $0<\epsilon \ll 1$. If one considers the methods mentioned above for CPD with $0<\eps \ll 1$,
the accuracy usually depends on $1/\eps^j$ for some $j>0$ and this makes the methods inefficient for small $\eps$.
Concerning the methods for CPD with strong regime $0<\epsilon \ll 1$, to the best of my knowledge, most existed algorithms are devoted to the accuracy or long time near conservation \cite{VP1,Zhao,VP4,VP5,VP6,Hairer2018,Lubich2020,Lubich2022,WZ}. Structure-preserving methods have not been considered and analysed for CPD with strong regime $0<\epsilon \ll 1$.
From the point of Hamiltonian system, it is well known that symplecticity is a very
important property, which has been
investigated by many researchers (\cite{Feng1985, Webb2014}). Various symplectic methods have been derived such as
symplectic Runge-Kutta methods
\cite{Sanz-Serna1988}, symplectic Runge-Kutta-Nystr\"{o}m methods
\cite{Suris}, and continuous-stage symplectic methods \cite{E. Hairer, Miyatake, Tang2018}.
However, as stated above, symplectic methods for CPD \cite{He2017, Y. Shi, Tao2016, Xiao,Xiao21,Zhang2016} have only been designed for
the system under normal magnetic fields and there is no numerical approximation or error analysis for strong magnetic fields.
In this article, we construct a new kind of symplectic methods for CPD with arbitrary electromagnetic fields and {establish} the rigorous error estimates. To formulate the methods, the ideas of exponential methods (see, e.g. \cite{bao-cai, Frenod, Hairer2002, M.Hochbruck2017, Mei2017, wang-2016}) and continuous-stage methods
will be employed. Moreover, the convergence of the obtained symplectic methods is rigorously researched, {and the error estimates show that some proposed methods {have} a uniform accuracy in the position w.r.t. $\eps$. This means that symplectic methods with uniform (w.r.t. $\eps$) error bounds are achieved {in this paper, and performances superior to the standard symplectic methods are
demonstrated on some examples.}
In the analysis of symplecticity, we consider the differential 2-form for the methods. However, since the
Hamiltonian is non-separable and the new methods are exponential type, some challenges and
difficulties emerge in the achievement of symplectic conditions. To make the derivation go smoothly,
we consider some transformations of the original system and the methods, and make well use of the antisymmetry of $\tilde{B}(x)$.
{These} transformations can keep the symplecticity of the methods, and thus
symplectic methods can be formulated by deriving the symplectic conditions for the transformed methods.
Concerning the convergence of the methods for CPD under nonhomogeneous magnetic fields,
exponential variation-of-constants formula unfortunately does not hold and thence the local errors of exponential methods cannot be derived.
In order to overcome this difficulty, we design a novel system called as linearized system and establish exponential variation-of-constants formula for it. The error between this linearized system and {the original} one is deduced and based on which, we
manage to derive the convergence of the methods. Moreover, following this approach, a method with uniform second order error bound and very good long time energy conservation is obtained for CPD under maximal ordering magnetic fields {\cite{scaling1,Lubich2020,scaling2}}. For the system of this case, two filtered Boris algorithms were presented in \cite{Lubich2020} {and they} were proved to have uniform second order error bound but without long time behaviour. An energy-preserving method was constructed in \cite{WZ} and uniform first order error estimate was established. Compared with {the methods derived in \cite{Lubich2020,WZ}}, the method proposed in this paper {has both uniform second order accuracy and symplecticity, and is} a brand-new method to CPD under maximal ordering magnetic fields.}
The rest of this paper is organised as follows. In Section \ref{sec1: methods}, for solving CPD with homogeneous magnetic field, we propose the
scheme of continuous-stage adapted exponential methods and derive its symplectic conditions. Then two practical symplectic methods up to order four {are presented in
Section \ref{sec:prac meth} and their properties including convergence and implementations are also studied there}. At the end of Section \ref{sec:prac meth}, { one
numerical experiment is provided and the results show the favorable behaviour of
the new methods in comparison with the Boris method and two symplectic Runge-Kutta methods}. Moreover, {the extension and application of the proposed symplectic methods to CPD with non-homogeneous magnetic fields are discussed in Section \ref{sec:general}. Three practical methods as well as their error estimates are proposed and two numerical tests are
given to demonstrate that they are sharp.} Section \ref{sec:conclusions} is
devoted to the conclusions of this paper.
\section{{Adapted exponential methods} and symplectic conditions}\label{sec1: methods}
In this section, we shall propose the novel methods for \eqref{charged-particle sts-cons}
in a homogeneous magnetic field $B=\big(B_1,B_2,B_3\big)^{\intercal}$ and analyse their symplectic conditions.
{We start by the variation-of-constants formula of the CPD \eqref{charged-particle sts-cons} which reads (see
\cite{Lubich2020})
\begin{equation}\label{VOC}
\begin{aligned}
x(t_n+h)&=x(t_n) +h\varphi_1( hM ) v(t_n)+h^2
\int_{0}^1(1-\tau) \varphi_1((1-\tau) hM ) F(x(t_n+h\tau))
d\tau,\\
v(t_n+h)&=\varphi_{0}( hM )v(t_n)+h
\int_{0}^1
\varphi_{0}((1-\tau) hM ) F(x(t_n+h\tau)) d\tau,
\end{aligned}
\end{equation}
for any step size $h\geq0$ and $t_n=nh$, where $v(t):=\dot{x}(t)$,
$M=\frac{\tilde{B}}{\epsilon}$ and the $\varphi$-functions are defined by (see
\cite{Hochbruck2005})
\begin{equation*}\label{phi}
\varphi_0(z)=e^{z},\ \ \varphi_k(z)=\int_{0}^1
e^{(1-\sigma)z}\frac{\sigma^{k-1}}{(k-1)!}d\sigma, \ \
k=1,2,\ldots.
\end{equation*}
To derive the novel methods, we consider the variation-of-constants formula \eqref{VOC} and utilize the idea of continuous-stage methods, which leads to the following definition.
\begin{mydef}
\label{scheme 1} \textbf{(Adapted exponential
methods)} By denoting
the numerical approximation $x_n\approx x(t_n),\, v_n\approx v(t_n)$ and choosing $v_0=\dot{x}_0$, a continuous-stage adapted exponential
method for solving \eqref{charged-particle sts-cons} is
defined {as}
\begin{equation}
\begin{aligned}\label{CSAEI}
&X_{\tau}=x_{n}+ \tau h\varphi_{1}(\tau hM ) v_{n}+
h^2 \int_{0}^{1}{\alpha}_{\tau \sigma}(hM)F (X_\sigma)d\sigma,\\
&x_{n+1}=x_{n}+ h\varphi_1(hM)
v_{n}+h^2 \int_{0}^{1}\beta_{\tau}(hM)F (X_{\tau})d\tau,\ \ v_{n+1}=\varphi_0(hM)v_{n}+h
\int_{0}^{1}\gamma_{\tau}(hM)F(X_{\tau})d\tau,
\end{aligned}
\end{equation}
where $h$ is the step size, $ X_{\tau} $ is a function about $\tau $, and
the coefficients $ {\alpha}_{\tau \sigma}(hM) $, $ \beta_{\tau}(hM) $ and $
\gamma_{\tau}(hM) $ are functions {depending} on $ hM$.
For the method adapted to the Hamiltonian system \eqref{H}, we
consider the continuous-stage adapted exponential
method \eqref{CSAEI} as well as a momentum calculation $$p_{n+1} = v_{n+1} -\frac{\tilde{B}}{2\epsilon}
x_{n+1}.$$
\end{mydef}
In what follows, we derive the symplectic conditions of the method given in Definition \ref{scheme 1}.
When applied to the non-separable Hamiltonian \eqref{H}, the method is called as symplectic
if it preserves the symplecticity exactly, i.e. (\cite{Hairer2002}),
\begin{equation*}\label{NMSD}
\sum\limits_{J=1}^{3}dx_{n+1}^{J} \wedge
dp_{n+1}^{J}=\sum_{J=1}^{3}dx_{n}^{J} \wedge dp_{n}^{J},
\end{equation*}
where the superscript $(\cdot)^{J}$ denotes the $J$th component of a vector.
To prove this statement, we first make some transformations of the system
\eqref{charged-particle sts-cons} and of the method \eqref{CSAEI}.
Since $\tilde{B}$ is a skew-symmetric matrix, it
can be expressed as $ \tilde{B}=K \Omega K^\textup{H}, $ where $K$
is a unitary matrix and $\Omega=
\textmd{diag}(-{\norm{B}}\mathrm{i},0,{\norm{B}}\mathrm{i}) $.
Now we denote the new variables as
\begin{equation}\label{change of variable}
\tilde{x}(t)= K^\textup{H} x(t),\quad \tilde{v}(t)= K^\textup{H} v(t).
\end{equation}
Then the system
\eqref{charged-particle sts-cons} becomes
\begin{equation*}\label{necharged-sts-first order}
\begin{array}[c]{ll}
\frac{d}{dt }\left(
\begin{array}{c}
\tilde{x} \\
\tilde{v} \\
\end{array}
\right) =\left(
\begin{array}{cc}
0 & I \\
0 & \tilde{\Lambda} \mathrm{i} \\
\end{array}
\right) \left(
\begin{array}{c}
\tilde{x} \\
\tilde{v} \\
\end{array}
\right)+\left(
\begin{array}{c}
0 \\
\tilde{F}(\tilde{x}) \\
\end{array}
\right),\quad \left(
\begin{array}{c}
\tilde{x}_0 \\
\tilde{v}_0 \\
\end{array}
\right)=\left(
\begin{array}{c}
K^\textup{H} x_0 \\
K^\textup{H} \dot{x}_0\\
\end{array}
\right),
\end{array}
\end{equation*}
where $\tilde{\Lambda}=
\textmd{diag}(-\tilde{\zeta},0,\tilde{\zeta})$ with
$\tilde{\zeta}=\frac{{\norm{B}}}{\epsilon}$ and
$\tilde{F}(\tilde{x})=K^\textup{H} F(K \tilde{x})=
-\nabla_{\tilde{x}} U(K\tilde{x})$.
It is noted that the vector $\tilde{x}$ is denoted by $\tilde{x}=(\tilde{x}^{1},\tilde{x}^{2},\tilde{x}^{3})^{\intercal}$ and all
the vectors in $\mathbb{R}^{3}$ or $\mathbb{C}^{3}$ use the same notation in this paper.
For the variables \eqref{change of variable} and the property of $K$, we
observe that
$\tilde{x}^{1}=\bar{\tilde{x}}^{3}$,
$\tilde{v}^{1}=\bar{\tilde{v}}^{3}$, and $\tilde{x}^{2}$,
$\tilde{v}^{2} \in \mathbb{R}$. We shall denote $\bar{\tilde{x}}^{3}$
as the conjugate of $\tilde{x}^{3}$ and the same expression is applied to other functions.
In the light of \eqref{change of variable}, the corresponding Hamiltonian system \eqref{H} is given by
\begin{equation}\label{new-H}
\begin{array}[c]{ll}{\dot{\tilde{x}}} =\nabla_{\tilde{p}} \tilde{H}(\tilde{x},\tilde{p})=\tilde{p}+\frac{1}{2}
\tilde{\Lambda} \mathrm{i}\tilde{x},
\ \ \ \ {\dot{\tilde{p}}}=-\nabla_{\tilde{x}} \tilde{H}(\tilde{x},\tilde{p})=-\frac{1}{2}
\big(\tilde{\Lambda} \mathrm{i}\big)^\textup{H} \big(\tilde{p}+\frac{1}{2}
\tilde{\Lambda} \mathrm{i}
\tilde{x}\big)-\nabla_{\tilde{x}}\tilde{U}(K\tilde{x}),
\end{array}
\end{equation}with
$\tilde{H}(\tilde{x},\tilde{p}) = \frac{1}{2} \norm{\tilde{p}+\frac{1}{2}
\tilde{\Lambda} \mathrm{i}\tilde{x}}^2+\tilde{U}(K\tilde{x}).$
For this transformed Hamiltonian system, the method \eqref{CSAEI} takes the {following transformed} form
\begin{equation}\label{TRM}
\begin{aligned}
&\tilde{X}_{\tau}=\tilde{x}_{n}+ h\tau \varphi_1\big(\tau W\big) \tilde{v}_{n}+
h^2 \int_{0}^{1}{\alpha}_{\tau \sigma}\big(W\big)\tilde{F} \big(\tilde{X}_\sigma\big)d\sigma,\\
& \tilde{x}_{n+1}=\tilde{x}_{n}+ h\varphi_1\big(W\big)
\tilde{v}_{n}+h^2 \int_{0}^{1}\beta_{\tau}\big(W\big)\tilde{F} \big(\tilde{X}_{\tau}\big)d\tau,\\
&\tilde{v}_{n+1}=\varphi_0\big(W\big)\tilde{v}_{n}+h
\int_{0}^{1}\gamma_{\tau}\big(W\big)\tilde{F}\big(\tilde{X}_{\tau}\big)d\tau,\ \ \tilde{p}_{n+1} = \tilde{v}_{n+1} -\frac{1}{2}
\tilde{\Lambda}\mathrm{i}\tilde{x}_{n+1},
\end{aligned}
\end{equation}
where $W=h\tilde{\Lambda}\mathrm{i}$.
Now we present symplectic conditions of the {transformed method \eqref{TRM} and then get the symplecticity of the method} given in Definition \ref{scheme 1}.
\begin{mytheo}\label{thm: symp} \textbf{(Symplecticity)}
{(a)} For solving {the transformed Hamiltonian {system \eqref{new-H}}, the map $(\tilde{x}_n,\tilde{p}_n ) \rightarrow (\tilde{x}_{n+1},\tilde{p}_{n+1})$ determined by the transformed method \eqref{TRM}} is
symplectic if the coefficients satisfy
\begin{equation}\label{17}
\begin{aligned}
(i)\quad &\gamma_{\tau}(W) -W\beta_{\tau}(W)
=d_{\tau} I, \ \ d_\tau\in \mathbb{C},\\
(ii)\quad &\gamma_{\tau}(W) ( \bar{\varphi}_{1}(W)-\tau\bar{\varphi}_{1}(\tau W) )
=\beta_{\tau}(W) ( e^{-W}+W
\bar{\varphi}_{1}(W)-\tau W \bar{\varphi}_{1}(\tau W) ),\\
(iii)\quad &\bar{\beta}_{\sigma}(W) \gamma_{\tau}(W) - W\bar{\beta}_{\sigma}(W)
\beta_{\tau}(W)/2 -\bar{{\alpha}}_{\tau \sigma}(W) ( \gamma_{\tau}(W)
-W\beta_{\tau}(W) ) \\
\ &=
\beta_{\tau}(W) \bar{ \gamma}_{\sigma}(W) +W \beta_{\tau}(W)
\bar{\beta}_{\sigma}(W)/2 -{\alpha}_{\sigma \tau}(W) ( \bar{\gamma}_{\sigma}(W)
+W\bar{\beta}_{\sigma}(W) ).
\end{aligned}
\end{equation}
{(b)} {If the transformed method \eqref{TRM} is symplectic, the map $(x_n,p_n ) \rightarrow (x_{n+1},p_{n+1})$ determined by Definition \ref{scheme 1} is
symplectic for solving the Hamiltonian system \eqref{H}.}
\end{mytheo}
\label{necharged-sts-first order}
\begin{proof}
{ We first prove the second statement {(b)}.
By the notation of differential
2-form and with the help of $K$,} we have
\begin{equation}
\begin{aligned}
&\sum\limits_{J=1}^{3}dx_{n+1}^{J} \wedge
dp_{n+1}^{J}=\sum\limits_{J=1}^{3}d\bar{x}_{n+1}^{J} \wedge
dp_{n+1}^{J}=\sum\limits_{J=1}^{3}d\big(\bar{K}\bar{\tilde{x}}_{n+1}\big)^{J}
\wedge
d\big(K\tilde{p}_{n+1}\big)^{J}\\
=& \sum\limits_{J=1}^{3}\bigg({ d}
\sum\limits_{i=1}^{3}\big(\bar{K}_{Ji}\bar{\tilde{x}}_{n+1}^{i}\big)\bigg)\wedge
\bigg( {
d} \sum\limits_{\iota=1}^{3}\big(K_{J\iota}\tilde{p}_{n+1}^{\iota}\big)\bigg)
= \sum\limits_{J=1}^{3}\bigg( \sum\limits_{i=1}^{3}\big(\bar{K}_{Ji}{
d}\bar{\tilde{x}}_{n+1}^{i}\big)\bigg)\wedge
\bigg( \sum\limits_{\iota=1}^{3}\big(K_{J\iota}{ d}\tilde{p}_{n+1}^{\iota}\big)\bigg)\\
= & \sum\limits_{J=1}^{3}\sum\limits_{i=1}^{3}\sum\limits_{\iota=1}^{3} \bar{K}_{Ji} K_{J\iota} \big({
d}\bar{\tilde{x}}_{n+1}^{i} \wedge { d}\tilde{p}_{n+1}^{\iota}\big)
= \sum\limits_{i=1}^{3} {
d}\bar{\tilde{x}}_{n+1}^{i} \wedge { d}\tilde{p}_{n+1}^{i}=
\sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n+1}^{J} \wedge {
d}\tilde{p}_{n+1}^{J}.\nonumber
\end{aligned}\label{td}
\end{equation}
Likewise, one gets $ \sum\limits_{J=1}^{3}dx_{n}^{J}
\wedge dp_{n}^{J}= \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n}^{J} \wedge { d}\tilde{p}_{n}^{J}.$ Hence {the second statement is hold.
For the first statement {(a)},
we need} to prove
$ \sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n+1}^{J} \wedge {
d}\tilde{p}_{n+1}^{J}= \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n}^{J} \wedge { d}\tilde{p}_{n}^{J},
$
i.e.,
\begin{equation}\label{imp res}
\begin{aligned} \sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n+1}^{J} \wedge {
d}\tilde{v}_{n+1}^{J}-\frac{1}{2} \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n+1}^{J} \wedge { d}
\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n+1}^{J}\big)=
\sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n}^{J} \wedge {
d}\tilde{v}_{n}^{J}-\frac{1}{2} \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n}^{J} \wedge { d}
\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n}^{J}\big).
\end{aligned}
\end{equation}
\textbf{(I) Computation of the left hand side of
\eqref{imp res}.}
It is noted that
$ \overline{{ d}\bar{\tilde{x}}_{n}^{J} \wedge { d}
\tilde{v}_{n}^{J}} ={ d} \tilde{x}_{n}^{J} \wedge { d}
\bar{\tilde{v}}_{n}^{J}$ and $ { d}\bar{\tilde{x}}_{n}^{J} \wedge {
d} \tilde{x}_{n}^{J} \in \mathrm{i} \mathbb{R}$, and
denote the method by{
\begin{equation}
\begin{aligned}
&\tilde{X}^{J}_{\tau}=\tilde{x}^{J}_{n}+ h\tau \varphi_1\big(\tau{W}^{J}\big) \tilde{v}^{J}_{n}+
h^2 \int_{0}^{1}{\alpha}_{\tau \sigma}\big({W}^{J}\big)\tilde{F}^{J}_{\sigma}d\sigma,\\
&\tilde{x}^{J}_{n+1}=\tilde{x}^{J}_{n}+
h\varphi_1({W}^{J})
\tilde{v}^{J}_{n}+h^2 \int_{0}^{1}\beta_{\tau}\big({W}^{J}\big)\tilde{F}^{J}_{\tau}d\tau,\ \ \tilde{v}^{J}_{n+1}=e^{{W}^{J}}\tilde{v}^{J}_{n}+h
\int_{0}^{1}\gamma_{\tau}\big({W}^{J}\big)\tilde{F}^{J}_{\tau}d\tau,
\end{aligned}\label{a}
\end{equation}}where the superscript $(\cdot)^{J}$ is the $J$th component of a vector
or a matrix and $\tilde{F}^{J}_{\tau}$ is the $J$th component of
$\tilde{F}\big(\tilde{X}_{\tau}\big)$.
Based on the above {expression}, in what follows we
calculate the left hand side of
\eqref{imp res}.
Inserting the scheme of \eqref{a} into \eqref{imp res}, it arrives that
\begin{equation}
\begin{aligned}\label{aaa1}
&\sum\limits_{J=1}^{3}\bigg(d\bar{\tilde{x}}_{n+1}^{J}\wedge d\tilde{v}_{n+1}^{J}-\frac{1}{2}d\bar{\tilde{x}}_{n+1}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n+1}^{J}\big)\bigg)
= \sum\limits_{J=1}^{3}\bigg(e^{{W}^{J}} d\bar{\tilde{x}}_{n}^{J} \wedge
d\tilde{v}_{n}^{J}
+h\int_{0}^{1}\Big(\gamma_{\tau}\big({W}^{J}\big)
d\bar{\tilde{x}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau\\
&+he^{{W}^{J}} \bar{\varphi}_{1}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{v}_{n}^{J}
+h^{2}\int_{0}^{1}\Big(\bar{\varphi}_{1}\big({W}^{J}\big)\gamma_{\tau}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau
+h^{2} e^{{W}^{J}} \int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{v}_{n}^{J}\Big)d\tau\\
&+h^{3}
\int_{0}^{1}\int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big) \gamma_{\sigma}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{F}_{\sigma}^{J}\Big)d\tau d\sigma
- \frac{1}{2}d\bar{\tilde{x}}_{n}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n}^{J}\big)
-\frac{1}{2} {W}^{J} \varphi_{1}\big({W}^{J}\big) d\bar{\tilde{x}}_{n}^{J}\wedge d\tilde{v}_{n}^{J}\\
&-\frac{1}{2}h {W}^{J} \int_{0}^{1}\Big(\beta_{\tau}\big({W}^{J}\big)
d\bar{\tilde{x}}_{n}^{J}\wedge d\tilde{F}_{\tau}^{J}\Big)d\tau
-\frac{1}{2}{W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{x}_{n}^{J}
-\frac{1}{2}h {W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) \varphi_{1}\big({W}^{J}\big)
d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{v}_{n}^{J}\\
&-\frac{1}{2}h^{2} {W}^{J} \varphi_{1}\big({W}^{J}\big) \int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{v}_{n}^{J}\Big)d\tau
-\frac{1}{2}h^{3} {W}^{J} \int_{0}^{1}\int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big) \beta_{\sigma}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{F}_{\sigma}^{J}\Big)d\tau d\sigma \\
&-\frac{1}{2}h^{2} {W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big)\int_{0}^{1}\Big(\beta_{\tau}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau
-\frac{1}{2}h{W}^{J} \int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{x}_{n}^{J}\Big)d\tau\bigg).
\end{aligned}
\end{equation}
Considering the properties of coefficient functions and exterior product, we obtain
\begin{equation*}
\begin{aligned}
&-\sum\limits_{J=1}^{3}\frac{1}{2}h
{W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{x}_{n}^{J}=-\sum\limits_{J=1}^{3}\frac{1}{2}h{W}^{J}
\varphi_{1}\big({W}^{J}\big) d\bar{\tilde{x}}_{n}^{J} \wedge d\tilde{v}_{n}^{J},\\
&-\sum\limits_{J=1}^{3}\frac{1}{2}h{W}^{J} \int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{x}_{n}^{J}\Big)d\tau
=-\sum\limits_{J=1}^{3}\frac{1}{2}h{W}^{J} \int_{0}^{1}\Big(\beta_{\tau}\big({W}^{J}\big)
d\bar{\tilde{x}}_{n}^{J}\wedge d\tilde{F}_{\tau}^{J}\Big)d\tau,\\
&-\sum\limits_{J=1}^{3}h^{2}\int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)e^{{W}^{J}} d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{v}_{n}^{J}\Big)d\tau
=\sum\limits_{J=1}^{3}h^{2}\int_{0}^{1}\Big(\beta_{\tau}\big({W}^{J}\big)e^{-{W}^{J}} d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau,\\
&-\sum\limits_{J=1}^{3}\frac{1}{2}h^{2} {W}^{J} \varphi_{1}\big({W}^{J}\big) \int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{v}_{n}^{J}\Big)d\tau
=-\sum\limits_{J=1}^{3}\frac{1}{2}h^{2} {W}^{J} \bar{\varphi}_{1}\big({W}^{J}\big) \int_{0}^{1}\Big(\beta_{\tau}\big({W}^{J}\big) d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau.\\
\end{aligned}
\end{equation*}
Thus \eqref{aaa1} can be simplified as{
\begin{equation}
\begin{aligned}
&\sum\limits_{J=1}^{3}\bigg(d\bar{\tilde{x}}_{n+1}^{J}\wedge d\tilde{v}_{n+1}^{J}-\frac{1}{2}d\bar{\tilde{x}}_{n+1}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n+1}^{J}\big)\bigg)\\
=&\sum\limits_{J=1}^{3}\bigg(d\bar{\tilde{x}}_{n}^{J} \wedge d\tilde{v}_{n}^{J}
- \frac{1}{2}d\bar{\tilde{x}}_{n}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n}^{J}\big)
+ h\int_{0}^{1}\Big( \gamma_{\tau}\big({W}^{J}\big)-{W}^{J} \beta_{\tau}\big({W}^{J}\big) \Big)
\big(d\bar{\tilde{x}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\big)d\tau\\
&+\Big(he^{{W}^{J}} \bar{\varphi}_{1}\big({W}^{J}\big)
-\frac{1}{2}h {W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) \varphi_{1}\big({W}^{J}\big)\Big)
d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{v}_{n}^{J}\\
&+h^{2}\int_{0}^{1}\Big(\bar{\varphi}_{1}\big({W}^{J}\big) \gamma_{\tau}\big({W}^{J}\big)
- \beta_{\tau}\big({W}^{J}\big) e^{-{W}^{J}}
-W^J \bar{\varphi}_{1}\big({W}^{J}\big) \beta_{\tau}\big({W}^{J}\big)\Big)
\big(d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\big)d\tau\\
&+h^{3}\int_{0}^{1}\int_{0}^{1}\Big(\bar{\beta}_{\tau}\big({W}^{J}\big)
\gamma_{\sigma}\big({W}^{J}\big) -\frac{1}{2}{W}^{J}
\bar{\beta}_{\tau}\big({W}^{J}\big)
\beta_{\sigma}\big({W}^{J}\big)\Big)
\big(d\bar{\tilde{F}}_{\tau}^{J} \wedge d\tilde{F}_{\sigma}^{J}\big)d\tau d\sigma\bigg),\\
\end{aligned}\label{21}
\end{equation}}where we have used the fact $e^{{W}^{J}} -{W}^{J}\varphi_{1}\big({W}^{J}\big)=I$
which is given by the definition of $\varphi$-functions.
{Further, it follows from the first formula of \eqref{a} that
$d\tilde{x}^{J}_{n} =d\tilde{X}^{J}_{\tau}-\tau h \varphi_1\big(\tau{W}^{J}\big)d\tilde{v}^{J}_{n}-h^2
\int_{0}^{1} \alpha_{\tau \sigma}\big({W}^{J}\big)d\tilde{F}_{\sigma}^{J}d\sigma. $
Then $
d\bar{\tilde{x}}^{J}_{n}\wedge d\tilde{F}^{J}_{\tau}
=d\bar{\tilde{X}}^{J}_{\tau}\wedge d\tilde{F}^{J}_{\tau} -
\tau h \bar{\varphi}_1\big(\tau{W}^{J}\big)d\bar{\tilde{v}}^{J}_{n}
\wedge d\tilde{F}^{J}_{\tau}-h^2
\int_{0}^{1}\big(\bar{{\alpha}}_{\tau \sigma}\big({W}^{J}\big)
d\bar{\tilde{F}}_{\sigma}^{J}\wedge d\tilde{F}^{J}_{\tau}\big)d\sigma. $
Therefore, the formula \eqref{21} can be rewritten as
\begin{align}
&\sum\limits_{J=1}^{3}d\bar{\tilde{x}}_{n+1}^{J}\wedge d\tilde{v}_{n+1}^{J}-\frac{1}{2}\sum\limits_{J=1}^{3}d\bar{\tilde{x}}_{n+1}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n+1}^{J}\big)
=\sum\limits_{J=1}^{3} d\bar{\tilde{x}}_{n}^{J} \wedge
d\tilde{v}_{n}^{J}
- \frac{1}{2}\sum\limits_{J=1}^{3} d\bar{\tilde{x}}_{n}^{J}\wedge d\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n}^{J}\big)\nonumber \\
&\qquad \qquad \quad
+ h\sum\limits_{J=1}^{3}\int_{0}^{1} \Big( \gamma_{\tau}\big({W}^{J}\big)
-{W}^{J}\beta_{\tau}\big({W}^{J}\big) \Big)
\Big(d\bar{\tilde{X}}^{J}_{\tau}\wedge d\tilde{F}^{J}_{\tau}\Big)d\tau \label{c} \\
&\qquad \qquad \quad
+\sum\limits_{J=1}^{3} \Big(he^{{W}^{J}} \bar{\varphi}_{1}\big({W}^{J}\big)
-\frac{1}{2}h {W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) \varphi_{1}\big({W}^{J}\big)\Big)
d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{v}_{n}^{J} \label{d} \\
&\qquad \qquad \quad
+h^{2}\sum\limits_{J=1}^{3}\int_{0}^{1} \bigg(\bar{\varphi}_{1}\big({W}^{J}\big) \gamma_{\tau}\big({W}^{J}\big)
- \beta_{\tau}\big({W}^{J}\big) e^{-{W}^{J}}-{W}^{J} \bar{\varphi}_{1}\big({W}^{J}\big) \beta_{\tau}\big({W}^{J}\big) \nonumber \\
&\qquad \qquad \quad
-\tau \bar{\varphi}_1\big(\tau{W}^{J}\big)\Big(\gamma_{\tau}\big({W}^{J}\big)
-{W}^{J}\beta_{\tau}\big({W}^{J}\big)\Big)\bigg)
\Big(d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau\label{e}\\
&\qquad \qquad \quad
+h^{3}\sum\limits_{J=1}^{3}\int_{0}^{1}\int_{0}^{1}\bigg(\bar{\beta}_{\sigma}\big({W}^{J}\big) \gamma_{\tau}\big({W}^{J}\big)
-\frac{1}{2}{W}^{J} \bar{\beta}_{\sigma}\big({W}^{J}\big) \beta_{\tau}\big({W}^{J}\big) \nonumber \\
&\qquad \qquad \quad
-\bar{{\alpha}}_{\tau\sigma}\big({W}^{J}\big)\Big( \gamma_{\tau}\big({W}^{J}\big) -{W}^{J} \beta_{\tau}\big({W}^{J}\big) \Big)
\bigg)\Big(d\bar{\tilde{F}}_{\sigma}^{J} \wedge d\tilde{F}_{\tau}^{J}\Big)d\tau d\sigma\label{f}.
\end{align}\label{21a}
\textbf{(II) Proof of the results.}
Next, we prove that all the formulae
\eqref{c}--\eqref{f} are equal to 0 under the conditions \eqref{17}.}
$\bullet$ \textbf{Prove that the term $\eqref{c}$ is
zero.}
In the light of the first condition
of \eqref{17}, ${F}(x)=-\nabla_{x} U(x)$ and \eqref{a}, we find $\sum\limits_{J=1}^{3}d\bar{\tilde{X}}^{J}_{\tau}\wedge d\tilde{F}^{J}_{\tau}=\sum\limits_{J=1}^{3} d
X^{J}_{\tau}\wedge dF^{J}_{\tau}$ which further implies
\begin{equation}
\begin{aligned}
&\sum\limits_{J=1}^{3} \Big( \gamma_{\tau}\big({W}^{J}\big)
-{W}^{J}\beta_{\tau}\big({W}^{J}\big) \Big)
d\bar{\tilde{X}}^{J}_{\tau}\wedge d\tilde{F}^{J}_{\tau}
=d_{\tau}\sum\limits_{J=1}^{3} d\bar{\tilde{X}}^{J}_{\tau}\wedge
d\tilde{F}^{J}_{\tau} =d_{\tau}\sum\limits_{J=1}^{3} d X^{J}_{\tau}\wedge
dF^{J}_{\tau}\\
=&-d_{\tau}\sum\limits_{J=1}^{3} dF^{J}_{\tau} \wedge d X^{J}_{\tau}
= -d_{\tau} \sum\limits_{J=1}^{3} d\bigg(-\frac{\partial{ U} }
{\partial{x} }(X_\tau) \bigg)^{J}\wedge d X_{\tau}^{J}
=d_{\tau} \sum\limits_{J,I=1}^{3}\frac{\partial^{2}{ U(X_\tau)} }
{\partial{x^{J}} \partial{x^{I}}} d X_{\tau}^{I} \wedge d
X_{\tau}^{J}=0.
\end{aligned}\nonumber
\end{equation}
$\bullet$ \textbf{Prove that the term $\eqref{d}$ is
zero.} Moreover, it can be checked that
\begin{equation}
\begin{aligned}
&\sum\limits_{J=1}^{3}\Big(he^{{W}^{J}} \bar{\varphi}_{1}\big({W}^{J}\big)
-\frac{1}{2}h {W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big) \varphi_{1}\big({W}^{J}\big)\Big)
d\bar{\tilde{v}}_{n}^{J} \wedge d\tilde{v}_{n}^{J}\\
=&\Big(he^{{W}^{1}}
\bar{\varphi}_{1}\big({W}^{1}\big) -\frac{1}{2}h
{W}^{1}\bar{\varphi}_{1}\big({W}^{1}\big)
\varphi_{1}\big({W}^{1}\big)\Big)
d\bar{\tilde{v}}_{n}^{1} \wedge d\tilde{v}_{n}^{1}\\
&+\Big(he^{{W}^{2}} \bar{\varphi}_{1}\big({W}^{2}\big)
-\frac{1}{2}h {W}^{2}\bar{\varphi}_{1}\big({W}^{2}\big) \varphi_{1}\big({W}^{2}\big)\Big)
d\bar{\tilde{v}}_{n}^{2} \wedge d\tilde{v}_{n}^{2}\\
&+\Big(he^{{W}^{3}} \bar{\varphi}_{1}\big({W}^{3}\big)
-\frac{1}{2}h {W}^{3}\bar{\varphi}_{1}\big({W}^{3}\big) \varphi_{1}\big({W}^{3}\big)\Big)
d\bar{\tilde{v}}_{n}^{3} \wedge d\tilde{v}_{n}^{3}.\\
\end{aligned}\nonumber
\end{equation}
According to the property of $\tilde{v}_{n}$, it yields
\begin{equation}
d\bar{\tilde{v}}_{n}^{1} \wedge d\tilde{v}_{n}^{1}=-d\bar{\tilde{v}}_{n}^{3} \wedge d\tilde{v}_{n}^{3}, \ \ \ d\bar{\tilde{v}}_{n}^{2} \wedge d\tilde{v}_{n}^{2}=0, \nonumber
\end{equation}
and
\begin{equation}
he^{{W}^{1}} \bar{\varphi}_{1}\big({W}^{1}\big)
-\frac{1}{2}h
{W}^{1}\bar{\varphi}_{1}\big({W}^{1}\big)
\varphi_{1}\big({W}^{1}\big)
=he^{{W}^{3}}
\bar{\varphi}_{1}\big({W}^{3}\big) -\frac{1}{2}h
{W}^{3}\bar{\varphi}_{1}\big({W}^{3}\big)
\varphi_{1}\big({W}^{3}\big).\nonumber
\end{equation}
Thus, we obtain
\begin{equation}
\sum\limits_{J=1}^{3}\Big(he^{{W}^{J}} \bar{\varphi}_{1}\big({W}^{J}\big)
-\frac{1}{2}h
{W}^{J}\bar{\varphi}_{1}\big({W}^{J}\big)
\varphi_{1}\big({W}^{J}\big)\Big) d\bar{\tilde{v}}_{n}^{J}
\wedge d\tilde{v}_{n}^{J}=0.\nonumber
\end{equation}
$\bullet$ \textbf{Prove that the terms
\eqref{e}--\eqref{f} are zero.} From the second and third formulae
of \eqref{17}, the last two terms \eqref{e} and \eqref{f} vanish.
\vspace{0.2cm} In the light of the above analysis, it is arrived at
\begin{equation}
\begin{aligned}
\sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n+1}^{J} \wedge {
d}\tilde{v}_{n+1}^{J}-\frac{1}{2} \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n+1}^{J} \wedge { d}
\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n+1}^{J}\big)=
\sum\limits_{J=1}^{3} { d}\bar{\tilde{x}}_{n}^{J} \wedge {
d}\tilde{v}_{n}^{J}-\frac{1}{2} \sum\limits_{J=1}^{3} {
d}\bar{\tilde{x}}_{n}^{J} \wedge { d}
\big(\tilde{\Lambda}^{J}\mathrm{i}\tilde{x}_{n}^{J}\big).
\end{aligned}\nonumber
\end{equation}
Therefore the method with the coefficients satisfying
\eqref{17} is symplectic. The proof is complete.
\end{proof}
\section{{Practical {symplectic methods} and numerical tests}}\label{sec:prac meth}
We are now ready to consider the construction of practical {symplectic methods}. In this section, we {shall} propose second-order and fourth-order continuous-stage symplectic adapted exponential methods based on the symplectic conditions \eqref{17} and {then rigorously study their error bounds and implementations. Finally, one numerical test is given to show the performance of the obtained methods.}
\subsection{Construction of the practical methods}\label{sec3.1}
In order to construct practical continuous-stage symplectic adapted exponential methods, we first present a class of
method \eqref{TRM} whose coefficients satisfy the symplectic conditions \eqref{17}. Then
symplectic methods {of} the form \eqref{CSAEI} are obtained since the transformed method \eqref{TRM} shares the same symplecticity with \eqref{CSAEI}.
\begin{mytheo}\label{thm: csxmthod
If the coefficients of \eqref{TRM} satisfy
\begin{align}
&\beta_{\tau}( W )=(1-\tau)\varphi_{1}\big((1-\tau)W\big),\qquad \gamma_{\tau}( W )=\varphi_{0}\big((1-\tau)W\big),\label{coebc}\\
& \bar{\alpha}_{\tau \sigma}(W)-\alpha_{\sigma\tau}(W)
=(\tau-\sigma)\varphi_{1}\big(-(\tau-\sigma) W\big),\label{coea}
\end{align}
then the method \eqref{TRM} is symplectic, {i.e.,} its coefficients satisfy the symplectic conditions \eqref{17}.
\end{mytheo}
\begin{proof}
By inserting \eqref{coebc} into the first formula of \eqref{17}, we have
\begin{equation*}
\begin{aligned}
&\gamma_{\tau}( W )-W\beta_{\tau}( W )
=\varphi_{0}\big((1-\tau)W\big)-W(1-\tau)\varphi_{1}\big((1-\tau)W\big)\\
=&e^{(1-\tau)W}-W(1-\tau)\big(e^{(1-\tau)W}-I\big)/\big((1-\tau)W\big)=I.
\end{aligned}
\end{equation*}
It follows from the second formula of \eqref{17} that
\begin{equation*}
\begin{aligned}
&\gamma_{\tau}( W )\big(\bar{\varphi}_{1}(W)-\tau\bar{\varphi}_{1}(\tau W)\big)-
\beta_{\tau}( W )\big(e^{-W}+W\bar{\varphi}_{1}(W)-\tau W\bar{\varphi}_{1}(\tau W)\big)\\
=&\varphi_{0}\big((1-\tau)W)\big)\big(\varphi_{1}(-W)-\tau\varphi_{1}(-\tau W)\big)-
(1-\tau)\varphi_{1}\big((1-\tau)W\big)\\
&\big(e^{-W}+W\varphi_{1}(-W)-\tau W\varphi_{1}(-\tau W)\big)\\
=& e^{(1-\tau)W}\big(e^{-\tau W}-e^{-W}\big)/W-e^{-\tau W}\big(e^{(1-\tau)W}-I\big)/W=0.\\
\end{aligned}
\end{equation*}
Furthermore, we substitute the coefficients \eqref{coebc} into the third formula of symplectic conditions \eqref{17} and compute the left hand side to get
\begin{equation*}
\begin{aligned}
&(1-\sigma)\varphi_{1}\big(-(1-\sigma) W\big)\varphi_{0}\big((1-\tau) W\big)
-\bar{\alpha}_{\tau \sigma}(W)\\
&-W(1-\sigma)\varphi_{1}\big(-(1-\sigma) W\big)(1-\tau)\varphi_{1}\big((1-\tau) W\big)/2\\
=&\big(-e^{(\sigma-\tau)W}+e^{(1-\tau)W}-e^{-(1-\sigma)W}+I\big)/(2W)-\bar{\alpha}_{\tau \sigma}(W).
\end{aligned}
\end{equation*}
Similarly, the right hand side of the third symplectic condition {in \eqref{17}} can be simplified as
\begin{equation*}
\begin{aligned}
&(1-\tau)\varphi_{1}\big((1-\tau) W\big)\varphi_{0}\big(-(1-\sigma) W\big)-\alpha_{\sigma\tau}(W)\\
&+W(1-\tau)\varphi_{1}\big((1-\tau) W\big)(1-\sigma)\varphi_{1}\big(-(1-\sigma) W\big)/2\\
=&\big(e^{(\sigma-\tau)W}-e^{-(1-\sigma)W}+e^{(1-\tau)W}-I\big)/(2W)-\alpha_{\sigma\tau}(W).
\end{aligned}
\end{equation*}
Combining the above two results, we get
\begin{equation*}
\begin{aligned}
\bar{\alpha}_{\tau \sigma}(W)-\alpha_{\sigma\tau}(W)
=\big(-e^{(\sigma-\tau)W}+I\big)/W
=(\tau-\sigma)\varphi_{1}\big(-(\tau-\sigma) W\big).
\end{aligned}
\end{equation*}
From the above analysis, {it is known} that the coefficients \eqref{coebc} and \eqref{coea} satisfy all the symplectic conditions \eqref{17}, which completes the proof of this theorem.
\end{proof}
As stated {in the beginning of this section}, once a
symplectic method determined by \eqref{TRM} with the coefficients \eqref{coebc} and \eqref{coea} is given, a
symplectic method \eqref{CSAEI} is derived immediately {whose coefficients are obtained by} replacing $W$ with $hM$ {in \eqref{coebc} and \eqref{coea}}.
In what follows, we present two practical continuous-stage symplectic adapted exponential methods \eqref{CSAEI} up to order four.
Actually, the result of $\beta_{\tau}( hM )$ and $\gamma_{\tau}( hM )$ is described by \eqref{coebc} of Theorem \ref{thm: csxmthod}. Therefore, we only need to make the choice of the coefficient
$\alpha_{\tau\sigma}(hM)$. The first case is given by the following algorithm which {will be proved to be of order two.}
\begin{algo}\label{algo3.2}
\textbf{(Second-order method)} Define a practical continuous-stage symplectic adapted exponential algorithm \eqref{CSAEI} with the following coefficients
\begin{equation}\label{coe2}
\begin{aligned}
&\alpha_{\tau\sigma}(hM)=\frac{\tau-\sigma}{2}\varphi_{1}((\tau-\sigma) hM),\ \
\beta_{\tau}( hM )=(1-\tau)\varphi_{1}((1-\tau)hM),\ \ \gamma_{\tau}( hM )=\varphi_{0}((1-\tau)hM).
\end{aligned}
\end{equation}
It is easily to check that the coefficient $\alpha_{\tau\sigma} $ satisfies \eqref{coea}, which means that
the method with the coefficients \eqref{coe2} is {symplectic.}
\end{algo}
In order to improve the accuracy, we consider the following choice of $\alpha_{\tau\sigma}(hM)$ which leads to a
fourth-order scheme.
\begin{algo}\label{algo3.3}
\textbf{(Fourth-order method)} If we take the following coefficients
\begin{equation}
\begin{aligned}\label{coe4}
&\alpha_{\tau\sigma}(hM)=\Big(\frac{1}{6}+\frac{\tau-\sigma}{2}\Big)\varphi_{1}((\tau-\sigma) hM),\\
& \beta_{\tau}( hM )=(1-\tau)\varphi_{1}((1-\tau)hM),\quad\gamma_{\tau}( hM )=\varphi_{0}((1-\tau)hM),
\end{aligned}
\end{equation}
then we get a continuous-stage symplectic adapted exponential method. Similarly, we can prove that the coefficients \eqref{coe4} satisfy all the symplectic conditions of {Theorem \ref{thm: csxmthod}.}
\end{algo}
{
\subsection{Convergence}
In this section, we shall study the convergence of the proposed two {algorithms}.
\begin{mytheo}\label{order condition} \textbf{(Convergence)}
It is assumed that $ F $ is locally Lipschitz-continuous with the Lipschitz constant $L$.
There {exist constants $h_0>0$ and $\widehat{C}>0$} independent of $\eps$, such that if the stepsize $h$ satisfies $ h \leq h_0$ and $ h \leq \widehat{C} \epsilon$, the global errors are estimated as
\begin{equation*}
\begin{aligned}&Algorithm\ \textmd{\ref{algo3.2}}:\ \ \norm{x(t_{n})-x_n}\leq C h^2,\ \ \ \ \ \ \
\norm{v(t_{n})-v_n}\leq C h^2/\epsilon,\\
&Algorithm\ \textmd{\ref{algo3.3}}:\ \ \norm{x(t_{n})-x_n}\leq C h^4/\eps^2,\ \ \
\norm{v(t_{n})-v_n}\leq C h^4/\epsilon^3,
\end{aligned}
\end{equation*}
for $nh\leq T$, where
$C>0$ is a generic constant independent of $\epsilon$ or $h$ or $n$ but depends on $\widehat{C}, L, T$ and $\norm{ \frac{d^s}{dx^s}F(x) }$ with $s=1,2$ for Algorithm \ref{algo3.2} and $s=1,2,3,4$ for Algorithm \ref{algo3.3}.
\end{mytheo}
\begin{proof}
$\bullet$ \textbf{Local errors.}
Local errors of the {method \eqref{CSAEI} are defined}
by inserting the exact solution
\eqref{VOC} into \eqref{CSAEI}, which leads to
\begin{equation*}
\begin{aligned}\label{PCSEEP1}
&x(t_{n}+\tau h)=x(t_{n})+\tau h\varphi_{1}(\tau hM )v(t_{n})+h^{2}\int_{0}^{1}
\alpha_{\tau \sigma}( hM )
\hat{F}(t_{n}+\sigma h)d \sigma +\triangle^x_{\tau},\\
&x(t_{n+1})=x(t_{n})+h\varphi_{1}( hM )v(t_{n})+h^{2}\int_{0}^{1}
\beta_{\tau}( hM )
\hat{F}(t_{n}+\tau h)d \tau+\delta^x_{n+1},\\
&v(t_{n+1})=\varphi_{0}( hM )v(t_{n})+h\int_{0}^{1}\gamma_{\tau}( hM )
\hat{F}(t_{n}+\tau h)d\tau+\delta^v_{n+1},
\end{aligned}
\end{equation*}
where $ \hat{F}(t):= F(x(t))$ and $ \triangle^x_{\tau}$, $ \delta^x_{n+1}$, $\delta^v_{n+1} $ are the discrepancies.
Combining with the variation-of-constants formula \eqref{VOC} and using Taylor series, we obtain
\begin{equation*}
\begin{aligned}
\triangle^x_{\tau}
&= {\tau}^2 h^2\int_{0}^{1}(1-z)\varphi_{1}(\tau(1-z) hM )\hat{F}(t_{n}+h \tau z)dz
-h^{2}\int_{0}^{1}\alpha_{\tau\sigma}( hM )\hat{F}(t_{n}+\sigma h)d\sigma\\
&=\sum\limits_{j=0}^{r-3}h^{j+2}\bigg({\tau}^{j+2}\int_{0}^{1}(1-\sigma)\varphi_{1}
(\tau(1-\sigma) hM )\frac{{\sigma}^{j}}{j!}d\sigma
-\int_{0}^{1}\alpha_{\tau\sigma}( hM )\frac{{\sigma}^{j}}{j!}d \sigma\bigg)\hat{F}^{j}(t_{n})+\mathcal{O}(h^{r}/\epsilon^{r-3})\\
&=\sum\limits_{j=0}^{r-3}h^{j+2}\bigg({\tau}^{j+2}\varphi_{j+2}(\tau hM )
-\int_{0}^{1}\alpha_{\tau\sigma}( hM )\frac{{\sigma}^{j}}{j!}d \sigma\bigg)\hat{F}^{j}(t_{n})+\mathcal{O}(h^{r}/\epsilon^{r-3}),
\end{aligned}
\end{equation*}
and similarly
\begin{equation*}
\begin{aligned}
\rho_{n+1}=&\sum\limits_{j=0}^{r-2}h^{j+2}\bigg(\varphi_{j+2}( hM )
-\int_{0}^{1}\beta_{\tau}( hM )\frac{{\tau}^{j}}{j!}d \tau\bigg)\hat{F}^{j}(t_{n})
+{\mathcal{O}(h^{r+1}/\epsilon^{r-2})},\\
{\rho}^{'}_{n+1}=&\sum\limits_{j=0}^{r-1}h^{j+1}\bigg(\varphi_{j+1}( hM )
-\int_{0}^{1}\gamma_{\tau}( hM )\frac{{\tau}^{j}}{j!}d \tau\bigg)\hat{F}^{j}(t_{n})
+{\mathcal{O}(h^{r+1}/\epsilon^{r-1})}.
\end{aligned}
\end{equation*}
where $ \hat{F}^{j}(t) $ denotes the $j$th order derivative of $ F(x(t)) $ with respect to $ t $ and $r$ is a positive integer which has different value for different algorithm.
Based on the coefficient functions chosen in Algorithms \ref{algo3.2}-\ref{algo3.3}, we have the following results for the discrepancies stated above
\begin{equation}\label{local error}
\begin{aligned}&Algorithm\ \textmd{\ref{algo3.2}}:\ \ r=2,\ \ \norm{\triangle^x_{\tau}}\leq Ch^{2},
\ \ \ \ \
\norm{\delta^x_{n+1}}\leq Ch^{3}, \ \ \ \ \ \ \norm{\delta^v_{n+1}}\leq Ch^{3}/\eps,\\
&Algorithm\ \textmd{\ref{algo3.3}}:\ \ r=4,\ \ \norm{\triangle^x_{\tau}}\leq Ch^{4}/\eps,
\ \
\norm{\delta^x_{n+1}}\leq Ch^{5}/\eps^2, \ \ \norm{\delta^v_{n+1}}\leq Ch^{5}/\eps^3.
\end{aligned}
\end{equation}
$\bullet$ \textbf{Global errors.} Denote the global errors of \eqref{CSAEI} by
\begin{equation*}
e_{n}^{x}=x(t_{n})-x_{n},\ \ \ e_{n}^{v}=v(t_{n})-v_{n},
\ \ \ E_{\tau}=x(t_{n}+\tau h)-X_{\tau},
\end{equation*}
and the error system is
\begin{equation}
\begin{aligned}\label{error}
&E_{\tau}=e_{n}^{x}+\tau h \varphi_1(\tau hM )e_{n}^{v}
+h^2\int_{0}^{1}\alpha_{\tau\sigma}( hM )\big(F(x(t_{n}+\sigma h))-F(X_{\sigma})\big)d\sigma+\Delta^x_{\tau},\\
&e_{n+1}^{x}=e_{n}^{x}+h \varphi_{1}( hM ) e_{n}^{v}
+h^{2}\int_{0}^{1}\beta_{\tau}( hM )\big(F(x(t_{n}+\tau h))-F(X_{\tau})\big)d\tau+\delta^x_{n+1},\\
&e_{n+1}^{v}=\varphi_{0}( hM ) e_{n}^{v}
+h\int_{0}^{1}\gamma_{\tau}( hM )\big(F(x(t_{n}+\tau h))-F(X_{\tau})\big)d\tau+\delta^v_{n+1},
\end{aligned}
\end{equation}
where the initial conditions are $ e_{0}^{x}=0 $, $ e_{0}^{v}=0 $.
With the uniform bound of the coefficients, the first equation of \eqref{error} is bounded by
\begin{equation*}
\norm{E_{\tau}} \leq \norm{e_{n }^{x}}+ \tau h \norm{e_{n }^{v}}+h^2 C L\norm{E_{\tau}}_{c}+{\norm{\Delta^x_{\tau}}_{c}},
\end{equation*}
which yields
\begin{equation*}
\norm{E_{\tau}}_{c} \leq \norm{e_{n }^{x}}+h \norm{e_{n }^{v}}+h^2 {C} L \norm{E_{\tau}}_{c}+\norm{\Delta^x_{\tau}}_{c},
\end{equation*}
where $\|\cdot\|_{c}$ denotes the maximum norm
$
\norm{E_{\tau}}_{c}=\max \limits_{\tau\in[0,1]} \norm{E_{\tau}}
$
for a continuous function $E_{\tau}$ on [0,1].
Under the condition that
$ h \leq \sqrt{\frac{1}{2 C L}}$ is satisfied, we have
\begin{equation*}\label{con6}
\norm{E_{\tau}}_{c} \leq 2\big(\norm{e_{n }^{x}}+h \norm{e_{n }^{v}}\big)+2\norm{\Delta_{\tau}^{x}}_{c}.
\end{equation*}
From this result and the last two equations of \eqref{error}, it is deduced that
\begin{equation*}
\begin{aligned}
\norm{e_{n+1}^{x}}
&\leq \norm{e_{n}^{x}}+h\norm{e_{n}^{v}}+Ch^2\big(\norm{e_{n}^{x}}+h\norm{e_{n}^{v}}+\norm{\Delta_{\tau}^{x}}_{c}\big)
+\norm{\delta^x_{n+1}},\\
\norm{e_{n+1}^{v}}
&\leq \norm{e_{n}^{v}}+Ch\big(\norm{e_{n}^{x}}+h\norm{e_{n}^{v}}+\norm{\Delta_{\tau}^{x}}_{c}\big)
+\norm{\delta^v_{n+1}}.
\end{aligned}
\end{equation*}
These {results, the local errors derived in \eqref{local error}} and Gronwall inequality immediately lead to the statement of this theorem, which completes the proof.
\end{proof}}
\subsection{{Implementation issues}}\label{sec:four}
It is noted that Algorithms \ref{algo3.2}--\ref{algo3.3} fail to be practical unless the integrals appearing in \eqref{CSAEI} are computed exactly or by using some numerical quadrature formulae. For most cases, those integrals cannot be solved exactly and usually quadrature formulae are needed in practical computations. In this section, we pay attention to this point and discuss the implementations of the obtained algorithms.
Let us start with the implementation of Algorithm \ref{algo3.2}.
In the computation, we consider the four-point Gaussian quadrature with the weights $b_{i}$ and abscissae $c_{i}$
and then derive the scheme as follows
\begin{equation*}
\begin{aligned}
&X_{c_{i}}=x_{n}+ c_{i} h\varphi_1(c_{i}hM) v_{n}
+h^2\sum_{j=1}^{4}\frac{b_{j}(c_{i}-c_{j})}{2}\varphi_1((c_{i}-c_{j})hM)F (X_{c_{j}}),\ \ i=1,2,\ldots,4,\\
&x_{n+1}=x_{n}+ h\varphi_1(hM) v_{n}+h^2\sum_{i=1}^{4}b_{i}(1-c_{i})\varphi_1((1-c_{i})hM)F (X_{c_{i}}),\\
&v_{n+1}=\varphi_0(hM)v_{n}+h\sum_{i=1}^{4}b_{i}\varphi_0((1-c_{i})hM)F (X_{c_{i}}),
\end{aligned}
\end{equation*}
where
\begin{equation*}
\begin{array}[c]{ll}
&\breve{a}=\frac{2\sqrt{30}}{35}, \qquad \ \ \ \ \ \breve{b}=\frac{\sqrt{30}}{36},\qquad \ \ \ \ \ \ b_{1}=b_{4}=\frac{1/2-\breve{b}}{2},\ \ \ \ \ \ b_{2}=b_{3}=\frac{1/2+\breve{b}}{2},\\
&c_{1}=\frac{1+\sqrt{3/7+\breve{a}}}{2},\ \ \ c_{2}=\frac{1+\sqrt{3/7-\breve{a}}}{2}, \ \ \
c_{3}=\frac{1-\sqrt{3/7-\breve{a}}}{2}, \ \ \ \ \ \ \ c_{4}=\frac{1-\sqrt{3/7+\breve{a}}}{2}.\\
\end{array}
\end{equation*}
We shall refer to this algorithm by SC1O2.
Obviously, this scheme is implicit and to get an explicit one, we consider the one-point Gaussian quadrature with
the weight $b_{1}=1$ and abscissa $c_{1}=\frac{1}{2}$. This yields
\begin{equation*}
\begin{aligned}\label{vv}
&X_{c_{1}}=x_{n}+ \frac{h}{2}\varphi_1(hM/2) v_{n},\ \ x_{n+1}=x_{n}+ h\varphi_1(hM) v_{n}+\frac{h^2}{2}
\varphi_1(hM/2)F (X_{c_{1}}),\\ &v_{n+1}=\varphi_0(hM)v_{n}+h\varphi_{0}(hM/2)F (X_{c_{1}}).
\end{aligned}
\end{equation*}
From this formulation, it is clear that this is
an explicit scheme and we
denoted it by SC2O2.
For the implementation of Algorithm \ref{algo3.3}, the same quadrature as used in SC1O2 is chosen here and the corresponding algorithm is referred as SC1O4. {To get the explicit scheme, we consider} the following scheme
\begin{equation*}
\begin{aligned}
&X_{c_{i}}=x_{n}+ c_ih\varphi_1(c_{i} hM) v_{n}
+h^2 \sum\limits_{j=1}^{i-1}a_{ij}(c_i-c_j)h\varphi_1((c_i-c_j)hM)F (X_j),\ \ i=1,2,3,\\
&x_{n+1}=x_{n}+ h\varphi_1(hM) v_{n}
+h^2\sum\limits_{i=1}^{3}b_i(1-c_i)\varphi_1((1-c_i)hM)F (X_i),\\
&v_{n+1}=\varphi_0(hM)v_{n}
+h\sum\limits_{i=1}^{3}b_i\varphi_0((1-c_i)hM)F (X_i),
\end{aligned}
\end{equation*}
where
\begin{equation*}
\begin{array}[c]{ll}
&a_{21}=\frac{4+2\sqrt[3]{2}+\sqrt[3]{4}}{6}, \qquad a_{31}=b_{1}=b_3=a_{21},\qquad a_{32}=\frac{-1-2\sqrt[3]{2}-\sqrt[3]{4}}{3},\\
&b_{2}=\frac{-1-2\sqrt[3]{2}-\sqrt[3]{4}}{3},\ \ \ \ \ \ c_{1}=b_{1}/2, \ \ \ c_{2}=\frac{1}{2}, \qquad c_{3}=1-c_1.\\
\end{array}
\end{equation*}
We shall call this method as SC2O4.
\subsection{Numerical test}\label{sec: experiment}
One numerical test is shown in this section to test the efficiency of the new methods
compared with some existing methods in the literature.
The methods for comparison are chosen as follows:
\begin{itemize}
\item BORIS: the Boris method of order two presented in \cite{Boris};
\item RKO2: a symplectic Runge-Kutta method of order two (implicit
midpoint rule) presented in \cite{Hairer2002};
\item SC1O2: the continuous-stage symplectic adapted exponential method of order two derived in Section \ref{sec:four};
\item SC2O2: the explicit continuous-stage symplectic adapted exponential method of order two derived in Section \ref{sec:four};
\item RKO4: a symplectic Runge-Kutta method of order four presented in
\cite{Sanz-Serna91};
\item SC1O4: the continuous-stage symplectic adapted exponential method of order four derived in Section \ref{sec:four};
\item SC2O4: the explicit continuous-stage symplectic adapted exponential method of order four derived in Section \ref{sec:four}.
\end{itemize}
For implicit methods, we choose fixed-point iteration and set $ 10^{-16} $ as the
error tolerance and 5 as the maximum number of each iteration.
To test the performance of all the methods, we compute the global errors: $error:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ and the energy error
$e_{H}:=\frac{|H(x_{n},v_n)-H(x_0,v_0)|}{|H(x_0,v_0)|}$ in {the} numerical experiment.
\noindent\vskip3mm \noindent\textbf{Problem 1. (Homogeneous magnetic field)}
This problem is devoted to the charged-particle system \eqref{charged-particle sts-cons} of \cite{Lubich2017} with an additional factor $1/\epsilon$ and a homogeneous magnetic field.
We take the scalar potential $U(x)=\frac{1}{100\sqrt{x_{1}^2+x_{2}^2}}$, the homogeneous magnetic field
$\frac{1}{\epsilon}B=\frac{1}{\epsilon}(0,0,1)^{\intercal}$ and the initial conditions as $x(0)=(0,0.2,0.1)^{\intercal}, v(0)=(0.09, 0.05, 0.2)^{\intercal}.$
\textbf{Order behaviour.} This system is integrated on $[0,1]$ with different $\epsilon$ and step sizes $h$ to
show the global {errors $error$} in Fig. \ref{fig:problem11}. According to the result, it can be seen that the global error lines of SC1O2 and SC2O2 are nearly parallel to the line with slope 2, which verifies that these two methods have second-order accuracy.
Similarly, the fourth-order accuracy can be observed for SC1O4 and SC2O4. Moreover, we can see that the global errors of our methods are smaller than the other three methods especially with small $\epsilon$. This demonstrates that our methods have better accuracy than others when solving CPD in a strong magnetic field.
\textbf{Uniform errors.} To display the influence of $\eps$ on the global errors,
we show the error $error2:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ for the second order methods in Fig. \ref{fig:problem11new1} and $error4:=\frac{\eps^2\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps^3\norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ for the fourth order ones in Fig. \ref{fig:problem11new2}. It can be observed from the results that our methods demonstrate the convergence stated in Theorem \ref{order condition} and behave much better than the others.
\textbf{Energy conservation.} Then we solve the problem
with $h=\frac{1}{100}$ on the interval $ [0,1000]$. Fig. \ref{fig:problem12} displays the results of energy conservation {$err_H$}.
It can be observed that for the normal magnetic field, all the methods have good energy conservation over long times and they behave similarly. For the case that $\eps$ is small, the energy error of those two Runge-Kutta methods (RKO2 and RKO4) increases slightly when $t$ goes large while the other methods have a good performance. Moreover, our methods display better accuracy in the energy preservation than the Boris method.
\textbf{Long-time behavior.} Finally, the problem is solved on $[0, 1000]$ with
$h=\frac{1}{2}$ and $\eps=0.1$. Fig. \ref{fig:problem11new3} presents the trajectory for the second order methods (BORIS, RKO2, SC1O2 and SC2O2) in [x y z] space and the results show that the methods SC1O2 and SC2O2 perform uniformly better than the Boris and RKO2 methods.
Meanwhile, we have noticed that the RKO2 method performs worse than expected. The reason
is that for the RKO2 method when solving CPD with small $\epsilon$, there is a
very strict requirement of the stepsize and the method usually does
not behave well.
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.45cm]{e11} & \includegraphics[width=4.7cm,height=4.45cm]{e12}&\includegraphics[width=4.7cm,height=4.45cm]{e13}
\end{tabular}
\caption{Problem 1. The global errors $error:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem11}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=3.8cm,height=4.45cm]{1Boris} &\includegraphics[width=3.8cm,height=4.45cm]{1RKO2} & \includegraphics[width=3.8cm,height=4.45cm]{1SC1O2} & \includegraphics[width=3.8cm,height=4.45cm]{1SC2O2}
\end{tabular}
\caption{Problem 1. The errors {$error2:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$} of {second order methods (BORIS, RKO2, SC1O2 and SC2O2)} with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem11new1}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=4.7cm,height=4.45cm]{1RKO4} &\includegraphics[width=4.7cm,height=4.45cm]{1SC1O4} & \includegraphics[width=4.7cm,height=4.45cm]{1SC2O4}
\end{tabular}
\caption{Problem 1. The errors {$error4:=\frac{\eps^2\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps^3\norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$} of {fourth order methods (RKO4, SC1O4 and SC2O4)} with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem11new2}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.45cm]{h11} & \includegraphics[width=4.7cm,height=4.45cm]{h12} & \includegraphics[width=4.7cm,height=4.45cm]{h13}
\end{tabular}
\caption{Problem 1. Evolution of the energy error $e_{H}:=\frac{|H(x_{n},v_n)-H(x_0,v_0)|}{|H(x_0,v_0)|}$ as function of time $t=nh$. }
\label{fig:problem12}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=3.8cm,height=4.45cm]{Boris-1tra3} &\includegraphics[width=3.8cm,height=4.45cm]{RKO2-1tra3} & \includegraphics[width=3.8cm,height=4.45cm]{SC1O2-1tra3} & \includegraphics[width=3.8cm,height=4.45cm]{SC2O2-1tra3}
\end{tabular}
\caption{Problem 1. The trajectory of second order methods (BORIS, RKO2, SC1O2 and SC2O2) in [x y z] space with $t=1000$, $h=1/2$ and $\epsilon=0.1$. }
\label{fig:problem11new3}
\end{figure}
\section{Extension to CPD with non-homogeneous magnetic field}\label{sec:general}
In this section, we consider {the extension of the above symplectic methods and corresponding error estimates} to the following CPD with a non-homogeneous magnetic field $B(x)$:
\begin{equation}\label{charged-particle}
\begin{array}[c]{ll}
\ddot{x}=\dot{x} \times \frac{B(x)}{\epsilon} +F(x), \quad
x(0)=x_0,\quad \dot{x}(0)=\dot{x}_0.
\end{array}
\end{equation}
We remark that the above methods cannot be directly applied and some novel modifications are needed in the construction of the new methods.
\subsection{Methods and their convergence}
\begin{algo} \label{alg:EPSX} \textbf{(Adapted exponential methods)}
For solving the CPD \eqref{charged-particle} in a non-homogeneous magnetic field $B (x)$, define the following continuous-stage
adapted exponential method (denoted by SG1O1)
{\begin{equation}\label{SG1O1}
\begin{aligned}
&X_{\tau}=x_{n}+ \tau h\varphi_{1}(\tau hM_n) v_{n}+
h^2 \int_{0}^{1}\frac{\tau-\sigma}{2}\varphi_{1}((\tau-\sigma) hM_n)F (X_\sigma)d\sigma,\\
&x_{n+1}=x_{n}+ h\varphi_1(hM_n)
v_{n}+h^2 \int_{0}^{1}(1-\tau)\varphi_{1}((1-\tau)hM_n)F (X_{\tau})d\tau,\\
&v_{n+1}=\varphi_0(hM_n)v_{n}+h
\int_{0}^{1}\varphi_{0}((1-\tau)hM_n)F(X_{\tau})d\tau,
\end{aligned}
\end{equation} with $M_n=\frac{1}{\epsilon}\tilde{B}(x_{n})$.
The second scheme is obtained by considering another approximation of $\tilde{B}$ which leads to}
\begin{equation}\label{SG1O2}
\begin{aligned}
&X_{\tau}=x_{n}+ \tau h\varphi_{1}(\tau h\tilde{M}_n) v_{n}+
h^2 \int_{0}^{1}\frac{\tau-\sigma}{2}\varphi_{1}((\tau-\sigma) h\tilde{M}_n)F (X_\sigma)d\sigma,\\
&x_{n+1}=x_{n}+ h\varphi_1(h\tilde{M}_n)
v_{n}+h^2 \int_{0}^{1}(1-\tau)\varphi_{1}((1-\tau)h\tilde{M}_n)F (X_{\tau})d\tau,\\
&v_{n+1}=\varphi_0(h\tilde{M}_n)v_{n}+h
\int_{0}^{1}\varphi_{0}((1-\tau)h\tilde{M}_n)F(X_{\tau})d\tau,\\
\end{aligned}
\end{equation}
where $\tilde{M}_n=\frac{1}{\epsilon}\tilde{B}(\frac{x_{n}+x_{n+1}}{2})$. We shall call it as SG1O2.
The third method is formulated as follows. Considering a Triple Jump splitting method \cite{Hairer2002} and denoting SG1O2 by $\Upsilon_{h}$, then we get a splitting scheme:
$ \Phi_h=\Upsilon_{\kappa_1h}\circ \Upsilon_{\kappa_2h}\circ\Upsilon_{\kappa_3h}$,
where $\kappa_1=\kappa_3=\frac{1}{2-\sqrt[3]{2}}$ and $\kappa_2=-\frac{\sqrt[3]{2}}{2-\sqrt[3]{2}}$. We denote this method by SG1O4.
\end{algo}
The following theorem states the convergence of the above three methods.
\begin{mytheo}\label{order condition-B(x)} \textbf{(Convergence)}
For the CPD
\eqref{charged-particle}, we assume that its solution is sufficiently smooth, and the function $ F$ is locally Lipschitz-continuous and
sufficient differentiable.
Under these conditions and $ h \leq C \epsilon$, the global errors of SG1O1, SG1O2 and SG1O4
are respectively given by
\begin{subequations}\label{RR}
\begin{align}&\textmd{SG1O1}:\quad\|x(t_{n+1})-x_{n+1}\|\leq C h,\qquad \qquad \|v(t_{n+1})-v_{n+1}\| \leq C h/\epsilon,\label{R0}\\
&\textmd{SG1O2}:\quad\|x(t_{n+1})-x_{n+1}\|\leq C h^2/\epsilon,\qquad \ \ \|v(t_{n+1})-v_{n+1}\| \leq C h^2/\epsilon^2,\label{R1}\\
&\textmd{SG1O4}:\quad\|x(t_{n+1})-x_{n+1}\|\leq C h^4/\epsilon^3,\qquad \ \|v(t_{n+1})-v_{n+1}\|\leq C h^4/\epsilon^4, \label{R2} \end{align}
\end{subequations} where $0<n<T/h$ and the
generic constant $C>0$ is independent of $\epsilon,\ h,\ n$.
\end{mytheo}
\begin{proof} {The proof is firstly given for the errors of SG1O2 and then the results of the other methods can be derived with
similar procedure.}
{ $\bullet$ \textbf{Proof of \eqref{R1}.}}
\textbf{Linearized system.} The first step is to rewrite the system \eqref{charged-particle} as
\begin{equation}\label{charged-particle:B(x)}
\begin{aligned}
\dot{x}=v,\ \ \dot{v}=v \times \frac{B(x)}{\epsilon} +F(x),\ \
x(0)=x_0,\ \ v(0)=v_0.
\end{aligned}
\end{equation}
For some $t=t_n+s$ with $n\geq 0$, we consider the following linearized system of \eqref{charged-particle:B(x)}
\begin{equation}\label{linearized system}
\begin{aligned}
\dot{\tilde{x}}_n(s)=\dot{\tilde{v}}_n(s),\ \
\dot{\tilde{v}}_n(s)=\frac{1}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})\tilde{v}_n(s) +F({\tilde{x}}_n(s)),\ \
{\tilde{x}}_n(0)=x(t_n),\ \ {\tilde{v}}_n(0)=v(t_n),\ \ \ 0<s\leq h,
\end{aligned}
\end{equation}
where {$x_{n+\frac{1}{2}}
\frac{x (t_{n})+x (t_{n+1})}{2} $}.
\textbf{Local errors between the original and linearized systems.}
Then we denote the local errors between the above two systems by
$$\xi^{x}_{n}(s)=x(t_{n}+s)-\tilde{x}_{n}(s), \ \ \xi^{v}_{n}(s)=v(t_{n}+s)-\tilde{v}_{n}(s),\ \ \ \ 0\leq n <\frac{T}{h}.$$
It is easily obtained from \eqref{charged-particle:B(x)} and \eqref{linearized system} that
\begin{equation}\label{er1}
\begin{aligned}
&\dot{\xi}^{x}_{n}(s)=\xi^{v}_{n}(s),\ \
\dot{\xi}^{v}_{n}(s)=\frac{1}{\epsilon} \tilde{B}(x_{n+\frac{1}{2}})\xi^{v}_{n}(s) +F(x(t_n+s))-F(\tilde{x}_{n}(s))
+\varsigma^{0}_{n}(s).
\end{aligned}
\end{equation}
Here we denote $\varsigma^{0}_{n}(s)=\frac{1}{\epsilon}\big(\tilde{B}\big(x(t_{n}+s)\big)-\tilde{B}\big(x_{n+\frac{1}{2}}\big)\big)v(t_{n}+s)$,
and using the Taylor expansion yields
\begin{equation*}
\begin{aligned}
&\tilde{B}(x(t_{n}+s))-\tilde{B}\big(x_{n+\frac{1}{2}}\big)
=\tilde{B}(x(t_n))+\frac{d\tilde{B}}{dx}(\dot{x}(t_n)s+\mathcal{O}(s^2/\epsilon))+{\mathcal{O}(s^3/\epsilon)}
-\tilde{B}(x(t_n))\\
&\ \ -\frac{d\tilde{B}}{dx}(\dot{x}(t_n)h/2+\mathcal{O}(h^2/\epsilon))-\mathcal{O}(h^2/\epsilon)
=\tilde{B}^{(1)}_n(s-h/2)+\mathcal{O}(h^2/\epsilon),
\end{aligned}
\end{equation*}
where $\tilde{B}^{(i)}_n=\frac{d^i}{dt^i}\tilde{B}(x(t))\mid_{t=t_n},\ \ i=1,2$.
Likewise, we get
$$v(t_n+s)=v(t_n)+\dot{v}(t_n)s+\mathcal{O}(s^2/\epsilon^2).$$
Combining the above two formulae, one arrives at
\begin{equation}\label{err0}
\begin{aligned}
\varsigma^{0}_{n}(s)= \frac{1}{\epsilon}\tilde{B}_{n}^{(1)}(s-h/2)v(t_n)+\frac{1}{\epsilon}\tilde{B}_{n}^{(1)} s (s-h/2)\dot{v}(t_n)+\mathcal{O}(h^2/\epsilon^2).\\
\end{aligned}
\end{equation}
Then taking the variation-of-constant formula of \eqref{er1} into account, we have that
\begin{equation*}
\begin{aligned}
&\xi^{x}_{n}(h)=\int_{0}^{h}\xi^{v}_{n}(s)ds,\ \
\xi^{v}_{n}(h)=\int_{0}^{h}e^{\frac{h-s}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}
\bigg(\int_{0}^{1}\nabla F(x(t_n+s)+(\sigma-1)\xi^{x}_{n}(s))\xi^{x}_{n}(s)d\sigma+\varsigma^{0}_{n}(s)\bigg)ds,
\end{aligned}
\end{equation*}
which further gives
\begin{equation}\label{xi}
\begin{aligned}
&\xi^{x}_{n}(h)=\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}
\int_{0}^{1}\nabla F(x(t_n+\delta)+(\sigma-1)\xi^{x}_{n}(\delta))\xi^{x}_{n}(s)d\sigma d\delta ds\\
&\ \ \ \ \qquad+\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}\varsigma^{0}_{n}(\delta)d\delta ds.\\
\end{aligned}
\end{equation}
Inserting the formula \eqref{err0} into the second part of \eqref{xi}, we get
\begin{equation*}
\begin{aligned}
&\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}\varsigma^{0}_{n}(\delta)d\delta ds
= \underbrace{\frac{1}{\epsilon}\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}B_{n}^{(1)}
(\delta-h/2)v(t_n)d\delta ds}_{\textmd{part}\ I}\\
& \quad +\underbrace{\frac{1}{\epsilon}\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}B_{n}^{(1)}
\delta(\delta-h/2)\dot{v}(t_n) d\delta ds}_{\textmd{part}\ II}+\mathcal{O}(h^4/\epsilon^2).
\end{aligned}
\end{equation*}
With some calculations, it is easy to have that
$$\textmd{part}\ I=\mathcal{O}(h^3/\epsilon),\quad\textmd{part}\ II=\mathcal{O}(h^5/\epsilon^2).$$
Considering further that $ h \leq C \epsilon$, we get
$
\int_{0}^{h}\int_{0}^{s}e^{\frac{s-\delta}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}\varsigma^{0}_{n}(\delta)d\delta ds=\mathcal{O}(h^3/\epsilon).$
Hence, the local error $\xi^{x}_{n}(h)$ is bounded by
\begin{equation}\label{xix}
\xi^{x}_{n}(h)=\mathcal{O}(h^3/\epsilon).
\end{equation}
In the same way, it is arrived that
\begin{equation*}
\begin{aligned}
&\int_{0}^{h}e^{\frac{h-s}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}\varsigma^{0}_{n}(s) ds
=\frac{1}{\epsilon}\int_{0}^{h}e^{\frac{h-s}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}v(t_n)
(s-h/2) B_{n}^{(1)}ds\\
&\quad+\frac{1}{\epsilon}\int_{0}^{h}e^{\frac{h-s}{\epsilon}\tilde{B}(x_{n+\frac{1}{2}})}\dot{v}(t_n)
s(s-h/2) B_{n}^{(1)}ds+\mathcal{O}(h^3/\epsilon^2)=\mathcal{O}(h^3/\epsilon^2),
\end{aligned}
\end{equation*}
and based on which, we thus have
\begin{equation}\label{xiv}
\xi^{v}_{n}(h)=\mathcal{O}(h^3/\epsilon^2).
\end{equation
\textbf{Global errors.}
With the above analysis, we shall estimate the global errors of the method \eqref{SG1O2}, i.e.,
$$e^{x}_{n+1}:=x(t_{n+1})-x_{n+1},\ \ \ e^{v}_{n+1}:=v(t_{n+1})-v_{n+1}.$$
These errors can be split into two parts
\begin{equation}\label{Global error}
e^{x}_{n+1}=\tilde{e}^{x}_{n}+\xi^{x}_{n}(h),\ \ \ e^{v}_{n+1}=\tilde{e}^{v}_{n}+\xi^{v}_{n}(h),
\end{equation}
and then converted to evaluate $$\tilde{e}^{x}_{n}:=\tilde{x}_{n}(h)-x_{n+1},\ \ \ \tilde{e}^{v}_{n}:=\tilde{v}_{n}(h)-v_{n+1}.$$
On the basis of the method \eqref{SG1O2}, we present the local error $\varphi_n^x$ and $\varphi_n^v$ as follows
\begin{equation}\label{truncation error}
\begin{aligned}
&\tilde{x}_{n}(\tau h)=\tilde{x}_{n}(0)+ \tau h\varphi_{1}\big(\tau h\tilde{P}\big)\tilde{v}_{n}(0)+
h^2 \int_{0}^{1}\frac{\tau-\sigma}{2}\varphi_{1}\big((\tau-\sigma) h\tilde{P}\big)
F (\tilde{x}_{n}(\sigma h))d\sigma+\varphi_n^\tau,\\
&\tilde{x}_{n}(h)=\tilde{x}_{n}(0)+ h\varphi_{1}\big(h\tilde{P}\big)
\tilde{v}_{n}(0)+h^2 \int_{0}^{1}(1-\tau)\varphi_{1}\big((1-\tau)h \tilde{P}\big)
F (\tilde{x}_{n}(\tau h))d\tau+\varphi_n^x,\\
&\tilde{v}_{n}(h)=\varphi_{0}\big( h\tilde{P}\big)\tilde{v}_{n}(0)+h
\int_{0}^{1}\varphi_{0}\big((1-\tau)h\tilde{P}\big)F(\tilde{x}_{n}(\tau h))d\tau+\varphi_n^v,\\
\end{aligned}
\end{equation}
where $\tilde{P}=\frac{1}{\epsilon}\tilde{B}\Big(\frac{\tilde{x}_{n}(0)+\tilde{x}_{n}(h)}{2}\Big)$.
It follows from the variation-of-constant formula of the linearized system \eqref{linearized system} that
\begin{equation}\label{VOC of linearized system}
\begin{aligned}
&\tilde{x}_{n}(\tau h)=\tilde{x}_{n}(0)+ \tau h\varphi_{1}(\tau hP)
\tilde{v}_{n}(0)+h^2 \int_{0}^{\tau}(\tau-\sigma)\varphi_{1}((\tau-\sigma)h P)
F (\tilde{x}_{n}(\sigma h))d\sigma,\\
&\tilde{x}_{n}(h)=\tilde{x}_{n}(0)+ h\varphi_{1}(hP)
\tilde{v}_{n}(0)+h^2 \int_{0}^{1}(1-\tau)\varphi_{1}((1-\tau)h P)
F (\tilde{x}_{n}(\tau h))d\tau,\\
&\tilde{v}_{n}(h)=\varphi_{0}( hP)\tilde{v}_{n}(0)+h
\int_{0}^{1}\varphi_{0}((1-\tau)hP)F(\tilde{x}_{n}(\tau h))d\tau,\\
\end{aligned}
\end{equation}
where $P=\frac{1}{\epsilon}\tilde{B}\Big(\frac{x(t_{n})+x(t_{n+1})}{2}\Big)$.
Considering the difference between \eqref{truncation error} and \eqref{VOC of linearized system}, we get
\begin{equation*}\label{}
\begin{aligned}
\varphi_n^\ta
&=\tau h\big(\varphi_{1}(\tau hP)-\varphi_{1}(\tau h\tilde{P})\big){v}(t_{n})
+\tau^2 h^2 \int_{0}^{1}(1-z)\varphi_{1}\big(\tau(1-z)h P\big)F \big(\tilde{x}_{n}( h \tau z)\big)dz\\
&\quad-h^2 \int_{0}^{1}\frac{(\tau-\sigma)}{2}\varphi_{1}\big((\tau-\sigma)h \tilde{P}\big)F \big(\tilde{x}_{n}(\sigma h)\big)d\sigma,\\
\varphi_n^x&=h\big(\varphi_{1}(hP)-\varphi_{1}(h\tilde{P})\big){v}(t_{n})
+h^2 \int_{0}^{1}(1-\tau)\big(\varphi_{1}\big((1-\tau)h P\big)-\varphi_{1}\big((1-\tau)h \tilde{P}\big)\big)
F \big(\tilde{x}_{n}(\tau h)\big)d\tau,\\
\varphi_n^v&=\big(\varphi_{0}(hP)-\varphi_{0}(h\tilde{P})\big){v}(t_{n})+h \int_{0}^{1}\big(\varphi_{0}\big((1-\tau)h P\big)-\varphi_{0}\big((1-\tau)h \tilde{P}\big)\big)
F \big(\tilde{x}_{n}(\tau h)\big)d\tau.
\end{aligned}
\end{equation*}
Similarly as the analysis of Theorem \ref{order condition}, it follows that
\begin{equation}\label{phixv}
\norm{\varphi_n^\tau}=\mathcal{O}(h^2),\qquad \ \norm{\varphi_n^x}=\mathcal{O}(h^3/\epsilon),\qquad
\ \norm{\varphi_n^v}=\mathcal{O}(h^3/\epsilon).
\end{equation}
Subtracting \eqref{SG1O2} from \eqref{truncation error} and by further using
\eqref{Global error}, we have
\begin{equation*}\label{}
\begin{aligned}
&{e}^{x}_{n+1}={e}^{x}_{n}+h\varphi_{1}(h\tilde{P}){e}^{v}_{n}+\varphi_n^x+\xi^{x}_{n}(h)+\gamma_{n}^{x},\ \ \ \ {e}^{v}_{n+1}=\varphi_{0}(h\tilde{P}){e}^{v}_{n}+\varphi_n^v+\xi^{v}_{n}(h)+\gamma_{n}^{v},
\end{aligned}
\end{equation*}
where
\begin{equation*}\label{}
\begin{aligned}
&\gamma_{n}^{x}=h\big(\varphi_{1}(h\tilde{P})-\varphi_1(h\tilde{M}_n)\big)v_n
+h^2\int_{0}^{1}(1-\tau)\varphi_{1}\big((1-\tau)h \tilde{P}\big)\big(F (\tilde{x}_{n}(\tau h))-F (X_{\tau})\big)d\tau\\
&\qquad+h^2\int_{0}^{1}(1-\tau)\big(\varphi_{1}\big((1-\tau)h \tilde{P}\big)-\varphi_{1}\big((1-\tau)h\tilde{M}_n\big)\big)F (X_{\tau})d\tau,\\
&\gamma_{n}^{v}=\big(\varphi_{0}(h\tilde{P})-\varphi_0(h\tilde{M}_n)\big)v_n
+h\int_{0}^{1}\varphi_{0}\big((1-\tau)h \tilde{P}\big)\big(F (\tilde{x}_{n}(\tau h))-F (X_{\tau})\big)d\tau\\
&\qquad+h\int_{0}^{1}\big(\varphi_{0}\big((1-\tau)h \tilde{P}\big)-\varphi_{0}\big((1-\tau)h\tilde{M}_n\big)\big)F (X_{\tau})d\tau.\\
\end{aligned}
\end{equation*}
Considering the same process of deduction in Theorem \ref{order condition}, one gets
\begin{equation*}
\begin{aligned}
&\norm{ \gamma_{n}^{x}}\leq \frac{h^2}{\epsilon}(\|e_n^x\|+h\|e_n^v\|+\|e_{n+1}^x\|+\|\xi_n^x(h)\|),
&\norm{ \gamma_{n}^{v}}\leq \frac{h}{\epsilon}(\|e_n^x\|+h\|e_n^v\|+\|e_{n+1}^x\|+\|\xi_n^x(h)\|).
\end{aligned}
\end{equation*}
We then obtain
\begin{equation*}\label{}
\begin{aligned}
\norm{ e_{n+1}^{x}}
&\leq \norm{ e_{n}^{x}}+h \norm{ e_{n}^{v}}+\norm{ \varphi_{n}^{x}}+\norm{ \xi_{n}^{x}(h)}+\norm{ \gamma_{n}^{x}}\\
& \leq \norm{ e_{n}^{x}}+h \big(\norm{ e_{n}^{x}}+\norm{ e_{n}^{v}}+\norm{ e_{n+1}^{x}}\big) +
\norm{ \varphi_{n}^{x}}+\norm{ \xi_{n}^{x}(h)},
\end{aligned}
\end{equation*}
and
\begin{equation*}\label{}
\begin{aligned}
\norm{ e_{n+1}^{v}}
&\leq \norm{ e_{n}^{v}}+\norm{ \varphi_{n}^{v}}+\norm{ \xi_{n}^{v}(h)}+\norm{ \gamma_{n}^{v}}\\
&\leq \norm{ e_{n}^{v}}+\frac{h}{\epsilon} \big(\norm{ e_{n}^{x}}+\norm{ e_{n}^{v}}+\norm{ e_{n+1}^{x}}\big) +
\norm{ \varphi_{n}^{v}}+\norm{ \xi_{n}^{v}(h)}.
\end{aligned}
\end{equation*}
Combining $\norm{ e_{n+1}^{x}}$ and $\epsilon \norm{ e_{n+1}^{v}}$, we have
\begin{equation*}\label{}
\begin{aligned}
&\norm{ e_{n+1}^{x}}+\epsilon \norm{ e_{n+1}^{v}}- \norm{ e_{n}^{x}}- \epsilon \norm{ e_{n}^{v}}\\
\leq & h(\|e_n^x\|+\|e_n^v\|+\|e_{n+1}^x\|)+\norm{ \varphi_{n}^{x}}+\epsilon \norm{ \varphi_{n}^{v}}
+\norm{ \xi_{n}^{x}(h)}+\epsilon \norm{ \xi_{n}^{v}(h)}.\\
\end{aligned}
\end{equation*}
Further, using recursion formula and $e_{0}^{x}=e_{0}^{v}=0$, it is arrived that
\begin{equation*}\label{}
\begin{aligned}
\norm{ e_{n+1}^{x}}+\epsilon \norm{ e_{n+1}^{v}}
\leq h\sum_{m=0}^{n}(\|e_m^x\|+\|e_m^v\|+\|e_{m+1}^x\|)+\sum_{m=0}^{n}(\norm{ \varphi_{m}^{x}}+\epsilon \norm{ \varphi_{m}^{v}}+\norm{ \xi_{m}^{x}(h)}+\epsilon \norm{ \xi_{m}^{v}(h)}).\\
\end{aligned}
\end{equation*}
By inserting the errors \eqref{xix}, \eqref{xiv} and \eqref{phixv} into the above formula and using the fact that $nh <T$, one has
\begin{equation*}\label{}
\begin{aligned}
\norm{ e_{n+1}^{x}}+\epsilon \norm{ e_{n+1}^{v}}
\leq h\sum_{m=0}^{n}(\|e_m^x\|+\|e_m^v\|+\|e_{m+1}^x\|)+Ch^2/\epsilon.
\end{aligned}
\end{equation*}
Using the Gronwall's inequality, we get
\begin{equation*}\label{}
\begin{aligned}
\norm{ e_{n+1}^{x}}+\epsilon \norm{ e_{n+1}^{v}}
\leq Ch^2/\epsilon.
\end{aligned}
\end{equation*}
Therefore, the global error \eqref{R1} of the method \eqref{SG1O2} is obtained immediately.
$\bullet$ \textbf{Proof of \eqref{R0}.}
This proof is also divided into three parts as the analysis of \eqref{R1}.
Firstly, we denote the linearized system of \eqref{charged-particle:B(x)} as
\begin{equation}\label{linearized system-R0}
\begin{aligned}
\dot{\tilde{x}}_n(s)=\dot{\tilde{v}}_n(s),\
\dot{\tilde{v}}_n(s)=\frac{1}{\epsilon}\tilde{B}(x(t_n))\tilde{v}_n(s) +F({\tilde{x}}_n(s)),\
{\tilde{x}}_n(0)=x(t_n),\ {\tilde{v}}_n(0)=v(t_n),\ 0<s\leq h.
\end{aligned}
\end{equation}
Then the local errors between the original \eqref{charged-particle:B(x)} and linearized systems \eqref{linearized system-R0} are given by
$$\xi^{x}_{n}(h)=\mathcal{O}(h^2),\qquad \xi^{v}_{n}(h)=\mathcal{O}(h^2/\epsilon),$$
which can be obtained in the same way by the proof of \eqref{R1}.
Finally, according to the above analysis, we get the final conclusion as follows
$$\norm{ e_{n+1}^{x}}+\epsilon \norm{ e_{n+1}^{v}}
\leq Ch,$$
which proves the result of \eqref{R0}.
$\bullet$ \textbf{Proof of \eqref{R2}.}
It is noted that the error bound between the system \eqref{charged-particle:B(x)} and linearized system \eqref{linearized system} is $\mathcal{O}(h^2)$. Thus higher order methods can not be obtained by considering algorithms with higher order for the linearized system \eqref{linearized system}. That is the reason why we use the idea of Triple Jump splitting method to construct the fourth-order method SG1O4. The convergence of this splitting can be derived by the standard analysis (see, e.g., \cite{Hairer2002}) and we omit it for brevity.
The proof of the theorem is finished
\end{proof}
For the special but important magnetic field, the so-called \emph{maximal ordering scaling} \cite{scaling1,Lubich2020,scaling2}:
\begin{equation}\label{ma0-charged-particle:B(x)}
\begin{aligned}
\dot{x}=v,\ \ \dot{v}=v \times \frac{ B(\epsilon x)}{\epsilon} +F(x),\ \
x(0)=x_0,\ \ v(0)=v_0,
\end{aligned}
\end{equation}
the improved convergent results can be obtained which are shown as follows.
\begin{mytheo}\label{order conditionN} \textbf{(Improved error bounds)}
Under the conditions of Theorem \ref{order condition-B(x)}, if the magnetic field has the \emph{maximal ordering scaling}, i.e., $B=B(\eps x)$, we have the following improved error bounds
\begin{subequations}\label{RRNEW}
\begin{align}&\textmd{SG1O1}: \quad \norm{x(t_{n+1})-x_{n+1}}\leq C h,\quad \quad \ \ \ \norm{v(t_{n+1})-v_{n+1}} \leq C h,\label{R0N}\\
&\textmd{SG1O2}: \quad \norm{x(t_{n+1})-x_{n+1}}\leq C h^2,\qquad \ \ \norm{v(t_{n+1})-v_{n+1}} \leq C h^2/\epsilon,\label{R1N}\\
&\textmd{SG1O4}: \quad \norm{x(t_{n+1})-x_{n+1}}\leq C h^4/\epsilon^2,\quad \ \norm{v(t_{n+1})-v_{n+1}}\leq C h^4/\epsilon^3. \label{R2N}
\end{align}
\end{subequations}
\end{mytheo}
\begin{proof}
$\bullet$ \textbf{Proof of \eqref{R0N}.}
Concern the linearized system of \eqref{ma0-charged-particle:B(x)} as follows
\begin{equation}\label{ma-linearized system}
\begin{aligned}
\dot{\tilde{x}}_n(s)=\dot{\tilde{v}}_n(s),\ \
\dot{\tilde{v}}_n(s)=\frac{1}{\epsilon}\tilde{B}(\epsilon x(t_n))\tilde{v}_n(s) +F({\tilde{x}}_n(s)),\ \
{\tilde{x}}_n(0)=x(t_n),\ \ {\tilde{v}}_n(0)=v(t_n),\ \ \ 0<s\leq h.
\end{aligned}
\end{equation}
Similarly as the proof of Theorem \ref{order condition-B(x)}, the errors between the systems \eqref{ma0-charged-particle:B(x)} and \eqref{ma-linearized system} are bounded as follows
\begin{equation*
\xi^{x}_{n}(h)=\mathcal{O}(h^3),\ \ \ \ \ \ \ \xi^{v}_{n}(h)=\mathcal{O}(h^3/\epsilon).
\end{equation*}
Then, we shall consider the global errors of the method \eqref{SG1O1} in the same way as Theorem \ref{order condition-B(x)}, and one gets
\begin{equation*}\label{R0N-error}
\begin{aligned}
&\norm{{e}^{x}_{n+1}}\leq \norm{{e}^{x}_{n}}+h\norm{{e}^{v}_{n}}+\norm{\varphi_n^x}+\norm{\xi^{x}_{n}(h)}+\norm{\gamma_{n}^{x}},\\
&\norm{{e}^{v}_{n+1}}\leq \norm{{e}^{v}_{n}}+\norm{\varphi_n^v}+\norm{\xi^{v}_{n}(h)}+\norm{\gamma_{n}^{v}}.
\end{aligned}
\end{equation*}
The notations $\varphi_n^x$, $\varphi_n^v$, $\gamma_{n}^{x}$ and $\gamma_{n}^{v}$ are totally the same as those in the proof of \eqref{R1} with some corresponding modifications in $B$. They are bounded by
$$
\norm{\varphi_n^x}=\mathcal{O}(h^3/\epsilon), \ \ \norm{\varphi_n^v}=\mathcal{O}(h^3/\epsilon), \ \
\norm{\gamma_n^x}\leq h^2(\norm{e_n^x}+h\norm{e_n^v}),\ \ \norm{\gamma_n^v}\leq h(\norm{e_n^x}+h\norm{e_n^v}).$$
The above results lead to
\begin{equation*}
\begin{aligned}
\norm{{e}^{x}_{n+1}}+\norm{{e}^{v}_{n+1}}-\norm{{e}^{x}_{n}}-\norm{{e}^{v}_{n}} \leq h(\norm{{e}^{x}_{n}}+\norm{{e}^{v}_{n}})
+\norm{\varphi_n^x}+\norm{\varphi_n^v}+\norm{\xi_n^x(h)}+\norm{\xi_n^v(h)}.
\end{aligned}
\end{equation*}
Therefore, we have
$$\norm{ e_{n+1}^{x}}+ \norm{ e_{n+1}^{v}} \leq Ch.$$
$\bullet$ \textbf{Proof of \eqref{R1N}--\eqref{R2N}.}
The proof of \eqref{R1N} and \eqref{R2N} can be obtained by the same progress with \eqref{R0N}, and so we
leave out the details for simplicity.
\end{proof}
\subsection{Numerical tests}
In what follows, we present two numerical experiments to show the behaviour of the derived methods. We still choose BORIS, RKO2 and RKO4 for comparison which are given in Section \ref{sec: experiment}. {To compare the first order method SG1O1 with some existing methods, the implicit Euler method (of order one) is chosen and we refer to it as Euler.}
\noindent\vskip3mm \noindent\textbf{Problem 2. (General magnetic field)}
For the charged-particle dynamics \eqref{charged-particle sts-cons}, the scalar potential $U(x)$ and non-homogeneous magnetic field $\frac{1}{\epsilon}B(x)$ are given by (\cite{Lubich2017})
$$ \frac{1}{\epsilon}B(x)=\nabla \times \frac{1}{3\epsilon}\Big(-x_2\sqrt{x_{1}^2+x_{2}^2},x_1\sqrt{x_{1}^2+x_{2}^2},0\Big)^{T}
=\frac{1}{\epsilon}\Big(0,0,\sqrt{x_{1}^2+x_{2}^2}\Big)^{\intercal}, \ \ \ U(x)=\frac{1}{100\sqrt{x_{1}^2+x_{2}^2}},$$
where initial values are chosen as
$x(0)=(0,1,0.1)^{\intercal},\ v(0)=(0.09,0.05,0.2)^{\intercal}.$
The problem is solved on $[0,1]$ with $h=1/2^k,$ where $k=3,\ldots,7$, and
{Figs. \ref{fig:problem31}--\ref{fig:problem31new2} show the results of global errors.} Then the system is integrated on the interval $ [0,100]$
with a step size $h=\frac{1}{100}$ and see Fig. \ref{fig:problem32} for the energy conservation.
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{e31} & \includegraphics[width=4.7cm,height=4.5cm]{e32}&\includegraphics[width=4.7cm,height=4.5cm]{e33}
\end{tabular}
\caption{Problem 2. The global errors $error:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{ \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$.}
\label{fig:problem31}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=4.7cm,height=4.5cm]{3Euler} &\includegraphics[width=4.7cm,height=4.5cm]{3SG1O1}
\end{tabular}
\caption{Problem 2. The errors $error1:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{ \eps \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of first order methods (Euler and SG1O1) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$.}
\label{fig:problem31new3}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{3boris} &\includegraphics[width=4.7cm,height=4.5cm]{3RKO2} & \includegraphics[width=4.7cm,height=4.5cm]{3SG1O2}
\end{tabular}
\caption{Problem 2. The errors $error2:=\frac{\eps\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps^2 \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of second order methods (BORIS, RKO2 and SG1O2) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem31new1}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=4.7cm,height=4.5cm]{3RKO4} &\includegraphics[width=4.7cm,height=4.5cm]{3SG1O4}
\end{tabular}
\caption{Problem 2. The errors $error4:=\frac{\eps^3\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps^4 \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of fourth order methods (RKO4 and SG1O4) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem31new2}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{h31} & \includegraphics[width=4.7cm,height=4.5cm]{h32} & \includegraphics[width=4.7cm,height=4.5cm]{h33}
\end{tabular}
\caption{Problem 2. Evolution of the energy error $e_{H}:=\frac{|H(x_{n},v_n)-H(x_0,v_0)|}{|H(x_0,v_0)|}$ as function of time $t=nh$.}
\label{fig:problem32}
\end{figure}
\noindent\vskip3mm \noindent\textbf{Problem 3. (Maximal ordering scaling)}
Consider the charged-particle dynamics \eqref{charged-particle sts-cons} from \cite{WZ} in a magnetic field with the maximal ordering scaling
$\frac{1}{\epsilon}B(\epsilon x)=\frac{1}{\epsilon}\left(
\begin{array}{c}
\cos(\epsilon x_2) \\
1+\sin(\epsilon x_3) \\
\cos(\epsilon x_1) \\
\end{array}
\right)+\left(
\begin{array}{c}
-x_1 \\
0 \\
x_3 \\
\end{array}
\right)$
and the scalar potential $U(x)=\frac{1}{\sqrt{x_{1}^2+x_{2}^2}}.$
We take $x(0)=(1/3, 1/4, 1/2)^{\intercal}, \ v(0)=(2/5, 2/3, 1)^{\intercal}$ as the initial values.
{Figs. \ref{fig:problem41}--\ref{fig:problem41new2} present the errors of all the methods} when solving this problem on $[0,1]$ with different $\epsilon$ and $h=1/2^{k}$, where $k=3,4,\ldots,7$. Finally, the energy errors are shown in Fig.
\ref{fig:problem42} on $[0,100]$ with $h=\frac{1}{20}$.
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{e41} & \includegraphics[width=4.7cm,height=4.5cm]{e42} & \includegraphics[width=4.7cm,height=4.5cm]{e43}
\end{tabular}
\caption{Problem 3. The global errors $error:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{ \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem41}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=4.7cm,height=4.5cm]{4Euler} &\includegraphics[width=4.7cm,height=4.5cm]{4SG1O1}
\end{tabular}
\caption{Problem 3. The errors $error1:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{ \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of first order methods (Euler and SG1O1) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$.}
\label{fig:problem41new3}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{4boris} &\includegraphics[width=4.7cm,height=4.5cm]{4RKO2} & \includegraphics[width=4.7cm,height=4.5cm]{4SG1O2}
\end{tabular}
\caption{Problem 3. The errors $error2:=\frac{\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of second order methods (BORIS, RKO2 and SG1O2) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$. }
\label{fig:problem41new1}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{cccc
\includegraphics[width=4.7cm,height=4.5cm]{4RKO4} &\includegraphics[width=4.7cm,height=4.5cm]{4SG1O4}
\end{tabular}
\caption{Problem 3. The errors $error4:=\frac{\eps^2\norm{x_{n}-x(t_n)}}{\norm{x(t_n)}}+\frac{\eps^3 \norm{v_{n}-v(t_n)}}{\norm{v(t_n)}}$ of fourth order methods (RKO4 and SG1O4) with $t=1$ and $h=1/2^{k}$ for $k=3,4,\ldots,7$ under different $\epsilon$.}
\label{fig:problem41new2}
\end{figure}
\begin{figure}[t!]
\centering\tabcolsep=0.4mm
\begin{tabular}
[c]{ccc
\includegraphics[width=4.7cm,height=4.5cm]{h41} & \includegraphics[width=4.7cm,height=4.5cm]{h42} & \includegraphics[width=4.7cm,height=4.5cm]{h43}
\end{tabular}
\caption{Problem 3. Evolution of the energy error $e_{H}:=\frac{|H(x_{n},v_n)-H(x_0,v_0)|}{|H(x_0,v_0)|}$ as function of time $t=nh$.}
\label{fig:problem42}
\end{figure}
Based on the above results, we can draw the following observations. For the accuracy, our methods agree with the results presented in Theorem \ref{order condition-B(x)} and \ref{order conditionN} and are more accurate than the Euler, Boris and Runge-Kutta methods. Concerning the energy conservation, it can be seen that our methods have a long time conservation not only for normal magnetic fields but also for strong ones.
\section{Conclusions} \label{sec:conclusions}
{
In this paper, symplectic methods
for solving the charged-particle dynamics (CPD) \eqref{charged-particle sts-cons} were presented and studied.
By employing some transformations of the system and methods, we derived symplecticity conditions for
a novel kind of adapted exponential methods, and based on which,
two symplectic methods up to order four were constructed for solving
CPD in a strong and homogeneous magnetic field.
Rigorous error estimates were presented and the proposed second order symplectic method was shown to have uniform error bound in the position w.r.t. the strength of the magnetic field.
Furthermore, the extension of the obtained symplectic methods to
the case of non-homogeneous magnetic fields was discussed. Three novel algorithms up to order four were constructed
and one method was proved to have uniform accuracy in the position.
Some numerical
tests on homogeneous and non-homogeneous magnetic fields were presented to confirm the theoretical results and to demonstrate the numerical behaviour in accuracy and energy conservation.
Last but not least, we point out that the numerical results of Problems 1--3 demonstrate a very good long time energy conservation for the methods presented in this paper. To theoretically prove this property, the strategies named as
backward error analysis \cite{Hairer2017,Lubich2017,Hairer2002,wangwu2020} and modulated Fourier expansion \cite{Hairer2018,Hairer2020,Hairer2002,Lubich2020,WZ} will be employed for CPD under normal and strong magnetic fields, respectively. The rigorous analysis on this topic will be considered in our next work. Another issue for future exploration is the study of uniform higher-order symplectic integrators. Besides, a combination of symplectic integrators and the Particle-In-Cell (PIC) approximation and its analysis are of great interests for Vlasov equations \cite{VP1,Zhao,VP4,VP5,VP6}.
}
|
\section{Introduction}
Generalized Bowen-Franks groups are abelian group invariants for equivalence relations in symbolic dynamics and toral endomorphisms. The original Bowen-Franks group arose as an invariant of flow equivalence for suspension flows of subshifts of finite type \cite{BF77}. Following the notation of Lind and Marcus \cite{LM21}, the Bowen-Franks group for $A\in M_n({\mathbb Z})$, an $n\times n$ integer matrix representing the subshift of finite type, is
\[ {\rm BF}(A) = {\mathbb Z}^n/{\mathbb Z}^n(I-A),\]
where $I-A$ acts on the right and the elements of ${\mathbb Z}^n$ are row vectors. Franks showed that ${\rm BF}(A)$, together with the flow equivalence invariant ${\rm det}(I-A)$ of Parry and Sullivan \cite{PS75}, are a complete set of computable invariants of shift equivalence for irreducible subshifts of finite type \cite{Fr84}. The Bowen-Franks group ${\rm BF}(A)$ has been an important tool in the classification of reducible subshifts \cite{Hu95} as well as general subshifts \cite{Ma01,Ma21}. More invariants of shift equivalence for subshifts of finite type were introduced by Kitchens that generalized the Bowen-Franks group \cite{Ki98}. These generalized Bowen-Franks groups are of the form
\[ {\rm BF}_g(A) = {\mathbb Z}^n/{\mathbb Z}^n g(A),\]
where $g\in {\mathbb Z}[t]$ is a polynomial with $g(0)$ being a unit in ${\mathbb Z}$, i.e., $g(0) = \pm 1$ \cite{Ki98, LM21}. The original Bowen-Franks group ${\rm BF}(A)$ corresponds to ${\rm BF}_g(A)$ for $g(t) = 1-t$. Another generalization of the Bowen-Franks group ${\rm BF}(A)$ was introduced by one of the authors and Sousa Ramos for toral endomorphisms of ${\mathbb T}^n$ which are represented by $A\in M_n({\mathbb Z})$ \cite{MR05}. These generalized Bowen-Frank groups ${\rm BF}_g(A)$ are invariants of topological conjugacy for $A$ where $g\in {\mathbb Z}[t]$. Topological conjugacy of two toral endomorphisms $A,B\in M_n({\mathbb Z})$ is equivalent to conjugacy by an element of ${\rm GL}_n({\mathbb Z})$, i.e., there exists $C\in {\rm GL}_n({\mathbb Z})$ such that $AC = CB$ (cf \cite{AP65}). For a toral endomorphism $A \in M_n({\mathbb Z})$, each generalized Bowen-Franks group ${\rm BF}_g(A)$ has the additional structure of a right $R$-module for a ring $R$ isomorphic to the matrix ring
\[ {\mathbb Z}[A] = \{ p(A) : p\in {\mathbb Z}[t]\}\]
(see Subsection \ref{sec:Module}). The structure of the finitely generated abelian group ${\rm BF}_g(A)$, with $A$ representing a subshift of finite type or a toral endomorphism, can be determined by computing the Smith normal form of $g(A)$. One of the authors and Sousa Ramos showed, for a given irreducible monic polynomial $f\in {\mathbb Z}[t]$ satisfying $f(0) = \pm 1$ and having square-free discriminant, that all $A\in {\rm GL}_n({\mathbb Z})$ having characteristic polynomial $f(t)$ have $R$-module isomorphic ${\rm BF}_g(A)$ for each $g\in {\mathbb Z}[t]$ \cite{MR05}. Chen showed in some cases that ${\rm BF}_g(A)$, as an abelian group, is determined by the zeta function of $A$ for those $g\in{\mathbb Z}[t]$ with $g(0) = \pm 1$, without assuming irreducibility of the characteristic polynomial of $A$ \cite{Ch09}. When $A\in {\rm GL}_n({\mathbb Z})$ is a hyperbolic toral automorphism (no eigenvalue of $A$ has modulus $1$) and $g_k(t) = t^k-1$, $k\in {\mathbb N}$, each generalized Bowen-Franks group,
\[ {\rm BF}_k(A) = {\mathbb Z}^n/{\mathbb Z}^n g_k(A) = {\mathbb Z}^n/{\mathbb Z}^n(A^k-I),\]
is a finite abelian group that is isomorphic by Pontryagin duality to the set ${\rm Per}_k(A)$ of $k$-periodic points of the left action of $A$ on ${\mathbb T}^n$, where the elements of ${\mathbb T}^n$ are column vectors. Note that the original Bowen-Franks group ${\rm BF}(A)$, corresponding to $g(t) = 1-t$, is isomorphic to ${\rm BF}_1(A)$, corresponding to $g_1(t) = t-1$. The main purpose of this paper is to show, when $A\in {\rm GL}_n({\mathbb Z})$ is hyperbolic, that the collection of {\it principal} generalized Bowen-Franks $R$-modules ${\rm BF}_k(A)$, $k \in {\mathbb N}$, is a complete set of invariants for an equivalence relation called profinite conjugacy.
Profinite conjugacy is a topological $R$-module isomorphism between two profinite $R$-modules. For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$, the principal Bowen-Franks $R$-modules ${\rm BF}_k(A)$, $k\in {\mathbb N}$, are parameterized by the collection ${\mathcal Q}^{\rm p}$ of polynomials $g_k$, $k\in {\mathbb N}$, and have an inverse limit $G_{ {\mathcal Q}^{\rm p},A}$ that is a profinite $R$-module, whose Pontryagin dual is the direct limit of ${\rm Per}_k(A)$, $k\in {\mathbb N}$, which is the rational torus. For another hyperbolic $B\in{\rm GL}_n({\mathbb Z})$ that is similar to $A$ over ${\mathbb Q}$ i.e., there exists $P\in {\rm GL}_n({\mathbb Q})$ such that $AP = PB$, so it has the same characteristic polynomial as that of $A$, the principal Bowen-Franks groups ${\rm BF}_k(B)$, $k\in {\mathbb N}$, have an inverse limit $R$-module $G_{ {\mathcal Q}^{\rm p},B}$. Similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ have a common ring $R$ to which both ${\mathbb Z}[A]$ and ${\mathbb Z}[B]$ are isomorphic (see Subsection \ref{sec:Module}), and are said to be strongly ${\rm BF}$-equivalent over ${\mathcal Q}^{\rm p}$ if ${\rm BF}_k(A)$ and ${\rm BF}_k(B)$ are $R$-module isomorphic, written ${\rm BF}_k(A) \cong_R {\rm BF}_k(B)$, for all $k\in {\mathbb N}$. We showed that if similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ are strongly ${\rm BF}$-equivalent over ${\mathcal Q}^{\rm p}$, then $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm p},B}$ are topologically $R$-module isomorphic, i.e., there exists a map $\Psi_{ {\mathcal Q}^{\rm p}}:G_{ {\mathcal Q}^{\rm p},A} \to G_{ {\mathcal Q}^{\rm p},B}$ that is an $R$-module isomorphism and a homeomorphism \cite{BM12}. The existence of such a map $\Psi_{ {\mathcal Q}^{\rm p}}:G_{ {\mathcal Q}^{\rm p},A}\to G_{ {\mathcal Q}^{\rm p},B}$ we call a profinite conjugacy over ${\mathcal Q}^{\rm p}$ between similiar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$. The main result of this paper includes the converse: for similar hyperbolic $A,B\in{\rm GL}_n({\mathbb Z})$, if $A$ and $B$ are profinitely conjugate over ${\mathcal Q}^{\rm p}$, then $A$ and $B$ are strongly ${\rm BF}$-equivalent over ${\mathcal Q}^{\rm p}$. This establishes the equivalence of strong ${\rm BF}$-equivalence over ${\mathcal Q}^{\rm p}$ and profinite conjugacy over ${\mathcal Q}^{\rm p}$.
One of the authors and Sousa Ramos proved a result in \cite{MR05} that implies for similar hyperbolic automorphisms $A,B\in {\rm GL}_n({\mathbb Z})$ with irreducible characteristic polynomial that if $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}^{\rm p}$, then $A$ and $B$ are strongly {\rm BF}-equivalent over the collection ${\mathcal Q}^{\rm a}$ of polynomials $g\in {\mathbb Z}[t]$ satisfying ${\rm det}(g(A))\ne 0$, i.e., ${\rm BF}_g(A) \cong_R {\rm BF}_g(B)$ for all $g\in {\mathcal Q}^a$. We show, as a corollary of the main result, that strong {\rm BF}-equivalence of $A$ and $B$ over ${\mathcal Q}^{\rm p}$ implies strong {\rm BF}-equivalence of $A$ and $B$ over ${\mathcal Q}^{\rm a}$ without assuming irreducibility of the characteristic polynomial of the similar hyperbolic $A$ and $B$. Hence, in the irreducible and reducible cases for similar hyperbolic toral automorphisms, the principal Bowen-Franks $R$-modules are the principal invariants in the class of generalized Bowen-Franks $R$-module invariants parameterized by ${\mathcal Q}^{\rm a}$.
For irreducible similar hyperbolic toral automorphisms, the principal generalized Bowen-Franks $R$-modules are a complete set of invariants for the equivalence relations of weak equivalence of the associated ideals and block conjugacy of the matrices as well as for profinite conjugacy. Here irreducible means the characteristic polynomial $f(t)$ for the similar $A,B\in {\rm GL}_n({\mathbb Z})$ is irreducible in ${\mathbb Z}[t]$. Fixing a root $\beta$ of $f(t)$, the selection of a right eigenvector $u = (u_1,\dots,u_n)^{\rm T}$ of $A$, i.e., $Au = \beta u$, with entries $u_i$ in the algebraic number field ${\mathbb Q}(\beta)$ determines a (fractional) ideal
\[ I_A = u_1{\mathbb Z} + \cdots + u_n {\mathbb Z}\]
in the ring ${\mathbb Z}[\beta]$. By the Lattimer-MacDuffey-Taussky one-to-one correspondence \cite{LM33,Ta78}, different choices of right eigenvectors determine arithmetically equivalent ideals in ${\mathbb Z}[\beta]$. Similarly the selection of a right eigenvector $v = (v_1,\dots,v_n)^{\rm T}$ for $B$ determines a (fractional) ideal $I_B$ in ${\mathbb Z}[\beta]$. Two (fractional) ideals $I$ and $J$ of ${\mathbb Z}[\beta]$ are weakly equivalent if there exists (fractional) ideals $X$ and $Y$ of ${\mathbb Z}[\beta]$ such that $IX = J$ and $JY = I$ \cite{DTZ62}. We showed that for similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ with an irreducible characteristic polynomial, if $I_A$ and $I_B$ are weakly equivalent, then $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}^{\rm a}$ \cite{BM12}. Since $A$ is hyperbolic, each $g_k(A) = A^k-I$ is invertible, and so $I_A$ and $I_B$ being weakly equivalent implies that $A$ and $B$ are strongly ${\rm BF}$-equivalent over ${\mathcal Q}^{\rm p}$. We showed that for similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ with an irreducible characteristic polynomial, if $A$ and $B$ are profinitely conjugate over ${\mathcal Q}^{\rm p}$, then $I_A$ and $I_B$ are weakly equivalent \cite{BMM16}. Similar $A,B\in {\rm GL}_n({\mathbb Z})$ (not assumed hyperbolic) with an irreducible characteristic polynomial are called block conjugate if there exists $A^\prime,B^\prime \in {\rm GL}_n({\mathbb Z})$ and $M,N\in {\rm GL}_{2n}({\mathbb Z})$ such that the block diagonal matrices $A\oplus A^\prime$, $B\oplus B$, $A\oplus A$, and $B\oplus B^\prime$ satisfy
\[ (A\oplus A^\prime)M = M(B\oplus B) {\rm\ and \ }(A\oplus A)N = N(B\oplus B^\prime).\]
We showed that similar $A,B\in {\rm GL}_n({\mathbb Z})$ with irreducible characteristic polynomial are block conjugate if and only if the associated ideals $I_A$ and $I_B$ are weakly equivalent \cite{BM19}. Applying this to similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ with irreducible characteristic polynomial gives $I_A$ and $I_B$ are weakly equivalent if and only if $A$ and $B$ are block conjugate. Putting these results together gives that the following are equivalent for irreducible similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$:
\begin{enumerate}
\item[(a)] $I_A$ and $I_B$ are weakly equivalent,
\item[(b)] $A$ and $B$ are block conjugate
\item[(c)] $A$ and $B$ are profinitely conjugate over ${\mathcal Q}^{\rm p}$,
\item[(d)] $A$ and $B$ are strongly ${\rm BF}$-equivalent over ${\mathcal Q}^{\rm p}$.
\end{enumerate}
This establishes the principal generalized Bowen-Franks $R$-modules as a set of complete invariants for the equivalence relations listed in items (a), (b), and (c) when $A$ and $B$ are irreducible similar hyperbolic toral automorphisms.
The structure of this paper is as follows. Section 2 states some basic definitions and theory for profinite groups, as found in \cite{RZ10}, that will be used in the proof of the main result. In particular profinite topologies, profinite completions, and cofinality are detailed. Section 3 applies the results from Section 2 to profinite completions of ${\mathbb Z}^n$ with particular attention to the full profinite topology of ${\mathbb Z}^n$ and its accompanying completion. In Section 4 we prove for ${\mathcal Q}$ being either ${\mathcal Q}^p$ or ${\mathcal Q}^a$ the main result that profinite conjugacy over ${\mathcal Q}$ implies strong ${\rm BF}$-equivalence over ${\mathcal Q}$. Prior to giving the proof in this section, the construction of the profinite $R$-modules $G_{{\mathcal Q},A}$ for ${\mathcal Q} = {\mathcal Q}^{\rm p}$ and ${\mathcal Q}^{\rm a}$ is given. We already have a proof of the main result over ${\mathcal Q}^p$ for {\it irreducible} similar hyperbolic toral automorphisms, as this follows from the implications (c) $\Rightarrow$ (a) $\Rightarrow$ (d) stated in the previous paragraph. The proof presented in Section 4 does not assume irreducibility and does not use weak equivalence. In Section 5 we present some properties of profinite conjugacies for similar hyperbolic toral automorphisms. We correct a characterization from \cite{BM12} of when profinitely conjugate similar hyperbolic toral automorphisms are conjugate. We also show how profinite conjugacy realizes the conjugacy invariant of cyclicity when applied to hyperbolic toral automorphisms.
\section{Profinite Theory}
\subsection{Profinite Topologies}
A profinite topology on an abelian group $G$ is a translation invariant topology determined by selecting a neighbourhood basis of the identity element $0$ of $G$ in the following way. Let ${\mathcal N}$ be a nonempty collection of finite index subgroups of $G$ that is filtered from below, i.e., for any $N_1,N_2\in{\mathcal N}$ there exists $N_3\in{\mathcal N}$ such that $N_3\subset N_1\cap N_2$. The filtered from below property makes ${\mathcal N}$ into a directed poset for the binary relation $\preceq$ on ${\mathcal N}$ defined by $N_1\preceq N_2$ if and only if $N_2 \subset N_1$. The ``directed'' part is that for any $N_1,N_2\in{\mathcal N}$ there exists $N_3\in{\mathcal N}$ such that $N_1\preceq N_3$ and $N_2\preceq N_3$. Viewing ${\mathcal N}$ as a fundamental system of neighbourhoods (which is the same as a neighbourhood basis) of the identity element $0$ of $G$ determines a profinite topology on $G$ (see \cite{Bo89}).
For a nonempty collection ${\mathcal N}$ of filtered from below finite index subgroups of $G$, the type of profinite topology on $G$ determined by ${\mathcal N}$ is described by the class ${\mathcal C}$ of finite abelian groups to which all the quotients $G/N$, $N\in {\mathcal N}$, belong. A specified class ${\mathcal C}$ of finite abelian groups always contains all isomorphic copies of the groups in it. The simplest structure on a class ${\mathcal C}$ is that of a formation. A class ${\mathcal C}$ of finite abelian groups is called a formation if it satisfies the following two properties: (1) if $K\in {\mathcal C}$ and $H\leq K$, then $K/H\in {\mathcal C}$, and (2) if $H_j\in {\mathcal C}$ for $j=1,\dots,r$ and $K$ is a subgroup of the direct product $H_1\times\cdots\times H_r$ such that for the canonical projections (epimorphisms)
\[ q_j: H_1\times\cdots \times H_j\times\cdots\times H_r\to H_j,\]
the restrictions $q_j\vert_K:K\to H_j$ are surjections for all $j=1,\dots,r$, then $K\in {\mathcal C}$. The first property is the closure of ${\mathcal C}$ under taking quotients. The second property is the closure of ${\mathcal C}$ under taking subdirect products. The second property implies the closure of ${\mathcal C}$ under taking direct products. The class ${\mathcal C}$ of all finite abelian groups is readily verified as a formation.
For a formation ${\mathcal C}$ of finite abelian groups, a pro-${\mathcal C}$ topology on $G$ is a profinite topology determined by a nonempty collection ${\mathcal N}$ of filtered from below finite indexed subgroups of $G$ that satisfies $G/N\in {\mathcal C}$ for all $N\in{\mathcal N}$. The full pro-${\mathcal C}$ topology on $G$ is the pro-${\mathcal C}$ topology determined by the nonempty filtered from below collection
\[ {\mathcal N}_{\mathcal C}(G) = \{ N \leq_f G : G/N \in {\mathcal C}\},\]
where $\leq_f$ means finite index subgroup.
\subsection{Profinite Completions}
A profinite completion of $G$ is induced by a profinite topology on $G$. Let ${\mathcal N}$ be a nonempty collection of finite index subgroups of $G$ that is filtered from below. For each $N\in{\mathcal N}$ equip the finite abelian quotient group $G/N$ with the discrete topology. Each $G/N$ is a compact, Hausdorff, totally disconnected topological group. For $N,M\in{\mathcal N}$ satisfying $M\preceq N$ and $M\ne N$, the natural epimorphism
\[ \varphi_{NM}:G/N \to G/M {\rm \ given\ by\ } x + N \mapsto x +M\]
is continuous. Denote the coset $x+N$ of $G/N$ by $[x]_N$. For each $N\in {\mathcal N}$ we set $\varphi_{NN}$ to be the identity map from $G/N$ to $G/N$. For all $N,M,P\in {\mathcal N}$ satisfying $N\preceq M\preceq P$ there holds $\varphi_{PN} = \varphi_{MN}\varphi_{PM}$. The continuous epimorphisms $\varphi_{NM}$ for $M\preceq N$ are called transitions maps. The collection $\{ G/N, \varphi_{NM}, {\mathcal N}\}$ is a surjective inverse system of topological finite abelian groups. The profinite completion of $G$ with respect to the profinite topology on $G$ determined by ${\mathcal N}$ is the inverse limit of the surjective inverse system,
\[ {\mathcal K}_{\mathcal N}(G) = \varprojlim_{N\in{\mathcal N}} G/N,\]
whose elements are the coherent tuples $\{[x_N]_N\}_{N\in{\mathcal N}}$ in $\prod_{N\in{\mathcal N}} G/N$, i.e., those tuples satisfying
\[ [x_M]_M = \varphi_{NM}([x_N]_N) = [x_N]_M\]
whenever $M\preceq N$. The profinite completion ${\mathcal K}_{\mathcal N}(G)$ is a nonempty, compact, Hausdorff, totally disconnected topological abelian group. The profinite completion ${\mathcal K}_{\mathcal N}(G)$ is called a profinite abelian group.
Several natural continuous homomorphisms are associated to the profinite abelian group ${\mathcal K}_{\mathcal N}(G)$. There are the continuous epimorphisms
\[ \varphi_{N}:{\mathcal K}_{\mathcal N}(G) \to G/N, \ \ \varphi_{N}( [x_M]_{M\in{\mathcal N}}) = [x_N],\ N\in {\mathcal N}.\]
The continuous epimorphisms $\iota_N:G\to G/N$, $N\in {\mathcal N}$, induce the continuous homomorphism $\iota_{\mathcal N}:G \to {\mathcal K}_{\mathcal N}(G)$ given by
\[ \iota_{\mathcal N}(x) = \{[x]_N\}_{N\in{\mathcal N}},\]
which take each $x\in G$ to the constant tuple $\{[x]_N\}_{N\in{\mathcal N}}$ in ${\mathcal K}_{\mathcal N}(G)$. The map $\iota_{\mathcal N}$ has the property that $\iota_{\mathcal N}(G)$ is dense in ${\mathcal K}_{\mathcal N}(G)$. Furthermore $\iota_{\mathcal N}$ is a monomorphism if and only if
\[ \bigcap_{N\in {\mathcal N}} N = \{0\}.\]
For a fixed formation ${\mathcal C}$ of finite abelian groups, a pro-${\mathcal C}$ completion of $G$ is the profinite abelian group ${\mathcal K}_{\mathcal N}(G)$ when $G/N\in {\mathcal C}$ for all $N\in {\mathcal N}$. The profinite completion of $G$ induced by the full profinite topology determined by ${\mathcal N}_{\mathcal C}(G)$ will be denoted by ${\mathcal K}_{\mathcal C}(G)$ instead of the more cumbersome ${\mathcal K}_{ {\mathcal N}_{\mathcal C}(G)}(G)$. The continuous homomorphism $\iota_{ {\mathcal N}_{\mathcal C}}(G):G \to {\mathcal K}_{\mathcal C}(G)$ will be denoted by the simpler $\iota_{\mathcal C}(G)$ as well. The profinite abelian group ${\mathcal K}_{\mathcal C}(G)$ is called residually ${\mathcal C}$ if
\[ \bigcap_{N\in {\mathcal N}_{\mathcal C}(G)} N = \{0\}.\]
This condition is equivalent to the continuous homomorphism $\iota_{\mathcal C}$ being a monomorphism.
We need a few additional notations to state the following result that relates the full pro-${\mathcal C}$ topology on $G$ with the topology on the profinite completion ${\mathcal K}_{\mathcal C}(G)$. The notation $\{ U: U\leq_o G;{\mathcal N}\}$ means the collection of all open subgroups $U$ of $G$ with respect to the profinite topology determined by a nonempty filtered from below collection ${\mathcal N}$ of finite index subgroups of $G$. For groups $H$ and $K$ with $H\leq K$ the notation $[K:H]$ means the index of $H$ in $K$. The notation $H\leq_o G$ means $H$ is an open subgroup of a topological group $G$. The following result (Proposition 3.2.2 in \cite{RZ10}) is presented in the abelian category, and will be used to prove the main result.
\begin{theorem}\label{fullproCtopology} Let ${\mathcal C}$ be a formation of finite abelian groups and suppose $G$ is a residually ${\mathcal C}$ group. Identify $G$ with its embedded image $\iota_{\mathcal C}(G)$ in its pro-${\mathcal C}$ completion ${\mathcal K}_{\mathcal C}(G)$. Let $\overline{X}$ denote the closure in ${\mathcal K}_{\mathcal C}(G)$ of a subset of $X$ in $G$.
\begin{enumerate}
\item[(a)] Let
\[ \Upsilon: \{ U: U \leq_o G; {\mathcal N}_{\mathcal C}(G)\} \to \{ V: V \leq_o {\mathcal K}_{\mathcal C}(G) \}\]
be the mapping that assigns to each open subgroup $H$ in $G$ $($with respect to the full profinite topology$)$ the closure $\Upsilon(H)=\overline{H}$ in ${\mathcal K}_{\mathcal C}(G)$. Then $\Upsilon$ is a one-to-one correspondence between the set of all open subgroups in the full pro-${\mathcal C}$ topology of $G$ and the set of all open subgroups of ${\mathcal K}_{\mathcal C}(G)$.
The inverse of $\Upsilon$ is
\[ V\to V\cap\iota_{\mathcal C}(G),\]
where in particular $\overline{ V\cap \iota_{\mathcal C}(G)} = V$ when
$V \leq_o {\mathcal K}_{\mathcal C}(G)$.
\item[(b)] The topology of ${\mathcal K}_{\mathcal C}(G)$ induces on $G$ its full pro-${\mathcal C}$ topology via the map $V \to V\cap \iota_{\mathcal C}(G)$ and the identification of $G$ with $\iota_{\mathcal C}(G)$.
\item[(c)] If $H,K \in \{ U : U \leq_o G; {\mathcal N}_{\mathcal C}(G)\}$ and $H \leq K$, then $K/H \cong \overline{K}/\overline{H}$ which implies $[K:H] = [\overline {K}:\overline{H}]$.
\end{enumerate}
\end{theorem}
\subsection{Cofinality}
Different choices of fundamental systems of neighbourhoods of the identity element may lead to topologically isomorphic profinite completions, where by topological isomorphism is meant an isomorphism that is also a homeomorphism. This can happen because of cofinality. Suppose ${\mathcal N}$ and ${\mathcal M}$ are nonempty filtered from below collections of finite index subgroups of $G$. The collection ${\mathcal M}$ is cofinal in ${\mathcal N}$ if ${\mathcal M} \subset {\mathcal N}$ and for every $N\in {\mathcal N}$ there exists $M\in {\mathcal M}$ such that $N \preceq M$, i.e., $M\subset N$. The following result (Lemma 1.1.9 in \cite{RZ10}) is presented in the notation of Subsection 2.2.
\begin{proposition}\label{topisoiota} For nonempty collections ${\mathcal N}$ and ${\mathcal M}$ of finite index subgroups of $G$ filtered from below, if ${\mathcal M}$ is cofinal in ${\mathcal N}$, then there exists a topological isomorphism $\Theta_{\mathcal{NM}}: {\mathcal K}_{\mathcal N}(G) \to {\mathcal K}_{\mathcal M}(G)$, given explicitly by
\[ \Theta_{ \mathcal{NM}} ( \{ [x_N]_N\}_{N\in{\mathcal N}}) = \{ [x_N]_N\}_{N\in{\mathcal M}},\]
that satisfies $\Theta_{\mathcal {NM}}\iota_{\mathcal N} = \iota_{\mathcal M}$.
\end{proposition}
The topological isomorphism $\Theta_{ {\mathcal N}{\mathcal M}}$ in Proposition \ref{topisoiota} is the inverse of the map constructed in Lemma 1.1.9 in \cite{RZ10}; the map $\Theta_{{\mathcal N}{\mathcal M}}$ ``forgets'' those $[x_N]_N$ in the coherent tuple $\{ [x_N]_N\}_ {N\in{ {\mathcal N}}}$ for $N\in {\mathcal N}\setminus{\mathcal M}$. A direct consequence of Proposition \ref{topisoiota} is that the injectivity of the homomorphisms $\iota_N$ and $\iota_M$ are related when ${\mathcal M}$ is cofinal in ${\mathcal N}$.
\begin{corollary}\label{iotainjective} For nonempty collections ${\mathcal N}$ and ${\mathcal M}$ of finite index subgroups of $G$ filtered from below, suppose ${\mathcal M}$ is cofinal in ${\mathcal N}$. Then $\iota_{\mathcal M}$ is a monomorphism if and only if $\iota_{\mathcal N}$ is a monomorphism.
\end{corollary}
A pro-${\mathcal C}$ topology on $G$ determined by ${\mathcal N}$ is the full pro-${\mathcal C}$ topology whenever ${\mathcal N}$ is cofinal in ${\mathcal N}_{\mathcal C}(G)$. This follows from Proposition \ref{topisoiota} and part (b) of Theorem \ref{fullproCtopology}. We present a proof of this that doesn't pass through the profinite completions.
\begin{proposition}\label{fullprofinite} Let ${\mathcal C}$ be a formation of finite abelian groups. If ${\mathcal N}$ is cofinal in ${\mathcal N}_{\mathcal C}(G)$, then the pro-${\mathcal C}$ topology on $G$ determined by ${\mathcal N}$ is the full pro-${\mathcal C}$ topology on $G$.
\end{proposition}
\begin{proof} By translation invariance it suffices to show for a subset $U$ containing $0$ that $U$ is open with respect to the pro-${\mathcal C}$ topology determined by ${\mathcal N}$ if and only if $U$ is open with respect to the full pro-${\mathcal C}$ topology.
Suppose $U$ is an open subset of $G$ containing $0$ with respect to the pro-${\mathcal C}$ topology determined by ${\mathcal N}$. For each $x\in U$ there exists $N_x\in {\mathcal N}$ such that $x\in x+N_x\subset U$, where $x+N_x$ is the translation of $N_x$ by $x$. Since ${\mathcal N}$ is cofinal in ${\mathcal N}_{\mathcal C}(G)$, we have ${\mathcal N}\subset {\mathcal N}_{\mathcal C}(G)$, so that $N_x\in {\mathcal N}_{\mathcal C}(G)$ for all $x\in U$. Thus for every $x\in U$ there exists $N_x\in {\mathcal N}_{\mathcal C}(G)$ such that $x\in x+N_x\subset U$, and thus $U$ is open in the full pro-${\mathcal C}$ topology.
Now suppose that $U$ is an open subset of $G$ containing $0$ with respect to the full pro-${\mathcal C}$ topology. For each $x\in U$ there exist $F_x\in {\mathcal N}_{\mathcal C}(G)$ such that $x\in x+F_x\subset U$. Since ${\mathcal N}$ is cofinal in ${\mathcal N}_{\mathcal C}(G)$ there exist $N_x\in {\mathcal N}$ such that $0\in N_x\subset F_x$. This implies that $x\in x+N_x\subset x+F_x\subset U$. Thus for every $x\in U$ there exist $N_x\in {\mathcal N}$ such that $x\in x+N_x\subset U$, and thus $U$ is open in the pro-${\mathcal C}$ topology determined by ${\mathcal N}$.
\end{proof}
\begin{remark}\label{applyfullproC} The hypotheses for Theorem \ref{fullproCtopology} are that $G$ is residually ${\mathcal C}$, i.e., the map $\iota_{\mathcal C}:G\to {\mathcal K}_{\mathcal C}(G)$ is a monomorphism, and that the pro-${\mathcal C}$ topology on $G$ is the full pro-${\mathcal C}$ topology, i.e., determined by ${\mathcal N}_{\mathcal C}(G)$. For a pro-${\mathcal C}$ completion ${\mathcal K}_{\mathcal N}(G)$, if ${\mathcal N}$ is cofinal in ${\mathcal N}_{\mathcal C}(G)$, then it follows from Proposition \ref{fullprofinite} that the pro-${\mathcal C}$ topology determined on $G$ by ${\mathcal N}$ is the full pro-${\mathcal C}$ topology, and it follows from Corollary \ref{iotainjective} that $\iota_{\mathcal N}$ is a monomorphism, i.e., $G$ is residually ${\mathcal C}$, hence we can apply Theorem \ref{fullproCtopology} with ${\mathcal K}_{\mathcal C}(G)$ replaced by ${\mathcal K}_{\mathcal N}(G)$, ${\mathcal N}_{\mathcal C}(G)$ replaced with ${\mathcal N}$, and $\iota_{\mathcal C}$ replaced with $\iota_{\mathcal N}$.
\end{remark}
\section{The Profinite Completion of ${\mathbb Z}^n$}
\subsection{The Abelian Group ${\mathbb Z}^n$} We set $G$ to be free abelian group ${\mathbb Z}^n$ of rank $n\geq 2$. We represent the elements of ${\mathbb Z}^n$ as row vectors in terms of the standard basis $\{e_1,e_2,\dots,e_n\}$ where $e_1 = (1,0,\dots,0)$, $e_2=(0,1,\dots,0)$, and $e_n = (0,0,\dots,1)$, so that each $x\in {\mathbb Z}^n$ is expressed as
\[ x = x_1 e_1 + x_2 e_2 + \cdots + x_n e_n\]
for $x_1,x_2,\dots,x_n\in{\mathbb Z}$.
\subsection{The Full Pro-${\mathcal C}$ Topology} For ${\mathcal C}$ fixed as the formation of all finite abelian groups, the full pro-${\mathcal C}$ topology on ${\mathbb Z}^n$ is determined by the nonempty collection of all finite index subgroups of ${\mathbb Z}^n$,
\[ {\mathcal N}_{\mathcal C}({\mathbb Z}^n) = \{N \leq_f {\mathbb Z}^n : {\mathbb Z}^n/N \in {\mathcal C}\} = \{N:N\leq_f {\mathbb Z}^n\}.\]
To simplify notation we denote ${\mathcal N}_{\mathcal C}({\mathbb Z}^n)$ by ${\mathcal F}$. The profinite completion of ${\mathbb Z}^n$ induced by ${\mathcal F}$ will be denoted by ${\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ and the continuous homomorphism ${\mathbb Z}^n \to {\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ will be denoted by $\iota_{\mathcal F}$.
To show that ${\mathbb Z}^n$ is residually ${\mathcal C}$ we use the collection
\[ {\mathcal D} = \{ d{\mathbb Z}^n: d\in {\mathbb N}\}\]
and Corollary \ref{iotainjective}. For each $d\in {\mathbb N}$ the subgroup $d{\mathbb Z}^n$ has the finite index $[{\mathbb Z}^n:d{\mathbb Z}^n] = d^n$. The collection ${\mathcal D}$ is filtered from below because for $d_1,d_2\in {\mathbb Z}$ the choice of $d_3 = {\rm lcm}(d_1,d_2)$ satisfies
\[ d_3{\mathbb Z}^n \subset d_1{\mathbb Z}^n \cap d_2{\mathbb Z}^n,\]
i.e., for $(d_3m_1,d_3m_2,\dots,d_3m_n)\in d_3{\mathbb Z}^n$ the relations $d_1\mid d_3$ and $d_2\mid d_3$ imply that $d_1\mid d_3 m_i$ and $d_2\mid d_3 m_i$ for all $i=1,2,\dots,n$. In proving that ${\mathcal D}$ is cofinal in ${\mathcal F}$ we make use of the result that states for a free abelian group $F$ of finite rank $n$ and a subgroup $H$ of rank $n$ there exists a basis $\{x_1,x_2,\dots,x_n\}$ of $F$ and positive integers $d_1,d_2,\dots,d_n$ satisfying $d_1\mid d_2\mid \cdots \mid d_n$ such that $\{ d_1x_1,d_2x_2,\dots,d_nx_n\}$ is a basis for the subgroup $H$ of $F$ (see Theorem 1.6 in \cite{Hu74}).
\begin{lemma}\label{kproduct} For each $n\geq 2$ the collection ${\mathcal D}$ is cofinal in ${\mathcal F}$.
\end{lemma}
\begin{proof} Clearly ${\mathcal D}\subset {\mathcal F}$. Let $N\in {\mathcal F}$. Since $N\leq_f {\mathbb Z}^n$, there exists a basis $\xi_1,\xi_2,\dots,\xi_n$ of ${\mathbb Z}^n$ and positive integers $d_1,d_2,\dots,d_n$ such that $d_1 \mid d_2 \mid \cdots \mid d_n$ and
\[d_1\xi_1, d_2\xi_2,\dots, d_n\xi_n\]
is a basis of $N$ . Since $d_1\mid d_2\mid \cdots\mid d_n$, all of the elements $d_n\xi_1,d_n\xi_2,\dots,d_n\xi_n$ belong to $N$. Set $d=d_n$. Since $\xi_1,\xi_2,\dots,\xi_n$ is a basis for ${\mathbb Z}^n$, we have
\[ d(m_1\xi_1+m_2\xi_2+\cdots+m_n\xi_n)\in N\]
for all $m_1,m_2,\dots,m_n\in{\mathbb Z}$. This implies that $d{\mathbb Z}^n \subset N$ where $d{\mathbb Z}^n \in {\mathcal D}$. Thus for every $N\in {\mathcal F}$ there exists $d{\mathbb Z}^n\in {\mathcal D}$ such that $N\preceq d{\mathbb Z}^n$, making ${\mathcal D}$ cofinal in ${\mathcal F}$.
\end{proof}
By Lemma \ref{kproduct} and Proposition \ref{fullprofinite}, the cofinal collection ${\mathcal D}$ induces the full pro-${\mathcal C}$ topology on ${\mathbb Z}^n$. Furthermore, by Proposition \ref{topisoiota} the pro-${\mathcal C}$ completion ${\mathcal K}_{\mathcal D}({\mathbb Z}^n)$ is topologically isomorphic to ${\mathcal K}_{\mathcal C}({\mathbb Z}^n)$.
\begin{lemma}\label{residualD} For each $n\geq 2$, the map $\iota_{{\mathcal D}}:{\mathbb Z}^n \to {\mathcal K}_{ {\mathcal D}}({\mathbb Z}^n)$ is a monomorphism.
\end{lemma}
\begin{proof} Suppose
\[ m \in \bigcap_{d\in{\mathbb N}} d{\mathbb Z}^n.\]
Then $m = (m_1,\dots,m_n) \in d{\mathbb Z}^n$ for all $d\in{\mathbb N}$. In particular each $m_i$ is divisible by every $d\in{\mathbb N}$. If $m_i\ne 0$, there is $d > \vert m_i\vert + 1$ for which $m_i$ is not divisible by $d$. This means $m_i=0$, implying that $m=0$.
\end{proof}
\begin{proposition}\label{residual} For each $n\geq 2$, the map $\iota_{{\mathcal F}}:{\mathbb Z}^n \to {\mathcal K}_{{\mathcal F}}({\mathbb Z}^n)$ is a monomorphism.
\end{proposition}
\begin{proof} Since ${\mathcal D}$ is cofinal in ${\mathcal F}$ by Lemma \ref{kproduct} and $\iota_{\mathcal D}$ is a monomorphism by Lemma \ref{residualD}, it follows by Corollary \ref{iotainjective} that $\iota_{\mathcal F}$ is a monomorphism.
\end{proof}
\begin{corollary}\label{Znresfin} The pro-${\mathcal C}$ group ${\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ is residually ${\mathcal C}$.
\end{corollary}
\section{Strong BF-Equivalence}
\subsection{Dynamically Defined Pro-${\mathcal C}$ Topologies}
For similar hyperbolic $A,B\in{\rm GL}_n({\mathbb Z})$, all of the generalized Bowen-Franks $R$-modules, viewed as groups, naturally correspond to a pair of pro-${\mathcal C}$ topologies on ${\mathbb Z}^n$. This pair of pro-${\mathcal C}$ topologies is parameterized by the collection of polynomials
\[ {\mathcal Q}^{\rm a} = \{ g\in {\mathbb Z}[t] : {\rm det}(g(A))\ne 0\}.\]
The invertibility of $g(B)$ follows from the invertibility of $g(A)$ and the similarity of $A$ with $B$, and so the invertibility of $g(B)$ is not explicitly stated in the definition of ${\mathcal Q}^{\rm a}$. Set $N_{g,A} = {\mathbb Z}^n g(A)$, $N_{g,B} = {\mathbb Z}^n g(B)$, ${\mathcal N}_{ {\mathcal Q}^{\rm a},A} = \{ N_{g,A}: g\in {\mathcal Q}^{\rm a} \}$, and ${\mathcal N}_{ {\mathcal Q}^{\rm a},B} = \{ N_{g,B}: g\in {\mathcal Q}^{\rm a}\}$.
\begin{lemma}\label{Qall} For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, the collections ${\mathcal N}_{ {\mathcal Q}^{\rm a},A}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm a},B}$ are pro-${\mathcal C}$ topologies on ${\mathbb Z}^n$.
\end{lemma}
\begin{proof} For each $g\in {\mathcal Q}^{\rm a}$, the invertibility of $g(A),g(B)\in M_n({\mathbb Z})$ implies by their Smith normal form that $N_{g,A}$ and $N_{g,B}$ are finite index subgroups of ${\mathbb Z}^n$. The filtered from below property follows from ${\mathcal Q}^{\rm a}$ being a multiplicative set, i.e., closed under multiplication. For $g_1,g_2\in {\mathcal Q}^{\rm a}$ there holds $g_1,g_2\in {\mathbb Z}[t]$ with $g_1(A)$ and $g_2(A)$ invertible, so that setting $g_3 = g_1g_2$ there holds $g_3\in {\mathbb Z}[t]$ with $g_3(A)$ invertible, hence $g_3\in {\mathcal Q}^{\rm a}$. If $g_1,g_2\in {\mathcal Q}^{\rm a}$ then for $g_3 = g_1g_2\in {\mathcal Q}^{\rm a}$ and $m\in {\mathbb Z}^n g_3(A)$ there is $z\in {\mathbb Z}^n$ such that $m = z g_3(A)$, hence there holds $m = (zg_2(A))g_1(A)$ and $m = (zg_1(A))g_2(A)$, implying that $N_{g_3,A} \subset N_{g_1,A}\cap N_{g_2,A}$. A similar argument gives $N_{g_3,B}\subset N_{g_1,B} \cap N_{g_2,B}$.
\end{proof}
For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, the principal generalized Bowen-Franks $R$-modules, viewed as groups, naturally correspond to another pair of pro-${\mathcal C}$ topologies on ${\mathbb Z}^n$. This pair of pro-${\mathcal C}$ topologies is parameterized by the collection of polynomials
\[ {\mathcal Q}^{\rm p} = \{ g_k(t) = t^k-1: k\in {\mathbb N}\}.\]
To simplify notation, set $N_{k,A} = N_{g_k,A}$ and $N_{k,B} = N_{g_k,B}$. Then ${\mathcal N}_{ {\mathcal Q}^{\rm p},A} = \{ N_{k,A}: k\in {\mathbb N}\}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm p},B} = \{ N_{k,B} : k\in {\mathbb N}\}$.
\begin{lemma}\label{Qprin} For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, there holds ${\mathcal Q}^{\rm p} \subset {\mathcal Q}^{\rm a}$, and the collections ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm p},B}$ are pro-${\mathcal C}$ topologies on ${\mathbb Z}^n$.
\end{lemma}
\begin{proof} For each $k\in {\mathbb N}$ there holds $g_k\in{\mathbb Z}[t]$ and the hyperbolicity of $A$ implies that $g_k(A)$ is invertible. Thus ${\mathcal Q}^{\rm p} \subset {\mathcal Q}^{\rm a}$. For each $k\in {\mathbb N}$, the similarity of $A$ and $B$ implies the invertibility of $g_k(B)$. By the Smith normal forms of $g(A)$ and $g(B)$, the subgroups $N_{k,A}$ and $N_{k,B}$ have finite index in ${\mathbb Z}^n$. The filtered from below property for ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ follows by (Lemma 2.2 in \cite{BM12}) wherein we showed that for $k_1,k_2\in {\mathbb N}$, the choice of $k_3 = k_1k_2$ leads to $g_{k_3} = q_1g_{k_1}$ and $g_{k_3} = q_2g_{k_2}$ for $q_1,q_2\in {\mathbb Z}[t]$, so that for $m\in {\mathbb Z}^n g_{k_3}(A)$ there exists $z\in {\mathbb Z}^n$ such that $m = zg_{k_3}(A)$, whence $m = zq_1(A)g_{k_1}(A) \in {\mathbb Z}^n g_{k_1}(A)$ and $m = z q_2(A)g_{k_2}(A) \in {\mathbb Z}^n g_{k_2}(A)$, implying that $N_{k_3,A} \subset N_{k_1,A}\cap N_{k_2,A}$. A similar argument gives $N_{k_3,B} \subset N_{k_1,B}\cap N_{k_2,B}$.
\end{proof}
For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, we show that ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm p},B}$ are each cofinal in ${\mathcal F}$, that ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ is cofinal in ${\mathcal N}_{ {\mathcal Q}^{\rm a},A}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm p},B}$ is cofinal in ${\mathcal N}_{ {\mathcal Q}^{\rm a},B}$. We only state the results for $A$, as those for $B$ are obtained by replacing $A$ with $B$.
\begin{lemma}\label{dksubset} Suppose $A\in {\rm GL}_n({\mathbb Z})$ is hyperbolic. For every $d\in {\mathbb N}$ there exists $k\in{\mathbb N}$ such that $N_{k,A}\subset d{\mathbb Z}^n$.
\end{lemma}
\begin{proof} For fixed but arbitrary $d\in{\mathbb N}$, consider the finite subset of ${\mathbb T}^n$ given by
\[ T_d = \{ x\in {\mathbb T}^n: dx = 0\}.\]
This finite set consists entirely of periodic points of the left action of the hyperbolic toral automorphism $A$ of ${\mathbb T}^n$. Let $k$ be the least common multiple of the periods of $x\in T_d$ under the left action of $A$. Then $(A^k-I)x = 0$ for all $x\in T_d$, so that $T_d\subset {\rm Per}_k(A)$. Each rational point of the form $(0,\dots,0,1/d,0,\dots,0)^{\rm T}$ belongs to $T_d$ and satisfies
\[ (A^k-I)(0,\dots,0,1/d,0,\dots,0)^{\rm T} = 0\]
because $T_d\subset {\rm Per}_k(A)$. This implies that each entry of $A^k-I$ is divisible by $d$. Thus ${\mathbb Z}^n(A^k-I)\subset d{\mathbb Z}^n$, i.e., $N_{k,A}\subset d{\mathbb Z}^n$.
\end{proof}
\begin{proposition}\label{cofinal} Suppose $A\in {\rm GL}_n({\mathbb Z})$ is hyperbolic. Then ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ is cofinal in ${\mathcal N}_{ {\mathcal Q}^{\rm a},A}$, and ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ is cofinal in ${\mathcal F}$.
\end{proposition}
\begin{proof} By Lemma \ref{Qprin} there holds ${\mathcal Q}^{\rm p}\subset {\mathcal Q}^{\rm a}$ which implies ${\mathcal N}_{{\mathcal Q}^{\rm p},A}\subset {\mathcal N}_{ {\mathcal Q}^{\rm a},A}$. Furthermore, by Lemma \ref{Qprin} there holds ${\mathcal N}_{{\mathcal Q}^{\rm p},A} \subset {\mathcal F}$. For $N\in {\mathcal N}_{ {\mathcal Q}^{\rm a},A}$ there holds $N\leq_f {\mathbb Z}^n$ by Lemma \ref{Qall}. For $N\in {\mathcal F}$ there holds $N\leq_f {\mathbb Z}^n$. In either case we have $N\leq_f {\mathbb Z}^n$. By Lemma \ref{kproduct} there is a $d\in{\mathbb N}$ such that $d{\mathbb Z}^n\subset N$. Associated to this $d\in {\mathbb N}$ there is by Lemma \ref{dksubset} a $k\in{\mathbb N}$ such that $N_{k,A}\subset d{\mathbb Z}^n$. Hence $N_{k,A} \subset d{\mathbb Z}^n \subset N$, i.e., $N\preceq N_{k,A}$.
\end{proof}
\begin{corollary}\label{GAisoZnbar} If $A\in {\rm GL}_n({\mathbb Z})$ is hyperbolic, then
\[ {\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm p},A} }({\mathbb Z}^n) = \varprojlim_{ k\in {\mathbb N}} {\mathbb Z}^n/N_{k,A}\]
is topologically isomorphic to ${\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ and to
\[ {\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},A}} ( {\mathbb Z}^n) = \varprojlim_{g\in {\mathcal Q}^{\rm a}} {\mathbb Z}^n/ N_{g,A}.\]
\end{corollary}
\begin{proof} Apply Proposition \ref{cofinal} and Proposition \ref{topisoiota}.
\end{proof}
By Corollary \ref{GAisoZnbar}, applied to similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, there are topological isomorphisms
\[ {\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},A}}({\mathbb Z}^n) \cong {\mathcal K}_{ {\mathcal N}_ { {\mathcal Q}^{\rm p},A}}({\mathbb Z}^n) \cong {\mathcal K}_{\mathcal F}({\mathbb Z}^n) \cong {\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm p}, B}}({\mathbb Z}^n) \cong {\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},B} }({\mathbb Z}^n).\]
Thus, for similar hyperbolic $A,B \in {\rm GL}_n({\mathbb Z})$, the pro-${\mathcal C}$ completions ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},A}}({\mathbb Z}^n)$, ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm p},A}}({\mathbb Z}^n)$, ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm p},B}}({\mathbb Z}^n)$, and ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},B}}({\mathbb Z}^n)$, are indistinguishable as topological abelian groups. We will show that what does distinguish ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm a},A}}({\mathbb Z}^n)$ and ${\mathcal K}_{ {\mathcal N}_{{\mathcal Q}^{\rm p},A}}({\mathbb Z}^n)$ from ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^{\rm p},B}}({\mathbb Z}^n)$ and ${\mathcal K}_{ {\mathcal N}_{ {\mathcal Q}^ {\rm a},B}}({\mathbb Z}^n)$ is an $R$-module structure on these pro-${\mathcal C}$ completions.
\subsection{Module Structure}\label{sec:Module}
For similar hyperbolic $A,B \in {\rm GL}_n({\mathbb Z})$, there is a commutative ring $R$ with identity to which both matrix rings ${\mathbb Z}[A]$ and ${\mathbb Z}[B]$ are isomorphic. When the common characteristic polynomial $f(t)$ of $A$ and $B$ is irreducible the ring $R$ is the integral domain given by quotient
\[ R = {\mathbb Z}[t] / \big(f(t)\big)\]
where $\big( f(t) \big) = f(t){\mathbb Z}[t]$ is the principal ideal generated by $f(t)$. The map $\vartheta_A: {\mathbb Z}[t]/\big( f(t)\big) \to {\mathbb Z}[A]$ given by $q(t) + \big(f(t)\big) \to q(A)$, where
\[ {\mathbb Z}[A] = \left\{ \sum_{i=0}^{n-1} c_i A^i : c_i\in {\mathbb Z}\right\},\]
is a ring isomorphism, and the similarly defined map $\vartheta_B:{\mathbb Z}[t]/\big(f(t)\big) \to {\mathbb Z}[B]$ is also a ring isomorphism. When $f(t)$ is irreducible, this gives ${\mathbb Z}[A] \cong R \cong {\mathbb Z}[B]$ as rings. When $f(t)$ is reducible, the map $\vartheta_A$ is a ring epimorphism, and so by the First Isomorphism Theorem for Rings it induces a ring isomorphism
\[ \bar{\vartheta}_A: \big({\mathbb Z}[t]/\big( f(t)\big)\big) / {\rm ker}(\vartheta_A) \to {\mathbb Z}[A].\]
Similarly $\vartheta_B$ is a ring epimorphism, and so it induces a ring isomorphism
\[ \bar{\vartheta}_B: \big({\mathbb Z}[t]/\big( f(t)\big)\big) / {\rm ker}\vartheta_B \to {\mathbb Z}[B].\]
Similarity of $A$ and $B$ implies that ${\rm ker}(\vartheta_A) = {\rm ker}(\vartheta_B)$. Thus it follows that
\[ {\mathbb Z}[A] \cong \big( {\mathbb Z}[t]/\big(f(t)\big)/{\rm ker}(\vartheta_A) = \big( {\mathbb Z}[t]/\big(f(t)\big)/{\rm ker}(\vartheta_B) \cong {\mathbb Z}[B].\]
When $f(t)$ is reducible, the ring $R$ to which ${\mathbb Z}[A]$ and ${\mathbb Z}[B]$ are both isomorphic is $\big( {\mathbb Z}[t]/\big(f(t)\big)/{\rm ker}(\vartheta_A)$, which is an integral domain when the minimal polynomial of the similar $A$ and $B$ is irreducible, and is not an integral domain when the minimal polynomial of the similar $A$ and $B$ is reducible. With ${\mathcal Q} = {\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, when we refer to the $R$-module structure it is understood that it is ${\mathbb Z}[A]$ on ${\mathcal N}_{ {\mathcal Q},A}$, ${\mathcal K}_{ {\mathcal N}_{{\mathcal Q}, A}}({\mathbb Z}^n)$, etc., and it is ${\mathbb Z}[B]$ on ${\mathcal N}_{ {\mathcal Q},B}$, ${\mathcal K}_{ {{\mathcal N}_{\mathcal Q},B}}({\mathbb Z}^n)$, etc.
For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ and ${\mathcal Q}$ equal to either ${\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, we describe the $R$-module structure on the pro-${\mathcal C}$ topology ${\mathcal N}_{{\mathcal Q},A} = \{ N_{g,A} : g\in {\mathcal Q}\}$, the pro-${\mathcal C}$ completion ${\mathcal K}_{ {\mathcal N}_ {{\mathcal Q},A}}( {\mathbb Z}^n)$, and the several natural homomorphisms associated to the pro-${\mathcal C}$ completion. (The $R$-module structure for hyperbolic $B$ similar to $A$ is described in like manner.) The ring $R\cong{\mathbb Z}[A]$ acts on ${\mathbb Z}^n$ via the automorphism $m\to mA$ for $m\in{\mathbb Z}^n$, making ${\mathbb Z}^n$ into a right ${\mathbb Z}[A]$-module. Since each $N_{g,A} \in {\mathcal N}_{ {\mathcal Q},A}$ is invariant under the right action of $A$, the subgroup $N_{g,A}$ is a right ${\mathbb Z}[A]$-module. Denoting the elements
\[ G_{g,A} = {\mathbb Z}^n/N_{g,A}\]
by $[m]_{g,A} = m + N_{g,A}$ for $m\in {\mathbb Z}^n$, the right ${\mathbb Z}[A]$-module structure on $G_{g,A}$ is induced by the topological automorphism $A_g$ of $G_{g,A}$ defined by
\[ A_g( [m]_{g,A}) = [mA]_{g,A}.\]
For $h\preceq g$ the transition maps are the continuous epimorphisms $\varphi_{g,h,A}:G_{g,A}\to G_{h,A}$ defined by
\[ \varphi_{g,h,A}([m]_{g,A}) = [m]_{h,A}\]
which are ${\mathbb Z}[A]$-module epimorphisms. The profinite completion of ${\mathbb Z}^n$ with respect to the pro-${\mathcal C}$ topology determined by ${\mathcal N}_{{\mathcal Q},A}$ we denote by
\[ G_{ {\mathcal Q},A} = \varprojlim_{g\in{\mathcal Q}} G_{g,A}\]
instead of ${\mathcal K}_{\mathcal N_{ {\mathcal Q},A}}({\mathbb Z}^n)$ for simplicity of notation.
The elements of $G_{ {\mathcal Q},A}$ consist of those coherent tuples $\{ [m_g]_{g,A}\}_{g\in{\mathcal Q}}$ lying in $\prod_{g\in{\mathcal Q}} G_{g,A}$ (where $m_g\in {\mathbb Z}^n$ for each $g\in {\mathcal Q}$), i.e., those tuples that satisfy
\[ [m_h]_{h,A} = \varphi_{g,h,A}( [m_g]_{g,A}) = [m_g]_{h,A}\]
whenever $h\preceq g$. The right ${\mathbb Z}[A]$-module structure on $G_{ {\mathcal Q},A}$ is induced by the topological automorphism $\Gamma_A$ of $G_{ {\mathcal Q},A}$ defined by
\[ \Gamma_A ( \{ [m_g]_{g,A}\}_{g\in{\mathcal Q}}) = \{ A_g ([m_g]_{g,A})\}_{g\in {\mathcal Q}}.\]
For each $g\in{\mathcal Q}$, the continuous epimorphism $\varphi_{g,A}:G_{ {\mathcal Q},A}\to G_{g,A}$ defined by
\[ \varphi_{g,A}( \{ [m_h]_{h,A}\}_{h\in{\mathcal Q}}) = [m_g]_{g,A}\]
is a ${\mathbb Z}[A]$-module epimorphism. By $R$-module isomorphism theory, each epimorphism $\varphi_{g,A}$ induces the canonical $R$-module isomorphism
\[ G_{{\mathcal Q},A}/{\rm ker} \varphi_{g,A} \cong_R G_{g,A}.\]
The continuous homomorphism $\iota_{ {\mathcal Q},A}:{\mathbb Z}^n \to G_{ {\mathcal Q},A}$ (we use $\iota_{{\mathcal Q},A}$ instead of $\iota_{ {\mathcal N}_{ {\mathcal Q},A}}$ for simplicity of notation) defined by
\[ \iota_{ {\mathcal Q},A}(m) = \{ [m]_{g,A}\}_{g\in{\mathcal Q}}\]
is also a ${\mathbb Z}[A]$-module homomorphism. The map $\iota_{ {\mathcal Q},A}$ is a monomorphism for both ${\mathcal Q} = {\mathcal Q}^{\rm p}$ and ${\mathcal Q} = {\mathcal Q}^{\rm a}$ by Corollary \ref{iotainjective} because ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ is cofinal in ${\mathcal F}$ and ${\mathcal N}_{ {\mathcal Q}^{\rm p},A}$ is cofinal in ${\mathcal N}_{ {\mathcal Q}^{\rm a},A}$ by Proposition \ref{cofinal}, and $\iota_{\mathcal F}$ is a monomorphism by Proposition \ref{residual}. We proved directly that hyperbolicity of $A$ implies the intersection of $N_{k,A}$, $k\in {\mathbb N}$, is $\{0\}$ (see Lemma 2.2 in \cite{BM12}), hence, since ${\mathcal N}_{ {\mathcal Q}^{\rm p},A} \subset {\mathcal N}_{ {\mathcal Q}^{\rm a},A}$ by Proposition \ref{cofinal}, that
\[ \bigcap_{k\in{\mathbb N} } N_{k,A} = \{0\} \ \Rightarrow\ \bigcap_{ g\in {\mathcal Q}^{\rm a}} N_{g,A} = \{0\},\]
also showing that $\iota_{ {\mathcal Q},A}$ is a monomorphism for both ${\mathcal Q} = {\mathcal Q}^{\rm p}$ and ${\mathcal Q} = {\mathcal Q}^{\rm a}$.
\subsection{Profinite Conjugacy}
\begin{definition} For ${\mathcal Q}$ being either ${\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, two similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ are profinitely conjugate over ${\mathcal Q}$ if there exists a topological $R$-module isomorphism $\Psi_{\mathcal Q}:G_{ {\mathcal Q},A}\to G_{ {\mathcal Q},B}$.
\end{definition}
Profinite conjugacy over ${\mathcal Q}$ for similar hyperbolic toral automorphisms is an equivalence relation, as is readily shown.
\begin{remark}\label{sufficient} We showed that strong {\rm BF}-equivalence over ${\mathcal Q}^{\rm p}$ implies that $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm p},B}$ are topologically $R$-module isomorphic (see Theorem 4.1 in \cite{BM12}) where we used the $R$-module version of Theorem 3.27 in \cite{RZ10} (up to $R$-module isomorphism, $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm p},B}$ have the same finite $R$-module quotients, namely ${\rm BF}_g(A) \cong_R {\rm BF}_g(B)$ for all $g\in {\mathcal Q}^{\rm p}$). The same argument shows that strong {\rm BF}-equivalence over ${\mathcal Q}^{\rm a}$ implies that $G_{ {\mathcal Q}^{\rm a},A}$ and $G_{ {\mathcal Q}^{\rm a},B}$ are topologically $R$-module isomorphic. Thus, for ${\mathcal Q} = {\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, a sufficient condition for when two similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ are profinitely conjugate over ${\mathcal Q}$, and hence in the same equivalence class, is strong {\rm BF}-equivalence of $A$ and $B$ over ${\mathcal Q}$, i.e., ${\rm BF}_g(A) \cong_R {\rm BF}_g(B)$ for all $g\in {\mathcal Q}$.
\end{remark}
\begin{lemma}\label{QminQ} For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ the profinite $R$-modules $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{{\mathcal Q}^{\rm a},A}$ are topologically $R$-module isomorphic.
\end{lemma}
\begin{proof} By Corollary \ref{GAisoZnbar} the profinite $R$-modules $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm a},A}$ are topologically isomorphic as topological abelian groups. The proof of Corollary \ref{GAisoZnbar} uses Proposition \ref{topisoiota} where an explicit topological isomorphism is given. It is straightforward to verify that this explicitly given topological isomorphism is an $R$-module map, making $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm a},A}$ topologically $R$-module isomorphic.
\end{proof}
For ${\mathcal Q}$ being either ${\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, the pro-${\mathcal C}$ completions $G_{{\mathcal Q},A}$ and $G_{ {\mathcal Q},B}$ have, of course, the closest connection with the generalized Bowen-Franks $R$-modules ${\rm BF}_g(A)$ and ${\rm BF}_g(B)$ for $g\in {\mathcal Q}$. This is because
\[ {\rm BF}_g(A) = {\mathbb Z}^n/{\mathbb Z}^n g(A) = G_{g,A} \cong_R G_A/{\rm ker}\varphi_{g,A},\ g\in {\mathcal Q},\]
with a similar statement holding with $A$ replaced by $B$. The pro-${\mathcal C}$ completions $G_{{\mathcal Q},A}$ and $G_{ {\mathcal Q},B}$ are both topologically isomorphic to the full pro-${\mathcal C}$ completion ${\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ by Corollary \ref{GAisoZnbar}, but not every finite index subgroup of ${\mathbb Z}^n$ that is used in the construction of ${\mathcal K}_{\mathcal F}({\mathbb Z}^n)$ is invariant by the corresponding right action of $A$ or $B$. Hence we use the dynamically defined $R$-modules $G_{{\mathcal Q},A}$ and $G_{ {\mathcal Q},B}$ to prove that profinite conjugacy over ${\mathcal Q}$ of similar hyperbolic $A$ and $B$ implies strong {\rm BF}-equivalence of $A$ and $B$ over ${\mathcal Q}$. We make use of the one-to-one correspondence in Theorem \ref{fullproCtopology}, via Remark \ref{applyfullproC}, from $\{ U: U\leq_o {\mathbb Z}^n; {\mathcal N}_{ {\mathcal Q},A}\}$ to $\{ V: V\leq_o G_{ {\mathcal Q},A}\}$ and its properties without explicit reference in what follows.
\subsection{The Converse} The main result is proven in this subsection. The idea of the proof of the main result is to mimic the quotient of the generalized Bowen-Franks $R$-modules in the profinite $R$-module setting and connect it with generalized Bowen-Franks $R$-modules through the isomorphism theory for $R$-modules. The first step in this direction is to show through a series of four lemmas that, for a pro-${\mathcal C}$ topology ${\mathcal N}_{ {\mathcal Q},A}$ on ${\mathbb Z}^n$ with ${\mathcal Q}$ being either ${\mathcal Q}^{\rm p}$ or ${\mathcal Q}^{\rm a}$, the closure of the image $g(\Gamma_A)(G_{ {\mathcal Q},A})$ is equal to ${\rm ker}\varphi_{g,A}$ for each $g\in {\mathcal Q}$.
\begin{lemma}\label{intersection} For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ and $g\in{\mathcal Q}$, there holds
\[\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))\subset g(\Gamma_A) (G_{ {\mathcal Q},A})\cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n).\]
\end{lemma}
\begin{proof} Suppose $x \in \iota_{ {\mathcal Q},A}( {\mathbb Z}^n g(A))$. Then $x=\iota_{ {\mathcal Q},A}(m g(A))$ for some $m\in{\mathbb Z}^n$. Hence $x = g(\Gamma_A)(\iota_{ {\mathcal Q},A}(m)) \in g(\Gamma_A)(G_{ {\mathcal Q},A})$. Since $mg(A)\in{\mathbb Z}^n$, we have $x \in \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$, and so
\[ x \in g(\Gamma_A)(G_{ {\mathcal Q},A}) \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n).\]
This implies that $\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A) )\subset g(\Gamma_A)(G_{ {\mathcal Q},A}) \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$.
\end{proof}
\begin{lemma}\label{contains} For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ and $g\in{\mathcal Q}$, there holds
\[ g(\Gamma_A)(G_{ {\mathcal Q},A}) \subset {\rm ker}\varphi_{g,A}.\]
\end{lemma}
\begin{proof} For $\{ [m_h]_{h,A}\}_{h\in{\mathcal Q}}\in G_{ {\mathcal Q},A}$, we have
\[ g(\Gamma_A)(\{ [m_h]_{h,A}\}_{h\in{\mathcal Q}}) = \{ [m_h g(A)]_{h,A}\}_{h\in{\mathcal Q}}.\]
Then
\[ \varphi_{g,A}( g(\Gamma_A) (\{ [m_h]_{h,A}\}_{h\in{\mathcal Q}})) = [m_g g(A)]_{g,A} = 0\]
because $m_g g(A)\in {\mathbb Z}^n g(A)$. Thus $g(\Gamma_A)(G_{ {\mathcal Q},A}) \subset {\rm ker}\varphi_{g,A}$.
\end{proof}
\begin{lemma}\label{kernel} For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ and $g\in{\mathcal Q}$, there holds
\[ {\rm ker}\varphi_{g,A} \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n) = \iota_{ {\mathcal Q},A}( {\mathbb Z}^n g(A)).\]
\end{lemma}
\begin{proof} Suppose $x\in \iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))$. Then $x=\iota_{ {\mathcal Q},A}(m g(A))$ for some $m\in{\mathbb Z}^n$. Hence $\iota_{ {\mathcal Q},A}(m g(A))\in \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$, and
\[ \varphi_{g,A} (x) = [mg(A)]_{g,A} = 0\]
because $m g(A)\in {\mathbb Z}^n g(A)$. Thus $\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))\subset {\rm ker}\varphi_{g,A}\cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$.
Now suppose that $x \in {\rm ker}\varphi_{g,A} \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$. Then $x=\iota_{ {\mathcal Q},A}(m)$ for some $m\in{\mathbb Z}^n$ and
\[ 0 = \varphi_{g,A} (\iota_{ {\mathcal Q},A}(m)) = [m]_{g,A}.\]
This says that $m\in {\mathbb Z}^n g(A)$, so that $x = \iota_{ {\mathcal Q},A}(m) \in \iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))$. This implies that ${\rm ker}\varphi_{g,A}\cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n) \subset \iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))$.
Therefore ${\rm ker}\varphi_{g,A} \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n) = \iota_{ {\mathcal Q},A}( {\mathbb Z}^n g(A))$.
\end{proof}
\begin{lemma}\label{open} For a hyperbolic $A\in {\rm GL}_n({\mathbb Z})$ and $g\in{\mathcal Q}$, there holds
\[ \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) } = {\rm ker}\varphi_{g,A}.\]
\end{lemma}
\begin{proof} By Lemma \ref{contains}, we have $\overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) } \subset {\rm ker}\varphi_{g,A}$ because ${\rm ker}\varphi_{g,A}$ is closed.
On the other hand, by Lemma \ref{kernel}, we have ${\rm ker}\varphi_{g,A} \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n) = \iota_{ {\mathcal Q},A}( {\mathbb Z}^n g(A))$. Since ${\rm ker}\varphi_{g,A} \leq_o G_{ {\mathcal Q},A}$, we have that
\[ \overline{\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A))} = \overline{ {\rm ker}\varphi_{g,A}\cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n)}= {\rm ker}\varphi_{g,A}.\]
By Lemma \ref{intersection}, we have $\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A)) \subset g(\Gamma_A)(G_{ {\mathcal Q},A})\cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n)$, so that
\begin{align*} \overline{\iota_{ {\mathcal Q},A}({\mathbb Z}^n g(A) )}
& \subset \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) \cap \iota_{ {\mathcal Q},A}({\mathbb Z}^n) } \\
& \subset \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) } \cap \overline{\iota_{ {\mathcal Q},A}({\mathbb Z}^n)} \\
& = \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) } \cap G_{ {\mathcal Q},A} \\
& = \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A})}.
\end{align*}
This gives ${\rm ker}\varphi_{g,A} \subset \overline{ g(\Gamma_A)(G_{ {\mathcal Q},A})}$.
Therefore $\overline{ g(\Gamma_A)(G_{ {\mathcal Q},A})} = {\rm ker}\varphi_{g,A}$.
\end{proof}
For any profinite conjugacy $\Psi_{\mathcal Q}:G_{ {\mathcal Q},A}\to G_{ {\mathcal Q},B}$ over ${\mathcal Q}$ there holds
\[ g(\Gamma_B)\Psi_{\mathcal Q} = \Psi_{\mathcal Q}g(\Gamma_A) {\rm\ for\ all\ }g\in{\mathcal Q},\]
because $\Psi_{\mathcal Q}$ is an $R$-module isomorphism, i.e., $ \Gamma_B\Psi_{\mathcal Q} = \Psi_{\mathcal Q}\Gamma_A $.
\begin{lemma}\label{congkpower} For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, if $\Psi_{\mathcal Q}:G_{{\mathcal Q},A}\to G_{ {\mathcal Q},B}$ is a profinite conjugacy over ${\mathcal Q}$, then for all $g\in{\mathcal Q}$ there holds
\[ \Psi_{\mathcal Q}\big( g(\Gamma_A)(G_{ {\mathcal Q},A}) \big) = g(\Gamma_B)(G_{ {\mathcal Q},B}) .\]
\end{lemma}
\begin{proof} Let $\Psi_{\mathcal Q}:G_{ {\mathcal Q},A}\to G_{ {\mathcal Q},B}$ be a topological $R$-module isomorphism.
Suppose $x\in g(\Gamma_A)(G_{ {\mathcal Q},A})$. Then there exists $y\in G_{ {\mathcal Q},A}$ such that $x = g(\Gamma_A)(y)$. Set $z= \Psi_{\mathcal Q}(y) \in G_{ {\mathcal Q},B}$. Then we have
\[ \Psi_{\mathcal Q}(x) = \Psi_{\mathcal Q} g(\Gamma_A) (y) = g(\Gamma_B)\Psi_{\mathcal Q} (y) = g(\Gamma_B)(z). \]
Thus $\Psi_{\mathcal Q}(x) \in g(\Gamma_B)(G_{ {\mathcal Q},B}) $, and so $\Psi_{\mathcal Q}( g(\Gamma_A)(G_{ {\mathcal Q},A}) ) \subset g(\Gamma_B) (G_{ {\mathcal Q},B}) $.
Now let $x\in g(\Gamma_B)(G_{ {\mathcal Q},B})$. Then there is $y\in G_{ {\mathcal Q},B}$ such that $x = g(\Gamma_B)(y)$. Because $\Psi_{\mathcal Q}$ is a bijection, there exists a unique $z\in G_{ {\mathcal Q},A}$ such that $\Psi_{\mathcal Q}(z) = y$. Set $w = g(\Gamma_A)(z)$. Then we have
\[ \Psi_{\mathcal Q} (w)= \Psi_{\mathcal Q}\big(g(\Gamma_A)(z) \big) = g(\Gamma_B)\Psi_{\mathcal Q} (z) = g(\Gamma_B)(y) = x.\]
This gives $x\in \Psi_{\mathcal Q}( g(\Gamma_A)(G_{ {\mathcal Q},A}) )$, and so $g(\Gamma_B)( G_{ {\mathcal Q},B}) \subset \Psi_{\mathcal Q} ( g(\Gamma_A)(G_{ {\mathcal Q},A}) )$.
Therefore $\Psi_{\mathcal Q}( g(\Gamma_A)(G_{ {\mathcal Q},A}) ) = g(\Gamma_B)(G_{ {\mathcal Q},B}) $.
\end{proof}
\begin{theorem}\label{main} Suppose that $A,B\in {\rm GL}_n({\mathbb Z})$ are similar and hyperbolic. If $A$ and $B$ are profinitely conjugate over ${\mathcal Q}$, then $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}$.
\end{theorem}
\begin{proof} Suppose $\Psi_{\mathcal Q}:G_{ {\mathcal Q},A}\to G_{ {\mathcal Q},B}$ is a profinite conjugacy over ${\mathcal Q}$. By Lemma \ref{congkpower}, we have for each $g\in{\mathcal Q}$ that $\Psi_{\mathcal Q}( g(\Gamma_A)(G_{ {\mathcal Q},A})) = g(\Gamma_B)(G_{ {\mathcal Q},B})$. Since $\Psi_{\mathcal Q}$ is a homeomorphism, we have for each $g\in {\mathcal Q}$ that
\[ \Psi_{\mathcal Q}\left(\overline{g(\Gamma_A)(G_{ {\mathcal Q},A})}\right) = \overline{ g(\Gamma_B)(G_{ {\mathcal Q},B}) }.\]
By Lemma \ref{open} for each $g\in {\mathcal Q}$ there holds $\overline{ g(\Gamma_A)(G_{ {\mathcal Q},A}) } = {\rm ker}\varphi_{g,A}$ and $\overline{ g(\Gamma_B)(G_{ {\mathcal Q},B}) } = {\rm ker}\varphi_{g,B}$, and so
\[ \Psi_{\mathcal Q}( {\rm ker}\varphi_{g,A}) = {\rm ker}\varphi_{g,B}.\]
Thus, by the isomorphism theory for $R$-modules, $\Psi_{\mathcal Q}$ induces an $R$-module isomorphism $G_{ {\mathcal Q},A}/{\rm ker}\varphi_{g,A} \cong_R G_{ {\mathcal Q},B}/{\rm ker}\varphi_{g,B}$ for each $g\in {\mathcal Q}$. Since $G_{g,A} \cong_R G_{ {\mathcal Q},A}/{\rm ker}\varphi_{g,A}$ and $G_{g,B}\cong_R G_{ {\mathcal Q},B}/{\rm ker}\varphi_{g,B}$, we obtain $G_{g,A} \cong_R G_{g,B}$ for all $g\in{\mathcal Q}$. Thus $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}$.
\end{proof}
\begin{corollary}\label{characterization} Suppose $A,B\in {\rm GL}_n({\mathbb Z})$ are similar and hyperbolic. Then $A$ and $B$ are profinitely conjugate over ${\mathcal Q}$ if and only if $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}$.
\end{corollary}
\begin{proof} Combine Theorem \ref{main} and Remark \ref{sufficient}.
\end{proof}
\begin{corollary}\label{prinall} Suppose $A,B\in {\rm GL}_n({\mathbb Z})$ are similar and hyperbolic. The following are equivalent.
\begin{enumerate}
\item[(a)] $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}^{\rm p}$.
\item[(b)] $A$ and $B$ are profinitely conjugate over ${\mathcal Q}^{\rm p}$.
\item[(c)] $A$ and $B$ are profinitely conjugate over ${\mathcal Q}^{\rm a}$.
\item[(d)] $A$ and $B$ are strongly {\rm BF}-equivalent over ${\mathcal Q}^{\rm a}$.
\end{enumerate}
\end{corollary}
\begin{proof} Parts (a) and (b) are equivalent by Corollary \ref{characterization} applied when ${\mathcal Q} = {\mathcal Q}^{\rm p}$ and parts (c) and (d) are equivalent by Corollary \ref{characterization} applied when ${\mathcal Q} = {\mathcal Q}^{\rm a}$. By Lemma \ref{QminQ} the profinite $R$-modules $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm a},A}$ are topologically $R$-module isomorphic and the profinite $R$-modules $G_{ {\mathcal Q}^{\rm p},B}$ and $G_{ {\mathcal Q}^{\rm a},B}$ are topologically $R$-module isomorphic. Hence $G_{ {\mathcal Q}^{\rm p},A}$ and $G_{ {\mathcal Q}^{\rm p},B}$ are profinite conjugate over ${\mathcal Q}^{\rm p}$ if and only if $G_{ {\mathcal Q}^{\rm a},A}$ and $G_{ {\mathcal Q}^{\rm a},B}$ are profinitely conjugate over ${\mathcal Q}^{\rm a}$, which is the equivalence of parts (b) and (c).
\end{proof}
The implications of Corollary \ref{prinall} are that profinite conjugacy over ${\mathcal Q}^{\rm p}$ and over ${\mathcal Q}^{\rm a}$ are equivalent and that the principal generalized Bowen-Franks $R$-modules are indeed the principal invariants among the generalized Bowen-Franks $R$-modules parameterized by ${\mathcal Q}^{\rm a}$. Going forward we need only be concerned with profinite conjugacy of similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$ over ${\mathcal Q}^p$. To aid with this, we simplify the notation by writing ${\mathcal N}_A = {\mathcal N}_{ {\mathcal Q}^{\rm p},A}$, $G_A = G_{ {\mathcal Q}^{\rm p},A}$, $\iota_A = \iota_{ {\mathcal Q}^{\rm P},A}$, an element of $G_A$ as $\{ [m_k]_{k,A}\}_{k\in {\mathbb N}}$, ${\mathcal N}_B = {\mathcal N}_{ {\mathcal Q}^{\rm p},B}$ $G_B = G_{ {\mathcal Q}^{\rm p},B}$, $\iota_B = \iota_{ {\mathcal Q}^{\rm p},B}$, and an element of $G_B$ as $\{ [m_k]_{k,B}\}_{ k\in {\mathbb N}}$. We also refer to a profinite conjugacy over ${\mathcal Q}^{\rm p}$ between $A$ and $B$ simply as a profinite conjugacy $\Psi:G_A\to G_B$ between $A$ and $B$, and to strong {\rm BF}-equivalence of $A$ and $B$ over ${\mathcal Q}^{\rm p}$ simply as strong {\rm BF}-equivalence of $A$ and $B$.
\section{Properties of Profinite Conjugacies}
\subsection{Two Characterizations of Conjugacy}
The existence of a profinite conjugacy between two similar hyperbolic toral automorphisms is insufficient to imply their conjugacy, as it is known that strong ${\rm BF}$-equivalence is insufficient to imply conjugacy (see \cite{BM19, MR05}). However, for similar hyperbolic toral automorphisms $A,B\in {\rm GL}_n({\mathbb Z})$, there are two characterizations of when profinite conjugacies between $A$ and $B$ gives conjugacy of $A$ and $B$ in terms of the embedded copies $\iota_A({\mathbb Z}^n)$ and $\iota_B({\mathbb Z}^n)$ of ${\mathbb Z}^n$, as we showed in (Theorem 5.1 in \cite{BM12}):
\begin{enumerate}
\item[(a)] $A$ and $B$ are conjugate,
\item[(b)] there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi(\iota_A({\mathbb Z}^n)) = \iota_B({\mathbb Z}^n)$, and
\item[(c)] there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi(\iota_A({\mathbb Z}^n))\cap \iota_B({\mathbb Z}^n)$ is $R$-module isomorphic to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$.
\end{enumerate}
We structured the proof of this in the cyclic format (a) $\Rightarrow$ (b), (b) $\Rightarrow$ (c), and (c) $\Rightarrow$ (a). The proof that (a) $\Leftrightarrow$ (b) follows from the proof of (a) $\Rightarrow$ (b) and an extra independent proof of (b) $\Rightarrow$ (a) included in \cite{BM12}, the latter highlighting the pro-${\mathcal C}$ topology ${\mathcal N}_B$ on ${\mathbb Z}^n$. But there was an error in the proof of (c) $\Rightarrow$ (a) in \cite{BM12}, which, using the notation
\[ \Delta_\Psi = \Psi( \iota_A({\mathbb Z}^n)) \cap \iota_B({\mathbb Z}^n),\]
leads to following correction of part (c):
\begin{enumerate}
\item[(c)] there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi^{-1}( \Delta_\Psi)$ is $R$-module isomorphic to the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ and $\Delta_\Psi$ is $R$-module isomorphic to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$.
\end{enumerate}
With the inclusion of the additional condition of the correction in part (c), here are the corrected parts of the proof of the two characterization of the conjugacy.
(b) $\Rightarrow$ (c). Suppose there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi(\iota_A({\mathbb Z}^n)) = \iota_B({\mathbb Z}^n)$. Then $\Delta_\Psi= \iota_B({\mathbb Z}^n)$. The map $\iota_B:{\mathbb Z}^n\to G_B$ from the right ${\mathbb Z}[B]$-module to the right ${\mathbb Z}[B]$-module $G_B$ is an $R$-module monomorphism. So the map $\iota_B^{-1}: \iota_B({\mathbb Z}^n)\to {\mathbb Z}^n$ is an $R$-module isomorphism from $\iota_B({\mathbb Z}^n) = \Delta_\Psi$ to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$. On the other hand, from $\Delta_\Psi = \Psi(\iota_A({\mathbb Z}^n))$ we have $ \Psi^{-1}( \Delta_\Psi) = \iota_A({\mathbb Z}^n)$. The map $\iota_A:{\mathbb Z}^n\to G_A$ is an $R$-module monomorphism from the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ to the right ${\mathbb Z}[A]$-module $G_A$. So the map $\iota_A^{-1}: \iota_A({\mathbb Z}^n)\to {\mathbb Z}^n$ is an $R$-module isomorphism from the right ${\mathbb Z}[A]$-module $\iota_A({\mathbb Z}^n) = \Psi^{-1}(\Delta_\Psi)$ to the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$. This gives (c).
(c) $\Rightarrow$ (a). The assumptions of part (c) imply the existence of $R$-module isomorphisms $h:{\mathbb Z}^n\to \Delta_\Psi$ and $g:{\mathbb Z}^n \to \Psi^{-1}(\Delta_\Psi)$ that satisfy
\[ B h^{-1} = h^{-1} \Gamma_B,\ \ gA = \Gamma_A g.\]
[The additional condition in the corrected version of part (c) gives the existence of $g$, while the original condition in part (c) gives the existence of $h$.] The mapping $h^{-1}\Psi g:{\mathbb Z}^n\to {\mathbb Z}^n$ is an $R$-module isomorphism from the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$. [The original incorrect proof of (c) $\Rightarrow$ (a) used the $R$-module mapping $h^{-1}\Psi \iota_A$ which is where the error occurred.] This implies that $h^{-1}\Psi g$ is an automorphism of ${\mathbb Z}^n$, and gives the existence of $D\in {\rm GL}_n({\mathbb Z})$ such that $D = h^{-1}\Psi g$. This is illustrated in the following diagram.
\[ \begin{CD}
\Psi^{-1}(\Delta_\Psi) @>\Psi>> \Delta_\Psi \\
@AgAA @ AAhA \\
{\mathbb Z}^n @>>D> {\mathbb Z}^n
\end{CD}
\]
Using the right $R$-module structures of $h^{-1}$, $\Psi$, and $g$, there holds
\[ DA = h^{-1}\Psi gA = h^{-1}\Psi \Gamma_A g = h^{-1}\Gamma_B \Psi g = Bh^{-1} \Psi g = BD.\]
Setting $C = D^{-1}$ gives $AC = CB$, hence (a).
We state the corrected version of the two characterizations of conjugacy for similar hyperbolic toral automorphisms.
\begin{theorem}\label{twochar}
For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, the following are equivalent.
\begin{enumerate}
\item[(a)] $A$ and $B$ are conjugate in ${\rm GL}_n({\mathbb Z})$, i.e., there is $C\in {\rm GL}_n({\mathbb Z})$ such that $AC=CB$,
\item[(b)] there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi(\iota_A({\mathbb Z}^n)) = \iota_B({\mathbb Z}^n)$, and
\item[(c)] there is a profinite conjugacy $\Psi:G_A\to G_B$ such that $\Psi^{-1}( \Delta_\Psi)$ is $R$-module isomorphic to the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ and $\Delta_\Psi$ is $R$-module isomorphic to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$.
\end{enumerate}
\end{theorem}
\subsection{The Embedded Copy of ${\mathbb Z}^n$}
The two characterizations of conjugacy in Theorem \ref{twochar} involve the embedded dense copies of ${\mathbb Z}^n$ through the $R$-module monomorphisms $\iota_A$ and $\iota_B$. The second characterization of conjugacy in part (c) of Theorem \ref{twochar} has the presence of the $R$-submodule $\Delta_\Psi=\Psi(\iota_A({\mathbb Z}^n))\cap\iota_B({\mathbb Z}^n)$. When similar hyperbolic $A$ and $B$ are conjugate, there exists a profinite conjugacy by Theorem \ref{twochar} for which $\Psi(\iota_A({\mathbb Z}^n)) = \iota_B({\mathbb Z}^n)$ so that $\Delta_\Psi = \iota_B({\mathbb Z}^n)$. In general, however, part (c) of Theorem \ref{twochar} poses the question of if it is possible for a profinite conjugacy $\Psi$ to send $\iota_A({\mathbb Z}^n)$ into a proper $R$-submodule of $\iota_B({\mathbb Z}^n)$. We answer this in the negative.
\begin{lemma}\label{properRsubmodule} For similar hyperbolic $A,B\in {\rm GL}_n({\mathbb Z})$, if $\Psi:G_A\to G_B$ is a profinite conjugacy between $A$ and $B$, then $\Psi(\iota_A({\mathbb Z}^n))$ is not a proper $R$-submodule of $\iota_B({\mathbb Z}^n)$.
\end{lemma}
\begin{proof} Suppose $\Psi:G_A\to G_B$ is profinite conjugacy for which $\Psi(\iota_A({\mathbb Z}^n))$ is a proper $R$-submodule of $\iota_B({\mathbb Z}^n)$. Since $\iota_A({\mathbb Z}^n)\leq_o {\mathbb Z}^n$ (with respect to the pro-${\mathcal C}$ topology on ${\mathbb Z}^n$ determined by ${\mathcal N}_A$ and the identification of $\iota_A({\mathbb Z}^n)$ with ${\mathbb Z}^n$), we have $\overline{\iota_A({\mathbb Z}^n)} \leq_o G_A$. Then $\Psi\left( \overline{\iota_A({\mathbb Z}^n)}\right) \leq_o G_B$, and so by the identification of ${\mathbb Z}^n$ with $\iota_B({\mathbb Z}^n)$ we have
\[ \Psi\left( \overline{\iota_A({\mathbb Z}^n)}\right) \cap \iota_B({\mathbb Z}^n)\leq_o {\mathbb Z}^n\]
(with respect to the pro-${\mathcal C}$ topology on ${\mathbb Z}^n$ determined by ${\mathcal N}_B$). Let $H\leq_o {\mathbb Z}^n$ be the set defined by
\[ \iota_B(H) = \Psi\left( \overline{\iota_A({\mathbb Z}^n)}\right) \cap \iota_B({\mathbb Z}^n).\]
Since $\Psi$ is a homeomorphism we have
\[ \iota_B (H) = \overline{ \Psi(\iota_A({\mathbb Z}^n))} \cap \iota_B({\mathbb Z}^n).\]
By hypothesis, $\Psi(\iota_A({\mathbb Z}^n))$ is a proper $R$-submodule of $\iota_B({\mathbb Z}^n)$. By the $1-1$ correspondence from $\{ U: U\leq_o {\mathbb Z}^n; {\mathcal N}_B\}$ to $\{ V:V\leq_o G_B\}$ there is only one open subgroup of ${\mathbb Z}^n$ (with respect to pro-${\mathcal C}$ topology determined by ${\mathcal N}_B$) whose closure is $G_B$, and this open subgroup is ${\mathbb Z}^n$. Since $\Psi(\iota_A({\mathbb Z}^n))\ne \iota_B({\mathbb Z}^n)$ it follows that $H$ is a proper open subgroup of ${\mathbb Z}^n$, for if $H={\mathbb Z}^n$, then
\[ \overline{ \Psi(\iota_A({\mathbb Z}^n)) }\cap \iota_B({\mathbb Z}^n) = \iota_B({\mathbb Z}^n)\]
which would imply that $\overline{\Psi(\iota_A({\mathbb Z}^n))} = G_B$, giving $\Psi(\iota_A({\mathbb Z}^n)) = \iota_B({\mathbb Z}^n)$, a contradiction. Hence $[{\mathbb Z}^n:H]>1$. Since $\overline{{\mathbb Z}^n}=G_B$, we have
\[ [G_B:\overline{H}] = [{\mathbb Z}^n:H]>1.\]
But since $\iota_A({\mathbb Z}^n)$ is dense in $G_A$ and since $\Psi$ is a homeomorphism, the image $\Psi(\iota_A({\mathbb Z}^n))$ is dense in $G_B$. Hence
\[ \overline{\Psi(\iota_A({\mathbb Z}^n))} = G_B,\]
which implies that $H = {\mathbb Z}^n$, so that $[G_B:\overline{H}] = 1$, a contradiction.
\end{proof}
\subsection{Cyclic Automorphisms}
For profinitely conjugate similar hyperbolic toral automorphisms $A$ and $B$ we show how $\iota_A({\mathbb Z}^n)$ and $\iota_B({\mathbb Z}^n)$ within $G_A$ and $G_B$ distinguish $A$ from $B$ when $A$ is cyclic and $B$ is not cyclic. An automorphism $A\in {\rm GL}_n({\mathbb Z})$ is called cyclic if there exists $\xi\in {\mathbb Z}^n$ such that the set $\{ \xi, \xi A,\dots, \xi A^{n-1}\}$ forms a basis for ${\mathbb Z}^n$, and $\xi$ is called a generator for $A$, i.e., the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ is cyclic, or equivalently, $\iota_A({\mathbb Z}^n)$ is a cyclic $R$-submodule of $G_A$. As is well known, a cyclic automorphism $A$ is conjugate to the companion matrix of its characteristic polynomial $f(t) = a_0 + a_1t + \cdots + a_{n-1} t^{n-1} + t^n \in {\mathbb Z}[t]$, i.e., to
\[ C_f=\begin{bmatrix}
0 & 1 & 0 & \dots & 0 & 0 \\
0 & 0 & 1 & \dots & 0 & 0 \\
0 & 0 & 0 & \dots & 0 & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \dots & 0 & 1 \\
-a_0 & -a_1 & -a_2 & \dots &-a_{n-2} & -a_{n-1}
\end{bmatrix}.\]
In particular, cyclicity is an invariant of the conjugacy class of the companion matrix. Profinite conjugacy realizes this invariant for hyperbolic toral automorphisms in a very explicit manner.
\begin{theorem}\label{cyclicity} Suppose that $A,B\in {\rm GL}_n({\mathbb Z})$ are similar and hyperbolic. If $A$ is cyclic with generator $\xi\in{\mathbb Z}^n$, and $B$ is not cyclic, then for any profinite conjugacy $\Psi:G_A\to G_B$ there holds
\[ \Psi(\iota_A(\xi))\not\in \iota_B({\mathbb Z}^n).\]
\end{theorem}
\begin{proof} Suppose $A$ is cyclic with generator $\xi\in{\mathbb Z}^n$. Then the set $\{\xi,\xi A,\dots,\xi A^{n-1}\}$ is a basis for ${\mathbb Z}^n$. Identifying $\xi$ with its image $\iota_A(\xi)$ in $G_A$, the set
\[ \big\{ \{ [\xi A^j]_{k,A}\}_{k\in {\mathbb N}}: j=0,1,\dots,n-1\big\}\]
is a basis for $\iota_A({\mathbb Z}^n)$. For a topological $R$-module isomorphism $\Psi:G_A\to G_B$,
set
\[ \{ [\eta_k]_{k,B}\}_{k\in{\mathbb N}} = \Psi( \{ [\xi]_{k,A}\}_{k\in{\mathbb N}}.\]
Since $\Psi$ is a $R$-module isomorphism, we have
\begin{align*}
\Psi( \{ [\xi A]_{k,A}\}_{k\in{\mathbb N}}
& = \Psi ( \{ A_k[\xi]_{k,A}\}_{k\in{\mathbb N}} \\
& = (\Psi\circ \Gamma_A)(\{ [\xi]_{k,A}\}_{k\in{\mathbb N}} \\
& = (\Gamma_B\circ \Psi) (\{ [\xi]_{k,A}\}_{k\in{\mathbb N}} \\
& = \Gamma_B ( \{ [\eta_k]_{k,B}\}_{k\in{\mathbb N}} \\
& = \{ B_k[\eta_k]_{k,B}\}_{k\in{\mathbb N}} \\
& = \{ [\eta_k B]_{k,B}\}_{k\in{\mathbb N}}.
\end{align*}
By induction we have for $1\leq r\leq n-1$ that
\[ \Psi(\{ [\xi A^r]_{k,A}\}_{k\in {\mathbb N}}) = \{ [\eta_k B^r]_{k,B}\}_{k\in{\mathbb N}}.\]
Suppose to the contrary that $\Psi(\iota_A(\xi))\in \iota_B({\mathbb Z}^n)$. Then $\{ [\eta_k]_{k,B}\}_{k\in{\mathbb N}}\in \iota_B({\mathbb Z}^n)$. Thus there is $m^\prime\in {\mathbb Z}^n$ such that $[\eta_k]_{k,B} = [m^\prime]_{k,B}$ for all $k\in{\mathbb N}$. For each $1\leq r\leq n-1$ we have
\[ \{ [\eta_k B^r]_{k,B}\}_{k\in{\mathbb N}} = \{ B_k^r[\eta_k]_{k,B}\}_{k\in{\mathbb N}} =
\{ B_k^r[m^\prime]_{k,B}\}_{k\in{\mathbb N}} = \{ [m^\prime B^r]_{k,B}\}_{k\in{\mathbb N}}.
\]
This implies that $\{ [n_k B^r]_{k,B}\}_{k\in{\mathbb N}}\in \iota_B({\mathbb Z}^n)$ for all $1\leq r\leq n-1$.
We show that $\Psi(\iota_A({\mathbb Z}^n)) \subset \iota_B({\mathbb Z}^n)$. Let $m\in {\mathbb Z}^n$. Since $\{ \xi, \xi A,\dots, \xi A^{n-1}\}$ is a basis for ${\mathbb Z}^n$ there exist $c_1,c_2,\dots,c_n\in{\mathbb Z}$ such that
\[ m = c_1\xi + c_2 \xi A + \cdots + c_n \xi A^{n-1}.\]
Then
\begin{align*}
\Psi(\iota_A(m))
& = \Psi ( \{ [m]_{k,A}\}_{k\in{\mathbb N}}) \\
& = \Psi ( \{ [c_1\xi + c_2\xi A + \cdots + c_n \xi A^{n-1}]_{k,A}\}_{k\in{\mathbb N}}) \\
& = \Psi ( \{ c_1[\xi]_{k,A} + c_2[\xi A]_{k,A} + \cdots + c_n [\xi A^{n-1}]_{k,A}\}_{k\in{\mathbb N}}) \\
& = \Psi ( c_1 \{ [\xi]_{k,A}\}_{k\in{\mathbb N}} + c_2 \{ [\xi A]_{k,A}\}_{k\in{\mathbb N}} + \cdots + c_n \{ [\xi A^{n-1}]_{k,A}\}_{k\in{\mathbb N}} ) \\
& = c_1\{ [m^\prime]_{k,B}\}_{k\in{\mathbb N}} + c_2\{ [m^\prime B]_{k,B}\}_{k\in{\mathbb N}} + \cdots + c_n\{ [m^\prime B^{n-1}]_{k,B}\}_{k\in{\mathbb N}} \\
& \in \iota_B({\mathbb Z}^n).
\end{align*}
Since $m\in{\mathbb Z}^n$ is arbitrary, we obtain $\Psi(\iota_A({\mathbb Z}^n)) \subset \iota_B({\mathbb Z}^n)$.
We show that $\Psi(\iota_A({\mathbb Z}^n)) \ne \iota_B({\mathbb Z}^n)$. Using $m^\prime$ from above, if every element of $\iota_B({\mathbb Z}^n)$ could be written in the form
\[ c_1\{ [m^\prime]_{k,B}\}_{k\in{\mathbb N}} + c_2\{ [m^\prime B]_{k,B}\}_{k\in{\mathbb N}} + \cdots + c_n\{ [m^\prime B^{n-1}]_{k,B}\}_{k\in{\mathbb N}}\]
for $c_i\in {\mathbb Z}$, $i=1,2,\dots,n$, then $B$ would be cyclic with generator $m^\prime$. But $B$ is not cyclic by hypothesis, and so $\Psi(\iota_A({\mathbb Z}^n)) \ne \iota_B({\mathbb Z}^n)$.
We now have $\Psi(\iota_A({\mathbb Z}^n))$ is a proper $R$-submodule of $\iota_B({\mathbb Z}^n)$. This contradicts Lemma \ref{properRsubmodule}. Therefore, $\Psi(\iota_A(\xi))\not\in \iota_B({\mathbb Z}^n)$.
\end{proof}
\subsection{Distinguishing Conjugacy Classes}
Profinite conjugacy distinguishes between the conjugacy classes of similar hyperbolic toral automorphisms according to the contrapositive of Theorem \ref{twochar} and the assumption of strong {\rm BF}-equivalence which by Corollary \ref{characterization} gives the existence of profinite conjugacies. For similar, hyperbolic, strongly {\rm BF}-equivalent $A,B\in {\rm GL}_n({\mathbb Z})$ the following are equivalent:
\begin{enumerate}
\item[(a)] $A$ and $B$ are not conjugate,
\item[(b)] {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ there holds $\Psi(\iota_A({\mathbb Z}^n)) \ne \iota_B({\mathbb Z}^n)$, and
\item[(c)] {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ either $\Psi^{-1}(\Delta_\Psi)$ is not $R$-module isomorphic to right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ or $\Delta_\Psi$ is not $R$-module isomorphic to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$ (where $\Delta_\Psi = \Psi(\iota_A({\mathbb Z}^n))\cap \iota_B({\mathbb Z}^n)$).
\end{enumerate}
Parts (a), (b), and (c) of this equivalence have particular realizations for similar, hyperbolic, strongly {\rm BF}-equivalent $A,B\in {\rm GL}_n({\mathbb Z})$ when $A$ is cyclic, with generator $\xi$, and $B$ is not cyclic. The realization of (a) follows because cyclicity is an invariant of conjugacy. The realization of (b) follows because {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ there holds $\Psi(\iota_A(\xi)) \not\in \iota_B({\mathbb Z}^n)$ by Theorem \ref{cyclicity}, which implies {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ that $\Psi(\iota_A({\mathbb Z}^n))\ne \iota_B({\mathbb Z}^n)$. The realization of (c) follows because {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ the $R$-submodules $\Psi^{-1}(\Delta_\Psi)$ and $\Delta_\Psi$ are $R$-module isomorphic, while the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ is cyclic and the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$ is not cyclic, implying {\it for all} profinite conjugacies $\Psi:G_A\to G_B$ that either $\Psi^{-1}(\Delta_\Psi)$ is not $R$-module isomorphic to the right ${\mathbb Z}[A]$-module ${\mathbb Z}^n$ or $\Delta_\Psi$ is not $R$-module isomorphic to the right ${\mathbb Z}[B]$-module ${\mathbb Z}^n$.
|
\section{Introduction}
\label{sec:introduction}
The interest in developing self-driving vehicles capable of navigating their environments without human intervention has significantly grown during the last decade. One of the main modules of a self-driving vehicle is the localization module~\cite{siegwart2011introduction}. A self-driving vehicle has to be able to estimate its position and orientation accurately in order to navigate its environment successfully. To achieve accurate localization, several exteroceptive sensors are usually used such as Global Positioning Systems (GPS)~\cite{drawil2012gps}, Cameras~\cite{leutenegger2015keyframe}, Light Detection And Ranging (LiDAR)~\cite{zhang2014loam}, and so on.
Cameras can be utilized for localization using Visual Odometry (VO) algorithms~\cite{scaramuzza2011visual, fraundorfer2012visual}. A VO algorithm estimates the incremental motion between successive frames captured by the camera. The position and orientation of the vehicle are then estimated by integrating these motion increments at each timestep. VO algorithms can be classified based on the type of camera into monocular~\cite{sabry2019ground}, stereo~\cite{fan2017stereo}, and RGB-D~\cite{sturm2012benchmark} visual odometries~\cite{aqel2016review}. It can also be classified based on the motion estimation approach as feature-based and direct VO. Feature-based VO relies on detecting and matching visual features in the successive frames. Such features are then used to solve the motion estimation problem.
A monocular visual odometry algorithm uses the captured frames from only one camera to estimate the motion of the vehicle. Such technique uses Structure from Motion (SFM) to estimate the incremental motion~\cite{tomono20053, hartley2003multiple}. However, since the algorithm uses only one camera, the incremental motion of the vehicle can only be estimated up to an unobservable scale which can be determined using another sensor such as a wheel encoder or an Inertial Measurement Unit (IMU).
Since VO algorithms rely on integration in estimating the position and orientation,
individual errors in each incremental motion lead to the accumulation of error in every time step.
Although the drift error in VO algorithms cannot be eliminated, several works have tried to reduce such drift. In~\cite{desai2016visual}, a new descriptor called the Syntheitc BAsis (SYBA) descriptor is used to reduce the falsely matched features between the successive frames. In order to achieve this, a sliding window approach is developed where the features detected in a given frame is not only matched to the prior frame but to a window of previously captured frames. In~\cite{peretroukhin2017reducing}, the drift in the estimation of a VO algorithm is reduced using a Bayesian Convolutional Neural Network that infers the direction of the sun. The sun direction then provides a global estimate of the orientation estimate which consequently reduces the drift in the visual odometry output.
Another approach to reduce the drift in the estimation of VO is through using sensor fusion techniques where the estimate from the VO algorithm is fused with other estimates from other sensors to reduce the error. Sensor fusion relies on probabilistic estimation approaches such as Extended Kalman Filter~\cite{thrun2002probabilistic}, Moving Horizon Estimation~\cite{osman2021generic} and pose graph optimization~\cite{tao2021multi}, and so on. In such approaches, the uncertainty in the position and orientation estimates need to be quantified using the covariance matrix of the measurements in order to enhance the accuracy of the estimation. In~\cite{osman2019novel, osman2018online}, a covariance estimation approach was developed utilizing a sensor not suffering from drift to estimate the covariance of the drift suffering odometry.
In this paper, a machine learning approach is proposed for reducing the drift in the incremental orientation estimates of a monocular VO algorithm which in turn results in the overall enhancement of the integrated position and orientation accuracy. By developing a neural network that can correlate the errors in the motion increments estimated by the VO algorithm with the motion of the features in the successive frames, the error in such increments can be reduced significantly. By using an accurate Realtime Kinematic GPS (RTK-GPS) during the training phase, the neural network can be trained to estimate the error in the visual odometry increments which can be then used to reduce the motion error in realtime during the operation of the vehicle. The proposed approach is validated using sequences from KITTI dataset with its ground-truth data~\cite{geiger2013vision}.
The remainder of the paper is organized as follows, Section~\ref{sec:drift_modeling} contains the mathematical formulation of the drift error in the visual odometry followed by our proposed approach for reducing the drift in Section~\ref{sec:data_driven}. In Section~\ref{sec:experimental_work} the experimental work executed to evaluate the proposed method is explained and the results are shown in Section~\ref{sec:results_discussion}. Finally, the conclusion and future work are stated in Section~\ref{sec:conclusion}.
\section{Error Modeling in Visual Odometry}
\label{sec:drift_modeling}
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{driftNeuralNetwork.pdf}
\caption{The overall feature-based visual odometry pipeline with the proposed drift reducing neural network. The image $I$ at the position and orientation $\mathbf{x}_{k}$ is passed to a features detection algorithm $h_{f}$ followed by a matching algorithm $h_{m}$ which matches the detected features to the features from the previous frame. The matched features are then used to estimate the motion increment using a motion estimation algorithm $h_{c}$. The statistical moments of the features motion in the camera frames are then passed to the NN along with the motion increment from the VO algorithm. Finally, the NN outputs a more accurate motion increment which is then used to calculate the current position and orientation of the vehicle through composition}
\label{fig:overall_pipeline}
\end{figure*}
As mentioned earlier, a visual odometry estimate the position and orientation of a vehicle by integrating the incremental motion. For feature-based VO algorithm, the first step is detecting the features in the frame, this can be modeled as
\begin{equation}
\label{equ:feature_model}
\mathcal{F}_{k} = \underbrace{h_{\text{f}}(\mathbf{x}_{k})}_{\hat{\mathcal{F}}_{k}} + \mathbf{w}_{k},
\end{equation}
where $\mathbf{x}_k = [\mathbf{p}^{\top}_k, \boldsymbol{\theta}^{\top}_k] \in \mathbb{R}^{3} \times \mathbb{M}$ is the state vector of the vehicle in the $k$-th time-step, which is composed of the position $\mathbf{p} \in \mathbb{R}^{3}$ and the orientation $\boldsymbol{\mathbf{\theta}} \in \mathbb{M}$ of the vehicle using an orientation representation (e.g. quaternions or Euler angles). $\mathcal{F}_{k} := \{f_{k}^{i}\}_{i=1}^{N_f} \in \mathbb{I}$ is the detected features set and $f := [u, v]^{\top}$ denotes a detected feature in the image. $h_{\text{f}}: \mathbb{R}^{3} \times \mathbb{M} \rightarrow \mathbb{I}$ is the model of the feature detection, and finally $\mathbf{w}_{k}$ is a noise term that models the errors in the feature detection process due to the noise affecting the camera.
After the detection of features in the $k$-th frame, they are matched with the features detected in the $(k-1)$-th frame. Using \eqref{equ:feature_model} and applying Taylor expansion, this step can be modeled as
\begin{align}
\label{equ:matching}
\begin{split}
\mathcal{M}_{k} &= h_{\text{m}}(\mathcal{F}_{k-1}, \mathcal{F}_{k}) + \varepsilon_{k}^{\text{m}} \\
&\approx \underbrace{h_{\text{m}}(\hat{\mathcal{F}}_{k-1}, \hat{\mathcal{F}}_{k})}_{\hat{\mathcal{M}}_{k}} +
\underbrace{\frac{\partial h_{\text{m}}}{\partial \hat{\mathcal{F}}_{k-1}} \mathbf{w}_{k-1} + \frac{\partial h_{\text{m}}}{\partial \hat{\mathcal{F}}_{k}} \mathbf{w}_{k}}_{\mathbf{w}^{\text{m}}_{k}} + \varepsilon_{k}^{\text{m}}, \\
\end{split}
\end{align}
where $\mathcal{M}_{k} := \{(f_{k-1}^{i}, f_{k}^{i})\}_{i=1}^{N_f} \in \mathbb{I}^{2}$ is the set of matched features (here we ignore all unmatched features), $h_{\text{m}}: \mathbb{I} \times \mathbb{I} \rightarrow \mathbb{I}^{2}$ is a mapping that represents the features matching algorithm, $\mathbf{w}^{\text{m}}_{k}$ is the effect of the white noise $\mathbf{w}_{k}$ on the matching, and $\varepsilon^{\text{m}}_{k} := \varepsilon^{\text{m}}(\mathcal{F}_{k-1}, \mathcal{F}_{k})$ is another noise term that is correlated to errors in the detection and matching of features in the successive frames. Such errors can result from false feature matching (wrong association), errors in the calibration of the intrinsic parameters of the camera, ... etc~\cite{yang2018challenges, calibration2008}. An example of matching errors in the VO pipeline is shown in Fig.~\ref{fig:mistmatch}.
Finally, depending on the type of camera being used, a motion estimation algorithm can be used to compute the motion increment from the matched feature pairs $\mathcal{M}_{k}$. This can be modeled similar to~\eqref{equ:matching} as
\begin{align}
\begin{split}
\Delta \mathbf{x}_{k} &= h_{\text{c}}(\hat{\mathcal{M}}_{k}) + \frac{\partial h_{\text{c}}}{\partial \hat{\mathcal{M}}_{k}} (\mathbf{w}^{\text{m}}_{k} + \varepsilon_{k}^{\text{m}}) \\
&= h_{\text{c}}(\hat{\mathcal{M}}_{k}) + \mathbf{w}_{k}^{\text{c}} + \varepsilon_{k}^{\text{c}},
\end{split}
\end{align}
where $h_{\text{c}}: \mathbb{I}^{2} \rightarrow \mathbb{R}^{3} \times \mathbb{M}$ is the motion estimation function, $\mathbf{w}_{k}^{\text{c}}$ is the colored noise resulting from the origin white noise in~\eqref{equ:feature_model}, and $\varepsilon_{k}^{\text{c}} := \varepsilon^{\text{c}}(\mathcal{F}_{k-1}, \mathcal{F}_{k})$ is the effect of the noise term $\varepsilon_{k}^{\text{m}}$ on the motion estimation module.
\begin{figure}[t]
\centering
\includegraphics[trim={1.5cm 4cm 4.8cm, 1.5cm}, clip, width=0.95\linewidth]{mismatching_example.eps}
\caption{An example of errors in mismatching of features between the 2180-th frame and the 2181-st frame in KITTI data set sequence 5. The two frames are shown in the figure using red and blue shades. The features in the first frame is represented by the green (+) marks and blue circles in the second frame. Finally, the matched features are joined by red lines.}
\label{fig:mistmatch}
\end{figure}
We aim at reducing the effect of the noise term $\varepsilon_{k}^{\text{c}}$ in order to increase the accuracy of the motion increments and ultimately reduce the drift in the VO algorithm. Using a Bayesian approach, this problem can be stated as the problem of estimating the distribution
\begin{equation}
\label{equ:bayes}
p(\Delta \mathbf{x}_{k} \mid \varepsilon_{k}^{\text{c}}, \mathcal{M}_{k}) = \eta \ p(\varepsilon_{k}^{\text{c}} \mid \Delta \mathbf{x}_{k}, \mathcal{M}_{k}) p(\Delta \mathbf{x}_{k} \mid \mathcal{M}_{k}).
\end{equation}
In~\eqref{equ:bayes}, the prior distribution $p(\Delta \mathbf{x}_{k} \mid \mathcal{M}_{k})$ is the visual odometry algorithm. The distribution $p(\varepsilon_{k}^{\text{c}} \mid \Delta \mathbf{x}_{k}, \mathcal{M}_{k})$ is the likelihood distribution of the $\varepsilon_{k}^{\text{c}}$. This distribution is hard to determine in practice because
(i) it is affected by many factors which are hard to model, (ii) the detection and matching of features as well as the motion estimation are actually algorithms that has no closed form expressions which makes the propagation of these errors to the motion increments infeasible.
Consequently, we propose the use of a feedforward Neural Network (NN) to estimate the error $\varepsilon_{k}^{\text{c}}$ using the output of a monocular VO algorithm $\Delta \mathbf{x}_{k}$ as well as the matched features pairs $\mathcal{M}_{k}$.
\section{Drift Reduction using Neural Networks}
\label{sec:data_driven}
In this section, the NN used to estimate the noise component $\varepsilon^{\text{c}_{k}}$ is described. A monocular VO algorithm estimates the incremental motion of the vehicle up to a scale which can be computed using an external sensor such as a wheel encoder or an IMU. the proposed work is only concerned with estimating the error component for the orientation increment $\Delta \theta$. The overall pipeline of the VO algorithm with the proposed Drift Reducing Neural Network (DRNN) is shown in Fig.~\ref{fig:overall_pipeline}.
The objective of the proposed DRNN is to estimate a refined incremental orientation $\Delta \theta$ using the output of the monocular VO as well as information about the matched features $\mathcal{M}_{k}$. As disucssed in the previous section, the error in the matched features result in errors in the estimated motion increments by the VO. Such errors can be detected using the motion of the features in a given frame $I(\mathbf{x}_{k})$ with respect to $I(\mathbf{x}_{k-1})$
\begin{equation}
\underbrace{\begin{bmatrix}
\Delta u^{i}_{k} \\ \Delta v^{i}_{k}
\end{bmatrix}}_{\Delta f^{i}_{k}} = f_{k}^{i} - f_{k-1}^{i},
\end{equation}
where $\Delta u$ and $\Delta v$ are the change in the position of the feature in the successive frames. Therefore, the DRNN can be trained using the position change of the features.
The space of the matched features in the successive frames is large and many different types of error can occur consequently, training the DRNN using all the matched features would require a very large amount of training samples which would be infeasible during the training phase of the DRNN. Alternatively, we use some of the statistical moments of the set $\{\Delta f_{k}^{i}\}_{i=1}^{N_f}$ to train the DRNN to reduce the error. In this paper, the information about the change in the positions is captured using three statistical moments (mean $\mu_{k}$, variance $\nu_{k}$, skewness $\kappa_{k}$) in addition to the Root Mean Square (RMS) $\rho_{k}$.
As for the orientation increment, the orientation representation used is the unit quaternions. It is well-known that the rotation space $\mathbb{M}$ is not an Euclidean space.
However, the tangent space of the rotation space is an Euclidean space ($\mathbb{R}^{3}$). Consequently, to avoid discontinuities in the cost function of the DRNN, the orientation deviation in the tangent space $\xi := \exp_{\text{q}}(\Delta \theta) \in \mathbb{R}^{3}$ is used where $\exp_{\text{q}}: \mathbb{M} \rightarrow \mathbb{R}^{3}$ is the exponential map that maps from the quaternion space to the tangent space.
Finally, the input layer of the DRNN is defined as $[(\xi^{-}_{k})^{\top}, \mu_{k}, \nu_{k} , \kappa_{k}, \rho_{k}]^{\top}$, where $\xi^{-}_{k} := \exp_{\text{q}}(\Delta \theta^{-}_{k})$ is the orientation deviation of the orientation increment estimated by the VO algorithm. The output layer of the DRNN is the refined orientation increment $\xi^{+}_{k}$. The overall position and orientation of the vehicle can finally be computed as
\begin{align}
\theta^{+}_{k} &= \theta^{+}_{k-1} \odot \Delta \theta^{+}_{k} \\
\mathbf{p}_{k} &= \mathbf{p}_{k-1} + R(\theta^{+}_{k}) \Delta \mathbf{p}_{k}
\end{align}
where $\odot$ denotes the quaternion product, and $R: \mathbb{M} \rightarrow SO(3)$ is a mapping from quaternion group to the special orthogonal group. Furthermore, to make sure that the DRNN does not overfit to the training data and is actually capable of reducing the VO errors during operation, a Bayesian regularization approach is used ~\cite{Foresee1997GaussNewtonAT}.
\section{Experimental Work}
In this section we discuss the details of the implemented monocular VO odometry algorithm and the executed experimental work in order to validate our approach. Using Mathworks' Matlab, a feature-based monocular VO algorithm is developed using Speeded Up Robust Features (SURF) detector~\cite{article},~\cite{surf}. The features are then matched using the approach introduced in~\cite{features2},
where only unique features are matched by using forward-backward matching.
The monocular VO implementation estimates the motion $\Delta \mathbf{x}$ using the epipolar constraint between frames
\begin{equation}
\begin{bmatrix}
x^{i}_{k-1} & y^{i}_{k-1}
\end{bmatrix} \mathbf{E} \begin{bmatrix}
x^{i}_{k} \\ y^{i}_{k} \\
\end{bmatrix} = 0, \ \ \forall \ 1 \leq i \leq N_f,
\end{equation}
where $(x^{i}, y^{i})$ is the position of the $i$-th feature in the camera frame, and $\mathbf{E} \in \mathbb{R}^{3 \times 3}$ is the essential matrix for the calibrated camera~\cite{hartley2003multiple}.
The essential matrix $\mathbf{E}$ can be estimated using the five-point algorithm~\cite{features3}. Furthermore, an outliers rejection approach was applied based on the M-estimator sample consensus (MSAC) algorithm~\cite{features4}.
The essential matrix $\mathbf{E}$ can then be decomposed to the translation (up to a scale) and rotation of the camera between two successive frames~\cite{hartley2003multiple}.
Since the monocular VO algorithm estimates the motion up to a scale, here we compute this scale using the ground-truth data for both the VO with and without the DRNN as we are only concerned with reducing the orientation drift in the VO output.
This does not affect the validation of the proposed approach since the scale is computed using the same method for both outputs.
The outcome of the VO is passed to a feed-forward NN with 3 layers. The input layer consists of 11 neurons for the inputs discussed in the previous section. The hidden and output layers consist of 30 and 3 neurons respectively.
The activation function used is a sigmoid function and
the weights of the NN are optimized using the Levenberg-Marquardt optimization algorithm~\cite{trainNN} and the training is done using 1500 training epochs.
Using the KITTI data set, the NN is trained with the first $60\%$ percent of sequences 1, 2, 5, 6, 7, 8, and 9 and is tested on the remaining $40\%$ percent of each sequence. Sequence $4$ is used completely for training as it is too short for partitioning. Finally, sequence $10$ is completely used for testing the DRNN. The training is accomplished by using the ground-truth data as targets for the NN.
The performance of the DRNN is validated and its output is compared to that of the monocular VO algorithm without the DRNN. For both outputs, the Root Mean Squared Error (RMSE) in translation and orientation are calculated using the ground-truth data provided for each sequence. The orientation error is computed as
\begin{equation}
R^{\text{e}, \pm}_k = R(\theta^{\text{GT}}_k)^{-1} R(\theta^{\pm}_k)
\end {equation}
where $ R^{\text{e, +}}_k, R^{\text{e, -}}_k \in SO(3)$ represents the error rotation matrix for the VO with and without the proposed DRNN, $\theta^{\text{GT}}_k$ represents the ground truth orientation in as a quaternion.
\label{sec:experimental_work}
\section{Results and Discussion}
\label{sec:results_discussion}
\begin{figure}[t]
\centering
\includegraphics[trim={0cm 0cm 0cm 0.85cm}, clip, width=0.95\linewidth]{seq-01.eps}
\caption{The computed trajectory by the VO with and without the DRNN along with the ground-truth trajectory of the testing portion of sequence 0 of KITTI dataset.}
\label{fig:seq1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{seq1error.eps}
\caption{The translation error in the position estimation of the vehicle in sequence 0 (top) and the incremental orientation error at every frame (bottom).}
\label{fig:seq1error}
\end{figure}
In this section the results of the testing scenarios are discussed as mentioned in Section \ref{sec:experimental_work}. Fig.~\ref{fig:seq1} shows the estimated trajectory by the VO algorithm with and without the proposed DRNN for the testing part of sequence 0 as well as the ground-truth trajectory. As can be seen in the figure, the computed trajectory using the VO algorithm without the DRNN suffers from significantly more drift compared to the VO algorithm with the DRNN.
\begin{figure}[t]
\centering
\includegraphics[trim={0cm 0cm 0cm 0.85cm}, clip, width=0.95\linewidth]{seq-05.eps}
\caption{The computed trajectory by the VO with and without the DRNN along with the ground-truth trajectory of the testing portion of sequence 5 of KITTI dataset.}
\label{fig:seq5}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[trim={0cm 0.2cm 0cm 0cm}, clip, width=0.95\linewidth]{seq5error.eps}
\caption{The translation error in the position estimation of the vehicle in sequence 5 (top) and the incremental orientation error at every frame (bottom).}
\label{fig:seq5error}
\end{figure}
Fig.~\ref{fig:seq1error} shows the translation error as well as the incremental orientation error for sequence 0. Using the proposed DRNN, the estimation of the incremental orientation change was persistently more accurate over the sequence frames which led to a better overall pose estimation. Although the VO output was relatively accurate at the beginning of the path however, with every turn that the vehicle took, the amount of drift in the estimated orientation increased which led to a root mean translation error of $94.19$ m compared to only $23.35$ m when using the DRNN which amounts to a $75\%$ improvement. This can be attributed to the enhanced orientation estimation which was enhanced from a RMSE of $15.22$ degrees to only $5.03$ degrees ($66\%$) when using the DRNN.
Fig.~\ref{fig:seq5} shows the results for the testing part of sequence 5. This sequence did not contain as many turns as sequence 0, which resulted in a slower drift compared to sequence 0 and a less error in the overall sequence (shown in Fig.~\ref{fig:seq5error}). That being said, by the end of the trajectory the VO output still suffered from a larger amount of drift compared to the VO output with the DRNN. In the case of using the DRNN, the RMSE in orientation for sequence 5 was reduced from $12.44$ degrees to $3.2$ degrees ($74\%$ improvement) which resulted in $82\%$ enhancement in the position accuracy. Here, the DRNN output was also persistently better than that of the VO output. Furthermore, notice that the DRNN managed to reduce the incremental orientation error at the error spikes. These spikes can lead to a significant increase in the overall error in the path due to the reliance on integration.
\begin{figure}[t]
\centering
\includegraphics[trim={0cm 0cm 0cm 0.7cm}, clip, width=0.95\linewidth]{seq-09.eps}
\caption{The computed trajectory by the VO with and without the DRNN along with the ground-truth trajectory of the testing portion of sequence 9 of KITTI dataset.}
\label{fig:seq9}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{seq9error.eps}
\caption{The translation error in the position estimation of the vehicle in sequence 9 (top) and the incremental orientation error at every frame (bottom).}
\label{fig:seq9error}
\end{figure}
Similarily, Fig.~\ref{fig:seq9} and~\ref{fig:seq9error} show the results for sequence 9 testing part of KITTI dataset which are similar to those of sequence 5. In this case, the orientation RMSE was reduced from $7.11$ degress to $2.08$ degrees ($70\%$ improvement) and the translation RMSE was reduced from $29.66$ m to $10.34$ m ($65\%$ improvement).
\begin{figure}[t]
\centering
\includegraphics[trim={0cm 0cm 0cm 0.7cm}, clip, width=0.95\linewidth]{seq10.eps}
\caption{The computed trajectory by the VO with and without the DRNN along with the ground-truth trajectory of the testing portion of sequence 10 of KITTI dataset.}
\label{fig:seq10}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{seq10error.eps}
\caption{The translation error in the position estimation of the vehicle in sequence 10 (top) and the incremental orientation error at every frame (bottom).}
\label{fig:seq10error}
\end{figure}
Fig.~\ref{fig:seq10} and~\ref{fig:seq10error} show the estimation results of the complete trajectory of sequence 10 as well as the translation and incremental orientation errors. Even though the sequence was completely used for testing without partitioning, the use of the DRNN resulted in significant enhancement in the translation and orientation estimation accuracy. In addition to the previous tests using the partitioned sequences, this sequences show the practical feasibility of the proposed DRNN. The idea behind the DRNN is that the drift can be reduced through training the NN during the development phase of the vehicle and then using the DRNN during operation to reduce the drift in the VO estimate. For sequence $10$, the orientation RMSE was reduced from $29.06$ degrees to $7.02$ degrees ($76\%$ improvement) which then led to a reduction in translation error from $160.56$ m to $23.85$ m ($85\%$ improvement).
Finally, Table~\ref{tab:results} shows the overall results for the used test sequences. The use of the DRNN resulted in a significant reduction in the translation and orientation errors for all of the testing sequences which shows that through using the proposed approach, the performance of the VO algorithm can be significantly enhanced.
\begin{table}[ht]
\centering
\caption{RMSE Values for the Kitti Dataset Sequences}
\label{tab:results}
\renewcommand{\arraystretch}{1.1}
\begin{tabular}{cccccc}
\hline \hline
\multirow{2}{*}{Seq. No.} & \multicolumn{2}{c}{Rotation {[}deg{]}} & \multicolumn{2}{c}{Translation {[}m{]}} & \multirow{2}{*}{Distance [m]} \\ \cline{2-5}
& NN & VO & NN & VO & \\ \hline
0 & \textbf{5.03} & 15.22 & \textbf{23.35} & 94.19 & 1648 \\ \hline
2 & \textbf{8.34} & 23.73 & \textbf{39.65} & 93.46 & 2108 \\ \hline
5 & \textbf{3.2} & 12.44 & \textbf{8.05} & 45.36 & 958 \\ \hline
6 & \textbf{3.79} & 15.26 & \textbf{6.96} & 49.88 & 491 \\ \hline
7 & \textbf{1.83} & 3.77 & \textbf{4.08} & 5.44 & 238 \\ \hline
8 & \textbf{6.88} & 16.8 & \textbf{35.86} & 81.72 & 1348 \\ \hline
9 & \textbf{2.08} & 7.11 & \textbf{10.34} & 29.66 & 706 \\ \hline
10 & \textbf{7.02} & 29.06 & \textbf{23.85} & 160.56 & 919 \\ \hline \hline
\end{tabular}
\end{table}
\section{Conclusion and future work}
\label{sec:conclusion}
In this paper, a drift reduction approach for the orientation estimation in VO using a neural network is proposed. The proposed approach estimates the error which is correlated to the errors in the feature detection and matching algorithms and consequently is able to refine the incremental orientation estimation of the VO odometry. Through training the neural network during the development of the vehicle, the performance of the VO can be significantly enhanced using the proposed approach. our proposed DRNN was validated using KITTI dataset, where 7 sequences where partitioned to training and testing sequences for validation. Furthermore, a complete sequence was used for further validation of the algorithm. The results show the efficacy of the proposed DRNN in significantly reducing the drift in the VO output.
In the future work, the proposed approach can be extended to work with different types of cameras such as RGB-D or stereo cameras as well as different VO algorithms such as direct VO. Furthermore, the effect of the proposed approach can also be integrated with visual simultaneous localization and mapping algorithms to achieve more accurate results and avoid significant drift in the pose estimates.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Robot path planning has been receiving increased interest within the research community due to the numerous emerging technologies it advances \cite{survHeuristic}. The most prominent of these technologies include robots \cite{robotPlanning}, autonomous vehicles \cite{surveyVehicles} and UAV path planning \cite{surveyUAV}. The problem of path planning is concerned with finding an optimal path between the source and destination. An optimal decision usually takes into consideration the following aspects:
\begin{itemize}
\item Minimizing traversed distance.
\item Reducing the magnitude of turning angles.
\item Avoiding both static and dynamic obstacles with a certain clearance.
\end{itemize}
Planning the path completely requires consideration of two layers of planning: The first layer develops a low-resolution high-level solution which is the goal of an \textit{offline path planner} or \textit{global path planning}. This is usually performed using an existing environment map which clearly does not take into consideration dynamic obstacles or unexpected events. To deal with the aforementioned dynamicity an \textit{online} or \textit{local path planner} is required which results in a high-resolution low-level path that guides the robot locally in a slice along the previously determined global path.
Depending on the space dimensionality, number of sensors acting as inputs, dynamicity of the environment, and different movement constraints such as maximum speed and turning angles, the problem of finding an optimal path can become increasingly complex \cite{complexDynamics}. To cope with the increased complexity of path planning modules, and to reduce the computation time required during operation, neural networks (NN) have been employed in various path planning solutions \cite{surveyNeuralPath, generalizedNeuralNetwork, neuralRrt, neuralRrt3d,vins, onTheTraining, segmentizedView, neuralDynamics2020}. In this paper we present the most recent works that utilize NNs in solving the robot path planning problem.
The paper is structured as follows. In section \ref{NRRT2D} we present the scheme proposed in \cite{neuralRrt} for 2D path planning which is based on rapidly random-exploring tree (RRT), followed by its generalization to 3D problem by the same authors which we present in section \ref{nrrt3d} based on \cite{neuralRrt3d}. Then, in section \ref{sec:vins}, a framework utilizing value iteration networks (VINs) for UAV 3D online path planning is described based on the work in \cite{vins}. The authors in \cite{onTheTraining} demonstrated the differences in performance when a multi-layer perceptron (MLP) network is trained using different offline-path planning algorithms and with different environment. We present their work and results in section \ref{sec:mlp}. In \cite{segmentizedView}, the authors consider the existence of moving obstacles and devise a scheme for online path planning which we present in \ref{sec:segmentized}. Finally, in \cite{neuralDynamics2020} the authors present the weaknesses in utilizing neural dynamics \cite{neuralDynamicsOld} to find paths, and propose modifications to overcome them. We present their findings and proposal in section \ref{sec:neuraldynamics} and then conclude our paper with section \ref{sec:conclusion}.
\section{Neural Rapidly Random-exploring Tree* (NRRT*)}\label{NRRT2D}
\subsection{RRT and RRT*}
RRT was first introduced in \cite{rrt} to design high-dimensional spaces by randomly building a space-filling tree. In the case of robot path planning, and when using uniform sampling distributions, such sampling methods like RRT and probabilistic roadmap (RPM) provide \textit{probabilistic completeness} in the sense that by allowing the sampling process to run infinitely, the probability of sampling all the points in the state space increases monotonically. Furthermore, they ensure \textit{asymptotic optimality} which means that the algorithm is guaranteed to converge to the optimal solution asymptotically.
Figure \ref{fig:rrt_sampling} demonstrates the process of RRT in the context of path planning. Specifically, the process begins with the starting node, $x_{\text{start}}$, where different points in space are sampled around it. If a sampled point, $x_{\text{rand}}$, does not overlap with any obstacle, then that sample is added to the tree as a leaf node to the nearest parent ($x_3$ in the figure). In the enhanced version RRT*, the procedure further includes a \textit{ChooseParent} and \textit{Rewire} processes. The former selects the parent for the new leaf node based on a defined cost minimization, and the latter process will test neighbor vertices to check if a path with less cost is now available through the new node.
\begin{figure}[!htp]
\centering
\centerline{\includegraphics[width=0.3\textwidth]{figures/rrt.eps}}
\caption{RRT sampling.}
\label{fig:rrt_sampling}
\end{figure}
Clearly, if a solution exists, then RRT* is guaranteed to find it. However, in real-life implementations, this process consumes a lot of time and memory until it reaches the optimal solution which motivated the authors in \cite{neuralRrt} to consider the utilization of NNs to speed-up the convergence such that the proposed mechanism can be used as an online-planning scheme.
\subsection{NRRT*}
Instead of uniformly sampling the sate space, the authors in \cite{neuralRrt} propose using a convolutional neural networks (CNNs) architecture, ResNet50 \cite{ResNet50}, to generate the sampling distribution of the state space.
\subsubsection*{Input} Specifically, the NN architecture receives as inputs the following:
\begin{enumerate}
\item The map of the environment whereas pixels are encoded according to
\begin{equation}
p\in
\begin{cases}
0 & \text{empty,}\\
1 & \text{obstacle,}\\
3 & \text{start position, }\\
4 & \text{goal position.}\\
\end{cases}
\end{equation}
\item The required step size, and clearance values.
\end{enumerate}
\subsubsection*{output} The output of the proposed neural architecture is a map with the same dimensions of the input map such that each pixels contains a value $\hat{p}\in[0,1]$ which denotes the probability of this pixel being chosen during the sampling process. To preserve the asymptotic optimality and probabilstic completeness characteristics of RRT*, the authors propose sampling with equal probabilities from two sources. The first being the results of the NNs, and the second is the standard uniform sampling distribution which considers the whole space.
\subsubsection*{Architecture} the map is fed into a convolutional layer which extracts a high level feature map, which is in turn fed into another convolutional layer to extract low level feature maps with smaller size. On the other path related to the step size and clearance, the result of the first fully connected layer is concatenated with the low-level feature map, and the result of the second layer connected to the first one is concatenated with the high-level feature map. The map which contains the previously mentioned low-level feature map is passed through an Atrous convolution layer which is a form of dilated convolution. Convolving with differnt dilution rate allows us to generate maps with information based on different scales. The result is then concatenated to the result from high-level features path which contains a Atrous spatial pyramid pooling (ASPP) layer followed by upsampling to match the size of the other path. Finally, the resulting map is fed to a decoding convolutional layer which outputs the probability distribution map.
\subsubsection*{Training} To train the NNs architecture, the authors use a number of maps with optimal solutions generated using the A* algorithm \cite{a*}. The cost function for generating the optimal path penalizes both the traversed distance and turning angle. After obtaining the optimal path using A*, a widened version of the path is encoded in pixels on the map which is used for training.
\subsubsection*{Results} The authors simulate their mechanism for different input maps with different clearance and step-size parameters and prove that utilizing NRRT* is indeed able to speed up the convergence towards the optimal solution substantially.
\section{NRRT* 3D}\label{nrrt3d}
The same authors of the mechanism introduced in the previous section have also presented another NRRT* algorithm but for the case where the space is 3-dimensional \cite{neuralRrt3d}. To tackle this problem, they propose using one of two architectures: the first is based on U-Net \cite{u-net} which also makes use of CNNs, and the second utilizes a conditional generative adversarial network (cGAN).
\subsubsection*{Input} Unlike the 2D case, in the proposed solution the input only includes two 3D maps, one encoding the obstacles, and the second encoding the start and goal locations. That is, the clearance and step size are no longer considered as inputs.
\subsubsection*{output} The output is again a 3D distribution of the sampling space whereas favoured pixels are given a binary value of one, while the rest are encoded by a zero. The authors suggest that the algorithm is run in two phases. The first phase favors sampling from the NN output with a probability of 0.9 vs. 0.1 for sampling uniformly in the 3D space. After finding a solution, the second phase begins where the solution is to be improved by sampling uniformly (i.e., equal probability) from both sampling sources. Similarly to before, this ensures the \textit{probabilistic completeness} and the \textit{asymptotic optimality} of the mechanism.
\subsubsection*{Architecture} In the case of U-Net, the network is composed on encoding and decoding layers. Each map is first passed through a different convolutional layer, and the results are then concatenated. Then 4 3x3x3 convolutional layers are applied. In the decoding stage that follows, 4 3x3x3 deconvolutions are performed. A channel is defined as the combination of two layers, one from the encoding stage and the other with the same size from the corresponding decoding stage. Unlike U-Net, the authors propose adding the results of the maps in each channel before passing them to the next decoding layer.
As for the case of GAN, two networks are required, a generator and a discriminator. The goal of the generator is to fool the discriminator by generating noisy output. On the other hand, the discriminator is trained to distinguish correct output from the generator. By adding conditional information a cGAN is able to generate data that satisfies the start and end location conditions. Briefly, the generator network proposed includes a third map which contains the 3D noise. All three maps are passed through 3 seperate convolutional layers, then concatenation layer, and then passed through an network similar to the one explained above. The results is a probability distribution map which is fed to the discriminator network next to the environment and locations maps. The three maps are then passed through three separate convolutional layers and the results are concatenated. Finally, 4 successive convolutional layers with reduced kernel size result in a binary value which translates to a decision of \textit{fake} or \textit{real}.
\subsubsection*{Training} Similarly to the previous section, the optimal solution for each training map is obtained using A* algorithm, and then the NNs are trained using a diluted version of these paths.
\subsubsection*{Results} The authors simulations prove that using NRRT* reduces the required number of sampling iterations by 20-30 times. Furthermore, they show that both NNs architecture are similar in performance, however, the cGAN model is able to improve slightly the accuracy and generalization ability.
\section{Online-planning with Value Iteration Networks (VINs)}\label{sec:vins}
The authors in \cite{vins} target the problem of UAV navigation by proposing a deep neural network framework which leverages VINs. They also test their mechanism with a UAV that is equipped with a ZED camera in both indoor and outdoor environment. They show impressive results in which the UAV is successfully able to navigate through obstacles towards a given goal location. The problem is formulated as a classification task with 26 possible actions (3 possible actions in each dimension corresponding to moving forward, backward, and staying still).
\subsubsection*{Input} The inputs for the UAV are obtained through processing the camera output such that two 3D maps are built: a local 3D map encoding the obstacles within the specified range, and another map which encodes the agent and goal locations. In the second map, after adding the goal position, and before encoding the current position, a 3D convolutional layer is applied to spread the goal information to all pixels.
\subsubsection*{Architecture} The proposed framework is composed of three modules. We briefly describe each module:
\begin{enumerate}
\item Decomposition module: this module slices the input maps along each 2D plane, with slices obtained around the local poisition in each map. Then a permutation of maps formed from the slices of each map and through all combinations of dimensions is obtained. Since only local information in each dimension is required, these slices are assumed to be sufficient for making the next decision.
\item Planning module: for each map pair, the reward is first approximated for each possible action using two convolutional layers. Then, the results are fed into a VIN network for K iterations. The results of this module are Q values approximations for each map pair. In terms of reinforcement learning, the Q value corresponding to each action indicates how profitable it is.
\item Composition module: this module concatenates the slices across plane and then concatenates the results again. Finally, the results are passed through convolutional layer, and then a fully connected layer which output the resulting Q value for each possible action.
\end{enumerate}
The action corresponding to the largest Q value in the final output layer is then chosen. The authors propose using two instances of the proposed framework with different parameters (i.e., size of the local map) such that if the first preferred network fails to generate a collision free path, the output from the others can be selected.
\subsubsection*{Training} using Dijkstra’s algorithm in virtual environments, the authors trained their frameworks through imitation learning.
\subsubsection*{Results} First, it was shown that the computational time required per prediction is around 3 ms, then, through grid-world simulations, the proposed algorithm was shown to achieve superior performance in terms of success rate when compared to RRT offline planning. Also, using virutal environments simulation, the authors showed that their algorithm resulted in shorter travelling distance when compared to RRT*. Finally, the authors tested their mechanism with a real UAV and showed the applicability of their work to both outdoor and indoor environments.
\section{Training a Multilayer Perceptron (MLP) for Online Path Planning}\label{sec:mlp}
To deal with the uncertainties during navigation within a given environment in a timely manner, a NN can be employed such that it will determine the local path of a vehicle that is following a general path that is previously obtained with an offline planner. That is, the NN acts as an online planner. In contrast to the previously existing literature which describes how such a planner can be built and trained, the authors in \cite{onTheTraining} consider how different training data can affect the performance of such planners.
\subsubsection*{NN Architecture} The authors utilize a six-layered NN with 50 neurons in each of the four hidden layers. The input layer consists of 3 neurons corresponding to the input, and the output layer is made of a single neuron. Finally, they use the recitified linear unit (ReLU) as the activation function.
\subsubsection*{Input} The vehicle state is represented by three parameters: $d_{obs}$, $a_{obs}$, and $a_{g}$ that denote the distance to the nearest obstacle, bearing angle to the nearest obstacle, and the bearing angle to goal direction, respectively.
\subsubsection*{Output} Given the state of the vehicle, $\Psi=\{d_{obs},\theta_{obs},\theta_{g}\}$, the NN outputs a value $\theta \in \{-\pi, \pi\}$ which serves as the steering angle deviating from the current direction.
\subsubsection*{Training} Similarly to the previous works, the authors train the NN with offline planning algorithms. They use two different algorithms and compare the achieved results to analyze the impact of algorithm selection. The first algorithm used is the Bellman-Ford (BF) algorithm which is commonly used to find the shortest path in a graph. Using this algorithm results in a discretized solution space obtained through decomposing the area into a grid whereas the vehicle can travel on 8 different paths from each vertex (i.e., through the square edges and diameters).
In contrast, the second algorithm used implements a quadratic program (QP) which solves the optimization problem that minimizes the distance between the current position and goal position at each time instance given the obstacle clearance, maximum velocity, and maximum acceleration constraints.
For each case, the NN is trained using the optimal paths outputs resulting from 10,000 different problems, each with 10 obstacles in the map.
\subsubsection*{Results} Then, the NNs were tested against two different sets of problems: the first one contained 30 maps with similar area and number of obstacles as the training set, and the second set contained 30 maps with double the area and with 80 obstacles instead of 10. It was found that using the BF algorithm with discrete solution space outperforms the QP scenario in both case. Specifically, the BF was able to succeed in avoiding all obstacles and finding the goal location in the first set of problem, whereas it failed at 5 out of 30 in the second more dense set. However, in the case of QP, the NN failed in 5 test maps of the first set, and 21 of the second set. An interesting observation is that when training the NN while using the BF algorithm, the NN was actually able to find paths with shorter distance than those obtained by the BF itself applied as an offline algorithm to the same maps. This is because the offline BF solution space is constrained to the grid waypoints, unlike the steering resulting from the NN.
\section{Path Planning with Moving Obstacles}\label{sec:segmentized}
The uncertainties in an environment can include moving obstacles as well, which was not considered by any of the previous works. However, the authors in \cite{segmentizedView} take into consideration the existence of obstacles moving with a constant speed less than a specified value in a random direction. As shown in figure \ref{fig:segmentized}, and at any time instant, the robot's field of view with an range of $150 ^{\circ}$ is divided into 5 segments (i.e., each segment contains $30 ^{\circ}$ of range) that are spread around the perpendicular line in the direction of the goal.
\begin{figure}[!htp]
\centering
\centerline{\includegraphics[width=0.3\textwidth]{figures/segmentized.eps}}
\caption{Navigating through moving obstacles.}
\label{fig:segmentized}
\end{figure}
The authors proceed by obtaining for each segment a \textit{critical obstacle} which is defined as the moving obstacle reaching the segment in the nearest time. That is, referring to figure \ref{fig:segmentized}, and by assuming that \textbf{O1} is a moving obstacle moving in the direction of \textbf{I}, then this obstacle is selected as the critical obstacle for segment \textbf{3I2}. Also, it is assumed that after each decision interval, the robot moves with a fixed distance of 20 units with a variable speed determined by the NN as will be explained below.
\subsubsection*{NN Architecture} The NN used is a 3-layer MLP with 25, 5, 5 neurons in the hidden, input, and output layers, respectively. As for the activation function, the authors use the sigmoid function.
\subsubsection*{Input} Each of the input layer neurons correspond to a different segment whereas the value of the input is as follows:
\begin{enumerate}
\item If the segment is already occupied by an obstacle, then the input is 0
\item Otherwise, the estimated time duration required for the critical obstacle to reach the segment is fed as an input.
\item A value of 1.5 is fed if there are no obstacles blocking the segment, or critical obstacles approaching it.
\end{enumerate}
\subsubsection*{Output} For each segment, the NN outputs the following:
\begin{enumerate}
\item 0 means no possible movement in that direction because it is blocked.
\item A value ranging from 0.1 to 0.5 mean that the critical obstacle disallows movement in the corresponding segment.
\item A value greater than 0.5 and less than 1.0 is interpreted as a decision for the robot to reach the destination within 0.5 seconds.
\item Whereas if the output is greater than 1.0, then this means that the robot can head to the corresponding location within a duration of 1s.
\end{enumerate}
If more than one segment is allowed, the robot chooses the one that moves it closer to the goal.
\subsubsection*{Training} The authors train the network with given environments by manually calculating the desired outputs for each state, and performing backpropagation to fix the weights of the network.
\subsubsection*{Results} The authors show that the robot is able to avoid obstacles successfully for two different maps, and they compare the performance of their mechanism to other algorithms mentioned in literature in terms of computational time. They show that unlike other algorithms which require a computational time of at least 190 ms, forward propagation through the NN takes only 50 ms to output a response.
\section{Path-planning Based on Neural Activity Dynamics}\label{sec:neuraldynamics}
The authors in \cite{neuralDynamicsOld} first presented a path-planning solution based on a different type of neural networks. These neural networks provide a nerual activity map. We briefly present the solution below.
\subsubsection*{Input} The input needed for the mechanism described below is a map of the complete environment with obstacles locations marked along the goal and start positions.
\subsubsection*{NN Architecture} The neural map generated from the environment map contains a grid of neurons where each neuron is characterized by a shunting equation \cite{shunting} which determines the neuron's activity dynamics. In this model, each neuron is connected to its neighbouring neurons, and both excitatory and inhibitory signals flow throughout the network through these connections. The activity of each neuron, $x_i$, is obtained through solving the following shunting differential equation
\begin{equation}
\frac{dx_i}{dt} = - A x_i+(B-x_i) \underbrace{\left([I_i]^+ +\sum^N_{j=1}w_{ij}[x_j]^+\right)}_{S^+_i} - (D+x_i)[I_i]^-
\end{equation}
where $w_{ij}$ is the weight of the connection between the $j-th$ and $i-th$ and it is proportional to the distance between them up to a certain value where futher neurons are excluded, $[I_i]^+$ and $[I_i]^-$ denote the excitatory and inhibitory inputs respectively, and finally, $A$, $B$, and $D$ represent the passive decay rate, upper bound, and lower bound of the neural activity, respectively. The summation between excitatory input and positive neighbouring neuron activity, $[x_j]^+$, is denoted by $S^+_i$. To model the path planning problem, the inputs at a given neuron (i.e., location) are defined as follows:
\begin{equation}
I_i=
\begin{cases}
E, & \text{if there is a target} \\
-E, & \text{if there is an obstacle}\\
0, & \text{otherwise}
\end{cases}
\end{equation}
whereas $E>>B$ is a large positive constant.
\subsubsection*{Output} Finally, after obtaining complete neural activity map values, and given the starting location, the robot selects the next neuron (i.e., spatial step) that has the maximum neural activity value taking into consideration an additional cost related to the turning angles of the robot. Therefore, the target \textit{attracts} the robot towards it through the propagation of excitatory input, while obstacles \textit{repel} the robot away.
However, the authors of the recent work \cite{neuralDynamics2020} proposed a modification to the proposed mechanism because it was found that the paths generated by this approach were sub-optimal in given situations. A plot of such a case which the authors presented is shown in figure \ref{fig:dynamic}
\begin{figure}[!htp]
\centering
\centerline{\includegraphics[width=0.2\textwidth]{figures/dynamic.eps}}
\caption{Sub-optimal path generated by neural dynamics.}
\label{fig:dynamic}
\end{figure}
To fix this problem of sub-optimal paths generation in challenging environments, the authors propose two modifications that include padding and averaging operations. The padding operation pads neurons to the boundary of the map such that the activity value of each neuron is equal to the average of the neighbouring neurons inside the map. Secondly, they modified the excitatory input modeling $S^+_i$ as follows:
\begin{equation}
S^+_i = [I_i]^+ + \frac{\left(\sum^N_{j=1}w_{ij}[x_j]^+\right)n-n_{\text{neg}}}{n_{\text{pos}}n}
\end{equation}
whereas $n_{\text{neg}}$ and $n_{\text{pos}}$ denote the number of neighbouring neurons with negative and positive activity, respectively, and $n$ is the total number of neighbouring neurons which also includes $n_\text{out}$ which is the number of neurons outside the map that are padded to the boundary of the map as explained above.
\subsubsection*{Results} The authors test their proposed algorithm in both dynamic (i.e., with moving obstacles) and static environments, and they compare their results with several other common path-planning solutions including those based on neural dynamics and neuro-fuzzy tools. The show cases where their proposed method was the only one able to find the optimal path with reduced number of turns and total turning angles. However, the clearance distance from obstacles resulting from their work is less on average from other solutions.
\section{Conclusions}\label{sec:conclusion}
As presented, the application of neural networks to path-planning problems provides a substantial degree of freedom in terms of the problems inputs, outputs, and environments. Different inputs can exist in practical situations (i.e., sensors, cameras, maps, etc.), and neural networks can be designed appropriately as presented above. Furthermore, the output can take several forms such as turning angle, speed, next step location, etc. Finally, the neural network architecture can be designed depending on the amount of information that is available about the environment such as the speed of obstacles, map, or camera input, etc.
\bibliographystyle{ieeetr}
|
\section{Introduction}
Galois covers of the projective line often give interesting examples
of algebraic curves of genus $g\geq 2$. Any such cover is a compact
Riemann surface $C$, endowed with an action of a finite group $G$ such
that $C /G \cong \PP^1$. Studying compact Riemann surfaces with a
$G$-action subject to the latter condition and Galois covers of the
line is equivalent.
To any $G$-action on a Riemann surface one can attach a topological
invariant, the \emph{topological type} (see Section \ref{sec:fam} for
the precise definitions). It turns out that for every fixed
topological type there is a sort of universal family containing all
the covers with the given topological type. These families are often
very interesting and represent interesting loci in the moduli space of
curves.
Moreover it follows from the existence and the properties of these
families that two Riemann surfaces with an action of the same group
are deformation equivalent (through Riemann surfaces with an action)
if and only if they have the same topological type (see \S \ref{fami} for
more details on this point). Thus, if one starts from a collection of
Riemann surfaces and constructs new algebraic varieties out of them,
the topological type controls the deformation equivalence class of the
varieties one obtains.
For these reasons it is very useful to have a list of all the
possibile topological types at least for reasonably small genus. This
is exactly the problem we address: for a positive integer $g \geq 2$,
describe explicitly the set of topological types of (faithful) actions
of a finite group on a Riemann surface of genus $g$ with quotient
isomorphic to ${\mathbb P}^1$. To our knowledge a complete answer is
known only for very small values of $g$. The papers
\cite{kurikuri,kuribayashi-akikazu-kimura} give a complete
classification of the topological types of group actions on Riemann
surfaces of genus $g \leq 5$ without any assumption on the quotient
$C/G$. For higher genus, we do not know of any classification result
of this type. An important contribution in this direction is the
systematic work of Paulhus \cite{paulhus}. She introduces a very
efficient algorithm that yields a database of pairs $(C,G)$ where $C$
is a Riemann surface of genus among $2$ and $15$ and $G$ is a finite
group acting on it with quotient $C/G$ of genus zero. Her database
contains representatives for all possible topological types of groups
actions in this range, but does not give a complete answer to the
question (up to genus $15$) since it does not indicate when two data
in the database share the same topological type. For instance, Paulhus'
database contains $174121$ data for $g=15$, which according to our
computations correspond to $768$ topological types.
Fix $g\geq 2$, a finite group $G$ and $r\geq 3$. It is well known
that
the topological types of holomorphic $G$-actions on a genus $g$
surface with $r$ branch points correspond bijectively to the points of
the quotient $\mathscr{D}^r_g(G) / {\HH}$ where $\mathscr{D}^r_g(G)$ is a finite (but
possibly huge) set and ${\HH}$ is an infinite group acting on it (this
fundamental fact is recalled with some details in Section
\ref{sec:fam}; see also \eqref{def-HH} for the definition of
${\HH}$). Therefore, the object of this paper is the description of
the quotient $\mathscr{D}^r_g(G)/{\HH}$ for fixed $g$ and $G$. Algorithms for
this kind of computation are already known and have been used in
several papers, like for example
\cite{BCGP,fgp,ContiGhigiPignatelli:SomeEvidence,gleissner}, just to
quote the ones closer to our approach. As the genus $g$ increases,
the set $\mathscr{D}^r_g(G)$ becomes quite large, and finding an economical
way of performing the computation becomes essential. In the present
paper we describe an efficient algorithm that computes
$\mathscr{D}^r_g(G)/{\HH}$ for given $g$ and $G$. The computation of
$\mathscr{D}^r_g(G)$ is based on our work
\cite{ContiGhigiPignatelli:SomeEvidence}, and uses some of the ideas
of \cite{paulhus}; the identification of the quotient is new.
We implemented the algorithm using {\texttt{MAGMA}}\xspace\cite{MA}; our implementation
is available at \cite{gullinbursti}. Running the code over several
months on a computer with 56 Intel Xeon 2.60GHz CPU and 128 GB of RAM
we have been able to compute $\mathscr{D}_g(G)/{\HH}$ for all groups $G$ and
$g\leq 39$, with the exception of only three cases. These exceptions
share the same group, $G= (\mathbb{Z}_3\times\mathbb{Z}_3)\rtimes\mathbb{Z} _2$,
and are respectively in genus $g=28,34,37$. See Table
\ref{table:numberofssgbyg} for an account of how many topological
types exist for each genus. See \S \ref{say:results} for some
perspectives on future work related to the group $G$ above and similar
ones.
The results of our computations are collected in a database, available at
\begin{quote}
\href{https://mate.unipv.it/ghigi/tipitopo}{\url{https://mate.unipv.it/ghigi/tipitopo}}.
\end{quote}
The database also contains several other data classifying actions on surfaces of
higher genus (up to $g=100$) under some further constraints. We refer
to the website for the exact indication of these constraints.
The paper is organized as follows: in \S \ref{sec:fam} we recall the
theoretical background, introducing the precise definitions of
$\mathscr{D}^r_g(G)$, ${\HH}$; in \S \ref {sec:algoritmo} we describe the
algorithm, and give some details on its implementation.
\smallskip {\bfseries \noindent{Acknowledgements.} }
The authors wish to thank Matteo Penegini, Fabio Perroni and Michael
Loenne for interesting discussions, and Fabio Gennai for crucial
technical help in the preparation of the website.
\section{Topological types and families of
\texorpdfstring{$G$}{G}-curves}
\label{sec:fam}
The goal of this section is to recall without proofs the mathematics
behind our computation.
A $G$-\emph{curve} is a smooth projective curve $C$ over the complex
numbers together with an effective algebraic action of the group $G$.
We always assume that the genus of $C$ is at least $ 2$ and hence that
$G$ is finite. We also assume throughout the paper that the quotient
$C/G$ is isomorphic to the projective line $ \PP^1$, i.e. the
projection $C \rightarrow C/G$ is a Galois cover of $\PP^1$. Under this
assumption it is completely equivalent to study $G$-curves or Galois
covers of the line.
\begin{definition}
If $C$ and $C'$ are two $G$-curves we say that they are
\emph{topologically equivalent} or that they have the same
\emph{topological type} if there exist $\eta \in \operatorname{Aut} G$ and an
orientation preserving homeomorphism $f: C \rightarrow C'$ such that
$f(g\cdot x) = \eta (g) \cdot f(x)$ for $x\in C'$ and $g\in G$. We say
that $C$ and $C'$ are $G$-\emph{isomorphic} if moreover $f$ is a
biholomorphism.
\end{definition}
These concepts are sometimes called \emph{unmarked} topological type
and isomorphisms, but we will drop the `unmarked' since we do not need
to consider their marked counterparts.
Let $\Gamma_r $ denote the group
\begin{gather*}
\Gamma_r = \langle \gamma_1, \dotsc, \gamma_r\, | \, \gamma_1 \cdots \gamma_r\rangle.
\end{gather*}
\begin{definition}
\label{def:data}
If $G$ is a finite group an $r$-datum is an epimorphism
$\theta : \Gamma_r \rightarrow G$ is such that $\theta(\gamma_i)\neq 1$ for
$i=1, \dotsc, r$.
\end{definition}
The \emph{signature} of $\theta$, denoted ${\mathbf{m}}(\theta)$ or simply
${\mathbf{m}}$, is the vector
\begin{gather*}
{\mathbf{m}}=(m_1,\dotsc, m_r)
\end{gather*}
where $m_i : = {\operatorname{ord}}\, \theta(\gamma_i)$. The \emph{genus} of $\theta$,
denoted by $g(\theta)$, is defined by the Riemann-Hurwitz formula:
\begin{gather}
\label{RH}
2(g(\theta)-1)=|G|\left(-2+\sum_{i=1}^{r}\left(1-\frac{1}{m_{i}}\right)\right)
\end{gather}
By covering theory if one chooses $r$ distinct points $x_1, \dotsc, x_r$
in $\PP^1$, a base point $x_0 \in \PP^1 \setminus \{x_1, \dotsc, x_r\}$
and an isomorphism among $\Gamma_r$ and
$\pi_1(\PP^1 \setminus \{x_1, \dotsc, x_r\}, x_0)$, then a datum
corresponds to a topological Galois covering of
$C^* \rightarrow \PP^1 \setminus \{x_1, \dotsc, x_r\} $ with structure group
$G$. By Riemann's Existence Theorem this compactifies to a
$G$-covering $C \rightarrow \PP^1$.
We let $\mathscr{D}^r_g(G)$ denote the set of all $r$-data of genus $g$
associated with the group $G$.
\begin{definition} \label{sayautstar} Denote by
$\Aut^* \Ga_r \subset \operatorname{Aut} \Gamma_r$ the subgroup of automorphisms $\nu$
satisfying:
\begin{enumerate}
\item for $i=1, \dotsc, n$ the element $\nu (\gamma_i)$ is conjugate to
$\gamma_j$ for some $j$;
\item the automorphism of $H^2(\Gamma_r,{\mathbb{Z}})$ induced by $\nu$ is the
identity.
\end{enumerate}
\end{definition}
The second condition means that $\nu$ is orientation-preserving: if we
identify $\Gamma_r $ with $ \pi_1(\PP^1-\{x_1, \dotsc, x_r\})$
appropriately (using a so-called geometric basis) then $\nu$ is
represented by an orientation preserving self-homeomorphism of
$\PP^1 \setminus \{x_1, \dotsc, x_r\}$ (Dehn-Nielsen-Baer theorem, see
e.g. \cite{fmarga,ivanov,zieschang}).
\begin{say}
\label{auta-azione}
The group $\operatorname{Aut}^*\Gamma_r \times \Aut G $ acts on the set
$\mathscr{D}^r_g(G) $ by the rule
\begin{gather*}
(\nu, \eta) \cdot \theta : = \eta \circ \theta \circ \nu^{-1},
\end{gather*}
where $(\nu, \eta) \in \operatorname{Aut}^*\Gamma_r\times \operatorname{Aut} G$ and
$\theta\in \mathscr{D}^r_g(G)$ is a datum. Moreover
$\operatorname{Inn} \Gamma_r \subset \operatorname{Aut}^* \Gamma_r$ and we set
\begin{gather*}
\Out^*\Ga_r: = \frac{\Aut^* \Ga_r}{\operatorname{Inn} \Gamma_r}.
\end{gather*}
This group has a presentation with generators
$\sigma_1, \dotsc, \sigma_{r-1} $ and relations
\begin{equation}
\label{umpf}
\begin{gathered}
\sigma_i \sigma_j = \sigma_j \sigma_i \quad \text{for} \, \,
|i-j|\geq 2, \qquad
\sigma_{i+1}\sigma_i\sigma_{i+1}=\sigma_i\sigma_{i+1}\sigma_i, \\
\sigma_1 \cdots \sigma_{r-2} \sigma_{r-1}^2 \sigma_{r-2} \cdots
\sigma_1 = 1,\qquad ( \sigma_1 \cdots \sigma_{r-1} )^r =1.
\end{gathered}
\end{equation}
Define $\tilde{\sigma}_i: \Gamma_r \rightarrow \Gamma_r$ as follows:
\begin{gather}
\label{deftztz}
\begin{gathered}
\tilde{\sigma}_i (\gamma_i) = \gamma_{i+1}, \quad \tilde{\sigma}_i
(\gamma_{i+1}) = \gamma_{i+1} ^{-1} \gamma_i \gamma_{i+1}, \\
\tilde{\sigma}_i (\gamma_j ) = \gamma_j \quad \text{for }j \neq i, i+1.
\end{gathered}
\end{gather}
Then $\tilde{\sigma}_i$ is an automorphism of $\Gamma_r$ and it is called the
\emph{$i$-th Hurwitz move}. (The same name also refers to
$\sigma_i \in \Out^*\Ga_r$.) The Hurwitz moves $ \tilde{\sigma}_1,$ $ \dotsc,$
$ \tilde{\sigma}_{r-1}\in\Aut^* \Ga_r$ satisfy the relations \eqref{umpf}. Thus there
is a (unique) morphism $\phi : \Out^*\Ga_r \rightarrow \Aut^* \Ga_r$ such that
$\phi(\sigma_i) : = \tilde{\sigma}_i$. Using $\phi$ we let
$\operatorname{Aut} G \times \Out^*\Ga_r$ act on $\mathscr{D}^r_g$: if
$(\eta,\sigma) \in \operatorname{Aut} G \times \Out^*\Ga_r$ and $\theta \in \mathscr{D}^r_g$,
then
\begin{gather*}
(\eta,\sigma) \cdot \theta := \eta \circ \theta \circ
\phi(\sigma)^{-1}.
\end{gather*}
The actions of $\operatorname{Aut} G \times \Aut^* \Ga_r $ and of $\operatorname{Aut} G\times \Out^*\Ga_r$ on
$\mathscr{D}^r_g$ have the same orbits, hence
\begin{gather*}
\mathscr{D}^r_g / \operatorname{Aut} G \times \Aut^* \Ga_r = \mathscr{D}^r_g / \operatorname{Aut} G \times \Out^*\Ga_r.
\end{gather*}
(For all this see \cite[\S 4.15]{gt}.) The orbits of the
$\Out^*\Ga_r \times \Aut G$--action are called \emph{Hurwitz equivalence
classes} and elements in the same orbit are said to be
\emph{Hurwitz equivalent}.
\end{say}
\begin{teo}
Fix $g \geq 2 $ and a finite group $G$, the topological types of
$G$-curves $C$ with $g(C)=g$, $g(C/G)= 0$ and $r$ branch points are
in bijection with the set
\begin{gather}
\label{set}
\mathscr{D}^r_g(G) / \left( \operatorname{Aut} G \times \Out^*\Ga_r \right).
\end{gather}
\end{teo}
A proof can be found for example in \cite[Section 5]{gt}.
\begin{say}\label{fami}
It often happens that one is not interested in a precise $G$-curve,
but rather in the whole family of $G$-curves of a given topological
type. Indeed, there is a sort of universal family containing all
$G$-curves of a given topological type. These ``universal''
families have been widely studied and used in the literature for
several purposes in the last decades, see
e.g. \cite{BCGP,baffo-linceo,clp,clp2,moonen-special,fgp,fn,friedv,penegini2013surfaces,perroni,volk}.
They were first constructed by Gonz\'alez-D\'{i}ez and Harvey
\cite{gabino} using Teichm\"uller theory. There are other ways to
construct it. Recently in \cite{gt} the second author and Tamborini
gave a different construction of these families and corrected an
inaccuracy in \cite{gabino}. The precise statement is a bit long
and there is no need to recall it here in full detail. At any rate,
the end result is that for any topological type there is a family
of $G$-curves whose fibres all have the given topological type, and
which is universal in the following sense: every $G$-curve with the
given topological type appears as a fibre of the family and it
appears at most a finite number of times. Moreover the base of this
family is an \'etale cover of the $\mathsf{M}_{0,r}$, so in
particular it is smooth and connected. This ``universal'' family is
not unique, but only unique up to the equivalence relation generate
by finite \'etale pull-backs. We refer to the Introduction in
\cite{gt} for full details.
An important consequence of this theorem is that topological types
and deformation equivalence coincide for $G$-curves. Indeed if $C$
and $C'$ are $G$-curves with the same topological type, then they
both appear as fibres of a common universal family and therefore
they are deformation equivalent (through $G$-curves). The converse
is obvious. This fact is very important in the applications to the
construction of new deformation types of algebraic varieties as in
\cite{BCGP,BP12,penegini2011surfaces,penegini2013surfaces,polizzi,
carnovale polizzi,mistretta polizzi,FP,gleissner,BP16,polizzi
pignatelli, pignatelli}. It follows from this discussion that the
classification of \emph{universal families} and of deformation
equivalence classes of $G$-curves are both equivalent to the
classification of topological types. Hence again these problems
boil down to studying the quotient in \eqref{set}. This is a strong
additional motivation --- in fact, our original motivation --- for
the classification of topological types.
\end{say}
\section{The algorithm}
\label{sec:algoritmo}
We illustrate an algorithm to attack the following:
\begin{problem}
\label{problem:classification}
Given a number $g_{\max}$ list all the topological types of
$G$-curves with $g(C) \leq g_{\max}$ and $C/G \cong \PP^1$.
\end{problem}
After fixing the genus $g$, the group $G$ and the number of branch
points $r$, this amounts to listing representatives of the quotient in
\eqref{set}.
\begin{say}
\label{ignore_order}
We use a refinement of the algorithm illustrated in
\cite{ContiGhigiPignatelli:SomeEvidence}, which lists $r$-data
forming counter-examples to the Coleman-Oort conjecture up to the
action of $\operatorname{Aut} G$. This algorithm uses signature as an invariant
for the classification, as was done in \cite{BCGP,fgp}. A
\emph{spherical system of generators} of the group $G$ is a list
$(g_1, \dotsc, g_r)$ such that (1) $g_i\neq 1$ for any $i$, (2) $G$ is
generated by $g_1, \dotsc, g_r$ and (3) $g_1\cdots g_r = 1$. Having
fixed a finite group $G$, giving an $r$-datum $\theta$ is equivalent
to giving a spherical systems of generators $(g_1, \dotsc, g_r)$ of
$G$: simply define $\theta$ or $g_i$ by the relation
$\theta(\gamma_i)=g_i$. From now on we will identify data and
spherical systems of generators and we will write a datum in
$\mathscr{D}^r_g(G)$ as $(g_1, \dotsc, g_r)$. Signature defines a map
\begin{equation}
\label{eqn:signaturemap}
\mathscr{D}^r_g(G) \rightarrow {\mathbb{N}}^r, \quad (g_1, \dots, g_r)\mapsto ( {\operatorname{ord}}(g_1), \dotsc, {\operatorname{ord}}(g_r)).
\end{equation}
With the interpretation of $\mathscr{D}^r_g(G)$ just described, we have
$\mathscr{D}^r_g (G) \subset G^r$, and the action of $\Out^*\Ga_r\times \Aut G$ on $\mathscr{D}^r_g(G)$
described in \eqref{umpf} and \eqref{deftztz} extends to an action
on $G^r$: $\operatorname{Aut} G$ acts componentwise,
\begin{gather*}
\eta\cdot (g_1,\dotsc, g_r)=(\eta(g_1),\dotsc, \eta(g_r),
\end{gather*}
while the generator $\sigma_ i $ of $\Out^*\Ga_r$ acts as follows
\begin{gather}
\label{azione}
\sigma_i \cdot (g_1, \dotsc, g_r) = (g_1,\dots,\underbrace{g_i
g_{i+1}g_i^{-1}}_i,\underbrace{g_{i}}_{i+1},\dotsc, g_r).
\end{gather}
Denote by $\Sigma_r$ the symmetric group. There is a surjective
morphism
\begin{equation}
\label{rhouta}
\rho : \outa} % {\mathbf{B_r} \rightarrow \Sigma_r,
\end{equation}
which maps the $i$-th Hurwitz move $\sigma_i$ to the transposition
$(i, i+1)$. The map \eqref{eqn:signaturemap} is $\rho$-equivariant:
if $\psi$ is in $\Out^*\Ga_r$ and $\sigma=\rho (\psi)$ then
$( \psi \cdot (g_1 , \dotsc, g_r) )$ is mapped to
$({\operatorname{ord}}(g_{\sigma_1}) , \dotsc, {\operatorname{ord}}(g_{\sigma_r}))$.
Recall that Hurwitz equivalence in $\mathscr{D}^r_g(G)$ is defined in terms
of the action of $\outa} % {\mathbf{B_r}\times\operatorname{Aut} G$; thus,
\eqref{eqn:signaturemap} maps Hurwitz equivalence classes onto
$\Sigma_r$-orbits in ${\mathbb{Z}}^r$. This shows that every Hurwitz
equivalence class has a representative with signature of the form
\[{\mathbf{m}}=(m_1,\dotsc, m_r), \quad m_1\leq \dots \leq m_r,\]
and ${\mathbf{m}}$ is uniquely determined by the equivalence class.
\end{say}
\begin{say}
As a first step we wish to determine the set of all possible
signatures. We iterate over the order $d=|G|$. For fixed $d$, let
$\mathfrak{S}_{d,g_{\max}}$ be the set of finite sequences
${\mathbf{m}}=(m_1,\dotsc, m_r)$ such that
\begin{enumerate}[label=(S\arabic*)]
\item \label{item:S1} $3\leq r\leq \frac{4(g_{\max}-1)}d+4$;
\item \label{item:S2} each $m_i$ is a divisor of $d$;
\item \label{item:S3} $2\leq m_i\leq d$;
\item \label{item:S4} $g$, determined by $d$ and ${\mathbf{m}}$ as in
\eqref{RH}, is an integer between $2$ and $g_{\max}$.
\item \label{item:S4b} if $r=3$, $d$ satisfies the bound of
\cite[Appendix 1]{Conder:Large group actions on surfaces} as long
as $g \leq 301$.
\item \label{item:S5}$m_1\leq \dots \leq m_r$;
\end{enumerate}
It follows from the same arguments as in the proof of Hurwitz
theorem that the signature of any $G$-curve satisfies \ref{item:S1},
see e.g. Lemma 3.2 (b) in \cite
{ContiGhigiPignatelli:SomeEvidence}. Notice that we do not exclude
$r=3$ or cyclic groups here, unlike in
\cite{ContiGhigiPignatelli:SomeEvidence}.
Computing $\mathfrak{S}_{d,g_{\max}}$ is not difficult (see
e.g. \cite[Algorithm 1]{ContiGhigiPignatelli:SomeEvidence}). In our
implementation, we found it convenient to store the resulting
signatures on disk for later retrieval; this allows us to iterate
through signatures with fixed genus $g$ at a later stage, in order
to compute the sets $\mathscr{D}_g(G)$. Upon retrieving signatures
with fixed genus, we make use of the inequalities $d\leq 84(g-1)$
(for $r=3$) and $d\leq 12(g-1)$ (for $r\geq 4$, see \cite[Lemma
3.2.(c)]{ContiGhigiPignatelli:SomeEvidence}), which do not appear in
the definition of $\mathfrak{S}_{d,g_{\max}}$ since they are
consequences of \eqref{RH}.
\end{say}
\begin{say}
Problem~\ref{problem:classification} can be addressed by iterating
through the signatures in $\mathfrak{S}_{d,g_{\max}}$. For each
signature, we iterate through isomorphism classes of groups $G$ of
order $d$. Some groups can be eliminated right away, namely:
\begin{itemize}
\item groups $G$ that do not contain elements of order $m_i$ for
some $m_i$ in the signature;
\item groups $G$ that contain elements of order greater than $4g+2$
(for $r=3$) or $4(g-1)$ (for $r>3$);
\item groups that cannot be generated by $r-1$ elements because
their abelianization cannot.
\end{itemize}
Similar exclusions are listed in \cite[Algorithm
2]{ContiGhigiPignatelli:SomeEvidence}, with some differences due to
the fact that we allow $r=3$ and cyclic groups here.
\end{say}
\begin{say}
What remains to be done is, in fact, the most complicated and most
novel part of our work, namely producing an algorithm to classify
spherical systems of generators for fixed $G$ and ${\mathbf{m}}$. The
construction of this algorithm will take the rest of this section.
Fix a group $G$ of order $d$ and a signature ${\mathbf{m}}$, and let $\data_{G,\mm}}%X_{G,\mathbf{m}}$
be the set of spherical systems of generators of $G$ with signature
${\mathbf{m}}$. The group
\[(\outa} % {\mathbf{B_r})_{\mathbf{m}}=\{g\in \outa} % {\mathbf{B_r}\mid \rho(g)\cdot{\mathbf{m}}={\mathbf{m}}\}\]
acts on $\data_{G,\mm}}%X_{G,\mathbf{m}}$, and so does $\operatorname{Aut} G$. We will need the following:
\end{say}
\begin{prop}
\label{prop:classify}
Given a group $G$ of order $d$ and ${\mathbf{m}}$ in $\mathfrak{S}_{d,g_{\max}}$, two elements
$\theta,\theta'$ of $\data_{G,\mm}}%X_{G,\mathbf{m}}$ are Hurwitz equivalent if and only if
they are in the same $(\Out^*\Ga_r)_{{\mathbf{m}}}\times \Aut G$-orbit.
\end{prop}
\begin{proof}
If $\theta$ and $\theta'$ are in the same
$(\Out^*\Ga_r)_{{\mathbf{m}}}\times \Aut G$-orbit, they are obviously Hurwitz
equivalent.
Conversely, suppose $\theta'=(\nu,\alpha)\cdot\theta$, with
$\nu\in\outa} % {\mathbf{B_r}$, $\alpha\in\operatorname{Aut} G$; by equivariance of
\eqref{eqn:signaturemap}, the signature of $\theta'$ is
${\mathbf{m}}=\rho(\nu)\cdot{\mathbf{m}}$, so $\nu$ lies in $(\outa} % {\mathbf{B_r})_{\mathbf{m}}$.
\end{proof}
\begin{say}
\label{say:determinesection}
Set for simplicity
\begin{gather}
\label{def-HH}
\mathbf{H}:= \Out^*\Ga_r \times \operatorname{Aut} G.
\end{gather}
Then $ \mathbf{H}_{\mathbf{m}} = (\Out^*\Ga_r)_{\mathbf{m}} \times \operatorname{Aut} G $. We are reduced to the
following problem: given a group $G$ of order $d$ and ${\mathbf{m}}$ in
$\mathfrak{S}_{d,g_{\max}}$, determine a \emph{section} for the action of $\HH_\mm$ on
$\data_{G,\mm}}%X_{G,\mathbf{m}}$, that is to say a subset of $\data_{G,\mm}}%X_{G,\mathbf{m}}$ that contains exactly one
element in each $\HH_\mm$-orbit.
The approach used in \cite{BCGP,fgp} was to iterate through lists of
elements in
\begin{gather*}
G^{\mathbf{m}}=\{(g_1,\dotsc, g_r)\in G^r\mid
{\operatorname{ord}}(g_1)=m_1,\dotsc, {\operatorname{ord}}(g_r)=m_r\}
\end{gather*}
and identify those for which $\prod g_i=1$ and
$\langle g_1,\dotsc, g_r\rangle =G$. This produces a set of
spherical systems of generators which can become quite large as $r$
or $|G|$ increase; the fact that we are ultimately interested in
extracting a representative for each orbit of $\HH_\mm$ suggests that an
alternative approach could be more suited to our goal.
A preliminary observation is that, whilst the group $\HH_\mm$ is
infinite, it can be replaced by its image in
$\operatorname{Perm}(\data_{G,\mm}}%X_{G,\mathbf{m}})$, the symmetric group over $\data_{G,\mm}}%X_{G,\mathbf{m}}$, which
is finite. In prior algorithms, precisely first in the paper
\cite{BCGP} and then in
\cite{BP12,fgp,fpp,FP,BP16,gleissner,ContiGhigiPignatelli:SomeEvidence},
the orbit of an element is calculated by a heavy recursive procedure
that builds an increasing chain of sets by the action of a fixed set
of generators of $\mathbf{H}$ and stops when the chain stabilizes. By
contrast, here we build first the image of $\HH_\mm$ as subgroup of
$\operatorname{Perm}(\data_{G,\mm}}%X_{G,\mathbf{m}})$, then compute directly all orbits
without any recursion.
After this first step, the problem boils down to extracting a
section for the action of a finite group on a finite set. This can
be achieved efficiently in {\texttt{MAGMA}}\xspace using \texttt{GSets}. It is then
the size of the finite group and of the finite set $\data_{G,\mm}}%X_{G,\mathbf{m}}$ that
determine memory use and execution time.
The basic idea to reduce both is to split the action of $\HH_\mm$ on
$G^{\mathbf{m}}$ into a large number of actions, each one by a much smaller
group acting on a much smaller set. The key observation is the
following. Consider the set $\mathcal{C}_G$ of conjugacy classes in
$G$, and consider the commutative diagram
\begin{equation}
\label{diag:Gr}
\xymatrix{ G^r\ar[r]^p\ar[dr]^\pi & (\mathcal{C}_G)^r\ar[d]\\
&(\mathcal{C}_G)^r/\Sigma_r}
\end{equation}
where $p(g_1,\dotsc, g_r)=([g_1],\dotsc, [g_r])$. Setting
$\mathcal{C}_G^{\mathbf{m}}:= p (G^{\mathbf{m}})$, we obtain a commutative
diagram
\begin{equation}
\label{diag:Gm}
\xymatrix{G^{\mathbf{m}}\ar[r]^p\ar[dr]^\pi & \mathcal{C}_G^{\mathbf{m}}\ar[d]\\
&\mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}}
\end{equation}
with $(\Sigma_r)_\mathbf{m}$ denoting the stabiliser of $\mathbf{m}$
in $\Sigma_r$. The group $\mathbf{H}$ acts also on $(\mathcal{C}_G)^r$ in
the following way:
\begin{equation*}
\begin{aligned}
\eta\cdot ([g_1],\dotsc, [g_r])&=(\eta([g_1]),\dotsc, \eta([g_r]), \qquad \eta \in \operatorname{Aut} G\\
\sigma_i\cdot ([g_1],\dotsc, [g_r])&=
([g_1],\dots,\underbrace{[g_{i+1}]}_i,\underbrace{[g_{i}]}_{i+1},\dotsc,
[g_r]).
\end{aligned}
\end{equation*}
This means that if $(\eta,\psi) \in \mathbf{H}$ and $ \sigma = \rho(\nu)$,
then
\begin{gather}
\label{pVppa}
(\eta,\nu)\cdot ([g_1],\dotsc, [g_r]) = ([\eta(g_{\sigma_1}], \dotsc,
[\eta(g_{\sigma_n}]).
\end{gather}
For $\tilde{c} \in (\mathcal{C}_G)^r$, let $\mathbf{H}_{\tilde {c}}$
denotes the stabilizer for this action.
Since $(\mathcal{C}_G)^r/\Sigma_r$ is the quotient of
$(\mathcal{C}_G)^r$ by $\outa} % {\mathbf{B_r}$, it has an induced action of
$\mathbf{H}$
with the factor $\outa} % {\mathbf{B_r}$ acting trivially. In other words $\operatorname{Aut}
G$ acts naturally on $(\mathcal{C}_G)^r/\Sigma_r$ and we let
$\Out^*\Ga_r$ act trivially on this set. With this understood the map
$p$ and the whole Diagram \eqref{diag:Gr} is
$\mathbf{H}$-equivariant, while \eqref{diag:Gm} is $\HH_\mm$-equivariant.
\end{say}
Notice that a section $S$ for the action of $\operatorname{Aut}
G$ on
$\mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}$ can be
constructed adapting \cite[Algorithm
4]{ContiGhigiPignatelli:SomeEvidence}. From this, we recover a
section for the action of $\HH_\mm$ on $G^{\mathbf{m}}$ as follows:
\begin{prop}
\label{prop:section}
Let $S\subset
\mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}$ be a section for
the action of $\operatorname{Aut} G$, and let $\tilde
S$ be a subset of
$\mathcal{C}_G^{\mathbf{m}}$ that projects one-to-one onto
$S$. Let $S'\subset G^{\mathbf
m}$ be the union of sections for the action of $\mathbf{H}_{\tilde
c}$ on $p^{-1}(\tilde c)$, as $\tilde c$ varies in $\tilde
S$. Then $S'$ is a section for the action of
$\HH_\mm$ on $\mathcal{C}_G^{\mathbf{m}}$.
\end{prop}
\begin{proof}
We need to show that $\HH_\mm\cdot S'=G^{\mathbf{m}}$ and that two elements $X,Y
\in S'$ that belong to the same $\HH_\mm$-orbit coincide.
To see that $\HH_\mm\cdot S'=G^{\mathbf{m}}$, pick $X=(g_1, \dotsc, g_r)$ in
$G^{\mathbf{m}}$. Up to the action of $\operatorname{Aut} G$, we can assume that
$\pi(X)\in S$. Then $p(X)=\sigma\cdot \tilde c$ for some
$\sigma\in(\Sigma_r)_{\mathbf{m}}$ and some $\tilde c\in \tilde S$. If
$\sigma=\rho(\nu)$, where $\rho$ is the map in \eqref{rhouta}, then
$\nu$ is in $(\Out^*\Ga_r)_{{\mathbf{m}}}$ and
$\nu^{-1} \cdot X =\nu^{-1}\cdot(g_1,\dotsc, g_r)$ is in
$p^{-1}(\tilde c)$, so its $\mathbf{H}_{\tilde c}$-orbit intersects
$S'$. As $\mathbf{H}_{\tilde c} \subset \HH_\mm$ we conclude that the
$\HH_\mm$-orbit of $X$ intersects $S'$, as desired.
Next assume that $X$, $Y$ are elements of $S'$ that belong to the
same $\HH_\mm$-orbit. By equivariance $\pi(X)$ and $\pi(Y)$ are in the
same $\HH_\mm$-orbit. But $(\Out^*\Ga_r)_{\mathbf{m}}$ acts trivially on
$\mathcal{C}_G^{\mathbf{m}}/(\Sigma_r)_{\mathbf{m}}$. Hence $\pi(X)$ and $\pi(Y)$ are
in the same $\operatorname{Aut} G$-orbit. By construction they also belong to $S$;
therefore, they coincide. Since $\tilde c=p(X)$ and $p(Y)$ are
elements of $\tilde S$ that lie over the same element of
$\mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}$, they also
coincide. By equivariance, $X$ and $Y$ are in the same $\HH_\mm$-orbit
only if they are in the same $\mathbf{H}_{\tilde c}$-orbit, which forces
them to be the same.
\end{proof}
A first advantage of this approach is that for some $G$ and
$\mathbf{m}$, $p^{-1}(\tilde c)$ can be considerably smaller than the
whole $G^{\mathbf{m}}$, leading to a reduced memory usage. In
addition, $\mathbf{H}_{\tilde c}$ has index
\[[\HH_\mm:\mathbf{H}_{\tilde c}]=[(\outa} % {\mathbf{B_r})_{\mathbf m} : (\outa} % {\mathbf{B_r})_{\tilde c}]
\cdot [\operatorname{Aut} G : (\operatorname{Aut} G)_{\tilde c}].\] This means that
$\mathbf{H}_{\tilde{c}}$ is typically smaller than $\HH_\mm$ and the image of
$\mathbf{H}_{\tilde c}$ in $\operatorname{Perm}(p^{-1}(\tilde c))$ is smaller
than the image of $\HH_\mm$ in $\operatorname{Perm}(G^{\mathbf{m}})$, in a
way that more than compensates for the fact that one has to iterate
through elements of $S$.
\begin{ex}
\label{example:Sigma4}
For example, take $G=\Sigma_4$ and
$\mathbf m=(\underbrace{2,\dots,2}_r)$. Then $G$ has nine elements
of order two, namely the conjugacy class of $(1,2)$ and the
conjugacy class of $(1,2)(3,4)$, call them $C_1$ and $C_2$. Then
$|C_1|=6$, $|C_2|=3$ and
\[G^{\mathbf m}=(C_1\cup C_2)^r\] has $9^r$ elements.
However, if
\begin{equation}
\label{eqn:tildecsigma4}
\tilde c=(\underbrace{C_1,\dotsc, C_1}_k,\underbrace{C_2,\dotsc, C_2}_{r-k}),
\end{equation}
up to the order, then $p^{-1}(\tilde c)$ has $6^k 3^{r-k}$ elements,
which is considerably less. It follows from \eqref{pVppa} that
$(\Out^*\Ga_r)_{\tilde c} = \rho^{-1} ((\Sigma_r)_ {\tilde c}) $, so
$\Out^*\Ga_r / (\Out^*\Ga_r)_{\tilde c} \cong \Sigma_r /(\Sigma_r)_{\tilde
c}$. By \eqref{eqn:tildecsigma4}
$(\Sigma_r)_{\tilde {c}} \cong \Sigma_k \times \Sigma_{r-k}$, so
$[\Out^*\Ga_r : (\Out^*\Ga_r)_{\tilde c} ] = \binom{r}{k}$. Moreover in this
special case $\operatorname{Aut} G$ fixes every conjugacy class, as every
automorphism of $G=\Sigma_4$ is inner \cite[vol. 1 Satz 5.5
p. 175]{Huppert} or \cite[Cor. 7.7 p. 159]{rotman}. Hence $\operatorname{Aut} G$
fixes $\tilde c$, and
$\mathbf{H}_{\tilde c}=(\outa} % {\mathbf{B_r})_{\tilde c}\times \operatorname{Aut} G$ has index
$\binom{r}{k}$ in $\HH_\mm$. Furthermore,
$\mathcal{C}_G^{\mathbf{m}} / (\Sigma_r)_{\mathbf{m}} $ has $r+1$
elements indexed by $k$ as above and $\operatorname{Aut} G$ acts trivially. We
have split the action into $r+1$ actions of much smaller size.
\end{ex}
\begin{say}
In order to apply Proposition~\ref{prop:section}, one needs to
compute the stabilizer $\mathbf{H}_{\tilde c}$ of an element $\tilde c$ in
$\mathcal{C}_G^{\mathbf{m}}$.
In the above example $\mathbf{H}_{\tilde{c}}$ splits as a product of
$(\Out^*\Ga_r)_{\tilde{c}}$ and $(\operatorname{Aut} G)_{\tilde{c}}$. In general
$(\Out^*\Ga_r)_{\tilde{c}} \times (\operatorname{Aut} G)_{\tilde{c}}$ is only a subgroup
of $\mathbf{H}_{\tilde{c}}$. In fact an element of $\Out^*\Ga_r$ can move the
conjugacy classes and element of $\operatorname{Aut} G$ can restore them to their
original order. The situation in the general case is described by
the exact sequence
\begin{equation}
\label{eqn:Hc}
1\to (\outa} % {\mathbf{B_r})_{\tilde c}\stackrel{\alpha}{\longrightarrow} \mathbf{H}_{\tilde c}\stackrel{\beta}{\longrightarrow} (\operatorname{Aut} G)_c\to 1,
\end{equation}
where $\alpha$ is the inclusion: $\alpha(\psi) = (\psi,1)$, $\beta$ is
the projection $\beta(\psi, \eta) = \eta$, and $c$ is the image of
$\tilde c$ in the quotient
$\quozient{\mathcal{C}_G^{\mathbf{m}}}{(\Sigma_r)_{\mathbf{m}}}$. Indeed if
$(\psi,\eta) \in \mathbf{H}_{\tilde{c}}$ and $\rho(\psi) = \sigma$, then by
\eqref{pVppa} we have
\begin{gather*}
(\psi,\eta)\cdot (\tilde{c}_1, \dotsc, \tilde{c}_r) =
(\eta(\tilde{c}_{\sigma(1)}), \dotsc, \eta(\tilde{c}_{\sigma(r)}) )
= (\tilde{c}_1, \dotsc, \tilde{c}_r).
\end{gather*}
Thus $\eta$ permutes the elements $\tilde{c}_i$ i.e. it fixes $c$.
This shows that $\beta$ lands in $(\Aut G)_c$. Obviously
$\beta \alpha = 1$. If $(\psi,\eta) \in \ker \beta$, then
$\eta=\operatorname{id}_G$ hence $\sigma =1$ and $\psi \in (\Out^*\Ga_r)_{\tilde{c}}$.
If $\eta \in (\Aut G)_c$, then $\eta(c_i) = c_{\sigma(i)}$. Since
$\rho$ is surjective there is $\psi \in \Out^*\Ga_r$ such that
$\rho(\psi) = \sigma^{-1}$ and then
$(\psi,\eta) \in \mathbf{H}_{\tilde{c}}$. Thus $\beta$ is onto.
In particular, one can obtain a set of generators for
$\mathbf{H}_{\tilde c}$ by choosing elements $\gamma_1,\dotsc, \gamma_k$
such that $\beta(\gamma_1),\dotsc, \beta(\gamma_k)$ generate
$(\operatorname{Aut} G)_c$ and adding a set of generators of
$(\outa} % {\mathbf{B_r})_{\tilde c}$.
So we need to find a set of generators of $(\outa} % {\mathbf{B_r})_{\tilde c}$. We
start with the following observation. An element of
$(\mathcal{C}_G)^r/\Sigma_r$ (or of
$\mathcal{C}_G^{\mathbf{m}} / \Sigma_{\mathbf{m}}$) is represented in our
implementation as a \emph{multiset}, i.e.
$(C_1^{r_1}, \dotsc, C_s^{r_s})$ where $C_i \in \mathcal{C}_G$ and
$r_i \geq 1$. To each multiset of conjugacy classes $c$ there
correspond several ordered sequences of conjugacy classes
$\tilde{c}$. We fix a total ordering on the set of conjugacy
classes $\mathcal{C}_G$. Then, among all sequences $\tilde{c}$
corresponding to the same $c$ there is a minimal one with respect to
the induced lexicographic order on $\mathcal{C}^r_G$, i.e. the only
nondecreasing one:
\begin{gather}
\label{pio}
\tilde{c}=(C_1,\dotsc, C_r) \quad \text{ with }\quad C_1\leq \dots
\leq C_r.
\end{gather}
For ease of notation, multisets will be represented by the
associated nondecreasing sequence in the following.
In order to find generators of $(\outa} % {\mathbf{B_r})_{\tilde c}$ one can
consider the exact sequence
\[0\to \mathcal{P}\outa} % {\mathbf{B_r} \to \outa} % {\mathbf{B_r} \stackrel{\rho}{\to} \Sigma_r\to
0,\] where $\mathcal{P}\outa} % {\mathbf{B_r}$ denote the group of pure braids,
i.e. $\ker \rho$. We also have the following exact sequence
obtained by restriction:
\[0\to \mathcal{P}\outa} % {\mathbf{B_r} \to (\outa} % {\mathbf{B_r})_{\tilde c} \to
(\Sigma_r)_{\tilde c}\to 0.\] Recall from \cite[p. 20]{birman}
that $\mathcal{P}\outa} % {\mathbf{B_r}$ is generated by the elements
\[A_{ij}=\sigma_{j-1}\sigma_{j-2}\dotsm\sigma_{i+1}\sigma_i^2\sigma_{i+1}^{-1}\dotsm
\sigma_{j-2}^{-1}\sigma_{j-1}^{-1}, \quad 1\leq i<j\leq r.\] Since
we chose $\tilde{c}$ with $C_1\leq \dots \leq C_r$, the subgroup
$(\Sigma_r)_{\tilde c}$ is generated by transpositions $(i\ i+1)$
where $i$ is such that $C_i=C_{i+1}$. It follows that
$(\outa} % {\mathbf{B_r})_{\tilde c}$ is generated by
\[\{A_{ij}\mid 1\leq i<j\leq r\}\cup \{\sigma_i \mid 1\leq i\leq
r-1, C_i=C_{i+1}\};\] notice however that when $C_i=C_j$, then
$C_i=C_{i+1}=\dots = C_j$, so $A_{ij}$ belongs to the group
generated by $\sigma_i,\dotsc, \sigma_{j-1}$, and it is redundant as
a generator of $(\outa} % {\mathbf{B_r})_{\tilde c}$. This gives
Algorithm~\ref{algo:Hc}.
\begin{algorithm}[th] \SetKwInOut{Input}{input}
\SetKwInOut{Output}{output} \Input{A group $G$ and an element
$\tilde c=(C_1,\dotsc, C_r)\in\mathcal{C}_G^r$,
$C_1\leq \dots\leq C_r$} \Output{A set of generators for
$\mathbf{H}_{\tilde c}$}
$\Gamma\leftarrow \{A_{ij}\mid 1\leq i<j\leq r, C_i\neq C_j\}\cup
\{\sigma_i \mid 1\leq i\leq r-1, C_i=C_{i+1}\}$\; \For{$\phi$ in a
set of generators for $(\operatorname{Aut} G)_{c}$}{ $\sigma\leftarrow$ a
permutation such that $ \phi\cdot \tilde c=\sigma\cdot\tilde c$\;
$\alpha_{i_1}\dotsm \alpha_{i_k}\leftarrow$ a decomposition of
$\sigma$ as a product of transpositions $\alpha_j=(j\,j+1)$\;
$\psi\leftarrow\sigma_{i_1}\dotsm \sigma_{i_k}$\; add
$(\psi^{-1},\phi)\in\outa} % {\mathbf{B_r}\times\operatorname{Aut} G$ to $\Gamma$ } \Return
$\Gamma$
\caption{Computing $\mathbf{H}_{\tilde c}$}\label{algo:Hc}
\end{algorithm}
\end{say}
\begin{say}
The problem considered in this paper is to compute effectively a
section for the action of $\mathbf{H}$ on $\mathscr{D}^r(G)$. Logically, the
problem can be split in two parts: first computing a section for
the action of $\mathbf{H}$ on $G^r$, next checking which elements of the
section are spherical systems of generators of $G$. For reasons of
efficiency, our algorithm does not attack the two problems in
sequence, but simultaneously.
Proposition~\ref{prop:section} reduces the first problem to
determining a section for the action of $\mathbf{H}_{\tilde c}$ on each
$p^{-1}(\tilde c)$. In view of the second part, two more
optimizations are important, already used by Breuer \cite{breuer}
and Paulhus \cite{paulhus}. Indeed, for some elements $c$, one
can ascertain \emph{a priori} that $\pi^{-1}(c)=p^{-1}(\tilde c)$
does not contain any system of generators at all!
\end{say}
This is based on a theorem of Frobenius, see
\cite[p. 406]{lando-zvonkin} for a proof and also \cite{jones} for a
generalization to higher genus.
\begin{teo}[Frobenius's formula] Given a finite group $G$ and
conjugacy classes $C_1,\dotsc, C_r$, the number of $r$-ples
$(g_1,\dotsc, g_r)\in C_1\times \dots \times C_r$ such that
$\prod g_i=1$ is
\[\frac{|C_1|\dotsm |C_r|}{|G|} \sum_{\chi} \frac{\chi(C_1)\dotsm
\chi(C_r)}{\chi(1)^{r-2}},\] where the sum is over characters
of irreducible representations of $G$.
\end{teo} Thus, $p^{-1}(C_1,\dotsc, C_r)$ can only contain a system
of spherical generators if
$\sum_{\chi} \frac{\chi(C_1)\dotsm \chi(C_r)}{\chi(1)^{r-2}}$ is
nonzero; in this case, we will say that $(C_1,\dotsc, C_r)$ passes
Frobenius' test. Notice that this condition is independent of the
order of the conjugacy classes $C_1, \dotsc, C_r$.
\begin{ex} In Example~\ref{example:Sigma4}, the fact that the
character of the alternating representation satisfies
$\chi(C_1)=-1$ and $\chi(C_2)=1$ easily implies that the Frobenius
formula evaluates to zero for $k$ odd in
\eqref{eqn:tildecsigma4}. This eliminates half of the elements of
$S$.
\end{ex}
A second condition is based on a theorem of Scott \cite[Theorem
1]{Scott}:
\begin{teo}[Scott]
\label{thm:Scott} Let $G$ be a group generated by
$g_1,\dotsc, g_n$ with $g_1\dotsm g_n=1$ and let $V$ be a
finite-dimensional representation of $G$ over any field. Then
\[\sum_{i=1}^n v(g_i)\geq v(G)+v(G^*),\] where $v(g_i)=\operatorname{codim}
V^{g_i}$, $v(G)=\operatorname{codim} V^G$, $v(G^*)=\operatorname{codim} (V^*)^G$.
\end{teo} Since $v(g_i)$ only depends on the conjugacy class of
$g_i$, and the order of the $g_i$ is irrelevant for the condition,
we see that Scott's theorem determines a test to identify the
$c\in \mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}$ which can
potentially have a system of generators in their preimage. We will say
that $c$ passes Scott's test over $\mathbb{K}$ if the condition of
Theorem~\ref{thm:Scott} is satisfied for every finite-dimensional
irreducible representation of $G$ over $\mathbb{K}$.
For any field $\mathbb{K}$, we define
\[\mathcal{C}_G^{{\mathbf{m}},\mathbb{K}}=\{\tilde c\in\mathcal{C}_G^{\mathbf{m}}\mid \tilde c
\text{ passes Frobenius' test and Scott's test over $\mathbb{K}$}\}.\] Our
implementation runs the test on a field $\mathbb{F}_q$, with $q$ a fixed
prime number greater than the group order. Since
$\mathcal{C}_G^{{\mathbf{m}},\mathbb{F}_q}$ is invariant under
$(\Sigma_r)_{\mathrm{m}}$ and $\operatorname{Aut} G$, this allows us to replace
$\mathcal{C}_G^{{\mathbf{m}}}$ with $\mathcal{C}_G^{{\mathbf{m}},\mathbb{F}_q}$ in
\S \ref{say:determinesection}. \medskip
\begin{ex} In Example~\ref{example:Sigma4}, consider the
two-dimensional representation of $\Sigma_4$ obtained by pulling
back the two-dimensional irreducible representation of $\Sigma_3$
(the morphism $\Sigma_4\to\Sigma_3$ being given by the quotient by
the group generated by $C_2$). Then
\[v(C_1)=1, v(C_2)=0, v(\Sigma_4)=2=v(\Sigma_4^*).\] Thus, Scott's
test eliminates all $\tilde c$ of the form
\eqref{eqn:tildecsigma4} such that $k < 4$.
\end{ex}
\begin{say} We also employed a technical optimization which deserves
to be mentioned.
The condition $\langle g_1,\dotsc, g_{r-1}\rangle=G$ only depends
on the set $\{g_1,\dotsc, g_{r-1}\}$, call it the \emph{underlying
set} of the $r$-ple $(g_1,\dotsc, g_r)$. Different elements in
$G^{\mathbf m}$ can have the same underlying set; indeed, when
$c,c'\in \mathcal{C}_G^{\mathbf{m}}/{(\Sigma_r)_\mathbf{m}}$
contain the same conjugacy classes, possibly with different
multiplicities, elements in $\pi^{-1}(c)$ and elements in
$\pi^{-1}(c')$ can have the same underlying set.
Thus, when iterating over the sets $p^{-1}(\tilde c)$ of potential
spherical systems of generators, it makes sense to store in memory
a list of underlying sets that are known to either generate $G$ or
not, and look up each underlying set in the list before actually
performing the (costly) test to see whether a given $r$-ple
actually generates the group. Since these lists can become quite
large, one can save memory by observing that an element of
$C_1\times\dots \times C_r$ cannot have the same underlying set as
an element of $C'_1\times\dots \times C'_r$ unless
$\{C_1,\dotsc, C_{r-1}\}=\{C'_1,\dotsc, C'_{r-1}\}$. Therefore,
given a section $F$ of $\mathcal{C}_G^{{\mathbf{m}},\mathbb{F}_q}$, we work
separately on each component $F_A$ of the partition
\begin{equation*}
\label{eqn:Fpartitioned} F=\bigsqcup_{A\subset \mathcal{C}_G} F_A,
\quad F_A=\{(C_1,\dotsc, C_r)\in F\mid \{C_1,\dotsc, C_{r-1}\}=A\}.
\end{equation*}
The partition is computed with Algorithm~\ref{algo:section}.
\end{say}
\begin{algorithm}[H] \SetKwInOut{Input}{input}
\SetKwInOut{Output}{output} \SetKwComment{tcp}{/*}{*/} \Input{A
group $G$, a signature ${\mathbf{m}}$, a prime $q\geq \abs{G}$} \Output{A
section $F$ of $\mathcal{C}_G^{{\mathbf{m}},\mathbb{F}_q}$, partitioned as in
\eqref{eqn:Fpartitioned}} $F_A\leftarrow\emptyset$ for all
$A\subset\mathcal{C}_G$\;
$K\leftarrow \{(C_1,\dotsc, C_r)\in\mathcal{C}_G^r\mid
\mbox{$C_1\leq\dots\leq C_r$} \text{ and some permutation }
(C_{\sigma_1},\dotsc, C_{\sigma_r}) \text{ is in }
\mathcal{C}^{\mathbf{m}}_G\}$\;
\tcp{$K=\mathcal{C}_G^{\mathbf{m}} / (\Sigma_r)_{\mathbf{m}}$, represented as in
\eqref{pio})} $S\leftarrow$
a section for the action of $\operatorname{Aut} G$ on $K$\\
\For{$(C_1,\dotsc, C_r)$ in $S$} {\If{$(C_1,\dotsc, C_r)$ passes
Frobenius' test and Scott's test over $\mathbb{F}_q$} {add the
sequence $(C_1,\dotsc, C_r)$ to the set
$F_{\{C_1,\dotsc, C_{r-1}\}}$} } \Return $F=\bigcup F_A$
\caption{Computing the section}\label{algo:section}
\end{algorithm}
\begin{say} In theory, a section in
$\data_{G,\mm}}%X_{G,\mathbf{m}} := \mathscr{D}^r(G)\cap G^{\mathbf{m}} \subset G^{\mathbf{m}} $ can be obtained by
computing $S'$ as in Proposition~\ref{prop:section}, then
verifying for each element $(g_1,\dotsc, g_r)$ whether it is a
spherical system of generators. A bit of experimenting shows that
it is better to identify the subset of $p^{-1}(\tilde c)$
consisting of generators, before applying the action of
$\mathbf{H}_{\tilde c}$ to extract a section.
We also point out that in a spherical system of generators
$(g_1,\dotsc, g_r)$, the last element is determined by the others;
thus, given
$\tilde c=(C_1,\dotsc, C_r)\in \mathcal{C}_G^{\mathbf m}$,
spherical generators in $p^{-1}(\tilde c)$ can be determined by
iterating in $C_1\times \dots \times C_{r-1}$, and testing for
each element if the inverse of the product is in $C_r$.
The number of iterations can therefore be reduced by choosing
$\tilde c=(C_1,\dotsc, C_r)$ in such a way that the last conjugacy
class is the biggest. Since we have $C_1\leq\dots\leq C_r$
relative to the fixed ordering of $\mathcal{C}_G$, it suffices to
choose the latter in such a way that conjugacy classes with more
elements come after.
This leads to Algorithm~\ref{algo:fixedGm}.
\begin{algorithm}[th] \SetKwInOut{Input}{input}
\SetKwInOut{Output}{output} \SetKw{Return}{return}
\SetKwComment{tcp}{//}{} \Input{A group $G$ of order $d$; a
signature ${\mathbf{m}}\in\mathfrak{S}_{d,g_{\max}}$} \Output{One
representative in each Hurwitz equivalence class of spherical
systems of generators of $G$ with signature ${\mathbf{m}}$}
\caption{Classifying spherical systems of generators for fixed
$G$, ${\mathbf{m}}$ \label{algo:fixedGm}} $q\leftarrow$ the smallest
prime number greater than $d$\; $F\leftarrow$ {a section of
$\mathcal{C}_G^{{\mathbf{m}},\mathbb{F}_q}$, partitioned as in
\eqref{eqn:Fpartitioned}}\; \For{$A\subset\mathcal{C}_G$}{
generating $\leftarrow \{\}$\; notgenerating
$\leftarrow \{\}$\;{ \For{$(C_1,\dotsc, C_{r})$ in $F_A$}{
$X\leftarrow\{\}$\; \For{$(g_1,\dotsc, g_{r-1})$ in
$C_1\times\dotsm \times C_{r-1}$}{
$g_r\leftarrow (g_1\dotsm g_{r-1})^{-1}$\;
\If{$g_r\in C_r$ and $\{g_1,\dotsc, g_{r-1}\}\notin$
notgenerating}{ \eIf{$\{g_1,\dotsc, g_{r-1}\}\in$
generating or
$\langle g_1,\dotsc, g_{r-1}\rangle =G$}{ add
$\{g_1,\dotsc, g_{r-1}\}$ to generating\; add
$(g_1,\dotsc, g_r)$ to $X$\; }{ add
$\{g_1,\dotsc, g_{r-1}\}$ to notgenerating\; } } }
\If{$X$ not empty}{ $\mathbf{H}_{\tilde c}\leftarrow $ stabilizer
of $(C_1,\dotsc, C_r)$ in $\HH_\mm$\; append to output a
section of $X$ for the action of $\mathbf{H}_{\tilde c}$ } } }
}
\end{algorithm}
\end{say}
\begin{say}
\label{say:dihedral}
For dihedral groups of order $4k+2$, our problem is the object of
one of the main results in \cite{baffo-linceo}: Theorem 2 of that
paper shows that given $\tilde c$ in $\mathcal{C}_G^{{\mathbf{m}}}$ there
is at most one orbit of systems of spherical generators in
$p^{-1}(\tilde c)$. Indeed, if the order is of the form $4k+2$,
the {\it numerical type} defined in \cite[Definition
2]{baffo-linceo} corresponds exactly to the class of $c$ modulo
the action of $\operatorname{Aut} G$.
Therefore, instead of computing the whole set of systems of
generators mapping to $\tilde c$ as in
Algorithm~\ref{algo:fixedGm}, it is sufficient to iterate through
$p^{-1}(\tilde c)$ and stop as soon as a system of generators is
found.
\end{say}
\begin{say} For abelian groups $G$, conjugacy classes contain a
single element and the map $p$ of Diagram~\ref{diag:Gr} is
injective, with the action of $\outa} % {\mathbf{B_r}$ reducing to an action of
$\Sigma_r$. Therefore, having computed a section $S$ for the
action of $\operatorname{Aut} G$ on $\mathcal{C}_G^{{\mathbf{m}}}/(\Sigma_r)_{\mathbf{m}}$ exactly
as in the nonabelian case, one only needs to determine for each
element of $S$ whether its preimage in $G^{{\mathbf{m}}}$ is a spherical
system of generators. The tests of Scott and Frobenius become
redundant here, since for fixed elements $(g_1,\dotsc, g_r)$ it is
more efficient to check the conditions $\prod g_i=1$ and
$\langle g_1,\dotsc, g_r\rangle =G$ directly.
\end{say}
\begin{say}
\label{say:results}
Running our implementation \cite{gullinbursti} of the
above-illustrated algorithms, we have been able to classify
topological types of holomorphic actions on Riemann surfaces
(equivalently of orientation-preserving actions on orientable
topological surfaces) of genus $g\leq 39$ with genus 0 quotient,
with only three exceptions:
\begin{align*}
G&=(\mathbb{Z}_3\times\mathbb{Z}_3)\rtimes\mathbb{Z}_2, & {{\mathbf{m}}}&= \{ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 \}, & g&=28;\\
G&=(\mathbb{Z}_3\times\mathbb{Z}_3)\rtimes\mathbb{Z}_2, & {{\mathbf{m}}}&= \{ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3 \}, & g&=34;\\
G&=(\mathbb{Z}_3\times\mathbb{Z}_3)\rtimes\mathbb{Z}_2, & {{\mathbf{m}}}&= \{ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 \}, & g&=37.
\end{align*}
In these cases, the number of spherical systems of generators is
too large to fit into the memory of the computer at our disposal.
The number of topological types by genus is summarized in
Table~\ref{table:numberofssgbyg}. A strict inequality such as
$>3580$ for $g=28$ refers to the fact that the program classifies
$3580$ topological types with groups and signatures distinct from
the offending group and signature, which gives rise to at least
one more topological type.
The three exceptions above, for which we were not able to finish
the computation, all share the same group,
i.e. $G=(\mathbb{Z}_3\times\mathbb{Z}_3)\rtimes\mathbb{Z}_2$. This is the semidirect
product of $A:= \mathbb{Z}_3 \times \mathbb{Z}_3$ with $\mathbb{Z}_2$ defined by the
morphism $[1]_2 \mapsto \phi \in \operatorname{Aut} A$, where $\phi(a) = -a$; it is denoted by \texttt{Smallgroup(18,4)} in the {\texttt{MAGMA}}\xspace
database. This group belongs to the family of so-called
\emph{generalized dihedral groups}, i.e. groups of the form
$G= A \rtimes \mathbb{Z}_2$, with $A$ abelian and morphism
$[1]_2\mapsto \phi$ as above. When the order of $A$ is odd, these
groups are naturally challenging for our algorithm, since they
have a very big conjugacy class, the complement of the index $2$
subgroup $A$. For $A$ cyclic, i.e. for $G$ a ``standard'' dihedral
group, we were able to avoid heavy computations by applying the
results of \cite{baffo-linceo}, as explained in \S
\ref{say:dihedral}. We suspect that a more complicated analysis
could yield a similar result also for more general groups of the
form $A\rtimes \mathbb{Z}_2$. Apart from the theoretical importance, this
would allow to reach the classification of topological types up to
39 or 40 without need of more computations. More precisely the
invariant in \cite{clp} might give results analogous to those in
\cite{baffo-linceo} for a wider class of generalized dihedral
groups. We hope to be able to give some results in this direction
in a forthcoming paper.
\end{say}
\begin{table}
\caption{\label{table:numberofssgbyg} Number of topological types
of Galois covers of the line with genus $g$}
\[
\begin{array}{cc}
g & \text{\# types}\\
\hline
2 & 19 \\
3 & 46 \\
4 & 65 \\
5 & 92 \\
6 & 95 \\
7 & 160 \\
8 & 129 \\
9 & 343 \\
10 & 289 \\
11 & 342 \\
12 & 317 \\
13 & 741 \\
14 & 323
\end{array}
\hspace{1cm}
\begin{array}{cc}
g & \text{\# types}\\
\hline
15 & 768 \\
16 & 687 \\
17 & 1473 \\
18 & 711 \\
19 & 1689 \\
20 & 881 \\
21 & 2790 \\
22 & 1546 \\
23 & 2178 \\
24 & 1852 \\
25 & 5955 \\
26 & 1881 \\
27 & 4351
\end{array}
\hspace{1cm}
\begin{array}{cc}
g & \text{\# types}\\
\hline
28 & > 3580 \\
29 & 8169 \\
30 & 3992 \\
31 & 8506 \\
32 & 4336 \\
33 & 16007\\
34 & > 6983 \\
35 & 11827 \\
36 & 8753 \\
37 & > 26712 \\
38 & 8486 \\
39 & 19099 \\
\hspace{0cm}
\end{array}
\]
\end{table}
|
\section{Introduction}
The design of the airfoil is a long-term development process \cite{DURU2022105312}.
To meet the needs of different scenarios, a variety of airfoil families have been developed for different flight tasks, such as low-speed flight missions generally use front circular and posterior pointed airfoils, front and rear pointed airfoils are generally used for high-speed missions.
With the development of airfoil design, many numerical representation methods of airfoil geometry have been designed \cite{Hicks, BOEHM198717, Sobieczky1999, Stephen9269}, such as the class function/shape function transformation (CST) method using a shape function and a class function to describe an airfoil.
In the early stages of aircraft design, the appropriate airfoil is usually selected from the existing airfoil family.
With the development of machine learning technology, some scholars also try to apply machine learning method to airfoil design optimization and aerodynamic shape optimization \cite{du2021rapid, li2021deep, wang2021dual, li2022low, wu2022missile}.
As an important reference index of the airfoil, geometric parameters, Reynolds number and angle of attack are closely related to the aerodynamic performance of airfoils.
In particular, obtaining the aerodynamic parameters around the airfoils is critical to the airfoil selection and aircraft design.
In order to obtain the aerodynamic coefficients around the airfoils, such as lift, resistance, pressure, velocity, etc., wind tunnel test and computational fluid dynamics (CFD) techniques are two common methods.
For wind tunnel testing, the design of the experimental process generally relies on the prior knowledge of experts, and the wind tunnel test process is not only long time required but also expensive.
Wind tunnel test is mainly used in the later stage of aircraft wing design, for comprehensive and accurate evaluation of the aerodynamic performance of aircraft wings.
With the rapid development of computer hardware, the computing speed and performance of CFD have also been greatly improved.
However, in terms of aerodynamics related design, many iterative processes need to be involved in the calculation of flow field with the CFD solver, such as the large eddy simulation (LES), direct numerical simulation (DNS) tasks, etc., which is a memory demanding, computationally expensive and time-consuming iterative process \cite{guo2016convolutional}.
In recent years, with the development of computer technology and the improvement of hardware resources, machine learning and deep learning technology have achieved great success in computer vision \cite{LocalNet, DBLP, ibtehaz2020multiresunet, liang2021fast}, natural language processing (NLP) \cite{wang2021covid, bragg2021flex, lewis2020retrieval, tetko2020state}, speech recognition translation \cite{wang2020transformer,han2020contextnet, ravanelli2020multi, zhang2020transformer} and other scenarios.
Due to the powerful learning capability of the neural network, machine learning technology as the fourth paradigm of studying aerodynamics, which has also attracted widespread attention in recent years.
Compared with the CFD, machine learning methods only needs to spend a certain amount of time in the early stage to train the neural network, then use the trained neural network model in a few seconds or even milliseconds to get the prediction results of the airfoil flow field.
Guo et al. \cite{guo2016convolutional} used convolutional neural networks (CNN) for variable geometry flow field prediction.
The test results show that CNN can effectively predict the whole velocity field of geometry.
When calculating the velocity field, CNN are four orders of magnitude faster than CPU-based solvers and two orders of magnitude faster than GPU-accelerated solvers.
Thuerey et al. \cite{thuerey2020deep} used U-Net deep learning models instead of Reynolds-Averaged Navier-Stokes (RANS) solvers to solve for pressure and velocity distributions around different airfoils.
The U-Net architecture is similar to a special codec, under the NVIDIA GTX 1080 GPU platform, U-Net takes about $5.53ms$ to calculate the flow field of airfoil, but OpenFOAM solver takes $40.4s$ to compute the same airfoil.
Chen et al. \cite{chen2020flowgan} used a generator based on U-Net architecture to generate predictions results of flow fields.
The multi-layer perceptron (MLP) is used to merge geometry information and flow parameters.
Combining conditional generative adversarial network (cGAN) and U-Net can establish the mapping relationship between geometry shape and flow fields.
The method obtains good prediction results on the large-scale test set.
Most CNN-based flow field prediction methods use a data pre-processing method that projects flow field data into a uniformly distributed Cartesian grid \cite{peng2020time, hui2020fast, li2021novel, hu2022mesh}.
This treatment is feasible for the flow field of interest not to contain geometric features.
However, when the flow area contains geometric features, pixelation will inevitably cause the lack of geometric information, making it difficult to characterize the flow field details of the near wall area, and even generating non-physical solutions inside the geometry.
To describe the geometric characteristics of the airfoil, Sekar et al. \cite{sekar2019fast} used deep learning techniques to replace the traditional airfoil parameterization process, using seventy parameters to characterize the airfoil, and the method has good generalization even as the number of airfoil samples increases.
Because the process of airfoil parameterization is independent of flow field prediction, therefore, it can debugged separately as a module.
In this study, we use a similar neural network architecture to characterize the airfoils in the UIUC airfoil coordinates database \cite{uiuc}.
The airfoil parameterization network is accomplished by PyTorch neural network framework \cite{PyTorch}.
For the problem of flow field prediction, a large number of training samples are needed to obtain more accurate prediction results.
For example, in order to predict the flow field around 110 NACA airfoils, Sekar et al. \cite{sekar2019fast} randomly obtained 5280 flow field cases under different Reynolds numbers and angles of attack for neural network model training.
If each case contains 12000 data points, the training data reaches a staggering $60 \times 20 \times 10^6$ data points, although more training samples will improve the generalization, but the huge training data increases the training time of neural network model.
According to the experimental results of Sekar et al., under the CPU architecture, the MLP training time up to 1440 hours.
Nagawkar et al. \cite{nagawkar2022multifidelity} use random forest (RF)-based algorithm to predict high-fidelity flow field, the results show that RF can well predict the pressure and skin friction coefficients of RAE2822 airfoil.
Most of the previous research work was carried out with sufficient training samples.
Nevertheless, due to the high expenditure of CFD simulations and wind tunnel tests, it is often impossible to obtain sufficient training samples covering all flow conditions \cite{haizhou2022generative}.
Such that the trained neural network cannot guarantee prediction accuracy when the amount of training data is insufficient.
In this work, we propose a multi-head perceptron (MHP) neural network to predict the incompressible steady flow field for the airfoil with sparse samples.
For small sample set with sparse data, if perform multi-variable prediction tasks, the traditional MLP method needs to balance multiple variables during the parameter update process of neural network back-propagation, which leads to model distraction.
The MHP with multiple sub-networks is used to predict aerodynamic parameters with different distribution characteristics. So that the proposed model will pay more attention to the sparse flow field parameters during the back propagation of neural network.
Secondly, by decoupling the prediction tasks of different parameters, the interference in the prediction process of different parameters is avoided.
Experiments are conducted to evaluate the airfoil flow field prediction accuracy and training time of MLP and MHP.
%
%
The rest of the paper is organized as follows.
Section \uppercase\expandafter{\romannumeral2} mainly describes the airfoil flow field prediction problem and deep learning methods of flow field prediction.
Section \uppercase\expandafter{\romannumeral3} discusses the data pre-processing problem in airfoil parameterization and flow field prediction.
Section \uppercase\expandafter{\romannumeral4} shows and discusses the results of neural network model training and prediction.
And the conclusion is given in Section \uppercase\expandafter{\romannumeral5}.
\section{Methodology}
\subsection{Problem description}
The traditional CFD methods to calculate the flow field needs to be meshed according to the initial airfoil coordinates.
The CFD solver such as Fluent, OpenFOAM, etc. is used to calculate the flow field information around the airfoils, and the calculation results can be displayed by post-processing software such as Tecplot, etc.
The above process only involves a single airfoil, and it may take one hour or more to obtain the final flow field.
Because CNN and MHP have powerful feature extraction capability and nonlinear fitting capability, in this study, they are used to make geometric parameterization and flow field prediction of airfoil, respectively.
Compared with the traditional CFD calculation method, the pre-trained MHP model can obtain the flow field prediction results of the airfoil in a few seconds. More comparison details can be found in Fig. \ref{overview}.
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.8 \linewidth]{MHP_images/overview.pdf}
\end{center} \vspace{-2mm}
\caption{{{Computational process of CFD simulation and deep learning method}
}} \label{overview}
\end{figure*}
\subsection{Deep learning methods}
\subsubsection{Convolutional neural network}
As the basic component of feature extraction, convolutional neural network have been widely used in well-known neural network architectures such as VGG16 \cite{simonyan2014very}, ResNet \cite{he2016deep}, Faster-RCNN \cite{ren2015faster}, YOLO \cite{redmon2016you}, and SSD \cite{liu2016ssd}.
As shown in Fig. \ref{convolution_net}(a), a typical convolutional neural network consists mainly of an input layer, a convolutional layer, an activation function layer, a pooling layer, a fully connected layer and an output layer.
In the Fig. \ref{convolution_net}(a), $W$, $H$ and $D$ represent the width, height and depth of the image, respectively.
$b$ represents the bias of the layer $i$ convolution kernel, in which the dotted line indicates that the image has made corresponding calculation operations through the convolution layer, activation function and max pooling layer.
The solid line indicates that the data has made relevant calculation operations through the fully connected layer.
The calculation process of the fully connected layer here is the same as that of the MLP.
For more details about the MLP, please refer to Section 2.2.3.
Figure \ref{convolution_net}(b) shows the convolution operation of the convolution kernel.
The convolution kernel calculates the image on multiple channels through the calculation method of the sliding window to obtain the image feature map.
The common convolution kernel size are 3 $\times$ 3 and 5 $\times$ 5.
Figure \ref{convolution_net}(c) shows the influence of step size selection on the calculation results during convolution operation.
The primary function of the activation function is to provide the nonlinear modeling capability of the network.
In this work, the ReLU activation function is used in the after convolutional layer and the Tanh activation function is used in the after fully connected layer.
Max pooling layer is used to extract the principal features of a certain region, reduce the number of parameters and prevent the model from over fitting.
More details can be found in Fig. \ref{convolution_net}(d). Because the size of the image will decrease after convolution, in order to convolute the image for many times, can fill a specific value around the input matrix, which is generally zero by default.
Therefore, padding values also affects the size of the output matrix, its width and height can be calculated by the following formula, where $W$ and $H$ represent the width and height of the input image, respectively.
The $P_i$ is the padding size.
The $K_i$ is the size of the corresponding convolutional kernel, and the $S_i$ is the stride of the convolution kernel:
\begin{equation}
\left\{
\begin{aligned}
H_{out} &= \frac{H + 2 \times P_0 - K_0}{S_0} + 1, \\
W_{out} &= \frac{W + 2 \times P_1 - K_1}{S_1} + 1.
\end{aligned}
\right.
\end{equation}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9 \linewidth]{MHP_images/convolution_net.pdf}
\end{center} \vspace{-2mm}
\caption{{{Typical convolutional neural network}
}} \label{convolution_net}
\end{figure*}
\subsubsection{Airfoil parameterization network}
Because CNN has the property of weight sharing, they have a greater advantage over MLP when performing airfoil image calculations.
In Fig. \ref{convolution_parametric}, the airfoil parameterization network is composed of CNN and MLP.
They are used to encode the input airfoil image into sixteen important geometric parameters and then decode these parameters into the $y$ coordinates corresponding to the current airfoil image.
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9 \linewidth]{MHP_images/cnn_parameters.pdf}
\end{center} \vspace{-2mm}
\caption{{{Geometric parameterization network of airfoil}
}} \label{convolution_parametric}
\end{figure*}
Refer to the experimental results of Sekar et al. \cite{sekar2019fast}, and the influence of the number of convolutional layers and fully connected layers on the model training results is not considered.
The details of the network architecture are shown in Tbl. \ref{cnn_architecture}.
$Cov_{i}$ represents the layer $i$ convolutional, and $Fcn_{j}$ represents the layer $j$ fully connected.
The first product in the fourth column represents the convolutional kernel size, and the second product represents the max pooling layer filter size.
Each convolutional layer is followed by a ReLU activation function, and a Tanh activation function followed the max pooling layer.
Using the mean squared error (MSE) as the loss function of the model, we define it as:
\begin{equation}
CNN_{loss} = \frac{1}{N} \sum_{i=1}^{N} \sum_{j=1}^{70} (y_{i,j}^t - y_{i,j}^p)^2
\end{equation}
\begin{table}[tb]
\caption{Airfoil parametrization convolutional neural network architecture} \label{cnn_architecture}
\begin{center}
\begin{tabular}{lccc}
\hline \hline
Layer type & In channels & Out channels & Kernel size \\ \hline
Cov1 & 1 & 32 & 4$\times$4, 3$\times$3 \\
Cov2 & 32 & 32 & 4$\times$4, 3$\times$3 \\
Cov3 & 32 & 64 & 4 $\times$4, 3$\times$3 \\
Cov4 & 64 & 64 & 4 $\times$4, 2$\times$2 \\
Cov5 & 64 & 128 & 4 $\times$4, 2$\times$2 \\
Fcn1 & 128 & 100 & \\
3$\times$Fcn2 & 100 & 100 & \\
Fcn3 & 100 & 16 & \\
Fcn4 & 16 & 100 & \\
3$\times$Fcn5 & 100 &100 & \\
Output &100 & 70 & \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
Equation (2) represents the loss function of a batch, which is set to 64.
$y_{i,j}^{t}$, $y_{i,j}^{p}$ represents the ground-truth and predicted values of the $y$ coordinate of the airfoil, respectively.
\subsubsection{Multi-layer perceptron}
A typical multi-layer neural network is shown in Fig. \ref{mlp}(a), which consists mainly of three parts: input layer, hidden layer and output layer.
The input of the MLP is twenty physical parameters.
$G_i$ represents the geometric parameters calculated by CNN parameterization network, which are used to characterize the geometric shape of different airfoils.
The Reynolds number along with the angle of attack is used to describe the physics field information in which the current airfoil is located.
The $x$ coordinate and $y$ coordinate are used to illustrate coordinate information for different points in the airfoil flow field.
The MLP neural network is fully connected between the different layers, that is, it connected any neurons in the upper layer to all neurons in the next layer.
MLP have three basic elements: weights, biases, and activation functions.
Weights control the strength of the connections between neurons, the size of which indicates the magnitude of the likelihood.
The bias is set to correctly classify the sample and is an important parameter in the model, which is to ensure that the output values calculated from the input values cannot be activated casually.
Activation functions act as nonlinear mappings that limit the output amplitude of neurons to a certain range, generally between $(-1, 1)$ or $(0, 1)$.
The output of the neural network is pressure and velocity in the $x$ and $y$ directions, respectively.
The prediction function of the MLP can be defined as:
\begin{equation}
f_{MLP}(p_{1},...,p_{16},Re,AOA,x,y) = (u,p,v) .
\end{equation}
The left side of the function represents the prediction model of the MLP, and the right side of the function represents the prediction result of the model.
MSE is used as the loss function of MLP, which is defined as:
\begin{equation}
MLP_{loss} = \frac{1}{3 \times N} \sum_{i=1}^{N}[(u_{i}^{t} - u_{i}^{p})^2 + (p_{i}^{t} - p_{i}^{p})^2 +(v_{i}^{t} - v_{i}^{p})^2] ,
\end{equation}
where $u_{i}^{t}$ represents the ground-truth of the velocity component in the $x$ direction, and $u_{i}^{p}$ represents the predicted values of the velocity component in the $x$ direction.
$v_{i}^{t}$ represents the ground-truth of the velocity component in the $y$ direction, and $v_{i}^{p}$ represents the predicted values of the velocity component in the $y$ direction.
$p_{i}^{t}$, $p_{i}^{p}$ represents the ground-truth and prediction value of pressure, respectively.
\subsubsection{Multi-head perceptron}
Most of the previous research works \cite{sekar2019fast, yu2019flowfield, pawar2021physics} has adopted the network architecture shown in Fig. \ref{mlp}(a).
But it can be found through related experimental results (more details can be found in Section IV), MLP has obvious drawbacks in processing sparse data.
Because for sparse data with uneven distribution, even if the training data is normalized, it is also difficult for MLP to produce enough observations.
Especially in the case of multiple outputs, in order to maximize the prediction of multiple target values, MLP will update the network parameters of each previous layer in back-propagation.
Due to the existence of sparse data, the final fitting effect of MLP is unsatisfactory.
In Fig. \ref{mlp}(b), in order to avoid the interference of sparse data on other aerodynamic parameters to be predicted, MHP is proposed to predict various physical parameters in the flow field, respectively.
In particular, firstly, the basic network is used to extract the flow field characteristics, and then the multi-head network is used to obtain the prediction output of the network for three physical parameters with different distribution characteristics.
The prediction function of the MHP can be defined as:
\begin{equation}
\left\{
\begin{aligned}
f_{MHP}(p_{1},...,p_{16},Re,AOA,x,y) &= (u) , \\
f_{MHP}(p_{1},...,p_{16},Re,AOA,x,y) &= (p) , \\
f_{MHP}(p_{1},...,p_{16},Re,AOA,x,y) &= (v) .
\end{aligned}
\right.
\end{equation}
The left side of the function represents the prediction model of the MHP, and the right side of the function represents the prediction result of the model.
MSE is also used as the loss function of the MHP. The loss function of each head of MHP is defined as:
\begin{equation}
\left\{
\begin{aligned}
MHP(u)_{loss}=\frac{1}{N}\sum_{i=1}^{N}(u_{i}^{t}-u_{i}^{p}), \\
MHP(p)_{loss}=\frac{1}{N}\sum_{i=1}^{N}(p_{i}^{t}-p_{i}^{p}), \\
MHP(v)_{loss}=\frac{1}{N}\sum_{i=1}^{N}(v_{i}^{t}-v_{i}^{p}).
\end{aligned}
\right.
\end{equation}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/mlpAndmhp.pdf}
\end{center} \vspace{-2mm}
\caption{{{Neural networks for flow field prediction}
}} \label{mlp}
\end{figure*}
\section{Data preparation}
\subsection{Airfoil dataset}
The CNN network architecture is used to predict the airfoil geometric parameters in the UIUC airfoil database. The airfoil shapes for CNN training is shown in Fig. \ref{uiuc_train}.
Because the number of $x$ coordinate and $y$ coordinate in the airfoil UIUC database is not uniform.
In order to facilitate the training of neural network models, each airfoil is first fitted using the nonuniform rational B-spline (NURBS) method \cite{lepine2001optimized,lepine2000wing}, and then seventy data points are randomly selected on the fitting airfoil curve as the new $x$ coordinate and $y$ coordinate of the current airfoil.
The fitting results of RAE2822 is shown in Fig. \ref{uiuc}(a).
Secondly, fix the $x$-coordinate along the chord length, each data point of $x$-coordinate is obtained by following calculation formula:
\begin{equation}
x_{j}=\frac{1}{n} \sum_{i=1}^{n} x_i
\end{equation}
Then select the normalized $y$ coordinate as the target label value of the airfoil image at model training, the numeric range of the normalized $y$ coordinate is $(-1,1)$.
The calculation formula is following:
\begin{equation}
y_{j} = \frac{y_i - y_{avg}}{y_{max} - y_{min}}
\end{equation}
In the above formula, $y_{avg}$ represents the global average, $y_{max}$ represents the global maximum, $y_{min}$ represents the global minimum.
The normalized $y$ coordinate and fixed $x$-coordinate can be used to obtain images of different airfoils and each airfoil image also needs to be normalized.
First of all, the gray scale image with a single channel size of 216 $\times$ 216 needs to be inverted and normalized, so that the pixel value on the airfoil geometry curve is $1$, the pixel value that is not on the airfoil geometry curve is $0$, and the rest of the pixel values are between $0$ and $1$.
The results of the airfoil image pre-processing is shown in Fig. \ref{uiuc}(b).
Eighty percent of the 1582 airfoils in UIUC airfoil database are used for model training set, $10\%$ for cross validation set, and the remaining $10\%$ are used as the test set.
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.6 \linewidth]{MHP_images/uiuc_train.pdf}
\end{center} \vspace{-2mm}
\caption{{{Airfoil shapes used in airfoil parameterization network.}
}} \label{uiuc_train}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.8 \linewidth]{MHP_images/uiuc.pdf}
\end{center} \vspace{-2mm}
\caption{{{Schematic diagram of data pre-processing. $(a)$ Fitting curve of RAE2822. $(b)$ Airfoil image for CNN training. $(c)$ Full flow field mixed mesh for CFD calculation. $(d)$ Structured mesh of airfoil near wall. }
}} \label{uiuc}
\end{figure*}
\subsection{Flow field dataset}
MHP and MLP is used to predict the flow field around the airfoils.
The airfoil flow field database is generated by CFD method.
To verify the prediction effect of MHP and MLP in the case of sparse samples.
Here for the NACA0006, NACA0008, NACA0012, NACA0024 four airfoils in the range of Reynolds number 1000 to 2000 (Re: 1000, 1200, 1800, 2000), angle of attack 0${}^{\circ}$ \, to 10${}^{\circ}$ \, (AOA: 0${}^{\circ}$, 2${}^{\circ}$, 4${}^{\circ}$, 6${}^{\circ}$, 8${}^{\circ}$, 10${}^{\circ}$).
Each airfoil generates $24$ cases, a total of $96$ cases as the training set and the test set of MLP and MHP (of which $80\%$ of the data is used as the training set, $10\%$ of the data is used as the cross-validation set, and $10\%$ of the data is used as the test set).
As shown in Fig. \ref{uiuc}(c) and Fig. \ref{uiuc}(d), only the flow field of the structured mesh around the airfoil is selected as the training data and test data of the model, and the far-field unstructured mesh data does not participate in the training and prediction of the model.
Since the input of MLP and MHP is $20$ parameters ($p_{1},..,p_{16}, Re, AOA, x, y$), and the flow field data of a single airfoil is about $8358$ rows, a total of $8358 \times 20$ data of a single airfoil participates, and the final full flow field data is about $8358 \times 20 \times 96$.
Similar to the parametric network model, in order to accelerate the convergence speed of the model during training, Reynolds number, angle of attack, $x$ coordinates, $y$ coordinates, $u-velocity$, $pressure$, $v-velocity$, are also normalized here.
The input values are normalized to the range of 0 to 1.
The sixteen geometric parameters of the airfoil obtained through the parametric network are not normalized in this research, because their numerical ranges are between -1 and 1.
In the experiments of Section 4.2, the flow field data of this section are used for relevant test work.
More details about training and prediction results of MLP and MHP can be found in Section \uppercase\expandafter{\romannumeral4}.
\section{Results and discussions}
\subsection{Airfoil parameterization}
CNN is used for parameterization of airfoils.
The model parameters are optimized using the Adam optimizer.
The initial learning rate is $2.5 \times 10^{-4}$, and the epoch is set to 5000, which means that the neural network traverses all the training data 5000 times during training.
The program is implemented using PyTorch deep learning framework, and the GPU (RTX3060) is used for the model training under the Linux platform.
From Tbl. \ref{cnn_device}, it can be found that the training speed of the model can be greatly improved by using the GPU.
The training time of the airfoil parametric network under the CPU is about 53.5h, while the training time with the GPU is only 1.5h.
Figure \ref{cnn_loss} shows the loss function curve of training set and cross-validation set during the CNN training.
After 1000 epochs, the model has basically reached the convergence state.
The loss function on the training set eventually converges to $1.7137\times10^{-5}$, and the loss function on the cross-validation set eventually converges to $1.8502\times10^{-4}$.
Figure \ref{cnn_relative_co}(a) and Fig. \ref{cnn_relative_co}(b) shows the fitting results of prediction values and ground truth of NACA0024 and NACA1412, respectively.
In Fig. \ref{cnn_relative_co}, the predicted value of CNN has a good fitting effect with the ground truth, indicating that the geometric parameters obtained in the airfoil parametric network can well characterize the current airfoil geometry shape.
The prediction accuracy of the CNN is further verified by using the correlation coefficient between ground truth and prediction value, which is defined as:
\begin{equation}
R=\frac{cov(T,P)}{\sigma_{T}\sigma_{P}}=\frac{\sum_{i=1}^{n}(T_{i}-\bar{T})(P_{i}-\bar{P})}{\sqrt{\sum_{i=1}^{n}(T_{i}-\bar{T})^2}\sqrt{\sum_{i=1}^{n}(P_{i}-\bar{P})^2}}
\end{equation}
In the above formula, $cov$ represent the covariance and $\sigma$ is the standard deviation.
$T$ and $P$ represent the ground truth of $y$ coordinates and the prediction values of CNN, respectively.
$\bar{T}$ and $\bar{P}$ represent the average of $T$ and $P$, respectively.
As shown in Fig. \ref{cnn_relative_co}(c), the correlation coefficient R=0.9999 between prediction value and ground truth of the NACA0024.
In Fig. \ref{cnn_relative_co}(d), the correlation coefficient R=0.9999 between prediction value and ground truth of NACA1412.
The discrete points of the two images are distributed near the diagonal, indicating that the degree of coincidence between the prediction value and the ground truth is high.
CNN has achieved a good prediction effect on the data in the airfoil database.
\begin{table}[tb]
\caption{Environment configuration of airfoil parameterization network} \label{cnn_device}
\begin{center}
\begin{tabular}{lcc}
\hline \hline
Name & version & train time \\ \hline
platform & Linux & \\
CPU & \begin{tabular}[c]{@{}l@{}}Intel i7-11700K 3.6GHz\\ \end{tabular} & 53.5h \\
GPU & RTX3060 & 1.5h \\
cuda & 11.6 & \\
PyTorch & 1.11.0+cu113 & \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.6 \linewidth]{MHP_images/CNN_loss.pdf}
\end{center} \vspace{-2mm}
\caption{{{CNN loss function curve}
}} \label{cnn_loss}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.8 \linewidth]{MHP_images/CNN_pre.pdf}
\end{center} \vspace{-2mm}
\caption{{{CNN prediction results and correlation curves. (a) CNN prediction result of NACA0024. (b) CNN prediction result of NACA1412. (c) Correlation curve of NACA0024. (d) NACA1412 correlation curve.}
}} \label{cnn_relative_co}
\end{figure*}
\subsection{Flow field prediction}
\subsubsection{MLP training results}
\begin{table}[tb]
\caption{Test network architectures with same nodes and different layers} \label{multi_layer mlp}
\begin{center}
\begin{tabular}{lccc}
\hline \hline
Name & MLP(6-100) & MLP(10-100) & MLP(20-100) \\ \hline
Input & 1 $\times$ 20 & 1 $\times$ 20 & 1 $\times$ 20 \\
Hidden & 6 $\times$ 100 & 10 $\times$ 100 & 20 $\times$ 100 \\
Output & 1 $\times$ 3 & 1 $\times$ 3 & 1 $\times$ 3 \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
\begin{table}[tb]
\caption{Test network architectures with different nodes and same layers} \label{multi_node mlp}
\begin{center}
\begin{tabular}{lccc}
\hline \hline
Name & MLP(10-60) & MLP(10-100) & MLP(10-180) \\ \hline
Input & 1 $\times$ 20 & 1 $\times$ 20 & 1 $\times$ 20 \\
Hidden & 10 $\times$ 60 & 10 $\times$ 100 & 10 $\times$ 180 \\
Output & 1 $\times$ 3 & 1 $\times$ 3 & 1 $\times$ 3 \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[tb]
\caption{Training time and loss function accuracy of MLP with different layers} \label{multi_layer_time}
\begin{center}
\begin{tabular}{c|clc}
\hline \hline
\multirow{2}{*}{Name} & \multicolumn{2}{c}{MSE} & \multicolumn{1}{l}{Time} \\ \cline{2-4}
& training & \multicolumn{1}{c}{validation} & \multicolumn{1}{l}{} \\ \hline
MLP(6-100) & $1.4281 \times10^{-5}$ & $2.4944 \times10^{-5}$ & 3h8m \\
MLP(10-100) & \multicolumn{1}{l}{$5.9263 \times10^{-6}$} & $1.8643 \times10^{-5}$ & 4h26m \\
MLP(20-100) & \multicolumn{1}{l}{$8.1577 \times10^{-6}$} & $1.4709 \times10^{-4}$ & 7h59m \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
\begin{table}[tb]
\caption{Training time and loss function accuracy of MLP with different nodes} \label{multi_node_time}
\begin{center}
\begin{tabular}{c|clc}
\hline \hline
\multirow{2}{*}{Name} & \multicolumn{2}{c}{MSE} & \multicolumn{1}{l}{Time} \\ \cline{2-4}
& training & \multicolumn{1}{c}{validation} & \multicolumn{1}{l}{} \\ \hline
MLP(10-60) & $3.3029 \times10^{-5}$ & $5.5689 \times10^{-5}$ & 4h35m \\
MLP(10-100) & \multicolumn{1}{l}{$5.9263 \times10^{-6}$} & $1.8643 \times10^{-5}$ & 4h26m \\
MLP(10-180) & \multicolumn{1}{l}{$1.2830 \times10^{-6}$} & $1.4593 \times10^{-5}$ & 4h29m \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
Firstly, the traditional MLP method is used to train the flow field of different geometry airfoils under different working conditions.
Figure \ref{hist} shows the histogram of all training data, in which the velocity value distribution is relatively uniform and the pressure value distribution is more sparse.
The weights of the MLP are trained using the Adam optimizer, the initial learning rate is set to 5 $\times$ $10^{-5}$.
To accelerate the convergence of the model, the learning rate is optimized using the learning rate scheduler.
The parameter gamma is set to 0.1, that is, the learning rate is multiplied by 0.1 for every 100 epochs passed.
The effects of different layers and different nodes on the model training results are tested.
The details of the test network are shown in Tbl. \ref{multi_layer mlp} and Tbl. \ref{multi_node mlp}.
The training set loss function curve and the validation set loss function curve of different models are shown in the Fig. \ref{multilayer mlp} and Fig. \ref{multilnode mlp}.
In Fig. \ref{multilayer mlp}, the effect of different layers on the loss function accuracy of the MLP is tested.
From Fig. \ref{multilayer mlp}(a) and Fig. \ref{multilayer mlp}(b), it can be found that the number of MLP neural network layers is increased from 6 layers to 10 layers.
The MSE curve can converge quickly, but the 10-layer MLP can get a smaller MSE and better prediction accuracy.
However, if the number of layers of MLP neural network continues to increase, the convergence rate of the MSE loss function curve is slower.
The MSE of 20-layer MLP is larger than that of 6-layer and 10-layer neural network.
And as can be seen from Tbl. \ref{multi_layer_time}, as the number of layers increases, the training time of the neural network will increase exponentially, but the loss function does not decrease.
Therefore, the 10-layer MLP are selected as the training architecture for subsequent models.
On the training set, the MSE of 10-layer MLP finally converges to $5.9263 \times 10^{-6}$.
On the cross-validation set, the MSE of 10-layer MLP finally converges to $1.8643 \times 10^{-5}$.
As shown in Fig. \ref{multilnode mlp}, the effect of different node numbers on the loss function accuracy of the MLP is tested.
In Fig. \ref{multilnode mlp}(a) and Fig. \ref{multilnode mlp}(b), the number of nodes in each layer of MLP is increased from 60 to 180.
In the initial training stage of MLP, the MSE decreases rapidly and becomes smaller.
After 100 epochs, MSE basically reached a stable state.
And as can be seen from Tbl. \ref{multi_node_time}, as the number of network nodes increases, the training time does not change much, but the loss function becomes smaller.
Therefore, in this research, MLP with 10 layers and 180 nodes in each layer is selected as the final flow field prediction neural network architecture.
In Fig. \ref{multilnode mlp}(a) and Fig. \ref{multilnode mlp}(b), the loss function of the MLP on the training set eventually converges to $1.2830 \times 10^{-6}$, while the loss function on the cross-validation set eventually converges to $1.4593 \times 10^{-5}$.
\begin{table}[tb]
\caption{Training time and loss function of MLP and MHP} \label{MLP and MHP}
\begin{center}
\begin{tabular}{cc|ccc}
\hline \hline
\multicolumn{2}{c|}{\multirow{2}{*}{Name}} & \multicolumn{2}{c}{MSE} & Time \\ \cline{3-5}
\multicolumn{2}{c|}{} & training & validation & \\ \hline
\multicolumn{1}{c|}{\multirow{3}{*}{MLP}} & u & $1.2830 \times 10^{-6}$ & $1.4593 \times 10^{-5}$ & 4h29m \\
\multicolumn{1}{c|}{} & p & $1.2830 \times 10^{-6}$ & $1.4593 \times 10^{-5}$ & 4h29m \\
\multicolumn{1}{c|}{} & v & $1.2830 \times 10^{-6}$ & $1.4593 \times 10^{-5}$ & 4h29m \\ \cline{1-2}
\multicolumn{1}{l|}{\multirow{3}{*}{MHP}} & u & $2.4273 \times 10^{-6}$ & $2.6691 \times 10^{-5}$ & 3h42m \\
\multicolumn{1}{l|}{} & p & $4.4581 \times 10^{-8}$ & $1.4126 \times 10^{-7}$ & 4h9m \\
\multicolumn{1}{l|}{} & v & $5.3882 \times 10^{-7}$ & $3.8955 \times 10^{-6}$ & 4h10m \\
\hline \hline
\end{tabular}
\end{center}
\vspace{-1.5em}
\end{table}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/hist.pdf}
\end{center} \vspace{-2mm}
\caption{Histogram of training set. (a) Distribution histogram of U-velocity. (b) Distribution histogram of pressure. (c) Distribution histogram of V-velocity.} \label{hist}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MLP_different_layer_loss.pdf}
\end{center} \vspace{-2mm}
\caption{{{MSE convergence curves of MLP with different layers. (a) MSE convergence curve of MLP with different layers on training set. (b) MSE convergence curve of MLP with different layers on cross-validation set}
}} \label{multilayer mlp}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MLP_different_node_loss.pdf}
\end{center} \vspace{-2mm}
\caption{{{MSE convergence curves of MLP with different nodes. (a) MSE convergence curve of MLP with different nodes on training set. (b) MSE convergence curve of MLP with different nodes on cross-validation set}
}} \label{multilnode mlp}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MHP_loss.pdf}
\end{center} \vspace{-2mm}
\caption{{{MHP neural network loss function curve. (a) Loss function curve of MHP on training set. (b) Loss function curve of MHP on cross-validation set. }
}} \label{MHP-net}
\end{figure*}
\subsubsection{MHP training results}
Secondly, MHP is used to train the flow field data in Section \uppercase\expandafter{\romannumeral3}.
The hyperparameters of the MHP during training, such as learning rate, batch size, number of neural network layers and number of neurons are the same as the MLP.
And it performed 500 iterations during the model training process.
Figure \ref{MHP-net} shows the loss function curve of MHP model during training.
As shown in Fig. \ref{MHP-net}(a), on the training set of the flow field, the MHP-U, MHP-P and MHP-V loss function curves all decreased rapidly at the beginning.
The MHP loss function curve decreases the fastest and the value after MSE finally converges is the smallest, about $4.4581\times10^{-8}$.
The comparison results of MLP and MHP training time and MSE are shown in Tbl. \ref{MLP and MHP}.
Compared with the MLP, the MSE of MHP-U did not change significantly, but the training time of MHP-U was reduced by 47 minutes, while the training time of MHP-P was reduced by 20 minutes and the MSE was reduced by 2 orders of magnitude compared with the MLP.
And the training time of the MHP-V was reduced by 19 minutes and the MSE was reduced by 1 order of magnitude compared with the MLP.
The comparison results show that MHP has more powerful prediction capability than MLP in the face of sparse flow field data.
The loss function curve of the MHP on the cross-validation dataset is shown in Fig. \ref{MHP-net}(b).
It is found that the MSE of the MHP-P achieves a good result at the beginning of the training.
And the reason why MHP-U and MHP-V curves oscillate before 100 epochs is that the initial learning rate is too large.
And then gradually stabilizes after 100 epochs because the learning rate scheduler is used in this work to automatically reduce the learning rate value during the model training.
Finally, the MHP loss function basically converges after 150 epochs.
The MSE of MHP-P on the cross-validation set eventually converges to $1.4126 \times 10^{-7}$.
On the cross-validation set, the loss function of MHP-U finally converges to $2.6691 \times 10^{-5}$.
The loss function of MHP-V on the cross-validation set eventually converges to $3.8955 \times 10^{-6}$.
\subsubsection{Flow field prediction results}
Test the flow field prediction effect of MLP and MHP by selecting test data that the model has never seen before.
In this research, NACA0012 airfoil at Re=1000 and AOA=6${}^{\circ}$ \, is randomly selected to test the prediction capability of MLP and MHP.
In Fig. \ref{MLP_MHP_U}, it can be found that the prediction results obtained by both MLP and MHP-U are consistent with the CFD calculation results.
In addition, the absolute error plot between CFD and MLP, MHP-U is also given.
Figure. \ref{MLP_MHP_U}(c) shows that the absolute error range between MLP and CFD is -0.02 to 0.002.
In Fig. \ref{MLP_MHP_U}(f), the absolute error range between CFD and MHP-U is -0.014 to 0.018.
Figure \ref{MLP_MHP_U_RES} uses the form of the histogram to show the error data distribution of Fig. \ref{MLP_MHP_U}(c) and Fig. \ref{MLP_MHP_U}(f).
In Fig. \ref{MLP_MHP_U_RES}(a), about 5000 error data are distributed around 0.
Figure \ref{MLP_MHP_U_RES}(b) there are about 6000 error data distributed in the numerical range of 0.
It can be seen from the test results that both MLP and MHP-U have a good prediction effect of u-velocity.
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MLP_MHP_U.pdf}
\end{center} \vspace{-2mm}
\caption{{{Comparison of MHP-U, MLP and CFD calculation results. (a), (d) CFD calculation results for u-velocity. (b) U-velocity prediction results of MLP. (e) MHP-U prediction results for u-velocity. (c) Absolute error for u-velocity between CFD and MLP. (f) Absolute error for u-velocity between CFD and MHP-U.}
}} \label{MLP_MHP_U}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9 \linewidth]{MHP_images/naca0012_U_error_hist.pdf}
\end{center} \vspace{-2mm}
\caption{{{Absolute error histogram between CFD and neural network. (a) Absolute error histogram of Fig. \ref{MLP_MHP_U}(c). (b) Absolute error histogram of Fig. \ref{MLP_MHP_U}(f)}
}} \label{MLP_MHP_U_RES}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MLP_MHP_P.pdf}
\end{center} \vspace{-2mm}
\caption{{{Comparison of MHP-P, MLP and CFD calculation results. (a), (d) CFD calculation results for pressure. (b) Pressure prediction results of MLP. (e) MHP-P prediction results for pressure. (c) Absolute error for pressure between CFD and MLP. (e) Absolute error for pressure between CFD and MHP-P.}
}} \label{MLP_MHP_P}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9 \linewidth]{MHP_images/naca0012_P_error_hist.pdf}
\end{center} \vspace{-2mm}
\caption{{{Absolute error histogram between CFD and neural networks. (a) Absolute error histogram of Fig. \ref{MLP_MHP_P}(c). (b) Absolute error histogram of Fig. \ref{MLP_MHP_P}(f).}
}} \label{MLP_MHP_P_RES}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/MLP_MHP_V.pdf}
\end{center} \vspace{-2mm}
\caption{{{Comparison of MHP-V, MLP and CFD calculation results. (a), (d) CFD calculation results for v-velocity. (b) MLP prediction results for v-velocity. (e) V-velocity prediction results of MHP-V. (c) Absolute error for v-velocity between CFD and MLP. (f) Absolute error for v-velocity between CFD and MHP-V.}
}} \label{MLP_MHP_V}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=0.9 \linewidth]{MHP_images/naca0012_V_error_hist.pdf}
\end{center} \vspace{-2mm}
\caption{{{Absolute error histogram between CFD and neural networks. (a) Absolute error histogram of Fig. \ref{MLP_MHP_V}(c). (b)Absolute error histogram of Fig. \ref{MLP_MHP_V}(f).}
}} \label{MLP_MHP_V_RES}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/naca0012_near_wall.pdf}
\end{center} \vspace{-2mm}
\caption{{{Comparison of variables distribution between CFD and MLP, MHP. (a) Distribution of CFD and MLP about variable U. (b) Distribution of CFD and MLP about variable P. (c) Distribution of CFD and MLP about variable V. (d) Distribution of CFD and MHP-U about variable U. (e) Distribution of CFD and MHP-P about variable P. (f) Distribution of CFD and MHP-V about variable V.}
}} \label{near-wall}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/naca0012_equal_line.pdf}
\end{center} \vspace{-2mm}
\caption{{{Comparison of local contour of CFD and MLP, MHP. The red dotted line is CFD, the blue dotted line is neural network. (a) U-velocity contour of MLP and CFD. (b) Pressure contour of CFD and MLP. (c) V-velocity contour of CFD and MLP. (d) U-velocity contour of CFD and MHP-U. (e) Pressure contour of CFD and MHP-P. (f) V-velocity contour of CFD and MHP-V.}
}} \label{MLP_MHP equal value}
\end{figure*}
\begin{figure*}[!h]
\begin{center}
\includegraphics[width=1 \linewidth]{MHP_images/naca0012_co.pdf}
\end{center} \vspace{-2mm}
\caption{{{Correlation curve of flow field prediction variables. (a) Correlation curve of CFD and MLP. (b) CFD and MLP correlation curve. (c) Correlation curve of CFD and MLP. (d) CFD and MHP-U correlation curve. (e) Correlation curve of CFD and MHP-P. (f) CFD and MHP-V correlation curve.}
}} \label{relative_coefficitive}
\end{figure*}
In Fig. \ref{MLP_MHP_P}, it can be found that the pressure prediction results of MHP-P is consistent with the CFD calculation results.
But the error between MLP prediction results and CFD calculation results is relatively large.
And it can also be seen from the error distribution plot of Fig. \ref{MLP_MHP_P}(c) and Fig. \ref{MLP_MHP_P}(f).
The error range between prediction values and ground truth of MLP is between -0.002 and 0.0016.
And the error range between prediction values and the ground truth of MHP-P is between -0.0008 and 0.0016.
Figure \ref{MLP_MHP_P_RES} uses the form of histogram to show the error data distribution of Fig. \ref{MLP_MHP_P}(c) and Fig. \ref{MLP_MHP_P}(f).
And it can be clearly seen from the histogram that prediction error of MHP-P is more concentrated in the numerical range of 0, while the prediction error of MLP is more dispersed.
According to the pressure prediction results, for sparse data, MHP-P achieves better prediction effect than MLP.
In Fig. \ref{MLP_MHP_V}, both MLP and MHP-V achieve a good prediction results for v-velocity.
As can be seen from Fig. \ref{MLP_MHP_V}(c), the error between prediction results of MLP and CFD calculation results is between -0.012 and 0.008.
In Fig. \ref{MLP_MHP_V}(f), the error between MHP-V prediction results and CFD calculation results is between -0.009 and 0.006.
Figure \ref{MLP_MHP_V_RES} uses the form of histogram to show the error data distribution of Fig. \ref{MLP_MHP_V}(c) and Fig. \ref{MLP_MHP_V}(f).
And it can be found from Fig. \ref{MLP_MHP_V_RES}(a) that for MLP prediction error data, there are about 5000 error data distributed in the numerical range of 0.
And for the MHP-V prediction error data, there are about 7000 error data distributed in the numerical range of 0.
This shows that under the same conditions, the prediction effect of MHP-V is better than MLP.
From Fig. \ref{near-wall}, in the near wall region of the airfoil, the prediction effect of MLP and MHP on the u-velocity and v-velocity is relatively good.
This is because the velocity distribution of the airfoil is relatively continuous, and the neural network can quickly learn the relevant distribution law of the data during training process.
However, it can be found from Fig. \ref{near-wall}(b) that due to the uneven distribution of pressure data, the curve obtained by MLP when performing the task of pressure prediction is not smooth, and the fitting effect with CFD is poor.
On the contrary, the pressure curve predicted by MHP is smoother.
The experimental results in Fig. \ref{near-wall}(e) shows that MHP-P still achieve a good prediction results in the face of sparse flow field data.
It shows that the network architecture of MHP has better generalization than MLP.
In Fig. \ref{MLP_MHP equal value}, for contours of u-velocity and v-velocity, both MLP and MHP have achieved a good prediction effect.
But due to the existence of sparse data, the prediction effect of MLP is slightly worse than that of MHP.
In particular, it can be found from Fig. \ref{MLP_MHP equal value}(b) and Fig. \ref{MLP_MHP equal value}(e) that this comparison is more obvious.
The pressure contours predicted by MLP is not smooth, resulting in poor fitting effect with CFD contours.
On the contrary, for sparse pressure data, MHP network architecture is decoupled, so the influence of sparse data on other parameters is avoided.
Moreover, MHP can focus more attention on sparse data, which makes the prediction effect of the model better and the generalization performance stronger in the face of sparse data.
Figure \ref{MLP_MHP equal value}(e) shows that the contours predicted by MHP-P is consistent with the CFD calculation results.
The accuracy of the model is further verified by using the correlation coefficient between the ground truth and prediction values of the flow field.
The correlation coefficient is calculated in the similar way to Equation (9), except that the parameters $T$ and $P$ here represents the ground truth and prediction values of the flow field, respectively.
In Fig. \ref{relative_coefficitive}(a), for u-velocity, the correlation coefficient between the prediction values and ground truth of MLP is $R=0.99984$.
From Fig. \ref{relative_coefficitive}(b), for $C_p$, the correlation coefficient between the prediction values and ground truth of MLP is $R=0.99823$.
Figure \ref{relative_coefficitive}(c) shows that for v-velocity, the correlation coefficient between the prediction values and the ground truth of MLP is $R=0.99973$.
From Fig. \ref{relative_coefficitive}(d), Fig. \ref{relative_coefficitive}(e) and Fig. \ref{relative_coefficitive}(f), it can be found that the correlation coefficients between the prediction values and the ground truth of MHP for u-velocity, pressure and v-velocity are $R=0.99992$, $R=0.99967$ and $R=0.99987$, respectively.
Secondly, it can be seen from Fig. \ref{relative_coefficitive} that most of the discrete points are distributed around the diagonal of image, indicating that the difference between ground truth and prediction values is small.
However, it can be found from Fig. \ref{relative_coefficitive}(b) that the correlation data points between the prediction values and the ground truth of MLP for pressure are scattered near the image diagonal, indicating that the prediction results of MLP for pressure is not accurate.
And for the same data, as shown in Fig. \ref{relative_coefficitive}(e), the correlation data between the prediction values and the ground truth of MHP-P are more closely distributed in the diagonal area of the image, indicating that the neural network architecture of MHP can obtained better prediction results even for sparse flow field data.
\section{Conclusions}
CNN is used to establish the mapping relationship between airfoil geometry shape and airfoil coordinates.
Firstly, the input airfoil image is encoded into sixteen geometric parameters by CNN.
And then uses a decoder-like network architecture to decode these geometric parameters into the $y$ coordinates of corresponding airfoil image.
The prediction capability of CNN network is tested on the test set of airfoil.
In the case of variable geometry, the correlation coefficient R=0.9999 between prediction values and ground truth of airfoils.
Compared with the traditional airfoil parameterization methods, the deep learning method is more flexible.
Moreover, based on the pre-training model, the training set can be further expanded to improve the generalization of CNN.
For sparse flow field data, multi-head perceptron neural network architecture is proposed to improve the accuracy and generalization of flow field prediction.
Firstly, the influence of the number of neural network layers and the number of neurons on the prediction accuracy of flow field is verified.
According to the experimental results, the network architecture of 10 layers and 180 neurons in each layer is selected as the basic network for MLP and MHP.
After experiment comparison, it can be found that for sparse flow field data, MHP can achieve better prediction results than MLP.
This is because in the multi-variable prediction task, the sparse flow field data will cause the neural network to pay insufficient attention to it in the training process due to the lack of data.
Moreover, in the multi-coupling architecture of MLP, multiple variables will interfere with each other in the process of neural network parameter updating.
Therefore, MHP decouples the multi-variable prediction task of flow field to avoid the interference of sparse data to the prediction results of other normal flow field data.
The loss function of MHP in training set and validation set can be reduced by two orders of magnitude compared with MLP.
And in order to test the prediction effect of MHP and MLP for different airfoil flow fields, Appendix A presents the flow field prediction results of MLP and MHP for NACA0024 at Re=1000 and AOA=8${}^{\circ}$.
\section *{Declaration of competing interest}
The authors declare that they are have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section *{}
|
\section{Introduction}
\vspace{-3 pt}
Nowadays, the vast majority of applications for person identity verification rely on Face Recognition Systems (FRSs), which match a human face to an entry from a database of faces. Modern FRSs have proven to be highly accurate when genuine faces are presented to the system \cite{grm2018strengths}. They are however prone to various attacks, whose aim is to gain illegal access by false authentication \cite{Huber_SYN-MAD_2022}.
Lately, face morphs have become a growing concern for the reliability of face verification systems. A face mor\-ph is a composite image generated from two (or more) facial images of distinct subjects.
Recent advances in generative deep models have enabled an almost effortless generation of realistic and high-quality morphed facial images. Such images can be utilized to verify all identities that have been used in the morph-generation process. A successful detection of \textit{face morphing attacks} is therefore critical for the prevention of illegal activities \cite{Naser_PW_MAD_2021}.
\begin{figure}[!t]
\begin{center}
\centering
\includegraphics[width=0.9\linewidth]{figures/teaser.pdf}
\end{center}
\vspace{-20pt}
\caption{To avoid privacy related concerns in the development of morphing attack detectors (MAD), we explore the idea of using synthetic training faces of non--existing people. Trained MADs are then tested on real--world datasets.\label{fig:teaser}}
\end{figure}
Various \textit{Morphing Attack Detection} (MAD) algori\-thms have been proposed over the years to automatically distinguish real from morphed faces.
However, regardless of the detection technique, the training of these models requires a massive database of genuine face images. Training protocols therefore raise various privacy-related concerns and limit the amount of publicly available training data that can be used to learn MAD models. In this paper, we address the privacy issues related to MADs by exploring the idea of using synthetic training data, as illustrated in Figure~\ref{fig:teaser}. For this purpose, we use the SMDD~\cite{Damer_SMDD_2022} dataset, where StyleGAN2 \cite{karras2020analyzing}, a state-of-the-art Generative Adversarial Network (GAN), was utilized to generate \textit{bona fide} face images of non-existing people. These images were then used for the generation of the face morphs.
With this dataset, we train two powerful binary classifiers, Xception and HRNet, and evaluate their detection performance on three real-world datasets -- FRLL-Morphs, FERET-Morphs and FRGC-Morphs. The results of the evaluation show that well-performing MAD models can be learned from synthetic data alone, and that the model generalize well over three diverse real-world morph datasets.
\vspace{-3 pt}
\section{Related work}
\vspace{-3 pt}
Existing morphing-attack-detection models can in general be categorized as single--image (S-MAD) or differential (D-MAD) MADs, depending on whether the face morph is examined independently or is compared to a reference sample. While D-MADs can be very accurate in closed--set problems, S-MADs aim to detect attacks without any prior knowledge about human identities. In this section, we only review S-MADs, since they are more closely related to our work.
Regardless of the face morphing technique used, the generated morphs usually contain image irregularities su\-ch as artifacts, noise, pixel discontinuity, distortions, spectrum discrepancies, inconsistent illumination, etc. In the past, shallow algorithms, that implement extraction of photo-response non-uniformity (PRNU) noise~\cite{Scherhag_PRNU_2019} or reflection analysis~\cite{Seibold_reflection_2018} have been successfully employed for the detection of morphing attacks. Some other hand--crafted MAD methods have also used texture--based descriptors, such as LBP~\cite{Ojala_1996_LBP}, LPQ~\cite{Ojansivu_LPQ_2008} or SUR
~\cite{Makrushin_SURF_2019}. Although these methods achieved promising results, they were shown to have limited generalization capabilities. Moreover, as the face morphing techniques improved over time, the performance of shallow methods became less competitive, as they struggled to detect modern, deep-learning generated or heavily post--processed face morphs.
More recent MAD models take advantage of the development of data--driven, deep-learning algorithms. Rag\-havendra \textit{et al.}~\cite{Raghavendra_transfer_2017} were amongst the first to propose transfer learning. In their work, attacks are detected with a simple, fully--connected binary classifier, fed with fused VGG19 and AlexNet features, pretrained on ImageNet. Wandzik \textit{et al.}~\cite{Wandzik_FRS_2018}, on the other hand, achieve highly accurate results with features from general--purpose face recognition systems (FRSs) combined with an SVM. Ramachandra \textit{et al.}~\cite{Ramachandra_Inception_2020} utilize Inception in a similar manner, while Damer \textit{et al.}~\cite{Naser_PW_MAD_2021} argue, that pixel--wise supervision, where each pixel is classified as a bona fide or a morphing attack, is superior, when used in addition to the binary, image--level objective.
Recently, MixFaceNet~\cite{Boutros_MixFaceNet_2021} by Boutros \textit{et al.} achieved state--of--the--art results in different face--related detection tasks, including face morphing detection~\cite{Damer_SMDD_2022}. This model represents a highly efficient architecture that captures different levels of face attack cues by using differently sized convolutional kernels.
\vspace{-3 pt}
\section{Methods}
\vspace{-3 pt}
We consider two different classification models, Xception and HRNet, to detect face morphing attacks in this study and train them using synthetic data only. The two models represent the entries from the University of Ljubljana to the recent \textit{Face Morphing Attack Detection Competition based on Privacy-aware Synthetic
Training Data} (SYN-MAD), held in conjunction with the 2022 International Joint Conference on Biometrics (IJCB 2022)~\cite{Huber_SYN-MAD_2022}, which achieved the best and third best overall performance among all submitted entries.
\textbf{Xception}~\cite{Xception} is a convolutional neural network (CNN) that updates and simplifies the architecture of the InceptionV3 model \cite{szegedy_rethinking_2015} by replacing the Inception modules with depth--wise separable convolutions.
We use Xception as a feature extractor, while the binary classification is performed by a fully connected two--layer network. The output layer consist of $2$ neurons, followed by a softmax activation function.
Similar to previous research, we use cross--entropy as the learning objective.
\textbf{HRNet}~\cite{Wang_HRNet_2019} is again a CNN that unlike other networks maintains high--resolution representations of the input sample through the whole feedforward process.
Su\-ch an architecture contributes to more descriptive image representations, which was proven to improve the results of different classification tasks.
In our experiments, we replace the classification head of HRNet with a two--layer classification module, to perform binary detection of bona fide images and morphing attacks. The output layer consists of only one neuron, followed by a sigmoid activation function. In the training phase, the parameters are optimized using the binary cross--entropy loss.
\begin{figure}[t]
\begin{center}
\centering
\includegraphics[width=\linewidth]{figures/dataset_examples_2.pdf}
\end{center}
\vspace{-20pt}
\caption{Examples of bona fide and morphing attack images from the SMDD~\cite{Damer_SMDD_2022} training dataset and the testing datasets FERET-Morphs, FRLL-Morphs and FRGC-Morphs~\cite{Sarkar2020_morphed_data}. \label{fig:datasets}}
\end{figure}
\begin{table}[t]
\caption{Number of bona fide images (BF), number of morphing attacks (MA) generated by morphing methods OpenCV (OCV), FaceMorpher (FM), StyleGAN (SG), AMSL, Webmorpf (WM) and image size of samples in each dataset.\vspace{-4mm}}
\label{tab1}
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ | l | c | c | c | c | c | c | c |}
\hline
\textbf{Dataset} & \textbf{Image size} & \textbf{BF} & \textbf{OCV} & \textbf{FM} & \textbf{SG} & \textbf{AMSL} & \textbf{WM} \\
\hline
FRLL-M & $1350\times1350$ & $204$ & $1221$ & 1222 & 1222 & 2175 & 1221 \\
FERET-M & $512\times768$ & $1.413$ & 529 & 529 & 529 & / & / \\%$1.587$\\
FRGC-M & $227\times277$ & $3.167$ & 964 & 964 & 964 & / & / \\%$2.892$\\
\hline
\end{tabular}}\vspace{-4mm}
\end{center}
\end{table}
\vspace{-3 pt}
\section{Experiments}
\vspace{-3 pt}
\subsection{Datasets}
We use one synthetic and three publicly available real--world datasets in this work. The training is done exclusively with the synthetic data, while the evaluation is performed on three commonly used face morphing datasets.
\textbf{Training data.}
For training, we use the SMDD data\-set \cite{Damer_SMDD_2022}, provided by the organizers of the SYN-MAD competition~\cite{Huber_SYN-MAD_2022}. The dataset consists of $25.000$ bona fide and $15.000$ morphed images of size $256\times256$ pixels.
Bona fide instances represent carefully selected images from a set of randomly generated StyleGAN2~\cite{karras2020analyzing,Karras_stylegan2_2020} faces. A separate, non -- overlapping StyleGAN2 image set was used for the generation of face morphing attacks. Face morphs were created using the landmark--based morphing technique from OpenCV\footnote{https://learnopencv.com/face-morph-using-opencv-cpp-python/}. A few selected samples from the SMDD dataset are presented in Figure~\ref{fig:datasets}.
\begin{figure*}[!t]
\begin{center}
\centering
\includegraphics[width=0.345\linewidth]{figures/FRLL_grid.pdf} \hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FERET_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FRGC_grid.pdf}
\includegraphics[width=0.345\linewidth,trim = 0mm 100mm 0 0, clip]{figures/legend_models.png}
\end{center}
\vspace{-25 pt}
\caption{ROC curves generated on FRLL-Morphs, FERET-Morphs and FRGC-Morphs for the tested models. Note that HRNet achieves very competitive results on FERET-Morphs and FRGC-Morphs, but performs the weakest on FRLL-Morphs. Xception on the other hand, consistently outperforms the baseline method MixFaceNet, on all considered datasets. \label{fig:ROC_by_datasets}}
\end{figure*}
\textbf{Testing data.}
The trained MAD models are tested on three diverse morphing datasets proposed by Sarkar \textit{et al.} in~\cite{Sarkar2020_morphed_data}, i.e. FRLL-
Morphs, FERET-Morphs and FRGC-Morphs. All face morphs were created by combining bona fide images from their respective face datasets, i.e. FRLL~\cite{debruine_jones_2017, Neubert_FRLL}, FERET~\cite{PHILLIPS_FERET} and FRGC~\cite{Phillips_FRGC}. To generate landmark--based morphs, the authors used OpenCV and FaceMorpher\footnote{https://github.com/alyssaq/face\_morpher}, while deep-learning--based morphs are generated with StyleGAN2. In addition to these methods, AMSL\cite{Neubert_FRLL} and Webmorph\footnote{https://webmorph.org/} are also used, but only for the images from the FRLL dataset. Information about image sizes and the number of samples per morphing method is given in Table~\ref{tab1}. Selected samples from all three datasets are presented in Figure~\ref{fig:datasets}.
\vspace{-5 pt}
\subsection{Experimental setup}
In our experiments, we first preprocess images from all datasets by cropping out the facial areas. Bounding boxes of the SMDD are provided by the authors of the dataset. For the other three databases, we use RetinaFace~\cite{RetinaFace}, to localize the facial region-of-interest. Prior to their use, cropped images are resized to $299\times299$ pixels for Xception and $256\times256$ pixels for HRNet. Additionally, the training data is augmented with horizontal flips to increase the amount of data available and avoid overfitting.
The CNNs were optimized using the Adam optimizer, with a learning rate of $0.0001$. The models were trained from scratch for $30$ full epochs, with a batch size of $16$. After each training epoch, the classification accuracy of the networks was calculated on a small holdout set of each test dataset. The best performing parameters on each of the three datasets were saved as the final model for that particular dataset.
The code was implemented in Python 3.8 with PyTorch 1.9 and CUDA 11.6. Experiments were run on a single GeForce GTX 1080 Ti. The computational complexity of Xception is $11$ GFLOPs, while HRNet has $34$ GFLOPs
\vspace{-3 pt}
\section{Results}
\vspace{-3 pt}
In Figure~\ref{fig:ROC_by_datasets} and Table~\ref{tab:quantitative_results} we present the results obtained with our two models, Xception and HRNet, and the baseline MixFaceNet-MAD from~\cite{Huber_SYN-MAD_2022}. The weights of MixFa\-ceNet-MAD, optimized on the SMDD dataset, were provided by the authors of the model. In our experiments, the best overall results were achieved by Xception, whose Equal Error Rates (EER) are $3.26\%$, $8.25\%$ and $9.75\%$ on FRRL--Morphs, FERET-Morphs and FRGC--Morphs, respectively (Table~\ref{tab:quantitative_results}). The runner-up, HRNet, achieves a similar performance on FERET-Morphs and FRGC-Mo\-rphs. However, among the tested models, HRNet is least successful on FRLL-Morphs, where it achieves an EER of $13.73\%$. On this dataset, MixFaceNet yields a slightly better performance than HRNet with an EER of $12.18\%$, but is outperformed by both, Xception and HRNet, on the other two databases, i.e. FERET-Morphs and FRGC-Morphs. The complete ROC curves of the experiments are visualized in Figure~\ref{fig:ROC_by_datasets} and show a similar picture as the discussed numerical results.
\begin{table}[t]
\caption{Detection results for MAD methods MixFaceNet (MFN), Xception (XN) and HRNet (HRN) on the real--world datasets FRLL-Morphs (FRLL-M), FERET-Morphs (FERET-M) and FRGC-Morphs (FRGC-M). Best scores per dataset are marked blue, while runner--up results are marked orange. All three models were trained on the synthetic SMDD dataset \cite{Damer_SMDD_2022}.
}
\vspace{-20pt}
\label{tab:quantitative_results}
\smallskip
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ | c | c | c | c | c | c | c | c | }
\hline
\multirow{2}{*}{\textbf{MAD}} & \multirow{2}{*}{\textbf{Test data}} & \multirow{2}{*}{\textbf{AUC(\%)}} & \multirow{2}{*}{\textbf{EER(\%)}} & \multicolumn{4}{c |}{\textbf{BPCER (\%) @ APCER =}} \\ \cline{5-8}
& & & & $0.10\%$ & $1.00\%$ & $10.00\%$ & $20.00\%$ \\ \hline
\hline
\multirow{3}{*}{MFN~\cite{Damer_SMDD_2022}} & FRLL-M & \textcolor{orange}{$\mathbf{95.43}$} & \textcolor{orange}{$\mathbf{12.18}$} & $100.0$ & $100.0$ & \textcolor{orange}{$\mathbf{15.20}$} & \textcolor{orange}{$\mathbf{5.88}$}\\\cline{2-8}
& FERET-M & $94.27$ & $10.65$ & $100.0$ & $100.0$ & $11.75$ & $6.51$\\\cline{2-8}
& FRGC-M & $91.42$ & $16.36$ & $100.0$ & $64.86$ & $25.89$ & $14.02$\\\hline \hline
\multirow{3}{*}{XN~\cite{Xception}} & FRLL-M & \color{blue} $\mathbf{99.17}$ & \color{blue} $\mathbf{3.26}$ & \color{blue} $\mathbf{85.29}$ & \color{blue} $\mathbf{28.92}$ & \color{blue} $\mathbf{0.49}$ & \color{blue} $\mathbf{0.0}$\\\cline{2-8}
& FERET-M & \textcolor{orange}{$\mathbf{96.84}$} & \color{blue} $\mathbf{8.25}$ & \color{blue} $\mathbf{79.62}$ & \color{blue} $\mathbf{43.31}$ & \color{blue} $\mathbf{7.29}$ & \textcolor{orange}{$\mathbf{4.03}$} \\\cline{2-8}
& FRGC-M & \color{blue} $\mathbf{96.63}$ & \color{blue} $\mathbf{9.75}$ & \color{blue} $\mathbf{58.19}$ & \color{blue} $\mathbf{35.11}$ & \color{blue} $\mathbf{9.44}$ & \color{blue} $\mathbf{4.23}$\\\hline \hline
\multirow{3}{*}{HRN~\cite{Wang_HRNet_2019}} & FRLL-M & $92.79$ & $13.73$ & $100.00$ & \textcolor{orange}{$\mathbf{42.84}$} & $18.65$ & $11.12$ \\\cline{2-8}
& FERET-M & \color{blue} $\mathbf{97.05}$ & \textcolor{orange}{$\mathbf{8.49}$} & \textcolor{orange}{$\mathbf{91.43}$} & \textcolor{orange}{$\mathbf{54.00}$} & \textcolor{orange}{$7.44$} & \color{blue} $\mathbf{2.27}$\\\cline{2-8}
& FRGC-M & \textcolor{orange}{$\mathbf{95.77}$} & \textcolor{orange}{$\mathbf{10.89}$} & \textcolor{orange}{$\mathbf{82.26}$} & \textcolor{orange}{$\mathbf{55.50}$} & \textcolor{orange}{ $\mathbf{12.14}$} & \textcolor{orange}{$\mathbf{4.46}$}\\%\cline{2-7}
\hline
\end{tabular}}\vspace{-3mm}
\end{center}
\end{table}
\begin{figure*}[!t]
\begin{center}
\centering
\includegraphics[width=0.345\linewidth]{figures/OpenCV_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FaceMorpher_grid.pdf}
\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/StyleGAN_grid.pdf}
\includegraphics[width=0.345\linewidth]{figures/AMSL_grid.pdf}
\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/Webmorph_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/legend_morph_types.png}
\end{center}
\vspace{-20 pt}
\caption{ROC curves generated for the tested MAD models for different types of face morphs. Note that the performance of the detectors differs quite considerably depending on the morphing procedure used.\label{fig:ROC_by_morph_types}}
\end{figure*}
To better understand the differences between the evaluated models, we additionally assess their performance on only one face morphing technique at a time. As can be seen from Figure~\ref{fig:ROC_by_morph_types}, Xception shows the greatest generalization capabilities, when it comes to detection of different face morphing methods. HRNet provides very competitive results on face morphs generated by OpenCV, FaceMorpher and StyleGAN. Nevertheless, AMSL and Webmorph attacks are too challenging for this model. We hypothesize, that this might be due to the structure of the training data. Synthetic face morphs from SMDD are generated using only one morphing method, i.e. OpenCV. With such training data, models are in general at higher risk of overfitting to one specific type of morph. Since HRNet has far more trainable parameters than Xception and MixFaceNet, it is also more prone to overfitting, when trained on smaller datasets like SMDD.
\vspace{-3 pt}
\section{Conclusion}
\vspace{-3 pt}
In this paper, we tackle the privacy issues associated with the datasets used for the development of face morphing detection algorithms. To address related privacy concerns, we explore the idea of using a training database with faces of non--existing people, generated by StyleGAN. Using this data, we train three different MAD models and evaluate their performance on three commonly used real--world datasets. Our experiments show that in general, MAD models
can be successfully trained on synthetic data and generalize well to real--world scenarios.
\footnotesize %
\bibliographystyle{ieee}
\section{Introduction}
\vspace{-3 pt}
Nowadays, the vast majority of applications for person identity verification rely on Face Recognition Systems (FRSs), which match a human face to an entry from a database of faces. Modern FRSs have proven to be highly accurate when genuine faces are presented to the system \cite{grm2018strengths}. They are however prone to various attacks, whose aim is to gain illegal access by false authentication \cite{Huber_SYN-MAD_2022}.
Lately, face morphs have become a growing concern for the reliability of face verification systems. A face mor\-ph is a composite image generated from two (or more) facial images of distinct subjects.
Recent advances in generative deep models have enabled an almost effortless generation of realistic and high-quality morphed facial images. Such images can be utilized to verify all identities that have been used in the morph-generation process. A successful detection of \textit{face morphing attacks} is therefore critical for the prevention of illegal activities \cite{Naser_PW_MAD_2021}.
\begin{figure}[!t]
\begin{center}
\centering
\includegraphics[width=0.9\linewidth]{figures/teaser.pdf}
\end{center}
\vspace{-20pt}
\caption{To avoid privacy related concerns in the development of morphing attack detectors (MAD), we explore the idea of using synthetic training faces of non--existing people. Trained MADs are then tested on real--world datasets.\label{fig:teaser}}
\end{figure}
Various \textit{Morphing Attack Detection} (MAD) algori\-thms have been proposed over the years to automatically distinguish real from morphed faces.
However, regardless of the detection technique, the training of these models requires a massive database of genuine face images. Training protocols therefore raise various privacy-related concerns and limit the amount of publicly available training data that can be used to learn MAD models. In this paper, we address the privacy issues related to MADs by exploring the idea of using synthetic training data, as illustrated in Figure~\ref{fig:teaser}. For this purpose, we use the SMDD~\cite{Damer_SMDD_2022} dataset, where StyleGAN2 \cite{karras2020analyzing}, a state-of-the-art Generative Adversarial Network (GAN), was utilized to generate \textit{bona fide} face images of non-existing people. These images were then used for the generation of the face morphs.
With this dataset, we train two powerful binary classifiers, Xception and HRNet, and evaluate their detection performance on three real-world datasets -- FRLL-Morphs, FERET-Morphs and FRGC-Morphs. The results of the evaluation show that well-performing MAD models can be learned from synthetic data alone, and that the model generalize well over three diverse real-world morph datasets.
\vspace{-3 pt}
\section{Related work}
\vspace{-3 pt}
Existing morphing-attack-detection models can in general be categorized as single--image (S-MAD) or differential (D-MAD) MADs, depending on whether the face morph is examined independently or is compared to a reference sample. While D-MADs can be very accurate in closed--set problems, S-MADs aim to detect attacks without any prior knowledge about human identities. In this section, we only review S-MADs, since they are more closely related to our work.
Regardless of the face morphing technique used, the generated morphs usually contain image irregularities su\-ch as artifacts, noise, pixel discontinuity, distortions, spectrum discrepancies, inconsistent illumination, etc. In the past, shallow algorithms, that implement extraction of photo-response non-uniformity (PRNU) noise~\cite{Scherhag_PRNU_2019} or reflection analysis~\cite{Seibold_reflection_2018} have been successfully employed for the detection of morphing attacks. Some other hand--crafted MAD methods have also used texture--based descriptors, such as LBP~\cite{Ojala_1996_LBP}, LPQ~\cite{Ojansivu_LPQ_2008} or SUR
~\cite{Makrushin_SURF_2019}. Although these methods achieved promising results, they were shown to have limited generalization capabilities. Moreover, as the face morphing techniques improved over time, the performance of shallow methods became less competitive, as they struggled to detect modern, deep-learning generated or heavily post--processed face morphs.
More recent MAD models take advantage of the development of data--driven, deep-learning algorithms. Rag\-havendra \textit{et al.}~\cite{Raghavendra_transfer_2017} were amongst the first to propose transfer learning. In their work, attacks are detected with a simple, fully--connected binary classifier, fed with fused VGG19 and AlexNet features, pretrained on ImageNet. Wandzik \textit{et al.}~\cite{Wandzik_FRS_2018}, on the other hand, achieve highly accurate results with features from general--purpose face recognition systems (FRSs) combined with an SVM. Ramachandra \textit{et al.}~\cite{Ramachandra_Inception_2020} utilize Inception in a similar manner, while Damer \textit{et al.}~\cite{Naser_PW_MAD_2021} argue, that pixel--wise supervision, where each pixel is classified as a bona fide or a morphing attack, is superior, when used in addition to the binary, image--level objective.
Recently, MixFaceNet~\cite{Boutros_MixFaceNet_2021} by Boutros \textit{et al.} achieved state--of--the--art results in different face--related detection tasks, including face morphing detection~\cite{Damer_SMDD_2022}. This model represents a highly efficient architecture that captures different levels of face attack cues by using differently sized convolutional kernels.
\vspace{-3 pt}
\section{Methods}
\vspace{-3 pt}
We consider two different classification models, Xception and HRNet, to detect face morphing attacks in this study and train them using synthetic data only. The two models represent the entries from the University of Ljubljana to the recent \textit{Face Morphing Attack Detection Competition based on Privacy-aware Synthetic
Training Data} (SYN-MAD), held in conjunction with the 2022 International Joint Conference on Biometrics (IJCB 2022)~\cite{Huber_SYN-MAD_2022}, which achieved the best and third best overall performance among all submitted entries.
\textbf{Xception}~\cite{Xception} is a convolutional neural network (CNN) that updates and simplifies the architecture of the InceptionV3 model \cite{szegedy_rethinking_2015} by replacing the Inception modules with depth--wise separable convolutions.
We use Xception as a feature extractor, while the binary classification is performed by a fully connected two--layer network. The output layer consist of $2$ neurons, followed by a softmax activation function.
Similar to previous research, we use cross--entropy as the learning objective.
\textbf{HRNet}~\cite{Wang_HRNet_2019} is again a CNN that unlike other networks maintains high--resolution representations of the input sample through the whole feedforward process.
Su\-ch an architecture contributes to more descriptive image representations, which was proven to improve the results of different classification tasks.
In our experiments, we replace the classification head of HRNet with a two--layer classification module, to perform binary detection of bona fide images and morphing attacks. The output layer consists of only one neuron, followed by a sigmoid activation function. In the training phase, the parameters are optimized using the binary cross--entropy loss.
\begin{figure}[t]
\begin{center}
\centering
\includegraphics[width=\linewidth]{figures/dataset_examples_2.pdf}
\end{center}
\vspace{-20pt}
\caption{Examples of bona fide and morphing attack images from the SMDD~\cite{Damer_SMDD_2022} training dataset and the testing datasets FERET-Morphs, FRLL-Morphs and FRGC-Morphs~\cite{Sarkar2020_morphed_data}. \label{fig:datasets}}
\end{figure}
\begin{table}[t]
\caption{Number of bona fide images (BF), number of morphing attacks (MA) generated by morphing methods OpenCV (OCV), FaceMorpher (FM), StyleGAN (SG), AMSL, Webmorpf (WM) and image size of samples in each dataset.\vspace{-4mm}}
\label{tab1}
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ | l | c | c | c | c | c | c | c |}
\hline
\textbf{Dataset} & \textbf{Image size} & \textbf{BF} & \textbf{OCV} & \textbf{FM} & \textbf{SG} & \textbf{AMSL} & \textbf{WM} \\
\hline
FRLL-M & $1350\times1350$ & $204$ & $1221$ & 1222 & 1222 & 2175 & 1221 \\
FERET-M & $512\times768$ & $1.413$ & 529 & 529 & 529 & / & / \\%$1.587$\\
FRGC-M & $227\times277$ & $3.167$ & 964 & 964 & 964 & / & / \\%$2.892$\\
\hline
\end{tabular}}\vspace{-4mm}
\end{center}
\end{table}
\vspace{-3 pt}
\section{Experiments}
\vspace{-3 pt}
\subsection{Datasets}
We use one synthetic and three publicly available real--world datasets in this work. The training is done exclusively with the synthetic data, while the evaluation is performed on three commonly used face morphing datasets.
\textbf{Training data.}
For training, we use the SMDD data\-set \cite{Damer_SMDD_2022}, provided by the organizers of the SYN-MAD competition~\cite{Huber_SYN-MAD_2022}. The dataset consists of $25.000$ bona fide and $15.000$ morphed images of size $256\times256$ pixels.
Bona fide instances represent carefully selected images from a set of randomly generated StyleGAN2~\cite{karras2020analyzing,Karras_stylegan2_2020} faces. A separate, non -- overlapping StyleGAN2 image set was used for the generation of face morphing attacks. Face morphs were created using the landmark--based morphing technique from OpenCV\footnote{https://learnopencv.com/face-morph-using-opencv-cpp-python/}. A few selected samples from the SMDD dataset are presented in Figure~\ref{fig:datasets}.
\begin{figure*}[!t]
\begin{center}
\centering
\includegraphics[width=0.345\linewidth]{figures/FRLL_grid.pdf} \hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FERET_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FRGC_grid.pdf}
\includegraphics[width=0.345\linewidth,trim = 0mm 100mm 0 0, clip]{figures/legend_models.png}
\end{center}
\vspace{-25 pt}
\caption{ROC curves generated on FRLL-Morphs, FERET-Morphs and FRGC-Morphs for the tested models. Note that HRNet achieves very competitive results on FERET-Morphs and FRGC-Morphs, but performs the weakest on FRLL-Morphs. Xception on the other hand, consistently outperforms the baseline method MixFaceNet, on all considered datasets. \label{fig:ROC_by_datasets}}
\end{figure*}
\textbf{Testing data.}
The trained MAD models are tested on three diverse morphing datasets proposed by Sarkar \textit{et al.} in~\cite{Sarkar2020_morphed_data}, i.e. FRLL-
Morphs, FERET-Morphs and FRGC-Morphs. All face morphs were created by combining bona fide images from their respective face datasets, i.e. FRLL~\cite{debruine_jones_2017, Neubert_FRLL}, FERET~\cite{PHILLIPS_FERET} and FRGC~\cite{Phillips_FRGC}. To generate landmark--based morphs, the authors used OpenCV and FaceMorpher\footnote{https://github.com/alyssaq/face\_morpher}, while deep-learning--based morphs are generated with StyleGAN2. In addition to these methods, AMSL\cite{Neubert_FRLL} and Webmorph\footnote{https://webmorph.org/} are also used, but only for the images from the FRLL dataset. Information about image sizes and the number of samples per morphing method is given in Table~\ref{tab1}. Selected samples from all three datasets are presented in Figure~\ref{fig:datasets}.
\vspace{-5 pt}
\subsection{Experimental setup}
In our experiments, we first preprocess images from all datasets by cropping out the facial areas. Bounding boxes of the SMDD are provided by the authors of the dataset. For the other three databases, we use RetinaFace~\cite{RetinaFace}, to localize the facial region-of-interest. Prior to their use, cropped images are resized to $299\times299$ pixels for Xception and $256\times256$ pixels for HRNet. Additionally, the training data is augmented with horizontal flips to increase the amount of data available and avoid overfitting.
The CNNs were optimized using the Adam optimizer, with a learning rate of $0.0001$. The models were trained from scratch for $30$ full epochs, with a batch size of $16$. After each training epoch, the classification accuracy of the networks was calculated on a small holdout set of each test dataset. The best performing parameters on each of the three datasets were saved as the final model for that particular dataset.
The code was implemented in Python 3.8 with PyTorch 1.9 and CUDA 11.6. Experiments were run on a single GeForce GTX 1080 Ti. The computational complexity of Xception is $11$ GFLOPs, while HRNet has $34$ GFLOPs
\vspace{-3 pt}
\section{Results}
\vspace{-3 pt}
In Figure~\ref{fig:ROC_by_datasets} and Table~\ref{tab:quantitative_results} we present the results obtained with our two models, Xception and HRNet, and the baseline MixFaceNet-MAD from~\cite{Huber_SYN-MAD_2022}. The weights of MixFa\-ceNet-MAD, optimized on the SMDD dataset, were provided by the authors of the model. In our experiments, the best overall results were achieved by Xception, whose Equal Error Rates (EER) are $3.26\%$, $8.25\%$ and $9.75\%$ on FRRL--Morphs, FERET-Morphs and FRGC--Morphs, respectively (Table~\ref{tab:quantitative_results}). The runner-up, HRNet, achieves a similar performance on FERET-Morphs and FRGC-Mo\-rphs. However, among the tested models, HRNet is least successful on FRLL-Morphs, where it achieves an EER of $13.73\%$. On this dataset, MixFaceNet yields a slightly better performance than HRNet with an EER of $12.18\%$, but is outperformed by both, Xception and HRNet, on the other two databases, i.e. FERET-Morphs and FRGC-Morphs. The complete ROC curves of the experiments are visualized in Figure~\ref{fig:ROC_by_datasets} and show a similar picture as the discussed numerical results.
\begin{table}[t]
\caption{Detection results for MAD methods MixFaceNet (MFN), Xception (XN) and HRNet (HRN) on the real--world datasets FRLL-Morphs (FRLL-M), FERET-Morphs (FERET-M) and FRGC-Morphs (FRGC-M). Best scores per dataset are marked blue, while runner--up results are marked orange. All three models were trained on the synthetic SMDD dataset \cite{Damer_SMDD_2022}.
}
\vspace{-20pt}
\label{tab:quantitative_results}
\smallskip
\begin{center}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ | c | c | c | c | c | c | c | c | }
\hline
\multirow{2}{*}{\textbf{MAD}} & \multirow{2}{*}{\textbf{Test data}} & \multirow{2}{*}{\textbf{AUC(\%)}} & \multirow{2}{*}{\textbf{EER(\%)}} & \multicolumn{4}{c |}{\textbf{BPCER (\%) @ APCER =}} \\ \cline{5-8}
& & & & $0.10\%$ & $1.00\%$ & $10.00\%$ & $20.00\%$ \\ \hline
\hline
\multirow{3}{*}{MFN~\cite{Damer_SMDD_2022}} & FRLL-M & \textcolor{orange}{$\mathbf{95.43}$} & \textcolor{orange}{$\mathbf{12.18}$} & $100.0$ & $100.0$ & \textcolor{orange}{$\mathbf{15.20}$} & \textcolor{orange}{$\mathbf{5.88}$}\\\cline{2-8}
& FERET-M & $94.27$ & $10.65$ & $100.0$ & $100.0$ & $11.75$ & $6.51$\\\cline{2-8}
& FRGC-M & $91.42$ & $16.36$ & $100.0$ & $64.86$ & $25.89$ & $14.02$\\\hline \hline
\multirow{3}{*}{XN~\cite{Xception}} & FRLL-M & \color{blue} $\mathbf{99.17}$ & \color{blue} $\mathbf{3.26}$ & \color{blue} $\mathbf{85.29}$ & \color{blue} $\mathbf{28.92}$ & \color{blue} $\mathbf{0.49}$ & \color{blue} $\mathbf{0.0}$\\\cline{2-8}
& FERET-M & \textcolor{orange}{$\mathbf{96.84}$} & \color{blue} $\mathbf{8.25}$ & \color{blue} $\mathbf{79.62}$ & \color{blue} $\mathbf{43.31}$ & \color{blue} $\mathbf{7.29}$ & \textcolor{orange}{$\mathbf{4.03}$} \\\cline{2-8}
& FRGC-M & \color{blue} $\mathbf{96.63}$ & \color{blue} $\mathbf{9.75}$ & \color{blue} $\mathbf{58.19}$ & \color{blue} $\mathbf{35.11}$ & \color{blue} $\mathbf{9.44}$ & \color{blue} $\mathbf{4.23}$\\\hline \hline
\multirow{3}{*}{HRN~\cite{Wang_HRNet_2019}} & FRLL-M & $92.79$ & $13.73$ & $100.00$ & \textcolor{orange}{$\mathbf{42.84}$} & $18.65$ & $11.12$ \\\cline{2-8}
& FERET-M & \color{blue} $\mathbf{97.05}$ & \textcolor{orange}{$\mathbf{8.49}$} & \textcolor{orange}{$\mathbf{91.43}$} & \textcolor{orange}{$\mathbf{54.00}$} & \textcolor{orange}{$7.44$} & \color{blue} $\mathbf{2.27}$\\\cline{2-8}
& FRGC-M & \textcolor{orange}{$\mathbf{95.77}$} & \textcolor{orange}{$\mathbf{10.89}$} & \textcolor{orange}{$\mathbf{82.26}$} & \textcolor{orange}{$\mathbf{55.50}$} & \textcolor{orange}{ $\mathbf{12.14}$} & \textcolor{orange}{$\mathbf{4.46}$}\\%\cline{2-7}
\hline
\end{tabular}}\vspace{-3mm}
\end{center}
\end{table}
\begin{figure*}[!t]
\begin{center}
\centering
\includegraphics[width=0.345\linewidth]{figures/OpenCV_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/FaceMorpher_grid.pdf}
\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/StyleGAN_grid.pdf}
\includegraphics[width=0.345\linewidth]{figures/AMSL_grid.pdf}
\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/Webmorph_grid.pdf}\hspace{-15 pt}
\includegraphics[width=0.345\linewidth]{figures/legend_morph_types.png}
\end{center}
\vspace{-20 pt}
\caption{ROC curves generated for the tested MAD models for different types of face morphs. Note that the performance of the detectors differs quite considerably depending on the morphing procedure used.\label{fig:ROC_by_morph_types}}
\end{figure*}
To better understand the differences between the evaluated models, we additionally assess their performance on only one face morphing technique at a time. As can be seen from Figure~\ref{fig:ROC_by_morph_types}, Xception shows the greatest generalization capabilities, when it comes to detection of different face morphing methods. HRNet provides very competitive results on face morphs generated by OpenCV, FaceMorpher and StyleGAN. Nevertheless, AMSL and Webmorph attacks are too challenging for this model. We hypothesize, that this might be due to the structure of the training data. Synthetic face morphs from SMDD are generated using only one morphing method, i.e. OpenCV. With such training data, models are in general at higher risk of overfitting to one specific type of morph. Since HRNet has far more trainable parameters than Xception and MixFaceNet, it is also more prone to overfitting, when trained on smaller datasets like SMDD.
\vspace{-3 pt}
\section{Conclusion}
\vspace{-3 pt}
In this paper, we tackle the privacy issues associated with the datasets used for the development of face morphing detection algorithms. To address related privacy concerns, we explore the idea of using a training database with faces of non--existing people, generated by StyleGAN. Using this data, we train three different MAD models and evaluate their performance on three commonly used real--world datasets. Our experiments show that in general, MAD models
can be successfully trained on synthetic data and generalize well to real--world scenarios.
\footnotesize %
\bibliographystyle{ieee}
|
\section{Introduction}
Drone small cells (DSCs) have gained popularity in recent years as a solution to wireless communication problems such as lack of fixed infrastructure (e.g. due to natural disasters) or the need for temporal capacity increase (e.g. to manage traffic in a mass-event). If properly deployed, and despite its inherent limitations, UAV-enabled wireless can leverage the performance of the network due to increased probability for line-of-sight (LOS) connectivity and reduced total path loss. Furthermore, due to their mobility, the location of the BS can adapt to the changes in ground user distribution to improve the total throughput of the system. DSCs can also be a cost-effective substitute for building expensive cellular towers and infrastructure where the need for such infrastructure is limited in terms of time or capacity \cite{Mozaffari2019}, or be deployed along with existing infrastructure in heterogeneous cellular networks \cite{Chakareski2019}.
Nonetheless, utilizing UAVs for serving ground users in a given area has its limitations. Limited onboard energy restricts the deployment duration and transmit power thereby limiting the communication range. Furthermore, in the case of multiple DSCs, severe co-channel interference can substantially degrade users' link qualities\cite{abeywickrama}. Therefore, the problem of 3D deployment and user allocation is a complex one when taking into consideration factors such as power minimization and interference mitigation. DSCs locations should minimize the path loss and maximize signal-to-interference-plus-noise-ratio (SINR) along with serving the highest possible number of users. Additionally, DSCs should be aware of each other to avoid inter-UAV collisions \cite{Mozaffari2019}.
\par The topic of DSCs deployment optimization gained a lot of focus in recent years [4-10]. The problem is commonly approached by optimizing a subset of all inherent aspects (e.g. locations or number of drones vs. coverage, transmit power vs. interference, etc.) without taking into account other factors such as initial deployment and user clustering. In this paper, we target the problem of power minimization, interference mitigation, DSC localization, and user clustering jointly along with deciding a minimum required number of drones to serve a given user set with known locations. Our centralized algorithm is based on Gaussian Mixture Models (GMM) and we call it Drone Users Clustering Expectation Maximization (DUCEM) algorithm. We believe it could also serve as an initial clustering step before applying further heuristic mechanisms to further improve system performance. That is, its results are more optimal than other clustering algorithms (e.g. K-means).
\par The paper continues as follows. Section \ref{Related Work} briefly summarizes related research found in the literature. The system model and optimization problem formulation are given in Section \ref{System Model}. We introduce EM algorithm in \ref{EM} and present our modified version for DSCs clustering in Section \ref{DUCEM}. Section \ref{Simulation} introduces random waypoint mobility (RWM) model and reference point mobility group (RPGM) model that are used for user mobility simulation. We also describe K-means algorithm which is chosen as the baseline with which we compare our algorithm's performance and present the results. Finally, in Section \ref{conclusion}, we give our final remarks and conclusion.
\section{Related Work} \label{Related Work}
The application of DSC has been investigated so far in various contexts. In particular, the authors in \cite{Hu2019} developed a machine learning scheme for predicting the number of required drones for load balancing in small cells based on historical data. In \cite{Lyu2015}, DSCs sequential placement is optimized to reduce the total number of drones and achieve maximum coverage. However, interference was not taken into account. The authors in \cite{Mozaffari2015} considered optimal deployment of two DSCs taking into consideration coverage and interference. In \cite{AffinityPropagation}, the authors proposed an interference mitigation scheme using affinity propagation which reduces the power of interfering drones and clusters the users using K-means algorithm without targeting the initial user clustering and allocation problem. In \cite{Kumbhar2018}, the authors mitigate DSC interference in LTE hetNets by applying 3GPP Release-10 enhanced inter-cell interference coordination (eICIC), and optimize UAV deployment using a genetic algorithm (GA). In \cite{Chakareski2019}, the authors optimize the transmit power and altitude of each DSC operating in a heterogeneous LTE network comprised of macro base stations, mmWave small base stations, and DSCs operating in the microwave band. The proposed algorithm determines subcarrier allocation using the Hungarian optimization method without targeting the problem of user-cell association. Using game theory, the authors in \cite{Li2020} formulated the problem in a mean-field game framework where the altitude of DSCs is controlled to improve total SINR. In \cite{Shi}, a solution based on Particle Swarm Optimization (PSO) algorithm was derived that maximizes coverage while maintaining link quality.
In general, the mechanisms reported in the literature do not jointly consider the following dynamic issues of DSCs deployment:
\begin{itemize}
\item number of required drones for a given area
\item interference mitigation between DSCs
\item limited energy of DSCs.
\end{itemize}
Our work presents a heuristic mechanism that tries to address all of the aforementioned aspects of DSCs deployment.
\section{Problem Formulation} \label{System Model}
Fig. 1 presents the considered scheme where $U$ ground users are randomly distributed and served by the set of $M$ drones, $ \mathcal{M}$. We consider a 3D Cartesian coordinates system in which each user $u$ is located at $\boldsymbol x_u= (x_u, y_u)$. The location of the $j$th drone is denoted by $\boldsymbol F_j=(x_j,y_j,h_j)$, where $h_j$ denotes the flying attitude, and $j \in \mathcal{M}$. In our system, we consider that all active drones are flying at the same fixed height $h = 10 \;m$. As discussed in \cite{10m}, $10 \;m$ is the optimal height for positioning a typical small cell antenna. Lower values cause possible coverage issues, and higher ones increase interference with neighbouring cells. Next, we assume drone $j\in \{1,..,M\}$ can transmit at a specified power $P_{T_j}\leq P_{\max}$, and $P_{\max}$ is the maximum transmit power of the DSC. In consequence, it will serve all $U_j \leq U$ users within its coverage area. As will be explained later in Section \ref{DUCEM}, $P_{T_j}$ is directly proportional to $\Sigma_j$, the covariance value of the Gaussian distribution associated with drone $j$.
\begin{figure}[htbp]
\centering
\includegraphics[width=3in]{Scenario.jpg}
\DeclareGraphicsExtensions.
\caption{Scenario of users served by DSCs.}
\label{fig_sim}
\end{figure}
As the drones are flying over a certain area, it is rational to assume that a LOS link exists between drones and users \cite{Wu2018}. Thus, we evaluate the free-space path loss model in our analysis. The distance between each drone and ground user is given by:
\begin{equation}
d_{j,u}=\sqrt{(x_j-x_u)^2+(y_j-y_u)^2+h^2},\label{eqn:dist}
\end{equation}
\noindent and the channel gain from drone $m$ to user $u$ can be defined as:
\begin{equation}
H_{j,u}=\alpha_0d^{-\lambda}_{j,u}.\label{eqn:plos}
\end{equation}
\noindent Here, $\lambda$ is the path loss exponent and $\alpha_0$ is the channel power at reference distance $d_0=1\;\mathrm{m}$. In consequence, the signal-to-noise ratio (SNR) between user $u$ and serving drone $j$ is then given by:
\begin{equation}
\mathrm{SNR}_{j,u}= \frac{P_{T_j} H_{j,u}}{\sigma_0^2}
\end{equation}
\noindent where $\sigma_0^2$ is the power of additive white Gaussian noise (AWGN) at the receiver.
For modeling the interference from other DSCs, the signal-to-interference-plus-noise ratio (SINR) of user $u$ served by DSC $j$, $\Gamma_{u,j}$, is given by:
\begin{equation}
\Gamma_{u,j} = \frac{P_{T_j} H_{j,u}}{\sigma_0^2+\sum_{i\neq j}^M P_{T_i} H_{i,u}}.
\end{equation}
In consequence, the theoretical capacity limit can be derived based on Shannon's formula as
\begin{align}
R_{\mathrm{total}}&=\sum^{M}_{j=1}\sum^U_{u=1} W(j,u)B \;\log_2(1+\Gamma_{u,j}),
\end{align}
where $W(j,u) = 1$ if user $u$ is served by drone $j$ and $0$ otherwise, and $B$ is the allocated bandwidth per user channel.
Focusing only on data communications, the total maximum power consumed by drones can be given by:
\begin{equation}
P_{total}=\sum^M_{i=1}P_{T_i},
\end{equation}
where we have neglected propulsion energy as it is irrelevant to our problem.
Thus, the energy efficiency of the whole system is defined as follows:
\begin{align}
EE&=\frac{R_{\mathrm{total}}}{P_{\mathrm{total}}}.
\end{align}
It is assumed that the transmitted power of each drone can be continuously varied up to a limit $P_{\max}$. Also, the maximum number of users served by each drone is also limited to $U_{\max}$. Furthermore, each user is guaranteed a minimum $\mathrm{SNR}$ value of ${\mathrm{SNR}}_T$. The optimization problem is then formulated as:
\begin{align}
\max_{P_{T_j}, F_j, U_j} &EE = \max_{P_{T_j}, F_j, U_j} \frac{R_{\mathrm{total}}}{P_{\mathrm{total}}}.
\end{align}
$s.t.$
\begin{enumerate}[a)]
\item $\mathrm{SNR}_{j,u}\geq{\mathrm{SNR}}_T \;\;\; \forall u\in \mathcal{U} $
\item $ U_j \;\;\; \leq U_{\max} \;\;\; \forall j\in \mathcal{M}$
\item $P_{T_j} \leq P_{\max} \;\;\; \forall j\in \mathcal{M} $.
\end{enumerate}
Next Sections present our heuristic approach to solving this problem.
\section{Expected-Maximization Algorithm} \label{EM}
Expectation Maximization (EM) algorithm is based on Gaussian Mixture Models (GMMs) where it is assumed that the set of samples is drawn from a "mixture" of Gaussian distributions \cite{MIT}. A GMM can be defined as:
\begin{equation}
\pi(\boldsymbol x;\theta)=\sum^M_{j=1} P(j)N(\boldsymbol x;\boldsymbol\mu_j,\boldsymbol\Sigma_j),
\end{equation}
where the parameters $\theta = \{P, \boldsymbol\mu, \boldsymbol\Sigma \}$ include mixing proportions, means of Gaussian components, and covariances respectively, and $N(\boldsymbol x;\boldsymbol\mu_j,\boldsymbol\Sigma_j)$ represents the $j$-th Gaussian distribution. \par
The EM algorithm learns the parameters $\theta$ given a set of sample points, $\boldsymbol X_t$, and number of mixtures, $M$ (As will be explained later, M denotes the number of drones). It consists of two steps (estimation step, called E-step, and maximization step, called M-step) performed iteratively:
\begin{enumerate}
\item (E-step) Evaluate the posterior assignment probabilities given by:
\begin{align}
p^{(l)}(j|t) &= p(j|\boldsymbol x_t,\theta^{(l)})=\nonumber\\
&=\frac{P^{(l)}(j)N\left( \boldsymbol x_t;\boldsymbol \mu_j^{(l)},\boldsymbol \Sigma_j^{(l)}\right)}{\sum^M_{j^\prime=1}P^{(l)}(j^\prime)N\left(\boldsymbol x_t;\boldsymbol \mu_{j^\prime}^{(l)},\boldsymbol \Sigma_{j^\prime}^{(l)}\right)}\nonumber \\
&=\frac{P^{(l)}(j)N\left(\boldsymbol x_t;\boldsymbol \mu_j^{(l)},\boldsymbol \Sigma_j^{(l)}\right)}{P(\boldsymbol x_t;\theta^{(l)})}
\end{align}
where $(l)$ denotes iteration number, and $\boldsymbol x_t = \left[ {\begin{array}{cc}
x_t \\
y_t \\
\end{array} } \right]$ is the sample $t = \{1,..,n\}$ where $n$ is the number of samples.
\item (M-step) Update parameters according to:
\begin{align}
p^{(l+1)}(j)&= \frac{\hat{n}(j)}{n}, \;\; \text{where } \hat{n}(j)= \sum^n_{t=1}p^{(l)}(j|t).\\
\boldsymbol \mu_j^{(l+1)}&= \frac{1}{\hat{n}(j)}\sum^n_{t=1}p^{(l)}(j|t)\boldsymbol x_t.\\
\boldsymbol \Sigma_j^{(l+1)}&=\\
\frac{1}{\hat{n}}\sum^n_{t=1}&p^{(l)}(j|t)
\left(\boldsymbol x_t-\boldsymbol \mu_j^{(l+1)}\right)\left(\boldsymbol x_t-\boldsymbol \mu_j^{(l+1)}\right)^T.
\nonumber
\end{align}
\end{enumerate}
As can be observed, the
E-step obtains the posterior probabilities that denote the probability of having the sample observed $x_t$ being drawn from the $j$-th Gaussian component. In the M-step, the parameters of each component $j$ are updated to increase the total likelihood of the GMM. It is worth noting that the EM algorithm is guaranteed to converge to a local optimal solution \cite{MIT}.
\section{Drone Users Clustering Expected Maximization Algorithm (DUCEM)} \label{DUCEM}
In this paper, we propose using a modified EM algorithm to tackle the problem of DSCs placement and user clustering along with considering the system energy efficiency.
\par The set of sample points, $\boldsymbol X$, is taken to be the set of users distributed on the ground to be served by DSCs. Furthermore, $\boldsymbol \mu_j$ indicates the $j$-th DSC location in the 2D $x,y$ Cartesian plane, and $\boldsymbol \Sigma_j$ is associated with the transmission power $P_{T_j}$ of drone $j$ (i.e. the values of each covariance matrix are proportional to the transmission power of each DSC). The posterior probabilities $p(j|t)$ bear no mapping to parameters in the real system but serve in determining the power and location of each DSC as evident in equations (12) and (13).
\par The reasoning behind our choice is that by examining equations (10-13) we observe that each sample point $\boldsymbol x_t$ (i.e. location of user $t$) contributes to the parameters of all components (drones). The magnitude of this contribution is determined by the posterior probability given in equation (10). Thus, the contribution of $\boldsymbol x_t$ to DSC $j$ parameters decreases as $p(k|t)_{k\neq j}$ increases. This is evident from (10) as $ N(\boldsymbol x;\boldsymbol \mu_k,\boldsymbol \Sigma_k)$ increases, $ p(j|t)$ decreases. This translates into the observation that user $t$, if served by DSC $k$, would contribute less to parameters updates of DSC $j$ in (11-13). Therefore, because of EM algorithm dynamics, it can be associated with the practical problem of DSC users clustering where each user requires to be served by one DSC.
Our DUCEM algorithm introduces several modifications to the standard EM to suit our considered problem:
\begin{enumerate}
\item Each covariance matrix $\boldsymbol \Sigma_j{_{M\times M}}$, is a diagonal matrix with equal diagonal values to map the DSCs circular coverage of ground users to a circular distribution of samples around means (power is transmitted equally through each axis around the DSC). By removing this constraint, coverage of DSCs will take a non-circular form (i.e. using beamforming). This will be a topic of our future work.
\item The diagonal values of $\boldsymbol \Sigma_j$, denoted as $\Sigma_j$, are limited to a maximum value $\Sigma_{\max}$ to resemble the real case scenario of limited transmit power of drones $P_{\max}$. This limits the number of sample points $\boldsymbol x_t$ to be associated with component $j$.
\item $\Sigma_{j}$ is only allowed to increase while $U_j \leq U_{\max}$. This ensures that the number of users served by DSC $j$ is bounded by $U_{\max}$.
\item The change of $\Sigma_{j}$ per DUCEM iteration is limited by a maximum value of $d\Sigma_{\max}$. This reduces overlapping between clusters and therefore the chance for a cluster to be completely contained in another one. Not limiting $d\Sigma_{\max}$ allows abrupt big jumps in $\Sigma_{j}$ initial updates where sample points far from $\boldsymbol \mu_{j}$ substantially affect the update in (13) because the corresponding posterior probability is large (other components have not yet contributed to the denominator in (10)). Fig. 2 shows the difference between setting a maximum $d\Sigma_{max}$ for a certain set of sample points and $M=4$. As observed, limiting $\Sigma_j$ change in equation (13) reduces overall overlapping. Clearly, this improves interference between DSCs. To summarize, the update in equation (13) is limited by $\boldsymbol \Sigma_j^{(l+1)} - \boldsymbol \Sigma_j^{(l)} \leq d\Sigma_{max}$, and $\Sigma_j^{(l+1)} \leq \Sigma_{\max} $.
\item Each user is associated to the cluster with smallest value of Standarized Euclidean Distance (SED) given by:
\begin{equation}
S\!E\!D_{j,u}=\frac{\sqrt{((x_j-x_u)^2+(y_j-y_u)^2+h^2)}}{\Sigma_j}.
\end{equation}
Users allocations are given by the matrix $\boldsymbol W_{M\times U}$ where $\boldsymbol W(j,u) = 1$ if the $u$-th user is associated with drone $j$ and $0$ otherwise.
\item Each DSC is centered at the mean location of its served users.
\item The algorithm begins with $m=1$ as the initial number of drones (components) and adds another drone while a solution that satisfies the requirements of the maximum number of users or link reliability is not yet found.
\item At each iteration, the total system EE is obtained. The setting that yields maximum EE ($\theta_{opt} = \{ \boldsymbol \mu_1, \Sigma_1 ,...,\boldsymbol \mu_M, \Sigma_M\}$ is saved and is chosen as the final solution. EM algorithm converges when the maximum change, $maxd\theta$, of any of the parameters $\boldsymbol \mu_j$ and $ \Sigma_j$ after iteration $i$ is limited by $\varepsilon_1$. Also, a new DSC is added when the number of DSCs is not enough to find a solution that satisfies the maximum number of users and maximum power constraints. Finally, $\varepsilon_2$ ensures that no new drones are added until EM updates are close to convergence.
\end{enumerate}
\begin{figure}[htbp]
\centering
\subfloat[High overlap for $d\Sigma_{max}=\infty$]{%
\centering
\includegraphics[clip]{dsigmaNone.jpg}%
}
\subfloat[Decreased overlap for $d\Sigma_{max}=0.1$]{%
\centering
\includegraphics[clip]{dsgima.jpg}%
}
\caption{Comparison of clustering overlap for a given set of test points with different $d\Sigma_{max}$ values.}
\end{figure}
\par Algorithm 1 shows the pseudocode for the DUCEM scheme. The algorithm takes as an input the drones maximum power $P_{max}$, maximum number of users per cluster $U_{max}$, and the locations of users $\boldsymbol X$. The location and transmitting power of each drone are initialized arbitrarily around the considered area. Furthermore, two thresholds are set: $\varepsilon_1$ which determines the minimum change in drone parameters ($\boldsymbol\mu, \boldsymbol \Sigma$) after which convergence is reached, and $\varepsilon_2$ which sets the threshold of parameter change below which a new drone would be added if no solution that satisfies power and capacity requirements is yet found. On each iteration the posterior probabilities are obtained and each drone parameters, $\theta$, are updated. Users are then assigned to the drone with lowest resulting $S\!E\!D$ given by (14). The drones 2D locations are then recalculated as the centers of assigned users clusters in step 6. The power of each drone is then obtained by satisfying the $S\!N\!R$ requirement. As long as no solution satisfying the constraints is found the algorithm reiterates. If the maximum change per parameters (step 29) is below $\varepsilon_2$, a new drone is added. The algorithm stops if the change in parameters is below $\varepsilon_1$ and a solution is found. It then returns the settings that result in the highest energy efficiency score (steps 23 - 27).
\begin{algorithm}[H]
\caption{Drone Users Clustering Expectation Maximization}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\REQUIRE $\boldsymbol X$, $U_{\max}$, $P_{\max}$
\ENSURE $M, \;\theta_{opt} = \{ \boldsymbol P,\boldsymbol \mu, \boldsymbol \Sigma\}$, $P_{T_j} \forall j\in \mathcal{M}$, $\boldsymbol W_{M \times U}$ \\
\textit{Initialisation} : Initialize $\theta = \{\boldsymbol P^{(0)},\boldsymbol \mu^{(0)}, \boldsymbol \Sigma^{(0)} \},i=0,$ $maxEE=0$, $\varepsilon_2<\varepsilon_1$
\\ \textit{LOOP Process}
\WHILE {($maxd\theta > \varepsilon_1$ or $SolutionFound=0$)}
\STATE Obtain posterior probabilities $p(j|t)\;\; \forall j\in \mathcal{M}$
\STATE Update $p(j), \boldsymbol \mu_j$ according to (11) and (12) $\forall j\in \mathcal{M}$
\STATE Obtain $\Sigma_{j_{new}}$ according to (13) $\;\;\forall j\in \mathcal{M}$
\STATE Allocate users to DSC with minimum SED given in (14) and obtain $\boldsymbol W^i$
\STATE $\boldsymbol \mu'_j=$ mean ($\boldsymbol W^i_j \boldsymbol X^T$) $\forall j\in\mathcal{M}$
\STATE calculate $P_{T_j}$ according to $\mathrm{SNR}_T \forall j\in\mathcal{M}$
\FOR{$j\in \mathcal{M}$}
\IF {($U_j \;\;\; \leq U_{\max}$ or ${\Sigma_{j_{new}} < \Sigma_j}$)}
\STATE Update $\Sigma_j = \Sigma_{j_{new}}$
\ENDIF
\ENDFOR
\IF {($maxd\theta < \varepsilon_2$ and $SolutionFound=0 $)}
\STATE Add new DSC: $M= M+1$
\STATE Go to step $2$
\ENDIF
\STATE $SolutionFound=1$
\IF {($U_j \;\;\; \geq U_{max}$ or $(P_j>P_{max}))\;\;\forall j\in \mathcal{M} $}
\STATE $SolutionFound=0$
\ENDIF
\IF{$SolutionFound=1$}
\STATE Obtain EE
\IF{$EE>maxEE$}
\STATE $maxEE=EE^i$
\STATE $\theta_{opt}=\theta^i$
\STATE $\boldsymbol W_{opt}=\boldsymbol W^i$
\ENDIF
\ENDIF
\STATE $maxd\theta=\max\{ \max_{j\in M}\{\mu^i_j-\mu_j^{i-1},\Sigma^i_j-\Sigma_j^{i-1}\}\}$
\STATE $i=i+1$
\ENDWHILE
\RETURN $m, \;\theta_{opt} = \{ \boldsymbol P,\boldsymbol \mu, \boldsymbol \Sigma\}$, $P_{T_j} \forall j\in \mathcal{M}$, $\boldsymbol W_{opt}$
\end{algorithmic}
\end{algorithm}
\section{Simulation and Results} \label{Simulation}
\subsection{Users Mobility Model}
For our simulations, we consider that users are moving within groups. This model reflects the scenario in which DSCs deployment is immediately required in mass-events (e.g. concerts, street gatherings, etc.). Other scenarios are also applicable. Each group has a leader moving according to the random waypoint mobility (RWM) model \cite{Bettstetter2003}. Each leader movement is determined by drawing from a memoryless stochastic process with three random variables $D_i,T_{p,i},$ and $V_i$ denoting destination point at instance $i$, pause time at the destination, and traveling speed respectively.
Furthermore, each leader is surrounded by group members that are moving according to the reference point group mobility model (RPGM) \cite{RPGM}. According to RPGM the movement of group members can be characterized by the respective motion vector:
\begin{equation}
\boldsymbol V^{(g)}_i(t)=\boldsymbol V^{(g)}_{\mathrm{leader}}(t)+ \boldsymbol {RM}
\end{equation}
where $i, g$ denote the member and group numbers respectively, $\boldsymbol V^g_{leader}(t)$ is the motion vector of group leader $g$ at time instant $t$, and $\boldsymbol {RM}$ is a random vector deviated by group member $i$ from its leader. The movement of group members can be represented by the angle and magnitude deviation between leaders and members' motion vectors. The equations that determine the deviation at time instant $t$ are as follows:
\begin{align}
|\boldsymbol V_{\mathrm{member}}(t)|&=|\boldsymbol V_{\mathrm{leader}}(t)| + r \; \phi_v dV_{max} \\
\theta_{member}(t)&=\theta_{leader(t)} + r\; \phi_\theta d\theta_{max}
\end{align}
where $r \in [0,1]$ is a random number drawn from a uniform distribution at each group member movement interval, $\phi_v$ and $\phi_\theta$ are the speed deviation ratio and angle deviation ratio respectively, and $dV_{max}$ and $d\theta_{max}$ are the group members maximum speed and angle deviation respectively. Arbitrary settings of $\phi_v$, $\phi_\theta$, $dV_{max},$ and $d\theta_{max}$ yield varied simulation scenarios.
\subsection{Simulation Parameters}
In our simulations we consider an area of $1200 \;\mathrm{m}^2$ for the users' mobility model. The simulation parameters are listed in Table 1. We compare our algorithm results in terms of link reliability and energy efficiency for different values of maximum number of users per cluster $U_{\max}$. The link reliability is given by:
\begin{equation}
L_{rel}= \Pr\{\Gamma_{u,j} \geq \Gamma_{Th}\}, \;\;\; \forall j\in \mathcal{M},
\end{equation}
where $\Gamma_{u,j}$ is the SINR between user $u$ and drone $j$ given by (4), and $\Gamma_{Th}$ is the SINR threshold above which the link is considered reliable.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.3}
\caption{\textsc{Simulation Parameters}}
\label{table_example}
\centering
\begin{tabular}{|c||c|}
\hline
\textbf{Parameter} & \textbf{Value}\\
\hline
DSCs altitude $h$ & 10 m\\
\hline
Path loss exponent $\lambda$ & 2\\
\hline
DSCs max transmit power $P_{\max}$ & 1 W\\
\hline
Noise power $\sigma_0$ & -100 dBm\\
\hline
Received power at reference distance 1 m $\alpha_0$ & -30 dB\\
\hline
$d\Sigma_{\max}$ & 0.1\\
\hline
Channel bandwidth $B$ & 10 MHz\\
\hline
$\mathrm{SNR}_{T}$ & 12 dB\\
\hline
\end{tabular}
\end{table}
\subsection{K-means Clustering}
K-means clustering is a special case of EM where all components have equal covariances $\Sigma_j$ and each point is assigned to the component with higher posterior probability $p(j|t)$ \cite{MIT}. The algorithm takes in the number of clusters and sample points $(X,m)$ as an input. For our simulations, we employ K-means clustering with simple modifications to better suit our application. Modifications concern the realistic restriction of the maximum number of users per cluster $U_{\max}$, and maximum transmission power $P_{\max}$.
\par K-means clusters users with the specified restrictions of $U_{\max}$ and $P_{max}$ that DUCEM satisfies. At each iteration, K-means algorithm is repeated 10 times with different initialization determined by Kmeans++ algorithm \cite{kmeans++}. The algorithm is presented in Algorithm 2. As an input, K-means receives the drones maximum power $P_{max}$, maximum number of users per cluster $U_{max}$, and the locations of users $\boldsymbol X$. The algorithm begins with one drone and adds a new one (steps 6- 8) if the constraints are not satisfied. Similarly to DUCEM, the transmission power for each drone is obtained by satisfying the $S\!N\!R$ constraint (step 5) after positioning the drone at the center of the associated users (step 4). Again, the settings that yields the highest $E\!E$ score is returned.
\begin{algorithm}[H]
\caption{K-means Clustering}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\REQUIRE $\boldsymbol X$, $U_{max}$, $P_{max}$\\
\ENSURE $\boldsymbol \mu$, $\boldsymbol W_{opt}$ \\
\textit{Initialisation}: Initialize $M=1$, $i=0$, $maxEE=0$
\\ \textit{LOOP Process}
\WHILE {($i<10$)}
\STATE Cluster users using K-means with inputs $(\boldsymbol X,M)$
\FOR{$j\in \mathcal{M}$}
\STATE $\boldsymbol \mu_j=$ mean ($\boldsymbol W^i_j X^T$)
\STATE calculate $P_j$ according to $\mathrm{SNR}_T$
\IF {($ U_j \; > U_{max}$) or ($P_j > P_{\max}$)$ \forall j\in \mathcal{M}$}
\STATE Cluster users in cluster $j$ using K-means with inputs $(\boldsymbol W^i_j \boldsymbol X^T,2)$
\STATE Check if sub-cluster $k$ satisfies conditions in $6$. If not, repeat $7$ for cluster $k$
\ENDIF
\ENDFOR
\IF{$EE>maxEE$}
\STATE $maxEE=EE^{(i)}$
\STATE $\boldsymbol W_{opt}=\boldsymbol W^i$
\ENDIF
\STATE $i=i+1$
\ENDWHILE
\RETURN $\boldsymbol \mu_j$, $\boldsymbol W_{opt}$
\end{algorithmic}
\end{algorithm}
\subsection{Results}
Fig. \ref{EE} shows the obtained EE for different values of $U_{max}$. On average, the EE of the system is improved by $25\%$. Furthermore, the achieved link reliability is obtained for two different values of $\Gamma_{Th}$. The results show that DUCEM provides, on average, $18.3\%$ better link reliability. Fig. \ref{Reliability} plots $L_{rel}$ results for $U_{max}=1500$.
\par We observe that DUCEM results are indeed more optimal than the widely used K-means clustering mechanism. The improvement in performance is almost constant for different values of $U_{max}$. This is expected since both algorithms do not directly optimize inter-cell interference which is expected to have greater effect with increasing number of drones (i.e. decreasing value of $U_{max}$). Employing DUCEM with the mechanisms mentioned in Section \ref{Related Work} is expected to speed up convergence and achieve better results which will be the topic of our future work.
\begin{figure}[htbp]
\centering
\includegraphics[width=2.9in]{Energy_Efficiency.jpg}
\DeclareGraphicsExtensions.
\caption{System energy efficiency (EE) against the maximum allowed number of users per cluster ($U_{max}$).}
\label{EE}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=2.9in]{Reliability.jpg}
\DeclareGraphicsExtensions.
\caption{Link reliability ($L_{rel}$) against SINR threshold ($\Gamma_{Th}$). $U_{max}=1500$.}
\label{Reliability}
\end{figure}
\section{Conclusion}
\label{conclusion}
DSCs provide a suitable solution for providing wireless coverage. However, their deployment problem requires consideration of various aspects such as co-channel interference, 3D placement, user clustering, and transmit power minimization. The problem of user clustering for DSCs has received limited attention. In this paper, a user clustering algorithm was proposed that takes into account the aforementioned aspects of the problem. The numerical results verified that the proposed algorithm enhances system performance in terms of energy efficiency and link reliability. The results of DUCEM can provide the initial clustering upon which further performance improving mechanisms can be implemented. To further improve the performance of DUCEM, the problem of inter-cell interference should be tackled and optimized. Other aspects to be considered and optimized for are users handover and the drone's propuslion energy.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
\input{introduction.tex}
\section{Related Work}
\label{sec:related_work}
\input{related_work.tex}
\section{Learning Interpretable Predictors by Composing Queries via Information Pursuit}
\label{section: Quantifying e-entropy}
\input{quantifying_explainable_entropy.tex}
\section{Information Pursuit using Variational Autoencoders and Unadjusted Langevin}
\label{section: Information Pursuit}
\input{information_pursuit.tex}
\section{Experiments}
\label{section: Experiments}
\input{experiments.tex}
\section{Conclusion}
\label{section: Conclusion}
\input{conclusion.tex}
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors thank Mar\'{\i}a P\'erez Ortiz and John Shawe-Taylor for their contributions to the design of the experiments on document classification presented in Section \ref{Section: Word-based Queries}. This research was supported by the Army Research Office under the Multidisciplinary University Research Initiative contract W911NF-17-1-0304 and by the NSF grant 2031985.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\section{}
In proofs of propositions and lemmas we rewrite the statement (un-numbered) for convenience.
\subsection{Characterizing the optimal strategy $\pi^*$}
\label{Appendix: Characterizing the optimal strategy}
In this subsection we characterize the optimal strategy for any such arbitrary query set chosen by the user.
\begin{proposition*}
Assuming $Q$ is finite and when $d$ is taken to be the KL-divergence then objective \eqref{eq: complexity definition} can be rewritten as,
\begin{align}
&H_Q^{\epsilon}(X; Y) := \min_{\pi} \mathbb{E}_{X}\left[|expl^{\pi}_Q(X)|\right] \\
&\textrm{s.t.} \; \sum_{k=1}^{\tau^{\pi}}I(Y; S^{\pi}_k(X) \mid S^{\pi}_{k - 1}(X)) \geq I(X; Y) - \epsilon \nonumber
\end{align}
where, $\tau^{\pi} = \max\{t^{\pi}(x): x \in {\mathcal{X}}\}$ and $t^{\pi}(X)$ is defined as the number of queries selected by $\pi$ for input $X$ until $q_{STOP}$. We define $S^{\pi}_{k}(X)$ as a random variable where any realization $S^{\pi}_{k}(x^{\textrm{obs}})$, $x^{\textrm{obs}} \in {\mathcal{X}}$, denotes the event
\[ S^{\pi}_k(x^{\textrm{obs}}) := \{x' \in {\mathcal{X}} \mid \{q_i, q_i(x^{\textrm{obs}})\}_{1:k} = \{q_i, q_i(x')\}_{1:k} \},\]
where $q_i$ is the $i^{th}$ query selected by $\pi$ for input $x^{\textrm{obs}}$.
Here we use the convention that $S^{\pi}_{0}(X) = \Omega$ (the entire sample space) and $S^{\pi}_{l}(X) = S^{\pi}_{t^{\pi}(X)}(X)$ $\; \forall l > t^{\pi}(X)$.
\end{proposition*}
\begin{proof}
We begin by reformulating our sufficiency constraint in \eqref{eq: complexity definition} in terms of entropy by taking the ``distance" between probability distributions as the KL divergence. The $\epsilon$-Sufficiency constraint can then be rewritten as,
\begin{equation*}
\begin{aligned}
\epsilon &\geq \mathbb{E}_X[KL\left(p(Y \mid X), p(Y \mid expl^{\pi}_Q(X)\right)] \\
&= \mathbb{E}_X\left[\mathbb{E}_{Y}\left[\log \frac{p(Y \mid X)}{p(Y \mid expl^{\pi}_Q(X))} \mid X\right]\right] \\
&= \mathbb{E}_X\left[\mathbb{E}_{Y}\left[\log \frac{p(Y \mid X)}{P(Y)} \mid X\right]\right] + \mathbb{E}_X\left[\mathbb{E}_{Y}\left[\log \frac{p(Y)}{P(Y \mid expl^{\pi}_Q(X))} \mid X\right]\right] \\
&= \mathbb{E}_{X, Y}\left[\log \frac{p(Y \mid X)}{P(Y)}\right] - \mathbb{E}_{Y, expl^{\pi}_Q(X)}\left[\log \frac{p(Y \mid expl^{\pi}_Q(X))}{P(Y)}\right] \\
&= I(X; Y) - I(expl^{\pi}_Q(X); Y) \\
&= H(Y \mid expl^{\pi}_Q(X)) - H(Y \mid X)
\end{aligned}
\end{equation*}
In the third equality we multiplied the term inside the $\log$ by the identity $\frac{P(Y)}{P(Y)} = 1$. The fifth inequality is by definition of mutual information. Thus, we can rewrite \eqref{eq: complexity definition} as,
\begin{align}
\label{eq: complexity definition 2}
H^{\epsilon}_Q(X; Y) := &\min_{\pi} \mathbb{E}_{X}\left[|expl^{\pi}_Q(X)|\right] \\
&\textrm{s.t.} \; H(Y \mid expl^{\pi}_Q(X)) - H(Y \mid X) \leq \epsilon \quad (\epsilon\textit{-Sufficiency)} \nonumber
\end{align}
Let's define $t^{\pi}(X)$ as the number of queries selected by $\pi$ for input $X$ until $q_{STOP}$. Define $\tau^{\pi} = \max\{t^{\pi}(X): X \in {\mathcal{X}}\}$. For the purpose of analysis we can vacuously modify $\pi$ such that for any given $x^{obs} \in {\mathcal{X}}$, $\pi$ asks a fixed $\tau^{\pi}$ number of queries by filling the remaining $\tau^{\pi} - t^{\pi}(x^{obs})$ queries with $q_{STOP}$. An immediate consequence of this modification is that $S^{\pi}_{l}(X) = S^{\pi}_{t^{\pi}(X)}(X)$ $\; \forall l > t^{\pi}(X)$.
We will now show that the sufficiency criteria $H(Y \mid expl^{\pi}_Q(X)) - H(Y \mid X)$ can be rewritten as a sum of successive mutual information terms.
\begin{equation}
\begin{aligned}
&H(Y \mid expl^{\pi}_Q(X)) - H(Y \mid X) \\
&= (H(Y) - H(Y \mid X)) - (H(Y) - H(Y \mid expl^{\pi}_Q(X)))\\
&= I(X; Y) - (H(Y) - H(Y \mid S^{\pi}_{\tau}(X))\\
\end{aligned}
\end{equation}
The third equality uses the fact that $H(Y \mid expl^{\pi}_Q(X)) = H(Y \mid S^{\pi}_{\tau}(X))$ since the $\forall x \in X$, $S^{\pi}_{\tau}(x) = S^{\pi}_{t^{\pi}(x)}(x)$.
We can now write $H(Y) - H(Y \mid S^{\pi}_{\tau}(X))$ as a telescoping series,
\begin{equation}
\begin{aligned}
H(Y) - H(Y \mid S^{\pi}_{\tau}(X) &= H(Y) - H(Y \mid S^{\pi}_1(X)) + H(Y \mid S^{\pi}_1(X))
- H(Y \mid S^{\pi}_1(X), S^{\pi}_2(X)) \\&+ H(Y \mid S^{\pi}_1(X), S^{\pi}_2(X)) \ldots + H(Y \mid S^{\pi}_1(X), S^{\pi}_2(X), \ldots, S^{\pi}_{\tau - 1}(X)) - H(Y \mid S^{\pi}_{\tau}(X)) \\
&= I(Y; S^{\pi}_1(X)) + I(Y; S^{\pi}_2(X) \mid S^{\pi}_1(X))) + \ldots + I(Y; S^{\pi}_{\tau}(X) \mid S^{\pi}_{\tau - 1}(X))
\end{aligned}
\end{equation}
The last equality is obtained by noticing that
\begin{enumerate}
\item $H(Y \mid S^{\pi}_1(X), \ldots, S^{\pi}_{k - 1}(X)) - H(Y \mid S^{\pi}_1(X), \ldots, S^{\pi}_{k}(X)) = I(Y; S^{\pi}_{k}(X) \mid S^{\pi}_{k - 1}(X))$,
\item $H(Y \mid S^{\pi}_{\tau}(X)) = H(Y \mid S^{\pi}_{1}(X), S^{\pi}_{2}(X), \ldots, S^{\pi}_{\tau}(X))$ since the events $\{S^{\pi}_{k}(x)\}_{1:\tau}$ are nested $\forall x \in {\mathcal{X}}$.
\end{enumerate}
Putting it all together we can rewrite \eqref{eq: complexity definition 2} as,
\begin{align}
&H_Q^{\epsilon}(X; Y) := \min_{\pi} \mathbb{E}_{X}\left[|expl^{\pi}_Q(X)|\right] \\
&\textrm{s.t.} \; I(X; Y) - \sum_{k=1}^{\tau^{\pi}}I(Y; S^{\pi}_k(X) \mid S^{\pi}_{k - 1}(X)) \leq \epsilon \nonumber
\end{align}
\end{proof}
\subsection{Approximation guarantees for \textrm{IP}}
\label{Appendix: proof of performance guarantee IP}
In Proposition \ref{prop: IP guarantees} we prove that the $\textrm{IP}$ strategy comes within $1$ bit of $H(Y)$ (the entropy of $Y$) under the assumption that one has implicit access to all possible binary functions of $Y$ as queries. Given such a query set, it is well-known that the optimal strategy $\pi^*$ is given by the Huffman Code for $Y$ which is also within $1$ bit of $H(Y)$. Thus, the result is immediate that $\textrm{IP}$ asks at most 1 query more than $\pi^*$ on average.
\begin{proposition} Let $Y$ be discrete. Let $\tilde{H}_Q(X; Y)$ be the expected description length obtained by the $\textrm{IP}$ strategy. If $H(Y | X) = 0$ and $Q$ is the set of all possible binary functions of $X$ such that $H(q(X) \mid Y) = 0$ $\forall q \in Q$, then $H(Y) \leq \tilde{H}_Q(X; Y) \leq H(Y) + 1$.\!\!
\end{proposition}
We make two remarks before turning to the proof.\\
\noindent
{\bf Remark 1:}
The task is to determine the true state of a latent variable $Y \in
\mathcal Y$ based on querying an observed data point $x^0$. We assume
$Y=f(X)$ with $f$ unknown (since $H(Y \mid X) = 0$). Were $Y$ observable, the natural queries
would be indexed by subsets $D
\subset \mathcal Y$, one query for every $D \subset \mathcal Y$, namely
$q(Y)=1$ if $Y \in D$ and 0 otherwise. (This is essentially the
classic ``twenty questions game'', but with an ``oracle'' and ``complete
tests''.) There are $2^{|\mathcal Y|}$ such queries and obviously they
collectively determine $Y$. Now since $Y=f(X)$, these queries are, at
least implicitly, functions of the data $X$, but we need {\it realizable}
functions, not requiring knowledge of $f$. So our fundamental
assumption is that for each subset $D \in \mathcal Y$ the
corresponding subset $D' \in \mathcal X$ ($D'=f^{-1}(D)$) can
be checked for inclusion of $X$, i.e., $Y \in D$
if and only if $X \in D'$. Or, what is the same, a binary query
$q(X)$ (and still denoted $q$ for simplicity) with
$q(X)=q(Y)$. In effect, we are assuming that whereas
we cannot determine $Y$ directly from $X$, we can answer simple
binary queries which determine $Y$ {\it and} can be expressed
as observable data features.
\noindent
{\bf Remark 2:} The sequence of queries $q_1,q_2,...$ generated by the
IP algorithm for a particular data point can be seen as one branch,
root to leaf, of a decision tree constructed by the standard machine
learning strategy based on successive reduction of uncertainty about $Y$ as
measured by mutual information: $q_1 =
\argmax_{q \in Q} I(q(X);Y), q_{k+1}= \argmax_{q \in
Q}I(A_q(X);Y|S^{\textrm{IP}}_k(x^0))$ where the $S^{\textrm{IP}}_k(x^0)$ is the event that for
the first $k$ questions the answers agree with those for $x^0$. We
stop as soon as $Y$ is determined. Whereas a decision tree
accommodates all $x$ simultaneously, the questions along the branch depends
on having a particular, fixed data point. But the learning problem
in the branch version (``active testing'') is exponentially
simpler.
\medskip
\begin{proof}
The lower bound $H(Y) \leq \tilde{H}_Q(X; Y)$ comes from Shannon's source coding theorem for stochastic source $Y$.\\ Now for the upper bound,
since $I(q(X);YS^{\textrm{IP}}_k(x^0))=H(q(X)|S^{\textrm{IP}}_k(x^0))-H(q(X)|Y,S^{\textrm{IP}}_k(x^0))$ and
since $Y$ determines $q(Y)$ and hence also $q(X)$, the second
entropy term is zero (since given $H(A_q(X)\mid Y) = 0)$. So our problem is maximize the conditional entropy
of the binary random variable $q(X)$ given $S^{\textrm{IP}}_k(x^0)$. So
the IP algorithm is clearly just
``divide and conquer'':
\[ q_1 = \argmax_{q \in Q} H(q(X)),\]
\[ q_{k+1}= \argmax_{q \in Q}H(q(X)|S^{\textrm{IP}}_k(x^0)).\]
Equivalently, since entropy of a binary random variable $\rho$ is maximized when $P(\rho) = \frac{1}{2}$,
\[ q_{k+1} = \arg \min_{q \in Q}|P(q(X)= 1|S^{\textrm{IP}}_k(x^0)) - \frac{1}{2}|.\]
Let ${\mathcal Y}_k$ be the set of ``active hypotheses'' after
$k$ queries (denoted as ${\mathcal{A}}_k$), namely those $y$ with positive posterior probability:
$P(Y=y|S^{\textrm{IP}}_k(x^0))>0$.
Indeed,
\begin{equation*}
\begin{aligned}
P(Y=y|S^{\textrm{IP}}_k(x^0)) &= \frac{P(S^{\textrm{IP}}_k(x^0)|Y=y)p(y)}{\sum_y P(S^{\textrm{IP}}_k(x^0)|Y=y)p(y)} \\
&= \frac{1_{{\mathcal Y}_k} p(k)}{\sum_{y \in {\mathcal A}_k}p(y)}
\end{aligned}
\end{equation*}
since
\[P(S^{\textrm{IP}}_k(x^0)|Y=y) = \left\{ \begin{array}{ll}
1, & \textrm{if $y \in {\mathcal A}_k$} \\
0, & y \notin {\mathcal A}_k \end{array}\right.\]
In particular, the classes in the active set
have the {\it same relative weights} as at the outset. In summary:
\begin{center}
$p(y|S^{\textrm{IP}}_k(x^0)) = \left\{ \begin{array}{ll}
p(y)/\sum_{{\mathcal A}_k}p(l), \,\,
y \in {\mathcal A}_k\\
0, & \textrm{otherwise}
\end{array}\right.$
\end{center}
The key observation to prove the theorem
is that if a hypothesis $y$ generates the same answers
to the first $m$ or more questions as $y^0$,
and hence is active at step $m$, then
its prior likelihood $p(y)$ is at most
$2^{-(m-1)},\,\, m =1,2,\ldots$.
This is intuitively clear: if $y$ has the same answer as
$y^0$ on the first question, and $p(y^0) > \frac{1}{2}$, then
only one question is needed and the active set is empty
at step two; if ${q_1}(y)={q_1}(y^0)$ and ${q_2}(y)={q_2}(y^0)$
and $p(y^0) > \frac{1}{4}$, then
only two question are needed and the active set is empty
at step three, etc.
Finally, since $C$, the code length,
takes values in the non-negative integers
$\{0,1,\ldots, \}$:
\begin{eqnarray*}
\tilde{H}_Q(X; Y) &:=& \mathbb{E}[C] \\ &=& \sum_{m=1}^\infty P(C \geq m) \\
&\leq& \sum_{m=1}^\infty P(p(Y) < 2^{-(m-1)}) \\
&=& \sum_{m=1}^\infty \sum_{y: p(y) < 2^{-(m-1)}} p(y)\\
&=& \sum_{y \in {\mathcal Y}} \sum_{m=1}^\infty
1_{\{p(y) < 2^{-(m-1)}\}}p(k) \\
&=& \sum_{y \in {\mathcal Y}}p(k) (1 - \log p(k)) \\
& = & H(Y) + 1
\end{eqnarray*}
\end{proof}
\subsection{Termination Criteria for $\textrm{IP}$}
\label{Appendix: termination criteria for IP}
We would first analyze the termination criteria for the exact case, that is, $\epsilon = 0$ in \eqref{eq: complexity definition}, and then move on to the more general case.
\myparagraph{Termination Criteria when $\boldsymbol{\epsilon = 0}$} Ideally for a given input $x^{\textrm{obs}}$, we would like to terminate ($\textrm{IP}$ outputs $q_{STOP}$) after $L$ steps if
\begin{equation}
p(y \mid x^{\textrm{obs}}) = p(y \mid x')
\; \forall x' \in S^{\textrm{IP}}_L(x^{\textrm{obs}}),\ y \in {\mathcal{Y}}
\label{eq: terminating criteria}
\end{equation}
Recall, $S^{\textrm{IP}}_L(x^{\textrm{obs}}) = \{x' \in {\mathcal{X}} \mid \{q_i, q_i(x')\}_{1:L} = \{q_i, q_i(x^{\textrm{obs}})\}_{1:L}\}$. In other words, its the event consisting of all $x' \in {\mathcal{X}}$ which share the first $L$ query-answer pairs with $x^{\textrm{obs}}$.
If \eqref{eq: terminating criteria} holds for all $x^{\textrm{obs}} \in {\mathcal{X}}$, the it is easy to see that this is equivalent to the sufficiency constraint in the case $\epsilon=0$,
\[p(y \mid x) = p(y \mid expl^{\textrm{IP}}_Q(x)) \; \forall (x, y) \times ({\mathcal{X}} \times {\mathcal{Y}})\]
where $p(y \mid expl^{\textrm{IP}}_Q(x)) := p(y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})) \; \forall (x, y) \in ({\mathcal{X}} \times {\mathcal{Y}})$ and $t^{\textrm{IP}}(x)$ is the number of iterations IP takes on input $x$ before termination.
Unfortunately, detecting \eqref{eq: terminating criteria} is difficult in practice. Instead we have the following lemma which justifies our stopping criteria for IP.
\begin{lemma}
For a given input $x^{\textrm{obs}}$ if event $S^{\textrm{IP}}_L(x^{\textrm{obs}})$ (after asking $L$ queries) satisfies the condition specified by \eqref{eq: terminating criteria} then for all subsequent queries $q_m$, $m \geq L$, $\max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) = 0$.
\label{lemma: termination_criterion}
\end{lemma}
Refer to Appendix \ref{Appendix: proof of terminating criteria} for a proof.
Inspired from Lemma \ref{lemma: termination_criterion} we formulate an optimistic stopping criteria as,
\begin{equation}
L = \inf\{k \in \{1, 2, ..., |Q|\}: \max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) = 0 \ \forall m \geq k, m \leq |Q|\}
\label{eq: terminating condition based on lemma exact}
\end{equation}
Evaluating \eqref{eq: terminating condition based on lemma exact} would be computationally costly since it would involve processing all the queries for every input $x$. We employ a more practically amenable criteria
\begin{equation}
q_{L+1} = q_{STOP} \quad \textrm{ if } \quad \max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) = 0 \; \forall m \in \{L, L + 1, ..., L + T\}
\label{eq: terminating condition based on lemma exact practical}
\end{equation}
$T > 0$ is a hyper-parameter chosen via cross-validation.
Note, it is possible that there does not exist any informative query in one iteration, but upon choosing a question there suddenly appears informative queries in the next iteration. For example, consider the XOR problem. $X \in {\mathbb{R}}^2$ and $Y \in \{0, 1\}$. Let $Q$ be the set to two axis-aligned half-spaces. Both half-spaces have zero mutual information with $Y$. However, upon choosing any one as $q_1$, the other half-space is suddenly informative about $Y$. Equation \eqref{eq: terminating condition based on lemma exact practical} ensures that we do not stop prematurely.\\
\myparagraph{Termination Criteria for general $\epsilon$ when $d$ is taken as the KL-divergence}
For a general $\epsilon > 0$ we would like IP to terminate such that on average,
\begin{equation}
\mathbb{E}_X[KL\left(p(Y \mid X), p(Y \mid expl^{\textrm{IP}}_Q(X)\right)] \leq \epsilon
\label{eq: random stuff}
\end{equation}
Detecting this is difficult in practice since IP is an online algorithm and only computes query-answers for a given input $x$. So it is not possible to know apriori when to terminate such that in expectation the KL divergence would be less than $\epsilon$. Instead we opt for the stronger requirement that,
\begin{equation}
KL\left(p(Y \mid x), p(Y \mid expl^{\textrm{IP}}_Q(x)\right) \leq \epsilon \quad \forall x \in {\mathcal{X}}.
\label{eq: stronger_requirement_general_epsilon}
\end{equation}
It is easy to see that \eqref{eq: stronger_requirement_general_epsilon} implies \eqref{eq: random stuff}.
As before, $p(y \mid expl^{\textrm{IP}}_Q(x)) := p(y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})) \; \forall (x, y) \in ({\mathcal{X}} \times {\mathcal{Y}})$ and $t^{\textrm{IP}}(x)$ is the number of iterations IP takes on input $x$ before termination. We have the following lemma (analogous to the $\epsilon = 0$ case).
\begin{lemma} We make the following assumptions:
\begin{enumerate}
\item ${\mathcal{Y}}$ is a countable set (recall $Y \in {\mathcal{Y}}$).
\item For any $x^{\textrm{obs}} \in {\mathcal{X}}$ and $x_1, x_2 \in S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})$, we have $p(Y \mid x_1)$ and $p(Y \mid x_2)$ have the same support.
\end{enumerate}
Then, for given input $x^{\textrm{obs}}$ if event $S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})$ (after asking $t^{\textrm{IP}}(x^{\textrm{obs}})$ queries) satisfies the condition specified by \eqref{eq: stronger_requirement_general_epsilon} then for all subsequent queries $q_m$, $m \geq t^{\textrm{IP}}(x^{\textrm{obs}})$, $\max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) \leq \epsilon'$, where $\epsilon' = C\epsilon$ for some constant $C > 0$.
\label{lemma: termination_criterion general}
\end{lemma}
Refer to Appendix \ref{Appendix: proof of terminating criteria general} for a proof. The assumption of ${\mathcal{Y}}$ being countable is typical for supervised learning (the scenario considered in this paper) where the set of labels is often finite. The third assumption intuitively means that $P(y \mid x_2) = 0 \implies P(y \mid x_1) = 0$ for any $y \in {\mathcal{Y}}$. This is a reasonable assumption since we envision practical scenarios in which $\epsilon$ is close to $0$ and thus different inputs which share the same query-answers until termination by $\textrm{IP}$ are expected to have very ``similar" posteriors.\footnote{We refer to the distribution $p(y \mid x)$ for any $x \in {\mathcal{X}}$ as the posterior distribution of $x$.}
Inspired from Lemma \ref{lemma: termination_criterion general} we formulate an optimistic stopping criteria $\forall x^{\textrm{obs}} \in {\mathcal{X}}$ as,
\begin{equation}
t^{\textrm{IP}}(x^{\textrm{obs}}) = \inf\{k \in \{1, 2, ..., |Q|\}: \max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) \leq \epsilon' \ \forall m \geq k, m \leq |Q|\}
\label{eq: terminating condition based on lemma}
\end{equation}
Evaluating \eqref{eq: terminating condition based on lemma} would be computationally costly since it would involve processing all the queries for every input $x^{\textrm{obs}}$. We employ a more practically amenable criteria
\begin{equation}
q_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1} = q_{STOP} \quad \textrm{ if } \quad \max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) \leq \epsilon' \; \forall m \in \{L, L + 1, ..., L + T\}
\end{equation}
$T > 0$ is a hyper-parameter chosen via cross-validation.
\subsection{Proof of Lemma \ref{lemma: termination_criterion}}
\label{Appendix: proof of terminating criteria}
\begin{proof}
Recall each query $q$ partitions the set ${\mathcal{X}}$ and $S^{\textrm{IP}}_L(x^{\textrm{obs}})$ is the event $\{x' \in {\mathcal{X}} \mid \{q_i, {q_i}(x^{\textrm{obs}})\}_{1:L} = \{q_i, {q_i}(x')\}_{1:L}\}$. It is easy to see that if $S^{\textrm{IP}}_L(x)$ satisfies the condition specified by \eqref{eq: terminating criteria} then
\begin{equation}
P(y \mid S^{\textrm{IP}}_m(x^{\textrm{obs}})) = P(y \mid x') \ \forall x' \in S^{\textrm{IP}}_m(x^{\textrm{obs}}) \ \forall m \geq L, \ \forall q \in Q
\label{eq: equivalence class partitioning}
\end{equation}
This is because subsequent query-answers partition a set in which all the data points have the same posterior distributions. Now, $\forall q \in Q, \ \forall a \in Range(q), \ y \in {\mathcal{Y}}$
\begin{equation}
\begin{aligned}
p({q}(X) = a, y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) = p({q}(X) = a \mid S^{\textrm{IP}}_m(x^{\textrm{obs}})) p(y \mid {q}(X) = a, S^{\textrm{IP}}_m(x^{\textrm{obs}}))
\end{aligned}
\label{eq: chain rule of prob}
\end{equation}
\eqref{eq: chain rule of prob} is just an application of the chain rule of probability. The randomness in $q(X)$ is entirely due to the randomness in $X$. For any $a \in Range(q)$, $y \in {\mathcal{Y}}$
\begin{equation}
\begin{aligned}
p(y \mid {q}(X) = a, S^{\textrm{IP}}_m(x^{\textrm{obs}})) &= \sum_{x'} p(y, X = x' \mid a, S^{\textrm{IP}}_m(x^{\textrm{obs}})) \\
&= \sum_{x'} p(y \mid X = x', a, S^{\textrm{IP}}_m(x^{\textrm{obs}}))p(X = x' \mid a, S^{\textrm{IP}}_m(x^{\textrm{obs}})) \\
&= \sum_{x'} p(y \mid X = x')p(X = x' \mid a, S^{\textrm{IP}}_m(x^{\textrm{obs}})) \\
&= p(y \mid S^{\textrm{IP}}_m(x^{\textrm{obs}}))\sum_{x'} p(X = x' \mid a, S^{\textrm{IP}}_m(x^{\textrm{obs}})) \\
&= p(y \mid S^{\textrm{IP}}_m(x^{\textrm{obs}}))
\end{aligned}
\label{eq: y indep a given sufficient history}
\end{equation}
The first equality is an application of the law of total probability, third due to conditional independence of the history and the hypothesis given $X=x'$ (assumption) and the fourth by invoking (\eqref{eq: equivalence class partitioning}).
Substituting \eqref{eq: y indep a given sufficient history} in \eqref{eq: chain rule of prob} we obtain $Y \perp \!\!\! \perp q(X) \mid S^{\textrm{IP}}_m(x^{\textrm{obs}}))$ $\forall m \geq L, q \in Q$. This implies that for all subsequent queries $q_m$, $m > L$, $\max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}}))) = 0$. Hence, Proved.
\end{proof}
\subsection{Proof of Lemma \ref{lemma: termination_criterion general}}
\label{Appendix: proof of terminating criteria general}
\begin{proof}
\myparagraph{Condition \eqref{eq: stronger_requirement_general_epsilon} implies bounded KL divergence between inputs on which IP has identical query-answer trajectories} Recall $S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}))}(x^{\textrm{obs}}))$ is the event $\{x' \in {\mathcal{X}} \mid \{q_i, {q_i}(x^{\textrm{obs}}))\}_{1:t^{\textrm{IP}}(x^{\textrm{obs}}))} = \{q_i, {q_i}(x')\}_{1:t^{\textrm{IP}}(x^{\textrm{obs}}))}\}$. If $S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}))}(x^{\textrm{obs}}))$ satisfies \eqref{eq: stronger_requirement_general_epsilon} then using Pinsker's inequality we conclude,
\begin{equation}
\delta(p(Y \mid x^{\textrm{obs}})), p(Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}))}(x^{\textrm{obs}}))) \leq \sqrt{\frac{\epsilon}{2}}
\label{eq: pinsker 1}
\end{equation}
Here $\delta$ is the total variational distance between the two distributions. Since $\delta$ is a metric we conclude for any $x_1, x_2 \in S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}))}(x^{\textrm{obs}}))$,
\begin{equation}
\begin{aligned}
\delta(p(Y \mid x_1), p(Y \mid x_2)) &\leq \delta(p(Y \mid x_1), p(Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})) + \delta(p(Y \mid x_2), p(Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})) \\
&\leq \sqrt{2\epsilon}
\end{aligned}
\label{eq: metric tv}
\end{equation}
Since ${\mathcal{Y}}$ is countable, define $\eta = \min\{p(y \mid \hat{x}): y \in {\mathcal{Y}}, p(y \mid \hat{x})) > 0, \hat{x} \in {\mathcal{X}}\}$. Then, by the reverse Pinsker's inequality we conclude,
\begin{equation}
KL(p(Y \mid x_1), p(Y \mid x_2))) \leq \frac{\epsilon}{\eta} =: \epsilon' \quad \forall x_1, x_2 \in S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})
\label{eq: revere pinsker}
\end{equation}
Note, the above upper bound holds since by assumption 3, $p(Y \mid x_1)$ and $p(Y \mid x_2))$ have the same support.
\myparagraph{Bounded KL divergence between inputs implies subsequent queries have mutual information bounded by $\boldsymbol{\epsilon}$}
For any subsequent query $q \in Q$ that IP asks about input $x^{\textrm{obs}}$ we have $\forall x \in S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})$,
\begin{equation}
KL\left(p(Y \mid x), p(Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}}))\right) = \sum_Y p(Y \mid x) \log p(Y \mid x) - \sum_Y p(Y \mid x) \log p(Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}}))
\label{eq. KL espression for statistic}
\end{equation}
where, $S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}}) := S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}}) \cap \{x' \in {\mathcal{X}}: q(x') = q(x)\}$. For brevity, we denote $S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}})$ as $\textrm{B}$,
\begin{equation}
\begin{aligned}
\sum_Y p(Y \mid x) \log p(Y \mid \textrm{B}) &= \sum_Y p(Y \mid x) \log \left[\sum_{x' \in \textrm{B}} p(Y \mid x', \textrm{B})P(x' \mid \textrm{B}) \right] \\
&= \sum_Y p(Y \mid x) \log \left[\sum_{x' \in \textrm{B}} p(Y \mid x')P(x' \mid \textrm{B}) \right] \\
&\geq \sum_Y p(Y \mid x) \sum_{x' \in \textrm{B}} p(x' \mid \textrm{B}) \log p(Y \mid x') \\
&= \sum_{x' \in \textrm{B}} p(x' \mid \textrm{B}) \sum_Y p(Y \mid x) \log p(Y \mid x')
\label{eq. second term of KL espression for statistic}
\end{aligned}
\end{equation}
In the third inequality Jensen's inequality was used.
Substituting \eqref{eq. second term of KL espression for statistic} in \eqref{eq. KL espression for statistic},
\begin{equation}
\begin{aligned}
KL\left(p(Y \mid x) || p(Y \mid \textrm{B})\right) &\leq \sum_{x' \in \textrm{B}} p(x' \mid \textrm{B}) \sum_Y p(Y \mid x) \log \frac{p(Y \mid x)}{ p(Y \mid x')} \\
&\leq \epsilon' \sum_{x'} p(x' \mid \textrm{B}) \\
&= \epsilon'
\label{eq: bounding KL divergence}
\end{aligned}
\end{equation}
In the second inequality we substituted from \eqref{eq: revere pinsker} since $x, x' \in \textrm{B} \subseteq S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}})}(x^{\textrm{obs}})$. In the third equality we used the identity $\sum_{x' \in \textrm{B}} p(x' \mid \textrm{B}) = 1$,
It is easy to see that \eqref{eq: bounding KL divergence} holds for all $x' \in S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}})$ and thus $I(X; Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}})) \leq \epsilon$
Since $Y \rightarrow X \rightarrow q(X)$ we can apply the data-processing inequality to obtain,
\[I(q(X); Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}})) \leq I(X; Y \mid S^{\textrm{IP}}_{t^{\textrm{IP}}(x^{\textrm{obs}}) + 1}(x^{\textrm{obs}})) \leq \epsilon' \quad \forall q \in Q.\]
This implies that for all subsequent queries $q_m$, $m > t^{\textrm{IP}}(x^{\textrm{obs}})$, $\max_{q \in Q} I({q}(X); Y|S^{\textrm{IP}}_m(x^{\textrm{obs}})) \leq \epsilon$. Hence, Proved.
\end{proof}
\subsection{Computing Mutual Information for \eqref{def. IP algorithm}}
\label{Computing Mutual Information for IP}
\subsubsection{Derivation for \eqref{eq: compute intermediate conditionals}}.
\begin{equation}
\begin{aligned}
p({q}(x), y \mid S^{\textrm{IP}}_k(x))
&= \int p({q}(x), z, y \mid S^{\textrm{IP}}_k(x))dz \\
&= \int p({q}(x) \mid z, y, S^{\textrm{IP}}_k(x))p(z \mid y, S^{\textrm{IP}}_k(x))p(y \mid S^{\textrm{IP}}_k(x))dz.\\
&= \int p({q}(x) \mid z, y)p(z \mid y, S^{\textrm{IP}}_k(x))p(y \mid S^{\textrm{IP}}_k(x))dz.
\end{aligned}
\end{equation}
The first equality is an application of the law of total probability. The last equality appeals to the assumption that $\{q(X), q \in Q\}$ are conditionally independent given $Y, Z$ \eqref{eq: factorization of query answers}.
\subsubsection{Recursive estimation of $p(y \mid S^{\textrm{IP}}_k(x))$}
Finally, $p(y \mid S^{\textrm{IP}}_k(x))$ (required for \eqref{eq: compute intermediate conditionals}) is computed recursively via the Bayes' theorem.
\begin{equation}
\label{eq: posterior y}
\begin{aligned}
p(y \mid S^{\textrm{IP}}_k(x)) &\propto p(y, S^{\textrm{IP}}_k(x)) \\ &= p({q_k}(x), y, S^{\textrm{IP}}_{k-1}(x)) \\ &\propto p({q_k}(x) \mid y, S^{\textrm{IP}}_{k-1}(x))p(y \mid S^{\textrm{IP}}_{k-1}(x))
\end{aligned}
\end{equation}
$S^{\textrm{IP}}_0(x) = \emptyset$ (since no evidence via queries has been gathered from $x$ yet) and so $p(y \mid S^{\textrm{IP}}_0(x)) = p_Y(y)$. $p(y \mid S^{\textrm{IP}}_k(x))$ is obtained by normalizing the last equation in \eqref{eq: posterior y} such that $\sum_y p(y|S^{\textrm{IP}}_k(x)) = 1$. $p({q_k}(x) \mid y, S^{\textrm{IP}}_{k-1}(x))$ can be estimated using \eqref{eq: compute intermediate conditionals}.
\subsection{Complexity of the Information Pursuit Algorithm}
\label{Appendix: complexity analysis}
For any given input $x$, the per-iteration cost of the IP algorithm is $\mathcal{O}(N + |Q|m)$, where $|Q|$ is the total number of queries, $N$ is the number of ULA iterations, and $m$ is cardinality of the product sample space $q(X) \times Y$. For simplicity we assume that the output hypothesis $Y$ and query-answers $q(X)$ are finite-valued and also that the number of values query answers can take is the same but our framework can handle more general cases.
More specifically, to compute $q_{k+1} = \argmax_{q \in Q} I({q}(X); Y \mid S^{\textrm{IP}}_{k}(x^{\textrm{obs}}))$. We first run ULA for $N$ iterations to get samples from $p(z \mid y, S^{\textrm{IP}}_k(x))$ which are then used to estimate the distribution $p(q(x), y \mid S^{\textrm{IP}}_k(x))$ (using \eqref{eq: compute intermediate conditionals}) for every query $q \in Q$ and every possible query-answer hypothesis, $(q(x), y)$, pair. This incurs a cost of $\mathcal{O}(N + |Q|m)$. We then numerically compute the mutual information between query-answer $q(X)$ and $Y$ given history for every $q \in Q$ as described in \eqref{eq: numerical MI}. This has a computational complexity $\mathcal{O}(|Q|m)$. Finally, we search over all $q \in Q$ to find the query with maximum mutual information (refer \eqref{def. IP algorithm}).
It is possible to reduce $N$ by using advanced MCMC sampling methods which converge faster to the required distribution $p(z \mid y, S^{\textrm{IP}}_k(x)))$. The linear cost of searching of all queries can also be reduced by making further assumptions about the structure of the query set. For example, we conjecture that this cost can be reduced to $\log |Q|$ using hierarchical query sets where answers to queries would depend upon to answers to queries higher up in the hierarchy. Note that if the query answers $q(X)$ and $Y$ are continuous random variables then we would need to resort to sampling to construct stochastic estimates of the mutual information between $q(X)$ and $Y$ instead of carrying our explicit numerical computations. We would explore these directions in future work.
\subsection{Network Architectures and Training Procedure}
\label{Appendix: Network architectures and training}
Here we describe the architectures and training procedures for the $\beta$-VAEs used to calculate mutual information as described in Section \ref{section: Information Pursuit}.
\subsubsection{Architectures}
\begin{figure}[H]
\centering
\includegraphics[scale=0.4]{figs/IP_VAE_Architectures-Binary_Images.pdf}
\caption{\textit{Binary Image VAE}}
\end{figure}
\myparagraph{Binary Image Classification} In the encoder, all convolutional layers use kernel size 3 and stride 1, and max pool layers use a pooling window of size 2. In the decoder all transposed convolutions use kernel size 3 and stride 1 unless otherwise noted. In both the encoder and decoder, all non-pooling layers are followed by a BatchNorm layer and LeakyReLU activation (with slope -0.3) except for the final encoder layer (no nonlinearities) and the final decoder layer (sigmoid activation).
\begin{figure}[H]
\centering
\begin{tabular}{c@{\hskip 1cm}c}
\includegraphics[width=0.45\textwidth]{figs/IP_VAE_Architectures-CUB.pdf} & \includegraphics[width=0.45\textwidth]{figs/IP_VAE_Architectures-HuffPost_News.pdf} \\
\small \textit{(a) CUB Concepts VAE} & \small \textit{(b) HuffPost News Headlines VAE}
\end{tabular}
\caption{\textit{CUB Concepts and HuffPost News Headlines VAEs}}
\end{figure}
\myparagraph{CUB Bird Species Classification and HuffPost News Headline Classification} For CUB and HuffPost News, we use essentially the same VAE architecture, only designed to handle different-sized inputs $X$ and one-hot labels $Y$. All layers are followed by a BatchNorm layer and ReLU activation except for the final encoder layer (no nonlinearities) and the final decoder layer (sigmoid activation).
\subsubsection{Training}
The $\beta$-VAE was trained by optimizing the Evidence Lower BOund (ELBO) objective
\begin{equation}
\label{eq:ELBO}
\begin{aligned}
\max_{\omega, \phi} \text{ELBO}(\omega, \phi) = \sum_{i=1}^n \left[ \mathbb{E}_{p_{\phi}(z \mid y^{(i)}, x^{(i)})}[\log p_{\omega}(x^{(i)} \mid z, y^{(i)})] - \beta D_{KL}(p_{\phi}(z \mid y^{(i)},
x^{(i)}) \| p(z)) \right]
\end{aligned}
\end{equation}
where $p_{\phi}(z \mid y^{(i)}, x^{(i)})$ denotes the encoder and $p_{\omega}(x^{(i)} \mid z, y^{(i)})$ the decoder. The prior over latents $p(z)$ is taken to be standard Gaussian.
For all binary image datasets, we trained our VAE for 200 epochs using Adam with learning rate $0.001$ and the $\beta$-VAE parameter $\beta = 5.0$. We also trained the CUB VAE for 200 epochs using Adam with learning rate $0.001$ but with $\beta = 1.0$. Finally, we trained the HuffPost News VAE for 100 epochs using Adam with learning rate $0.0002$ and $\beta = 1.0$.
\section{}
\subsection{HuffPost News Task and Query Set} \label{Appendix: Tasks and Query Sets}
The Huffington Post News Category dataset consists of 200,853 articles published by the Huffington Post between 2012 and 2018. Each datapoint contains the article ``headline'', a ``short description" (a one to two-sentence-long continuation of the headline), and a label for the category/section it was published under in the newspaper. We concatenate the article headline and short description to form one extended headline. Additionally, many of the 41 category labels are redundant (due to changes in how newspaper sections were named over the years), are semantically ambiguous and HuffPost-specific (e.g. ``Impact'', ``Fifty'', ``Worldpost''), or have very few articles. Therefore, we combine category labels for sections with equivalent names (e.g. ``Arts \& Culture'' with ``Culture \& Arts''), remove ambiguous HuffPost-specific categories, and then keep only the 10 most frequent categories to ensure that each category has an adequate number of samples. This leaves us with a final dataset of size 132,508 and category labels ``Entertainment'', ``Politics'', ``Queer Voices'', ``Business'', ``Travel'', ``Parenting'', ``Style \& Beauty'', ``Food \& Drink'', ``Home \& Living'', and ``Wellness''.
\subsection{Comparison Models} \label{Appendix: Comparison Models}
\subsubsection{MAP using $Q$} For IP, we use ULA to sample from $p(z \mid y, S^{\textrm{IP}}_k(x))$, but now that we have access to the all the query answers, $Q(x)$, we can improve performance by making use of the VAE's encoder instead. Following equation \ref{eq: marginalize z}, we draw many samples from the encoder $p(z \mid Q(x), y)$ and then decode these samples to estimate the VAE's posterior distribution $p(y | Q(x))$. For each problem, we set the number of samples to be the same as what we draw during each iteration of IP (12,000 for binary image tasks, 12,000 for CUB, 10,000 for HuffPost News). We expect the accuracy of this model to serve as an upper bound for what we can achieve with IP given that it uses all queries. Our task-specific VAE architectures can be found in Appendix \ref{Appendix: Network architectures and training}.
\subsubsection{Black-Box Using $Q$} We also compare to non-interpretable supervised models which receive all queries $Q(x)$ as input and try to predict the associated label $y$. This allows a comparison between the accuracy of the posterior of our generative model and traditional supervised approaches on the chosen interpretable query set.
For the binary image datasets, we use a simple CNN where all convolutional layers use kernel size $3 \times 3$ and a stride of 1, all max pooling uses a kernel size of 2. For CUB and HuffPost News we use simple MLPs.
\begin{table}[H]
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\caption{\textit{Binary Image CNN Architecture}}
\begin{tabular}{| l | l | l | l |}
\hline
Layer & Input Size/Channels & Output Size/Channels & Nonlinearity \\
\hline
Convolution & 3 & 32 & BatchNorm + ReLU \\
Convolution & 32 & 64 & BatchNorm + ReLU + MaxPool \\
Convolution & 64 & 128 & BatchNorm + ReLU \\
Convolution & 128 & 256 & BatchNorm + ReLU + MaxPool + Global Avg Pool \\
Fully-connected & 256 & 2048 & BatchNorm + ReLU \\
Fully-connected & 2048 & 10 & Sigmoid \\
\hline
\end{tabular}
\endgroup
\end{table}
\begin{table}[H]
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\caption{\textit{CUB Attributes MLP Architecture}}
\begin{tabular}{| l | l | l | l |}
\hline
Layer & Input Size/Channels & Output Size/Channels & Nonlinearity \\
\hline
Fully-connected & 312 & 100 & ReLU \\
Fully-connected & 100 & 25 & ReLU \\
Fully-connected & 25 & 200 & Sigmoid \\
\hline
\end{tabular}
\endgroup
\end{table}
\begin{table}[H]
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\caption{\textit{HuffPost Bag-of-Words MLP Architecture}}
\begin{tabular}{| l | l | l | l |}
\hline
Layer & Input Size/Channels & Output Size/Channels & Nonlinearity \\
\hline
Fully-connected & 1000 & 100 & ReLU \\
Fully-connected & 100 & 25 & ReLU \\
Fully-connected & 25 & 10 & Sigmoid \\
\hline
\end{tabular}
\endgroup
\end{table}
\subsubsection{Black-Box} \label{Appendix: Black-Box} Since we ourselves pre-processed the cleaned 10-class version of HuffPost News, there are no reported accuracies in the literature to compare IP with. Therefore, as a strong black-box baseline, we fine-tune a pre-trained Bert Large Uncased Transformer model \cite{devlin2018bert} with an additional dropout layer (with dropout probability 0.3) and randomly initialized fully-connected layer. Our implementation is publicly available at https://www.kaggle.com/code/stewyslocum/news-classification-using-bert.
\newpage
\subsection{Additional Example Runs} \label{Appendix: Additional Example Runs}
\subsubsection{IP with Various Patch Scales} \label{Appendix: IP Patch Scales}
For binary image classification, we also experimented with patch queries of sizes other than $3 \times 3$, from single pixel $1 \times 1$ queries up to $4 \times 4$ patches.
\begin{figure}[H]
\centering
\includegraphics[width=0.7\textwidth]{figs/IP_in_action_MNIST_extra_v1.pdf}
\caption{\textbf{IP on MNIST with different sized patch queries.} In each subfigure, the top row displays the test image with red boxes denoting the current queried patch and blue boxes denoting previous patches. The second row shows the revealed portion of the image that IP gets to use at each query. The final row shows the model's estimated posteriors at each query. For conciseness in (a), we only display the 8 iterations of IP with highest KL divergence between successive posteriors (i.e. the most influential iterations). Observe that at all patch sizes, the queries chosen by IP cover roughly the same parts of the image, illustrating the importance of this region during classification. Reaching the stopping criteria of $99\%$ posterior confidence takes 20 queries with $1\times1$ patches, 7 queries with $2\times2$ patches, 4 queries with $3\times3$ patches, and 4 queries with $4\times4$ patches.}
\label{fig: effect of scale}
\end{figure}
\begin{table}[H]
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\caption{\textit{Number of queries and pixels gleaned by IP (until termination) using query sets of different patch scales on MNIST.}}
\begin{tabular}{| l | l | l | l | l |}
\hline
Patch Size & $1\times1$ & $2\times2$ & $3\times3$ & $4\times4$ \\
\hline
\# Queries & 21.1 & 9.6 & 5.2 & 4.6 \\
\# Pixels & 21.1 & 32.8 & 44.7 & 54.7 \\
\% Pixels & 2.7 & 4.2 & 5.7 & 6.9 \\
\hline
\end{tabular}
\endgroup
\end{table}
While $1\times1$ patch queries use the smallest total number of pixels and most similarly resemble existing post-hoc attribution maps, they lead to a large number of scattered queries that are hard to interpret individually. On the other extreme, as the patch size grows larger, the number of total queries decreases, but queries becomes harder to interpret since each patch would contain many image features. On the MNIST dataset, we found $3\times3$ patches to be a sweet spot where explanations tended to be very short, but were also at a level of granularity where each patch could be individually interpreted as a single edge or stroke. Remarkably, at all chosen patch scales, only a very small fraction (2-7\%) of the image needs to be revealed to classify the images with high confidence.
\newpage
\subsubsection{IP for Binary Image Classification}
Now we provide additional example runs of IP for each of the three binary image classification datasets.
As in Figure \ref{Fig: IP in action} in the main paper, in each plot, the top row displays the test image with red boxes denoting the current queried patch and blue boxes denoting previous patches. The second row shows the revealed portion of the image that IP gets to use at each query. The final row shows the model's estimated posteriors at each query.
\begin{figure}[H]
\begin{center}\includegraphics[height=4.8cm]{figs/IP_in_action_MNIST1.pdf}\end{center}
\textit{(a) Each query reveals the patch with maximum mutual information with $Y$, conditioned on query history. This is initially independent of the particular image and asks for the pixel intensities in the center patch (see k = 1 in row 1). After the first query reveals that the center patch is all black, the posterior concentrates on ``0'' and ``7''. After observing a white corner in the bottom left (which would be black for a ``7''), the model becomes confident that the image is a ``0'', and even more so after one final query when it reaches termination.}
\begin{center}\includegraphics[height=4.8cm]{figs/IP_in_action_MNIST2.pdf}\end{center}
\textit{(b) The first query reveals a vertical white stroke in the center of the image, leading to a concentration of the posterior on ``1''. In the next two queries, IP determines that there is a single long vertical stroke center stroke taking up the entire height of the image, and so it reaches a 99\% confidence of the image being a ``1''.}
\begin{center}\includegraphics[height=4.8cm]{figs/IP_in_action_MNIST3.pdf}\end{center}
\textit{(c) As in the top left example, the first query is all black, and the posterior mass shifts onto ``0''. Because the answer to this first query is identical as in the top left example, the second query chosen, in the bottom left of the image, is also the same. The response to this query is also a white corner as in the top left example, and so the posterior continues to concentrate on ``0'', and the third query is also in the same left area of the image. However, this third query reveals a black patch, indicating that the image might be a ``5'' instead of a ``0''. In the remaining four queries, IP discovers other portions of the ``5'' digit and finally arrives at the right answer with high confidence.}
\caption{\textit{Additional Examples of IP on MNIST}}
\end{figure}
\newpage
Recall that in order to improve performance on the KMNIST and FashionMNIST datasets (at the expense of asking a few more queries) we modified IP's termination criteria to include a stability condition: terminate when the original criterion ($\max_Y p(Y | S^{\textrm{IP}}_k(x)) \geq 0.99$) is true for 5 queries in a row.
\begin{figure}[H]
\begin{center}\includegraphics[width=\textwidth]{figs/KMNIST_classes.pdf}\end{center}
\textit{(a) KMNIST is a 10-class dataset of handwritten Japanese Hiragana characters. To assist the reader in understanding the examples below, we display each typed character along with its romanized name.}
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_KMNIST1.pdf}\end{center}
\textit{(b) For each image, IP selects the first query to be in the middle right of the image, where several characters are likely to have a stroke. Upon finding none, IP rules out ``o'', ``tsu'', and ``ya'' but otherwise distributes probability mass rather equally. On the second query, IP discovers a closed loop in the bottom of the character, a clear sign of ``na'' and ``ma'', which increase the most. The discovery of the double crosses in the remaining queries concentrate the posterior on ``ma'' until termination. For conciseness, we only display the 8 iterations of IP with highest KL divergence between successive posteriors (i.e. the most influential iterations).}
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_KMNIST2.pdf}\end{center}
\textit{(c) In the first query, IP discovers a left edge, hinting at the presence of a large loop on the right side of the image, features of ``o'' and ``tsu''. The second query is likely intended to disambiguate these two characters as ``o'' contains white strokes in this region. Upon finding a black patch here, IP is already confident, and reaches 99\% confidence in just four queries.}
\caption{\textit{Additional Examples of IP on KMNIST}}
\end{figure}%
\newpage
\begin{figure}[ht]\ContinuedFloat
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_KMNIST3.pdf}\end{center}
\textit{(d) The first query reveals a right edge, suggesting a slightly smaller loop on the right side of the image, a feature shared by several characters. However, most of these queries have a busy center region except for ``tsu'', whose probability increases after the second query reveals a black patch. The next two queries outline the shape of the loop which is very large, a distinctive characteristic of ``tsu''.}
\caption{\textit{Additional Examples of IP on KMNIST}}
\end{figure}
\newpage
\begin{figure}[H]
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_FashionMNIST1.pdf}\end{center}
\textit{(a) On this dataset, IP always selects the first query to be the patch in the top center, which being all black in this case rules out the possibility of the object being a type of pant or upper body garment, which would take up the entire height of the image. Over the next several queries, IP focuses on queries that would allow it to distinguish types of shoes from each other, in particular finding the shoe to have a high top and a small heel, eventually causing the posterior to concentrate on the correct ``Ankle Boot" category. For conciseness, we only display the 8 iterations of IP with highest KL divergence between successive posteriors (i.e. the most influential iterations).}
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_FashionMNIST2.pdf}\end{center}
\textit{(b) The first query detects a white corner in the top center of the image, which hints at the presence of a collar, causing the posterior mass to move to the ``Coat'' and ``Shirt'' categories. Determining between these two categories is relatively difficult however, especially with binary images. But in general, coats tend to be bulkier than shirts. Therefore, after finding a white patch in iteration $k=2$, the probability of ``Coat'' slightly increases, but as more partially black queries along the outside of the shirt are revealed and the slimmer outline of the shirt comes into view, the posterior converges on the correct category of ``Shirt''.}
\begin{center}\includegraphics[height=6cm]{figs/IP_in_action_FashionMNIST3.pdf}\end{center}
\textit{(c) Detecting an all white patch in the top center, IP rules out the possibility of the image being some type of shoe. The second query in the lower left returns a black patch, suggesting that the image is also not an upper body garment, which would take up the width of the image. In query $k=3$, IP queries the center bottom of the image, discovering the space in between the two legs of the trouser, causing the posterior probability of ``Trouser'' to jump to nearly 100\%, which remains stable over the last few queries.}
\caption{\textit{Additional Examples of IP on FashionMNIST}}
\end{figure}
\newpage
\subsubsection{IP for Bird Species Classification}
As in the main text, for each plot, on the left, we show the input image and on the right we have a heatmap of the estimated class probabilities per iteration. For readability, we only show the top 10 most probable classes out of the 200. To the right, we display the queries asked at each iteration, with red indicating a ``no'' response and green a ``yes'' response.
\begin{figure}[H]
\begin{center}\includegraphics[height=5cm]{figs/IP_in_action_CUB1_done.pdf}\end{center}
\textit{(a) The first few queries narrow down the potential species from 200 down to a small number. Since we only show the 10 most probable classes, the first few queries increase the probability of all shown classes. The first queries that distinguish among these classes concern bill length (which rules out the short-billed Nighthawk) and size (which rule out the smaller Mangrove Cuckoo, Geococcyx, Kingfisher, Northern Flicker, and Brown Creeper). The remaining birds are quite similar, all medium-sized brown water birds. The next two color queries suggest that the bird in question is a Black-footed Albatross, which is confirmed by the answers to the next few queries, which all match up with the characteristics of that bird.}
\begin{center}\includegraphics[height=5cm]{figs/IP_in_action_CUB2_done.pdf}\end{center}
\textit{(b) Again, the first few queries narrow down the likely species into the top 10 displayed classes. In just two queries (bill length and bill shape), IP distinguishes among these similar-looking, black, plain-headed birds that are hard for non-expert humans to differentiate between. Again, the last few queries serve to confirm the posterior prediction that the bird is a Groove-billed Ani.}
\begin{center}\includegraphics[height=5cm]{figs/IP_in_action_CUB3_done.pdf}\end{center}
\textit{(c) After establishing the top few most probable classes, IP converges on the class Ringed Kingfisher after just 7 queries. The last four queries simply serve to increase its confidence in its prediction.}
\caption{\textit{Additional Examples of IP on CUB Bird Species Identification}}
\end{figure}
\subsubsection{IP for HuffPost News Headline Category Classification}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figs/IP_in_action_HuffPost_done_all.pdf}
\caption{\textbf{Additional Examples of IP on HuffPost News Headline Classification.} As before, when more than 20 queries were asked, we only display the 20 queries that led to greatest KL divergence between successive posteriors.
\textbf{(a)} Because of the sparse structure of natural language, it typically takes a significant number of queries before the first word that is present in the sentence is found. Until this point, no query is particularly informative, and the posterior distribution remains mostly unchanged from the prior. However, at query 60, IP asks the word ``trailer'', which is present in the extended headline. Naturally, the posterior shifts heavily towards ``Entertainment'', and a few queries later IP reaches its termination criteria. Analyzing this run, we can say that IP reached its decision primarily because of the presence of the word ``trailer'', leading us to say that this is a reasonable and trustworthy prediction.
\textbf{(b)} This is an example of a relatively short explanation as IP happens to discover words present in the sentence after just two queries. Initially, the presence of ``photo'' causes the categories ``Travel'', ``Home \& Living'', and ``Style \& Beauty'' to become more probable. Several words later however, the word ``style'' is found, which is very strongly associated with the ``Style \& Beauty'' category.
\textbf{(c)} The posterior remains mostly unchanged until IP discovers the word ``life'', which reasonably, shifts probability mass onto the ``Wellness'' category. However, several queries later, ``star'' is found to be present, which shifts the posterior away from ``Wellness'' onto ``Entertainment'' and ``Queer Voices''. After discovering several more words that are present, ``race'', ``change'', ``word'', ``reflect'', the posterior progressively converges on ``Queer Voices'', but still with relatively high uncertainty, likely because it never came across identifying words such as ``drag'', which was not present in the vocabulary.}
\end{figure}
\subsection{Connections to information theory}
\subsubsection{Binary Image Classification with Patch Queries}
\myparagraph{Task and query set.} We start with the simple task of binary image classification. We consider three popular datasets -- MNIST \cite{lecun1998gradient}, Fashion-MNIST\cite{xiao2017fashion} and KMNIST \cite{clanuwat2018deep}. We choose a threshold for binarizing these datasets since they are originally grayscale.
We choose the query set $Q$ as the set of all $w \times w$ overlapping patch locations in the image. The answer $q(X)$ for any $q \in Q$ is the $w^2$ pixel intensities observed at the patch indexed by location $q$. This choice of $Q$ reflects the user's desire to know which parts of the input image are most informative for a particular prediction, a common practice for explainability in vision tasks \cite{chen2019looks}. We conduct experiments for multiple values of $w$ and conclude that $w=3$ provides a good trade-off between the required number of queries and the interpretability of each query. Note that when $w>1$ the
factorization in \eqref{eq: factorization of query answers} that we use to model $p(Q(x), y, z)$ and compute mutual information no longer holds as the overlapping queries $q(X)$ are now \textit{causally related} (and therefore dependent even when conditioned on $Z$, making them unable to be modeled by a VAE). So instead of training a VAE to directly model the query set $p(Q(x) \mid y, z)$, we train a VAE to model the pixel distribution $p(x \mid y, z)$, and then compute the probability distribution over the patch query $p(q(x) \mid z, y)$ as the product of the probabilities of all pixels in that patch
\myparagraph{IP in action.}
Fig.~\ref{Fig: IP in action}(a) illustrates the decision-making process of IP using $3\times 3$ patch queries on an image $x^{\textrm{obs}}$ of a $6$ from the MNIST test set. The first query is near the center of the image; recall from \eqref{def. IP algorithm} that this choice is independent of the particular input image and represents the patch whose pixel intensities have maximum mutual information with $Y$ (the class label). The updated posterior, $p\left(Y \mid S^{\textrm{IP}}_1(x^{\textrm{obs}})\right)$, concentrates most of its mass on the digit ``1'', perhaps because most of the other digits do not commonly have a vertical piece of stroke at the center patch. However, the next query (about three pixels below the center patch) reveals a horizontal stroke and the posterior mass over the labels immediately shifts to $\{2, 3, 6, 8\}$. The next two queries are well-suited to discerning between these four possibilities and we see that after asking $4$ questions, IP is more than $90\%$ confident that the image is a $6$. Such rich explanations in terms of querying informative patches based on what is observed so far and seeing how the belief $p\left(Y \mid S^{\textrm{IP}}_k(x^{\textrm{obs}})\right)$ of the model evolves over time is missing from post-hoc attribution methods which output static importance scores for every pixel towards the black-box model's final prediction.
\myparagraph{Explanation length vs. task complexity.} Fig. \ref{fig:accuracy_minimality_plot} shows that IP requires an average of $5.2$, $12.9$ and $14.5$ queries of size $3\times 3$ to predict the label with $99\%$ confidence ($\epsilon = 0.01$ in \eqref{eq: practical terminating condition for deterministic cases}) on MNIST, KMNIST and FashionMNIST, respectively. This reflects the intuition that more complex tasks require longer explanations. For reference, state-of-the-art deep networks on these datasets obtain test accuracies in order MNIST $\geq$ KMNIST $\geq$ FashionMNIST (see last row in Table \ref{fig: results table}).
\myparagraph{Effect of patch size on interpretability.}
We also run IP on MNIST with patch sizes of $1\times1$ (single pixels), $2\times2$, $3\times3$, and $4\times4$. We observed that IP terminates at 99\% confidence after $21.1$, $9.6$, $5.2$, and $4.6$ queries on average, respectively. While this suggests that larger patches lead to shorter explanations, we note that explanations with larger patches use more pixels (e.g. on MNIST, IP uses $21.1$ pixels on average for $1\times 1$ patches and $54.7$ pixels on average for $4\times4$ patches). That being said, very small patch queries are hard to interpret (see Fig. \ref{fig:IP_in_action_MNIST_1x1}) and very large patch queries are also hard to interpret since each patch contains many image features. Overall, we found that $3 \times 3$ patches represented the right trade-off between interpretability in terms of edge patterns and minimality of the explanations. Specifically, single pixels are not very interpretable to humans but the explanations generated are more efficient in terms of \textit{number of pixels needed to predict the label}. On the other extreme, using the entire image as a query is not interesting from an interpretability point of view since it does not help us understand which parts of the image are salient for prediction. We refer the reader to Appendix \ref{Appendix: IP Patch Scales} for additional patch size examples and quantitative analysis.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figs/IP_in_action_MNIST0_1x1.pdf}
\caption{\textbf{IP with $1 \times 1$ patches on MNIST.} Through the first 6 iterations, IP asks queries in the same center vertical region as in Fig. \ref{Fig: IP in action}(a) (which uses $3\times3$ queries), outlining the distinctive loop in the bottom of the ``6''. However, reaching 99\% confidence requires a total of 21 $1\times1$ queries as opposed to just 4 $3\times3$ ones. For conciseness, we show only the 6 queries that led to the greatest KL divergence between successive posterior class probabilities.}
\label{fig:IP_in_action_MNIST_1x1}
\end{figure}
\subsubsection{Concept-Based Queries}
\label{Sec: Concept-Based Queries Birds}
\myparagraph{Task and query set.} What if the end-user is interested in a more semantic explanation of the decision process in terms of high-level concepts? This can be easily incorporated into our framework by choosing an appropriate query set $Q$. As an example we consider the challenging task of bird species classification on the Caltech-UCSD Birds-200-2011 (CUB) dataset \cite{wah2011caltech}. The dataset contains images of 200 different species of birds. Each image is annotated with $312$ binary attributes representing high-level concepts, such as the colour and shape of the beak, wings, and head. Unfortunately, these attribute annotations are very noisy. We follow \cite{koh2020concept} in deciding attribute labels by majority voting. For example, if more than 50\% of images in a class have black wings, then we set all images in that class to have black wings. We construct $Q$ by choosing a query for asking the presence/absence of each of these $312$ binary attributes. Unfortunately, attribute annotations are not available at test time. To remedy this, we train a CNN (see \cite{koh2020concept} for details) to answer each query using the training set annotations, which is then used to answer queries at test time. Subsequently, we learn a VAE to model the joint distribution of query-answers supplied by this CNN (instead of the ground truth annotations) and $Y$, so our generative model can account for any estimation errors incurred by the CNN. Finally, we carry out IP as explained in \S\ref{section: Information Pursuit}.
\myparagraph{IP in action.} Consider the image of a \textit{Great Crested Flycatcher} in Fig.~\ref{Fig: IP in action}(b). IP proceeds by asking most informative queries about various bird attributes progressively making the posterior over the species labels more and more peaked. After $5$ queries, IP has gathered that the image is of a bird that has a perching-like shape, all-purpose beak and yellow belly, but does not have a yellow throat nor yellow upperparts. This results in a posterior concentrated on just $4$ species that exhibit these characteristics. IP then proceeds to discount \textit{Green Jay} and \textit{Scott Oriole} which have black breasts with query $6$. Likewise, \textit{Tropical Kingbirds} have grayish back and is segregated from \textit{Great Crested Flycatchers} which have buff-coloured backs with query $7$. Finally after $9$ queries, IP is $99\%$ confident about the current class. Such concept-based explanations are more accessible to non-experts, especially on fine-grained classification datasets, which typically require domain expertise. On average IP takes $14.7$ queries to classify a given bird image with $\epsilon = 0.007$ as the stopping criteria (See \eqref{eq: practical terminating condition}).
\subsubsection{Word-based Queries}
\label{Section: Word-based Queries}
\myparagraph{Task and query set.} Our framework can also be successfully applied to other domains like NLP. As an example we consider the task of topic identification from newspaper
extended headlines (headline + short description field) using the the Huffington Post News Category Dataset \cite{huffpostdataset}. We adopt a simple query set that consists of binary queries probing the existence of words in the extended headline. The words are chosen from a pre-defined vocabulary obtained by stemming all words in the HuffPost dataset and choosing the top-1,000 according to their tf-idf scores \cite{lavin2019analyzing}. We process the dataset to merge redundant categories (such as \textit{Style \& Beauty} and \textit{Beauty \& Style}), remove semantically ambiguous, HuffPost-specific categories (e.g. \textit{Impact} or \textit{Fifty}) and remove categories with few samples, arriving at 10 final categories (see Appendix \ref{Appendix: Tasks and Query Sets}).
\myparagraph{IP in action.} Fig. \ref{Fig: IP in action}(c) shows an example run of IP on the HuffPost dataset. Note that positive responses to queries are very sparse, since each extended headline only contains $8.6$ words on average out of the 1,000 in the vocabulary. As a result, IP asks $125$ queries before termination. As discussed in \S \ref{sec:related_work}, such long decision paths would be impossible in decision trees due to data fragmentation and memory limitations. For clarity of presentation we only show the $20$ queries with the greatest impact on the estimated posterior (as measured by KL-divergence from previous posterior). Upon reaching the first positive query ``eat'', the probability mass concentrates on the categories \textit{Food \& Drink} and \textit{Wellness} with little mass on \textit{Travel}. However, as the queries about the existence of ``citi'', ``visit'', ``york'', and ``bar'' in the extended headline come back positive, the model becomes more and more confident that ``Travel'' is the correct class. IP requires about $199.3$ queries on average to predict the topic of the extended headline with $\epsilon = 10^{-3}$ as the stopping criteria (See \eqref{eq: practical terminating condition}). Additional details on the HuffPost query set are in Appendix \ref{Appendix: Tasks and Query Sets}.
Further examples of IP performing inference on all tasks can be found in Appendix \ref{Appendix: Additional Example Runs}.
\subsection{Quantitative Evaluation}
\subsubsection{Classification Accuracy} \label{Section: Experiments.Models}
We compare the classification accuracy of our model's prediction based on the query-answers gathered by IP until termination with several other baseline models. For each of the models considered, we first give a brief description and then comment on their performance with respect to IP. All the results are summarized in Table \ref{fig: results table}.
\textsc{Decision Tree} refers to standard classification trees learnt using the popular CART algorithm \cite{breiman2017classification}. In the Introduction, we mentioned that classical decision trees learnt using $Q$ to supply the node splitting functions will be intepretable by construction but are not competitive with state-of-the-art methods. This is illustrated in our results in Table \ref{fig: results table}. Across all datasets, IP obtains superior performance since it is based on an underlying generative model (VAE) and only computes the branch of the tree traversed by the input data in an online manner, thus it is not shackled by data fragmentation and memory limitations.
\textsc{MAP using $Q$} refers to the Maximum A Posteriori estimate obtained using the posterior distribution over the labels given the answers to all the queries in $Q$ (for a given input). Recall, IP asks queries until the stopping criteria is reached (Equation \eqref{eq: practical terminating condition} \& Equation \eqref{eq: practical terminating condition for deterministic cases}). Naturally, there is a trade-off between the length of the explanations and the predictive performance observed. If we ask all the queries then the resulting explanations of length $|Q|$ might be too long to be desirable. The results for IP reported in Table \ref{fig: results table} use different dataset-specific stopping criteria according to the elbow in their respective accuracy vs. explanation length curves (see Fig.~\ref{fig:accuracy_minimality_plot}). On the binary image datasets, (MNIST, KMNIST, and FashionMNIST) IP obtains an accuracy within $~3\%$ of the best achievable upon seeing all the query-answers with only about $2\%$ of the total queries in $Q$. Similarly for the CUB and Huffpost datasets, IP achieves about the same accuracy as \textsc{MAP using $Q$} but asks less than $5\%$ and $20\%$ of total possible queries respectively.
\begin{figure*}
\centering
\includegraphics[width=0.8\linewidth]{figs/Acc_Minimality_Plot.pdf}
\caption{\textbf{Trade-off between predictive performance and explanation length} Different points along the curves correspond to different values of $\epsilon$ as the stopping criteria \eqref{eq: practical terminating condition} is varied. The colored dotted vertical line in each plot indicates the avg. explanation length v/s test accuracy at the $\epsilon$ value used as the stopping criteria for reporting results for the IP strategy in this work.}
\label{fig:accuracy_minimality_plot}
\end{figure*}
\textsc{Black-Box using $Q$} refers to the best performing deep network model we get by training on features supplied by evaluating all $q \in Q$ on input data from the various training datasets. For the binary image datasets, this is just a 4-layer CNN with ReLU activations. For CUB we use the results reported by the sequential model in \cite{koh2020concept}. For HuffPost, we found a single hidden layer with ReLU non-linearity give the best performance. Further architectural and training details are in Appendix \ref{Appendix: Comparison Models}.
In Table \ref{fig: results table} we show that across all datasets, the predictive performance obtained by \textsc{MAP Using $Q$} is on par with the best performance we obtained using black-box expressive non-interpretable networks \textsc{Black-Box using $Q$}. Thus, our generative models, which form the backbone for IP, are competitive with state-of-the-art prediction methods.
\textsc{Black-Box} refers to the best performing black-box model on these datasets in terms of classification accuracy as reported in literature; to the best of our knowledge. In Table \ref{fig: results table}, we see a performance gap in each dataset when compared with \textsc{MAP using $Q$} which uses an interpretable query set. This is expected since explainability can be viewed as an additional constraint on learning. For example, on FashionMNIST we see an almost $8.5\%$ relative fall in accuracy due to binarization. This is because it is harder to decipher between some classes like shirts and pullovers at the binary level. On the other hand, binary patches are easily interpretable as edges, foregrounds and backgrounds. Similarly, there is a relative drop of accuracy of about $17\%$ for the HuffPost dataset since our queries regarding the existence of different words ignore their arrangement in sentences. Thus we lose crucial contextual information used by state-of-the-art transformer models \cite{devlin2018bert}. Ideally, we would like query sets to be easily interpretable, lead to short explanations and be sufficient to solve the task. Finding such query sets is nontrivial and will be explored in future work.
\begin{table}[h]
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\caption{\textbf{Classification accuracy of our model (Information Pursuit) relative to baselines on different test sets.} See \ref{Section: Experiments.Models} for details on each model.}
\label{fig: results table}
\resizebox{\columnwidth}{!}{\begin{tabular}{| l | c | c | c | c | c |}
\hline
\textbf{Model} & \textbf{MNIST} & \textbf{KMNIST} & \textbf{Fashion} & \textbf{CUB} & \textbf{HuffPost} \\
\hline
\textsc{Information Pursuit} & 96.78\% & 91.02\% & 85.60\% & 76.73\% & 71.21\% \\
\textsc{Decision Tree}\cite{breiman2017classification} & 90.23\% & 78.00\% & 80.80\% & 68.80\% & 63.00\% \\
\hline
\textsc{MAP using $Q$} & 99.05\% & 94.25\% & 87.56\% & 76.80\% & 71.72\% \\
\textsc{Black-box using $Q$} & 99.15\% & 95.10\% & 88.43\% & 76.30\% & 71.48\% \\
\hline
\textsc{Black-box} & 99.83\%\cite{hu2018squeeze} & 98.83\% \cite{clanuwat2018deep} & 96.70\% \cite{fashion_github} & 82.70\% \cite{koh2020concept} & 86.45\% \tablefootnote{We fine-tuned a Bert Large Uncased Transformer model \cite{devlin2018bert} with the last layer replaced with a linear one. See Appendix \ref{Appendix: Black-Box} for details.} \\
\hline
\end{tabular}}
\endgroup
\end{table}
\subsubsection{Comparison to current attribution methods} \label{section: comparison to attribution methods}
At first glance, it might seem that using attribution methods/saliency maps can provide the same insights as to which parts of the image or more generally which queries in $Q$ were most influential in a decision made by a black-box model trained on input features supplied by all the query-answers. However, the unreliability of these methods in being faithful to the model they try to explain brings their utility into question \cite{adebayo2018sanity, yang2019benchmarking, shah2021input}. We conjecture that this is because current attribution methods are not designed to generate explanations that are sufficient statistics of the model's prediction. We illustrate this with a simple experiment using our binary image classification datasets.
For each input image $x$, we compute the corresponding attribution map $e(x)$ for the model's predicted class using two popular attribution methods, Integrated gradients (IG) \cite{sundararajan2017axiomatic} and DeepSHAP \cite{lundberg2017unified}. We then compute the $L$ most important $3 \times 3$ patches, where $L$ is the number of patches queried by IP for that particular input image. For computing the attribution/importance of a patch we average the attributions of all the pixels in that patch (following \cite{yang2019benchmarking}). We proceed as follows: {\it(i)} Given $e(x)$, compute the patch with maximum attribution and add these pixels to our explanation, {\it(ii)} Zero-out the attributions of all the pixels in the previously selected patch and repeat step {\it(i)} until $L$ patches are selected. The final explanation consists of $L$ possibly overlapping patches. Now, we evaluate the sufficiency of the generated explanation for the model's prediction by estimating the MAP accuracy of the posterior over labels given the intensities in the patches included in this explanation. This is done via a VAE trained to learn the joint distribution over image pixels and class labels. We experiment with both the raw attribution scores returned by IG and DeepSHAP and also the absolute values of the attribution scores for $e(x)$. The results are reported in Table \ref{Table: comparison_to_current_attrib_methods}. In almost all cases (with the exception of DeepSHAP on FashionMNIST), IP generates explanations that are more predictive of the class label than popular attribution methods.
\begin{table}[h]
\centering
\renewcommand{\arraystretch}{1.5}
\caption{\textbf{MAP accuracy of explanations generated by Information Pursuit (IP) v/s other attribution methods.} IP explanations (in almost all cases) achieve a higher classification accuracy than explanations of the same length generated using baseline attribution methods. The (absolute) method refers to explanations generated using absolute values of the attribution map scores. On MNIST and KMNIST, IP explanations achieve a $10\%$ and $2.38\%$ relative improvement respectively over the best performing baseline method. On FashionMNIST, IP explanations are second best with a relative decrease of about $3.12\%$ from the best performing baseline.}
\label{Table: comparison_to_current_attrib_methods}
\begin{tabular}{| @{\;\;}l@{\;\;} | @{\;\;}c@{\;\;} | @{\;\;}c@{\;\;} | @{\;\;}c@{\;\;} |}
\hline
\textbf{Explanation Method} & \textbf{MNIST} & \textbf{KMNIST} & \textbf{Fashion-MNIST} \\
\hline
\textsc{Information Pursuit} & \textbf{96.78\%} & \textbf{91.02\%} & 85.60\% \\
\textsc{IG} & 78.48\% & 84.87\% & 78.49\% \\
\textsc{IG (absolute)} & 70.39\% & 84.72\% & 64.95\% \\
\textsc{DeepSHAP} & 87.98\% & 88.90\% & \textbf{88.36}\% \\
\textsc{DeepSHAP (absolute)} & 84.80\% & 84.56\% & 84.35\% \\
\hline
\end{tabular}
\end{table}
\subsection{Tasks and Query Sets} \label{Section: Tasks and Query Sets}
\stewy{There is a lot to describe concerning the tasks + query sets as well as the baseline models we compare to in the following section. I will definitely cut the length down, but in this draft, I wanted to start with the longest and most detailed version.}
\myparagraph{Image Patch Queries} When we set queries to be the pixel values of image patches, we recover the familiar feature attribution-type explanations of post-hoc methods, although with a guarantee that the explanations are faithful. We apply this query set to the simple task of classifying binary images from MNIST (digit classification), KMNIST \cite{clanuwat2018deep} (Japanese character classification), and FashionMNIST \cite{xiao2017fashion} (clothing classification).
We set $Q$ to be the set of queries corresponding to all $3 \times 3$ overlapping patches in the image. All three datasets consist of images of size $28 \times 28$, so we have $26^2$ patches of size three, giving $Q = \{q_1, q_2, ..., q_{676}\}$. For any $q_i \in Q$, the answer $q_i(X) \in \{0, 1\}^9$ is simply the $9$ pixel values observed in image $X$ at the patch indexed by $i$. However, our factorization \eqref{eq: factorization of query answers} that we use to model $p(Q(x), z, y)$ no longer holds as the overlapping queries $q(X)$ are now \textit{causally related} (and therefore dependent even when conditioned on $Z$, making them unable to be modeled by a VAE). So instead we train a VAE to model the pixel distribution, and then compute the probability distribution over the patch query $p(q(x) | z, y)$ as the product of the probabilities of all pixels in that patch. See Appendix \ref{Appendix: Network architectures and training} for further training details.
\myparagraph{Concept-Based Queries} Now we consider a more complex classification problem with a concept-based query set. We use the Caltech-UCSD Birds-200-2011 (CUB) dataset \cite{wah2011caltech}, which contains 11,788 images of birds each belonging to one of 200 species. This fine-grained bird-species identification task typically requires domain expertise, but the concept-level explanations generated by our method make understanding the decision-making process accessible, even for a non-expert user.
Each image in the CUB dataset is annotated with 312 binary attributes representing high-level concepts in the image, such as the color and shape of the beak, wings, and head. However, these attribute labels are noisy, either because a part of the bird corresponding to that attribute is occluded or because they were simply labeled incorrectly. We follow \cite{koh2020concept} in deciding attribute labels by majority voting. For example, if more than 50\% of images in a class have black wings, then we set all images in that class to have black wings. It is natural to assume that birds in the same species will share the same attribute descriptions.
Our query set $Q$ consists of all 312 binary attributes. So as described in Section \ref{section: Information Pursuit}, we learn a VAE to model $p(Q(x), z, y)$, the distribution of concepts. However, we do not assume that images at test-time come with concept-level annotations $Q(x)$. So following \cite{koh2020concept}, we train a separate model $p_\theta(Q(x) | x)$, called a \textit{concept network}, that learns to predict the concepts $Q(x)$ in an image $x$. During inference for a particular image $x$, we use the concept network to answer the concept queries asked by IP. \textcolor{blue}{Missing some details here that will be provided by Aditya on sequential vs independent concept bottleneck stuff}
\myparagraph{Word-Level Queries} For NLP tasks, words make a natural first candidate for queries, being the fundamental semantic units of text. So for a vocabulary of words $V = \{ v_1, ..., v_n \}$, we consider queries $Q = \{q_{v_1}, ..., q_{v_n}\}$ where $q_{v_i}(x)$ is a boolean function determining if word $v_i$ is present in document $x$. We evaluate this query set on the Huffington Post News Category Dataset \cite{huffpostdataset}, a document classification task consisting of newspaper headlines and the corresponding categories they were published under in the Huffington Post newspaper. We stem all words in the dataset and choose the top-1,000 according to tf-idf score to be our vocabulary $V$.
\subsection{Models} \label{Section: Experiments.Models}
In Table \ref{fig: results table}, we compare our method with several other baselines intended to reflect the performance characteristics of our method.
The first two models in the table, \textsc{IP} and \textsc{Decision Tree}, seek minimal subsets of the query set to make predictions on, while the following two models, \textsc{VAE Posterior} and \textsc{Baseline on Q}, use the full query set. \textsc{Decision Tree} is a CART decision tree, with the information gain splitting criterion. While both methods form shallow decision trees that branch on responses to queries, the CART decision tree approximates mutual information gain from data while IP estimates it with a generative model, significantly improving performance in terms of both explanation length and final accuracy. See Appendix XX for more details on how stopping criterion $\epsilon$ was chosen for each model.
\textsc{VAE Posterior} uses the same generative model to make predictions as IP but has access to the full query set. For IP, we use ULA to sample from $p(z | s_k^{\pi_{IP}}(x), y)$, but now that we have access to the full query set $s_k^{\pi_{IP}}(x) = Q(x)$, we can improve performance by making use of the VAE's encoder instead. Following equation \ref{eq: marginalizing z}, we draw many samples from the encoder $p(z | Q(x), y)$ and then decode these samples to estimate the VAE's posterior distribution $p(y | Q(x))$. \textcolor{blue}{Appendix? For each problem, we set the number of samples to be the same as what we draw during each iteration of IP (12,000 for binary image tasks, 12,000 for CUB, 10,000 for HuffPost News). We call this model \textsc{VAE Posterior}}.
\textsc{Baseline} is a standard supervised neural network that receives $x$ as input. For MNIST, KMNIST, and FashionMNIST, \textsc{Baseline} is a convolutional network. For CUB, it is an MLP and for Huffpost News, it is an LSTM. Its counterpart \textsc{Baseline on $Q$} is also a supervised neural network but receives the full query set $Q(x)$ as input instead, illustrating the effect that using a particular query set has on performance. For MNIST, KMNIST, and FashionMNIST this is the same convolutional network as \textsc{Baseline} and for CUB and HuffPost News, it is an MLP. See Appendix XX for architectural and training details.
\begin{figure*}
\centering
\begingroup
\renewcommand{\arraystretch}{1.5}
\begin{tabular}{| l | c c | c c | c c | c c | c c |}
\hline
\multirow{2}{*}{Model} & \multicolumn{2}{c}{MNIST} & \multicolumn{2}{c}{KMNIST} & \multicolumn{2}{c}{FashionMNIST} & \multicolumn{2}{c}{CUB Birds} & \multicolumn{2}{c}{HuffPost News} \\
& Acc & $L$ & Acc & $L$ & Acc & $L$ & Acc & $L$ & Acc & $L$ \\
\hline
\textsc{IP} & 96.78\% & 5.16 & 91.02\% & 16.02 & 85.40\% & 18.10 & 76.37\% & 11.08 & 71.21\% & 199.36 \\
\textsc{Decision Tree} & 90.23\% & 12.16 & 78.00\% & 13.86 & 80.80\% & 9.02 & & & 63.00\% & 112.02 \\
\hline
\textsc{VAE Posterior} & 98.97\% & - & 94.25\% & - & 87.56\% & - & & & 71.72\% & - \\
\textsc{Baseline on $Q$} & 99.15\% & - & 95.10\% & - & 86.96\% & - & 75.70\% & - & 71.48\% & - \\
\hline
\textsc{Baseline} & 99.15\% & - & 95.10\% & - & 86.96\% & - & 82.70\% & - & & - \\
\hline
\end{tabular}
\endgroup
\caption{Classification accuracies and explanation length $L = \mathbb{E}_X [|expl_Q^\pi(X)|]$ on a variety of methods and datasets. See Section \ref{Section: Experiments.Models} for more details on the models in the table. \stewy{Cells that are blank correspond to baselines that are not quite done yet. Dashes mean that the metric is not applicable to that particular method.}}
\label{fig: results table}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.4 \textwidth]{figs/Acc_Minimality_Plot.pdf}
\caption{Accuracy vs. IP's Average Explanation Length as the mutual information stopping threshold $\epsilon$ varies. \stewy{Should be a plot of relative test accuracy instead}}
\label{fig:accuracy_minimality_plot}
\end{figure}
\subsection{Performance and Minimality}
The performance of IP relative to traditional non-interpretable methods depends on three factors: 1) the sufficiency of the query set 2) the posterior accuracy of the generative model and 3) the value of the $\epsilon$ parameter governing the performance-minimality trade-off. We perform experiments to measure the relative impact of each of these factors across tasks and query sets. In this section, we also study the length of generated explanations and find that our method generates surprisingly short explanations...?
\myparagraph{1. Query Set Sufficiency} Since IP receives queries as input, it can only perform as well as the query set is sufficient for the task. In Table \ref{fig: results table}, we compare the performance of IP with other baselines, including standard non-interpretable models. We can get a sense of how nearly sufficient the query set $Q$ is for a task by comparing the accuracies of the non-interpretable \textsc{Baseline} (which has access to the full input $x$) with those of \textsc{Baseline on $Q$} (which only has access to query responses $Q(x)$).
In the case of image patch queries, the query set is trivially sufficient as it reconstructs the image itself. Therefore, the accuracies of \textsc{Baseline} and \textsc{Baseline on $Q$} are identical on the binary image classification datasets. On CUB and HuffPost News however, there is a sizable drop in accuracy when only the query set $Q(x)$ is observed. For HuffPost News, the query set is the presence/absence of words from a vocabulary of size 1,000. \stewy{Need to finish RNN text baseline experiment here:} So the query set loses ordering information, which causes a ... drop in accuracy .... For CUB, as mentioned in Section \ref{Section: Tasks and Query Sets}, we pre-process the attribute labels for each image in such a way that the class is known to be fully determined by these attributes. However, at test time, we do not have access to attribute labels, and estimate them from the raw image using a trained concept network. While this process leads to a decrease in performance between \textsc{Baseline} and \textsc{Baseline on $Q$}, it enables powerful, human-interpretable concept-level queries. Choosing the right query set for a task is non-trivial and affects both the quality of final explanations as well as the performance of the method. We cover further design considerations in \ref{Section: Domain Intelligibility}.
\myparagraph{2. Posterior Accuracy of the Generative Model} Another significant difference between our method and traditional approaches is that we use the posterior of a generative model to make predictions $p(y | s_{k}^{\pi_{IP}}(x))$. Therefore, the performance of the VAE's posterior $p(y | Q(x))$ given the full query set gives an upper bound to the performance of IP, which only has access to a portion of the queries. Table \ref{fig: results table} shows that the VAE's posterior accuracy is within 2\% of its standard supervised counterpart \textsc{Baseline on $Q$} across all problems tested, validating our approach of using the posterior of a conditional VAE as a predictive model.
\myparagraph{3. Accuracy-Minimality Tradeoff}
Finally, unlike standard supervised methods, we seek to use only a small subset of queries to make a prediction. IP terminates on a given input once a certain metric reaches a stopping threshold $\epsilon$. For the binary image tasks, we terminate once the probability of the most confident class exceeds 99\%, or equivalently once the remaining uncertainty in the prediction is less than 0.01. For CUB and HuffPost News, we terminate once the highest mutual information of any remaining query falls below a certain threshold (0.05 and $10^{-3}$ respectively).
As described in Section \ref{section: info theory connections}, this $\epsilon$ corresponds to how approximately sufficient the generated code is for task $Y$, thereby controlling an accuracy-minimality tradeoff. When $\epsilon$ is large, the stopping criteria is quickly reached, so IP produces short explanations, but with a loss in accuracy since there is high uncertainty in the model's predictions. When $\epsilon$ is small, the generated explanations are more nearly lossless codes for $Y$, but may be prohibitively long.
\stewy{Describe minimality plots here (Figure \ref{fig:accuracy_minimality_plot}) - still TODO}
\begin{figure*}
\centering
\begin{tabular}[]{c}%
\subfloat[MNIST digit classification - Image Patch Queries]{
\includegraphics[width=0.5 \textwidth]{figs/IP_in_action_MNIST.eps}
\label{fig:IP MNIST in action}
} \\
\subfloat[CUB bird identification - Concept-level Queries]{
\includegraphics[width=0.65 \textwidth]{figs/IP_in_action_CUB.eps} \label{fig:IP CUB in action}
}
\end{tabular}
\quad
\subfloat[HuffPost News document classification - Word-level Queries]{
\includegraphics[width=0.27\textwidth]{figs/IP_in_action_HuffPost.eps}
\label{fig:IP Text in action}
}%
\caption{Visualizations of IP's decision-making process across several tasks. In Figure \ref{fig:IP MNIST in action}, the top row displays the test image with red boxes denoting the current selected patch and blue boxes denoting previous patches. The second row shows the revealed portion of the image that the predictive model gets to use at each query. The final row shows the model's estimated posteriors at each query, beginning at a nearly uniform prior before converging on the true digit ``6'' after 4 queries. The left image in Figure \ref{fig:IP CUB in action} is the input image and on right we have a heatmap of the estimated class probabilities at each iteration. We only show the top 10 most probable classes out of the 200. To the right, we display the queries asked at each iteration, with red indicated a ``false'' response and green a ``true'' response. In Figure \ref{fig:IP Text in action}, we again show a heatmap of estimated class probabilities at each iteration. Words colored in red are absent from the sentence while words in green are present.}
\label{Fig: IP in action}
\end{figure*}
\subsection{Domain Intelligibility} \label{Section: Domain Intelligibility}
\stewy{Need a short intro/topic sentence here to orient the reader. This subsection should probably also appear towards the beginning of the experiments section.} Figure \ref{fig:IP MNIST in action} illustrates a typical run of our algorithm on a test image from MNIST, choosing $3 \times 3$ patches until termination. The first query is near the center of the image; recall from \eqref{def. IP algorithm} that this choice is independent of the particular image and represents the patch whose pixel values have maximum mutual information with $Y$. After seeing a vertical white line through this center patch, the updated posterior concentrates its mass on ``1''. By the fourth query, the model expects the digit to be either a ``6'' or ``8''. The next query is well-suited to separating sixes and eights, being a patch in a region that should be black for an ``8'' but could contain a vertical white stroke for a ``6''. Upon discovering such a vertical stroke in that patch, the algorithm reaches high confidence in the image being a ``6'' and terminates. From Table \ref{fig: results table}, the average explanation length is only 5.16 queries. Paired with visualizations of the posteriors at each iteration, these patch explanations are concise and easy to interpret on MNIST digit classification.
However, Table \ref{fig: results table} shows that even slightly harder classification tasks like KMNIST and FashionMNIST require significantly more patch queries. Low-level patch queries are ill-suited to complex vision tasks, where they may be hard to interpret and lead to very long explanations.
In contrast, on the much more complex 200-class CUB bird identification task, IP with concept-level queries converges on the correct class after just 11.08 queries on average. Figure \ref{fig:IP CUB in action} shows IP's inference process. This example illustrates a common emergent query selection strategy. After converging on just four possible bird species out of the 200 by query 5, IP chooses queries that would verify/descredit specific classes among its highest probability predictions. Green Jays and Scott Orioles have black breasts, so they are ruled out after query 6. Likewise, Tropical Kingbirds are buff-colored (a yellow ochre) while Great Crested Flycatchers are brown. These concept queries lead to concise and easily-understood explanations, even on a fine-grained dataset that typically requires expert knowledge to perform accurate classification.
Lastly, in Figure \ref{fig:IP Text in action} we display an example run of IP on the HuffPost News Category Dataset. The query responses are much sparser, since each sentence only contains a few words out of the 1,000 in the vocabulary. As a result, most queries are not very informative, so IP requires more queries, 125 in total. However, due to the sequential nature of our algorithm, by plotting how the predicted class probabilities change with each query, we can gain insight into the prediction, even with a comparatively large number of queries. Upon reaching its first positive query ``eat'', the probability mass concentrates on the categories ``Food \& Drink'' and ``Wellness''. However, as the queries ``citi'', ``visit'', ``york'', and ``bar'' come back positive, the model becomes more and more confident that ``Travel'' is the correct class.
In general, we seek query sets that lead to concise, intelligible explanations that are also sufficient for the task. Choosing a good query set for a given problem is non-trivial and requires an understanding of the important structure in the data. In particular, concept-level queries are promising as an approach towards interpretable, concise, faithful explanations for complex machine learning tasks.
|
\section{Introduction}
\IEEEPARstart{W}{ideband} receivers are essential for numerous applications such as opportunistic radios and high-throughput, multi-channel links. Exposing receiver circuits to a very wide spectrum, however, makes them vulnerable to strong interferers. This can result in desensitization, deteriorated linearity performance, and possibly damaging the receiver. Consequently, a diverse amount of technologies and techniques are researched to detect and suppress interferers in wideband systems.
Signal processing techniques, analog and digital, have been implemented to detect and suppress interferers for various wideband applications such as UWB \cite{uwb_digital_suppression,gnss_digital_interf_detec,dsp_beamforming,mixer_detection}. Such techniques benefit from the agility of digital systems, or the speed of analog circuits. To operate properly, however, those techniques rely on the linearity, and the power reliability, of the RF front-end, which could be a limitation if a high-power interferer exists. The operating bandwidth is also typically limited.
RF power limiters have shown a relatively fast response (nanoseconds in \cite{abbas_resonator}) while protecting the rest of the RF chain. This is favorable in pulsed interferers, where it is necessary to attenuate a short high-energy pulse. Wideband attenuators are available in various technologies such as plasma \cite{abbas_resonator}, MEMS \cite{mems_self_actuating_limiter} or solid-state \cite{pin_limiter,gan_robust}. The use of wideband attenuators results in an adverse suppression of the whole operating band, suppressing the signal of interest. Alternatively, widely tunable narrow band limiters \cite{abbas_resonator} can single out the interferer in the attenuation. This, however, still requires prior knowledge of the interfering signal, and limits the instantaneous operating bandwidth of the RF chain. Furthermore, limiters frequently suffer from limited linearity, and recovery times of up to 10--40 ms \cite{gan_robust}.
Another set of technologies to combat interferers in wideband systems is frequency-selective limiters. They typically protect the receiver circuits, while leaving the rest of the channel intact. The innovations in this area include adaptive power limit and tunable frequency. For example, the work in \cite{mansour_self_actuation}, utilizes an intelligent self-actuation of MEMS switches, where a dc bias controls the actuation power level. This provides a frequency-dependent power limiting, which is implemented with a tunable filter. The limited response time of MEMS switches can be resolved with the use of diode-based nonlinear filters \cite{freq_selective_lim_nonlinear_bsf, guyette_sig_track, naglich_freq_selec,hcplr_diode_limiter}. Such power-dependent responses typically have limited power dynamic range (10--20 dB). A wider dynamic range can be achieved with active coupling techniques \cite{wei_selective}, which, similar to \cite{mansour_self_actuation,freq_selective_lim_nonlinear_bsf}, requires prior knowledge of the interfering signal frequency. Spectrum scanning methods in \cite{ak_diplexer,ak_jammer_suppression} are effective in locating the interferer in the band of interest, at the cost of extended response time.
The interference detection method presented in this paper is capable of operating over a multiple octaves, while maintaining sub-$\mu$s response time. The detection method is based on weakly coupling into the incoming signal, and measuring the standing wave pattern of the resulting waveform. The initial concept is presented by the authors in \cite{ak_spiky}. This work significantly expands the on the previous work by analyzing the design and optimistic the tradeoffs. In addition, several new concepts and methods are introduced that result in 1) increasing the instantaneous operating bandwidth by two octaves (from 2--8 GHz to 1--16 GHz), 2) increasing the detected power dynamic range by 20 dB (from 20 dB to 40 dB), 3) halving the response time (from 1 $\mu s$ to 0.5 $\mu s$), and 4) demonstrate the ability to detect and suppress multiple interferers. The interference detection method is implemented and fitted with different types of tunable bandstop filters (BSFs) to demonstrate its versatility. The unmatched measured results represents a feasible solution for wideband receivers operating over multi-octave bands simultaneously.
\begin{figure*}[!t]
\centering
\includegraphics[width=6.5in]{block_diagram.eps}
\caption{The interferer detection concept relies on resistive sampling of the signal (using $R_C$). The amplitude of the sampled signal is then conditioned (with a programmable attenuator and an amplifier) to drive an open stub. The voltage at the open end is proportional to the interferer power. Additional measurement points (at $l1$ and $l2$) are used to measure the standing wave voltage, which carries the interferer frequency information. Multiple point are used to increase the detectable frequency range. A look-up-table is used to relate the voltage readings to interferer power and frequency.}
\label{fig_concept}
\end{figure*}
\section{Detection Theory of Operation}
The initial detection concept is briefly summarized in \cite{ak_spiky}. Some parts below are repeated for completeness.
\subsection{Detection Concept}
\label{sec_structure}
In the presence of multiple signals in a channel, the waveform is dominated by that of the highest power signal. As a result, the frequency and power of the incoming signal are monitored and used to determine whether the strongest signal is an interferer that should be suppressed. The concept of monitoring the power and frequency of an incoming signal is summarized in Fig. \ref{fig_concept}.
A weak resistive coupling into the received signal path produces a replica of the signal into the monitoring part. A resistor, as opposed to reactive components, is preferred since it results in a frequency-independent coupling. In addition, reactive coupling typically suffers from self-resonance when operating over multi-octave bandwidth. The coupled signal then goes through a programmable attenuator, which maintains the signal to within a manageable level for subsequent stages. This also serves in increasing the power dynamic range of the signal monitoring part. The signal is then amplified to a detectable level.
While, from a noise perspective, the amplifier is expected to be placed before the attenuator, the attenuator is placed first since it typically cannot handle the high power from the amplifier in the case of a strong interferer. This also helps in maintaining a linear gain from the amplifier.
The signal from the amplifier then drives an open circuit stub. The magnitude of the voltage at the open end is (assuming sufficiently large sensing resistor $R_D$ at the open end)
\begin{equation}
\left|V_{OC} \right| = \sqrt{8P_{STUB}Z_{0S}}= \sqrt{8CL_{Att}G_{PA}P_{in}Z_{0S}}.
\label{eq_voc}
\end{equation}
Where $P_{STUB}$ is the power incident on the open stub, $Z_{0S}$ is its characteristic impedance, $C$ is the power coupling from the $R_C$, $L_{Att}$ is the loss from the attenuator, $G_{PA}$ is the amplifier gain, and $P_{in}$ is the RF input power. The analysis of the circuit model of the resistive coupling (shown in in Fig. \ref{fig_concept}) reveals that the coupling $C$ is
\begin{equation}
C (dB) = 20 \log \left( \frac{2Z_0}{2R_C+3Z_0}\right).
\label{eq_cplg}
\end{equation}
Where $Z_0$ is the system impedance ($50~\Omega$ here). From \eqref{eq_voc} and \eqref{eq_cplg}, a smaller coupling resistance provides stronger coupling and larger voltage to detect in the monitoring part of the design. This, however, can deteriorate the insertion loss (S21) and the matching condition (S11). These conditions are described by
\begin{subequations}\label{eq_snp}
\begin{align}
S11 (dB) = 20 \log \left(\frac{Z_0}{2R_C+3Z_0}\right),\label{eq_snp_s11}\\
S21 (dB) = 20 \log \left(\frac{2R_C+2Z_0}{2R_C+3Z_0}\right).\label{eq_snp_s21}
\end{align}
\end{subequations}
To properly choose the coupling resistor value, the insertion loss condition is used here since the matching is less sensitive to $R_C$. For example, for a 0.8 dB insertion loss condition, $R_C=210~\Omega$, the coupling $C$=-15 dB, and the matching S11= -21 dB.
The relationship between $V_{OC}$ and $P_{in}$ in \eqref{eq_voc} presents a power monitoring method. This is used in a comparison with a programmable power threshold in order to decide whether to take action against the interferer or not. The minimum and maximum powers are generally dictated by the power detector limits.
To monitor the frequency, the standing wave pattern is used.
The normalized magnitude of the standing wave pattern is
\begin{equation}
\left| \frac{V\left(l\right)}{V_{OC}} \right| = \left| \cos \left( \frac{\pi}{2} \frac{f}{f_{MAX}} \right) \right|.
\label{eq_ratio}
\end{equation}
Where $l$ is the length from the open end of the stub, $f$ is the frequency of the strongest signal in the path (potentially the interferer), $f_{MAX}$ is the maximum operating frequency ($c/4l$) where the length $l$ is a quarter wavelength ($\lambda/4$). Beyond this frequency, the relationship between $V(l)$ and the frequency is no longer bijective, and the ratio no longer indicates a unique frequency.
The relationship in \eqref{eq_ratio} dictates the maximum detectable frequency. The minimum detectable frequency, however, is dictated by the resolution of the monitoring system (details in Section \ref{sec_resol}).
To widen the frequency detection range, a second monitoring node is used at a further distance from the open end ($l2$ in Fig. \ref{fig_concept}). The further monitoring node is used to measure lower frequencies up to its maximum, then the nearer one is used ($l1$). Switching the decision between monitoring $l1$ or $l2$ is based on the value of $l1$ reaching the minimum detectable frequency.
The relationship in \eqref{eq_ratio} works as the frequency monitoring method. Combined with \eqref{eq_voc}, a two-dimensional calibration table is now formed to identify the frequency and power of the strongest signal in the RF path, which is potentially an interferer as depicted in in Fig. \ref{fig_concept}.
\subsection{Resolution Analysis}
\label{sec_resol}
The ratio in \eqref{eq_ratio} shows small variations at lower frequencies. For example, at $f=0.1 f_{MAX}$ and $f=0.2f_{MAX}$, the values are 0.987 and 0.951, respectively (a difference of 3.7$\%$). While the ratios at $f=0.8 f_{MAX}$ and $f=0.9 f_{MAX}$ are 0.31 and 0.156, respectively (a difference of 15$\%$). In other words, at low frequencies, the monitoring system reaches a point where the ratio no longer indicates frequencies accurately. As a result, it is essential to analyze the resolution of the system to be able to know where to place the secondary sensor, and to quantify the minimum detectable frequency as a specification of the system.
For the analysis below, $V_{OC}$ is assumed to be held constant at the maximum voltage of the Analog to Digital Converter (ADC). This assumption can be validated by using the programmable attenuator as can be inferred from \eqref{eq_voc}. Furthermore, the equations take into consideration that the employed detectors shown in Fig. \ref{fig_concept} are logarithmic power detectors, to cover a wide sensing dynamic range.
Using the assumptions above, \eqref{eq_ratio} becomes
\begin{equation}
\log \left( V\left(l1,2\right) \right) - \log \left(V_{OC} \right) = \log \left( \cos \left( \frac{\pi}{2} \frac{f}{f_{MAX}} \right) \right).
\label{eq_ratio_log}
\end{equation}
The output voltage of the logarithmic power detector, $V_{DET}$, is given by
\begin{equation}
V_{DET}= A\log \left( V\left( l1,2 \right) \right) +B.
\label{eq_ratio_log2}
\end{equation}
From \eqref{eq_ratio_log} and \eqref{eq_ratio_log2}, we get
\begin{equation}
V_{DET}= A\log \left( \cos \left( \frac{\pi}{2} \frac{f}{f_{MAX}} \right) \right) +D.
\label{eq_ratio_log3}
\end{equation}
Where $A$, $B$, and $D$ are constants that can be extracted from the power detector properties and $R_D$.
Then, \eqref{eq_ratio_log3} is differentiated versus frequency, leading to
\begin{equation}
\frac{\partial V_{DET}}{\partial f}= -\frac{A \pi }{2ln(10)f_{MAX}} \tan \left( \frac{\pi}{2} \frac{f}{f_{MAX}} \right).
\label{eq_dv_df}
\end{equation}
To find the frequency resolution $ \left| \Delta f_{min} \right|$, \eqref{eq_dv_df} is rearranged, and the derivative is approximated as a finite step division, or
\begin{equation}
\left| \Delta f_{min} \right|= \frac{\Delta V_{min}}{\frac{A \pi }{2ln(10)f_{MAX}} \tan \left( \frac{\pi}{2} \frac{f}{f_{MAX}} \right)}.
\label{eq_res}
\end{equation}
Where $\Delta V_{min}$ is the minimum change in $V_{DET}$ that is detectable by the ADC. As predicted earlier in this section, it is clear from \eqref{eq_res} that the frequency resolution worsens at low frequencies. As a result, once the minimum detectable frequency change is decided, the further detector (at $l2$) is placed such that its maximum detectable frequency coincides with that minimum acceptable frequency from the nearer detector (at $l1$). These conclusions are illustrated in Fig. \ref{fig_resolution}.
For example, in Fig. \ref{fig_resolution}, if the maximum frequency for $l1$ is 16 GHz, and the maximum accepted $\Delta f$ is 0.25$\%$, then the sensing node $l2$ is placed such that its maximum frequency is at 8 GHz (for 12-bit ADC). Following the 12-bit curve for $l2$ results in a minimum frequency of approximately 3.8 GHz. More sensing nodes can be added to detect even lower frequencies. This, however, increases the length of the sensing stub.
The minimum detectable frequency ($\Delta f$) in \eqref{eq_res} is predicted to be infinitesimally small at $f_{MAX}$. This, however, assumes that the power detector has an unlimited power detection dynamic range. As a result, a finite value of $\Delta f$ is expected in practice.
While a higher number of bits in an ADC can result in a finer frequency resolution, it typically comes at the cost of slower operation. In other words, the response time has to be taken into consideration when deciding the acceptable resolution.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{resolution_theory.eps}
\caption{The resolution of the sensing method (a) in GHz and (b) in percent of the frequency, as a function of frequency, using various ADC resolution. The percentage numbers are necessary for designing subsequent stages such as BSFs. If the resolution using $l1$ is not accceptable, the systems can switch the sensing to $l2$. The plots are based on A=0.4.}
\label{fig_resolution}
\end{figure}
\section{Design Consideration}
\subsection{Power Limitations}
The applications of the detection method include suppressing high-power incoming signals. As a result, a few power limitations are studied below.
The coupling resistor dissipates part of the incoming power. The value of this resistor is typically dictated by the monitoring design as discussed in Section \ref{sec_structure}. As a result, the maximum allowable dissipated power by $R_C$ is dictated by its technology and package. Fig. \ref{fig_power_lim} demonstrated the input power effect on the power dissipated in $R_C$ along with typical power limitations of various packages. This plot can be used to determine the necessary package of $R_C$ for a particular application.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{rc_power.eps}
\caption{The power dissipated in $R_C$ versus the input power at multiple values. The package for $R_C$ dictates the maximum power it can handle, which in turn limits the maximum input average power.}
\label{fig_power_lim}
\end{figure}
High input power can also saturate the power detector, leading to an erroneous frequency read. This can be concluded from \eqref{eq_ratio}, where a saturated power detector leads to $V_{OC}$ remaining nearly constant, while $V_{l}$ is increasing with power. The programmable attenuator can mitigate this issues by keeping the signal within the linear regime of the detector.
\subsection{Directional Coupler}
\label{sec_cplr}
Resistive coupling provides a wideband, compact solution for the detection method. In the presence of a highly reflective filter after it, however, there is a chance that the operation of the detector and the filter becomes unstable. This instability, demonstrated in Fig. \ref{fig_cplr}(a), results from the filter reflecting the interferer, which creates a standing wave between the filter and the preceding circuits including the detection method. If the resistor is at a location where the standing wave has a low value, it will then falsely detect that the interferer is no longer present and the controller can disengage the filter. At which point, the interferer is detected again.
This instability can be resolved with various methods including reflectionless filters \cite{wei_selective, guyette_absorptive, ak_dual}, and isolators \cite{3d_printed_isolator,isolator_w_diodes}. The bandwidth of these technologies, however, is typically limited to one octave. As a result, a directional coupler is also demonstrated in lieu of $R_C$ in case a highly reflective filter is used. The applied concept is shown in Fig. \ref{fig_cplr}(b).
To minimize the size of the directional coupler, an asymmetric five-section coupled line design is used \cite{cplr_tables}. A picture of the directional coupler is shown in Fig. \ref{fig_cplr}(c). The utilized striplines are sandwiched between two substrates to minimize the mismatch between the even and odd mode propagations in the coupled lines. The radial stub enhances the termination at high frequencies.
The coupler is fabricated on a 25-mil RO3006 substrate with $\epsilon_r$=6.15. While higher dielectric constants can shrink the overall size, it makes the stripline widths smaller and more sensitive to fabrication tolerances. On the other hand, lower dielectric constants dictate smaller gaps between the coupled lines, which can also be sensitive to tolerances. As a result, an intermediate dielectric constant is used to maintain the widths and gaps reasonably large to be more immune to variations.
The simulated and measured results show approximately 15 dB of coupling, and better than 1.6 dB insertion loss up to 14 GHz as shown in Fig. \ref{fig_cplr}(d). The directivity is above 6 dB throughout its operating band, which is sufficient for normal operation of the detector.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{coupler.eps}
\caption{(a) If the employed filter is highly reflective, it might cause the detection method to falsely decide that the interferer is no longer present, due to the standing wave pattern created by the filter reflection. (b) A directional coupler can also be used to distinguish the traveling wave from the reflected one. This ensures a stable operation with highly reflective filters at the cost of size. (c) The implemented directional coupler with and without the substrate to cover the striplines. An asymmetric design is used to minimize the footprint. (d) Simulated and measured results of the directional coupler.}
\label{fig_cplr}
\end{figure}
\section{Implementation and Measurements}
\subsection{Designed Structure}
The interference detection method in Fig. \ref{fig_concept} is prototyped with both coupling methods, resistive and using the directional coupler. The rest of the circuits are similar. The implemented PCBs are shown in Fig. \ref{fig_photos} along with relevant dimensions. Table \ref{table_components} summarizes the components used in this design. The 12-bit ADC is running at 5 MSPS.
The presented prototype consumes 280 mA from a 3.3 V power supply, and 68 mA from an 8 V power supply for the amplifier. Since the power detectors and ADCs are electronic circuits, further power reduction can be achieved with an integrated implementation.
The detector is designed such that the maximum operating frequency is 16 GHz for the first detector (at $l1$), and 5 GHz for the second detector (at $l2$). The dimensions $l1$ and $l2$ deviate slightly from the theoretical values ($\lambda_g/4$) primarily due to the loading effect from the finite resistors $R_D$. The final values are optimized in simulations. This configuration results in a minimum detectable frequency of 1 GHz. In the prototype with the directional coupler, however, the maximum frequency is limited to 14 GHz as shown in Fig. \ref{fig_cplr}(d). The results from either coupling method are otherwise identical.
\begin{table}[!t]
\caption{List of components in the interference detection prototype}
\centering
\begin{tabular}{|l|c||l|c|}
\hline
$R_C$ & 220 $\Omega$ & $R_D$ & 180 $\Omega$\\
\hline
Amplifier & HMC460LC5 & Power detector & HMC1094LP3E\\
\hline
Attenuator & RFSA2113 & ADC & LTC2315CTS8\\
\hline
\end{tabular}
\label{table_components}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{photos.eps}
\caption{The implemented interference detection with (a) resistive coupling, and (b) directional coupler. The directional coupler prevents reflected signals to cause potential false readings. The two designs are identical otherwise.}
\label{fig_photos}
\end{figure}
\subsection{Characterization}
To quantify the performance of the interference detection method, first, its effect on the insertion loss is studied. Since the results from the directional coupler prototype are discussed in Section \ref{sec_cplr}, the resistive coupling is discussed below.
Fig. \ref{fig_insertion}(a) shows the S-parameters from $RF_{IN}$ to $RF_{OUT}$ with and without the inclusion of the coupling resistor $R_C$. As expected from \eqref{eq_snp_s11} and \eqref{eq_snp_s21} a small attenuation in S21 appears as a result of the coupling, while S11 is nearly unaffected. Fig. \ref{fig_insertion}(b) compares the achieved attenuation with \eqref{eq_snp_s21}. The minor deviation is attributed to the finite reflection of the programmable attenuator, and the frequency-dependent value of the resistors due to package parasitics.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{insertion_loss.eps}
\caption{(a) Measured and simulated S-parameters of the transmission line in the detection board with and without the coupling resistor $R_C$. The coupling resistor shows low loss and minimal effect on reflection. (b) Zoomed in measured and simulated insertion loss with the resistive coupling compared to the analytical value in \eqref{eq_snp_s21}.}
\label{fig_insertion}
\end{figure}
The readouts from the ADCs are then investigated versus the input power and frequency (as an interferer). Fig. \ref{fig_f_lf_hf}(a) shows the RF input power versus the reading of the ADC at the open end of the stub. To avoid the saturation of the power detectors, the attenuator is activated at relatively high input power. The power dynamic range is limited by the linear range of the power detectors. The linear (in logarithmic scale) relationship complies with the \eqref{eq_voc}. The frequency-dependent response of the detector does not compromise the frequency/power detection. This is because the detection method also relies on the frequency reading. In other words, the system picks the proper curve in Fig. \ref{fig_f_lf_hf}(a) when determining the power.
For the frequency detection, Fig. \ref{fig_f_lf_hf}(b) shows the normalized power detector readings at $l1$ and $l2$, along with the theoretically predicted values in \eqref{eq_ratio_log3}. It is clear that at lower frequencies, the readings from $l1$ show no significant changes. As a result, the readings from $l2$ are used since they are more accurate. At higher frequencies, on the other hand, readings from $l2$ are no longer bijective as discussed in Sections \ref{sec_structure}, which means they can only be detected from $l1$ readings. The analytical values in Fig. \ref{fig_f_lf_hf}(b) reach a theoretical negative infinite value. This is naturally not achievable in practice since the power detectors saturate at low powers.
Fig. \ref{fig_f_lf_hf}(a) and Fig. \ref{fig_f_lf_hf}(b) form the calibration look-up-table shown in Fig. \ref{fig_concept}.
The resolution of the presented prototype is shown in Fig. \ref{fig_f_lf_hf}(c), in GHz, and Figs. \ref{fig_f_lf_hf}(d), in percentage, as derived from the results in Fig. \ref{fig_f_lf_hf}(b). The theoretical value in \eqref{eq_res} is also plotted with the 12-bit ADC used in this implementation. Deviations from the theoretical values are contributed to factors such as modeling accuracy and the frequency-dependent response of the components. The results also show that the minimum detectable frequency is below 3$\%$ for most of the frequency range, which is within the capabilities of numerous filter technologies \cite{microstrip_example,ak_vib,saw_example,ak_temp}.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{pwr_lf_hf.eps}
\caption{(a) Readings from the ADC at the open end of the stub. The voltage is proportional to the input power and is used for power monitoring. The programmable attenuator is activated around 0 dBm to maintain the linear region of the power detector. (b) Analytical and measured values of $V_{DET}$ in \eqref{eq_ratio_log2} at the low frequency point $l2$ and high frequency point $l1$. This represents the frequency monitoring capability. (c) and (d) The frequency resolution of the measured monitoring data in (b) in GHz and in percentage of operating frequency, respectively. All the measurements above are taken from -20 to 20 dBm.}
\label{fig_f_lf_hf}
\end{figure}
Interferers typically have a finite bandwidth due to modulation. Fig. \ref{fig_mod} shows the reading from the interference detection with continuous wave (CW) signals, in addition to modulated signals with various bandwidths (using 16QAM modulation). The shown measurements are a combination of $l1$ and $l2$, as shown in Fig. \ref{fig_f_lf_hf}(b). The resulting frequency reading shows a weak dependence on the modulation bandwidth, with most of the band being within 2$\%$ error and a maximum deviation of less than 5$\%$. As a result, the detection method is showing a reasonable immunity against interferer bandwidth. While many other combinations of modulation types and bandwidths can be tested, this is beyond the scope of this paper.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{modulation.eps}
\caption{Frequency readings (similar to the one in Fig. \ref{fig_f_lf_hf}(b)) for CW signal, in addition to 16QAM signals modulated at 5 and 12 MSPS. The monitoring method shows reasonable robustness versus a modulated signal.}
\label{fig_mod}
\end{figure}
\subsection{Application Measurements}
The response time for interference detection is a critical measure to ensure the input energy into a receiver is maintained below the damage level. Furthermore, it is a measure of the downtime of a jammed channel. Fig. \ref{fig_response_time} shows the measurement setup for the response time of the implemented prototype. In summary, an RF pulse is injected into the interference detector, and the output is measured with a wideband envelop detector (ADL6012). The controller in the interference detector also generates a signal indicating the presence of the interferer when it is above a predetermined threshold. The two signals are then measured with an oscilloscope.
As shown in the plots in Fig. \ref{fig_response_time}, the response time of the interference detector is lagging the RF pulse by 500 ns on average. The delay is primarily due to the ADC (1--2 sample times) and the controller (1 clock time). Faster ADCs can mitigate that if sacrificing the resolution or high power consumption is acceptable.
To test the response time in a more realistic fashion, the interference detector is fitted with a tunable BSF as shown in the block diagram in Fig. \ref{fig_response_time_eva}(a). The filter is designed with evanescent-mode resonators to achieve a wide tuning range \cite{ak_vib,bsf_liu}. The tuning is achieved with PIN diodes, resulting in a tuning time of 10s ns. The schematic and the photo the implemented filter are also shown in Fig. \ref{fig_response_time_eva}(a).
The envelop of the output signal is shown in Fig. \ref{fig_response_time_eva}(b). Compared to the case when the BSF is disabled, the signal is suppressed after the 500 ns delay. When the pulse ends, the filter remains at the pulse frequency for an additional 500 ns before it tunes away from the pulse frequency. The progression of the output spectrum is shown in Fig. \ref{fig_response_time_eva}(c)--(f).
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{response_time.eps}
\begin{picture}(0,0)
\put(-28,147){\fontsize{6pt}{1em}\sffamily(Fig. \ref{fig_concept})}
\end{picture}
\caption{The measurement setup for the response time (top). The measured results show an average of 500 ns response time, independent of interferer frequency. Oscilloscope signals are vertically scaled to clarify the delay time.}
\label{fig_response_time}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.8in]{response_time_w_eva.eps}
\begin{picture}(0,0)
\put(-49,523){\fontsize{6pt}{1em}\sffamily(Fig. \ref{fig_concept})}
\end{picture}
\caption{(a) Measurement setup for the response time with a BSF. PIN diode tunable evanescent-mode filters are used to suppress the interferer. (b) Measured envelop of the signal with and without a filter. Initially, the signal has low power and no filter is necessary. This spectrum is shown in (c). Then, a high-power pulse exceeds the threshold as shown in the spectrum in (d). The controller then engages the filter after 500 ns, suppressing the interferer as shown in (e). When the RF pulse ends, the filter remains engaged momentarily, resulting in a spectrum shown in (f). The detector controller then realizes that the pulse is below the threshold, disengaging the filter. The resulting spectrum is also shown in (c)}
\label{fig_response_time_eva}
\end{figure}
The interferer detection method can be employed in an automated and programmable power limiter, where it tunes a BSF to the frequency of the interferer. The setup is similar to the one shown in Fig. \ref{fig_response_time_eva}(a). The measurements in Fig. \ref{fig_limiter}(a) is the result using an evanescent-mode-based filter and Fig. \ref{fig_limiter}(b) a commercial YIG-based filter (MLFR-0220 from Micro Lambda Wireless), respectively. This demonstrates the agility of the presented work versus the filtering technique. The results also demonstrate the ability to program the threshold power between -20 to 20 dBm regardless of the operating frequency. The suppression of the evanescent-mode filter weakens at higher powers due to the PIN diodes being no longer in the off state. This can be mitigated with higher reverse bias.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{limiters.eps}
\caption{Measured signal suppression versus its input power from using the detection method in a programmable limiter. The results are achieved by using (a) evanescent-mode-based filter, and (b) YIG-based filter. The power threshold to engage the filter is programmable between -20 to 20 dBm.}
\label{fig_limiter}
\end{figure}
In the presence of multiple strong interferers, the presented method can detect them sequentially when cascaded. In such a setup, as shown in Fig.\ref{fig_2_interferer}(a), the first detector/filter combination suppresses the highest power interferer, and subsequent stages follow the power order. This is demonstrated with two interferers as shown in Fig.\ref{fig_2_interferer}(b)--(e) with various combinations.
\begin{figure}[!t]
\centering
\includegraphics[width=3in]{multi_interf.eps}
\caption{(a) Cascading the detectors and filters can suppress multiple interferers. (b)--(e) The output spectra at various power combinations for two interferers at 6 GHz and 12 GHz. The high suppression in (c) and (d) is due to the use of YIG filter at 12 GHz.}
\label{fig_2_interferer}
\end{figure}
\section{Conclusion}
This article presented an interference detection method with a state-of-the-art four-octave frequency detection range, and a response time of 500 ns. The frequency range and resolution of the system are analytically derived for a general implementation. The detection method has the flexibility to interface with a variety of tunable BSFs. This detection method is an enabling technology for wide band receivers, where high-power interferers are a limiting factor.
\section*{Acknowledgments}
This work was sponsored by NSF, award number: 2030257.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Speech Emotion Recognition (SER) task is a well-studied field in the domain of affective computing and has been essential for computers understanding human's mood state \cite{anagnostopoulos2015features, AKCAY202056}. The SER has attracted many researchers' attention and been applied to plenty of fields such as automated customer service systems, smart voice assistants, human psychotherapy, and so on \cite{poria2017review, yeh2019interaction}. However, SER is an extraordinarily complicated task even for human beings because speech emotion is elusive.
Graph Neural Network (GNN) has been an active research field for the last ten years and made significant advancements in graph representation learning \cite{scarselli2008graph, bronstein2017geometric, battaglia2018relational}. GNNs broadly follow a recursive neighborhood aggregation (or message passing) scheme, where each node aggregates the feature vectors of its neighbors to compute its new feature vector \cite{xu2018representation, gilmer2017neural}. In recent years, some methods based on GNN have been applied in audio processing fields, including few-shot audio classification \cite{zhang2019few}, anti-spoofing \cite{tak2021graph} and so on. Su et al. \cite{su2020improving} proposed a framework in imposing a graph attention mechanism on a gated recurrent unit network (GA-GRU) to improve utterance-based SER. Shirian et al. \cite{shirian2021compact} constructed a compact Graph Convolution Network architecture which for the first time only utilized GNN for SER task. Liu et al. \cite{liu2021graph} proposed a novel LSTM-GIN model, which applies Graph Isomorphism Network (GIN) \cite{xu2018powerful} on LSTM outputs for global emotion modeling in the non-Euclidean space. These SER methods based on GNN all model the time sequence features of speech signals as graphs and treat the SER as a graph classification task.
There are many GNN variants with different neighborhood aggregation, such as GCN \cite{kipf2017semi}, GAT \cite{velivckovic2018graph}, PATCHY-Diff \cite{ying2018hierarchical}. In this paper, we choose GIN as the basic GNN structure because it possesses the discriminative power over other GNNs \cite{xu2018powerful}. Nevertheless, in the abovementioned papers, GNNs all adopt a single aggregator in a GNN layer, which does not extract enough information from the nodes’ neighbourhoods and thus limit their expressive power and learning abilities. Corso et al. \cite{corso2020principal} have mathematically proved the necessity of adopting multiple aggregators. We adopt weighted multiple aggregations in the GIN structure and choose three kinds of aggregators: sum, mean and softmax aggregations. We refer to our proposed GNN-based SER structure as Weighted Multiple Aggregators GIN (WMA-GIN).
Many research fields utilizing GNN, including SER, require the interaction between nodes that are not directly connected. They achieve this by stacking multiple GNN layers for long-range sequence problems. However, as the number of layers increases, the number of nodes in each node’s receptive field grows exponentially. This phenomenon is named as \(over$-$squashing\) by U. Alon and E. Yahav \cite{alon2020bottleneck}. For alleviating this problem in our network, we transform the neighbourhood relationship to the Full-Adjacent (FA) at one layer of the WMA-GIN. Furthermore, we have made the ablation experiments to demonstrate which layer of the WMA-GIN should be introduced into the FA for the SER task.
A Bi-GRU layer extracts global context sequence information before the stacked WMA-GIN layers. Moreover, multi-phase attention (MPA) is adopted to extract emotional information in underutilization. And a multi-stage loss training strategy is employed the handle the output information from different stages in our network. The experimental results show the effectiveness of our proposed method.
The contributions of this paper are summarized as follows:
(1) We propose a Weighted Multiple Aggregators GIN (WMA-GIN) for the SER task, and the experimental results show its effectiveness for obtaining more accurate information from neighbour nodes.
(2) We introduce a Full-Adjacent (FA) layer into the WMA-GIN for relieving the \(over$-$squashing\) problem. And we have explored which layer of WMA-GIN is more suitable to adopt the FA for the SER task.
(3) Our proposed WMA-GIN achieves 72.48\% of WA and 67.72\% of UA on the IEMOCAP dataset, outperforming other state-of-the-art GNN-based SER methods.
\section{Proposed Method}
In this section, we will introduce our proposed WMA-GIN network architecture, the graph construction from the audio utterances, GIN, and weighted multiple aggregators. At last, we discuss the Fully-Adjacent layer and multi-stage loss training strategy.
\subsection{WMA-GIN Architecture}
The architecture of WMA-GIN is illustrated in Fig.~\ref{proposed network}. A Bi-GRU layer is used for extracting the global context information of the features among long-range time series, and the hidden dimension of each direction is 128. Then, the output \(\mathbf{g}\) of Bi-GRU is fed into stacked WMA-GIN layers for extracting the features of higher resolution and learning the information with richer emotional characteristics. The number of WMA-GIN layers is set to 4. There are residual connections between each two adjacent WMA-GIN layers. Moreover, multi-phase attention (MPA) is employed for extracting emotional features that might be omitted from different phases of the network. Similar to self-attention \cite{vaswani2017}, each of three linear projections is performed on the input features, the output of Bi-GRU, and output of the last WMA-GIN layer, respectively, such that they are transformed as \(query\), \(key\) and \(value\). The output dimension of the linear projection is equal to the hidden size of the WMA-GIN layer.
Finally, the outputs of the Bi-GRU layer, WMA-GIN1, WMA-GIN2, WMA-GIN3, and MPA are fed into the corresponding linear layer to produce the predicted results \(a\), \(b\), \(c\) \(d\), \(e\), respectively. As shown in Figure~\ref{proposed network}, all outputs of predicted results are summed up as \(G\). Then, the predicted emotion label is obtained via a softmax layer.
\\
\noindent \textbf{Graph Construction}
Following the prior works \cite{shirian2021compact,liu2021graph}, we adopt frame-to-node transformation for our graph construction. The node feature \(\mathbf{s}_{i} \in \mathbb{R}^{h}\) is obtained from the corresponding audio frame with the feature dimension of \(h\). A graph \(\mathcal{G}=(\mathcal{V}, \mathcal{E})\), where \(\mathcal{V} \in\left\{v_{i}\right\}_{i=1}^{n}\) is the set of \(n\) nodes, and \(\mathcal{E}\) is the set of all edges of each nodes. So the feature matrix of each graph is denoted as \(\mathbf{S} \in \mathbb{R}^{n \times h}\). The adjacency matrix of \(\mathcal{G}\) is denoted by \(\mathbf{A} \in \mathbb{R}^{n \times n}\) where an element \((\mathbf{A})_{i j}\) denotes the edge weight connecting node \(v_{i}\) and \(v_{j}\). Here, each node has two neighboring nodes corresponding to the previous and following one frame. The relationship of node neighborhood is defined as a cycle construction where the node of the first time frame is connected with the last time frame.\\
\noindent \textbf{Graph Isomorphism Network (GIN)}
The basic GIN is a GNN architecture proposed by \cite{xu2018powerful} that is provably the most expressive among the class of GNNs and is as powerful as the Weisfeiler-Lehman graph isomorphism test. The GIN has achieved excellent performance for graph classification and node classification tasks. The GIN also follows a neighborhood aggregation scheme, where the representation vector of a node is computed by recursively aggregating and transforming representation vectors of its neighboring nodes \cite{xu2018powerful}. The original GIN adopts a sum aggregation strategy for aggregating information representation from the neighboring nodes. The calculation of sum aggregation is as follow:
\begin{equation}
sum _{i}\left(X^{l}\right)=\sum_{j \in N(i)} X_{j}^{l}
\end{equation}
where \(X^{l}\) denotes the nodes’ features at layer \(l\), and \(N(i)\) the neighbourhood of node \(i\).
\subsection{Weighted Multiple Aggregators (WMA)}
While a single aggregator is not enough for differentiating between neighbourhood messages. For example, a node of graph \(a\) receives two messages \($\{$0.2, 0.2$\}$\) from two neighbour nodes. And the received messages of another node of graph \(b\) are \($\{$0, 0.4$\}$\). In this case, the sum aggregator cannot distinguish between graph \(a\) and graph \(b\). For alleviating the abovementioned information confusion, we introduce two other aggregators: mean aggregator and softmax aggregator. Their calculations are as follows:
\begin{equation}
mean_{i}\left(X^{l}\right)=\frac{1}{d_{i}} \sum_{j \in N(i)} X_{j}^{l}
\end{equation}
\begin{equation}
softmax_{i}\left(X^{l}\right)=\sum_{j \in N(i)} \frac{X_{j}^{l} \exp \left(X_{j}^{l}\right)}{\sum_{k \in N(i)} \exp \left(X_{k}^{l}\right)}
\end{equation}
where \(d_{i}=|N(i)|\) which denotes the number of neighbour nodes. The messages from different nodes can be assigned different weights through softmax aggregation. During the model training, the information will constantly flow to the part with salient emotional features. As the graph structure in our proposed SER method is relatively simple where each node gets two neighbour nodes, we do not choose to scale the results of aggregations as \cite{corso2020principal} does. Finally, the node representation in WMA-GIN updates as:
\begin{equation}
\begin{aligned}
&X_{i}^{(l)}=\operatorname{MLP}^{(l)}( ( 1+\epsilon^{(l)} ) \cdot X_{i}^{(l-1)}\\
&+\alpha \sum_{j \in \mathcal{N}(i)} X_{j}^{(l-1)}+\beta \frac{1}{d_{i}} \sum_{j \in N(i)} X_{j}^{(l-1)}\\&+\gamma \sum_{j \in N(i)} \frac{X_{j}^{(l-1)} \exp X_{j}^{(l-1)}}{\sum_{k \in N(i)} \exp X_{k}^{(l-1)}})
\end{aligned}
\end{equation}
where MLP denotes Multilayer Perceptron that is a linear layer, \(X_{i}^{(l)}\) denotes feature representation of node \(i\) in the \(l-th\) hidden layer. And \(X_{i}^{(0)}\) is initialized as \(g_{i}\) which denotes the output of Bi-GRU layer. \(\epsilon\) is a learnable parameter. \(\alpha\), \(\beta\), \(\gamma\) are three hyperparameters of weights for three aggregators.
\subsection{Fully-Adjacent layer}
To ease the \(over$-$squashing\) problem in all GNNs, we modified the second WMA-GIN layer to be a Fully-Adjacent (FA) layer. The FA layer is a WMA-GIN layer in which every pair of nodes is connected by an edge. This does not change the type of WMA-GIN layer nor add trainable parameters but only changes the adjacency relationship of the nodes in a single layer. After adding the FA layer, only the second WMA-GIN layer allows the topology-aware node-representations to interact directly and consider nodes far beyond their original neighbors \cite{alon2020bottleneck}.
\subsection{Multi-stage Loss}
A multi-stage loss training strategy is adopted in our proposed SER method. The outputs of four WMA-GIN layers and the final output are defined as five different stages in the SER network. We denote them as \(b\), \(c\), \(d\), \(e\), \(G\) in Fig.~\ref{proposed network}. Generally, the stage deeper, the obtained emotional information more salient. Thus the output \(G\) is assigned with the larger weights. Cross-entropy (CE) loss is used as the loss function. Note that the multi-stage loss training strategy is only used during the training phase. The calculation of the final loss function is as follows:
\begin{equation}
Loss=\frac{\sum_{i=1}^{I} i \cdot L_{CE}}{\sum_{i=1}^{I} i}
\end{equation}
where \(i/\sum_{i=1}^{I} i\) is the weight of loss at the \(i-th\) stage and \(I\) is the number of total stages.
\section{Experiments}
\subsection{Dataset}
We performed the experiments on the Interactive Emotional Dyadic Motion Capture (IEMOCAP) dataset \cite{busso2008}, which is a popular benchmark dataset for emotion recognition. The corpus contains approximately 12 hours of data over five dyadic sessions with ten subjects. Each interaction conversation is around 5 minutes long and segmented into multiple sentences. The single categorical emotional label was assigned to every utterance, with over two of three annotators agreeing on the emotional labels. We performed the experiments on the task of four-class emotion classification for a fair comparison with other methods. The samples contain 5531 utterances (1,103 angry, 1636 happy merged with excitement, 1,084 sad, 1,708 neutral).
\subsection{Node features}
Follow two prior work \cite{liu2021graph, su2020improving}, we extract 78-dimensional frame-level Low Level Descriptors (LLDs) from \cite{schuller2010interspeech} using the openSMILE toolkit \cite{10.1145/1873951.1874246}. We also make the experiments using 128-dimensional log-Mel spectrograms as \cite{zhong2020lightweight} does. For each sample, we use a sliding window of length 25ms (with a stride length of 10ms) to extract the LLDs locally. We set each graph length to 120 as \cite{shirian2021compact} and \cite{liu2021graph} do, that means each graph contains 120 nodes. The graph label is the same as its original utterance. An utterance may be cut as several graphs because the lengths of some utterances are much longer than 120. Padding is used to make the samples of equal length.
\subsection{Experimental Settings}
In the experiments, we performed 5-fold cross-validation in the speaker-independent environment, the proportion of training set, validation set, the test set was set with 8:1:1. The Adam optimizer \cite{kingma2014adam} was used with the initial learning rate of 1e-4 and weight decay set to 1e-8. Moreover, the early stopping was set during training. The batch size was set to 128. We employ unweighted accuracy (UA) and weighted accuracy (WA) as evaluation metrics following the previous studies.
\subsection{The discussion about FA}
\begin{figure}[htbp]
\centering
\includegraphics[width=6cm]{888.eps}
\caption{The comparison among the different numbers of WMA-GIN layer along with different hidden sizes}
\label{hidden size}
\end{figure}
Fig.~\ref{hidden size} shows the UA and WA scores of our proposed method adapting the different numbers of WMA-GIN layer along with different hidden sizes. Alon et.al \cite{alon2020bottleneck} provide the minimal hidden size for different numbers of GIN layers to fit the training data perfectly. Therefore, we set the lower bound of hidden size to 64, 128, and 256 for four, five, and six WMA-GIN layers. The results of WA are all higher than that of UA, and both of them increase first and then decrease with the increase of hidden size. Generally, the more WMA-GIN layer we use, the worse the results are for WA or UA. The model suffers from \(over$-$squashing\) that results in underfitting, which prevents the model from distinguishing between different examples. As a result, we employ the Fully-Adjacent layer in our proposed model once in the following experiments, and adopt the WMA-GIN layers of 4 with the hidden size of 256 according to the result in Fig.\ref{hidden size}.
We further explore which WMA-GIN layer should adopt the FA. Tab.\ref{FA layer} shows three situations: the FA is applied at the last, penultimate, and antepenultimate layer. The model with the FA applied at the antepenultimate layer performs best. Those three situations of applying the FA layers can effectively ease information flow and prevent the \(over$-$squashing\) problem. The model with the FA applied at the antepenultimate layer achieves a better balance between the information interaction with the emotional features extraction. So, in the following experiments, the FA was all applied at the antepenultimate layer in our proposed WMA-GIN model.
\begin{table}[h]
\caption{The ablation of FA layer}\label{FA layer}
\centering
\begin{tabular}{ccc}
\hline
FA layer& WA& UA\\
\hline
Last FA& 70.67 &65.80\\
Penultimate layer& 69.68&66.41\\
Antepenultimate layer& \textbf{72.48}& \textbf{67.72}\\
\hline
\end{tabular}
\end{table}
\subsection{The ablation of Weighted Multiple Aggregators}
\begin{table}[h]
\caption{The ablation of Weighted Multiple Aggregators}\label{Multi-Aggregators}
\centering
\begin{tabular}{ccc|c|c}
\hline
\multicolumn{3}{c|}{\textbf{Aggregator weight}} & \multicolumn{1}{c|}{\multirow{2}{*}{\textbf{WA(\%)}}} & \multicolumn{1}{c}{\multirow{2}{*}{\textbf{UA(\%)}}}\\
\cline{1-3}
\multicolumn{1}{c|}{\textbf{softmax}} & \multicolumn{1}{c|}{\textbf{sum}} & \multicolumn{1}{c|}{\textbf{mean}} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{}\\
\hline
\multicolumn{1}{c|}{1/2} & \multicolumn{1}{c|}{1/4} & 1/4& 68.46& \multicolumn{1}{c}{66.52} \\
\multicolumn{1}{c|}{1/4} & \multicolumn{1}{c|}{1/2} & 1/4 & 67.67 & \multicolumn{1}{c}{65.31} \\
\multicolumn{1}{c|}{1/4} & \multicolumn{1}{c|}{1/4} & 1/2 & 67.26 & 65.30 \\
\multicolumn{1}{c|}{3/5} & \multicolumn{1}{c|}{1/5} & 1/5 & 67.88 & 65.78 \\
\multicolumn{1}{c|}{1/5} & \multicolumn{1}{c|}{1/5} & 3/5 & 70.76 & 66.24 \\
\multicolumn{1}{c|}{1/5} & \multicolumn{1}{c|}{3/5} & 1/5 & 69.31 & 65.15 \\
\multicolumn{1}{c|}{1/3} & \multicolumn{1}{c|}{1/3} & 1/3 & \textbf{72.48} & \textbf{67.72} \\
\multicolumn{1}{c|}{1} & \multicolumn{1}{c|}{0} & 0 & 71.42 & 66.59 \\
\multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{1} & 0 & 71.26 & 66.38 \\
\multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{0} & 1 & 69.86 & 65.89 \\
\hline
\end{tabular}
\end{table}
In this section, we explore different weights of three aggregators for aggregating the nodes' information more precisely from different graphs. Furthermore, we also performed experiments with three kinds of single aggregators. As shown in Tab.\ref{Multi-Aggregators}, the model performs the best when three aggregators' weights are equal. That means three aggregators are equally crucial in alleviating information confusion when aggregating node information. As for the single aggregator, the model with the softmax aggregator does the best for the softmax aggregator can allow an asymmetric message passing in the direction of the strongest signal \cite{corso2020principal}.
\subsection{Comparison with other methods}
As shown in Tab.\ref{graph-based}, the results of GCN \cite{kipf2017semi}, GAT \cite{velivckovic2018graph} are reported by \cite{liu2021graph}, and we reproduced the result of PATCHY-Diff \cite{ying2018hierarchical}. The results of other compared methods are reported in their published papers. Our proposed WMA-GIN outperforms all compared graph-based methods, especially on WA. Compared with LSTM-GIN \cite{liu2021graph} which also adopts GIN structure, WMA-GIN achieves the improvements by 12.1\% and 3.3\% on WA and UA, respectively. Without applying FA, the performance decreases. It proves that the FA layer can effectively alleviate the information compression problem for the SER task. The results in Tab.\ref{graph-based} also indicate that adopting the Multi-stage Loss, assigning different weights to the features at different stages of the network has a positive influence. And the log-Mel spectrogram features are inferior to the LLDs. Furthermore, we also compare six recent advanced methods which do not adopt any GNN structures as shown in Table \ref{non-graph-based}. WMA-GIN utilizing LLDs features outperforms all listed methods on WA.
\begin{table}[h]
\caption{Comparison with other graph-based methods}\label{graph-based}
\centering
\setlength{\tabcolsep}{1.0pt}
\begin{tabular}{lcccc}
\hline
\textbf{Graph-based}& \textbf{Feature sets}&\textbf{Para.}&\textbf{WA(\%)}& \textbf{UA(\%)}\\
\hline
GCN \cite{kipf2017semi}&78-LLDs&78K&61.16 &62.21\\
GAT \cite{velivckovic2018graph}&78-LLDs&-&60.93&62.09\\
PATCHY-Diff \cite{ying2018hierarchical}&78-LLDs&68K& 63.23& 58.71\\
GA-GRU \cite{su2020improving}&78-LLDs&- &62.27& 63.80\\
Amir et.al \cite{shirian2021compact}&35-LLDs&30K& 63.69& 59.87\\
L-GrIN \cite{shirian2021dynamic}&35-LLDs&92K& 65.50& N/A\\
LSTM-GIN \cite{liu2021graph}&78-LLDs&0.89M&64.65& 65.53\\
\hline
\noindent\textbf{WMA-GIN} & 78-LLDs&0.98M&\textbf{72.48} & \textbf{67.72}\\
\quad w/o FA & 78-LLDs&-&70.19 & 65.55\\
\quad w/o Multi-stage Loss &78-LLDs&-&70.25 & 66.47\\
\noindent\textbf{WMA-GIN}& 128-Log-Mel&0.98M& 65.61&60.26\\
\hline
\end{tabular}
\end{table}
\begin{table}[h]
\caption{Comparison with other non-graph-based methods}\label{non-graph-based}
\centering
\setlength{\tabcolsep}{4.5pt}
\begin{tabular}{lcc}
\hline
\textbf{Non-Graph-based}&\textbf{WA(\%)}& \textbf{UA(\%)}\\
\hline
SegCNN \cite{mao2019deep}&64.53 &62.34\\
LSTM + CTC \cite{zhao2019attention}& 69.00&67.00\\
Fine-tune Context Network \cite{xia2021temporal}&66.90& 65.40\\
Zhong et.al \cite{zhong2020lightweight}&70.39& 71.72\\
HNSD \cite{cao2021hierarchical}&70.50& 72.50\\
DAAE+CNN+Attention \cite{gao2021domain} &70.07& 70.67\\
\hline
\textbf{WMA-GIN}& \textbf{72.48}& 67.72\\
\hline
\end{tabular}
\end{table}
\section{Conclusions}
In this paper, we propose a network based on stacked Weighted Multiple Aggregators Graph Isomorphism Network (WMA-GIN) for SER. The experimental results demonstrate that the WMA can effectively improve the GIN structure performance. Moreover, the Full-Adjacent (FA) layer is proved to help alleviate \(over$-$squashing\) problem in the SER task. Finally, with the assistance of multi-phase attention (MPA) and multi-stage loss training strategy, WMA-GIN surpasses other graph-based methods and achieves comparable performance to some advanced non-graph-based methods. In the future, we will focus on exploring different graph construction for modeling longer speech utterances in a single graph.
\section{Acknowledgements}
This work is supported by National Natural Science Foundation of China (NSFC) (U1903213), Tianshan Innovation Team Plan Project of Xinjiang (202101642)
\bibliographystyle{IEEEtran}
|
\section{Introduction}
One of the most natural and successful ways to study an infinite object is via finite approximations to it.
A natural way to approximate an infinite group is to consider larger and larger finite quotients of it.
The object that encodes all finite quotients of a group $G$ is its profinite completion $\widehat{G}$. Indeed, this is the categorical limit of all finite quotients and the natural maps between them.
In order for $\widehat{G}$ to be a faithful approximation to $G$, the latter must be residually finite, in which case $G$ embeds in $\widehat{G}$.
We restrict ourselves to the study of these groups.
One then may wonder to what extent a residually finite group is determined by its profinite completion.
The best possible answer is known as \emph{absolute profinite rigidity}: if $\widehat{G}\cong \widehat{H}$ then $G\cong H$.
It is straightforward to see that finite groups and finitely generated abelian groups are absolutely profinitely rigid.
Beyond that, things get difficult very quickly.
There are examples of virtually cyclic groups (\cite{Baumslag}) and of finitely generated torsion-free nilpotent groups of class 2 (\cite{GrunewaldScharlau}) that are not absolutely profinitely rigid.
Perhaps the boldest question in this area is that posed by Remeslennikov (\cite[Problem 5.48]{Kourovka}): \emph{Are free non-abelian groups absolutely profinitely rigid?
}
The answer to this question still seems remote, given current knowledge.
Nevertheless, some progress has been made on variants of this question: free groups and surface groups are determined, among limit groups, by their profinite completion (\cite[Corollary D]{Wi18}, \cite[Theorem 2]{Wilton_comptes}); certain
Kleinian groups and triangle groups are absolutely profinitely rigid (\cite{BMRS18, BMRS20}).
Here we study another variant of this question, namely, whether one can detect a free factor of a group from its profinite completion.
It is an exercise in the definition of free profinite products (\cite[9.1.1]{RibesZalesskii}) to show that if $G$ is the free product of subgroups $H, K$ then $\widehat{G}$ is the free profinite product of the closures $\overline{H},\overline{K}$.
One might hope that the converse is also true:
\begin{Question}\label{qu:recognise_free_factors_profinitely}
Let $G$ be a finitely generated residually finite group and $H\leq G$.
If $\overline{H}$ is a free profinite factor of $\widehat{G}$, must $H$ be a free factor of $G$?
\end{Question}
This question is surprisingly tricky.
The only results so far require $G$ to be a free group, and then the answer is ``yes''.
This case was first shown by Parzanchevski and Puder \cite{PP15} and later, using different methods, by Wilton \cite{Wi18}.
Each of the two proofs rely on beautiful and deep results.
The first uses detailed calculations for the expected number of common fixed points of the image of a finitely generated $H$ under a uniformly random homomorphism $G\rightarrow \mathrm{Sym}(n)$, for each $n$.
In particular, Puder and Parzanchevski obtain that if $w\in G$ is not a primitive word then, for large enough $n$, the average number of fixed points of the image of $w$ under a uniformly random homomorphism $G\rightarrow \mathrm{Sym}(n)$ is strictly larger than $1$.
Wilton's result is a consequence of another result of his that partially answers a question of Gromov: if the fundamental group of a non-trivial graph of free groups with
cyclic edge groups is one-ended and hyperbolic then it contains a surface subgroup.
Here we use entirely different methods to answer Question \ref{qu:recognise_free_factors_profinitely} positively when $G$ is virtually free.
We need some notation.
Let $H$ be a subgroup of a finitely generated group $G$ and $\gamma:H\rightarrow P$ a homomorphism to a finite group $P$.
If $H$ is a free factor of $G$, then the number $h(G,H,\gamma,P)$ of homomorphisms $\tilde{\gamma}:G\rightarrow P$ that extend $\gamma$ is independent of the choice of $\gamma$; it is simply the number of homomorphisms from the free complement of $H$ to $P$.
It turns out that the converse is also true when $G$ is virtually free.
\begin{teo}\label{teo:vfree}
Let $G$ be a finitely generated virtually free group. Let $H$ be a finitely generated subgroup of $G$ and $\overline H$ the closure of $H$ in $\widehat G$.
The following are equivalent:
\begin{enumerate}[label=\textnormal{(\alph*)}]
\item $H$ is a free factor of $G$;
\item for every finite group $P$, the number $h(G,H,\gamma,P)$ is constant on $\Hom (H, P)$;
\item $\overline H$ is a free profinite factor of $\widehat G$.
\end{enumerate}
\end{teo}
We have already explained, (a)$\implies$ (b).
Now notice that $\Hom(G,P)$ can be identified with the continuous morphisms $\Hom_{\text{cont.}}(\widehat{G},P)$, for every finite group $P$.
If $\widehat{H}=\overline{H}$ then $\Hom(H,P)$ can also be identified with $\Hom_{\text{cont.}}(\overline{H},P)$.
In this case condition (b) is equivalent to
\smallskip
(b') for every finite group $P$, the number $h(\widehat{G},\overline{H},\gamma,P)$ is constant on $\Hom_{cont.}(\overline{H},P)$.
\smallskip
As a first step we show in Proposition \ref{prop:profinitefactor} that (b') is equivalent to (c).
This is a completely general result.
The only condition in passing from (b) to (b') is that $H$ be closed in the profinite topology of $G$. This is ensured, for instance, if $G$ is subgroup separable (every finitely generated subgroup is closed in the profinite topology), also termed LERF.
The meat of the paper is the proof of (c)$\implies$ (a).
Condition (c) implies that there is a non-inner derivation (or crossed-homomorphism) $d:\widehat{G}\rightarrow k[[\widehat{G}]]$ whose kernel is precisely $\overline{H}$, where $k$ is a finite commutative ring.
This, and the related definitions, are explained in Section \ref{sec:prelims}.
The analogous statement in the discrete setting is almost, but not quite, equivalent to (a).
In fact, a result of Dicks following Dunwoody \cite{Di81} states that if there is a derivation $d':G\rightarrow M$ to a projective $k[G]$-module $M$ with kernel $H$, then $G$ splits as a fundamental group of a finite graph of groups, one of whose vertex groups is $H$, and all of whose edge groups are finite, with conditions on their sizes related to the ring $k$.
We must therefore do two things: first obtain some $d'$ from $d$, and second show that the graph of groups decomposition of $G$ actually makes $H$ a free factor.
The second thing is done in Section \ref{sec:Dicksproof}.
The first thing is done in Section \ref{sec:limits_projective_modules}, where it is shown (Corollary \ref{locvfree}) that, for $G$ virtually free, and $k=\mathbb{Z}/n\mathbb{Z}$ with square-free $n$ the left $k[G]$-module $k[[\widehat G]]$ is isomorphic to a direct union of projective $k[G]$-modules.
This means, as $G$ is finitely generated, that the image $d(G)$ is contained in a finitely generated projective $k[G]$-module.
In order to obtain Corollary \ref{locvfree}, we give in Theorem \ref{teo:Rhatflat_iff_Rcoherent} a criterion for the profinite completion $\widehat{R}$ of a ring $R$ satisfying certain conditions to be a flat $R$-module.
By a result of \cite{CoFR}, the criterion implies that every finitely generated submodule of $\mathbb{F}_p[[\widehat{F}]]$ is free, where $F$ is a finitely generated free group.
The main theorem could be generalised to wider classes of groups if we can answer the following question.
\begin{Question}
For which finitely generated groups $G$ is
the profinite completion of $\mathbb{F}_p[G]$ a subgroup of a direct union of projective $\mathbb{F}_p[G]$-modules?
\end{Question}
Putting together Theorem \ref{teo:Rhatflat_iff_Rcoherent} and deep results on virtually polycyclic groups (\cite{Jategaonkar,Roseblade73, Roseblade}) yields that the completed group ring $\mathbb{F}_p[[\widehat{G}]]$ of a virtually polycyclic group $G$ is a flat $\mathbb{F}_p[G]$-module.
Although this is unrelated to the rest of the paper, it seems an interesting phenomenon in its own right.
\begin{Question}
For which finitely generated groups $G$ is
the profinite completion of $\mathbb{F}_p[G]$ a flat $\mathbb{F}_p[G]$-module?
\end{Question}
\section*{Acknowledgments}
The authors acknowledge partial financial support from Spain's Ministry of Science and Innovation, through grant PID2020-114032GB-I00 and the Severo Ochoa Programme for Centres of Excellence in R \& D, (CEX2019-000904-S).
\section{Preliminaries}\label{sec:prelims}
All rings in this paper have a unit, denoted 1, and all ring homomorphisms send the unit to the unit.
By a module we will mean a left module if we do not say the opposite.
A nonempty class of finite groups $\mathcal C$ is
a {\bf pseudovariety} if it is closed under taking subgroups, homomorphic images and
finite direct products.
The pro-$\mathcal C$ completion of a group $G$ is denoted by ${G}_{\widehat {\mathcal C}}$ and by $\widehat{G}$ we denote the profinite completion of $G$.
The closure of a subgroup $H\leq G$ inside $\widehat{G}$ is denoted $\overline{H}$.
The abstract free product of groups $G_1,G_2$ is denoted by $G_1\ast G_2$, while $G_1\coprod_{\mathcal C} G_1$ denotes the free pro-$\mathcal C$ product of two pro-$\mathcal C$ groups $G_1, G_2$.
%
Recall that $G_1\coprod_{\mathcal C} G_2$ is characterized up to isomorphism by the following universal property: there are continuous homomorphisms $\varphi_i:G_i\rightarrow G_1\coprod_{\mathcal C} G_2$ ($i=1,2$), and for any pair of continuous homomorphisms $\psi_i: G_i\rightarrow P$ ($i=1,2$) to a finite group $P$ in $\mathcal{C}$ there is a unique continuous homomorphism $\psi:G_1\coprod_{\mathcal C} G_2 \rightarrow P$ such that $\psi\circ\varphi_i=\psi_i$, $i=1,2$.
The basic properties of free pro-$\mathcal C$ products can be found in \cite[Section 9]{RibesZalesskii}.
If $\mathcal C$ is the pseudovariety of all finite groups, then we write $A\coprod B$ instead of $A\coprod_{\mathcal C} B$.
Let $G$ be a pro-$\mathcal C$-group and $H$ a closed subgroup of $G$. We say that $H$ is a {\bf pro-$\mathcal C$ free factor} of $G$ if there exists a closed subgroup $K$ of $G$ such that $G= H\coprod_{\mathcal C} K$.
If $k$ is a ring and $G$ is a group, we denote by $k[G]$ the group ring of $G$ over $k$.
The {\bf augmentation ideal} of $k[G]$ is the kernel of the canonical map $k[G]\to k$, and it is denoted by $I_{k[G]}$.
If $G$ is a profinite group and $k$ is a finite ring the {\bf completed group ring} $k[[G]]$ of $G$ over $k$ is the inverse limit of $k[G/U]$, where $U$ is a normal open subgroup of $G$.
The {\bf augmentation ideal} of $k[[G]]$ is the kernel of the canonical map $k[[G]]\to k$, and it is denoted by $I_{k[[G]]}$.
It is a free $k$-module on the profinite space $\{g-1 \mid g\in G\}$ and, if $T$ is a profinite space generating $G$ and containing the identity, then $I_{k[[G]]}$ is generated as a profinite $k[[G]]-$module by $\{t-1 \mid t\in T\}$ (see \cite[6.3.2]{RibesZalesskii}).
If $A$ is a closed subgroup of $G$ we denote by $I_{k[[A]]}^G$ the left closed ideal of $I_{k[[G]]}$ generated by $I_{k[[A]]}$.
Let $G$ be a group, $k$ a ring and $M$ a $k[G]$-module.
A \textbf{derivation} or \textbf{crossed homomorphism} is a map $d:G\rightarrow M$ such that $d(xy)=xd(y)+d(x)$ for all $x,y\in G$.
The derivation is \textbf{inner} if there is some $m\in M$ such that $d=\mathrm{ad_m}: x\rightarrow (1-x)m$.
Let $\Gamma$ be a profinite group, $k$ a profinite ring and $A$ a discrete $k[[\Gamma]]$-module.
A \textbf{continuous derivation} $d:\Gamma\rightarrow A$ is a continuous map that satisfies $d(gh)=gd(h)+d(g)$ for all $g,h\in\Gamma$.
Denote by $\mathrm{Der}(\Gamma, A)$ the set of continuous derivations from $\Gamma$ to $A$.
The following Lemma is well-known to experts, but we could not find a reference with the exact statement.
It uses similar ideas to \cite[Proposition 9.3.8]{RibesZalesskii}.
\begin{lem}\label{lem:freeprofprod_implies_idealsdecompose}
Let $k$ be a finite ring and $\mathcal C$ an extension-closed pseudovariety of finite groups containing the group $(k,+)$. Let $\Gamma$ be a pro-$\mathcal C$ group and assume that $\Gamma=\Delta_1\coprod_{\mathcal C} \Delta_2$.
Then $$I_{k[[\Gamma]]}=I_{k[[\Delta_1]]}^{\Gamma}\oplus I_{k[[\Delta_2]]}^{\Gamma}.$$
\end{lem}
\begin{proof}
To simplify notation, write $I:=I_{k[[\Gamma]]}$ and $J_i:=I_{k[[\Delta_i]]}^{\Gamma}$, for $i=1,2$.
Since $\Gamma$ is generated by $\Delta_1$ and $\Delta_2$, the augmentation ideal $I$ is generated as a profinite $k[[\Gamma]]$-module by $\{\delta-1 \mid \delta \in \Delta_1 \cup \Delta_2\}$.
So evaluating in $I$ gives a canonical surjective $k[[\Gamma]]$-morphism $\pi: J_1\oplus J_2 \rightarrow I$.
To prove the claim, we will find an inverse $\rho:I\rightarrow J_1\oplus J_2$ of $\pi$.
For $i=1,2$ and each open normal subgroup $U\unlhd_o\Gamma$, denote by $J_{i,U}:=I_{k[[\Delta_iU/U]]}^{\Gamma/U}$ the image of $J_i$ in $k[\Gamma/U]$.
Note that $J_1\oplus J_2=\varprojlim_{U\unlhd_o\Gamma}J_{1,U}\oplus J_{2,U}$.
Since $\Gamma$ acts on $J_{1,U}\oplus J_{2,U} \in \mathcal{C}$, the group $H_U:=(J_{1,U}\oplus J_{2,U})\rtimes \Gamma$ is a pro-$\mathcal C$ group.
For $i=1,2$, define $\phi_{i,U}:\Delta_i\rightarrow H_U$ by $\delta \mapsto (\delta U- U, \delta )$ where $\delta U-U \in J_{i,U}$.
By definition of $J_{i,U}$, these are continuous homomorphisms.
The universal property of free pro-$\mathcal C$ products yields a unique continuous homomorphism $\psi_U:\Gamma \rightarrow H_U$ extending $\phi_{1,U}$ and $\phi_{2,U}$.
Denote, respectively, by $p_J$ and $p_{\Gamma}$ the projections of $H_U$ to $J_{1,U}\oplus J_{2,U}$ and $\Gamma$.
Since $p_{\Gamma}\circ\phi_{i,U}=\iid_{\Delta_i}$, the universal property of free pro-$\mathcal C$ products applied to $\iid_{\Gamma}$ ensures that $p_{\Gamma}\circ\psi_U=\iid_{\Gamma}$.
Now define $\rho_U:I\rightarrow J_{1,U}\oplus J_{2,U}$ by $\gamma -1 \mapsto p_J(\psi_U(\gamma))$ for $\gamma \in \Gamma$ and extending by $k$-linearity.
This is indeed a profinite $k[[\Gamma]]$-module morphism because, as $\psi_U$ is a homomorphism,
for every $\gamma, \epsilon\in \Gamma$ we have
\begin{align*}
\rho_U(\epsilon(\gamma -1))&=\rho_U(\epsilon\gamma-1 - (\epsilon -1)) =p_J(\psi_U(\epsilon\gamma))-p_J(\psi_U(\epsilon))\\
&=p_{\Gamma}(\psi_U(\epsilon))p_{J}(\psi_U(\gamma))+p_J(\psi_U(\epsilon))-p_J(\psi_U(\epsilon)) \\
&=\epsilon p_J(\psi_U(\gamma)) =\epsilon \rho_U(\gamma).
\end{align*}
Let $\rho:I\rightarrow J_1\oplus J_2$ be the map given by the universal property of inverse limits for $J, \{J_{1,U}\oplus J_{2,U} \mid U\unlhd_o\Gamma\}$ and $\{\rho_U\mid U\unlhd_o \Gamma\}$.
Observe that, for $i=1,2,$ and every $U\unlhd_o\Gamma$ and $\delta \in \Delta_i$, we have $\rho_U(\pi(\delta-1))=\rho(\delta-1)=\delta U-U$, so $\rho_U\circ\pi$ is the projection map $J_1\oplus J_2 \rightarrow J_{1,U}\oplus J_{2,U}$ and thus $\rho\circ \pi=\iid_{\Gamma}$, as required.
\end{proof}
Recall that a \textbf{supernatural number} is a formal product
$n=\prod_{p}p^{n(p)}$ where $p$ runs through all prime numbers and $n(p)\in \mathbb{N}\cup \{\infty\}$.
By convention, $n<\infty$ and $\infty+n=n+\infty=\infty+\infty=\infty$ for all $n\in\mathbb{N}$.
A supernatural number $m=\prod_p p^{m(p)}$ is said to \textbf{divide} another one $n=\prod_p p^{n(p)}$ if $m(p)\leq n(p)$ for every prime $p$.
The \textbf{lowest common multiple} of a collection of supernatural numbers $\{n_i=\prod_{p}p^{n(i,p)} \mid i\in I\}$ is defined as $\lcm\{n_i : i\in I\}:=\prod_p p^{n(p)}$ where $n(p)=\max\{n(i,p) \mid i\in I\}$.
If $H$ is a closed subgroup of a profinite group $\Gamma$, we put $|\Gamma:H|=\lcm\{|\Gamma:HN|\colon N\unlhd_o \Gamma\} $.
Let $\Gamma$ be a profinite group with closed subgroups $A,B\leq \Gamma$ and $p$ a prime number.
The subgroup $A$ acts continuously on $\mathbb{F}_p[[\Gamma/B]]=\varprojlim_{N\unlhd_o\Gamma}\mathbb{F}_p[\Gamma/BN]$ by left coset multiplication: $am=a(m_N)_{N\unlhd_o\Gamma}=(am_N)_{N\unlhd_o\Gamma}$ where $$am_N=a\left(\sum_{gBN\in\Gamma/BN}f_{N,g}gBN\right)=\sum_{gBN\in\Gamma/BN}f_{N,g}agBN, \quad f_{N,g}\in \mathbb{F}_p.$$
The above then means that, for each $N\unlhd_o \Gamma$, $\mathbb{F}_p[\Gamma/BN]$ decomposes as an $A$-module into $$\mathbb{F}_p[\Gamma/BN]\cong \prod_{AgBN\in A\backslash\Gamma/B}\mathbb{F}_p[AgBN]\cong\prod_{AgBN\in A\backslash \Gamma/B}\mathbb{F}_p[A/A\cap B^gN] $$
(since we are using left cosets, conjugation is done on the left: $B^g=gBg^{-1}$).
Taking the inverse limit, we obtain that the $A$-module (and even the $\mathbb{F}_p[[A]]$-module) $\mathbb{F}_p[[\Gamma/B]]$ decomposes as
\begin{equation}\label{eq:G/B as prod of A orbits}
\mathbb{F}_p[[\Gamma/B]]\cong \prod_{AgB\in A\backslash \Gamma / B} \mathbb{F}_p[[A/A\cap B^g]].
\end{equation}
Denote by $\mathbb{F}_p[[\Gamma/B]]^A$ the set $\{m\in \mathbb{F}_p[[\Gamma/B]]:\ am=m \textrm{\ for every\ }a\in A\}$ of fixed points of $A$.
By \eqref{eq:G/B as prod of A orbits}, this set is reduced to 0 if and only if $\mathbb{F}_p[[A/A\cap B^g]]^A=\{0\}$ for every $g\in \Gamma$.
This observation will help in proving the following.
\begin{lem}\label{lem:no_fixed_points_modH}Let $\Gamma$ be a profinite group with closed subgroups $A,B\leq \Gamma$ and $p$ a prime number.
Then $\mathbb{F}_p[[\Gamma/B]]^A=\{0\}$ if and only if
$p^{\infty}$ divides $ |A:A\cap B^g | $
for every $g\in \Gamma$.
\end{lem}
In order to prove the lemma we need the following result on finite groups.
\begin{lem}\label{finitefix}
Let $G$ be a finite group, $H\leq G$ and $N\unlhd G$.
The image of the map
$\mathbb{F}_p[G/H]^G\to \mathbb{F}_p[G/NH]^G$ is trivial if and only if $p$ divides $|NH:H|$.
\end{lem}
\begin{proof}
Observe that every element of $\mathbb{F}_p[G/H]^G$ is a constant function $\displaystyle{f\sum_{gH\in G/H} gH}$ for some $ f\in \mathbb{F}_p$ and that its image in
$\mathbb{F}_p[G/NH]^G$ is $\displaystyle{\sum_{gNH\in G/NH}(|NH:H|f)gNH}$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:no_fixed_points_modH}]
By the observation made before the statement, we can suppose without loss of generality that $A=\Gamma$.
For ease of exposition, assume also that $\Gamma$ is countably based, so that $\mathbb{F}_p[[\Gamma/B]]=\varprojlim_{i\in\mathbb{N}}\mathbb{F}_p[\Gamma/BN_i]$ where $\Gamma = N_0\geq N_1\geq \dots$ is a countable chain of open normal subgroups such that $\bigcap_{i\in\mathbb{N}}N_i=\{1\}$.
The general case is not difficult, just more tedious to notate.
Assume first that $p^{\infty}$ divides $|\Gamma:B|$.
Let $m=(m_i)_{i\in\mathbb{N}} \in \mathbb{F}_p[[\Gamma/B]]^\Gamma$, so $m_i\in \mathbb{F}_p[[\Gamma/BN_i]]^\Gamma$
for each $i\in\mathbb{N}$.
Since $p^{\infty}$ divides $|\Gamma:B|$, $p$ divides $|BN_i:BN_{i+1}|$ for infinitely many values of $i$. By Lemma \ref{finitefix}, for these $i$, $m_i=0$. Thus, $m=0$.
Conversely, if the $p$-part of $|\Gamma:B|$ is finite, there must be some $k$ such that $p$ does not divide $|BN_k:BN_{i}|$ for all $i\ge k$.
Put $$m_i= \frac 1{ |BN_k:BN_{i}|}\sum_{gBN_i\in \Gamma/BN_i} gBN_i \in \mathbb{F}_p[\Gamma/BN_i] \textrm{\ and \ } m=(m_i)\in \mathbb{F}_p[[\Gamma/B]].$$ It is clear that $0\ne m\in \mathbb{F}_p[[\Gamma/B]]^\Gamma$.
\end{proof}
We finish the section with another auxiliary result that we will need later.
\begin{lem}\label{lem:chain_char_sgps}
Let $\mathcal{C}$ be a pseudovariety of finite groups and $G$ a finitely generated pro-$\mathcal{C}$ group.
Then there is a chain $G>N_1>N_2>\dots$ of open characteristic subgroups of $G$ with trivial intersection such that for every $i$ and $\phi_i\in\Sur(G,G/N_i)$ we have that $\ker\phi_i=N_i$.
\end{lem}
\begin{proof}
For each $i\in\mathbb{N}$, let $N_i$ be the intersection of all open normal subgroups of $G$ of index at most $i$.
Since $G$ is finitely generated, there are finitely many such subgroups, so $N_i$ is a closed subgroup of finite index and therefore open.
By construction, each $N_i$ is characteristic and they form a descending chain whose intersection is trivial.
Now, for any $i\in\mathbb{N}$, if $\phi\in\Sur(G,G/N_i)$, then $N_i\leq \ker \phi$ because $\phi$ is a homomorphism, and $\ker\phi \leq N_i$ as it is onto $G/N_i$.
\end{proof}
\section{Free profinite factors}
Throughout this section, $\mathcal{C}$ denotes an extension-closed pseudovariety of finite groups.
Subgroups of pro-$\mathcal{C}$ groups are closed and homomorphisms are continuous, unless stated otherwise.
The main result of this section is a general criterion for a given closed subgroup $H$ of a pro-$\mathcal C$ group $G$ to be a free pro-$\mathcal{C}$ factor in terms of extending homomorphisms $H\rightarrow P\in \mathcal{C}$ to $G$.
It is the key to showing that (b) and (c) in Theorem \ref{teo:vfree} are equivalent, as explained in the introduction.
Recall that given $H\leq G$ and a homomorphism $\gamma:H\rightarrow P$ to a finite group $P\in \mathcal{C}$, we denote by $h(G,H,\gamma,P)$ the number of homomorphisms $\tilde{\gamma}:G\rightarrow P$ such that $\tilde{\gamma}|_H=\gamma$.
The number of epimorphisms $G\twoheadrightarrow P$ that extend $\gamma$ is denoted by $e(G,H,\gamma,P)$.
\begin{pro}\label{prop:profinitefactor}
Let $G$ be a finitely generated pro-$\mathcal C$ group and $H$ a closed finitely generated subgroup of $G$.
Then $H$ is a free pro-$\mathcal C$ factor of $G$ if and only if for every group $P$ from $\mathcal C$, the number $h(G,H,\gamma,P)$ is independent of the choice of $\gamma\in\Hom (H, P)$.
\end{pro}
In the language of \cite{PP15}, the result says that $H$ is a free pro-$\mathcal{C}$ factor of $G$ if and only if $H$ is measure-preserving, meaning that for every $P\in \mathcal{C}$, if $\varphi\in \Hom(G,H)$ is a random homomorphism chosen uniformly, then $\varphi_{|H}$ is uniformly distributed in $\Hom(H,P)$.
In order to prove Proposition \ref{prop:profinitefactor} we will study whether an isomorphism between closed subgroups of a profinite group can be extended to an automorphism of the ambient group.
\begin{pro} \label{prob}
Let $G$ be a finitely generated pro-$\mathcal C$ group and $\alpha:H_1\to H_2$ a continuous isomorphism between two closed subgroups of $G$.
Then the following are equivalent.
\begin{enumerate}
\item[(a)] there exists $\tilde \alpha\in \Aut(G)$ such that $\tilde \alpha_{|H_1}=\alpha$;
\item [(b)] for every group $P\in \mathcal C$ and every $\gamma\in \Hom (H_2,P)$,
$$h(G,H_1,\gamma\circ\alpha,P)=h(G,H_2,\gamma,P);$$
\item[(c)] for every group $P\in \mathcal C$ and every $\gamma\in \Hom (H_2,P)$, $$e(G,H_1,\gamma\circ\alpha,P)=e(G,H_2,\gamma,P);$$
\item [(d)] for every group $P\in \mathcal C$ and every $\gamma\in \Hom (H_2,P)$,
$$ \textrm{if $e(G,H_2,\gamma,P)\ne 0$, then $e(G,H_1,\gamma\circ\alpha,P)\ne 0$}.$$\end{enumerate}
\begin{rem*}
Since $G$ is finitely generated, by a well-known result of Nikolov and Segal \cite{NS07}, any homomorphism in $\Hom(G,P)$ is continuous. Thus, for non-continuous $\gamma$ the conditions (b), (c) and (d) hold trivially as all numbers involved are 0.
\end{rem*}
\end{pro}
\begin{proof} The implications $(a)\Longrightarrow (b)$ and $(c)\Longrightarrow (d)$ are immediate.
For a homomorphism $\beta\in \Hom(H, P)$ and a subgroup $\im \beta \le Q\le P$, we denote by $\beta^Q$
the corestriction of $\beta$ to $Q$; that is, the homomorphism in $\Hom(H, Q)$ such that $\beta^Q(h)=\beta(h)$ for every $h\in H$.
The implication $(b)\Longrightarrow (c)$ is proved by induction on the order of $P$ and taking into account that
$$h(G,H_1,\gamma\circ\alpha,P)=
\sum_{Q\le P} e(G,H_1,\gamma^Q\circ \alpha,Q) \textrm{\ and\ } h(G,H_2,\gamma,P )=
\sum_{Q\le P} e(G,H_2,\gamma^Q,Q).$$
Let us prove the implication $(d)\Longrightarrow (a)$.
By Lemma \ref{lem:chain_char_sgps}, there is a chain $G>N_1>N_2>\dots$ of characteristic open subgroups of $G$ with trivial intersection such that for each $i$, if $\phi_i\in\Sur(G,G/N_i)$ then $\ker \phi_i=N_i$.
For each $i\in\mathbb{N}$, put $P_i=G/N_i$, denote by $\delta_i:G\rightarrow P_i$ the canonical map and by $\gamma_i=(\delta_i)_{|H_2}$ the natural map $H_2\rightarrow P_i$.
By assumption, there exists $\phi_i\in\Sur(G,P_i)$ such that $\phi_{|H_1}=\gamma_i\circ \alpha$.
Since $\ker\phi_i=N_i$ we obtain an automorphism of $P_i$, denoted $\overline{\phi_i}$, which satisfies that $\overline{\phi_i}\circ(\delta_i)_{|H_1}=\gamma_i\circ\alpha$.
This shows that the set $S_i=\{\tau\in\Aut(P_i): \tau\circ(\delta_i)_{|H_1}=\gamma_i\circ\alpha\}$ is not empty.
Now, for $j>i$, the maps $\delta_{ji}:P_j\rightarrow P_i$ induce maps $S_j\rightarrow S_i$ that turn $(S_i)_{i\in\mathbb{N}}$ into a directed system of non-empty sets.
Their inverse limit $S=\varprojlim_{i}S_i$
is therefore not empty and it is easy to check that any $\tilde{\alpha}\in S$ is an automorphism of $G$ such that $\tilde{\alpha}_{|H_1}=\alpha$, as required.
\end{proof}
\begin{proof} [Proof of Proposition \ref{prop:profinitefactor}]
The ``only if "part is clear. Let us show the ``if " part.
Let us first assume that $H$ is finite.
Let $P$ be a group from $\mathcal C$ and $\gamma\in \Hom (H, P)$.
Then, since $h(G,H,\gamma,P)$ is constant on $\Hom (H, P)$, \begin{equation}\label{eq:numberhoms}
h(G,H,\gamma,P)\cdot |\Hom(H,P)|=|\Hom(G, P)|.
\end{equation}
Let $\alpha:K\to H$ be an isomorphism, $\gamma\in \Hom(H,P)$ and consider the free pro-$\mathcal C$ product $\Gamma=G\coprod_{\mathcal C} K$.
By the universal property, every homomorphism $\Gamma \rightarrow P$ is determined by two homomorphisms $G\rightarrow P$ and $K\rightarrow P$.
This observation gives the first and last equalities in the following:
\begin{multline*}
h(\Gamma,H,\gamma,P)=h(G,H,\gamma,P)\cdot |\Hom(K,P)|\\=h(G,H,\gamma,P)\cdot |\Hom(H,P)| \overset{\eqref{eq:numberhoms}}{=} |\Hom(G, P)|=h(\Gamma,K,\gamma\circ\alpha,P).
\end{multline*}
By Proposition \ref{prob}, there exists $\tilde \alpha\in \Aut(\Gamma)$ that extends $\alpha$.
In particular, there exists a subgroup $G^\prime$ in $\Gamma$ such that $\Gamma=G^\prime\coprod_{\mathcal C} H$.
By \cite[Corollary 7.1.3 (a)]{Ribesprofinitetree}, there exists $g\in\Gamma$ such that $K^g\le G^\prime$ or $K^g\le H$. The second possibility cannot occur because the normal subgroup of $\Gamma$ generated by $K$ has trivial intersection with $H$.
Thus $K^g\le G^\prime$, and so,
$G\cong \Gamma/\overline{ \langle K^{\Gamma}\rangle}\cong (G^\prime/\overline{ \langle K^{ G^\prime}\rangle})\coprod_{\mathcal C} H.$
Hence $H$ is a free pro-$\mathcal C$ factor of $G$.
Now assume that $H$ is arbitrary. Let $N$ be an open normal subgroup of $G$.
We put $$H_N^\prime=H/(N\cap H) \textrm{\ and \ } G_N^\prime =G/\overline{ \langle (N\cap H)^{ G}\rangle}.$$ We see $H_N^\prime$ as a subgroup of $G_N^\prime$. Let $P\in \mathcal C$ and $\gamma\in \Hom (H_N^\prime, P)$.
We denote by $\tilde \gamma \in \Hom (H, P)$ the canonical lift of $\gamma$.
Then we have that
$h(G_N^\prime, H_N^\prime,\gamma,P)=h(G,H,\tilde \gamma,P)$.
Hence $h(G_N^\prime, H_N^\prime,\gamma,P)$ is constant on $\Hom (H_N^\prime, P)$.
Using what we have proved before, $H_N^\prime$ is a free pro-$\mathcal C$ factor of $G^\prime_N$. Let $d$ be the number of generators of $G/\overline{H^G}$. Then we have just proved that the set
$$S_N=\{(h_1,\ldots, h_d)\in (G^\prime_N)^d\colon G^\prime_N=\overline {\langle h_1,\ldots, h_d\rangle }\coprod_{\mathcal C} H^\prime_N\}$$ is not empty.
Observe that $\{S_N\}$ form an inverse system of compact sets. Thus, there exists $(g_1,\ldots, g_d)\in \varprojlim_N S_N\subset G^d$. It is clear that $G=\overline {\langle g_1,\ldots, g_d\rangle }\coprod_{\mathcal C} H$.
\end{proof}
\section{Profinitely flat modules}\label{sec:flat_rings}
In this section we establish a connection between coherence of rings and flatness of profinite completions. In the context of commutative rings the relation between completions, flatness and coherence was studied in \cite{Ha11}.
Recall that $R$ is called {\bf right (resp. left) coherent} if every finitely generated submodule of a free right (resp. left) $R$-module is finitely presented.
It is \textbf{coherent} if it is both right and left coherent.
If $M$ is a (right/left) $R$-module, its {\bf profinite completion} $\widehat{M}$ is the (right/left) $R$-module $\varprojlim_{N\leq_f M} M/N$ where $N$ ranges over all submodules of $M$ of finite index (as a subgroup). We denote by $\delta_M : M\to \widehat M$ the natural map from the group to its profinite completion. If $M$ is residually finite we will not distinguish between the elements of $m\in M$ and of $\delta_M(m)\in \widehat M$.
Observe that if $M$ is a left (right) $R$-module, then $\widehat M$ is a left (right) $\widehat R$-module: given $r=(r_I)$, where $r_I\in R/I$ and $I$ is an ideal of $R$ of finite index, and $m=(m_N)\in \widehat{M}$, where $m_N\in M/N$ and $N$ is a submodule of $M$ of finite index, define $ r\cdot m=( r_{\Ann(M/N)}\cdot m_N)$ ($ m\cdot r=(m_N\cdot r_{\Ann(M/N)})$. In the case $M=R$, the completion $\widehat{R}$ is an $R$-bimodule.
The goal of this section is to prove the following:
\begin{teo}\label{teo:Rhatflat_iff_Rcoherent}
Let $R$ be a finitely generated residually finite ring such that all finitely presented right $R$-modules are residually finite.
The following are equivalent:
\begin{enumerate}[label=\textnormal{(\arabic*)}]
\item\label{it:Rhatflat+FPMRF} The profinite completion $\widehat{R}$ of $R$ is a flat left $R$-module.
\item\label{it:Rcoherent+FPMRF} $R$ is right coherent.
\end{enumerate}
\end{teo}
Before proving the theorem, we need some auxiliary results.
\begin{lem}[See Proposition 3.2.5 of \cite{RibesZalesskii}] \label{indpro} Let $ A\to B\to C\to 0$ be an exact sequence of right $R$-modules.
Then the induced sequence of right $\widehat R$-modules
$ \widehat A\to \widehat B \to \widehat C\to 0$ is also exact.
\end{lem}
\begin{lem} \label{lem:alpha}Let $M$ be a right $R$-module and $\alpha_M: M\otimes_R\widehat R \to \widehat M$, $m\otimes r \mapsto \delta_M(m)\cdot r$ the natural map of right $\widehat R$-modules.
\begin{enumerate}
\item If $M$ is finitely generated, then $\alpha_M$ is onto.
\item If $M$ is finitely presented, then $\alpha_M$ is an isomorphism.
\item A (right) $R$-module morphism $\gamma: M_1\to M_2$ induces the following commutative diagram
$$\begin{array}{ccc}
M_1\otimes_R \widehat R &\xrightarrow{\gamma\otimes \Id} & M_2\otimes_R\widehat R\\
{\Big\downarrow}{\alpha_{M_1}} &&{\Big\downarrow}{\alpha_{M_2}}\\
\widehat{M_1} &\xrightarrow{\widehat \gamma} &\widehat {M_2}\end{array} .$$
\end{enumerate}
\end{lem}
\begin{proof}
Suppose that $M$ is finitely generated, say $M=m_1R + \dots + m_dR$.
Then $\alpha_M(M\otimes 1)=\delta_M(M)$ and $\alpha_M(M \otimes_R \widehat{R})=\delta_M(m_1)\widehat R + \dots + \delta_M(m_d)\widehat{R}$.
As each $\delta_M(m_i)\widehat{R}$ is closed in $\widehat M$, their sum is a closed subset of $\widehat M$ that contains $\delta_M(M)$, so must be all of $\widehat M$.
Let $M$ be finitely presented, with free presentation $R^e\rightarrow R^d \rightarrow M \rightarrow 0$ and $e, d$ finite.
Tensoring with $\widehat R$ and using Lemma \ref{indpro} we obtain the following commutative diagram with exact top and bottom rows:
$$\begin{array}{cccccc}
R^e\otimes_R \widehat{R}& \rightarrow & R^d\otimes_R \widehat{R} &\rightarrow & M\otimes_R \widehat{R}&\to 0\\
{\big\downarrow}&& {\big\downarrow}&&{\Big\downarrow}{\scriptstyle\alpha_M}&\\
\widehat{R}^e&\rightarrow &\widehat{R}^d&\rightarrow & \widehat M&\to 0
\end{array}.$$
The two left downward arrows are isomorphisms, because tensor products and completions commute with finite direct sums.
A diagram chase shows that $\alpha_M$ is also an isomorphism.
The last item is equivalent to $\widehat \gamma$ being a $\widehat R$-homomorphism, which is clear.
\end{proof}
The next lemma shows that the condition that all finitely presented right $R$-modules are residually finite plays an analogous role to the Artin-Rees Lemma for $I$-adic completions of finitely generated modules over commutative Noetherian rings.
In particular, this lemma is classically used to show that if $0\rightarrow A\rightarrow B\rightarrow C\rightarrow 0$ is a short exact sequence of finitely generated $R$-modules where $R$ is commutative and Noetherian and $I$ is some proper ideal of $R$, then the
$I$-adic completion of $A$ is the same as the closure of $A$ in the $I$-adic topology of $B$.
This guarantees that the induced sequence $0\rightarrow \widehat{A}\rightarrow\widehat B\rightarrow \widehat C\rightarrow 0$ is exact.
Moreover, it is also used to show that the $I$-adic completion of $R$ is flat as an $R$-module.
\begin{lem}\label{lem:fpMrf_implies_lerf}
Let $R$ be a finitely generated ring such that all finitely presented right $R$-modules are residually finite.
Then, for every $d>0$ and every finitely generated submodule $L\leq R^d$, the profinite topology of $L$ coincides with the subspace topology from the profinite topology of $R^d$.
In particular, given an exact sequence $0\to L\to R^d\to M\to 0$ of right $R$-modules with $L$ finitely generated, the induced sequence
$$0\to \widehat L\to \widehat R^d\to \widehat M\to 0$$ of right $\widehat R$-modules, is exact.
\end{lem}
\begin{proof}
The second statement follows directly from the first one and Lemma \ref{indpro}, because the kernel of the map to $\widehat M$ is the closure of $L$ in $R^d$.
To prove the first statement, let $L\leq R^d$ be a finitely generated submodule.
Then $R^d/L$ is finitely presented, so, by assumption, residually finite.
In other words, $L$ is the intersection of all finite index submodules of $R^d$ that contain $L$.
%
We need to show that for every $K\leq L$ of finite index, there is some $U\leq R^d$ of finite index such that $L\cap U\leq K$.
Since $R$ is a finitely generated ring, $K$ is a finitely generated submodule.
Therefore $R^d/K$ is residually finite and, as $L/K$ is finite, there is some $U\geq K$ of finite index in $R^d$ avoiding any set of non-trivial coset representatives of $K$ in $L$.
This means that $U\cap L=K$, as required.
\end{proof}
\begin{proof}[Proof of Theorem \ref{teo:Rhatflat_iff_Rcoherent}]
\ref{it:Rhatflat+FPMRF} $\implies$ \ref{it:Rcoherent+FPMRF}.
Let $L$ be a finitely generated right submodule of a free right module $R^k$.
We want to show that $L$ is also finitely presented.
Consider the following exact sequence of right $R$-modules
$$0\to L\to R^k\to M\to 0.$$
Since $\widehat R$ is left flat and taking into account Lemma \ref{indpro} we obtain the commutative diagram
$$\begin{array}{cccccccccc}
0 &\to & L \otimes_R \widehat{R} & \rightarrow & R^k \otimes_R \widehat{R}&\rightarrow& M \otimes_R \widehat{R}&\to 0\\
& &{\Big\downarrow}{\scriptstyle\alpha_L}&& {\Big\downarrow}{\scriptstyle\alpha_{R^k}}&&{\Big\downarrow}{\scriptstyle\alpha_M}&\\
& & \widehat L&\xrightarrow{\beta} &\widehat R^k&\rightarrow& \widehat M&\to 0.
\end{array}$$
By Lemma \ref{lem:alpha}, $\alpha_L$ is onto. Since $\alpha_{R^k}$ is bijective, the commutativity of the diagram implies that $\alpha_L$ should be injective, and so, bijective.
Assume that $L$ is generated by $d$ elements. Therefore we have the following short exact sequence of right $R$-modules
$$0\to I \to R^d \to L\to 0.$$
Tensoring with $\widehat{R}$ and noting that $\widehat{R}$ is flat gives the short exact sequence
%
$$0\to I\otimes_R \widehat{R} \rightarrow R^d\otimes_R \widehat{R} \rightarrow L\otimes_R \widehat{R} \to 0.$$
Let $\overline I$ be the closure of $I$ in $\widehat R^d$ and denote by $\overline {\alpha_I}: I\otimes_R \widehat{R} \to \overline I$ the composition of $\alpha_I$ and the map $\widehat I\to \overline I$.
Then we also have, upon taking completions and applying Lemmas \ref{indpro} and \ref{lem:alpha} the commutative diagram
$$\begin{array}{cccccccccc}
0 &\to & I\otimes_R \widehat{R} & \rightarrow & R^d\otimes_R \widehat{R} &\rightarrow& L\otimes_R \widehat{R}&\to 0\\
& &{\Big\downarrow}{\scriptstyle\overline{\alpha_I}}&& {\Big\downarrow}{\scriptstyle\alpha_{R^d}}&&{\Big\downarrow}{\scriptstyle\alpha_L}&\\
0&\to & \overline I&\rightarrow &\widehat R^d&\rightarrow & \widehat L&\to 0
\end{array}$$
where $\alpha_{R^d}$ and $\alpha_L$ are isomorphisms and $\gamma, \phi$ are continuous.
The commutativity of the diagram implies that $\overline{\alpha_I}$ is also an isomorphism.
As $R$ is countable, $I$ is also countable.
Enumerate the elements of $I=\{i_k\}_{k\in \mathbb{N}}$ and define, for each $k\in \mathbb{N}$ the sets
$$I_k=\sum_{j=1}^ki_jR \text{ and } T_k= \sum_{j=1}^k i_j\otimes \widehat{R}.$$
Thus, $\overline I=\cup_{k\in \mathbb{N}}\overline{\alpha_I}(T_k)$ is a countable union of compact submodules.
Since $\overline I$ is compact, by the Baire category theorem, there exists $n\in \mathbb{N}$ such that $\overline{\alpha_I}(T_n)$ contains an open subset of $\overline I$.
Hence for some $m\in \mathbb{N}$, $\overline{\alpha_I} (T_m)=\overline I$.
This implies that the closure of $I_m=\overline{ \alpha_I}( I_m\otimes 1)$ in $R^k$ contains $I=\overline{\alpha_I}(I\otimes 1)$.
As $R^k/I_m$ is finitely presented, it is residually finite, by assumption on $R$.
In other words, $I_m$ is closed in $R^k$, which yields $I=I_m$ and thus $L$ is finitely presented.
\ref{it:Rcoherent+FPMRF}$\implies $\ref{it:Rhatflat+FPMRF}.
We wish to show that $\Tor_1^R(M,\widehat{R})=0$ for every right $R$-module $M$.
Since every module is a direct limit of finitely presented modules, and $\Tor_1^R$ commutes with direct limits (in both variables), we may assume without loss of generality that $M$ is finitely presented.
Suppose that $M$ is defined by the short exact sequence
$$0\to L \to R^d \to M \to 0$$
where $L$ is finitely generated.
As $R$ is right coherent, $L$ is in fact finitely presented.
Tensoring this exact sequence with $\widehat{R}$ we obtain the long exact sequence
$$0 \to \Tor_1^R(M,\widehat{R}) \to L\otimes_R \widehat{R} \xrightarrow{i} R^d \otimes_R \widehat{R} \xrightarrow{p} M\otimes_R \widehat{R}\to 0. $$
%
Then by Lemmas \ref{lem:alpha} and \ref{lem:fpMrf_implies_lerf}, we obtain the following commutative diagram with exact top and bottom rows
$$\begin{array}{cccccccccc}
0 & \to \Tor_1^R(M,\widehat R)&\xrightarrow{\beta} & L\otimes_R \widehat{R} &\xrightarrow{i}& R^d \otimes_R \widehat{R} &\xrightarrow{p} & M \otimes_R \widehat{R}&\to 0\\
& & &{\Big\downarrow}{\scriptstyle\alpha_L}&& {\Big\downarrow}{\scriptstyle\alpha_{R^d}}&&{\Big\downarrow}{\scriptstyle\alpha_M}&\\
& 0&\to & \widehat L&\xrightarrow{\eta} &\widehat R^d&\xrightarrow{\pi} & \widehat M&\to 0,
\end{array}$$
where $\alpha_L$,
$\alpha_M$ and $\alpha_{R^d}$ are isomorphisms, because the right $R$-modules $L$, $M$ and $R^d$ are finitely presented.
A diagram chase shows that $\Tor_1^R(M,\widehat R)=0$, as required.
\end{proof}
The condition on $R$ ``each finitely presented module is residually finite" can also be encoded in a condition of flatness of some module.
\begin{pro}\label{pro:profinitely_flat_iff_coherent_and_fprf}
Let $R$ be a finitely generated residually finite ring.
The following are equivalent:
\begin{enumerate}
\item the quotient module $\widehat{R}/R$ is left flat;
\item every finitely presented right $R$-module is residually finite and $\widehat{R}$ is left flat;
\item\label{it:fprf+coherent} every finitely presented right $R$-module is residually finite and $R$ is right coherent.
\end{enumerate}
\end{pro}
\begin{proof}
The last two items are equivalent by Theorem \ref{teo:Rhatflat_iff_Rcoherent}, so we show that the first two are equivalent.
Assume that $\widehat R/R$ is left flat.
Since $R$ is also flat, $\widehat R$ is flat as well.
Let $M$ be a finitely presented right $R$-module.
Tensoring the short exact sequence $$0\to R\to \widehat R \to \widehat{R}/R \to 0$$ with $M$
yields the following commutative diagram with exact upper sequence:
$$\begin{array}{cccccccccc}
\Tor_1^R(M,\widehat R) & \to &\Tor^R_1(M,\widehat R/R)&\to &M\otimes_R R&\xrightarrow{\gamma} & M\otimes_R \widehat{R} &\to & M\otimes_R (\widehat{R}/R) &\to 0\\ &&&
& {\Big\downarrow}{\scriptstyle\sigma}&&{\Big\downarrow}{\scriptstyle\alpha_M}&& &\\
&&&& M&\xrightarrow{\delta_M} &\widehat M& & &
\end{array}$$
where $\gamma$ and $\sigma$ are the canonical maps.
By Lemma \ref{lem:alpha}, $\alpha_M$ is an isomorphism.
It is obvious that the diagram commutes.
Since $\widehat{R}$ and $\widehat R/R$ are flat, $\Tor_1^R(M,\widehat{R})= \Tor^R_1(M,\widehat R/R)=0$.
Thus, $\delta_M$ is an embedding, and so $M$ is residually finite.
Now assume that every finitely presented right $R$-module is residually finite and $\widehat R$ is left flat.
This means that in the previous diagram $\Tor_1^R(M,\widehat{R})=0$ and, as $\delta_M $ is an embedding, $\Tor_1^R(M,\widehat{R}/R)=0$ as well.
\end{proof}
\begin{rem} Let $G$ be a topologically finitely generated profinite group and $R=\mathbb{F}_p[[G]]$.
By the result of Nikolov and Segal \cite{NS07}, the map $\delta_R: R\to \widehat R$ is an isomorphism.
So $\widehat{R}/R=0$ is flat, but obviously $R$ is not always coherent.
Thus, the condition that $R$ is finitely generated is not superficial.
\end{rem}
We will see in the next section that, if $G$ is a virtually free group, then $\widehat{\mathbb{F}_p[[G]]}$ is flat, for every prime $p$.
Although we will not need it in the rest of the paper, it is interesting to note that the same is true for virtually polycylic groups:
\begin{pro}\label{pro:polycyclic_prof_flat}
Let $G$ be a virtually polycyclic group and $p$ a prime number, then $\mathbb{F}_p[G]$ satisfies the conditions of Theorem \ref{teo:Rhatflat_iff_Rcoherent} and of Proposition \ref{pro:profinitely_flat_iff_coherent_and_fprf}.
In particular, $\widehat{\mathbb{F}_p[[G]]}$ is a flat left $\mathbb{F}_p[G]$-module.
\end{pro}
\begin{proof}
This is a corollary of a result proved independently by Jategaonkar \cite{Jategaonkar} and Roseblade \cite{Roseblade} (based on a previous result of Roseblade \cite{Roseblade73}), answering a question of P. Hall; namely, every finitely generated abelian-by-polycylic-by-finite group is residually finite (see also \cite[7.2.1]{LennoxRobinson} for a proof).
This implies that every finitely generated right $R$-module is residually finite, where $R=\mathbb{F}_p[G]$ for prime $p$ and $G$ is virtually polycyclic.
It follows from \cite[Theorem 1]{PHall1954} that $k[G]$ is right Noetherian (and therefore right coherent) whenever $k$ is right Noetherian and $G$ is virtually polycyclic.
Since $R$ is finitely generated, because $G$ is, we conclude that $R$ enjoys the properties in Theorem \ref{teo:Rhatflat_iff_Rcoherent} and in Proposition \ref{pro:profinitely_flat_iff_coherent_and_fprf}(\ref{it:fprf+coherent}).
\end{proof}
\section{Completed group algebras of completions of virtually free groups}\label{sec:limits_projective_modules}
We now apply the results of the previous section, together with results from \cite{CoFR}, to show that the completed group algebra $\mathbb{F}_p[[\widehat{G}]]$ of a finitely generated virtually free group is a direct union of free $\mathbb{F}_p[G]$-modules.
This will be a key ingredient in the next section, to show that if $\widehat{G}$ splits as a profinite free product, then so does $G$.
\begin{definition}
A \textbf{free left ideal ring}, or\textbf{ left fir} for short, is a ring in which all left ideals are free left $R$-modules of unique rank.
\textbf{Right firs} are defined correspondingly and a \textbf{fir} is a left and right fir.
\end{definition}
Note that if $R$ is a fir then
every finitely generated right (or left) ideal is a finitely presented right (or left) $R$-module.
\begin{pro}[\cite{CoFR} Theorem A.6 and Corollary A.7, p. 554]\label{prop:coherent_equivalent}
For a ring $R$, the following are equivalent:
\begin{enumerate}
\item $R$ is right coherent;
\item every finitely generated right ideal of $R$ is a finitely presented right $R$-module.
\end{enumerate}
Furthermore, over a right coherent ring, the intersection of any pair of finitely generated submodules of a free right module is finitely generated.
\end{pro}
\begin{pro}
Let $F$ be a finitely generated free group and $p$ a prime.
Then $\mathbb{F}_p[[\widehat{F}]]$ is a flat left $\mathbb{F}_p[F]$-module.
\end{pro}
\begin{proof}
The ring $\mathbb{F}_p[F]$ is a fir (\cite{Be74}).
Thus, it is also coherent.
We claim that every finitely presented $\mathbb{F}_p[F]$-module is residually finite.
If this holds, then, observing that $\widehat{\mathbb{F}_p[F]}\cong \mathbb{F}_p[[\widehat F]]$, Theorem \ref{teo:Rhatflat_iff_Rcoherent} implies that $\mathbb{F}_p[[\widehat{F}]]$ is a flat left $\mathbb{F}_p[F]$-module.
To show the claim, write $R:=\mathbb{F}_p[F]$, let $M\cong R^n/I$ be a finitely presented right $R$-module, with $n\in \mathbb{N}$ and $I$ finitely generated.
We wish to show that given any $z\in R^n\setminus I$ there is a right submodule $L\supseteq I$ such that $z\notin L$ and $R^n/L$ is finite.
We proceed by induction on $n$.
If $n=1$, then \cite[Theorem 4.3]{RR94} implies that for any $z\in R \setminus I$ there is a finitely generated ideal $L\supseteq I$ such that $z\notin L$ and $R/L$ is finite.
Suppose the claim true for $n-1\geq 1$ and let $z\in R^n \setminus I$.
Denote by $R_1$ the first summand in $R^n$.
If $z\notin R_1+I$, we consider the quotient $R^n/(R_1+I)\cong R^{n-1}/I_1$ where $R^{n-1}=R^n/R_1$ and $I_1=R_1+I/R_1$.
Since $R_1$ and $I$ are finitely generated right $R$-modules, $I_1$ is a finitely generated right submodule of $R^{n-1}$ that does not contain the image $z_1$ of $z$ in $R^{n-1}$.
By inductive hypothesis, there is a finitely generated right submodule $L_1\supseteq I_1$ of $R^{n-1}$ such that $z_1\notin L_1$ and $R^{n-1}/L_1$ is finite.
The preimage of $L_1$ in $R^n$ is the desired right submodule of $R^n$.
%
If $z\in R_1+I$, consider its image $z_1$ under the isomorphism $R_1+I/I\cong R_1/R_1\cap I$, so $z_1\notin R_1\cap I$.
Since $\mathbb{F}_p[F]$ is coherent, Proposition \ref{prop:coherent_equivalent} guarantees that $R_1\cap I$ is a finitely generated right ideal of $R_1$.
Theorem 4.3 of \cite{RR94} then gives a finitely generated right ideal $L_1\supseteq R_1\cap I$ of $R_1$ not containing $z_1$ and such that $R_1/L_1$ is finite.
The preimage of $L_1$ in $R_1+I$ is our desired $L$.
%
\end{proof}
\begin{pro}[See Proposition 1.4.5 in \cite{CoFR}]\label{prop:flat iff all fg submodules free}
Let $R$ be a fir and $M$ a left $R$-module.
Then $M$ is flat if and only if every finitely generated left submodule of $M$ is free.
\end{pro}
Putting together the last two results yields the following:
\begin{cor}\label{locfree}
Let $F$ be a finitely generated free group.
Then every finitely generated submodule of the left $\mathbb{F}_p[F]$-module $\mathbb{F}_p[[\widehat F]]$ is free.
\end{cor}
\begin{cor}\label{locvfree}
Let $G$ be a finitely generated virtually free group.
Then the left $\mathbb{F}_p[G]$-module $\mathbb{F}_p[[\widehat G]]$ is isomorphic to a direct union of free left $\mathbb{F}_p[G]$-modules.
\end{cor}
\begin{proof}
Let $F$ be a free subgroup of $G$ of finite index and denote by $\overline F$ the closure of $F$ in $\widehat G$.
Then $\overline F\cong \widehat F$ because $F$ is of finite index in $G$.
Now, $\mathbb{F}_p[[\widehat G]]\cong \mathbb{F}_p[G]\otimes_{\mathbb{F}_p[F]}\mathbb{F}_p[[\overline F]]$.
By Corollary \ref{locfree}, $\mathbb{F}_p[[\widehat F]]=\mathbb{F}_p[[\bar F]]$ is isomorphic to a direct union of free $\mathbb{F}_p[F]$-modules, therefore, as tensor products commute with direct unions, $\mathbb{F}_p[[\widehat G]]\cong \mathbb{F}_p[G]\otimes_{\mathbb{F}_p[F]}\mathbb{F}_p[[\overline F]]$ is isomorphic to a direct union of free $\mathbb{F}_p[G]$-modules.
\end{proof}
\begin{cor} \label{cor:completedgroupalg=directunionprojectives}
Let $G$ be a finitely generated virtually free group.
Let $n$ be a product of a finite set of different primes and $k=\mathbb{Z}/n\mathbb{Z}$.
Then the $k[G]$-module $k[[\widehat G]]$ is isomorphic to a direct union of projective $k[G]$-modules.
\end{cor}
\section{Free factors of virtually free groups}\label{sec:Dicksproof}
This section is devoted to the proof of the equivalence of items a) and c) in Theorem \ref{teo:vfree}.
Recall that this equivalence is the following:
\begin{teo}\label{teo:profinite_freefactor_iff_abstract_freefactor}
Let $G$ be a finitely generated group that is virtually free and let $H\leq G$ be finitely generated. Denote by $\overline{H}$ the closure of $H$ in the profinite completion $\widehat{G}$ of $G$. Then $\widehat{G}$ splits as a free profinite product $\widehat{G}=\overline{H}\coprod K$ if and only if $G=H\ast L$ splits as an abstract free product.
\end{teo}
The ``if" part follows from the facts that ``profinite completion commutes with free product"; i.e., $\widehat{H}\coprod \widehat{L}=\widehat{H\ast L}$ and that the profinite topology of $H\ast L$ induces the full profinite topologies on $H$ and $L$ (see \cite[Corollary 3.1.6]{RibesZalesskii}).
The rest of this section is devoted to the ``only if" implication.
It relies on a decomposition theorem due to Dicks and Dunwoody (\cite{Di81,Dicks_book}).
To set notation, we recall the relevant Bass--Serre theory, mostly following \cite{Dicks_book}.
\subsection{Prerequisites from Bass--Serre Theory}
A \textbf{graph} $Y=V(Y) \sqcup E(Y)$ is the disjoint union of a set $V=V(Y)$ of \textbf{vertices} and a set $E=E(Y)$ of \textbf{edges}, given with two maps $i,t:E\rightarrow V$ that associate to each edge its \textbf{initial} $i(e)$, respectively, \textbf{terminal} $t(e)$ vertex.
Loops ($i(e)=t(e)$) are allowed.
In other words, we consider directed multigraphs.
$Y$ is \textbf{connected} if it is connected as an undirected graph.
A \textbf{tree} is a connected graph which has no undirected cycles.
By Zorn's Lemma, every connected graph has a maximal subtree.
For our purposes, $Y$ will always be finite.
A \textbf{graph of groups} $(\mathcal{G},Y)$ consists of a graph $Y$, vertex groups $\{\mathcal{G}(v) : v\in V(Y)\}$ and edge groups $\{\mathcal{G}(e) : e\in E(Y)\}$ and, for each $e\in E(Y)$, two group morphisms $i_e, t_e: \mathcal{G}(e)\rightarrow \mathcal{G}(i(e)), \mathcal{G}(t(e))$.
If all the morphisms $i_e$,$t_e$, $e\in E(Y)$ are injective, the graph of groups is called \textbf{faithful.}
Given a connected graph of groups $(\mathcal{G},Y)$ and a maximal subtree $T$ of $Y$, the \textbf{fundamental group} $G=\pi(\mathcal{G},Y,T)$ of $(\mathcal{G},Y)$ with respect to $T$ is the group with presentation:
\begin{itemize}
\item generators $\{\mathcal{G}(v), q_e: v\in V(Y), e\in E(Y)\}$;
\item relations of each $\mathcal{G}(v), v\in V(Y)$;
\item relations $(i_e(g))^{q_e}=t_e(g)$ for each $g\in \mathcal{G}(e), e\in E(Y)$;
\item relations $q_e=1$ if $e\in E(T)$
\end{itemize}
One of the first results of Bass--Serre theory (see \cite[II.2.4]{Dicks_book}, \cite[I.5.20]{Serre_trees}) is that any other maximal subtree yields an isomorphic fundamental group.
Another result (\cite[II.2.7]{Dicks_book}) states that if $(\mathcal{G},Y)$ is connected and faithful, then the canonical maps $\mathcal{G}(v)\rightarrow G, v\in V(Y)$ are injective and we may identify each $\mathcal{G}(v)$ with its image.
\begin{rem}\label{rem:edge_groups_are_not_vertex_groups}
Let $(\mathcal{G},Y)$ be a connected faithful graph of groups with maximal subtree $T$.
Suppose there is some edge $e$ in $T$ with $i(e)=v, t(e)=w$ and such that one of $i_e$ or $t_e$ is an isomorphism.
Without loss of generality, assume that $i_e$ is an isomorphism, so that we may identify $\mathcal{G}_e=\mathcal{G}_v$.
Then, in the fundamental group $\pi(\mathcal{G},Y,T)$ we have that $(\mathcal{G}_v)^{q_e}=\mathcal{G}_v=t_e(\mathcal{G}_v)\leq \mathcal{G}_w$.
Consider the graph $Y'$ and maximal tree $T'$ obtained from $Y$ by identifying $v$ and $w$ along $e$ (leaving the other possible edges between $v$ and $w$ as loops).
Then $\pi(\mathcal{G},Y,T)=\pi(\mathcal{G},Y',T')$ because $\mathcal{G}_v\leq \mathcal{G}_w$ in $\pi(\mathcal{G},Y,T)$.
Doing this procedure for all edges $e$ of $T$ as above, we may assume that if $G=\pi(\mathcal{G},Y,T)$, then for every edge $e$ of $T$, neither $i_e$ nor $t_e$ are isomorphisms.
Since choosing a different maximal subtree does not change the isomorphism type of the group, we may further assume that $G$ is the fundamental group of a connected faithful graph of groups such that for every edge $e$ joining different vertices neither $i_e$ nor $t_e$ are isomorphisms.
%
%
\end{rem}
Let $G$ act on a tree $X$ and let $S\subset X$ be a \textbf{connected transversal} for the $G$-action.
This is a subset of $X$ (in general, not a subgrap
) that is in bijection with $G\backslash X$, any two of whose vertices are connected by a path in $S$ and such that $i(e)\in S$ for every $e\in S$.
For each $e\in E(S)$, there is a unique vertex of $S$ in the same orbit as $t(e)$, so we choose $q_e\in G$ such that $q_e\cdot t(e) \in S$, taking $q_e=1$ if $t(e)\in S$.
The collection $(q_e : e\in E(S))$ is a \textbf{connecting family} for $S$.
For each $s\in S$, denote by $G_s$ its stabiliser in $G$.
Note that $G_{q_e\cdot t(e)}=q_eG_{t(e)}q_e^{-1}$ and that for every $e\in E(S)$, $G_e=G_{i(e)}\cap G_{t(e)}=G_{i(e)}\cap q_e^{-1}G_uq_e$ where $u=q_e\cdot t(e)$, so that $G_e^{q_e}\leq G_{q_e\cdot t(e)}$.
These are exactly the conjugation relations that appear in the definition of the fundamental group of a graph of groups.
Let $(\mathcal{G},Y)$ be a connected graph of groups with maximal subtree $T$.
Write $\pi=\pi(\mathcal{G},Y,T)$.
The \textbf{standard tree} of $(\mathcal{G},Y,T)$ is the graph $\tilde{X}=\tilde{X}(\mathcal{G},Y,T)$ with, respectively, vertex and edge sets $$V(\tilde{X})=\bigsqcup_{v\in V(Y)}\pi/\mathcal{G}(v), \qquad E(\tilde{X})=\bigsqcup_{e\in E(Y)}\pi/G_{e}$$
and incidence maps given by $$i(g\mathcal{G}(e))=g\mathcal{G}(i(e)), \quad t(g\mathcal{G}(e))=gq_e\mathcal{G}(t(e)).$$
That $\tilde{X}$ is indeed a tree is a key result of Bass--Serre theory (\cite[I.5.3]{Dicks_book}, \cite[I.5.12]{Serre_trees}).
\begin{teo}[Structure Theorem of Bass--Serre theory]
Let $G$ act on a connected graph $X$ with connected transversal $ S\subset X$, maximal subtree $T\subset S$ and connecting family $(q_e\in G: e\in E(S))$.
Then there is a faithful connected graph of groups $(\mathcal{G}, G\backslash X)$ defined by $\mathcal{G}(s)=G_s$ for $s\in S$ and $i_e, t_e: G_e\rightarrow G_{i(e)}, G_{(q_e\cdot t(e))}$ given, respectively, by $g \mapsto g, g^{q_e}$, for each $e\in E(S)$.
There is a surjective homomorphism $\varphi:\pi=\pi(\mathcal{G},G\backslash X, T) \rightarrow G$, uniquely determined by the inclusion maps $\mathcal{G}(v)\rightarrow G$ and $q_e\rightarrow q_e$, $v\in V(Y), e\in E(S)$.
There is an equivariant universal covering map from the standard tree $\tilde{X}(\mathcal{G},G\backslash X, T)$ to $X$, given by $g\mathcal{G}(s)\mapsto \varphi(g)\cdot s, g\in\pi, s\in S$.
The map $\tilde{X}\rightarrow X$ is bijective if and only if $\varphi$ is an isomorphism.
\end{teo}
Let $G=\pi(\mathcal{G},Y,T)$ be the fundamental group of a connected faithful graph of groups, with respect to a maximal subtree $T$.
Since we can identify each $\mathcal{G}(v), v\in V(Y)$ with its image in $G$, we write $G_y:=\mathcal{G}(y)$ for each $y\in Y$.
For every $e\in E(Y)$, write $i(G_e)\leq G_{i(e)}$ and $t(G_e)\leq G_{t(e)}$ for the respective images of $G_e$ under $i_e, t_e$.
Conjugating by $q_e$ induces an isomorphism $i(G_e)\rightarrow t(G_e)$, which is the identity if $e\in T$, in which case we identify $G_e$ with $i(G_e)=t(G_e)$.
\begin{pro}[Proposition II.3.1 of \cite{Dicks_book}]\label{pro:intersections_of_vertex_groups}
With the above notation, $\mathcal{G}_{i(e)}\cap \mathcal{G}_{t(e)}^{q_e}=i_e(\mathcal{G}(e))$ for any $e\in E(Y)$.
Given any $v, w\in V(Y)$, let $P$ be the set of edges of $T$ in the geodesic joining $v, w$. Then $G_v\cap G_w=\bigcap_{e\in P}G_e$.
\end{pro}
\begin{rem}\label{rem:suff_cond_free_factor}
Suppose that $G=\pi(\mathcal{G},Y,T)$ is the fundamental group of some connected faithful graph of groups and that $H\leq G$ is one of the vertex groups.
$H$ is a free factor of $G$ if all edge groups incident with $H$ have trivial image in $H$.
By Proposition \ref{pro:intersections_of_vertex_groups},
this holds if $H\cap G_v^g$ is trivial for every vertex group distinct from $H$ and every $g\in G$ and if $H\cap H^g$ is trivial for every $g\in G\setminus H$.
\end{rem}
The last result that we need is the crucial link between derivations and decompositions into fundamental groups of graphs of groups.
A derivation $d: G\rightarrow P$ is \textbf{inner} if there is some $p\in P$ such that $d(g)=(1-g)p$ for all $g\in G$.
\begin{teo}[Theorem III.4.6 of \cite{Dicks_book}]\label{teo:dicks_derivation_to_decomposition}
Let $G$ be a group and $R$ a ring.
Suppose that $P$ is a projective $R[G]$-module, and $d:G \rightarrow P$ is a derivation such that $G$ is generated by $\ker d$ together with finitely many elements.
Then $G$ decomposes as the fundamental group of a faithful connected finite graph of groups with finite edge groups, $\ker d$ as one of the vertex groups, and such that the derivation $d$ is inner when restricted to each vertex group.
\end{teo}
Note that if $d$ is inner when restricted to vertex groups, then $d$ is also inner on conjugates of vertex groups: for every $f\in G$ and $g\in G_v$ we have
$d(fgf^{-1})=(1-fgf^{-1})d(f)+fd(g)=(1-fgf^{-1})(d(f)+fp).$
\subsection{Proof of Theorem \ref{teo:profinite_freefactor_iff_abstract_freefactor}, continued}
Suppose that $G$ is finitely generated, virtually free and $H\leq G$ satisfies that $\widehat{G}=\overline{H}\coprod K$ for some closed subgroup $K$ of $\widehat{G}$.
Let $U$ be a free, normal subgroup of finite index in $G$ and let $k=\mathbb{Z}/n\mathbb{Z}$ where $n$ is the product of all prime factors of $|G:U|$.
Lemma \ref{lem:freeprofprod_implies_idealsdecompose} yields a continuous homomorphism $$\phi:I_{k[[\widehat G]]}\rightarrow I_{k[[ K]]}^{\widehat{G}}$$ of $k[[\widehat{G}]]$-modules, which is simply projection to the second component.
Note that these are also $k[G]$-modules.
Since $G$ is finitely generated, the image of $\{1-g:g\in G\}\subseteq I_{k[[\widehat G]]}$ under $\phi$ is a finitely generated $k[G]$-module.
By Corollary \ref{cor:completedgroupalg=directunionprojectives}, this finitely generated $k[G]$-module lies in a direct union of projective $k[G]$-modules (which we can take to be finitely generated, since every module is the direct union of its finitely generated submodules).
Let $P$ be a projective finitely generated $k[G]$-module containing $\phi(I_{k[ G]})$ and denote by $\varphi:I_{k[ G]}\rightarrow P$ the restriction of $\phi$ to $I_{k[ G]}$, with image in $P$.
Then $\varphi$ induces a derivation $$d:G\rightarrow P, \ g\mapsto \varphi(1-g).$$
Notice that $\varphi(1-g)=0$ exactly when $g\in \overline{H}\cap G$.
Since $G$ is virtually free, all finitely generated subgroups are closed in the profinite topology (that is, $G$ is LERF), so $\overline{H}\cap G=H$.
We have thus obtained a derivation $d:G\rightarrow P$ to a projective $k[G]$-module whose kernel is exactly $H$.
Theorem \ref{teo:dicks_derivation_to_decomposition} then implies that $G$ decomposes as the fundamental group of a faithful connected finite graph of groups $Y=(V,E)$, with finite edge groups, with $H=G_u$ for some $u\in V$ and such that the derivation $d$ is when restricted to any conjugate of any vertex group.
Note that by Remark \ref{rem:edge_groups_are_not_vertex_groups}, we can assume that for every $e\in E$ with distinct endpoints $v,w$, neither of the injections $G_e\rightarrow G_v,G_w$ is surjective.
Following Remark \ref{rem:suff_cond_free_factor}, we will show that $H\cap G_v^g=1$ for every $g\in G$ and $H\cap H^g=1$ for every $g\in G\setminus H$.
\
\
\noindent\textbf{Case 1: $v\neq u$ and $G_v$ is infinite.}
First we reduce to the case where $H$ is finite.
Recall that $U$ was taken to be a free, normal subgroup of finite index in $G$.
Then $H':=H/(H\cap U)$ is finite and, since $H\cap U$ is torsion-free, it intersects trivially all edge groups, so that $G':=G/\langle H\cap U\rangle^G$ has the same presentation as $G$, with the only addition that $H$ is replaced by the finite group $H'$.
In particular, each vertex group $G_v$ distinct from $H$ maps isomorphically to a subgroup of $G'$, which, abusing notation, we shall also call $G_v$.
We then have $\widehat{G'}=\overline{H'}\coprod K=H'\coprod K$ and so $I_{k[[\widehat{G'}]]}=I_{k[[H']]}^{\widehat{G'}}\oplus I_{k[[K]]}^{\widehat{G'}}$.
This gives the following commutative diagram:
$$\begin{array}[c]{ccccccc}
G & \hookrightarrow & \widehat{G} &\hookrightarrow & I_{k[[\widehat{G}]]} & \stackrel{\phi}{\rightarrow} & I_{k[[K]]}^{\widehat{G}}\\
\downarrow && \downarrow && \downarrow && \downarrow \\
G' & \hookrightarrow & \widehat{G'} & \hookrightarrow & I_{k[[\widehat{G'}]]} & \stackrel{\phi'}{\rightarrow} & I_{k[[K]]}^{\widehat{G}}
\end{array}$$
where the downward arrows are the quotient maps by the appropriate image of $\langle H\cap U\rangle ^G$.
Note that $d$ is the composition of all maps on the top row of the diagram.
Denote by $d'$ the composition of the maps on the bottom row.
This is in particular a derivation on $G'$, whose kernel is $H'$, and that is inner on every conjugate of $G_v\leq G'$ for every $v\in V$.
Fix now some $v\neq u$ such that $G_v$ is infinite and denote by $L$ any conjugate of $G_v$.
Because $d
$ is inner on $L$, there is some $m\in I_{k[[\widehat {G^\prime}]]} $ such that $d'(g)=(1-g)m$ for every $g\in L$.
Recalling that $1-g=d'(g)+\left(I_{k[H]}^{\widehat{G^\prime}}-\text{part of }1-g\right)$, we have that $1-g \equiv(1-g)m$ modulo $I_{k[H]}^{\widehat{G^\prime}}$ for every $g\in L$.
Equivalently, $(1-g)(1-m)\equiv 0$ modulo $I_{k[H^\prime]}^{\widehat{G^\prime}}$ for every $g\in L$, so the image of $1-m$ in $k[[\widehat{G'}/H']]$ is a fixed point of $L$.
Any fixed point of $L$ in $k[[\widehat{G'}/H']]$ is also a fixed point of the closure $\overline{L}$ in $\widehat{G'}$, because $L$ is dense in $\overline{L}$ (as $G'$ is residually finite) and the action of $\widehat{G'}$ on $k[[\widehat{G'}/H']]$ is continuous.
Now, $\overline{H'}=H'$ is finite, so for every $N\unlhd_o\widehat{G'}$ and $g\in \widehat{G'}$ we have
$$|\overline{L}:\overline{L}\cap N|=|\overline{L}:\overline{L}\cap (H')^gN|\cdot |\overline{L}\cap (H')^gN: N|\leq |\overline{L}:\overline{L}\cap (H')^gN| \cdot |H'|.$$
Therefore $p^\infty$ divides $\lcm_{N\unlhd_o\widehat{G'}} |\overline{L}:\overline{L}\cap (H')^gN|$ if and only if it divides $\lcm_{N\unlhd_o\widehat{G'}}|\overline{L}:\overline{L}\cap N|=|\overline{L}|$.
Because $G'$ is LERF, the profinite completion of $L$ coincides with its closure $\overline{L}$ in $\widehat{G'}$.
As $L$ is virtually torsion-free, $\overline{L}$ contains a copy of $\widehat{\mathbb{Z}}$; in particular, $p^\infty$ divides $|\overline{L}|$.
Now Lemma \ref{lem:no_fixed_points_modH} implies that $\overline{L}$ has no non-trivial fixed points when it acts on $k[[\widehat{G'}/H']]$.
In particular, $(1-m)$ is congruent to 0 in $k[[\widehat{G'}/H']]$, which is equivalent to $1-m \in I_{k[H^\prime]}^{\widehat{G^\prime}}$.
Recall that the kernel of $d'$ is $H'$, so $d'(h)=(1-h)m=0$, that is, $hm=m$, for all $h\in L\cap H'$.
Let $G'/N$ be a finite quotient of $G'$ in which $H'$ injects. The image $\widetilde{m}\in k[G'/N]$ of $m$ is a finite sum $\widetilde{m}=\sum_{\gamma\in G'/N}r_{\gamma}\gamma$.
Since $m\equiv 1$ in $k[[\widehat{G'}/H']]$, the coefficients of $\gamma\in H'N/N\cong H'$ in $\widetilde{m}$ must sum to 1: $\sum_{\gamma\in H'N/N}r_{\gamma}=1$.
Moreover, because $\widetilde{m}$ is invariant under $H'\cap L$, we must have that $r_{\gamma}=r_{h\gamma}$ for all $h\in H'\cap L$ and $\gamma\in G'/N$.
In particular, $$1=\sum_{\gamma\in H'N/N}r_{\gamma}=\sum_{\delta\in H'N/(H'\cap L)N}\left(r_{\delta}|H'\cap L|\right),$$
so $|H'\cap L|$ is invertible in $k$.
By the choice of $k$, the invertible elements are those that are coprime with all primes that divide $|G:U|$.
But $|H'\cap L|$ divides $|H'|=|HU:U|$, which divides $|G:U|$.
So $H'\cap L=1$.
This means that the image of $H\cap L\leq G$ in $G'$ is trivial, which implies that $H\cap L\leq H\cap U$.
But $U$ is torsion-free, so in fact $H\cap L=1$.
\
\noindent\textbf{Case 2: $v\ne u$ and $G_v$ is finite.}
Like in the previous case, denote by $L$ any conjugate of $G_v$ in $G$.
By \cite[Theorem 9.5.1]{RibesZalesskii}, since $L$ is finite, it is conjugate in $\widehat{G}$ to a subgroup of $\overline{H}$ or $K$.
If the latter, we are done, as any conjugate of $K$ intersects $\overline{H}$ trivially , by \cite[Corollary 3.1.2]{MelnikovZalesskii}.
The same result guarantees that a conjugate of $\overline{H}$ distinct from $\overline{H}$ has trivial intersection with $\overline{H}$ (see also \cite[Theorem 9.1.12]{RibesZalesskii}).
So we are left with the case $L\leq \overline{H}$, which means that $L\leq H$
because $H$ is closed in the profinite topology of $G$.
Suppose that $L=G_v^g$, so that $L$ is the stabiliser of $gv$ in the standard tree.
Let $gw$ be the next vertex in the path between $gv$ and $u$ in the standard tree.
Since $G_v^g\leq H=G_u$, we also have that $G_v^g\leq G_w^g$ and therefore $G_v\leq G_w$, where inclusion is induced by the edge homomorphisms of the edge $e\in E$ between $v,w\in V$.
This contradicts the assumption on $G$ that we made following Remark \ref{rem:edge_groups_are_not_vertex_groups}.
\
\
\noindent\textbf{Case 3: $v=u$.}
Here we must show that $H^g\cap H=1$ for every $g\notin H$.
No such $g$ is contained in $\overline{H}$, as $H=G\cap \overline{H}$.
By \cite[Theorem 9.1.12]{RibesZalesskii}, this implies that $\overline{H}^g\cap \overline{H}=1$, which finishes the argument.\hfill \qedsymbol
|
\section{Introduction}
Einstein gravity is a successful gravitational theory from many aspects. However, one can still consider modifying it at the energy scale $E>100$ MeV, where it has not been fully tested. In fact, in the spirit of effective field theory, it is permissible to add all the possible diffeomorphism invariant curvature terms to the Einstein--Hilbert (EH) action, thus modifying the original Einstein gravity. Indeed, such higher curvature terms can be naturally generated by quantum fluctuations at short distances \cite{Donoghue:1994dn}, or from the low energy limit of the string theory \cite{Boulware:1985wk}. Over the past decades, higher curvature theories of gravity have attracted much attention. In AdS/CFT, these higher curvature terms correspond to the large N corrections to CFTs. And in cosmology, it has a lot of implications on the scenario of early inflation and the present acceleration of the universe \cite{staro,Clifton:2011jh}.
Generally, when the higher curvature terms are added to the EH action, the original Schwarzschild black hole (SBH) can not be the exact solution of the theory anymore. And accordingly, the black hole thermodynamics (i.e., entropy, temperature and radius of the horizon) also receives corrections. For example, in Guass--Bonnet gravity (or more generally the Lanczos-Lovelock gravity), some exact, spherically symmetric black hole solutions have been found, and the black hole thermodynamics has been analyzed \cite{Boulware:1985wk,Cai:2001dz}. Despite the few cases, for most of the higher curvature theories of gravity, it is very difficult to obtain exact black hole solutions. However, one can always solve them perturbatively and iteratively. To be concrete, let us consider the effective action of the form
\begin{align}
I_{eff}= \int d^4 x \sqrt{-g}\left (\frac{R}{16\pi}+\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma}) \right),\label{action}
\end{align}
where the first part is the classical EH term, and the second part is the diffeomorphism invariant higher curvature term, which is constructed by the metric $g_{\mu \nu}$, the operator $\nabla_{\mu}$ and Riemann tensor $R_{\mu\nu\rho\sigma}$. The parameter $\alpha$ controls the overall effects of the higher curvature terms, and its value should be small to avoid a direct conflict with existing experiments.
Conventionally, one should vary the effective action with respect to the metric $g_{\mu \nu}$ in order to get the gravitational field equations. Then the static and spherical symmetric black hole solution is to be solved under the ansatz
\begin{align}
ds^2=-A(r)dt^2 +\frac{1}{B(r)} dr^2 +r^2 d \Omega_2, \label{ansatz}
\end{align}
where $A(r)$ is usually written as $A(r)=N(r) B(r)$. As mentioned, the metric can be solved perturbatively around SBH, with the first-order corrections to the original SBH metric denoted by $\alpha\,a(r)$ and $\alpha\,b(r)$. Then the metric at first order is solved and given by
\begin{align}
A(r)& \doteq 1-\frac{2M}{r}+\alpha\,a(r), \label{mar}\\
B(r)& \doteq 1-\frac{2M}{r}+\alpha\,b(r). \label{mbr}
\end{align}
Throughout the paper, we use the symbol ``$\doteq$" to represent that the expression is valid merely at the first order in $\alpha$, omitting the higher order components $\mathcal{O}(\alpha^2)$. Then, with the metric at hand, the effects of higher curvature terms on the black hole thermodynamics can be derived. Concretely speaking, the horizon radius can be obtained from $A(r_h)=B(r_h)=0$; the temperature from the conical singularity of the metric; and the entropy from the Wald formula or other available techniques.
Though the above processes can work in principle, practically it is rather tedious to realize. For most higher curvature terms, the variation problem in deriving the gravitational field equations has already been complicated and even formidable, not to mention solving it. See, for instance, Ref.\cite{Biswas:2013cha,Xiao:2021ewv}. So there is often a long way to go before really getting the corrections to the black hole thermodynamics.
In this paper, we present an interesting fact that, at least at the first order level, the corrections to the SBH thermodynamics can be easily obtained, without solving the modified black hole metric. The paper is organized as follows. First, we describe the procedure of our approach. Then we examine some specific examples, which produces the known results in the literature, as well as those previously unknown. Next, we provide some exact results inspired and obtained by generalizing the first-order analysis. In particular, a novel relation between the Euclidean integrals of the EH term and the higher curvature term is found and proved. We conclude with a summary and discussion, while the explanations of some assertions of the paper are left to the Appendix.
\section{The procedure of our approach}\label{sec2}
We mainly use the Euclidean approach to derive the black hole thermodynamics. In this approach, one should do the Euclidean integral of the action and take it as the partition function of the thermodynamics. The spirit of the section is that the \emph{first-order} partition function can be calculated by using only the \emph{zeroth-order} metric, i.e., the original SBH metric. In the following, we shall describe the procedure, and to avoid diverging too much from the main thread, the proofs of some assertions are left to the Appendix.
First, we evaluate the Euclidean integral of the EH part of the effective action \eqref{action}, that is,
\begin{align}
I_1\equiv \beta \int d^3 x \sqrt{-g}\frac{R}{16\pi}. \label{I1def}
\end{align} The temporal integral has given the time periodicity of black hole space-time, which is the inverse of the temperature $\beta\equiv 1/T$. The spatial integral runs from the horizon radius to infinity. For now, we only need to know the existence of the horizon, but don't need to know its precise location. As explained in the Appendix, if $A(r)=B(r)$ in the metric, the integral can be done exactly, which gives
$I_1=\frac{\beta M}{2}+\pi r_h^2-\frac{\beta r_h}{2}$, where $r_h$ is the horizon radius. If $A(r)\neq B(r)$, the formula becomes approximately valid only at the first order in $\alpha$. Thus, whatever the case is, we always have
\begin{align}
I_1\doteq\frac{\beta M}{2}+\pi r_h^2-\frac{\beta r_h}{2}.\label{eu1}
\end{align}
Second, the Euclidean action implicitly includes the contribution from the Gibbs-Hawking-York (GHY) term or other necessary boundary terms. For the SBH and its modifications in the higher curvature gravity, we can prove the boundary term always takes the exact form
\begin{align}
I_2= -\frac{\beta M}{2}.\label{eu2}
\end{align}
Adding $I_1$ and $I_2$ together, there is $I_1+I_2\doteq \pi r_h^2-\frac{\beta r_h}{2}$. In order to reduce the number of variables for the convenience of computation, it can be further re-expressed in the form
\begin{align}
I_1+I_2\doteq -\frac{\beta ^2}{16 \pi }. \label{act12}
\end{align}
Note that the parameter $\alpha$ doesn't explicitly present even though it is a first-order expression. In fact, the expression $ -\frac{\beta ^2}{16 \pi }$ is the same partition function for the original SBH in Einstein gravity, and was occasionally borrowed without illustration when Einstein gravity is modified \cite{El-Menoufi:2017kew,Xiao:2021zly}. However, whether and to what extent it could be safely applied were far from clear. In the Appendix, we prove that it is still applicable to the evaluation of $I_1+I_2$ at the first order level, just meeting our current needs. A further analysis can show that it fails at second and higher orders.
Third, we evaluate the contribution from the higher curvature terms, that is,
\begin{align}
I_3\equiv \beta \int d^3 x \sqrt{-g}\, \alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma}). \label{I3def}
\end{align}
Because there has already been a coefficient $\alpha$, we only need to use the original SBH metric to calculate the integrand and do the integration. It yields
\begin{align}
I_3\doteq\beta \,\alpha F(M,r_h)\doteq\beta \,\alpha F(\beta) \label{I3},
\end{align}
with a little abuse of the notation $F(\cdots)$. In the last step, we used the relation between $M$, $r_h$ and $\beta$ of SBH to reduce the number of variables.
In total, the partition function is
\begin{align}
I\doteq -\frac{\beta ^2}{16 \pi }+\beta \,\alpha F(\beta) \label{Itotal}.
\end{align}
The energy and entropy of the black hole can be derived by the standard formulas
\begin{align}
&M(\beta)=-\frac{\partial I}{\partial \beta}, \label{mass}\\
&S_E(\beta)=\beta M+I. \label{euentr}
\end{align}
The shift of the horizon radius can also be determined. The Wald entropy formula \cite{Wald:1993nt,Iyer:1994ys} provides an independent way to compute the entropy, with the form
\begin{align}
S_{W} = -2\pi \oint
(\frac{\partial \mathcal{L}}{\partial R_{\mu\nu\rho\sigma}})^{(0)} \epsilon_{\mu\nu} \epsilon_{\rho\sigma} d\Sigma. \label{wald}
\end{align}
Taking the derivative of $\mathcal{L}=\frac{R}{16\pi}+\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})$ with respect to Riemann curvature is a standard calculation \cite{Maeda:2010bu}. The EH term always leads to the area entropy $\pi r_h^2$. And as above, when evaluating the output of the higher curvature term, we can straightforwardly employ the curvatures of the original SBH. It gives
\begin{align}
S_W(r_h) \doteq \pi r_h^2+\alpha F(r_h). \label{waldentr}
\end{align}
The non-trivial agreement between the Euclidean and Wald entropy has been studied, for example, in \cite{Dutta:2006vs}. Thus, by requiring $S_E(\beta)=S_W(r_h)$, the relation between $r_h$ and $\beta$ can be solved. Interestingly, the Wald entropy \eqref{waldentr} looks more like a geometric quantity compared to the thermodynamic Euclidean entropy \eqref{euentr}.
If needed, the entropy, temperature and horizon radius can also be re-expressed as the functions of $M$, the quantity that can be soundly defined and measured as the conserved energy of the space-time. Obviously, from eq.\eqref{mass}, $\beta$ can be inversely solved as the function of $M$.
After the above procedures, the information about the first order corrections to black hole thermodynamics has been revealed.
\section{some examples} \label{sec3}
In this section, we generate results for some concrete examples and compare them with those existing in the literature.
First, consider the example
\begin{align}
\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})= \frac{\alpha}{m_\Lambda^2} R^{\mu\nu}_{\ \ \alpha\beta} R^{\alpha\beta}_{\ \ \rho\sigma}R^{\rho\sigma}_{\ \ \mu\nu},
\end{align}
where $m_\Lambda$ denotes the energy scale where the effects of the added term become salient. But we shall set $m_\Lambda=1$ for brevity from now on. According to eqs.\eqref{I3} and \eqref{Itotal}, the partition function can be evaluated as
\begin{align}
I\doteq-\frac{\beta ^2}{16 \pi }+ \alpha \frac{512 \pi ^4 }{\beta ^2}.
\end{align}
Then, following the procedure of last section, we get the thermodynamic quantities
\begin{align}
T &\doteq \frac{1}{8 \pi M}+ \alpha \frac{1}{4 M^5}, \label{curT}\\
S &\doteq 4 \pi M^2+\alpha \frac{8 \pi ^2 }{M^2},\label{curS}\\
r_h &\doteq 2 M- \alpha \frac{10 \pi }{M^3} \label{curR},
\end{align}
where the second terms in the above expressions represent the first order corrections to the black hole thermodynamics due to the higher curvature term of the action. The results \eqref{curT}, \eqref{curS} and \eqref{curR} are exactly the same as eqs.(16), (18) and (20) of \cite{Calmet:2021lny} which exploited the conventional approach.
For higher and higher curvature terms, the conventional approach becomes much harder to work through. The variation problem in deriving the gravitational field equations has already been a hard work, before really solving the metric and the corresponding thermodynamics. In contrast, our approach is still simple and easily executed. Without too much effort, for the general situation
\begin{align}
\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})= \alpha \overbrace{ R^{\mu\nu}_{\ \ \alpha\beta} R^{\alpha\beta}_{\ \ \centerdot\,\centerdot} \cdots R^{\centerdot\,\centerdot}_{\ \ \mu\nu}}^n,
\end{align}
we find the correction terms as
\begin{align}
T^{(c)} &\doteq \alpha\frac{4^{2-n} (n-2) \left( 2^n+2 (-1)^n \right) M^{1-2 n}}{3 (n-1)} ,\\
S^{(c)} &\doteq \alpha \frac{ 2^{9-2 n} \left( 2^n + 2 (-1)^n \right) \pi ^2 M^{4-2 n}}{3 (n-1)},\\
r_h^{(c)} &\doteq \alpha \frac{ 2^{5-2 n} [8 (-1)^n- 2^n ( 3 n^2- 3 n-4 ) ] \pi M^{3-2 n}}{3 (n-1)}.
\end{align}
By the way, if using $r_h$ as the variable, instead of $M$, the expression of entropy becomes
\begin{align}
S\doteq \pi r_h^2+\alpha 2^{n+3} n \pi ^2 r_h^{4-2 n}.
\end{align}
Obviously, the case $n=3$ recovers to what we just have given. And the case $n=2$ corresponds to the Gauss-Bonnet gravity (the scalar curvature $R$, Ricci tensor $R_{\mu\nu}$ of SBH vanish and do not contribute at the first order) in $D=4$ dimensions, where the temperature and horizon radius remain unchanged and the entropy correction is a constant.
Next we turn to another example
\begin{align}
\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})=\alpha \sum_{n=0}^{\infty} f_n R^{\mu\nu\rho\sigma}\Box^n R^{\mu\nu\rho\sigma}, \label{boxes}
\end{align}
where the right hand side comes from a Taylor expansion of some operator $f(\Box)=\sum f_n \Box^n$ \cite{Biswas:2013cha}. The gravitational field equations are extremely involved and complicated, and its first-order corrections to the black hole thermodynamics were studied recently in \cite{Xiao:2021ewv} using the conventional approach. Once again, now we can avoid the arduous computations therein. Following eqs.\eqref{I3} and \eqref{Itotal}, the partition function can be evaluated as
\begin{align}
I \doteq -\frac{\beta ^2}{16 \pi }+64 \pi ^2 \alpha f_0 - \alpha f_1\frac{1536 \pi ^4 }{\beta ^2}+\alpha f_2 \frac{49152 \pi ^6 }{\beta ^4}+\cdots.
\end{align}
Thus the thermodynamic quantities are
\begin{align}
T &\doteq \frac{1}{8 \pi M } - \frac{3 \alpha f_{1}}{4 M^5} + \frac{3 \alpha f_{2}}{4 M^7} +\cdots,\label{temmis}\\
S &\doteq 4 \pi M^2 +64\pi^2 \alpha f_{0} - \frac {24 \pi^2 \alpha f_{1}}{M^2} + \frac {12 \pi^2 \alpha f_{2}}{M^4}\cdots,\\
r_h &\doteq 2M + \frac{6 \pi \alpha f_{1} }{M^3} - \frac{15 \pi \alpha f_{2} }{M^5} +\cdots.
\end{align}
Though the general expressions for the thermodynamic quantities up to any value of $n$ of eq.\eqref{boxes} haven't been found, the interested readers can rapidly generate further results by a simple program following our procedure.
\section{Some exact results}
\subsection{An exact relation between $I_1$ and $I_3$}
In computing the thermodynamic quantities, a lot of data can be easily generated and accumulated, from which we have a chance of finding some rules or patterns. In particular, we find an interesting relation between $I_1$ and $I_3$. This is an exact relation, meaning that it is applicable to the exact black hole solutions, or perturbative solutions to any order.
As an example, substituting eqs.\eqref{curT} and \eqref{curR} into the expression for $I_1$ and $I_3$, we find
\begin{align}
I_1 &\doteq -\frac{\beta ^2}{16 \pi }-(-\frac{\beta M}{2}) \doteq \alpha \frac{8 \pi ^2 }{M^2} \\
I_3 &\doteq \alpha \frac{512 \pi ^4 }{\beta ^2} \doteq \alpha \frac{8 \pi ^2 }{M^2}.
\end{align}
Obviously, there is $I_1=I_3$. At first sight, it is not transparent why this should be true, since the definitions for $I_1$ and $I_3$ look so different. But it is not that curious from another perspective. For the original SBH in Einstein gravity, $I_1=0$; after the higher curvature term is added, the metric gets shifted a little, causing $I_1$ nonzero and closely related to the value of $I_3$.
Generally, we find
\begin{align}
I_1=(m-2)I_3,
\end{align} where $m$ counts the number of curvatures $R$, $R_{\mu\nu}$ $R_{\mu\nu\rho\sigma}$, and also half the number of the operator $\nabla_\mu$'s that are explicitly included in $ F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})$. Taking a random example, $R R_{\alpha \beta }R^{\alpha \nu\rho\sigma} \Box R^{\beta}_{\ \nu\rho\sigma}$, one counts $m=5$. We further explore this issue for SBHs in $D \geq 4$ dimensions, and the formula turns out to be
\begin{align}
I_1=-\frac{D-2m}{D-2}I_3. \label{I13}
\end{align}
A careful examination of various examples shows that the formula \eqref{I13} is actually an exact formula applicable to any orders, so there must be some mathematical structure hidden behind it. Surely, we can figure out a proof for the formula \eqref{I13} as below.
Varying the action \eqref{action}, we find that the gravitational field equations always have the structure
\begin{align}
\frac{1}{16 \pi} (R^{\mu \nu}-\frac{1}{2} g^{\mu \nu}R)=-\alpha ( f^{\mu \nu}-\frac{1}{2} g^{\mu \nu}F) - \alpha \nabla^\mu K^\nu, \label{eofpr}
\end{align}
where $F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})$ is abbreviated as $F$. On the right hand side, $\frac{1}{2} g^{\mu \nu}F$ comes from $\delta(\sqrt{-g})\, F$, and $-\alpha ( f^{\mu \nu}+ \nabla^\mu K^\nu)$ comes from $\sqrt{-g}\,\delta F$. Specifically, $f_{\mu\nu}$ has a useful property $g_{\mu \nu}f^{\mu\nu}=m F$, and $\nabla^\mu K^\nu$ represents a total derivative term.
In order to understand eq.\eqref{eofpr} better, consider the typical example $F=R R_{\alpha \beta }R^{\alpha \nu\rho\sigma} \Box R^{\beta}_{\ \nu\rho\sigma}$, and write it as
\begin{align}
\begin{split}
F=& g^{\mu_1 \mu_2} g^{\nu \mu_3}g^{\rho \mu_4}g^{\sigma \mu_5} g^{\mu_6 \mu_7}\\
\bullet\, & R_{\mu_1 \mu_2} R_{\alpha \beta} R^{\alpha}_{\ \mu_3\mu_4\mu_5} \nabla_{\mu_6} \nabla_{\mu_7} R^{\beta}_{\ \nu\rho\sigma}.
\end{split}
\end{align}
Then $\delta F $ can be written in the form $- ( f^{\mu \nu}+ \nabla^\mu K^\nu) \delta g_{\mu \nu}$. The five terms deduced from the five different $\delta g^{\alpha \beta}= -g^{\alpha \mu} g^{\beta\nu} \delta g_{\mu\nu}$ are collected in $- f^{\mu \nu}$. The terms deduced from $\delta R^{\alpha \beta }$, $\delta R^{\beta}_{\ \nu\rho\sigma}$ and $\delta \nabla_\mu\sim \delta \Gamma$ are always total derivative terms and collected in $-\nabla^\mu K^\nu $. Moreover, when contracted with $g_{\mu \nu}$, each of the five terms in $f^{\mu \nu}$ gives back to the same form $F$, so $g_{\mu \nu}f^{\mu\nu}=5 F$ in this case.
Let's continue to do the proof. With eq.\eqref{eofpr} at hand, contracting it with $g_{\mu \nu}$ and do the Euclidean integral from the horizon radius to infinity, we get
\begin{align}
(1-\frac{D}{2}) \int_\mathcal{M} \frac{ R}{16 \pi}=- (m-\frac{D}{2}) \int_\mathcal{M} \alpha F- \int_\mathcal{M} \alpha \nabla_\mu K^\mu,\label{pr1}
\end{align}
where $\int_\mathcal{M} \centerdot \equiv \beta \int d^3 x\sqrt{-g}\,\centerdot$. In fact, we can get rid of the last term, because it simply vanishes \footnote{In the case $m \neq \frac{D}{2}$, $ \int_\mathcal{M} \nabla_\mu K^\mu$ is dimensionful by direct counting, so it must vanish. If not, it would be a function of $M$; imagining adding $ \alpha \int d^4 x \sqrt{-g} \nabla_\mu K^\mu$ to the action, it would not change the black hole solutions while changing the thermodynamics, which is physically unreasonable. The case $m=\frac{D}{2}$ is another story. Now $ \int_\mathcal{M} \nabla_\mu K^\mu$ is dimensionless, so in principle it could be some number denoted by $c_1$. Because of eq.\eqref{pr1}, $I_1= \frac{2 \alpha}{D-2}c_1$. Then the partition function is $I\doteq -\frac{\beta M(\beta)}{D-2}+\text{const.}$, from which we can solve the thermodynamics and finally get $I_1= 0$. So there must be $c_1=0$. The fact of $I_1\equiv \int_M \frac{R}{16\pi}=0$ ($m=\frac{D}{2}$) is trivial in $D=4$ dimensions, because $\mathcal{O}(R^2)$ terms don't
t change the SBH metric where $R=0$. And this is non-trivial in higher dimensions, since the SBH metric can be changed along with $R\neq 0$.}. Thus, eq.\eqref{pr1} can be re-expressed as
\begin{align}
\int_\mathcal{M} \frac{ R}{16 \pi}=-\frac{D-2m}{D-2} \int_\mathcal{M} \alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma}),
\end{align}
which finishes our proof of eq.\eqref{I13}, applying the definition of $I_1$ and $I_3$.
Clearly, when $F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})$ is a sum of individual higher curvature terms $\mathcal{O}(R^m)$, the formula becomes
\begin{align}
I_1=-\sum_m\frac{D-2m}{D-2}I_3^{(m)} .
\end{align}
\subsection{On the Gauss-Bonnet gravity in higher dimensions}
The Gauss-Bonnet gravity is simple in that its field equations only contain second derivatives of the metric. Thus, the exact black hole solutions are not hard to obtain in higher dimensions \cite{Boulware:1985wk,Cai:2001dz}, and there is no necessity to derive the thermodynamics from our approach. But maybe it is still interesting to see how the exact thermodynamic quantities emerge with only partial information of the metric. We shall work in $D=5$ dimensions as the example.
Assume we have known $A(r)=B(r)$ and their zeroth order approximation, i.e., the original SBH metric $A(r)=B(r)
\simeq 1-\frac{8 M}{3 \pi r^2}$. Note that the $D=5$ generalizations of eqs.\eqref{eu1} and \eqref{eu2} are
\begin{align}
I_1 = \frac{\beta M}{3}+\frac{\pi ^2 r_h^3}{2}-\frac{3}{8} \pi \beta r_h^2, \label{gbI1}
\end{align}
and $I_2=-\frac{\beta M}{3}$. The Gauss--Bonnet term is
\begin{align}
\alpha F(g_{\mu\nu},\nabla_\mu,R_{\mu\nu\rho\sigma})=\frac{\alpha}{16 \pi}(R^2-4 R_{\mu \nu}R^{\mu \nu}+R_{\mu \nu \rho \sigma }R^{\mu \nu \rho \sigma }).
\end{align}
Substituting the metric into it, the Euclidean integral gives
$\beta\frac{3}{4} \pi \alpha [2 r (A(r)-1) A'(r)+A(r)^2-2 A(r)]$.
As $r\rightarrow \infty$, the expression reduces to $-\beta \frac{3\pi \alpha}{4}$. At horizon, it can be simplified by $A(r_h)=0$, $A'(r_h)=4\pi/\beta$. Then the subtraction between the infinity and the horizon radius leads to $I_3=6 \pi ^2 \alpha r_h-\beta \frac{3\pi \alpha}{4}$. In total, the partition funciton is
\begin{align}
I_1+I_2+I_3=\frac{\pi ^2 r_h^3}{2}+6 \pi ^2 \alpha r_h - \beta \left( \frac{3 \pi \alpha }{4}+\frac{3 \pi r_h^2}{8}\right).
\end{align}
Comparing with the standard form $\ln Z=S-\beta M$, we read
\begin{align}
M= \frac{3 \pi r_h^2}{8}+\frac{3 \pi \alpha }{4},\label{gbM}\\
S=\frac{\pi ^2 r_h^3}{2}+6 \pi ^2 \alpha r_h.
\end{align}
The temperature is calculated by
\begin{align}
T= \frac{\partial M}{\partial S}=\frac{r_h}{8 \pi \alpha +2 \pi r_h^2}.\label{gbT}
\end{align}
These are surely the exact results known for the Gauss-Bonnet black hole \cite{Cai:2001dz}.
Incidentally, we verify the formula \eqref{I13} again. Substituting eqs.\eqref{gbM} and \eqref{gbT} into the expressions of $I_1$ and $I_3$, and using $r_h$ as the independent variable, we have $I_1=\frac{2 \pi ^2 \alpha ^2}{r_h}-\frac{3}{2} \pi ^2 \alpha r_h $ and $I_3=\frac{9}{2} \pi ^2 \alpha r_h-\frac{6 \pi ^2 \alpha ^2}{r_h} $. As expected, there is
\begin{align}
I_1=-\frac{1}{3}I_3,
\end{align}
which exactly matches the case $D=5$, $m=2$ of eq.\eqref{I13}.
\section{conclusions}
In this paper, we present an interesting fact that, at least at the first order level, the corrections to the thermodynamic quantities of SBH, due to higher curvature terms in the action, can be easily obtained. We have recovered some known first-order results existing in the literature, and provided more results which may be hard (if not possible) to obtained from the conventional approach. Furthermore, we have found an exact relation between the Euclidean integrals of the EH term $I_1$ and higher curvature term $I_3$, and also proved it.
It is regretted that such an elegant procedure can \emph{not} be applied to higher orders in $\alpha$. And we also considered black holes in AdS background, though we didn't present it in the paper. Unless under some severe conditions (such as, the pure AdS metric is still the solution of the modified gravitational field equations), our approach can't be further applied to the AdS black holes. In short, it seems that our approach just provides a tiny window for us to take a rapid look at some of the thermodynamic properties of the black holes in higher curvature theories of gravity. To gain more, one should go back to the conventional powerful techniques.
\section*{Acknowledgments}
YX would like to thank Longting Zhang for helpful discussions. This work was supported in part by NSF of Hebei province with Grant No. A2021201022.
\section*{Appendix}
In this Appendix, we provide a brief proof for eqs.\eqref{eu1}, \eqref{eu2} and \eqref{act12}.
In the case $A(r)=B(r)$, $I_1\equiv \beta \int d^3 x \sqrt{-g} \frac{R}{16 \pi}$ can be integrated out explicitly. The outcome is $-\frac{1}{4} \beta r \left(r A'(r)+2 A(r)-2\right)$, where $A(r)=1-\frac{2M}{r} +\cdots$, with ``$\cdots$" representing the $r^{-n}$ ($n\geq 2$) corrections to the SBH metric due to the higher curvature terms. As $r\rightarrow \infty$, $r\left(2A(r)-2 \right)\rightarrow -4M$, $r^2 A'(r)\rightarrow 2M$, so the expression gives $\frac{\beta M}{2}$. At the horizon, it can be simplified by $A(r_h)=0$,
$A'(r_h)=4 \pi/{\beta}$. Then, the subtraction between the infinity and the horizon radius leads to $I_1=\frac{\beta M}{2}+\pi r_h^2-\frac{\beta r_h}{2}$. In the case $A(r)\neq B(r)$, unfortunately the integration can't yield exact expressions, so the analysis can only be done at the first order in $\alpha$. Consequently, in general, the above formula of $I_1$ holds only at the first order level. Because the steps are more involved, we don't intend to present it here, but leave it as an exercise.
Next we explain that the boundary terms lead to $I_2= -\frac{\beta M}{2}$. The boundary terms are important to ensure the field equations are well-defined. In particular, the boundary term corresponding to the EH term is the GHY term. And the boundary terms corresponding to the higher curvature terms are often complicated and not generally known. So, in order to get the contributions from the boundary terms, we adopt another strategy. We first consider the black holes in the AdS space-time. There is a subtraction procedure for calculating the Euclidean action of AdS black holes without the necessity of considering the boundary terms \cite{hawkingpage,Dutta:2006vs}, even if higher curvature terms have been included in the action. Then, by taking the limit $\Lambda \rightarrow 0$, we will obtain the Euclidean action for the asymptotic-flat black holes and identify the contributions from the boundary terms. Concretely speaking, for Einstein gravity, the boundary term can be calculated by $ \beta \lim\limits_{r\rightarrow\infty}
[\frac{ r^3\Lambda}{6}- (1+\frac{3 M}{ r^3 \Lambda} ) \frac{ r^3 \Lambda}{6} ] =-\frac{\beta M}{2}$. When higher curvature terms are added, the result can be sketched as the form $-\beta (\frac{M}{2} + \alpha \Lambda M +\cdots )$. Taking the limit $\Lambda\rightarrow 0$, we get neatly $I_2= -\frac{\beta M}{2}$.
In the end, we have $I_1+I_2\doteq \pi r_h^2-\frac{\beta r_h}{2}$. Assume $r_h \doteq 4\pi\beta +\alpha f_1(\beta)$ and substitute it into the expression to reduce the number of variables. Then the unknown function $f_1(\beta)$ cancels; the formula becomes $I_1+I_2\doteq -\frac{\beta ^2}{16 \pi }$.
\baselineskip=1.6pt
|
\section{Introduction}
Symbolic regression (SR) is a sub-field of machine learning concerned with discovering a model of the given data in the form of a mathematical expression (or equation)~\citep{koza1994genetic,schmidt2009distilling}.
For example, consider having measurements of planet masses $m_1$ and $m_2$, the distance $r$ between them, and the respective gravitational force $F$.
Then, an SR algorithm would ideally re-discover the well-known expression (or an equivalent formulation thereof) $F=G \times \frac{m_1 m_2}{r^2}$, with $G = 6.6743 \times 10^{-11}$, by opportunely combining the mathematical operations (here, of multiplication and division) with the variables and constant at play.
The appeal of learning models as mathematical expressions goes beyond obtaining predictive power alone, as is commonplace in machine learning.
In fact, SR models are particularly well suited for human interpretability and in-depth analysis~\citep{otte2013safe,virgolin2021model,la2021contemporary}.
This aspect enables a safe and responsible
use of machine learning models for high-stakes societal applications,
as requested in the AI acts by the European Union and the United States~\citep{euact,usact,jobin2019global}.
Moreover, it enables scientists to gain deeper knowledge about the phenomena that underlie the data.
Consequently, SR enjoys wide applicability: SR has successfully been applied to
astrophysics~\citep{lemos2022rediscovering}, chemistry~\citep{hernandez2019fast}, control~\citep{derner2020constructing}, economics~\citep{verstyuk2022machine}, mechanical engineering~\citep{kronberger2018predicting}, medicine~\citep{virgolin2020machine}, space exploration~\citep{martens2022symbolic}, and more~\citep{matsubara2022rethinking}.
As we will describe in \Cref{sec:related}, many different algorithms have been proposed to address SR, ranging from genetic algorithms to deep learning ones.
Existing algorithms either lack optimality guarantees or heavily restrict the space of SR models to consider.
In fact, there is a wide belief in the community that SR is an NP-hard problem\footnote{\cite{lu2016using} state that SR is NP-hard but provide no reference nor proof.}~\citep{lu2016using,petersen2019deep,udrescu2020ai,li2022console}.
However, to the best of our knowledge, this belief had yet to be solidified in the form of a proof prior to the advent of this paper.
Indeed, we prove that there exist instances of the SR problem for which one cannot discover the best-possible mathematical expression in polynomial time.
Id est, SR is an NP-hard problem.
\section{Background: Existing SR algorithms}
\label{sec:related}
The introduction of SR is generally attributed to John R.~Koza (e.g., \cite{zelinka2005analytic} make this claim); however, the problem of finding a mathematical expression or equation that explains empirical measurements was already considered in earlier works~\citep{gerwin1974information,langley1981data,falkenhainer1986integrating}.
Such works build mathematical expressions by iterative application of multiple heuristic tests on the data.
\begin{wrapfigure}{r}{0.3\textwidth}
\vspace{-.3cm}
\centering
\includegraphics[width=0.9\linewidth]{tree.pdf}
\caption{Example of a tree that encodes $f(\mathbf{x})=\left( \sin(x_1)+x_2 \right) \times x_3 / x_1$.}
\label{fig:graph-sr}
\vspace{-.25cm}
\end{wrapfigure}
Koza is best known for his pioneering work on genetic programming (GP), i.e., the form of evolutionary computation where candidate solutions are variable-sized and represent programs~\citep{koza1989hierarchical,koza1990genetic,koza1994genetic}.
Early forms of GP where proposed by~\cite{cramer1985representation,hicklin1986application}.
Koza showed that GP can be used to discover SR models by encoding mathematical expressions as computational trees (see \Cref{fig:graph-sr}).
In such trees, internal nodes represent functions (e.g., $+$, $-$, $\times$, etc.) that are drawn from a pre-decided set of possibilities, and leaf nodes represent variables or constants (e.g., $x_1$, $x_2$, \dots, $-1$, $\pi$, etc.).
GP evolves a population of trees by initially sampling random trees, and then conducts the following steps: (1) stochastic replacement and recombination of their sub-trees; (2) evaluation of the fitness by executing the trees and assessing their output; and (3) stochastic survival of the fittest.
Recently, \cite{la2021contemporary} proposed \emph{SRBench}, a benchmarking platform for SR that includes more than $20$ algorithms have been applied on more than 250 data sets.
SRBench shows that several state-of-the-art algorithms for SR are GP-based.
Among these, at the time of writing, \emph{Operon} by~\cite{burlacu2020operon} was found to perform best in terms of discovering accurate SR models; and \emph{GP-GOMEA} by~\cite{virgolin2021improving} was found to perform best in terms of discovering decently-accurate and relatively-simple SR models (i.e., shorter mathematical expressions).
Other forms of GP, such as \emph{strongly-typed} GP~\citep{montana1995strongly}, \emph{grammar-guided} GP~\citep{mckay2010grammar}, and \emph{grammatical evolution}~\citep{o2001grammatical}, are often used to tackle \emph{dimensionally-aware} SR, i.e., the search of mathematical expressions with constraints to achieve meaningful combinations of units of measurement.
SR has been addressed with many other types of algorithms than genetic ones, oftentimes in order to obtain a deterministic behavior.
\cite{worm2013prioritized} and \cite{kammerer2020symbolic} proposed enumeration algorithms which make SR tractable by restricting the space of possible models to consider and including dynamic programming and pruning strategies.
\cite{cozad2014data,cozad2018global} showed how SR can be addressed with mixed integer nonlinear programming.
\cite{mcconaghy2011ffx} proposed \emph{FFX}, which generates a linear combination of many functions that are linearly-independent from each other, and then fits its coefficients with the \emph{elastic net}~\citep{zou2005regularization} to promote sparsity.
\cite{olivetti2018} and \cite{rivero2022dome} propose greedy algorithms that start from small mathematical expressions and iteratively expand them, by replacing existing components with larger ones from a set of possibilities.
Lastly, recent years have seen the proposal of deep learning-based algorithms for SR.
\cite{petersen2020deep} cast the SR problem as a reinforcement learning one and train a recurrent neural network to generate accurate SR models.
\cite{udrescu2020ai} leverage neural networks in order to test for symmetries and invariances in the data that are then used to prune the space of possible SR models.
An end-to-end approach is taken by~\cite{kamienny2022end} and~\cite{vastl2022symformer}, who train deep neural transformers to produce SR models directly from the data.
\cite{li2022console} seek SR models by proposing a convexified formulation of deep reinforcement learning.
In summary, existing SR algorithms are either heuristics, which do not guarantee optimality (e.g., genetic, greedy, or deep learning-based algorithms), or they are exact algorithms that achieve optimality but only over a small subset of all possible SR models, to limit the runtime (e.g., dynamic programming and mixed-integer nonlinear programming algorithms).
This strongly hints to the fact that SR is NP-hard.
As mentioned earlier, no proof has yet been given.
\section{Preliminaries}
\label{sec:preliminaries}
We will hereon refer to SR models as \emph{functions} when appropriate, as this is their fundamental nature.
To begin, let us recall the concept of function composition, which is central to SR.
\begin{definition}
\emph{Function composition}
Given two functions $f : \mathbb{A} \rightarrow \mathbb{B}$ and $g : \mathbb{B} \rightarrow \mathbb{C}$, function composition, which we denote by $g \circ f$, is the operation that produces a third function $h : \mathbb{A} \rightarrow \mathbb{C}$, such that $h(x)=g(f(x))$.
\end{definition}
Thanks to function composition, we can now define the concept of \emph{search space} of an SR problem.
\begin{definition}
\label{def:searchspace}
\emph{Search space of SR}
Let $\mathcal{P}$ be a set of functions and variables.
The search space of SR is the function space $\mathcal{F}$ that contains all functions that can be formed by composition of the elements of $\mathcal{P}$ and their compositions.
\end{definition}
To better understand what \Cref{def:searchspace} states, consider that $\mathcal{P}$ can be set to contain a mix of functions that perform basic algebraic operations such as addition, subtraction, multiplication, and division; transcendental functions such as $\sin$, $\cos$, $\log$, $\exp$; constant functions (or simply \emph{constants}), such as $c_{42}(x)=42$ and $c_\pi(x)=\pi$ for any $x$; and variables of interest for the problem at hand, such as $x_1, x_2, x_3$.
$\mathcal{P}$ is typically referred to as the \emph{primitive set}, and its elements as \emph{primitives}~\citep{poli2008field}.
Once $\mathcal{P}$ has been decided, $\mathcal{F}$ is determined.
For example, choosing $\mathcal{P} = \{+(\cdot,\cdot), -(\cdot, \cdot), \times(\cdot,\cdot), x_1, x_2, -1, +1\}$ means that $\mathcal{F}$ will contain a subset of all possible polynomials of arbitrary degree in $x_1$ and $x_2$.
In particular, $\mathcal{F}$ is a subset because only some coefficients can be expressed, by composing constants with addition, subtraction, and multiplication.
Let us clarify a point regarding constants in particular.
Normally, one would include constants considered to be relevant to the instance of SR at hand.
For example, if the unknown phenomenon for which an SR model is sought is suspected to have sinusoidal components, it may be advisable to include multiples of $\pi$ in $\mathcal{P}$.
Moreover, $\mathcal{P}$ can be set to contain special elements that represent probability distributions from which constants can be sampled (see the concept of \emph{ephemeral random constant} described by~\cite{koza1994genetic,poli2008field}).
We denote one such element by $\mathfrak{R}$ and, e.g., $\mathfrak{R}$ can be chosen to represent the uniform distribution between two numbers, or the normal distribution with a certain mean and variance.
When an SR algorithm picks $\mathfrak{R}$ from $\mathcal{P}$ to compose an SR model, a constant is sampled from the distribution identified by $\mathfrak{R}$.
Here (more specifically, in \Cref{th:corollary-const}) we will generously assume that any constant can be sampled directly from $\mathfrak{R}$, and therefore that there is no need for a real-valued optimizer to be part of the SR algorithm.
For example, having $\mathcal{P} = \{+(\cdot,\cdot), -(\cdot, \cdot), \times(\cdot,\cdot), x_1, x_2, \mathfrak{R}\}$ will mean that $\mathcal{F}$ contains \emph{all} polynomials of arbitrary degree in $x_1$ and $x_2$.
We can now proceed by providing a definition of the SR problem.
While this definition can be extended to other domains, we focus on handling real-valued numbers as the majority of the works takes place in this domain, and subsets thereof.
\begin{definition}
\label{def:sr}
\emph{Symbolic Regression (SR) problem}
Given a set $\mathcal{P}$ of functions and variables,
a metric $\mathcal{L} : \mathbb{R}^n \times \mathbb{R}^n \rightarrow \mathbb{R}$,
vectors $\mathbf{x}_i = \left( x_{1,i}, \dots, x_{d,i} \right) \in \mathbb{R}^d$
and scalars $y_i \in \mathbb{R}$, for $i=1, \dots, n$,
the SR problem asks for finding a function $f^\star$ such that:
\begin{equation}
\label{eq:sr}
f^\star = \argmin_{f \in \mathcal{F}} \mathcal{L} \left( \mathbf{y}, f(\mathbf{x}) \right),
\end{equation}
where $\mathcal{F}$ is the search space that is defined by $\mathcal{P}$.
\end{definition}
We provide some remarks concerning the proposed definition of the SR problem.
Firstly, let us map the objects provided in the definition to terms familiar to a machine learning audience.
The pair $(\mathbf{x}_i, y_i)$ is normally what is referred to as \emph{observation}, \emph{data point}, \emph{example}, or \emph{sample}, where $x_{j,i}$ is the value of the $j$th \emph{feature} or \emph{variable} for the $i$th observation, and $y_i$ is the value of the \emph{label} or \emph{target variable} for the same observation.
The set that contains the observations upon which $\mathcal{L}$ is computed, i.e., $\mathcal{D}=\{ (\mathbf{x}_i, y_i) \}^n_{i=1}$, is called \emph{training set}.
Moreover, the metric $\mathcal{L}$ is called \emph{loss function}.
Normally, we actually desire $f$ to \emph{generalize} to \emph{new} (or also called \emph{unseen}) observations, i.e., observations which are similar to those in $\mathcal{D}$ but not exactly the same (they come from the same underlying and unknown distribution).
In other words, it is not sufficient that $f^\star$ is a best-possible function with respect to the training set, as the loss should remain minimal also for new observations that are not available to us.
Still, considering a ``pure optimization'' formulation, as given in \Cref{eq:sr}, can be considered to be a pre-requisite for being able to machine-learn accurate models from the data; in fact, much literature that concerns the generation of provably-optimal models provides proofs with respect to the training set alone (see, e.g., results for decision trees~\citep{hu2019optimal}).
In a similar fashion, here we will consider the case of minimizing the loss with respect to the training set $\mathcal{D}$ and show that this is already problematic for any SR algorithm.
As loss function, we consider $\mathcal{L}$ to be a metric (i.e., distance) which operates between the output $f(\mathbf{x}_i)$ and the label $y_i$ across $i=1,\dots,n$.
Commonly-used loss functions such as the \emph{mean absolute error}, \emph{mean squared error}, and \emph{root mean squared error} fit this definition.
However, certain works include regularization terms in the loss function, such as $\lambda \times C(f)$, where $\lambda \in \mathbb{R}$ controls the regularization strength and $C : \mathcal{F} \rightarrow \mathbb{R}$ is a function of the complexity of $f$.
Typical goals of such regularization terms are improving generalization (by limiting effects akin to Runge's phenomenon~\citep{fornberg2007runge}) and improving the interpretability of $f$.
For the latter, implementations of $C$ range from weighed counting of the number of primitives that constitute $f$~\citep{ekart2001selection,hein2018interpretable}, to machine learning models trained from human feedback to predict $f$'s interpretability~\citep{virgolin2020learning,virgolin2021model}.
Here, for simplicity, we focus on $\mathcal{L}$ being a plain metric as stated in \Cref{def:sr} or, equivalently put, we consider $\lambda=0$.
Lastly on \Cref{def:sr}, we consider only cases in which $f$ is \emph{not} a recursive function, which to the best of our knowledge is the case for the majority of the literature on SR.
Recursive function discovery is an interesting topic in general (see, e.g.,~\cite{d2022deep}), but it is not interesting here because recursive functions can take exponential time to compute (consider, e.g., Fibonacci's sequence).
Therefore, it is obvious that the SR problem cannot be solved in polynomial time if certain recursive functions can be considered.
Here, we will assume that computing $f(\mathbf{x})$ and $\mathcal{L}(\mathbf{y},f(\mathbf{x}))$ can be done in polynomial time.
Regarding $\mathcal{L}$, our assumption is met for all commonly-used metrics by which $\mathcal{L}$ is implemented (mean absolute error, mean squared error, variants thereof with margins, etc.).
In fact, computing losses of such form takes $O(n)$ time, i.e., the runtime is linear in the number of observations.
Regarding the computation of $f(\mathbf{x})$, $f$ itself can be implemented as a directed acyclic graph, where nodes represent the functions and variables from $\mathcal{P}$, and edges represent compositions.
To compute $f(\mathbf{x})$, it suffices to visit each node of the graph for each observation, thus requiring $O(\ell \times n)$, where $\ell$ is the number of primitives in $f$.
\Cref{fig:graph-sr} shows an example of such a graph, especially in the form of a \emph{tree}, which is perhaps the most common way of encoding mathematical expressions in SR (see, e.g., the SR algorithms benchmarked by~\cite{la2021contemporary}).
We conclude this section with the following important definition.
\begin{definition}
\label{def:sr-dec}
\emph{Decision version of the SR problem (SR-Dec)}
Given an SR instance and an $\epsilon \in \mathbb{R}^+_0$, SR-Dec outputs \textsf{YES} if and only if:
\begin{equation}
\label{eq:sr-dec}
\exists f \in \mathcal{F} : \mathcal{L} \left( \mathbf{y}, f(\mathbf{x}) \right) \leq \epsilon.
\end{equation}
\end{definition}
Essentially, \Cref{def:sr-dec} is the problem of deciding whether there exists a function $f$ in the search space such that its loss is smaller than a chosen threshold $\epsilon$.
\section{The result}
We proceed directly by providing the main result of this paper.
\begin{theorem}
\label{th:sr-nphard}
The SR problem is NP-hard.
\end{theorem}
\begin{proof}
Let us begin by stating that SR-Dec is in NP.
Recall that the computations of $f(\mathbf{x})$ and $\mathcal{L}(\mathbf{y},f(\mathbf{x}))$ take polynomial time (see \Cref{sec:preliminaries}).
Of course, the check $\leq \epsilon$ takes $O(1)$ time.
Thus, if $f$ is guessed by an oracle, then we can provide an answer to SR-Dec in polynomial time.
We proceed by considering the unbounded subset sum problem (USSP).
USSP is a similar problem to the unbounded knapsack problem, where a same item can be put in the knapsack an arbitrary number of times, and the weight of an item corresponds exactly to the profit gained by including that item in the knapsack.
The decision version of USSP, USSP-Dec, is defined as follows.
Given $j=1, \dots, k$ ($k$ items), $w_j \in \mathbb{N}$ (weight of that item), and $t \in \mathbb{N}$ (the target), USSP-Dec asks:
\begin{equation}
\label{eq:ussp}
\exists \mathbf{m} : \sum_{j=1}^k w_j m_j = t?
\end{equation}
where $m_j \in \mathbb{N}_0$ (multiplicity with which an item is picked).
USSP-Dec is known to be NP-complete~\citep{kellerer2004introduction}.
To prove that SR-Dec is NP-complete, we show that any instance of USSP-Dec can be reduced to some instance of SR-Dec in polynomial time.
To this end, we will restrict SR-Dec as follows:
(1) We pick the set of primitives $\mathcal{P}$ to be $\mathcal{P}=\{+, x_1, \dots, x_d \}$;
(2) We set $\epsilon = 0$.
In other words, we set the search space $\mathcal{F}$ to contain only linear sums of the features in the data set $\mathcal{D}$, i.e., functions of the form $f(\mathbf{x}) = \sum^d_{j=1} x_j m_j$ with $m_j \in \mathbb{N}_0$.
SR-Dec will output \textsf{YES} if and only if there exists such a function in $\mathcal{F}$ that achieves zero loss, i.e., it perfectly interpolates all observations in $\mathcal{D}$.
Next, we craft $\mathcal{D}$ to have a single observation ($n=1$) and $k$ features ($d=k$).
For the only observation in $\mathcal{D}$ (dropping the index for the observation number, since there is only one), we set $x_1 = w_1, x_2 = w_2, \dots, x_k = w_k$, and $y = t$.
Then, the following holds:
\begin{align}
\label{eq:derivation}
\emph{(\Cref{eq:sr-dec}) \ \ } & \exists f \in \mathcal{F} : \mathcal{L} \left( y, f(\mathbf{x}) \right) \leq \epsilon? \\
\emph{(Choosing $\epsilon = 0$) \ \ } & \exists f \in \mathcal{F} : \mathcal{L} \left( y, f(\mathbf{x}) \right) \leq 0? \\
\emph{($\mathcal{L} \left( y, f(\mathbf{x}) \right)=0 \iff f(\mathbf{x})=y$) \ \ } &
\exists f \in \mathcal{F} : f(\mathbf{x}) = y? \\
\emph{(Equivalence $y=t$ due to $\mathcal{D}$) \ \ } &
\exists f \in \mathcal{F} : f(\mathbf{x}) = t? \\
\emph{(Expanding $\mathcal{F}$ based on choice of $\mathcal{P}$) \ \ } &
\exists f \in \left\{ \sum_{j=1}^d x_j m_j : m_j \in \mathbb{N}_0 \right\} : f(\mathbf{x}) = t? \\
\emph{(Equivalence $x_j=w_j, d=k$ due to $\mathcal{D}$) \ \ } &
\exists f \in \left\{ \sum_{j=1}^k w_j m_j : m_j \in \mathbb{N}_0 \right\} : f(\mathbf{x}) = t? \\
\emph{(Re-formulating in terms of $\mathbf{m}$) \ \ } &
\exists \mathbf{m} : \sum_{j=1}^k w_j m_j = t? \label{eq:last-step}
\end{align}
In other words, there exist some instances of SR-Dec that can be re-formulated as USSP-Dec (cfr.~\Cref{eq:ussp,eq:last-step}).
Now, since assembling $\mathcal{P}$ as stated above takes linear time in $k$, picking $\epsilon=0$ takes $\mathcal{O}(1)$ time, and constructing $\mathcal{D}$ as stated above takes linear time in $k$, then any instance of USSB-Dec can be reduced to some instance of SR-Dec in polynomial time: SR-Dec is NP-complete.
We conclude the proof with a \emph{reductio ab absurdum}.
Let us assume that there exists an algorithm to compute an optimal $f^\star$ for the SR problem (\Cref{def:sr}) in polynomial time.
An optimal $f^\star$ is the one for which the loss is minimal, which means that using $f^\star$ in \Cref{eq:sr-dec} allows us to immediately answer SR-Dec.
Since verifying that $\mathcal{L}(\mathbf{y},f^\star(\mathbf{x})) \leq \epsilon$ takes polynomial time, we conclude that if the SR problem can be solved in polynomial time, then we can also solve SR-Dec in polynomial time.
Therefore, the SR problem is NP-hard.
\end{proof}
We remark that, in the proof of \Cref{th:sr-nphard}, we construct $\mathcal{P}$ so as not to contain $\mathfrak{R}$ (nor any constant).
Some readers might disagree with this quite broad definition of SR.
In fact, some SR algorithms heavily rely on the presence of constants as well as on their optimization (e.g., \emph{FFX} by~\cite{mcconaghy2011ffx} and \emph{FEAT} by~\cite{la2018learning}).
Not allowing for arbitrary constants to be present in the functions of the search space might be seen as a violation of the very definition of SR.
In other words, some might think that $\mathcal{P}$ \emph{must} contain $\mathfrak{R}$.
We next show that SR remains NP-hard in this special case.
\begin{corollary}
\label{th:corollary-const}
The SR problem is NP-hard even when $\mathcal{P}$ must include $\mathfrak{R}$.
\end{corollary}
\begin{proof}
We follow a similar construction of the proof of \Cref{th:sr-nphard}.
Namely, the only difference from before is in the way we pick $\mathcal{P}$ and construct $\mathcal{D}$.
This time, we set $\mathcal{P}$ to additionally contain $\mathfrak{R}$, i.e., $\mathcal{P}=\{+,x_1,x_2,\dots,x_d,\mathfrak{R}\}$.
This means that the function space $\mathcal{F}$ now contains functions of the form $f(\mathbf{x}) = c + \sum^d_{j=1} x_j m_j$ with $m_j \in \mathbb{N}_0$ and $c \in \mathbb{R}$ (sampled from $\mathfrak{R}$).
As to $\mathcal{D}$, we will now include two observations instead of a single one.
The first observation is set as before, i.e.,
$x_{1,1} = w_1, x_{2,1} = w_2, \dots, x_{k,1} = w_k$ ($d=k$) and $y_1 = t$.
As to the second observation, we set
$x_{1,2} = 0, x_{2,2} = 0, \dots, x_{k,2} = 0$ and $y_2=0$, i.e., the value of all features and of the label are set to zero. Now, $\mathcal{L}(\mathbf{y}, f(\mathbf{x}))=0 \iff f(\mathbf{x}_i) = y_i$ for \emph{both} $i=1,2$.
For $f(\mathbf{x}_2) = y_2 = 0$, since any $f$ has the form
$f(\mathbf{x}) = c + \sum^d_{j=1} x_j m_j$ and $x_{j,2} = 0$, for all $j$, then $f(\mathbf{x}_2) = c + \sum^d_{j=1} 0 \times m_j = c$.
But $f(\mathbf{x}_2) = y_2 = 0 \iff c = 0$.
In other words, we know that every $f$ for which $c \neq 0$ is one for which SR-Dec outputs \textsf{NO}.
Therefore, by construction, we can immediately ignore all of those functions, and consider only the subset of $\mathcal{F}$ that contains functions of the form $f(\mathbf{x}) = 0 + \sum^d_{j=1} x_j m_j = \sum^d_{j=1} x_j m_j$.
For every one of such functions, the loss for the second observation is by construction zero and we can therefore ignore it.
Consequently, we are now back to the same setting considered in \Cref{th:sr-nphard}, which concludes the proof.
\end{proof}
\section{Conclusion}
Our main contribution here was to prove that
symbolic regression (SR), i.e., the problem of discovering an accurate model of data in the form of a mathematical expression, is in fact NP-hard.
In particular, we have provided formal definitions of what SR should entail, and showed how the decision version of the unbounded subset sum problem can be reduced to a decision version of the SR problem.
Except for the general definition of SR we considered, we have additionally shown that SR remains NP-hard even when the set of primitives must contain distributions from which constants can be sampled.
We hope that this note inspires more works on lower and upper bounds of different SR variants.
\section*{Acknowledgements}
We thank Marc Schoenauer (INRIA, France) for discussions on early works on symbolic regression.
\bibliographystyle{abbrvnat}
|
\section{Introduction}
\label{sec:Introduction}
One of the oldest problems of philosophy is whether the universe had an ultimate origin. In the context of relativistic cosmology, that is usually translated into the question of whether there was a past singularity. In standard Friedmann Robertson Walker (FRW) big bang models, causal development of the initial curvature singularity is indeed a past boundary of spacetime. By contrast, the once-popular steady-state cosmological model, involving the continuous creation of particles \cite{HoyleNarlikar1964:1,HoyleNarlikar1964:2,HoyleNarlikar1966}, is eternal. Criticism of the big bang theory by advocates of the steady-state theory were summed up by the aphorism that ‘it merely claims things are as they are because they were as they were.’ In other words, the observed nature of the universe is simply relegated to unexplained initial conditions, and specifically to its ordered, low-entropy, initial state. However, in the steady state model the theory is also incomplete, but for a different reason. The metric for a FRW space is given as:
\begin{equation}
ds^2 = - dt^2 + f^2(t) \left( \frac{dr^2}{1-kr^2}+ r^2 d\Omega^2 \right) \label{FRW metric defn}
\end{equation}
with $f$ being the scale factor and $k$ being the spatial constant sectional curvature, usually normalized to $k \in \{ 0, +1, -1 \}$, and $d \Omega^2$ being the usual spherical metric. The steady state universe is described as a patch of de Sitter space with $f\left( t \right) = e^{H t}$ and $k=0$. The steady state universe, which expands at all times, occupies one-half of the complete de Sitter manifold (topologically $\mathbb{R}^1_1 \times S^{n-1}$). A coordinate system that covers the entire de Sitter spacetime is given by the metric of Eq.~\ref{FRW metric defn} with $f = \mathrm{cosh} \left( H t \right)$ and $k=+1$. It describes a universe that contracts to a minimum radius and then expands in a time-symmetric manner. Consider a world line $r = \mathrm{constant}$ in the latter system. Tracing back in time from future infinity, the world line traverses the steady state patch and eventually crosses the edge of the coordinate system, and hence out of the steady state universe. Clearly the steady state universe is geodesically incomplete. Recognizing this shortcoming, Hoyle and Narlikar proposed a modified theory \cite{HoyleNarlikar1964:1,HoyleNarlikar1964:2,HoyleNarlikar1966} using the entire spacetime, which involved a contraction followed by an expansion. This so-called C-field cosmology approaches the steady state in the far future and far past and can be made time symmetric by assuming matter creation to be oppositely directed in the regions $t > 0$ and $t < 0$. A closely analogous issue arises in the case of eternal inflation, in which the ‘units of creation’ are not subatomic particles but ‘bubble universes.’ A simple model of eternal inflation uses the steady state metric outside of the bubbles, and is therefore geodesically incomplete. However, this model can be amended with a minimum spacelike diameter, just like the full de Sitter case, which is complete.
Some confusion arises from the ambiguous use of the term `universe.' One definition is the entire spacetime, including all extensions, e.g. metric $f = \mathrm{cosh} \left( H t \right)$ and $k = +1$ in the foregoing example. Another definition is a congruence of worldlines orthogonal to a spacelike foliation bounded by horizons, e.g. steady state metric in the foregoing. A steady state observer on a worldline $r = \mathrm{constant}$ in this metric perceives an eternally expanding ‘universe’ with an event horizon, but the `universe' (i.e. the timelike foliated coordinate patch shown in Fig.~\ref{fig1.png}) is past incomplete.
Here we discuss how to characterize the geodesic completeness of a generalization of FRW spacetimes, and show that the aforementioned heuristic example holds for any inflationary spacetime. Our result is purely geometric, holding for any metrical theory of gravity.
\section{Inflation and Geodesic Completeness}
\label{sec:Inflation and Geodesic C}
Among the most compelling models of the early universe is the inflationary universe paradigm~\cite{Guth:1980zm}.
Shortly after its conception the question naturally emerged: could inflation be eternal, in particular, eternal into the past? If so, the universe itself could be eternal without the need for a definite beginning \cite{Linde:1986fc,Linde:1986fd,Goncharov:1987ir,Borde:1993xh,Borde2003,Mithani:2012ii}.
A critical component needed to answer
this question is a thorough understanding of singularities in general
spacetimes, see \emph{e.\!\!~g.}~\cite{Hawking1975,Beem1996,ONeill1983}, namely, spacetimes without any further assumptions on structure.
If a spacetime contains a singularity (curvature singularity or otherwise) it cannot inflate eternally.
One should ascertain whether a spacetime admits even a single \emph{incomplete} geodesic, namely a geodesic which does not exist for all values of affine parameter. If such an incomplete geodesic exists, the spacetime is singular; an observer on such a timelike geodesic trajectory experiences a catastrophic halt to their proper time, as their worldline cannot be extended past a particular spacetime event.
In the case of a curvature singularity an observer's clock stops because a spatial defect is encountered.
Venturing forward with pedantic mathematical concern, one must restrict to the class of \emph{maximal} geodesics--curves which cannot be further extended. A maximal geodesic $\gamma$ is defined as
\begin{equation}
\gamma : \left( a , b \right) \rightarrow \mathcal{M} \label{geodesic eqn}
\end{equation}
where $a,b$ is are real numbers in \emph{open} interval $\left( a , b \right)$ and $\mathcal{M}$ is the spacetime in question.
Critical information about the singularity structure of a given spacetime may be elucidated by studying the world lines of observers on maximal geodesic trajectories. In particular, if $\gamma$ is complete the domain is defined on all $\mathbb{R}$.
In a general model with very
little structure, it can be difficult to classify all geodesics in
a spacetime.~~Even if structure is introduced allowing
such a classification of geodesics, the geodesic equation may not have an analytic solution: thus advanced methods
vis \`{a} vis hyperbolic differential equations and functional analysis
are required.~~However, a sufficient condition for an incomplete
geodesic can be found utilizing Jacobi fields to find conjugate
or focal points, often easier to calculate than solving
all possible geodesic equations: this is the basis for the singularity theorems of \cite{Hawking1975}.
Demonstrating that a spacetime is free of incomplete geodesics
does not mean that the spacetime is physically reasonable. While free of singularities, Such models may violate reasonable energy conditions with `exotic' sources required to ensure that all
geodesics are complete.~~Nonetheless, spacetimes
which are geodesically complete are void of singularities at
the geometric level and one may hope that a reasonable metric theory of
gravity exists, capable of supporting these spacetimes.
Returning to the discussion of eternal inflation, \cite{Borde2003}, claims that any inflationary spacetime must be geodesically incomplete.~~Below, we provide an uncountably infinite cardnality of \emph{monsters} -- see \cite{Lakatos1976} --
counter-examples, showing that this cannot be the case.
These simple monsters are well defined within the realm of GR, although in that context,
they violate energy conditions for an arbitrarily short time.~~We begin
with a brief discussion of \cite{Borde2003} elucidating some of the contentious points therein.
We then introduce the class of Generalized Friedmann Robertson
Walker spacetimes (GFRW), and discuss their geodesic completeness. Next
we discuss properties of scale functions yielding inflationary monsters which satisfy the $H_{avg}>0$
condition of \cite{Borde2003} yet are geodesically
complete. Finally, we speculate on how quantum gravity may ultimately constrain the domain of discourse with respect to physically reasonable scale functions.
\section{What is Inflation?}
\label{sec:inflation}
Much of the contention surrounding $H_{avg}$ in \cite{Borde2003} stems from ambiguous underlying assumptions. We remedy this by attempting to explicitly state the underlying assumptions. Fittingly, we begin by discussing what is meant by the term \emph{inflationary}.
In the introduction of \cite{Guth:2007ng}, inflationary theory pioneer Alan Guth aptly states, ``the term \emph{inflation} encompasses a wide range of detailed theories'', and refrains from defining inflation in a mathematically precise manner. Popular definitions of inflation include ``a quasi-de Sitter period'', ``a period when gravity acted as a repulsive force'' or ``a period during which the Hubble sphere is shrinking''. Several authors prefer to discuss inflation in terms of \it what it does\rm, avoiding the need for a precise or rigorous definition. In \cite{Liddle:2000cg}, the authors define inflation as ``any epoch during which the scale factor of the Universe is accelerating'', and provide the defining equation
\begin{definition}
$
\mathrm{INFLATION} \Longleftrightarrow \ddot{f} > 0 \,, \label{inflation defn}
$
\end{definition}
\noindent
where we take $f$ to be the scale factor and ``$\cdot$'' denotes differentiation with respect to cosmic time $t$.
This definition assumes that the metric can be expressed as a warped product from \cite{Bishop1969} as
\begin{equation}
g = d \pi_1^* g_1 + f^2 \, d \pi_2^* g_2 \label{warped product defn}
\end{equation}
where topologically the manifold in question is a Cartesian product $\mathcal{M}_1 \times \mathcal{M}_2$ of two semi-Riemannian manifolds $\left( \mathcal{M}_1 , g_1 \right)$ and $\left( \mathcal{M}_2 , g_2 \right)$ and $d \pi_1^*, d \pi_2^* $ are the metric pullbacks of the canonical projections $\pi_1 : \mathcal{M} \rightarrow \mathcal{M}_1$, $\pi_2 : \mathcal{M} \rightarrow \mathcal{M}_2$ in the Cartesian product construction. Additionally, $f$ is a strictly positive smooth function ``warping'' the contribution of the second metric $g_2$ to the overall metric $g$.
Relating this to the standard Friedmann Robertson Walker (FRW) model of
cosmology and cosmogony, one has already foliated out a privileged
time dimension by assumption and by defining a scale function $f$ with respect to this
time dimension. The spacetime metric is that of Eq. \ref{FRW metric defn}.
In fact, the results apropos scale functions discussed in this paper are applicable to the broader class of Generalized Friedmann Robertson Walker (GFRW) spacetimes. These GFRW spacetimes are defined as the warped product
\begin{equation}
\mathbb{R}^1_1 \times_f \Sigma \label{GFRW defn}
\end{equation}
where $\left( \Sigma , g_\Sigma \right)$ is a complete\footnote{Any differentiable manifold admits a Riemannian metric - which can be pulled up from the underlying Euclidean space via the atlas, and any Riemannian metric is conformal to a complete metric~\cite{Nomizu1961}. Thus, given any (Hausdorff, second countable) reasonable space $\Sigma$, it can be endowed with a complete Riemannian metric.} Riemannian manifold -- taken to be the spacelike leaf -- and $\mathbb{R}^1_1$ is the timelike foliation. Note that the subscript is used to denote a negative definite index of the metric in $\mathbb{R}^1_1$ as in \cite{Beem1996,ONeill1983}. Finally a positive definite smooth warping function $f$ is used in the $\times_f$ notation for specification of the warped product metric $g = -dt^2 + f^2 \, g_\Sigma$. Additionally, the $d\pi_t^*, \, d\pi_\Sigma^*$ are unambiguous and have been suppressed.
One of the goals of \cite{Borde2003} is to generalize concepts from a highly structured FRW spacetime, so one may ask if a spacetime which cannot admit a global foliation of a time direction should be considered inflationary or not. In fact, if a dimension 3 or higher open neighborhood $U \subset \mathcal{M}$ admits a curl-free, local Killing vector field it is isometric to a warped product spacetime -- see \cite{Beem1996} Lemma 3.78. Are such spacetimes to be considered inflationary? In particular, it is much easier for local Killing vector fields to exist than global ones, so many spacetimes which cannot globally foliate out a time direction fall into this category.
An additional concern with respect to Defn.~\ref{inflation defn} are bouncing cosmologies. Clearly, we must have $\ddot f>0$ for a time during the transition from contraction to expansion; however, the contracting phase strictly violates the definition of \cite{Liddle:2000cg}. Although strict definitions of a \emph{spacetime} varies subtly between authors, the canonical literature typically demands that the Lorentzian manifold be time oriented, thus it makes sense to talk about $\ddot{f}$ being explicitly positive or negative. Because there can only be a single choice of time orientation, once this time orientation is fixed what constitutes inflation or deflation is fixed. Given the above, we feel that it is myopic to exclude the possible bouncing phenomenology from inflationary models and a rigorous definition for inflation should encompass bouncing cosmologies (in the cosmological bounce the co-moving hubble radius decreases, as it does during inflation).
Given these considerations, we feel that the inflation definition of Defn.~\ref{inflation defn} is overly restrictive. We propose a new definition for an inflationary spacetime:
\begin{definition}
Let $\left( \mathcal{M}, g \right)$ be an $n$ dimensional spacetime which admits a connected open neighborhood $U \subset \mathcal{M}$ which is isometric to $ \left( a,b \right) \times_f V$ as a warped product open submanifold, where $\left( a,b \right)$ is a timelike codimension $n-1$ embedded submanifold, $V$ is a spacelike codimension 1 embedded submanifold and no assumptions are made concerning $f$ except that it is a well defined function between sets. The spacetime $\left( \mathcal{M}, g \right)$ is an \emph{inflationary spacetime} if there exists some $t_0 \in \left( a,b \right)$ such that (assuming that it exists and is well defined), $\ddot{f} \left( t_0 \right) > 0$ where derivatives of $f$ are taken with respect to the timelike $\left( a,b \right)$ coordinate and the sign of $\ddot{f}$ is given with respect to the time orientation of $\mathcal{M}$. \label{my inflation defn}
\end{definition}
In this definition, no assumptions about the continuity of $f$ or its derivatives are made. There is debate in the literature about what precise assumptions functions in metrical theories of gravity must have for well defined curvatures to exist -- see \cite{Hawking1975} Chapter 3.1 for example -- we do not comment on this issue here. Additionally, we do not assume that $f>0$, as in \cite{Bishop1969,Beem1996,ONeill1983}, so as not to exclude the possibility of an initial singularity caused by $f\rightarrow 0$.
There are several technicalities apropos Defn.~\ref{my inflation defn}. The first is a local global relationship of the local inflationary neighborhood $ \left( a,b \right) \times_f V$ to the global spacetime $\left( \mathcal{M}, g \right)$. As written in Defn. \ref{my inflation defn}, the inflationary canonical time axis must be timelike with respect to the global metric and $V$ must be spacelike with respect to this metric as well. This assumes the existence and causal knowledge of $\left( \mathcal{M}, g \right)$ non-locally. If one holds the view that inflation is purely a local phenomenon, Defn. \ref{my inflation defn} is overly restrictive, excluding the possibility of two distant observers undergoing inflation orthogonal to each other, each believing the other's time axes to be spacelike - see Fig.~\ref{fig1.png}. In fact, upon accepting these arguments one could even theorize that having a well defined metric is a local property! If one demands causality -- even exotic causality such as in closed timelike curves or vicious spacetimes\footnote{A spacetime is vicious at a point if the causal future / past encompasses the entire spacetime, namely $\mathcal{J}^\pm \left( p \right)= \mathcal{M}$. A spacetime is totally vicious if it is vicious everywhere -- see \cite{Beem1996,Minguzzi:2006sa}} -- this possibility must be excluded as it is in the definition. However, by examining various inflationary neighborhoods within a single spacetime under Defn.~2, it gives valuable information as to the causal structure, and subsequently the overall metric of the global spacetime, especially as one extrapolates upon a distant inflationary neighborhood which may exceed one's observable universe.
\begin{figure}[ht]
\includegraphics[scale=.37]{fig1.png}
\caption{An example spacetime which does not conform to Defn. \ref{my inflation defn}. The external rectangle is schematic for overall spacetime $\left( \mathcal{M}, g \right)$ which contains two potential inflationary neighborhoods $U,V$. Neighborhood $V$ has a canonical time axis which is timelike with respect to overall metric $g$ and is indeed inflationary with respect to the above definition. However, neighborhood $U$ has a canonical ``time'' axis which, in its native immersed Lorentzian (causal) structure is taken to be timelike; however, it is spacelike with respect to overall metric $g$. Such a situation is excluded by Defn. \ref{my inflation defn}. Light cones are shown with respect to overall metric $g$.}
\label{fig1.png}
\end{figure}
We now discuss various subtleties concerning the immersion versus the embedding of the inflationary neighborhood. Should a spacetime like that of Fig.~\ref{fig2.png} be licit under the definition of inflation? As was mentioned in the previous paragraph, the reader's opinion of local versus global concerns drives possible definitions of inflation. If one believes inflation to be observer dependent, an immersed submanifold definition is superior and the spacetime of Fig.~\ref{fig2.png} is allowed because what occurs outside of the observer's observable universe is absurd. This has the drawback that every observer could potentially have a different immersed topology. However, if one believes that the ``top down'' structure of the entire spacetime is paramount, the embedded submanifold definition -- this is what is assumed in Defn.~\ref{my inflation defn} -- and all possible observers must be compatible with the entire spacetime's topology, and Fig.~\ref{fig2.png} is excluded.
\begin{figure}[tbhp]
\includegraphics[scale=.37]{fig2.png}
\caption{An example spacetime $\left( \mathcal{M}, g \right)$ conformal to $\mathbb{R}^1_1 \times S^1 \times S^2$. The vertical arrowed lines are identified
along the perimeter and glued aligning the arrows, as in a standard quotient topology diagram. The cylinder admits a single inflationary neighborhood $U$ with canonical time axis wrapped around the cylinder. In the topology of the overall space one considers a neighborhood $V$ - circle neighborhood in the center, but the intersection with the inflationary subspace topology yields a disconnected subset - one connected component shaded dark. In the (embedded) overall topology all points in $V$ are ``close'', however in the (immersed) subspace topology the connected components are ``far away''. Defn. \ref{my inflation defn} excludes this possibility.}
\label{fig2.png}
\end{figure}
What about the maximality of inflationary neighborhoods? As discussed in the introduction, it might be possible to find an extension of an inflationary neighborhood enlarging it. Thus, the length of a geodesic to the boundary of an inflationary neighborhood can vary depending upon the arbitrary selection of neighborhood boundary. However, if one chooses to use the \emph{maximal inflationary neighborhood} and a geodesic intersects the boundary in finite length then the space is incomplete. Defn. \ref{my inflation defn} does not comment on this phenomena, and it is possible one could be interested in inflationary phenomena in a particular coordinate system without wanting maximality. However, if one desires to compute geodesic completeness in an inflationary neighborhood, it must be maximal.
Does an inflationary neighborhood need to be connected? The explicit use of the term ``connected'' in Defn. \ref{my inflation defn} is equivalent to the connectedness of the spatial section $V$ in the Cartesian product construction of $\left( a , b \right) \times_f V$. In order for the definition to correlate with observation it is necessary that the inflationary neighborhood be connected: how would one observe a distinct\footnote{The use of the term ``distinct'' implies each spatial foliation is disjoint from every other one.} distant connected component? However, in some theories - in particular those which admit pocket universes - it is possible that inflation is occurring in multiple distinct distant places at once, and as long as the geometry of the inflating regions is uniformly scaled by the same $f$ the inflationary patch could be disconnected (topologically $\left( a , b \right) \times_f \coprod_{i \in \mathcal{I}} V_i$ for some index set $\mathcal{I}$). One could consider the union of all inflating regions - even those with differing scale functions - as a union: this could be useful in calculations of inflating region volumes \cite{Borde2003,Borde:1993xh,Mithani:2012ii}. All of the aforementioned concerns are built on the edifice of Defn. \ref{my inflation defn}; the definition itself, however, only defines the geometry of a single connected inflating region akin to our observable universe.
Finally one can consider the dimension of the spacelike foliation $V$. In standard metric theories of gravity, this must be a codimension 1 submanifold, akin to a hypersurface. If, however, one would like to accommodate a theory with extra dimensions, it is conceivable that other codimensions might be considered, perhaps for various D-branes populating the bulk. This is not accommodated by Defn. \ref{my inflation defn} in order to agree with FRW historical convention. It would not be difficult to alter the definition to include this phenomenology.
Defn.~\ref{my inflation defn} broadly expands the phenomenology which is considered inflationary. In the absence of a unique model of the nascent universe supported by experimental evidence, we choose to be overly accommodating with respect to the inflationary paradigm, which is consistent with known observations. Additionally, there is precedent with defining a categorization of a spacetime with respect to a quantity at a single point -- consider the general condition of \cite{Hawking1975}. One may also discuss how much inflation a spacetime is experiencing by using statements like it is ``completely inflationary'' where it inflates for all time, or only inflates on $\left( a,b \right)$, to accommodate realistic models and also bouncing models.
We adopt Defn.~\ref{my inflation defn} as a working definition for inflation, attempting to address some of the topological machinery enthymeme in Defn.~\ref{inflation defn}. Naturally, depending on one's bias apropos phenomenology, there are other potential definitions for inflation which might be superior. For example, the topological criteria for an inflationary neighborhood to exist -- namely the existence of a curl free local Killing vector field -- might uniquely specify a definition upon further investigation and proof. We leave such exploration for future work.
\section{Discussion of $H_{avg}$}
\label{sec:h_avg disc}
We now compare our proposed definition for an inflationary spacetime, Defn. \ref{my inflation defn} with a discussion of $H_{avg}$ in \cite{Borde2003}. There, the authors propose a gedanken experiment of measuring the tidal forces in a
general spacetime $\left(\mathcal{M},g\right)$ along a geodesic $\gamma$ with velocity
$V^{\mu}=\frac{\mathrm{d}\gamma}{\mathrm{d}\lambda}$ by subsequent
measurements of a massive test particle moving along some timelike
geodesic $\alpha\,:\,J\rightarrow\mathcal{M}$ with velocity
$U^{\mu}=\frac{\mathrm{d}\alpha}{\mathrm{d}\lambda}$ in a locally Minkowski neighborhood.~~The quantity
\begin{equation}
H^{\gamma} = -\frac{V_{\mu}\left(\lambda\right)}{\beta^{2}\left(\lambda\right)+\kappa}\frac{\mathrm{D}U^{\mu}\left(\lambda\right)}{\mathrm{D}\lambda}\label{eq:H BGV}
\end{equation}
along geodesic
$\gamma$, where a spatially dependent generalized Lorentz factor
is given by $\beta^{2}\left(\lambda\right)=V_{\mu}U^{\mu}$, and $\kappa=V_{\mu}V^{\mu}$
is taken to be $\kappa=-1$ for $\gamma$ being a timelike geodesic, and $\kappa=0$
for $\gamma$ being a null geodesic.\footnote{We use metric signature convention $\left\{ -1,+1,\cdots,+1\right\} $ opposed to
$\left\{ +1,-1,\cdots,-1\right\} $ used in \cite{Borde2003}.}~~Additionally,
we denote the parallel transport of $U^{\mu}$ along $\gamma$ as
$\frac{\mathrm{D}U^{\mu}}{\mathrm{D}\lambda}$. The above may be written succinctly in geometers notation as
\begin{equation}
H^{\gamma} = - \frac{g \left( V , \nabla_{V} U \right)}{g \left( U , V \right) + \kappa} \label{eq:H BGV geometer}
\end{equation}
where $g$ is the metric and $\nabla$ is the Levi-Civita connection.
Note that Eq.~\ref{eq:H BGV}
reduces to the usual definition of the Hubble parameter
\begin{equation}
H\left(t\right)=\frac{\dot{f}\left(t\right)}{f\left(t\right)}\label{eq:H FRW}
\end{equation}
in the special case that the spacetime is inflationary, and in the inflationary neighborhood, $V^\mu$ is purely spacelike and $U^\mu$ is purely timelike\footnote{This calculation makes use of Corollary 12.8 in \cite{ONeill1983}.}. In fact, in a gedanken experiment where timelike observer $V = a \frac{\partial}{\partial t} + A$ measures a timelike test particle $U = b \frac{\partial}{\partial t} + B$ with both $A , B$ purely spacelike, we find
\begin{equation}
H^\gamma = - \frac{b \frac{\dot{f}}{f} \lVert A \rVert ^2_g - a^2 \frac{\partial b}{\partial t} + g \left( A , d \pi_\Sigma \nabla_A B \right) }{g \left( A , B \right) - ab - 1} \label{havg calc full}
\end{equation}
where $d \pi_\Sigma \nabla_A B $ is the projection of the $\Sigma$-tangential components of $\nabla_A B$ onto spacelike slice $\Sigma$ and $\lVert A \rVert^2_g = g \left( A , A \right)$. By utilizing the fact that in warped product constructions $\xi^2_A \equiv f^2 \lVert A \rVert^2_g$ is a constant -- which corresponds to the initial speed of a test particle with respect to cosmic time, not a N{\"o}ether conserved quantity -- Eq. \ref{havg calc full} can be expressed as
\begin{widetext}
\begin{equation}
H^\gamma = \frac{\dot{f}}{f} \left( \frac{\sqrt{1+\frac{\xi^2_B}{f^2}}\left( \frac{\xi^2_A}{f^2} \right) + \left( 1 + \frac{\xi^2_A}{f^2} \right) \frac{\xi^2_B}{f^2} \frac{1}{\sqrt{1+\frac{\xi^2_B}{f^2}}} }{1 + \sqrt{1+\frac{\xi^2_A}{f^2}}\sqrt{1+\frac{\xi^2_B}{f^2}} - g \left( A , B \right)} \right) + \frac{g \left( A , d \pi_\Sigma \nabla_A B \right)}{1 + \sqrt{1+\frac{\xi^2_A}{f^2}}\sqrt{1+\frac{\xi^2_B}{f^2}} - g \left( A , B \right)} \label{havg calc final}
\end{equation}
\end{widetext}
which differs from the expected Hubble parameter $\frac{\dot{f}}{f}$. If we assume spatial curvature is negligible - namely $d \pi_\Sigma \nabla_A B \sim 0$ and the directions of spatial propagation between observer and test particle are orthogonal, the above equation simplifies to
\begin{equation}
H^\gamma \sim \frac{\dot{f}}{f} \left( \frac{\sqrt{1+\frac{\xi^2_B}{f^2}}\left( \frac{\xi^2_A}{f^2} \right) + \left( 1 + \frac{\xi^2_A}{f^2} \right) \frac{\xi^2_B}{f^2} \frac{1}{\sqrt{1+\frac{\xi^2_B}{f^2}}} }{1 + \sqrt{1+\frac{\xi^2_A}{f^2}}\sqrt{1+\frac{\xi^2_B}{f^2}}} \right) \label{havg calc final slow}
\end{equation}
which is schematic to
\begin{equation}
H^\gamma \sim \frac{\dot{f}}{f} \times \mathcal{O} \left( \xi^2_A , \xi^2_B \right) \label{havg calc final schematic}
\end{equation}
Hence, we arrive at the undesirable result, that $H^\gamma$ is irreducibly convoluted with the particular speeds $\xi^2_A, \xi^2_B$ of the test particles in question. It would be ideal to find an expression which yields $\frac{\dot{f}}{f}$ independent of the particular speeds of the particle, which Eq. \ref{eq:H BGV} is not. However, in the case that both particles are relativistic with respect to cosmic time one recovers $H^\gamma \sim \frac{2 \sqrt{2}}{3} \frac{\dot{f}}{f}$, but for non-relativistic particles this pre-factor varies by choice of test particle.
It would be ideal to calculate the Hubble parameter $\frac{\dot{f}}{f}$ in an inflationary spacetime independent of frame and particular velocity. One proposal is the quantity
\begin{equation}
H^\gamma = \frac{g \left( \frac{\partial}{\partial t} , \mathrm{II} \left( X , Y \right) \right)}{g \left( X , Y \right)}
\label{H II}
\end{equation}
where $X,Y$ are purely spacelike (horizontal) vector fields and $\mathrm{II} \left( X , Y \right) = \mathrm{nor} \nabla_X Y$ is the second fundamental form yielding extrinsic curvature. In a GFRW
\begin{equation}
\mathrm{II} \left( X , Y \right) = - \frac{\dot{f}}{f} g \left( X , Y \right) \frac{\partial}{\partial t}
\label{II GFRW}
\end{equation}
thus, for any $X,Y$ as stated above, one reaps $H^\gamma = \frac{\dot{f}}{f}$. There are of course drawbacks to this definition: ambiguity of precisely knowing $\frac{\partial}{\partial t}$ or that $\mathrm{II}$ is an abstract geometric quantity and $X,Y$ are spacelike; neither of which can be directly measured.
We now state what we consider to be the main result of
\cite{Borde2003}:
\begin{thm} \label{thm:bgv}
\emph{\noun{Inflationary spacetimes are past
incomplete}} - Let $\left(\mathcal{M},g\right)$ be a spacetime, and
$\gamma\,:\,\left[\lambda_{i},\lambda_{f}\right]\rightarrow\mathcal{M}$
be some causal geodesic.~~If one computes the quantity
\begin{equation}
H_{avg}^{\gamma}=\frac{1}{\lambda_{f}-\lambda_{i}}\int_{\lambda_{i}}^{\lambda_{f}}H^{\gamma}\left(\alpha\right)d\alpha\label{eq:havg bgv}
\end{equation}
to be strictly positive along the image of $\gamma$, the spacetime
is geodesically incomplete.
\end{thm}
First we study the implications of this theorem, and
discuss possible implications of its results. In particular, we will express an uncountably infinite number of classical monsters with respect to Thm.~\ref{thm:bgv} which are geodesically complete despite having $H_{avg} > 0$. Here we consider only metrical theories of gravity;
and while cursory computations of the Einstein Field Equations (EFE)
will be discussed within the context of energy conditions, all proofs that follow
require only assumptions at the \emph{geometric} level.~~Our aforementoned GFRW spacetime counter-examples exist
within the scope of general relativity, with only an arbitrarily small
necessity for energy-condition-violating exotic matter.~~Any further speculations on relationships
between geometrical quantities and mass-energy is beyond the scope of this paper.
Our first criticism of Eq.~\ref{eq:havg bgv}, involves the implicit assumption that the integral
is computed over a \it compact \rm interval. In this way,
the integral is predestined to converge, assuming the minimum condition that
$H^{\gamma}\left(\lambda\right)$ is never infinite on a set of non-zero
(Borel) measure.~~Canonically,
properties of geodesic completeness are formulated with respect
to \it maximal \rm geodesic rays of domain $\left[0,b\right)$. Thus, by defining geodesics in this manner it is difficult to compare Thm. \ref{thm:bgv} to the standard mathematical literature.~~In particular,
if one defines $H^{\gamma}$ by Eq.~\ref{eq:havg bgv},
the suppression by a finite $\frac{1}{\lambda_{f}-\lambda_{i}}$ makes it
much more difficult for $H_{avg}\rightarrow0$ in the limit, yielding
false positives for geodesic incompleteness. This is arguably equivalent to an a priori assumption of incompleteness in many cases.~~Considering
this, it is not surprising the aforementioned use of $H_{avg}$ finds so many spacetimes to be incomplete.
Another interesting subtext of
of Thm.~\ref{thm:bgv} is the dependence of $H_{avg}^{\gamma}$ on a
particular choice of geodesic $\gamma$.~~In
particular, the geodesic which one selects might not happen
to yield a $H_{avg}^{\gamma}>0$: in order to properly classify geodesic completeness
of $\mathcal{M}$ one must examine $\sup_{\gamma\in\mathrm{MGeo}\left(\mathcal{M}\right)}H_{avg}^{\gamma}$,
where $\mathrm{MGeo}\left(\mathcal{M}\right)$ is the space of maximal
geodesics over $\mathcal{M}$. Utilizing methods of analysis and the fact that the space of complete maximal geodesics is closed in the $\mathcal{C}^0$ Whitney topology -- see Theorem 4.7 in \cite{Sanchez1998} -- one should be able to calculate the range of $H^\gamma_{avg}$ when viewed as a functional on the space of geodesics over a (fixed) spacetime, which we leave to future work. Finally, the fact that both the observer and test particle are assumed to be geodesics has not been used in the derivation of $H^\gamma$.
To accommodate the above issues, we propose
Eq.~\ref{eq:havg bgv} should be modified to
\begin{equation}
H_{avg}=\lim_{a,b\rightarrow\infty}\frac{1}{b-a}\int_{a}^{ b}H^{\gamma}\left(\alpha\right)d\alpha\label{eq:Havg EL}
\end{equation}
This limit naturally incorporates the fact that, if
a geodesic $\gamma$ is complete it will have domain of all $\mathbb{R}$. From
this point on we will take Eq.~\ref{eq:Havg EL} to be the definition
of $H_{avg}$.
\section{Monsters}
\label{sec:counter-examples}
We now present an uncountably infinite class of classical monsters which have $H_{avg} > 0$ but are geodesically complete. We would like to remind the reader that this a purely geometric result: it applies to any metrical theory of gravity. For any GFRW $\mathbb{R}^1_1 \times_f \Sigma$ with a smooth scale function $f$ -- thus guaranteeing there will not be any curvature singularities -- the geodesic completeness of GFRWs is enumerated by the following \cite{Sanchez1994,Sanchez1998}:
\begin{thm}{Geodesic Completeness Criterion for GFRW Spacetimes - }
Let $\mathcal{M} = \mathbb{R}^1_1 \times_f \Sigma$ be a GFRW spacetime
\begin{enumerate}
\item The spacetime $\mathcal{M}$ is future timelike complete iff $ \int_{t_0}^\infty \frac{f \left( t \right) dt}{\sqrt{\left( f \left( t \right) \right)^2 + 1}}$ diverges for all $t_0 \in \mathbb{R}$.
\item The spacetime $\mathcal{M}$ is future null complete iff $ \int_{t_0}^\infty f \left( t \right) dt$ diverges for all $t_0 \in \mathbb{R}$.
\item The spacetime $\mathcal{M}$ is future spacelike complete iff $\mathcal{M}$ is future null complete or if $\int_{t_0}^\infty f \left( t \right) dt < \infty$ then $f$ is unbounded; for all $t_0 \in \mathbb{R}$.
\item The GFRW is past timelike / null / spacelike complete if, for items 1-3 above, upon reversing the limits of integration from $\int_{t_0}^\infty$ to $\int_{-\infty}^{t_0}$ the word ``future'' is replaced by ``past''.
\item The spacetime $\mathcal{M}$ is \emph{geodesically complete} iff it is both future and past timelike, null, and spacelike geodesically complete.
\end{enumerate}
\label{sanchezthm}
\end{thm}
Thus, one can definitively calculate the geodesic completeness of a given GFRW with Thm.~\ref{sanchezthm} and compare the predictions to Thm. \ref{thm:bgv}. In particular, if the condition that
\begin{equation}
\inf_\mathbb{R} f = a > 0
\label{f inf}
\end{equation}
holds then the GFRW spacetime is geodesically complete. One can imagine a class of GFRW spacetimes with a smooth scale function obeying $\inf_\mathbb{R} f = a > 0$, however for $t>t_0$ with respect to some given $t_0 \in \mathbb{R}$ behaves as de Sitter space with $f = e^{Ht}$. Such scale functions can be constructed with smooth partitions of unity. If one then calculates the desired $H_{avg}$ -- where Eq. \ref{eq:H FRW} is used -- the final calculation, which depends on the de Sitter phase to generate non-zero measure, yields $$H_{avg} \geq H$$ Thus, all that matters to construct this class of counter examples is that the tail of $f$ behaves as de Sitter, and it can be smoothly combined with any other function of positive infimum on the interval $\left( -\infty , t_0 \right)$ with a smooth partition of unity over a compact interval. The size of the function space over this interval enumerates a cohort of classical monsters, see Fig.~\ref{GFRW counter-example}. Additionally, because the dimensionality of $\mathcal{C}^\infty \left( \left( -\infty , t_0 \right) \right)$, the space of smooth functions on the interval $\left( -\infty , t_0 \right)$, is uncountably infinite, there are an uncountably infinite number of GFRW spacetimes which have $H_{avg} > 0$ but are geodesically complete.
\begin{figure*}[ht]
\centering \includegraphics[width=6in]{fig3.png}
\caption{The scale function for a counter-example GFRW $\mathbb{R}^1_1 \times_f \Sigma$ is shown; here $f$
loiters in a Minkowski phase for $t < -1$ and enters a de Sitter phase for $t \geq 0$. The transition zone, for $t \in \left[ -1 , 0 \right]$, consists of both phases being convoluted with a smooth bump function. By careful construction the spacetime model is
free of curvature singularities commensurate with the smoothness of $f$ and the completeness of $\left( \Sigma , g_\Sigma \right)$; the GFRW
is geodesically complete as per Thm.~\ref{sanchezthm}, despite having $H_{avg} = \frac{1}{2}$. On the right, the shaded region where $\dot{H} \geq 0$, indicates the NEC (and WEC) is violated.
\label{GFRW counter-example}
}
\end{figure*}
The question now emerges, are such classical monsters physically reasonable? In \cite{Mithani:2011en} the authors discuss one element in the above cohort, namely $f\left( t \right) = a_0 \left( 1 + e^{H_0 t} \right)$ and argue that such a scale function possesses non-zero quantum mechanical tunneling probability when treated with the Wheeler-DeWitt equation, thus eliminating the viability of an eternally loitering phase prior to inflation and expansion. In particular, they argue that the class of aforementioned classical monsters (belonging to the ``emergent universe'' scenario \cite{Ellis:2002we,Ellis:2003qz,Mulryne:2005ef,delCampo:2011mq,Graham:2011nb}) should be excluded from the domain of discourse of physically reasonable scale factors (see also \cite{Aguirre:2013kea}). In fact, the implications of cosmologies which infinitely loiter in a steady state generates the prime antinomy of \emph{temporal Copernicanism}: why do we exist NOW even though it is infinitely likely we should exist at some other time? For the current discussion, however, we have little to contribute.
We do feel it is important to point out that calculational use of the Wheeler-DeWitt equation, predicated on the Arnowitt Deser and Misner (ADM) formalism -- see \cite{DeWitt:1967yk} and \cite{Arnowitt:1959ah} -- has somewhat of a contentious past, see \emph{e.\!\!~g.}~ \cite{Kiriushcheva2011,Landsman:1995gh}. In particular, resolving the factor ordering problem eliminates the consistent canonical quantization of most general Hamiltonians in curved spacetimes~\cite{Dirac1964}, a detail which has simply been assumed to have been resolved in \cite{DeWitt:1967yk}. With respect to the current problem at hand, one criticism of the calculation of \cite{Mithani:2011en} is the semi-classical treatment and subsequent elimination of time-ordering terms from the potential; a well known point of controversy apropos the Wheeler-DeWitt equation. For future work, it would be interesting to repeat the calculation including these terms, possibly utilizing methods contained in \cite{Steigl2006}. Many semi-classical physical quandaries are resolved when treated in a complete quantum formalism.
Besides the above concern, it is not possible to have such a loitering universe exit the loitering phase or a bounce into an expanding phase in GR while obeying the weak energy condition (WEC): $T_{\mu\nu} t^\mu t^\nu \geq 0$, for timelike vectors $t^\sigma$. As shown in Fig.~\ref{GFRW counter-example} the WEC is violated very briefly in the transition between the two phases during $t \in \left[ -1 , 0 \right]$. Indeed, even the null energy condition (NEC) is violated in this region. However, this transition zone can be made arbitrarily small, perhaps even smaller than a Planck time, so the need for any such exotic matter such as galileons \cite{Nicolis:2008in} which can stably support such a NEC violation \cite{Creminelli:2006xe}, or any other new physics such as strings to accommodate a loitering phase \cite{Alexander:2000xv,Brandenberger:2001kj}, is fleeting. Of course, even the standard inflationary scenario necessarily violates the strong energy condition (SEC); indeed, our current period of accelerated expansion indicates the SEC is being violated even today, and it is well known that all of the energy conditions can be violated at the quantum level.
Because of their ability to violate the NEC, galileon fields have been used to construct monsters of the type discussed above, including loitering, inflationary and bouncing cosmologies \cite{Creminelli:2006xe,Creminelli:2010ba,Kobayashi:2010cm,Qiu:2011cy,Easson:2011zy,Cai:2012va} and other exotic cosmological solutions \cite{Bains:2015gpv,Easson:2016klq,Easson:2018qgr}.
Stability arguments in such theories are complicated and depend on external background matter \cite{Easson:2013bda}. Even properly defining energy conditions in non-canonical scalar-tensor theories is non-trivial~\cite{Chatterjee:2012zh}. The devil is in the details, and it is unclear how similar quantum stability arguments of \cite{Mithani:2011en} apply to these models.
\section{Conclusion}
\label{sec:conclusion}
For a mathematical model to aspire to be a plausible description of the actual universe (or multiverse), additional criteria are often brought to bear. Such criteria may depend on the theoretical framework adopted in the treatment. Within classical theories of gravity, a frequently invoked question is whether a proposed solution is ‘generic’. More precisely, if the solution involves either a parameter value or initial conditions that must be chosen from a set of measure zero, the model is often dismissed as unphysical. A simple example is in Newtonian cosmology, where an initial density singularity will exist if the distribution of matter is exactly spherically symmetric. However, if the coefficients of the higher multipoles are not all strictly 0, infinite densities may not occur. In general relativity, however, departures from sphericity may not serve to remove the singularity if certain energy conditions apply, but it may be that most matter will ‘miss’ the singularity and the corresponding world lines will be geodesically complete in the past. In this paper, we have presented a set of geodesically complete solutions of inflationary metric theories of gravity that belong to an uncountable continuum, which is therefore a set of non-zero measure in the space of (continuous) initial conditions. As such, these ‘monsters’ are generic and thus are stable to perturbations in initial conditions, which are impossible to know precisely. We note that the measure of the set of initial conditions being described here is predicated on the assumption of the continuity of spacetime, which may be transcended in the context of quantum cosmology.
Another criterion that might be used as a filter for physically plausible theories is the second law of thermodynamics, generalized to include cosmological event horizons. Such horizons are a feature of inflationary cosmological models. In the case of de Sitter space, the event horizon area is constant, but in FRW models where $f$ is concave upward as $t$ tends to infinity, the event horizon area will shrink and will generally imply a violation of the generalized second law of thermodynamics. Such models might then be regarded as unphysical. There may, of course be additional criteria beyond a solution being generic and not violating the generalized second law. The point we wish to make is that the plausibility or otherwise of a mathematical model to describe the real universe goes beyond it merely being a correct solution to a set of accepted dynamical equations.
In this paper, we offered a critical discussion of the $H_{avg} > 0$ geodesic completeness criterion for inflationary spacetimes with respect to the arguments presented in \cite{Borde2003}. Our first area of discussion involved the definition of inflationary spacetimes. In the voluminous body of literature on the subject, what is rigorously meant by an \emph{inflationary spacetime} varies between authors. By introducing Defn.~\ref{my inflation defn}, we have suggested a standardized definition for inflation which underpins the geometrical requirements while broadly encompassing physically reasonable phenomenology. We have further introduced $H^\gamma$ and $H^\gamma_{avg}$ and discussed both advantages and disadvantages to this definition, culminating with the generalization of Eq.~\ref{eq:Havg EL}. Drawing on this equation, we discussed an uncountably infinite cohort of classical monsters which
are geodesically complete despite having $H_{avg} > 0$, in contradiction to the arguments found in~\cite{Borde2003}. These results are purely geometric and apply for any metrical theory of gravity.
The solutions presented necessarily violated traditional energy conditions within the context of pure GR; although it is possible, such solutions may exist within non-canonical scalar field theories in a stable way as discussed above. Again within the context of GR the solutions may suffer from quantum considerations, as discussed in \cite{Mithani:2011en}, leading to the possible exclusion of this class of classical monsters from the domain of discourse for physically reasonable scale factors; however, we feel a deeper understanding of quantum gravity is needed to definitively make such a statement.
For the time being, the question of the viability of eternal inflation, and the controversy of $H_{avg}$ is far from settled, leaving many possibilities for future work.
\section*{Acknowledgements}
We are delighted to thank George Ellis, Brett Kotschwar, Roger Penrose, and Tuna Yidrim for helpful discussions.
DE is supported in part by the U.S. Department of Energy, Office of High Energy Physics, under Award No.~DE-SC0019470, and the Foundational Questions Institute under Grant No.~FQXi-MGB-1927.
\newpage
|
\section{INTRODUCTION}
A basic postulate of Maxwell's electromagnetism as well as Einstein's theory of special relativity is the constant speed $c$,
in vacuum, of all electromagnetic radiation, which implies that the photon should be massless. Determining the rest mass of
the photon has therefore been one of the most enduring efforts on testing the validity of this postulate. However, it is
unfeasible to prove experimentally that the photon rest mass is exactly zero. According to the uncertainty principle,
the ultimate upper limit on the photon rest mass would be $m_\gamma \leq \hbar / \Delta t c^2 \approx {10}^{-69}$ kg, using
the age of the universe of about $10^{10}$ years \citep{1971RvMP...43..277G,2005RPPh...68...77T}. The optimal experimental
strategy is therefore to set ever tighter upper bounds on $m_\gamma$ and push the results more closely towards the ultimate
bounds of measurement uncertainty.
From a theoretical perspective, a nonzero photon rest mass can be accommodated in a unique way by changing the inhomogeneous
Maxwell's equations to the Proca equations. Using them, it is possible to consider some far-reaching implications associated
with massive photons, such as variations of the speed of light with frequency, deviations in the behavior of static
electromagnetic fields, the existence of longitudinal electromagnetic waves, and so on. All of these effects have been
employed to set upper limits on the photon rest mass via various terrestrial and extra-terrestrial approaches
\citep{1971RvMP...43..277G,1973PhRvD...8.2349L,1997sref.book.....Z,2005RPPh...68...77T,2006AcPPB..37..565O,2010RvMP...82..939G,
2011EPJD...61..531S}. Over the past several decades, the common approaches for determining the photon mass include
measurement of the frequency dependence in the speed of light \citep{1964Natur.202..377L,1969Natur.222..157W,
PhysRevLett.82.4964,Wu_2016,2016PhLB..757..548B,2017PhLB..768..326B,ZHANG201620,PhysRevD.95.123010,Wei_2017,
2018JCAP...07..045W,2020RAA....20..206W,2021FrPhy..1644300W,Xing_2019,2021PhLB..82036596W}, null tests of Coulomb's
inverse square law \citep{PhysRevLett.26.721}, tests of Amp\`ere's law \citep{PhysRevLett.68.3383}, torsion balance
\citep{PhysRevLett.80.1826,2003PhRvL..90h1801L}, Jupiter magnetic field \citep{PhysRevLett.35.1402}, magnetohydrodynamic
phenomena of the solar wind \citep{1997PPCF...39...73R,2007PPCF...49..429R,2016APh....82...49R}, cosmic magnetic fields
\citep{1959PThPS..11....1Y,1976UsFiN.119..551C,2007PhRvL..98a0402A}, gravitational deflection of massive photons
\citep{PhysRevD.8.2349,PhysRevD.69.107501}, suppermassive black hole spin \citep{PhysRevLett.109.131102}, pulsar spindown
\citep{Yang_2017}, and so on. Among these approaches, the most direct one is to measure a possible frequency dependence
in the velocity of light.
According to Einstein's special relativity, the energy of the photon with a nonzero rest mass $m_\gamma$ can be written as
\begin{equation}
E=h\nu=\sqrt{p^2c^2+m^2_\gamma c^4}\;.
\end{equation}
The relation between the massive photon group velocity $\upsilon$ and the frequency
$\nu$ then takes the form:
\begin{equation}\label{velocity}
\upsilon=\frac{\partial E}{\partial p}=c\sqrt{1-\frac{m^2_\gamma c^4}{E^2}}=c\sqrt{1-A\nu^{-2}}\;,
\end{equation}
where $A=m^2_\gamma c^4/h^2$. It is evidently clear from Equation (\ref{velocity}) that the lower frequency, the slower
the massive photon propagates in vacuum. The photon mass can therefore be constrained by comparing the arrival-time differences
of photons with different frequencies originating from the same source. Moreover, it is easy to understand that measurements
of shorter arrival-time differences between lower energy bands from sources at longer propagation distances are particularly
sensitive to the photon mass. Thanks to their fine time structures, radio emissions, and cosmological distances,
fast radio bursts (FRBs) provide the current best celestial laboratory for constraining the photon mass
\citep{Wu_2016,2016PhLB..757..548B,2017PhLB..768..326B,PhysRevD.95.123010,Xing_2019,2020RAA....20..206W,2021FrPhy..1644300W,
2021PhLB..82036596W}. The first attempts to place upper limits on the photon mass using FRBs were presented in \cite{Wu_2016}
and \cite{2016PhLB..757..548B}. Adopting the controversial redshift $z=0.492$ for FRB 150418, \cite{Wu_2016} set a strict
upper limit of $m_\gamma \leq 5.2 \times {10}^{-50}$ kg (see also \citealt{2016PhLB..757..548B}). Subsequently,
\cite{2017PhLB..768..326B} used the reliable redshift measurement of FRB 121102 to derive a robust limit of
$m_\gamma \leq 3.9 \times {10}^{-50}$ kg. \cite{PhysRevD.95.123010} developed a Bayesian framework to constrain the photon mass
with a catalog of 21 FRBs, yielding $m_\gamma \leq 8.7 \times {10}^{-51}$ kg at $68\%$ confidence level. By analyzing
the time-frequency structure of subbursts in FRB 121102, \cite{Xing_2019} obtained a tighter limit on the photon mass of
$m_\gamma \leq 5.1 \times {10}^{-51}$ kg. Since the plasma and photon mass contributions to the dispersion measure (DM) have
different redshift dependences, \cite{2020RAA....20..206W} proved that they can be distinguished by measurements of nine FRB
redshifts, enabling the sensitivity to $m_\gamma$ to be improved (i.e., $m_\gamma \leq 7.1 \times {10}^{-51}$ kg at $68\%$
confidence level). \cite{2021PhLB..82036596W} used a catalog of 129 FRBs in a Bayesian framework to derive a combined limit
of $m_\gamma \leq 3.1 \times {10}^{-51}$ kg at $68\%$ confidence level.
Although photon mass limits obtained through the dispersion method have reached high precision, all current investigations
considered the first-order Taylor expansion of the dispersion relation only
(i.e., $\Delta t_{m_{\gamma},1}\propto {m^{2}_\gamma}/{\nu^2}$).
If we keep the second-order derivative of Taylor series, the frequency-dependent dispersion due to a nonzero photon
mass would be $\Delta t_{m_{\gamma}}=\Delta t_{m_{\gamma},1}+\Delta t_{m_{\gamma},2}\propto a_{1}{m^{2}_\gamma}/{\nu^2}+a_{2}{m^{4}_\gamma}/{\nu^4}$.
From observations, all radio emission signatures show an indisputable $\nu^{-2}$-dependent time delay,
$\Delta t_{\rm obs,1}\propto {\rm DM}\cdot\nu^{-2}$,
which is in good agreement with the first-order Taylor expansion of the dispersions expected from both the plasma
($\Delta t_{\rm DM,1}$) and nonzero photon mass ($\Delta t_{m_{\gamma},1}$, if it exists) effects. That is,
$\Delta t_{\rm obs,1}=\Delta t_{\rm DM,1}+\Delta t_{m_{\gamma},1}$. Assuming that the dedispersed time delay
($\Delta t_{\rm obs}-\Delta t_{\rm obs,1}$) is attributed solely to the second-order term arising from the nonzero
photon mass effect ($\Delta t_{m_{\gamma},2}$), we can therefore obtain a new upper limit on the photon mass
$m_{\gamma,2}$.
With the dedispersed time delays from radio observations in hand, it is interesting to investigate what level of
$m_{\gamma,2}$ limits can be obtained by taking into account the second-order Taylor expansion of the dispersion due to
a nonzero photon mass. In this work, we make use of the dedispersed pulses from the Crab pulsar \citep{Hankins_2007} and
FRB 180916B \citep{2021ApJ...911L...3P}, for the first time, to study the possible second-order photon mass effect.
Throughout this paper a flat $\Lambda$CDM cosmological model with $H_{0}=67.36$ km $\rm s^{-1}$ $\rm Mpc^{-1}$,
$\Omega_m=0.315$, and $\Omega_{\Lambda}=0.685$ is adopted \citep{2020A&A...641A...6P}.
\section{Theoretical Framework}
\subsection{Dispersion from a Nonzero Photon Mass}
With Equation (\ref{velocity}), it is straightforward to show that two massive photons emitted simultaneously from
a same source would reach us at different times if they have different frequencies. For a cosmic source at redshift $z$,
the rest-frame time delay ($\Delta{t_{m_{\gamma},z}}$) between two massive photons with different rest-frame frequencies
($\nu_{l,z}<\nu_{h,z}$) can be obtained by using second-order approximation of Taylor expansion:
\begin{equation}\label{eq:tmrz}
\begin{aligned}
\Delta{t_{m_{\gamma},z}}&=\int \frac{{\rm d}l}{c}\left[\left(1-A\nu_{l,z}^{-2}\right)^{-1/2}-\left(1-A\nu_{h,z}^{-2}\right)^{-1/2}\right]\\
&\simeq\int \frac{{\rm d}l}{c} \left[\frac{1}{2}A\left(\nu_{l,z}^{-2}-\nu_{h,z}^{-2}\right)+\frac{3}{8}A^{2}\left(\nu_{l,z}^{-4}-\nu_{h,z}^{-4}\right)\right]\;.
\end{aligned}
\end{equation}
In the observer frame, the observed time delay is $\Delta{t_{m_{\gamma}}}=\Delta{t_{m_{\gamma},z}}\times(1+z)$ and
the observed frequency is $\nu=\nu_{z}/(1+z)$. Noticing
\begin{equation}
{\rm d}l=\frac{1}{1+z}\frac{c}{H_0}\frac{{\rm d}z}{\sqrt{\Omega_{m}(1+z)^3+\Omega_{\Lambda}}}
\end{equation}
for a flat $\Lambda$CDM cosmological model, so Equation~(\ref{eq:tmrz}) can be modified as
\begin{equation}
\begin{aligned}
\Delta t_{m_\gamma}&=\frac{A}{2H_0}\left(\nu_l^{-2}-\nu_h^{-2}\right)H_1(z)+\frac{3A^2}{8H_0}\left(\nu_l^{-4}-\nu_h^{-4}\right)H_2(z)\\
&=\Delta t_{m_{\gamma},1}+\Delta t_{m_{\gamma},2}\;,
\end{aligned}
\end{equation}
where
\begin{equation}
H_1(z)=\int_0^z\frac{(1+z')^{-2}dz'}{\sqrt{\Omega_m(1+z')^3+\Omega_\Lambda}}
\end{equation}
and
\begin{equation}
H_2(z)=\int_0^z\frac{(1+z')^{-4}dz'}{\sqrt{\Omega_m(1+z')^3+\Omega_\Lambda}}\;.
\end{equation}
For a nearby source at distance $d$, we do not need to take into account the cosmological expansion. The observed time delay
induced by the nonzero photon mass effect can be simplified as
\begin{equation}
\begin{aligned}
\Delta t_{m_\gamma}&=\frac{d}{\upsilon_l}-\frac{d}{\upsilon_h}\\
&\approx \frac{dA}{2c}\left(\nu^{-2}_l-\nu^{-2}_h\right)+\frac{3dA^2}{8c}\left(\nu^{-4}_l-\nu^{-4}_h\right)\\
&=\Delta t_{m_{\gamma},1}+\Delta t_{m_{\gamma},2}\;,
\end{aligned}
\end{equation}
where $\Delta t_{m_{\gamma},1}=\frac{dA}{2c}\left(\nu^{-2}_l-\nu^{-2}_h\right)$ and
$\Delta t_{m_{\gamma},2}=\frac{3dA^2}{8c}\left(\nu^{-4}_l-\nu^{-4}_h\right)$ correspond to
the first- and second-order terms, respectively.
\subsection{Dispersion from the Plasma Effect}
Due to the dispersive nature of plasma, the group velocity of electromagnetic waves propagating
through the ionized median would have a frequency dependence, i.e.,
\begin{equation}
\upsilon_p=c\sqrt{1-\frac{{\nu_p}^2}{\nu^2}}\;,
\end{equation}
where the plasma frequency $\nu_{p}=(n_{e}e^{2}/4\pi^{2} m_{e}\epsilon_{0})^{1/2}$ with
$n_{e}$ the electron number density in the plasma, $m_{e}$ and $e$ the mass and charge of
an electron, respectively, and $\epsilon_{0}$ the permittivity. The arrival of a radio signal
with frequency $\nu$ travelling across the plasma along the line of sight would be delayed by
\citep{1979rpa..book.....R}
\begin{equation}
\tau=\int\frac{{\rm d}l}{c}\left[\left(1-\frac{{\nu_p}^2}{\nu^2}\right)^{-1/2}-1\right]
\end{equation}
with respect to the arrival time had the signal travelled through vacuum.
The delay time between two wave packets with different frequencies, which caused by the plasma effect,
can then be obtained by using second-order approximation of Taylor expansion:
\begin{equation}\label{eq:tDM}
\begin{aligned}
\Delta{t_{\rm DM}}&\simeq\int \frac{{\rm d}l}{c} \left[\frac{1}{2}\nu^{2}_{p}\left(\nu_l^{-2}-\nu_h^{-2}\right)+\frac{3}{8}\nu^{4}_{p}\left(\nu_l^{-4}-\nu_h^{-4}\right)\right]\\
&=\Delta{t_{{\rm DM},1}}+\Delta{t_{{\rm DM},2}}\;,
\end{aligned}
\end{equation}
where $\Delta{t_{{\rm DM},1}}$ and $\Delta{t_{{\rm DM},2}}$ correspond to the first- and second-order
terms, respectively. The first-order term $\Delta{t_{{\rm DM},1}}$ can be further reduced to
\begin{equation}
\Delta{t_{{\rm DM},1}}=\frac{e^{2}}{8\pi^{2} m_{e}\epsilon_{0}c}\left(\nu_l^{-2}-\nu_h^{-2}\right){\rm DM}\;,
\end{equation}
where ${\rm DM}=\int n_{e}{\rm d}l$ is the dispersion measure, which is given in the absence of a photon mass
by the integrated electron column density along the propagation path. In a cosmological setting,
${\rm DM}=\int n_{e,z}(1+z)^{-1}{\rm d}l$, where $n_{e,z}$ is the electron number density in the rest-frame
and $z$ is the redshift \citep{Deng_2014}.
\subsection{The Residual/dedispersed Time Delay}
In our analysis, we suppose that the total time delay of two photons with different frequencies is attributed to
the following components:
\begin{equation}
\label{obs_delay}
\Delta t_{\rm obs} = \Delta t_{\rm int} + \Delta{t_{{\rm DM},1}}+\Delta{t_{{\rm DM},2}} + \Delta t_{m_{\gamma},1}+\Delta t_{m_{\gamma},2}\;,
\end{equation}
where $\Delta t_{\rm int}$ is the intrinsic time delay that depends on the geometry and radiation processes of the source, and
both $\Delta t_{\rm DM}$ and $\Delta t_{m_\gamma}$ include the first- and second-order terms.
In practice, the de-dispersion could remove all the time delay of $\nu^{-2}$ behavior, i.e., the first-order terms in Equation~(\ref{obs_delay}).
Thus, after dedispersed, the residual delay time is
\begin{equation}
\begin{aligned}
\Delta t_{\rm res}&=\Delta t_{\rm obs}-\Delta{t_{{\rm DM},1}}-\Delta t_{m_{\gamma},1}\\
&= \Delta t_{\rm int} +\Delta{t_{{\rm DM},2}} +\Delta t_{m_{\gamma},2}\;,
\end{aligned}
\end{equation}
If we assume that $\Delta t_{\rm int}+\Delta{t_{{\rm DM},2}}\geq 0$, then
$\Delta t_{m_{\gamma},2}=\Delta t_{\rm res}-(\Delta t_{\rm int}+\Delta{t_{{\rm DM},2}})\leq\Delta t_{\rm res}$.
Thus, $\Delta t_{\rm res}$ is the upper limit of $\Delta t_{m_{\gamma},2}$, providing an upper limit on the photon mass.
For a nearby source at distance $d$, the photon mass can then be constrained by the second-order effect as
\begin{equation}
m_{\gamma,2} \leq hc^{-2}\left[\frac{8c\Delta t_{\rm res}}{3d\left(\nu_l^{-4}-\nu_h^{-4}\right)}\right]^{1/4}\;,
\end{equation}
which can be simplified as
\begin{equation}\label{mass}
m_{\gamma,2}\leq\left(0.93\times{10}^{-46}\;{\rm kg}\right)\left\{\frac{\frac{\Delta t_{\rm res}}{\rm ns}}{\left[\left({\frac{\nu_l}{\rm GHz}}\right)^{-4}-\left({\frac{\nu_h}{\rm GHz}}\right)^{-4}\right]\frac{d}{\rm kpc}}\right\}^{1/4}\;.
\end{equation}
For a cosmic source at redshift $z$, the photon mass derived from the second-order effect is given by
\begin{equation}
m_{\gamma,2} \leq hc^{-2}\left[\frac{8H_0\Delta t_{\rm res}}{3\left(\nu_l^{-4}-\nu_h^{-4}\right)H_2(z)}\right]^{1/4}\;,
\end{equation}
which can be further reduced to
\begin{equation}\label{zmass}
m_{\gamma,2} \leq \left(2.02\times{10}^{-48}\;{\rm kg}\right)\left\{\frac{\frac{\Delta t_{\rm res}}{\rm ns}}{\left[\left({\frac{\nu_l}{\rm GHz}}\right)^{-4}-\left({\frac{\nu_h}{\rm GHz}}\right)^{-4}\right]H_2(z)}\right\}^{1/4}\;.
\end{equation}
\section{Second-order Photon mass limits from the dedispersed pulses}
Here we show that the nanosecond-long giant pulse observed from the Crab pulsar and the burst from FRB 180916B
detected at very low frequencies (below 150\,MHz) can provide strict limits on the second-order photon mass $m_{\gamma,2}$.
Even though the Crab pulsar giant pulse is in our Milky Way galaxy, the effect of the ultra short duration (e.g.,
$\Delta t_{\rm res}\leq0.4$ ns) overcompensates the deficit in distance, and results in a stricter limit on $m_{\gamma,2}$ than extragalactic ms-duration FRBs observed at the similar frequency range ($\sim$ GHz). Furthermore,
even though the burst durations ($\sim40$--160 ms observed at 150\,MHz) of FRB 180916B are relatively large,
the effect of the very-low-frequency emission overcompensates the deficit in duration, and leads to a much more stringent limit on $m_{\gamma,2}$ than the Crab pulsar giant pulse.
\subsection{The Crab Pulsar}
Giant pulses are one of the most conspicuous phenomena of radio pulsars, with extremely high fluxes (exceeding MJy) and
very short durations. Their typical durations are a few microseconds, but occasional pulses shorter than one nanosecond
(the so-called ``nanoshots'') have also been detected \citep{Hankins_2007}. To date, they have been observed from
the Crab pulsar \citep{2004ApJ...612..375C,2005AdSpR..35.1166J,2010A&A...524A..60J,Hankins_2007}
and some other pulsars \citep{2001ApJ...557L..93R,2003ApJ...590L..95J,2004ApJ...616..439S,2005ApJ...625..951K}.
\cite{Hankins_2007} reported a giant pulse from the Crab pulsar that showed an extremely intense nanoshot with a flux
exceeding 2 MJy and an unresolved dedispersed duration, $\Delta t_{\rm res}\leq0.4$ ns. This pulse was recorded at 9.25 GHz
center frequency over a 2.2 GHz bandwidth. The distance of the Crab pulsar is $d=2$ kpc. With the above information of
the 0.4-nanosecond giant pulse of the Crab pulsar, a severe limit on the photon mass from Equation~(\ref{mass}) is
\begin{equation}
m_{\gamma,2} \leq5.7\times10^{-46}\;{\rm kg}\simeq3.2\times10^{-10}\; {\rm eV}/c^{2}\;.
\end{equation}
\subsection{FRB 180916B}
FRB 180916B is a well-studied repeating FRB source, located at a redshift of $z=0.0337$ \citep{2020Natur.577..190M}.
\citet{2021ApJ...911L...3P} reported on the lowest-frequency detection to date of 18 bursts from FRB 180916B,
observed at 110--188 MHz with LOFAR. One of these bursts detected between 116.9 MHz and 188 MHz has a dedispersed
duration of $\Delta t_{\rm res} < 60$ ms. From Equation~(\ref{zmass}), we can tighten the constraint on the photon mass to
\begin{equation}
m_{\gamma,2} \leq5.1\times10^{-47}\;{\rm kg}\simeq2.9\times10^{-11}\; {\rm eV}/c^{2}
\end{equation}
for the dedispersed burst of FRB 180916B.
\section{CONCLUSION}
The frequency-dependent time delays of radio waves from astrophysical sources have been widely used to constrain the rest mass
of the photon. In this work, we make use of the dedispersed time delays between different energies from the Crab pulsar and
FRB 180916B, for the first time, to test the zero-mass hypothesis for the photon. This is a step forward in the investigation
of the photon mass, since all current studies considered the first-order Taylor expansion of the dispersion arising from
a nonzero photon mass only and the second-order effect $\Delta t_{m_{\gamma},2}$ has not yet been explored.
Assuming that the dedispersed time delay is mainly caused by the second-order term of the nonzero photon mass dispersion,
we place robust upper limits on the photon mass: $m_{\gamma,2} \leq5.7\times10^{-46}\;{\rm kg}$
(or equivalently $m_{\gamma,2} \leq3.2\times10^{-10}\; {\rm eV}/c^{2}$) for the 0.4-nanosecond giant pulse of the Crab pulsar and $m_{\gamma,2} \leq5.1\times10^{-47}\;{\rm kg}$ (or equivalently $m_{\gamma,2} \leq2.9\times10^{-11}\; {\rm eV}/c^{2}$) for the burst of the order of tens of milliseconds from FRB 180916B. It is notable that the ms-duration pulse of FRB 180916B detected at 116.9--188 MHz provides a much more stringent limit on $m_{\gamma,2}$, improving by at lest one order of magnitude from
the result based on the 0.4-nanosecond giant pulse of the Crab Pulsar observed at 8.15--10.35 GHz.
Our best limit is four orders of magnitude worse than previous limits
($m_{\gamma,1} \leq{10}^{-51}$ kg) obtained through the first-order Taylor expansion of the dispersion relation
\citep{2020RAA....20..206W,2021PhLB..82036596W}. While the dedispersed time delays of the Crab pulsar and FRB 180916B do not
currently have the best sensitivity to photon mass limits, there is nonetheless merit to the result. This is the first time to explore the possible second-order photon mass effect. More stringent constraints on $m_{\gamma,2}$ can be expected as our
analysis method is applied to larger numbers of nanoseconds-long astrophysical pulses observed at lower frequencies.
\begin{acknowledgments}
This work is partially supported by the National Natural Science Foundation of China
(grant Nos.~11725314 and 12041306), the Key Research Program of Frontier Sciences (grant No.
ZDBS-LY-7014) of Chinese Academy of Sciences, the Major Science and Technology
Project of Qinghai Province (2019-ZJ-A10), ACAMAR Postdoctoral Fellow, China Postdoctoral Science Foundation (Grant No. 2020M681758), the Natural Science Foundation of Jiangsu Province (grant No. BK20210998), and the China Manned Space Project (CMS-CSST-2021-B11).
\end{acknowledgments}
|
\section{Introduction}
Context information plays an important role in ASR, especially in scenes that require inter-sentential information such as conversation since semantically related words, or phrases often reoccur across sentences~\cite{kim2019cross}. Typically, traditional hybrid acoustic-language ASR models usually rely on rich language models to model contextual information~\cite{mikolov2010recurrent,mikolov2012context,mnih2007three, ji2015document,liu2017dialog, xiong2018session}. Meanwhile, there are also several researches adopting context information particularly in end-to-end ASR by adding additional context to the decoder or simply concatenate multiple consecutive utterances as the input of an end-to-end model~\cite{kim2018dialog, masumura2019large, masumura2021hierarchical}.
Transformer~\cite{vaswani2017attention}, as the most successful attention-based end-to-end model, has recently received more attention due to its superior performance on a wide range of tasks including ASR~\cite{wang2019learning, raganato2018analysis, dong2018speech, karita2019comparative, luo2021simplified, gulati2020conformer}.
However, since the computational and memory cost of self-attention is quadratic w.r.t the input sequence length, Transformer is hard to process long sequences and mainly models independent utterances.
Several studies in natural language processing (NLP) have been explored to utilize the long contextual information for Transformer~\cite{dai2019transformer, rae2019compressive, beltagy2020longformer, zhou2020informer}. Inspired by above studies in the NLP task, some approaches were also proposed to incorporate contextual information across successive input sequences in Transformer-based ASR~\cite{masumura2021hierarchical,hori2020transformer}, but these methods do not solve the problem of the high computational and memory cost, or have high model complexity.
In this study, we propose a novel Transformer-based architecture to explicitly model the inter-sentential information for conversational ASR. Inspired by~\cite{he2020realformer}, we include a residual attention module in the encoder, which accelerates the convergence speed and well models the long-range global dependencies within each input sequence. Besides, to further transfer the contextual information of previous sentences, we also propose a novel context-aware residual attention module, which transfers contextual information through attention scores. For the decoder part, we use an additional context module to learn more inter-sentential information. By using the methods above, we introduce inter-sentential contextual information in the popular Transformer ASR model. We demonstrate the superiority of our approach on two dialogue benchmarks (speech from two speakers) HKUST and Switchboard, a lecture benchmark TED-LIUM2, and a dialog dataset DATATANG-dialog, with obvious error rate reduction and neglectable increase of computational cost and model complexity.
\section{Transformer and Conformer}
Transformer~\cite{vaswani2017attention} is an attention-based end-to-end model, which consists of multi-block stacked encoder and decoder. Each block can be characterized by a multi-headed attention (MHA) module, a position-wise feed-forward (FFN) module, layer normalization (LN) layers, and residual connections.
The MHA module calculates the score through the vectors values $(\textbf{V}^h)$ and keys $(\textbf{K}^h)$, and assigns values to the output embeddings queries $(\textbf{Q}^h)$~\cite{he2020realformer}:
\begin{equation}
\text{MHA}(\textbf{Q}^h,\textbf{K}^h,\textbf{V}^h) = \text{Concat} (head_{1},...,head_{m})\textbf{W}^{O} ,
\end{equation}
where $head_{i}=\text{Attn}(\textbf{QW}^{Q}_{i},\textbf{KW}^{K}_{i},\textbf{VW}^{V}_{i})$.
Here, $\textbf{Q}$, $\textbf{K}$ and $\textbf{V}$ are matrices with dimension $d_{k}$,$d_{k}$ and $d_{v}$. $\textbf{W}^{Q}_{i},\textbf{W}^{K}_{i}$ and $\textbf{W}^{V}_{i}$ are matrices that maps three vectors to the \emph{i}-th head in the multi-head attention space. $\textbf{W}^{O}$ is a linear layer to transform the output after the stitching.
In the attention module, we use the traditional scaled dot-product attention module~\cite{vaswani2017attention} to calculate the attention scores, as shown below:
\begin{equation} \label{eq:dot_product}
\text{Attn}(\textbf{Q},\textbf{K},\textbf{V})=
\text{Softmax}(\frac{\textbf{Q}\textbf{K}^{T}}{\sqrt[]{d_{k}}})\textbf{V},
\end{equation}
where $\frac{\textbf{Q}^h\textbf{K}^h{T}}{\sqrt[]{d_{k}}}$ is a matrix representing the attention scores for each query and key.
The FFN module is composed of two fully-connected layers with a ReLU activation in between, as follows:
\begin{equation}
\text{FFN}(x)= \text{ReLU} (x\textbf{W}_{1}+b_{1})\textbf{W}_{2}+b_{2},
\end{equation}
In addition to the two sub-layers in each encoder block, the decoder inserts a third sub-layer, which performs multi-headed attention over the source and target sequences.
Recently, Gulati et al.~\cite{gulati2020conformer} combined Transformer and Convolutional Neural Networks (CNN) as Conformer. The Conformer encoder adds a convolution module between MHA and FFN in Transformer encoder blocks, simultaneously capturing local and global contextual information and leading to superior performance in ASR tasks.
\section{Proposed Method}
\subsection{Residual Multi-headed Attention}
Residual multi-headeded attention (ResMHA) closely follows the same Post LN strategy as in~\cite{vaswani2017attention}, which normalizes the output at the end of each MHA or FFN module~\cite{he2020realformer}. ResMHA connects MHA of adjacent layers through attention scores, as shown in Fig. 1 (a). Formally, it uses the attention score \emph{Prev} of the previous layer as a conditional input to calculating the attention score of the current layer’s MHA. In particular, \emph{Prev} is the attention score before the Softmax operation.:
\begin{equation}
\begin{split}
\text{ResMHA} & (\textbf{Q}^h,\textbf{K}^h,\textbf{V}^h, \emph{Prev}) = \\
& \text{Concat}(head_{1},...,head_{m})\textbf{W}^{O},
\end{split}
\end{equation}
where $head_{i}=\text{ResAttn}(\textbf{QW}^{Q}_{i},\textbf{KW}^{K}_{i},\textbf{VW}^{V}_{i}, \emph{Prev}_{i})$.
Like $head_{i}$, $\emph{Prev}_{i}$ is the slice of $\emph{Prev}$. Then these residual attention scores, which corresponding to the MHA heads, will be added to the current attention calculation through the Residual attention (ResAttn) module:
\begin{equation}
\begin{split}
\text{ResAttn} &(\textbf{Q},\textbf{K},\textbf{V},\emph{Prev})=
\text{Softmax}(\frac{\textbf{Q}\textbf{K}^{T}}{\sqrt[]{d_{k}}}+\emph{Prev}_i)\textbf{V}.
\end{split}
\end{equation}
We apply it to speech Transformer, accelerating the convergence speed of the model during training and improving the final speech recognition accuracy.
\subsection{Context-aware Multi-headeded Residual Attention}
Context-aware multi-headeded residual attention is designed on the basis of Residual attention~\cite{he2020realformer}, which adds a skip edge to connect multi-headed attention (MHA) modules adjacent layers, as shown in Fig. 1~(a).
To capture contextual information across different consecutive utterances, an intuitive idea is to simply concatenate previous inputs with the current input to the encoder. Although, such method can give a slight performance improvement, it will also confront a large increment of memory cost and computation complexity. Thus, we propose a context-aware multi-headeded attention, which transfers the attention hidden states of the previous sentence in time order ($\emph{StateLS}$) to the current sentence and straightforwardly includes more contextual information during the training. Fig.~\ref{fig:realattention} (b) shows the details of our method. When we recognize the m-th sentence $X_{m}$ in the conversation and use the previous sentence $X_{m-1}$, or each $head_i$, the context-aware residual attention (CtxResAttn) can be formulated as:
\begin{equation}
\begin{split}
\text{CtxResAttn}(\textbf{Q},\textbf{K},\textbf{V}, \emph{Prev}, \emph{PrevLS} (X_{m-1})) = \\
\text{Softmax}(\frac{\textbf{Q}\textbf{K}^{T}}{\sqrt[]{d_{k}}}+ \emph{Prev}+\alpha \emph{PrevLS}(X_{m-1}))\textbf{V},
\end{split}
\end{equation}
where the CtxResAttn is our proposed context-aware residual attention module, $\emph{PrevLS}(X_{m-1})$ is the correlation attention score of previous one sentence before current input sentence $X_{m}$, and $\alpha$ is an interpolation factor to adjust the weight of historical information. The correlation attention score is calculated from pre-Softmax attention scores $X_{m-1}$ and $X_{m}$ as follows:
\begin{equation}
\begin{split}
\emph{PrevLS}(X_{m-1})= \text{LAN}(s(X_{m-1}), s(X_{m})),
\end{split}
\end{equation}
where LAN is a linear layer, and $s(X_{m})$ is the pre-Softmax attention score of input $X_{t}$. Similar to the ResAttn, the new attention scores are applied on multi-head attention and passed over to the next layer.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{realattention.pdf}
\caption{Details of the residual attention and context-aware residual attention.}
\label{fig:realattention}
\end{figure}
\subsection{Conditional Decoder}
The decoder of Transformer contains a lot of linguistic information. In this part, we will describe our proposed conditional decoder, which adds the embedded vector to current input vector through an iterative attention block. As shown in Fig. 2, we recursively get the information of the previous vector, and then incorporate it with the current embedded input vector. Supposing we will look backward $n$ historic sentences when recognizing the $k$-th sentence in the conversation, the label texts $\mathbf{Y_{m-n}}$ and $\mathbf{Y_{m-n+1}}$ will be first processed by the Context-previous attention layer (CtxPrevAttn).
\begin{equation}
\begin{split}\label{c1}
&\text{CtxPrevAttn}(Y_{m-n}, Y_{m-n+1})= \\
& \text{Attn}(\text{Embed}(Y_{m-n}), \text{Embed}(Y_{m-n+1}), \text{Embed}(Y_{m-n+1})),
\end{split}
\end{equation}
where the Embed is the word embedding layer and Attn is the dot-product attention mechanism, as described in Eq.~\ref{eq:dot_product}. Next, the output of the attention layer will be sent to a liner layer to get the position information:
\begin{equation}
\begin{split}\label{c2}
& \text{Context}(Y_{m-n}, Y_{m-n+1})= \\
& \text{LAN}(\text{CtxPrevAttn}(Y_{m-n}, Y_{m-n+1}), \text{Embed}(Y_{m-n+1})),
\end{split}
\end{equation}
where LAN is a liner transform layer. Then, we get the output of
$\text{Context}(\text{Context}(Y_{m-n},Y_{m-n+1}),\text{Embed}(Y_{m-n+2}))$ in turn until the current $m$-th sentence. Finally, we send the contextual vector to the decoder of Transformer.
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{decoder.pdf}
\caption{Conditional decoder with n previous sentences contextual information.}
\label{fig:decoder}
\end{figure}
\section{Experiments}
\subsection{Dataset}
We conduct experiments on three dialogue datasets, including HKUST~\cite{liu2006hkust}, Switchboard~\cite{godfrey1992switchboard}, DATATANG-dialog, and a lecture dataset TED-LIUM2. Table 1 summarizes the information of each dataset. The 80-dimensional log-Mel filterbank(fbank) acoustic features plus 3-dimensional pitch features are used as the input features.
For each corpus, the detail configurations of our Transformer model are same as ESPnet Transformer recipes~\cite{watanabe2018espnet} $(Enc =
12, Dec = 6, d^{ff} = 2048, H = 4, d^{att} = 256)$.
We use 1996 byte-pair-encoding (BPE) tokens~\cite{kudo2018sentencepiece} as output units for the English corpora Switchboard and TED-LIUM2, and characters for the Mandarin corpora HKUST and DATATANG-dialog.
\begin{table}[htbp]\centering
\caption{Dataset information.}
\begin{tabular}{l|l l l}
\hline
dataset & lang & hours & test\_sets \\ \hline
Switchboard & en & 260 & callhome/swbd \\
TED-LIUM2 & en & 213 & dev/test \\
HKUST & ch & 200 & train\_dev/dev \\
DATATANG-dialog & ch & 176 & train\_dev/dev \\ \hline
\end{tabular}
\end{table}
\begin{table*}[ht]\label{res}\centering
\caption{Results of proposed method. Numbers indicate WERs (\%) for SWITCHBOARD and TED-LIUM2, and CERs (\%) for HKUST and DATATANG(DATATANG-dialog) }
\begin{tabular}{c|c|c|c|c|c|c|c|c}
\hline
& \multicolumn{2}{c|}{Switchboard} & \multicolumn{2}{c|}{TED-LIUM2} & \multicolumn{2}{c|}{HKUST} & \multicolumn{2}{c}{DATATANG} \\ \hline
&
\multicolumn{1}{c|}{callhome} &
\multicolumn{1}{c|}{swbd} &
\multicolumn{1}{c|}{dev} &
\multicolumn{1}{c|}{test} &
\multicolumn{1}{c|}{train\_dev} &
\multicolumn{1}{c|}{dev} &
\multicolumn{1}{c|}{train\_dev} &
\multicolumn{1}{c}{dev} \\ \hline
Transformer baseline & 17.3 & 8.5 & 11.2 & 9.4 & 24.2 & 23.6 & 23.9 & 25.1 \\ \hline
+Proposed method & \textbf{16.3} & \textbf{8.3} & \textbf{9.6} & \textbf{8.7} & \textbf{23.5} & \textbf{22.9} & \textbf{23.0} & \textbf{23.9} \\ \hline
Conformer baseline & 15.6 & 8.4 & 10.2 & 9.0 & 20.8 & 20.0 & 17.4 & 18.1 \\ \hline
+Proposed method & \textbf{15.1} & \textbf{8.0} & \textbf{9.7} & \textbf{8.7} & \textbf{19.9} & \textbf{19.7} & \textbf{17.0} & \textbf{17.7} \\ \hline
\end{tabular}
\end{table*}
\subsection{Experimental Setup}
We train the Transformer and Conformer~\cite{gulati2020conformer} models with the open-source end-to-end speech processing toolkit ESPnet~\cite{watanabe2018espnet, guo2020recent}. We use speed perturbation at ratio 0.9, 1.0, 1.1 for all corpora.
We also apply SpecAugment~\cite{park2019specaugment} for additional data augmentation. Moreover, the Switchboard dataset is trained with more epochs than the other datasets.
The baseline results are trained on independent sentence level, without speaker and context information.
We find that random input training, in other words, in the process of model training, the input sentences are not sent into the model in chronological order, will get better results than the time-ordered input. So instead of training with time-ordered input, we keep additional lists of dialog information during training. We send the features of previous sentences to the encoder to get the attention score.
Since there is no need for backpropagation, the demand for computing resources is not significantly increased.
For the decoder,
we prepare two historical sentences, which means $n=2$, for decoding every sentence.
In order to be consistent with the training input, in the decoding stage, we use the decode results of previous sentences to act as the previous text. For the first sentence in dialog, we just repeat this sentence to replace the position of the previous text. We use two previous sentences information in this paper when training models and the context-aware residual attention has a constant $\alpha$ value of 0.1.
When decoding, we average the best five models based on the validation loss for recognition on HKUST and DATATANG-dialog, while 10 models for Switchboard and TED-LIUM2. Besides, we use Long Short Term Memory Network language models (LSTM LM) to improve the recognition accuracy. The LM consist of four LSTM layers with 1024 units. The LM of HKUST, TED-LIUM2 and DATATANG-dialog are trained with the transcripts of training sets, and the Switchboard language model is trained with additional Fisher texts. ASR performance is measured by character error rate (CER) or word error rate (WER) depending on the language.
\subsection{Results and Analysis}
The results are shown in Table 2. For each dataset, the proposed method reduces ASR errors compared with the baseline, and the relative error rate reduction is up to 14\%. The DATATANG-dialog and TED-LIUM2 datasets get more error rate reduction. We suppose that those two datasets have greater topic coherence, which enhances the learning ability of the model to the corresponding keywords of specific domains.
\subsection{Ablation Study}
We also conduct ablation experiments on HKUST and DATATANG-dialog datasets. In Table 3, $add\_1sen$ means using the proposed conditional decoder with $n=1$, real means using ResAttn and con means using CtcResAttn. Numbers indicate CERs (\%).
\begin{table}[ht]\centering
\caption{Ablation study on HKUST and DATATANG-dialog.}
\begin{tabular}{l|l|l|l|l}
\hline
& \multicolumn{2}{c|}{HKUST} & \multicolumn{2}{c}{DATATANG} \\ \hline
& train\_dev & dev & train\_dev & dev \\ \hline
Transformer Baseline & 24.2 & 23.6 & 23.9 & 25.1 \\ \hline
+add\_1sen & 24.0 & 23.5 & 23.7 & 25.2 \\ \hline
+add\_2sen & 23.8 & 23.3 & 23.1 & 24.3 \\ \hline
+add\_2sen+real & 23.6 & 23.3 & 23.1 & 24.1 \\ \hline
+add\_2sen+con & \textbf{23.5} & \textbf{22.9} & \textbf{23.0} & \textbf{23.9} \\ \hline
\end{tabular}
\end{table}
\subsubsection{Conditional Decoder}
From the second row and the third row of Table 3, we can find that the attention decoder input with two previous sentences significantly improved the recognition accuracy. However, the attention decoder with one sentence gets negligible improvement and even the negative effect on the DATATANG-dialog dataset. We come to the conclusion that the longer contextual information, especially in the sentences with both sides of a conversation, can help the decoder to learn more language information.
To verify the hypothesis, we performe speaker-dependent and speaker-independent experiments on dataset DATATANG-dialog. We add the previous sentence information of all speakers in the conversation in the speaker-independent training, and add the previous sentence information of current speaker in the speaker-dependent training. We verify the hypothesis in two different input methods, one is the time-order input of the dialogue, and the other one is the shuffle input. In the time-order training, we sort the input by speaker or simply sort by time to distinguish speaker-dependent or not. In the shuffle training, we keep a list of history context additionally, because we cannot get historical information directly based on the shuffled input. We can see from Table 4 that, the results of speaker-dependent are worse than speaker-independent, both in model trained with time-ordered training dataset and shuffle training dataset. We can draw the conclusion that sentences that span both sides are better than the historical information on either side of the conversation.
\begin{table}[htbp]\centering
\caption{Speaker information study on DATATANG-dialog, SI is the speaker-independent training and SD is the speaker-dependent training, indicate CERs (\%).}
\begin{tabular}{c|l|c|c|c|c}
\hline
& & \multicolumn{2}{c|}{time-order} & \multicolumn{2}{c}{shuffle} \\ \hline
& & train\_dev & dev & train\_dev & dev \\ \hline
\multirow{2}{*}{baseline} & SI & 25.5 & 26.8 & 23.9 & 25.1 \\ \cline{2-6}
& SD & 25.9 & 27.0 & 24.6 & 25.7 \\ \hline
\multirow{2}{*}{proposed} & SI & \textbf{23.7} & \textbf{24.8} & \textbf{23.0} & \textbf{23.9} \\ \cline{2-6}
& SD & 24.2 & 25.3 & 23.8 & 24.6 \\ \hline
\end{tabular}
\end{table}
\subsubsection{Context-aware Residual Attention}
The residual attention in encoder does not add any multiplication operations to the computational graph and improves the accuracy of the ASR task. Meanwhile, it achieves competitive results on ASR training with only 90\% of the number of epochs of the baseline. As shown in the forth row and the fifth row of Table 3, context-aware residual attention has greatly improved the recognition accuracy of our models. The context-aware residual attention adds previous context information and does not introduce too much redundant information into the encoder.
\subsection{The Impact of I-vector}
Previous work has shown that simply concatenating speaker related features, e.g. i-vector, with acoustic feature benefits conversational ASR ~\cite{audhkhasi2017direct,fan2019speaker}. I-vector can introduce additional speaker context and reduce the speaker mismatch between the training set and the test set. We further study our approach on the DATATANG-dialog dataset to see if there is still space for performance improvement when i-vector is adopted.
The i-vector estimator is trained with all the training data of DATATANG-dialog dataset. The training process follows the SRE08 recipe in Kaldi toolkit~\cite{povey2011kaldi}. A 2048 diagonal component universal background model (UBM) is first trained, and then 200-dimensional i-vectors are extracted and further compressed to 100 dimension by linear discriminant analysis (LDA) followed by length normalization. We concat the input fbank-pitch feature with 100-dimensional i-vector vector and send it to the network for training. It can be seen from Table 5 that the use of i-vector do lead to substantial performance gain, and our proposed method still can achieve extra performance improvement after adding i-vector as input
\begin{table}[]\centering
\caption{CER (\%) results of i-vector and our proposed method on DATATANG.}
\label{tab:my-table}
\begin{tabular}{l|l|l}
\hline
& train\_dev & dev \\ \hline
Baseline & 23.9 & 25.1 \\ \hline
+proposed method & 23.0 & 23.9 \\ \hline
+i-vector & 21.3 & 21.9 \\ \hline
+i-vector+proposed method & \textbf{20.6} & \textbf{21.2} \\ \hline
\end{tabular}
\end{table}
\subsection{Attention Mechanism in Conditional Decoder}
We also analyze the capture ability improvement of recognition ability of keywords after adding the conditional decoder context module. We compare the attention scores of baseline decoder and conditional decoder in the third layers. We use the attention score of the first head in the decoder layer to plot the figure. Fig. 3 shows an example.
We can find that the dark color in Fig. 3(b) is concentrated on the diagonal while Fig. 3(a) has not yet formed a reasonable dissemination of attention. We can draw a conclusion that the conditional decoder with the context module can get accurate language information faster at a shallower level. Moreover, the dark blocks are better focused on the keywords mentioned in the previous conversation, which means the attention layer in Figure 3(b) improves the ability to perceive the keywords mentioned above.
\begin{figure}[htbp]
\centering
\subfigure[baseline]{\includegraphics[width=0.23\textwidth]{attention2.pdf}}
\subfigure[conditional decoder]{\includegraphics[width=0.23\textwidth]{attention.pdf}}
\caption{Attention scores of 3-th decoder layers, a darker color indicates a higher score for the character.}
\label{fig:1}
\end{figure}
\section{Conclusions}
In this paper, we design context-aware residual attention to get the contextual information without extra modules and parameters to the encoder. Moreover, the conditional decoder takes the text information from previous speech and improves the ability of the model to capture long contextual information. The experiments on four datasets demonstrated the effectiveness of our method in enhancing the prediction capacity in dialog ASR tasks. We will improve the decoding speed of our method and context-sensitive decoding strategies in our future work.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Consider the fixed-effects panel data regression model:
\begin{eqnarray}\label{model}
y_{it} = x_{it}^\prime \beta +\mu_i +\nu_{it}, \ \textrm{for} \ i=1,\ldots,n;\ t=1,\ldots,T,
\end{eqnarray}
where $i$ indexes the cross-sectional (individual) units and $t$ the time-series observations. The dependent variable is $y_{it}$ and $x_{it}$ denotes the exogenous regressors of dimension $k\times 1$, $\beta$ is the corresponding $k\times 1$ vector of parameters, $\mu_i$
denotes the time-invariant individual effect, which could be correlated with the regressors $x_{it}$. Throughout the paper, the number of covariates $k$ is fixed while the two dimensions $n$ and $T$ may grow to infinity.
Let $\nu_t = (\nu_{1t},\ldots,\nu_{nt})^\prime$ be the panel-wise
error vector at time $t$. In this paper, we assume that the error vectors
$\nu_1,\ldots, \nu_T$ are homoscedastic so that the cross-sectional covariance matrix
$\Sigma_n=\cov(\nu_t)$ is independent of $t$.
Our primary focus is on testing the sphericity of the covariance matrix of the disturbances:
\begin{eqnarray}\label{null}
H_0: \Sigma_n = \sigma_{\nu}^2 \mathbf{I}_n, \quad \textrm{vs} \quad H_1: \Sigma_n \neq \sigma_{\nu}^2 \mathbf{I}_n,
\end{eqnarray}
where $\sigma_{\nu}^2$ is an unknown positive constant and $\mathbf{I}_n$ the identity matrix of order $n$.
Testing for sphericity of symmetric errors in the linear model has a long history. Among others, the problems of testing for serial correlation, spatial correlation, unit roots, and stationarity can all be cast in this framework \citep{broda2019testing}. The null of sphericity expresses that the disturbances are cross-sectionally uncorrelated (independent if they are normal) and have the same variance (homoscedasticity). Rejecting the null means having cross-sectional correlation among the individual units of observations, heteroskedasticity, or both. This test is important for applied panel data work, given the numerous applications that report fixed effects estimation ignoring cross-section dependence or heteroscedasticity \citep{baltagi2011testing}.
Much of the existing theory about the test for sphericity has been explored first in \cite{mauchly1940significance} about Gaussian likelihood ratio test and later in \cite{john1971some} and \cite{john1972distribution} about the invariant John's test, and also in textbooks like \cite{anderson2003introduction}. It is well known that classical multivariate procedures are generally challenged by large dimensional data \citep{paul2014random, yao2015sample}. The likelihood ratio test requires the dimension to be smaller than the sample size; otherwise, the likelihood ratio is identically null. Some corrections to traditional likelihood ratio test are proposed for large dimensional data; see \cite{wang2013sphericity} and \cite{li2016testing}. \cite{ledoit2002some} corrected John's test to cope with the high-dimensional context with normality assumption. \cite{wang2013sphericity} removed such a normality restriction and proved that the robustness of John's test is general. \cite{chen2010tests} proposed to use a family of well-selected U-statistics to test the sphericity; however, this test is very time-consuming \citep{li2016testing} and is illustrated to be slightly lower power than the corrected John's test proposed by \cite{wang2013sphericity}. \cite{li2016testing} studied John's test under an ultra-high-dimensional regime where the dimension is of a higher order than the sample size.
These existing statistical tests for raw data do not apply directly to the sphericity test in panel regression models since the disturbances are unobservable. \cite{baltagi2011testing} extended \cite{ledoit2002some}'s result on John's test to the fixed effects panel data model. They corrected for the bias due to substituting the within residuals for the actual disturbances. \cite{baltagi2017asymptotic} studied the corresponding asymptotic power under the alternative of factor model. However, these results rely on the normality assumption of disturbances, a restriction for real data analysis . Building on the work of \cite{chen2010tests}, \cite{baltagi2015testing} extended the U-statistics to test for sphericity of the disturbances without the normality assumption. However, this kind of U-statistics carries a burden of extensive computations, and its consistency under the alternative is unknown.
This paper focuses on John's test due to its robustness proved by existing literature. We make the following main contributions.
\begin{enumerate}
\item {\bf We consider general populations of disturbances}. Without normality restriction on disturbances, we derive the null distribution of John's test based on residuals for the sphericity test in (\ref{null}) and conduct power analysis.
\item {\bf We provide theoretical proof of the dimension-proof property of John's test in the large panel data model}.
This paper studies the panel data model (\ref{model}) in large dimensions, that is, the number of cross-section units is commensurate with or much larger than the number of times series observations. Accordingly, by imposing conditions on the relative speed at which $n$ and $T$ go to infinity, we define two asymptotic schemes:
\begin{itemize}
\item Large-panel asymptotic (LPA): the number of cross-sectional units $n$ and the number of time-series observations $T$ tend to infinity in such a way that $c_T = n/T \to c \in (0,\infty)$.
\item Ultra-large panel asymptotic (ULPA): the number of cross-sectional units $n$ and the number of time-series observations $T$ tend to infinity in such a way that $c_T = n/T \to \infty$, $T^\delta /n = O_p(1), 1<\delta<2$.
\end{itemize}
The LPA scheme is appropriate for macroeconomic applications in which typically $n$ and $T$ are both large and of comparable magnitudes. The ULPA scheme is appropriate for microeconomic applications where typically $n$ is much larger than $T$.
John's test is studied under both of LPA and ULPA schemes defined above, and then its potent dimension-proof property is established, that is, John's test based on the residuals of the panel data model keeps the same null distribution under different $(n, T)$-asymptotics, i.e., the small or medium panel regime $n/T\to 0$ as $T\to \infty$, the large panel regime $n/T\to c \in (0,\infty)$ as $ T\to \infty$ and the ultra-large panel regime $n/T\to \infty, T^\delta/n =O_p(1), 1<\delta<2$ as $T\to \infty$.
\item {\bf We conduct a comprehensive power analysis}. We study the asymptotic power of John's test under the factor model alternative with different settings. Without normality assumption on the disturbances and following \cite{baltagi2017asymptotic}, we investigate the asymptotic power of John's test for the panel model (\ref{model}) under the alternative of a weak factor model with fixed number of factors. Moreover, we conduct the power analysis under the following three scenarios, which are proposed as open questions by \cite{baltagi2017asymptotic}:
\begin{enumerate}
\item the asymptotic power as $\frac{n}{T}\to \infty$, i.e., under the ULPA scheme;
\item the factors are neither strong nor weak;
\item the factors are weak, but the number of factors goes to infinity jointly with $n$ and $T$.
\end{enumerate}
\end{enumerate}
The rest of the paper is organized as follows. Section 2 includes a literature review on John's test for raw data and panel regression model, respectively. The null distributions of residual-based John's test under both LAP and ULAP schemes for general populations of disturbances are presented in Section 3. Section 4 conducts the power analysis under the factor model alternative with different settings, followed by concluding remarks in Section 5. All technical proofs are relegated to the Appendix.
\section{John's test, residual-based John's test}
\subsection{John's test for raw data}
\cite{john1971some} proposed a test statistic for the null hypothesis of sphericity in (\ref{null}) defined as follows:
\begin{eqnarray}\label{U}
U = \frac{1}{n} tr \left[\left(\frac{1}{n} tr S_ T\right)^{-1} S_T -\mathbf{I}_n \right]^2 = \left(\frac{1}{n} tr S_T \right)^{-2} \left(\frac{1}{n}tr S_T^2 \right) -1,
\end{eqnarray}
where $S_T$ is the {\em sample error covariance matrix} with order $n$. For the panel regression model (\ref{model}), $\displaystyle S_T=\frac{1}{T}\sum_{t=1}^T \nu_t \nu_t^\prime$. Notice that the statistic $U$ is independent of the scale parameter $\sigma_{\nu}^2$ under the null. Therefore, we assume w.l.o.g. $\sigma_{\nu}^2=1$ when dealing with the null distribution of $U$.
\subsubsection{Large sample theory}
\cite{john1971some} proved that when the observations $\{\nu_{it}\}$ are normal, the sphericity test based on $U$ is a locally most powerful invariant test. Under these conditions, \cite{john1972distribution} established the large sample limiting distribution of $U$ under $H_0$, when $n$ is fixed and $T\to \infty$,
\begin{eqnarray}\label{john1972}
TU-n \overset{\mathcal{D}}{\longrightarrow} \frac{2}{n}\chi^2_{n(n+1)/2-1} -n.
\end{eqnarray}
This is referred to as John's test. It has been noticed for a while that John's test does not suffer from large dimensions, and this $\chi^2$ limit is quite accurate even when the ratio $n/T$ is not small.
\subsubsection{High-dimensional theory}
\cite{ledoit2002some} studied the null distribution of $U$ under the LPA scheme. Assume that the observations $\{\nu_{it}\}$ are normal-distributed, under $H_0$ and the LPA scheme,
\begin{eqnarray}\label{ledoit}
TU-n \overset{\mathcal{D}}{\longrightarrow} N(1,4).
\end{eqnarray}
Meanwhile, if we let $n\to \infty$ on the right-hand side of (\ref{john1972}), it is not hard to see that
\begin{eqnarray*}
\frac{2}{n}\chi^2_{n(n+1)/2-1} -n \overset{\mathcal{D}}{\longrightarrow} N(1,4).
\end{eqnarray*}
In other words, \cite{ledoit2002some} extended the classical $T$-asymptotic theory (where $n$ is fixed) to the LPA scheme where $n$ goes to infinity proportionally with $T$.
Subsequently, \cite{wang2013sphericity} relaxed the normality restriction and proved that, assume $\{\nu_{it}\}$ are i.i.d., satisfying $E \nu_{it}=0, E|\nu_{it}|^2=1, \gamma_4 = E|\nu_{it}|^4<\infty$, under $H_0$ and the LPA scheme,
\begin{eqnarray}\label{Wangyao}
TU-n \overset{\mathcal{D}}{\longrightarrow} N(\gamma_4-2,4).
\end{eqnarray}
\subsubsection{Ultra high-dimensional theory}
\cite{li2016testing} further considered the asymptotic behavior of John's test statistic under the ULPA framework. Assume that $\{\nu_{it}\}$ are i.i.d., satisfying $E (\nu_{it})=0, E|\nu_{it}|^2=1, \gamma_4 = E|\nu_{it}|^4<\infty$, under $H_0$ and the ULPA scheme,
\begin{eqnarray}\label{liyao}
TU-n \overset{\mathcal{D}}{\longrightarrow} N(\gamma_4-2,4),
\end{eqnarray}
which is the same as the result~(\ref{Wangyao}) under the LPA scheme.
Since $\gamma_4=3+o(1)$ for normal distribution, the current results confirm each other. Besides, if the data has a non-normal distribution but has the same first four moments as the normal distribution, we have again $TU-n \overset{\mathcal{D}}{\longrightarrow} N(1,4)$ in (\ref{ledoit}). Another striking fact is that the limiting distribution of $U$ is independent of the dimension-to-sample ratio $\lim n/T$. John's test possesses a remarkable {\it dimension-proof} property, making it a very competitive candidate for sphericity testing regardless of the magnitude of $n,T$. Therefore, this paper focuses on John's test for sphericity test of the disturbances in the panel regression model (\ref{model}).
\subsection{John's test for panel data model}
In the panel regression model (\ref{model}), as $\{\nu_{it} \}$ are unobserved, the test statistic $U$ is modified using regression residuals. Consider the OLS estimator of the slope parameter $\beta$,
\begin{eqnarray*}
\hat{\beta}= \left(\sum_{t=1}^T \sum_{i=1}^n \tilde{x}_{it}\tilde{x}_{it}^\prime \right)^{-1} \left(\sum_{t=1}^T \sum_{i=1}^n \tilde{x}_{it}\tilde{y}_{it} \right),
\end{eqnarray*}
where $\tilde{x}_{it} = x_{it}-\frac{1}{T}\sum_{t=1}^T x_{it}$ and $\tilde{y}_{it} = y_{it}-\frac{1}{T}\sum_{t=1}^T y_{it}$. It is well established that under mild conditions, $\hat{\beta}$ is a consistent estimator of $\beta$ no matter whether $n$ is fixed or tends to infinity jointly with $T$ \citep{baltagi2011testing,baltagi2017asymptotic}. The regression residuals $\{ \hat{\nu}_{it}\}$ are thus,
\begin{eqnarray}
\hat{\nu}_{it} = \tilde{y}_{it} - \tilde{x}_{it}^\prime \hat{\beta} =\tilde{\nu}_{it} -\tilde{x}_{it}^\prime \left(\hat{\beta}-\beta \right),
\end{eqnarray}
where $\tilde{\nu}_{it}=\nu_{it}-\bar{\nu}_{i.}$ with $\bar{\nu}_{i.}=\frac{1}{T}\sum_{t=1}^T \nu_{it}$. The {\em sample residual covariance matrix} is $\displaystyle \hat{S}_T = \frac{1}{T}\sum_{t=1}^T \hat{\nu}_t \hat{\nu}_t^\prime$, where $\hat{\nu}_t = (\hat{\nu}_{1t},\ldots,\hat{\nu}_{nt})^\prime$ for $t=1,\ldots, T$, and the modified John's test statistic based on $\hat{S}_T$ is
\begin{eqnarray}\label{U_hat}
\hat{U} = \left(\frac{1}{n} tr \hat{S}_T \right)^{-2} \left(\frac{1}{n}tr \hat{S}_T^2 \right) -1.
\end{eqnarray}
which is referred to as the residual-based John's test statistic.
\subsubsection{The null distribution under LPA with normality}
The asymptotics of $\hat{U}$ can be derived from the asymptotics of $U$ by estimating their difference, that is to examine the effect of replacing the errors $\{\nu_{it} \}$ with the residuals $\{\hat{\nu}_{it} \}$ on the asymptotics. Based on the result (\ref{ledoit}) of \cite{ledoit2002some}, \cite{baltagi2011testing} proved that, when the disturbances $\{\nu_{it} \}$ are normal-distributed, under $H_0$ and the LPA scheme, $\hat{U}$ follows asymptotically a normal distribution,
\begin{eqnarray}\label{b2011}
T\hat{U}-n-\frac{n}{T-1} \overset{\mathcal{D}}{\longrightarrow} N(1,4).
\end{eqnarray}
Compared to (\ref{ledoit}), the use of residuals introduces a drift $\frac{n}{T-1}$ in the asymptotic mean of John's statistic. The test based on the asymptotic normal distribution given in the equation (\ref{b2011}) will be referred to as the {\em residual-based John's (RJ) test}.
\subsubsection{Asymptotic power of the RJ test}
\cite{baltagi2017asymptotic} studied the asymptotic power of the RJ test under the alternative of factor model, i.e., the disturbances $\{\nu_{it}\}$ follow a factor model,
\begin{eqnarray}\label{factor_model}
\nu_{it} = \sum_{j=1}^r \xi_{ij}f_{tj} +\epsilon_{it},
\end{eqnarray}
where $\xi_{ij}$ is the factor loading of individual $i$ for factor $j$, $f_{tj}$ is the factor $j$ in period $t$, and $r$ is the known number of factors. To ensure the identification of the model, constraints are introduced as follows. Let $\xi_j = (\xi_{1j},\ldots,\xi_{nj})^\prime$, and $\Xi=(\xi_1,\ldots,\xi_r)$.
\begin{itemize}
\item[(a)] The factors $\{f_{tj} \}$ are i.i.d. with mean $0$ and variance $\sigma_j^2<\infty$ across time, the correlation coefficient between factors $f_{tj}$ and $f_{tk}$ is zero, for all $j, k$ and $t$.
\item[(b)] The vectors of factor loading $\xi_j$ are orthogonal to each other, that is, the matrix $\Xi^\prime \Xi$ is diagonal with distinct diagonal elements.
\item[(c)] The idiosyncratic errors $\{\epsilon_{it} \}$ are i.i.d. $N(0, \sigma_{\epsilon}^2)$, and independent of all factors.
\end{itemize}
Then, the population covariance matrix of the disturbances $\{\nu_{it}\}$ is
\begin{eqnarray} \label{Sigma_factor}
\Sigma_n = \sigma_\epsilon^2 \left(\sum_{j=1}^r h_j e_je_j^\prime +\mathbf{I}_n \right),
\end{eqnarray}
where $h_j = \frac{\sigma_j^2}{\sigma_\epsilon^2}||\xi_j||^2$, $e_j = \frac{\xi_j}{||\xi_j||}$, and $||e_j||=1$. And the covariance matrix can be viewed as a rank-$r$ perturbation of the null case. Let $\Sigma_n =\Gamma_n D_n \Gamma_n^\prime$, where $D_n = diag(\lambda_1, \ldots, \lambda_n)$ with the eigenvalues of $\Sigma_n$, $\lambda_j = \sigma_\epsilon^2 (1+h_j)$ for $j=1,\ldots, r$, and $\lambda_j=\sigma_\epsilon^2$ for $j=r+1,\ldots,n$. The matrix $\Gamma_n$ is the eigenvector matrix. This is known as spiked population model introduced in \cite{johnstone2001distribution}.
\cite{baltagi2017asymptotic} studied two kinds of alternatives: $h_j \to c_j \in (0, \infty)$ and $h_j/n \to c_j \in (0,\infty)$, which correspond to the weak and strong factor cases considered by \cite{onatski2012asymptotics} and \cite{johnstone2009consistency}. They found that the RJ test is consistent under the strong factor alternative while inconsistent under the weak factor alternative. Under the factor model alternative (\ref{factor_model}) with weak factors $h_j \to c_j \in (0, \infty)$, constraints (a)-(c), and the LPA scheme,
\begin{eqnarray}\label{alter_dist_weak}
T\hat{U}-n-\frac{n}{T-1}-\frac{T \sum_{j=1}^r h_j^2}{n} \overset{\mathcal{D}}{\longrightarrow} N(1,4).
\end{eqnarray}
Since $r$ is fixed, the limiting distribution of $\hat{U}$ under this alternative is just a rightward mean shift of its null distribution in (\ref{b2011}).
\section{Extension of the RJ test}
The extension is twofold: we relax the normality assumption on the disturbances and investigate the RJ test under both LPA and ULPA schemes. The powerful dimension-proof property of John's test for the panel data model is presented here.
To derive the null distribution of the residual-based John's test statistic $\hat{U}$ in (\ref{U_hat}), we need the following assumptions to general populations of disturbances in the panel data model (\ref{model}).
\begin{assumption}\label{A1}
The regressors $\{x_{it} \}$ are independent of the
disturbances $\{\nu_{it} \}$, and have uniformly bounded fourth moments, that is
\[
\sup_{i,t} E|x_{it}|^4\le C_1, \quad
\text{for some positive constant } C_1.
\]
\end{assumption}
\begin{assumption}\label{A2}
The panel-wise error vectors $\nu_1,\ldots, \nu_T$ are i.i.d. with
mean zero and uniformly bounded eighth moments, that is
\[
\sup_{i,t} E|\nu_{it}|^8\le C_2, \quad
\text{for some positive constant } C_2.
\]
\end{assumption}
\begin{assumption}\label{A3}
There is a doubly infinite array of i.i.d. random variables $(z_{ij}), i,j\geq 1$, satisfying
\[E(z_{11})=0, \ E(z_{11}^2)=1, \ E(z_{11}^4) = \gamma_4 <\infty,\]
such that for each n, T, letting $Z_T=(z_{it} )_{1\leq i \leq n, 1\leq t \leq T}$, the idiosyncratic error vectors can be represented as $\nu_t = \Sigma_n^{1/2}z_j$ where $z_j=(z_{ij})_{1\leq i\leq n}$ denotes the $j$th column of $Z_T$.
\end{assumption}
\begin{assumption}\label{high_sigma}
The population spectral distribution $H_n$ of $\Sigma_n$ weakly converges to a probability distribution $H$, as $n\to \infty$, and the sequence of spectral norms $(||\Sigma_n||)$ is bounded.
\end{assumption}
Assumption \ref{A1} is standard in panel data models.
Under Assumption \ref{A2}, the matrix $\Sigma_n=\cov(\nu_t)$ is constant over time. Under Assumptions \ref{A1} and \ref{A2}, $\hat{\beta}$ is a consistent estimator of $\beta$. Based on Assumption \ref{A3}, the sample error covariance matrix can be represented as $ S_T =T^{-1}\sum_{t=1}^T \nu_t \nu_t^\prime= T^{-1}\Sigma_n^{1/2}Z_TZ_T^\prime \Sigma_n^{1/2}$, which is motivated by the random matrix theory, and it is generic enough for a precise analysis of the sphericity test. Under the null, set $\sigma_{\nu}^2=1$, then $\Sigma_n=\mathbf{I}_n$ and thus $\nu_{it}=z_{it}$.
The following theorem establishes the asymptotic normality of $\hat{U}$ under the LPA scheme. \begin{theorem2}\label{theorem_high}
Suppose Assumptions 1-4 hold for the panel data model (\ref{model}). Under $H_0$ and the LAP scheme, when $(n,T)\to \infty$, $c_T=n/T\to c \in(0, \infty),$ we have
\begin{eqnarray}\label{null_gene}
T\hat{U}-n - \left(\hat{\gamma}_4+c_T-2\right) \overset{\mathcal{D}}{\longrightarrow} N(0,4),
\end{eqnarray}
where $\hat{\gamma}_4$ is the fourth moment of the residuals, i.e. $\hat{\gamma}_4=\frac{1}{nT}\sum_{i=1}^n \sum_{t=1}^T |\hat{\nu}_{it}|^4$, and $\hat{\gamma}_4=\gamma_4+o_p(1)$ as in the following Proposition \ref{diff_between_U}.
\end{theorem2}
Since $\hat{\gamma}_4=3+o_p(1)$ for normally distributed disturbances, this result reduces to the result (\ref{b2011}) for the normal case. In addition, if the errors have a non-normal distribution but have the same first fourth moments as the normal distribution, we have (\ref{b2011}) again as in the normal case. Therefore, Theorem \ref{theorem_high} constitutes a natural extension of the null distribution of $\hat{U}$ from Gaussian context to general population ones. The test based on the asymptotic normal distribution given in the equation (\ref{null_gene}) will be referred to as the {\em general residual-based John's (GRJ) test}.
Theorem \ref{theorem_high} is obtained by the combination of the result (\ref{Wangyao}), a general CLT for $U$ of the sample error covariance matrix of (unobserved) disturbances $\{\nu_{it}\}$, and the difference between $U$ and $\hat{U}$ of the sample residual covariance matrix from residuals $\{\hat{\nu}_{it} \}$ in the following proposition.
\begin{proposition}\label{diff_between_U}
With the same assumptions as in Theorem \ref{theorem_high}, we have
\begin{eqnarray}
T(\hat{U}-U) - c_T =O_p\left(\frac{1}{n} \right),\label{prop1}\\
\hat{\gamma_4}-\gamma_4 = O_p\left(\frac{1}{T} \right).\label{prop2}
\end{eqnarray}
\end{proposition}
The proof of (\ref{prop1}) is similar to that of Proposition 4.3 in \cite{baltagi2011testing} for normal disturbances but is recalled in the Appendix together with the proof of (\ref{prop2}). This proposition implies that for the residual-based John's test statistic $\hat{U}$, the additional noise contained in the within residuals $\{\hat{\nu}_{it} \}$ will accumulate into a constant $c_T$, and the fourth moment $\hat{\gamma}_4$ of within residuals is a consistent estimator of the fourth moment $\gamma_4$ of the disturbances.
The following theorem establishes the asymptotic normality of $\hat{U}$ under the ULPA scheme.
\begin{theorem2}\label{ultra_null}
Suppose Assumptions 1-4 hold for the panel data model (\ref{model}). Under $H_0$ and the ULPA scheme, when $(n, T)\to \infty, c_T=n/T\to \infty, T^\delta/n=O_p(1), 1<\delta<2$, we have
\begin{eqnarray}
T\hat{U}-n-(\hat{\gamma}_4+c_T-2) \overset{\mathcal{D}}{\longrightarrow} N(0,4).
\end{eqnarray}
\end{theorem2}
Theorem \ref{ultra_null} is obtained by the combination of the result (\ref{liyao}), a general CLT for $U$ of the sample error covariance matrix under ULPA, and the difference between $U$ and $\hat{U}$ of the sample residual covariance matrix. The proof is skipped as it is the same as that of Theorem \ref{null_gene}, i.e., the results under the LPA scheme in Proposition~\ref{diff_between_U} still hold under the ULPA scheme.
Theorem \ref{theorem_high} and Theorem \ref{ultra_null} together indicate the {\it dimension-proof} property of the GRJ test (or John's test for panel data model), i.e., regardless of normality, under any $(n,p)$-asymptotic, $n/T\to (0,\infty]$, has been testified. From the application perspective, where $n$ and $T$ are explicitly known, it becomes tough to decide which asymptotic scheme to use, namely, ``$n$ fixed, $T\to \infty$", ``$n/T\to c\in(0,\infty)$, $n,T\to \infty$", or ``$n/T\to \infty$, $n,T\to \infty$"etc. Therefore, the GRJ test possesses the powerful dimension-proof property, making it recommended over other procedures, as most test statistics constructed under the ``$n$ fixed, $T\to \infty$" asymptotic scheme would have different limiting distributions under ``$n/T\to c\in(0,\infty)$" and ``$n/T\to \infty$" asymptotic schemes.
\section{Asymptotic power}
In this section, we comprehensively study the GRJ test's asymptotic power. This is done under the alternative hypothesis of the factor model~(\ref{factor_model}) with the following three different settings:
\begin{itemize}
\item[S1.] weak factors with a fixed number of factors;
\item[S2.] weak factors with a divergent number of factors;
\item[S3.] neither weak nor strong factors with a fixed number of factors.
\end{itemize}
The setting S1 has been studied by \cite{baltagi2017asymptotic} for normal-distributed disturbances. Without normality assumption, the constraints for the factor model (\ref{factor_model}) keep the same except that
\begin{itemize}
\item[(c')] The idiosyncratic errors $\{\epsilon_{it} \}$ are i.i.d. $(0, \sigma_{\epsilon}^2)$, and independent of all factors.
\end{itemize}
Under setting S1, the factors are weak and the number of factors is fixed, so the spectral norm of the covariance matrix~(\ref{Sigma_factor}) is bounded, while its spectral norm is unbounded under the other two settings, S2 and S3. In the literature of random matrix theory, the asymptotic results of linear spectral statistics of covariance matrix with unbounded spectral norm are different from that of covariance matrix with bounded spectral norm. Therefore, the setting S1 and settings S2-S3 will be studied, respectively, in the following two subsections.
\subsection{Covariance matrix with bounded spectral norm}
We first study the asymptotic power of the GRJ test under the alternative with setting S1 under the LPA scheme.
\begin{theorem2}\label{high_weak}
Suppose Assumptions 1, 3 and 4 hold for the panel data model (\ref{model}), and the disturbances follow the factor model (\ref{factor_model}) with constraints (a), (b) and (c'). Under the LAP scheme and the weak factor alternative S1 with $h_j\to d_j\in (0,\infty)$, $j=1,\ldots,r$, when $(n, T)\to \infty$, $c_T=n/T\to c\in (0, \infty)$, we have
\begin{eqnarray}
T\hat{U} -n - \left( \hat{\gamma}_4+c_T-2 \right) - \frac{1}{c_T}\sum_{j=1}^r h_j^2 \overset{\mathcal{D}}{\longrightarrow} N(0,4).
\end{eqnarray}
\end{theorem2}
Theorem \ref{high_weak} implies that under the weak factor alternative with a fixed number of factors, the limiting distribution of the residual-based John's test statistic is just a mean shift of its null distribution (\ref{null_gene}). When the disturbances are normal or $\gamma_4\to 3$, this result reduces to (\ref{alter_dist_weak}) for normal population. From Theorem \ref{theorem_high} and Theorem \ref{high_weak}, it is easy to obtain the power function of the GRJ test under the LPA scheme, that is,
\begin{eqnarray}\label{eqn_power_weak}
P_1 (H_1) = 1-\Phi \left(Z_\alpha -\frac{1}{2c_T} \sum_{j=1}^r h_j^2 \right),
\end{eqnarray}
which coincides with the result for the normal case. $P_1(H_1)\geq \alpha$, and the power of the test increases as spikes $\{h_j\}$ become large but decreases as the ratio $c_T$ of $n$ and $T$ increases. Therefore, under the LAP scheme, the GRJ test is inconsistent under setting S1 regardless of the distribution of the disturbances.
Theorem \ref{high_weak} is obtained by combining the limiting distribution of $U$ and the difference between $U$ and $\hat{U}$ in the following propositions.
\begin{proposition}\label{U_weak_alter}
With the same assumptions as in Theorem \ref{high_weak}, the asymptotic normality of $U$ is
\begin{eqnarray*}
T{U} -n - \left( \frac{1}{c_T}\sum_{j=1}^r h_j^2 -{\gamma}_4+2 \right) \overset{\mathcal{D}}{\longrightarrow} N(0,4).
\end{eqnarray*}
\end{proposition}
\begin{proposition}\label{diff_between_U_weak}
With the same assumptions as in Theorem \ref{high_weak}, we have
\begin{eqnarray}\label{prop3}
T(\hat{U}-U) - c_T =O_p\left(\frac{1}{n} \right),\\
\hat{\gamma_4}-\gamma_4 = O_p\left(\frac{1}{T} \right).
\end{eqnarray}
\end{proposition}
Next, we study the asymptotic power of the GRJ test under the alternative with setting S1 under the ULPA scheme.
\begin{theorem2}\label{ultra_high_weak}
Suppose Assumptions 1,3 and 4 hold for the panel data model (\ref{model}), and the disturbances follow the factor model (\ref{factor_model}) with constraints (a), (b) and (c'). In addition to Assumption \ref{high_sigma} of $\Sigma_n$, assume the following limits exits,
\begin{itemize}
\item[(a)] $\eta_1 = \lim_{n\to \infty} \frac{1}{n} tr (\Sigma_n)$,
\item[(b)] $\eta_2= \lim_{n\to \infty} \frac{1}{n} tr(\Sigma_n^2)$,
\item[(c)] $\eta_3 = \lim_{n\to \infty} \frac{1}{n}\sum_{i=1}^n (\Sigma_{ii})^2$, where $\Sigma_{ii}$ denotes the $i$-th diagonal entry of $\Sigma_n$.
\end{itemize}
Under the ULAP scheme and the weak factor alternative S1 with $h_j\to d_j\in (0,\infty)$, $j=1,\ldots,r$, when $(n, T)\to \infty$, we have
\begin{eqnarray}
T\hat{U} -n - \left(\frac{\eta_2}{\eta_1^2} -1\right) T - \frac{\eta_2+\eta_3(\hat{\gamma}_4-3)}{\eta_1^2} -c_T \overset{\mathcal{D}}{\longrightarrow} N\left(0,\frac{4\eta_2^2}{\eta_1^4}\right).
\end{eqnarray}
\end{theorem2}
Note that if let $\Sigma_n = \sigma_{\nu}^2\mathbf{I}_n$, then $\eta_1=\sigma_{\nu}^2, \eta_2=\eta_3=\sigma_{\nu}^4$, this result reduces to the null distribution of John's test statistic under the ULPA scheme in Theorem \ref{ultra_null}.
From Theorem \ref{ultra_null} and Theorem \ref{ultra_high_weak}, it is easy to obtain the power function of the GRJ test under the ULPA scheme, that is,
\begin{eqnarray*}
P_2(H_1) \sim 1- \Phi \left(\frac{\eta_1^2}{\eta_2}Z_\alpha +\frac{\eta_1^2(\hat{\gamma}_4-2)-\eta_2-\eta_3(\hat{\gamma}_4-3)}{2\eta_2} +\frac{(\eta_1^2-\eta_2)T}{2\eta_2} \right) \to 1,
\end{eqnarray*}
as $\eta_1^2<\eta_2$. Therefore, under setting S1, the GRJ test is consistent under the ULPA scheme.
Theorem \ref{ultra_high_weak} is obtained by combining the limiting distribution of $U$ and the difference between $U$ and $\hat{U}$ in the following propositions.
\begin{proposition}\label{U_ulpa}
With the same assumptions as in Theorem \ref{ultra_high_weak}, we have
\begin{eqnarray}
TU-n -\left(\frac{\eta_2}{\eta_1^2} -1\right) T -\frac{\eta_2+\eta_3(\gamma_4-3)}{\eta_1^2} \overset{\mathcal{D}}{\longrightarrow} N\left( 0, \frac{4\eta_2^2}{\eta_1^4} \right).
\end{eqnarray}
\end{proposition}
\begin{proposition}\label{conjection_general_ulpa}
With the same assumptions as in Theorem \ref{ultra_high_weak}, when $(n,T)\to \infty, n/T=c_T\to c \in (0,\infty)$, we have
\begin{eqnarray}
T(\hat{U}-U) - c_T =O_p\left(\frac{1}{n} \right),\\
\hat{\gamma_4}-\gamma_4 = O_p\left(\frac{1}{T} \right).
\end{eqnarray}
\end{proposition}
In conclusion, under Setting S1, the GRJ test is inconsistent under the LPA scheme where $n/T\to c\in(0, \infty)$, but consistent under the ULPA scheme where $n/T \to \infty$.
\begin{remark}
Based on the above power analysis under the weak factor alternative with the LPA scheme, we can investigate the power behavior of the GRJ test under a general form of covariance matrix with bounded spectral norm, which covers the weak factor case as a special case; see the Supplementary Material.
\end{remark}
\subsection{Covaraince matrix with unbounded spectral norm}
In this subsection, we study the asymptotic power of the GRJ test under the other two settings S2 and S3, where the spectral norm of the covariance matrix $\Sigma_n$ of disturbances is unbounded. To achieve this goal, we consider John's test as a way to test the null hypothesis $H_0$ against an alternative hypothesis of the form:
\[H_1^\ast: { spec}(\Sigma_n)= \underbrace{\{\lambda_1,\ldots\lambda_{r_n}\}}_{G_1}\cup\underbrace{\{\lambda_{r_n+1},\ldots,\lambda_n\}}_{G_2}, \]
where the number of factors $r_n$ in $G_1$ can keep fixed (i.e., $r_n=r$) or diverges as the dimensions $n, T$ increase.
\begin{assumption}\label{rn}
The eigenvalues in $G_1$ satisfy $\displaystyle \sup_i \max_{1\leq i \leq r_n} \lambda_i =O(n^\alpha)$ and the size of $
G_1$ is $r_n/n\to \tau$ for some $0\leq \alpha< 1$ and $0\leq \tau <1$.
\end{assumption}
In Assumption \ref{rn}, the eigenvalues (factors) are weak (or bounded) when $\alpha=0$, and the eigenvalues (factors) are neither weak nor strong when $0<\alpha<1$. The number of factors is fixed when $\tau=0$, otherwise divergent.
We first give the asymptotic Gaussian of $U$ under the alternative $H_1^\ast$.
\begin{theorem2}\label{J_H1}
Suppose that Assumption \ref{A3} with $E(|z_{11}|^6)<\infty$ and Assumption \ref{rn} hold when $4\alpha+\tau\leq 1$ or Assumption \ref{A3} with $E(|z_{11}|^{16})<\infty$ and Assumption \ref{rn} hold when $4\alpha+\tau>1$, then under $H_1^\ast$ and the LPA scheme, when $c_T=n/T \to c \in (0, \infty)$ as $n, T\to \infty$, we have
\begin{eqnarray}
T\left( U- \mu \right)/\sigma \stackrel{D}{\rightarrow} N(0, 1),
\end{eqnarray}
where
\[\mu= \frac{c_T \left\{(\gamma_4-3) tr(\Sigma_n \circ \Sigma_n)+(T+1)tr(\Sigma_n^2) \right\}}{(tr \Sigma_n)^2} +c_T-1, \]
and
\[\sigma^2 \sim \frac{1}{c_T^2}\left\{\frac{4 T^4 \theta_2^2}{\theta_1^6}\omega_{1}-\frac{4 T^4 \theta_2}{\theta_1^5}\omega_{2} +\frac{T^4}{\theta_1^4}\omega_{3} \right\}, \]
with
\begin{eqnarray*}
\theta_1 &=& tr \Sigma_n,\\
\theta_2 &=& T^{-1}\left\{(\gamma_4-3) tr(\Sigma_n \circ \Sigma_n) +(tr \Sigma_n)^2 +(T+1)tr(\Sigma_n^2) \right\},\\
\omega_{1} &=& T^{-1}\left\{(\gamma_4-3) tr(\Sigma_n \circ \Sigma_n)+2(tr \Sigma_n)^2 \right\},\\
\omega_{2} &=& T^{-2}\left\{4 tr \Sigma_n^2 tr \Sigma_n + 2 (\gamma_4-3) tr(\Sigma_n \circ \Sigma_n)tr \Sigma_n +2T(\gamma_4-3) tr(\Sigma_n \circ \Sigma_n^2)+4Ttr \Sigma_n^3 \right\},\\
\omega_{3} &=& T^{-3} \Big\{8 tr \Sigma_n^2 (tr \Sigma_n)^2 + 4 (\gamma_4-3) (tr \Sigma_n)^2tr(\Sigma_n \circ \Sigma_n) + 16 T tr \Sigma_n tr \Sigma_n^3\\
&& \quad \quad + 4 T (tr \Sigma_n^2)^2 + 8 T (\gamma_4-3) tr(\Sigma_n \circ \Sigma_n^2) tr \Sigma_n + 4 T^2 (\gamma_4-3) tr(\Sigma_n^2 \circ \Sigma_n^2) + 8 T^2 tr \Sigma_n^4 \Big\}.
\end{eqnarray*}
\end{theorem2}
Theorem \ref{J_H1} implies that the limiting distribution of $U$ under $H_1^\ast$ relies on the magnitudes of the terms involving $\Sigma_n$. To illustrate the consistency of John's test, we analyze the power functions under a simplified factor model, where the covariance matrix $\Sigma_n$ is diagonal,
\[\Sigma_n=diag(\lambda_1, \ldots, \lambda_{r_n}, \lambda_{r_n+1}, \ldots, \lambda_p). \]
\begin{itemize}
\item Under setting S2: the eigenvalues in $G_1$ are weak (bounded), but the size of $G_1$ (the number of factors) is divergent, i.e., $\alpha=0, 0<\tau<1$, all trace terms involving $\Sigma_n$, e.g. $tr \Sigma_n$, $tr \Sigma_n^2$, $tr(\Sigma_n \circ \Sigma_n) (=tr \Sigma_n^2)$, $tr \Sigma_n^3$ and $tr \Sigma_n^4$, are an order of $n$. Then it is easy to verify that
\[\mu = c_T-1+\frac{B_2}{B_1^2}\(1+\frac{1+\nu_4}{T} \), \ \textrm{and} \ \sigma^2=O(n), \]
where $B_1=\frac{1}{n}tr\Sigma_n$ and $B_2=\frac{1}{n}tr\Sigma_n^2$. The power of John's test under this setting is
\[P_1(H_1^\ast)=1-\Phi\(\frac{2Z_\alpha+(1+\nu_4)(1-\tau^{-1})+n(1-\frac{B_2}{B_1^2})}{\sigma}. \) \]
As $B_2>B_1^2$ since all eigenvalues can't be equal, $1-\frac{B_2}{B_1^2}<0$, $n(1-\frac{B_2}{B_1^2})/\sigma\to -\infty$ as $n\to \infty$, thus the power function converges to 1. Therefore, John's test is consistent.
\item Under setting S3: the eigenvalues in $G_1$ are divergent, but the size of $G_1$ is fixed, i.e., $\alpha>0, \tau=0$, set $r_n=r$, we have $tr \Sigma_n=O( r n^{\alpha})$, $tr \Sigma_n^2=tr(\Sigma_n \circ \Sigma_n)=O(r n^{2\alpha})$, $tr \Sigma_n^3=O(r n^{3\alpha})$ and $tr \Sigma_n^4=O(r n^{4\alpha})$. Then it is easy to verify that
\[ \mu=c_1 n, \ \textrm{and} \ \sigma^2=c_2 n^3, \]
where $c_1$ and $c_2$ are some positive constants. The power of John's test is
\[P_2(H_1^\ast)=1-\Phi\left( \frac{2 Z_{\alpha}+n+(\gamma_4-2)-T\mu}{\sigma}\right). \]
As $\frac{-T\mu}{\sigma}=-O(\sqrt{n})$, this power function converges to 1 as $n\to \infty$.
Therefore, John's test is consistent.
\end{itemize}
Therefore, John's test for raw data is consistent under the alternative $H_1^\ast$.
Next, we examine the consistency of the GRJ test.
\begin{theorem2}\label{divergent_factors}
Suppose Assumptions \ref{A1}, \ref{A3} and \ref{rn} hold for the panel data model \eqref{model} and the disturbances $\{\nu_{it}\}$ follow the factor model \eqref{factor_model} with constraints (a), (b) and (c'). Under the LPA scheme and the alternative $H_1^\ast$ with a divergent number of weak factors, the GRJ test is consistent.
\end{theorem2}
Theorem \ref{divergent_factors} is obtained by combining Theorem \ref{J_H1} and the following proposition.
\begin{proposition}\label{diff_between_U_diverge_factor}
With the same assumptions as in Theorem \ref{divergent_factors},
\begin{equation*}
T(\hat{U}-U)=c_3c_T+c_4\tau+O_p\left(\frac{1}{\sqrt{n}}\right),
\end{equation*}
where $c_3$ and $c_4$ are some constants.
\end{proposition}
Proposition \ref{diff_between_U_diverge_factor} implies that under the weak factor with divergent number alternative, additional noise obtained in $T\hat{U}-TU$ is $O_p(1)$. This magnitude is smaller than $O_p(T)$, the magnitude of $TU$. Thus, $T\hat{U}-TU$ is asymptotically dominated by $TU$, leading to the GRJ test's consistency.
\begin{theorem2}\label{not_weak}
Suppose that for the panel data model \eqref{model}, Assumption \ref{A1}, Assumption \ref{A3} with $E(|z_{11}|^6)<\infty$ and Assumption \ref{rn} hold when $4\alpha+\tau\leq 1$, or Assumption \ref{A1}, Assumption \ref{A3} with $E(|z_{11}|^{16})<\infty$ and Assumption \ref{rn} hold when $4\alpha+\tau>1$, the disturbances $\{\nu_{it}\}$ follow the factor model \eqref{factor_model} with constraints (a), (b) and (c'). Under the LPA scheme and the alternative $H_1^\ast$ with a fixed number of neither strong nor weak factors, the GRJ test is consistent.
\end{theorem2}
Theorem \ref{not_weak} is obtained by combining Theorem \ref{J_H1} and the following proposition.
\begin{proposition}\label{diff_between_U_not_weak}
With the same assumptions as in Theorem \ref{not_weak},
\begin{equation*}
TU=O_p\left(Tn^{2\alpha-1}\right)+O_p(T).
\end{equation*}
When $0<\alpha<\frac{1}{2}$,
\begin{equation*}
T(\hat{U}-U)=c_T+O_p\left(\frac{1}{\sqrt{T}}\right),
\end{equation*}
When $\frac{1}{2}\leq\alpha<1$,
\begin{equation*}
T(\hat{U}-U)=O_p\left(n^{2\alpha-1}\right).
\end{equation*}
\end{proposition}
Proposition \ref{diff_between_U_not_weak} implies that under neither strong nor weak factor model, the magnitude of the additional noise contained in $T(\hat{U}-U)$ is smaller than $O_p(Tn^{2\alpha-1})+O_p(T)$, the magnitude of $TU$. Thus $T\hat{U}=TU+T(\hat{U}-U)$ is asymptotically dominated by $TU$, leading to the consistency of the GRJ test.
\begin{remark}
In Theorem \ref{J_H1}, the limiting result still holds when $\alpha=1$, which corresponds to the strong factor case. Thus, together with the results under the strong factor alternative in Theorems 4 \& 5 of \cite{baltagi2017asymptotic}, we can conclude that the GRJ test is consistent under the strong factor alternative regardless of the distribution of disturbances.
\end{remark}
\begin{remark}
In Theorem \ref{J_H1}, the factors and the number of factors can divergent simultaneously with $(n, T)$, i.e., $\alpha>0$ and $\tau>0$. We call it the fourth setting: neither weak nor strong factors with a divergent number of factors. We believe that the GRJ test is consistent under this setting, although the proof is omitted.
\end{remark}
So far, we have conducted the power analysis of the GRJ test under the alternative of unbounded-norm covariance with the LPA scheme. The corresponding power analysis for the ULPA scheme is missing as there is no asymptotic result for linear spectral statistics of unbounded-norm covariance under the ULPA scheme. But based on the above power analyses, it is natural to conjecture that the GRJ test is still consistent under the ULPA scheme as it is consistent under both bounded-norm covariance with the scheme and unbounded-norm covariance with the LPA scheme.
\section{Conclusion}
This paper studies John's test for sphericity of the covariance matrix in a large panel data model with general populations of disturbances. The limiting distributions of the residual-based John's test are derived under both LPA and ULPA schemes. The results show that John's test possesses the powerful dimension-proof property with different $(n,T)$- asymptotic, making it recommended over other procedures. Moreover, the asymptotic power of John's test is conducted comprehensively under the factor model alternative with different settings. John's test is always consistent under the ULPA scheme, divergent number of weak factors and fixed number of non-weak factors, except under the weak factor alternative with a fixed number of factors under the LPA scheme.
However, when the cross-sectional units $n$ increases sufficiently fast with time-series size $T$, say $n$ has an order of $T^2$ or larger, the limiting distribution of the residual-based John's test statistic $\hat{U}$ is unknown, which goes beyond the scope of the existing literature including this paper, and will require developing new mathematical techniques worthing much research effect in the future.
\section*{Supplementary Material}
The Supplementary Material contains the power analysis of the GRJ test under the general form of covariance matrix with bounded spectral norm under the LPA scheme.
\section*{Acknowledgments}
The author thanks Jianfeng Yao and Weiming Li for their helpful comments on the contents. Li's research was supported by NSFC grant 11901492.
\bibliographystyle{plainnat}
\section{Asymptotic power under general form of covariance}
Based on the power analysis of the GRJ test under the weak factor alternative in Section 4.1, i.e., the disturbances follow a factor model with fixed number of weak factors, we continue to explore the power behavior of the GRJ test under general form of covariance with bouned spectral norm $\Sigma_n\neq \sigma_{\nu}^2 \mathbf{I}_n$ which covers the weak factor alternative as a special case. We first develop the limiting distribution of $U$ for general covariance matrix and investigate the asymptotic power of John's test for raw data $\{\nu_{it} \}$. We will write $a_n\sim b_n$ for positive numbers $a_n$ and $b_n$ if $a_n/b_n\to 1$ as $n\to \infty$.
\begin{lemma}\label{high_alter}
Suppose the moments of $H_n$, the ESD of covariance matrix ${\Sigma}_n$, converge to the corresponding moments of $H$ defined as,
\[\theta_i = \int t^{i} d H(t) \quad \textrm{for} \ i=1,2,3,4 . \]
Define the first and second moments of the finite counterpart of limiting spectral distribution $S_T$ $F^{c_T, H_n}$ of the sample error covariance matrix $S_T$ as,
\[\vartheta_1= \int x d F^{c_T, H_n},\quad \vartheta_2= \int x^2 d F^{c_T, H_n}. \]
Suppose Assumptions 1, 2 and 4 hold for the fixed effects panel data model (1.1). Under the alternative $H_1$ and the LPA scheme, when$(n,T)\to \infty$, $c_T=n/T\to c\in(0,\infty)$, we have
\begin{enumerate}
\item If in addition $\gamma_4=3$, then
\begin{eqnarray}
s_1^{-1} \left[TU-T\left(\frac{\vartheta_2+\theta_2}{\vartheta_1^{2}}-1\right)\right] \overset{\mathcal{D}}{\longrightarrow} N(0,1),
\end{eqnarray}
where
\begin{eqnarray*}
s_1^2 &\sim & \frac{8\theta_4/c+4\theta_2^2+8c \theta_1^2 \theta_2+16\theta_1\theta_3}{\vartheta_1^4} -\frac{\left(16\theta_3/c+16\theta_1\theta_2\right)(\vartheta_2+\theta_2)}{\vartheta_1^5}\\
&& \hspace{1.5cm}+\frac{8\theta_2(\vartheta_2+\theta_2)^2}{c \vartheta_1^6}.
\end{eqnarray*}
\item If $\Sigma_n$ is diagonal for all $n$ large, then
\begin{eqnarray}
s_2^{-1}\left[TU -T \left(\frac{\vartheta_2+\theta_2+\hat{\gamma}_4-3}{\vartheta_1^2}-1 \right) \right] \overset{\mathcal{D}}{\longrightarrow} N(0,1),
\end{eqnarray}
where
\begin{eqnarray*}
s_2^2 &\sim & (\gamma_4-1)\Bigg( \frac{4\theta_4/c+2\theta_2^2+4 c \theta_1^2\theta_2+8\theta_1\theta_3}{\vartheta_1^4}+\frac{4\theta_2(\vartheta_2+\theta_2+\gamma_4-3)^2}{c \vartheta_1^6} \\
&& \hspace{3cm}-\frac{\left(8\theta_3/c+8\theta_1\theta_2\right)(\vartheta_2+\theta_2+\gamma_4-3)}{\vartheta_1^5} \Bigg) .
\end{eqnarray*}
\end{enumerate}
\end{lemma}
When neither $\gamma_4=3$ nor $\Sigma_n$ is diagonal, the limiting distribution will become more complex, depending on the eigenvectors of $\Sigma_n$, see \cite{pan2008central}. From the above theorem, we know that the limiting distribution of John's test statistic $U$ depends on the moments of limiting spectral distribution of $\Sigma_n$ and the first two moments of $F^{c_T, H_n}$.
\begin{corollary}\label{high_power_general}
With the same assumptions as in Lemma \ref{high_alter}, when $(n,T)\to \infty, n/T=c_T\to c \in (0,\infty)$, the power of John's test under the alternative of general covariance $\Sigma_n$ is
\item If in addition $\gamma_4=3$, then
\begin{eqnarray*}
P_1(H_1) \sim 1-\Phi \left(\frac{2 Z_\alpha +1+T\left(c+1 - \frac{\vartheta_2+\theta_2}{\vartheta_1^2} \right)}{s_1} \right),
\end{eqnarray*}
where $\alpha$ is the nominal test level, and $Z_\alpha, \Phi(\cdot)$ are the alpha upper quantile and cdf of standard normal distribution, respectively.
\item If $\Sigma_n$ is diagonal for all $n$ large, then
\begin{eqnarray*}
P_2(H_1) \sim 1-\Phi \left(\frac{2 Z_\alpha +\hat{\gamma}_4-2+T\left(c+1 - \frac{\vartheta_2+\theta_2 +\gamma_4-3}{\vartheta_1^2} \right)}{s_2} \right).
\end{eqnarray*}
\end{corollary}
Clearly, the consistency of John's test depends on the magnitude of $c+1$ compared to that of $\frac{\vartheta_2+\theta_2+\gamma_4-3}{\vartheta_1^2}$. John's test is consistent when $c+1<\frac{\vartheta_2+\theta_2+\gamma_4-3}{\vartheta_1^2}$, otherwise inconsistent.
\begin{conjecture}\label{conjection_general}
With the same assumptions as in Lemma \ref{high_alter}, when $(n,T)\to \infty, n/T=c_T\to c \in (0,\infty)$, the difference between $\hat{U}$ and $U$ is
\[T(\hat{U}-U )=O_p(1). \]
\end{conjecture}
It is natrual to get this conjecture since, as a special case of general bounded covariance matrix $\Sigma_n$, the difference $T(\hat{U}-U)$ is $O_p(1)$ under weak factor alternative.
Therefore, combining the above results, we can conclude that the consistency of the GRJ test is not guaranteed when the covariance matrix $\Sigma_n$ has a bounded norm under the LPA scheme.
\section{Proof of Lemma \ref{high_alter}}
Following the calculations in \cite{tian2015robust}, when $\gamma_4=3$, we get the following CLT.
\begin{proposition}
Let $\{\ell_i \}_{1\leq i\leq n}$ be the eigenvalues of the sample error covariance matrix $S_T$, under the conditions of Lemma \ref{high_alter}, if $\gamma_4=3$, we have
\begin{eqnarray*}
T \left(\begin{array}{l}
n^{-1}\sum_{i=1}^n \ell_i^2 -(\vartheta_2+\theta_2) \\ n^{-1}\sum_{i=1}^n \ell_i -\vartheta_1
\end{array} \right) \overset{\mathcal{D}}{\longrightarrow} N\left( \left(\begin{array}{c} 0 \\0
\end{array} \right), \mathbf{V_1} \right),
\end{eqnarray*}
with
\begin{eqnarray*}
\mathbf{V_1} = \left(\begin{array}{cc}
8 \theta_4/c+4\theta_2^2+8c \theta_1^2 \theta_2 +16\theta_1\theta_3 & 4\theta_3/c +4\theta_1\theta_2 \\ 4\theta_3/c +4\theta_1\theta_2 & 2\theta_2/c
\end{array} \right).
\end{eqnarray*}
\end{proposition}
Define the function $f(x,y) = \frac{x}{y^2}-1$, then $U= f\left(n^{-1} \sum_{i=1}^n \ell_i^2, n^{-1} \sum_{i=1}^n \ell_i \right)$. We have
\begin{eqnarray*}
&&\frac{\partial f}{\partial x}\left(\vartheta_2+\theta_2, \vartheta_1 \right) = \frac{1}{\vartheta_1^2},\\
&&\frac{\partial f}{\partial y}\left(\vartheta_2+\theta_2, \vartheta_1 \right) = \frac{-2(\vartheta_2+\theta_2)}{\vartheta_1^3},\\
&& f \left(\vartheta_2+\theta_2, \vartheta_1 \right) = \frac{\vartheta_2+\theta_2}{\vartheta_1^2}-1.
\end{eqnarray*}
By the delta method,
\begin{eqnarray*}
T \left(U-f\left(\vartheta_2+\theta_2, \vartheta_1 \right) \right) \overset{\mathcal{D}}{\longrightarrow} N(0,s_1^2),
\end{eqnarray*}
where
\begin{eqnarray*}
s_1^2 = \lim \left\{ \left( \begin{array}{c}
\frac{\partial f}{\partial x}\left(\vartheta_2+\theta_2, \vartheta_1 \right) \\ \frac{\partial f}{\partial y}\left(\vartheta_2+\theta_2, \vartheta_1 \right)
\end{array}
\right)^T \cdot \mathbf{V}_1 \cdot \left( \begin{array}{c}
\frac{\partial f}{\partial x}\left(\vartheta_2+\theta_2, \vartheta_1 \right) \\ \frac{\partial f}{\partial y}\left(\vartheta_2+\theta_2, \vartheta_1 \right)
\end{array}
\right) \right\}.
\end{eqnarray*}
For the second result, the central limit theorem in \cite{pan2008central} is applied to functions $f(x)=x^2$ and $g(x)=x$.
\begin{proposition}
Let $\{\ell_i \}_{1\leq i\leq n}$ be the eigenvalues of the sample covariance matrix $S_T$, under the conditions of Theorem \ref{high_alter}, if $\Sigma_n$ is diagonal for all $n$ large, we have
\begin{eqnarray*}
T \left(\begin{array}{l}
n^{-1}\sum_{i=1}^n \ell_i^2 -(\vartheta_2+\theta_2+\gamma_4-3) \\ n^{-1}\sum_{i=1}^n \ell_i -\vartheta_1
\end{array} \right) \overset{\mathcal{D}}{\longrightarrow} N\left( \left(\begin{array}{c} 0 \\0
\end{array} \right), \mathbf{V_2} \right),
\end{eqnarray*}
with
\begin{eqnarray*}
\mathbf{V_2} = (\gamma_4-1)\left(\begin{array}{cc}
8 \theta_4/c+4\theta_2^2+8c \theta_1^2 \theta_2 +16\theta_1\theta_3 & 4\theta_3/c +4\theta_1\theta_2 \\ 4\theta_3/c +4\theta_1\theta_2 & 2\theta_2/c
\end{array} \right).
\end{eqnarray*}
\end{proposition}
Similarly, we can get,
\begin{eqnarray*}
T \left(U-f\left(\vartheta_2+\theta_2+\gamma_4-3, \vartheta_1 \right) \right) \overset{\mathcal{D}}{\longrightarrow} N(0,s_2^2),
\end{eqnarray*}
where
\begin{eqnarray*}
s_1^2 = \lim \left\{ \left( \begin{array}{c}
\frac{\partial f}{\partial x}\left(\vartheta_2+\theta_2+\gamma_4-3, \vartheta_1 \right) \\ \frac{\partial f}{\partial y}\left(\vartheta_2+\theta_2+\gamma_4-3, \vartheta_1 \right)
\end{array}
\right)^T \cdot \mathbf{V}_2 \cdot \left( \begin{array}{c}
\frac{\partial f}{\partial x}\left(\vartheta_2+\theta_2+\gamma_4-3, \vartheta_1 \right) \\ \frac{\partial f}{\partial y}\left(\vartheta_2+\theta_2+\gamma_4-3, \vartheta_1 \right)
\end{array}
\right) \right\}.
\end{eqnarray*}
The proof of Lemma \ref{high_alter} is complete.
\bibliographystyle{plainnat}
|
\section{Introduction}
It was conjectured that every continuous action on the circle by a higher rank lattice must
factor through a finite group action (the so called $1$-dimensional Zimmer's rigidity conjecture). Burger-Monod \cite{Bu} and Ghys \cite{Gh99}
proved independently the existence of finite orbits for higher rank lattice actions on the circle,
which translates the conjecture into the equivalent form: no higher rank lattice admits a total ordering which is invariant by left translations.
The latter was answered affirmatively by Witte-Morris for arithmetic groups of higher $\mathbb{Q}$-rank \cite{Wi2} and by Deroin-Hurtado
for general higher rank lattices \cite{DH}. Recently there has been a great progress on the Zimmer program for smooth higher
rank lattice actions on manifolds with dimensions $\geq 2$. We do not plan to list all the related results here and just suggest the readers to consult
\cite{Br, Fi} for the surveys.
\medskip
There has been intensively studied around group actions on dendrites very recently. One motivation is that
dendrites can appeare as the limit sets of some Klein groups, the structures of which
are closely related to the geometric properties of $3$-dimensional hyperbolic manifolds (see e.g. \cite{Bo, Mi}). Also,
the compactifications of the Cayley graphs of free groups are dendrites, which are important for understanding the algebraic
properties of free groups. Group actions on the circle have been
systematically investigated during the past few decades \cite{Gh01, Na1}. However, group actions on general curves
lack of the same depth of understanding. Dendrites and the circle lie on two
opposite ends of Peano curves in topologies. So, studying group actions on dendrites is the starting point for better understanding group actions
on curves or continua with higher dimensions. Some people studied the Ghys-Margulis' alternative for dendrite homeomorphism groups (\cite{Shi, Mal, DM2}).
One may consult \cite{AN, GM, SY} for the discussions around the structures of minimal sets for group actions on dendrites. The algebraic structures of dendrite
homeomorphism groups were investigated in \cite{DM1}
\medskip
Duchesne-Monod proved the existence of finite orbits for
higher rank lattice actions on dendrites (\cite{DM2}). This motivates us to consider further
the structures of higher rank lattice actions on dendrites. Though there are examples indicating that the exact analogy
to the rigidity results by Witte-Morris and Deroin-Hurtado mentioned above do not hold anymore for higher rank lattice actions on dendrites,
the authors showed in \cite{SX} that such actions are very restrictive. The aim of the paper is to
give a more detailed description of the structures of higher rank lattice actions on dendrites. Explicitly, we obtained
the following theorem.
\begin{thm}\label{main theorem 1}
Let $\Gamma$ be a higher rank lattice acting on a nondegenerate dendrite $X$ with no infinite order points. Then
there exists a nondegenerate subdendrite $Y$ which is $\Gamma$-invariant and satisfies the following items:
(1) There is an inverse system of finite actions $\{(Y_i, \Gamma):i=1,2,3,\cdots\}$ with monotone bonding maps $\phi_i: Y_{i+1}\rightarrow Y_i$ and with each $Y_i$ being a dendrite, such that $(Y, \Gamma|Y)$ is topologically conjugate to the inverse limit $(\underset{\longleftarrow}{\lim}(Y_i, \Gamma), \Gamma)$.
(2) The first point map $r:X\rightarrow Y$ is a factor map from $(X, \Gamma)$ to $(Y, \Gamma|Y)$; if $x\in X\setminus Y$, then $r(x)$ is an end point of $Y$ with infinite orbit; for each $y\in Y$, $r^{-1}(y)$ is contractible, that is there is a sequence $g_i\in \Gamma$ with ${\rm diam}(g_ir^{-1}(y))\rightarrow 0$..
\end{thm}
The proof of Theorem \ref{main theorem 1} relies on the non-left-orderability of higher rank lattices by Deroin-Hurtado, the elementarity
of higher rank lattice actions on dendrites by Duchesne-Monod, and a connection between semilinear-left-orderability of a group and its actions on dendrites
observed by the authors.
\medskip
We will not introduce the related notions and properties around group actions and dendrites. One may consult \cite{SX} for the details.
\section{General facts about semilinear preorders}
Recall that a binary relation $\preceq$ on a set $X$ is a {\it preorder} if it satisfies that
\medskip
\begin{itemize}
\item[(O1)] $x\preceq x$ for any $x\in X$;
\item[(O2)] if $x\preceq y$ and $y\preceq z$ then $x\preceq z$ for any $x,y,z\in X$.
\end{itemize}
\medskip
\noindent If, in addition, $\preceq$ satisfies that
\medskip
\begin{itemize}
\item[(O3)] if $x\preceq y$ and $y\preceq x$ then $x=y$ for any $x,y\in X$,
\end{itemize}
\medskip
\noindent then $\preceq$ is a {\it partial order} on $X$.
\medskip
A preorder (resp. partial order) on $X$ is called a {\it total preorder} ({resp. \it total order}) on $X$ if
\medskip
\begin{itemize}
\item[(O4)] any $x,y\in X$ are comparable; that is either $x\preceq y$ or $y\preceq x$.
\end{itemize}
\medskip
Let $G$ be a group with $e$ being the unit. A preorder/ partial order/ total order $\preceq$ on $G$ is said to be {\it left-invariant} if
\medskip
\begin{itemize}
\item[(O5)] for any $x,y\in X$ and $g\in G$, $gx\preceq gy$ whenever $x\preceq y$.
\end{itemize}
\medskip
Finally, we say $\preceq$ is a {\it semilinear left preorder} (resp. {\it semilinear left partial order}) if it satisfies (O1), (O2), (O5) (resp. (O1),(O2),(O3),(O5)) and
\medskip
\begin{itemize}
\item[(O6)] for any $x\in G$, any two elements of $\{ y\in G: y\preceq x\}$ are comparable;
\item[(O7)] for any $x,y\in G$, there is some $z\in G$ with $z\preceq x$ and $z\preceq y$.
\end{itemize}
\medskip
The following characterization of semilinear left preorder is similar to \cite[Theorem 1.1]{Kop}.
\begin{prop}\label{char of semi}
Let $\preceq$ be a semilinear left preorder on a group $G$. Then the positive cone $P:=\{g\in G: e\preceq g\}$ of $\preceq$ has the following properties:
\begin{itemize}
\item[(P1)] $P$ is a semigroup;
\item[(P2)] $P\cap P^{-1}$ is a subgroup of $G$;
\item[(P3)] $G=P^{-1}\cdot P$;
\item[(P4)] $P\cdot P^{-1}\subseteq P\cup P^{-1}$.
\end{itemize}
If $\preceq$ is further a partial order, then
\begin{itemize}
\item[(P2')] $P\cap P^{-1}=\{e\}$;
\end{itemize}
is further a total preorder, then
\begin{itemize}
\item[(P5)] $G=P\cup P^{-1}$.
\end{itemize}
Conversely, a subset $P$ of $G$ satisfying (P1)-(P4) (resp. (P1)(P2')(P3)(P4)) determines a semilinear left preorder (resp. semilinear left partial order) on $G$.
\end{prop}
\begin{proof}$(\Longrightarrow)$
(P1) and (P2) are direct from the definition of semilinear left preorder.
\medskip
To prove (P3), let $g\in G$. Take $u\in G$ with $u\preceq e$ and $u\preceq g$ by (O7). Then $g=u(u^{-1}g)$ and $e\preceq u^{-1}g$.
So, $G\subset P^{-1}\cdot P$. The inclusion $P^{-1}\cdot P\subset G$ is clear.
\medskip
To prove (P4), let $x, y\in P$. Then $x^{-1}\preceq e$ and $y^{-1}\preceq e$. From (O6), either $x^{-1}\preceq y^{-1}$ or
$y^{-1}\preceq x^{-1}$, which implies $xy^{-1}\in P\cup P^{-1}$.
\medskip
(P2') and (P5) are clear.
\medskip
$(\Longleftarrow)$ Define $x\preceq y$ if $x^{-1}y\in P$. We only check (O6) and (O7), the others are direct.
\medskip
For (O6), let $x, y, z\in G$ be such that $x\preceq z$ and $y\preceq z$. Then $x^{-1}z\in P$ and $z^{-1}y\in P^{-1}$.
By (P4), $x^{-1}y=x^{-1}z\cdot z^{-1}y\in P\cup P^{-1}$. This means either $x\preceq y$ or $y\preceq x$.
\medskip
For (O7), let $x, y\in G$. By (P3), we have $x^{-1}y=a\cdot b$, where $a\in P^{-1}$ and $b\in P$.
Then $a\preceq e$ and $a\preceq x^{-1}y$, which implies $xa\preceq x$ and $xa\preceq y$.
\end{proof}
According to Proposition \ref{char of semi}, we also say that a subsemigroup $P$ of $G$ satisfying (P1)-(P4) is a {\it semilinear left preorder} on $G$.
\begin{lem}\label{preorder}
Let $P$ be a semilinear left preorder on $G$.
\begin{itemize}
\item[(1)] For any $q\in P$, we have $q(P\cup P^{-1})q^{-1}\subset P\cup P^{-1}$.
\item[(2)] For any $g\in G$, we have $\{x\in G: x\preceq g\}=gP^{-1}$.
\item[(3)] Let $w$ be a nontrivial word composed by some elements of $P$. If $w$ is in $P\cap P^{-1}$ then every letter occurring in $w$ is in $P\cap P^{-1}$.
\end{itemize}
\end{lem}
\begin{proof}
(1) For any $x\in P$, we have $qxq^{-1}\in qPq^{-1}\subset P\cdot P^{-1}\subset P\cup P^{-1}$ by (P4). For any $x\in P^{-1}$, we have $xq^{-1}\in P^{-1}$. Thus $qP^{-1}q^{-1}\subset P\cdot P^{-1}\subset P\cup P^{-1}$.
\medskip
(2) If $x\preceq g$ then $g^{-1}x\in P^{-1}$ and hence $x=g(g^{-1}x)\in g P^{-1}$. For any $y\in P^{-1}$, we have $gy\preceq g$. Thus (2) holds.
\medskip
(3) Set $H=P\cap P^{-1}$ and $w=g_1\cdots g_n$ with $g_1,\cdots,g_n\in P$.
To the contrary, assume that $g_i\notin H$ for some $i\in\{1,\cdots,n\}$. Then $g_i\succ e$ and
\begin{eqnarray*}
w&=&g_1\cdots g_n\succeq g_1\cdots g_{n-1}\succeq\cdots\succeq g_1\cdots g_{i}\\
&\succ& g_1\cdots g_{i-1}\succeq e.
\end{eqnarray*}
This contradicts that $w$ is in $H$. Hence each $g_i$ is in $H$.
\end{proof}
\section{Left orderability and semilinear left preorders}
\begin{defn}
Let $\preceq$ be a semilinear preorder on a set $X$. A subset $F\subset X$ is {\it coinitial} if for any $x\in X$ there is some $y\in F$ with $y\preceq x$.
\end{defn}
For a subset $S$ of a group $G$, we use ${\rm sgr}(S)$ to denote the semigroup generated by $S$.
It was shown in \cite[Theorem 2.7]{Kop} that a group admitting a semilinear left partial order is left-orderable. Now we generalize it to the case of preorder.
\begin{prop} \label{key}
Let $G$ be a group admitting a semilinear left preorder $\preceq$ and let $P$ be its positive cone. Set
$H=P\cap P^{-1}$ and $\widetilde{H}=\bigcup_{p\in P}\bigcap_{q^{-1}\preceq p^{-1}} q^{-1}Hq.$
Then
\begin{itemize}
\item[(1)] $\widetilde{H}$ is a normal subgroup of $G$;
\item[(2)] if $G$ is finitely generated and $H\neq G$, then $\widetilde{H}\neq G$;
\item[(3)] the quotient $G/\widetilde{H}$ is left-orderable.
\end{itemize}
\end{prop}
\begin{proof}
(1) From (P2), $H$ is a subgroup of $G$; thus $x\in \widetilde{H}$ implies $x^{-1}\in \widetilde{H}$ by the definition. Let $x,y\in \widetilde{H}$. Suppose that $x\in \bigcap_{q^{-1}\preceq p_{1}^{-1}} q^{-1}Hq$ and $y\in \bigcap_{q^{-1}\preceq p_{2}^{-1}} q^{-1}Hq$ for some $p_1,p_2\in P$. Since $p_1^{-1}$ and $p_2^{-1}$ are comparable
by (O6), we may assume that $p_1^{-1}\preceq p_{2}^{-1}$. Thus $y\in \bigcap_{q^{-1}\preceq p_{2}^{-1}} q^{-1}Hq\subseteq \bigcap_{q^{-1}\preceq p_{1}^{-1}} q^{-1}Hq$ and hence $xy\in \bigcap_{q^{-1}\preceq p_{1}^{-1}} q^{-1}Hq\subset \widetilde{H}$. So $\widetilde{H}$ is a group.
For each $p\in P$ and $g\in G$, from Lemma \ref{preorder} (2),
\begin{eqnarray*}
&& g\left(\cap_{q^{-1}\preceq p^{-1}} q^{-1}Hq\right)g^{-1}=g\left(\cap_{x\in p^{-1}P^{-1}} xHx^{-1}\right)g^{-1}\\
&&= \cap_{x\in p^{-1}P^{-1}} gxH(gx)^{-1}
=\cap_{x\in gp^{-1}P^{-1}} xHx^{-1} \\
&& = \cap_{x\preceq gp^{-1}} xHx^{-1} \subset \cap_{x\preceq y^{-1}} xHx^{-1}\\
&&\subset \widetilde{H},
\end{eqnarray*}
where $y^{-1}$ is some element in $P^{-1}$ with $y^{-1}\preceq gp^{-1}$. Thus $\widetilde{H}$ is normal in $G$.
\medskip
(2) Suppose that $G$ is finitely generated and $\{g_1,\cdots, g_n\}$ is a finite set of generators. To the contrary, assume that $G=\widetilde{H}$. Then, for each $i\in\{1,\cdots,n\}$, there is some $p_{i}\in P$ such that $g_i\in \cap_{q^{-1}\preceq p_{i}^{-1}}q^{-1}Hq$. Take some $p\in P$ with $p^{-1}\preceq p_{i}^{-1}$ for each $i\in\{1,\cdots,n\}$. Then we have
\[\{g_1,\cdots, g_n\}\subset \cap_{q^{-1}\preceq p^{-1}}q^{-1}Hq\subset p^{-1}Hp.\]
Thus $G\subset p^{-1}Hp$ and hence $G=H$. So (2) holds.
\medskip
(3) We may assume that $H\neq G$; otherwise, $\widetilde{H}=G$ and the conclusion is trivial.
\medskip
{\bf Claim 1.} For any finitely many $x_1,\cdots, x_n\in G\setminus\widetilde{H}$, there are some $\varepsilon_1,\cdots, \varepsilon_n\in\{-1,1\}$ such that
\begin{equation*}
{\rm sgr}(x_1^{\varepsilon_1},\cdots,x_n^{\varepsilon_n})\cap \widetilde{H}=\emptyset.
\end{equation*}
\medskip
We show the Claim 1 by induction on $n$. Given $x\in G\setminus \widetilde{H}$, suppose that there is a positive integer $k$ such that $x^{k}\in \widetilde{H}$. Then $x^{k}\in \bigcap_{q^{-1}\preceq p^{-1}} q^{-1}Hq$, by the definition of $\widetilde{H}$, for some $p\in P$ with $p^{-1}\preceq x$. Let $q^{-1}\preceq p^{-1}$ be given. Then $q^{-1}\preceq x$ and hence $qxq^{-1}\in P\cup P^{-1}$ by (P4). WLOG, we may assume that $qxq^{-1}\in P$. Then $qx^{k}q^{-1}\in H$ implies that $qxq^{-1}\in H$, by Lemma \ref{preorder} (3). Thus $x\in \bigcap_{q^{-1}\preceq p^{-1}} q^{-1}Hq$ whence $x\in \widetilde{H}$. This contradicts the assumption. So the Claim 1 holds for $n=1$.
Now assume that $n\geq 2$ and the Claim 1 holds for any $y_1,\cdots, y_m\in G\setminus\widetilde{H}$ with $m<n$.
By (O7), there is some $p\in P$ with
\begin{equation*}
p^{-1}\preceq x_1, \cdots, p^{-1}\preceq x_n.
\end{equation*}
Then $qx_1,\cdots, qx_n\in P$, for each $q^{-1}\preceq p^{-1}$. According to (P4) in Proposition \ref{char of semi}, we have $\{q x_1q^{-1}, \cdots, qx_nq^{-1}\}\subset P\cup P^{-1}$.
Thus there are some $\vec{\varepsilon}(q)=(\varepsilon_1(q),\cdots, \varepsilon_n(q))\in\{-1,1\}^{n}$ such that
$\{q x_1^{\varepsilon_1(q)}q^{-1}, \cdots, qx_n^{\varepsilon_n(q)}q^{-1}\}\subset P$.
For each $\vec{\varepsilon}=(\varepsilon_1,\cdots, \varepsilon_n)\in\{-1,1\}^{n}$, let
\[Q(\vec{\varepsilon})=\left\{q^{-1}\preceq p^{-1}: \{q x_1^{\varepsilon_1}q^{-1}, \cdots, qx_n^{\varepsilon_n}q^{-1}\}\subset P\right\}.\]
Then $\{q^{-1}: q^{-1}\preceq p^{-1}\}=\bigcup_{\vec{\varepsilon}\in \{-1,1\}^{n}}Q(\vec{\varepsilon})$.
Let $P^{++}=\{g\in G: g\succ e\}$ and $P^{--}=\{g\in G: g\prec e\}$ be the strictly positive and negative cones respectively. Now we discuss into two cases.
\medskip
{\bf Case 1}. There is an $\vec{\varepsilon}=(\varepsilon_1,\cdots, \varepsilon_n)\in\{-1,1\}^{n}$ and a coinitial set $Q\subset Q(\varepsilon)$ such that for any $q^{-1}\in Q$,
\[\{q x_1^{\varepsilon_1}q^{-1}, \cdots, qx_n^{\varepsilon_n}q^{-1}\}\subset P^{++}.\]
Noting that any word composed of $qx_1^{\varepsilon_1}q^{-1},\cdots, qx_n^{\varepsilon_n}q^{-1}$ will lie in $P^{++}$ for any $q^{-1}\in Q$ and $Q$ is a coinitial set, the semigroup $ {\rm sgr}(x_1^{\varepsilon_1},\cdots,x_n^{\varepsilon_n})$ has empty intersection with $\widetilde{H}$. Then the Claim holds.
\medskip
{\bf Case 2}. For any $\vec{\varepsilon}=(\varepsilon_1,\cdots, \varepsilon_n)\in\{-1,1\}^{n}$, there is some $p^{-1}(\vec{\varepsilon})\preceq p^{-1}$ such that for any $q^{-1}\in\{q^{-1}\in Q(\vec{\varepsilon}): q^{-1}\preceq p^{-1}(\vec{\varepsilon})\}$, we have
\[ \left\{i\in\{1,\cdots,n\}: qx_{i}^{\varepsilon_i}q^{-1}\in H\right\}\neq \emptyset.\]
Thus, for each $q^{-1}\in Q(\vec{\varepsilon})$, there is a partition $\{1,\cdots,n\}=A(q, \varepsilon)\cup B(q, \varepsilon) $ such that
\[ \{ qx_{i}^{\varepsilon_i}q^{-1}: i\in A(q, \varepsilon)\}\subset H\ \ \text{and }\ \ \{ qx_{i}^{\varepsilon_i}q^{-1}: i\in B(q, \varepsilon)\}\subset P^{++}.\]
Since $H\neq G$ by the assumption, we have $P^{++}=P\setminus H\neq \emptyset$ and hence $P^{++}$ is infinite. Note that there are only finitely many partitions of $\{1,\cdots, n\}$. So there is an $\vec{\varepsilon}=(\varepsilon_1,\cdots, \varepsilon_n)\in\{-1,1\}^{n}$, a coinitial set $Q\subset Q(\vec{\varepsilon})$ and a partition $\{1,\cdots,n\}=A \cup B $ with $A\neq\emptyset, B\neq\emptyset$ such that
\[ \{ qx_{i}^{\varepsilon_i}q^{-1}: i\in A \}\subset H\ \ \text{and }\ \ \{ qx_{i}^{\varepsilon_i}q^{-1}: i\in B \}\subset P^{++},\]
for each $q^{-1}\in Q$. WLOG, we may assume that $A=\{1,\cdots,k\}$ and $B=\{k+1,\cdots,n\}$ for some $k\in\{1,\cdots, n-1\}$.
Now by the induction hypothesis, there is some $\vec{\eta}=(\eta_1,\cdots,\eta_{k})\in\{-1,1\}^{k}$ such that
\[ {\rm sgr}(x_1^{\eta_1},\cdots,x_k^{\eta_k})\cap \widetilde{H}=\emptyset. \]
Then we conclude that $(\eta_1,\cdots,\eta_{k}, \varepsilon_{k+1},\cdots, \varepsilon_{n})\in\{-1,1\}^{n}$ satisfies
\[ {\rm sgr}(x_1^{\eta_1},\cdots,x_k^{\eta_k}, x_{k+1}^{\varepsilon_{k+1}}, \cdots, x_{n}^{\varepsilon_{n}})\cap \widetilde{H}=\emptyset. \]
Indeed, let $w$ be a word composed of $x_1^{\eta_1},\cdots,x_k^{\eta_k}, x_{k+1}^{\varepsilon_{k+1}}, \cdots, x_{n}^{\varepsilon_{n}}$. If $x_{k+1}^{\varepsilon_{k+1}}, \cdots, x_{n}^{\varepsilon_{n}}$ do not occur in $w$, then the choice of $\vec{\eta}$ implies that $w$ is not in $\widetilde{H}$. If there are some letters of $x_{k+1}^{\varepsilon_{k+1}}, \cdots, x_{n}^{\varepsilon_{n}}$ occur in $w$, then for each $q^{-1}\in Q$, $qwq^{-1}\in P^{++}$ and hence $w$ is not in $\widetilde{H}$ by the coinitiality of $Q$. Thus we complete the proof of the Claim 1.
\medskip
Now (3) is followed from some standard arguments (see \cite[Lemma 2.2.3]{Glass}). For convenience of the readers, we afford a detailed proof here.
\medskip
By the principe of compactness, the Claim 1 implies the following directly.
\medskip
{\bf Claim 2}. There is a map $\varepsilon: G\setminus \widetilde{H}\rightarrow \{-1,1\}$ such that for any finite $g_1,\cdots,g_n\subset G\setminus \widetilde{H}$, ${\rm sgr}\left(g_1^{\varepsilon(g_1)}, \cdots, g_{n}^{\varepsilon(g_n)}\right)\cap \widetilde{H}=\emptyset$.
\medskip
Let $\widetilde{P}=\{ g\in G\setminus\widetilde{H}: \varepsilon(g)=1\}$. Then it is easy to verify that $\widetilde{P}$ is a subsemigroup of $G$ and $G=\widetilde{P}\cup \widetilde{H}\cup\widetilde{P}^{-1}$.
\medskip
{\bf Claim 3}. $\widetilde{P}=\widetilde{H}\widetilde{P}\widetilde{H}$.
\medskip
First $\widetilde{P}=e\widetilde{P}e\subset \widetilde{H}\widetilde{P}\widetilde{H}$. To show the converse, we conclude that for any $h\in \widetilde{H}$ and $q\in \widetilde{P}$, $hq\in \widetilde{P}$. Otherwise, $\varepsilon(hq)=-1$. Then $h^{-1}\in {\rm sgr}((hq)^{-1}, q)\cap \widetilde{H}$, which contradicts the choice of $\varepsilon$. Similarly, $qh\in \widetilde{P}$. Thus for any $h_1,h_2\in \widetilde{H}$ and $q\in \widetilde{P}$, $h_1qh_2\in \widetilde{P}$ and hence $\widetilde{H}\widetilde{P}\widetilde{H}\subset \widetilde{P}$. Thus Claim 3 holds.
\medskip
We define an order $\leq$ on the quotient group $G/\widetilde{H}=\{g\widetilde{H}: g\in G\}$ by
\[ f\widetilde{H}<g\widetilde{H}\ \ \ \text{ if and only if }\ \ \ f^{-1}g\in \widetilde{P}.\]
It is well defined by Claim 3 and is a total order by the equality $G=\widetilde{P}\cup \widetilde{H}\cup \widetilde{P}^{-1}$. It is obvious that $\leq$ is $G$-invariant, i.e. for any $g,x,y\in G$, $gx\widetilde{H}<gy\widetilde{H}$ whenever $x\widetilde{H}<y\widetilde{H}$. Thus the quotient $G/\widetilde{H}$ is left-orderable.
\end{proof}
\section{Existence of pointwise fixed arcs}
The following is the well-known Margulis' Normal Subgroup Theorem (see e.g. \cite[Theorem 8.1.2]{Zim}).
\begin{lem}\label{NST}
Let $G$ be a connected real semisimple Lie group with finite center and no compact factors, let $\Gamma$ be an irreducible lattice of $G$. Assume that $\mathbb{R}$-{\rm rank}$(G)\geq 2$. Then every normal subgroup of $\Gamma$ either is contained in the center of $G$ and hence is finite or has finite index in $\Gamma$.
\end{lem}
The following is the Lemma 2.1 in \cite{SX}
\begin{lem} \label{fixed point}
Let $X$ a nondegenerate dendrite and $h:X\rightarrow X$ be a homeomorphism. If $h$ fixes an end point $e\in X$, then $h$ fixes another point $o\not=e$.
\end{lem}
The following is key to establish the structure theorem.
\begin{thm}\label{end fixed}
Let $\Gamma$ be a higher rank lattice acting on a nondegenerate dendrite $X$. If $\Gamma$ fixes some end point $z$ of $X$, then there is another point $s\in X$ such that $\Gamma$ fixes the arc $[z,s]$ pointwise.
\end{thm}
\begin{proof}
Now take a finite set $\{g_1,\cdots,g_n\}$ of generators of $\Gamma$. For each $i\in \{1,\cdots, n\}$, by Lemma \ref{fixed point}, we can take $x_i\in X\setminus\{z\}$ fixed by $g_i$. Let $t$ be the point such that $[z,t]=\cap_{i=1}^{n}[z,x_i]$. Define a positive cone $P$ of $\Gamma$ by
\[ P=\left\{g\in \Gamma: g^{-1}(t)\in [z,t]\right\}.\]
Then $P$ leads to a preorder $\preceq$ on $\Gamma$. Let $H=P\cap P^{-1}$ and $\widetilde{H}=\bigcup_{p\in P}\bigcap_{q^{-1}\preceq p^{-1}} q^{-1}Hq$. Note that $H$ is just the stabilizer of $t$ in $\Gamma$.
\medskip
{\bf Claim}. There is some point $s\in (z,t]$ fixed by $\Gamma$.
\medskip
We discuss into two cases for the proof of the Claim.
\medskip
{\bf Case 1}. There is a sequence $(f_i)$ in $P^{-1}$ such that $f_i(t)\rightarrow z$ as $i\rightarrow \infty$.
Then $P$ gives rise to a semilinear left preorder $\preceq$ on $\Gamma$: for $g, h\in\Gamma$,
$$g\preceq h\ \mbox{if and only if}\ [z, g(t)]\subset [z, h(t)].$$
By Proposition \ref{key}, $\Gamma$ admits a left-orderable quotient $\Gamma/\widetilde{H}$. By Lemma \ref{NST}, either $\widetilde{H}$ is contained in the center of $G$ or $\Gamma/\widetilde{H}$ is a finite group, where $G$ is the ambient Lie group of $\Gamma$. In the former case, $\Gamma/\widetilde{H}$ is also a higher rank lattice, which
contradicts Deroin-Hurtado's theorem in \cite{DH}. In the latter case, $\Gamma/\widetilde{H}$ is a finite group. Then the orderability implies that it is trivial and hence $\widetilde{H}=\Gamma$. Then we have $H=\Gamma$, by Proposition \ref{key} (2). Thus $t$ is fixed by $\Gamma$ and take $s=t$.
\medskip
{\bf Case 2}. There is no sequence $(f_i)$ in $P^{-1}$ with $f_i(t)\rightarrow z$ as $i\rightarrow \infty$.
Fix a canonical ordering $<$ on $[z,t]$ with $z<t$. Let $s=\inf_{g^{-1}\in P^{-1}} g(t)$. Then $s\in (z,t]$ and we claim that $s$ is fixed by $\Gamma$. Indeed, let $g^{-1}\in P^{-1}$ be given. Suppose that $s=\lim\limits_{n\rightarrow\infty} \gamma_{n}^{-1}(t)$ for some sequence $(\gamma_n^{-1})$ in $P^{-1}$. Then $g^{-1}(s)\geq s$ with respect to the canonical ordering $<$ on $[z,t]$. If $g^{-1}(s)\neq s$ then $g^{-1}\gamma_n^{-1}(t)>\gamma_n^{-1}(t)$, for all sufficiently large $n$. Then $g\gamma_n^{-1}(t)<\gamma_n^{-1}(t)\leq t$ and hence $g\gamma_{n}^{-1}\in P^{-1}$. Thus
\[ s\leq \lim_{n\rightarrow\infty}g\gamma_{n}^{-1}(t)\leq \lim_{n\rightarrow\infty}\gamma_{n}^{-1}(t)=s.\]
So $g(s)=s$ and hence $g^{-1}(s)=s$. This contradiction shows that $s$ is fixed by each element in $P$.
Since for each $i$, either $g_i\in P$ or $g_i^{-1}\in P$, $s$ is fixed by $\Gamma$.
\medskip
Thus the Claim holds. Now, by Denrion-Hurtado's theorem, $\Gamma$ fixes the $[z,s]$ pointwise.
\end{proof}
\section{Proof of the main theorem}
Now we are ready to prove the main theorem.
\begin{lem}\cite[Proposition 5.11]{DK}\label{index}
Let $G$ be a finitely generated group. For any $n\in\mathbb{N}$, there are finitely many subgroups of $G$ with indices less than $n$.
\end{lem}
\begin{proof}[Proof of the main theorem] Along the same lines of the proof of Theorem 1.2 in \cite{SX}, Theorem \ref{main theorem 1} follows from Theorem \ref{end fixed}
and Lemma \ref{index} (We need only use Theorem \ref{end fixed}
and Lemma \ref{index} instead of Proposition 5.3 and Lemma 5.5 in \cite{SX} respectively).
\end{proof}
|
\section{Introduction}
Optomechanical crystal (OMC) cavities are micro- or nanofabricated structures that take advantage of the coupling between optical and acoustic waves to generate superposition between mechanical and optical cavity modes~\cite{Eichenfield2009, Aspelmeyer2014}, allowing amplification, generation and detection of photons and phonons~\cite{Kippenberg2005, Arcizet2006, Chan2011, Weitz2013, Teufel2009, Galland2014}. Most applications of OMC cavities rely on the capability of laser cooling the mechanical resonator to its ground state~\cite{Chan2011}, where a true quantum-coherent interactions may take place. In principle, ground state cooling may be achieved even at room temperature, by strongly pumping the cavity with a red-detuned laser. In such a scenario the optical cavity suppress Stokes scattering while amplifying the anti-Stokes scattering, which effectively drains energy from the mechanical resonator. The OMC laser cooling is efficient if the optomechanical coupling rate ($g_0$) is large enough to overcome the photons and phonon dissipation rates. This condition is commonly referred to as high-cooperativity regime ($C=4n_pg_0^2/(\kappa \Gamma)>1$)~\cite{Aspelmeyer2014}, where $n_p$ is the number of pump photons inside the cavity and $\kappa$ and $\Gamma$ are the cavity optical and mechanical decay rates. Another condition for reaching efficient cooling is related to the capability of the optical cavity to distinguish between Stokes and anti-Stokes, referred to as the sideband-resolved regime ($\omega_m\gg\kappa$), where $\omega_m$ is the resonance frequency of the mechanical oscillator.
Diamond OMC cavities are specially interesting for optomechanics due to the remarkable mechanical and optical properties of diamond~\cite{Burek2016, Chia2021}. For instance, the large bandgap of diamond, $\SI{5.45}{\eV}$, combined with its high thermal conductivity and small thermal expansion coefficient, allows fabrication of nanobeams operating with large number of pump photons ($n_c>10^4$)~\cite{Burek2016}. At the same time the large Young's modulus of diamond, combined with its low intrinsic mechanical dissipation (e.g., low thermoelastic damping) enables the fabrication of mechanical resonators with high resonance frequencies and high quality factors ($Q_m > 10^5$)~\cite{Tao2014}. At last, diamond also have been seen as a promising material for quantum networks since it can host resilient single-spin color centers~\cite{Gruber1997} to store quantum information~\cite{Hausmann2013, Burkard2014}. OMC cavities can be used to manipulate diamond color centers with light at telecom wavelengths, without relying on qubit optical transitions~\cite{Shandilya2021}.
Diamond devices are however not simple to process~\cite{Castelletto2017}. Standard planar fabrication techniques do not work for bulk diamond~\cite{Piracha2016} and single crystal diamond (SCD) grown by Chemical Vapor Deposition (CVD) are not uniform over large areas~\cite{Burek2012}. SCD usually needs to be built with a relatively larger thickness and then thinned to the device thickness~\cite{Lee2012}. Since only the upper surface of the SCD has good optical quality, it must be either thinned from bottom to up with quasi-isotropic etching, mounted upside down over another material through wafer bonding, or flipped using electrochemical lift-off~\cite{Khanaliloo2015}. Resonators in mono-crystalline CVD grown SCD can be produced by removing the lower-quality nucleation layer with high-energy ion implantation, leaving the cavity in the highest quality layer at the diamond surface~\cite{Riedrich-Moller2012}. A quasi-isotropic etch technique has been applied to undercut the resonator structure achieving optical quality factor above $1.4\times10^4$~\cite{Mouradian2017}. The optical quality factor in these cases is affected by the residual damage of the etching in the structure. Resonators in polycrystalline diamond (PCD) thin films, in contrast, are much simpler and cheaper to process since they can be grown with the resonator thickness. However, they have limited optical quality factors due to absorption and light scattering across grain boundaries ($Q_o\sim10^5$)~\cite{Abdou2018}.
In this work, we are interested in developing an optimized PCD-based nanobeam design capable of operating at both the high-cooperativity and the resolved-sideband regime. These regimes have already been reached in SCD-based microdisks~\cite{Mitchell2016} and triangular nanobeams~\cite{Burek2016}. PCD-based device design optimization yet has to consider the limited optical quality factor of the material, using strategies to boost cavity performance~\cite{Babinec2011} while aiming in geometries with high optomechanical coupling to increase the cooperativity. A widely spread design optimization method developed for Si-based nanobeams~\cite{Chan:2012iy} focuses on maximizing $Q_og_0^2$ by combining finite-element (FEM) simulations with numerical optimizations methods. When applied to diamond devices, however, they lead to designs with a low mechanical quality factor~\cite{Cady2019}. Because diamond has a lower refraction index ($n=2.45$) and a higher Young's modulus than silicon, it has a weaker localization of the optical cavity mode and a stronger localization of the mechanical mode. As consequence, to increase the overlap between optical and mechanical modes, numerical methods tend to converge to geometries with low confinement of mechanical modes. Under these circumstances, we had to develop a new optimization method considering optical and mechanical losses. Some advantages of our method are: (i) it leads to a better solution after a relatively small number of iterations due to the hybrid optimization process that starts with a genetic algorithm (GA) followed by a pattern search (PS) for local optimization; (ii) it returns a set of locally optimized solutions, instead of a single solution; (iii) it restricts solutions to a specific-frequency range, instead of fixing geometric parameters; (iv) our output geometries already have the resonant optical mode in the optical communication wavelength bands and don't need to be rescaled.
We present here, two distinct robust geometries with high single-photon cooperativity ($C_0>10^{-2}$) for diamond nanobeams, that were delivered by our optimization method.
\section{Cavity design}
Our optimized cavity is a nanobeam with rectangular cross-section (width $w=\SI{664}{\nm}$ and thickness $t=\SI{450}{\nm}$) patterned with elliptical holes (Fig.~\ref{fig:nanobeam}\textbf{a}). The nanobeam design consists of a sequence of unit cells (rectangular blocks with an oval hole in the middle, as depicted in Fig.~\ref{fig:nanobeam}\textbf{b}). From the central cell to any of the borders, there are $13$ unit cells. The first $N=8$ are defects cells followed by 5 mirror cells.
\begin{figure*}[ht!]
\begin{center}
\includegraphics[width=12.5cm]{Fig1_vfinal.pdf}
\caption{\small{\textbf{a1)} Representation of the rectangular nanobeam pattered with quasi-periodic elliptical holes~\cite{Chan2012}. \textbf{a2)} Ilustration of the central defect ($j=0$) and mirror mirror (j=8) nominal unit cells, with thickness $t$, width $w$, lattice parameter $a_j$ and hole diameters $h_{xj}$ and $h_{yj}$. \textbf{a3)} Defect cells parameters variation for an optimized solution of the nanobeam.
\textbf{b)} Mechanical band diagram~(\textbf{b2}-right) and the variation of the unit cell mechanical frequency at the $\Gamma$-point for each defect from the center ($j=0$) to the mirrors ($j=8$) (\textbf{b2}-left). The blue, orange and green dots represent the zero-, first-~(\textbf{b1}-left) and second-order~(\textbf{b1}-right) symmetric modes, respectively. The zero-order mode (blue dots) at the $\Gamma$-point comes down to a translation of the unit cell. The continuous blue lines represent the non-symmetric modes and the transparent red area represents the bandgap at the mirror. The parameters used for the defects and mirrors (\textbf{a3}) were obtained from a solution of the nanobeam's optimization, and the dashed line represents the frequency of the confined optical mode for such nanobeam.
\textbf{c)} Optical band diagram~(\textbf{c2}-left) and the variation of the unit cell optical frequency at the X-point for each defect from the mirror ($j=8$) to the center ($j=0$)~(\textbf{c2}-right). The blues and orange dots represent the zero-~(\textbf{c1}-right) and first-order~(\textbf{c1}-left) modes. The gray area represents a continuum of radiation. The dashed line represents the frequency of the lower-energy optical mode confined at the center of the optimized nanobeam.}}
\label{fig:nanobeam}
\end{center}
\end{figure*}
For our nanobeam, the design optimization boils down to defining, in a way to maximize the cooperativity, the unit cell area (cell width and lattice parameter) and its hole size and shape, for the mirrors and the central defect cells, together with how they change along the nanobeam. The mirrors act as a waveguide with defined propagating modes and a bandgap between them. The defects are fabricated in a way that the energy of a specific mode deepens into the mirror bandgap, confining it locally. Figs.~\ref{fig:nanobeam}~\textbf{b} and \textbf{c} show the mechanical and optical band diagrams for propagating waves in a unit cell with Floquet periodicity at the boundaries and how the resonance frequency of the modes changes from the mirror to the defect cell. The parameters used for the defects fabrication are the same as those expressed in Fig.~\ref{fig:nanobeam}~\textbf{a3} and were obtained from the optimization process.
The optical mode is confined at the X-point. The zero- and first-order modes (Fig.~\ref{fig:nanobeam}~\textbf{c1}) are both confined along the x-direction, parallel to the lattice parameter direction. While the first-order mode is confined mainly over the holes, where there is no material strain, the zero-order mode is distributed over the diamond, which tends to yield a higher value of the photoelastic component of the optomechanical coupling. For these modes, reducing the lattice parameter decreases the confinement region, increasing the resonance frequency. Thus defects were parameterized in a way that their lattice parameter decreases closer to the center of the nanobeam, creating a confining potential for the zero-order optical mode: $a_j=(1-d_j)a_N$, where the index $j$ refers to the number of the defect counting from the central defect ($j=0$) to the mirror defect $j=N$, $a_j$ is the $j$-th defect's lattice parameter, and $d_j$ is a function that softly varies from 1 to a defined value $d_0$. For the optical mode, the choice of $a_N$ is related to the frequency of the modes at the mirror, while $d_0$ determines how deeply the defect frequency penetrates in the bandgap. This amount can be increased or balanced by the variation of the hole area, which affects the effective cell refraction index. Increasing the hole reduces the refraction index of the defect which has the opposite effect of decreasing the lattice parameter.
The mechanical mode is confined at the $\Gamma$-point, where the zero-order mode (blue dots of Fig.~\ref{fig:nanobeam}~\textbf{b2}) becomes a simple displacement of the cell. The first mode (orange dots) is a ``breathing'' mode and normally is the mode that leads to the highest cooperativity since it can have high quality factor and optomechanical coupling with the zero-order optical mode. We have though observed less confined high-order modes with high-optomechanical coupling in the nanobeam simulation. The advantage of the breathing mode for confinement control is that its frequency strongly depends on the hole shape. If on one hand, the breathing mode has periodicity along the x-direction and therefore its frequency increases with the lattice parameter decreasing, on other hand, it is characterized by an increase of stress above the edge of the hole (Fig.~\ref{fig:nanobeam}~\textbf{b1}), thus squeezing the hole in the y-direction increases the stressed area decreasing the mode frequency. There is however a catch specific to this diamond nanobeam design that must be considered, that is the fact that for confining optical modes in the telecommunication C-band ($\sim\SI{200}{THz}$) the lattice parameter must be close to the $\SI{450}{nm}$ of thickness. This means that squeezing the holes in the x-direction may lead to modes with high enough energy to be confined only at the nanobeam surface, which may have a significant impact on the optomechanical coupling.
We opted to give the optimization method freedom to test different mechanical modes and confining strategies. The limitation was restricted only by the number of modes the simulation was allowed to find and by the symmetry constraints (only symmetric modes were simulated during the optimization process). The $x$ and $y$ components of the elliptical hole diameters were parameterized as $h_{xj} = (a_j-2\delta)\eta_j+\delta$ and $h_{yj} = (w-3\delta)\gamma_j+\delta$, where $\eta_j$ and $\gamma_j$ can vary between the defined parameters $\eta_0$, $\gamma_0$, $\eta_N$ and $\gamma_N$ (limited between 0 and 1) and $\delta=\SI{50}{nm}$, is a constant used to ensure that all simulated geometries are bounded by realistic fabrication limits. More details of the parameterization are given in the Supplement 1.
The high number of parameters to adjust, the way each of them influences at the same time the optical and mechanical modes, the different possible strategies for confinement, and the interplay between the mechanical confinement and the optomechanical coupling in diamond, make the nanobeam design optimization a complex problem with no trivial solutions.
\section{Nanobeam optimization}
To optimize the nanobeam design, we combined FEM simulations performed in COMSOL$^\text{\textregistered}$ with numerical optimization methods performed by Dakota$^\text{\textregistered}$. The simulations account for one-eighth of the nanobeam inside a computational domain (vacuum) bounded by PMLs, using proper mirror symmetry conditions to ensure that only symmetric mechanical modes are sought. The optical and mechanical studies are computed individually and used to evaluate the optomechanical coupling rate, which is composed of a photoelastic~\cite{Ramachandran1947} and a moving boundary component~\cite{Johnson2002}. The PML region ensures that radiative optical and mechanical damping are accurately calculated (the PML typical absorption wavelength is changed according to the ongoing study).
Optimization methods that focus on maximizing the optomechanical coupling only, don't need to take losses into account. This simplifies a lot the optimization process because optimization is preserved if the nanobeam is rescaled. More than that, the optimization is preserved if small scaling is performed only in the $xy$-plane keeping the nanobeam thickness constant. This is usually done to adjust the design to the desired operating frequency. However, when dealing with the cooperativity, such scaling does not preserve the optimization, since losses do not scale together with the optomechanical coupling and the spectral distance between the defect modes and other bands which do not scale linearly with the geometry becomes a relevant factor. Consequently, the nanobeam must be optimized at a specific optical frequency. Therefore, we introduce an optical bandpass frequency filter $F(\nu_{o})=[\erf(\nu_{o}-\SI{183.9}{THz})-\erf(\nu_{o}-\SI{198.4}{THz})]/2$ when defining the optimization goal variable ($\erf$ is the error function), forcing solutions with optical resonance $\nu_o$ between a desirable frequency range, while leaving all geometric parameters free for optimization, except for the nanobeam thickness. Therefore, our design optimization consists on maximizing $q \equiv C_0 F(\nu_{o})$, where $C_0$ is the cooperativity of a single photon.
The optimization process is performed by Dakota$^\text{\textregistered}$ in two steps starting with a Genetic Algorithm (GA). In the second step, some solutions from the GA are selected and locally optimized using a Pattern Search (PS). Due to the complexity of the problem and the strong dependence on initial conditions, the combination of these two methods results in better solutions than the ones achievable by a single local optimization algorithm, in a short convergence time compared to global optimization algorithms.
The GA is initially fed with random geometries containing a confined zero-order optical mode within the frequency range delimited by the frequency filter. This is ensured by a pre-start process where randomly generated geometries are tested. The geometries containing modes with a frequency above the frequency threshold are rescaled and reparemeterized. The ones without confined modes or with the zero-order mode frequency beyond the minimum threshold are discarded. For the rectangular nanobeam, approximately $45\%$ of the randomly generated geometries were used as the initial population of the GA and $70\%$ of them had to be rescaled.
To find the nanobeam zero-order optical mode frequency with a low computational cost, we simulate the defect unit cell with scattering bound conditions before the nanobeam study. The unit cell is much smaller and faster to simulate than the nanobeam, and its modes frequencies will be closer to the ones of the confined modes in the whole cavity, reducing the number of optical modes we have to simulate to find the zero-order mode.
The PS method was applied for 26 distinct solutions from the GA. The solutions were automatically selected by our algorithm, which excludes solutions too close to each other in the parameter space. The distance between two solutions were defined as $D \equiv \sum_{i} [(p_{1i}-p_{2i})/(p_{1i}+p_{2i})]^2$, where $p_{1i}$ and $p_{2i}$ refers to each one of the geometry parameters of the solutions. After discarding the solutions with the lowest $q$ value the algorithm compares pairs of solutions and, if they are too close to each other in the parameter space, the one with the lowest $q$ value is discarded. The threshold for $D$ to decide if a solution will be or not discarded is automatically varied until at least 20 solutions are selected.
By the end, the program returns the geometry parameters and all evaluated parameters (such as optomechanical coupling, quality factors, and modes frequency) for each tested solution.
\section{Results and Discussions}
\begin{figure*}[ht!]
\begin{center}
\includegraphics[width=12.5cm]{Fig2_vfinal.pdf}
\caption{\small{\textbf{a)} Value of the optimization parameter $q$ at each step of the optimization process. The first region (light gray region) are the 480 steps of the GA, the other 26 colors represent the PS applied for different initial parameters selected from the GA. The colored dots on the first region represent the distinct selected solutions from the GA for the PS. \textbf{b)} Hole diameters at the center ($j=0$) and mirror ($j=n$) for all tested geometry. The blue and orange dots refers to solutions of the GA and PS methods respectively. \textbf{c)} Optical (left) and mechanical (right) quality factor for each tested geometry. The color of the dots represent the value of the single photon cooperativity $C_0$. \textbf{d)} Ilustration of the optical and mechanical modes for two different geometries. The geometry represented in \textbf{d1} has two well-confined symmetric mechanical modes.}}
\label{fig:results}
\end{center}
\end{figure*}
\begin{table*}[ht!]
\centering
\begin{tabular}{c|c|c|c|c|c|c}
& $\nu_o$~(THz) & $\nu_m$~(GHz) & $Q_o$ & $Q_m$ & $g_0/(2\pi)$~(kHz) & $C_0$ \\
\hline
1 & 194.28 & 8.37 & $2.3\times10^5$ & $\geq 10^6$ & 213.95 & $2.59\times10^{-2}$\\
2 & 200.43 & 12.05 & $8.4\times10^4$ & $\geq 10^6$ & 306.74 & $1.31\times10^{-2}$\\
\end{tabular}
\caption{\small{Optical and mechanical frequencies ($\nu_o$ and $\nu_m$) and quality factors ($Q_o$ and $Q_m$), optomechanical coupling ($g_0$) and single-photon cooperativity ($C_0$) for two distinct solutions of the design optimization method. Both where optimized for $C_0 F(\nu_o)$. The first is the most optimized solution and the second is a particular solution using a distinct strategy to confine the mechanical mode.}}
\vspace{-6pt}
\label{tab:results}
\end{table*}
For the results presented in this paper we ran the GA with an initial population of 48 geometries and a limit of 480 evaluations. At the end, the PS was applied to 26 selected solutions with a minimum of 30 evaluations for each of them. We constrained an upper bound for both optical and mechanical quality factors of $1\times10^6$ to account for realistic material losses that ultimately limit these quantities. The graphics on Fig.~\ref{fig:results}~\textbf{a-c} summarize the results for all tested geometries. Fig.~\ref{fig:results}~\textbf{c} shows the optomechanical coupling, optical and mechanical quality factor and the cooperativity for the modes with the highest cooperativity in each tested geometry. While we were expecting solutions with the mechanical breathing mode confined between the zero- and first-order mirror modes, as shown in Fig.~\ref{fig:nanobeam}~\textbf{b2}, an unforeseen result was a solution with high cooperativity where the first-order mechanical mode was confined in a band formed between the first- and second-order modes. The Table~\ref{tab:results} contains the relevant properties of such solution compared with the solutions with highest $q=C_0 F(\nu_o)$. The most optimized solution presents two symmetric modes with high-cooperativity. The second mode is the one with the highest cooperativity since we limited the mechanical quality factor, and by being less localized, it has a larger overlap with the optical field and consequently a higher $g_0$. This mode has $\omega_m/\kappa = 43$, far in the resolved-sideband regime, and $C_0 = 2.59\times 10^{-2}$, implying that around 40 photons would be enough to reach the high-cooperativity regime, which is orders of magnitude bellow previously obtained experimentally in triangular diamond nanobeams~\cite{Burek2016}. Having two modes may be interesting from a fundamental point of view to study cooling in different regimes. The distance in frequency between these modes is around $\SI{228}{kHz}$ and depending on the mechanical quality factor of real devices they may overlap. There is also an antisymmetric mechanical mode for geometry 1 between the first and second presented modes that is not shown in Fig.~\ref{fig:results}~\textbf{d1}. We omitted the antisymmetryc mode because while it expands on one side of the nanobeam, it contracts in the opposite side. This makes the contribution to the $g_0$ from one half of the nanobeam to cancel with the other one, thus the total optomechanical coupling is negligible.
The solution for the second geometry has a lower optical quality factor, but a higher optomechanical coupling, which is interesting in a scenario where the material losses limit the optical quality factor. As for the first geometry, this solution can also reach the high-cooperativity regime at room temperature. Fig.~\ref{fig:results}~\textbf{d1} and \textbf{d2} show the geometry and the optical and mechanical modes for the most optimized (geometry 1) and the alternative (geometry 2) solutions, respectively. The difference between the first and second geometries resides in the way the hole shape changes.
At last, we performed a resilience test against imprecision of the size of the holes that are expected from the fabrication process. We introduce such imprecision by adding normal-distributed random value with null mean and standard deviation $\sigma = \SI{10}{nm}$ to $h_{xj},h_{yj}$ (Fig.~\ref{fig:resilience}~\textbf{a}). Due to the added imperfections, the mechanical modes change and it is important to define a way of tracking the breathing modes from Fig.~~\ref{fig:results}~\textbf{d}. We did this by computing, for each mode, the inner product of the displacement vectors from the ``perfect'' geometry with the ones of the ``imperfect'' geometry, at the nanobeam upper surface, which is the nanobeam surface that remains unchanged when adding imperfections to the holes. By normalizing this inner product we obtain the projection (at the upper surface) of a unknown mode over the modes we want to track:
\begin{equation}
p_{i,j} = \frac{\langle m_i,m_j \rangle}{\|m_i\|\|m_j\|}
\end{equation}
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=6.375cm]{Fig3_vfinal.pdf}
\caption{\small{\textbf{a)} Illustration of the possible introduced fluctuations of the x and y components of the holes diameter. Each component of each hole is changed individually by a random amount that can be positive or negative, with normal distribution $N(\mu=0,\sigma = \SI{10}{nm})$.
\textbf{b)} Probability of successfully tracking well confined modes for geometries 1 and 2 according to the classification criteria which considers $p_{i,j}>0.75$ and $Q_m \geq 3\times10^4$ to confirm the right mode were tracked and it is still confined despite the introduced perturbations in the geometry.
\textbf{c)} Statistical distribution of optical and mechanical quality factors, optomechanical coupling, and cooperativity normalized by the values of the ``perfect'' design (left) and the statistical distribution of the optical and mechanical frequencies (right).}}
\label{fig:resilience}
\end{center}
\end{figure}
We thus track the mechanical modes of the perturbed geometries by selecting the ones with the maximum projections. Since there are cases in which one or all modes we are tracking are heavily changed it is important to define thresholds for $p_{i,j}$ and $q$ where we assume these modes are not the same modes we had before the perturbation or they are not confined anymore. We used $p_{i,j}>0.75$ and $Q_m \geq 3\times10^4$ as our thresholds and then classified the perturbed geometries as having 2, 1 or no tracked modes (Fig.~\ref{fig:resilience}~\textbf{b}). For the first geometry, we successfully tracked both modes in most cases, although mode 1 (Fig.~\ref{fig:resilience}~\textbf{c1}) is more resilient. This is consequence of the fact that the first mechanical mode has simulated mechanical quality factor fair above $10^6$ and it remains above this value when introducing perturbations. For the second geometry, the mechanical quality factor is less resilient to perturbations due to the narrower bandgap, therefore, the mode confinement is weaker.
Concerning the cooperativity, for the first geometry it drops down around one order of magnitude on average for the first mechanical mode due to a drop in the optical quality factor (Fig.~\ref{fig:resilience}~\textbf{c1-2}) and a bit more for the second mechanical mode that experiments more losses, but the high-cooperativity regime is still reachable at room temperature for both modes at many simulated devices (from all simulated devices 44.2\% had $C_0>10^{-4}$). For the second geometry most of the tested devices did not match the tracking criteria (Fig.~\ref{fig:resilience}~\textbf{b}). For the devices in which we found the first-order confined mechanical mode, there is a drop down in the mechanical quality factor (Fig.~\ref{fig:resilience}~\textbf{c3}) probably due to the narrower mechanical bandgap, the lower mechanical confinement however makes $g_0$ very resilient and 62.0\% of these devices (23.6\% of all simulated devices) have $C_0>10^{-4}$.
In summary, we have developed an efficient method to OMC cavities design optimization. The mixed search of our optimization scheme ensures a broad scan of the parameter space, leading to optimized solutions with distinct properties. For the rectangular nanobeam we investigated two solutions with single photon cooperativity $C_0 > 10^{-2}$ and distinct strategies to confine the mechanical modes. Resilience test suggests the high-cooperativity regime can be reached using both geometries. The developed optimization method is not limited to diamond devices neither to the rectangular nanobeam geometry.
\section*{Funding}
This work was supported by S{\~a}o Paulo Research Foundation (FAPESP) through grants 2020/00100-9, 2020/00119-1, 2018/15580-6, 2018/15577-5, 2018/25339-4, 2021/10249-2, Coordena{\c c}\~ao de Aperfei{\c c}oamento de Pessoal de N{\'i}vel Superior - Brasil (CAPES) (Finance Code 001), Conselho Nacional de Desenvolvimento Cient{\'i}fico e Tecnol{\'o}gico, and Financiadora de Estudos e Projetos (Finep).
\section*{Disclosures}
The authors declare no conflicts of interest.
\section*{Data availability}
Data underlying the results presented in this paper are available at ZENODO$^\text{\textregistered}$ repository (\href{https://doi.org/10.5281/zenodo.6560537}{10.5281/zenodo.6560537})~\cite{Dataset}, including FEM simulations, scripts files for generating figures and data related to the optimization method.
\section*{Supplemental document}
See Supplement 1 for supporting content.
\section{Optimization process}
\subsection{Nanobeam parameterization}
The geometric parameters of the nanobeam are:
\begin{itemize}
\item $w$ the nanobeam width,
\item $t$ the nanobeam thickness,
\item $N$ the number of defects from the central defect to the mirrors,
\item $M$ the number of holes (defects + mirrors),
\item $a_j$ the j-th cell's lattice parameter,
\item $h_{xj}$ the j-th cell's x-component of the hole diameter,
\item $h_{yj}$ the j-th cell's y-component of the hole diameter.
\end{itemize}
The four first parameters from the list were used as they are and the nanobeam width was the only one from them used as an optimization variable. The last three parameters were parameterized in function of:
\begin{itemize}
\item $a_N$ the mirrors lattice parameter,
\item $d_0$ to define the central defect's lattice parameter,
\item $\gamma_N$ to define the x-component of the hole diameter at the mirror,
\item $\gamma_0$ to define the x-component of the hole diameter at the central defect,
\item $\eta_N$ to define the y-component of the hole diameter at the mirror,
\item $\eta_0$ to define the y-component of the hole diameter at the central defect,
\item $\delta = \SI{50}{nm}$ the fabrication precision limit.
\end{itemize}
To define the geometry of a unit cell between the mirrors and the central defect we define $x_j\equiv j/N$ where $j$ is the cell index counting from the central defect ($j=0$), to the first mirror ($j=8$) and a function $f(x_j)=2x_j^3-3x_j^2+1$, which is limited between 1 and 0 for our domain and has a null derivative at 0 and 1. The j-th lattice parameter is defined as:
\begin{align}
a_j &= (1-d_j)a_N\label{eq:first}\\
d_j &= d_0 f(x_j)
\end{align}
The equations to parameterize $h_{xj}$ and $h_{yj}$ are:
\begin{align}
h_{xj} &= (a_j-2\delta)\eta_j+\delta\\
h_{yj} &= (w-3\delta)\gamma_j+\delta\\
\eta_j &= \eta_N-(\eta_N-\eta_0)f(x_j)\\
\gamma_j &= \gamma_N-(\gamma_N-\gamma_0)f(x_j)\label{eq:last}
\end{align}
They ensure $\delta \leq h_{xj} \leq a_j$ and $\delta \leq h_{yj} \leq w - 2\delta$, which means no hole or wall is smaller than $\delta$. This parameterization requires $\gamma_N$, $\gamma_0$, $\eta_N$, $\eta_0$ and $d_0$ to be bounded between 0 and 1.
\subsection{Simulation}
All simulations were performed in Comsol$^\text{\textregistered}$ using built-in diamond (100) and air as materials. Diamond has refractive index $n=2.417$ and Young's modulus given by $E=(C_{11}-C_{12})(C_{11}+2*C_{12})/(C_{11}+C_{12})=\SI{1050}{GPa}$. The elasticity matrix is defined as:
\begin{equation*}
D=
\begin{pmatrix}
C_{11} & C_{12} & C_{12} & 0 & 0 & 0\\
C_{12} & C_{11} & C_{12} & 0 & 0 & 0\\
C_{12} & C_{12} & C_{11} & 0 & 0 & 0\\
0 & 0 & 0 & C_{44} & 0 & 0\\
0 & 0 & 0 & 0 & C_{44} & 0\\
0 & 0 & 0 & 0 & 0 & C_{44}
\end{pmatrix}
\end{equation*}
with $C_{11}=\SI{1076}{GPa}$, $C_{12}=\SI{125}{GPa}$ and $C_{44}=\SI{578}{GPa}$. We also used the diamond photoelastic coefficients $(\rho_{11},\rho_{12},\rho_{44})=(-0.25,0.043,-0.172)$ to compute the photoelastic component of the optomechanical coupling.
The unit cell simulations account for a quarter of the cell using proper symmetry conditions at the internal boundaries, Floquet periodicity at the cell limits, and scattering boundary conditions (SBC) at the external boundaries as a filter for optical modes. The nanobeam simulations account for one eight of the nanobeam and were performed using perfectly matched layers (PML) instead of SBC at the external boundaries to improve accuracy on losses evaluation.
The geometries were created using the previously described parameterization through Comsol$^\text{\textregistered}$ methods. Optical and mechanical studies are computed independently by another Comsol$^\text{\textregistered}$ method, and the value of the integrals of optical fields are stored so they don't need to be recalculated during the evaluation of the optomechanical coupling for each mechanical mode, which strongly reduces each simulation's computational time. The simulation files are available at Zenodo$^\text{\textregistered}$ through the DOI:~\href{https://doi.org/10.5281/zenodo.6560537}{10.5281/zenodo.6560537}.
\subsection{Initial parameters generation}
The optimization method has two steps: a global optimization with a genetic algorithm (GA) and a multi-start local optimization using a pattern search (PS) method. To improve the efficiency of the GA we fed it with an initial population containing geometries capable of confining optical modes between $\SI{183.9}{THz}$ and $\SI{198.4}{THz}$. To generate such an initial population we first create a file with randomly generated geometries using Dakota LHS sampling with restricted conditions as described in Table~\ref{tab:pars_bounds}. Each generated geometry was tested by simulating first the central defect unit cell with periodic conditions at the transversal plane (yz) and scattering bound conditions at the other planes to identify the expected frequency for the nanobeam confined mode. After this we simulate the whole nanobeam and ran the optical study, looking for modes close to the expected value. At this point, the program discards all geometries that don't have optical modes with $Q_o>300$ and the geometries with low-frequency modes ($\nu_o<183.9$). The geometries with $\nu_o>198.4$ are then rescaled and reparameterized. Geometries with low frequency are not rescaled because they could make walls or holes to be thinner than the fabrication resolution limit.
\begin{table}[hb]
\centering
\begin{tabular}{c c c}
parameter & lower bound & upper bound\\
\hline
$w$ & $\SI{400}{nm}$ & $\SI{800}{nm}$\\
$a_N$ & $\SI{500}{nm}$ & $\SI{950}{nm}$\\
$d_0$ & 0.05 & 0.6\\
$\eta_N$ & 0.1 & 1\\
$\gamma_N$ & 0.1 & 1\\
$\eta_0$ & 0 & 1\\
$\gamma_0$ & 0 & 1\\
\end{tabular}
\caption{Bounds restriction for GA initial population generation.}
\label{tab:pars_bounds}
\end{table}
To scale the geometry to find a confined optical mode in the expected frequency range we estimate an expected scaling factor by dividing the frequency of the mode with the highest quality factor by the maximum allowed frequency and start to test scaling factors $S_f$ in a range between the expected factor more or less 10\%. We multiply all geometric parameters, except the nanobeam thickness and the number of holes, by the scaling factor and if a mode is found with an allowed frequency we reparameterize the geometry in a way the mirrors and the central defect cells are kept the same but the geometry can again be written by equations~\ref{eq:first}-\ref{eq:last}. The reparameterization equations are:
\begin{align}
w' & = w S_f\\
a' & = aS_f\\
\eta_m' & = \frac{(a_N-2\delta)\eta_m S_f+\delta (S_f-1)}{a_N S_f-2 \delta}\\
\eta_0' & = \frac{[a_N*(1-d_0)-2\delta]\eta_0 S_f+\delta (S_f-1)}{a_N*(1-d_0) S_f-2 \delta}\\
\gamma_m' &= \frac{(w-3\delta)\gamma_m S_f+\delta (S_f-1)}{w S_f-3 \delta}\\
\gamma_0' &= \frac{(w-3\delta)\gamma_0 S_f+\delta (S_f-1)}{w S_f-3 \delta}
\end{align}
After this, the new list containing the scaled geometries is used as input parameters for the GA. The full process is described in the flowchart of Fig.~\ref{fig:flowchart}.
\begin{figure*}[ht]
\centering
\includegraphics[width=12.5cm]{flowchart.pdf}
\caption{Flowchart of the optimization process.}
\label{fig:flowchart}
\end{figure*}
\subsection{GA output selection}
The GA randomly mutates geometries looking for solutions with a higher value of the optimization goal $q$. Solutions with a low $q$-value are discarded and the remaining ones mutate again to generate new solutions. After a while, it is common though to reach a branch of solutions with high $q$ derived from the same solution and very similar to each other. Performing local optimizations in many solutions of the same branch is therefore an overkill process. Hence, to find solutions with different optimization strategies, we define clusters of solutions in the parameters space and select only the ones with the highest $q$. Two solutions are considered to belong to the same cluster if the normalized distance between then $D \equiv \sum_{i} [(p_{1i}-p_{2i})/(p_{1i}+p_{2i})]^2$, where $p_{1i}$ and $p_{2i}$ refers to each one of the geometry parameters of the solutions, is smaller than a minimum distance $D_{min}$. To define $D_{min}$ we initially discard all solutions with $q<0.3(q_{max}-q_{min})+q_{min}$, then we start the selection with $D_{min}=0$ and we increase $D_{min}$ by 0.03 until reach less than 20 selected solutions. At least we start to slowly decrease $D_{min}$ again by 0.001 until reach at least 20 selected solutions.
\section{Output geometries}
The parameter values for the geometries presented in the paper are listed in Table~\ref{tab:build_pars}
\begin{table}[hb]
\centering
\begin{tabular}{c c c}
parameter & geometry 1 & geometry 2\\
\hline
$N$ & $8$ & $8$\\
$M$ & $13$ & $13$\\
$t$ & $\SI{450}{nm}$ & $\SI{450}{nm}$\\
$w$ & $\SI{633.9}{nm}$ & $\SI{584.3}{nm}$\\
$a_N$ & $\SI{610.8}{nm}$ & $\SI{569.2}{nm}$\\
$d_0$ & 0.184 & 0.178\\
$\eta_N$ & 0.513 & 0.557\\
$\gamma_N$ & 0.839 & 0.677\\
$\eta_0$ & 0.733 & 0.357\\
$\gamma_0$ & 0.561 & 0.773\\
\end{tabular}
\caption{Parameters of the optimized solutions presented in the paper.}
\label{tab:build_pars}
\end{table}
\section*{Acknowledgement}
This work was supported by S{\~a}o Paulo Research Foundation (FAPESP) through grants 2020/00100-9, 2020/00119-1, 2021/10249-2, 2018/15580-6, 2018/15577-5, 2018/25339-4, Coordena{\c c}\~ao de Aperfei{\c c}oamento de Pessoal de N{\'i}vel Superior - Brasil (CAPES) (Finance Code 001), Conselho Nacional de Desenvolvimento Cient{\'i}fico e Tecnol{\'o}gico, and Financiadora de Estudos e Projetos (Finep).
\end{document} |
\section{Examples and questions}\label{sec-qe}
If the topological group $G$ contains a \v Cech complete subspace, then $G$ is \v Cech complete (Corollary 5.4 \cite{ArhangelskiiChoban2011}).
\begin{question}(A.V.~Arhangelskii)\label{qqe-1-2}
Let $G$ be a right topological group.
Is it true that if $G$ contains a dense \v Cech complete subspace, then $G$ is \v Cech complete?
Is it true that if $G$ is a separable metrizable and contains a dense Polish subspace, then $G$ is Polish?
\end{question}
\begin{example}\label{eqe1} The van Mill space ${\mathbf{S}}$ and the van Dawn space ${\mathbf{T}}$ have the following properties:
\begin{itemize}
\item
are strongly homogeneous \cite{vanMill1980};
\item
absolutely Borel \cite{vanMill1980,vanEngelen1986};
\item
coset sets of some absolutely Borel groups \cite{vanMill2007};
\item
have a dense Polish subspace and are not Polish \cite{vanMill1980};
\item
precompact strongly rectifiable (Theorem \ref{trexscsc-main});
\item
${\mathbf{T}}$ is not a right topological group (Proposition \ref{phsnrtg1});
\item
${\mathbf{S}}$ is not a topological group (Proposition \ref{phsnrtg2});
\end{itemize}
\end{example}
\begin{question}\label{qqe-1-1}
Is the van Mill space ${\mathbf{S}}$ a right topological group?
\end{question}
\begin{question}\label{qqe-1}
Let $X\in\sset{{\mathbf{S}},{\mathbf{T}}}$.
\begin{enumerate}
\item
Is the space $X\times {\mathbf{Q}}$ homeomorphic to a (boolean) (precompact) (right) topological group?
\item
Is the space $X\times {\mathbf{P}}$ homeomorphic to a (boolean) (precompact) (right) topological group?
\item
Is there a \smz/ space $Y$ such that $X\times Y$ is homeomorphic to a (boolean) (precompact) (right) topological group?
\end{enumerate}
\end{question}
Let $\mu$ be the Lebesgue measure on $\R$.
There is a dense subgroup $G\subset \R$ (Theorem 5.1 \cite{vanMill1992},\cite{vanDouwen1984}) for which the condition
\begin{itemize}
\item[]
if $A,B\subset \R$ are open sets and $A\cap G$ is homeomorphic to $B\cap G$, then $\mu(A)=\mu(B)$.
\end{itemize}
\begin{example}\label{eqe2}
\smz/ the group $G$ is not strongly homogeneous, it suffices to take two segments $A,B\subset \R$ of different lengths, whose ends lie in $\R\setminus G$. The group $G$ is not homeomorphic to a precompact rectifiable space, a finite number of shifts of the set $(-1,1)\cap G$ cannot cover $G$ (see Lemma \ref{lrexscs1}).
\end{example}
\begin{question}\label{qqe1}
Which of the following spaces (1) are strongly homogeneous; (2) are homeomorphic to a precompact group; (3) are a pre-compact (strongly) rectifiable space:
$G\times {\mathbf{Q}}$, $G\times P$, $G\times {\mathbf{C}}$, $G^\omega$?
\end{question}
\begin{question}\label{qqe2}
Which of the spaces listed below are homeomorphic to a Boolean group:
$G$, $G\times {\mathbf{Q}}$, $G\times P$, $G\times {\mathbf{C}}$, $G^\omega$?
\end{question}
\begin{question}\label{qqe3}
Let $X$ be the \smz/ space.
\begin{enumerate}
\item
Which of the following spaces are (1) strongly homogeneous; (2) (strongly) (precompactly) rectifiable spaces; (3) is homeomorphic to a right topological group:
$X^\omega$, $X^\omega\times {\mathbf{Q}}$?
\item
Is the space $X^\omega\times {\mathbf{Q}}$ homeomorphic to a topological (boolean, precompact) group?
\item
Is there a \smz/ space $Y$ such that $X\times Y$ is homeomorphic to a (right) topological group?
\end{enumerate}
\end{question}
\begin{question}\label{qqe4}
Let $X$ be a homogeneous \smz/ space.
\begin{enumerate}
\item
Will the space $X$ be a (strongly) (precompact) rectifiable space?
\item
Is the space $X\times {\mathbf{Q}}$ homeomorphic to a (precompact) (right) topological group?
\item
Which of the following spaces are (1) strongly homogeneous; (2) (strongly) (precompactly) rectifiable spaces; (3) is homeomorphic to a right topological group:
$X^\omega$, $X^\omega\times {\mathbf{Q}}$?
\item
Is the space $X^\omega\times {\mathbf{Q}}$ homeomorphic to a topological (boolean, precompact) group?
\end{enumerate}
\end{question}
\begin{question}\label{qqe5}
In the class \smz/ spaces, distinguish the following classes:
rectifiable; strongly straightened; pre-compact straightening; strongly precompactly rectifiable.
\end{question}
\section{Non-Archimedean spaces}\label{sec-nas}
In this section, we prove a more detailed and specialized version of Theorem 7 (version 2) from \cite{grs1997} and apply it to ${\mathbf{C}}$.
Let $X$ be a zero-dimensional space. Denote by $B(X)$ the set of all finite subsets of $X$.
We consider $B(X)$ as an adiative Boolean group with zero $\varnothing$ and the group operation of addition is a symmetric difference.
Denote by $\Gamma(X)$ the family of all open partitions of $X$. For $\gamma\in \Gamma(X)$ we set
\[
H(\gamma)=\set{g\in B(X): |g\cap U|\text{ is even for }U\in\gamma}.
\]
The set $H(\gamma)$ is a subgroup of $B(X)$. Let $\mathcal{G}\subset \Gamma(X)$. Denote by $B_z(X,\mathcal{G})$ the group $B(X)$ with the group topology whose prebase at unity is formed by sets of the form $H(\gamma)$ for $\gamma\in\mathcal{G}$. The set $X$ naturally embeds in $B(X)$: $i: X\to B(X),\ x\mapsto \sset x$. This embedding is topological and $i(X)$ is closed in $B(X)$ if and only if $\bigcup \mathcal{G}$ is a subbase of $X$. In what follows we will identify $X$ and $i(X)$.
Denote by $B_z(X)=B_z(X,\Gamma(X))$. The space $X$ is close embedded in $B_z(X)$. Let $Y\subset X$. Denote
\begin{align*}
B_z^0(X)&=\set{g\in B_z(X): |g|\text{ is even}},
&
B_z^1(X)&=\set{g\in B_z(X): |g|\text{ is odd}},
\\
B_z(X|Y)&=\set{g\in B_z(X): g\subset Y},
\\
B_z^0(X|Y)&=B_z(X|Y) \cap B_z^0(X),
&
B_z^1(X|Y)&=B_z(X|Y) \cap B_z^1(X).
\end{align*}
The set $B_z^0(X)=H(\sset X)$ is a clopen subgroup of index two.
The set $B_z^1(X)=B_z(X)\setminus B_z^0(X)$ is openly closed in $B_z(X)$ and is a coset of the group $B_z^0(X)$ \cite{grs1997}.
Let $X$ be a space with non-Archimedean base ${\mathcal B}$ is a tree by reverse inclusion and some linear order $<$ consistent with base ${\mathcal B}$.
For ${\mathcal B}'\subset {\mathcal B}$ denote by $\lla_*({\mathcal B}')$ the set of maximal by inclusion elements ${\mathcal B}'$. Then
$\lla_*({\mathcal B}') \subset {\mathcal B}'$;
the family ${\mathcal B}'$ is inscribed in $\lla_*({\mathcal B}')$;
$\lla_*({\mathcal B}')$ is an open partition of $\bigcup {\mathcal B}'$.
For a family $\gamma$ of open subsets $X$, denote $\Lambda(\gamma)=\lla_*({\mathcal B}')$, where ${\mathcal B}'=\set{U\in{\mathcal B}: U\subset V\text { for some }V\in\gamma}$. Then $\Lambda(\gamma)$ is inscribed in $\gamma$; $\Lambda(\gamma)$ is an open partition of $\bigcup\gamma$; if $V\in{\mathcal B}$ and $V\subset W$ for some $W\in\gamma$, then $V\subset U$ for some $U\in \Lambda(\gamma)$. If $\gamma$ is an open cover of $X$, then $\Lambda(\gamma)\in\Gamma(X)$ and $\Lambda(\gamma)$ are inscribed in $\gamma$. Therefore, $B_z(X)=B_z(X,{\mathcal P})$, where
\[
{\mathcal P}=\set{\gamma\in \Gamma(X): \gamma \subset {\mathcal B}}.
\]
For an open $W\subset X$, denote $\Lambda(W)=\Lambda(\sset W)$. Then $\Lambda(W)$ is an open partition of $W$; if $V\in{\mathcal B}$ and $V\subset W$, then $V\subset U$ for some $U\in \Lambda(W)$.
Recall the construction of the retraction $r: B_z^1(X)\to X$ from \cite{grs1997}. For $g\in B_z^1(X)$ we set $\Upsilon(g)=\lla_*(\gamma)$, where $\gamma=\set{U\in{\mathcal B}: g\cap U\neq\varnothing \text{ and }|g\cap U|\text{ is even}}$. Let $R(g)=g\setminus \bigcup \Upsilon(g)$ and $r(g)=\min R(g)$. Theorem 7 (version 2) \cite{grs1997} proves that $r$ is a continuous retraction.
\begin{lemma}\label{lnas1}
The mapping $r$ is open. If $Y\subset X$ and $S=B_z^1(X|Y)$, then the mapping
$\restr rS: S\to Y$ is a continuous open retraction of $S$ to $Y$.
\end{lemma}
\begin{proof}
Let $g\in B_z^1(X)$ and $W\subset B_z(X)$ be a neighborhood of the point $g$. Then $g+H(\gamma)\subset W$ for some $\gamma\in \Gamma(X)$. Let $x=r(g)$. Let $x\in V \in \gamma$. The family $\Upsilon(g)$ is finite and $x\notin \bigcup \Upsilon(g)$. Take $U\in{\mathcal B}$ such that $x\in U \subset V\setminus \bigcup \Upsilon(g)$.
Let $y\in U$ and $h=g + \sset{x,y}$. Then $\Upsilon(g)=\Upsilon(h)$, $R(h)=R(g) + \sset{x,y}$ and $r(h)=y$.
Let $M=\set{g + \sset{x,y}: y\in U}$. Then $M\subset g+H(\gamma)$ and $r(M)=U$. Hence $r(g)\in U \subset r(W)$.
Let $g\in Y$, $Q=U\cap Y$ and $L=\set{g + \sset{x,y}: y\in Q}$. Then $L=M\cap S$ and $r(L)=Q$. Hence $r(g)\in Q \subset r(W\cap S)$.
\end{proof}
Lemma \ref{lnas1} and Theorem 7 (version 2) \cite{grs1997} imply the following assertion.
\begin{theorem}\label{tnas1}
Let $X$ be a non-Archimedean space. Then there exists a continuous open retraction $r: B_z^1(X)\to X$ such that for any $Y\subset X$ and $S=B_z^1(X|Y)$ the map $\restr rS:S\to Y$ is a continuous open retraction.
\end{theorem}
If $X$ is a compact space, then $B_z(X)$ is a precompact group. The set $\Gamma(X)$ is countable. Hence $B_z({\mathbf{C}})$ is a precompact \smz/ group.
Let's apply the theorem \ref{tnas1} to ${\mathbf{C}}$.
\begin{theorem}\label{tnas2}
There exists a continuous open retraction $r: B_z^1({\mathbf{C}})\to {\mathbf{C}}$ such that for any $X\subset {\mathbf{C}}$ and $S=B_z^1({\mathbf{C}}|X)$ the map $\restr rS: S\to X$ is a continuous open retraction.
The set $B_z^1({\mathbf{C}})$ is a clopen subset of the precompact \smz/ group $B_z({\mathbf{C}})$. Let $y\in Y$ and
\[
h_y: B_z^1({\mathbf{C}})\to B_z^0({\mathbf{C}}),\ g \mapsto g + y.
\]
The mapping $h_y$ is a homeomorphism and $h_y(B_z^1({\mathbf{C}}|X))=B_z^0({\mathbf{C}}|X)$, i.e. $X$ is an open continuous retract of the precompact Boolean \smz/ group $B_z^ 0({\mathbf{C}}|X)$.
\end{theorem}
\begin{cor}\label{cnas1}
Any \smz/ space is a continuous open retract of a precompact Boolean group.
\end{cor}
\section{Mal'tsev operations on the Cantor set ${\mathbf{C}}$}\label{sec-rexscsc}
\begin{theorem}\label{trexscsc1}
Let $X$ be the \smz/ space. Then $X$ is a pre-compact Mal'tsev space.
Moreover, there is a precompact Mal'tsev operation $\Phi$ on ${\mathbf{C}}$, so that $\Phi(X^3)=X$ for any $X\subset {\mathbf{C}}$.
\end{theorem}
\begin{proof}
Let $r: B_z^1({\mathbf{C}})\to {\mathbf{C}}$ as in Theorem \ref{tnas2}. Let $\Phi(x,y,z)=r(\sset{x,y,z})$. Then $\Phi(x,y,z)\in\sset{x,y,z}$ and hence $\Phi$ is a Mal'tsev operation.
\end{proof}
\begin{theorem}\label{trexscsc-main}
Let $X$ be a strongly homogeneous \smz/ space. Then $X$ is a precompact strongly rectifiable space.
Moreover, there exists a precompact strongly Malcev operation $\Phi$ on $X$, so that $\wh{(X,\Phi)}$ is homeomorphic to ${\mathbf{C}}$.
\end{theorem}
\begin{proof}
There is a compact zero-dimensional extension $bX$ of the space $X$ homeomorphic to ${\mathbf{C}}$, so that for any non-empty open-closed $U,V\subset bX$, $U,V\neq Y$ there exists $f\in\aut{ Y|X}$ so $f(U)=V$, Lemma 2.2 \cite{vanMill2007}.
We will assume that $bX={\mathbf{C}}$.
\begin{lemma}\label{lrexscsc-1}
For every non-empty open-closed $U\subset {\mathbf{C}}$ there exists a homeorphism $f_U:{\mathbf{C}}\to U$, so that $f_U(X)=U\cap X$.
\end{lemma}
\begin{proof}
Let $V_1$ and $V_2$ be a partition of ${\mathbf{C}}$ into two non-empty open-closed subsets and $U_1$ and $U_2$ be a partition of $U$ into two non-empty open-closed subsets. For $i=1,2$, let $f_i: V_i\to U_i$ be a homeorphism such that $f_i(V_i\cap X)=U_i\cap X$. Then $f_U$ is the union of $f_1$ and $f_2$.
\end{proof}
Denote $\cC^*=\mathcal{C}\setminus \sset{\mathbf{C}}$.
For $U,V\in\cC^*$ we set
\[
\fuv{U,V}=f_U\circ f_V^{-1}: V\to U.
\]
It follows from the construction
\begin{itemize}
\item[{\rm ($\Theta_1$)}]
$\fuv{U,V}: V\to U$ is a homeomorphism and $\fuv{U,V}(V\cap X)=U\cap X$ for $U,V\in\cC^*$;
\item[{\rm ($\Theta_2$)}]
$\fuv{U,U}=\operatorname{id}_U$ for $U\in\cC^*$;
\item[{\rm ($\Theta_3$)}]
$\fuv{U,V} \circ \fuv{V,W} =\fuv{U,W}$ for $U,V,W\in\cC^*$.
\end{itemize}
Let $x=\sqnn x\in{\mathbf{C}}$, $k\in\omega$.
Let's put
\begin{align*}
U_k(x)&=\set{y\in{\mathbf{C}}:x|k+1=y|k+1}=B(x|k+1),
\\
V_k(x)&=\set{y\in{\mathbf{C}}:x|k=y|k\text{ and }x(k)\neq y(k)}=U_{k-1}(x)\setminus U_k(x)=B(\tilde x),
\end{align*}
where $\tilde x=(x_0,...,x_{k-1},1-x_k)$. Note that
\begin{itemize}
\item[{\rm ($U_1$)}]
$U_k(x),V_k(x)\in\cC^*$;
\item[{\rm ($U_2$)}]
the family $\set{V_n(x): n\leq k}$ is an open partition of the set ${\mathbf{C}}\setminus U_k(x)$;
\item[{\rm ($U_3$)}]
the family $\set{V_n(x): {n\in\om} }$ is an open partition of the set ${\mathbf{C}}\setminus \sset x$;
\item[{\rm ($U_4$)}]
the family $\set{U_n(x): {n\in\om} }$ is the base at the point $x$;
\item[{\rm ($U_5$)}]
$U_k(x)=U_k(y)$, $V_k(x)=V_k(y)$ and $x\in U_k(y)$ if $y\in U_k(x)$;
\item[{\rm ($U_6$)}]
$V_k(x)=U_k(y)$, $V_k(y)=U_k(x)$ and $x\in V_k(y)$ if $y\in V_k(x)$.
\end{itemize}
For $x,y\in {\mathbf{C}}$ we define the mapping $h_{x,y}\in\aut {\mathbf{C}}$. Let $z\in {\mathbf{C}}$. Let's put
\[
h_{x,y}(z)=\begin{cases}
x,& z=y \\
\fuv {V_k(x),V_k(y)}(z),&\text{if $z\in V_k(x)$ for some $k\in\omega$}
\end{cases}
\]
It follows from the construction
\begin{itemize}
\item[{\rm ($h_1$)}]
$h_{x,y}\in\aut {\mathbf{C}}$ for $x,y\in {\mathbf{C}}$ and $h_{x,y}\in \aut{{\mathbf{C}}|X}$ for $x,y\in X$;
\item[{\rm ($h_2$)}]
$h_{x,x}=\operatorname{id}_{\mathbf{C}}$;
\item[{\rm ($h_3$)}]
$h_{x,y} \circ h_{y,z}=h_{x,z}$;
\item[{\rm ($h_4$)}]
$h_{x,y}(y)=x$;
\item[{\rm ($h_5$)}]
$h_{x,y}(U_k(y))=U_k(x)$ and $h_{x,y}(V_k(y))=V_k(x)$.
\item[{\rm ($h_6$)}]
if $x'\in U_k(x)$ and $y'\in U_k(y)$ then
$\restr{h_{x,y}}{V}=\restr{h_{x',y'}}{V}$, where $V=V_k(y)$.
\end{itemize}
Let's put
\[
\Phi: {\mathbf{C}}^3\to {\mathbf{C}},\quad (x,y,z) \mapsto h_{x,y}(z).
\]
The condition ($\Phi_1$) follows from ($h_2$) and ($h_4$). The condition ($\Phi_2$) follows from ($h_3$).
Hence the operation $\Phi$ is a strong Mal'tsev operation.
Let us prove that $\Phi$ is continuous. Let $x,y,z,u\in {\mathbf{C}}$, $\Phi(x,y,z)=u$, $U$ be a neighborhood of the point $u$. We need to find a neighborhood $W\subset {\mathbf{C}}^3$ of the point $(z,y,z)$, so that $\Phi(W)\subset U$.
Consider the case $y=z$. Then $x=u$.
Then $S=U_k(x)\subset U$ for some $k\in\omega$.
Let $V=U_k(y)$ and $W=S\times V \times V$. Let's check $\Phi(W)\subset U$. Let $(x',y',z')\in W$. From ($U_5$) and ($h_5$) it follows that $h_{x',y'}(V)=S$. Hence $\Phi(x',y',z')\in U$.
Consider the case $y\neq z$. Then $z\in V_m(y)$ for some $m\in \omega$. Let $S=U_m(x)$ and $V=U_m(y)$. Since the mapping $h_{x,y}$ is continuous, then $h_{x,y}(Q')\subset U$ for some neighborhood $Q'$ of the point $z$. Let $Q=Q'\cap V_m(y)$ and $W=S\times V \times Q$. Let's check $\Phi(W)\subset U$. Let $(x',y',z')\in W$. From ($U_6$) it follows that $\restr{h_{x',y'}}Q=\restr{h_{x,y}}Q$. Hence $\Phi(x',y',z')\in U$.
The continuity of $\Phi$ is proved. From ($h_1$) it follows that $\Phi(X^3)=X$.
\end{proof}
\begin{proposition}\label{prexscsc1}
Let $X$ be a zero-dimensional first countable space. Then $X^\omega\times{\mathbf{C}}$ is strongly homogeneous.
\end{proposition}
\begin{proof}
Let $Y=X\times\Delta$. Then the spaces $X^\omega\times{\mathbf{C}}$, $Y^\omega$ and $Y^\omega \times\Delta$ are homeomorphic. Therefore, $Y^\omega$ has a proper open-closed subspace homeomorphic to $Y^\omega$. Proposition 24 (5) \cite{Medini2011} implies that $Y^\omega$ is strongly homogeneous.
\end{proof}
\begin{proposition}\label{prexscsc2}
Let $X$ be the \smz/ space. There is a space \smz/ space $Y$ such that $X\times Y$ is strongly homogeneous and strongly rectifiable.
\end{proposition}
\begin{proof}
Let $Y=X^\omega\times{\mathbf{C}}$. Then, by the Proposition \ref{prexscsc1}, $Y$ is strongly homogeneous. The space $X\times Y$ is homeomorphic to $Y$. It remains to apply the Theorem \ref{trexscsc-main}.
\end{proof}
\section{Introduction}
In this paper, we study what algebraic structures are possible on subspaces of the Cantor set ${\mathbf{C}}$, that is, on separable metrizable zero-dimensional (\smz/) spaces. Any \smz/ space is Mal'tsev, moreover, it is a retract of the topological group, that is, a retral space \cite{grs1997}.
We are primarily interested in Mal'tsev, rectifiable, retral, homogeneous \smz/ spaces, (right) topological groups.
It is also studied when a \smz/ space with algebraic structure can be embedded in \smz/ a compact space with the same structure. In particular, when the \smz/ space can be embedded in ${\mathbf{C}}$ with the algebraic structure preserved.
Main results of the work: each \smz/ space is an open retract of a Boolean precompact \smz/ group (Theorem \ref{tnas1}); every \smz/ strongly homogeneous space is rectifiable (Theorem \ref{trexscsc-main}).
Rectifiable spaces are also characterized as spaces on which there is a Mal'tsev operation with some additional identity ($\Phi_3$), the so-called homogeneous Mal'tsev operations (Definition \ref{drexscs1}).
A strongly Mal'tsev operation is introduced and studied, which defines strongly rectifiable spaces. This class lies between rectifiable spaces and topological groups.
In the \ref{sec-hsnrtg} section, an example of a strongly homogeneous \smz/ space is constructed which do not admit the structure of a right topological group.
\section{Mal'tsev and rectifiable spaces}\label{sec-rexscs}
Let $X$ be a space with the operation $\Phi:X^3\to X$.
Consider the following conditions for the operation $\Phi$:
\begin{itemize}
\item[{\rm ($\Phi_1$)}]
$\Phi(x,y,y)=\Phi(y,y,x)=x$;
\item[{\rm ($\Phi_2$)}]
$\Phi(x,y,\Phi(y,z,u))=\Phi(x,z,u)$;
\item[{\rm ($\Phi_3$)}]
$\Phi(x,y,\Phi(y,x,u))=u$.
\end{itemize}
for $x,y,z,u\in X$.
The $\Phi$ operation is a Mal'tsev operation if ($\Phi_1$) is satisfied.
\begin{definition}\label{drexscs1}
Let $X$ be a space and $\Phi:X^3\to X$ be a continuous map.
We call a mapping $\Phi$ \term{homogeneous Mal'tsev operation} if $\Phi$ satisfies the conditions {\rm ($\Phi_1$)} and {\rm ($\Phi_3$)}.
We call a mapping $\Phi$ a \term{strong Mal'tsev operation} if the conditions {\rm ($\Phi_1$)} and {\rm ($\Phi_2$)} are satisfied for $\Phi$. A space $X$ is said to be \term{strongly rectifiable} if there exists a continuous strong Mal'tsev operation on $X$.
\end{definition}
\begin{proposition}\label{prexscs1}
The space $X$ is rectifiable if and only if there exists a homogeneous Mal'tsev operation $\Phi$ on $X$.
\end{proposition}
\begin{proof}
Assume that $X$ is a rectifiable space. Let the operations $p$ and $q$ define the structure of a homogeneous algebra, $\Psi$ is a rectification, $\Psi(x,y)=(x,p(x,y))$. We put $\Phi(x,y,z)=q(x,p(y,z))$. $q(x,e)=x$ and $p(x,x)=e$ imply $\Phi(x,y,y)=x$. $q(x,p(x,y))=y$ implies $\Phi(y,y,x)=x$. Let's check ($\Phi_3$).
Since $p(y,q(y,p(x,u)))=p(x,u)$ and $q(x,p(x,u))=u$, then
\[
\Phi(x,y,\Phi(y,x,u))=q(x,p(y,q(y,p(x,u))))=u.
\]
Suppose that $\Phi$ satisfies the conditions ($\Phi_1$) and ($\Phi_3$). Take $e\in X$ arbitrarily.
Let $p(x,y)=\Phi(e,x,y)$ and $q(x,y)=\Phi(x,e,y)$. From ($\Phi_1$) follows $p(x,x)=e$ and $q(x,e)=x$.
From ($\Phi_3$) it follows
\begin{align*}
q(x,p(x,y))&=\Phi(x,e,\Phi(e,x,y))=y,
\\
p(x,q(x,y))&=\Phi(e,x,\Phi(x,e,y))=y.
\end{align*}
\end{proof}
\begin{proposition}\label{prexscs2}
If $X$ is strongly rectifiable, then $X$ is rectifiable.
\end{proposition}
\begin{proof}
The condition ($\Phi_3$) follows from ($\Phi_1$) and ($\Phi_2$), it suffices to substitute $z=x$ into ($\Phi_3$). It remains to apply the Proposition \ref{prexscs1}.
\end{proof}
\begin{proposition}\label{prexscs2+1}
Let $Y$ be a Hausdorff space, $X\subset Y=\cl X$, mapping $\wh\Phi: Y^3\to Y$ is continuous, and $\Phi(X^3)=X$.
Denote $\Phi=\restr{\wh\Phi}{X^3}$.
\begin{enumerate}
\item
If $\Phi$ is a Mal'tsev operation, then $\wh\Phi$ is a Mal'tsev operation.
\item
If $\Phi$ is a homogeneous Mal'tsev operation, then $\wh\Phi$ is a homogeneous Mal'tsev operation.
\item
If $\Phi$ is a strong Mal'tsev operation, then $\wh\Phi$ is a strong Mal'tsev operation.
\end{enumerate}
\end{proposition}
\begin{proof}
If the identities ($\Phi_1$), ($\Phi_2$), and ($\Phi_3$) hold on a dense subspace, then they hold everywhere.
\end{proof}
\begin{proposition}\label{prexscs2+2}
Let $X$ be a space with continuous Mal'tsev operation $\Phi$, $Y_1$ and $Y_2$ be compact Hausdorff extensions of $X$, the operation $\Phi$ extends to $Y_1$ and $Y_2$ up to continuous operations $\Phi_1$ and $\Phi_2$, respectively. Then $Y_1$ and $Y_2$ coincide, that is, there exists a homeomorphism $f:Y_1\to Y_2$ for which $\restr f{X}=\operatorname{id}_X$.
\end{proposition}
\begin{proof}
We identify $X$ with $\set{(x,x)\in Y_1\times Y_2: x\in X}$. Let $Y_3$ be the closure of $X$ in $Y_1\times Y_2$, $\pi_i$ be the projection of $Y_1\times Y_2$ onto $Y_i$, $f_i=\restr{\pi_i}{Y_3}$ for $i =1.2$, $\Phi_3=\restr{\Phi_1\times\Phi_2}{X^3}$. Then $\Phi_3(X^3)=X$ and, by the proposition \ref{prexscs2+1}, the operations $\Phi_i$ for $i=1,2,3$ are a Mal'tsev operation and $f_i$ are morphisms with respect to them. Since the spaces $Y_i$ are compact, the mappings $f_i$ are factorial. Therefore, mappings $f_i$ are open (Mal'tsev's
Theorem 4.11 \cite{ReznichenkoUspenskij1998}, \cite{malcev1954}). An open continuous map that is a homeomorphism on a dense set is itself a homeomorphism. Hence the mappings $f_i$ are homeomorphisms. Therefore, the mapping $f=f_1\circ f_2^{-1}: Y_1\to Y_2$ is a homeomorphism.
\end{proof}
\begin{note}\label{nrexscs1}
Propositions \ref{prexscs2+1}, \ref{prexscs2+2} and their proofs remain valid if we consider separately continuous Mal'tsev operations.
\end{note}
\begin{definition}\label{drexscs2}
Let $X$ be a space and $\Phi:X^3\to X$ be a continuous Mal'tsev operation. An operation $\Phi$ is called \term{precompact} if there exists a Hausdoff compact extension $bX$ of the space $X$ onto which the mapping $\Phi$ extends to the mapping $\wh \Phi: bX^3\to bX$.
A compact extension $bX$ is uniquely defined up to isomorphism (Proposition \ref{prexscs2+2}), we denote it by $\wh X$ or $\wh{(X,\Phi)}$.
\begin{enumerate}
\item
A space $X$ is called \term{precompact Mal'tsev} if there exists a continuous precompact Mal'tsev operation on $X$.
\item
A space $X$ is called \term{precompact rectifiable} if there exists a continuous precompact homogeneous Mal'tsev operation on $X$.
\item
We call a space $X$ \term{precompact strongly rectifiable} if there exists a continuous precompact strong Mal'tsev operation on $X$.
\end{enumerate}
\end{definition}
\begin{proposition}\label{prexscs3}
If $X$ is a topological group, then $X$ is strongly rectifiable.
If $X$ is a precompact topological group, then $X$ is precompact and strongly rectifiable.
\end{proposition}
\begin{proof}
We put $\Phi(x,y,z)=xy^{-1}z$. The operation $\Phi$ is the standard Mal'tsev operation on a group. Let's check ($\Phi_2$).
\[
\Phi(x,y,\Phi(y,z,u))=xy^{-1}(yz^{-1}u)=xz^{-1}u=\Phi(x,z,u ).
\]
If $G$ is precompact, then $\Phi$ extends to the completion of $G$.
\end{proof}
\begin{proposition}\label{prexscs4}
Let $X$ be a Mal'tsev precompact space. Then
$w(\wh X)=w(X)=\chi(X)$.
\end{proposition}
\begin{proof}
Let $\tau=\chi(X)$.
The compact Mal'tsev space $\hat X$ is a Dugundzhi compact (Theorem 1 \cite{uspenskij1989}).
The Dugunji compactum is a dyadic compactum.
If a compact dyadic compact contains a dense subspace whose character does not exceed $\tau$, then its weight does not exceed $\tau$ (\cite{efimov1965}).
\end{proof}
\begin{lemma}\label{lrexscs1}
Let $\Phi$ be a precompact homogeneous Mal'tsev operation on the space $X$ and $e\in X$.
For any non-empty open $U\subset X$ there exists a finite set $M\subset X$ so that
$X=\bigcup_{y\in M} \Phi(e,y,U)$.
\end{lemma}
\begin{proof}
Let $W\subset \wh X$ be open and $W\cap X=U$. Take a non-empty open $V\subset \cl V \subset W$. The family $\set{\Phi(e,y,V):y\in\wh X}$ forms an open cover of $\wh X$. Then $\bigcup_{i=1}^n \Phi(e,x_i,V)$ for some $x_1,x_2,...,x_n\in \wh X$. Then $\Phi(e,x_i,V)\subset \Phi(e,y_i,U)$ for some $y_i\in X$. Let $M=\sset{y_1,y_2,...,y_n}$.
\end{proof}
\section{Definitions and notation}
By a space we understand a regular topological space.
The weight and character of the space $X$ will be denoted by $w(X)$ and $\chi(X)$, respectively.
Denote by $\aut X$ the set of all homeomorphisms of the space $X$ onto itself, $\aut{X|M}=\set{f\in\aut X: f(M)=M}$ for $M\subset X $. Denote by $\operatorname{id}_X$ the identity mapping of $X$ onto itself.
A space $X$ is called \term{strongly homogeneous} if every non-empty clopen subset of $X$ is homeomorphic to $X$. Every strongly homogeneous \smz/ space is homogeneous.
If $f:X\to Y$ is a mapping and $A\subset X$, then $\restr fA$ denotes the restriction of $f$ to $A$.
\subsection{Non-Archimedean spaces}
A subset $M$ of a linearly ordered set $X$ is called convex if $a\leq b\leq c$ holds for $a,c\in M$ and $b\in X$, then $b\in M$.
A family of sets ${\mathcal B}$ is called \term{non-Archimedean} if for $U,V\in{\mathcal B}$ the following holds: if $U\cap V\neq\varnothing$, then either $U\subset V$ or $ V\subset U$. \term{Non-Archimedean base} is the base ${\mathcal B}$ of the space $X$, which is a non-Archimedean family. Non-Archimedean bases called bases of rank 1 were introduced and studied in \cite{arhangelskii1962,arhangelskii1963}.
Every non-Archimedean space has a base which is a tree by reverse inclusion (Theorem 2.9 \cite{Nyikos1999}).
Each a base which is a tree by reverse inclusion is a non-Archimedean base.
A linear order $\leq$ on a space $X$ is said to be \term{consistent with a non-Archimedean base ${\mathcal B}$} if any element of $U\in{\mathcal B}$ is convex with respect to the order $\leq$.
\subsection{Cantor set}
Denote $\Delta=\sset{0,1}$, $\D^{<\om}=\bigcup_{n\in\om} \Delta^n$. Here $\Delta^0=\sset\varnothing$.
Denote by ${\mathbf{C}}$ the Cantor set, ${\mathbf{C}}=\Delta^\omega$.
Let $x=\sqnn x\in{\mathbf{C}}$, $k\in\omega$ and $c=\sq{c_n}{n<k}\in\Delta^k \subset \D^{<\om}$. Denote
\begin{align*}
x|k&=(x_0,...,x_{k-1}), &
x(k)&=x_k, &
B(c)&=\set{x\in{\mathbf{C}}:x|k=c}.
\end{align*}
The family \[\mathcal{C}=\set{B(c):c\in \D^{<\om}}\] is the standard non-Archimedean base ${\mathbf{C}}$.
The base $\mathcal{C}$ is a tree by reverse inclusion.
On ${\mathbf{C}}$ we will consider the lexigraphic order: for different $x,y\in {\mathbf{C}}$, $x<y$ if $x|k=y|k$, $x(k)=0$ and $y( k)=1$ for some $k\in\omega$. The lexigraphic order is consistent with the non-Archimedean base $\mathcal{C}$.
Denote $\operatorname{supp} x=\set{{n\in\om}: x(n)=1}$ for $x\in{\mathbf{C}}$, ${\mathbf{Q}}=\set{x\in{\mathbf{C}}: |\operatorname{supp} x|<\omega }$ and ${\mathbf{P}}={\mathbf{C}}\setminus {\mathbf{Q}}$. The space ${\mathbf{Q}}$ is homeomorphic to rational numbers, and the space ${\mathbf{P}}$ is homeomorphic to irrational numbers.
\subsection{Rectifiable, Maltsev, retral spaces}
Let $X$ be a space, $e\in X$, $\Psi: X^2\to X^2$. A mapping is called \term{rectification} if it is a homeomorphism and there exists a mapping $p: X^2\to X$ such that $\Psi(x,y)=(x,p(x,y))$ and $p (x,x)=e$ for $x,y\in X$.
A space $X$ is called \term{rectifiable} if there exists a rectification on $X$.
We define the operation $q: X^2\to X$ by the identity $q(x,p(x,y))=y$. The operation $q$ is continuous and the identities $q(x,e)=x$ and $p(x,q(x,y))=y$ hold for it. Also, $\Psi^{-1}(x,y)=(x,q(x,y))$.
The operations $p$ and $q$ are called \term{homogeneous algebra} \cite{choban1992,ArhangelskiiChoban2010}.
A biternary algebra on a space $X$ is a pair of ternary continuous operations $\alpha,\beta: X^3 \to X$ such that
\[
\alpha (y,y,x)=\alpha(\beta(x,y,z),y,z)=\beta(\alpha(x,y,z),y,z)=x,
\]
for all $x,y,z \in X$ (see \cite{malcev1954}).
For a space $X$ the following conditions are equivalent: $X$ is a rectifiable space;
$X$ is homeomorphic to a homogeneous algebra; there exists a structure of a biternary algebra on $X$ \cite{choban1992}.
A mapping $\Phi: X^3\to X$ is called a \term{Maltsev operation} if $\Phi(x,y,y)=\Phi(y,y,x)=x$ for $x,y\ inX$.
A space $X$ is called \term{Mal'tsev} if there exists a continuous Mal'tsev operation on $X$.
Rectifiable spaces are Maltsev: $\Phi(x,y,z)=q(x,p(y,z))$.
A space $X$ is called \term{retral} if there exists a topological group $G$ into which $X$ is embedded in such a way that $X$ is a retract of $G$ \cite{arkhangelskii1987}. The retral space is Mal'tsev, the reverse is not true \cite{grs1997}.
\section{Homogeneous spaces that are not right topological groups}\label{sec-hsnrtg}
Van Mill in \cite{vanMill1982} constructed a homogeneous space $A$, a subspace of the real numbers, which is not a topological group. In fact, the space $A$ is not a right topological group either.
We call a set $A\subset X$ be a \term{clumsy} if $h(A)\cap A$ is non empty for any homeomorphism $h:X\to X$.
\begin{theorem}\label{thsnrtg1}
Let $X$ be an space and $A\subset X$ be a clumsy subset. If $|A|<|X|$, then $X$ is not a right topological group.
\end{theorem}
\begin{proof}
From the contrary. Take $g\in X \setminus A^{-1}A$. Then $Ag\cap A=\varnothing$. Contradiction.
\end{proof}
Theorem \ref{thsnrtg1} generalizes Lemma 3.4. from \cite{vanMill1982} and the proof of the theorem is much simpler.
There is a countable dense set $E\subset A$ such that $E$ is clumsy in $A$ (Theorem 3.3 \cite{vanMill1982}). Therefore, by the Theorem \ref{thsnrtg1}, the space $A$ from \cite{vanMill1982} is not a right topological group.
In \cite{Banakh2008} Taras Banakh noticed that the zero-dimensional homogeneous first-countable van Douwen compactum \cite{vanDouwen1984} contains a countable clumsy set and, therefore, is not a right topological group.
Let ${\mathbf{T}}={\mathbf{P}}\times {\mathbf{P}}\cup {\mathbf{Q}}\times {\mathbf{Q}}$ is van Douwen space and ${\mathbf{S}}={\mathbf{C}}\times {\mathbf{T}}$ is van Mill space. Spaces ${\mathbf{T}}$ and ${\mathbf{S}}$ are \smz/ strongly homogeneous spaces \cite{vanEngelen1986,vanMill1980}.
\begin{assertion}\label{ahsnrtg1}
${\mathbf{Q}}\times {\mathbf{Q}}$ is clumsy in ${\mathbf{T}}$.
\end{assertion}
\begin{proof}
Assume the contrary, that is, $h({\mathbf{Q}}\times {\mathbf{Q}})\cap {\mathbf{Q}}\times {\mathbf{Q}}=\varnothing$ for some homeomorphism $h:{\mathbf{T}}\to {\mathbf{T}}$.
Let $F={\mathbf{Q}}\times \{0\}$. Then $F$ is closed in ${\mathbf{T}}$ and homeomorphic to ${\mathbf{Q}}$. Hence $h(F)$ is closed in ${\mathbf{P}}\times {\mathbf{P}}$ and homeomorphic to ${\mathbf{Q}}$. Contradiction.
\end{proof}
The Theorem \ref{thsnrtg1} and Statement \ref{ahsnrtg1} imply
\begin{proposition}\label{phsnrtg1}
${\mathbf{T}}$ is not a right topological group.
\end{proposition}
It is not clear about the van Mill space ${\mathbf{S}}$ whether the structure a right topological group is allowed on it.
\begin{proposition}[Corollary 5.4 \cite{ArhangelskiiChoban2011}]\label{phsnrtg2}
If a topological group $G$ contains a dense \v Cech complete subspace, then $G$ is \v Cech complete.
\end{proposition}
Recall that Polish spaces are exactly \v Cech complete separable metrizable spaces.
The space ${\mathbf{S}}$ contains a dense subspace homeomorphic to ${\mathbf{P}}$ and is not Polish. The following sentence follows from the Proposition \ref{phsnrtg2}.
\begin{proposition}\label{phsnrtg3}
${\mathbf{S}}$ is not a topological group.
\end{proposition}
|
\section{Preface and Preliminaries}
Since rendezvous of spacecraft has been a hot topic from the early sixties to the present, a brief motivation from the engineering point of view seems adequate.
The first successful automated docking manoeuvre was carried out in 1967 between \textit{Cosmos 186} and \textit{Cosmos 188} in the context of the Soviet space program.
In the subsequent years more and more such docking manoeuvres became relevant as for bringing crew members to a space station (successful in 1971 for the first time), supplying a space station as well as for service missions for, e.g., the Hubble Space Telescope.
In view of the recent space mission to Mars (rover ``Perseverance" and helicopter ``Ingenuity") due to the finite speed of information travelling through space-time, the farer away from the earth spacecraft are, the more and more \textit{feedback control schemes} become relevant in launching, landing and docking manoeuvres. \\
However, back to earth, in daily life contexts there is use of docking manoeuvres as well, as for instance, in automatic production processes, linking up two parts of a train or charging of autonomous vehicles to name but a few possible fields of application.
\\
Reviewing the current literature on feedback based output reference tracking, one gets the impression that \textit{exact tracking} is an old problem which has only recently been addressed, and \textit{exact tracking in finite time} represents a chapter that has hardly been written yet.
\\
Referring to the results in~\cite{ByrnWill84}, in~\cite{IlchRyan02b} it is shown that the proposed funnel control scheme can achieve global asymptotic stabilization for a class of linear multi-input multi-output (MIMO) systems of relative degree one.
A generalization to a class of nonlinear relative degree one MIMO systems is proposed in~\cite{RyanSang09}.
In~\cite{NuneHsu09} an extended sliding mode controller is proposed which achieves asymptotic tracking of linear single-input single-output (SISO) systems.
This controller is extended to linear MIMO systems in~\cite{NunePeix14}.
In~\cite{Davi13} backstepping is combined with feedback linearization techniques and higher order sliding modes to design a controller which achieves exponential tracking.
In~\cite{OlivPeix13} a high-gain based sliding mode controller is introduced, where the peaking related to the high-gain observer is obviated introducing a dwell-time activation scheme.
This controller achieves asymptotic tracking for a class of nonlinear SISO systems of arbitrary relative degree, where the reference signal is generated by a reference model.
To the price of a discontinuous control, asymptotic tracking for nonlinear MIMO systems is achieved in~\cite{VergDima19,VergDima21}.
In~\cite{LeeTren19} a funnel control scheme is proposed, which achieves asymptotic tracking for a class of nonlinear relative degree one MIMO systems.
This result is extended in the recent work~\cite{BergIlch21}, where it is shown that the proposed funnel controller achieves asymptotic tracking of nonlinear MIMO systems with arbitrary relative degree whereas the tracking error has prescribed transient behaviour.
\\
Now, we turn from asymptotic tracking towards exact tracking in finite time.
In~\cite{BasiPana16} sliding mode control concepts and results from~\cite{BhatBern05} are used to establish control schemes which achieve finite time stabilization for linear (SISO \& MIMO) systems.
In~\cite{EstrFrid16} backstepping and higher order sliding mode control are combined to construct a control scheme which achieves exact output tracking in finite time for nonlinear MIMO systems in nonlinear block controllable form.
Similar to the prescribed performance controller in~\cite{DimaBech20} this controller suffers from the proper initialization problem, where it is not clear how large to choose the involved parameters.
The controller in~\cite{EstrFrid16}, along with limiting conditions on the system class, presumes knowledge of the system's functions and explicitly involves inverses of some.
In~\cite{YangDing18} a controller is introduced which achieves exact tracking in finite time for a class of nonlinear SISO systems satisfying a certain homogeneity assumption. This controller relies on estimating techniques of the external disturbances, where the problem of proper initialization is avoided by assuming explicit knowledge of the bounds of the disturbances and the reference.
The control scheme explicitly involves (parts of) the system's right hand side and is of relatively high complexity.
As far as we understand, in the control schemes for exact tracking in finite time discussed above, the final time, this is the time when the desired state should be matched, cannot be prescribed; only the existence of such a finite time is ensured.
Contrary, in~\cite{RodrSanc17} a control scheme is introduced which solves a \textit{predefined-time exact tracking problem} for the class of fully actuated mechanical (relative degree two) systems.
The controller relies on a backstepping procedure and consists of a \textit{predefined-time stabilization function} and involves the system's equations explicitly.
In~\cite{Tren19} a funnel controller is introduced which achieves asymptotic tracking as well as convergence to zero of the tracking error in finite time for a class of relative degree one SISO systems.
\\
Circumventing the drawbacks mentioned above, in the present article we introduce a funnel control scheme which achieves \textit{exact tracking in predefined finite time}.
Since the control scheme is of funnel type it inherits the advantages of robustness with respect to noise, being model free, i.e., no knowledge of the system's parameters (roughly speaking no knowledge of the right hand side) is assumed - except the relative degree $r \in \mathbb{N}$ and the input/output dimension $m \in \mathbb{N}$ -, and moreover, the tracking error evolves within prescribed bounds.
The system class under consideration is the same as in~\cite{BergIlch21}.
\subsection{Nomenclature}
Throughout the present article we use the following notation, where $I \subseteq \mathbb{R}$ denotes an interval.
$\mathbb{N}$ the set of positive integers;
$[a,b], [a,b), (a,b)$ a closed, half-open, and open interval for~$a,b \in \mathbb{R}$,~$a < b$;
$\mathbb{R}_{\geq 0} := [0, \infty)$;
$ \langle \cdot, \cdot \rangle $ inner product in~$\mathbb{R}^n$;
$\| x \| := \sqrt{ \langle x, x \rangle }$ Euclidean norm of $x \in \mathbb{R}^n$;
$\mathcal{L}_{\rm loc}^\infty (I ; \mathbb{R}^p)$ the set of locally essentially bounded functions $ f: I \to \mathbb{R}^p$;
$\mathcal{L}^\infty (I ; \mathbb{R}^p)$ the set of essentially bounded functions $ f: I \to \mathbb{R}^p$ ;
$\| f \|_{\infty} := \esssup_{t \in I} \|f(t)\| $ norm of $f \in \mathcal{L}^\infty(I ; \mathbb{R}^p)$;
$\mathcal{W}^{k,\infty}(I ; \mathbb{R}^p)$ the set of $k$-times weakly differentiable functions $ f : I \to \mathbb{R}^p$ such that $f,\dot f,\ldots,f^{(k)} \in \mathcal{L}^\infty(I ; \mathbb{R}^p)$;
$\mathcal{C}^k( I ; \mathbb{R}^p) $ the set of $k$-times continuously differentiable functions $f : I \to \mathbb{R}^p$, $\mathcal{C}(I; \mathbb{R}^p) = \mathcal{C}^0(I; \mathbb{R}^p)$;
$f|_{J}$ the restriction of $f : I \to \mathbb{R}^n$ to $J \subseteq I$, $I,J$ intervals.
\subsection{System class}
To introduce the class of systems under consideration properly, first of all we provide some definitions.
We introduce a class of operators to which the operator appearing in the system class under consideration belongs.
This definition is taken from~\cite[Sec.~1.2]{BergIlch21}.
\begin{Definition} \label{Def:OP-class}
If the operator $\textbf{T} : \mathcal{C}([-\sigma,\infty) ; \mathbb{R}^{n}) \to \mathcal{L}_{\rm loc}^\infty(\mathbb{R}_{\geq 0} ; \mathbb{R}^q)$ has following properties
\begin{enumerate}[label = (\alph{enumi}), ref=(\alph{enumi}),leftmargin=*]
\item $\textbf{T}$ maps bounded trajectories to bounded trajectories, i.e., for all $c_1 > 0$, there exists $c_2>0$ such that for all
$\xi \in \mathcal{C}([-\sigma,\infty) ; \mathbb{R}^n)$,
\[
\sup_{t \in [-\sigma,\infty)} \| \xi(t) \| \le c_1 \ \Rightarrow \ \sup_{t \in [0,\infty)} \| \textbf{T}(\xi)(t) \| \le c_2,
\]
\item $\textbf{T}$ is causal, i.e., for all $t \ge 0$ and all $\zeta, \xi \in \mathcal{C}([-\sigma,\infty) ; \mathbb{R}^n)$,
\[
\zeta |_{[-\sigma,t)} = \xi|_{[-\sigma,t)} \ \Rightarrow \ \textbf{T}(\zeta)|_{[0,t)} \overset{a.a.}{=} \textbf{T}(\xi)|_{[0,t)},
\]
\item $\textbf{T}$ is locally Lipschitz continuous in the following sense: for all $t \ge 0 $ and all $\xi \in \mathcal{C}([-\sigma,t] ; \mathbb{R}^n)$
there exist $\Delta, \delta, c > 0$ such that for all
$\zeta_1, \zeta_2 \in \mathcal{C}([-\sigma,\infty) ; \mathbb{R}^n)$ with $\zeta_1|_{[-\sigma,t]} = \xi $, $\zeta_2|_{[-\sigma,t]} = \xi $
and $\|\zeta_1(s) - \xi(t)\| < \delta$, $\|\zeta_2(s) - \xi(t)\| < \delta$ for all $s \in [t,t+\Delta]$ we have
\begin{align*}
\esssup_{s \in [t,t+\Delta]} & \| \textbf{T}(\zeta_1)(s) - \textbf{T}(\zeta_2)(s) \| \\
& \le c \ \rm{sup}_{s \in [t,t+\Delta]} \| \zeta_1(s) - \zeta_2(s)\| ,
\end{align*}
\end{enumerate}
then we say the operator~$\textbf{T}$ belongs to the operator class~$\mathcal{T}_\sigma^{n,q}$.
\end{Definition}
To introduce the notion of a \textit{solution} to an initial value problem, let~$\mathcal{D} \subseteq \mathbb{R}_{\geq 0} \times \mathbb{R}^{n} $ a non-empty, connected and relatively open set in~$\mathbb{R}_{\geq 0} \times \mathbb{R}^n$.
Let~$q \in \mathbb{N}$ and assume that $F: \mathcal{D} \times \mathbb{R}^q \to \mathbb{R}^n$ is a \textit{Carathéodory function} as in~\cite[App.~B]{IlchRyan09}.
Let~$\textbf{T} \in \mathcal{T}_\sigma^{n,q}$, $t_0 \in \mathbb{R}_{\geq 0}$ and consider the initial value problem
\begin{align} \label{eq:ODE-t}
\tfrac{\text{\normalfont d}}{\text{\normalfont d}t} x(t) &= F \big( t,x(t),\textbf{T}(x)(t) \big), \\
x|_{[-\sigma,t_0]} &= x^0 \in \mathcal{C}([-\sigma,t_0] ; \mathbb{R}^n), \ (t_0, x^0(t_0)) \in \mathcal{D}. \nonumber
\end{align}
Then, in virtue of~\cite[App. B]{IlchRyan09} we have the following definition.
\begin{Definition}
A \textit{solution} of~\eqref{eq:ODE-t} is a function $x \in \mathcal{C}(I ; \mathbb{R}^n)$ on an interval of the form~$ I = [-\sigma, \bar \omega] $, for $t_0 < \bar \omega < \infty$,
or ${I = [-\sigma, \omega)}$, for $\omega \le \infty$, such that, for $J := I\setminus [-\sigma, t_0]$, we have $x|_{[-\sigma,t_0]} = x^0$, $x|_J$ is absolutely continuous with
$(t,x(t)) \in \mathcal{D}$ for all $t \in J$, and
\begin{equation} \label{def:solution}
\forall\, t \in J \,: \ x(t) = x(t_0) + \int_{t_0}^{t} F\big(s,x(s),\textbf{T}(x)(s)\big) \text{\normalfont d}s .
\end{equation}
A solution is $ \rm maximal$ if it has no proper right extension that is also a solution.
\end{Definition}
\begin{Remark} \label{Rem:localisation}
As pointed out in~\cite[p.~234]{Ilch13} the interpretation of the ``local Lipschitz property" in Definition~\ref{Def:OP-class}~$(c)$ and so the interpretation of a solution $x \in \mathcal{C}(I ; \mathbb{R}^{n})$ of~\eqref{eq:ODE-t},
where~$I = [-\sigma, \tilde \omega]$ or~$I = [-\sigma,\omega)$ for finite~$\omega$, requires special attention since the operator~$\textbf{T}$ has~$\mathcal{C}([-\sigma,\infty) ; \mathbb{R}^{n})$ as its domain.
Let~$J := I\setminus [-\sigma, 0)$ and~$x \in \mathcal{C}(I ; \mathbb{R}^{n})$.
For each~$\tau \in J$ define the function~$x_\tau \in \mathcal{C}([-\sigma,\infty) ; \mathbb{R}^{n})$ by
\begin{equation*}
x_\tau(t) =
\begin{cases}
x(t), & t \in [-\sigma,\tau], \\
x(\tau), & t>\tau.
\end{cases}
\end{equation*}
Now, we associate with the operator~$\textbf{T} \in \mathcal{T}_{\sigma}^{n,q}$ its ``localisation", namely the operator $\tilde{\textbf{T}} : \mathcal{C}(I ; \mathbb{R}^{n}) \to \mathcal{L}_{\rm loc}^{\infty}(J ; \mathbb{R}^{n})$ defined as
\begin{equation*}
\forall \, \tau \in J \, : \ \tilde{\textbf{T}}(x)|_{[0,\tau]} = \textbf{T}(x_\tau)|_{[0,\tau]}.
\end{equation*}
Note, that by Definition~\ref{Def:OP-class}~$(b)$ the operator~$\tilde{\textbf{T}}$ is well defined.
Together, a replacement of~$\textbf{T}$ by~$\tilde{\textbf{T}}$ in~\eqref{def:solution} yields the correct interpretation of a solution; however, in what follows we will (notationally) not distinguish between $\textbf{T} \in \mathcal{T}_{\sigma}^{n,q}$ and~$\tilde{\textbf{T}}$.
\end{Remark}
Throughout the present article we consider the following functional differential equation
\begin{equation} \label{eq:system}
\begin{aligned}
y^{(r)}(t) &= f\left( d(t), \textbf{T}(y,\dot y,\ldots, y^{(r-1)})(t), u(t) \right), \\
y|_{[-\sigma,0]} &= y^0 \in \mathcal{C}^{r-1}([-\sigma,0] ; \mathbb{R}^m),
\end{aligned}
\end{equation}
where~$\sigma>0$ is the ``memory" of the system, the parameter~$r \in \mathbb{N}$ is related to the concept of relative degree; furthermore, the dimension~$m \in \mathbb{N}$ of the input~$u$ and the output~$y$ of the system coincide.
\\
In order to characterize the class of admissible nonlinearities~$f$ in~\eqref{eq:system} we recall the definition of ``high-gain property" from~\cite[Sec.~1.2]{BergIlch21}.
\begin{Definition} \label{Def:high-gain-prop}
For~$p,q,m \in \mathbb{N}$ a function~$f \in \mathcal{C}(\mathbb{R}^p \times \mathbb{R}^q \times \mathbb{R}^m ; \mathbb{R}^m)$ satisfies the \textit{high-gain property}, if there exists~$\rho \in (0,1)$ such that, for every compact~$K_p \subset \mathbb{R}^p$ and compact~$K_q \subset \mathbb{R}^q$ the continuous function
\begin{equation*}
\begin{aligned}
\chi : \mathbb{R} & \to \mathbb{R}, \\
s &\mapsto \min \setdef{ \langle v, f(\delta,z,-s v \rangle }{ \begin{array}{l}
(\delta,z) \in K_p \times K_q, \\
v \in \mathbb{R}^m, \rho \le \| v \| \le 1
\end{array} }
\end{aligned}
\end{equation*}
is such that $\sup_{s \in \mathbb{R}} \chi(s) = \infty$.
\end{Definition}
For a detailed discussion and equivalent conditions of the high-gain property we refer to~\cite[Rem.~1.3~\&~1.4]{BergIlch21}.
Now, with the definitions so far, we may introduce the system class under consideration, which is the same class of systems as in~\cite{BergIlch21}.
\begin{Definition}
For~$m,r \in \mathbb{N}$ a system~\eqref{eq:system} is said to belong to the system class~$\mathcal{N}^{m,r}$ if for some~$p,q \in \mathbb{N}$ the ``disturbance" is bounded, i.e., $d \in \mathcal{L}^\infty(\mathbb{R}_{\geq 0} ; \mathbb{R}^p)$, the function $f \in \mathcal{C}(\mathbb{R}^p \times \mathbb{R}^q \times \mathbb{R}^m ; \mathbb{R}^m)$ satisfies the high-gain property from Definition~\ref{Def:high-gain-prop} and for~$\sigma > 0$ the operator~$\textbf{T}$ belongs to~$\mathcal{T}^{rm,q}_{\sigma}$; we write~$(d,f,\textbf{T}) \in \mathcal{N}^{m,r}$.
\end{Definition}
\subsection{Control objective and feedback law}
We aim to find a control structure which achieves exact output reference tracking in the following sense.
For every system~\eqref{eq:system} belonging to the class~$\mathcal{N}^{m,r}$ and any reference trajectory~$y_{\rm ref} \in \mathcal{W}^{r,\infty}([0,T) ; \mathbb{R}^m)$, $T > 0$, the output approaches the reference within the interval~$[0,T)$, whereas for a given function~$\varphi(\cdot)$ introduced next, the error~$e(\cdot) := y(\cdot) - y_{\rm ref}(\cdot)$ evolves within the so called ``performance funnel" defined by
\begin{equation*}
\mathcal{F}_\varphi := \setdef{ (t,e) \in [0,T) \times \mathbb{R}^m}{ \varphi(t) \|e\| < 1 }.
\end{equation*}
Moreover, the control achieves exact tracking in finite time with predefined accuracy, this is
\begin{equation} \label{eq:control-objective}
\begin{aligned}
& \forall \, i=0,\ldots,r-1 , \forall \, t \in [0,T) \, : \\
& (t,e(t)) \in \mathcal{F}_\varphi, \text{ and } \lim_{t \to T} \|e^{(i)}(t)\| = 0,
\end{aligned}
\end{equation}
where~$e^{(i)}(\cdot)$ denotes the~$i^{\rm th}$ derivative of~$e(\cdot)$.
In order to establish the control structure, we introduce the following funnel control parameters.
Choose the final time $T > 0$ and some~$c>0$, then we define the funnel function
\begin{subequations} \label{def:control_functions}
\begin{equation} \label{def:vp}
\varphi(t) = \frac{1}{c} \frac{1}{T-t}, \ t \in [0,T),
\end{equation}
where we highlight~$\lim_{t \to T} \varphi(t) = \infty$, further choose
\begin{equation}
N \in \mathcal{C}(\mathbb{R}_{\geq 0} ; \mathbb{R}) \ \text{a surjection},
\end{equation}
and for~$c>0$ from~\eqref{def:vp} choose
\begin{equation} \label{def:alp}
\alpha \in \mathcal{C}^{r-1}( [0,1) ; [c(r+1) , \infty)) \ \text{a bijection},
\end{equation}
\end{subequations}
e.g. $\alpha(s) = c(r+1)/(1-s)$.
For a detailed comment on the surjection~$N$ and possible simplifications we recommend reading~\cite[Rem.~1.8]{BergIlch21}.
\begin{Remark} \label{Rem:choice-vp}
Compared to the funnel control schemes proposed in the common literature, see e.g.~\cite{IlchRyan02b,Ilch13,BergLe18a,BergIlch21} to name but a few, the explicit choice of~$\varphi(\cdot)$ in~\eqref{def:vp} seems to be quite restrictive.
Anticipating the initial condition~\eqref{eq:initial} in Theorem~\ref{Thm:Exact-tracking-in-finte-time}, the choice of the funnel function~$\varphi(\cdot)$ in~\eqref{def:vp} reflects the intuition that the shorter the final time~$T$ is chosen, the better the initial guess has to be.
Further, the tuning parameter~$c$ in~\eqref{def:vp} links the funnel function to the gain function~$\alpha(\cdot)$ in~\eqref{def:alp}.
If the final time~$T$ is short and the error between the initial state of system~\eqref{eq:system} and the reference, however, can not be influenced to be small, a large~$c$ can ensure the initial condition~\eqref{eq:initial} to be satisfied; this in turn causes a larger lower bound of the gain function~$\alpha(\cdot)$ which means that small tracking errors cause high input values.
\end{Remark}
\begin{Remark}
A careful inspection of the proof of~\cite[Thm.~1.9]{BergIlch21} yields that a crucial property of the feasible funnel functions is a growth condition, namely
\begin{equation*}
\exists \, d > 0 \, : \ |\dot \varphi(t) | \le d \big( 1+\varphi(t) \big) \text{ for almost all } t \ge 0.
\end{equation*}
This condition prevents a ``blow up" in finite time, i.e., $\varphi(\cdot)$ is bounded on $\mathbb{R}_{\geq 0} \setminus \{\infty\}$.
With this, however, exact tracking in finite time via funnel control is impossible.
Contrary, the funnel functions~$\varphi(\cdot)$ defined in~\eqref{def:vp} do not satisfy this growth condition.
Therefore, in order to achieve exact tracking in finite time via funnel control, we propose a modified feedback law in the present article.
\end{Remark}
We introduce the control law which achieves the previously formulated control objective~\eqref{eq:control-objective}.
For~$i=0,\ldots,r-1$ we set~$e^{(i)}(\cdot) := y^{(i)}(\cdot) - y_{\rm ref}^{(i)}(\cdot)$ and recursively define for~$k=1,\ldots,r$ with~$\alpha(\cdot)$ from~\eqref{def:alp} and~$\gamma_0(\cdot) \equiv 0$ the functions
\begin{subequations} \label{def:control_variables}
\begin{align}
e_k(t) &= \varphi(t) e^{(k-1)}(t) + \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-1-i)}(t), \label{def:e_k} \\
\alpha_k(t) &= \alpha(\|e_k(t)\|^2), \label{def:alp_k} \\
\gamma_k(t) &= \alpha_k(t) e_k(t). \label{def:gam_k}
\end{align}
\end{subequations}
Then, with the functions introduced in~\eqref{def:control_functions},\eqref{def:control_variables} we define the feedback law $u : \mathbb{R}_{\geq 0} \to \mathbb{R}^m$ as
\begin{equation} \label{def:u}
u(t) = (N \circ \alpha_r)(t) \, e_r(t) = (N \circ \alpha)(\|e_r(t)\|^2) \, e_r(t) .
\end{equation}
\begin{Remark} \label{Rem:Gam_k}
We comment on the computation of the expressions recursively given in~\eqref{def:control_variables}.
We define the set
\begin{equation*}
\mathcal{D}_{0} := \setdef{ \zeta \in \mathbb{R}^m }{ \| \zeta \| < 1 } ,
\end{equation*}
and, for~$\alpha(\cdot)$ from~\eqref{def:alp}, the function~$\Gamma_{0}: \mathcal{D}_{0} \to \mathbb{R}^m$ by
\begin{equation*}
\Gamma_{0}(t,\zeta) := \alpha(\|\zeta\|^2) \, \zeta.
\end{equation*}
For~$k=1,\ldots,r-1$ we recursively define the sets~$\mathcal{D}_{k}$ and the functions $\Gamma_{k} : [0,T) \times \mathcal{D}_{k} \to \mathbb{R}^m$ by
\begin{align} \label{def:Gam_k}
\mathcal{D}_k &:= \underbrace{\mathcal{D}_0 \times \cdots \times \mathcal{D}_0}_{k-\rm times} \times \mathbb{R}^m, \nonumber \\
& \Gamma_{k}(t,\zeta_1,\ldots,\zeta_{k+1}) := \frac{\partial \Gamma_{k-1}(t,\zeta_1,\ldots,\zeta_{k})}{\partial t} \\
& \!+\! \sum_{i=1}^{k} \frac{\partial \Gamma_{k-1}(t,\zeta_1,\ldots,\zeta_{k})}{\partial \zeta_i}
\Big( \! \varphi(t) \left( c \zeta_i \! -\! \Gamma_{0}(\zeta_i) \right) \! +\! \zeta_{i+1} \!\Big). \nonumber
\end{align}
Then, with~$e_k(\cdot)$ from~\eqref{def:e_k} and $\gamma_k(\cdot)$ from~\eqref{def:gam_k} we obtain
\begin{equation*}
\gamma_k^{(j)}(t) = \Gamma_j \big( t,e_k(t),\ldots,e_{k+j}(t) \big), \ 0 \le j \le r-k,
\end{equation*}
which can be seen via a brief induction over~$k$ using equations~\eqref{def:control_variables}.
\end{Remark}
\begin{Remark} \label{Rem:Control-structure-as-in-BergLe18a}
Due to the recursive structure, the implementation of the control scheme~\eqref{def:u} in the present article is not as simple as for the control scheme in~\cite[Eq.~(9)]{BergIlch21}.
However, with the explicit recursion~\eqref{def:Gam_k} given in Remark~\ref{Rem:Gam_k} calculation of the required expressions can be done completely algorithmically.
\begin{comment}
Comparing the control scheme~\eqref{def:u} with the control scheme in~\cite[Eq.~$(5)$]{BergLe18a} a certain similarity is recognizable; both control schemes involve recursively defined signals from which the input is constructed.
In view of the findings in the recent work~\cite{BergIlch21} and the resulting simplified control scheme this similarity gives reason to hope that as well for the control objective formulated in the present article there is a simpler (non-recursive) control scheme.
\end{comment}
\end{Remark}
We complete the preliminaries by recording the following lemma.
\begin{Lemma} \label{Lem:x_le_M}
If for~$x \in \mathcal{C}^1([\tau,T) ; \mathbb{R}^m)$, $m \in \mathbb{N}$, there exists~$M \ge \|x(\tau)\| \ge 0$ such that
\begin{equation} \label{eq:ddt_x_le_0}
\forall \, t \in [\tau,T) \, : \ \left(
\|x(t)\| \ge M \ \Rightarrow \ \tfrac{\text{\normalfont d}}{\text{\normalfont d}t} \|x(t)\|^2 \le 0 \right),
\end{equation}
then
\begin{equation} \label{eq:x_le_M}
\forall\, \ t \in [\tau,T) \, : \ \|x(t)\| \le M.
\end{equation}
\end{Lemma}
\begin{proof}
The proof follows the ideas in~\cite[Thm.~4.3]{Lanz21}.
Seeking a contradiction we assume that there exists~$t_1 \in (\tau,T)$ such that $\|x(t_1)\| > M$.
Then, by continuity, there exists $t_0 := \max \setdef{ t \in [\tau,t_1)}{ \|x(t)\| = M}$, and hence we have $\|x(t)\| \ge M$ for all~$t \in [t_0,t_1]$.
Then, using~\eqref{eq:ddt_x_le_0} we obtain
\begin{equation*}
\|x(t_1)\|^2 - \|x(t_0)\|^2 = \int_{t_0}^{t_1} \left( \tfrac{\text{\normalfont d}}{\text{\normalfont d}t} \tfrac{1}{2} \|x(t)\|^2 \right) \text{\normalfont d}t \le 0,
\end{equation*}
and we find the contradiction
\begin{equation*}
M^2 < \|x(t_1)\|^2 \le \|x(t_0)\|^2 = M^2.
\end{equation*}
Therefore,~\eqref{eq:x_le_M} holds for all~$t \in [\tau,T)$.
\end{proof}
\section{Main result} \label{Sec:Main-result}
In this section we show that the application of the control scheme~\eqref{def:u} to a system~\eqref{eq:system} yields an initial value problem that has a solution; the input and output signals are bounded and in particular, we show that the introduced control scheme achieves exact output tracking in finite time with prescribed transient behaviour of the tracking error.
As highlighted in earlier works, see e.g.~\cite{Ilch13,BergLe18a,BergIlch21} some care is required when showing boundedness of the involved signals since the bijection~$\alpha(\cdot)$ may introduce a singularity.
Moreover, in the present context expressions involving the unbounded funnel function~$\varphi(\cdot)$ demand particularly high attention.
\begin{Theorem} \label{Thm:Exact-tracking-in-finte-time}
Consider a system~\eqref{eq:system} with~$(d,f,\textbf{T}) \in \mathcal{N}^{m,r}$ and initial data~$y^0 \in \mathcal{C}^{r-1}([-\sigma,0] ; \mathbb{R}^m)$, for $m,r \in \mathbb{N}$.
Let~$T>0$ and~$y_{\rm ref} \in \mathcal{W}^{r,\infty}([0,T) ; \mathbb{R}^m)$.
Assume that for the control design parameters introduced in~\eqref{def:control_functions} and~$e_k(\cdot)$ defined in~\eqref{def:control_variables} the initial condition
\begin{equation} \label{eq:initial}
\forall\, k=1,\ldots,r \, : \ \|e_k(0)\| < 1
\end{equation}
is satisfied.
Then, the funnel control scheme~\eqref{def:u} applied to~\eqref{eq:system} yields an initial value problem which has
a solution and every maximal solution~$y: [-\sigma, \omega) \to \mathbb{R}^m$ has the following properties
\begin{enumerate}[label=\roman*)]
\item $\omega = T$,
\item $u \in \mathcal{L}^\infty([0,T) ; \mathbb{R}^m)$, $y \in \mathcal{W}^{r,\infty}([-\sigma, T) ; \mathbb{R}^m)$,
\item the tracking error~$e(t) = y(t) - y_{\rm ref}(t)$ evolves within the performance funnel~$\mathcal{F}_\varphi$, this is,
\begin{equation*}
\forall \, t \in [0,T) \,: \ \varphi(t) \| e(t) \| < 1,
\end{equation*}
\item the tracking of the reference and its derivatives is exact at~$t=T$, this is,
\begin{equation*}
\forall\, i=0,\ldots,r-1 \, : \ \lim_{t \to T} \|e^{(i)}(t)\| = 0.
\end{equation*}
\end{enumerate}
\end{Theorem}
The proof is relegated to the appendix.
\begin{Remark}
Assertion~$i)$ in Theorem~\ref{Thm:Exact-tracking-in-finte-time}, namely~$[0,T)$ being the maximal solution interval, naturally raises the question of a global solution in time.
If the system's equations~\eqref{eq:system} are available and~$y_{\rm ref}(\cdot)$ is defined on~$\mathbb{R}_{\geq 0}$, then $y(t) - y_{\rm ref}(t) \equiv 0$ for all time~$t \ge T$ can be achieved by asking the reference to satisfy~\eqref{eq:system} for~$ t \ge T$, with~$u(\cdot) \equiv 0$ and ``initial" conditions~$y_{\rm ref}(T) = y(T), \dot y_{\rm ref}(T) = \dot y(T),\ldots, y_{\rm ref}^{(r-1)}(T) = y^{(r-1)}(T) $.
\end{Remark}
\begin{Remark} \label{Rem:practical-exact}
We highlight the following aspect of Theorem~\ref{Thm:Exact-tracking-in-finte-time} which is of interest from a practical point of view.
For any given~$\varepsilon > 0$ there exists a time~$T_\varepsilon < T$ such that the norm of each of the first~$r-1$ derivatives of the error~$e(\cdot) = y(\cdot) - y_{\rm ref}(\cdot)$ is smaller than (or equal to)~$\varepsilon$ for all~$t \in [T_\varepsilon , T)$, this is, for all~$k=0,\ldots,r-1$
\begin{equation*}
\forall \, \varepsilon > 0 \ \exists \, T_\varepsilon < T \ \forall t \in [T_\varepsilon, T) \, : \ \|e^{(k)}(t)\| \le \varepsilon;
\end{equation*}
in fact we have~$\|e^{(k)}(t)\| < \varepsilon$ for all~$t \in (T_\varepsilon, T)$.
This property is relevant, e.g. if during a docking manoeuvre the demanded accuracy changes.
\end{Remark}
\section{Example} \label{Sec:Simulation}
In this section we simulate an application of the control scheme~\eqref{def:u} to a system~\eqref{eq:system} with $(d,f,\textbf{T}) \in \mathcal{N}^{m,r}$.
As indicated in the introduction, docking of two spaceships is a red-hot application of the findings of the present article.
Therefore, we look at such a situation as an example.
Consider a passive target space station, e.g., the international space station~ISS, in a circular orbit around the earth and an active spacecraft chasing the first.
We assume the passive space station to be on a constant altitude~$r_s$ with constant angular velocity~$\omega = \sqrt{\mu / (r_e + r_s)^3}$, where~$\mu \approx 3.986 \cdot 10^{14} \, \rm{m^3/s^2}$ is the standard gravitational parameter and~$r_e = 6378137 \, \rm{m}$ the radius of the earth.
To analyse the motion of the spacecraft, we use Hill's local-vertical-local-horizontal coordinate frame introduced in~\cite{Hill78}, see Figure~\ref{Fig:Hills_Frame}.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.22]{coordinate_frame.jpg}
\caption{Hill's local-vertical-local-horizontal coordinate frame.}
\label{Fig:Hills_Frame}
\end{center}
\end{figure}
Within this frame we use the commonly used Clohessy-Wiltshire model for satellite rendezvous proposed in~\cite{ClohWilt60} and elaborated on in~\cite{Klue99} for instance.
Let~$r(t)$ denote the altitude of the chasing spacecraft at time~$t$.
Then, for~$x(\cdot) = r(\cdot) - r_s(\cdot)$ the component of relative distance along the radial direction,
$y(\cdot)$ the downtrack component along satellite's circular orbit,
and~$z(\cdot)$ the distance component along the satellite's angular momentum,
and setting~$ \zeta(\cdot) := (\zeta_1(\cdot),\zeta_2(\cdot),\zeta_3(\cdot))^\top := (x(\cdot),y(\cdot),z(\cdot))^\top$ we obtain the Clohessy-Wiltshire equations
\begin{equation*}
\begin{aligned}
\ddot \zeta_1(t) &= 3 \omega^2 \zeta_1(t) + 2\omega \dot \zeta_2(t) + u_x(t), \\
\ddot \zeta_2(t) &= -2\omega \dot \zeta_1(t) + u_y(t), \\
\ddot \zeta_3(t) &= -\omega^2 \zeta_3(t) + u_z(t).
\end{aligned}
\end{equation*}
Defining the control input $u(\cdot) := (u_1(\cdot),u_2(\cdot),u_3(\cdot))^\top = (u_x(\cdot), u_y(\cdot), u_z(\cdot))^\top$, and
$f: \mathbb{R}^3 \times \mathbb{R}^3 \to \mathbb{R}^3$, $(\xi, \nu) \mapsto (3 \omega^2 \xi_1 + 2\omega \nu_2, -2 \omega \nu_1, -\omega^2 \xi_3)^\top $
with~$B = I_3 \in \mathbb{R}^{3 \times 3}$ the equations of motion above with output~$ \zeta(\cdot)$ can compactly be written as
$\ddot \zeta(t) = f(\zeta(t), \dot \zeta(t)) + Bu(t)$,
which is a system of relative degree two and belongs to the system class~$\mathcal{N}^{3,2}$.
For simulation purposes we choose~$r_s =415000 \, \rm m $ ($\approx$ altitude ISS) which yields~$\omega \approx 0.00113 \, \rm{s^{-1}}$ corresponding to an orbital period of approximately~$93$~minutes.
Since we aim to simulate a docking manoeuvre we choose the reference trajectory~$\zeta_{\rm ref}(\cdot)$ such that~$\zeta(T) = (0,0,0)^\top$ in Hill's coordinate frame.
To this end, we choose
\begin{equation*}
\zeta_{\rm ref}(t) = \zeta(0) \Big( 1- \sin \Big(\frac{\pi}{2} \frac{t}{T} \Big) \Big),
\end{equation*}
guiding the spacecraft smoothly to the satellite, with~$\zeta_{\rm ref}(T) = \dot \zeta_{\rm ref}(T) = (0,0,0)^\top$.
We take the initial conditions as in~\cite{Klue99} $x(0)=-y(0) = 1000\, \rm m$, and additionally~$z(0)=250 \, \rm m$;
and~$\dot x(0) = - 0.1 \, \rm{ms^{-1}}$, $\dot y(0)= 1.69 \, \rm{ms^{-1}}$ and~$\dot z(0)=-0.05 \, \rm{ms^{-1}}$.
As docking time we choose~$T = 1800 \, \rm s$, this is docking within thirty minutes.
For the funnel parameters we choose~$N : s \mapsto -s \cos(10^{-2} \, s)$ and $\alpha: s \mapsto (r+1)c/(1-s)$, where with~$c = 1$ the initial conditions~\eqref{eq:initial} are satisfied.
Since~$\lim_{t \to T} \varphi(t) = \infty$ simulation is possible only for~$[0,t_{\rm max}]$ with~$t_{\rm max} = T - \rm eps$ for a predefined~$\rm eps > 0$.
Now, since~$\varphi(t) \|e(t)\| < 1$ for all~$t \in [0,T)$ we choose~$\rm eps$ such that a certain upper bound of the spatial error at final time~$t_{\rm max}$ is guaranteed, this is
\begin{align*}
\| e(t_{\rm max}) \| < \frac{1}{\varphi(t_{\rm max})} &= c(T-t_{\rm max}) \le \rm{eps} \, \\
&\Rightarrow \, t_{\rm max} \ge T - \rm{eps}/c.
\end{align*}
For the simulation we choose~$\rm eps = 10^{-10} \, \rm m$ which means a spatial accuracy of \AA ngstr\"om (range of size of atoms).
This seems to be a unnecessary high accuracy since in real applications the required rendezvous distance is about centimetres, then magnetic docking structures become active; however, if these fail unexpectedly, the feedback control still is capable to perform a docking manoeuvre.
The simulation has been performed in \textsc{Matlab} (solver: \textsf{ode23tb}).
\begin{figure}[ht]
\centering
\includegraphics[width=0.75\linewidth]{Relative_motion_3D_earth-eps-converted-to.pdf}
\caption{Motion of the chasing spacecraft and the satellite with indicated earth (not true to scale) as center.}
\label{Fig:Relative_motion_3D}
\end{figure}
Figure~\ref{Fig:Relative_motion_3D} shows the docking manoeuvre with indicated earth (not true to scale) as center.
Figure~\ref{Fig:Errors_and_funnel} shows the error between the reference and the respective coordinates.
It can be seen that the docking manoeuvre is successful within the predefined finite time~$T$ and the errors evolve within the prescribed funnel boundary.
\begin{figure}[h]
\centering
\includegraphics[width=0.78\linewidth]{Error_and_funnel-eps-converted-to.pdf}
\caption{Errors and funnel.}
\label{Fig:Errors_and_funnel}
\end{figure}
In Figure~\ref{Fig:Errors_velocities} the errors of the respective velocities are depicted.
As expectable from Theorem~\ref{Thm:Exact-tracking-in-finte-time}, the errors of the velocities tend to zero for~$t \to T$.
\begin{figure}[h]
\centering
\includegraphics[width=0.78\linewidth]{Error_Velocities-eps-converted-to.pdf}
\caption{Error of the respective velocities.}
\label{Fig:Errors_velocities}
\end{figure}
The control input is depicted in Figure~\ref{Fig:Control}; note that only little control effort is necessary to reach the prescribed docking position.
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=0.78\linewidth]{Control_detail-eps-converted-to.pdf}
\caption{Control input.}
\label{Fig:Control_detail}
\end{subfigure}
\hfill
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=0.78\linewidth]{Control_detail_end-eps-converted-to.pdf}
\caption{Control input in the very last moments.}
\label{Fig:Control_detail_end}
\end{subfigure}
\caption{Control input.}
\label{Fig:Control}
\end{figure}
Figure~\ref{Fig:Control_detail_end} shows the control input in the very last moments before docking; where the largest input signals are generated, which may not be needed if the magnetic docking tools are activated and work as intended.
\section{Conclusion and outlook}
We introduced a funnel control scheme which achieves exact tracking in predefined finite time while the error between the reference and the output evolve within the boundaries of a prescribed funnel.
We proved that the application of the proposed control scheme yields a closed loop system which has a maximal solution, all signals are bounded, and the error as well as all its relevant derivatives are zero at the predefined final time.
\begin{ack}
I am deeply indebted to Thomas Berger (University of Paderborn) for many fruitful discussions and helpful advices.
\end{ack}
\section*{Appendix: Proof of Theorem~\ref{Thm:Exact-tracking-in-finte-time}}
\begin{proof}
The proof is subdivided in several steps. \\
\textit{Step one.}
We show existence of a solution of~\eqref{eq:system},\eqref{def:u}.
To this end, we aim to reformulate~\eqref{eq:system},\eqref{def:u} as an initial value problem of the form
\begin{equation} \label{eq:IVP}
\begin{aligned}
\dot x(t) &= F\left(t,x(t), \textbf{T}(x)(t) \right), \\
x (0) &= \left(y^0(0), \dot y^0(0),\ldots, (\tfrac{\text{\normalfont d}}{\text{\normalfont d}t})^{r-1}y^0(0) \right),
\end{aligned}
\end{equation}
where we set~$n = rm $, and
\begin{equation*}
x(t) = \left( y(t), \dot y(t), \ldots, y^{(r-1)}(t) \right).
\end{equation*}
\begin{comment}
\\
We define the set
\begin{equation*}
\mathcal{D}_{0} := \setdef{ \zeta \in \mathbb{R}^m }{ \| \zeta \| < 1 } ,
\end{equation*}
and for~$\alpha(\cdot)$ from~\eqref{def:alp} we define the function~$\Gamma_{0}: \mathcal{D}_{0} \to \mathbb{R}^m$
\begin{equation*}
\Gamma_{0}(\zeta) := \alpha(\|\zeta\|^2) \, \zeta.
\end{equation*}
For~$k=1,\ldots,r-1$ we recursively define the sets~$\mathcal{D}_{k}$ and the functions~$\Gamma_{k} : [0,T) \times \mathcal{D}_{k} \to \mathbb{R}^m$
\begin{equation} \label{def:Gam_k}
\begin{aligned}
\mathcal{D}_k &:= \underbrace{\mathcal{D}_0 \times \cdots \times \mathcal{D}_0}_{k-\rm times} \times \mathbb{R}^m, \\
\Gamma_{k}(t,\zeta_1,\ldots,\zeta_{k+1}) &:= \frac{\partial \Gamma_{k-1}(t,\zeta_1,\ldots,\zeta_{k})}{\partial t} \\
& \ + \sum_{i=1}^{k} \frac{\partial \Gamma_{k-1}(t,\zeta_1,\ldots,\zeta_{k})}{\partial \zeta_i} \left( \varphi(t) \left( c \zeta_i - \Gamma_{0}(t,\zeta_i) \right) + \zeta_{i+1} \right).
\end{aligned}
\end{equation}
We choose some interval~$I \subseteq [0,T)$ with~$0 \in I$ such that $(e_1,\ldots,e_r) : I \to \mathbb{R}^{rm}$ satisfy the relations in~\eqref{def:control_variables} and be such that for all~$t \in I$ and all~$i=0,\ldots,r-1$ we have~$e_i(t),\ldots,e_i(t) \in \mathcal{D}_0$.
Then, with $\gamma_k(\cdot) := \alpha(\|e_k(\cdot)\|^2) e_k(\cdot)$ from~\eqref{def:gam_k} we obtain for~$t \in I$
\begin{equation*}
\gamma_k^{(j)}(t) = \Gamma_j \big( t,e_k(t),\ldots,e_{k+j}(t) \big), \ 0 \le j \le r-k,
\end{equation*}
which can be seen via a brief induction.
\end{comment}
\begin{comment}
We define the set
\begin{subequations}
\begin{equation}
\mathcal{D}_{k,0} := \setdef{ (t,e_k) \in \mathbb{R}_{\geq 0} \times \mathbb{R}^m }{ (t,e_k) \in \Psi }
\end{equation}
and the function~$\Gamma_{k,0} : \mathcal{D}_{k,0} \to \mathbb{R}^m$
\begin{equation}
\Gamma_{k,0}(t,e_k) := \alpha(\|e_k\|^2) e_k.
\end{equation}
\end{subequations}
For~$k=1,\ldots,r-1$ and~$1 \le j \le r-k$ we recursively define the sets~$\mathcal{D}_{k,j}$ and the functions~$\Gamma_{k,j} : \mathcal{D}_{k,j} \to \mathbb{R}^m$
\begin{equation}
\begin{aligned}
\mathcal{D}_{k,j} & := \setdef{ (t,e_k,\ldots,e_{k+j}) \in \mathbb{R}_{\geq 0} \times \mathbb{R}^m \times \cdots \times \mathbb{R}^m }{ (t,e_i) \in \Psi, \ i=k,\ldots,k-1+j }, \\
\Gamma_{k,j}(t,e_k,\ldots,e_{k+j}) &:= \frac{\partial \Gamma_{k,j-1}(t,e_k,\ldots,e_{k-1+j})}{\partial t} \\
& \ + \sum_{i=k}^{k-1+j} \frac{\partial \Gamma_{k,j-1}(t,e_k,\ldots,e_{k-1+j})}{\partial e_i} \big( \varphi(t)(e_i - \Gamma_{i,0}(t,e_i)) + e_{i+1} \big).
\end{aligned}
\end{equation}
We choose some interval~$I \subseteq \mathbb{R}_{\geq 0}$ with~$0 \in I$.
Let $(e_1,\ldots,e_r) : I \to \mathbb{R}^{rm}$ be such that for all~$t \in I$ and all~$k=1,\ldots,r-1$ and~$1\le j \le r-k$ we have $(t,e_k,\ldots,e_{k+j}) \in \mathcal{D}_{k,j}$,
and $(e_1,\ldots,e_{r})$ satisfies the relations in~\eqref{def:control_variables}.
Then, for~$\gamma_k(\cdot) := \alpha(\|e_k(\cdot)\|^2) e_k(\cdot)$ from~\eqref{def:gam_k} we obtain
\begin{equation}
\gamma_k^{(j)}(t) = \Gamma_k,j \big( t,e_k(t),\ldots,e_{k+j}(t) \big), \quad 1\le j \le r-k,
\end{equation}
\red{{>>> which can be seen via a brief induction ??? <<<TODO!!!}}
\end{comment}
Setting~$\mathcal{D}_0 := \setdef{ v \in \mathbb{R}^m}{\|v\| < 1}$ we choose some interval~$I \subseteq [0,T)$ with~$0 \in I$ such that $(e_1,\ldots,e_r) : I \to \mathbb{R}^{n}$ satisfy the relations in~\eqref{def:control_variables} and be such that for all~$t \in I$ we have $e_1(t),\ldots,e_{r-1}(t) \in \mathcal{D}_0$, which is possible via the initial condition~\eqref{eq:initial}.
Then, with the aid of~\eqref{def:Gam_k} we have for all $k=1,\ldots,r-1$
\begin{equation*}
\gamma_k^{(j)}(t) = \Gamma_j \big( t,e_k(t),\ldots,e_{k+j}(t) \big), \ 0 \le j \le r-k, \ t \in I.
\end{equation*}
Next, we define the function
\begin{equation*}
\begin{aligned}
\tilde e_1 : [0,T) \times \mathbb{R}^m &\to \mathbb{R}^m, \\
(t,\xi_0) & \mapsto \varphi(t) \left( \xi_0 - y_{\rm ref}(t) \right), \\
\end{aligned}
\end{equation*}
and the set
\begin{equation*}
\tilde \mathcal{D}_1 := \setdef{ (t,\xi_0) \in [0,T) \times \mathbb{R}^m }{ \tilde e_1(t,\xi_0) \in \mathcal{D}_0 } .
\end{equation*}
With this, we recursively define for~$k=2,\ldots,r$ the functions
\begin{equation*}
\begin{aligned}
\tilde e_k \, : \tilde \mathcal{D}_{k-1} \times \mathbb{R}^m &\to \mathbb{R}^m, \\
(t,\xi_0,\ldots,\xi_{k-1}) & \mapsto \varphi(t) \left( \xi_{k-1} - y_{\rm ref}^{(k-1)}(t) \right) \\
& \quad + \varphi(t) \sum_{i=1}^{k-1} \Gamma_{k-1-i} \big( t,\tilde e_i,\ldots,\tilde e_{k-1} \big),
\end{aligned}
\end{equation*}
where for the sake of better legibility we omit the arguments of~$\tilde e_j$, $j=1,\ldots,k-1$; and further, we define the sets
\begin{equation*}
\begin{aligned}
\tilde \mathcal{D}_{k} := \big\{ &(t,\xi_0,\ldots,\xi_{k-1}) \in \tilde \mathcal{D}_{k-1} \times \mathbb{R}^m \ \vline \\
& \quad \tilde e_{i}(t,\xi_0,\ldots,\xi_{k-1}) \in \mathcal{D}_0, i=1,\ldots,k \big\} .
\end{aligned}
\end{equation*}
Together with~$\alpha(\cdot), N(\cdot)$ as in~\eqref{def:control_functions}
we define for~$t \in I$
\begin{equation*}
N_r(t) := \big( N \circ \alpha \big) \left( \| \tilde e_r \big( t,y(t),\dot y(t),\ldots,y^{(r-1)}(t) \big) \|^2 \right).
\end{equation*}
Then, the control~$u(\cdot)$ defined in~\eqref{def:u} reads
\begin{equation*}
\begin{aligned}
u(t) = N_r(t) \cdot \tilde e_r \big( t,y(t),\ldots,y^{(r-1)}(t) \big), \ t \in I.
\end{aligned}
\end{equation*}
Lastly, we define the function~$F : \tilde \mathcal{D}_{r-1} \times \mathbb{R}^q \to \mathbb{R}^{n}$ by
\begin{equation*}
\begin{aligned}
(t,\xi_0,\ldots, \xi_{r-1}, \eta) \mapsto \left( \xi_1,\ldots,\xi_{r-1}, f \left( d(t), \eta, N_r(t) \cdot \tilde e_r \right) \right),
\end{aligned}
\end{equation*}
where in $\tilde e_r = \tilde e_r(t,\xi_0,\ldots,\xi_{r-1})$ we omit the arguments for the sake of better legibility.
Together, the initial value problem~\eqref{eq:system},\eqref{def:u} is equivalent to~\eqref{eq:IVP}.
In particular, we have~$(0,x(0)) \in \tilde \mathcal{D}_{r-1}$, the function~$F$ is measurable in the variable~$t$, continuous in~$(\xi_0,\ldots,\xi_{r-1}, \eta)$ and locally essentially bounded.
Hence, invoking Remark~\ref{Rem:localisation}, adapted to the current context,~\cite[Thm.~B.1]{IlchRyan09} yields the existence of a maximal solution~$x: [-\sigma,\omega) \to \mathbb{R}^{n}$ of~\eqref{eq:IVP}, $0 < \omega \le T$.
In particular, the graph of the solution of~\eqref{eq:IVP} is not a compact subset of~$\tilde \mathcal{D}_{r-1} $.
\\
\textit{Step two.}
For the functions~$e_k(\cdot)$ introduced in~\eqref{def:e_k} we show
that for all~$k=1,\ldots,r-1$ there exists~$\varepsilon_k \in (0,1)$ such that~$\| e_k(t)\| \le \varepsilon_k$ for all~$t \in [0,\omega)$.
We observe that for~$t \in [0,\omega)$ and~$k = 1,\ldots,r$ we have
\begin{equation*}
e_k(t) - \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-1-i)}(t) = \varphi(t) e^{(k-1)}(t).
\end{equation*}
With this, and using~$ \dot \varphi(t) = c \varphi(t)^2$, we calculate for $k=1,\ldots,r-1$
\begin{equation} \label{eq:ddt_ek}
\begin{aligned}
\dot e_k(t) &= \dot \varphi(t) e^{(k-1)}(t) + \varphi(t) e^{(k)}(t) \\
& + \dot \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-1-i)}(t) + \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-i)}(t) \\
& = \frac{\dot \varphi(t)}{\varphi(t)} \Big( e_k(t) - \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-1-i)}(t) \Big) \\
&+ \Big( e_{k+1}(t) - \varphi(t) \sum_{i=1}^{k} \gamma_{i}^{(k-i)}(t) \Big) \\
& + \dot \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-1-i)}(t) + \varphi(t) \sum_{i=1}^{k-1} \gamma_{i}^{(k-i)}(t) \\
& = (c-\alpha_k(t) ) \varphi(t) e_k(t) + e_{k+1}(t),
\\
\dot e_r(t) &= c \varphi(t) e_r(t) + \varphi(t) e^{(r)}(t) + \varphi(t) \sum_{i=1}^{r-1} \gamma_{i}^{(r-i)}(t).
\end{aligned}
\end{equation}
Further, using the definitions of~$\alpha_k(\cdot)$ and~$\gamma_k(\cdot)$, we record for later use
\begin{align} \label{eq:ddt_alp}
\dot \gamma_k(t) &= \tfrac{\text{\normalfont d}}{\text{\normalfont d}t} ( \alpha_k(t) e_k(t) ) \\
& = 2 \alpha'(\|e_k(t)\|^2) \langle e_k(t), \dot e_k(t) \rangle e_k(t) + \alpha_k(t) \dot e_k(t) . \nonumber
\end{align}
We observe~$e_k(t) = \tilde e_k(y(t),\ldots,y^{(k-1)}(t))$ and hence, since~$\tilde e_k(\cdot) \in \mathcal{D}_0$ due to the initial conditions~\eqref{eq:initial}, we have
\begin{equation*}
\forall \, k=1,\ldots,r \ \forall \, t \in [0,\omega) \,: \ \|e_k(t)\| < 1.
\end{equation*}
We set $\hat \varepsilon_k := \| e_k(0)\|^2 < 1$ and $\lambda := \varphi(0) = \inf_{s \in [0,T)} \varphi(s)> 0$.
Let~$\varepsilon$ be the unique point in~$(0,1)$ such that $\alpha(\varepsilon) \varepsilon = (1 + c \lambda)/\lambda$ and define $\varepsilon_k := \max\{\varepsilon,\hat \varepsilon_k\} < 1$.
We show that
\begin{equation} \label{eq:ek_le_vek}
\forall\, k\in\{1,\ldots,r-1\} \ \forall \, t \in [0,\omega) \, : \ \|e_k(t)\|^2 \le \varepsilon_k.
\end{equation}
Seeking a contradiction we suppose this is false for at least one~$\ell \in \{1,\ldots,r-1\}$.
Then $\|e_\ell(t_1)\|^2 > \varepsilon_\ell$ for some~$t_1 \in (0,\omega)$ and
we define
\begin{equation*}
t_0 := \max \setdef{ t \in [0,t_1)}{ \|e_\ell(t)\|^2 = \varepsilon_\ell}.
\end{equation*}
With this we have
\begin{equation*}
\forall\, t \in [t_0,t_1]\,: \ \varepsilon \le \varepsilon_\ell \le \|e_\ell(t)\|^2,
\end{equation*}
which gives, invoking monotonicity of~$\alpha(\cdot)$, the following relation
\begin{equation*}
\forall\, t \in [t_0,t_1]\,: \ \alpha(\varepsilon) \le \alpha(\|e_\ell(t)\|^2) = \alpha_\ell(t).
\end{equation*}
Hence,
\begin{equation*}
\forall\, t \in [t_0,t_1]\,: \ \alpha_\ell(t) \|e_\ell(t)\|^2 \ge \alpha(\varepsilon) \varepsilon = \frac{1+ c \lambda}{\lambda}.
\end{equation*}
With this, using~$\alpha_\ell(\cdot) \ge c$ via~\eqref{def:alp}, and the relations in~\eqref{eq:ddt_ek}, we calculate for~$t \in [t_0,t_1]$
\begin{equation*}
\begin{aligned}
\tfrac{\text{\normalfont d}}{\text{\normalfont d}t} \tfrac{1}{2} \|e_\ell(t)\|^2 &= \langle e_\ell(t), (c-\alpha_\ell(t) ) \varphi(t) e_\ell(t) + e_{\ell+1}(t) \rangle \\
& = \! - \varphi(t) (\alpha_\ell(t) \!-\! c) \|e_\ell(t)\|^2 \! + \! \langle e_\ell(t), e_{\ell+1}(t) \rangle \\
& < -\varphi(t)(\alpha_\ell(t) - c) \|e_\ell(t)\|^2 + 1 \\
& \le - \lambda (\alpha_\ell(t) - c) \|e_\ell(t)\|^2 + 1 \\
& < -\lambda \alpha_\ell(t) \|e_{\ell}(t)\|^2 + c\lambda + 1 \le 0 ,
\end{aligned}
\end{equation*}
which implies the contradiction
\begin{equation*}
\varepsilon_\ell < \|e_\ell(t_1)\|^2 < \|e_\ell(t_0)\|^2 = \varepsilon_\ell.
\end{equation*}
Therefore~\eqref{eq:ek_le_vek} holds.
This implies boundedness of~$\alpha_k$ (bounded by $\alpha(\varepsilon_k)$) and boundedness of~$\gamma_k$ (bounded by $\alpha(\varepsilon_k) \sqrt{\varepsilon_k}$) for all $k=1,\ldots,r-1$.
\\
\textit{Step three.}
Since the functions~$e_k(\cdot)$ defined in~\eqref{def:e_k} involve higher derivatives of the functions~$\gamma_k(\cdot)$ we aim to show boundedness of the latter.
In order to do so, recalling the definition of~$\gamma_k(\cdot)$ we establish boundedness of higher derivatives of~$\alpha_k(\cdot)$ on~$[0,\omega)$, which in turn involve higher derivatives of~$e_k(\cdot)$.
Hence, we show boundedness of higher derivatives of~$e_k(\cdot)$ on~$[0,\omega)$; more precise, we show boundedness of~$e_k^{(r-k)}(\cdot)$ on~$[0,\omega)$ for $k=1,\ldots,r-1$.
Recalling the definition of~$\varphi(\cdot)$ in~\eqref{def:vp} we have $\varphi^{(j)}(t) = c^{j} j! \varphi(t)^{j+1}$ for~$j\in\mathbb{N}$.
Using the generalized Leibniz rule, we obtain via~\eqref{eq:ddt_ek} for~$k=1,\ldots,r-1$ and~$1 \le j \le r-k$ the recursion
\begin{scriptsize}
\begin{equation} \label{eq:j_th-derivative-ek}
\begin{aligned}
e_k^{(j)}(t) &= \big( (c-\alpha_k(t)) \varphi(t) e_k(t) \big)^{(j-1)} + e_{k+1}^{(j-1)}(t) \\
&= \sum_{j_1+j_2+j_3 = j-1} \frac{(j-1)!}{j_1! j_2! j_3!} (c-\alpha_k(t))^{(j_1)} \varphi^{(j_2)}(t) e_k^{(j_3)}(t) \\
& + e_{k+1}^{(j-1)}(t) \\
&= \sum_{j_1+j_2+j_3 = j-1} \frac{(j-1)!}{j_1! j_3!} (c-\alpha_k(t))^{(j_1)} c^{j_2} \varphi(t)^{j_2+1} e_k^{(j_3)}(t) \\
& + e_{k+1}^{(j-1)}(t) .
\end{aligned}
\end{equation}
\end{scriptsize}
We expatiate on the expression above for~$j=2$:
\begin{equation*}
\begin{aligned}
\ddot e_k(t)
& = -2\alpha'(\|e_k(t)\|^2) \langle e_k(t), \dot e_k(t) \rangle \varphi(t) e_k(t) \\
& + (c-\alpha_k(t))c \varphi^2 e_k(t) \\
& + (c\!-\! \alpha_k(t)) \left( (c \!- \! \alpha_k(t)) \varphi(t)^2 e_k(t) \!+\! \varphi(t) e_{k+1}(t) \right) \\
& + (c-\alpha_{k+1}(t)) \varphi(t) e_{k+1}(t) + e_{k+2}(t).
\end{aligned}
\end{equation*}
This recursion successively leads to the following observations. Since~$j_1+j_2+j_3 = j-1$
\begin{itemize}
\item for~$j_1=0$ the expression~$\varphi(\cdot)^{j_2+1} e_k^{(j_3)}(\cdot)$ involves at most the~$j-1^{\rm st}$ derivative of~$e_k(\cdot)$,
and at most the~$j^{\rm th}$ power of~$\varphi(\cdot)$;
the other terms involve (at most) derivatives and powers of the form~$\varphi(\cdot)^{j_2+1} e_k^{(j-1-j_2)}(\cdot)$ for~$j\le r-k$,
\item $e_k^{(j)}(\cdot)$ involves $e_{k+1}^{(j-1)}(\cdot)$ which itself involves~$e_{k+2}^{(j-2)}(\cdot)$ and so forth; therefore~$e_k^{(j)}(\cdot)$ involves~$e_{k+j}(\cdot)$,
\item the highest derivative of~$\alpha_k(\cdot)$ appearing in $e_k^{(j)}(\cdot)$ is~$\alpha_k^{(j-1)}(\cdot)$, which itself involves at most the~~$j-1^{\rm st}$ derivative of~$e_k(\cdot)$.
\end{itemize}
These observations together with the fact that
\begin{equation} \label{eq:vp+j_ek-vp+j-1_ek}
\begin{aligned}
\forall\, j \in \mathbb{N}\,:\ & \varphi(\cdot)^{j} e_k(\cdot) \in \mathcal{L}^{\infty}([0,\omega) ; \mathbb{R}^m) \\
& \Rightarrow \varphi(\cdot)^{j-1} e_k(\cdot) \in \mathcal{L}^{\infty}([0,\omega) ; \mathbb{R}^m)
\end{aligned}
\end{equation}
yield that boundedness of~$e_k^{(j)}(\cdot)$ on~$[0,\omega)$ can be established by showing boundedness of~$\varphi(\cdot)^{r-k} e_k(\cdot)$ for all~$k=1,\ldots,r-1$.
In order to show this we initially establish the following:
for all~$k=2,\ldots,r-1$
\begin{equation} \label{eq:vp+r-k_e_k-bounded-vp+r-k+1_e_k-1-bounded}
\begin{aligned}
& \varphi(\cdot)^{r-k} e_k(\cdot) \in \mathcal{L}^{\infty}([0,\omega) ; \mathbb{R}^m) \\
& \Rightarrow \varphi(\cdot)^{r-k+1} e_{k-1}(\cdot) \in \mathcal{L}^{\infty}([0,\omega) ; \mathbb{R}^m).
\end{aligned}
\end{equation}
To see this, let $\varphi(\cdot)^{r-k} e_k(\cdot) \in \mathcal{L}^{\infty}([0,\omega) ; \mathbb{R}^m)$ and set $M_k := \sup_{s \in [0,\omega)} \| \varphi(s)^{r-k} e_k(s)\| < \infty$.
Then we consider for~$t \in [0,\omega)$
\begin{small}
\begin{equation*}
\begin{aligned}
& \tfrac{\text{\normalfont d}}{\text{\normalfont d}t} \tfrac{1}{2} \| \varphi(t)^{r-k+1} e_{k-1}(t)\|^2 \\
&= \langle \varphi(t)^{r-k+1} e_{k-1}(t), \varphi(t)^{r-k+1} \dot e_{k-1}(t) \rangle \\
& \qquad + \langle \varphi(t)^{r-k+1} e_{k-1}(t), c(r-k+1) \varphi(t)^{r-k+2} e_{k-1}(t) \rangle \\
& = \langle \varphi(t)^{r-k+1} e_{k-1}(t), \varphi(t)^{r-k+1} \big( (c-\alpha_{k-1}(t))\varphi(t)e_{k-1}(t) \rangle \\
& \qquad + \langle \varphi(t)^{r-k+1} e_{k-1}(t), c(r-k+1) \varphi(t)^{r-k+2} e_{k-1}(t)\rangle \\
& \qquad + \langle \varphi(t)^{r-k+1} e_{k-1}(t), \varphi(t)^{r-k+1} e_k(t) \rangle \\
& \le - \varphi(t) \big( c(r+1) - c(r-k+2) \big) \|\varphi(t)^{r-k+1} e_{k-1}(t)\|^2 \\
& \qquad + \varphi(t) M_{k} \|\varphi(t)^{r-k+1} e_{k-1}(t)\| \\
&\le -\varphi(t) \Big( c(k-1) \|\varphi(t)^{r-k+1} e_{k-1}(t)\| \\
& \qquad \qquad \qquad \qquad \qquad - M_{k} \Big) \cdot \|\varphi(t)^{r-k+1} e_{k-1}(t)\|
\end{aligned}
\end{equation*}
\end{small}
which is non-positive for~$\|\varphi(t)^{r-k+1} e_{k-1}(t)\| \ge \tfrac{M_k}{c(k-1)} \ge 0$ and hence Lemma~\ref{Lem:x_le_M} yields boundedness of~$\varphi(\cdot)^{r-k+1} e_{k-1}(\cdot)$ on~$[0,\omega)$.
A successive application of~\eqref{eq:vp+r-k_e_k-bounded-vp+r-k+1_e_k-1-bounded} yields
\begin{equation} \label{eq:vp+r-k-e_k-bounded}
\forall \, k=1,\ldots,r-1\, : \ \varphi(\cdot)^{r-k} e_k(\cdot) \in \mathcal{L}^\infty([0,\omega) ; \mathbb{R}^m).
\end{equation}
In particular, via~\eqref{eq:vp+j_ek-vp+j-1_ek} we have~$\varphi(\cdot) e_{k}(\cdot) \in \mathcal{L}^\infty([0,\omega) ; \mathbb{R}^m)$ for all~$k=1,\ldots,r-1$ from which boundedness of~$\dot e_k(\cdot)$ follows, which in turn implies boundedness of~$\dot \alpha_k(\cdot)$ and~$\dot \gamma_k(\cdot)$ on~$[0,\omega)$.
Now, via~\eqref{eq:j_th-derivative-ek} and~\eqref{eq:vp+j_ek-vp+j-1_ek} boundedness of~$e^{(j)}_k(\cdot)$ successively follows for all~$j \le r-k-1$, from which we may deduce boundedness of~$ \alpha_k^{(j)}(\cdot)$ and~$ \gamma_k^{(j)}(\cdot)$ for~$j \le r-k-1$,
this is, for all~$k=1\ldots,r-1$ and~$j\le r-k-1$ there exists $ \bar \gamma_k^{j} := \sup_{s \in [\tau,\omega)} \gamma_k^{(j)}(s) < \infty $.
\\
\textit{Step four.} We show boundedness of~$x(\cdot)$ on~$[0,\omega)$.
Recalling the definition of~$e_k(\cdot)$ we see that for all $k=1,\ldots,r$ we have via the previous steps
\begin{equation} \label{eq:e+k-1-bounded}
\begin{aligned}
\forall\, t \in [\tau,\omega) : \, \|e^{(k-1)}(t)\| & \le \left\|\frac{e_{k}(t)}{\varphi(t)} \right\| + \left\|\sum_{i=1}^{k-1} \gamma_i^{(k-1-i)}(t) \right\| \\
& \le \frac{1}{\lambda} + \sum_{i=1}^{k-1} \bar \gamma_i^{k-1-i} < \infty.
\end{aligned}
\end{equation}
Therefore, since $x(\cdot) = (y(\cdot),\ldots,y^{(r-1)}(\cdot)) = (e(\cdot) + y_{\rm ref}(\cdot), \ldots, e^{(r-1)}(\cdot) + y_{\rm ref}^{(r-1)}(\cdot)) $
and the reference $y_{\rm ref} \in \mathcal{W}^{r,\infty}([0,T) ; \mathbb{R}^m)$, we have $x \in \mathcal{L}^\infty([0,\omega) ; \mathbb{R}^{rm})$.
\\
\textit{Step five.}
We show boundedness of~$\alpha_r(\cdot)$ on~$[0,\omega)$.
Invoking the previous steps, in particular boundedness of~$x(\cdot)$ on~$[0,\omega)$, and the properties of the operator class $\mathcal{T}^{n,q}_\sigma$ we deduce the existence of a compact $K_q \subset \mathbb{R}^q$ such that $T(x)(t) \in K_q$ for~$t \in [0,\omega)$; furthermore, since $d \in L^\infty(\mathbb{R}_{\geq 0} ; \mathbb{R}^p)$ there exists a compact $K_p \subset \mathbb{R}^p$ such that $d(t) \in K_p$ for~$t \in [0,\omega)$.
Via the high gain property there exists~$\rho \in (0,1)$ such that the continuous function
\begin{equation*}
\begin{aligned}
\chi: \mathbb{R} \to \mathbb{R}, \quad s \mapsto \min \big\{ &\langle v,f(\delta,z,-sv) \rangle \ \vline \\
&(\delta,z,v) \in K_q \times K_p \times V \big\}
\end{aligned}
\end{equation*}
is unbounded from above;
where we defined the compact set
$
V := \setdef{ v \in \mathbb{R}^m}{ \rho \le \|v\| \le 1}.
$
We show boundedness of~$\alpha_r(\cdot)$ by contradiction.
Since
$N : \mathbb{R}_{\geq 0} \to \mathbb{R}$ is surjective, the set~$\setdef{ \kappa > \rho_0 }{ N(\kappa) = \rho_1}$ is non-empty for every~$\rho_0 \in \mathbb{R}_{\geq 0}$ and every~$\rho_1 \in \mathbb{R}$.
Following the proof in~\cite[pp.~188-190]{BergIlch21},
we choose a real sequence~$(s_j)$ such that the corresponding sequence~$\left(\chi(s_j)\right)$ is positive, strictly increasing and in particular unbounded.
We initialize a sequence~$(\kappa_j)$ by choosing $\kappa_1 > \alpha(\rho^2) + \alpha_r(0)$ such that~$N(\kappa_1) = s_1$, and hereinafter define the strictly increasing sequence~$(\kappa_j)$ via
\begin{equation*}
\kappa_{j+1} := \inf \setdef{ \kappa > \kappa_j }{ N(\kappa) = s_{j+1} },
\end{equation*}
which obviously yields that
$
\lim_{j \to \infty} \chi(N(\kappa_j)) = \lim_{j \to \infty} \chi(s_j) = \infty.
$
Now, since we assumed~$\alpha_r(\cdot)$ to be unbounded and we have~$\kappa_{j+1} > \kappa_1 > \alpha_r(0)$ for all~$j \in \mathbb{N}$, we may define the sequence
\begin{equation*}
\tau_j := \inf \setdef{ t \in [0,\omega) }{ \alpha_r(t) = \kappa_{j+1} }, \ j \in \mathbb{N}_0,
\end{equation*}
which lies within~$(0,\omega)$.
Note that~$(\tau_j)$ is strictly increasing and we have $N(\alpha_r(\tau_j)) = N(\kappa_{j+1}) = s_{j+1}$ for each~$j \in \mathbb{N}_0$.
Next, we define a second sequence in~$(0,\omega)$
\begin{equation*}
\sigma_j = \sup \setdef{ t \in [\tau_{j-1} , \tau_j] }{ \chi(N(\alpha_r(t))) = \chi(s_j) }, \ j \in \mathbb{N}.
\end{equation*}
With this, since the sequence~$(\chi(s_j))$ is strictly increasing, we obtain for all~$j \in \mathbb{N}$
\begin{equation*}
\begin{aligned}
& \chi(N(\alpha_r(\sigma_j))) = \chi(s_j) < \chi(s_{j+1}) = \chi(N(\alpha_r(\tau_j))),
\end{aligned}
\end{equation*}
and therefore,
\begin{subequations} \label{eq:sigma_le_tau+chi_sj_le_chi_sj1}
\begin{equation}
\forall\, j \in \mathbb{N} \,: \ \sigma_j < \tau_j,
\end{equation}
as well as for all~$j \in \mathbb{N}$
\begin{equation}
\forall\, t \in (\sigma_j , \tau_j ] : \chi(N(\alpha_r(\sigma_j))) = \chi(s_j) < \chi(N(\alpha_r(t))).
\end{equation}
\end{subequations}
Next, we show by contradiction that for all~$j \in \mathbb{N}$ and for all~$t \in [\sigma_j,\tau_j]$ we have~$e_r(t) \in V$.
To this end, we first show - by contradiction - that for all~$j \in \mathbb{N}$ we have~$\alpha_r(t) \ge \kappa_j$ for~$t \in [\sigma_j, \tau_j]$.
Suppose that~$\alpha_r(t) < \kappa_j$ for some~$t \in [\sigma_j,\tau_j]$.
Then by~$\alpha_r(\tau_j) = \kappa_{j+1} > \kappa_j$ and by continuity of~$\alpha_r$ there exists~$\tilde t \in (\sigma_j, \tau_j)$ such that~$\alpha_r(\tilde t) = \kappa_j$.
Hence, we find
$
\chi(N(\alpha_r(\tilde t))) = \chi(N(\kappa_j)) = \chi(s_j),
$
which contradicts the definition of~$\sigma_j$.
Therefore, $\alpha_r(t) \ge \kappa_j$ for all~$t \in [\sigma_j,\tau_j]$.
Now, suppose~$e_r(t) \notin V$, this is, since for all~$t \in [0,\omega)$ we have~$\|e_r(t)\| < 1$, we suppose~$\|e_r(t)\| < \rho$ for some~$[\sigma_j,\tau_j]$.
This, together with~$\alpha_r(t) \ge \kappa_j$, leads to the contradiction
$
\alpha(\rho^2) < \kappa_1 \le \kappa_j \le \alpha_r(t) = \alpha(\|e_r(t)\|^2) < \alpha(\rho^2).
$
Hence, we deduce
\begin{equation*}
\forall \, j \in \mathbb{N} \ \forall \, t \in [\sigma_j, \tau_j] \, : \ e_r(t) \in V.
\end{equation*}
Now, since~$d(t) \in K_p$ and~$T(x)(t) \in K_q$ for~$t \in [0,\omega)$ we obtain, using~\eqref{eq:sigma_le_tau+chi_sj_le_chi_sj1}, for all~$j \in \mathbb{N}$ and~$t \in [\sigma_j, \tau_j]$
\begin{equation} \label{eq:er_pari_f}
\begin{aligned}
&\langle e_r(t), f\big( d(t), (Tx)(t), u(t) \big) \rangle \\
& = - \langle -e_r(t) , f\big( d(t), \textbf{T}(x)(t), -N(\alpha_r(t)) (-e_r(t)) \rangle \\
& \le - \min \setdef{ \langle v, f(\delta, z, -N(\alpha_r(t)) v }{ \begin{array}{l}
(\delta , z, v) \\ \in K_p \times K_q \times V
\end{array} } \\
& = - \chi(N(\alpha_r(t))) \\
& \le - \chi(s_j) .
\end{aligned}
\end{equation}
Since~$y_{\rm ref} \in \mathcal{W}^{r,\infty}([0,T) ; \mathbb{R}^m)$ we may set $c_{\rm ref} := \sup_{s \ge 0}\|y_{\rm ref}^{(r)}(s)\| < \infty$, and we recall $\sum_{i=1}^{r-1} \bar \gamma_i^{r-i} < \infty$ from the previous steps.
Furthermore, we observe~$\sigma_1 > 0$ and therefore, by properties of~$\varphi(\cdot)$ we may define $0 < \inf_{s \in [\sigma_1, T)} \varphi(s) =: c_{\varphi}$.
Then, with the aid of~\eqref{eq:ddt_ek} and~\eqref{eq:er_pari_f} for all~$j \in \mathbb{N}$ and~$t \in [\sigma_j, \tau_j]$ we obtain
\begin{equation*}
\begin{aligned}
&\tfrac{\text{\normalfont d}}{\text{\normalfont d}t} \tfrac{1}{2} \|e_r(t)\|^2 \\
&= \langle e_r(t), c \varphi(t) e_r(t) \rangle \\
& \quad + \langle e_r(t) \varphi(t) \big( f(d(t), (Tx(t), u(t)) - y_{\rm ref}^{(r)}(t) \big) \rangle \\
& \qquad + \langle e_r(t), \varphi(t) \sum_{i=1}^{r-1} \gamma_i^{(r-i)}(t) \rangle \\
& < \varphi(t) \Big( c + c_{\rm ref} + \sum_{i=1}^{r-1} \bar \gamma_i^{r-i} - \chi(s_j) \Big).
\end{aligned}
\end{equation*}
Thus, still seeking a contradiction, we may choose~$J \in \mathbb{N}$ large enough such that for~$t \in [\sigma_J, \tau_J]$ we have
\begin{equation*}
\begin{aligned}
& \varphi(t) \Big( c + c_{\rm ref} + \sum_{i=1}^{r-1} \bar \gamma_i^{r-i} - \chi(s_J) \Big) \\
& \le - c_{\varphi} \Big( \chi(s_J) - \Big(c + c_{\rm ref} + \sum_{i=1}^{r-1} \bar \gamma_i^{r-i} \Big)\Big) < 0,
\end{aligned}
\end{equation*}
which yields
$
\|e_r(\tau_J)\|^2 < \| e_r(\sigma_J)\|^2,
$
which in turn gives for~$t \in [\sigma_J, \tau_J]$
\begin{equation*}
\alpha_r(\tau_J) = \alpha(\|e_r(\tau_J)\|^2) < \alpha(\|e_r(\sigma_J)\|^2) = \alpha_r(\sigma_J).
\end{equation*}
This, however, contradicts the definition of~$\tau_J$, by which we have~$\alpha_r(t) < \alpha_r(\tau_J)$ for all~$t \in [0,\tau_J)$.
Therefore, the assumption of an unbounded~$\alpha_r(\cdot)$ cannot be true.
As a direct consequence thereof, we may infer the existence of~$\varepsilon_r \in (0,1)$ such that
\begin{equation*}
\forall \, t \in[0,\omega)\,: \ \|e_r(t)\|^2 \le \varepsilon_r.
\end{equation*}
\textit{Step six.} We show~$\omega = T$.
Via the previous steps we have for all~$k=1,\ldots,r$
\begin{equation*}
\forall \, t \in [0,\omega) \, : \ \|e_k(t)\| \le \varepsilon^* := \sqrt{\max\{\varepsilon_1,\ldots,\varepsilon_r\}} < 1,
\end{equation*}
by which the set
\begin{equation*}
\begin{aligned}
\hat \mathcal{D} := & \setdef{ (\zeta_1,\ldots,\zeta_r) \in \mathbb{R}^{rm} }{ \|\zeta_i\| \le \varepsilon^*, i=1,\ldots,r}
\end{aligned}
\end{equation*}
is a compact subset of~$\tilde \mathcal{D}_{r-1} $.
Assume~$\omega < T$. Then,
\begin{equation*}
\forall \, t \in [0,\omega) \, : \ x(t) \in \hat \mathcal{D} \subset \tilde \mathcal{D}_{r-1} .
\end{equation*}
So, by compactness of~$\hat \mathcal{D}$ the closure of the graph of the solution~$x(\cdot)$ of~\eqref{eq:IVP} on~$[0,\omega)$ is a compact subset of~$\tilde \mathcal{D}_{r-1} $ which contradicts the findings of \textit{Step one}. Thus, $\omega = T$.
\\
\textit{Step seven.} Assertion~$ii)$ is a direct consequence of \textit{Step four} and \textit{Step six}; and assertion~$iii)$ follows from \textit{Step two} and \textit{Step six}.
\\
\textit{Step eight.}
We show that the tracking error~$e(\cdot)$ and its derivatives tend to zero as~$t \to T$, this is, we show
\begin{equation} \label{eq:e-exact}
\forall\, k=1,\ldots,r \, : \ \lim_{t\to T} \|e^{(k-1)}(t)\| = 0.
\end{equation}
Note that the estimation in~\eqref{eq:e+k-1-bounded} is too rough to show~\eqref{eq:e-exact}.
Recalling the definition~\eqref{def:gam_k} of~$\gamma_k(\cdot) = \alpha_k(\cdot) e_k(\cdot)$ and exemplary its derivative~\eqref{eq:ddt_alp} we see that by \textit{Step three} not only~$\gamma_k^{(j)}(\cdot)$ is bounded on~$[0,\omega)$ for $j\le r-k-1$ but with the aid of~\eqref{eq:vp+r-k-e_k-bounded} even the product~$\varphi(\cdot) \, \gamma_k^{(j)}(\cdot)$ is bounded on~$[0,\omega)$,
i.e., for all $\ell=1\ldots,r-1$ there exists $ \hat \gamma_\ell^{j} := \sup_{s \in [\tau,\omega)} \varphi(s) \gamma_\ell^{(j)}(s) < \infty $ for~$0 \le j \le r-\ell - 1$.
Invoking \textit{Step two} and \textit{Step five} we may improve estimation~\eqref{eq:e+k-1-bounded} for all~$k=1,\ldots,r$ and all~$t \in [\tau, \omega) $ as follows
\begin{equation*}
\begin{aligned}
\|e^{(k-1)}(t)\| &\le \frac{\|e_{k}(t)\|}{\varphi(t)} + \frac{1}{\varphi(t)} \left\|\sum_{i=1}^{k-1} \varphi(t)\gamma_i^{(k-1-i)}(t) \right\| \\
& \le \frac{\sqrt{\varepsilon_k} + \sum_{i=1}^{k-1} \hat \gamma_i^{k-1-i} }{\varphi(t)}. \\
\end{aligned}
\end{equation*}
From this, since~$\omega=T$ by \textit{Step six}, and~$\lim_{t\to T} \varphi(t) = \infty$ we obtain~\eqref{eq:e-exact} for all~$k=1,\ldots,r$, which shows assertion~$iv)$ of the theorem
and completes the proof.
\end{proof}
|
\section{Introduction}
We consider the Navier--Stokes--Fourier system in the time-space cylinder $ Q=(0,T) \times \Om $, where $ T>0 $ and $ \Om $ is a bounded domain of $ \R^d $ with $ d=2 $ or $ 3 $. The time evolution of the \emph{density} $ \varrho =\varrho(t,x) $, \emph{velocity} $ \vc{u}=\vc{u}(t,x) $ and the \emph{absolute temperature} $ \vartheta=\vartheta(t,x) $ of a compressible, viscous and heat conducting fluid is given by the following system of field equations which describes conservation of mass, momentum and internal energy:
\begin{align}
&\partial_t \varrho + \text{div}_x (\varrho \vc{u}) =0,\label{NSF-c}\\
&\partial_t (\varrho \vc{u}) + \text{div}_x (\varrho \vc{u} \otimes \vc{u}) + \nabla_x p(\varrho,\vartheta) = \text{div}_x \bb{S} + \varrho \vc{g},\label{NSF-m}\\
&\partial_t (\varrho e(\varrho,\vartheta)) + \text{div}_x (\varrho e(\varrho,\vartheta)\vc{u} ) + \text{div}_x \vc{q} = \bb{S}\colon \bb{D}_x \vc{u} - p(\varrho , \vartheta) \text{div}_x \vc{u} ,\label{NSF-e}
\end{align}
where $\bb{D}_x \vc{u}=\frac{1}{2}( \nabla_x \vc{u} + \nabla_x^t \vc{u}) $, $ p= p(\varrho,\vartheta) $ is the pressure and $ e=e(\varrho, \vartheta) $ is the internal energy and $ \vc{g} $ is an external force.
The pressure and the internal energy is interrelated by means of \emph{Gibbs' equation}
\begin{align}\label{GE}
\vartheta Ds= De+pD\left(\frac{1}{\varrho}\right),
\end{align}
where $ s=s(\varrho,\vartheta) $ is the \emph{entropy} and $D= \begin{pmatrix}
\frac{\partial}{\partial \varrho} \\
\frac{\partial}{\partial \vartheta}\\
\end{pmatrix}.$
We assume the fluid is Newtonian, hence the \emph{viscous stress tensor}$ (\bb{S}) $ is given by
\begin{align}\label{visc_str}
\mathbb{S}(\nabla_x \mathbf{u})=\mu(\varrho , \vartheta) \bigg(\frac{\nabla_x \mathbf{u} + \nabla_x^{T} \mathbf{u}}{2}-\frac{1}{d} (\text{div}_x\mathbf{u})\mathbb{I} \bigg) + \lambda(\varrho , \vartheta) (\text{div}_x \mathbf{u}) \mathbb{I},
\end{align}
where $ \mu $ is the \emph{shear viscosity} coefficient and $ \lambda $ is the \emph{bulk viscosity} coefficient with $ \mu>0 $ and $ \lambda\geq 0 $. The \emph{heat flux} $ \vc{q} $ is given by the \emph{Fourier law},
\begin{align}\label{Fou_law}
\vc{q}(\varrho,\vartheta, \nabla_x \vartheta) = - \kappa(\varrho,\vartheta) \nabla_x \vartheta,
\end{align}
where $ \kappa $ is the \emph{heat conductivity} coefficient.
Since we consider the bounded domain $ \Om $, our goal is to discuss the solvability of the initial-boundary value problem for the system \eqref{NSF-c}-\eqref{NSF-e} endowed with the constitutive relations \eqref{visc_str} and \eqref{Fou_law} where we consider \emph{homogeneous Dirichlet boundary condition} in velocity and \emph{in-homogeneous Dirichlet boundary condition} in the temperature
\begin{align}
&\vc{u}|_{\partial \Om} = 0, \label{velocity_bdry}\\
&\vartheta|_{\partial \Om} = \vartheta_B>0. \label{temp_bdry}
\end{align}
\par
The consideration of \eqref{temp_bdry} is closely related to the celebrated \emph{Rayleigh-Benard problem}.
For the above system \eqref{NSF-c}-\eqref{NSF-e} with the consitutive relations \eqref{GE}-\eqref{temp_bdry} and the pressure following Boyle's law, the existence of a\emph{ local in time} strong solution was proved by Valli and Zajaczkowski \cite{VZ1986}. For the global in time weak solutions there are many articles and monographs by P.L. lions,\cite{PLL1996} Bresch and Desardin\cite{BD2007}, Feireisl and Novotn\'y \cite{FN2009b}, Bresch and Jabin \cite{BJ2018} which focus mainly on space-periodic domains or energy-conserving boundary conditions for the temperature. Also for thermodynamically open systems, Feireisl and Novotn\'y \cite{FN2021} prove the existence of weak solutions, but also require the control of the internal (heat) energy flux $ \vc{q} $ in $ \partial \Om $. Recently, the concept of weak solution was introduced for the particular boundary condition \eqref{temp_bdry} in \cite{CF2021} and it also considers very general in flow-outflow boundary data. This weak solution approach is slightly different from the earlier considerations that we will discuss in Section 2. This similar idea is adapted by Pokorn\' y in \cite{P2022} for steady flows.\par
The concept of measure valued solutions in the context of inviscid and incompressible fluids has been used by several authors, beginning with the seminal work of DiPerna \cite{D1985} and DiPerna and Majda \cite{DM1987}, see also related results of Kr\"oner and Zajaczkowski \cite{KJ1996}, Necasova and Novotn\'y\cite{NN1994} and Neustupa \cite{Nes1993}. B\v rezina, Feireisl and Novotn\'y\cite{BFN2020} also introduce the measure valued solution for compressible viscous heat conducting fluids for the "no flux" boundary condition. Their definition follows the approach of considering measure valued solution as a general object as described by Brenier, De Lellis and Szekelyhidi \cite{BDS11}, in contrast to the approach described by Malek et. al. \cite{MNRR1996}, where they identify a measure valued solution as a weak limit of weak solutions. We follow here the approach proposed by \cite{BFN2020}.\par
For a generalized (weak or measure valued) solution, it is quite important to prove the \emph{generalized-strong uniqueness principle}. The principle asserts that given the same initial data a weak solution will coincide with the strong or classical solution if the latter exists. For such properties in the context of weak solutions, see Feireisl and Novotn\'y\cite{FN2012} and Feireisl \cite{F2012}. For measure valued solution for the system was avalaible in B\v rezina et al\cite{BFN2020}. The main idea here is to use the \emph{relative energy inequality}. It was first introduced by Daefermos \cite{D1979} for scalar conservation law and the for compressible fluid a suitable adaptation is available in Feireisl and Novotn\'y \cite{FN2009b}. \par
An important application of measure valued solutions is their identification as limits of numerical schemes. Interesting results in numerical analysis have been obtained by Fjordholm et al. \cite{FjMT2012}, Feireisl and Luk\'a\v cov\'a-Medvid'ov\'a \cite{FL2018} and Feireisl, Luk\'a\v cov\'a-Medvid'ov\'a and Mizerov\'a\cite{FLM2020i}. Together with the existing weak(measure valued)-strong uniqueness principle in the class of measure valued solutions, one can show that numerical solutions converge strongly to a strong solution of the system as long as the latter exists. In the context of compressible Navier-Stokes-Fourier system with no-flux boundary condition see Feireisl et al\cite{FLMS2021} and a detailed discussion is available in the monograph by Feireisl et al. \cite{FLMS2021book}. \par
The plan for the article is as follows:
\begin{itemize}
\item At first in Section \ref{s2}, we devote ourselves to the weak formulation of the problem along with a proper definition of the measure valued solution.
\item In Section \ref{s3}, adapting the relative energy inequality suitably for the measure valued solution, we derive the relative energy inequality for the measure valued solution.
\item We present our main results in Section \ref{s4}. We conclude the weak(measure valued)-strong uniqueness property for the system. Our first two theorems, Theorem \ref{thc1} and Theorem \ref{thc2}, requires some additional hypothesis on measure valued solutions, while Theorem \ref{thu} is does not need any of such extra assumption on solutions, although we impose some physically relevant structural assumptions on transport coefficients $ \mu, \lambda $ and $ \kappa $.
\item Finally, in the Section \ref{s5}, we discuss briefly the limitation of our results, some comments on existence and validation of the definition.
\end{itemize}
\section{Measure valued solution}\label{s2}
\subsection{Weak formulation: Revisit}
At first we quickly recall the weak formulation described in Feireisl and Novotn\'y \cite[Chapter 2]{FN2009b}. Although it deals with a no-flux boundary condition for temperature, i.e.
\begin{align}\label{temp-bdry-2}
\vc{q} \cdot \vc{n}=0 \text{ on } \partial \Om.
\end{align}
The weak formulation of the problem depends on the weak formulation of continuity equation, momentum equation and replaces the \emph{internal energy equation} \eqref{NSF-e} by the \emph{entropy inequality}
\begin{align}\label{ent_ineq}
\begin{split}
\partial_t (\varrho s(\varrho, \vartheta)) + \text{div}_x (\varrho s(\varrho, \vartheta) \vc{u}) + \text{div}_x \left( \frac{ \vc{q}(\varrho,\vartheta, \nabla_x \vartheta ) }{\vartheta} \right)\nonumber \\ \geq \frac{1}{\vartheta}
\left( \mathbb{S} (\varrho,\vartheta, \bb{D}_x \vc{u} ) : \bb{D}_x \vc{u} - \frac{\vc{q}(\varrho,\vartheta, \nabla_x \vartheta) \cdot \nabla_x \vartheta }{\vartheta} \right),
\end{split}
\end{align}
along with the weak form of \emph{total energy balance} :
\begin{align}
\frac{\text{d} }{\;\text{d}t\,} &\intO{ \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e \right] } = \intO{ \varrho \vc{u} \cdot \vc{g} } .
\end{align}
Instead of boundary condition \eqref{temp-bdry-2}, if we consider inhomogeneous Dirichlet boundary condition for the temperature \eqref{temp_bdry}, we need to proceed as prescribed \cite[Section 2.4]{CF2021}. In this case, assuming all quantities in consideration are smooth, we have
\begin{align}
\partial_t \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e \right] + \text{div}_x\left( \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e +p(\varrho,\vartheta)\right] \vc{u} \right) + \text{div}_x \vc{q}=\text{div}_x\left(\bb{S}\vc{u}\right)+\varrho \vc{u} \cdot \vc{g},
\end{align}
and consequently
\begin{align}
\frac{\text{d} }{\;\text{d}t\,} &\intO{ \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e \right] } +
\int_{\partial \Omega} \vc{q} \cdot \vc{n} \ \text{d} \sigma_x
= \intO{ \varrho \vc{u} \cdot \vc{g} }
\label{w1}
\end{align}
where $ \sigma_x $ is the surface measure on the boundary $ \partial \Om $. Therefore, the total energy balance is unavailable. But, considering a smooth function $ \tilde{\vartheta} $ such that \[ \tilde{\vartheta} >0 \text{ in } (0,T) \times \Om \text{ with } \tilde{\vartheta} = \vartheta_B \text{ on }\partial \Om,\]
and multiplying the entropy inequality \eqref{ent_ineq} by $\tilde{\vartheta}$ along with performing the integrating by parts formula, we obtain
\begin{align}
- \frac{\text{d} }{\;\text{d}t\,} &\intO{ \tilde{\vartheta} \varrho s } - \int_{\partial \Omega} \vc{q} \cdot \vc{n}\, \text{d} \sigma_x
\nonumber\\ &\leq - \intO{ \frac{\tilde{\vartheta}}{\vartheta} \left( \mathbb{S} : \bb{D}_x \vc{u} - \frac{\vc{q} \cdot \nabla \vartheta }{\vartheta} \right) }
- \intO{ \left[ \varrho s \left( \partial_t \tilde{\vartheta} + \vc{u} \cdot \nabla \tilde{\vartheta} \right) + \frac{\vc{q}}{\vartheta} \cdot \nabla \tilde{\vartheta} \right] }.
\label{w2}
\end{align}
Adding \eqref{w1} and \eqref{w2} we obtain the \emph{ballistic energy inequality}
\begin{align*}
\frac{\text{d} }{\;\text{d}t\,} &\intO{ \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right] } + \intO{ \frac{\tilde{\vartheta}}{\vartheta} \left( \mathbb{S} : \bb{D}_x \vc{u} - \frac{\vc{q} \cdot \nabla \vartheta }{\vartheta} \right) }\nonumber \\
&\leq \intO{ \varrho \vc{u} \cdot \vc{g} }- \intO{ \left[ \varrho s \left( \partial_t \tilde{\vartheta} + \vc{u} \cdot \nabla \tilde{\vartheta} \right) + \frac{\vc{q}}{\vartheta} \cdot \nabla \tilde{\vartheta} \right] }.
\end{align*}
We recall that for some smooth function $ \widetilde{\Theta} $, the ballistic energy is denoted by $H_{\widetilde{\Theta}}(\varrho,\vartheta) $ and it reads as \[ H_{\widetilde{\Theta}}(\varrho,\vartheta) = \varrho [e(\varrho,\vartheta)- \widetilde{\Theta} s(\varrho,\vartheta)] .\]
Therefore we consider the weak formulation with ballistic energy inequality instead of energy balance.
In this paper our goal is to provide a suitable definition of a measure valued solution based on the above discussion for boundary condition \eqref{temp_bdry}. To define it, we take motivation from B\v rezina, Feireisl and Novotn\' y \cite{BFN2020} that covers the the boundary condition \eqref{temp-bdry-2}. This definition will give in terms of Young measure and suitable defect measures. For a simpler consideration, from now on we consider $ \vc{g}=0 $.
\subsubsection{Phase space and Young measure} A natural candidate for the phase space is given by the state variables $ [\varrho,\vc{u},\vartheta] $. Since we are looking for a more general class of solutions and $ \nabla_x \vc{u} $ and $ \nabla_x \vartheta $ are present in the system \eqref{NSF-c}-\eqref{NSF-e}, thus gradient of velocity and temperature have been included along with the natural choices. Hence a proper phase space is
\begin{align}
\mathcal{F}=\{[\varrho,\vc{u},\vartheta,\mathbb{D}_{\mathbf{u}}, \vc{D}_{\vartheta}] \big | \varrho\in [0,\infty),\; \mathbf{u}\in \R^d,\; \vartheta \in [0,\infty),\; \mathbb{D}_{\mathbf{u}}\in \R^{d\times d}_{\text{sym}},\; \vc{D}_\vartheta \in \R^{d} \}.
\end{align}
We consider a Young measure $ \mathcal{V} $ such that $\mathcal{V} \equiv \{ \mathcal{V}_{t,x} \}_{(t,x)\in (0,T)\times \Omega}$ and
\begin{align}\label{mv-ym}
\mathcal{V} \in L^{\infty}_{\text{weak-(*)}} \big( (0,T)\times \Om ;\mathcal{P}(\mathcal{F} )\big),
\end{align}
\subsubsection{Initial data}
Let $ \mathcal{V}_0 \in L^{\infty}_{\text{weak-(*)}} \big( \Om ;\mathcal{P}(\mathcal{F}_0 )\big) $ such that
$ \mathcal{F}_0= \{ (\varrho_0, \vc{u}_0, \vartheta_0)| \varrho_0 \in [0,\infty),\, \vc{u}_0\in \R^d,\, \vartheta_0\in [0,\infty)\} $ with finite energy, i.e.,
\begin{align}\label{mv-ic}
\int_{ \Om} \left\langle \mathcal{V}_{0,x}; \left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta}(0,x) \varrho s(\varrho,\vartheta)\right) \right\rangle \;\!\text{d}x < \infty
\end{align}
for each $0< \widetilde{\Theta} \in C^1([0,T] \times \Ov{\Om}) $ with $ \widetilde{\Theta}= \vartheta_B $ on $\partial \Om $.
\subsubsection{Compatibility relations for the Young measure}
We consider a very general phase space for the Young measure. It satisfies the following compatibility conditions.
\begin{itemize}
\item \textbf{Velocity compatibility:} The identity
\begin{align}\label{mv-vc}
-\int_{0}^{T} \int_{ \Om}\left\langle \mathcal{V}_{t,x}; \vc{u} \right \rangle \cdot \text{div}_x \bb{T} \;\!\text{d}x \;\text{d}t\, = \int_{0}^{T} \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \bb{D}_{\vc{u}} \right \rangle \;\!\text{d}x \;\text{d}t\,
\end{align}
holds for any $ \bb{T} \in C^1(\Ov{Q_T}; \R^{d\times d}_{\text{sym}}) $.
\item \textbf{Temperature compatibility:} The identity
\begin{align}\label{mv-tc}
-\int_{0}^{T} \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \vartheta - \widetilde{\Theta} \right \rangle \text{div}_x \psi\;\!\text{d}x \;\text{d}t\, = -\int_{0}^{T} \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \vc{D}_\vartheta - \nabla_x \widetilde{\Theta} \right \rangle \psi \;\!\text{d}x \;\text{d}t\,
\end{align}
holds for any $ \psi \in C^1([0,T]\times \Ov{\Om}) $ and $ \widetilde{\Theta} \in C^1([0,T]\times \Ov{\Om}) $ with $ \widetilde{\Theta}= \vartheta_B $ on $\partial \Om $.
\end{itemize}
\subsubsection{Field equations with defect}
\begin{itemize}
\item \textbf{Equation of continuity:} For a.e. $\tau \in (0,T) $ and $\psi \in C^{1}([0,T]\times \bar{\Om})$, the following equation holds:
\begin{align} \label{mv-cont-eqn}
\begin{split}
&\int_{ \Om} \langle \mathcal{V}_{\tau,x} ; \varrho \rangle \psi(\tau, \cdot) \;\!\text{d}x - \int_{ \Om} \langle \mathcal{V}_{0,x} ; \varrho \rangle \psi(0, \cdot) \;\!\text{d}x \\
&\quad = \int_{0}^{\tau} \int_{ \Om} \big[ \langle \mathcal{V}_{t,x}; \varrho \rangle \partial_{t}\psi + \langle \mathcal{V}_{t,x}; \varrho \vc{u} \rangle \cdot \nabla_x \psi \big] \;\!\text{d}x \;\text{d}t\,.
\end{split}
\end{align}
\item \textbf{Momentum equation:} There exists a \textit{measure} $r^{M} \in L^\infty_{\text{weak-(*)}}(0,T;\mathcal{M}(\bar{\Om};\R^{d\times d}))$ such that for a.e. $\tau \in (0,T)$ and every $\pmb{\varphi}\in C^{1}([0,T]\times \bar{\Om};\R^d)$, $\pmb{\varphi}|_{\partial \Om} =0$ the following equation holds:
\begin{align}\label{mv-mom-eqn}
\begin{split}
&\int_{ \Om} \langle \mathcal{V}_{\tau,x}; \varrho \vc{u} \rangle \cdot \pmb{\varphi}(\tau,\cdot) \;\!\text{d}x - \int_{ \Om} \langle \mathcal{V}_{0,x} ;\varrho \vc{u} \rangle \cdot \pmb{\varphi}(0,\cdot) \;\!\text{d}x\\
&= \int_{0}^{\tau} \int_{ \Om} \big[ \langle \mathcal{V}_{t,x}; \varrho \vc{u} \rangle \cdot \partial_{t} \pmb{\varphi} + \langle \mathcal{V}_{t,x}; \varrho \vc{u} \otimes \mathbf{u}) \rangle : \nabla_x \pmb{\varphi} + \langle \mathcal{V}_{t,x}; p(\varrho,\vartheta) \rangle \text{div}_x \pmb{\varphi} \big] \;\!\text{d}x \;\text{d}t\,\\
& - \int_{0}^{\tau} \int_{ \Om} \langle \mathcal{V}_{t,x};\mathbb{S}(\varrho,\vartheta,\mathbb{D}_{\mathbf{v}}) \rangle : \nabla_x \pmb{\varphi} \;\!\text{d}x \;\text{d}t\, + \int_{0}^{\tau} \langle r^M;\nabla_x \pmb{\varphi} \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\,.
\end{split}
\end{align}
\item \textbf{Entropy inequality:} For a.e. $\tau \in (0,T) $ and $0\leq \phi \in C_c^{1}([0,T]\times {\Om})$, we have the following inequality:
\begin{align}\label{mv-ent-ineq}
\begin{split}
&\int_{ \Om} \langle \mathcal{V}_{\tau,x} ; \varrho s(\varrho,\vartheta) \rangle \phi(\tau,x) \;\!\text{d}x - \int_{ \Om} \langle \mathcal{V}_{0,x} ; \varrho s(\varrho,\vartheta) \rangle \phi(0,x) \;\!\text{d}x \\
&\geq \int_0^\tau \int_{ \Om} \langle \mathcal{V}_{t,x} ; \varrho s(\varrho,\vartheta) \rangle \partial_t \phi(t,x) \;\!\text{d}x \;\text{d}t\, \\
& + \int_0^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \varrho s(\varrho,\vartheta) \vc{u} - \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vc{D}_\vartheta \right\rangle \cdot \nabla_x \phi(t,x) \;\!\text{d}x \;\text{d}t\, \\
&+ \int_0^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \frac{1}{\vartheta} \left( \bb{S}(\varrho,\vartheta,\nabla_{x}\vc{u}) \colon \bb{D}_x \vc{u} + \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vert \vc{D}_\vartheta \vert^2 \right) \right\rangle \phi(t,x) \;\!\text{d}x \;\text{d}t\,
\end{split}
\end{align}
\item \textbf{Ballistic energy inequality:} For any $\widetilde{\Theta} \in C^1([0,T]\times \Ov{\Om}),\, \widetilde{\Theta}>0,\; \widetilde{\Theta}|_{\partial\Om}= \vartheta_B$, there exists a \emph{dissipation defect} $\mathcal{D}_{\widetilde{\Theta}}$ such that
\begin{align*}
\mathcal{D}_{\widetilde{\Theta}}\in L^{\infty}(0,T),\; \mathcal{D}_{\widetilde{\Theta}}\geq 0,
\end{align*} and the following inequality holds:
\begin{align}\label{Bal-eng-ineq}
\begin{split}
&\int_{ \Om} \left \langle\mathcal{V}_{t,x} ; \left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta} \varrho s(\varrho,\vartheta)\right) \right \rangle \;\!\text{d}x \\
&+ \int_0^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \frac{1}{\vartheta} \left( \bb{S}(\varrho,\vartheta,\nabla_{x}\vc{u}) \colon \bb{D}_x \vc{u} + \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vert \vc{D}_\vartheta \vert^2 \right) \right\rangle \widetilde{\Theta}(t,x) \;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\widetilde{\Theta}}(\tau) \\
&\leq \int_{ \Om} \left\langle \mathcal{V}_{0,x}; \left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta} \varrho s(\varrho,\vartheta)\right) \right\rangle \;\!\text{d}x \\
&- \int_0^\tau \int_{ \Om} \langle\mathcal{V}_{t,x} \varrho s(\varrho,\vartheta) \rangle \partial_{t} \widetilde{\Theta}\;\!\text{d}x \;\text{d}t\, - \int_0^\tau \int_{ \Om} \langle\mathcal{V}_{t,x} \varrho s(\varrho,\vartheta) \vc{u} \rangle \cdot \nabla_x \widetilde{\Theta}\;\!\text{d}x \;\text{d}t\, \\
&+\int_0^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \frac{\kappa(\varrho,\vartheta) }{\vartheta}\mathbf{D}_\vartheta\right\rangle \cdot \nabla_x \widetilde{\Theta} \;\!\text{d}x \;\text{d}t\,.
\end{split}
\end{align}
\end{itemize}
\subsubsection{Compatibility of defect measures}
For any $\widetilde{\Theta} \in C^1([0,T]\times \Ov{\Om}),\, \widetilde{\Theta}>0,\; \widetilde{\Theta}|_{\partial\Om}= \vartheta_B$, we have
\begin{equation}\label{def-m-comp}
\vert \langle r^M(\tau);\nabla_x \pmb{\varphi} \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d\times d})\}} \vert \leq \xi(\tau) \mathcal{D}_{\widetilde{\Theta}}(\tau) \Vert \pmb{\varphi} \Vert_{C^1(\bar{\Om})}
\end{equation}
where $ \xi \in L^1(0,T) $.
\subsubsection{Generalized Korn--Poincar\'e inequality}
The following version of Korn-Poincar\' e inequality is true:
\begin{align}\label{mv-KP}
\int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \vert \vc{u} - \widetilde{\vc{U}} \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\, \leq C_p \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \vert \bb{D}_0(\bb{D}_\vc{u}) - \bb{D}(\nabla_x \widetilde{\vc{U}} ) \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\,
\end{align}
for any $ \widetilde{\vc{U}} \in L^2(0,T; W^{1,2}_0(\Om; \R^d)) $.
\begin{Rem}
We must note that the ballistic energy inequality is given for a large class of functions $ \widetilde{\Theta} $ and in general the dissipation defect $ \mathcal{D}_{\widetilde{\Theta}} $ is dependent of $ \tilde{\Theta} $. On the other hand the defect measure $ (r^M) $ in \eqref{mv-mom-eqn} is independent of $ \widetilde{\Theta} $. Therefore, the \eqref{def-m-comp} is very strong assumption. Although for some physical equation state we will able to conclude that $ \mathcal{D}_{\widetilde{\Theta}} $ is independent of $ \widetilde{\Theta} $, we will discuss it in Section \ref{s5}.
\end{Rem}
\begin{Rem}
Analogously, on can think of to have a defect measure in the right hand side of inequality \eqref{Bal-eng-ineq}. We are avoiding it. An explanation is available in \ref{s5}.
\end{Rem}
\subsection{Definition of a measure valued solution}
Here now we provide the definition of the measure valued solution
\begin{Def}\label{def:m}
Let $ \vartheta_B \in C^1((0,T)\times \partial \Om) $ with $ \vartheta_B >0 $ and $ \widetilde{\Theta} $ belongs to the class $ \{ \widetilde{\Theta} \in C^1([0,T]\times \Ov{\Om})\,| \widetilde{\Theta}>0,\; \widetilde{\Theta}|_{\partial\Om}= \vartheta_B\} $. Moreover, the initial condition $ \mathcal{V}_0 $ satisfies \eqref{mv-ic}. Then $ \{\mathcal{V},\mathcal{D}_{\widetilde{\Theta}}\} $ is a measure valued solution for the system \eqref{NSF-c}-\eqref{NSF-e} with \eqref{GE}, \eqref{visc_str}, \eqref{Fou_law}, \eqref{velocity_bdry} and \eqref{temp_bdry} if it satisfies \eqref{mv-cont-eqn}-\eqref{Bal-eng-ineq} with the compatibility condition \eqref{mv-vc},\eqref{mv-tc}, \eqref{def-m-comp} and \eqref{mv-KP}.
\end{Def}
\section{Relative energy inequality }\label{s3}
Following Feireisl and Novotn\' y in \cite[Chapter 9]{FN2009b}, we consider the relative energy with the help of Ballistic energy and it is given by
\begin{align*}
E(\vr,\vt,\vc{u}\,\vert\, \tvr,\tvt,\tu) = \frac{1}{2} \varrho \vert \vc{u} -\tilde{\vc{u}} \vert^2 + \left( H_{\tilde{\vartheta}} (\varrho,\vartheta) - \frac{\partial H_{\tilde{\vartheta}} (\tilde{\varrho},\tilde{\vartheta}) }{\partial \varrho } (\varrho -\tilde{\varrho}) - H_{\tilde{\vartheta}} (\tilde{\varrho},\tilde{\vartheta}) \right),
\end{align*}
where $ (\tvr,\tvt,\tu) $ are smooth functions such that they satisfy
\begin{align}
\tilde{\varrho}>0 , \tilde{\vc{u}} \text{ with }\tilde{\vc{u}}\vert_{\partial \Om}=0 \text{ and } \tilde{\vartheta}>0 \text{ with } \tilde{\vartheta}\vert_{\Om} = \vartheta_B .
\end{align}
As observed in \cite[Section 1.2]{FN2021}, if the relative energy is interpreted in terms of the conservative entropy variables $(\varrho, S = \varrho s, \vc{m} = \varrho \vc{u})$, represents a Bregman distance/divergence associated with the energy functional
\[
E(\varrho, S, \vc{m}) = \frac{1}{2} \frac{|\vc{m}|^2}{\varrho} + \varrho e(\varrho, S).
\]
To obtain that the reltive enrgy functional is non-negative, Indeed we need the \emph{hypothesis of thermodynamic stability}, i.e.
\begin{align}\label{ther-stab}
\frac{\partial p(\varrho, \vartheta)}{\partial \varrho} > 0,\; \frac{\partial e(\varrho, \vartheta)}{\partial \vartheta} > 0,
\end{align}
which in turn yields the convexity of the internal energy $\varrho e(\varrho,S)$ with respect to the variables $(\varrho,S)$. In addition,
\begin{equation} \label{ws2}
\frac{\partial (\varrho e(\varrho, S))}{\partial \varrho} = e - \vartheta s + \frac{p}{\varrho},\
\frac{\partial (\varrho e(\varrho, S))}{\partial S} = \vartheta.
\end{equation}
Thus the relative energy expressed in the conservative entropy variable may be interpreted as
\[
E \left( \varrho, S, \vc{m} \Big| \tilde{\varrho} , \tilde{S}, \tilde{\vc{m}} \right) = E(\varrho, S, \vc{m}) - \left< \partial E(\tilde{\varrho}, \tilde{S}, \tilde{\vc{m}}) ; (\varrho - \tilde{\varrho}, S - \tilde{S}, \vc{m} - \tilde{\vc{m}}) \right> - E(\tilde{\varrho}, \tilde{S}, \tilde{\vc{m}}).
\]
\begin{comment}
Using Gibbs relation we can rewrite the above expression as
\begin{align*}
E(\vr,\vt,\vc{u}\,\vert\, \tvr,\tvt,\tu) = &\frac{1}{2} \varrho \vert \vc{u} \vert^2 + H_{\tilde{\vartheta}}(\varrho,\vartheta) - \varrho \vc{u} \cdot \tilde{\vc{u}} \\
& + \varrho \left( \frac{1}{2} \vert \tilde{\vc{u}} \vert^2 - \frac{\partial H_{\tilde{\vartheta}} (\tilde{\varrho},\tilde{\vartheta}) }{\partial \varrho} \right) + p(\varrho,\vartheta) .
\end{align*}
\end{comment}
The time evolution of the relative energy is given by \[\mathcal{E}(\vrthu\,\vert\, \tildevrthu) = \int_{ \Om} E(\vrthu\,\vert\, \tildevrthu) \;\!\text{d}x .\]
Given a measure valued solution $ \{\mathcal{V}_{t,x}\}_{(t,x)\in Q_T} $ of Navier-Stokes-Fourier system, we adapt the relative energy as
\begin{align}\label{rel1}
\mathcal{E}_{mv}(\tau):= \int_{ \Om} \left\langle \mathcal{V}_{\tau,x} ; E(\vrthu\,\vert\, \tildevrthu) \right \rangle \;\!\text{d}x .
\end{align}
At first, using the standard expansion, we have
\begin{align}\label{rel2}
\begin{split}
\mathcal{E}_{mv}(\tau)= &\int_{ \Om} \left\langle \mathcal{V}_{\tau,x} ; \frac{1}{2} \varrho \vert \vc{u} \vert^2 + H_{\tilde{\vartheta}}(\varrho,\vartheta) \right \rangle \;\!\text{d}x- \int_{ \Om} \left\langle \mathcal{V}_{\tau,x} ; \varrho \vc{u} \right \rangle \cdot \tilde{\vc{u}} \;\!\text{d}x \\
& + \int_{ \Om} \left\langle \mathcal{V}_{\tau,x} ; \varrho \right \rangle \left( \frac{1}{2} \vert \tilde{\vc{u}} \vert^2 - \frac{\partial H_{\tilde{\vartheta}} (\tilde{\varrho},\tilde{\vartheta}) }{\partial \varrho} \right) \;\!\text{d}x + \int_{ \Om} p(\tilde{\varrho},\tilde{\vartheta})(\tau,\cdot) \;\!\text{d}x = \Sigma_{i=1}^{4} \mathcal{L}_i.
\end{split}
\end{align}
In the identity \eqref{rel2}, the term $ \mathcal{L}_1 $ is associated with the ballistic energy inequality \eqref{Bal-eng-ineq}, while for the terms $ \mathcal{L}_2 $ and $ \mathcal{L}_3 $ we use the momentum equation \eqref{mv-mom-eqn} and the continuity equation \eqref{mv-cont-eqn}, respectively. For the term $ \mathcal{L}_4 $ we use the identity
\[ \int_0^ \tau \int_{ \Om} \partial_{t} p(\tilde{\varrho},\tilde{\vartheta}) \;\!\text{d}x \;\text{d}t\, = \left[ \int_{ \Om} p(\tilde{\varrho},\tilde{\vartheta})(t,\cdot) \;\!\text{d}x\right]_{t=0}^{t=\tau} .\]
A suitable application of the Gibbs' relation \eqref{GE} yields
\begin{align}\label{rel3}
\begin{split}
\mathcal{E}_{mv}(\tau) & + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \frac{\tilde{\vartheta}(t,x)}{\vartheta}\bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \colon \bb{D}_{\vc{u}} \right \rangle + \tilde{\vartheta}(t,x) \left\langle \mathcal{V}_{t,x} ; \frac{\kappa(\varrho,\vartheta)}{\vartheta^2} \vert \vc{D}_\vartheta \vert^2\right \rangle \right] \!\!\;\!\text{d}x \;\text{d}t\,\\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} \colon \bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \right \rangle \colon \nabla_x \tilde{\vc{u}}(t,x) \;\!\text{d}x \;\text{d}t\, + \mathcal{D}(\tau)\\
\leq &\mathcal{E}_{mv}(0)- \int_0^ \tau \int_{ \Om} \bigg[ \left\langle \mathcal{V}_{t,x} ; \varrho s(\varrho,\vartheta) \right \rangle \partial_{t} \tilde{\vartheta} (t,x) \\
& \hspace{ 22mm } +\left\langle \mathcal{V}_{t,x} ; \left( \varrho s(\varrho,\vartheta)\vc{u} - \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vc{D}_\vartheta \right) \right \rangle \cdot \nabla_x \tilde{\vartheta} (t,x) \bigg] \;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \big[ \left\langle \mathcal{V}_{t,x} ; \varrho ( \tilde{\vc{u}}(t,x)-\vc{u}) \right \rangle \cdot \partial_{t} \tilde{\vc{u}}(t,x) \\
& \hspace{ 22mm } + \left\langle \mathcal{V}_{t,x} ; \varrho( \tilde{\vc{u}}(t,x)-\vc{u}) \otimes \vc{u} \right \rangle \colon \nabla_x \tilde{\vc{u}}(t,x) \big] \;\!\text{d}x \;\text{d}t\,\\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; p(\varrho,\vartheta ) \right \rangle \text{div}_x \tilde{\vc{u}}(t,x) \;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \varrho \right \rangle \partial_{t} \tilde{\vartheta}(t,x) s(\tilde{\varrho},\tilde{\vartheta})(t,x) + \left\langle \mathcal{V}_{t,x} ; \varrho \vc{u} \right \rangle \cdot \nabla_x \tilde{\vartheta}(t,x) \, s(\tilde{\varrho},\tilde{\vartheta})(t,x) \right]\\
& + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \tilde{\varrho}(t,x)- \varrho \right \rangle \frac{1}{\tilde{\varrho}} \partial_{t} p(\tilde{\varrho},\tilde{\vartheta})(t,x) - \left\langle \mathcal{V}_{t,x} ; \varrho \vc{u} \right \rangle \cdot \frac{1}{\tilde{\varrho}} \nabla_x \partial_{t} p(\tilde{\varrho},\tilde{\vartheta})(t,x) \right]\\
& + \int_{0}^{\tau} \langle r^M;\nabla_x \tilde{\vc{u}}(t,x) \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\,.
\end{split}
\end{align}
Our main goal is to establish the weak (measure valued)-- strong uniqueness property. To obtain this result we choose $ (\tilde{\varrho},\tilde{\vc{u}},\tilde{\vartheta}) $ as a strong solution of the problem emanating from the same initial data $ \mathcal{V}_{0,x} $ and they share same boundary condition. Thus, suppose that $ (\tilde{\varrho},\tilde{\vc{u}},\tilde{\vartheta}) $ is smooth and $ \varrho, \vartheta >0 $ in $ (0,T)\times \Om $, then the inequality\eqref{rel3} reduces to
\begin{align*}
\mathcal{E}_{mv}(\tau) & + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \frac{\tilde{\vartheta}(t,x)}{\vartheta}\bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \colon \bb{D}_{\vc{u}} \right \rangle \right] \;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta}(t,x) \left\langle \mathcal{V}_{t,x} ; \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vc{D}_\vartheta \left( \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right)\right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} \colon \bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \right \rangle \colon \nabla_x \tilde{\vc{u}}(t,x) \;\!\text{d}x \;\text{d}t\, \\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; (\bb{D}_\vc{u}- \bb{D}(\tilde{\vc{u}})) \right \rangle \colon \bb{S}(\tilde{\varrho},\tilde{\vartheta},\nabla_x \tilde{\vc{u}}) \;\!\text{d}x \;\text{d}t\, + \mathcal{D}(\tau)\\
\leq &\mathcal{E}_{mv}(0) - \int_0^ \tau \int_{ \Om} \tilde{\varrho} \left\langle \mathcal{V}_{t,x} ; s(\varrho,\vartheta)-s(\tilde{\varrho},\tilde{\vartheta}) \right \rangle \left( \partial_{t} \tilde{\vartheta} + \tilde{\vc{u}} \cdot \nabla_x \tilde{\vartheta} \right) \;\!\text{d}x \;\text{d}t\, \\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \vc{u}- \tilde{\vc{u}} \right \rangle \cdot \nabla_x p(\tilde{\varrho},\tilde{\vartheta}) \;\!\text{d}x \;\text{d}t\, - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; p(\varrho,\vartheta ) \right \rangle \text{div}_x \tilde{\vc{u}}(t,x) \;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \tilde{\varrho}(t,x)- \varrho \right \rangle \frac{1}{\tilde{\varrho}} \partial_{t} p(\tilde{\varrho},\tilde{\vartheta})(t,x) - \left\langle \mathcal{V}_{t,x} ; \varrho \vc{u} \right \rangle \cdot \frac{1}{\tilde{\varrho}} \nabla_x p(\tilde{\varrho},\tilde{\vartheta})(t,x) \right]\\
& + \int_{0}^{\tau} \langle r^M;\nabla_x \tilde{\vc{u}}(t,x) \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\, + \int_0^ \tau \int_{ \Om} R_1 \;\!\text{d}x \;\text{d}t\,,
\end{align*}
where
\begin{align*}
R_1= &\left\langle \mathcal{V}_{t,x}; \varrho (\tilde{\vc{u}}(t,x) - \vc{u}) \otimes (\vc{u}- \tilde{\vc{u}}(t,x)) \right \rangle \colon \bb{D}_x \tilde{\vc{u}}(t,x) \\
&+ \left\langle \mathcal{V}_{t,x}; \left( \frac{\varrho}{\tilde{\varrho}(t,x)}-1\right)(\tilde{\vc{u}}(t,x) -\vc{u}) \right \rangle \cdot \text{div}_x \bb{S}(\tilde{\vartheta}, \bb{D}_x \tilde{\vc{u}})(t,x)\\
& +\left\langle \mathcal{V}_{t,x}; \left( \frac{\varrho}{\tilde{\varrho}(t,x)}-1\right)(\tilde{\vc{u}}(t,x) -\vc{u}) \right \rangle \cdot \nabla_x p(\tilde{\varrho},\tilde{\vartheta})(t,x) \\
& - \left\langle \mathcal{V}_{t,x}; \varrho (s(\varrho,\vartheta)- s(\tilde{\varrho},\tilde{\vartheta})) (\tilde{\vc{u}}-\vc{u})\right \rangle \cdot \nabla_x \tilde{\vartheta} \\
&- \left\langle \mathcal{V}_{t,x}; (\varrho-\tilde{\varrho})(s(\varrho,\vartheta)-s(\tilde{\varrho},\tilde{\vartheta})) \right \rangle \left( \partial_{t} \tilde{\vartheta} + \tilde{\vc{u}} \cdot \nabla_x \tilde{\vartheta} \right).
\end{align*}
Moreover, adjusting a few terms suitably, the above inequality becomes
\begin{align*}
\mathcal{E}_{mv}(\tau) & + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \frac{\tilde{\vartheta}(t,x)}{\vartheta}\bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \colon \bb{D}_{\vc{u}} \right \rangle \right] \;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta}(t,x) \left\langle \mathcal{V}_{t,x} ; \frac{\kappa(\varrho,\vartheta)}{\vartheta} \bb{D}_\vartheta \left( \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right)\right \rangle \;\!\text{d}x \;\text{d}t\,\\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} \colon \bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \right \rangle \colon \nabla_x \tilde{\vc{u}}(t,x) \;\!\text{d}x \;\text{d}t\, \\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; (\bb{D}_\vc{u}- \bb{D}(\nabla_x \tilde{\vc{u}})) \right \rangle \colon \bb{S}(\tilde{\varrho},\tilde{\vartheta},\nabla_x \tilde{\vc{u}}) \;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \tilde{\vartheta}-\vartheta \right \rangle \text{div}_x\left( \frac{\kappa(\tilde{\varrho},\tilde{\vartheta}) \nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right)\;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left(1 - \frac{\vartheta}{\tilde{\vartheta}}\right) \right \rangle \left( \bb{S}(\tilde{\varrho},\tilde{\vartheta},\bb{D}(\nabla_x \tilde{\vc{u}}))\colon \bb{D}(\nabla_x) \tilde{\vc{u}} + \frac{\kappa(\tilde{\varrho},\tilde{\vartheta}) \nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right) \;\!\text{d}x\;\text{d}t\, \\
& + \int_{0}^{\tau} \langle r^M;\nabla_x \tilde{\vc{u}}(t,x) \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\, + \int_0^ \tau \int_{ \Om} R_2 \;\!\text{d}x \;\text{d}t\,,
\end{align*}
where
\begin{align*}
R_2=&R_1+ \left\langle \mathcal{V}_{t,x}; \left( 1- \frac{\varrho}{\tilde{\varrho}(t,x)}\right)(\tilde{\vc{u}}(t,x) -\vc{u}) \right \rangle \cdot \nabla_x p(\tilde{\varrho},\tilde{\vartheta})(t,x) \\
&+ \left\langle \mathcal{V}_{t,x} ; p(\tilde{\varrho},\tilde{\vartheta}) - \frac{\partial p(\tilde{\varrho},\tilde{\vartheta})}{\partial \varrho} (\tilde{\varrho}-\varrho)- \frac{\partial p(\tilde{\varrho},\tilde{\vartheta})}{\partial \vartheta} (\tilde{\vartheta} -\vartheta) - p(\varrho,\vartheta) \right \rangle \text{div}_x \tilde{\vc{u}} \\
& + \left\langle \mathcal{V}_{t,x} ; s(\tilde{\varrho},\tilde{\vartheta}) - \frac{\partial s(\tilde{\varrho},\tilde{\vartheta})}{\partial \varrho} (\tilde{\varrho}-\varrho)- \frac{\partial s(\tilde{\varrho},\tilde{\vartheta})}{\partial \vartheta} (\tilde{\vartheta} -\vartheta) - s(\varrho,\vartheta) \right \rangle \tilde{\varrho} \left( \partial_{t} \tilde{\vartheta} + \tilde{\vc{u}} \cdot \nabla_x \tilde{\vartheta} \right).
\end{align*}
More precisely, we have
\begin{align}\label{R2}
\begin{split}R_2=&\left\langle \mathcal{V}_{t,x}; \varrho (\tilde{\vc{u}}(t,x) - \vc{u}) \otimes (\vc{u}- \tilde{\vc{u}}(t,x)) \right \rangle \colon \bb{D}_x \tilde{\vc{u}}(t,x) \\
&+ \left\langle \mathcal{V}_{t,x}; \left( \frac{\varrho}{\tilde{\varrho}(t,x)}-1\right)(\tilde{\vc{u}}(t,x) -\vc{u}) \right \rangle \cdot \left(\text{div}_x \bb{S}(\tilde{\vartheta}, \bb{D}_x \tilde{\vc{u}})(t,x)+ \nabla_x p(\tilde{\varrho},\tilde{\vartheta})(t,x)\right)\\
& - \left\langle \mathcal{V}_{t,x}; \varrho (s(\varrho,\vartheta)- s(\tilde{\varrho},\tilde{\vartheta})) (\tilde{\vc{u}}-\vc{u})\right \rangle \cdot \nabla_x \tilde{\vartheta} \\
&- \left\langle \mathcal{V}_{t,x}; (\varrho-\tilde{\varrho})(s(\varrho,\vartheta)-s(\tilde{\varrho},\tilde{\vartheta})) \right \rangle \left( \partial_{t} \tilde{\vartheta} + \tilde{\vc{u}} \cdot \nabla_x \tilde{\vartheta} \right)\\
&+\left\langle \mathcal{V}_{t,x}; \left( 1- \frac{\varrho}{\tilde{\varrho}(t,x)}\right)(\tilde{\vc{u}}(t,x) -\vc{u}) \right \rangle \cdot \nabla_x p(\tilde{\varrho},\tilde{\vartheta})(t,x) \\
&+ \left\langle \mathcal{V}_{t,x} ; p(\tilde{\varrho},\tilde{\vartheta}) - \frac{\partial p(\tilde{\varrho},\tilde{\vartheta})}{\partial \varrho} (\tilde{\varrho}-\varrho)- \frac{\partial p(\tilde{\varrho},\tilde{\vartheta})}{\partial \vartheta} (\tilde{\vartheta} -\vartheta) - p(\varrho,\vartheta) \right \rangle \text{div}_x \tilde{\vc{u}} \\
& + \left\langle \mathcal{V}_{t,x} ; s(\tilde{\varrho},\tilde{\vartheta}) - \frac{\partial s(\tilde{\varrho},\tilde{\vartheta})}{\partial \varrho} (\tilde{\varrho}-\varrho)- \frac{\partial s(\tilde{\varrho},\tilde{\vartheta})}{\partial \vartheta} (\tilde{\vartheta} -\vartheta) - s(\varrho,\vartheta) \right \rangle \tilde{\varrho} \left( \partial_{t} \tilde{\vartheta} + \tilde{\vc{u}} \cdot \nabla_x \tilde{\vartheta} \right)
\end{split}
\end{align}
Now we use the compatibility of the Young measure \eqref{mv-vc} and \eqref{mv-tc} to deduce
\begin{align}\label{REineq2}
\begin{split}
\mathcal{E}_{mv}(\tau) & + \int_0^ \tau \int_{ \Om} \left[ \left\langle \mathcal{V}_{t,x} ; \frac{\tilde{\vartheta}}{\vartheta}\bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \colon \bb{D}_{\vc{u}} \right \rangle + \left\langle \mathcal{V}_{t,x}; \frac{\vartheta}{\tilde{\vartheta}} \right \rangle \bb{S}(\tilde{\varrho}.\tilde{\vartheta}, \nabla_x \tilde{\vc{u}})\right] \;\!\text{d}x\;\text{d}t\,\\
& - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} \colon \bb{S}(\varrho,\vartheta, \bb{D}_{\vc{u}}) \right \rangle \colon \nabla_x \tilde{\vc{u}} \;\!\text{d}x\;\text{d}t\, - \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \bb{D}_\vc{u} \right \rangle \colon \bb{S}(\tilde{\varrho},\tilde{\vartheta},\nabla_x \tilde{\vc{u}}) \;\!\text{d}x \;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \frac{\kappa(\varrho,\vartheta)}{\vartheta} \vc{D}_\vartheta \left( \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right)\right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \kappa(\tilde{\varrho},\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; \vartheta \left( \frac{\nabla_x \vartheta}{\vartheta}- \frac{\vc{D}_\vartheta}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + \int_{0}^{\tau} \langle r^M;\nabla_x \tilde{\vc{u}}(t,x) \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\, + \int_0^ \tau \int_{ \Om} R_2 \;\!\text{d}x \;\text{d}t\,,
\end{split}
\end{align}
where the term $ R_2 $ is given by \eqref{R2} and it consists of quadratic error terms.
\begin{comment}
+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \tilde{\vartheta}-\vartheta \right \rangle \text{div}_x\left( \frac{\kappa(\tilde{\vartheta}) \nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right)\;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left(1 - \frac{\vartheta}{\tilde{\vartheta}}\right) \right \rangle \left( \bb{S}(\tilde{\varrho},\tilde{\vartheta},\bb{D}_x \tilde{\vc{u}})\colon \bb{D}_x \tilde{\vc{u}} + \frac{\kappa(\tilde{\vartheta}) \nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right) \;\!\text{d}x\;\text{d}t\, \\
\end{comment}
Let us now fix some notation to write \eqref{REineq2} more precisely. For $ \bb{A}\left(=(a_{ij})_{i,j=1}^{d} \right)\in \R^{d\times d} $, we consider the \textit{symmetric part} and the \textit{traceless part} of $ \bb{A} $ as
\begin{align*}
\bb{D}(\bb{A})= \frac{\bb{A}+\bb{A}^T}{2}\text{ and }
\bb{D}_0(\bb{A}) =\frac{\bb{A}+\bb{A}^T}{2}-\frac{1}{d}\text{Tr}(\bb{A})\, \bb{I},
\end{align*}
respectively, where $ \text{Tr}(\bb{A}) =\sum\limits_{i=1}^d a_{ii}$.
We write the Newtonian stress tensor as
\begin{align*}
\mathbb{S}(\nabla_x \mathbf{u})&=\mu(\varrho , \vartheta) \bigg(\frac{\nabla_x \mathbf{u} + \nabla_x^{T} \mathbf{u}}{2}-\frac{1}{d} (\text{div}_x\mathbf{u})\mathbb{I} \bigg) + \lambda(\varrho , \vartheta) (\text{div}_x \mathbf{u}) \mathbb{I},\\
&= \mu(\varrho , \vartheta) \bb{D}_0 \vc{u} + \lambda(\varrho,\vartheta) \text{div}_x \vc{u} \bb{I}.
\end{align*}
We rewrite the inequality \eqref{REineq2} as
\begin{align}\label{REineq3}
\begin{split}
\mathcal{E}_{mv}(\tau) &+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; (\mu(\varrho,\vartheta)- \mu(\tilde{\varrho},\tilde{\vartheta}) )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \lambda(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \text{div}_x \tilde{\vc{u}} \left\langle \mathcal{V}_{t,x}; (\lambda(\varrho,\vartheta)- \lambda(\tilde{\varrho},\tilde{\vartheta}) )\left( \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} (\text{div}_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(\varrho,\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \kappa(\tilde{\varrho},\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta-\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \\
&- \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\kappa(\varrho,\vartheta)-\kappa(\tilde{\varrho},\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + \int_{0}^{\tau} \langle r^M;\nabla_x \tilde{\vc{u}} \rangle_{\{\mathcal{M}(\bar{\Om};\R^{d \times d}),C(\bar{\Om};\R^{d \times d})\}} \;\text{d}t\, + \int_0^ \tau \int_{ \Om} R_2 \;\!\text{d}x \;\text{d}t\,.
\end{split}
\end{align}
The inequality \eqref{REineq3} is called the \emph{relative energy inequality} associated with the problem. Therefore we summarize the above discussion in the following lemma:
\begin{Lemma}\label{lem-1}
Let the transport coefficients $ \kappa(\varrho,\vartheta) $, $ \mu(\varrho,\vartheta) $ and $ \lambda(\varrho,\vartheta) $ be continuously differentiable and positive for $ \varrho >0,\; \vartheta >0 $. Let the thermodynamic functions satisfy Gibbs equation\eqref{GE} and the thermodynamic stability assumption\eqref{ther-stab}. Let $ \{\mathcal{V}, \mathcal{D}_{\tilde{\vartheta}}\}$ be a measure valued solution of the system \eqref{NSF-c}-\eqref{NSF-e} with initial data $ \mathcal{V}_0 $ and $ \{\tilde{\varrho}, \tilde{\vc{u}},\tilde{\vartheta}\} $ be a strong solution with sufficient regularity and initial data. Then we have the inequality \eqref{REineq3} holds, where the remainder term $ R_2 $ is given by \eqref{R2}.
\end{Lemma}
\subsection{A suitable reduction of relative energy inequality}
In this subsection we will try to reduce the \eqref{REineq3}. Since we already notice that the relative energy is a non-negative functional and the remainder term $ R_2 $ contains certain quadratic terms. Therefore we introduce the next part to have a close look on relative energy.
\subsubsection{Essential and residual part of a function}\label{ess-res-ss}
At first, we introduce a cut-off function
$\chi_\delta $ such that
\begin{align}\nonumber
\begin{split}
\chi_\delta \in C_c^{\infty} \left((0, \infty)^2\right),\; 0 \leq \chi_\delta \leq 1,\; \chi_\delta(\varrho,\vartheta) = 1 \mbox{if}\ \delta \leq \varrho \leq \frac{1}{\delta} \text{ and }\delta \leq \vartheta \leq \frac{1}{\delta} \text{ for some }\delta>0.
\end{split}
\end{align}
For a function $H = H(\varrho,\vartheta,\vc{u}, \bb{D}_{\vc{u}}, \vc{D}_\vt)$, we set
\begin{align}\nonumber
\begin{split}
[H]_{\text{ess}}=\chi_\delta(\varrho,\vartheta) H(\varrho,\vartheta,\vc{u}, \bb{D}_{\vc{u}}, \vc{D}_\vt) ,\; [H]_{\text{res}}=(1- \chi_\delta(\varrho,\vartheta)) H(\varrho,\vartheta,\vc{u}, \bb{D}_{\vc{u}}, \vc{D}_\vt).
\end{split}
\end{align}
If $ \tilde{\varrho}, \tilde{\vartheta} $ is strictly positive, bounded above and bounded below, $ \vert \tilde{\vc{u}} \vert $ is also bounded, then with the help of the above notation, we have
\begin{align}\label{RE_ess_res}
\begin{split}
E (\vr,\vt,\vc{u}\mid \tvr,\tvt,\tu)
\geq c(\delta,\tilde{\varrho},\tilde{\vartheta},\tilde{\vc{u}}) \bigg( & \left[ \left\vert \varrho-\tilde{\varrho} \right \vert^2+\left\vert \vartheta-\tilde{\vartheta} \right \vert^2+\left\vert \vc{u}-\tilde{\vc{u}} \right \vert^2 \right]_{\text{ess}} \\
& + \left[ 1 +\varrho+ \varrho \vert s(\varrho, \vartheta) \vert+ \varrho e(\varrho,\vartheta) + \varrho \vert \vc{u} \vert^2 \right]_{\text{res}} \bigg).
\end{split}
\end{align}
\begin{Rem}
The inequality \eqref{RE_ess_res} implies that the relative energy functional is coercive.
\end{Rem}
At this point, our goal is either to control the remainder term $ R_2$ of \eqref{REineq3}, by the integral of the relative energy or to absorb it into a non-negative term on the left-hand side of \eqref{REineq3}.
\begin{Lemma}
Let the hypothesis in Lemma \ref{lem-1} remains true. Then with the help of \eqref{RE_ess_res}, the inequality \eqref{REineq3} reduces to
\begin{align}\label{REineq4}
\begin{split}
\mathcal{E}_{mv}(\tau) &+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; (\mu(\varrho,\vartheta)- \mu(\tilde{\varrho},\tilde{\vartheta}) )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \lambda(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \text{div}_x \tilde{\vc{u}} \left\langle \mathcal{V}_{t,x}; (\lambda(\varrho,\vartheta)- \lambda(\tilde{\varrho},\tilde{\vartheta}) )\left( \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} (\text{div}_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(\varrho,\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
& + \int_0^ \tau \int_{ \Om} \kappa(\tilde{\varrho},\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta-\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \\
&- \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\kappa(\varrho,\vartheta)-\kappa(\tilde{\varrho},\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + C(\delta,\tvr,\tvt,\tu) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + \int_{0}^\tau \mathcal{D}_{\tilde{\vartheta}}(t) \;\text{d}t\, \\
& + \int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[\vartheta +\vert p(\varrho,\vartheta) \vert + \vert \vc{u}-\tilde{\vc{u}} \vert + \varrho \vert s(\varrho, \vartheta) \vert\; \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\, .
\end{split}
\end{align}
\end{Lemma}
\begin{proof}
The proof is quite straight forward, since $ R_2 $ contains quadratic terms, therefore the essential parts of $ R_2 $ will be controlled by $ C(\delta,\tvr,\tvt,\tu) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, $. For the residual parts of $ R_2 $, we control some of the terms that are compatible with the right hand side of \eqref{RE_ess_res} and we keep the other terms we in the right hand side of the \eqref{REineq4}.
\end{proof}
\section{Main results: Weak (measure valued)-strong uniqueness }\label{s4}
\subsection{Conditional Weak(measure valued)--strong uniqueness}
By the term `conditional', we mean that, along with the measure valued solution if we assume some additional hypothesis, them we will able to achieve the desired \emph{Weak(measure valued)--strong uniqueness} property.
\subsubsection{First conditional result}
First conditional result when density and temperature are uniformly bounded.
\begin{Th}\label{thc1}
Let the transport coefficients $ \kappa(\varrho,\vartheta) $, $ \mu(\varrho,\vartheta) $ and $ \lambda(\varrho,\vartheta) $ be continuously differentiable and positive for $ \varrho >0,\; \vartheta >0 $. Let the thermodynamic functions satisfy Gibbs equation \eqref{GE} and the thermodynamic stability \eqref{ther-stab}. Further assume that $ (\tilde{\varrho}, \tilde{\vc{u}}, \tilde{\vartheta} )$ be a strong solution of the system in $ [0,T]\times \Om $ emanating from the initial data $ (\varrho_0, \vc{u}_0, \vartheta_0 )$ with $ \varrho_0>0, \vartheta_0 $. Assume that $ \mathcal{V} $ be a measure valued solution of the same problem following the Definition \ref{def:m} and it satisfies \begin{align}\label{hyp_con1}
\mathcal{V}_{t,x} \left\{0< \underline{\varrho} < \varrho < \Ov{\varrho},\; 0<\underline{\vartheta}< \vartheta < \overline{\vartheta} \right\} =1 {\text{ for a.e. }} (t,x)\in Q_T,
\end{align}
for some constants $ \underline{\varrho}, \Ov{\varrho},\Un{\vartheta},\Ov{\vartheta} >0 $ and initial data coincides, i.e.
\begin{align*}
\mathcal{V}_{0,x}= \delta_{[\varrho_0,\vc{u}_0,\vartheta_0]} \text{ for a.e. } x\in \Om.
\end{align*}
Then \begin{align*}
\mathcal{V}_{t,x}= \delta_{[ \tilde{\varrho}(t,x), \tilde{\vc{u}}(t,x), \tilde{\vartheta}(t,x), \bb{D}_x \tilde{\vc{u}}(t,x), \nabla_x \tilde{\vartheta}(t,x)]}\text{ for a.e. } (t,x) \in Q_T.
\end{align*}
\end{Th}
\begin{proof}
At first we note that there exists a $ \delta >0 $, such that $ \delta \leq \min\{ \Un{\varrho}, \Un{\vartheta}\} $ and $ \max{\{\Ov{\varrho},\Ov{\vartheta}\}} \leq \frac{1}{\delta}$. There for we consider a cut-off function $ \chi_\delta $ as mentioned in \ref{ess-res-ss}. The hypothesis \eqref{REineq4} implies \begin{align*}
\int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[\vartheta +\vert p(\varrho,\vartheta) \vert + \vert \vc{u}-\tilde{\vc{u}} \vert + \varrho \vert s(\varrho, \vartheta) \vert\; \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\, =0.
\end{align*}
Also, form the assumption \eqref{hyp_con1}, we observe
\begin{align*}
& \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \lambda(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(\varrho,\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
&\geq \underline{C}(\delta, \tvr,\tvt,\tu) \int_0^ \tau \int_{ \Om} \bigg\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \\
&\hspace{42mm}+\left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2
+ \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \bigg\rangle \;\!\text{d}x\;\text{d}t\, \\
\end{align*}
and
\begin{align*}
&+ \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; (\mu(\varrho,\vartheta)- \mu(\tilde{\varrho},\tilde{\vartheta}) )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
&+ \int_0^ \tau \int_{ \Om} \text{div}_x \tilde{\vc{u}} \left\langle \mathcal{V}_{t,x}; (\lambda(\varrho,\vartheta)- \lambda(\tilde{\varrho},\tilde{\vartheta}) )\left( \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} (\text{div}_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
&- \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\kappa(\varrho,\vartheta)-\kappa(\tilde{\varrho},\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \kappa(\tilde{\varrho},\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta-\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \\
&\leq C(\epsilon, \tvr,\tvt,\tu) \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \varrho-\tilde{\varrho} \right \vert^2+\left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\, \\
&+ \epsilon \int_0^ \tau \int_{ \Om} \bigg\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \\
&\hspace{42mm}+\left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2
+ \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \bigg\rangle \;\!\text{d}x\;\text{d}t\, ,
\end{align*}
where $ \underline{C} (\delta,\tvr,\tvt,\tu)>0 $.
Therefore, choosing $ \epsilon $ suitably, we deduce
\begin{align}\label{ws1,1}
\begin{split}
\mathcal{E}_{mv}(\tau) &+ \frac{1}{2}\underline{C} (\delta,\tvr,\tvt,\tu) \int_0^ \tau \int_{ \Om} \bigg\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \\
&\hspace{42mm}+\left\vert \text{Tr} (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \text{div}_x \tilde{\vc{u}} \right\vert^2
+ \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \bigg\rangle \;\!\text{d}x\;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + C(\delta,\tvr,\tvt,\tu) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + \int_{0}^\tau \mathcal{D}_{\tilde{\vartheta}}(t) \;\text{d}t\, .
\end{split}
\end{align}
Employing Gr\" onwall's lemma we obtain the desired result.
\end{proof}
\subsubsection{Second conditional result}
Our next result is for a more physical situation with where pressure law is boyle's law. But the weak-strong uniqueness we prove is imposing certain condition on entropy and additional hypothesis on transport coefficients.
\begin{itemize}
\item The bulk and shear viscosity coefficient are given by
\begin{align}\label{tc1}
\mu(\varrho,\vartheta)= C_\mu (1+\vartheta) \text{ and } \lambda(\varrho,\vartheta)= C_\lambda (1+\vartheta) \text{ with } C_\mu>0,\; \C_\lambda \geq 0.
\end{align}
\item The heat conductivity coefficient is given by
\begin{align}\label{tc2}
\kappa(\varrho,\vartheta) = \kappa(1+\vartheta) \text{ with } \kappa >0.
\end{align}
\end{itemize}
\begin{Th}\label{thc2}
Let the transport coefficients $ \kappa(\varrho,\vartheta) $, $ \mu(\varrho,\vartheta) $ and $ \lambda(\varrho,\vartheta) $ be given by \eqref{tc1} and \eqref{tc2}. Let the thermodynamic functions satisfy Gibbs equation and the thermodynamic stability assumption with \begin{align}\label{pr-law-per}
p(\varrho,\vartheta)=\varrho \vartheta,\; e(\varrho,\vartheta)=c_v \vartheta ,\; s(\varrho,\vartheta)=\log \left(\frac{\vartheta^{c_v}}{\varrho}\right),\; c_v>1.
\end{align}
Further assume that $ (\tilde{\varrho}, \tilde{\vc{u}}, \tilde{\vartheta}) $ be a strong solution of the system in $ [0,T]\times \Om $ emanating from the initial data $ (\varrho_0, \vc{u}_0, \vartheta_0) $ with $ \varrho_0>0, \vartheta_0 $. Assume that $ \mathcal{V} $ be a measure valued solution of the same problem following the Definition \ref{def-m-comp} and it satisfies \begin{align}\label{hyp_con2}
\mathcal{V}_{t,x} \left\{ \vert s(\varrho,\vartheta) \vert \leq \bar{s} \right\} =1 {\text{ for a.e. }} (t,x)\in Q_T,
\end{align}
for some $ \bar{s}>0 $ and
\begin{align*}
\mathcal{V}_{0,x}= \delta_{[\varrho_0,\vc{u}_0,\vartheta_0]} \text{ for a.e. } x\in \Om.
\end{align*}
Then \begin{align*}
\mathcal{V}_{t,x}= \delta_{[ \tilde{\varrho}(t,x), \tilde{\vc{u}}(t,x), \tilde{\vartheta}(t,x), \bb{D}_x \tilde{\vc{u}}(t,x), \nabla_x \tilde{\vartheta}(t,x)]}\text{ for a.e. } (t,x) \in Q_T.
\end{align*}
\end{Th}
\begin{proof}
For the sake of simplicity we assume $ C_\lambda=0 $. Using the hypothesis \eqref{tc1}, we get
\begin{align*}
& \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& \geq \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \frac{1}{2}C_\mu (1+\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \frac{1}{4} \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& - \int_0^ \tau \int_{ \Om} \frac{1}{4}\left\langle \mathcal{V}_{t,x}; \tilde{\vartheta} \left\vert \left(1-\frac{\vartheta}{\tilde{\vartheta}}\right) \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\,
\end{align*}
and
\begin{align*}
&\left\vert \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; (\mu(\varrho,\vartheta)- \mu(\tilde{\varrho},\tilde{\vartheta}) )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\, \right \vert \\
& \leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ C(\epsilon, \tilde{\vartheta},\tilde{\vc{u}} ) \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\,.
\end{align*}
Also, \eqref{tc2} implies
\begin{align*}
& \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(1+\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, \geq C(\tilde{\vartheta}) \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\,.
\end{align*}
Similarly, using Young's inequality, we infer the following estimates
\begin{align*}
&\left\vert \int_0^ \tau \int_{ \Om} \kappa(\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta-\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \right \vert \\
&\leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, + C(\epsilon, \tilde{\vartheta} ) \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\,
\end{align*}
and
\begin{align*}
&\left\vert \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\kappa(\varrho,\vartheta)-\kappa(\tilde{\varrho},\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\,\right \vert\\
&\leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, + C(\epsilon, \tilde{\vartheta} ) \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\,.
\end{align*}
Now, for pressure law \eqref{pr-law-per}, we note that
$$ \vert p(\varrho,\vartheta) \vert \leq 1+ \varrho e(\varrho,\vartheta) + \varrho \vert s(\varrho,\vartheta) \vert .$$ Then using compatibility condition \eqref{mv-KP} and \eqref{RE_ess_res} we obtain
\begin{align}\label{ws2}
\begin{split}
\mathcal{E}_{mv}(\tau) &+ \frac{1}{2} \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; C_\mu(1+\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \frac{1}{4} \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(1+\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, + \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + C(\delta) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + \int_{0}^\tau \mathcal{D}_{\tilde{\vartheta}}(t) \;\text{d}t\, \\
& + \int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[\vartheta^2 + \varrho \vert s(\varrho, \vartheta) \vert\; \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\, +\epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, \\
&+ \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, .
\end{split}
\end{align}
Now using \eqref{hyp_con2} we have $ \vartheta^{c_v} \leq C(\bar{s}) \varrho$ and $ \vert \varrho s \vc{u} \vert \leq C(\bar{s} )\left (\varrho + \varrho \vert \vc{u} \vert^2\right)$.
This implies
\[ \vartheta^{c_v+1} \leq \varrho \vartheta = \frac{1}{c_v} \varrho e (\varrho,\vartheta). \]
By proper choice of $ \epsilon $, inequality \eqref{ws2} reduces to
\begin{align*}
\mathcal{E}_{mv}(\tau) &+ \frac{1}{4} \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; C_\mu(1+\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \frac{1}{8} \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& +\frac{1}{2} \int_0^ \tau \int_{ \Om} \tilde{\vartheta} \left\langle \mathcal{V}_{t,x} ; \kappa(1+\vartheta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\,\\
&+ \mathcal{D}_{\tilde{\vartheta}}(\tau)\\
\leq &\mathcal{E}_{mv}(0) + C(\delta) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + \int_{0}^\tau \mathcal{D}_{\tilde{\vartheta}}(t) \;\text{d}t\, .
\end{align*}
Again, invoking Gr\" onwall's lemma, we complete the proof.
\end{proof}
\subsection{Unconditional weak(measure valued)--strong uniqueness}
Here we consider a general pressure law and some structural assumption on the transport coefficients. We call it unconditional result because, here we will not assume any further information on measure valued solution.
Let us first consider the pressure law in the following way:
\begin{align}\label{Pr1}
p(\varrho,\vartheta)= p_M(\varrho,\vartheta)+ p_R(\varrho,\vartheta),
\end{align}
where $ p_M $ stands for the \emph{molecular pressure }and $ p_R $ is the \emph{radiation pressure}. The relation between the molecular pressure $ p_M $ and the associated internal energy $ e_M $ is is given by
\begin{align}\label{pm,em}
p_M(\varrho,\vartheta)= \frac{2}{3} \varrho e_M(\varrho,\vartheta).
\end{align}
From the Gibb's relation \eqref{GE}, we have
\[p_M(\varrho,\vartheta) = \vartheta^{\frac{5}{2}} P\left( \frac{\varrho}{\vartheta^{\frac{3}{2}}}\right) ,\]
for some function $ P $. Moreover, following \cite{FN2009b}, we assume that $ P\in C^1[0,\infty) \cap C^5(0,\infty)$,
\begin{align}
\begin{split}
& P(0)=0, \; P^\prime(q)>0,\text{ for all } q>0,\\
&0< \frac{\frac{5}{3} P(q)- P^\prime(q)q }{q} <c \text{ for all }q < 0, \lim\limits_{q\rightarrow \infty } \frac{P(q)}{q^{\frac{5}{3}}}= \overline{p}>0.
\end{split}
\end{align}
We rewrite the internal energy $ (e_M) $ associated with molecular pressure as \[ e_M(\varrho)= \frac{3}{2} \frac{\vartheta^{\frac{5}{2}}}{\varrho} P\left( \frac{\varrho}{\vartheta^{\frac{3}{2}}}\right), \]
and, again using Gibbs relation, we have
\begin{align}\label{sm}
s_M= S\left( \frac{\varrho}{\vartheta^{\frac{3}{2}}}\right),
\end{align}
for some function $ S $ with the property
\begin{align}
S^\prime (q)= -\frac{3}{2} \frac{\frac{5}{3} P(q)- P^\prime(q)q }{q^2} <0 .
\end{align}
Finally, we impose the third law of thermodynamics in the form
\begin{align}\label{3rd law}
\lim\limits_{q\rightarrow \infty } S(q) =0.
\end{align}
the structural assumptions on transport coefficient reads as
\begin{align}\label{k,mu,lamda}
\begin{split}
&\kappa(\varrho,\vartheta)= \kappa_1 + \kappa_2 \vartheta^\beta, \text{with} \kappa_1>0,\; \kappa_2 \geq 0 \text{ and } \beta \leq 2 \\
& \mu(\varrho,\vartheta) = \mu_0+ \mu_1 \vartheta,\; \mu_0,\mu_1\geq 0,\\
&\lambda(\varrho,\vartheta) = \lambda_0+ \lambda_1 \vartheta,\; \lambda_0,\lambda_1\geq 0.
\end{split}
\end{align}
\begin{Th}\label{thu}
Let the pressure follows \eqref{Pr1} with $ p_M, e_M $ and $ s_M $ given by \eqref{pm,em} and \eqref{sm}, respectively, $ \kappa,\mu, \lambda $ satisfy \eqref{k,mu,lamda} and the radiation pressure along with associated internal energy and entropy are given by \begin{align}\label{rad_p1}
p_R= a \vartheta^2,\; e_R= a \frac{\vartheta^2}{\varrho} \text{ and } s_R= 2a\frac{\vartheta}{\varrho} \text{ with } a>0.
\end{align}
Assume that $ [\tilde{\varrho},\tilde{\vc{u}},\tilde{\vartheta}] $ is a classical solution to the system in $ [0,T] \times \Om $ emanating from initial data $ [\varrho_0,\vc{u}_0,\vartheta_0] $ with $ \varrho_0,\vartheta_0>0 $ in $ \Ov{\Om} $. Assume further that $ \{ \mathcal{V}_{t,x}\}_{(t,x)\in Q_T} $ is a measure valued solution of the same problem following the Definition \ref{def-m-comp} such that \[ \mathcal{V}_{0,x}= \delta_{[\varrho_0,\vc{u}_0,\vartheta_0]} \text{ for a.e. } x \in \Om .\]
Then \begin{align}
\mathcal{V}_{t,x}= \delta_{[ \tilde{\varrho}(t,x), \tilde{\vc{u}}(t,x), \tilde{\vartheta}(t,x), \bb{D}_x \tilde{\vc{u}}(t,x), \nabla_x \tilde{\vartheta}(t,x)]}\text{ for a.e. } (t,x) \in Q_T.
\end{align}
\end{Th}
\begin{proof}
At first, using the structural assumption \eqref{k,mu,lamda}, we have
\begin{align*}
& \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu(\varrho,\vartheta) \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; (\mu(\varrho,\vartheta)- \mu(\tilde{\varrho},\tilde{\vartheta}) )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
&= \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu_0 \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
& + \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu_1 \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; \mu_1(\vartheta -\tilde{\vartheta} )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\,\\
&= \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu_0 \frac{\tilde{\vartheta}}{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \frac{\vartheta}{\tilde{\vartheta}} \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu_1 \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&- \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; \mu_1(\vartheta -\tilde{\vartheta} )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\, \\
&+ \int_0^ \tau \int_{ \Om} \mu_1 \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; \frac{1}{\tilde{\vartheta}}\mu_1(\vartheta -\tilde{\vartheta} )^2 \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right \rangle\;\!\text{d}x\;\text{d}t\,.
\end{align*}
Using Young's inequality, we obtain
\begin{align*}
& \int_0^ \tau \int_{ \Om} \bb{D}_0(\nabla_x \tilde{\vc{u}}) \colon \left\langle \mathcal{V}_{t,x}; \mu_1(\vartheta -\tilde{\vartheta} )\left( \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right) \right \rangle\;\!\text{d}x\;\text{d}t\, \\
&\leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \mu_1 \tilde{\vartheta} \left\vert \bb{D}_0 (\bb{D}_{\vc{u}}) - \bb{D}_0 (\nabla_x \tilde{\vc{u}} ) \right\vert^2 \right \rangle \;\!\text{d}x\;\text{d}t\, \\
&+C(\nabla_x \tilde{\vc{u}} ) \frac{1}{4\epsilon} \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ;\left[ \left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right]_{\text{ess}} + \left[ 1+ \varrho e_R(\varrho,\vartheta ) \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\, .
\end{align*}
Similarly, we derive the following inequalities:
\begin{align*}
&\left\vert \int_0^ \tau \int_{ \Om} \kappa(\tilde{\vartheta}) \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta-\tilde{\vartheta}) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\, \right \vert \\
&\leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, + C(\epsilon, \tilde{\vartheta} ) \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left\vert \vartheta-\tilde{\vartheta} \right \vert^2 \right \rangle \;\!\text{d}x \;\text{d}t\,
\end{align*}
and
\begin{align*}
&\left\vert \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\kappa(\varrho,\vartheta)-\kappa(\tilde{\varrho},\tilde{\vartheta})) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\,\right \vert\\
&=\left\vert \int_0^ \tau \int_{ \Om} {\nabla_x \tilde{\vartheta}} \cdot \left\langle \mathcal{V}_{t,x}; (\vartheta^\beta-\tilde{\vartheta}^\beta) \left( \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}- \frac{\vc{D}_{\vartheta}}{\vartheta} \right) \right \rangle \;\!\text{d}x \;\text{d}t\,\right \vert\\
&\leq \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x} ; (1+\vartheta^\beta) \left\vert \frac{\vc{D}_\vartheta}{\vartheta} - \frac{\nabla_x \tilde{\vartheta}}{\tilde{\vartheta}}\right\vert^2 \right \rangle \!\!\;\!\text{d}x \;\text{d}t\, \\
&+ C(\epsilon, \tilde{\vartheta} ) \bigg( \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; [ |\vartheta-\tilde{\vartheta} |^2]_{\text{ess}} \right \rangle \;\!\text{d}x \;\text{d}t\,
+ \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x};[1+\vartheta^\beta]_{\text{res}} \right \rangle \;\!\text{d}x \;\text{d}t\, \bigg).
\end{align*}
Since $ \beta \leq 2 $, the second term s controlled by radiation pressure $ p_R $.
Therefore, our choice of $ p $ yields
\begin{align*}
&\int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[\vartheta^2 +\vert p(\varrho,\vartheta) \vert + \vert \vc{u}-\tilde{\vc{u}} \vert + \varrho \vert s(\varrho, \vartheta) \vert\; \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\,\\
&\leq C(\epsilon) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \vert \vc{u}-\tilde{\vc{u}} \vert ^2 \right \rangle \;\!\text{d}x \;\text{d}t\, \\
&+ \int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[ \varrho \vert s_M(\varrho, \vartheta) \vert^2\; + \varrho s_R \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\,.
\end{align*}
Furthermore, for any $ \epsilon>0 $, we get the following inequality
\begin{align}\label{E1}
\varrho s_R \vert \vc{u} \vert \leq C(\tilde{\vc{u}}) (\vartheta \vert \vc{u}-\tilde{\vc{u}} \vert+ \varrho s_R) \leq C(\tilde{\vc{u}}) \left( \frac{1}{4\epsilon} \vartheta^2 + \epsilon \vert \vc{u}-\tilde{\vc{u}} \vert^2 + \varrho s_R \right).
\end{align}
On the other hand, the structural assumption on $ p_M $ and $ S_M $ gives us
\begin{align}\label{E2}
\varrho \vert s_M(\varrho, \vartheta) \vert^2 \leq C(1+ \varrho + \varrho e_M).
\end{align}
Finally, we conclude that
\begin{align*}
&\int_{0}^\tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \left[\vartheta +\vert p(\varrho,\vartheta) \vert + \vert \vc{u}-\tilde{\vc{u}} \vert + \varrho \vert s(\varrho, \vartheta) \vert\; \vert \vc{u} \vert \right]_{\text{res}} \right \rangle \;\!\text{d}x\;\text{d}t\,\\
&\leq C(\epsilon) \int_{0}^\tau \mathcal{E}_{mv}(t) \;\text{d}t\, + 2 \epsilon \int_0^ \tau \int_{ \Om} \left\langle \mathcal{V}_{t,x}; \vert \vc{u}-\tilde{\vc{u}} \vert ^2 \right \rangle \;\!\text{d}x \;\text{d}t\, .
\end{align*}
Now proper choice of $ \epsilon $, generalized Korn-Poincar\' e inequality and Gr\" onwall's argument gives us the desired result.
\end{proof}
\begin{Rem}
We prove the weak strong uniqueness result for $ \kappa(\vartheta)= \kappa_1+\kappa_2\vartheta^\beta $, $ 0\leq \beta \leq 2 $. This method does not allow us to prove the weak (measure valued)-strong uniqueness when $ \beta>2 $.
\end{Rem}
\section{Comments on measure valued solution}\label{s5}
We try to obtain a priori bounds for the system in which the pressure law follows \eqref{Pr1} with \eqref{pm,em} and \eqref{rad_p1} and the transport coefficients $ \mu, \;\lambda $ and $ \kappa $ follow
\begin{align}\label{k,mu.lam2}
\begin{split}
0 < \underline{\mu} \left(1 + \vartheta \right) &\leq \mu(\vartheta) \leq \overline{\mu} \left( 1 + \vartheta \right),\
|\mu'(\vartheta)| \leq c \ \mbox{for all}\ \vartheta \geq 0,\\
0 &\leq \eta(\vartheta) \leq \overline{\eta} \left( 1 + \vartheta \right), \\
0 < \underline{\kappa} \left(1 + \vartheta^\beta \right) &\leq \kappa(\vartheta) \leq \overline{\kappa} \left( 1 + \vartheta^\beta \right),\ \beta \geq 2.
\end{split}
\end{align}
We also invoke the third law of thermodynamics \eqref{3rd law}.
\subsection{A priori estimate}\label{appest}
Let us first recall the ballistic energy inequality
\begin{align}\label{ap1}
\begin{split}
\frac{\text{d} }{\;\text{d}t\,} &\intO{ \left[ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right] } + \intO{ \frac{\tilde{\vartheta}}{\vartheta} \left( \mathbb{S} : \bb{D}_x \vc{u} - \frac{\vc{q} \cdot \nabla \vartheta }{\vartheta} \right) }\\
&\leq \intO{ \varrho \vc{u} \cdot \vc{g} }- \intO{ \left[ \varrho s \left( \partial_t \tilde{\vartheta} + \vc{u} \cdot \nabla \tilde{\vartheta} \right) + \frac{\vc{q}}{\vartheta} \cdot \nabla \tilde{\vartheta} \right] },
\end{split}
\end{align}
where $ \tilde{\vartheta} $ is a smooth function such that \[ \tilde{\vartheta} >0 \text{ in } (0,T) \times \Om \text{ with } \tilde{\vartheta} = \vartheta_B \text{ on }\partial \Om. \]
In order to control the last integral in \eqref{ap1}, we proceed analogously as in \cite[Section 4.1] {CF2021}. Hence, we consider the extension $\tilde{\vartheta}$ to be the unique solution of the Laplace equation
\begin{equation} \nonumber
\Delta \tilde{\vartheta} (\tau, \cdot) = 0 \ \mbox{in}\ \Omega,\ \tilde{\vartheta}(\tau, \cdot)|_{\partial \Omega} = \vartheta_B \ \mbox{for any}\ \tau \in [0,T].
\end{equation}
The maximum principle for the Laplace equation gives
\[
\min_{[0,T] \times \partial \Omega} \vartheta_B \leq \tilde{\vartheta} (t,x) \leq \max_{[0,T] \times \partial \Omega} \vartheta_B \ \mbox{for any}\ (t,x) \in (0,T) \times \Omega.
\]
Let us denote this particular extension by $\widehat{\vartheta_B}$.
At first we note that
\[
- \intO{ \frac{\vc{q}}{\vartheta} \cdot \nabla_x\widehat{\vartheta_B }} = \intO{ \frac{\kappa(\vartheta)}{\vartheta} \nabla_x \vartheta \cdot \nabla_x \widehat{\vartheta_B }} =
\intO{ \nabla_x K(\vartheta) \cdot \nabla_x \widehat{\vartheta_B }} = \int_{\partial \Omega} K(\vartheta_B) \nabla_x \vartheta_B \cdot \vc{n},
\]
where $K'(\vartheta) = \frac{\kappa (\vartheta)}{\vartheta}$.
Next, as $\widehat{\vartheta_B}$ is continuously differentiable in time, we obtain
\[
- \intO{ \varrho s \partial_t \vartheta_B } \leq \left[ (1 + \intO{ \left( \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right) } \right].
\]
Therefore, for the term
\begin{equation} \nonumber
- \intO{ \varrho s \vc{u} \cdot \nabla_x \widehat{\vartheta B} } = - \intO{ \varrho \mathcal{S} \left( \frac{\varrho}{\vartheta^{\frac{3}{2} } } \right) \vc{u} \cdot \nabla_x \widehat{\vartheta_B }} - {2a} \intO{ \vartheta \vc{u} \cdot \nabla_x\widehat{\vartheta_B }},
\end{equation}
using the inequality in \cite[Section 4, formula (4.6)]{FN2012}, we have
\[ \varrho \left\vert \mathcal{S} \left( \frac{\varrho}{\vartheta^{\frac{3}{2} } } \right) \right \vert \leq \left( \varrho + \varrho |\log(\varrho)| + \varrho [ \log(\vartheta)]^+ \right) \ \mbox{for any} \ \varrho \geq 0,\ \vartheta \geq 0.\]
Consequently, it yields
\begin{align*}
\left| \intO{ \varrho \mathcal{S} \left( \frac{\varrho}{\vartheta^{\frac{3}{2} } } \right) \vc{u} \cdot \nabla_x \widehat{\vartheta_B} } \right| &\leq
\left( \intO{ \varrho |\vc{u}|^2 } + \intO{\varrho \mathcal{S}^2\left( \frac{\varrho}{\vartheta^{\frac{3}{2} } } \right) } \right) \\ &\leq
\left( 1 + \intO{ \varrho |\vc{u}|^2 } + \intO{ \varrho^{\frac{5}{3} } } +\intO{\varrho \vartheta }+ \intO{\vartheta^2} \right) \\
&\leq \left( 1 + \intO{ \left( \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right) } \right).
\end{align*}
For the other term, we notice that
\begin{align*}
\left| \intO{ \vartheta \vc{u} \cdot \nabla \widehat{\vartheta_B} } \right| &\leq \epsilon \| \vc{u} \|^2_{W^{1,2}(\Omega; R^d)} + c(\epsilon) \int_{ \Om} \vartheta^2 \;\!\text{d}x .
\end{align*}
From the definition of the radiation pressure, we conclude
\[ \int_{ \Om} \vartheta^2 \;\!\text{d}x \leq C(\epsilon) \int_{ \Om}(\varrho e - \tilde{\vartheta} \varrho s) \;\!\text{d}x + c(\vartheta_B) + \epsilon \| \vc{u} \|^2_{W^{1,2}(\Omega; R^d)} .\]
Therefore, chossing $ \epsilon $ properly, we are able to get
\begin{align*}
&\intO{ \left( \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right) (\tau, \cdot) } \\
&+
\inf_{[0,T] \times \partial \Omega}\{ \vartheta_B \} \int_0^\tau \intO{ \left( \| \vc{u} \|^2_{W^{1,2}(\Omega; R^d)} + \frac{\kappa (\vartheta) |\nabla_x \vartheta|^2 }{\vartheta^2} \right) }\; \;\text{d}t\, \\
&\leq
\intO{ \left( \frac{1}{2} \varrho_0 |\vc{u}_0 |^2 + \varrho_0 e(\varrho_0, \vartheta_0) - \tilde{\vartheta} \varrho_0 s (\varrho_0, \vartheta_0) \right) } \\
& + c( \vartheta_B ) \left[ 1 +
\int_0^\tau \intO{ \left( \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s \right) }\; \;\text{d}t\, \right].
\end{align*}
Now we can use Gr\"onwall's argument to deduce a priori estimate for $ \frac{1}{2} \varrho |\vc{u} |^2 + \varrho e - \tilde{\vartheta} \varrho s $ .
Moreover for this particular pressure law \eqref{Pr1} with \eqref{pm,em} and \eqref{rad_p1} along with the third law of thermodynamics \eqref{3rd law}, following Feireisl and B\v rezina\cite{BF2018a}, we are able to conclude that
\begin{align*}
&\text{ess} \sup_{(0,T)} \Vert \varrho s \Vert_{L^q(\Om)} + \text{ess} \sup_{(0,T)} \Vert \varrho s \vc{u}\Vert_{L^q(\Om)} \leq C
\end{align*}
for some $ q >1$. Although they consider only for the molecular pressure, the radiation pressure case is quite straight forward. Therefore we obtain the following a priori estimate :
\begin{align*}
&\text{ess} \sup_{(0,T)} \int_{ \Om} \left[\varrho + \varrho \vert \vc{u} \vert^2 + \varrho e(\varrho, \vartheta) + |\varrho \vert s(\varrho,\vartheta)|^q \vert \right] \;\!\text{d}x \\
&+\int_{0}^T \int_{ \Om} \left[ \underline{\mu}\left( 1+\frac{1}{ \vartheta }\right) \left\vert \nabla_x \vc{u} + \nabla_x^T \vc{u} -\frac{2}{d} \text{div}_x \vc{u} \right\vert^2 + \frac{\lambda( \vartheta)}{2 \vartheta} \vert \text{div}_x \vc{u} \vert ^2\right] \;\!\text{d}x \;\text{d}t\,\\
&+ \int_{0}^T \int_{ \Om} \underline{\kappa}\left(\frac{1}{\vartheta^2}+ \vartheta^{\beta-2}\right) \vert \nabla_x \vartheta \vert^2 \;\!\text{d}x \;\text{d}t\, \leq C( \vartheta_B),
\end{align*}
for some $ q>1 $.
\subsection{Comments on compatibility conditions}
Now suppose we assume a young measure $ \mathcal{V} $ is generated by a family of sequences $ (\varrho_\epsilon, \vc{u}_\epsilon, \bb{D}_x \vc{u}_\epsilon, \vartheta_\epsilon , \nabla_x \vartheta_\epsilon )_{\{\epsilon>0\}}$ that satisfies the bound
\begin{align}\label{uni-est-ep}
\begin{split}
&\text{ess} \sup_{(0,T)} \int_{ \Om} \left[\varrho_\epsilon + \varrho_\epsilon \vert \vc{u}_\epsilon \vert^2 + \varrho_\epsilon e(\varrho_\epsilon, \vartheta_\epsilon) + |\varrho_\epsilon \vert s(\varrho_\epsilon,\vartheta_\epsilon)|^q \vert \right] \;\!\text{d}x \\
&+\int_{0}^T \int_{ \Om} \left[ \frac{\mu(\varrho_\epsilon, \vartheta_\epsilon)}{2 \vartheta_\epsilon } \left\vert \nabla_x \vc{u}_\epsilon + \nabla_x^T \vc{u}_\epsilon -\frac{2}{d} \text{div}_x \vc{u}_\epsilon \right\vert^2 + \frac{\mu(\varrho_\epsilon, \vartheta_\epsilon)}{2 \vartheta_\epsilon } \vert \text{div}_x \vc{u}_\epsilon \vert ^2\right] \;\!\text{d}x \;\text{d}t\,\\
&+ \int_{0}^T \int_{ \Om} \kappa(\varrho_\epsilon,\vartheta_\epsilon) \vert \log \vartheta_\epsilon \vert^2 \;\!\text{d}x \;\text{d}t\, \leq C,
\end{split}
\end{align}
uniformly with respect to $ \epsilon $ and $ \mu, \lambda \text{ and } \kappa $ follows \eqref{k,mu.lam2} and $ q>1 $. Clearly the bound \eqref{uni-est-ep} is motivated from the discussion we have in Section \ref{appest}.
Then following \cite[Section 2.4]{BFN2020}, we are able to deduce the velocity and temperature compatibility.
Also, Generalized Korn-Poincar\'e inequality can be derived in the similar way.
\subsubsection*{Defect measures and its compatibility}
The bound \eqref{uni-est-ep}, gives an uniform estimate of $ \left( \frac{1}{2}\varrho_\epsilon \vert \mathbf{u}_\epsilon \vert^2 + \varrho e(\varrho_\epsilon,\vartheta_\epsilon) - \widetilde{\Theta} \varrho s(\varrho_\epsilon,\vartheta_\epsilon)\right)_{\epsilon>0} $ with respect to $ \epsilon $ in $ L^\infty(0,T;L^1(\Om)) $.
We denote the weak* limit of in $ L^\infty_{\text{weak-(*)}}(0,T;\mathcal{M}(\Om)) $ by $$ \overline{\left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta} \varrho s(\varrho,\vartheta)\right)}. $$.Next, we define the corresponding \emph{defect measure} as
\begin{align*}
\mathcal{D}_{\widetilde{\Theta}} = \overline{\left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta} \varrho s(\varrho,\vartheta)\right)}- \left \langle\mathcal{V}_{t,x} ; \left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) - \widetilde{\Theta} \varrho s(\varrho,\vartheta)\right) \right \rangle.
\end{align*}
Now from the estimate \eqref{uni-est-ep} for entropy, we are able to conclude
\[ \left\langle \mathcal{V}_{t,x}; \varrho s(\varrho,\vartheta) \right\rangle= \overline{ \varrho s(\varrho,\vartheta) } \text{for a.e.} (t,x) \in (0,T)\times \Om \]
where $ \overline{ \varrho s(\varrho,\vartheta) } $ is a weak limit of $ \varrho_\epsilon s_\epsilon $.
This implies $ \mathcal{D}_{\widetilde{\Theta}} $ is independent of $ \widetilde{\Theta} $ and eventually we have
\begin{align*}
\mathcal{D} = \overline{\left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) \right)}- \left \langle\mathcal{V}_{t,x} ; \left( \frac{1}{2}\varrho \vert \mathbf{u} \vert^2 + \varrho e(\varrho,\vartheta) \right) \right \rangle.
\end{align*}
From the comparison lemma for defect measures in Feireisl et. al \cite[Lemma 2.1]{FPAW2016}, we can claim the compatibility of defects \eqref{def-m-comp}, since $ r^M $ is only related to the term $ \varrho \vc{u} \otimes \vc{u} + p(\varrho,\vartheta) \bb{I} $. Therefore, this particular pressure law \eqref{Pr1} allows us to have weak convergence for the terms $ \varrho s $ and $ \varrho s \vc{u} $ in some $ L^q(0,T;L^q(\Om)) $, for some $ q>1 $. Hence, we avoid defect measure for entropy inequality and eventually for ballistic energy inequality.
\section*{Concluding remarks}
Here we briefly discuss the possible extensions and limitations of the problem:
\begin{itemize}
\item We have given no proof of the existence of a measure valued solution. One way to prove existence is to use \emph{consistent approximations}. By consistent approximation, we mean an approximated system of the original system with error terms that vanish in the limiting case. The approximate solutions with some uniform bounds of state variables generate a Young measure, which is eventually a measure valued solution of the system. For a detailed discussion, see \cite[Chapter 5, Section 5.3.1]{FLMS2021book}.
\item Another way to construct measure valued solution is to consider it as a weak limit of weak solutions. From \cite{CF2021}, we know that there exists a weak solution for
\begin{align*}
0 < \underline{\kappa} \left(1 + \vartheta^\beta \right) &\leq \kappa(\vartheta) \leq \overline{\kappa} \left( 1 + \vartheta^\beta \right),\ \beta > 6,\\
& p_R = a \vartheta^4,
\end{align*}
Therefore, at least for system \eqref{NSF-c}-\eqref{temp_bdry} with \eqref{k,mu,lamda} and pressure following \eqref{pm,em}, \eqref{rad_p1}, we expect to obtain a measure valued solution as a suitable limit of this weak solutions. If we consider a modified radiation pressure and the heat conductivity coefficient as
\begin{align*}
p_{\epsilon,R} = a \vartheta^2 + \epsilon \vartheta^4
\end{align*}
and
\begin{align*}
\kappa_\epsilon(\vartheta)=1+ \vartheta^\beta + \epsilon \vartheta^\gamma, \text{ with }\beta\leq 2 \text{ and }\gamma> 6.
\end{align*}
Then one can expect to generate a measure value solution by sequence of weak solutions $ \{\varrho_\epsilon,\vc{u}_\epsilon, \vartheta_\epsilon \}_{\{\epsilon>0\}} $.
\item Navier-slip boundary condition for velocity: Instead of boundary condition \eqref{velocity_bdry}, we consider the Navier-Slip boundary condition
\begin{align*}
\vc{u} \cdot \mathbf{n}|_{\partial \Om}=[\mathbb{S}(\nabla_{x}\vc{u})\cdot \mathbf{n}]_{\text{tan}}|_{\partial\Om}=0.
\end{align*}
We can provide a similar definition by modifying the Korn-Poincare inequality suitably, see \cite[Section 1.3.6]{C2021}. A similar weak(measure valued)-strong uniqueness result is expected to be true.
\item Limitation with radiation pressure following Stefan-Boltzman law: Unfortunately, we are not able to prove our results for the radiation pressure following Stefan-Boltzman law $ p_R(\vartheta) = a \vartheta^4 $ with $ a>0 $. The main difficulty is to deal the term $\left[ \varrho s_R \vert \vc{u} \vert \right]_{\text{res}} $. Following Feireisl \cite{F2012}, we notice that, the estimation of the term needs certain Sobolev embedding which is missing in our definition. Our choice of $ p_R $ in \eqref{rad_p1} is motivated from the models of Neutron star, see Lattimer et al. \cite{LRPP1994}.
\end{itemize}
\centerline{ \bf Acknowledgement}
\vspace{2mm}
The work of N. Chaudhuri was partly supported by EPSRC Early Career Fellowship no. EP /V000586/1. The author would like to thank Prof. E. Feireisl and Dr. E. Zatorska for their valuable comments.
\def\cprime{$'$} \def\ocirc#1{\ifmmode\setbox0=\hbox{$#1$}\dimen0=\ht0
\advance\dimen0 by1pt\rlap{\hbox to\wd0{\hss\raise\dimen0
\hbox{\hskip.2em$\scriptscriptstyle\circ$}\hss}}#1\else {\accent"17 #1}\fi}
\bibliographystyle{abbrv}
|
\section{Introduction}
The study of quantum many-body systems often rely on approximate descriptions such as mean-field \cite{Bogoliubov1947,Bardeen1957, BruusFlensberg} and variational treatments \cite{Griffiths2005,Landau_QM}, or on extensive numerical analyses such as Quantum Monte Carlo \cite{Metropolis1949,Kalos1962,Hammond1994,Boninsegni2006,Kolorenc2011}. For this reason, instances in which the many-body dynamics can be solved exactly \cite{Katsura1974,Baxter1982,Andrei1995,Braak2011} are important to get precise insights into these systems, and may offer new ways to approximate related models . An important class of such systems are given by Bethe lattices [Fig. \ref{fig.bethe_lattices}], whose geometry is uniquely defined by the number of nearest neighbors. The name of these fascinating fractal structures are given in honor of the groundbreaking work of Hans Bethe on the Bethe ansatz \cite{Bethe1935}, being exact descriptions of e.g. ferro- and antiferromagnetism in these particular lattices \cite{Katsura1974}. Related studies have shown that the free quantum walk of a single particle in Bethe lattices may also be solved exactly \cite{Brinkman1970,Mahan2001,Eckstein2005,Economou2006,Nagy2017,Aryal2020}, which has been used to model amorphous solids \cite{Weaire1971,Weaire1971,Thorpe1971,Thorpe1973,Joannopoulos1974,Yndurain1975,Allan1980} and the hopping of ions in ice \cite{Chen1974}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig1.pdf}
\end{center}
\caption{\textbf{Single dopant in Bethe lattices}. A single hole (grey circle) is initialized at the center of an otherwise perfectly ordered antiferromagnetic Bethe lattice, showcased by coordination numbers $q = 1, 2, 3, 4$ nearest neighbors in panels (a) through (d). Here, red (blue) circles denote spin-$\uparrow$ (spin-$\downarrow$) fermions.}
\label{fig.bethe_lattices}
\end{figure}
In this article, we fuse these ideas to describe the motion of a single dopant in an antiferromagnetic Bethe lattice. In particular, we show that the non-equilibrium dynamics of a single hole hopping with amplitude $t$ in an antiferromagnetic environment with nearest neighbor spin-$z$ coupling $J_z$ can be solved exactly at zero temperature. Previously, an exact solution for the local hole Green's function, corresponding to the lowest order coefficient in the many-body wave function, was found \cite{Chernyshev1999,Wrzosek2021}. Here, we show that the full many-body wave function can be calculated, not only in the non-equilibrium case, but also for an important set of many-body eigenstates, i.e. the magnetic polaron ground state and the so-called string excitations \cite{Bulaevskii1968,Shraiman1988,Kane1989,Martinez1991,Liu1991,Reiter1994}. The appearance of magnetic polarons is a result of the inherent competition between the delocalization of the hole and the emergent magnetic frustrations. Such processes also give rise to induced interactions between two holes, which may provide a mechanism for high-temperature superconductivity \cite{Emery1987,Schrieffer1988,Dagotto1994}. Due to recent advances in quantum simulation using optical lattices \cite{Bakr2009,Esslinger2010,Sherson2010,Haller2015,Boll2016,Cheuk2016,Mazurenko2017,Hilker2017,Brown2017,Chiu2018,Brown2019,Koepsell2019,Chiu2019,Sanchez2020,Vijayan2020,Hartke2020,Sanchez2021,Ji2021,Gall2021,Yang2021}, systems in which these quasiparticles arise can now be realized and probed with unprecedented detail \cite{Koepsell2019,Chiu2019,Koepsell2020}. Consequently, magnetic polarons and the associated motion of holes in antiferromagnetic environments are also receiving renewed theoretical inquiries \cite{Carlstrom2016,Grusdt2018_2,Grusdt2018,Bohrdt2019,Bohrdt2020,Blomquist2020,Wang2021,Nielsen2021,Nielsen2022}. In particular, our present work is related to the so-called string theory of magnetic polarons \cite{Grusdt2018_2}, in which the hole is effectively described as moving in a Bethe lattice.
The exact solvability of the anisotropic $t$-$J_z$ model in Bethe lattices presented in this work comes about as a result of the fractal self-similarity of the Bethe lattices. In particular, every time the hole hops, the system in the forward direction of motion looks the same as at the previous site. This underlying structure means that the wave function, expressed as a superposition of states with an increasing number of spin excitations, also attains a self-similar form, in which the coefficients of the wave function are related by quite simple recursion relations. This structure is very closely related to the approximate magnetic polarons states \cite{Reiter1994,Ramsak1998,Nielsen2021} and non-equilibrium wave function \cite{Nielsen2022} for the full $t$-$J$ model on square lattices. In fact, the present solution can also be understood in the context of the retraceable path approximation \cite{Brinkman1970}, which becomes exact in Bethe lattices.
The article is organized as follows. Section \ref{sec.bethe_ising_antiferromagnet} describes the structure of the antiferromagnetic ground state, and the hopping Hamiltonian in the presence of holes. Section \ref{sec.holstein_primakoff} introduces the Holstein-Primakoff transformation, making it possible to give a concise description of the many-body states. Section \ref{sec.exact_solution} gives the derivation of the exact solution for the many-body wave function, before Sec. \ref{sec.results} characterizes the dynamics as a function of inverse interaction strength $J_z / t$ and coordination number $q$.
\section{Ising antiferromagnets in Bethe lattices} \label{sec.bethe_ising_antiferromagnet}
We consider spin $1/2$ fermions hopping in Bethe lattice structures with $q$ nearest neighbors, as exemplified in Fig. \ref{fig.bethe_lattices}. We assume that the nearest neighbor spin-spin interactions are of the Ising type,
\begin{equation}
\hat{H}_J = J_z \sum_{\braket{{\bf i}, {\bf j}}} \left[\hat{S}^{(z)}_{\bf i}\hat{S}^{(z)}_{\bf j} - \frac{\hat{n}_{\bf i} \hat{n}_{\bf j}}{4} \right],
\label{eq.H_J}
\end{equation}
and antiferromagnetic ($J_z > 0$). The Schwinger-fermion representation of spin $1/2$ as usual reads
\begin{equation}
\mathbf{S}_{\bf j}=\frac{1}{2}\sum_{\sigma,\sigma'}\hat{c}_{{\bf j},\sigma}^\dagger\boldsymbol{\sigma}_{\sigma\sigma'}\hat{c}_{{\bf j},\sigma'}
\end{equation}
with $\boldsymbol{\sigma}=(\sigma_x,\sigma_y,\sigma_z)$ a vector of the Pauli matrices. The antiferromagnetic ground state is, thus, achieved by having every neighboring spin pointing in opposite directions. Choosing a specific lattice site as the central reference point, we then speak of the depth $d$ of a given site, as the number of hops it takes to go to that site. Correspondingly, the total depth $d_{\rm tot}$ of the lattice is the maximum number of jumps that can be made from the central site. In Fig. \ref{fig.bethe_lattices}(c) and (d), the cases of $q = 3$ and $q = 4$ respectively, the total depth of the shown lattices is $d_{\rm tot} = 5$. The total number of sites in a Bethe lattice with $q$ nearest neighbors and total depth $d_{\rm tot}$ is
\begin{equation}
N(q,d_{\rm tot}) = 1 + q \sum_{j = 0}^{d_{\rm tot} - 1} (q - 1)^j = 1 + \frac{(q - 1)^{d_{\rm tot}} - 1}{q - 2}.
\label{eq.number_of_sites}
\end{equation}
In the case of $q = 2$, the Bethe lattice collapses to a one-dimensional chain [Fig. \ref{fig.bethe_lattices}(b)], and the expression $N = 1 + d_{\rm tot}$ can be found from Eq. \eqref{eq.number_of_sites} by applying l'Hospital's rule to the limit $q \to 2$. Generally, the number of nearest neighbor links is simply $N - 1$, and the antiferromagnetic energy contribution from each of these is $-J_z/2$. Therefore, the ground state energy for an Ising antiferromagnet in a Bethe lattice is simply
\begin{equation}
E_0(q,d_{\rm tot}) = - \left[N(q,d_{\rm tot}) - 1\right]\frac{J_z}{2}.
\label{eq.ground_state_energy_antiferromagnet}
\end{equation}
Since the number of sites at a depth $d + 1$ is $q - 1$ times larger than the number of sites at depth $d$, and since the spins at $d+1$ point opposite to the spins at $d$, the lattice has a nonzero total spin
\begin{align}
S^{(z)}_{\rm tot}(q, d_{\rm tot}) &= -\frac{1}{2}\left[1 - q\sum_{j = 0}^{d_{\rm tot} - 1} (1 - q)^j \right] \nonumber \\
&= -\frac{1}{2}(1-q)^{d_{\rm tot}},
\label{eq.S_z_tot_ground_state}
\end{align}
assuming that the central site has $S^{(z)} = -1 / 2$. This characterizes the underlying antiferromagnetic ground state. Below half-filling, we assume that the fermionic particles can hop between nearest neighbor sites with amplitude $t$
\begin{equation}
\hat{H}_t = - t\sum_{\braket{{\bf i}, {\bf j}},\sigma} \left[\tilde{c}^\dagger_{{\bf i}, \sigma}\tilde{c}_{{\bf j}, \sigma} + {\rm H.c.} \right],
\label{eq.H_t}
\end{equation}
where $\tilde{c}^\dagger_{{\bf j},\sigma} = \hat{c}^\dagger_{{\bf j},\sigma}(1 - n_{\bf j})$, and the factor $1 - n_{\bf j}$ restrains the Hilbert space to states with maximally one fermion per lattice site.
\section{The Holstein-Primakoff transformation} \label{sec.holstein_primakoff}
The Holstein-Primakoff transformation is performed to give a more efficient description of the system just below half-filling, in terms of bosonic spin excitation operators $\hat{s}_{\bf i}$, and fermionic hole operators $\hat{h}_{\bf i}$. To perform the transformation, it is first useful to define two sublattices corresponding to the sites that in the absence of holes host spins pointing up and down respectively. More precisely, fermions on every site at an odd-numbered depth, $d = 1, 3, 5, \dots$, initially has spin-$\uparrow$, and is defined to lie on sublattice A. Similarly, every site at an even-numbered depth, $d = 0, 2, 4, \dots$, belong to sublattice B, having spin-$\downarrow$ fermions. \\
We may, then, define the Holstein-Primakoff transformation according to
\begin{align}
{\rm {\bf A}}: \hat{S}^{-}_{\bf i} &= \hat{s}_{\bf i}^\dagger F(\hat{h}_{\bf i}, \hat{s}_{\bf i}), \; \tilde{c}_{{\bf i},\downarrow}=\hat{h}^\dagger_{\bf i}\hat{S}_{\bf i}^{+}, \nonumber \\
\tilde{c}_{{\bf i},\uparrow} &=\hat h^\dagger_{\bf i} F(\hat{h}_{\bf i}, \hat{s}_{\bf i}), \; \hat{S}_{\bf i}^z= \left[\frac{1}{2} - \hat{s}^\dagger_{\bf i}\hat{s}_{\bf i}\right] [1-\hat{h}_{\bf i}^\dagger \hat{h}_{\bf i}]. \nonumber \\
{\rm {\bf B}}: \hat{S}^{+}_{\bf j} &= \hat{s}_{\bf j}^\dagger F(\hat{h}_{\bf j}, \hat{s}_{\bf j}), \; \tilde{c}_{{\bf j},\uparrow}=\hat{h}^\dagger_{\bf j}\hat{S}_{\bf j}^{-}, \nonumber \\
\tilde{c}_{{\bf j},\downarrow} &= \hat h^\dagger_{\bf j} F(\hat{h}_{\bf j}, \hat{s}_{\bf j}), \; \hat{S}_{\bf j}^z= \left[\hat{s}^\dagger_{\bf j}\hat{s}_{\bf j} - \frac{1}{2}\right] [1-\hat{h}_{\bf j}^\dagger \hat{h}_{\bf j}],
\end{align}
with $F(\hat{s},\hat{h}) = \sqrt{1 - \hat{s}^\dagger \hat{s} - \hat{h}^\dagger \hat{h}}$. Consequently, the spin-spin interaction is rewritten as
\begin{align}
\hat{H}_J = -J_z \sum_{\braket{{\bf i}, {\bf j}}} & [1-\hat{h}_{\bf i}^\dagger \hat{h}_{\bf i}] \left[\left(\frac{1}{2} - \hat{s}^\dagger_{\bf i}\hat{s}_{\bf i}\right)\left(\frac{1}{2} - \hat{s}^\dagger_{\bf j}\hat{s}_{\bf j}\right) + \frac{1}{4} \right] \nonumber \\
\cdot &[1-\hat{h}_{\bf j}^\dagger \hat{h}_{\bf j}].
\label{eq.H_J_holstein_primakoff}
\end{align}
This expression fully accounts for the presence of holes, in which case the nearest neighbor spin coupling is naturally $0$. Similarly, the hopping Hamiltonian becomes
\begin{align}
\hat{H}_t = t \sum_{\braket{{\bf i}, {\bf j}}} \! F(\hat{s}_{\bf i})F(\hat{s}_{\bf j}) \left[ \hat{h}^\dagger_{\bf j}\hat{h}_{\bf i} \hat{s}_{\bf j} + \hat{h}^\dagger_{\bf i}\hat{h}_{\bf j} \hat{s}_{\bf i} \right] + {\rm H.c.}
\label{eq.H_t_holstein_primakoff}
\end{align}
Here, due to the fermionic statistics of the holes, we may use that $F(\hat{s}_{\bf i},\hat{h}_{\bf i}) \hat{h}_{\bf i} = F(\hat{s}_{\bf i},0) \hat{h}_{\bf i}$. Additionally, we simplify the notation by writing $F(\hat{s}) = F(\hat{s},0) = \sqrt{1 - \hat{s}^\dagger\hat{s}}$. Equation \eqref{eq.H_t_holstein_primakoff} shows that as the hole hops through the lattice, it may do so by absorbing or emitting spin excitations. The factors of $F(\hat{s}_{\bf i})F(\hat{s}_{\bf j})$ contrains this process, so that there is always either $0$ or $1$ spin excitation on each site, yielding the exact hardcore constraint.
\section{The exact many-body wave function} \label{sec.exact_solution}
We are now ready to efficiently formulate the exact non-equilibrium wave function. We will focus on Bethe lattices with more than 1 nearest neighbors, $q\geq 2$, as the case of $q = 1$ corresponds to a system of only two sites. We assume that the hole is initially located at the central site of the Bethe lattice. Note that in the thermodynamic limit, $d_{\rm tot} \to \infty$, any site in the lattice is equivalent. Therefore, this choice for the initial state is actually the general starting point for quenching the system from the antiferromagnet ground state into a state with a single localized hole. A lattice site at depth $d$ is written as ${\bf j}_d = 0, j_1, \dots, j_d$. Here, $j_1 = 1, 2, \dots, q$, and $j_l = 1, 2, \dots, q-1$ for $l\geq 2$ describe the sites at each depth, similar to Ref. \cite{Katsura1974}. In this manner, ${\bf j}_d$ gives the full path from the central site ${\bf j}_0 = 0$ to the specific site of interest at depth $d$. Two sites ${\bf j}_d$ and ${\bf l}_{d+1}$ at depth $d$ and $d+1$ are, thus, nearest neighbors only if ${\bf l}_{d+1} = {\bf j}_d, l_{d+1}$. This construction allows us to write the full non-equilibrium wave function at time $\tau$ as
\begin{align}
\ket{\Psi(\tau)} =\; & C^{(0)}(\tau) \cdot \hat{h}^{\dagger}_{0} \ket{{ \rm AF }} + C^{(1)}(\tau) \sum_{{\bf j}_1} \hat{h}^\dagger_{{\bf j}_1} \hat{s}^\dagger_{0} \ket{{ \rm AF }} \nonumber \\
& + C^{(2)}(\tau) \sum_{{\bf j}_2} \hat{h}^\dagger_{{\bf j}_2} \hat{s}^\dagger_{0}\hat{s}^\dagger_{{\bf j}_1} \ket{{ \rm AF }} + \dots \nonumber \\
=\; & C^{(0)}(\tau) \hat{h}^{\dagger}_{0} \ket{{ \rm AF }} + \sum_{d = 1}^{d_{\rm tot}} C^{(d)}(\tau) \sum_{{\bf j}_d} \hat{h}^\dagger_{{\bf j}_d} \prod_{l = 0}^{d-1} \hat{s}^\dagger_{{\bf j}_l} \ket{{ \rm AF }}. \!\!
\label{eq.exact_many_body_wave_function}
\end{align}
In this expression, the lattice site at depth $d$ is always constrained to be next to the one at depth $d - 1$: ${\bf j}_d = {\bf j}_{d-1}, j_d$.
Note that the coefficients of the wave function $C^{(d)}$ \emph{cannot} depend on the exact path to the lattice point, ${\bf j}_d = j_1, \dots, j_d$, only the depth $d$ of that point. The underlying reason is that the system is symmetric in all these paths, and that the initial wave function, $\ket{\Psi(\tau = 0)} = \hat{h}^{\dagger}_{0} \ket{{ \rm AF }}$, is as well. This gives a remarkable simplification of the treatment, as it describes the dynamics in a system whose size grows exponentially with the depth [see Eq. \ref{eq.number_of_sites}], in terms of a linear number of coefficients $C^{(0)}, \dots, C^{(d_{\rm tot})}$. \\
\subsection{General time-dependent solution} \label{subsec.general_solution}
To progress on solving for the coefficients of the non-equilbrium wave function, we introduce the retarded and advanced wave functions
\begin{align}
\ket{\Psi(\tau)} &= \ket{\Psi^{\rm R} (\tau)} + \ket{\Psi^{\rm A}(\tau)} \nonumber \\
&= {\rm e}^{-\eta|\tau|}\left[\theta(\tau)\ket{\Psi(\tau)} + \theta(-\tau)\ket{\Psi(\tau)}\right],
\label{eq.retarded_and_advanced_state}
\end{align}
allowing us to regularize the Fourier transformation to frequency space with the positive infinitesimal $\eta > 0$, and the Heaviside step function $\theta(\tau)$. Consequently, we may express the Schr{\"o}dinger equation, $i\partial_{\tau} \ket{\Psi(\tau)} = \hat{H}\ket{\Psi(\tau)}$, as
\begin{align}
(\omega + i\eta) \ket{\Psi^{\rm R}(\omega)} &= +i\ket{\Psi(\tau = 0)} + \hat{H} \ket{\Psi^{\rm R}(\omega)}, \nonumber \\
(\omega - i\eta) \ket{\Psi^{\rm A}(\omega)} &= -i\ket{\Psi(\tau = 0)} + \hat{H} \ket{\Psi^{\rm A}(\omega)}.
\label{eq.schrodinger_equation_frequency_space}
\end{align}
As these equations are complex conjugate of each other, it follows that $\ket{\Psi^{\rm A}(\omega)} = [\ket{\Psi^{\rm R}(\omega)}]^*$. We denote the coefficients of the retarded state $R^{(d)}(\omega)$. The dynamical coefficients in Eq. \eqref{eq.exact_many_body_wave_function} are then retrieved from $R^{(d)}(\omega)$ by a Fourier transformation
\begin{equation}
C^{(d)}(\tau) = \int \frac{{\rm d}\omega}{2\pi} {\rm e}^{-i(\omega + i\eta)\tau} \cdot 2{ \rm Re }[R^{(d)}(\omega)].
\label{eq.C_d_and_R_d}
\end{equation}
Using Eq. \eqref{eq.schrodinger_equation_frequency_space}, the equations of motion for the retarded state coefficients become
\begin{align}
\!\!\!\!\!\!\!\!(\omega + i\eta) R^{(0)}(\omega) &= i + E_J^{(0)} R^{(0)}(\omega) + qt R^{(1)}(\omega), \!\!\!\nonumber \\
\!\!\!\!\!\!\!\!(\omega + i\eta) R^{(d)}(\omega) &= E_J^{(d)} R^{(d)}(\omega) + t R^{(d-1)}(\omega) \!\!\!\nonumber \\
\!\!\!\!\!\!\!\!&\phantom{=} + (q-1) t R^{(d+1)}(\omega), \!\!\!\nonumber \\
\!\!\!\!\!\!\!\!(\omega + i\eta) R^{(d_{\rm tot})}(\omega) &= E_J^{(d_{\rm tot})} R^{(d_{\rm tot})}(\omega) + t R^{(d_{\rm tot}-1)}(\omega). \!\!\!
\label{eq.equations_of_motion}
\end{align}
The term $i$ in the equation for $R^{(0)}(\omega)$ comes directly from the term $i\ket{\Psi(\tau = 0)}$ in Eq. \eqref{eq.schrodinger_equation_frequency_space}. Furthermore, the magnetic energy cost $E_J^{(d)}$ comes from applying the spin-spin interaction Hamiltonian in Eq. \eqref{eq.H_J_holstein_primakoff} to the $d$th term in the wave function. Finally, the hopping Hamiltonian simply relates the coefficient at depth $d$ to the coefficients at depth $d - 1$ and $d + 1$. At $d = 0$, there are $q$ nearest neighbors at depth $d = 1$. All subsequent depths $d \leq d_{\rm tot} - 1$, however, only has $q-1$ nearest neighbors at depth $d+1$ and a single neighbor one depth up, $d - 1$. This leads to the terms $(q-1) \cdot t \cdot R^{(d + 1)}(\omega)$ and $1 \cdot t \cdot R^{(d-1)}(\omega)$ for $1 \leq d \leq d_{\rm tot}$. Finally, by assumption the bottom depth is $d_{\rm tot}$, and so $R^{(d)} = 0$ for $d > d_{\rm tot}$. \\
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig2.pdf}
\end{center}
\caption{\textbf{Broken antiferromagnetic spin bonds}. As the hole moves into the Bethe lattice [panels (a)-(d)], a number of antiferromagnetic spin bonds are broken along the way (red lines), illustrated here in the case of $q = 3$. (a) At $d = 0$, there are $q = 3$ broken bonds, giving a spin energy of $E_J^{(0)} = E_0 + qJ_z / 2$. (b) For $d = 1$, there are $q - 1 = 2$ further broken bonds, so $E_J^{(1)} = E_J^{(0)} + (q-1)J/2$. (c) and (d) For $d\geq 2$, the spin bonds along the path of the hole are repaired. As a result, there are $q-1$ new spin bonds broken and $1$ repaired, yielding $q - 2 = 1$ in the present case. Therefore, $E_J^{(d+1)} = E_J^{(d)} + (q-2)J_z/2$ for $d\geq 1$.}
\label{fig.bethe_lattices_broken_bonds}
\end{figure}
The value of $E_J^{(d)}$ may be understood recursively, starting from $d = 0$. Here, the hole is located at the central site, and there are $q$ broken antiferromagnetic spin bonds of strength $J_z / 2$, as indicated in Fig. \ref{fig.bethe_lattices_broken_bonds}(a). Note that Eq. \eqref{eq.H_J_holstein_primakoff} implies that a hole at a site has the same magnetic energy cost as a spin excitation at that site. Therefore, $E_J^{(0)} = E_0(q,d) + qJ_z / 2$, where $E_0(q,d)$ is the ground state energy of the Ising antiferromagnet [Eq. \eqref{eq.ground_state_energy_antiferromagnet}]. When the hole is at depth $d = 1$, as shown in Fig. \ref{fig.bethe_lattices_broken_bonds}(b), there are $q-1$ additional broken bonds. So $E_J^{(1)} = E_J^{(0)} + (q-1)J_z / 2$. At every subsequent depth $d\geq 2$ [Fig. \ref{fig.bethe_lattices_broken_bonds}(c)-(d)], the spin bonds along the path of the hole are repaired. Therefore, each hop is associated with $q-1$ new broken bonds and $1$ repaired bond. The magnetic energy cost for $d\geq 2$ is, thus, $(q-2) J_z / 2$ for every step in depth. In this way,
\begin{align}
E_J^{(d)} &= \left[(q-1) + (d-1) (q-2)\right]\frac{J_z}{2},
\label{eq.E_J}
\end{align}
for $1\leq d \leq d_{\rm tot} - 1$. Here, we let $E_J^{(d)} \to E_J^{(d)} - E_J^{(0)}$, measuring all energies with respect to $E_J^{(0)}$. Finally, at depth $d_{\rm tot}$ the absence of neighbors at $d_{\rm tot} + 1$ in principle results in a surface effect, in which there is no additional magnetic energy cost for the hole to hop to the surface, $E_J^{(d_{\rm tot})} = E_J^{(d_{\rm tot} - 1)}$. While this presumably has an effect on the dynamics for small lattices, we focus on the behavior for large total depths, $d_{\rm tot} \gg 1$, and ignore the surface effects. With the magnetic energy costs for the hole to hop to the $d$th depth in place [Eq. \eqref{eq.E_J}], we can now solve the equations of motion in Eq. \eqref{eq.equations_of_motion} iteratively. Specifically, the equation for $R^{(d_{\rm tot})}$ simply yields
\begin{align}
R^{(d_{\rm tot})}(\omega) &= \frac{t}{\omega + i\eta - E_J^{(d_{\rm tot})}} R^{(d_{\rm tot} - 1)}(\omega) \nonumber \\
&= t \cdot G^{(d_{\rm tot})}(\omega) \cdot R^{(d_{\rm tot} - 1)}(\omega).
\label{eq.R_d_tot_omega}
\end{align}
Continuing this process for $d_{\rm tot} \geq d \geq 1$, we get the recursion relation
\begin{align}
R^{(d)}(\omega) &= t \cdot G^{(d)}(\omega) \cdot R^{(d - 1)}(\omega), \nonumber \\
G^{(d)}(\omega) &= \frac{1}{\omega + i\eta - E_J^{(d)} - (q-1)t^2 \cdot G^{(d+1)}(\omega)},
\label{eq.recursion_relation}
\end{align}
defining $G^{(d_{\rm tot}+1)}(\omega) = 0$. We refer to $G^{(d)}(\omega)$ as the retarded Green's function at depth $d$. Finally, inserting the solution from $d = 1$ into the equation for $R^{(0)}$, we get $R^{(0)}(\omega) = iG^{(0)}(\omega)$, with
\begin{equation}
G^{(0)}(\omega) = \frac{1}{\omega + i\eta - q t^2 \cdot G^{(1)}(\omega)}.
\label{eq.G_0}
\end{equation}
Using the recursion relation in Eq. \eqref{eq.recursion_relation} in reverse, we obtain the general result for the coefficients of the retarded wave function in frequency space
\begin{align}
R^{(0)}(\omega) &= iG^{(0)}(\omega), \nonumber \\
R^{(d)}(\omega) &= iG^{(0)}(\omega) \cdot t^{d} \cdot \prod_{l = 1}^d G^{(l)}(\omega), \; 1 \leq d \leq d_{\rm tot}.
\label{eq.R_d_general}
\end{align}
We end this subsection by connecting the coefficients of the wave function to a specific set of many-body correlation functions. Specifically, we have that
\begin{align}
R^{(0)}(\tau) &= \theta(\tau) \bra{{ \rm AF }}\hat{h}_0 {\rm e}^{-iH\tau} \hat{h}_0^\dagger \ket{{ \rm AF }}, \nonumber \\
R^{(d)}(\tau) &= \theta(\tau) \bra{{ \rm AF }}\hat{h}_{{\bf j}_d} \prod_{l=0}^{d-1} \hat{s}_{{\bf j}_l} {\rm e}^{-iH\tau} \hat{h}_0^\dagger \ket{{ \rm AF }}.
\label{eq.correlation_functions}
\end{align}
This more clearly demonstrates that the $d$th coefficient of the wave function is the probability amplitude of the hole to be at depth $d$ with $d-1$ spin excitations after a time $\tau$.
\subsection{Thermodynamic limit} \label{subsec.thermodynamic_limit}
In the thermodynamic limit, $d_{\rm tot}\to\infty$, the recursion relation for the retarded Green's functions in Eq. \eqref{eq.recursion_relation} can be written as a self-consistency equation for a single Green's function $G_1(\omega)$. Specifically, we can write $G^{(d)}(\omega) = G_1(\omega - E_J^{(d)})$ for $d\geq 1$, where
\begin{equation}
G_1(\omega) = \frac{1}{\omega - (q-1)t^2\, G_1(\omega - (q-2)\frac{J_z}{2})}.\!
\label{eq.G_1_self_consistency}
\end{equation}
For brevity, in this subsection we let $\omega + i\eta\to\omega$. Similarly, we get the Green's function for the central site, $d = 0$, to be
\begin{equation}
G_0(\omega) = \frac{1}{\omega - qt^2 \, G_1(\omega - (q-1)\frac{J_z}{2})}.
\label{eq.G_0_self_consistency}
\end{equation}
Inspired by Ref. \cite{Chernyshev1999}, we anticipate that the $G_1$ Green's function can be expressed in terms of Bessel functions. In fact, using the ansatz $G_1(\omega) = -1/(\sqrt{q-1}t) \cdot Y(\omega) / Y(\omega + (q-2)J/2)$, the resulting recursive relation for the $Y$ functions coincide with that for the Bessel functions of the first kind. This results in
\begin{equation}
G_1(\omega) = -\frac{1}{\sqrt{q-1}t} \frac{J_{\Omega(\omega)}\left(\frac{4\sqrt{q-1}t}{(q-2)J_z}\right)}{J_{\Omega(\omega) - 1}\left(\frac{4\sqrt{q-1}t}{(q-2)J_z}\right)},
\label{eq.G_1_solution}
\end{equation}
for $q\geq 3$, with $\Omega(\omega) = -2\omega / (q-2)J_z$. This generalizes the result in Ref. \cite{Chernyshev1999}, where the $q = 4$ Bethe lattice is effectively used as an approximate description of a two-dimensional square lattice, and coincides with what is found in Ref. \cite{Wrzosek2021}. We can now use the simple fraction of Bessel functions in Eq. \eqref{eq.G_1_solution} together with $G^{(d)}(\omega) = G_1(q, \omega - E_J^{(d)})$ for $d\geq 1$, to find the explicit expression for the coefficients of the non-equilibrium many-body wave function in the thermodynamic limit
\begin{equation}
R^{(d)}(\omega) = \frac{iG_0(\omega)}{(-\sqrt{q-1})^d}\cdot \frac{J_{\Omega(\omega - E_J^{(d)})}\left(\frac{4\sqrt{q-1}t}{(q-2)J_z}\right)}{J_{\Omega(\omega - E_J^{(1)}) - 1}\left(\frac{4\sqrt{q-1}t}{(q-2)J_z}\right)},
\label{eq.R_d_thermodynamic_limit}
\end{equation}
where the magnetic energy cost for a hole at depth $d$, $E_J^{(d)}$, is given in Eq. \eqref{eq.E_J}. In the limit of infinitely strong interactions, $J_z / t \to 0^+$, Eq. \eqref{eq.G_1_self_consistency} leads to a second order equation for $G_1$, yielding
\begin{gather}
G_1(\omega) \!\overset{J \to 0^+}{=}\! \frac{\omega}{2(q-1)t^2} \left(1 - \sqrt{1 - \frac{4(q-1)t^2}{\omega^2}}\right).
\label{eq.G_1_J_0_limit}
\end{gather}
This results in the hole Green's function
\begin{gather}
G_0(\omega) \overset{J \to 0^+}{=} \! \frac{1}{\omega \left[1 - \frac{q}{2(q-1)}\left(1 - \sqrt{1-\frac{4(q-1)t^2}{\omega^2}}\right)\right]}.
\label{eq.J_0_limit}
\end{gather}
Eqs. \eqref{eq.G_1_J_0_limit}-\eqref{eq.J_0_limit} together with Eq. \eqref{eq.R_d_general} describes a free quantum walk of a single particle in Bethe lattices \cite{Eckstein2005}, as one might expect in this extreme limit, where spin interactions play no role.
\subsection{Exact ground state and string excitations} \label{subsec.many_body_eigenstates}
The present methodology allows us to extract a certain set of many-body eigenstates on top of the non-equilibrium wave function derived in Subsec. \ref{subsec.general_solution}. This constitutes the \emph{depth symmetric} states
\begin{equation}
\ket{\Psi_n} = C^{(0)}_n \hat{h}^{\dagger}_{0} \ket{{ \rm AF }} + \sum_{d = 1}^{d_{\rm tot}} C^{(d)}_n \sum_{{\bf j}_d} \hat{h}^\dagger_{{\bf j}_d} \prod_{l = 0}^{d-1} \hat{s}^\dagger_{{\bf j}_l} \ket{{ \rm AF }},
\label{eq.Psi_n}
\end{equation}
which are all $d_{\rm tot} + 1$ states for which the coefficients of the wave function only depend on the depth $d$ of the hole, and not the particular point ${\bf j}_d$ at which it is located. Put in another manner, they are all the states that can be reached dynamically by immersing a single hole at the center of the lattice, i.e. the ones for which $\braket{\Psi_n|\Psi(\tau)} \neq 0$. In the limit of weak interactions, $J_z \gg t$, the many-body ground state is just the lowest order term, $\hat{h}^{\dagger}_{0} \ket{{ \rm AF }}$, as the hopping of the hole is strongly suppressed. Conversely, in the limit of infinite interactions, $J_z / t \to 0^+$, the energy should go to that of a free particle, which in a Bethe lattice is $-2\sqrt{q-1}t$. This is indeed the case for our solution, shown explicitly in Subsec. \ref{subsec.strong_interaction_limit}. Therefore, we expect $\ket{\Psi_0}$ to be the true many-body ground state for any value of $J_z / t \geq 0$. This is in contrast to the case of regular lattices, in which this type of states, so-called string excitations, do not approach the correct limiting value. Instead, a ferromagnetic polaron is expected to emerge \cite{White2001}. \\
The static Schr{\"o}dinger equation,
\begin{equation}
\varepsilon_n \ket{\Psi_n} = \hat{H} \ket{\Psi_n},
\end{equation}
simply corresponds to removing the $i\ket{\Psi(\tau = 0)}$ term in Eq. \eqref{eq.schrodinger_equation_frequency_space}, and replacing the frequency $\omega + i\eta$ with the eigenstate energy $\varepsilon_n$. As a result, we simply have to remove the term $i$ in the equations of motion in Eq. \eqref{eq.equations_of_motion} for the non-equilibrium case, and the same recursive relation in Eq. \eqref{eq.recursion_relation} for the non-equilibrium coefficients, therefore, applies to these eigenstates. Explicitly,
\begin{equation}
C^{(d)}_n = t \cdot G^{(d)}(\varepsilon_n) \cdot C^{(d - 1)}_n,
\label{eq.recursion_relation_eigenstate}
\end{equation}
with the Green's functions $G^{(d)}$ evaluated at the quasiparticle pole $\varepsilon_n$. Inserting the result at depth $d = 1$ into the equation of motion at depth $d = 0$, we get
\begin{equation}
\varepsilon_n \cdot C^{(0)}_n = qt^2 \cdot G^{(1)}(\varepsilon_n) \cdot C^{(0)}_n \Rightarrow \varepsilon_n = \Sigma^{(0)}(\varepsilon_n).
\label{eq.equation_for_epsilon_n}
\end{equation}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig3.pdf}
\end{center}
\caption{\textbf{Spectral functions}, $A(\omega)$, for 4 indicated values of the number of nearest neighbors, $q$. (a) For $q = 2$, the particles sit in a 1D chain [Fig. \ref{fig.bethe_lattices}(b)], and there is a continuous spectrum above the quasiparticle pole at energy $\varepsilon_0 = J_z / 2 - \sqrt{(J_z / 2)^2 + (2t)^2}$. For higher values of $q$ [panels (b)-(d)], the spectrum consists of a series quasiparticle peaks, scaling with $(J_z / t)^{2/3}$ at strong interactions (colored lines). }
\label{fig.spectral_functions}
\end{figure}
Here, $\Sigma^{(0)}(\omega) = qt^2 \cdot G^{(1)}(\omega)$ is the self-energy associated with the Green's function $G^{(0)}(\omega)$ in Eq. \eqref{eq.G_0}. This equation simply states the fact that the poles of $G^{(0)}$ are eigenstate energies of the system, the lowest of which gives the true ground state of the system. To find the lowest order coefficient of the $n$th state, $C^{(0)}_n$, we compute the full norm of the wave function, $\ket{\Psi_n}$. Repeated use of the recursion relation in Eq. \eqref{eq.recursion_relation_eigenstate} yields
\begin{align}
&1 = \braket{\Psi_n|\Psi_n} = \left|C^{(0)}_n\right|^2 + q \sum_{d = 1}^{d_{\rm tot}} (q-1)^{d-1} \left|C^{(d)}_n\right|^2 \nonumber \\
&= \left|C^{(0)}_n\right|^2 \Big[1 + qt^2 \big(G^{(1)}(\varepsilon_n)\big)^2 \nonumber \\
&\phantom{=\left|C^{(0)}_n\right|^2 \Big[} + q(q-1)t^4 \big(G^{(1)}(\varepsilon_n)G^{(2)}(\varepsilon_n)\big)^2 + \dots \Big] \nonumber \\
&= \left|C^{(0)}_n\right|^2\! \Big[1 + qt^2 \big(G^{(1)}(\varepsilon_n)\big)^2 \big[1 + (q-1)t^2 G^{(2)}[1 + \dots] \big] \Big] \nonumber
\end{align}
Investigating the quasiparticle residue, $Z_n = [1 - \partial_{\omega}\Sigma^{(0)}(\omega)|_{\omega = \varepsilon_n}]^{-1}$, shows that the expression within the square brackets above is simply $Z^{-1}_n$. In turn, $C^{(0)}_n = \sqrt{Z_n}$, whereby we may write the eigenstate coefficients as
\begin{align}
C^{(0)}_n &= \sqrt{Z_n}, \nonumber \\
C^{(d)}_n &= \sqrt{Z_n} \cdot t^{d} \cdot \prod_{l = 1}^d G^{(l)}(\varepsilon_n), \; 1 \leq d \leq d_{\rm tot}.
\label{eq.C_d_n_general}
\end{align}
In the thermodynamic limit, these coefficients can, in complete analogy to the non-equilibrium case investigated in Subsec. \ref{subsec.thermodynamic_limit}, also be written as fractions of Bessel functions. This has previously been established for the ground state in Ref. \cite{Bieniasz2019}, where the Bethe lattice description is used as a variational ansatz for the square lattice case. In Fig. \ref{fig.spectral_functions}, we plot the spectral function $A(\omega) = -2{ \rm Im }\, G_0(\omega)$, computed in the thermodynamic limit from Eq. \eqref{eq.G_0_self_consistency}. This reveals the poles of $G_0$, and thereby the eigenstate energies $\varepsilon_n$. In the special case of $q = 2$ \cite{Batista2000}, the system is one-dimensional and the spectral function splits into a quasiparticle pole at
\begin{equation}
\varepsilon_0 = \frac{J_z}{2} - \sqrt{4t^2 + \frac{J_z^2}{4}}.
\label{eq.epsilon_0_q_2}
\end{equation}
and a continuum of states residing between $J_z/2 \pm 2t$. See also Fig. \ref{fig.spectral_functions}(a). This describes the effect known as spin-charge separation, which has been intensely studied both theoretically and experimentally in one-dimensional systems \cite{Tomonaga1950,Kim1996,Segovia1999,Auslaender2005,Kim2006,Jompol2009,Jayadev2020}. In the present case, as the charge degree of freedom moves (the hole), a single spin excitation is created and remains at the original site of the hole. Due to intense previous studies of these effects, we will not go any further with it in the present paper.
For more nearest neighbors, $q \geq 3$, a series of quasiparticles peaks arises [see Fig. \ref{fig.spectral_functions}(b)-(d)]. In Subsec. \ref{subsec.strong_interaction_limit}, we show that at strong interactions, their energies all scale with $(J_z / t)^{2/3}$. The origin of this scaling is that the system in this limit can be rephrased as a continuum model, in which the hole moves in a linear potential with strength $J_z$. While such a rephrasing in a regular lattice is approximate \cite{Bulaevskii1968,Kane1989,Grusdt2018_2}, this becomes exact in Bethe lattices as we shall unfold in Subsec. \ref{subsec.strong_interaction_limit}. In the limit of weak interactions on the other hand, $J_z \gg t$, the hole becomes immobile and the $n$th eigenstate is completely localized at the $n$th depth with an energy simply given by the magnetic energy cost, $\varepsilon_n \to E_J^{(n)} = [(q-1) + (n-1) (q-2)]J_z/2$ [Eq. \eqref{eq.E_J}] for $n\geq 1$. In fact, to order $\mathcal{O}(t^2 / J_z)$, only the magnetic polaron ground state and the first string excitation are affected and attain the energies
\begin{equation}
\varepsilon^{(\pm)} = \frac{(q-1)J_z}{4}\left[1 \pm \left(1 + \frac{q}{2}\left[\frac{4t}{(q-1)J_z}\right]^2\right)\right].
\label{eq.perturbative_energies}
\end{equation}
Here, $\varepsilon_0 = \varepsilon^{(-)}$ and $\varepsilon_1 = \varepsilon^{(+)}$ are the ground and lowest string excitation energies respectively. \\
The method used here to find the explicit coefficients of these eigenstates is analogous to Refs. \cite{Ramsak1998,Reiter1994}, where the approximate magnetic polaron eigenstates within the SCBA \cite{Kane1989,Martinez1991,Liu1991} are found. We finally note that these many-body eigenstates can in principle be used to completely characterize the full non-equilibrium dynamics. Specifically, $\braket{\Psi_n|\Psi(\tau)} = \bra{\Psi_n} {\rm e}^{-i\hat{H}\tau}\hat{h}^\dagger_0\ket{{ \rm AF }} = {\rm e}^{-i\varepsilon_n\tau}\!\bra{\Psi_n}\hat{h}^\dagger_0\ket{{ \rm AF }} = \sqrt{Z_n} \cdot {\rm e}^{-i\varepsilon_n\tau}$. Therefore, the dynamics is the quantum interference of the polaron ground state and the string excitations. While this lends insight into the conceptual nature of the dynamics, it does not actually simplify the characterization of the underlying hole motion.
\subsection{Continuum limit for strong interactions} \label{subsec.strong_interaction_limit}
In this subsection, we describe in detail the continuum limit arising for $J_z/t\to 0^+$ and a number of nearest neighbors $q\geq 3$. We begin by transforming the equations of motion to an effective one-dimensional setting valid for the depth symmetric eigenstates found in the previous Subsec. \ref{subsec.many_body_eigenstates}. To this end, we define $\psi(0) = C^{(0)}$, and $\psi(d) = (-1)^d \sqrt{q(q-1)^{d-1}} \cdot C^{(d)}$ for $d \geq 1$, similar to the approach in Ref. \cite{Grusdt2018_2}. In this manner, the wave function fulfills a one-dimensional normalization condition: $\sum_d |\psi(d)|^2 = 1$. With this in hand, the equations of motion become
\begin{align}
\!\!\!\varepsilon \psi(0) &= -\sqrt{q}t \psi(1), \nonumber \\
\!\!\!\varepsilon \psi(1) &= E_J^{(1)} \psi(1) - \sqrt{q}t \psi(0) - \sqrt{q-1}t \psi(2), \nonumber \\
\!\!\!\varepsilon \psi(d) &= E_J^{(d)} \psi(d) - \sqrt{q-1}t \left[\psi(d-1) + \psi(d+1)\right],
\label{eq.equations_of_motion_1}
\end{align}
where the lower equation is valid for $d\geq 2$. We may now formulate the continuum model by taking this lower equation and rephrase it as a continuous equation in the limit of $J_z \ll t$. This is possible, because the wave function spreads out over an ever increasing number of lattice sites in this limit, as we shall explicitly see in Sec. \ref{sec.results}. Here, it is first beneficial to rewrite this equation according to
\begin{align}
\tilde{\varepsilon}\psi(d) =& \, \frac{(q-2)J_z}{2\sqrt{q-1}t } d \cdot \psi(d) \nonumber \\
& - \left[\psi(d-1) - 2\psi(d) + \psi(d+1)\right],
\label{eq.equations_of_motion_2}
\end{align}
using Eq. \eqref{eq.E_J}, and defining $\sqrt{q-1}t \cdot \tilde{\varepsilon} = \left[\varepsilon + 2\sqrt{q-1}t \right]$. Next, we let $d = x / \lambda$ and introduce a rescaled wave function $\phi(x) = \psi(x / \lambda) / \sqrt{\lambda}$. Inserting this in Eq. \eqref{eq.equations_of_motion_2} and dividing out $\lambda^2$, we get
\begin{align}
\frac{\tilde{\varepsilon}}{\lambda^2}\phi(x) =& \, \frac{(q-2)J_z}{2\sqrt{q-1}t} \frac{x}{\lambda^3} \phi(x) \nonumber \\ &- \frac{\phi(x-\lambda) - 2\phi(x) + \phi(x+\lambda)}{\lambda^2}.
\label{eq.equations_of_motion_3}
\end{align}
To eliminate $J_z / t$, we set $\lambda = [(q-2)J_z/(2\sqrt{q-1}t)]^{1/3}$. For $J_z / t \to 0^+$, the fraction in the second line of Eq. \eqref{eq.equations_of_motion_3} approaches the second order derivative. Setting $a = \tilde{\varepsilon} / \lambda^2$, we, hereby, obtain
\begin{align}
&a\phi(x) = x \cdot \phi(x) - \frac{{\rm d}^2\phi}{{\rm d}x^2},
\label{eq.equations_of_motion_4}
\end{align}
which is accurate up to order $\lambda^2 \propto (J_z / t)^{2/3}$. The wave function $\phi(x)$ fulfills the continuous normalization condition: $\int_{0}^\infty\!{\rm d}x |\phi(x)|^2$. Letting $y = x - a$, we get the Airy equation
\begin{equation}
0 = y f(y) - \frac{{\rm d}^2f}{{\rm d}y^2},
\label{eq.airy_equation}
\end{equation}
for $y \geq - a$, where $f(y) = \phi(y + a)$. The solutions to this equation can thus be written as a superposition of the Airy functions: $f(y) = A \cdot {\rm Ai}(y) + B \cdot {\rm Bi}(y)$. Since ${\rm Bi}(y)$ blows up for $y \to \infty$, $B = 0$. Furthermore, the continuous one-dimensional description, which is exact for $J_z / t \to 0^+$, entails that $\phi(x)$ must vanish for $x < 0$, and therefore also at $x = 0$. In turn, $-a$ must be a zero of the Airy function ${\rm Ai}(y)$. In this way, we obtain the sought set of eigenstates in the strongly interacting limit
\begin{align}
\psi_n(d) &= \sqrt{\lambda}\cdot \phi_n\left(\lambda d\right) = \sqrt{\lambda} \cdot A_n {\rm Ai}\left(\lambda d - a_n\right).
\label{eq.eigenstates_strong}
\end{align}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig4.pdf}
\end{center}
\caption{\textbf{Eigenstates for strong interactions.} The hole density is plotted as function of depth for four indicated interaction strengths [panels (a) through (d)] and $q = 4$ nearest neighbors. We compare the numerically calculated eigenstates from Eq. \eqref{eq.C_d_n_general} to the strongly interacting limit, Eq. \eqref{eq.eigenstates_strong} with $d \to d - d_0$, for the three lowest string excitation states in red (ground state), green (first excited state) and blue (second excited state).}
\label{fig.eigenstates_strong}
\end{figure}
The $n$th eigenstate is, thus, defined by the $n$th order zero of the Airy function, $-a_n$. Also, $A_n^{-2} = \int_0^\infty {\rm d}x\, |{\rm Ai}\left(x - a_n\right)|^2$ ensures normalized eigenstates. For any nonzero value of $J_z / t$, the eigenstates remain nonzero at the origin, $d = 0$. This turns out to yield a correction term linear in $J_z / t$. To accommodate for this, we let $d \to d - d_0$ in Eq. \eqref{eq.eigenstates_strong} and use $d_0$ as a variational parameter. A rather lengthy calculation (see Appendix \ref{app.strong_eigenstates_d0}) yields the variational energy for the $n$th eigenstate
\begin{align}
\!\varepsilon_n^{\rm var} &= - 2\sqrt{q-1}t\left(1 - \frac{a_n}{2} \lambda^2\right) + c_q(d_0) J_z,
\label{eq.eigenenergies_order_1}
\end{align}
with $c_q(d_0) = (q-2)[((q-2)^{-1} + d_0) + 2(\sqrt{q(q-1)^{-1}} - 1) d_0(1-d_0) + (d_0+1/2)^2]$. The two first terms yield the exact asymptotic behavior \cite{Zhong1995} with the dominant $\lambda^2 \propto (J_z / t)^{2/3}$ scaling. The term proportional to $J_z$ is variationally determined, yielding a value of $d_0 = 1 / (2 - 3\sqrt{q(q-1)^{-1}})$. The resulting energies of the three lowest energies are plotted as colored lines in Fig. \ref{fig.spectral_functions} for $q = 3$, $4$ and $5$. The corresponding eigenstates for $q = 4$ are plotted in Fig. \ref{fig.eigenstates_strong}, showing the approach to the strongly interacting limit. Whereas the behavior at large depths is already captured very well at $J_z / t = 0.16$, the convergence of the full wave function requires very strong interactions of around $J_z / t = 0.02$. \\
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig5.pdf}
\end{center}
\caption{\textbf{Residue for strong interactions.} The residue is plotted for three lowest energy states as a function of interaction strength: the ground state in red, the first excited state in green, and the second excitated state in blue. This is compared to the linear behavior in Eq. \eqref{eq.residue_asymptote} (black dashed), confirming the universal linear trend at strong enough interactions.}
\label{fig.residue_strong}
\end{figure}
The presence of a nonzero $d_0$ allows us to to calculate the residues, $Z_n$, in this asymptotic strong coupling limit. This yields
\begin{align}
Z_n = |\psi_n(d = 0)|^2 \simeq \lambda \cdot \left[A_n {\rm Ai}\left(-\lambda d_0 - a_n\right)\right]^2
\label{eq.residue_approximate}
\end{align}
This can be greatly simplified by expanding the Airy function. In fact, ${\rm Ai}\left(-\lambda d_0 - a_n\right) \to (-\lambda d_0) {\rm Ai}'(-a_n)$, for $\lambda \propto (J_z / t)^{1/3} \to 0$. Furthermore, using the integral relation $\int_0^\infty {\rm d}x [{\rm Ai}(x - x_0)]^2 = x_0 [{\rm Ai}(-x_0)]^2 + [{\rm Ai}'(-x_0)]^2 = [{\rm Ai}'(-a_n)]^2$, and that $A_n^{-2} = \int_0^\infty {\rm d}x [{\rm Ai}(x - a_n)]^2 ) = [{\rm Ai}'(-a_n)]^2$ for $(J_z / t)^{1/3} \to 0$, we get the very simple asymptotic behavior of the residues
\begin{align}
Z_n &\to \lambda \cdot \left[A_n (-\lambda d_0) {\rm Ai}'(-a_n)\right]^2 = \lambda^3 d_0^2 \nonumber \\
&= \frac{q-2}{2\sqrt{q-1}}\frac{1}{\left(2 - 3 \sqrt{\frac{q}{q-1}}\right)^2} \cdot \frac{J_z}{t},
\label{eq.residue_asymptote}
\end{align}
which is independent of the eigenstate $n$, essentially because the energy shifts due to a nonzero $d_0$, $\varepsilon_n^{\rm var} - \varepsilon_n$ is independent of $n$. Consequently, this variational approach strongly suggests that the residues of all string excitation states approach this universal value, linear in $J_z / t$, which is confirmed by our numerical calcuations in Fig. \ref{fig.residue_strong} for $q = 4$ nearest neighbors. Consequently, the quasiparticles remain well-defined for any $J_z/t>0$.
\section{Numerical results} \label{sec.results}
In this section, we describe the main results for the non-equilibrium dynamics. In Subsection \ref{subsec.hole_dynamics}, we investigate the hole dynamics, whereas the non-equilibrium spin-spin correlation dynamics is studied in Subsec. \ref{subsec.spin_dynamics}.
\subsection{Hole dynamics} \label{subsec.hole_dynamics}
The density of holes at certain depth $d$ is readily computed from the coefficients of the many-body wave function as
\begin{align}
n_h(d = 0, \tau) &= |C^{(0)}(\tau)|^2, \nonumber \\
n_h(d, \tau) &= q(q-1)^d \cdot |C^{(d)}(\tau)|^2, \;\; d\geq 1.
\label{eq.n_h_d_and_C_d}
\end{align}
Here, $C^{(d)}(\tau)$ is calculated from Eq. \eqref{eq.C_d_and_R_d}. The results for $q = 4$ and $J_z = 0.2t$ is shown in Fig. \ref{fig.local_densities} in the vicinity of the initial position of the hole. Since $C^{(d)}(\tau)$ is the Fourier transform of $2{ \rm Re }[R^{(d)}(\omega)]$ [given in Eq. \eqref{eq.R_d_general}], the appearance of heavy oscillations in these local densities can be understood from the presence of a plethora of spectral peaks, Fig. \ref{fig.spectral_functions}(d), corresponding to the string excitations found in Subsec. \ref{subsec.many_body_eigenstates}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig6.pdf}
\end{center}
\caption{\textbf{Local densities}. We plot the local density of holes in the vicinity of its original position, $d = 0$, for 4 indicated depths (insets) in the case of $q = 4$ nearest neighbors, and a spin-spin coupling $J_z = 0.2t$. This is further compared to the limit of infinite interactions, $J / t \to 0^+$, where the hole moves like a free particle. While the initial stage shows a clear dampening of regular oscillations, associated with a free quantum walk, for any nonzero $J_z /t$ large aperiodic oscillations start to kick at later times, in the present case of $J_z / t = 0.2$ around $\tau = 10 / t$.}
\label{fig.local_densities}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig7.pdf}
\end{center}
\caption{\textbf{Time-averaged hole distribution}. In black squares is shown the time-averaged hole density distribution $\braket{n_h(d)}_\tau$ [Eq. \eqref{eq.average_hole_density}] compared to the hole distribution for the first 3 eigenstates (colored markers) as a function of the depth $d$ and for 4 indicated interaction strengths. The coordination number is set to $q = 4$.}
\label{fig.ave_density_distributions}
\end{figure}
To better understand this complex many-body scenario, we calculate the average depth of the hole,
\begin{equation}
l_s(\tau) = \sum_d d\cdot n_h(d,\tau).
\label{eq.string_length}
\end{equation}
We denote it $l_s$, as it also gives the average \emph{length of overturned spins} at time $\tau$. This is plotted in Fig. \ref{fig.string_length_q}(a) as a function of time for three indicated coordination numbers. Solving the equations of motion at short times, reveals that the initial dynamics is a free quantum walk independent of the inverse interaction strength $J_z / t$, with
\begin{equation}
l_s = qt^2 \cdot \tau^2 + \mathcal{O}[(t\cdot\tau)^4].
\label{eq.l_s_short_times}
\end{equation}
It is actually fairly easy to show \cite{Nielsen2022} that the initial motion of an initially localized hole within the $t$-$J$ model has to be that of a free quantum walk, even in the presence of anisotropic spin-couplings.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig8.pdf}
\end{center}
\caption{\textbf{String length vs. coordination number}. (a) The string length $l_s(\tau) = \sum_d d \cdot n_h(d, \tau)$ as a function of time for three different coordination numbers, $q = 4, 8, 16$. For $\tau \ll 1 / \sqrt{q}t$, all curves collapse to an initial ballistic behavior of the hole, corresponding to $l_s = qt^2\cdot \tau^2$. At long times, the string length has heavy oscillations around a well-defined mean value, $l_s^\infty$, shown in dashed lines. This mean value is plotted in panel (b) as a function of the coordination number $q$ for three different values of $J_z / t$, and compared to the weak coupling result (dashed lines).}
\label{fig.string_length_q}
\end{figure}
At long times, our results reveal that the string length undergoes heavy oscillations around a well-defined mean value,
\begin{equation}
l_s^\infty = \sum_d d\cdot \braket{n_h(d)}_\tau,
\label{eq.string_length_long_time_average}
\end{equation}
where
\begin{equation}
\braket{n_h(d)}_\tau = \lim_{T\to\infty}\frac{1}{T} \int_0^T {\rm d}\tau \; n_h({\bf d}, \tau).
\label{eq.average_hole_density}
\end{equation}
Physically, this finite asymptote reflects that the hole remains bound to its initial position. The time-averaged hole distribution in Eq. \eqref{eq.average_hole_density} is shown in Fig. \ref{fig.ave_density_distributions} for a set of indicated inverse interaction strengths $J_z / t$, and compared to the hole distribution for the polaron ground state and the two lowest string states [see Subsec. \ref{subsec.many_body_eigenstates}]. It is evident that for strong coupling, $J_z \ll t$, the dynamical wave function is significantly more spread out than its equilibrium counterparts. This is a natural consequence of the fact that the average energy of the quenched system $\bra{\Psi(\tau)} H \ket{\Psi(\tau)} = 0$ (relative to $E_J^{(0)} = qJ_z / 2 + E_0(q,d)$) is much larger than the ground state energy $\sim-2\sqrt{q-1}t$. Additionally, the shape of the time-averaged distribution changes quite dramatically with decreasing $J_z / t$. Indeed, below $J_z = 0.4t$ the hole is no longer found with the highest probability at its original site, but rather one of its nearest neighbors. In Fig. \ref{fig.string_length_q}(b), we further compare the asymptotic string length $l_s^\infty$ to the weak coupling result,
\begin{equation}
l_s^\infty \to \frac{8q}{(q-1)^2}\cdot \left(\frac{J_z}{t}\right)^{-2},
\label{eq.l_s_weak_coupling}
\end{equation}
valid for $J_z / t \gg 1$. To further investigate the dependency on the interaction strength, we plot the string length dynamics in Fig. \ref{fig.string_length_J}(a) for several indicated values of $J_z / t$. As can be expected, the hole travels further into the Bethe lattice for decreasing values of $J_z / t$. The motion is generally aperiodic, due to the irregular spacing of the energy levels $\varepsilon_n$ evident in Fig. \ref{fig.spectral_functions}. The only exception is in the limit of very weak interactions, $J_z \gg t$, in which the hole is restricted to hop back and forth between depths $d = 0$ and $d = 1$, resulting in periodic motion with an angular frequency given by the energy difference between the polaron ground state and the lowest string excitation,
\begin{equation}
\varepsilon_1 - \varepsilon_0 \to (q-1)\frac{J_z}{2}\left(1 + \frac{q}{2}\left[\frac{4t}{(q-1)J_z}\right]^2\right),
\label{eq.perturbative_angular_frequency}
\end{equation}
approaching the magnetic energy cost of going to depth $d = 1$, $E_J^{(1)} = (q-1)J_z / 2$, as $J_z \gg t$. \\
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig9.pdf}
\end{center}
\caption{\textbf{String length vs. interaction strength}. (a) The string length is plotted as a function of time for $q = 4$ for four indicated inverse interaction strengths, $J_z / t$, with long time averages, $l_s^{\infty}$, shown as dashed lines. In the limit of infinitely strong interactions, $J_z / t = 0^+$, the hole motion is a free quantum walk [black solid line]. (b) The long time averages, $l_s^{\infty}$, is plotted as a function of interaction strength and further compared to the string length in the magnetic polaron ground state $l_s^{0}$ [panel (c)] for three indicated values of the coordination number $q$. At weak coupling, $l_s^{\infty} = 2 \cdot l_s^0 \propto (J_z / t)^{-2}$ (black short-dashed lines). At strong coupling, they feature different power-law scalings (black long-dashed lines) with $l_s^{\infty} \propto (J_z / t)^{-1}$ and $l_s^0 \propto (J_z / t)^{-1/3}$.}
\label{fig.string_length_J}
\end{figure}
In the opposite extreme of $J_z / t = 0^+$, the dynamics is characterized by a free quantum walk of the hole as anticipated by Eqs. \eqref{eq.G_1_J_0_limit}-\eqref{eq.J_0_limit}. In Figures \ref{fig.string_length_J}(b)-\ref{fig.string_length_J}(c), we further characterize the full dependency on the inverse interaction strength, $J_z / t$. Whereas the two string lengths are simply proportional in the weak coupling limit with $l_s^\infty = 2\cdot l_s^0$, they feature remarkably different scaling behaviors for strong coupling. In fact, for any number of nearest neighbors, $q \geq 3$, our power-law fits at strong interactions, $J_z / t \ll 1$, reveal that $l_s^\infty = f^\infty(q) \cdot (J_z / t)^{-1}$. On the contrary, the scaling-law for the eigenstates are dramatically different, as we may derive explicitly from the strong coupling states derived in Subsec. \ref{subsec.strong_interaction_limit}. Explicitly, we can use that the $n$th eigenstate is asymptotically given by $\psi_n(d) = \sqrt{\lambda}\phi_n(\lambda (d-d_0)) = \sqrt{\lambda}A_n \cdot {\rm Ai}(\lambda(d-d_0)-a_n)$. Here, we also include the effect of a nonzero shift $d_0$. We then get
\begin{align}
l_s^n &= \sum_d d \cdot |\psi_n(d)|^2 = \lambda \sum_d d \cdot |\phi_n(\lambda (d-d_0))|^2 \nonumber \\ &= \sum_{x\geq -\!\lambda d_0} \!\!\!\left(\frac{x}{\lambda} + d_0 \right) |\phi_n(x)|^2 \Delta x. \nonumber
\end{align}
Here, we use $x = \lambda(d-d_0)$, whereby $\Delta x = \lambda$. The term proportional to $d_0$ is simply the normalization of the wave function, and so just yields $d_0$. The remaining terms can, in the limit of strong interactions $\lambda \propto (J_z/t)^{1/3} \to 0^+$, be rephrased as an integral. This yields
\begin{align}
l_s^n &\to d_0 + \frac{1}{\lambda}\int_0^\infty {\rm d}x \, x |\phi_n(x)|^2 \nonumber \\
&= d_0 + \frac{1}{\lambda}\int_0^\infty {\rm d}x \, x A_n^2 [{\rm Ai}(x-a_n)]^2 \nonumber \\
&= d_0 + \frac{1}{\lambda} \cdot \frac{2a_n}{3} A_n^2 [{\rm Ai}'(-a_n)]^2 = d_0 + \frac{2a_n}{3\lambda}.
\label{eq.l_s_n}
\end{align}
In the last line, we first use an integral relation for the Airy functions: $\int_0^\infty {\rm d}x \, x A_n^2 [{\rm Ai}(x-x_0)]^2 = (2x_0 [{\rm Ai}(-x_0)]^2 - {\rm Ai}(-x_0){\rm Ai}'(-x_0) + 2 x_0 [{\rm Ai}'(-x_0)]^2)/3$, and that $-x_0 = -a_n$ is a zero of the Airy function. Finally, we use that the normalization constant is given by $A_n^{-2} = [{\rm Ai}'(-a_n)]^2$. This expression, hereby, yields a dominant $\lambda^{-1} \propto (J_z/t)^{-1/3}$ scaling of the string length for all eigenstates. Additionally, the increase in string length for eigenstates with higher energy is simply linearly related to the increase in the zeroes of the Airy function, $a_n$. In Fig. \ref{fig.string_length_J}, Eq. \eqref{eq.l_s_n} is compared to the numerically obtained string length for the ground state for three different values of the number of nearest neighbors, showing excellent agreement at strong coupling. \\
We note that to get converging results for the thermodynamic limit in the case of $q = 4$ and very strong interactions of $J_z = 0.05t$, we need to go to a total depth of at least $d_{\rm tot} = 200$. In this case, the Bethe lattice consists of $N(q = 4, d_{\rm tot} = 200) \simeq 10^{95}$ sites [Eq. \eqref{eq.number_of_sites}]. This far exceeds the total number of atoms in the observable universe \cite{LifeScience_atoms_in_obs_universe}, and exemplifies the enormity of the simplification achieved when reducing the description of an exponential number of sites in the Bethe lattice with just a linear number of coefficient, $C^{(d)}$.
\subsection{Spin dynamics} \label{subsec.spin_dynamics}
In the present subsection, we investigate the dynamics of the spin-spin correlation function
\begin{equation}
C_S(d, \tau) = 4\bra{\Psi(\tau)}\hat{S}^{(z)}_{0}\hat{S}^{(z)}_{{\bf j}_d}\ket{\Psi(\tau)}.
\label{eq.spin_correlator}
\end{equation}
This describes the tendency of the spin at the origin, $d = 0$, to align ($C_S > 0$) or antialign ($C_S < 0$) with a spin at depth $d$. Note that the depth symmetry of the dynamics entails that $C_S$ only depends on the depth $d$ of the second spin. The advent of quantum simulation platforms enables the study of such quantities, as has been seen in two-dimensional square lattices both in \cite{Koepsell2019} and out of equilibrium \cite{Ji2021}. On the other hand, the actual computation of these correlators often present an astonishing theoretical feat. In the presently studied Bethe structures, however, the full knowledge of the many-body wave function enables the precise and efficient investigation of the spin-spin correlator in Eq. \eqref{eq.spin_correlator}. \\
To see this more concretely, we link $C_S$ to the coefficients of the many-body wave function. In the absence of a hole, the system is a perfect antiferromagnetic state, resulting in $C_S^{(0)}(d) = 4\bra{{ \rm AF }}\hat{S}^{(z)}_{0}\hat{S}^{(z)}_{{\bf j}_d}\ket{{ \rm AF }} = (-1)^{d}$. This overall sign expresses the perfectly staggered antiferromagnetism. In the presence of a hole, we now link $C_S(d, \tau)$ to the hole density. Consider, then, first the case where the hole is located between depths $0$ and $d$, $0 < d_h < d$. In this case, the $z$ component of the spin at $d = 0$ is just $+ 1 / 2$, while the $z$ component of the spin at depth $d$ is $(-1)^{d-1} / 2$. In turn, we get a contribution of $(-1)^{d-1} \cdot P(0 \!<\! d_h \!<\! d, \tau)$ to $C_S(d, \tau) $. Here, $P(0\!<\!d_h\!<\!d,\tau) = \sum_{0<d_h<d} n_h(d_h, \tau)$ is the probability to find the hole between depths $0$ and $d$ at time $\tau$.
Next, if the hole has passed depth $d$, the above correlation flips sign if the hole has passed the specific site ${\bf j}_d$. If not, then the correlation does not flip. The relative probability to have passed ${\bf j}_d$ is just $1 / q(q-1)^{d-1}$. If the hole passes ${\bf j}_d$, there is, thus, a contribution of $(-1)^{d} \cdot P(d_h \!>\! d, \tau) / q(q-1)^{d-1}$. If it does \emph{not} pass ${\bf j}_d$, it contributes with $(-1)^{d-1} \cdot P(d_h \!>\! d, \tau) (1 - 1/ q(q-1)^{d-1})$. Here, $P(d_h \!>\! d, \tau) = \sum_{d_h > d}n_h(d_h, \tau)$ is the probability for the hole to have passed depth $d$.
Finally, if the hole is at the specific depth $d$, the correlator $C_S(d, \tau)$ vanishes if the hole is at site ${\bf j}_d$. The contribution from this scenario is, therefore, only $(-1)^{d-1} \cdot P(d_h = d, \tau) (1 - 1/ q(q-1)^{d-1})$, coming from the instance where the hole \emph{is not} at ${\bf j}_d$. In total then, the non-equilibrium spin-spin correlator in Eq. \eqref{eq.spin_correlator} is
\begin{align}
&C_S(d, \tau) = (-1)^{d-1} \Bigg[P(0 < d_h < d, \tau) \nonumber \\
&\phantom{C_S(d, \tau) =}+ P(d_h > d, \tau)\left(1 - \frac{2}{q(q-1)^{d-1}}\right) \nonumber\\
&\phantom{C_S(d, \tau) =}+ P(d_h = d, \tau)\left(1 - \frac{1}{q(q-1)^{d-1}}\right) \Bigg] \nonumber \\
&\!= \!(-1)^{d-1} \Bigg[1 \!-\! n_h(0, \tau) \!-\! \frac{n_h(d, \tau) + 2\sum_{d_h\! > d} n_h(d_h, \tau)}{q(q-1)^{d-1}} \Bigg].
\label{eq.spin_correlator_result}
\end{align}
This expression is valid for any $d\geq 1$. For $d = 0$, we simply have $C_S(d = 0, \tau) = 1 - n_h(d = 0,\tau)$. In Eq. \eqref{eq.spin_correlator_result}, we use that the total probability of finding the hole away from the origin is $1$ minus the hole density at $d = 0$: $P(0 \!<\! d_h \!<\! d, \tau) + P(d_h \!=\! d, \tau) + P(d_h \!>\! d, \tau) = P(d_h \!>\! 0, \tau) = 1 - n_h(d_h = 0, \tau)$. At $\tau = 0$, it follows that $C_S(d, \tau = 0) = 0$, which also has to be the case physically, because there is no spin at $d = 0$ initially. At later times, as $n_h(0, \tau)$ diminishes, the spin-spin correlations can be strongly affected in the vicinity of $d = 0$. If e.g. the hole is entirely located at $d = 1$, $C_S(d = 1, \tau) = 1 - 1 / q > 0$. This has the opposite sign of the spin correlations in the absence of holes, and is simply a result of removing the original spin-$\downarrow$ fermion at $d = 0$, and letting the resulting hole travel to $d = 1$ [see Fig. \ref{fig.bethe_lattices_broken_bonds}].
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\columnwidth]{fig10.pdf}
\end{center}
\caption{\textbf{Spin dynamics}. The time-dependent spin-spin correlation function $C_S(d,\tau) = 4\bra{\Psi(\tau)}\hat{S}^{(z)}_0\hat{S}^{(z)}_{{\bf j}_d}\ket{\Psi(\tau)}$ [Eqs. \eqref{eq.spin_correlator} and \eqref{eq.spin_correlator_result}] relative to the spin correlation in the absence of a hole $C_S^{(0)}(d) = (-1)^d$ for indicated depths $d$ through panels (a) through (d). This is shown in the case of weak ($J = 2t$, blue lines), intermediate ($J = 0.5t$, green lines), strong ($J = 0.2t$, orange lines), and very strong interactions ($J = 0.1t$, red lines), as well as in the quantum walk limit of $J / t \to 0^+$ (black lines). Note that for all $d\geq 1$, the spin correlation has flipped sign with respect to the value in the absence of holes.}
\label{fig.spin_dynamics}
\end{figure}
We investigate this mechanism in more detail in Fig. \ref{fig.spin_dynamics} by plotting the full dynamics of the spin-spin correlator [Eq. \eqref{eq.spin_correlator_result}] relative to the spin correlator in the absence of holes. Throughout the entire dynamics, we observe the mentioned flip in correlation for any $d\geq 1$. Furthermore, for weak to intermediate interactions we observe heavy oscillations originating in the density oscillations [Fig. \ref{fig.local_densities}]. At very strong interactions, approaching the free quantum walk of the hole, the relative spin correlation reaches an asymptotic value of $C_S(d,\tau) / C_S^{(0)}(d) \to -1 + 2 / q(q-1)^{d-1}$ at long timescales, $\tau \gg 1 / t$. This is because the hole in this case will always leave any finite region of the Bethe lattice, so that $\sum_{d_h > d} n_h(d_h, \tau) \to 1$ in Eq. \eqref{eq.spin_correlator_result}, while $n_h(0,\tau), n_h(d,\tau) \to$. Finally, by carefully analyzing the possible extremal values of $C_S(d,\tau) / C_S^{(0)}(d)$, we find that
\begin{align}
\frac{C_S(d = 0,\tau)}{C_S^{(0)}(d = 0)} &\in [0, 1],\nonumber \\
\frac{C_S(d = 1,\tau)}{C_S^{(0)}(d = 1)} &\in [-1 + 1 / q, 0],\nonumber \\
\frac{C_S(d \geq 2,\tau)}{C_S^{(0)}(d \geq 2)} &\in [-1, 0],
\label{eq.C_S_limits}
\end{align}
used as the axis limits on the second axes in Fig. \ref{fig.spin_dynamics}. This result is not limited to the $t$-$J_z$ model investigated in the present paper, but holds in general. It only depends on the depth symmetry of the wave function in Eq. \eqref{eq.exact_many_body_wave_function}, and may be derived by varying $C_S(d,\tau) / C_S^{(0)}(d)$ with respect to the coeffients $C^{(d)}$ of the wave function given that the norm of the wave function is preserved, $\braket{\Psi|\Psi} = 1$. Indeed, we see that $C_S$ does not necessarily explore all the possible values, evident in the cases of $d = 1$ and $d = 2$, Figs. \ref{fig.spin_dynamics}(b) and \ref{fig.spin_dynamics}(c) respectively. \\
In this way, we see how we may we characterize both the hole and spin dynamics exactly and very efficiently in these Bethe lattice structures.
\section{Conclusions and outlook}
In this article, we have found exact solutions to the non-equilibrium many-body dynamics and a certain class of eigenstates of a single hole in antiferromagnetic Bethe lattices, described by the fully anisotropic $t$-$J_z$ model. The found eigenstates include the magnetic polaron ground state as well as the ubiquitous string excitations. The latter are in this case exact many-body eigenstates with a vanishing spectral width in contrast to the $t$-$J_z$ model in regular crystal lattices \cite{Wrzosek2021}, as well as in the presence of transverse spin coupling present in the full $t$-$J$ model \cite{Kane1989,Martinez1991}.
As our methodology yields the full many-body wave function, any correlation function can be calculated very efficiently, illustrated by the investigated spin-spin correlation dynamics. The exact solvability of the model is a result of the fractal self-similarity of Bethe lattices, which we have shown leads to simple recursion relations for the coefficients of the wave function. We anticipate that it should be possible to extend the present methodology to nonzero temperatures. In this case, we see two possible routes forward, either by expressing the system dynamics in terms of a full density matrix, or by translating the methodology to finite temperature quantum field theory. In either case, understanding the impact of temperature in these highly idealized lattices may further our understanding of the same phenomena in regular crystal lattices. Finally, the exploration of pairing of two holes is essential to improve our understanding of the mechanisms behind high-temperature superconductivity. The massive simplification found in the Bethe lattices for a single hole in the present work, may lead to interesting new insights into this scenario, which we hope to explore in the future. \\
\acknowledgments
KKN would like to thank Georg M. Bruun for helpful and valuable input on the manuscript. KKN also thanks Jens Havgaard Nyhegn and Thomas Pohl for fruitful discussions. This work has been supported by the Carlsberg Foundation through a Carlsberg Internationalisation Fellowship and the Danish National Research Foundation through the Center of Excellence “CCQ” (Grant agreement no.: DNRF156).
|
\section{Overview}
A generalized diagonal of a polygonal billiard is an orbit which connects two
vertices of the polygon. One of the characteristics of billiard dynamics is the complexity function
$P_n$, which counts the number of generalized diagonals of length no greater than $n$. Here
by length we mean the number of reflections. See also [1], [4], [5], [6] for other notions of billiard complexity, related to our definition.
\
\
Katok [9] proved the subexponential estimate.
\
\
\begin{thm} (Katok, 1987). For any polygon: $\lim \frac{ln(P_n)}{n}=0$
\end{thm}
\
Masur [11], [12] using Teichmuller theory proved quadratic estimates for any
rational-angled polygon:
\begin{thm} (Masur, 1990). For any polygon with angles in $\pi \mathbb{Q}$ there are constants $C_1, C_2 > 0$ such that: $C_1 · n^2<P_n<C_2 n^2$
\end{thm}
\
Due to the difficulty of the further progress in non-rational case, Katok formulated the following
question, which he included in his list “Five most resistant problems in dynamics”[10]:
\
\
\textit{Find the upper and lower estimates for} $P_n$.
\
\
The key difficulty here is the luck of structure for general polygonal billiards. In the rational case the billiard flow reduces to the geodesic flow on the translation surface which allows its study by methods of Teichmuller theory. However in irrational case the billiard is only equivalent to the geodesic flow on the flat sphere with finite number of conical singularities, with infinite holonomy group. For the zero measure set of billiards with extremely well-approximated angles ergodicity results were obtained by Katok [8] and then, more explicitly by Vorobets[15]. Recently Chaika and Forni [2] gave an example of a weak-mixing billiard, essentially using rational methods.
\
\
Regarding billiard complexity, in works [13], [14] we developed a bootstrapping and indexed partition technique allowing in its strongest form the stretched exponential estimate for typical triangular billiards.
\
\begin{thm}(S, 2012) For a typical (in the Lebesgue measure sense) triangular billiard and any $\epsilon>0$ the following estimate takes place: $P_n<Ce^{n^{\sqrt{3}-1+\epsilon}}$.
\end{thm}
\
\begin{thm}(S, 2020) For a typical (in the Lebesgue measure sense) triangular billiard and any $\epsilon>0$ the following estimate takes place: $P_n<Ce^{n^{\epsilon}}$.
\end{thm}
\
In the current paper we find a \textit{lower} estimate on $P_n$ for right triangular billiards using indexed partitions, however the geometric idea differs from that of [14]. Informally speaking the key idea of [14] was based on some delicate interplay between the abundance of \textit{long} generalized diagonals and \textit{short} ones. Here we roughly show that relatively small number of generalized diagonals implies the existence of rather short periodic orbits, which in turn produce new \textit{unwanted} generalized diagonals, which gives a contradiction.
The aim of the current paper is to prove the following estimate:
\begin{thm}For a typical (in the Lebesgue measure sense) right triangular billiard and for any $\epsilon>0$ the complexity function $P_n$ satisfies: $\limsup \frac{P_n}{n^{\frac{2}{\sqrt{3}}-\epsilon}}>0$
\end{thm}
\
\section{Development along the trajectory and geodesic flow on the flat punctured sphere}
We remind a useful development construction associated to any
polygonal billiard [7]. We fix a polygon on the plane and consider a time moment
when a particular billiard orbit hits a polygon side. Then instead of reflecting
the orbit we continue it as a straight line and reflect the polygon along the
line. As we continue this process indefinitely the sequence of polygons obtained
this way is called development of the polygon along the orbit. See fig. $1$, showing triangle development.
\
\
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-6.554861)(14.78236,3.0313833)
\pspolygon[linecolor=black, linewidth=0.02](2.1789713,-5.581379)(4.3850155,-1.5864611)(7.1300006,-2.409296)
\pspolygon[linecolor=black, linewidth=0.02](2.2026267,-5.586878)(4.402627,-1.5668777)(2.2226267,0.29312226)
\psline[linecolor=black, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(0.3426268,-2.1668777)(14.742627,-2.1868777)
\pspolygon[linecolor=black, linewidth=0.02](2.1826267,-5.566878)(0.0226268,-1.5468777)(2.2026267,0.31312224)
\pspolygon[linecolor=black, linewidth=0.02](4.7995067,2.9812706)(4.3936834,-1.5833442)(7.136144,-2.4145544)
\pspolygon[linecolor=black, linewidth=0.02](4.8162565,2.978476)(8.409924,0.1657267)(7.161407,-2.4136543)
\pspolygon[linecolor=black, linewidth=0.02](12.866853,-0.9558881)(8.420342,0.15269212)(7.1740117,-2.4277458)
\pspolygon[linecolor=black, linewidth=0.02](12.865252,-0.9518255)(9.523942,-4.060145)(7.1721635,-2.4227266)
\pspolygon[linecolor=black, linewidth=0.02](12.890776,-0.9414156)(9.5456295,-4.0735707)(11.035921,-6.521229)
\pspolygon[linecolor=black, linewidth=0.02](12.895707,-0.93856764)(13.69782,-5.431074)(11.047332,-6.52053)
\end{pspicture}
}
\end{figure}
Fig. 1. Development of triangular billiard along the trajectory.
\
\
Another useful construction, associated with the polygonal billiard is the geodesic flow on the flat sphere with a finite number of conical singularities. Namely we take a copy $B$ of the polygon $A$ and glue them along corresponding sides. The resulting object is a sphere $S$ with a flat metric on it and a finite number of singularities, corresponding to the vertices of $A$. We slightly modify billiard flow $T_t$ on the polygon $A$. Namely, having a billiard orbit $T_t$ in $A$, we continue it in its copy $B$, after it hits the side of $A$. Analogous thing we make with the billiard in $B$. So the resulting flow "jumps" between $A$ and $B$ after each reflection, thus giving a geodesic flow $\overline{T_t}$ on $S$.
\
\
\textbf{Remark 1.} On has to be a bit careful about relation between \textit{periodic} orbits of flows $T_t$ and $\overline{T_t}$. Namely if $T$ is a periodic orbit of $T_t$ of \textit{even} period (meaning an even number of reflections) then it does correspond to the periodic orbit $\overline{T}$ of $\overline{T_t}$, which is not the case for periodic orbits of an odd period. The issue here is that any orbit on $S$ "jumps" between $A$ and $B$, so there has to be an even number of reflections ( or "side-crossings" on $S$) in order for it to return to the "correct" side.
\
\
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-5.979097)(14.034969,1.0727091)
\definecolor{colour0}{rgb}{0.0,0.5019608,0.0}
\pspolygon[linecolor=black, linewidth=0.03](0.02895874,-1.3140969)(2.0089588,1.005903)(4.888959,0.285903)(5.508959,-2.414097)(3.2889588,-5.954097)(3.2889588,-5.954097)
\pspolygon[linecolor=black, linewidth=0.03](8.528958,-1.274097)(10.508959,1.045903)(13.388959,0.32590303)(14.008959,-2.3740969)(11.788959,-5.914097)(11.788959,-5.914097)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(2.8089588,-3.514097)(5.2689586,-1.374097)
\psline[linecolor=blue, linewidth=0.03, linestyle=dashed, dash=0.17638889cm 0.10583334cm, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(11.308959,-3.534097)(13.768959,-1.394097)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(13.748959,-1.374097)(8.8289585,-0.934097)
\rput[bl](2.3289587,-0.21409698){A}
\rput[bl](10.608959,-0.21409698){B}
\psline[linecolor=red, linewidth=0.03, linestyle=dashed, dash=0.17638889cm 0.10583334cm, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(5.3089585,-1.354097)(0.38895875,-0.914097)
\psline[linecolor=colour0, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(0.36895874,-0.894097)(0.38895875,-1.754097)
\end{pspicture}
}
\end{figure}
Fig. 2. Two copies of the billiard table $A$ and $B$ are glued together along sides and form flat sphere with conical singularities at vertices. Blue geodesic trajectory at the side $A$ of the sphere instead of reflecting from the boundary, continues as a red trajectory at the side $B$ of the sphere. Then red trajectory instead of reflecting from the boundary of $B$, continues as a green trajectory on the side $A$ of the sphere.
\
\
\begin{Lemma} Let $M$ be a $2$-dimensional oriented manifold with a flat metric, $x_0\in M$ and $v_0\in T_{x_0}(M)$. Let us also assume that the geodesic $\gamma_0$ of length $L$ which starts at point $x_0$ in the direction of the vector $v_0$ is closed. Let $x=x(t)$, $t\in[0, 1]$ be a continuous curve on $M$ and $x(0)=x$. Let $v_t$ be a parallel translation of the vector $v_0$ along the curve $x(t)$ and let $\gamma_t$ be a geodesic of length $L$ which starts at the point $x_t$ in the direction $v_t$. Then $\gamma_t$ is closed for $t\in [0, 1]$.
\end{Lemma}
\begin{proof} For any point $y\in M$, a vector $v\in T_y(M)$ and a constant $L>0$ we consider a shift map $S_L$, defined as follows. Consider a small enough disk $U$ centered at $y$ and a point $a\in U$. Let
$u\in T_aM$ be a parallel translate of $v$ and $\gamma_u(s)$, $s\in[0, L]$ is a geodesic of length $L$ from $a$ into direction $u$. Then we define $S_L(a)=\gamma_u(L)$. It is easily seen that $S_L$ is a local isometry.
\
\
Let now $x_t, v_t$, $t\in[0, 1]$ be as in the conditions of the Lemma. Let $C\subseteq[0, 1]$ be such that for $t\in C$ the trajectory $\gamma_t$ of length $L$ is a closed geodesic. As $S_L$ is a local isometry then $C$ is open. But of course $C$ is closed as the limit of closed geodesics of length $L$ is also a closed geodesic of length $L$. As $C$ is non-empty, it implies that $C=[0, 1]$.
\end{proof}
Fig. 3. Reflecting triangle about right angle 3 times, we make an associated rhombus. Any billiard trajectory in a rhombus naturally projects to the triangle billiard trajectory. For any reflection of the trajectory in the rhombus, there are at most 2 more reflections of the triangular trajectory, corresponding to "diagonal-crossings".
\
\
Another useful lemma will be used later on and will allow us to consider a rhombus instead of triangle which turns out to be slightly easier from technical point of view. We first take a right triangle and reflect it 3 times about the right angle to get a rhombus. As fig.3 shows then the billiard in the rhombus naturally "projects' to the billiard in triangle. As one can see, each trajectory of rhombus billiard can cross at most 2 triangle sides between the two consecutive reflections. That means that any generalized diagonal of rhombus billiard of algebraic length $n$ corresponds to the generalized diagonal of a triangle billiard of length no more than $3n$.
\
\
\textbf{Remark 2.}. The generalized diagonal of a rhombus may possibly pass through the center of the rhombus on its way between rhombus vertices. In this case it \textit{still} gives a generalized diagonal for the triangle billiard, only of the smaller algebraic length.
\
\
Summarizing the arguments above, we have the following useful statement. Let $P_n^{\triangle}$ be a complexity function for a right triangle, and $P_n^{\diamondsuit}$ be a complexity function for a corresponding rhombus.
\begin{Lemma} $P_{3n}^{\triangle}\geq P_n^{\diamondsuit}$
\end{Lemma}
\
\
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-2.945)(10.46,0.965)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer](5.23,-0.995)(5.23,1.95)
\psline[linecolor=black, linewidth=0.03](0.02,-1.005)(10.4,-1.005)
\psline[linecolor=black, linewidth=0.03](5.22,-2.925)(5.22,0.955)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(1.38,-1.485)(5.66,0.755)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.4,arrowinset=0.0]{->}(5.64,0.755)(6.56,-2.405)
\end{pspicture}
}
\end{figure}
Fix a rhombus vertex and consider the set of trajectories emanating from the vertex through the fixed angular segment $I$, with a fixed algebraic length $n$, such that no trajectory hits another vertex up to $n$ reflections. Such a set is called \textit{angular} $I$-\textit{beam}. As angular $I$-beam does not hit a vertex( except its boundary rays) , the development up to $n$ reflections is the same for all the trajectories inside it.(see Fig.4)
\
\
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-5.916918)(13.9420595,1.8355953)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-84.83724](2.938032,-0.32323158)(1.19,2.95)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-40.73992](3.9232721,-2.2663481)(1.19,2.95)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-356.49136](6.0084157,-2.9724472)(1.19,2.95)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-312.136](7.9975796,-2.0427017)(1.19,2.95)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-356.0042](9.96758,-1.1072338)(1.19,2.95)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-39.876133](12.050726,-1.7501189)(1.19,2.95)
\psline[linecolor=blue, linewidth=0.04, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.04090305,-0.5539782)(8.75407,-1.1519033)(8.75407,-1.1519033)
\psline[linecolor=blue, linewidth=0.04, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.05434477,-0.57886434)(12.903196,-2.4951653)
\psline[linecolor=blue, linewidth=0.02, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.035180837,-0.5731421)(12.985305,-2.2900841)
\psline[linecolor=blue, linewidth=0.02, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.021739118,-0.548256)(13.065415,-2.021789)
\psline[linecolor=blue, linewidth=0.02, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.04090305,-0.5539782)(13.164691,-1.7592163)
\psline[linecolor=blue, linewidth=0.02, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.021739118,-0.548256)(13.263966,-1.4966435)
\end{pspicture}
}
\end{figure}
Fig. 4. Trajectories of the angular beam have the same development up to $n$ reflections. On the figure above $n=6$.
\
\
Similar notion is the \textit{parallel} $\mu$-\textit{beam}. Here we take a set of parallel trajectories of width $\mu$, emanating from some segment on the rhombus side and not hitting a vertex up to $n$-reflections. By length of the parallel $\mu$-beam we mean the length of its middle trajectory. As the parallel $\mu$-beam does not hit a vertex, all the trajectories of a beam have the same development up to $n$ reflections.
\
\
\section{Interval partitions}
We consider a rhombus, a fixed vertex and the corresponding angular segment located at the vertex, which we naturally associate with an interval $I = [0, 1]$ using the normalized angular distance on it. Points on the interval then correspond to directions of rays emanating from the vertex. We introduce a useful reduced quantity $Q_n$ as a number of generalized diagonals of algebraic length no greater than $n$ emanating from the fixed vertex .
Now let us create a decreasing sequence of finite indexed partitions $\zeta_n $ of $I$ on subintervals as follows. The intervals of $\zeta_n$ are formed by the points of $I$ corresponding to the generalized diagonals of algebraic length no greater than $n$ and the endpoints of $I$. We define \textsl{the index} of a cutting point of $\zeta_n$ as the algebraic length of the corresponding generalized diagonal.
Immediate observation is that $Q_n$ equals to the number $|\zeta_n|$ of cutting points of $\zeta_n$. By the diameter \textbf{diam}$(\zeta_n)$ we mean the maximal length of the division intervals.
\begin{Lemma} Let $\zeta_n$ be a sequence of indexed partitions of the interval $[0, 1]$ satisfying the following properties:
\
\
1) There exists $\gamma>0$ such that for all $n$ large enough $|\zeta_n|<n^{\gamma+1}$
\
\
2) There exists $c>0$ such that diam$(\zeta_n)< \frac{c}{n}$
\
\
Then there exist constants $P, K>0$ such that for all $n$ large enough there exists an interval $I\in \zeta_n$ such that:
\
\
1)$|I|>\frac{P}{n^{\gamma+1}}$
\
\
2) Left and right indices of $I$ are greater than $K n^{\frac{1}{1+\gamma}}$
\end{Lemma}
\begin{proof}
Assuming that $n$ is large enough we consider a division of the interval $[0, 1]$ on the intervals $I_1=[0, \frac{c}{n}]$, $I_2=[\frac{c}{n}, \frac{2c}{n}]$,$\ldots$, $I_k=[\frac{(k-1)c}{n},\frac{kc}{n}]$, $\ldots$,
$k\in\{1,[\frac{n}{c}]\}$ and one (possibly consisting of one point) short extra-interval
$[1-[\frac{n}{c}]\frac{c}{n}, 1]$.
By condition 2 we can pick some points $x_k\in \zeta_n\cap I_k$, $k\in\{1,[\frac{n}{c}]\}$ and form a finite set $F_n=\{x_1, x_2,\ldots, x_k, \ldots, x_{[\frac{n}{c}]}\}$ of cardinality $[\frac{n}{c}]$.
Then we choose a subset $G_n\subseteq F_n$, $G_n=\{x_1, x_3, x_5, x_7,\ldots\}$ which can be easily seen to satisfy the following two conditions for all $n$ large enough:
\
$1)$ The cardinality $|G_n|> \frac{n}{3c}$
\
$2)$ $\frac{c}{n}\leq Diam(G_n) \leq \frac{3c}{n}$.
\
Let us now consider the set of intervals $J_n=\{[x_1, x_3], [x_3, x_5], [x_5, x_7],\ldots \}$. The cardinality $|J_n|>\frac{n}{3c}$ for all $n$ large enough.
\
\
Let us now divide a set $J_n$ on two subsets $X$ and $Y$ characterized by the following properties:
\
\
$1)$ An interval $J=[a, b] \in J_n$ belongs to $X$ if the cardinality $\zeta_n\cap (a, b)< 6cn^{\gamma}$
\
$2)$ An interval $J=[a, b] \in J_n$ belongs to $Y$ if the cardinality $\zeta_n\cap (a, b)\geq 6cn^{\gamma}$
\
\
As the cardinality $|\zeta_n|<n^{\gamma+1}$ it implies that $|Y|6cn^{\gamma}<n^{\gamma+1}$ and so $|Y|<\frac{n}{6c}$. As $X=J_n-Y$ it implies that $|X|>\frac{n}{3c}-\frac{n}{6c}=\frac{n}{6c}$. As each interval $J$ from $X$ has less than $6cn^{\gamma}$ points of $\zeta_n$ it follows that inside $J$ there is an interval $I\in\zeta_n$ of length at least $\frac{1}{6n^{\gamma+1}}$. Let $Z$ be a set of such intervals $I$. Clearly $|Z|\geq\frac{n}{6c}$.
\
Let $R_n\subseteq\zeta_n$ be a set of points, whose indices are less than ${[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}$. Then, by definition, $R_n\subseteq {\zeta}_{{[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}}$ and so
the cardinality $|R_n|<|{\zeta}_{{[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}}|<{[{[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}]}^{\gamma+1}=\frac{n}{24c}$.
\
It follows that at least $|Z|-\frac{n}{24c}$ left ends of intervals from $Z$ have indices greater than ${[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}$ and the same is true for right ends. This implies that at least $|Z|-\frac{n}{12c}$ intervals from $Z$ have both left and right indices greater than ${[\frac{n}{24c}]}^{\frac{1}{\gamma+1}}$. As $|Z|>\frac{n}{6c}$ the conclusion of the lemma follows.
\end{proof}
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-5.539643)(10.022228,0.90750915)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-10.855984](4.7328897,-2.2885551)(1.44,3.25)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-59.48238](7.22244,-3.7842927)(1.44,3.25)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-322.41183](1.9824399,-2.9642928)(1.44,3.25)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.0676646,-5.156424)(4.287665,-5.156424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.0076647,-4.856424)(4.5076647,-4.856424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.8476648,-4.356424)(4.887665,-4.316424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.6676648,-3.516424)(5.5076647,-3.436424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.4876647,-2.676424)(6.1076646,-2.556424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.767665,-3.916424)(5.207665,-3.896424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.5876648,-3.0964239)(5.767665,-3.016424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.5876646,-5.216424)(5.287665,-5.196424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.847665,-4.876424)(5.5876646,-4.856424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(5.5676646,-3.8364239)(6.227665,-3.8364239)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(5.887665,-3.436424)(6.5276647,-3.436424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(6.1076646,-3.076424)(6.727665,-3.036424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(6.5076647,-2.576424)(7.207665,-2.556424)
\psline[linecolor=blue, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(5.187665,-4.416424)(5.9876647,-4.396424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.4476647,-2.476424)(6.0676646,-2.3764238)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.3276649,-2.016424)(5.9676647,-1.8964239)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(3.8476648,-1.2964239)(5.787665,-1.196424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.307665,-0.63642395)(5.687665,-0.5964239)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.707665,-0.01642395)(5.5076647,-0.01642395)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(4.9876647,0.34357604)(5.4676647,0.34357604)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.1676648,-0.99642396)(0.8676648,-0.976424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.2476648,-1.4364239)(1.0276648,-1.4364239)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.3876648,-1.956424)(1.1476648,-1.956424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.5276648,-2.476424)(1.3276649,-2.476424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.6276648,-3.016424)(1.3876648,-2.996424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.7476648,-3.436424)(1.5276648,-3.456424)
\psline[linecolor=red, linewidth=0.03, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(0.8276648,-3.796424)(1.5676647,-3.776424)
\psline[linecolor=red, linewidth=0.05](5.307665,0.88357604)(3.3476648,-2.036424)(3.4876647,-2.636424)
\psline[linecolor=blue, linewidth=0.05](3.4876647,-2.676424)(4.1276646,-5.4764237)
\end{pspicture}
}
\end{figure}
Fig.5. Blue part of $X_n$ represents those points of $X_n$ which are mapped to $X_{n+1}$ by the rhombus development map $\mathcal{R}$. Correspondingly red part of $X_n$ is mapped to $X_{n-1}$ by $\mathcal{R}$.
\section{Rhombus development map}
\
Our construction of rhombus development map is partially motivated by [3], where authors have shown the abundance of perpendicular periodic orbits for right triangular billiards. On the coordinate $xy$-plane with counterclockwise orientation we fix some rhombus $R_0$ and arbitrarily choose a pair of opposite vertices $A$ and $B$. Let $\alpha$ be a rhombus angle, adjacent to the vertex $A$. Let us denote $R_n$ the rhombus, obtained by rotation of $R_0$ to the angle $n\alpha$, $n\in\mathbb{Z}$. Assuming that $\alpha$ is irrational, all $R_n$ are different up to a parallel translation. On the boundary of $R_n$ we consider the set $X_n$ of all points $x$, such that the vector $v=(1, 0)$ located at $x$, points inside $R_n$. As the boundary of $R_n$ consists of two pairs of parallel sides, it is clear that $X_n$ is the union of two adjacent sides of $R_n$.
\
\
Let $X=\cup X_n$ and let us define a rhombus development map $\mathcal{R}: X\longrightarrow X$. Picking a point $x\in X_n$ we shoot a billiard trajectory from $x$ along the vector $v=(1, 0)$. Depending on where this trajectory hits the boundary of $R_n$ we get a point $\mathcal{R}(x)$ belonging to $X_{n-1}$ or $X_{n+1}$ up to a parallel translation. Clearly $\mathcal{R}$ is an invertible map, well defined on $X$ except the countable number of points which hit vertices. This countable number of points will make no difference for our considerations, so slightly abusing the notation we will assume that $\mathcal{R}$ is defined on the whole set $X$.(See Fig. 5.)
\
\
Let $\lambda_n$ be a measure on $X_n$ defined as the Lebesgue measure of projection to the vertical direction. Clearly all $\lambda_n$ then define a measure $\lambda$ on $X$ preserved by $\mathcal{R}$.
\
\
Let us now consider a rotation $R_{\alpha}: S^1\longrightarrow S^1$ on irrational angle $\alpha$ on the circle $S^1=\mathbb{R}/\mathbb{Z}$.
For $\mu\in (0, 0.5)$ we define a \textit{hitting function} $L(\mu)$ as follows. $L(\mu)$ is a minimal integer such that the finite trajectory
$\{x, R_{\alpha}x, \ldots, R^{L(\mu)}_{\alpha}x\}$ is $\mu$-dense, meaning that any interval of length $\mu$ has a point of this finite set inside.
\begin{Lemma} For a typical angle $\alpha\in [0, \pi]$ and any $\epsilon>0$ there is a constant $C=C(\alpha, \epsilon)>0$ such that $L(\mu)< \frac{C}{{\mu}^{2+\epsilon}}$.
\end{Lemma}
\begin{proof} Let $q_n$ be a sequence of continued fraction denominators for a number $\alpha$. By classical theorem of Khintchin, we have typically $\lim\limits_{n \to \infty} \frac{\ln(q_n)}{n}=\beta$. It implies that for $n$ large enough: $ e^{n(\beta-\epsilon)}<q_n< e^{n(\beta+\epsilon)}$. By well-known estimate : $\frac{1}{2q_{n+1}}<\frac{1}{q_{n+1}+q_n}< ||\alpha q_n||<\frac{1}{q_{n+1}}$. Now we look for an integer $n$, such that
$\frac{1}{q_{n+1}}<\mu$. Using the estimate above we have: $e^{(n+1)(\beta+\epsilon)}> q_{n+1}>\frac{1}{\mu}$ which implies that $n=\frac{\ln (\frac{1}{\mu})}{\beta-\epsilon}$ works. Applying $\alpha$-rotation $q_n$ times we have $\frac{1}{2q_{n+1}}< ||\alpha q_n||<\frac{1}{q_{n+1}}<\mu$ which in turn immediately implies that applying $\alpha$-rotation $2q_{n+1}q_n$ times to any point $x\in S^1$ we obtain a $\mu$-dense set.
\
From the considerations above we obtain a final estimate $L(\mu)< 2q_{n+1}q_n< 2e^{(n+1)(\beta+\epsilon)}e^{n(\beta+\epsilon)}<Ce^{2n(\beta+\epsilon)}
<Ce^{2\frac{\beta+\epsilon}{\beta-\epsilon}\ln(\frac{1}{\mu})}<\frac{C}{\mu^{2+\epsilon}}$
\end{proof}
\begin{Lemma} For a typical rhombus billiard and any $\epsilon>0$ there is a constant $C=C(\epsilon)>0$ such that any parallel $\mu$-beam of length greater than $T(\mu)=\frac{C}{{\mu}^{3+\epsilon}}$ has a parallel trajectory, contained inside.
\end{Lemma}
\
\
\textbf{Remark.} The parallel trajectory, contained inside the parallel $\mu$-beam should not necessarily be one of the beam trajectories ,nor should it have its ends on the beam sides. It literally means a segment,
geometrically contained inside the beam as a subset, which represents a periodic trajectory of a rhombus billiard.
\
\
\begin{proof}Taking a parallel $\mu$-beam we assume that it starts from some rhombus $R_0$. Let us call $I\in X_0$ an initial side of the beam. It follows that the development of $R_0$ along the beam produces a sequence $R_{a_n}$ where $\{a_n\}$ is a sequence of integer numbers with the property $|a_{n+1}-a_n|=1$. Let us denote $I_n={\mathcal{R}}^n(I)\in X_n$.
\
It follows that there exist positive integers $N_{-}$ and $N_{+}$ such that the gap for passing from $R_{N_{-}}$ to $R_{-N_{-}-1}$ has width less than $\mu$ and the gap for passing from $R_{N_{+}}$ to $N_{+}+1$ position has with less than $\mu$ (see fig.6). From lemma 4.1 it follows that $N_{-}$ and $N_{+}$ both satisfy
$N_{-},N_{+}<\frac{C}{\mu^{2+\epsilon}}$
\
\begin{figure}
\psscalebox{1.0 1.0}
{
\begin{pspicture}(0,-5.555309)(8.018796,3.8170772)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-57.04132](3.4626632,-1.5938071)(1.41,3.96)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-96.293755](4.082663,-4.153807)(1.41,3.96)
\psline[linecolor=red, linewidth=0.04](0.99266326,-2.6738071)(1.0526633,-2.593807)
\psline[linecolor=red, linewidth=0.06](0.9726633,-2.6738071)(2.6926632,-0.43380708)
\psline[linecolor=red, linewidth=0.06](2.6926632,-0.43380708)(6.732663,0.5261929)
\psline[linecolor=blue, linewidth=0.06](0.15266328,-3.6938071)(0.93266326,-2.6938071)
\psdiamond[linecolor=black, linewidth=0.03, dimen=outer, gangle=-17.778597](1.3426633,0.046192933)(1.41,3.96)
\psline[linecolor=black, linewidth=0.06](0.95266324,-2.6938071)(4.232663,-2.733807)
\end{pspicture}
}
\end{figure}
Fig. 6. Blue part of $X_{N_{+}}$ is mapped to $X_{N_{+}+1}$ by $\mathcal{R}$ and the red part of $X_{N_{+}}$ is mapped to $X_{N_{+}-1}$ by $\mathcal{R}$. So no parallel $\mu$-beam can pass to $X_{N_{+}+1}$ without splitting if the vertical width of the blue gap is less than $\mu$.
\
\
But as the union $\bigcup\limits_{n=-N_{-}}^{N_{+}} X_n$ has a measure $\lambda(\bigcup\limits_{n=-N_{-}}^{N_{+}} X_n)< N_{+}-N_{-}<\frac{2C}{\mu^{2+\epsilon}}$ and as $\mathcal{R}$ preserves $\lambda$ and $\lambda(I)=\mu$ it follows that there exist
$p, q: 1\leq p\neq q\leq \frac{2C}{\mu^{3+\epsilon}} $, such that $I_p\cap I_q\neq \varnothing$. Choosing a point $x\in I_n\cap I_m$ and connecting $x\in R_{a_n}$ to $x\in R_{a_m}$ we get a periodic trajectory of length no greater than $\frac{2C}{\mu^{3+\epsilon}}$ inside $\mu$-beam (see fig. 7).
\
\
What is important for us is that such a periodic trajectory necessarily has an \textit{even} period. Indeed assume that development from $R_{a_p}$ to $R_{a_q}$ has $u$ clockwise and $v$ counterclockwise $\alpha$-rotations. But then the rhombus $R_{a_q}$ would be obtained from $R_{a_p}$ by rotation to the angle $(u-v)\alpha$ up to a parallel translation. As the rhombi $R_{a_p}$ and $R_{a_q}$ are parallel and $\alpha$ is irrational, it implies that $u=v$, and so the trajectory has an even period.
\
\end{proof}
\
\
\begin{figure}
\psscalebox{0.8 0.8}
{
\begin{pspicture}(0,-6.358333)(13.25,3.918333)
\definecolor{colour0}{rgb}{0.8,0.2,0.2}
\psdiamond[linecolor=black, linewidth=0.04, dimen=outer](1.46,-2.101667)(1.46,4.24)
\psdiamond[linecolor=black, linewidth=0.04, dimen=outer, gangle=38.0](4.04,-1.2216668)(1.46,4.24)
\psdiamond[linecolor=black, linewidth=0.04, dimen=outer](6.62,-0.32166687)(1.46,4.24)
\psdiamond[linecolor=black, linewidth=0.04, dimen=outer, gangle=-38.0](9.2,-1.2216668)(1.46,4.24)
\psdiamond[linecolor=black, linewidth=0.04, dimen=outer](11.79,-2.1183336)(1.46,4.24)
\psline[linecolor=black, linewidth=0.04](10.32,-2.061667)(0.26,-1.3816669)
\psline[linecolor=black, linewidth=0.04](10.8,-0.7816669)(0.74,-0.10166687)
\psline[linecolor=colour0, linewidth=0.1](0.26,-1.3616669)(0.7,-0.10166687)
\psline[linecolor=colour0, linewidth=0.1](10.32,-2.041667)(10.78,-0.7816669)
\psline[linecolor=colour0, linewidth=0.1](10.34,-2.061667)(10.8,-0.80166686)
\psline[linecolor=blue, linewidth=0.1](0.3,-1.1816669)(10.7,-1.1416669)
\end{pspicture}
}
\end{figure}
Fig.7. Periodic orbit is inside the $\mu$-beam connecting parallel rhombi $R_{a_p}$ and $R_{a_q}$ as soon as $I_p\cap I_q\neq \varnothing$. ($I_p$ is a left red segment and $I_q$ is a right red segment.)
\
\
\begin{thm} For a typical right triangular billiard and for any $\epsilon>0$ the complexity function $P_n$ satisfies: $\limsup \frac{P_n}{n^{\frac{2}{\sqrt{3}}-\epsilon}}>0$
\end{thm}
\begin{proof} Let us assume that there are positive constants $\gamma<\frac{2}{\sqrt{3}}-1$ and $C$ such that $P_n<Cn^{1+\gamma}$ for all $n$. By reflecting the triangle about the right angle we make a rhombus and by Lemma 2.2 we may assume that the rhombus complexity function also satisfies $P_n<Cn^{1+\gamma}$ ( with different constant $C$). Then we choose any vertex of the rhombus and consider the reduced complexity function for the vertex, which also then satisfies $P_n<Cn^{1+\gamma}$. Consider the sequence of associated indexed partitions for the vertex. Fix $n$ large enough.
\
Then by Lemma 3.1 there is an angular segment $I$ and positive constants $P, K$ ( both independent on $n$) such that:
\
\
$1)$ $|I|> Pn^{-\gamma-1}$
\
\
$2)$ The indices of left and right ends of $I$ are greater than $Kn^{\frac{1}{\gamma+1}}$.
\
\
Taking angular beam corresponding to the segment $I$ we see that the length of the large angular segment, corresponding to the both left and right indices of $I$ has length greater than
$cn^{\frac{1}{\gamma+1}-\gamma-1}$ for some constant $c$ independent on $n$.
\
\
Fig.8. Depending on the direction of the periodic trajectory inside the $\mu$-beam, one can drag it either to the original vertex or to one of the two boundary vertices of the interval $I$. As the dragged orbit is also periodic, it implies the existence of a singular point inside the angular beam, which is absurd.
\
\
Let us pick $\mu=\frac{c}{2}n^{\frac{1}{\gamma+1}-\gamma-1}$ and $\epsilon$ small enough. We are interested to see if we can find a parallel $\mu$-beam of length $T(\mu)=\frac{C}{{\mu}^{3+\epsilon}}$ inside the angular $I$-beam up to the lower index of $I$ for $n$ large enough. Taking into account that the length of the angular $I$-beam is greater then
$n^{\frac{1}{\gamma+1}}$ we get the sufficient condition:
\
\
\begin{figure}
\psscalebox{0.6 0.6}
{
\begin{pspicture}(0,-8.329899)(23.558989,6.0282817)
\psline[linecolor=black, linewidth=0.04](0.21,1.9190906)(14.37,1.8990905)
\psline[linecolor=black, linewidth=0.04](0.19,1.9390906)(14.03,5.9990907)
\psline[linecolor=black, linewidth=0.04](5.03,1.9390906)(5.91,2.6990905)(8.07,2.6990905)(7.07,1.9190906)
\psline[linecolor=red, linewidth=0.04](5.15,2.0790906)(7.61,2.3590906)
\psline[linecolor=red, linewidth=0.04](0.17,1.9190906)(2.63,2.1990905)
\psline[linecolor=blue, linewidth=0.04, linestyle=dashed, dash=0.17638889cm 0.10583334cm, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(6.45,2.3390906)(2.27,2.0990906)
\psline[linecolor=black, linewidth=0.04](0.03,-3.4009094)(15.33,0.91909057)
\psline[linecolor=black, linewidth=0.04](0.03,-3.4009094)(15.61,-3.3809094)
\psline[linecolor=black, linewidth=0.04](10.35,-3.3609095)(9.97,-1.8409095)(13.43,-1.8009094)
\psline[linecolor=black, linewidth=0.04](13.41,-1.7609094)(13.93,-3.3409095)
\psline[linecolor=red, linewidth=0.04](10.31,-3.3009095)(13.43,-1.8609095)
\psline[linecolor=red, linewidth=0.04](12.21,-0.5209094)(15.33,0.91909057)
\psline[linecolor=blue, linewidth=0.04, linestyle=dashed, dash=0.17638889cm 0.10583334cm, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(11.57,-2.6409094)(13.83,0.13909058)
\psdots[linecolor=black, dotsize=0.14](15.31,0.85909057)
\psline[linecolor=black, linewidth=0.04](15.29,0.91909057)(15.63,-3.3609095)
\psdots[linecolor=black, dotsize=0.14](12.21,-0.5209094)
\psdots[linecolor=black, dotsize=0.14](23.49,-2.8209095)
\psdots[linecolor=black, dotsize=0.14](0.17,1.9190906)
\psdots[linecolor=black, dotsize=0.14](2.61,2.2190905)
\psline[linecolor=black, linewidth=0.04](0.01,-8.28091)(15.07,-8.28091)
\psline[linecolor=black, linewidth=0.04](0.03,-8.24091)(14.37,-5.8409095)
\psline[linecolor=black, linewidth=0.04](7.79,-8.260909)(7.99,-7.620909)(10.79,-7.600909)
\psline[linecolor=black, linewidth=0.04](10.77,-7.5809093)(10.61,-8.28091)
\psline[linecolor=red, linewidth=0.04](7.97,-7.6609097)(10.63,-8.16091)
\psdots[linecolor=black, dotsize=0.14](15.05,-8.260909)
\psline[linecolor=black, linewidth=0.04](14.35,-5.8209095)(15.05,-8.220909)
\psline[linecolor=red, linewidth=0.04](12.39,-7.7409096)(15.05,-8.24091)
\psdots[linecolor=black, dotsize=0.14](12.39,-7.7409096)
\psline[linecolor=blue, linewidth=0.04, linestyle=dashed, dash=0.17638889cm 0.10583334cm, arrowsize=0.05291667cm 2.0,arrowlength=1.42,arrowinset=0.0]{->}(8.89,-7.7409096)(13.29,-8.00091)
\end{pspicture}
}
\end{figure}
\
\
$(-3-\epsilon)(\frac{1}{\gamma+1}-\gamma-1)<\frac{1}{\gamma+1}$.
\
\
Looking for maximal $\gamma$ satisfying inequality above we take a limit case $\epsilon=0$ and get equation: $3(\gamma+1-\frac{1}{\gamma+1})=\frac{1}{\gamma+1}$ which gives a critical value
${\gamma}_{crit}=\frac{2}{\sqrt{3}}-1$. As by theorem conditions $\gamma<{\gamma}_{crit}$ then picking $\epsilon$ small enough we will be able to find a parellel $\mu$-beam of length $L(\mu)$ and so to satisfy the condition of the Lemma 4.2.
\
\
Then by Lemma 4.2 we are able to find a periodic trajectory $T$ of even period inside the parallel $\mu$-beam and so inside an angular $I$-beam.
\
\
Let $\alpha\in(-\frac{\pi}{2},\frac{\pi}{2})$ be an angle which $T$ forms with horizontal direction, and $\beta$ be the angle for $I$-beam. We have three cases:
\
\
$1)$ $\alpha\in(-\frac{\pi}{2}, 0)$
\
\
$2)$ $\alpha\in[0, \beta]$
\
\
$3)$ $\alpha\in(\beta, \frac{\pi}{2})$
\
\
Depending on the case we are able to drag the trajectory $T$ inside the angular $I$-beam or to the left cut-point of $I$ or to the right cut-point of $I$ or to the vertex of $I$. By Lemma 2.1 the dragged trajectory is also periodic, as periodic rhombus trajectories of even period correspond to the periodic trajectories on the associated flat sphere (See Remark 1.). But that would give a contradiction because in each of the three cases we would get a new vertex inside the angular $I$ segment, as the dragged periodic trajectory starts from a vertex and so should end in a vertex.
\end{proof}
\
\newpage
|
\section{Introduction}
\renewcommand{\theequation}{\arabic{section}.\arabic{equation}}
\setcounter{equation}{0}
The one-dimensional compressible Navier-Stokes system in Lagrangian coordinates read
\begin{equation}\label{ns}
\left\{
\begin{array}{ll}
\di v_t-u_x=0,\\
\di u_t+p_x=\mu\left(\frac{u_x}{v}\right)_x,\\
\di \left(e+\frac{u^2}{2}\right)_t+\left(p u\right)_x=\left(\kappa \frac{\t_x}{v}+\mu\frac{uu_x}{v}\right)_x
\end{array}
\right.
\end{equation}
for $x\in\mathbb{R}=(-\infty,+\infty)$, $t>0$, where $v(x,t)>0$, $u(x,t)$, $\t(x,t)>0$, $e(x,t)>0$ and $p(x,t)$
are the specific volume, fluid velocity, absolute temperature, internal energy
and pressure, respectively, while the positive constants $\mu$ and $\k$ denote
the viscosity and heat conduction coefficients, respectively. Here we study the ideal
polytropic fluids so that $p$ and $e$ are given by the state equations
\begin{equation*}
\di p=\frac{R\t}{v}=Av^{-\gamma}\exp\left(\frac{\gamma-1}{R}s\right),\quad e=c_{\nu}\t+\mathrm{const}.,
\end{equation*}
where $s$ is the entropy, $\gamma>1$ is the adiabatic exponent, $c_{\nu}=\frac{R}{\gamma-1}$ is
the specific heat, and $A$ and $R$ are both positive constants.
We consider the Cauchy problem to the system \eqref{ns} supplement with the following initial
and far field conditions:
\begin{equation}\label{initial}
\left\{
\begin{array}{ll}
\di (v,u,\t)(x,0)=(v_0,u_0,\t_0)(x), &\di x\in \mathbb{R},\\
\di (v,u,\t)(\pm\infty,t)=(v_{\pm},u_{\pm},\t_{\pm}), &\di t>0,
\end{array}
\right.
\end{equation}
where $v_{\pm}(>0)$, $u_{\pm}$ and $\t_{\pm}(>0)$ are given constants, and we assume
$\inf_{\mathbb{R}}v_0>0$, $\inf_{\mathbb{R}}\t_0>0$, and $(v_0,u_0,\t_0)(\pm\infty)=(v_{\pm},u_{\pm},\t_{\pm})$
as compatibility conditions. When the far field states are the same, i.e., $v_+=v_-$, $u_+=u_-$, $\t_+=\t_-$,
there has been considerable progress on the global existence of the solutions to the system \eqref{ns} since 1977,
see \cite{jiang,jiang-2,kazhikhov,kazhi-she,L-L} and the reference therein. In particular, Jiang \cite{jiang, jiang-2} first obtained some
interesting results on the large-time behavior of solutions, however the temperature is only shown to be locally bounded in space. More recently, Li and Liang \cite{L-L} improved Jiang's results by proving the temperature is uniformly bounded.
The existence and large time behavior of solutions to the system \eqref{ns} with different end states
become much more complicated. It is noted that, if the dissipation effects are neglected, i.e.,
$\mu=\k=0$, the system \eqref{ns} is reduced to the compressible Euler equations as follows
\begin{equation}\label{euler}
\left\{
\begin{array}{ll}
\di v_t-u_x=0,\\
\di u_t+p_x=0,\\
\di \left(e+\frac{u^2}{2}\right)_t+\left(p u\right)_x=0,
\end{array}
\right.
\end{equation}
which is the most important hyperbolic system of conservation laws. It is well known that
the system \eqref{euler}
has rich wave phenomena. Indeed, it contains three basic wave patterns (see \cite{Smoller}), two nonlinear waves:
shock and rarefaction wave, and a linearly degenerate wave: contact discontinuity. When we consider the Riemann initial data
\begin{equation}\label{Riemann}
(v,u,\t)(x,0)=\left\{
\begin{array}{ll}
\di (v_-,u_-,\t_-),&\di x<0,\\
\di (v_+,u_+,\t_+),&\di x>0,
\end{array}
\right.
\end{equation}
the solutions consist of the above three wave patterns
and their superpositions, called by Riemann solutions, and govern both the local and large time
asymptotic behavior of general solutions of the system \eqref{euler}. It is of great importance and interest
to study the large-time behavior of the viscous version of these basic wave patterns and their superpositions
to the compressible Navier-Stokes system \eqref{ns}.
There has been extensive literature on the stability analysis of viscous wave pattern to system \eqref{ns},
meanwhile new phenomena has been discovered and new techniques have been developed. We refer to
\cite{Huang-matsumura,KM,MN-85} for the shock wave, \cite{MN-86,MN-92,nishi-yang-zhao} for the rarefaction wave,
\cite{huang-ma-shi,Huang-Matsumura-Xin,Huang-Xin-Yang,Huang-Yang,huang-zhao} for the viscous contact discontinuity, and the reference therein.
However, the stability of the superposition of several wave patterns is more complicated and challenging due to
the fact that the stability analysis essentially depends on the underlying properties of basic wave pattern and these frameworks are not compatible with each other. Besides, the wave interaction between different families of wave patterns is complicated. Recently, Huang-Matsumura in \cite{Huang-matsumura} showed that the superposition of the two viscous shock profiles
for the Navier-Stokes system \eqref{ns} is asymptotically stable without the zero initial mass condition.
This result was extended in \cite{Huang-Li-matsumura} to the combination of
viscous contact discontinuity with rarefaction waves by deriving new estimates on the heat kernel.
The time-asymptotic stability of other cases is still open!
It is noted that all results mentioned above are concerned with the small perturbation around the viscous
wave pattern. In another word, they are ``local" stability. A nature problem is, whether or not these basic wave patterns and their linear superpositions
are stable even for large perturbation. This is equivalent to study the global stability of the viscous
wave patterns to the system \eqref{ns}, which is challenging because the nonlinear terms play leading role in the large solutions, while the linearized system around wave patterns is essential for the local stability. Along this direction, Nishihara-Yang-Zhao in \cite{nishi-yang-zhao} first proved the
rarefaction waves for the system \eqref{ns} are stable with ``partially" large perturbation with the condition that the adiabatic exponent $\gamma$ is
closing enough to ``1". Precisely speaking, the amplitude of initial perturbation is reciprocal to $\gamma-1$. That is, $\gamma-1$ is smaller, the perturbation around the rarefaction wave can be larger. This result is extended by Huang-Zhao in \cite{huang-zhao} to the case of single viscous contact wave and the combination of viscous contact wave and rarefaction
waves for a free boundary value problem, and further by Hong in \cite{Hong} for the Cauchy problem.
Note that the condition on $\gamma$ plays essential role in \cite{Hong,huang-zhao,nishi-yang-zhao}, however, it is not natural in the physical setting. The main aim of this paper is to remove the non-physical condition of the adiabatic exponent $\gamma$.
\
Before stating the main results, we first recall the viscous contact wave $(V,U,\Theta)$
for the compressible Navier-Stokes system \eqref{ns} introduced in \cite{Huang-Xin-Yang}. For the Riemann
problem \eqref{euler}-\eqref{Riemann}, it is known that the contact discontinuity solution takes the form
\begin{equation}
(\tilde V,\tilde U,\tilde\Theta)(x,t)=\left\{
\begin{array}{ll}
\di (v_-,u_-,\t_-),&\di x<0,t>0,\\
\di (v_+,u_+,\t_+),&\di x>0,t>0,
\end{array}
\right.
\end{equation}
provide that
\begin{equation}\label{RH}
\di u_-=u_+,\quad p_-\triangleq\frac{R\theta_-}{v_-}=p_+\triangleq\frac{R\t_+}{v_+}.
\end{equation}
We assume that $u_-=u_+=0$ without loss of generality.
Due to the effect of heat conductivity, the contact discontinuity $(\tilde V,\tilde U,\tilde \Theta)$
is smoothed and behaves as a diffusion wave, called by "viscous contact wave". The viscous contact wave $(V,U,\Theta)$ can be constructed
as follows. Since the pressure for the profile $(V,U,\Theta)$ is expected to be
constant asymptotically, we set
$$
\frac{R\Theta}{V}=p_+,
$$
which indicates the leading part of the energy equation $\eqref{ns}_3$ is
\begin{equation}\label{Theta t}
\di c_{\nu}\Theta_t+p_+U_x=\k\left(\frac{\Theta_x}{V}\right)_x.
\end{equation}
The equation \eqref{Theta t} and $\eqref{ns}_1$ lead to a nonlinear diffusion equation,
\begin{equation}\label{1.8}
\di \Theta_t=a\left(\frac{\Theta_x}{\Theta}\right)_x,\quad
\Theta(\pm\infty,t)=\theta_{\pm},\quad a=\frac{\kappa p_+(\gamma-1)}{\gamma R^2}>0,
\end{equation}
which has a unique self-similar solution $\Theta(x,t)=\Theta(\xi)$, $\xi=\frac{x}{\sqrt{1+t}}$
due to \cite{xiao-liu}. Furthermore, $\Theta(\xi)$ is a monotone function,
increasing if $\theta_+>\theta_-$ and decreasing if $\theta_+<\theta_-$. On the other hand,
there exists some positive constant $\delta$, such that for $\d=|\t_+ -\t_-|$, $\Theta$ satisfies
\begin{equation}\label{1.9}
\di (1+t)|\Theta_{xx}|+(1+t)^{\frac{1}{2}}|\Theta_x|+|\Theta-\theta_{\pm}|
=O(1)\delta e^{-\frac{c_1x^2}{1+t}} \quad \mathrm{as}~ |x|\rightarrow\infty,
\end{equation}
where $c_1$ is positive constant depending only on $\t_{\pm}$. Once $\Theta$ is determined,
the contact wave profile $(V,U,\Theta)(x,t)$ is then defined as follows:
\begin{equation}\label{contact}
\di V=\frac{R}{p_+}\Theta,\quad U=\frac{\k(\gamma-1)}{\gamma R}\frac{\Theta_x}{\Theta},
\quad \Theta=\Theta.
\end{equation}
The contact wave $(V, U,\Theta)(x,t)$ solves the compressible Navier-Stokes system \eqref{ns}
time asymptotically, that is,
\begin{equation}\label{contact equation}
\left\{
\begin{array}{ll}
\di V_t-U_x=0,\\
\di U_t+\left(\frac{R\Theta}{V}\right)_x=\mu\left(\frac{U_x}{V}\right)_x+R_1,\\
\di c_{\nu}\Theta_t+p(V,\Theta)U_x=\left(\kappa \frac{\Theta_x}{V}\right)_x+\mu\frac{U^2_x}{V}+R_2,
\end{array}
\right.
\end{equation}
where
\begin{equation}
\di \widetilde R_1=U_t-\mu\left(\frac{U_x}{V}\right)_x,\quad \widetilde R_2=-\mu\frac{U_x^2}{V}.
\end{equation}
We first study the global stability of single viscous contact wave $(V, U,\Theta)$ for arbitrary $\gamma>1$. For this, we put the perturbation $(\phi,\psi,\zeta)(x,t)$ by
\begin{equation}
(\phi,\psi,\zeta)(x,t)=(v-V,u-U,\t-\Theta)(x,t).
\end{equation}
The precise statement of the first result is
\begin{theorem}{\bf (Viscous contact wave)}\label{theorem1}
For any given left end state $(v_-,u_-,\theta_-)$, suppose that the right end state $(v_+,u_+,\t_+)$ satisfies \eqref{RH}.
Let $(V,U,\Theta)$ be the viscous contact wave defined in \eqref{contact} with strength
$\d=|\t_+-\t_-|$. There exist a function $m(\d)$ satisfying $m(\d)\to +\infty$, as $\d\to 0$ and a small constant $\delta_0$ such that if $\d<\d_0$ and the initial data satisfies
\begin{equation}\label{data}
\left\{
\begin{array}{ll}
\di v_0(x), \t_0(x)\geq m_0^{-1},\quad m_0=:m(\d_0),\\[0.2cm]
\|(v_0(x)-V(x,0),u_0(x)-U(x,0),\t_0(x)-\Theta(x,0))\|_{H^1(\mathbb{R})}\le m_0,
\end{array}
\right.
\end{equation}
then the Cauchy problem \eqref{ns}-\eqref{initial} admits
a unique global solution $(v,u,\t)$ satisfying
$$
(v-V,u-U,\theta-\Theta)(x,t)\in C\big((0,+\infty);H^1(\mathbb{R})\big);
$$
$$
(v-V)_x(x,t)\in L^2(0,+\infty;L^2\big(\mathbb{R})\big);
$$
$$
(u-U,\t-\Theta)_x(x,t)\in L^2\big(0,+\infty;H^1(\mathbb{R})\big).
$$
Furthermore,
\begin{equation}\label{behavior-1}
\lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}}\left|(v-V,u-U,\t-\Theta)(x,t)\right|=0.
\end{equation}
\end{theorem}
\begin{remark}
Theorem 1.1 means that if the strength of contact wave is smaller, the initial perturbation can be larger. In particular, when $\d=0$, that is, the asymptotic state is a constant one $(\bar{v},\bar{u},\bar{\t})$ instead of a wave pattern, then for any initial data $(v_0-\bar{v},u_0-\bar{u},\t_0-\bar{\t})(x)\in H^1(\mathbb{R})$, there always exists a small constant $\delta_0$ such that
\eqref{data} holds, in which the second term is replaced by $\|(v_0(x)-\bar{v},u_0(x)-\bar{u},\t_0(x)-\bar{\t})\|_{H^1(\mathbb{R})}\le m_0$. This coincides with the one in Li-Liang \cite{L-L}.
\end{remark}
\begin{remark}
Theorem 1.1 holds for any $\gamma>1$ and thus removes the condition that $\gamma$ is close to 1 in
Nishihara-Yang-Zhao \cite{nishi-yang-zhao} and also in \cite{Hong,huang-zhao}.
\end{remark}
When the relation \eqref{RH} fails, the basic theory of hyperbolic systems of conservation laws
implies that for any given constant state $(v_-, u_-, \t_-)$ with $v_->0$, $\t_->0$ and $u_-\in\mathbb{R}$,
there exists a suitable neighborhood $\Omega(v_-,u_-,\t_-)$ of $(v_-, u_-, \t_-)$ such that for any
$(v_+, u_+, \t_+)\in\Omega(v_-,u_-,\t_-)$, the Riemann problem of the Euler system \eqref{euler}, \eqref{Riemann}
has a unique solution. In this paper, we only consider the case of the superposition of the viscous
contact wave and rarefaction waves with
\begin{equation}\label{zheng}
\di (v_+,u_+,\t_+)\in R_1CR_3(v_-,u_-,\t_-)\subset\Omega(v_-,u_-,\t_-),
\end{equation}
where
\begin{equation*}
\begin{array}{ll}
\di R_1CR_3(v_-,u_-,\t_-)\triangleq\Bigg\{(v,u,\t)\in\Omega(v_-,u_-,\t_-)\Bigg|s\neq s_-,\\
\di u\geq u_--\int_{v_-}^{e^{\frac{\g-1}{R\g}(s_--s)}v}\lambda_-(\eta,s_-)d\eta,
u\geq u_--\int^v_{e^{\frac{\g-1}{R\g}(s_--s)}v_-}\lambda_+(\eta,s)d\eta\Bigg\}
\end{array}
\end{equation*}
and
$$
s=\frac{R}{\g-1}\ln\frac{R\t}{A}+R\ln v,\quad s_{\pm}=\frac{R}{\g-1}\ln\frac{R\t_{\pm}}{A}+R\ln v_{\pm},~~
\lambda_{\pm}(v,s)=\pm\sqrt{A\g v^{-\g-1}e^{\frac{\g-1}{R}s}}.
$$
By the standard argument (e.g. \cite{Smoller}), there exists a unique pair of points $(v_-^m, u^m, \t_-^m)$
and $(v_+^m, u^m, \t_+^m)$ in $\Omega(v_-,u_-,\t_-)$ satisfying
$$
\frac{R\t_-^m}{v_-^m}=\frac{R\t_+^m}{v_+^m}\triangleq p^m,
$$
the points $(v_-^m,u^m,\t_-^m)$ and $(v_+^m,u^m,\t_+^m)$ belong to the 1-rarefaction wave
curve $R_-(v_-,u_-,\t_-)$ and the 3-rarefaction wave curve $R_+(v_+,u_+,\t_+)$, respectively, where
$$
R_{\pm}(v_{\pm},u_{\pm},\t_{\pm})=\left\{(v,u,\t)\Bigg|s=s_{\pm},
u=u_{\pm}-\int_{v_{\pm}}^v\lambda_{\pm}(\eta,s_{\pm})d\eta,v>v_{\pm}
\right\}.
$$
Without loss of generality, we assume $u^m=0$ in what follows.
The 1-rarefaction wave $(v_-^r,u_-^r,\t_-^r)(\frac{x}{t})$ (respectively the 3-rarefaction wave $(v_+^r,u_+^r,\t_+^r)(\frac{x}{t})$)
connecting $(v_-,u_-,\t_-)$ and $(v_-^m,0,\t_-^m)$ (respectively $(v_+^m,0,\t_+^m)$ and $(v_+,u_+,\t_+)$) is the
weak solution of the Riemann problem of the Euler system \eqref{euler} with the following initial Riemann data
\begin{equation}\label{Riemann2}
\di(v_{\pm},u_{\pm},\t_{\pm})(x,0)=\left\{
\begin{array}{ll}
(v_{\pm}^m,0,\t_{\pm}^m), &\quad\pm x<0,\\
(v_{\pm},u_{\pm},\t_{\pm}), &\quad\pm x>0.
\end{array}
\right.
\end{equation}
Since the rarefaction wave $(v_{\pm}^r,u_{\pm}^r,\t_{\pm}^r)$ are weak solutions, it is
convenient to construct approximate rarefaction wave which is smooth. Motivated by \cite{MN-86}, the smooth solutions of Euler
system \eqref{euler}, $(V_{\pm}^r,U_{\pm}^r,\T_{\pm}^r)$, which approximate $(v_{\pm}^r,u_{\pm}^r,\t_{\pm}^r)$,
are given by
\begin{equation}\label{appro-rare}
\di\left\{
\begin{array}{ll}
\di\lambda_{\pm}(V_{\pm}^r(x,t),s_{\pm})=w_{\pm}(x,t),\\
\di U_{\pm}^r=u_{\pm}-\int_{v_{\pm}}^{V_{\pm}^r(x,t)}\lambda_{\pm}(\eta,s_{\pm})d\eta,\\
\T_{\pm}^r=\t_{\pm}(v_{\pm})^{\g-1}(V_{\pm}^r)^{1-\g},
\end{array}
\right.
\end{equation}
where $w_-$ (respectively $w_+$) is the solution of the initial problem for the typical Burgers equation:
\begin{equation}\label{burgers}
\di\left\{
\begin{array}{ll}
\di w_t+ww_x=0,\quad(x,t)\in\mathbb{R}\times(0,\infty),\\
\di w(x,0)=\frac{w_r+w_l}{2}+\frac{w_r-w_l}{2}\tanh x,
\end{array}
\right.
\end{equation}
with $w_l=\lambda_-(v_-,s_-)$, $w_r=\lambda_-(v_-^m,s_-)$ (respectively $w_l=\lambda_+(v_+^m,s_+)$, $w_r=\lambda_+(v_+,s_+)$).
Let $(V^{cd},U^{cd},\T^{cd})(x,t)$ be the viscous contact wave constructed in \eqref{1.8} and \eqref{contact}
with $(v_{\pm},u_{\pm},\t_{\pm})$ replaced by $(v_{\pm}^m,0,\t_{\pm}^m)$, respectively.
To describe the strengths of the viscous contact wave and rarefaction waves for later use, we set
\begin{equation*}
\begin{array}{ll}
\d^{r_1}=|v_-^m-v_-|+|0-u_-|+|\t_-^m-\t_-|, \quad \d^{cd}=|\t_+^m-\t_-^m|,\\
\d^{r_3}=|v_+^m-v_+|+|0-u_+|+|\t_+^m-\t_+|
\end{array}
\end{equation*}
and $\d=\min(\d^{r_1},\d^{cd},\d^{r_3})$. If
\begin{equation}\label{same-order}
\d^{r_1}+\d^{cd}+\d^{r_3}\leq C\d,\quad {\rm as }\quad \d^{r_1}+\d^{cd}+\d^{r_3}\rightarrow 0
\end{equation}
holds for a positive constant $C$, we call the strengths of the wave patterns ``small with the same order". In this case, we have
\begin{equation}
\d^{r_1}+\d^{cd}+\d^{r_3}\leq C|(v_+-v_-,u_+-u_-,\t_+-\t_-,)|.
\end{equation}
In what follows, we always assume \eqref{same-order}. We define
\begin{equation}\label{ansatz}
\left(
\begin{array}{l}
\di V\\
\di U \\
\di \T
\end{array}
\right)(x,t) = \left(
\begin{array}{l}
\di V^{cd}+V_-^r+V_+^r\\
\di U^{cd}+U_-^r+U_+^r\\
\di \T^{cd}+\T_-^r+\T_+^r
\end{array}\right)(x,t)-\left(
\begin{array}{l}
\di v_-^m+v_+^m\\
\di \quad~~ 0\\
\di \t_-^m+\t_+^m
\end{array}\right),
\end{equation}
and
$$
(\phi,\psi,\zeta)(x,t)=(v-V,u-U,\t-\T)(x,t).
$$
The precise statement of the second result is
\begin{theorem}{\bf (Composite waves)}\label{theorem2}
For any given left end state $(v_-,u_-,\theta_-)$, let $(V,U,\Theta)$ be defined in \eqref{ansatz} with strength
satisfying \eqref{same-order}. Then there exists a function $m(\d)$ satisfying $m(\d)\to +\infty$,
as $\d\to 0$ and a small constant $\delta_0$ , such that if $|(v_+-v_-,u_+-u_-,\t_+-\t_-)|<\d_0$ and the initial data satisfies
\begin{equation}\label{data2}
\left\{
\begin{array}{ll}
\di v_0(x), \t_0(x)\geq m_0^{-1},\quad m_0=:m(\d_0),\\[0.2cm]
\|(v_0(x)-V(x,0),u_0(x)-U(x,0),\t_0(x)-\Theta(x,0))\|_{H^1(\mathbb{R})}\le m_0,
\end{array}
\right.
\end{equation}
then the Cauchy problem \eqref{ns}-\eqref{initial} admits
a unique global solution $(v,u,\t)$ satisfying
$$
(v-V,u-U,\theta-\Theta)(x,t)\in C\big((0,+\infty);H^1(\mathbb{R})\big);
$$
$$
(v-V)_x(x,t)\in L^2(0,+\infty;L^2\big(\mathbb{R})\big);
$$
$$
(u-U,\t-\Theta)_x(x,t)\in L^2\big(0,+\infty;H^1(\mathbb{R})\big),
$$
and
\begin{equation}\label{wt-2}
\lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}}\left|(v-V,u-U,\t-\Theta)(x,t)\right|=0,
\end{equation}
where the $(v_-^r,u_-^r,\t_-^r)(x,t)$ and $(v_+^r,u_+^r,\t_+^r)(x,t)$ are the 1-rarefaction and
3-rarefaction waves uniquely determined by \eqref{euler}, \eqref{Riemann2}, respectively.
\end{theorem}
\begin{remark}
By $(iv)$ of Lemma \ref{rare-pro}, Theorem \ref{theorem2} implies
\begin{equation*}\label{behavior-2}
\lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}}
\left(
\begin{array}{l}
|(v-v_-^r-V^{cd}-v_+^r+v_-^m+v_+^m)(x,t)|\\[2mm]
\quad\quad|(u-U^{cd}-u_-^r-u_+^r)(x,t)|\\[2mm]
|(\t-\t_-^r-\T^{cd}-\t_+^r+\t_-^m+\t_+^m)(x,t)|\\[2mm]
\end{array}
\right)=0.
\end{equation*}
\end{remark}
We now explain the main strategy of this paper. It is noted that in \cite{Hong,huang-zhao,nishi-yang-zhao},
the smallness of $\gamma-1$ is used to control the lower and upper bound of the absolute temperature $\t$.
To remove the smallness condition of $\gamma-1$, the key point is to derive the uniform bound of $\t$, which is
also closely related to the uniform bound of the specific volume $v$.
Motivated by \cite{jiang,kazhi-she,L-L}, we first obtain the basic energy estimate (see Lemma \ref{basic-lemma}),
which is independent of the time $t$, with the help of the new estimates on the heat kernel developed in \cite{Huang-Li-matsumura},
provided the strengths of the waves are suitable small.
It should be emphasized that the basic energy estimate is nontrivially obtained, while it is trivial for the case of small initial perturbation or the far-field condition being a constant one $(\bar{v},\bar{u},\bar{\t})$.
In fact, we essentially use the structure of wave patterns to control the terms involving the derivative of perturbation around the wave patterns.
Secondly, the specific volume $v$ is shown uniformly bounded from below and above with respect to space and time through
delicate analysis based on the basic energy estimate and a cut-off technique. Finally, we manipulate some weighted estimates on the perturbation
around the wave patterns to derive the uniform bound of $\t$. We remark that the underlying structures of viscous contact wave and rarefaction waves are essentially used throughout the whole proof, and the idea may not be valid
for shock wave whose structure is quite different from those of
viscous contact wave and rarefaction waves.
\
This paper is organized as follows. In the next section, we collect some useful lemmas and
fundamental facts concerning the viscous contact wave as well as rarefaction waves.
The main proof of Theorem \ref{theorem1} and \ref{theorem2} are completed in Section 3 and 4, respectively.
\
\emph{Notations.} Throughout this paper, generic positive
constants are denoted by $c$ and $C$ without confusion.
For function spaces, $L^{p}(\Omega),1\leq p\leq
\infty$ denotes the usual Lebesgue space on $\Omega
\subset\mathbb{R}=(-\infty,\infty)$ with its norm given by
$$
\|f\|_{L^{p}(\Omega)}:=\left(\int_{\Omega}|f(x)|^{p}dx\right)^{\frac{1}{p}},
\quad 1\leq p<\infty, \quad \parallel
f\parallel_{L^{\infty}(\Omega)}:=\mbox{ess.sup}_{\Omega} |f(x)|.
$$
$H^{k}(\Omega)$ denotes the $k^{th}$ order Sobolev space with its
norm
$$
\|f\|_{H^{k}(\Omega)}:=\left(\sum ^{k}_{j=0}
\parallel \partial^{j}_{x}f\parallel^{2}(\Omega)\right)^{\frac{1}{2}}, \quad \mathrm{when} \parallel
\cdot
\parallel=\parallel \cdot
\parallel_{L^{2}(\Omega)}.
$$
The domain $\Omega$ will be often abbreviated without confusion.
\
\section{Preliminaries}
\setcounter{equation}{0}
The properties of the viscous contact wave $(V,U,\Theta)$ defined by \eqref{contact} are useful in the following sections.
\begin{lemma}\label{decay}
Assume that $\delta=|\t_+-\t_-|\leq \d_0$ for a small positive constant $\d_0$. Then the viscous contact wave $(V,U,\Theta)$ defined by \eqref{contact} has the following properties:
\begin{itemize}
\item [(1)]
$$
|V-v_{\pm}|+|\Theta-\t_{\pm}|\leq O(1)\d e^{-\frac{c_1x^2}{1+t}},
$$
\item [(2)]
$$
|\partial^k_x V|+|\partial^{k-1}_x U|+|\partial^k_x\Theta|\leq
O(1)\delta(1+t)^{-\frac{k}{2}}e^{-\frac{c_1x^2}{1+t}}, \quad k\geq 1.
$$
\end{itemize}
\end{lemma}
Therefore, we have
\begin{equation}
\di \widetilde R_1=O(1)\d(1+t)^{-\frac{3}{2}}e^{-\frac{c_1x^2}{1+t}},
\quad \widetilde R_2=O(1)\d(1+t)^{-2}e^{-\frac{c_1x^2}{1+t}}.
\end{equation}
The following two lemmas play important roles to obtain the basic energy
estimate, the proofs can be found in \cite{Huang-Li-matsumura}, we omit them for brevity.
\begin{lemma}\label{hlm}
For $0<T\leq+\infty$, suppose that $h(x,t)$ satisfies
$$
h\in L^{\infty}(0,T;L^2(\mathbb{R})),\quad h_x\in L^{2}(0,T;L^2(\mathbb{R})),\quad h_t\in L^{2}(0,T;H^{-1}(\mathbb{R})).
$$
Then
\begin{equation}
\int_0^T\int h^2w^2dxdt\leq 4\pi\|h(0)\|^2+4\pi\alpha^{-1}\int_0^T\|h_x\|^2dt
+8\alpha\int_0^T<h_t,hg^2>_{H^{-1}\times H^1}dt
\end{equation}
for $\alpha>0$, and
$$
w(x,t)=(1+t)^{-\frac{1}{2}}\exp\left(-\frac{\alpha x^2}{1+t}\right),\quad
g(x,t)=\int_{-\infty}^x w(y,t)dy.
$$
\end{lemma}
\begin{lemma}\label{lemma5}
For $\alpha\in(0,\frac{c_1}{4}]$ and $w$ defined in Lemma \ref{hlm}, there exists some positive
constant $C$ depending on $\alpha$, such that the following estimate holds
\begin{equation}\label{important}
\int_0^t\int(\phi^2+\psi^2+\zeta^2)w^2 dxds\leq C\left(
1+\int_0^t\int (\phi_x^2+\psi_x^2+\zeta_x^2)dxds
\right).
\end{equation}
\end{lemma}
Next, we state the following properties of the solution to the problem \eqref{burgers}
due to \cite{MN-86}.
\begin{lemma}\label{bur-pro}
For given $w_l\in\mathbb{R}$ and $\bar w>0$, let $w_r\in\{0<\tilde w\triangleq w-w_l<\bar w\}$.
Then the problem \eqref{burgers} has a unique smooth global solution in time satisfying the following properties.
\begin{itemize}
\item[(i)] $w_l<w(x,t)<w_r$, $w_x>0$ $(x\in\mathbb{R},t>0)$.
\item[(ii)] For $p\in[1,\infty]$, there exists some positive constant $C=C(p,w_l,\bar w)$ such that
for $\tilde w\geqq0$ and $t\geqq0$,
$$
\|w_x(t)\|_{L^p}\leq C\min\{\tilde w,\tilde w^{1/p}t^{-1+1/p}\},\quad
\|w_{xx}(t)\|_{L^p}\leq C\min\{\tilde w,t^{-1}\}.
$$
\item[(iii)] If $w_l>0$, for any $(x,t)\in(-\infty,0]\times[0,\infty)$,
$$
|w(x,t)-w_l|\leq\tilde we^{-2(|x|+w_lt)},\quad
|w_x(x,t)|\leq2\tilde we^{-2(|x|+w_lt)}.
$$
\item[(iv)] If $w_r<0$, for any $(x,t)\in[0,\infty)\times[0,\infty)$,
$$
|w(x,t)-w_r|\leq\tilde we^{-2(x+|w_r|t)},\quad
|w_x(x,t)|\leq2\tilde we^{-2(x+|w_r|t)}.
$$
\item[(v)] For the Riemann solution $w^r(x/t)$ of the scalar equation \eqref{burgers} with the Riemann
initial data
\begin{equation*}
w(x,0)=\left\{
\begin{array}{ll}
w_l,&\quad x<0,\\
w_r,&\quad x>0,
\end{array}
\right.
\end{equation*}
we have
$$
\lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}}|w(x,t)-w^r(x/t)|=0.
$$
\end{itemize}
\end{lemma}
Finally, we divide $\mathbb{R}\times(0,t)$ into
three parts, that is $\mathbb{R}\times(0,t)=\Omega_-\cup\Omega_c\cup\Omega_+$ with
$$
\Omega_{\pm}=\big\{(x,t)\big|\pm2x>\pm\lambda_{\pm}(v_{\pm}^m,s_{\pm})t\big\},
$$
and
$$
\Omega_{c}=\big\{(x,t)\big|\lambda_-(v_-^m,s_-)t\leq2x\leq\lambda_{+}(v_{+}^m,s_{+})t\big\}.
$$
Then Lemma \ref{decay} and Lemma \ref{bur-pro} lead to
\begin{lemma}\label{rare-pro}
For any given left end state $(v_-,u_-,\t_-)$, we assume that \eqref{zheng}
and \eqref{same-order} hold. Then the smooth rarefaction waves $(V_{\pm}^r,U_{\pm}^r,\T_{\pm}^r)$
constructed in \eqref{appro-rare} and the viscous contact wave $(V^{cd},U^{cd},\T^{cd})$ constructed in \eqref{contact} satisfying
the following:
\begin{itemize}
\item[(i)] $(U_{\pm}^r)_x\geq0$,~$(x\in\mathbb{R},t>0)$.
\item[(ii)] For $p\in[1,\infty]$, there exists a positive constant $C=C(v_-,u_-,\t_-,\d)$
such that for $\d$ satisfying \eqref{same-order},
$$
\|\big((V_{\pm}^r)_x,(U_{\pm}^r)_x,(\T_{\pm}^r)_x\big)(t)\|_{L^p}
\leq C\min\Big\{\d,\d^{1/p}t^{-1+1/p}\Big\}
$$
and
$$
\|\big((V_{\pm}^r)_{xx},(U_{\pm}^r)_{xx},(\T_{\pm}^r)_{xx}\big)(t)\|_{L^p}
\leq C\min\Big\{\d,t^{-1}\Big\}.
$$
\item[(iii)] There exists a positive constant $C=C(v_-,u_-,\t_-,\d)$
such that for
$$
c_0=\frac{1}{10}\min\Big\{|\lambda_-(v_-^m,s_-)|,\lambda_+(v_+^m,s_+),c_1\lambda_-^2(v_-^m,s_-),
c_1\lambda_+^2(v_+^m,s_+),1\Big\},
$$
we have in $\Omega_c$
$$
(U_{\pm}^r)_x+|(V_{\pm}^r)_x|+|V_{\pm}^r-v_{\pm}^m|+|(\T_{\pm}^r)_x|+|\T_{\pm}^r-\t_{\pm}^m|
\leq C\d e^{-c_0(|x|+t)},
$$
and in $\Omega_{\mp}$
$$
\left\{
\begin{array}{ll}
|V^{cd}-v_{\mp}^m|+|V^{cd}_x|+|\T^{cd}-\t_{\mp}^m|+|U^{cd}_x|+|\T^{cd}_x|
\leq C\d e^{-c_0(|x|+t)},\\
(U_{\pm}^r)_x+|(V_{\pm}^r)_x|+|V_{\pm}^r-v_{\pm}^m|+|(\T_{\pm}^r)_x|+|\T_{\pm}^r-\t_{\pm}^m|
\leq C\d e^{-c_0(|x|+t)}.
\end{array}
\right.
$$
\item[(iv)] For the rarefaction waves $(v_{\pm}^r,u_{\pm}^r,\t_{\pm}^r)(x/t)$ determined by
\eqref{euler}\eqref{Riemann2}, it holds
$$
\lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}}
\big|(V_{\pm}^r,U_{\pm}^r,\T_{\pm}^r,)(x,t)-(v_{\pm}^r,u_{\pm}^r,\t_{\pm}^r,)(x/t)\big|=0.
$$
\end{itemize}
\end{lemma}
\section{Proof of Theorem \ref{theorem1}}
\setcounter{equation}{0}
Substituting \eqref{contact equation} into \eqref{ns}, \eqref{initial} yields
\begin{equation}\label{perturb}
\left\{
\begin{array}{ll}
\di \phi_t-\psi_x=0,\\
\di \psi_t+\left(p-p_+\right)_x=\mu\left(\frac{u_x}{v}-\frac{U_x}{V}\right)_x-\widetilde R_1,\\
\di c_{\nu}\zeta_t+pu_x-p_+U_x=\kappa\left(\frac{\t_x}{v}-\frac{\Theta_x}{V}\right)_x
+\mu\left(\frac{u_x^2}{v}-\frac{U^2_x}{V}\right)-\widetilde R_2,\\
\di (\phi,\psi,\zeta)(x,0)=(\phi_0,\psi_0,\zeta_0)(x),\quad x\in \mathbb{R}.
\end{array}
\right.
\end{equation}
We shall prove Theorem \ref{theorem1} by the local existence and the a priori estimate.
We look for the solution $(\phi,\psi,\zeta)$ in the solution space $X([0,+\infty))$,
\begin{equation*}
\begin{array}{l}
\di X([0,T])=\Big\{(\phi,\psi,\zeta)\Big| v, ~\theta \geq M^{-1},
\quad \sup_{0\leq t\leq T}\|(\phi,\psi,\zeta)\|_{H^1}\leq M
\Big\}
\end{array}
\end{equation*}
for some $0<T\leq +\infty$, where the constants $M$ will be determined later.
Since the local existence of the solution is well known (for example, see \cite{huang-ma-shi}),
to prove the global existence part of Theorem \ref{theorem1}, we only need to establish the
following a priori estimates.
\begin{proposition}\label{prop}
\textbf{(A priori estimates)} Assume that the conditions of Theorem \ref{theorem1} hold, then there exists
a positive constant $\d_0$ such that if $\d< \d_0$ and $(\phi,\psi,\zeta)\in X([0,T])$,
\begin{equation}
\begin{array}{l}\label{p-1}
\di \sup_{0\leq t\leq T}\|(\phi,\psi,\zeta)(t)\|_{H^1}^2+\int_0^T(\|\phi_x\|^2+\|(\psi_x,\zeta_x)\|^2_{H^1})ds
\leq C_0,
\end{array}
\end{equation}
where $C_0$ denotes a constant depending only on $\mu$, $\kappa$, $R$, $c_{\nu}$, $v_{\pm}$, $u_{\pm}$, $\t_{\pm}$ and $m_0$.
\end{proposition}
Once Proposition \ref{prop} is proved, we can extend the unique local solution $(u,v,\t)$ which can be
obtained as in \cite{huang-ma-shi} to $T=\infty$. Estimate \eqref{p-1}
and the equations \eqref{perturb} (respectively \eqref{perturb2}) imply that
\begin{equation}
\di\int_0^{+\infty}\left(\|(\phi_x, \psi_x,\zeta_x)(t)\|^2+\left|\frac{d}{dt}\|(\phi_x, \psi_x,\zeta_x)(t)\|^2\right|\right)dt< \infty,
\end{equation}
which, together with \eqref{p-1} and the Sobolev's inequality, easily leads to the large time behavior of the solutions,
that is, \eqref{behavior-1} (resp. \eqref{wt-2}).
Proposition \ref{prop} will be finished by the following lemmas. First, we give the basic energy estimate,
which is nontrivially obtained, compared with the case of small initial perturbation or the far-field condition
being a constant one $(\bar v,\bar u,\bar\t)$.
\begin{lemma}\label{basic-lemma}
There exist some positive constant $C_0$ and $\d_0$ such that if $\d<\d_0$, it holds that
\begin{equation}\label{basic}
\begin{array}{ll}
\di \Big\|\left(\psi,\sqrt{\Phi\left(\frac{v}{V}\right)},\sqrt{\Phi\left(\frac{\t}{\Theta}\right)}\right)(t)\Big\|^2
+\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right)dxds
\leq C_0.
\end{array}
\end{equation}
\end{lemma}
\textbf{Proof}: The proof of the Lemma \ref{basic-lemma} consists of two steps.
\underline{ Step 1. }\quad Similar to \cite{huang-ma-shi}, multiplying $\eqref{ns}_1$ by
$-R\Theta(v^{-1}-V^{-1})$, $\eqref{ns}_2$ by $\psi$ and $\eqref{ns}_3$ by $\zeta\theta^{-1}$, then
adding the resulting equations together, we get
\begin{equation}
\begin{array}{ll}
\di\left(\frac{\psi^2}{2}+R\Theta\Phi\left(\frac{v}{V}\right)+c_{\nu}\Theta\Phi\left(\frac{\theta}{\Theta}\right)\right)_t
+\frac{\mu\Theta}{\theta v}\psi_x^2+\frac{\kappa\Theta}{\theta^2 v}\zeta_x^2
+H_{x}+Q=-\psi \widetilde R_1-\frac{\zeta}{\theta}\widetilde R_2
\end{array}
\label{ba1}
\end{equation}
with
$$
\Phi(z)=z-\ln z-1, \quad z>0
$$
and
\begin{equation}\label{H}
\di H=(p-p_+)\psi-\mu\left(\frac{u_x}{v}-\frac{U_x}{V}\right)\psi-\frac{\k\zeta}{\theta}\left(\frac{\t_x}{v}-\frac{\Theta_x}{V}\right),
\end{equation}
\begin{equation}
\begin{array}{ll}
\di Q&\di=p_+\Phi\left(\frac{V}{v}\right)U_x+\frac{p_+}{\gamma-1}\Phi\left(\frac{\Theta}{\theta}\right)U_x
+\mu\left(\frac{1}{v}-\frac{1}{V}\right)\psi_xU_x\\[3mm]
&\di -\frac{\zeta}{\theta}(p_+-p)U_x-\frac{\k\Theta_x}{\theta^2v}\zeta\zeta_x
-\frac{\k\Theta\Theta_x}{\theta^2vV}\zeta_x\phi+\frac{\k\Theta_x^2}{\t^2vV}\zeta\phi\\[3mm]
&\di-\frac{2\mu U_x}{\theta v}\psi_x\zeta+\frac{\mu U_x^2}{\theta vV}\zeta\phi.
\end{array}
\end{equation}
Since
\begin{equation}
\di Q\leq \frac{\mu\Theta}{4\theta v}\psi_x^2+\frac{\k\Theta}{4\theta^2v}\zeta_x^2
+C(M)(\phi^2+\zeta^2)(|U_x|+\Theta_x^2),
\end{equation}
where $C(M)$ denotes a constant depending on $M$. Recalling Lemma \ref{decay}, we have
\begin{equation}
\begin{array}{ll}
\di \left|\int_0^t\int\widetilde R_1\psi dxds\right|&\di\leq O(1)\d \int_0^t\int (1+s)^{-3/2}e^{-\frac{c_1x^2}{1+s}}|\psi|dxds\\[3mm]
&\di \leq O(1)\d\int_0^t(1+s)^{-5/4}\|\psi\|ds\\[3mm]
&\di \leq O(1)\d\int_0^t(1+s)^{-5/4}\|\psi\|^2ds+O(1)\d
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\di \left|\int_0^t\int\widetilde R_2\frac{\zeta}{\t} dxds\right|
\leq C(M)\d\int_0^t(1+s)^{-7/4}\left\|\sqrt{\Phi\left(\frac{\t}{\T}\right)}\right\|^2ds+C(M)\d.
\end{array}
\end{equation}
Then integrating \eqref{ba1} over $\mathbb{R}\times(0,t)$, choosing $\a=\frac{c_1}{4}$ in Lemma \ref{hlm}
and $\d$ suitable small, it follows from Lemma \ref{hlm}-\ref{lemma5} and Gronwall's inequality that
\begin{equation}\label{ba2}
\begin{array}{ll}
\di \Big\|\left(\psi,\sqrt{\Phi\left(\frac{v}{V}\right)},\sqrt{\Phi\left(\frac{\t}{\Theta}\right)}\right)(t)\Big\|^2
+\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right)dxds\\[3mm]
\di
\leq C_0+ C(M)\d \int_0^t\int(1+s)^{-1}(\phi^2+\zeta^2)e^{-\frac{c_1x^2}{1+s}}dxds\\
\di \leq C_0+ C(M)\d \int_0^t\int\frac{\t\phi_x^2}{v^3}dxds.
\end{array}
\end{equation}
\underline{ Step 2. }\quad Following \cite{MN-92}, we introduce a new
variable $\di\tilde v=\frac{v}{V}$. Then $\eqref{perturb}_2$ can be rewritten by the
new variable as
\begin{equation}\label{newperturb}
\di \left(\mu\frac{\tilde v_x}{\tilde v}-\psi\right)_t-p_x=\widetilde R_1.
\end{equation}
Multiplying \eqref{newperturb} by $\di\frac{\tilde v_x}{\tilde v}$, we have
\begin{equation}\label{phi-x1}
\begin{array}{ll}
\di \left(\frac{\mu}{2}\left(\frac{\tilde v_x}{\tilde v}\right)^2-\psi\frac{\tilde v_x}{\tilde v}\right)_t
+\left(\psi\frac{\tilde v_t}{\tilde v}\right)_x+\frac{R\t}{v}\left(\frac{\tilde v_x}{\tilde v}\right)^2
-\frac{R}{v}\zeta_x\frac{\tilde v_x}{\tilde v}\\[3mm]
\di\quad +\frac{R\t}{v}\left(\frac{1}{\Theta}-\frac{1}{\t}\right)\Theta_x\frac{\tilde v_x}{\tilde v}
=\frac{\psi_x^2}{v}+\psi_xU_x\left(\frac{1}{v}-\frac{1}{V}\right)+\widetilde R_1\frac{\tilde v_x}{\tilde v}.
\end{array}
\end{equation}
The Cauchy's inequality yields that
\begin{equation}
\begin{array}{ll}
\di \left|\frac{R}{v}\zeta_x\frac{\tilde v_x}{\tilde v}\right|
+\left|\psi_xU_x\left(\frac{1}{v}-\frac{1}{V}\right)\right|+\frac{\psi_x^2}{v}\\[3mm]
\di \leq \frac{R\t}{4v}\left(\frac{\tilde v_x}{\tilde v}\right)^2+
C(M)\left(\frac{\zeta_x^2}{\t^2 v}+\frac{\psi_x^2}{\t v}\right)
+C(M)\phi^2U_x^2,
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\di \left|\frac{R\t}{v}\left(\frac{1}{\Theta}-\frac{1}{\t}\right)\Theta_x\frac{\tilde v_x}{\tilde v}\right|
+\left|\widetilde R_1\frac{\tilde v_x}{\tilde v}\right|\\[3mm]
\di\leq \frac{R\t}{4v}\left(\frac{\tilde v_x}{\tilde v}\right)^2
+C(M)(\zeta^2\Theta_x^2+\widetilde R_1^2).
\end{array}
\end{equation}
Note that
\begin{equation*}
\begin{array}{ll}
\di \frac{\phi_x^2}{2v^2}-C(M)\phi^2\Theta_x^2\leq \left(\frac{\tilde v_x}{\tilde v}\right)^2
\leq \frac{\phi_x^2}{v^2}+C(M)\phi^2\Theta_x^2.
\end{array}
\end{equation*}
Integrating \eqref{phi-x1} over $\mathbb{R}\times(0,t)$, we have
\begin{equation}
\begin{array}{ll}
\di\int\frac{\phi_x^2}{v^2}dx+\int_0^t\int\frac{\t\phi_x^2}{v^3} dxds\leq C_0+C\|\psi\|^2\\[3mm]
\di\quad+C(M)\d^2\left\|\sqrt{\Phi\left(\frac{v}{V}\right)}\right\|^2
+C(M)\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right) dxds\\[3mm]
\di\quad +C(M)\int_0^t\int(\phi^2+\zeta^2)(U_x^2+\Theta_x^2) dxds.
\end{array}
\end{equation}
By Lemma \ref{lemma5}, \eqref{ba2} and choosing $\d$ suitable small, we have
\begin{equation}\label{phi-x3}
\begin{array}{ll}
\di \int\frac{\phi_x^2}{v^2}dx+\int_0^t\int\frac{\theta\phi_x^2}{v^3}dxds
\leq C_0+C(M)\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\theta^2v}\right) dxds.
\end{array}
\end{equation}
Then the proof of Lemma \ref{basic-lemma} is completed by substituting \eqref{phi-x3} into \eqref{ba2},
and choosing $\d$ suitable small.
\hfill $\Box$
In Lemma \ref{basic-lemma}, the smallness of $\delta$ is used to guarantee that the basic energy (\ref{basic}) is only bounded by the initial data. Based on
the basic energy estimate, we shall show the specific volume $v$ and the absolute temperature $\t$ are uniformly bounded from below and above, which in turn decides
how small for $\delta$. That is why need the initial condition (\ref{data}). To prove Theorem \ref{theorem1}, we first try to get the uniform bound of $v(x,t)$. We have
\begin{lemma}
Let $\a_1$, $\a_2$ be the two positive roots of the equation $y-\ln y-1=C_0$ and the constant $C_0$ be the same in \eqref{basic}. Then
\begin{equation}\label{root}
\di \a_1\leq \int_k^{k+1}\tilde v(x,t)dx, \quad\int_k^{k+1}\tilde \t(x,t)dx\leq \a_2,\quad t\geq 0,
\end{equation}
and for each $t\geq 0$ there are points $a_k(t)$, $b_k(t)\in[k,k+1]$ such that
\begin{equation}\label{root2}
\alpha_1\leq \tilde v(a_k(t),t), \tilde \t(b_k(t),t)\leq \alpha_2,\quad t\geq 0,
\end{equation}
where $k=0,\pm1,\pm2, \cdots$.
\end{lemma}
\textbf{Proof}: From \eqref{basic}, we see that
\begin{equation}\label{4.1}
\di \int_k^{k+1}(\tilde v(x,t)-\ln\tilde v(x,t)-1)dx, \quad\int_k^{k+1}(\tilde \t(x,t)-\ln\tilde \t(x,t)-1)dx\leq C_0,
\end{equation}
where $\tilde v=\frac{v}{V}$, $\tilde \t=\frac{\t}{\Theta}$, and $k=0,\pm1,\pm2,\cdots$. If we apply Jessen's inequality to the convex function
$y-\ln y-1=C_0$, we obtain
$$
\di \int_k^{k+1}\tilde v(x,t)dx-\ln\int_k^{k+1}\tilde v(x,t)dx-1,
\quad\int_k^{k+1}\tilde \t(x,t)dx-\ln\int_k^{k+1}\tilde \t(x,t)dx-1\leq C_0,
$$
which gives
$$
\di \a_1\leq \int_k^{k+1}\tilde v(x,t)dx, \quad\int_k^{k+1}\tilde \t(x,t)dx\leq \a_2,
$$
where $\alpha_1$, $\alpha_2$ are two positive roots of the equation $y-\ln y-1=C_0$.
Moreover, in view of mean value theorem, for each $t\geq 0$, there are points
$a_k(t)$, $b_k(t)\in [k,k+1]$ such that
\begin{equation}
\di 0<\alpha_1\leq \tilde v(a_k(t),t), \tilde \t(b_k(t),t)\leq \alpha_2,\quad t\geq 0.
\end{equation}
\hfill $\Box$
The following Lemma can be found in Jiang \cite{jiang} Lemma 2.3.
\begin{lemma}\label{jiang-lemma}
For each $x\in[k.k+1]$, $k=0,\pm1,\pm2,\cdots$, it follows from \eqref{ns}$_2$ that
\begin{equation}\label{v-repre}
\di v(x,t)=B(x,t)Y(t)+\frac{R}{\mu}\int_0^t\frac{B(x,t)Y(t)}{B(x,s)Y(s)}\t(x,s) ds,
\end{equation}
where
\begin{equation}\label{B}
\di B(x,t)=v_0(x)\exp\left(\frac{1}{\mu}\int_x^{+\infty}\big(u_0(y)-u(y,t)\big)\beta(y)dy\right),
\end{equation}
\begin{equation}\label{Y}
\di Y(t)=\exp\left(\frac{1}{\mu}\int_0^t\int_{k+1}^{k+2}\s(y,s)dyds\right),
\end{equation}
\begin{equation}\label{sigma}
\di \sigma(x,t)=\left(\mu\frac{u_x}{v}-R\frac{\t}{v}\right)(x,t),
\end{equation}
and
\begin{equation}\label{beta}
\di \beta(x)=\left\{
\begin{array}{ll}
1,& x\leq k+1,\\
k+2-x, & k+1\leq x\leq k+2,\\
0, & x\geq k+2.
\end{array}
\right.
\end{equation}
\end{lemma}
By Cauchy's inequality and \eqref{basic}, we have
\begin{equation}\label{b-bound}
\di \underline{B}(C_0)\leq B(x,t)\leq \overline{B}(C_0), \quad \forall x\in[k,k+1],\quad t\geq 0,
\end{equation}
where $\underline{B}(C_0)$, $\overline{B}(C_0)$ are two constants depending on $C_0$.
\begin{lemma}\label{v-bound}
There are two positive constants $\underline{v}(C_0)$, $\bar{v}(C_0)$ such that
\begin{equation}\label{v}
\di \underline{v}(C_0) \leq v(x,t)\leq\bar{v}(C_0),\quad \forall x\in\mathbb{R},\quad t\geq0,
\end{equation}
where $\underline{v}(C_0)$, $\bar{v}(C_0)$ depending on $C_0$, independent of $x$, $t$.
\end{lemma}
\textbf{Proof}: From now on, we always assume that $\t_-<\t_+$ for convenient. So from the properties of
viscous contact wave, we have $\t_-<\T(x,t)<\t_+$ and $v_-<V(x,t)<v_+$.
For each $t\geq0$, there exists at least one point $x=x_{k+1}(t)\in[k+1,k+2]$ such that
$$
\di \inf_{x\in[k+1,k+2]}\tilde\t(x,t)=\tilde\t(x_{k+1}(t),t).
$$
By Cauchy's inequality, \eqref{basic}, \eqref{root}, and choosing $\d$ suitable small, we see that
\begin{equation}\label{4.2}
\begin{array}{ll}
\di\left|\int_s^t\int_{b_{k+1}(\tau)}^{x_{k+1}(\tau)}\frac{\tilde\t_y}{\tilde\t}(y,\tau) dyd\tau \right|=
\left|\int_s^t\int_{b_{k+1}(\tau)}^{x_{k+1}(\tau)} \left(\frac{\zeta_y}{\t}-\frac{\zeta\Theta_y}{\t\Theta}\right)dyd\tau \right|\\[3mm]
\di\quad\leq\int_s^t\left(\int_{k+1}^{k+2}\frac{\zeta_x^2}{\t^2v}dx\right)^{\frac{1}{2}}\left(\int_{k+1}^{k+2}vdx\right)^{\frac{1}{2}}d\tau
+\int_s^t\int_{k+1}^{k+2}\left|\frac{\zeta\Theta_x}{\t\Theta}\right|dxd\tau\\[3mm]
\di\quad\leq\int_s^t\left(\int_{k+1}^{k+2}\frac{\zeta_x^2}{\t^2v}dx\right)^{\frac{1}{2}}\big(v_+\a_2\big)^{\frac12}d\tau
+C(M)\int_s^t\left(\int_{k+1}^{k+2}\zeta^2\Theta_x^2 dx\right)^{\frac12}d\tau\\[3mm]
\di\quad \leq C\left(\int_s^t\int\frac{\zeta_x^2}{\t^2v} dyd\tau\right)^{\frac12}\sqrt{t-s}
+C(M)\left(\int_s^t\int\zeta^2\Theta_x^2 dxd\tau\right)^{\frac12}\sqrt{t-s}\\[3mm]
\di\quad \leq C_0\sqrt{t-s}.
\end{array}
\end{equation}
We apply Jessen's inequality to the convex function $e^x$, and utilize \eqref{basic}, \eqref{root2},
\eqref{4.2} to obtain that for $t\geq s\geq 0$,
\begin{equation}
\begin{array}{ll}
\di\int_s^t\inf_{x\in[k+1,k+2]}\tilde \t(\cdot,\tau) d\tau=
\int_s^t\tilde \t(x_{k+1}(\tau),\tau) d\tau=
\int_s^t\exp\left(\log\tilde\t(x_{k+1}(\tau),\tau)\right) d\tau\\[4mm]
\di\quad\geq(t-s)\exp\left(\frac{1}{t-s}\int_s^t\log\tilde\t(x_{k+1}(\tau),\tau) d\tau\right)\\[4mm]
\di\quad=(t-s)\exp\left(\frac{1}{t-s}\int_s^t\left[\log\frac{\tilde\t(x_{k+1}(\tau),\tau)}{\tilde\t(b_{k+1}(\tau),\tau)}
+\log\tilde\t(b_{k+1}(\tau),\tau)\right] d\tau\right)\\[4mm]
\di\quad=(t-s)\exp\left(\frac{1}{t-s}\int_s^t\left[\int_{b_{k+1}(\tau)}^{x_{k+1}(\tau)}\frac{\tilde\t_y}{\tilde\t} dy
+\log\tilde\t(b_{k+1}(\tau),\tau)\right]d\tau\right)\\[4mm]
\di\quad\geq(t-s)\exp\left(\log\a_1-\frac{1}{t-s}\left|\int_s^t\int_{b_{k+1}(\tau)}^{x_{k+1}(\tau)}\frac{\tilde\t_y}{\tilde\t}dyd\tau\right|\right)\\[4mm]
\di\quad\geq C(t-s)e^{-\frac{C_0}{\sqrt{t-s}}}.
\end{array}
\end{equation}
Noticing that $\t_-\leq\Theta\leq\t_+$, so we have
\begin{equation}\label{int-theta}
-\int_s^t\inf_{x\in[k+1,k+2]}\t(\cdot,\tau) d\tau\leq
\left\{
\begin{array}{ll}
0,&0\leq t-s\leq 1,\\
-C_0(t-s),&t-s\geq 1.
\end{array}
\right.
\end{equation}
Applying Cauchy's inequality and Jessen's inequality for the function $\frac{1}{x}(x>0)$, using
\eqref{basic}, \eqref{int-theta}, and noting that
\begin{equation}
\di\left\{
\begin{array}{ll}
C_0, &0\leq t-s\leq 1,\\
C_0-\frac{t-s}{C_0}, & t-s\geq 1
\end{array}
\right.
\leq C_0-\frac{t-s}{C_0},\quad t\geq s\geq 0,
\end{equation}
we obtain
\begin{equation}\label{int-sigma}
\begin{array}{ll}
\di\int_s^t\int_{k+1}^{k+2}\sigma(x,\tau)dxd\tau=
\int_s^t\int_{k+1}^{k+2}\left(\mu\frac{\psi_x}{v}-R\frac{\t}{v}+\mu\frac{U_x}{v}\right)(x,\tau)dxd\tau\\[3mm]
\di\quad\leq C\int_s^t\int_{k+1}^{k+2}\frac{\psi_x^2}{\t v}dxd\tau-\frac{R}{2}\int_s^t\int_{k+1}^{k+2}\frac{\t}{v}dxd\tau
+\mu\int_s^t\int_{k+1}^{k+2}\frac{U_x}{v}dxd\tau\\[3mm]
\di\quad\leq C_0-\frac{R}{2}\int_s^t\inf_{x\in[k+1,k+2]}\t\left(\int_{k+1}^{k+2}v^{-1}dx\right)d\tau
+C(M)\left(\int_s^t\int_{k+1}^{k+2} U_x^2 dxd\tau\right)^{\frac{1}{2}}(t-s)^{\frac12}\\[3mm]
\di\quad\leq C_0-\frac{R}{2}\int_s^t\inf_{x\in[k+1,k+2]}\t\left(\int_{k+1}^{k+2}vdx\right)^{-1}d\tau
+C(M)\d(t-s)^{\frac12}\\[3mm]
\di\quad\leq C_0-\frac{R}{2\a_2v_+}\int_s^t\inf_{x\in[k+1,k+2]}\t d\tau
+C(t-s)^{\frac12}\\[3mm]
\di\quad\leq C_0-\frac{t-s}{C_0}.
\end{array}
\end{equation}
It follows from the definition of $Y(t)$ and \eqref{int-sigma} that
\begin{equation}\label{Y-e}
\di 0\leq Y(t)\leq C_0e^{-t/C_0},\quad
\frac{Y(t)}{Y(s)}\leq C_0e^{-(t-s)/C_0},
\end{equation}
which, together with \eqref{v-repre} and \eqref{b-bound}, gives,
\begin{equation}\label{v-repre2}
\di v(x,t)\leq C_0+C_0\int_0^t\t(x,s)e^{-(t-s)/C_0} ds.
\end{equation}
On the other hand, we have
\begin{equation}
\begin{array}{ll}
\di |\tilde \t^{\frac{1}{2}}(x,t)-\tilde\t^{\frac{1}{2}}(b_k(t),t)|
\leq\int_{k}^{k+1}\tilde\t^{-\frac{1}{2}}|\tilde\t_x|dx
\leq\int_k^{k+1}\left(\frac{\Theta}{\theta}\right)^{\frac{1}{2}}
\left(\left|\frac{\zeta_x}{\Theta}\right|+\left|\frac{\zeta\Theta_x}{\Theta^2}\right|\right)dx\\[3mm]
\di\quad\leq\t_{-}^{-\frac{1}{2}}\int_k^{k+1}\frac{|\zeta_x|}{\sqrt{\t}}dx
+C(M)\int_k^{k+1}|\zeta\Theta_x|dx\\[3mm]
\di\quad\leq\t_{-}^{-\frac{1}{2}}\left(\int_k^{k+1}\frac{\zeta_x^2}{\t^2v}dx\right)^{\frac{1}{2}}
\left(\int_k^{k+1}\t vdx\right)^{\frac{1}{2}}+C(M)\left(\int_k^{k+1}\zeta^2\Theta_x^2dx\right)^{\frac{1}{2}}\\[3mm]
\di\quad\leq\t_{-}^{-\frac{1}{2}}\left(\int_k^{k+1}\frac{\zeta_x^2}{\t^2v}dx\right)^{\frac{1}{2}}
\left(\int_k^{k+1}\t dx\right)^{\frac{1}{2}}\max_{x\in[k,k+1]}v(\cdot,t)^{\frac{1}{2}}
+C(M)\left(\int_k^{k+1}\zeta^2\Theta_x^2dx\right)^{\frac{1}{2}}\\[3mm]
\di\quad\leq\sqrt{\frac{\a_2\t_+}{\t_-}}\left(\int\frac{\zeta_x^2}{\t^2v}dx\right)^{\frac{1}{2}}\max_{x\in[k,k+1]}v(\cdot,t)^{\frac{1}{2}}
+C(M)\left(\int\zeta^2\Theta_x^2dx\right)^{\frac{1}{2}}\quad \mathrm{for}\quad x\in[k,k+1]
\end{array}
\end{equation}
and $k=0$, $\pm1$, $\pm2, \cdots$, which, along with \eqref{root2}, leads to
\begin{equation}\label{v-theta}
\begin{array}{ll}
\di \frac{\a_1\t_-}{3}-\a_2\frac{\t_+^2}{\t_-}\left(\int_{\mathbb{R}}\frac{\zeta_x^2}{\t^2v}\right)\max_{x\in\mathbb{R}}v(\cdot,t)
-C(M)\int_{\mathbb{R}}\zeta^2\Theta^2_x\\[3mm]
\di\quad\leq \theta(x,t)\leq 3\a_2\t_+ +3\a_2\frac{\t_+^2}{\t_-}\left(\int_{\mathbb{R}}\frac{\zeta_x^2}{\t^2v}\right)\max_{x\in\mathbb{R}}v(\cdot,t)
+C(M)\int_{\mathbb{R}}\zeta^2\Theta^2_x, \quad \forall x\in\mathbb{R}.
\end{array}
\end{equation}
Hence, substituting \eqref{v-theta} into \eqref{v-repre2}, applying Gronwall's inequality and \eqref{basic}, one has,
\begin{equation}\label{v-upper}
\di v(x,t)\leq C_0,\qquad \forall x\in\mathbb{R},\quad t\geq 0.
\end{equation}
Integrating \eqref{v-repre} over $[k,k+1]$ with respect to $x$, we obtain
\begin{equation}
\begin{array}{ll}
\di v_-\a_1&\di\leq C_0e^{-t/C_0}+C_0\int_0^t\frac{Y(t)}{Y(s)}\int_k^{k+1}\theta(x,s)dxds\\
&\di \leq C_0e^{-t/C_0}+C_0\int_0^t\frac{Y(t)}{Y(s)}ds.
\end{array}
\end{equation}
This directly yields that
\begin{equation}\label{Yt-Ys}
\di \int_0^t\frac{Y(t)}{Y(s)} ds\geq C_0-C_0e^{-t/C_0}.
\end{equation}
From \eqref{basic},\eqref{v-repre}, \eqref{v-theta}, \eqref{v-upper} and \eqref{Yt-Ys}, and choosing $\d$
suitable small, we have
\begin{equation}\label{wt-1}
\begin{array}{ll}
\di v(x,t)&\di\geq C_0\int_0^t\frac{Y(t)}{Y(s)}\theta(x,s)ds\\
&\di\geq C_0\int_0^t\frac{Y(t)}{Y(s)}ds-C_0\left(\int_0^{t/2}+\int_{t/2}^t\right)\frac{Y(t)}{Y(s)}\int\frac{\zeta_x^2}{\t^2v} dxds\\
&\di\qquad-C(C_0,M)\int_0^t\int\zeta^2\Theta_x^2 dxds\\
&\di\geq C_0-C_1(C_0)e^{-t/C_0}-C_0e^{-t/2C_0}\int_0^t\int\frac{\zeta_x^2}{\theta^2v} dxds
-C_0\int_{t/2}^t\int\frac{\zeta_x^2}{\theta^2v}dxds\\
&\di\geq C_0/2, \qquad \forall x\in\mathbb{R}, ~t\geq T_0,
\end{array}
\end{equation}
where $C_1(C_0)$ is some positive constant depending on $C_0$, and $T_0$, $C_0$ are positive constants independent of $t$.
Next we consider the lower bound of $v(x,t)$ on $[0,T]$ for a positive constant $T>0$. From \cite{kazhikhov},
for any $x\in[k,k+1]$, $k=0,\pm1,\pm2,\cdots$, it holds that
\begin{equation}\label{v-ka}
\di v(x,t)=\frac{1}{\widetilde Y(t)\widetilde B(x,t)}
\left(v_0(x)+\frac{R}{\mu}\int_0^t\widetilde Y(s)\widetilde B(x,s)\t(x,s) ds\right),
\end{equation}
where
$$
\di\widetilde Y(t)=\frac{v_0(a_k(t))}{v(a_k(t),t)}\exp\left(\frac{R}{\mu}\int_0^t\frac{\t}{v}(a_k(t),s)ds\right)
$$
and
$$
\di\widetilde B(x,t)=\exp\left(\frac{1}{\mu}\int_{a_k(t)}^x\Big(u_0(y)-u(y,t)\Big)dy\right)
$$
with $a_{k}(t)$ is the same as in \eqref{root2}.
It follows from \eqref{v-ka} that
\begin{equation}\label{v-ka2}
\di \widetilde Y(t)v(x,t)=\frac{1}{\widetilde B(x,t)}
\left(v_0(x)+\frac{R}{\mu}\int_0^t\widetilde Y(s)\widetilde B(x,s)\t(x,s) ds\right).
\end{equation}
Integrating \eqref{v-ka2} over $[k,k+1]$ with respect to $x$, we obtain
\begin{equation}
\begin{array}{ll}
\di \a_1v_- \widetilde Y(t)&\di\leq C_0+C_0\int_0^t\widetilde Y(s)\int_k^{k+1}\t(x,s) dxds\\
&\di\leq C_0+C_0\int_0^t\widetilde Y(s)ds,
\end{array}
\end{equation}
which, together with the Gronwall's inequality, yields
$$
\widetilde Y(t)\leq C(C_0,T).
$$
Then from \eqref{v-ka}, one has
\begin{equation*}
\di v(x,t)\geq \frac{v_0(x)}{\widetilde Y(t)\widetilde B(x,t)}\geq C(C_0,T),
\qquad \forall x\in \mathbb{R},
\end{equation*}
which, together with \eqref{wt-1} and \eqref{v-upper}, completes the proof of Lemma \ref{v-bound}.
\hfill $\Box$
\
Motivated by \cite{L-L}, we shall show the uniform bound of the absolute temperature $\t$
from below and above with respect to space and time. We have
\begin{lemma}\label{lemma-ll-1}
There exists some positive constants $C_0$ such that for any given $T>0$,
\begin{equation}\label{new-nergy}
\di\sup_{0\leq t\leq T}\int(\zeta^2+\psi^4)dx+\int_0^T\int((\t+\psi^2)\psi_x^2+\zeta_x^2) dxdt
\leq C_0.
\end{equation}
\end{lemma}
\textbf{Proof}: The proof of the Lemma \ref{lemma-ll-1} consists of the following steps.
\underline{ Step 1. }\quad First, for $t\geq0$, and $a>1$, denoting
$$
\Omega_a(t)\triangleq \left\{x\in\mathbb{R}\Bigg|\frac{\t}{\Theta}(x,t)>a\right\}
=\{x\in\mathbb{R}|\zeta(x,t)>(a-1)\Theta(x,t)\}.
$$
We derive from \eqref{basic} that $\Omega_a$ is bounded since
\begin{equation}\label{omega-bound}
\di a|\Omega_a|<\sup_{0\leq t\leq T}\int_{\Omega_a}\frac{\t}{\Theta}dx
\leq C(a)\sup_{0\leq t\leq T}\int_{\mathbb{R}}\Phi\left(\frac{\t}{\Theta}\right)dx\leq C(a,C_0).
\end{equation}
Next, multiplying $\eqref{perturb}_3$ by $(\zeta-\Theta)_+=\max\{\zeta-\Theta,0\}$, then integrating the resulted equation
over $\mathbb{R}\times[0,t]$, one has
\begin{equation}\label{5.1}
\begin{array}{ll}
\di\frac{c_{\nu}}{2}\int(\zeta-\Theta)_+^2dx+\k\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
=\frac{c_{\nu}}{2}\int(\zeta_0(x)-\Theta(x,0))_+^2dx\\[2mm]
\di\quad-\int_0^t\int\frac{R\zeta+R\Theta}{v}\psi_x(\zeta-\Theta)_+ dxds
-\int_0^t\int\frac{R\zeta-p_+\phi}{v}U_x(\zeta-\Theta)_+ dxds\\[2mm]
\di\quad+\k\int_0^t\int_{\Omega_2}\frac{\zeta_x\Theta_x}{V}dxds
-\k\int_0^t\int_{\Omega_2}\frac{\phi\Theta_x^2}{vV}dxds
+\mu\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds\\[2mm]
\di\quad+2\mu\int_0^t\int\frac{\psi_xU_x}{v}(\zeta-\Theta)_+ dxds-\mu\int_0^t\int\frac{\phi U_x^2}{vV}(\zeta-\Theta)_+ dxds\\[2mm]
\di\quad-\int_0^t\int\widetilde R_2(\zeta-\Theta)_+ dxds-c_{\nu}\int_0^t\int\partial_t\Theta(\zeta-\Theta)_+ dxds.
\end{array}
\end{equation}
We multiply $\eqref{perturb}_2$ by $2\psi(\zeta-\Theta)_+$, and integrate the resulting equation over
$\mathbb{R}\times[0,t]$ to get
\begin{equation}\label{5.2}
\begin{array}{ll}
\di \int\psi^2(\zeta-\Theta)_+dx+2\mu\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
=\int\psi_0^2(x)(\zeta_0(x)-\Theta(x,0))_+dx\\[2mm]
\di\quad+2\int_0^t\int\frac{R\zeta-p_+\phi}{v}\psi_x(\zeta-\Theta)_+dxds
+2\int_0^t\int_{\Omega_2}\frac{R\zeta-p_+\phi}{v}\psi\zeta_xdxds\\[2mm]
\di\quad-2\int_0^t\int_{\Omega_2}\frac{R\zeta-p_+\phi}{v}\psi\Theta_xdxds
+2\mu\int_0^t\int\frac{\phi U_x}{vV}\psi_x(\zeta-\Theta)_+ dxds\\[2mm]
\di\quad-2\mu\int_0^t\int_{\Omega_2}\frac{\psi\psi_x\zeta_x}{v}dxds
+2\mu\int_0^t\int_{\Omega_2}\frac{\phi\psi U_x}{vV}\zeta_xdxds\\[2mm]
\di\quad+2\mu\int_0^t\int_{\Omega_2}\frac{\psi\psi_x\Theta_x}{v}dxds
-2\mu\int_0^t\int_{\Omega_2}\frac{\phi\psi}{vV}U_x\Theta_xdxds\\[2mm]
\di\quad-2\int_0^t\int\psi\widetilde R_1(\zeta-\Theta)_+ dxds
+\int_0^t\int_{\Omega_2}\psi^2\partial_t\zeta dxds-\int_0^t\int_{\Omega_2}\psi^2\partial_t\Theta dxds.
\end{array}
\end{equation}
Adding \eqref{5.2} into \eqref{5.1}, using $\eqref{perturb}_3$, we have
\begin{equation}
\begin{array}{ll}\label{long}
\di\int\left(\frac{c_{\nu}}{2}(\zeta-\Theta)_+^2+\psi^2(\zeta-\Theta)_+\right)dx
+\mu\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds+\k\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds\\[3mm]
\di=\int\left(\frac{c_{\nu}}{2}(\zeta_0(x)-\Theta(x,0))_+^2+\psi_0^2(x)(\zeta_0(x)-\Theta(x,0))_+\right)dx
+\int_0^t\int\frac{R\zeta-2p_+\phi-R\Theta}{v}\psi_x(\zeta-\Theta)_+ dxds\\[3mm]
\di-\int_0^t\int \frac{R\zeta-p_+\phi}{v}U_x(\zeta-\Theta)_+dxds
+\k\int_0^t\int_{\Omega_2}\frac{\zeta_x\Theta_x}{V}dxds
-\k\int_0^t\int_{\Omega_2}\frac{\phi\Theta_x^2}{vV}dxds
\\[2mm]
\di+2\mu\int_0^t\int\frac{\psi_xU_x}{V}(\zeta-\Theta)_+dxds
-\mu\int_0^t\int\frac{\phi U_ x^2}{vV}(\zeta-\Theta)_+dxds
+2\int_0^t\int_{\Omega_2}\frac{R\zeta-p_+\phi}{v}\psi\zeta_xdxds\\[3mm]
\di-2\int_0^t\int_{\Omega_2}\frac{R\zeta-p_+\phi}{v}\psi\Theta_x dxds
-2\mu\int_0^t\int_{\Omega_2}\frac{\psi\psi_x\zeta_x}{v}dxds
+2\mu\int_0^t\int_{\Omega_2}\frac{\phi\psi U_x}{vV}\zeta_xdxds\\[3mm]
\di+2\mu\int_0^t\int_{\Omega_2}\frac{\psi\psi_x\Theta_x}{v}dxds
-2\mu\int_0^t\int_{\Omega_2}\frac{\phi\psi}{vV}U_x\Theta_xdxds
-2\int_0^t\int\psi\widetilde R_1(\zeta-\Theta)_+dxds\\[3mm]
\di-\int_0^t\int\widetilde R_2(\zeta-\Theta)_+dxds
-c_{\nu}\int_0^t\int\partial_t\Theta(\zeta-\Theta)_+dxds
-\int_0^t\int_{\Omega_2}\psi^2\partial_t\Theta dxds\\[3mm]
\di+\frac{\mu}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left(\frac{\psi_x^2+2\psi_xU_x}{v}-\frac{\phi U_x^2}{vV}\right)dxds
-\frac{1}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\widetilde R_2dxds\\[3mm]
\di-\frac{1}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left (\frac{R\zeta+R\Theta}{v}\psi_x+\frac{R\zeta-p_+\phi}{v}U_x\right)dxds
+\frac{\k}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left(\frac{\t_x}{v}-\frac{\Theta_x}{V}\right)_xdxds\\[3mm]
\di\triangleq \int\left(\frac{c_{\nu}}{2}(\zeta_0(x)-\Theta(x,0))_+^2+\psi_0^2(x)(\zeta_0(x)-\Theta(x,0))_+\right)dx
+\sum_{i=1}^{20}I_i.
\end{array}
\end{equation}
We will estimate \eqref{long} term by term. Recalling \eqref{basic}, \eqref{v} and \eqref{omega-bound}, it holds that
\begin{equation}
\begin{array}{ll}\label{i1}
\di|I_1|&\di=\left|\int_0^t\int\frac{R\zeta-2p_+\phi-R\Theta}{v}\psi_x(\zeta-\Theta)_+ dxds\right|\\[2mm]
&\di\leq \frac{\mu}{4}\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
+C_0\int_0^t\int(\zeta^2+\phi^2+1)(\zeta-\Theta)_+dxds\\[2mm]
&\di\leq \frac{\mu}{4}\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
+C_0\int_0^t\int(\zeta^2+\phi^2\zeta)(\zeta-\Theta)_+dxds\\
&\di\leq \frac{\mu}{4}\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
+C_0\int_0^t\int(\zeta+\phi^2)\left(\zeta-\frac{1}{2}\Theta\right)_+^2dxds\\[2mm]
&\di\leq \frac{\mu}{4}\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
+C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\int_{\{\zeta>\frac{\Theta}{2}\}}(\zeta+\phi^2)dxds\\
&\di\leq \frac{\mu}{4}\int_0^t\int\frac{\psi_x^2}{v}(\zeta-\Theta)_+ dxds
+C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2ds.
\end{array}
\end{equation}
It follows from \eqref{important}, \eqref{basic}, \eqref{phi-x3} and Cauchy's inequality, one has
\begin{equation}
\begin{array}{ll}\label{i3}
\di|I_2|+|I_3|+|I_4|\\
\di\leq C_0\int_0^t\int(\phi^2+\zeta^2)|U_x|dxds+
\k\int_0^t\int_{\Omega_2}\left|\frac{\zeta}{\Theta}\right|\left(\left|\frac{\zeta_x\Theta_x}{V}\right|
+\left|\frac{\phi\Theta_x^2}{vV}\right|\right) dxds\\
\di\leq \frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int(\phi^2+\zeta^2)(|U_x|+\Theta_x^2)dxds\\
\di\leq \frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds+C_0.
\end{array}
\end{equation}
Similarly,
\begin{equation}
\begin{array}{ll}\label{i4}
\di|I_5|+|I_6|+|I_{10}|+|I_{11}|+|I_{12}|\\
\di\leq C_0\int_0^t\int\frac{\psi_x^2}{\t}dxds
+\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds\\
\di\quad+C(C_0,M)\int_0^t\int(\phi^2+\psi^2+\zeta^2)(U_x^2+\Theta_x^2)dxds\\
\di\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds+C_0.
\end{array}
\end{equation}
By Cauchy's inequality, \eqref{basic}, \eqref{v} and \eqref{omega-bound}, it holds that
\begin{equation}
\begin{array}{ll}\label{i7}
\di|I_7|\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int_{\Omega_2}(\zeta^2+\phi^2)\psi^2dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int_{\Omega_2}(\zeta^2\psi^2+\phi^2\psi^4+\phi^2)dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int_{\Omega_2}(\zeta^2(\phi^2+\psi^2)+\phi^2\psi^4)dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int_{\Omega_2}\left(\left(\zeta-\frac{1}{2}\Theta\right)^2(\phi^2+\psi^2)+\phi^2\psi^4\right)dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)\int(\phi^2+\psi^2)dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds.
\end{array}
\end{equation}
Similarly, one has
\begin{equation}
\begin{array}{ll}\label{i8}
\di|I_8|\leq C_0\int_0^t\int_{\Omega_2}(\psi^2+(\phi^2+\zeta^2)\Theta_x^2)dxds\\
\di\quad\leq C_0\int_0^t\int_{\Omega_2}(\zeta^2\psi^2+(\phi^2+\zeta^2)\Theta_x^2)dxds\\
\di\quad\leq C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\int\psi^2dxds
+C_0\int_0^t\int(\phi^2+\zeta^2)\Theta_x^2dxds\\
\di\quad\leq C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2ds+C_0.
\end{array}
\end{equation}
Using Cauchy's inequality, it holds that
\begin{equation}
\begin{array}{ll}\label{i9}
\di|I_9|\leq \frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds
+C_0\int_0^t\int\psi^2\psi_x^2dxds.
\end{array}
\end{equation}
Recalling Lemma \ref{decay}, \eqref{basic} and \eqref{phi-x3}, one has
\begin{equation}
\begin{array}{ll}\label{i10}
\di|I_{13}|+|I_{14}|+|I_{15}|+|I_{16}|+|I_{18}|\\
\di\leq C\d \int_0^t\int (1+s)^{-1}(\psi^2+\zeta^2)e^{-\frac{c_1x^2}{1+s}}dxds
\leq C_0.
\end{array}
\end{equation}
By Cauchy's inequality, and using \eqref{basic}, \eqref{phi-x3} and Lemma \ref{decay}, we obtain
\begin{equation}
\begin{array}{ll}\label{i17}
\di|I_{17}|\leq C_0\int_0^t\int(\psi^2\psi_x^2+\psi^2|\psi_x||U_x|+\psi^2|\phi|U_x^2)dxds\\
\di\quad\leq C_0\int_0^t\int(\psi^2\psi_x^2+\psi^2U_x^2+\psi^2|\phi|U_x^2)dxds\\
\di\quad\leq C_0\int_0^t\int\psi^2\psi_x^2dxds+C(C_0,M)\int_0^t\int\psi^2U_x^2dxds\\
\di\quad\leq C_0\int_0^t\int\psi^2\psi_x^2dxds+C_0.
\end{array}
\end{equation}
Similarly,
\begin{equation}
\begin{array}{ll}\label{i19}
\di|I_{19}|\leq\int_0^t\int_{\Omega_2}(\psi^2\psi_x^2+\psi^2\zeta^2+\psi^4|U_x|+(\zeta^2+\phi^2)|U_x|)dxds\\
\di\leq C_0\int_0^t\int\psi^2\psi_x^2dxds
+C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\int\psi^2dxds\\
\di\quad+\int_0^t\max_{x\in\mathbb{R}}\psi^4\int|U_x|dxds+C_0\\
\di\leq C_0\int_0^t\int\psi^2\psi_x^2dxds
+C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds
+C_0.
\end{array}
\end{equation}
Finally, for
\begin{equation}\label{cut-off}
\di \varphi_{\eta}(z)=\left\{
\begin{array}{ll}
1, & z>\eta,\\
z/\eta, & 0< z\leq \eta,\\
0 & z\leq 0.
\end{array}
\right.
\end{equation}
Integrating by parts shows
\begin{equation}
\begin{array}{ll}\label{i20}
\di I_{20}=\frac{\k}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left(\frac{\t_x}{v}-\frac{\Theta_x}{V}\right)_xdxds\\
\di=\frac{\k}{c_{\nu}}\lim_{\eta\rightarrow0+}\int_0^t\int\varphi_{\eta}(\zeta-\Theta)\psi^2\left(\frac{\zeta_x-\Theta_x}{v}\right)_xdxds\\
\di\quad+\frac{\k}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left(\frac{2\Theta_x}{v}-\frac{\Theta_x}{V}\right)_xdxds
\triangleq I_{20}^1+I_{20}^2.
\end{array}
\end{equation}
We have
\begin{equation}
\begin{array}{ll}\label{i201}
\di I_{20}^1&\di=-\frac{2\k}{c_{\nu}}\lim_{\eta\rightarrow 0+}\int_0^t\int\varphi_{\eta}(\zeta-\Theta)\psi\psi_x\frac{\zeta_x-\Theta_x}{v}dxds\\[3mm]
&\di\quad-\frac{\k}{c_{\nu}}\lim_{\eta\rightarrow 0+}\int_0^t\int\varphi_{\eta}'(\zeta-\Theta)\frac{\psi^2(\zeta_x-\Theta_x)^2}{v}dxds\\[3mm]
&\di\leq C\int_0^t\int_{\Omega_2}\frac{|\psi\psi_x\zeta_x|+|\psi\psi_x\Theta_x|}{v} dxds\\[3mm]
&\di\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds+C_0\int_0^t\int\psi^2\psi_x^2dxds\\[3mm]
&\di\quad+\int_0^t\int\frac{\psi_x^2}{\t v}dxds+C(C_0,M)\int_0^t\int\psi^2\Theta_x^2dxds\\[3mm]
&\di\leq\frac{\k}{8}\int_0^t\int_{\Omega_2}\frac{\zeta_x^2}{v}dxds+C_0\int_0^t\int\psi^2\psi_x^2dxds+C_0,
\end{array}
\end{equation}
where in the second inequality we have used both $\varphi_{\eta}(z)\in [0,1]$ and $\varphi_{\eta}'(z)\geq 0$.
Similarly,
\begin{equation}
\begin{array}{ll}\label{i202}
\di I_{20}^2=\frac{\k}{c_{\nu}}\int_0^t\int_{\Omega_2}\psi^2\left(\frac{2\Theta_{xx}}{v}-\frac{\Theta_{xx}}{V}
-\frac{2\Theta_xv_x}{v^2}+\frac{\Theta_xV_x}{V^2}\right) dxds\\
\di\quad\leq C_0\int_0^t\int\psi^2(|\Theta_{xx}|+\Theta_x^2)+\psi^2|\phi_x||\Theta_x| dxds\\
\di\quad\leq C_0+C_0\int_0^t\max_{x\in\mathbb{R}}\psi^4\left(\int|\Theta_x|dx\right)ds
+C(C_0,M)\d\int_0^t\int \frac{\theta\phi_x^2}{v^3}dxds\\
\di\quad\leq C_0+C_0 \int_0^t\max_{x\in\mathbb{R}}\psi^4ds.
\end{array}
\end{equation}
Noticing that
\begin{equation}
\begin{array}{ll}\label{5.4}
\di\int_0^t\int(\t\psi_x^2+\zeta_x^2)dxds=\int_0^t\int_{\{\zeta>2\Theta\}}(\t\psi_x^2+\zeta_x^2)dxds
+\int_0^t\int_{\{\zeta\leq 2\Theta\}}(\t\psi_x^2+\zeta_x^2)dxds\\[3mm]
\di\quad\leq\int_0^t\int_{\{\zeta>2\Theta\}}\left(\frac{3}{2}\psi_x^2\zeta+C_0\frac{\zeta_x^2}{v}\right) dxds
+\int_0^t\int_{\{\zeta\leq 2\Theta\}}\left(\frac{\psi_x^2}{\t}+\frac{\zeta_x^2}{\t^2}\right)\t^2dxds\\[3mm]
\di\quad\leq\int_0^t\int_{\{\zeta>2\Theta\}}\left(3\psi_x^2(\zeta-\Theta)+C_0\frac{\zeta_x^2}{v}\right) dxds
+C\int_0^t\int_{\{\zeta\leq 2\Theta\}}\left(\frac{\psi_x^2}{\t}+\frac{\zeta_x^2}{\t^2}\right)dxds\\[3mm]
\di\quad\leq C_0\int_0^t\int_{\Omega_2}\left(\frac{\psi_x^2}{v}(\zeta-\Theta)_+ +\frac{\zeta_x^2}{v}\right)dxds+C_0.
\end{array}
\end{equation}
Substituting the estimates \eqref{i1}-\eqref{i202} into \eqref{long}, and using \eqref{5.4}, we have
\begin{equation}
\begin{array}{ll}\label{5.3}
\di \int(\zeta-\Theta)_+^2 dx+\int_0^t\int(\t\psi_x^2+\zeta_x^2)dxds\leq C_0\\
\di\quad +C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds
+C_0\int_0^t\int\psi^2\psi_x^2dxds.
\end{array}
\end{equation}
\
\underline{ Step 2. }\quad To estimate the last term on the right hand side of \eqref{5.3},
we multiply $\eqref{perturb}_2$ by $\psi^3$, and integrate the resulted equation over $\mathbb{R}\times[0,t]$ to get
\begin{equation}
\begin{array}{ll}\label{5.5}
\di\frac{1}{4}\int\psi^4dx+3\mu\int_0^t\int\frac{\psi^2\psi_x^2}{v} dxds
=\frac{1}{4}\int\psi_0^4dx+3R\int_0^t\int\frac{\zeta\psi^2\psi_x}{v}dxds\\[3mm]
\di\quad-3p_+\int_0^t\int\frac{\phi\psi^2\psi_x}{v}dxds+3\mu\int_0^t\int\frac{\phi U_x}{vV}\psi^2\psi_x dxds
-\int_0^t\int\widetilde R_1\psi^3dxds\\
\di\quad\triangleq\frac{1}{4}\int\psi_0^4dx+\sum_{i=1}^4J_i.
\end{array}
\end{equation}
It follows from \eqref{basic} and \eqref{v} that,
\begin{equation}
\begin{array}{ll}\label{j1}
\di|J_1|&\di=3R\int_0^t\int_{\{\zeta>\Theta\}}\frac{\zeta\psi^2\psi_x}{v}dxds
+3R\int_0^t\int_{\{\zeta\leq\Theta\}}\frac{\zeta\psi^2\psi_x}{v}dxds\\[3mm]
&\di\leq \mu\int_0^t\int_{\{\zeta>\Theta\}}\frac{\psi^2\psi_x^2}{v}dxds+C_0\int_0^t\int_{\{\zeta>\Theta\}}\zeta^2\psi^2dxds\\[3mm]
&\di\quad+\int_0^t\int_{\{\zeta\leq\Theta\}}\psi_x^2dxds+C_0\int_0^t\int_{\{\zeta\leq\Theta\}}\zeta^2\psi^4dxds\\[3mm]
&\di\leq\mu\int_0^t\int\frac{\psi^2\psi_x^2}{v}dxds+C_0\int_0^t\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\left(\int\psi^2dx\right)ds\\[3mm]
&\di\quad+C\int_0^t\int_{\{\zeta\leq\Theta\}}\frac{\psi_x^2}{\t}dxds
+C_0\int_0^t\max_{x\in\mathbb{R}}\psi^4\left(\int_{\{\zeta\leq\Theta\}}\zeta^2dx\right)ds\\[3mm]
&\di\leq\mu\int_0^t\int\frac{\psi^2\psi_x^2}{v}dxds
+C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds+C_0.
\end{array}
\end{equation}
Recalling \eqref{basic}, \eqref{v}, and using Cauchy's inequality, it holds that
\begin{equation}
\begin{array}{ll}\label{j2}
\di|J_2|&\di\leq \v\int_0^t\int\psi_x^2dxds+C(\v^{-1},C_0)\int_0^t\int\phi^2\psi^4dxds\\
&\di\leq\v\int_0^t\int\psi_x^2dxds+C(\v^{-1},C_0)\int_0^t\max_{x\in\mathbb{R}}\psi^4\left(\int\phi^2dx\right)ds\\
&\di\leq\v\int_0^t\int\psi_x^2dxds+C(\v^{-1},C_0)\int_0^t\max_{x\in\mathbb{R}}\psi^4ds.
\end{array}
\end{equation}
From \eqref{basic}, \eqref{important} and \eqref{phi-x3}, one has
\begin{equation}
\begin{array}{ll}\label{j3}
\di|J_3|&\di\leq \mu\int_0^t\int\frac{\psi^2\psi_x^2}{v}dxds
+C_0\int_0^t\int\phi^2\psi^2U_x^2dxds\\
&\di\leq\mu\int_0^t\int\frac{\psi^2\psi_x^2}{v}dxds
+C(C_0,M)\int_0^t\int\phi^2U_x^2dxds\\
&\di\leq\mu\int_0^t\int\frac{\psi^2\psi_x^2}{v}dxds+C_0.
\end{array}
\end{equation}
By Lemma \ref{decay}, \eqref{basic}, \eqref{important} and \eqref{phi-x3}, we have
\begin{equation}
\begin{array}{ll}\label{j4}
\di|J_4|&\di\leq O(1)\d\int_0^t\int(1+s)^{-1}|\psi|^3e^{-\frac{c_1x^2}{1+s}}dxds\\
&\di\leq C(M)\d\int_0^t\int(1+s)^{-1}\psi^2e^{-\frac{c_1x^2}{1+s}}dxds\leq C_0.
\end{array}
\end{equation}
Putting the estimates \eqref{j1}-\eqref{j4} into \eqref{5.5} gives
\begin{equation}
\begin{array}{ll}\label{5.6}
\di\int\psi^4dx+\int_0^t\int\psi^2\psi_x^2dxds\leq C_0+C_0\v\int_0^t\int\psi_x^2dxds\\
\di\qquad +C(\v^{-1},C_0)\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds.
\end{array}
\end{equation}
Noticing that
\begin{equation}\label{psi-x}
\di 2\int_0^t\int\psi_x^2dxds\leq \int_0^t\int\frac{\psi_x^2}{\t}dxds+\int_0^t\int\t\psi_x^2dxds
\leq C_0+C_0\int_0^t\int\t\psi_x^2dxds.
\end{equation}
Combining \eqref{5.3} and \eqref{5.6}, choosing $\v$ suitable small, we have
\begin{equation}
\begin{array}{ll}\label{5.7}
\di \int((\zeta-\Theta)_+^2+\psi^4) dx+\int_0^t\int((\psi^2+\t)\psi_x^2+\zeta_x^2)dxds\\
\di\quad \leq C_0+C_0\int_0^t\left(\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\max_{x\in\mathbb{R}}\psi^4\right)ds.
\end{array}
\end{equation}
\
\underline{ Step 3. }\quad It remains to estimate the last two terms on the right hand side of \eqref{5.7}.
For $x\in\mathbb{R}$,
\begin{equation}
\begin{array}{ll}
\di\left(\zeta-\frac{1}{2}\Theta\right)_+^2=\int_{-\infty}^x2\left(\zeta-\frac{1}{2}\Theta\right)_+\left(\zeta_x-\frac{1}{2}\Theta_x\right)dx\\[3mm]
\di\quad\leq C\int\left(\zeta-\frac{1}{2}\Theta\right)_+(|\zeta_x|+|\Theta_x|)dx\\[3mm]
\di\quad\leq \v\int\left(\zeta-\frac{1}{2}\Theta\right)_+^2\t dx
+\frac{C}{\v}\int_{\{\zeta>\frac{\Theta}{2}\}}\left(\frac{\zeta_x^2}{\t}+\frac{\Theta_x^2}{\t}\right)dx\\[4mm]
\di\quad\leq 3\v\int\left(\zeta-\frac{1}{2}\Theta\right)_+^2\zeta dx
+\frac{C}{\v}\int\frac{\zeta_x^2}{\t}dx+\frac{C}{\v}\int_{\{\zeta>\frac{\Theta}{2}\}}\zeta^2\Theta_x^2dx\\[3mm]
\di\quad\leq 3\v\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\int_{\{\zeta>\frac{\Theta}{2}\}}\zeta dx
+\frac{C}{\v}\int\frac{\zeta_x^2}{\t}dx+\frac{C}{\v}\int\zeta^2\Theta_x^2dx\\[3mm]
\di\quad\leq \v C_0\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2+\frac{C}{\v}\int\frac{\zeta_x^2}{\t}dx
+\frac{C}{\v}\int\zeta^2\Theta_x^2dx.
\end{array}
\end{equation}
This yields
\begin{equation}\label{max-theta}
\di\max_{x\in\mathbb{R}}\left(\zeta-\frac{1}{2}\Theta\right)_+^2\leq C_0\int\frac{\zeta_x^2}{\t}dx
+C_0\int\zeta^2\Theta_x^2dx.
\end{equation}
\begin{equation}
\begin{array}{ll}
\di\psi^4=\int_{-\infty}^x4\psi^3\psi_xdx\leq4\int_{\{\zeta>\Theta\}}|\psi|^3|\psi_x|dx
+4\int_{\{\zeta\leq\Theta\}}|\psi|^3|\psi_x|dx\\[3mm]
\di\quad\leq\v\int_{\{\zeta>\Theta\}}|\psi|^5\sqrt{\t}dx+\frac{C}{\v}\int_{\{\zeta>\Theta\}}\psi_x^2\frac{|\psi|}{\sqrt{\t}}dx
+\v\int_{\{\zeta\leq\Theta\}}\psi^6\t dx+\frac{C}{\v}\int_{\{\zeta\leq\Theta\}}\frac{\psi_x^2}{\t}dx\\[3mm]
\di\quad\leq\v\max_{x\in\mathbb{R}}\psi^4\int_{\{\zeta>\Theta\}}(\psi^2+\theta)dx
+C\v\max_{x\in\mathbb{R}}\psi^4\int_{\{\zeta\leq\Theta\}}\psi^2dx
+\frac{C}{\v}\int\psi_x^2\left(\frac{|\psi|}{\sqrt{\t}}+\frac{1}{\t}\right)dx\\[3mm]
\di\quad\leq \v C_0\max_{x\in\mathbb{R}}\psi^4+\frac{C}{\v}\int\psi_x^2\left(\frac{|\psi|}{\sqrt{\t}}+\frac{1}{\t}\right)dx,
\end{array}
\end{equation}
which directly gives
\begin{equation}\label{max-psi}
\di\max_{x\in\mathbb{R}}\psi^4\leq C_0\int\psi_x^2
\left(\frac{|\psi|}{\sqrt{\t}}+\frac{1}{\t}\right)dx.
\end{equation}
Substituting \eqref{max-theta} and \eqref{max-psi} into \eqref{5.7},
recalling \eqref{basic}, \eqref{important} and \eqref{phi-x3}, and choosing $\v$ suitable small, it holds that
\begin{equation}
\begin{array}{ll}\label{5.8}
\di \sup_{0\leq t\leq T}\int((\zeta-\Theta)_+^2+\psi^4) dx+\int_0^T\int((\psi^2+\t)\psi_x^2+\zeta_x^2)dxdt\\
\di\quad \leq C_0+C_0\int_0^T\int\left(\frac{\zeta_x^2}{\t}+\psi_x^2\left(\frac{|\psi|}{\sqrt{\t}}+\frac{1}{\t}\right)\right)dxdt\\
\di\quad\leq C_0+\frac{1}{2}\int_0^T\int(\zeta_x^2+\psi^2\psi_x^2)dxdt
+C_0\int_0^T\int\left(\frac{\zeta_x^2}{\t^2}+\frac{\psi_x^2}{\t}\right)dxdt\\
\di\quad\leq C_0+\frac{1}{2}\int_0^T\int(\zeta_x^2+\psi^2\psi_x^2)dxdt.
\end{array}
\end{equation}
From\eqref{basic}, \eqref{omega-bound}, we have
\begin{equation}\label{5.9}
\di\int_{\{\zeta\leq2\Theta\}}\zeta^2dx\leq C\int_{\mathbb{R}}\Phi\left(\frac{\t}{\Theta}\right)dx\leq C_0,
\end{equation}
and
\begin{equation}\label{5.10}
\di\int_{\{\zeta>2\Theta\}}\zeta^2dx\leq4\int_{\{\zeta>2\Theta\}}(\zeta-\Theta)^2dx
\leq 4\int(\zeta-\Theta)_+^2dx.
\end{equation}
Thus combining \eqref{5.8}-\eqref{5.10}, the proof of Lemma \ref{lemma-ll-1} is completed.
\hfill $\Box$
\
\begin{lemma}\label{high-order}
Suppose that $(\phi,\psi,\zeta)\in X([0,T])$ satisfies $\d=|\t_+-\t_-|\leq\d_0$
with suitable small $\d_0$, it holds
\begin{equation}\label{high-deri}
\di\sup_{0\leq t\leq T}\int(\phi_x^2+\psi_x^2+\zeta_x^2)dx
+\int_0^T\int(\t\phi_x^2+\psi_{xx}^2+\zeta_{xx}^2)dxdt\leq C_0.
\end{equation}
\end{lemma}
\textbf{Proof}:\quad Due to \eqref{basic}, \eqref{v} and \eqref{new-nergy}, some terms of \eqref{phi-x1} can be considered more carefully, that is,
\begin{equation}
\begin{array}{ll}
\di \left|\frac{R}{v}\zeta_x\frac{\tilde v_x}{\tilde v}\right|
+\left|\psi_xU_x\left(\frac{1}{v}-\frac{1}{V}\right)\right|+\frac{\psi_x^2}{v}\\[3mm]
\di \leq \frac{R\t}{4v}\left(\frac{\tilde v_x}{\tilde v}\right)^2+
C_0\frac{\zeta_x^2}{\t}+C_0\psi_x^2+C_0\phi^2U_x^2\\[3mm]
\di \leq \frac{R\t}{4v}\left(\frac{\tilde v_x}{\tilde v}\right)^2+
C_0\left(\zeta_x^2+\frac{\zeta_x^2}{\t^2}+\t\psi_x^2+\frac{\psi_x^2}{\t}\right)
+C_0\phi^2U_x^2.
\end{array}
\end{equation}
The other terms in \eqref{phi-x1} can be estimated the same as in step 2 in Lemma \ref{basic-lemma}.
Integrating \eqref{phi-x1} over $\mathbb{R}\times(0,t)$, recalling \eqref{basic} and \eqref{new-nergy}, we have
\begin{equation}\label{new-phi-x}
\di \sup_{0\leq t\leq T}\int\phi_x^2 dx+\int_0^T\int\t\phi_x^2dxdt\leq C_0.
\end{equation}
Multiplying $\eqref{perturb}_2$ by $-\psi_{xx}$, integrating the resulted equation over $\mathbb{R}\times(0,t)$,
and noticing that
\begin{equation*}
\begin{array}{ll}
\di(p-p_+)_x&\di=\left(\frac{R\zeta-p_+\phi}{v}\right)_x=\frac{R\zeta_x-p_+\phi_x}{v}
-\frac{R\zeta-p_+\phi}{v^2}\phi_x-\frac{R\zeta-p_+\phi}{v^2}V_x\\[3mm]
&\di=\frac{R\zeta_x}{v}-\frac{R\t\phi_x}{v^2}-\frac{R\zeta-p_+\phi}{v^2}V_x.
\end{array}
\end{equation*}
Then we have
\begin{equation}
\begin{array}{ll}\label{6.1}
\di\int\frac{\psi_x^2}{2}dx+\mu\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
=\int\frac{\psi_{0x}^2}{2}dx+\int_0^t\int\left(\frac{R\zeta_x}{v}-\frac{R\t\phi_x}{v^2}
-\frac{R\zeta-p_+\phi}{v^2}V_x\right)\psi_{xx}dxds\\
\di\quad-\mu\int_0^t\int\psi_x\left(\frac{1}{v}\right)_x\psi_{xx}dxds
+\mu\int_0^t\int\left(\frac{U_x}{V}-\frac{U_x}{v}\right)_x\psi_{xx} dxds
+\int_0^t\int\widetilde R_1\psi_{xx}dxds.
\end{array}
\end{equation}
In the following, each term on the right hand side of \eqref{6.1} will be estimated.
From \eqref{important}, \eqref{new-nergy} and \eqref{new-phi-x}, one has
\begin{equation}
\begin{array}{ll}\label{psi-xx-1}
\di\left|\int_0^t\int\left(\frac{R\zeta_x}{v}-\frac{R\t\phi_x}{v^2}
-\frac{R\zeta-p_+\phi}{v^2}V_x\right)\psi_{xx}dxds\right|\\
\di\leq \frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\int\zeta_x^2+\theta^2\phi_x^2+(\phi^2+\zeta^2)\Theta_x^2dxds\\
\di\leq \frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0+\max_{x,t}\t\int_0^t\int\t\phi_x^2dxds\\
\di\leq \frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0+C_0\max_{x,t}\t.
\end{array}
\end{equation}
By Cauchy's inequality and Sobolev's inequality, and recalling \eqref{basic}, \eqref{new-nergy},
\eqref{psi-x} and \eqref{new-phi-x}, we obtain
\begin{equation}
\begin{array}{ll}\label{psi-xx-2}
\di\left|\mu\int_0^t\int\psi_x\left(\frac{1}{v}\right)_x\psi_{xx}dxds\right|
\leq C\int_0^t\int\left|\frac{\psi_x\phi_x\psi_{xx}}{v^2}\right|+\left|\frac{\psi_xV_x\psi_{xx}}{v^2}\right|dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\int \psi_x^2\phi_x^2+\psi_x^2\Theta_x^2dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\|\psi_x\|^2_{L^{\infty}}\|\phi_x\|^2ds
+C_0\int_0^t\int\psi_x^2dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\|\psi_x\|\|\psi_{xx}\|ds
+C_0\int_0^t\int\psi_x^2dxds\\
\di\leq\frac{\mu}{4}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\int\psi_x^2dxds\\
\di\leq\frac{\mu}{4}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds+C_0.
\end{array}
\end{equation}
Similarly,
\begin{equation}
\begin{array}{ll}\label{psi-xx-3}
\di\mu\int_0^t\int\left(\frac{U_x}{V}-\frac{U_x}{v}\right)_x\psi_{xx} dxds\\
\di=\mu\int_0^t\int\left(\frac{U_{xx}}{V}-\frac{U_{xx}}{v}-\frac{U_xV_x}{V^2}+\frac{v_xU_x}{v^2}\right)\psi_{xx}dxds\\
\di=\mu\int_0^t\int\left(\frac{\phi U_{xx}}{vV}-U_xV_x\frac{\phi(\phi+2V)}{v^2V^2}+\frac{\phi_xU_x}{v^2}\right)\psi_{xx}dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0\int_0^t\int(\phi^2(U_{xx}^2+\Theta_x^2U_x^2)+\phi^4U_x^2\Theta_x^2+\phi_x^2U_x^2)dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds
+C_0+C(C_0,M)\d^2\int_0^t\int\t\phi_x^2dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds+C_0,
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}\label{psi-xx-4}
\di\left|\int_0^t\int\widetilde R_1\psi_{xx}dxds\right|
\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds+C_0\int_0^t\int\widetilde R_1^2dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds+C_0\d^2\int_0^t\int(1+s)^{-3}e^{-\frac{c_1x^2}{1+s}}dxds\\
\di\leq\frac{\mu}{8}\int_0^t\int\frac{\psi_{xx}^2}{v}dxds+C_0.
\end{array}
\end{equation}
Substituting \eqref{psi-xx-1}-\eqref{psi-xx-4} into \eqref{6.1} shows
\begin{equation}\label{psi-xx}
\di\sup_{0\leq t\leq T}\int\psi_x^2dx+\int_0^T\int\psi_{xx}^2dxdt
\leq C_0+C_0\max_{x,t}\t.
\end{equation}
Multiplying $\eqref{perturb}_3$ by $-\zeta_{xx}$, then integrating the resulted equation over
$\mathbb{R}\times(0,t)$, we have
\begin{equation}
\begin{array}{ll}\label{6.2}
\di\frac{c_{\nu}}{2}\int\zeta_x^2dx+\k\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds
=\frac{c_{\nu}}{2}\int\zeta_{0x}^2dx+\int_0^t\int(pu_x-p_+U_x)\zeta_{xx}dxds\\[3mm]
\di\quad-\k\int_0^t\int\zeta_x\left(\frac{1}{v}\right)_x\zeta_{xx}dxds
-\k\int_0^t\int\left(\frac{\Theta_x}{v}-\frac{\Theta_x}{V}\right)_x\zeta_{xx}dxds\\[3mm]
\di\quad-\mu\int_0^t\int\left(\frac{u_x^2}{v}-\frac{U_x^2}{V}\right)\zeta_{xx}dxds
+\int_0^t\int\widetilde R_2\zeta_{xx}dxds.
\end{array}
\end{equation}
We will estimate \eqref{6.2} one by one. First, we have
\begin{equation}
\begin{array}{ll}\label{zeta-xx-1}
\di\int_0^t\int(pu_x-p_+U_x)\zeta_{xx}dxds=
\int_0^t\int\left(\frac{R\t}{v}\psi_x+\frac{R\zeta-p_+\phi}{v}U_x\right)\zeta_{xx}dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\int(\theta^2\psi_x^2+(\phi^2+\zeta^2)U_x^2)dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\max_{x,t}\t\int_0^t\int\theta\psi_x^2dxds+
C_0\int_0^t\int(\phi^2+\zeta^2)U_x^2dxds\\
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\max_{x,t}\t+C_0.
\end{array}
\end{equation}
It follows from Cauchy's inequality, \eqref{new-nergy} and \eqref{new-phi-x} that
\begin{equation}
\begin{array}{ll}\label{zeta-xx-2}
\di-\k\int_0^t\int\zeta_x\left(\frac{1}{v}\right)_x\zeta_{xx}dxds
\leq C\int_0^t\int\frac{|\zeta_x\phi_x\zeta_{xx}|+|\zeta_xV_x\zeta_{xx}|}{v^2} dxds\\[3mm]
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\int\zeta_x^2\phi_x^2+\zeta_x^2\Theta_x^2dxds\\[3mm]
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\sup_{0\leq s\leq t}\|\phi_x\|^2\int_0^t\|\zeta_{x}\|\|\zeta_{xx}\|ds+
C_0\int_0^t\int\zeta_x^2dxds\\[3mm]
\di\quad\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\|\zeta_{x}\|\|\zeta_{xx}\|ds+C_0\\[3mm]
\di\quad\leq\frac{\k}{4}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+C_0.
\end{array}
\end{equation}
Recalling Lemma \ref{decay} and \eqref{new-nergy}, and choosing $\delta$ suitable small, we have
\begin{equation}
\begin{array}{ll}\label{zeta-xx-3}
\di
-\k\int_0^t\int\left(\frac{\Theta_x}{v}-\frac{\Theta_x}{V}\right)_x\zeta_{xx}dxds\\[3mm]
\di=-\k\int_0^t\int\left(\frac{\Theta_{xx}}{v}-\frac{\Theta_{xx}}{V}
-\frac{\Theta_{x}v_x}{v^2}+\frac{\Theta_{x}V_x}{V^2}\right)\zeta_{xx}dxds\\[3mm]
\di=-\k\int_0^t\int\left(\frac{-\phi\Theta_{xx}}{vV}-\frac{\phi_x\Theta_x}{v^2}
+\frac{\phi(\phi+2V)}{v^2V^2}\Theta_xV_x\right)\zeta_{xx}dxds\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\int(\phi^2(\Theta_{xx}^2+\Theta_x^4)+\phi_x^2\Theta_x^2+\phi^4\Theta_x^4)dxds\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0+C(C_0,M)\d^2\int_0^t\int\t\phi_x^2dxds+C(C_0,M)\int_0^t\int\phi^2\Theta_x^4dxds\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+C_0.
\end{array}
\end{equation}
Similarly,
\begin{equation}
\begin{array}{ll}\label{zeta-xx-4}
\di-\mu\int_0^t\int\left(\frac{u_x^2}{v}-\frac{U_x^2}{V}\right)\zeta_{xx}dxds=
-\mu\int_0^t\int\left(\frac{\psi_x^2+2\psi_xU_x}{v}-\frac{\phi U_x^2}{vV}\right)\zeta_{xx}dxds\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\int(\psi_x^4+\psi_x^2U_x^2+\phi^2U_x^4)dxds\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\int_0^t\|\psi_x\|^3\|\psi_{xx}\|ds+C_0\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
\int_0^t\int\psi_{xx}^2dxds+C_0\sup_{0\leq s\leq t}\|\psi_x\|^4\int_0^t\int\psi_x^2dxds+C_0\\[3mm]
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+
C_0\max_{x,t}\t^2+C_0,
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}\label{zeta-xx-5}
\di
\int_0^t\int\widetilde R_2\zeta_{xx}dxds\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds
+C_0\int_0^t\int\widetilde R_2^2dxds\\
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds
+C_0\d^2\int_0^t\int(1+s)^{-4}e^{-\frac{c_1x^2}{1+s}}dxds\\
\di\leq\frac{\k}{8}\int_0^t\int\frac{\zeta_{xx}^2}{v}dxds+C_0.
\end{array}
\end{equation}
Substituting estimates \eqref{zeta-xx-1}-\eqref{zeta-xx-5} into \eqref{6.2} shows
\begin{equation}\label{zeta-xx}
\di\sup_{0\leq t\leq T}\int\zeta_x^2dx+\int_0^T\int\zeta_{xx}^2dxdt
\leq C_0+C_0\max_{x,t}\t^2.
\end{equation}
By Sobolev's inequality and \eqref{new-nergy}, \eqref{zeta-xx}, we have
\begin{equation}
\di \|\zeta\|^2_{L^{\infty}}\leq C\|\zeta\|\|\zeta_x\|\leq C_0+C_0\max_{x,t}\t.
\end{equation}
Noticing that
\begin{equation}
\di\max_{x,t}\t^2\leq 2\max_{x,t}\zeta^2+2\max_{x,t}\Theta^2\leq C_0+C_0\max_{x,t}\t.
\end{equation}
This yields
\begin{equation}\label{theta-upper}
\max_{x,t}\t\leq C_0,
\end{equation}
which, together with \eqref{new-phi-x}, \eqref{psi-xx} and \eqref{zeta-xx},
completes the proof of Lemma \ref{high-order}.
\hfill $\Box$
\
Finally, it follows from \eqref{new-nergy}, \eqref{high-deri} and equation $\eqref{perturb}_3$ that
$$
\di\int_0^{+\infty}\|\zeta_x\|^2+\left|\frac{d}{dt}\|\zeta_x\|^2\right| dt\leq C_0,
$$
which, together with the Sobolev's inequality gives
\begin{equation}
\di\lim_{t\rightarrow \infty}\|\zeta\|^2_{L^{\infty}}
\leq C\lim_{t\rightarrow \infty}\|\zeta\|\|\zeta_x\|
\leq C_0\lim_{t\rightarrow \infty}\|\zeta_x\|=0.
\end{equation}
Hence there exists some $T_0>0$ such that for all $(x,t)\in\mathbb{R}\times[T_0,+\infty)$, it holds that
$$
-\frac{\t_-}{2}<\zeta<\frac{\t_+}{2}.
$$
This directly yields, for all $(x,t)\in\mathbb{R}\times[T_0,+\infty)$,
\begin{equation}\label{theta-}
\di \theta=\Theta+\zeta>\theta_- -\frac{\theta_-}{2}=\frac{\theta_-}{2},
\end{equation}
and
\begin{equation}
\di \theta=\Theta+\zeta<\theta_+ +\frac{\theta_+}{2}=\frac{3\theta_+}{2}.
\end{equation}
Finally, it follows from $\eqref{ns}_3$ that,
\begin{equation}
\di \t_t+\frac{R\t}{c_{\nu}}\frac{u_x}{v}-\frac{\k}{c_{\nu}}\frac{\t_{xx}}{v}+\frac{\k}{c_{\nu}}\frac{\t_xv_x}{v^2}\geq 0.
\end{equation}
Define
\begin{equation}
\di \bar\t=\t \exp\left(\frac{R}{c_{\nu}}\int_0^t\left\|\frac{u_x}{v}\right\|_{L^{\infty}}ds\right).
\end{equation}
We find
\begin{equation}
\di \bar\t_t-\frac{\k}{c_{\nu}}\frac{\bar\t_{xx}}{v}+\frac{\k}{c_{\nu}}\frac{\bar\t_xv_x}{v^2}\geq
\frac{R\bar\t}{c_{\nu}}\Big(\left\|\frac{u_x}{v}\right\|-\frac{u_x}{v}
\Big)\geq 0.
\end{equation}
By the minimum principle of the parabolic equation, we obtain
\begin{equation}
\di\inf_{x,t}\bar\t\geq\inf_{x,t}\bar\t\Big|_{t=0}=\inf_{x\in\mathbb{R}}\t_0\geq m^{-1}_0,
\end{equation}
which directly yields
\begin{equation}
\begin{array}{ll}\label{thata-lower}
\di\inf_{x,t}\t(x,t)&\di\geq m^{-1}_0e^{-\frac{R}{c_{\nu}}\int_0^t\|\frac{u_x}{v}\|_{L^{\infty}}ds}\\
&\di\geq m_0^{-1}e^{-C_0\int_0^t\|\psi_x\|_{L^{\infty}}+\|U_x\|_{L^{\infty}}ds}\\
&\di\geq m_0^{-1} e^{-C_0\int_0^t\|\psi_x\|^{1/2}\|\psi_{xx}\|^{1/2}}e^{-C_0\int_0^t\|U_x\|_{L^{\infty}}ds}\\
&\di\geq m_0^{-1} e^{-C_0\int_0^t\|\psi_{xx}\|^{1/2}}e^{-C_0\d t}\\
&\di\geq m_0^{-1} e^{-C_0(t^{3/4}+\d t)}\geq C_0e^{-C_0t}.
\end{array}
\end{equation}
Thus from \eqref{theta-}, we have $\t>\min\{\frac{\t_-}{2},C_0e^{-C_0T_0}\}$
for all $(x,t)\in\mathbb{R}\times(0,+\infty)$. By Lemma \ref{basic-lemma},
\ref{v-bound}-\ref{high-order}, Proposition \ref{prop} is completed.
\hfill $\Box$
\
\section{Proof of Theorem \ref{theorem2}}
\setcounter{equation}{0}
It is sufficient to show the same a priori estimate as Proposition \ref{prop}. Noticing that
$(V_{\pm}^r,U_{\pm}^r,\T_{\pm}^r)$ satisfies Euler system \eqref{euler} and $(V^{cd},U^{cd},\T^{cd})$
satisfies $\eqref{ns}_1$ \eqref{Theta t}, we rewrite the Cauchy problem \eqref{ns}\eqref{initial} as
\begin{equation}\label{perturb2}
\left\{
\begin{array}{ll}
\di \phi_t-\psi_x=0,\\
\di \psi_t+\left(p-P\right)_x=\mu\left(\frac{u_x}{v}-\frac{U_x}{V}\right)_x+F,\\
\di c_{\nu}\zeta_t+pu_x-PU_x=\kappa\left(\frac{\t_x}{v}-\frac{\Theta_x}{V}\right)_x
+\mu\left(\frac{u_x^2}{v}-\frac{U^2_x}{V}\right)+G,\\
\di(\phi,\psi,\zeta)(\pm\infty,t)=0,\\
\di (\phi,\psi,\zeta)(x,0)=(\phi_0,\psi_0,\zeta_0)(x),\quad x\in \mathbb{R},
\end{array}
\right.
\end{equation}
where
$$
P=\frac{R\T}{V},\quad P_{\pm}=\frac{R\T_{\pm}^r}{V_{\pm}^r},
$$
$$
F=(P_-+P_+-P)_x+\left(\mu\frac{U_x}{V}\right)_x-U_t^{cd},
$$
and
$$
\begin{array}{ll}
G=(p^m-P)U^{cd}_x+(P_--P)(U_-^r)_x+(P_+-P)(U_+^r)_x\\
\di\quad\quad+\mu\frac{U_x^2}{V}+\kappa\left(\frac{\T_x}{V}-\frac{\T_x^{cd}}{V^{cd}}\right)_x.
\end{array}
$$
Similar to Lemma \ref{basic-lemma}, the following key estimate holds.
\begin{lemma}\label{basic2}
For $(\phi,\psi,\zeta)\in X([0,T])$, we assume \eqref{same-order} holds, then there exist some positive
constants $C_0$ and $\d_0$ such that if $\d<\d_0$, it follows that for $t\in[0,T]$,
\begin{equation}
\begin{array}{ll}
\di \Big\|\left(\psi,\sqrt{\Phi\left(\frac{v}{V}\right)},\sqrt{\Phi\left(\frac{\t}{\Theta}\right)}\right)(t)\Big\|^2
+\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right)dxds\\[3mm]
\di+\int_0^t\int P\left(\Phi\left(\frac{\t V}{\T v}\right)+\g\Phi\left(\frac{v}{V}\right)\right)\big((U_-^r)_x+(U_+^r)_x\big)dxds
\di\leq C_0,
\end{array}
\end{equation}
where $C_0$ denotes a constant depending only on $\mu$, $\kappa$, $R$, $c_{\nu}$,
$v_{\pm}$, $u_{\pm}$, $\t_{\pm}$ and $m_0$.
\end{lemma}
\textbf{Proof}: First, multiplying $\eqref{perturb2}_2$ by $\psi$ leads to
\begin{equation}\label{4.17}
\begin{array}{ll}
\di \left(\frac{\psi^2}{2}\right)_t+\left[(p-P)\psi-\mu\left(\frac{u_x}{v}-\frac{U_x}{V}\right)\psi\right]_x
-\frac{R\zeta}{v}\psi_x\\
\di-R\T\left(\frac{1}{v}-\frac{1}{V}\right)\phi_t+\mu\frac{\psi_x^2}{v}
+\mu\left(\frac{1}{v}-\frac{1}{V}\right)U_x\psi_x=F\psi.
\end{array}
\end{equation}
Next, we multiply $\eqref{perturb2}_3$ by $\z\t^{-1}$ to get
\begin{equation}\label{4.18}
\begin{array}{ll}
\di\frac{R}{\g-1}\frac{\z\z_t}{\t}-\left[\k\left(\frac{\t_x}{v}-\frac{\T_x}{V}\right)\frac{\z}{\t}\right]_x
+\frac{R\z}{v}\psi_x+(p-P)U_x\frac{\z}{\t}\\[2mm]
\di+\k\frac{\T\z_x^2}{\t^2v}-\k\frac{\z_x\z\T_x}{\t^2v}-\k\frac{\phi\T_x\T\z_x}{\t^2vV}
+\k\frac{\phi\z\T_x^2}{\t^2vV}\\[2mm]
\di-\mu\frac{\z\psi_x^2}{\t v}-2\mu\frac{\psi_xU_x\z}{\t v}+\mu\frac{\phi\z U_x^2}{\t vV}=G\frac{\z}{\t}.
\end{array}
\end{equation}
Noticing that
\begin{equation}\label{4.19}
\di-R\T\left(\frac{1}{v}-\frac{1}{V}\right)\phi_t=\left[R\T\Phi\left(\frac{v}{V}\right)\right]_t
-R\T_t\Phi\left(\frac{v}{V}\right)+\frac{P\phi^2}{vV}V_t,
\end{equation}
\begin{equation}
\di\frac{\z\z_t}{\t}=\left[\T\Phi\left(\frac{\t}{\T}\right)\right]_t+\T_t\Phi\left(\frac{\T}{\t}\right),
\end{equation}
\begin{equation}
\begin{array}{ll}
-R\T_t&\di=(\g-1)P_-(U_-^r)_x+(\g-1)P_+(U_+^r)_x-p^mU_x^{cd}\\[2mm]
&\di=(\g-1)P(U_-^r)_x+(\g-1)P(U_+^r)_x+(\g-1)(P_--P)(U_-^r)_x\\[2mm]
&\di+(\g-1)(P_+-P)(U_+^r)_x-p^mU_x^{cd},
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
\di-R\T_t\Phi\left(\frac{v}{V}\right)+\frac{P\phi^2}{vV}V_t+\frac{R}{\g-1}\T_t\Phi\left(\frac{\T}{\t}\right)
+(p-P)U_x\frac{\z}{\t}\\
=Q_1\big((U_-^r)_x+(U_+^r)_x\big)+Q_2,
\end{array}
\end{equation}
where
\begin{equation}
\begin{array}{ll}
Q_1&\di=(\g-1)P\Phi\left(\frac{v}{V}\right)+\frac{P\phi^2}{vV}-P\Phi\left(\frac{\T}{\t}\right)
+\frac{\z}{\t}(p-P)\\
&\di=P\left(\Phi\left(\frac{\t V}{\T v}\right)+\g\Phi\left(\frac{v}{V}\right)\right),
\end{array}
\end{equation}
and
\begin{equation}
\begin{array}{ll}\label{q2}
Q_2&\di=U_x^{cd}\left(\frac{P\phi^2}{vV}-p^m\Phi\left(\frac{v}{V}\right)+\frac{p^m}{\g-1}\Phi\left(\frac{\T}{\t}\right)
+\frac{\z}{\t}(p-P)\right)\\[3mm]
&\di+(\g-1)(P_--P)(U_-^r)_x\left(\Phi\left(\frac{v}{V}\right)-\frac{1}{\g-1}\Phi\left(\frac{\T}{\t}\right)\right)\\[3mm]
&\di+(\g-1)(P_+-P)(U_+^r)_x\left(\Phi\left(\frac{v}{V}\right)-\frac{1}{\g-1}\Phi\left(\frac{\T}{\t}\right)\right).
\end{array}
\end{equation}
Combining \eqref{4.18} and \eqref{4.19}, it follows from \eqref{4.19}-\eqref{q2} that
\begin{equation}\label{2-basic}
\begin{array}{ll}
\di\left(\frac{\psi^2}{2}+R\Theta\Phi\left(\frac{v}{V}\right)+\frac{R}{\g-1}\Theta\Phi\left(\frac{\theta}{\Theta}\right)\right)_t
+\frac{\mu\Theta}{\theta v}\psi_x^2+\frac{\kappa\Theta}{\theta^2 v}\zeta_x^2\\
\di+H_{x}+Q_1\big((U_-^r)_x+(U_+^r)_x\big)+\tilde Q=F\psi+G\frac{\zeta}{\theta}
\end{array}
\end{equation}
with $H$ the same as in \eqref{H}, and
\begin{equation}\label{tilde-q}
\begin{array}{ll}
\tilde Q&\di=Q_2-\k\frac{\z_x\z\T_x}{\t^2v}-\k\frac{\phi\T_x\T\z_x}{\t^2vV}
+\k\frac{\phi\z\T_x^2}{\t^2vV}\\[2mm]
&\di-\mu\frac{\phi U_x\psi_x}{vV}-2\mu\frac{\psi_xU_x\z}{\t v}+\mu\frac{\phi\z U_x^2}{\t vV}.
\end{array}
\end{equation}
Recalling (iii) in Lemma \ref{rare-pro}, we can compute
\begin{equation}
\begin{array}{ll}
\di|(P_--P)(U_-^r)_x|\\
\di\leq C\big(|\T^{cd}-\t_-^m|+|\T^{r}_+-\t_+^m|+|V^{cd}-v_-^m|+|V^{r}_+-v_+^m|\big)|(U_-^r)_x|\\
\di\leq C\big(|\T^{cd}-\t_-^m|+|\T^{r}_+-\t_+^m|+|V^{cd}-v_-^m|+|V^{r}_+-v_+^m|\big)\big|_{\Omega_-}
+C|(U_-^r)_x|\big|_{\Omega_c\cap\Omega_+}\\
\di\leq C\d e^{-c_0(|x|+t)},
\end{array}
\end{equation}
which leads to
\begin{equation}
\di|Q_2|\leq C(M)|U_x^{cd}|(\phi^2+\zeta^2)+C(M)\d e^{-c_0(|x|+t)}(\phi^2+\zeta^2)
\end{equation}
and
\begin{equation}
\di|\tilde Q|\leq |Q_2|+\frac{\mu\T\psi_x^2}{4\t v}+\frac{\k\T\zeta_x^2}{4\t^2 v}
+C(M)(\phi^2+\zeta^2)(\T_x^2+U_x^2).
\end{equation}
Note that
\begin{equation}
\begin{array}{ll}
(\phi^2+\zeta^2)(\T_x^2+U_x^2)\leq C(\phi^2+\zeta^2)((\T_x^{cd})^2+(U^r_-)_x^2+(U^r_+)_x^2)\\
\quad \leq C(1+t)^{-1}(\phi^2+\zeta^2)\e^{-\frac{c_1x^2}{1+t}}+C\d Q_1((U^r_-)_x+(U^r_+)_x).
\end{array}
\end{equation}
Following the same calculations as in \cite{Huang-Li-matsumura}, it holds that
\begin{equation}
\di \|(F,G)\|_{L^1}\leq C\d^{1/8}(1+t)^{-7/8}.
\end{equation}
Then we have
\begin{equation}\label{int-F}
\begin{array}{ll}
\di\int_0^t\int\left(F\psi+G\frac{\z}{\t}\right)dxds\leq C(M)\int_0^t\|(F,G)\|_{L^1}\|(\psi,\zeta)\|_{L^{\infty}}ds\\
\di\quad\leq C(M)\d^{1/8}\int_0^t(1+s)^{-7/8}\|(\psi,\zeta)\|^{1/2}\|(\psi_x,\zeta_x)\|^{1/2}ds\\
\di\quad\leq \int_0^t\int\left(\frac{\mu\T\psi_x^2}{4\t v}+\frac{\k\T\zeta_x^2}{4\t^2v}\right)dxds\\
\di\quad+C(M)\d^{1/6}\int_0^t(1+s)^{-7/6}\left(1+\left\|\left(\psi,\sqrt{\Phi\left(\frac{\t}{\T}\right)}\right)\right\|^2\right)ds.
\end{array}
\end{equation}
Integrating \eqref{2-basic} over $\mathbb{R}\times(0,t)$ and using Gronwall's inequality,
we deduce from \eqref{tilde-q}-\eqref{int-F} that
\begin{equation}\label{7.18}
\begin{array}{ll}
\di \Big\|\left(\psi,\sqrt{\Phi\left(\frac{v}{V}\right)},\sqrt{\Phi\left(\frac{\t}{\Theta}\right)}\right)(t)\Big\|^2
+\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right)dxds\\[3mm]
\di+\int_0^t\int Q_1\big((U_-^r)_x+(U_+^r)_x\big)dxds \leq C_0+C(M)\d^{1/6}\\[3mm]
\di+C(M)\d\int_0^t(1+s)^{-1}\int_{\mathbb{R}}(\phi^2+\zeta^2)e^{-\frac{c_1x^2}{1+s}}dxds.
\end{array}
\end{equation}
Finally, due to the fact that
\begin{equation}\label{7.19}
\begin{array}{ll}
\di\int_0^t\int(\phi^2+\psi^2+\zeta^2)w^2dxds
&\di\leq C(M)+C(M)\int_0^t\|(\phi_x,\psi_x,\zeta_x)\|^2ds\\
&\di+C(M)\int_0^t\int(\phi^2+\zeta^2)\big((U_-^r)_x+(U_+^r)_x\big)dxds\\
&\di\leq C(M)+C(M)\int_0^t\int\left(\frac{\t\phi_x^2}{v^3}+\frac{\psi_x^2}{\t v}
+\frac{\zeta_x^2}{\t^2v}\right)dxds\\
&\di+C(M)\int_0^t\int Q_1\big((U_-^r)_x+(U_+^r)_x\big)dxds,
\end{array}
\end{equation}
whose proof can be found in \cite{Huang-Li-matsumura} and $w$ is defined as in Lemma \ref{hlm}, substituting \eqref{7.19} into \eqref{7.18} and choosing $\d$ suitable small imply
\begin{equation}
\begin{array}{ll}
\di \Big\|\left(\psi,\sqrt{\Phi\left(\frac{v}{V}\right)},\sqrt{\Phi\left(\frac{\t}{\Theta}\right)}\right)(t)\Big\|^2
+\int_0^t\int\left(\frac{\psi_x^2}{\t v}+\frac{\zeta_x^2}{\t^2 v}\right)dxds\\[3mm]
\di+\int_0^t\int Q_1\big((U_-^r)_x+(U_+^r)_x\big)dxds
\leq C_0+C(M)\d\int_0^t\int\frac{\t\phi_x^2}{v^3}dxds.
\end{array}
\end{equation}
Thus we can finish the proof of Lemma \ref{basic2} in the similar way as in Lemma \ref{basic-lemma}. We omit the details for brevity.
\hfill $\Box$
It is easy to check that the other estimates for single viscous contact wave still hold for the case in which the composite waves are the combination of viscous contact wave with rarefaction waves.
Thus, we complete the proof of Proposition \ref{prop}, and finally prove Theorem \ref{theorem2}.
\small
|
\section{Introduction} \label{sec:introduction}
Images of the solar disk taken in the line core of the \mbox{H$\alpha$}\ line display
dramatic elongated dark and bright features called fibrils and
mottles. They are present everywhere, except in the quietest Sun away
from concentrations of magnetic field. Their elongated shape and
presence wherever the magnetic field is strong, suggest that
fibrils and mottles outline the plane-of-the-sky component of the
magnetic field.
There is, however, very little direct observational evidence for this
suggestion in network areas and quiet Sun regions. The \mbox{H$\alpha$}\ line
itself is ill-suited for measuring chromospheric magnetic fields
\citep{2004ApJ...603L.129S}.
The plane-of-the sky magnetic field around sunspots was measured by
\citet{2011A&A...527L...8D}
using spectropolarimetry in the \CaII\ 854.2 nm\ line. This line has a similar
formation height as \mbox{H$\alpha$}\
\citep{2013ApJ...772...90L}
and shows similar fibril structure. Their conclusion is that most of
the time, but not always, fibrils appear to be aligned with the
plane-of-the sky magnetic field. To our knowledge no studies in quiet
Sun have been done so far.
On the modelling side only
\citet{2012ApJ...749..136L}
have reproduced fibril-like structures in synthetic \mbox{H$\alpha$}\ images
computed from a snapshot of a radiation-MHD (RMHD) simulation of the
solar atmosphere. They found that the fibril-like structures in the
simulation indeed trace out the horizontal component of the magnetic
field, but did not measure the alignment of the 3D magnetic field
vector and the fibril-like structure. It is thus not clear whether
fibrils trace out single field lines, or that they actually intersect
a set of different field lines that happen to have a horizontal
component parallel to the fibril.
For a subset of dark elongated \mbox{H$\alpha$}\ features, there is compelling
evidence based on their dynamical behaviour that they align with the
magnetic field in all three dimensions: The tops of so-called 'dynamic fibrils' in
active regions and mottles in quiet Sun show parabolic trajectories as
function of time. RMHD models show that they can be
explained as slow-mode acoustic shocks that propagate only along the
direction of the magnetic field
\citep{2006ApJ...647L..73H,2007ApJ...655..624D,2007ApJ...660L.169R,2013ApJ...776...56R}.
Another class of fibrils is present in network areas and active
regions. They appear as dense rows of elongated dark features that
span the chromosphere between areas with opposite magnetic polarity in
the photosphere. The horizontal size of these fibrils ($\sim30$~Mm) is
much larger than the extent of the optically thick \mbox{H$\alpha$}\ core
emission above the solar continuum limb
\citep[$< \sim$5~Mm, see, e.g,][]{1966ApJ...143...38S},
and estimates of the typical \mbox{H$\alpha$}\ formation height of 0.5~Mm to 3~Mm
\citep{1981ApJS...45..635V,2012ApJ...749..136L}
Therefore, in a simplified view where we assume magnetic field lines
to be semi-circular, we do not expect these fibrils to trace the
magnetic field in 3D, because the loop height is much larger than
the \mbox{H$\alpha$}\ formation height. While field lines in reality will not
be semi-circular, the relation between fibrils and field lines remains
unclear.
Understanding this relationship is relevant because \mbox{H$\alpha$}\ is one of
the most-used diagnostics of the chromosphere, so improving our
knowledge of its line formation will help understanding the physics of
the chromosphere. In addition, \mbox{H$\alpha$}\ is used as a constraint on
magnetic field extrapolations
\citep[e.g.,][]{2008SoPh..247..249W,2011ApJ...739...67J},
which requires understanding the
fibril-field line relation. Finally, transverse oscillations in
fibrils are used in chromospheric seismology.
Chromospheric seismology aims to derive properties of the chromosphere through
measuring properties of observed oscillations and interpreting them as
a certain types of waves. Oscillations of fibrils seen in the
\mbox{H$\alpha$}\ line core have been interpreted as {\it kink waves}:
transverse incompressible oscillations of a tube of high mass density
that is aligned with the magnetic field and is embedded in a medium
with lower mass density. In this view, measurement of the periods $P$,
wave displacement amplitudes $A$ and phase speeds of perturbation of
the \mbox{H$\alpha$}\ fibrils, and possibly their variation along the length of
individual fibrils, then allows derivation of chromospheric
properties. The period and displacement can be converted to velocity
amplitudes $V$ if one assumes sinusoidal motion. The velocity
amplitude together with an assumed mass density allows computation of
the energy density of the wave. If phase speeds are measured then one
can also derive the strength of the magnetic field and the energy flux
carried by the wave.
In an observation of quiet Sun including some network boundary
\citet{2012ApJ...750...51K}
measured lifetimes of 42 mottles/fibrils and the period and amplitude of their
transverse oscillations in a time-series of \mbox{H$\alpha$}\ line-center
images.
\citet{2013ApJ...768...17M}
analyzed the same dataset in more detail and derived improved
statistics on the period and displacement of transverse fibril
oscillations, finding an average period of 94~s and an average displacement
of 71~km.
\citet{2013ApJ...779...82K} looked at a different quiet Sun dataset
and presented three fibrils for which apparent phase speeds were
measured. They found evidence of apparent phase speeds increasing from
40 \mbox{km s$^{-1}$}\ to 120 \mbox{km s$^{-1}$}\ along one fibril, and a fibril with signal of
opposite phase speeds of +101 \mbox{km s$^{-1}$}\ and -79 \mbox{km s$^{-1}$}\ respectively,
suggestive of two waves traveling in opposite direction. They also
reported on a fibril oscillation with an apparent phase speed of 350
\mbox{km s$^{-1}$}, which they interpret as a standing wave pattern caused by
oppositely directed waves.
In a recent paper,
\citet{2014ApJ...784...29M}
derived a velocity power spectrum of transverse \mbox{H$\alpha$}\ oscillations
in a quiet-Sun region and an active region and compared the power
spectrum with the coronal transverse motion power spectrum derived
from observations with the Coronal Multi-channel Polarimeter
instrument
\citep{2009ApJ...697.1384T},
to estimate the frequency-dependent damping of transverse waves between the
chromosphere and corona.
The above discussion shows that \mbox{H$\alpha$}\ fibril seismology has the
potential to be a powerful tool for diagnosis of chromospheric
physical properties and the transport of energy by transverse
oscillations from the photosphere into the outer solar
atmosphere. However, the validity of this approach depends critically
on two related assumptions:
\begin{enumerate} \itemsep1pt \parskip0pt \parsep0pt
\item fibrils trace single magnetic field lines at any given instant in time;
\item fibrils trace the same field line during their lifetime.
\end{enumerate}
The fibrils most likely do trace out the horizontal component of the
magnetic field, but as we have argued, it is not known whether fibrils
also follow the vertical component of the field.
For the second assumption there is no observational evidence. We argue
that it not at all clear that this is the case:
\citet{2012ApJ...749..136L}
show that low \mbox{H$\alpha$}\ line-core intensity typically corresponds to a
large formation height, and that the line opacity is proportional
to the mass density. Their simulation shows that fibrils correspond to
mountain-ridge like structures of increased mass density. Dark fibrils
are thus most likely structures that are over-dense in the horizontal
direction compared to their surroundings. Even if a fibril traces out
a single field line at a given instant in time, it is thinkable that
this field line is drained of its mass while a neighbouring field line
gets filled with mass. The \mbox{H$\alpha$}\ fibril will then appear to move,
while the actual field lines do not.
The main aim of this paper is to investigate the relation between
magnetic field lines in a RMHD simulation of the solar atmosphere and
the fibril-like structures (from now on we call them just fibrils) in a time series of synthetic
\mbox{H$\alpha$}\ images computed from the simulation. We use the results of
the investigation to see to what extent the assumptions made in
chromospheric seismology are valid, by analysing the time
evolution of the fibrils and magnetic field lines that thread
through them. Finally we analyse the evolution of the atmosphere along
those fibril-threading field lines to investigate the dynamics of the
simulated chromosphere.
In Section~\ref{sec:simulations} we describe the radiation-MHD
simulation and the subsequent radiative transfer
modelling. Section~\ref{sec:observations} describes the observations
that we use. In Section~\ref{sec:fibrilosc} we treat the fibrils in the synthetic
images as observations to establish that they show the same
oscillatory behaviour as the observations. Section~\ref{sec:reduct}
describes the reduction of the RMHD simulation data to extract
properties along selected magnetic field lines and in
Section~\ref{sec:analysis} we analyse the relation between the fibrils
and the magnetic field lines. We finish with a discussion and our
conclusions in Section~\ref{sec:discussion}.
\section{Simulations and radiative transfer} \label{sec:simulations}
\begin{figure} \itemsep1pt \parskip0pt \parsep0pt
\includegraphics[width=\columnwidth]{./fig1a.eps}
\includegraphics[width=\columnwidth]{./fig1b.eps}
\includegraphics[width=\columnwidth]{./fig1c.eps}
\caption{Magnetic field configuration in the RMHD simulation. The
wire frame outlines the border of the computational domain. Each panel shows
the same set of magnetic field lines, color coded with the local
temperature, blue for $T < 20$~kK, red for $T \ge 20$~kK. {\it
Top and middle panel:} side and top view of the domain. The
gray-scaled plane displays the vertical magnetic field at
z=0~Mm. {\it Bottom panel:} The gray-scaled surface is the
\mbox{H$\alpha$}\ line-core height of optical depth unity, color coded
with the corresponding \mbox{H$\alpha$}\ vertically emergent
intensity. Only the section of the field lines above the surface
of optical depth unity is displayed. The visualization was created with
VAPOR \citep{2007NJPh....9..301C}.
\label{fig:snap60}}
\end{figure}
We performed a 3D radiation-MHD simulation of a part the solar
atmosphere using the {\it Bifrost}\ code
\citep{2011A&A...531A.154G}.
This code solves the equations of resistive MHD on a staggered
Cartesian mesh. In addition to the MHD equations the simulation
included optically thick non-LTE radiative transfer in the photosphere
and low chromosphere, parameterized radiative losses and gains in the
upper chromosphere, transition region and corona, thermal conduction
along field lines and an equation of state that takes the
non-equilibrium ionization of hydrogen into account.
The simulation was run on a grid of $504 \times 504 \times 496$ grid
cells, with an extent of $24 \times 24 \times 16.8$ Mm. In the
vertical direction the grid extends from 2.4 Mm below to 14.4 Mm
above average optical depth unity at 500 nm, encompassing the upper
convection zone, photosphere, chromosphere and lower corona. The $x$
and $y$-axes are equidistant with a grid spacing of 48 km. The
$z$-axis is non-equidistant. It has a grid spacing of 19 km between
$z=-1$ and $z=5$ Mm, while the spacing increases towards the upper
and lower boundaries to a maximum of 98 km at the coronal
boundary. The simulation has periodic boundary conditions in the two
horizontal directions and open upper and lower boundaries.
The magnetic field configuration is mainly bipolar, chosen such that
it leads to a small network-like configuration. It was created by
specifying the magnetic field at the bottom boundary and using a
potential field extrapolation to compute the field throughout the
computational domain. The magnetic field was inserted into a
relaxed hydrodynamical simulation and was then allowed to evolve
freely. The simulation was subsequently run for 3000 s of solar time using
LTE ionization, and then the non-equilibrium hydrogen ionization was
switched on. Afterwards the simulation was run for 2240 s of solar time. We
use a dataset of 1200 s of solar time (121 snapshots taken at 10 s
intervals) starting 830 s after non-equilibrium ionization was
switched on. The first snapshot of our time series is the same as was
used in
\citet{2012ApJ...749..136L,2012ApJ...758L..43S,%
2013ApJ...772...89L,2013ApJ...772...90L}
and
\citet{2013ApJ...778..143P}.
The snapshots are publicly available for
download\footnote{\url{http://sdc.uio.no/search/simulations}}.
The whole simulation is described in more detail in Carlsson et al. (2015).
The synthetic \mbox{H$\alpha$}\ imagery was computed following
\citet{2012ApJ...749..136L}.
In short, we use the radiative transfer code {\it Multi3d}\
\citep{2009ASPC..415...87L}
to solve the non-LTE radiative transfer problem in 3D geometry for a
5-level-plus-continuum hydrogen atom. Owing to the strong scattering
in the Lyman and Balmer lines it is essential to include 3D
effects. The Lyman-$\alpha$ and Lyman-$\beta$ lines are treated with
Doppler absorption profiles only to approximate PRD effects. In order
to save computation time we down-sampled the {\it Bifrost}\ snapshots to
half resolution in the horizontal direction, and solved the radiative
transfer only every second snapshot, i.e, 61 snapshots at 20~s cadence with a grid
size of $252 \times 252 \times 496$ cells. This is a slight
improvement over \citet{2012ApJ...749..136L}, where the atmosphere was
also down-sampled in the vertical direction. We performed a test
computation at the full resolution of the RMHD simulation, and found
that the down-sampling reduces image sharpness but otherwise does not
affect our results.
In Figure~\ref{fig:snap60} we show a characterisation of the magnetic
field in the simulation at $t=600$~s in our time series. The magnetic
field exhibits a large-scale bipolar structure in the photosphere; as
a consequence the chromosphere and corona are pervaded by loop-like
structures. The majority of the magnetic field closes in the
chromosphere: the average unsigned field at $z=0$ is 29.3~G, at
z=0.8~Mm it reaches only 50\% of this value and at z=3.3~Mm 25\%. The
bottom panel of Figure~\ref{fig:snap60} shows an iso-surface of the
\mbox{H$\alpha$}\ line-core height of optical depth unity, colored with the
corresponding \mbox{H$\alpha$}\ intensity, illustrating that the fibrils
roughly align with the horizontal component of the magnetic field.
\section{Observations} \label{sec:observations}
\begin{figure*}
\includegraphics[width=17cm]{./fig2.eps}
\caption{Example \mbox{H$\alpha$}\ images of the active region dataset taken
on June 15, 2008 (top) and the quiet Sun dataset taken on May 5,
2011 (bottom). The left-hand column shows images in the red wing
of the line, the right-hand column the profile-minimum intensity.
The subregions used in Figure~\ref{fig:obs_sim_fibrils} are
indicated by white boxes. }
\label{fig:AR_QS_ims}
\end{figure*}
In Section~\ref{sec:fibrilosc} we compare the synthetic
\mbox{H$\alpha$}\ imagery with observations. These observations were obtained
with the CRisp Imaging SpectroPolarimeter
\citep[CRISP,][]{2008ApJ...689L..69S}
at the Swedish 1-m Solar Telescope
\citep[SST,][]{2003SPIE.4853..341S}
on La Palma.
We use a dataset of the remains of active region AR10998 taken on June
15, 2008 at 7:53 UT, with a duration of 27 minutes at 6.8 s
cadence. The data were taken close to disk-center, at $\mu=0.98$.
The \mbox{H$\alpha$}\ line was sampled at 25 positions, equidistant
with 0.1~\AA\ steps between
$-1.6$~\AA\ and $+0.8$~\AA\ from line center. We refer to this dataset
as Active Region (AR).
A second dataset was acquired on May 5, 2011 at 7:38 UT. The
\mbox{H$\alpha$}\ spectral line was sampled with 35 line positions, from
$-$2.064~\AA\ to $+$1.290~\AA, and equidistant 86~m\AA\ stepping
between $\pm$1.290~\AA. The time to complete a full line scan was 8~s.
The target area was a quiet 60$\times$60\arcsec region at
$\mu=1.0$. We refer to this dataset as Quiet Sun (QS).
Both datasets were captured during excellent seeing conditions, and
the image quality
was improved by the
adaptive optics system of the SST
\citep{2003SPIE.4853..370S}
and post-processing using Multi-Object Multi-Frame Blind Deconvolution
\citep[MOMFBD,][]{2005SoPh..228..191V}
image restoration. For the data reduction we followed early versions
of parts of the CRISPRED data processing pipeline
\citep{2014arXiv1406.0202D}.
Figure~\ref{fig:AR_QS_ims} shows example images of the AR and QS
datasets.
\section{Comparison of observed and simulated \mbox{H$\alpha$}\ fibrils} \label{sec:fibrilosc}
\subsection{\mbox{H$\alpha$}\ as velocity diagnostic of fibrils}
\begin{figure}
\includegraphics[width=\columnwidth]{./fig3.eps}
\caption{Scaled joint probability density functions of the Doppler
shift of the \mbox{H$\alpha$}\ profile minimum (positive is redshift) and
the vertical velocity at optical depth unity at that Doppler shift
(positive is down flow). Each column in the panels is scaled to
maximum contrast to increase visibility. The inner blue contour
includes 50\% of all pixels, the outer contour 90\%. The red line
denotes the line $v_z = \Delta \lambda$.}
\label{fig:dv_vz_corr}
\end{figure}
Our SST observations include a
sampling of the line profile and we can thus use them to compare
Doppler-shifts in the observed and simulated datasets. Because fibrils
are most clearly visible in line-core images, we use the Doppler-shift
of the profile minimum as their velocity diagnostic.
\citet{2013ApJ...772...90L} showed that the Dopplershift of the
central line-profile minima of the \ion{Mg}{2}\ h\&k lines exhibit
remarkable correlation with the vertical velocity at the height of
optical depth unity. They are thus excellent velocity diagnostics. The
high correlation for the \ion{Mg}{2}\ lines is caused by the
small ($\simeq 2.5$~\mbox{km s$^{-1}$}) Doppler width of the absorption
profile and the steep increase of optical depth with geometric depth
causing only a small height range to be sampled. The \mbox{H$\alpha$}\ line has
five times larger Doppler width. Numerical simulations indicate that its
contribution function to intensity is often non-zero over a height
range of 1 Mm or more
\citep{2012ApJ...749..136L}.
We thus expect \mbox{H$\alpha$}\ to perform somewhat worse as a velocity
diagnostic than \ion{Mg}{2}\ h\&k.
We computed the correlation between the Doppler shift of the
line-profile minimum and the vertical velocity at optical depth unity
at the wavelength of the profile minimum based on all pixels in our synthetic
time-series.
The results are displayed in
Figure~\ref{fig:dv_vz_corr}. The correlation is indeed not as good as
for \ion{Mg}{2}~h\&k, but its Pearson correlation coefficient is still very
high at 0.96. We conclude therefore that the \mbox{H$\alpha$}\ line is a good
velocity diagnostic for those layers that are sampled by the profile
minimum. Test with the observational datasets show that traditional
Dopplergrams created by subtracting and normalizing intensity images
at equal wavelength separation from line center are visually very
similar to the Doppler-shift of the profile minimum. The advantages of
using profile-minimum Dopplershift maps is that it provides the
vertical velocity in absolute units, and that the profile-minimum
intensity provides a direct visualization of the structures -- fibrils
in this paper -- for which the velocity is measured.
\subsection{Comparison of the appearance of synthetic and observed \mbox{H$\alpha$}\ fibrils}
\label{sec:compsyntobs}
\begin{figure*}
\includegraphics[width=17cm]{./fig4.eps}
\caption{Comparison of images and space-time diagrams of fibrils in
an active region (left), quiet Sun (middle) and the simulation
(right). Rows, from top to bottom display \mbox{H$\alpha$}\ profile-minimum
intensity images, \mbox{H$\alpha$}\ line-core Doppler-shift images,
intensity space-time diagrams, and Doppler-shift space-time
diagrams. The dashed lines specify the location of the space-time
cut in the images and vice versa. Blue color in the Doppler-shift
panels corresponds to upflow. Time is given in units of
hecto-seconds ($10^2$s ) in this and all following figures.}
\label{fig:obs_sim_fibrils}
\end{figure*}
The RMHD simulation incorporates many of the physical processes that
act in the solar chromosphere and indeed shows fibrilar structure
reminiscent of observations. However, the separation of the opposite
polarity patches in the photosphere is only $\sim 10$~Mm, considerably
smaller than the typical size of a supergranule
\citep[$\sim 30$~Mm, e.g.][]{1964ApJ...140.1120S},
and the maximum length of the simulated fibrils is thus shorter than
possible on the sun. The finite resolution of the simulation (48~km
horizontal, 19~km vertical) does not allow horizontal structures
smaller than $\sim 100$~km to form. It is therefore worthwhile to
analyze in more detail how the simulated fibrils appear and behave in
comparison with observed ones.
In Figure~\ref{fig:obs_sim_fibrils} we compare fibrils in the AR and
QS observations and the simulation. The panels showing the
observations display the same spatial and temporal extent as the
simulation to allow for fair comparison.
The AR area has a cluster of photospheric magnetic elements in the
upper-left and lower-right corners of the field-of-view (FOV). Between
these elements a carpet of long and finely structured fibrils is
present. The QS image displays a few broad and short
dark fibrils
\citep[also called mottles, see ][and references therein]{2012SSRv..169..181T}
that extend towards to top of the image. The corresponding \mbox{H$\alpha$}\ wing
image shows a small group of photospheric magnetic elements in the
lower part of the image (see Figure~\ref{fig:AR_QS_ims}).
The simulated fibrils appear as a cross between the AR and QS ones,
they typically are not as wide as in the QS, but also not as long,
slender and densely packed as in the AR area. The simulated intensity
contrast is similar to the observed one.
The second row shows images of the Dopper-shift of the line
core. The overall structuring is similar to the intensity images. Both
the AR and QS images show upward and downward moving fibrils, but the
AR panel shows Dopper-shift variation perpendicular to the magnetic
field on smaller spatial scales.
The two bottom rows compare time-distance diagrams of the line-core
intensity and Doppler-shift. The line-core panels (third row) show the
characteristic swaying motions of the fibrils. The swaying fibril
tracks are ubiquitously present in all three datasets. The AR dataset
typically shows tracks that remain visible during a longer time than
in the QS panel, and the simulation again shows characteristics
somewhere in between the AR and QS data.
In the bottom row we show the time evolution of the profile-minimum Doppler
shift along the same
cut across the fibrils. All three panels show
elongated structure only at the location of the very darkest
fibrils. The rest of the panel is dominated by oscillations with
periods of 3 to 5 min and a coherency length along the slit of $\sim
4$ Mm in the AR and $\sim 10$ Mm in the QS observations. These
coherent oscillations are most likely the result of buffeting from
below by internetwork acoustic shocks excited by $p$-modes and
granular buffeting
\citep{2008SoPh..251..533R}.
The simulation displays an oscillation with a typical period of 4 min
that is coherent along the whole 24~Mm cut. This is the signature of
shocks driven by global radial box oscillations -- the simulation
counterpart of $p$-modes -- excited by the convection
\citep{2001ApJ...546..585S}.
The AR panel shows smaller Doppler-shift amplitude ($\pm 7$ \mbox{km s$^{-1}$}, RMS
of 1.7 \mbox{km s$^{-1}$}) than the QS (amplitude -8 to +14 \mbox{km s$^{-1}$}, RMS of 2.5 \mbox{km s$^{-1}$})
and simulated (amplitude -12 to +11 \mbox{km s$^{-1}$}, RMS of 2.9 \mbox{km s$^{-1}$}) data.
It is unclear what is causing the difference in pattern between the
intensity and Doppler shift time-distance diagrams. It is possible
that many of the fibrils have an optical thickness of close to unity,
so that the Doppler-shift of the profile minimum is partially
determined by the sub-canopy domain, which is full of acoustic waves
and shocks. Another possibility is that we see the actual upward
and downward motion of the fibrils themselves, and that this is
residual motion caused by the reflection of sub-canopy waves at the
$\beta=1$ surface, as demonstrated in the simulations of
\citet{2002ApJ...564..508R}.
The reflection of the acoustic waves can have a significant signal
into layers where $\beta<0.1$, where fibrils are presumably
located. This scenario is supported by the fact that the Doppler-shift
amplitude in the AR dataset, where \mbox{H$\alpha$}\ forms much higher above
the $\beta=1$ surface, are smaller than in the QS, where
\mbox{H$\alpha$}\ samples closer to the $\beta=1$ surface. Of course another
mechanism altogether remains possible too.
Focusing back on our present purpose: we have established that the
simulation qualitatively reproduces many of the features of the
appearance of fibrils in network and quiet Sun, in both intensity and
Doppler shift. The fibril density and width in the simulations are
intermediate between AR and QS, the simulated Dopplershifts are closer
to the QS properties than to the properties of the AR.
\subsection{Transverse oscillations of synthetic fibrils}
\begin{figure}
\includegraphics[width=8.8cm]{./fig5.eps}
\caption{Top panel: synthetic \mbox{H$\alpha$}\ line-core image at t=600~s.
The location of the time-distance diagrams used to measure fibril
oscillation motions are overplotted. Bottom panel: time-distance
diagram along the green cut in the upper panel, with fibril tracks
overplotted in red.}
\label{fig:cuts}
\end{figure}
\begin{figure*}
\includegraphics[width=17cm]{./fig6.eps}
\caption{Histogram of the properties of the oscillations of the
simulated fibrils. Left: maximum transverse displacement; middle:
period; right: maximum transverse velocity.}
\label{fig:fibril_histo}
\end{figure*}
We now turn our attention to the periods, displacement and velocity
amplitudes of the synthetic fibrils. After manual inspection of the
data we placed five cuts through the data that showed clear
oscillations of dark fibrils (see Figure~\ref{fig:cuts}). There are
many more fibrils that appear to sway when playing the image sequence
as a movie, but whose time-distance diagram does
not show an unambiguous swaying
track. After selecting the cuts with clear swaying we manually traced
the swaying fibril tracks. We also examined automated detection
algorithms after unsharp masking of the data to bring out contrast,
but found that this leads to many false detections. These are in
particular connected with the bright-dark pattern of the box
oscillations, that after unsharp masking often give rise to sinusoidal
tracks that are not associated with dark fibrils.
In the cuts we found 13 fibrils that could be unambiguously followed
for one period or more. Because we found that the fibrils show
oscillatory behavior, but rarely with constant amplitude and period,
we also measured periods and amplitudes by hand. For each space-time
track we removed a linear trend. Then we measured the space and time
coordinate of each extremum ($t_i,x_i$). The time difference between
two consecutive maxima or minima gives one measurement of a period
$P$: $P_i=t_{i+1}-t_{i-1}$. The spatial difference between two
consecutive extrema yields a measurement of the transverse
displacement $d$: $d_i=0.5(x_{i+1}-x_i)$. Finally we computed the
velocity amplitude $v$ as $v_i=\pi (d_{i-1}+d_{i}) / P_i$. In total we
measured 40 displacements, 31 periods and 31 velocity amplitudes.
The results of our measurements are shown in
Figure~\ref{fig:fibril_histo}. The average (standard deviation) of the
displacement is 158~km (98 km), for the period it is 221~s (70~s) and
for the velocity amplitude it is 4.5~\mbox{km s$^{-1}$}\ (1.8~\mbox{km s$^{-1}$}).
\citet{2013ApJ...768...17M}
measure smaller average displacements and periods, 71~km and 94~s. We
believe this to be at least partially caused by an event selection
bias caused by the higher spatial and temporal resolution of their
observations and their automated versus our manual event
selection. The observations used by
\citeauthor{2013ApJ...768...17M}
have a spatial pixel size of 50~km and a temporal resolution of 7.7~s,
compared to 96~km and 20~s for the simulations. The observations thus
in principle allow detection of oscillations at half the amplitude and
2.6 times smaller period. The automated detection algorithm used by
\citet{2013ApJ...768...17M}
also picks up weak signals, while we focus
on the oscillations of the most prominent dark fibrils. Another possible
source of the difference is the resolution of the simulation, that
might not support oscillations below a certain displacement amplitude.
The average velocity amplitude that we find is the same as
\citet{2013ApJ...768...17M},
but the synthetic distribution lacks the tail of high amplitude events
detected in the observations.
In summary, we find that the swaying fibrils in the synthetic
\mbox{H$\alpha$}\ imagery have periods, maximum displacements and velocity
amplitudes that are consistent with those observed. The lack of
short-period and low-displacement events are possibly caused by finite
spatial resolution of the simulation, the spatial resolution and
cadence of the synthetic images and a selection bias.
\subsection{Phase speeds} \label{sec:phase_speeds}
\begin{table}
\caption{Approximate phase speeds in synthetic fibrils (\mbox{km s$^{-1}$})}
\label{table:phase_speeds}
\centering
\begin{tabular}{ccc}
\hline\hline
54 & 18 & 49 \\
33 & 60 & 56 \\
320 & 27 & 16 \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\includegraphics[width=8.8cm]{./fig7.eps}
\caption{Phase speeds in a simulated fibril. Top panel:
\mbox{H$\alpha$}\ image of the dark fibril at $t=600$~s, with the
cross-cuts used in the bottom panel overplotted in red. Bottom
panel: The black curves show the displacement of the fibril along
each of the cross cuts in the top panel as function of time, with
a linear trend removed. The track labelled (1) corresponds to the
cut with the same label in the upper panel. The other tracks have
been shifted vertically by the distance between the corresponding
cross-cut and the cut labeled (1) in the upper panel. Blue lines:
linear fits to selected extrema. Red curves: third-order
polynomial fits to the strongest extrema.}
\label{fig:sim_phase6}
\end{figure}
\begin{figure}
\includegraphics[width=8.8cm]{./fig8.eps}
\caption{Same as Figure~\ref{fig:sim_phase6}, but for a another fibril.}
\label{fig:sim_phase8}
\end{figure}
In our synthetic \mbox{H$\alpha$}\ time series we found a few instances of
strong fibrils for which it was possible to measure phase speeds. We
placed a number of cross-cuts intersecting the fibrils at different
places along their length and measured the $st$-track along each
cut. We then subtracted a linear fit from the tracks to remove
systematic drift. In Figures~\ref{fig:sim_phase6}
and~\ref{fig:sim_phase8} we show two strong, long-lived fibrils and
their swaying along the cross-cuts. The phase speeds of the extrema
are typically not constant, but show acceleration, deceleration and
sometimes even a reversal of direction. The fibril in
Figures~\ref{fig:sim_phase8} shows clear indication of a minimum
traveling in one direction along the fibril, and the following maximum
in the other direction. The complexity of these phase diagrams is
suggestive of an interference pattern caused by waves traveling in
opposite directions. For those instances where the phase speed was
nearly constant over most of the length of the fibrils we show the
measured speeds in Table~\ref{table:phase_speeds}, they range from 16
\mbox{km s$^{-1}$}\ to 320 \mbox{km s$^{-1}$}. This is consistent with
\citet{2013ApJ...779...82K},
who report phase speeds of 79~\mbox{km s$^{-1}$}, 101~\mbox{km s$^{-1}$}\ and 360~\mbox{km s$^{-1}$}.
We conclude that the simulation reproduces the observed properties
sufficiently well to assume that the simulation contains the essential
physics driving transverse oscillations of fibrils. In
Section~\ref{sec:reduct} we take a closer look at the MHD structures
and their dynamics that give rise to the synthetic \mbox{H$\alpha$}\ fibrils.
\section{Reduction of the RMHD data} \label{sec:reduct}
We chose to reduce the data using a magnetic field line based
approach. The rationale for this approach is that fibrils form in the
chromosphere in a low plasma $\beta$ environment. Both slow waves and
Alv\'en waves therefore propagate parallel to the field lines, and
analysis of the data along field lines then provides a natural way to
bring out those waves.
We planted 80 seed points in the simulation at $t=600$~s
along a number of fibrils in the corresponding \mbox{H$\alpha$}\ image. The $x$
and $y$ coordinates of the seed points thus follow the fibrils, and
are spaced about 0.5~Mm apart. As $z$-coordinate of the points we
chose the height of optical depth unity for the profile minimum. From
each seed point we traced a magnetic field line. In order to trace the
same field line over time we advected the field line apex both forward
and backward in time, used the new location to trace a new field line
and repeated this until we obtained the time evolution of the entire
field line. In Figure~\ref{fig:flines} we show the location
of the seed points and the field lines traced from them.
\begin{figure} \itemsep1pt \parskip0pt \parsep0pt
\includegraphics[width=\columnwidth]{./fig9a.eps}
\includegraphics[width=\columnwidth]{./fig9b.eps}
\includegraphics[width=\columnwidth]{./fig9c.eps}
\caption{Three-dimensional visualisation of selected field
lines that thread through \mbox{H$\alpha$}\ fibrils at $t=600$ s in the
simulation. Top: Surface of optical depth unity, colour coded with
the emergent \mbox{H$\alpha$}\ intensity, with red line segments
indicating the field line seed points and the local direction of
the magnetic field. Middle: Same as top panel, but now showing
the part of the field lines that extends above the optical depth
unity surface. Bottom: The same field lines, but now shown on top
of the vertical magnetic field at $z=0$ Mm. The field lines are
colour coded with temperature: dark blue for $T<17$~kK; red for
$T>100$~kK; other colours indicate intermediate values. The labels
a, b and c indicate fibrils and associated flux bundles shown in
Figure~\ref{fig:fibril_class}.
\label{fig:flines}}
\end{figure}
\begin{figure}
\includegraphics[width=8.8cm]{./fig10.eps}
\caption{Distribution of the magnetic Reynolds number as function of
height at t=600~s in the Bifrost simulation.The red curves
indicate the 25th, 50th and 75th percentile.}
\label{fig:jpdf_z_rm}
\end{figure}
The validity of this approach is limited by the finite time resolution
of our snapshot series (10~s cadence), which limits the ability to
accurately follow test particles in time, and the magnetic Reynolds number
$R_\mathrm{m}$, which determines how well the assumption of frozen-in
magnetic flux holds. Bifrost employs a space and time variable,
non-isotropic magnetic diffusion coefficient
$\vect{\eta}(\vect{r},t)$, so we cannot give a single value for
$R_\mathrm{m} =(LU)/\eta$. Instead we show the joint probability
density distribution in Figure~\ref{fig:jpdf_z_rm}, assuming a typical
length scale $L$ of 1~Mm and a typical velocity $U$ of 10~\mbox{km s$^{-1}$}. The
Reynolds number is always larger than unity. Even at heights where
it is smallest, 75\% of the grid cells has $R_\mathrm{m}>10$ and 50\%
has $R_\mathrm{m}>30$. The frozen-in condition thus holds for a large
fraction of the domain, but we also expect locations where diffusion
is important. In Section~\ref{sec:analysis} we show that many field
lines that we analyse remain anchored in the same photospheric field
concentrations for the 1200~s duration that we consider, and evolve
smoothly in time, indicating sufficient time resolution and validity
of the frozen-in assumption.
However, there are also instances where the field lines evolve in
jumps, indicating insufficient time resolution,strong diffusion
or reconnection. In this paper we focus on wave dynamics and we
simply ignore such field lines. We note however that this approach,
improved by proper tracking of the advecting velocity field with test
particles, will yield interesting insights also in case of strong
diffusion and reconnection.
As we are interested in the velocity transverse to the magnetic field
we need to project the velocity onto appropriate axes. A convenient
choice is the Frenet-Serret frame
\citep{1851S,1852F}.
A magnetic field line can be defined as curve $\vect{r}(s)$ through
three-dimensional space as function of the arc length $s$. The curve
is defined by
\begin{equation}
\frac{\mathrm{d} \vect{r}}{\mathrm{d} s} = \frac{\vect{B}}{| \vect{B} |} = \vect{b},
\end{equation}
given a seed point $\vect{r}_0(s_0)$, where we introduced
$\vect{b}$ as the unit vector pointing in the direction of
$\vect{B}$. The Frenet-Serret frame is an orthogonal coordinate system
at each point along the curve with unit vectors $\vect{T}$,
$\vect{N}$ and $\vect{P}$ given by
\begin{eqnarray}
\vect{T} & = & \frac{\mathrm{d} \vect{r}}{\mathrm{d} s} = \vect{b}, \\
\vect{N} & = & \frac{\mathrm{d} \vect{T}}{\mathrm{d} s} \left/ {\left| \frac{\mathrm{d} \vect{T}}{\mathrm{d} s} \right| }\right. , \\
\vect{P} & = & \vect{T} \times \vect{N}.
\end{eqnarray}
As an example one can consider a semicircular magnetic field line with
radius $R$ in a plane perpendicular to the photosphere, the latter we take
to lie in the $xy$ plane at $z=0$:
\begin{equation}
\vect{r}(s) = \left\{
\begin{array}{ll}
R \cos (s/R) \,\hat{x} \\
0\, \hat{y}\\
R \sin (s/R)\, \hat{z}
\end{array}
\right.
\end{equation}
The $\vect{N}$ vector along each point of the field line then points
towards the center of the circle at the origin and $\vect{P}$ is
the unit vector in the $y$-direction. If one would observe this field
line at solar disk center and observe transverse oscillations, then
these oscillations would be along $\vect{P}$. For field lines with
more complex 3D geometry, as we have in our simulation, we generally
do not expect that transverse waves that are initially excited only
along either $\vect{N}$ or $\vect{P}$ will remain like that. For our
analysis this will not pose problems: Models of
transverse oscillations along curved loops developed by
\citet{2009SSRv..149..299V}
indicate that the eigenfrequencies, and thus phase speeds of the two
wave modes are nearly identical. We thus expect a
transverse wave to change its amplitude along $\vect{N}$ and
$\vect{P}$ as it propagates, but not to show significant dispersion
between those two directions.
Finally, one should consider the meaning of the length coordinate of a
field line. A field line is not a physical entity, but merely a
convenient mathematical abstraction
\citep[e.g.,][]{2005LRSP....2....7L}.
At any given instant in time a length coordinate $s$ along the
field line can be unambiguously defined. We define $s=0$ to be at the
location where the field line crosses the $z=0$ plane and the magnetic
field vector $\vect{B}$ points upward (i.e, towards the corona), with
$s$ increasing in the direction of $\vect{B}$. This coordinate system
does not remain fixed in time, because the field lines are moving with
the velocity field. The length of the infinitesimal line element
$\delta \mathbf{l}$ between two points $\mathbf{l}$ and
$\mathbf{l}+\delta \mathbf{l}$ lying on the same field line, with
$\delta \mathbf{l} =\delta s \mathbf{b}$ evolves as
\begin{equation}
\frac{d \ln \delta s}{dt} = \sum_{i,j} S_{ij} b_i b_j
\end{equation}
where $b_i$ are the components of $\vect{b}$ and
$S_{ij}$ is the symmetric part of the velocity gradient tensor
$\partial u_i/ \partial x_j$
\citep[see for example][]{1992JFM...236..415D}.
In practice this means that one needs to be careful when interpreting
signal speed in $st$-diagrams of a field line as expansion and
contraction of the field line changes the apparent signal speed.
\section{Analysis of fibril-threading fieldlines} \label{sec:analysis}
\subsection{Do fibrils trace field lines in 3D?} \label{sec:ffspace}
\begin{figure*}
\centering
\includegraphics[width=14.2cm]{./fig11a.eps}
\includegraphics[width=14.2cm]{./fig11b.eps}
\includegraphics[width=14.2cm]{./fig11c.eps}
\caption{Visualization of flux bundles threading through three
different fibrils at $t=600$~s. The labels a, b and c correspond to the labels
in Figure~\ref{fig:flines}.
The left panels show the top view of the field
lines, superimposed on the \mbox{H$\alpha$}\ line-core image. The panels
on the right side show the $xz$ and $yz$ side views. The field
lines are colour coded: the part of the field lines more than 200
km below optical depth unity in the line-core are green. The part
more than 200 km above optical depth unity is blue, and the part
within 200 km of optical depth unity is red. The red part of the
field lines can thus be considered as visible in the
\mbox{H$\alpha$}\ line core. {\it A movie showing the time evolution of
the fibrils and field lines can be found in the online
material}.
\label{fig:fibril_class}}
\end{figure*}
In Figure~\ref{fig:flines} we show the seed points and associated
field lines that we selected for further analysis. The seed points
follow a number of fibrils, mostly very prominent ones, but also one
fibril that appears almost transparent at $t=600$ s but is more visible before
and after this moment. The $z$-coordinate of the seed point is the
height of optical depth unity at that $(x,y)$ location.
The middle panel shows the part of the field lines above the height of
optical depth unity. The field lines in the centre area
remain low, while
the other field lines reach large heights. The bottom panel shows that
the field lines that point away from the centre of the domain typically
reach coronal temperatures, while the central ones have chromospheric
temperatures, with the latter defined as $T<17$~kK. The field lines
typically end in strong photospheric field concentrations, but there
are a few instances where one endpoint lies just next to a
concentration.
These photospheric endpoints do not visually appear to be associated with a
fibril.
The ending of fibril-threading field lines in photospheric
magnetic field concentrations is consistent with the observational
evidence that \mbox{H$\alpha$}\ fibrils are found as if emanating
from
photospheric magnetic concentrations
\citep[e.g.,][]{2001SoPh..199...61C,2003A&A...402..361T,2007ApJ...660L.169R},
and further corroborates the idea that fibrils trace at
least the horizontal components of field lines that start in
photospheric magnetic field concentrations. Also note that the field
lines show a strong suggestion of twist and shear, indicating that the
simulated chromosphere is not in a force-free state.
Figure~\ref{fig:fibril_class} displays for three different cases (a, b and c, also labeled in Figure~\ref{fig:flines}) how the fibrils and the field
lines intersect each other, and which part of the field lines can be
considered visible in \mbox{H$\alpha$}\ line-core images. The definition of
visible is necessarily somewhat imprecise, because the \mbox{H$\alpha$}\
contribution function is typically non-zero over a range of several
hundred kilometers, and can even have both a chromospheric and
photospheric component
\citep{2012ApJ...749..136L}.
We settled for defining a point $\mathbf{r}=(x,y,z)$ on a field line
as visible if it lies within 200 km from the height of optical depth
unity along the same column $z_{\tau=1}(x,y)$.
The top row of panels of Figure~\ref{fig:fibril_class} shows case a: a short
fibril in the centre of the computational domain. All field lines except one
form a very tight bundle, and are visible along the whole length of
the fibril.
The exception is the highest-reaching field line, whose
seed point was on one endpoint of the fibril and can thus be
considered an intermediate case between fibril and non-fibril field
lines. For this fibril the assumption that fibrils trace the magnetic
field in all three dimensions is true. This finding is confirmed too
in the middle panel of Figure~\ref{fig:flines} where the field lines
can be seen to be almost parallel to the $\tau=1$ surface. The
temperature along the field lines is chromospheric. Note that the
field lines show a characteristic sigmoid shape in the top view of Figure~\ref{fig:fibril_class}.
In the middle row of Figure~\ref{fig:fibril_class} we show case b, an example
of a longer fibril that does not trace the magnetic field in three
dimensions. In the middle panel of Figure~\ref{fig:flines} the
field lines can be seen to thread through the optical depth unity
surface of the fibril at an angle. Different field lines are visible
at different locations along the length of the fibril. At one end of
the fibril at $(x,y)=(17,11)$ Mm the field lines all emanate from a
single magnetic element, whereas at the other end the field lines
connect to different magnetic elements in a patch of about
$1$ Mm$\times 3$ Mm.
Also these field lines remain chromospheric and
do not protrude into the corona.
Finally, the bottom row of Figure~\ref{fig:fibril_class} shows case c: field
lines threading through a short fibril at $(x,y)=(20,17)$ Mm. This
case is very different from the previous two. The optical depth unity
surface in the fibril is nearly horizontal, and the field lines
intersect this surface at large angle (more than $45^{\circ}$). In the
$yz$-view at $y=17$ Mm the field lines are seen to form a fan-like
structure. The field line apexes reach transition region and coronal
temperatures as shown in the bottom panel of Figure~\ref{fig:flines}.
\subsection{Time evolution of fibrils and field lines} \label{sec:fftime}
We now turn our attention to the question whether fibrils trace the
same field line (or parts of field lines) over their lifetime. In the
supplemental material we provide movies of the time evolution of the
three fibrils and associated field lines shown in
Figure~\ref{fig:fibril_class}. In each case the fibrils and test
field lines intersect at $t=600$~s. In all three cases the field
lines do not start out tracing a fibril at $t=0$~s, but instead seem
to ''drift'' into the fibril around $t=600$~s. After that time the field lines
drift out of the fibril, while the same apparent fibril remains
visible. Typically the field lines remain visible for about
100--200 s, which is of the same order as a typical observed wave
period.
\subsection{Analysis of the dynamics of single field lines}
\begin{figure*}
\centering
\includegraphics[width=15cm]{./fig12.eps}
\caption{Evolution of fibril-threading field line A. The upper left
panel shows a top view of the computational domain, with the
location of the field line at $t=0$, $t=600$ and $t=1200$~s. The
location and time evolution of the footpoints are indicated in
red ($\vect{b}$ points up towards the corona) and green
($\vect{b}$ points down into the convection zone). The red cross
marks the location of the original seed point from which the
field line and its time evolution are computed. All other panels
show $st$-diagrams: the time evolution (abscissa) of quantities
as function of the loop length $s$ (ordinate). The displayed
quantity and its range are given above each panel. The zero
point of $s$ is the intersection of the field line with the $z=0$
plane where $\vect{b}$ points upwards. It is indicated with the
red line. The green curve is the $s$-coordinate of the other
footpoint. Areas of the field line within the dark blue curves
are visible in the \mbox{H$\alpha$}\ line core. The orange curves
indicates plasma $\beta=1$. The light blue curves are wave
trajectories, i.e, a signal propagating with the wave speed along
the loop would follow the indicated trajectory. The wave speed is
the sound speed for the panels labeled $v_\mathrm{s}$, $v_\mathrm{para}$
and $(\mathrm{div} \, u)_\mathrm{para}$ and the Alfv\'en speed for all
other panels.
\label{fig:fibril_050_st}}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=15cm]{./fig13.eps}
\caption{Same as Figure~\ref{fig:fibril_038_st} but for field line B.
\label{fig:fibril_038_st}}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=15cm]{./fig14.eps}
\caption{Same as Figure~\ref{fig:fibril_050_st} but for field line C.
\label{fig:fibril_019_st}}
\end{figure*}
Figures~\ref{fig:fibril_050_st}--\ref{fig:fibril_019_st} show the time
evolution of various quantities along the length $s$ of three
fibril-threading field lines, which we label A, B and C, one from each
of the three cases a,b, and c shown in Figure~\ref{fig:flines} and~\ref{fig:fibril_class}. The seed
points of these field lines were in the centre of the fibrils at
$t=600$~s. In the online material we provide movies of the time
evolution of the 3D position of the three field lines.
The $st$-panels show the evolution of temperature
$T$ and density $\rho$ in the first row; magnetic field magnitude
$|\vect{B}|$, sound speed $v_\mathrm{s}$ and Alfv\'en speed
$v_\mathrm{A}$ in the second row; plasma flow speed parallel to the field
line $u_{||}=\vect{u} \cdot \vect{b}$,
and along the two
transverse directions $u_{\vect{P}}=\vect{u} \cdot \vect{P}$ and
$u_{\vect{N}}=\vect{u} \cdot \vect{N}$ in the third row. In the fourth
row we show the divergence of the plasma flow
parallel to the magnetic field
\begin{equation}
\nabla_{||} \cdot \vect{u} = \partial u_{||}/ \partial s,
\end{equation}
the divergence of the flow field
perpendicular to the field
\begin{equation}
\nabla_{\perp} \cdot \vect{u} = \nabla \cdot \vect{u}
- \partial u_{||}/ \partial s,
\end{equation}
and the vorticity parallel to the magnetic field
\begin{equation}
\omega_{||}=(\nabla \times \vect{u}) \cdot \vect{b}.
\end{equation}
\subsection{Time evolution of purely chromospheric field lines} \label{sec:chromfl}
Figure~\ref{fig:fibril_050_st} shows the evolution of field line A
that forms a small loop in the centre of the field-of-view. The
\mbox{H$\alpha$}\ fibril at $t=600$ s traces the field lines in this loop very
well (See Section~\ref{sec:ffspace}). The $\beta=1$ curves indicate
that the magnetic pressure is dominant throughout the chromospheric
part of the field line. The blue curves, that indicate whether the
field line is visible show that the whole loop apex is visbile in
\mbox{H$\alpha$}\ for a duration $\simeq 200$ s around t=600 s. At earlier
times the field line is buried too deep in the atmosphere, and at
later times the loop has risen above the optical depth unity
surface. The field line has typical chromospheric temperatures between
3~kK and 15~kK. As it rises over time the density of the loop apex is
steadily decreasing. The magnetic field remains rather constant over
time. The sound speed just depends on the square root of the
temperature and varies between 5.8~and 17~\mbox{km s$^{-1}$}\ in the
chromosphere. In contrast, the Alfv\'en speed shows a much large
variation. At the apex of the field line, it increases over time from
50 kms to 800 \mbox{km s$^{-1}$} as the field line rises and its density
decreases. The light-blue wave-speed signal tracks illustrate
this. The footpoint-to-footpoint loop crossing time for sound waves is
more than 600 s, whereas the Alfv\'en crossing time is of the order of
100 s.
The $u_{||}$ and $\nabla_{||} \cdot \vect{u}$ panels reveal a pattern
of diagonal stripes from the $\beta=1$ curve towards the loop
apex. They are most clearly seen in the $\nabla_{||} \cdot \vect{u}$
panel as dark stripes, indicating compression of the plasma along the
direction of the loop. They are the signature of
photospherically-excited slow-mode waves that only propagate parallel
to the field lines in the low-$\beta$ regime. Their slopes (and thus
propagation speeds) are comparable with the sound-speed, as can be
seen from comparing them with the sound-speed tracks. They are not
exactly equal because of loop expansion and contraction, variation in
space and time of the sound speed and the presence of background flows
which modify the wave speed in the observer's frame. Interestingly,
the slow-mode waves to not cross the loop from one end to the
other. Instead they typically fade before they reach the loop apex. A
preliminary analysis of these waves (not shown here) shows that the
radiative cooling time in the shock front is of the order of ~200
s. The waves are thus damped within a sound-speed crossing time.
The $u_{||}$ panel shows another interesting phenomenon. For
600~s $\lesssim t \lesssim 800$~s
it shows a white patch around $s=5$ Mm and a black
patch around
$s=2$ Mm. These are flows along the field line from the apex
towards both footpoints and thus represent the field line draining of
mass as it rises through the chromosphere.
Finally we look at the plasma motion orthogonal to the field line in
the panels $u_{\vect{P}}$ and $u_{\vect{N}}$. The field line is
approximately semi-circular, so $u_{\vect{P}}$ can roughly be thought
of as vertical oscillation, and $u_{\vect{N}}$ as horizontal
oscillation. Careful inspection of both panels shows a pattern of
perturbations propagating with the Alfv\'en speed. The perturbations
originate in the $\beta=1$ layer in both footpoints and travel all
along the field line to the other footpoint with little or no apparent
damping. There appears to be no wave reflection. The $\nabla_{\perp}
\cdot \vect{u}$ panel shows that these perturbations are only weakly
compressive. Only the perturbation around $t=5$ hs leaves a clearly
recognizable imprint of weak compression orthogonal to the field line.
The panel that shows $\omega_{||}$
exhibits a strong imprint of perturbations traveling with the Alfv\'en
speed and thus provide a hint of the presence of torsional
waves. However, true confirmation requires a more thorough
investigation of multiple field lines.
Field line B, displayed in Figure~\ref{fig:fibril_038_st}, shows by and
large the same features as Figure~\ref{fig:fibril_050_st}: the field
line remains at chromospheric temperatures, it carries slow-mode waves
that are damped before they cross to the other footpoint and it
shows an interference pattern of transverse waves and possibly
torsional waves that propagate at Alfv\'en speed. The biggest
difference is the velocity amplitude of the transverse waves, it
reaches up to 15 \mbox{km s$^{-1}$}\ for this case, while field line A in
Figure~\ref{fig:fibril_050_st} showed a maximum amplitude of only
7 \mbox{km s$^{-1}$}.
\subsection{Time evolution of a coronal field line}
Field line C shown in Figure~\ref{fig:fibril_019_st} behaves
differently from the chromospheric field lines discussed in
Section~\ref{sec:chromfl}. The loop is $\sim20$ Mm long, twice the
length of the chromospheric field lines A and B. A large fraction of
the loop has coronal temperatures, and only a short part close to the
footpoints are chromospheric. The temperature and mass density
panels show
the formation of a cold, dense condensation at $t=600$~s and
$s=6$~Mm. After its formation, it falls towards the chromosphere,
hitting it at $t=1000$~s. These properties are similar to what is
observed for coronal rain, even though the spatial scale in the
simulation is much smaller
\citep[see for example][and references
therein]{2011ApJ...736..121A,2012SoPh..280..457A}.
Slow-mode waves propagating along this field line ($u_\mathrm{para}$ and
$\nabla_{||} \cdot \vect{u}$ panels) do travel the whole length of the
field line. This is caused by the quadratic density dependence of the
radiative cooling.
The wave crosses the chromosphere too fast
to be completely damped.
Once the wave front reaches the corona, the
radiative cooling timescale goes down dramatically and becomes longer
than the wave crossing time. \edt{Both loop footpoints are sources of slow-mode waves, and their lack of damping in the corona leads to an interference pattern in $u_\mathrm{para}$ and $\nabla_{||} \cdot \vect{u}$, but the latter is not as clear in Figure~\ref{fig:fibril_019_st} because of their small compression compared to the slow-modes in the chromosphere.}
The $u_{\vect{P}}$ and $u_{\vect{N}}$ again show transverse
oscillations running at Alfv\'en speed along the field line
\edt{in both directions}.
The
vorticity panel ($\omega_{||}$) shows a
hint of torsonial waves.
Note that the Alfv\'en crossing time for this coronal field line is
$\sim40$~s assuming a length of 20 Mm and a speed of 500 \mbox{km s$^{-1}$},
so that we barely resolve the propagation at the 10 s cadence of our
time series.
\subsection{Velocity power spectrum}
\begin{figure*}
\includegraphics[width=8.8cm]{./fig15a.eps}
\includegraphics[width=8.8cm]{./fig15b.eps}
\includegraphics[width=8.8cm]{./fig15c.eps}
\includegraphics[width=8.8cm]{./fig15d.eps}
\caption{Top row and bottom left: Velocities as function of time
parallel ($u_{||}$, green) and perpendicular to the magnetic
field ($u_N$ in red and $u_P$ in blue) along the three field
lines A,B,C. Subpanels show the velocities at different locations
along the field line, where 100\% means the length from one
footpoint to the other. Lower right panel: power
spectrum of the velocity components at the loop apexes averaged
over all 80 field lines. Only positive frequencies are
shown.
\label{fig:power_spect}}
\end{figure*}
For completeness we further quantify the flow velocities along the
field lines A, B and C in Figure~\ref{fig:power_spect}. The velocities
increase from the footpoints towards the field line midpoints. The
parallel and perpendicular velocities have similar magnitude at
similar locations along the field lines. The parallel velocity
$u_{||}$ varies on longer time scales than the two perpendicular
components. This slow variation is associated with field lines loading
and draining of mass.
In the lower right panel of Figure~\ref{fig:power_spect} we show the
power spectrum of the three velocity components at the midpoint of the
field lines (i.e., halfway between the footpoints). These power
spectra were obtained by computing a power spectrum of each velocity
component at 11 points equally spaced between 40\% and 60\% of the
length of each field line. For each velocity component these spectra
were then averaged over the 11 points and over all field lines. The
averaging over 11 points was done to beat down the noise.
The high power of $u_{||}$ at the lowest resolvable frequencies again
represents loop mass loading and draining. There are weak maxima of
power at 5~mHz in all three velocity components. Given the limited
sample size and duration of the time series, we do not know whether
these peaks are significant. If they are, they might correspond to the
familiar 3-minute oscillations. The typical power of 1-2 km$^2$
s$^{-2}$ mHz$^{-1}$ are the same order of magnitude as those found
based on fibril oscillations in Figure~4 of
\citet{2014ApJ...784...29M}.
\section{Discussion and conclusions} \label{sec:discussion}
In this paper we investigated the relation between fibrils, dark elongated features seen in images of the Sun taken in the core of the chromospheric \mbox{H$\alpha$}\ line, and magnetic field lines that thread through the chromosphere.
We compared synthetic \mbox{H$\alpha$}\ imagery
computed from an RMHD simulation of an enhanced network area on the
Sun with observations. The simulation behaves similar in terms of
intensity contrast, Doppler shifts and time evolution. A difference is
that the fibrils in the synthetic imagery appear as a cross between
active region and quiet Sun fibrils. Comparison of the properties of
the oscillatory behaviour of fibrils in the synthetic line core
imagery with those observed by
\citet{2012ApJ...750...51K,2013ApJ...779...82K}
and
\citet{2014ApJ...784...29M}
showed that the simulated fibrils have periods, amplitudes, and phase
speeds consistent with the observations.
We then proceeded with an extensive analysis of the relation between
fibrils and fibril-threading field lines in our simulation. This
relation is one of complex interplay between field line dynamics on
the one hand, and the formation of the \mbox{H$\alpha$}\ line on the other
hand.
The evolution of the field lines themselves is extremely
intricate.
The slow motion of their footpoints by the evolving granulation
pattern leads to a slow migration of the field line paths through the
3D computational volume on time scales of several hundred seconds. In
our examples, the field lines tend to evolve from lower
to
higher-reaching loops. The field lines can in general thus not be
thought of as a static background along which waves propagate.
Footpoint buffeting by radial box oscillations (corresponding to
solar $p$-modes) and granules cause excitation of different types
of waves.
We identified the presence of compressive longitudinal slow-mode
waves, that steepen to shocks at chromospheric heights and propagate
with speeds close to the sound speed. The plasma $\beta$ is smaller
than unity along the field lines, so that these waves propagate
parallel to the field lines.
If these field lines are close
to horizontal, then the large angle of the field lines with the
vertical combined with the frozen in condition might prevent the loop
from draining quickly. For those field lines that do not reach coronal
temperatures, the slow-modes are typically damped before the wave
front reaches the other footpoint. This damping is reduced once the
shock front reaches to the corona, and such waves travel through the
corona to the other footpoint.
The second wave type that we identified is a nearly-incompressible
transverse wave that propagates with the Alfv\'en speed. These waves
are ubiquitously present in all investigated field lines along both
transverse Frenet-Serret directions ($\vect{N}$ and $\vect{P}$). They
have maximum velocity amplitudes of 5--20 \mbox{km s$^{-1}$}. The field lines
typically carry waves travelling in both directions along the field.
We do not find evidence of standing waves or reflection. We also find
signal propagating with Alfv\'en speed in the vorticity parallel to
the magnetic field $\omega_{||}$. Without further analysis we cannot
confirm whether the vorticity signal indicates a torsional wave-like
phenomenon, such as reported in
\citet{2012ApJ...752L..12D,2014Sci...346D.315D}
or whether it is caused by a form of motion comparable to helical kink
waves in idealised linear wave analysis
\citep{2008ApJ...683L..91Z}.
In the latter case they merely represent another way to view the
transverse waves along $\vect{N}$ and $\vect{P}$.
Our field line based analysis approach only brings forward waves that
travel parallel to the magnetic field. The third classical MHD wave,
the fast mode, can propagate at an angle to the magnetic field \edt{and can therefore not be easily identified in our field-line based approach}. While
these waves are present in the simulation, we did not try to identify
them.
From an \mbox{H$\alpha$}\ line formation point of view, the fibrils are dark
elongated features. They appear dark because they reach optical depth
unity at larger height than their surroundings. The \mbox{H$\alpha$}\ opacity
is relatively insensitive to temperature, and depends mainly on mass
density. The location of optical depth unity in the line core is
therefore roughly equal to the height were the integrated mass density
(column mass) along the line of sight reaches a certain value
($3\times 10^{-5}$~g~cm$^{-3}$ in the analysis by
\citet{2012ApJ...749..136L}).
Whether (a part of) a given magnetic field line is visible as part of
a fibril therefore depends on the mass density along the field line,
and the column mass density above it. The column mass density at which
optical depth unity is located is just set by the characteristics of the
radiative transfer in the \mbox{H$\alpha$}\ line, and has no special
significance in terms of properties of the atmosphere.
The combination of the movement of field lines through the atmosphere
and the filling and draining of mass along field lines thus determines
whether a part of a given field line is visible in \mbox{H$\alpha$}\ line core
images. From the analysis of the time evolution of the fibrils and
field lines in the movies associated with
Figure~\ref{fig:fibril_class} we find that a fibril in our simulation
tracks the same set of field lines for about 200~s. The fibrils
themselves keep their identity typically for a longer time (500~s to
1000~s). In this respect our simulation behaves more like the active
region fibrils analysed in Figure~2 of
\citet{2014ApJ...784...29M},
where fibrils retain their identity for about 500~s, whereas the quiet
Sun fibrils are identifiable over shorter periods
(200~s\,--\,300~s). A big open question is whether true solar field
lines migrate through the atmosphere as fast as in our simulation.
Returning to the assumptions behind using fibril oscillations in
coronal seismology in Section~\ref{sec:introduction}, we find the
following: The simulation shows instances where fibrils outline single
magnetic field lines (top row of Figure~\ref{fig:flines}, but also
examples where the fibril
roughly follows a field line bundle,
but different field lines are visible at different locations along the
fibril (middle row of Figure~\ref{fig:flines}) and examples where the
fibril intersects field lines at a large angle (bottom row of
Figure~\ref{fig:flines}). In the latter case the fibril does not
follow single field lines at all. Whether a given fibril traces single
field lines cannot be determined from the \mbox{H$\alpha$}\ observations alone.
The simulation also shows that field lines evolve in time though the
computational domain, and do not intersect a fibril for longer than $~
200$~s. This is comparable to typical transverse wave periods.
We conclude that our simulation only partially supports the
assumptions typically made in chromospheric seismology. There will be
instances where a fibril indeed traces out a single field line over
more than one transverse-wave period, but this cannot be expected to
be generally true. We therefore urge caution when interpreting results
obtained from seismology, as they might contain systematic biases
caused by the partial mapping of fibril oscillations to actual plasma
and field line motion.
The work presented in the manuscript also leads to a number of
intriguing new questions:
What are the precise excitation mechanisms
of the transverse waves along fibril threading field lines? How can
fibrils retain their identity for a time that is longer than the time
that they intersect the same field lines? How much of the transverse
and longitudinal wave energy is deposited in the chromosphere? What is
the dissipation mechanism of the waves? How exactly is mass loaded
into fibril-threading field lines? Which forces act to prevent
the mass-loaded fibrils to sink back deeper into the atmosphere? Why
are fibrils in active regions narrower than in the quiet Sun? We
intend to address these questions in future publications.
\begin{acknowledgements}
The research leading to these results has received funding from the European Research Council
under the European Union's Seventh Framework Programme (FP7/2007-2013) / ERC Grant
agreement nr. 291058. This research was supported
by the Research Council of Norway through
the grant ``Solar Atmospheric Modelling'' and through grants of
computing time from the Programme for Supercomputing. It was also
supported by the Swedish Knut and Alice Wallenberg foundation. The
Swedish 1-m Solar Telescope is operated by the Institute for Solar
Physics of Stockholm University in the Spanish
Observatorio del Roque de los Muchachos of the Instituto de
Astrof\'{\i}sica de Canarias. The authors recognize support from the International Space Science Institute in Bern.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:Introduction}
The $SO(3)$ symmetric spin-1 anti-ferromagnetic Heisenberg chain, which was shown by Haldane~\cite{Haldane-1983-a,Haldane-1983-b,Haldane-1985} to have a symmetry preserving gapped ground state, provides the oldest known example of a bosonic symmetry-protected topological (SPT) state in one dimension.
The topological character of this state is captured by a topological $\theta$-term present in the non-linear sigma model effective action describing long-wavelength degrees of freedom.~\cite{Haldane-1983-a,Haldane-1983-b,Haldane-1985}
An insightful account of the properties of this anti-ferromagnetic spin chain was given by Affleck, Kennedy, Lieb, and Tasaki (AKLT), who constructed a Hamiltonian in the same phase as the Heisenberg model, where the $S=1$ spins emerge from the composition of underlying $S=1/2$ degrees of freedom.~\cite{Affleck-1987,Affleck-1988}
With periodic boundary conditions, the AKLT model has a non-degenerate ground state that does not break any symmetries and is separated from the first excited state by a finite gap.
With open boundary conditions, on the other hand, the AKLT model makes it manifest that each edge supports a ``free" $S=1/2$ degree of freedom contributing to a 2-fold degeneracy per edge.
Interestingly, while the Hamiltonian is $SO(3)$ symmetric and the bulk degrees of freedom transform linearly under this symmetry, the effective $S=1/2$ spins on the edges transform projectively under the action of the spin rotation symmetry: an initial $S=1/2$ spin state rotated by $360$ degrees about an arbitrary axis is mapped into itself up to a minus sign.
The inter-connection among the topological $\theta$-term action, the ground state degeneracy with open boundary conditions and the projective representation of the global symmetry on the edge degrees of freedom makes this system a non-trivial gapped phase of matter.
Inspired by the example of the anti-ferromagnetic chain, there have been recent proposals to classify gapped
SPT phases of matter protected by a global symmetry $G$ using various mathematical frameworks such as group cohomology,~\cite{Chen-2012-a,Chen-2013} which generalizes the concept of projective representations,
topological field theories~\cite{Lu-2012,Chen-2012-b,Vishwanath-2013,Metliski-2013-a,Sule-2013,C-Wang-2014,Ye-2013,Santos-2014,J-Wang-2014}
and non-linear sigma models in the presence of a topological $\theta$-term action compatible with the
global symmetry $G$.~\cite{Bi-2013,Xu-2013}
Recently, a number of microscopic models of bosonic SPT states have been
studied, which help to shed light on the role played by physical interactions in bringing about SPT phases.~\cite{Chen-2012-a,Chen-2013,Chen-2012-b,Santos-2014,J-Wang-2014,Chen-2011,Levin-2012,Chen-2014,Burnell-2013,Fidkowski-2013,Lu-2014,Chen-2014-2,Geraedts-2014}
The purpose of this paper is to provide a framework for constructing microscopic models capable of describing bosonic SPT states.
As we shall see, some of the exactly solvable models
previously studied~\cite{Levin-2012,Chen-2014,Geraedts-2014}
will be identified as special cases of a large class of models to be constructed here.
We shall also be able to construct parent Hamiltonians for two dimensional
$
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
$
paramagnets,
whose effective edge theory was shown in Ref.~\onlinecite{J-Wang-2014}
to be in direct relation to non-trivial $3$-cocycles.
The classes of gapped bosonic insulators protected a by global symmetry $G$ that we shall be concerned with have, on a
$\mathrm{d}$-dimensional closed manifold, a non-degenerate ground state
\begin{equation}
\label{eq: SPT ground state}
\begin{split}
\ket{\Psi_{G}}
=
\frac{1}{\sqrt{Z(\beta)}}\,
\sum_{ s }\,
e^
{
-\frac{\beta}{2}\,E_{G}(s)
+
i\,W_{G}(s)
}
\,
\ket{s}
\,,
\end{split}•
\end{equation}•
where $\{ \ket{s} \}$ denotes an orthonormal many-body basis, $E_{G}(s) \in \mathbb{R}$ is a non-universal local function related to the decay of correlations of local operators in the ground state and the phase $W_{G}(s) \in \mathbb{R}$ is a universal piece that endows the ground state~(\ref{eq: SPT ground state}) with its non-trivial topological properties.
$W_{G}(s)$ plays, at the microscopic level considered here, a role analogous to the topological $\theta$-term.~\cite{Chen-2012-a,Chen-2013,Bi-2013,Xu-2013}
When $W_{G}(s)= 0$, one obtains from Eq.~(\ref{eq: SPT ground state})
the nodeless ground state
\begin{equation}
\label{eq: RK ground state}
\begin{split}
\ket{\Phi_{G}}
=
\frac{1}{\sqrt{Z(\beta)}}\,
\sum_{ s }\,
e^
{
-\frac{\beta}{2}\,E_{G}(s)
}
\,
\ket{s}
\,.
\end{split}•
\end{equation}•
The form of the ground state Eq.~(\ref{eq: RK ground state}) is very appealing; for
equal time correlation functions of operators in the diagonal representation $\{ \ket{s} \}$,
\begin{equation}
\label{eq: Correlation function of diagonal operators}
\begin{split}
&\,
\bra{\Phi_{G}}\,
\hat{\mathcal{O}}_{a}(\hat{s})\,
\hat{\mathcal{O}}_{b}(\hat{s})\,
\ket{\Phi_{G}}
=
\sum_{ s }\,
\mathcal{O}_{a}(s)\,
\mathcal{O}_{b}(s)\,
\frac
{
e^
{
-\beta\,E_{G}(s)
}
}
{
Z(\beta)
}
\,,
\end{split}•
\end{equation}•
can be interpreted as equal time correlation functions
of an equilibrium
$\mathrm{d}$-dimensional
statistical mechanical system with classical configurations $\{ s \}$,
each one occurring with probability
$
p^{(0)}_{s}
=
e^
{
-\beta\,E_{G}(s)
}
/
Z(\beta)
$,
where the real parameter $\beta$ acquires the natural interpretation of an effective inverse temperature and the normalization factor of the ground state,
\begin{equation}
\label{eq: Normalization of the ground state}
\begin{split}
Z(\beta)
=
\sum_{ s }\,
e^
{
-\beta\,E_{G}(s)
}
\,,
\end{split}•
\end{equation}•
is interpreted as the partition function of the classical system in ``thermal" equilibrium.
Hence, if the associated classical model described by the partition function
Eq.~(\ref{eq: Normalization of the ground state}) is in the ``disordered" phase,
then typical correlation functions of local operators distant by
$|\bs{r}|$ behave as
$
e^
{
-|\bs{r}|/\xi
}
$,
for some finite correlation length
$
\xi
$,
and the representation~(\ref{eq: RK ground state}) can be associated with a
quantum many-body ground state in its gapped phase.
In fact the foregoing classical-quantum correspondence can be made more precise.~\cite{Rokhsar-1988,Henley-1997,Ardonne-2004,Henley-2004,Castelnovo-2005}
Eq.~(\ref{eq: RK ground state}) is recognized as the zero energy ground state of a class of
quantum dimer-like models at the so-called Rokhsar-Kivelson (RK) point.~\cite{Rokhsar-1988}
In Ref.~\onlinecite{Rokhsar-1988}, it was noted that dimer-dimer correlation functions of the square lattice quantum dimer model at the RK point can be computed exactly from the
corresponding classical dimer problem.~\cite{Fisher-1963}
In Ref.~\onlinecite{Ardonne-2004}, Ardonne, Fendley and Fradkin have established
the quantum-classical correspondence to more general classes of RK Hamiltonians beyond dimer models.
In Ref.~\onlinecite{Henley-2004},
Henley has observed that any stochastic classical system described by a real transition rate matrix $M$
can be interpreted, via a similarity transformation, as an RK Hamiltonian.
In Ref.~\onlinecite{Castelnovo-2005}, Castelnovo, Chamon, Mudry and Pujol have shown that the reverse is true, namely, that given a quantum RK Hamiltonian in a ``preferred basis" [the basis in which the ground state is expressed as a linear combination of
same-sign coefficients as in Eq.~(\ref{eq: RK ground state})], there exists an associated stochastic classical model whose spectrum of relaxation rates is (up to an overall minus sign) the same as the energy spectrum of the quantum RK Hamiltonian and whose equilibrium probability distribution is the square of the coefficients in the expansion of the RK quantum ground state
Eq.~(\ref{eq: RK ground state}).
In light of the above arguments, if one considers the configurations $\{ s \} $ to be made of spins on a lattice, then the ground state
Eq.~(\ref{eq: RK ground state}) offers a natural representation of a paramagnetic
state, provided the corresponding classical system is chosen to have a spectrum of relaxation rates with a finite gap and correlation functions of local
operators, Eq.~(\ref{eq: Correlation function of diagonal operators}), exhibiting
short-range behavior.
As for the role played by symmetries, we now let the quantum system be invariant under a global symmetry group $G$, whose action on the basis
$
\ket{ s }
$
is represented by
\begin{equation}
\label{eq: Action of the symmetry on the basis elements}
\widehat{S}_{G}\,\ket{ s }
=
\ket{ g\,s }
\,.
\end{equation}•
It is then clear that the ground state Eq.~(\ref{eq: RK ground state}) is a unique and
$G$ invariant state provided the local ``classical energy"
$
E_{G}(s)
$
is symmetric under the transformation
$
\{ s \}
\rightarrow
\{ g\,s \}
$:
\begin{equation}
\label{eq: Transformation of E under G}
\begin{split}
E_{G}(g\,s)
=
E_{G}(s)
\,.
\end{split}•
\end{equation}•
The central point of this paper is the observation, which will be supported by concrete examples, that in a \textit{closed} manifold, the SPT
ground state Eq.~(\ref{eq: SPT ground state}) can be obtained from the
trivial insulator ground state Eq.~(\ref{eq: RK ground state}) via a \textit{global} symmetry-preserving unitary transformation
$
\mathbb{W}_{G}
$,
whose action on the many-body basis $\{ \ket{s} \}$ is
\begin{subequations}
\label{eq: Definition of unitary mapping}
\begin{equation}
\label{eq: Action of W on the basis elements}
\begin{split}
\mathbb{W}_{G}\,\ket{s}
:=
e^
{
i\,W_{G}(s)
}
\,
\ket{s}
\,,
\end{split}•
\end{equation}•
hence,
\begin{equation}
\label{eq: Unitary mapping between Psi and Phi}
\begin{split}
\ket{\Psi_{G}}
=
\mathbb{W}_{G}\,
\ket{\Phi_{G}}
\,.
\end{split}•
\end{equation}•
\end{subequations}•
For the SPT ground state Eq.~(\ref{eq: SPT ground state}) to be invariant under the symmetry $G$, it is required that
\begin{subequations}
\begin{equation}
\label{eq: Transformation of E and W under G unitary}
\begin{split}
&\,
E_{G}(g\,s)
=
E_{G}(s)
\,,
\\
&\,
W_{G}(g\,s)
=
W_{G}(s)
~
\textrm{mod}~2\pi
\,,
\end{split}•
\end{equation}•
if $G$ is an unitary global symmetry, and
\begin{equation}
\label{eq: Transformation of E and W under G anti-unitary}
\begin{split}
&\,
E_{G}(g\,s)
=
E_{G}(s)
\,,
\\
&\,
W_{G}(g\,s)
=
- W_{G}(s)
~
\textrm{mod}~2\pi
\,,
\end{split}•
\end{equation}•
\end{subequations}•
if $G$ is an anti-unitary global symmetry.
Now let $H_{G}$ and $\mathcal{H}_{G}$ be,
respectively, the quantum Hamiltonians whose non-degenerate ground states are
$\ket{\Phi_{G}}$ and $\ket{\Psi_{G}}$.
Then the unitary mapping Eq.~(\ref{eq: Unitary mapping between Psi and Phi}) establishes
\begin{equation}
\label{eq: Unitary mapping between Hamiltonians}
\mathcal{H}_{G}
=
\mathbb{W}_{G}
\,
H_{G}
\,
\mathbb{W}_{G}^{-1}
\,.
\end{equation}•
Starting, thus, from a parent Hamiltonian $H_{G}$ for the trivial gapped state
Eq.~(\ref{eq: RK ground state}),
one can construct a parent Hamiltonian $\mathcal{H}_{G}$ for the SPT ground state
Eq.~(\ref{eq: SPT ground state})
using Eq.~(\ref{eq: Unitary mapping between Hamiltonians}), if the unitary transformation connecting the two ground states,
Eq.~(\ref{eq: Definition of unitary mapping}), is known.
This paper is organized as follows.
In Sec.~(\ref{sec: Review of the classical-quantum mapping}) we review
the relevant points about the mapping between stochastic classical systems
and quantum RK Hamiltonians.
In Sec.~(\ref{sec: SPT states in one dimension}) we construct parent Hamiltonians of one dimensional bosonic SPT states with $\mathbb{Z}_{n} \times \mathbb{Z}_{n}$ symmetry, where we shall use the concept of entanglement spectrum degeneracy to determiner the unitary transformation
Eq.~(\ref{eq: Definition of unitary mapping})
that maps the trivial ground state to other $n-1$ topological phases.
In Sec.~(\ref{sec: Z2T SPT state in one dimension}) we discuss the one
dimensional SPT state with anti-unitary time-reversal symmetry
$\mathbb{Z}^{T}_{2}$.
In Sec.~(\ref{SPT states in two dimensions}) we construct two dimensional microscopic models that account for all the $8$ possible classes of $\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ SPT paramagnets.
We show that the $\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetry transformation projected onto the one dimensional edge acquires a non-onsite form, which was studied in Ref.~\onlinecite{J-Wang-2014} in connection with non-trivial $3$-cocycles in the group cohomology.
Finally we draw some conclusions and point to future directions in
Sec.~(\ref{sec:Conclusions and future directions}).
\section{Classical-quantum mapping}
\label{sec: Review of the classical-quantum mapping}
In order to make the discussion self-contained, we review the essential aspects of the relationship between quantum RK Hamiltonians and stochastic classical systems.~\cite{Henley-2004,Castelnovo-2005} The content of
Secs.~(\ref{sec: Rokhsar-Kivelson Hamiltonians})
and~(\ref{sec: Relation between quantum RK Hamiltonians and stochastic classical systems})
closely follows Ref.~\onlinecite{Castelnovo-2005}, which the reader may consult for further details.
In Sec.~(\ref{sec: SPT-RK Hamiltonians}) we give the general form of the SPT-RK Hamiltonians describing bosonic
SPT states in $\mathrm{d}$-dimensional space obtained from the unitary mapping
Eq.~(\ref{eq: Unitary mapping between Hamiltonians}).
\subsection{Rokhsar-Kivelson Hamiltonians}
\label{sec: Rokhsar-Kivelson Hamiltonians}
A quantum RK Hamiltonian satisfies three properties~\cite{Castelnovo-2005}:
\begin{enumerate}
\item
The orthonormal elements of the basis
$
\Sigma
=
\{
\ket
{
s
}
\}
$,
which span the Hilbert space,
\begin{equation}
\label{eq: Definition of orthonormal basis sigma}
\begin{split}
\bra{s} s' \rangle
=
\delta_{s,s'}
\,,
\quad
\openone
=
\sum_{s}\,
\ket{s}
\bra{s}
\,,
\end{split}•
\end{equation}•
form a countable set.
\item
The quantum Hamiltonian can be decomposed into a sum of positive-semidefinite projector-like Hermitian operators $P_{s,s'}$
[Eq.~(\ref{eq: Definition of the RK Hamiltonian})].
\item
The ground state [Eq.~(\ref{eq: RK ground state 2})] is annihilated by every $P_{s,s'}$ and the normalization constant of the ground state can be interpreted as the partition function of a classical system [Eq.~(\ref{eq: Definition of Z})].
\end{enumerate}•
The RK Hamiltonian takes the form
\begin{subequations}
\label{eq: Definition of the RK Hamiltonian}
\begin{equation}
\label{eq: RK Hamiltonian}
H_{\textrm{RK}}
=
\frac{1}{2}\,
\sum^{s \neq s\rq{}}_{(s,s')}\,
\omega_{s,s'}\,P_{s,s'}
\,,
\end{equation}•
where
\begin{equation}
\omega_{s,s'} \in \mathbb{R}
\,,
\quad
\omega_{s,s'} > 0
\,,
\end{equation}•
\begin{equation}
\label{eq: Definition of P}
\begin{split}
&\,P_{s,s'}
=
-
\ket{s} \bra{s'}
-
\ket{s'} \bra{s}
\\
&\,
+
e^
{
-\frac{\beta}{2}\,
\left[
E(s') - E(s)
\right]
}
\,
\ket{s} \bra{s}
+
e^
{
-\frac{\beta}{2}\,
\left[
E(s) - E(s')
\right]
}
\,
\ket{s'} \bra{s'}
\,,
\end{split}•
\end{equation}•
\end{subequations}•
where $\beta \in \mathbb{R}$, $E(s) \in \mathbb{R}$ and
$P_{s,s'}$ is a projector-like positive semi-definite Hermitian operator satisfying
\begin{equation}
\begin{split}
P^{2}_{s,s'}
=
2\,
\cosh
{
\left\{
\frac{\beta}{2}\,
\left[
E(s')
-
E(s)
\right]
\right\}
}
\,
P_{s,s'}
\,.
\end{split}•
\end{equation}•
One easily verifies that the nodeless state
\begin{equation}
\label{eq: RK ground state 2}
\begin{split}
\ket{\Phi_\textrm{RK}}
=
\frac
{
1
}
{
\sqrt
{
Z(\beta)
}
}
\,
\sum_{s}\,
e^
{
-\frac{\beta}{2}\,E(s)
}
\,
\ket
{
s
}
\,,
\end{split}•
\end{equation}•
with normalization constant
\begin{equation}
\label{eq: Definition of Z}
Z(\beta)
=
\sum_{s}\,
e^
{
-\beta\,E(s)
}
\,,
\end{equation}•
satisfies
\begin{equation}
\begin{split}
P_{s,s'}\,\ket{\Phi_\textrm{RK}}
=
0\,,
\quad
\forall~(s,s')
\,.
\end{split}•
\end{equation}•
Thus Eq.~(\ref{eq: RK ground state 2}) is the ground state of the RK Hamiltonian Eq.~(\ref{eq: Definition of the RK Hamiltonian}) with energy zero.
The normalization constant Eq.~(\ref{eq: Definition of Z}) can be interpreted as the partition function of a classical system with classical energy $E(s)$ at the effective inverse temperature $\beta$.
In the ``infinite temperature" limit $\beta = 0$, the Boltzmann factors in the partition function tend to unity and $Z(0)$ counts the number of allowed configurations.
\subsection{Relation between quantum RK Hamiltonians and stochastic classical systems}
\label{sec: Relation between quantum RK Hamiltonians and stochastic classical systems}
In order to show the relation between quantum RK Hamiltonian and
stochastic classical systems, one considers the real-valued matrix $M$ defined by~\cite{Henley-2004,Castelnovo-2005}
\begin{equation}
\label{eq: Definition of M}
M_{s s'}
\equiv
-
e^
{
-\frac{\beta}{2}\,
\left[
E(s)
-
E(s')
\right]
}
\,
\left(
H_\textrm{RK}
\right)_{ss'}
\,,
\end{equation}•
where
$
\left(
H_\textrm{RK}
\right)_{ss'}
=
\langle\,s\,|\,H_\textrm{RK}\,|\,s'\,\rangle
$
denotes the matrix elements of $H_\textrm{RK}$.
From the fact that $H_\textrm{RK}$ is a real and Hermitian operator, it follows that
\begin{equation}
\label{eq: Symmetry of the RK Hamiltonian}
\left(
H_\textrm{RK}
\right)_{ss'}
=
\left(
H_\textrm{RK}
\right)_{s's}
\,
\end{equation}•
and, from Eqs.~(\ref{eq: Definition of the RK Hamiltonian}) and~(\ref{eq: Definition of M}), that the matrix $M$ satisfies:
\begin{subequations}
\label{eq: Properties of M}
\begin{equation}
\label{eq: Property of M (a)}
M_{s s'} > 0
\,,
\quad
\textrm{if}~
s \neq s'
\,,
\end{equation}•
and
\begin{equation}
\label{eq: Property of M (b)}
M_{s s}
=
-
\sum^{s' \neq s}_{s'}\,
M_{s' s}
\,.
\end{equation}•
\end{subequations}•
One then considers a classical system with a phase space formed
by configurations $\{ s \}$, which, as a function of time $\tau$, can be visited stochastically with probability $p_{s}(\tau)$ evolving according to the master equation
\begin{equation}
\label{eq: Master equation}
\begin{split}
\frac{d\,p_{s}(\tau)}{d\,\tau}\,
&\,=
\sum_{s'}\,
M_{s s'}\,p_{s'}(\tau)
\\
&\,
=
\sum^{s' \neq s}_{s'}\,
\Big[
M_{s s'}\,p_{s'}(\tau)
-
M_{s' s}\,p_{s}(\tau)
\Big]
\,,
\end{split}•
\end{equation}•
where Eq.~(\ref{eq: Property of M (b)}) has been used to achieve the last equality of Eq.~(\ref{eq: Master equation}).
The first term on the r.h.s. accounts for the transitions out of the configurations $s'$ into the configuration $s$, while the second term on the r.h.s. accounts for transitions out of the configuration $s$ into the configurations $s'$.
Moreover Eqs.~(\ref{eq: Definition of M}) and~(\ref{eq: Symmetry of the RK Hamiltonian}) are easily seen to imply, for every pair of indices $(s,s')$,
\begin{subequations}
\label{eq: Detailed balance}
\begin{equation}
M_{s s'}\,p^{(0)}_{s'}
=
M_{s' s}\,p^{(0)}_{s}
\,,
\end{equation}•
where
\begin{equation}
p^{(0)}_{s}
\equiv
\frac
{
1
}
{
Z(\beta)
}
\,
e^
{
-\beta\,E(s)
}
\,.
\end{equation}•
\end{subequations}•
Eq.~(\ref{eq: Detailed balance}) implies the condition of detailed balance
on the matrix $M$ as well as that $p^{(0)}_{s}$ is the equilibrium probability distribution associated with the classical dynamics Eq.~(\ref{eq: Master equation}).
Denoting by $\lambda_{n}$ and $\psi^{(R;n)}_{s}$, respectively, the right-eigenvalues and right-eigenvectors of $M$,
\begin{equation}
\label{eq: Righ-eigenvalue equation of W}
\sum_{s'}\,
M_{s s'}\,
\psi^{(R;n)}_{s'}
=
\lambda_{n}\,
\psi^{(R;n)}_{s}
\,,
\end{equation}•
then the time dependent solution of Eq.~(\ref{eq: Master equation}) can be expressed as
\begin{equation}
\label{eq: Probability amplitudes as a function of time}
p_{s}(\tau)
=
\sum_{n}\,
a_{n}(0)\,
e^
{
\lambda_{n}\,\tau
}
\,
\psi^{(R;n)}_{s}
\,,
\end{equation}•
where $a_{n}(0)$ are coefficients determined by the initial conditions.
Since Eq.~(\ref{eq: Definition of M}) establishes, up to an overall minus sign,
a similarity transformation between $H_{\textrm{RK}}$ and $M$, the spectrum of
relaxation rates $\{ \lambda_{n} \}$ of $M$ and the energy spectrum
$\{ \varepsilon_{n} \}$ of $H_{\textrm{RK}}$ are simply related:
\begin{equation}
\label{eq: Relation between eigenvalues of H and M}
\varepsilon_{n}
=
-\lambda_{n}
> 0
\,.
\end{equation}•
When the classical system whose dynamics is described by Eq.~(\ref{eq: Master equation}) has a spectrum of relaxation rates such that the largest characteristic time scale associated with the decay into the equilibrium configuration is finite in the thermodynamic limit, then it follows from Eq.~(\ref{eq: Relation between eigenvalues of H and M}) that the many-body energy spectrum of the quantum Hamiltonian possess a finite energy gap for excitations above ground state.
\subsection{SPT-RK Hamiltonians}
\label{sec: SPT-RK Hamiltonians}
We now give the general form of the SPT-RK Hamiltonians with ground state given by
Eq.~(\ref{eq: SPT ground state}).
Let the $\mathrm{d}$-dimensional quantum system,
protected by global symmetry $G$, in its trivial phase be described by an
RK Hamiltonian
$
H_{G}
$
of the form
Eq.~(\ref{eq: Definition of the RK Hamiltonian})
with the non-degenerate ground state
$
\ket
{
\Phi_{G}
}
$,
Eq.~(\ref{eq: RK ground state}),
where we impose the symmetry constraint
Eq.~(\ref{eq: Transformation of E under G})
upon $E_{G}(s)$.
Then the unitary transformation
Eq.~(\ref{eq: Definition of unitary mapping})
yields the SPT ground state
$
\ket
{
\Psi_{G}
}
$,
Eq.~(\ref{eq: SPT ground state}),
and
Eq.~(\ref{eq: Unitary mapping between Hamiltonians})
yields the SPT-RK Hamiltonian
$
\mathcal{H}_{G}
$:
\begin{widetext}
\begin{subequations}
\label{eq: Definition of the SPT Hamiltonian}
\begin{equation}
\mathcal{H}_{G}
=
\mathbb{W}_{G}\,
H_{G}\,
\mathbb{W}_{G}^{-1}\,
=
\frac{1}{2}\,
\sum^{s' \neq s}_{(s,s')}\,
\omega_{s,s'}\,\mathcal{P}_{s,s'}
\,,
\end{equation}•
where
\begin{equation}
\omega_{s,s'} \in \mathbb{R}
\,,
\quad
\omega_{s,s'} > 0
\,,
\end{equation}•
\begin{equation}
\label{eq: Definition of mathcal P}
\begin{split}
&\,\mathcal{P}_{s,s'}
=
-
e^
{
i\,
\left[
W_{G}(s)
-
W_{G}(s')
\right]
}
\,
\ket{s} \bra{s'}
-
e^
{
i\,
\left[
W_{G}(s')
-
W_{G}(s)
\right]
}
\,
\ket{s'} \bra{s}
+
e^
{
-\frac{\beta}{2}\,
\left[
E_{G}(s') - E_{G}(s)
\right]
}
\,
\ket{s} \bra{s}
+
e^
{
-\frac{\beta}{2}\,
\left[
E_{G}(s) - E_{G}(s')
\right]
}
\,
\ket{s'} \bra{s'}
\,.
\end{split}•
\end{equation}•
\end{subequations}•
\end{widetext}
Before we proceed to discuss specific SPT systems, we close this section with a few important remarks:
\begin{enumerate}[label=(\alph*)]
\item
In Secs.~(\ref{sec: Rokhsar-Kivelson Hamiltonians})
and~(\ref{sec: Relation between quantum RK Hamiltonians and stochastic classical systems})
we saw that the structure of the RK ground state is determined only by
$
\beta\,E(s)
$,
while the energy spectrum is determined solely by the couplings
$
\omega_{s,s'}
$.~\cite{Henley-2004,Castelnovo-2005}
\item
If the couplings
$
\omega_{s,s'}
$
are such that the energy spectrum has a non-degenerate gapped ground state
(which is the only case considered in this work),
then Eq.~(\ref{eq: RK ground state 2})
parametrizes a large class of ground states, whereby
$
\beta\,E(s)
$
controls the correlation length associated with the decay of correlation functions
such as
Eq.~(\ref{eq: Correlation function of diagonal operators}).
\item
In the particular case $\beta=0$, the ground state
Eq.~(\ref{eq: RK ground state 2})
acquires the form of a trivial product state.
It will prove useful to work in this
\lq\lq{infinite temperature}\rq\rq~
limit, where the correlation length is zero, in order to determine
the form of the unitary transformation
Eq.~(\ref{eq: Definition of unitary mapping}).
\item
While the unitary mapping
Eq.~(\ref{eq: Definition of unitary mapping})
leaves the energy spectrum unchanged,
it introduces a non-trivial entanglement structure in the
SPT ground state
Eq.~(\ref{eq: SPT ground state})
that is responsible for its universal topological properties.
\end{enumerate}
\section{SPT states in one dimension}
\label{sec: SPT states in one dimension}
In this section we derive parent Hamiltonians of one dimensional bosonic
SPT states protected by
$
\mathbb{Z}_{n}\times\mathbb{Z}_{n}
$
symmetry.
In a chain with periodic boundary conditions, each of these phases is described by a non-degenerate gapped symmetric ground state. In a chain with open boundary conditions, on the other hand, there remains a trivial phase with a non-degenerate ground state, while $n-1$ phases have $n$-fold degeneracy per edge, accounting for a total $n^2$-fold degeneracy of the ground state manifold in the thermodynamic limit.
Our aim is to construct the unitary transformations
Eq.~(\ref{eq: Definition of unitary mapping})
connecting the trivial ground state and the $n-1$ non-trivial SPT ground states.
Our strategy in deriving such unitary mappings is to draw on the notion of entanglement spectrum degeneracy as follows.
As we pointed out in the remark (c) of
Sec.~(\ref{sec: SPT-RK Hamiltonians}),
in the ``infinite temperature" limit $\beta = 0$
the ground state of the trivial SPT chain reduces to a
product state.
The entanglement structure of the product state is as simple
as it gets, for the Schmidt decomposition with respect to any partition
contains only one eigenvalue (equal to $1$).
We shall find the unitary transformation
Eq.~(\ref{eq: Definition of unitary mapping})
by demanding that, in the limit $\beta = 0$,
the entanglement spectrum of the non-trivial SPT
ground state, for any partition of the chain, acquires
an $n$-fold degeneracy.
Remarkably, we shall verify that these unitary mappings, via
Eq.~(\ref{eq: Unitary mapping between Hamiltonians}), endow the parent Hamiltonians of non-trivial SPT chains with the required $n$-fold degeneracy of the energy spectrum per edge.
Once the unitary transformation
Eq.~(\ref{eq: Definition of unitary mapping})
is derived, we can obtain the most general form of the SPT
ground state
Eq.~(\ref{eq: SPT ground state})
by allowing $\beta \neq 0$
without changing either the gapped nature of the many-body energy spectrum or
the topological properties of the ground state.
That this is true can be seen perturbatively: moving away from the $\beta=0$ limit
with $\beta << 1$ and $E_{G}(s)$ a local function
in Eq.~(\ref{eq: Definition of the SPT Hamiltonian}), amounts to adding small local symmetry-preserving perturbations to the gapped $\beta=0$ theory, which therefore,
cannot immediately destroy the SPT phase.
Based on this we expect that, for $0< \beta \leq \beta_{c}$,
the ground state Eq.~(\ref{eq: SPT ground state}) describes a class of many-body SPT
ground states adiabatically connected to the $\beta=0$ limit.
\subsection
{
$\mathbb{Z}_{2}\,\times\,\mathbb{Z}_{2}$ SPT states in $\mathrm{d} = 1$
}
\label{sec: Z2 times Z2 in one dimension}
We consider a spin chain,
with an even number $N_{\mathrm{s}}$ of sites,
divided into even and odd sublattices and
$
\mathbb{Z}_{2}\,\times\mathbb{Z}_{2}
$
symmetry generated by
\begin{equation}
\label{eq: Z2 times Z2 symmetry in one dimension}
\begin{split}
\widehat{S}^{(1)}_
{
\mathbb{Z}_{2}
}
=
\prod_{j \in \mathrm{even}}\,
X_{j}
\,,
\quad
\widehat{S}^{(2)}_
{
\mathbb{Z}_{2}
}
=
\prod_{j \in \mathrm{odd}}\,
X_{j}
\,.
\end{split}•
\end{equation}•
In the following,
$
\{ X, Y, Z\}
$
denote the three Pauli matrices.
We start by examining a trivial spin chain in the special limit where it is represented by a product state,
\begin{equation}
\label{eq: Trivial Z2 times Z2 ground state in one dimension}
\begin{split}
\ket{\Phi_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}}
=
\prod^{N_{\mathrm{s}}}_{j =1}
\,
\ket{\rightarrow}_{j}
=
\frac
{
1
}
{
2^{N_{\mathrm{s}}/2}
}
\sum_{ s }
\,
\ket{s}
\,,
\end{split}
\end{equation}
where
$
X_{j}\,\ket{\rightarrow}_{j}
=
\ket{\rightarrow}_{j}
$
and $\{\,\ket{s}\,\}$ represents the $2^{N_{\mathrm{s}}}$ many-body spin states in the eigenbasis of $Z$ operators:
$
Z_{j}\,\ket{s_{j}}
=
s_{j}\,\ket{s_{j}}
$,
for
$
s_{j}
\pm
1
$.
We recognize
Eq.~(\ref{eq: Trivial Z2 times Z2 ground state in one dimension}) as the
ground state Eq.~(\ref{eq: RK ground state}) in the $\beta = 0$ limit.
It has the parent Hamiltonian
\begin{equation}
\label{eq: Trivial Hamiltonian for the Z2 times Z2 state in one dimension}
\begin{split}
H_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}
=
\omega_{0}
\,
\sum^{N_{\mathrm{s}}}_{j=1}\,
\left(
\,
1 - X_{j}
\,
\right)
\,,
\end{split}•
\end{equation}•
which is an RK Hamiltonian of the form Eq.~(\ref{eq: Definition of the RK Hamiltonian})
with $\beta = 0$
and
$
\omega_{s,s'}
=
\omega_{0}
>
0
$,
where the sum in Eq.~(\ref{eq: RK Hamiltonian}) extends
over pairs of states $(s,s')$ which differ by a single spin flip.
The product state form of the ground state
Eq.~(\ref{eq: Trivial Z2 times Z2 ground state in one dimension})
implies that its Schmidt decomposition, with respect to any partition $\Sigma_{i}$ between sites $i$ and $i+1$ of the lattice, contains a single Schmidt eigenvalue
(equal to $1$).
Now, for $\theta_{j} \in \mathbb{R}$, we introduce the unitary mapping
\begin{equation}
\label{eq: General form of the unitary mapping Z2 times Z2 in one dimension}
\begin{split}
\mathbb{W}^{\theta}_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
\equiv
\prod_{j}\,
e^
{
i\,\theta_{j}\,
\left(
1
-
Z_{j}\,Z_{j+1}
\right)/2
}
\,
\end{split}•
\end{equation}•
and the SPT state
\begin{equation}
\label{eq: Non-trivial Z2 times Z2 ground state in one dimension}
\begin{split}
\ket{\Psi_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}}
&\,
\equiv
\mathbb{W}^{\theta}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}\,
\ket{\Phi_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}}
\\
&\,
\equiv
\frac
{
1
}
{
2^{N_{\mathrm{s}}/2}
}
\sum_{ s }
\,
e^
{
i\,W_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}(s)
}
\,
\ket{s}
\\
&\,
=
\frac
{
1
}
{
2^{N_{\mathrm{s}}/2}
}
\sum_{ s }
\,
e^
{
i\,
\sum_{j}\,
\theta_{j}\,
\left(
1
-
s_{j}\,s_{j+1}
\right)/2
}
\ket{s}
\,.
\end{split}
\end{equation}
The unitary transformation
Eq.~(\ref{eq: General form of the unitary mapping Z2 times Z2 in one dimension})
endows the state
Eq.~(\ref{eq: Non-trivial Z2 times Z2 ground state in one dimension})
with an amplitude $e^{i\,\theta_{j}}$ for every domain wall between neighbor spins in the many-body configuration $\ket{s}$.
We seek to find $\theta_{j}$ for which
Eq.~(\ref{eq: Non-trivial Z2 times Z2 ground state in one dimension})
describes the SPT ground state.
Due to the product state nature of
Eq.~(\ref{eq: Trivial Z2 times Z2 ground state in one dimension})
and the pairwise entanglement induced by the unitary mapping
Eq.~(\ref{eq: General form of the unitary mapping Z2 times Z2 in one dimension}),
one effortlessly finds that, for any partition $\Sigma_{i}$, the reduced density operator obtained by tracing over one of the subsystems is given by the $2 \times 2$ matrix
\begin{equation}
\rho_{i}
=
\frac{1}{2}\,
\begin{pmatrix}
1 & \cos{(\theta_{i})}
\\
\cos{(\theta_{i})} & 1
\end{pmatrix}•
\,.
\end{equation}•
For $\theta_{i}=0$ the above density matrix has a single non-zero Schmidt
eigenvalue.
The existence of $2$ degenerate Schmidt eigenvalues (equal to $1/2$) is verified for
\begin{equation}
\label{eq: Value of theta for the Z2 times Z2 in one dimension}
\theta_{i}
=
\pm \pi/2
\,.
\end{equation}•
Moreover, imposing that the unitary transformation
Eq.~(\ref{eq: General form of the unitary mapping Z2 times Z2 in one dimension})
commutes with either of the $\mathbb{Z}_{2}$ symmetries
in Eq.~(\ref{eq: Z2 times Z2 symmetry in one dimension}) yields the final form
\begin{equation}
\label{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}
\begin{split}
\mathbb{W}_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
=
\prod_{j}\,
e^
{
i\,\frac{\pi}{4}\,
\left(
1
-
Z_{2\,j-1}\,Z_{2\,j}
\right)
}
\,
e^
{
-i\,\frac{\pi}{4}\,
\left(
1
-
Z_{2\,j}\,Z_{2\,j+1}
\right)
}
\,.
\end{split}•
\end{equation}•
One then finds
\begin{equation}
\begin{split}
&\,
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}
\,
X_{2j}
\,
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}^{-1}
=
X_{2j}\,
Z_{2j-1}
\,
Z_{2j+1}
\,,
\\
&\,
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}
\,
X_{2j+1}
\,
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}^{-1}
=
X_{2j+1}\,
Z_{2j}
\,
Z_{2j+2}
\,.
\end{split}•
\end{equation}•
The operator
$
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}
\,
X_{j}
\,
\mathbb{W}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}^{-1}
$
can be regarded as a modified Pauli spin operator (since it is obtained
from $X_{j}$ by a unitary transformation) which is ``dressed" by the domain wall operator
$
Z_{j-1}\,Z_{j+1}
$
with support on the opposite sublattice.
So, under the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}),
the SPT Hamiltonian at zero correlation length is
\begin{equation}
\label{eq: Non-trivial Z2 times Z2 Hamiltonian in one dimension}
\begin{split}
\mathcal{H}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
&\,
\equiv
\mathbb{W}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
\,
H_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
\,
\mathbb{W}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}^{-1}
\\
&\,=
\omega_{0}
\,
\sum^{N_{\mathrm{s}}}_{j=1}
\,
\Big(
1
-
Z_{j-1}
\,
X_{j}
\,
Z_{j+1}
\Big)
\,.
\end{split}•
\end{equation}•
We note that the model
Eq.~(\ref{eq: Non-trivial Z2 times Z2 Hamiltonian in one dimension})
has been constructed in Ref.~\onlinecite{Chen-2014} using the concept of
decorated domain walls, while we have arrived on it by appealing to the notion
of entanglement spectrum via the unitary mapping
Eq.~(\ref{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}).
Ground state degeneracy can be easily attested by studying this model
with open boundary conditions, where
\begin{equation}
\begin{split}
\mathcal{H}^{\textrm{open}}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
&\,=
\omega_{0}
\,
\sum^{N_{\mathrm{s}-1}}_{j=2}
\,
\Big(
1
-
Z_{j-1}
\,
X_{j}
\,
Z_{j+1}
\Big)
\,.
\end{split}•
\end{equation}•
The fact that the above Hamiltonian commutes with $Z_{1}$ and $Z_{N_{\mathrm{s}}}$
implies that there are $2$-fold degenerate states associated with the left and the right edges corresponding to states with
$
s_{1}
=
\pm
1
$
and
$
s_{N_{\mathrm{s}}}
=
\pm
1
$.
Thus the universal properties of the SPT state studied here
are encoded in the unitary mapping
Eq.~(\ref{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}).
With the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}),
we can obtain the more general form of the SPT
ground state
Eq.~(\ref{eq: SPT ground state})
by allowing $\beta \neq 0$
without changing either the gapped nature of the many-body energy spectrum or
the topological properties of the ground state.
In this regard, the Hamiltonian
Eq.~(\ref{eq: Non-trivial Z2 times Z2 Hamiltonian in one dimension}) is a particular example of a larger class of SPT models described in
Eq.~(\ref{eq: Definition of the SPT Hamiltonian}), with the phase factors
$
e^
{
i\,
W_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}(s)
}
$
given by acting with the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Z2 times Z2 in one dimension}).
on the state $\ket{s}$.
\subsection
{
$\mathbb{Z}_{3}\,\times\,\mathbb{Z}_{3}$ SPT states in $ \mathrm{d} = 1$
}
\label{sec: Z3 times Z3 in one dimension}
We consider a spin chain, with an even number $N_{\mathrm{s}}$ of sites and
$
\mathbb{Z}_{3}\,\times\mathbb{Z}_{3}
$
symmetry generated by
\begin{equation}
\label{eq: Z3 times Z3 symmetry in one dimension}
\begin{split}
\widehat{S}^{(1)}_
{
\mathbb{Z}_{3}
}
=
\prod_{j \in \mathrm{even}}\,
\tau_{j}
\,,
\quad
\widehat{S}^{(2)}_
{
\mathbb{Z}_{3}
}
=
\prod_{j \in \mathrm{odd}}\,
\tau_{j}
\,.
\end{split}•
\end{equation}•
where, at each site $j$, we consider the operators
\begin{equation}
\begin{split}
\tau_{j}
=
\begin{pmatrix}
0 & 0 & 1
\\
1 & 0 & 0
\\
0 & 1 & 0
\end{pmatrix}•
\,,
\quad
{\sigma}_{j}
=
\begin{pmatrix}
1 & 0 & 0
\\
0 & \omega & 0
\\
0 & 0 & \omega^2
\end{pmatrix}•
\,,
\end{split}•
\end{equation}•
satisfying
$
\tau^{3}_{j}
=
{\sigma}^{3}_{j}
=
1
$,
$
\tau^{\dagger}_{j}
\,
{\sigma}_{j}
\,
\tau_{j}
=
\omega
\,
{\sigma}_{j}
$,
where
$\omega = e^{i\,2\pi/3}$.
Let $\ket{s_{i}}$,
for
$
s_{i}
=
0,1,2
$,
be the eigenstates of ${\sigma}_{i}$:
$
{\sigma}_{i}
\,
\ket{s_{i}}
=
\omega^{s_{i}}\,
\ket{s_{i}}
$.
As in
Sec.~(\ref{sec: Z2 times Z2 in one dimension}),
we start with a trivial $\mathbb{Z}_{3} \times \mathbb{Z}_{3}$
paramagnet described by the ground state
\begin{equation}
\label{eq: Trivial Z3 times Z3 ground state in one dimension}
\begin{split}
\ket{\Phi_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}}
\equiv
\prod_{j}\,
\Big(
\frac
{
\ket{0_{j}}
+
\ket{1_{j}}
+
\ket{2_{j}}
}
{
\sqrt{3}
}
\Big)
=
\frac
{
1
}
{
3^{N_{\mathrm{s}}/2}
}
\,
\sum_{\{ s \}}\,
\ket{s}
\,,
\end{split}•
\end{equation}•
where $\{\,\ket{s}\,\}$ represents the $3^{N_{\mathrm{s}}}$ many-body spin states in the eigenbasis of ${\sigma}$ operators.
We recognize Eq.~(\ref{eq: Trivial Z3 times Z3 ground state in one dimension}) as the
RK ground state Eq.~(\ref{eq: RK ground state}) in the $\beta = 0$ limit.
It has the parent Hamiltonian
\begin{equation}
\label{eq: Trivial Hamiltonian for the Z3 times Z3 state in one dimension}
H_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}
=
\omega_{0}
\,
\sum^{N_{s}}_{j=1}\,
\left[
2
-
\left(
\tau_{j}
+
\tau^{\dagger}_{j}
\right)
\right]
\,,
\end{equation}•
which is an RK Hamiltonians of the form Eq.~(\ref{eq: Definition of the RK Hamiltonian})
with $\beta = 0$,
$
\omega_{s,s'}
=
\omega_{0}
>
0
$,
where the sum in
Eq.~(\ref{eq: RK Hamiltonian})
extends over pairs of states $(s,s')$ which differ by a single $\mathbb{Z}_{3}$ spin flip.
We now consider, for $\theta_{j} \in \mathbb{R}$, the unitary transformation
\begin{equation}
\label{eq: General form of the unitary mapping Z3 times Z3 in one dimension}
\begin{split}
\mathbb{W}^{\theta}_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}
=
\prod_{j}\,
\exp
{
\left[
i\,\theta_{j}
\sum^{2}_{a=1}\,
\frac
{
1-({\sigma}^{\dagger}_{j}\,{\sigma}_{j+1})^{a}
}
{
(
\bar{\omega}^a - 1
)
\,
(
\omega^{a} - 1
)
}
\right]
}
\,
\end{split}•
\end{equation}•
and the SPT state
\begin{equation}
\label{eq: Non-trivial Z3 times Z3 ground state in one dimension}
\begin{split}
&\,
\ket{\Psi_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}}
\equiv
\mathbb{W}^{\theta}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}\,
\ket{\Phi_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}}
\\
&\,
\equiv
\frac
{
1
}
{
3^{N_{\mathrm{s}}/2}
}
\sum_{ s }
\,
e^
{
i\,W_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}(s)
}
\,
\ket{s}
\\
&\,
=
\frac
{
1
}
{
3^{N_{\mathrm{s}}/2}
}
\sum_{ s }
\,
\exp
{
\left[
\sum_{j}\,
i\,\theta_{j}
\sum^{2}_{a=1}\,
\frac
{
1
-
\omega^
{
a\,\left(
s_{j+1}
-
s_{j}
\right)
}
}
{
(
\bar{\omega}^a - 1
)
\,
(
\omega^{a} - 1
)
}
\right]
}
\ket{s}
\,.
\end{split}
\end{equation}
The unitary transformation
Eq.~(\ref{eq: General form of the unitary mapping Z3 times Z3 in one dimension}) endows the state
Eq.~(\ref{eq: Non-trivial Z3 times Z3 ground state in one dimension})
with an amplitude $e^{i\,\theta_{j}}$ for every
pair of neighbor spins $j$ and $j+1$ for which
$
s_{j}
\neq
s_{j+1}
~
(\textrm{mod}~3)
$
in the many-body configuration $\ket{s}$.
Due to the product state nature of
Eq.~(\ref{eq: Trivial Z3 times Z3 ground state in one dimension})
and the pairwise entanglement induced by the unitary mapping
Eq.~(\ref{eq: General form of the unitary mapping Z3 times Z3 in one dimension}),
one effortlessly verifies that, for any partition $\Sigma_{i}$, the reduced density operator obtained by tracing over one of the subsystems is given by the $3 \times 3$ matrix
\begin{equation}
\rho_{i}
=
\frac{1}{3}
\,
\left(
\begin{array}{ccc}
1 & f(\theta_{i}) & f(\theta_{i}) \\
f(\theta_{i}) & 1 & f(\theta_{i}) \\
f(\theta_{i}) & f(\theta_{i}) & 1 \\
\end{array}
\right)
\,,
\end{equation}•
where
$
f(\theta)
=
\frac{1}{3} (1+2 \cos{(\theta)} )
$.
For $\theta_{i}=0$ the above density matrix has a single non-zero Schmidt
eigenvalue.
The existence of $3$ degenerate Schmidt eigenvalues (equal to $1/3$) is
then verified for
\begin{equation}
\theta^{(p)}_{i}
=
\frac{2\pi}{3}\,p
\,,
\quad
p = 1, 2
\,.
\end{equation}•
Moreover, imposing that the unitary transformation
Eq.~(\ref{eq: General form of the unitary mapping Z3 times Z3 in one dimension})
commutes with either of the $\mathbb{Z}_{3}$ symmetries
in Eq.~(\ref{eq: Z3 times Z3 symmetry in one dimension}) yields the final form
\begin{widetext}
\begin{equation}
\label{eq: Particular form of the unitary mapping Z3 times Z3 in one dimension}
\begin{split}
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
&\,=
\prod_{j}\,
\exp
{
\Big[
i\,\frac{2 \pi p}{3}
\,
\sum^{3}_{a=1}\,
\frac
{
1
-
(
{\sigma}^{\dagger}_{2j-1}\,
{\sigma}_{2j}
)^a
}
{
(
\omega^a - 1
)
\,
(
\bar{\omega}^a - 1
)
}
\Big]
}
\,
\exp
{
\Big[
-i\,\frac{2 \pi p}{3}
\,
\sum^{2}_{a=1}\,
\frac
{
1
-
(
{\sigma}^{\dagger}_{2j}\,
{\sigma}_{2j+1}
)^a
}
{
(
\omega^a - 1
)
\,
(
\bar{\omega}^a - 1
)
}
\Big]
}
\,,
\quad
p = 1,2
\,.
\end{split}•
\end{equation}•
\end{widetext}
Eq.~(\ref{eq: Particular form of the unitary mapping Z3 times Z3 in one dimension})
establishes two unitary mappings between the trivial SPT ground state and the two non-trivial SPT ground states.
Moreover we find that
\begin{equation}
\begin{split}
&\,
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\,
\tau_{2j}
\,
\left(
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\right)^{-1}
=
\tau_{2j}
\,
(
{\sigma}_{2j-1}\,{\sigma}^{\dagger}_{2j+1}
)^{p}
\,,
\\
&\,
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\,
\tau_{2j+1}
\,
\left(
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\right)^{-1}
=
\tau_{2j+1}
\,
(
{\sigma}^{\dagger}_{2j}\,
{\sigma}_{2j+2}
)^{p}
\,,
\end{split}•
\end{equation}•
so that the operator $\tau_{j}$ gets ``dressed" by a domain wall like operator --
carrying $\mathbb{Z}_{3}$ charge $p = 1,2$ -- with support on the opposite sublattice.
The SPT Hamiltonians in the $\beta=0$ limit then read
\begin{widetext}
\begin{equation}
\label{eq: Non-trivial Z3 times Z3 Hamiltonian in one dimension}
\begin{split}
\mathcal{H}^{(p)}_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}
&\,=
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\,
H_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}
\,
\left(
\mathbb{W}_{\mathbb{Z}_{3}\times\mathbb{Z}_{3}}^{(p)}
\right)^{-1}
\\
&\,
=
\omega_{0}
\,
\sum_{j}\,
\Big\{
\left[
\,
1
-
\tau_{2j}
\,
(
{\sigma}_{2j-1}\,
{\sigma}^{\dagger}_{2j+1}
)^{p}
\,
\right]
+
\textrm{H.c.}
\Big\}
+
\omega_{0}
\,
\sum_{j}\,
\Big\{
\left[
\,
1
-
\tau_{2j+1}
\,
(
{\sigma}^{\dagger}_{2j}\,
{\sigma}_{2j+2}
)^{p}
\,
\right]
+
\textrm{H.c.}
\Big\}
\,,
\end{split}
\end{equation}
\end{widetext}
for
$
p = 1,2.
$
As in Sec.~(\ref{sec: Z2 times Z2 in one dimension}),
degeneracy of the ground state energy manifold can be checked by placing this system in an open chain, in which case, one finds that the SPT Hamiltonians commutes with ${\sigma}_{1}$ and ${\sigma}_{N_{\mathrm{s}}}$, thus implying a $3$-fold degenerate state associated to the left and right edges.
Thus the universal properties of the SPT state studied here
are encoded in the unitary mapping
Eq.~(\ref{eq: Particular form of the unitary mapping Z3 times Z3 in one dimension}).
With the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Z3 times Z3 in one dimension})
we can obtain the more general form of the SPT ground state
Eq.~(\ref{eq: SPT ground state})
by allowing $\beta \neq 0$ without changing either the gapped nature of the many-body energy spectrum or the topological properties of the ground state.
In this regard, the Hamiltonian
Eq.~(\ref{eq: Non-trivial Z3 times Z3 Hamiltonian in one dimension}) provides a particular example of a larger class of models described in
Eq.~(\ref{eq: Definition of the SPT Hamiltonian}), with the phase factors
$
e^
{
i\,
W_{\mathbb{Z}_{3} \times \mathbb{Z}_{3}}(s)
}
$
given by acting with the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Z3 times Z3 in one dimension}).
on the state $\ket{s}$.
\subsection
{
$\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}$ SPT states in $ \mathrm{d} = 1$
}
\label{sec: Zn times Zn in one dimension}
We generalize the findings of Secs.~(\ref{sec: Z2 times Z2 in one dimension})
and~(\ref{sec: Z3 times Z3 in one dimension})
to $\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}$
SPT states (see the Appendix for definitions and useful formulas).
We arrive at $n-1$
$\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}$
symmetric unitary transformations, labeled by
$
p = 1, ... , n-1
$,
\begin{widetext}
\begin{equation}
\label{eq: Particular form of the unitary mapping Zn times Zn in one dimension}
\begin{split}
\mathbb{W}_{\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}}^{(p)}
&\,=
\prod_{j}\,
\exp
{
\Big[
i\,\frac{2 \pi p}{n}
\,
\sum^{n-1}_{a=1}\,
\frac
{
1
-
(
{\sigma}^{\dagger}_{2j-1}\,
{\sigma}_{2j}
)^a
}
{
(
\omega^a - 1
)
\,
(
\bar{\omega}^a - 1
)
}
\Big]
}
\,
\exp
{
\Big[
-i\,\frac{2 \pi p}{n}
\,
\sum^{n-1}_{a=1}\,
\frac
{
1
-
(
{\sigma}^{\dagger}_{2j}\,
{\sigma}_{2j+1}
)^a
}
{
(
\omega^a - 1
)
\,
(
\bar{\omega}^a - 1
)
}
\Big]
}
\,.
\end{split}•
\end{equation}•
\end{widetext}
It can be shown, using
Eqs.~(\ref{eq: Zn algebra appendix}),
~(\ref{eq: Algebra Zn operators})
and~(\ref{eq: Identity Zn operators}), that
\begin{equation}
\begin{split}
&\,
\mathbb{W}_{\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}}^{(p)}
\,
\tau_{2j}
\,
\left(
\mathbb{W}_{\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}}^{(p)}
\right)^{-1}
=
\tau_{2j}
\,
(
{\sigma}_{2j-1}\,{\sigma}^{\dagger}_{2j+1}
)^{p}
\,,
\\
&\,
\mathbb{W}_{\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}}^{(p)}
\,
\tau_{2j+1}
\,
\left(
\mathbb{W}_{\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}}^{(p)}
\right)^{-1}
=
\tau_{2j+1}
\,
(
{\sigma}^{\dagger}_{2j}\,
{\sigma}_{2j+2}
)^{p}
\,,
\end{split}•
\end{equation}•
for
$
p = 1, ... , n-1
$.
The
$
\mathbb{Z}_{n}\,\times\,\mathbb{Z}_{n}
$
SPT Hamiltonians in the $\beta = 0$ limit are of the same form as
Eq.~(\ref{eq: Non-trivial Z3 times Z3 Hamiltonian in one dimension})
with
$
p = 1, ..., n-1
$.
We notice that this class of one dimensional SPT models has been studied in
Ref.~\onlinecite{Geraedts-2014} without reference to the connection
between unitary transformations and ground state entanglement spectrum that we
are exploring here.
Moreover, in our formalism, we do not need to be restricted to the $\beta=0$
limit for, with the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Zn times Zn in one dimension}),
we can construct a large class of SPT-RK Hamiltonians of the form
Eq.~(\ref{eq: Definition of the SPT Hamiltonian})
with the phase factors
$
e^
{
i\,
W_{\mathbb{Z}_{n} \times \mathbb{Z}_{n}}(s)
}
$
given by acting with the unitary transformation
Eq.~(\ref{eq: Particular form of the unitary mapping Zn times Zn in one dimension})
on the basis state $\ket{s}$.
\section{
SPT state in one dimension with time-reversal symmetry
}
\label{sec: Z2T SPT state in one dimension}
We now study the one dimensional bosonic SPT states protected by time-reversal symmetry which, being an anti-unitary symmetry, requires conditions
Eq.~(\ref{eq: Transformation of E and W under G anti-unitary}) to be satisfied.
The action of time-reversal symmetry shall be represented by the anti-unitary operator
\begin{subequations}
\label{eq: Z2T symmetry in one dimension}
$
\Theta
$,
\begin{equation}
\Theta\,\Theta^{-1}
=
1
\,,
\quad
\Theta^2 = 1
\,,
\end{equation}•
where we work with the representation
\begin{equation}
\begin{split}
\Theta
&\,=
\Big(
\,
\prod_{j}\,X_{j}
\,
\Big)
K
\,,
\end{split}•
\end{equation}•
\end{subequations}•
with $K$ denoting the complex conjugation operator.
A one dimensional chain of $N_{\mathrm{s}}$ sites
in the trivial
$
\mathbb{Z}^{T}_{2}
$
insulating phase can be described by the Hamiltonian
\begin{equation}
\label{eq: Trivial Hamiltonian for the Z2T state in one dimension}
\begin{split}
H_{\mathbb{Z}^{T}_{2}}
&\,=
\omega_{0}\,
\sum^{N_{\mathrm{s}}}_{j=1}\,
\left(
1 - X_{j}
\right)
\,,
\end{split}•
\end{equation}•
which has the time-reversal symmetric ground state
\begin{equation}
\label{eq: Trivial Z2T ground state in one dimension}
\begin{split}
\ket{\Phi_{\mathbb{Z}^{T}_{2}}}
=
\frac
{
1
}
{
2^{N_{\mathrm{s}}/2}
}
\,
\sum_{ s }
\,
\ket
{s}
\,,
\end{split}•
\end{equation}•
where
$\{\,\ket{s}\,\}$ represents the $2^{N_{\mathrm{s}}}$ many-body spin states in the eigenbasis of $Z$ operators:
$
Z_{j}\,\ket{s_{j}}
=
s_{j}\,\ket{s_{j}}
$,
for
$
s_{j}
=
\pm
1
$.
We recognize
Eq.~(\ref{eq: Trivial Z2T ground state in one dimension})
as the
ground state Eq.~(\ref{eq: RK ground state}) in the $\beta = 0$ limit
and
Eq.~(\ref{eq: Trivial Hamiltonian for the Z2T state in one dimension})
as an RK Hamiltonian of the form Eq.~(\ref{eq: Definition of the RK Hamiltonian})
with $\beta = 0$
and
$
\omega_{s,s'}
=
\omega_{0}
>
0
$,
where the sum in Eq.~(\ref{eq: RK Hamiltonian}) extends
over pairs of states $(s,s')$ which differ by a single spin flip.
We now obtain the unitary mapping $\mathbb{W}_{\mathbb{Z}^{T}_{2}}$, which,
applied on the trivial ground state
Eq.~(\ref{eq: Trivial Z2T ground state in one dimension}),
yields the non-trivial SPT state
\begin{equation}
\label{eq: Non-trivial Z2T ground state in one dimension}
\begin{split}
\ket{\Psi_{\mathbb{Z}^{T}_{2}}}
\equiv
\mathbb{W}_{\mathbb{Z}^{T}_{2}}\,
\ket{\Phi_{\mathbb{Z}^{T}_{2}}}
\equiv
\frac
{
1
}
{
2^{N_{\mathrm{s}}/2}
}
\,
\sum_{ s }
\,
e^
{
i\,W_{\mathbb{Z}^{T}_{2}}(s)
}
\,
\ket
{s}
\,,
\end{split}•
\end{equation}•
wherein the action of $\mathbb{W}_{\mathbb{Z}^{T}_{2}}$
on the many-body basis
$\{ \ket{s} \}$ is defined as
\begin{equation}
\label{eq: Action of W for the Z2T state on the basis elements}
\begin{split}
\mathbb{W}_{\mathbb{Z}^{T}_{2}}\,\ket{s}
=
e^
{
i\,W_{\mathbb{Z}^{T}_{2}}(s)
}
\,
\ket
{s}
\,.
\end{split}•
\end{equation}•
Invariance of the SPT ground state
$
\ket{\Psi_{\mathbb{Z}^{T}_{2}}}
$
under time-reversal symmetry
Eq.~(\ref{eq: Z2T symmetry in one dimension}),
according to
Eq.~(\ref{eq: Transformation of E and W under G anti-unitary}),
implies
\begin{equation}
\label{eq: Z2T condition on W}
\begin{split}
W_{\mathbb{Z}^{T}_{2}}(s)
=
-
W_{\mathbb{Z}^{T}_{2}}(-s)
+
2 \pi \times \textrm{integer}
\,.
\end{split}•
\end{equation}•
In light of the discussion in Sec.~(\ref{sec: SPT states in one dimension}),
if we assume for $\mathbb{W}_{\mathbb{Z}^{T}_{2}}$ an ansatz that leads to the entanglement of nearest neighbor spins, than
the only transformation (up to a trivial global phase) consistent with
Eq.~(\ref{eq: Z2T condition on W}) has
\begin{equation}
\label{eq: W_{G}(s) for the Z2T state in one dimension}
W_{\mathbb{Z}^{T}_{2}}(s)
=
\frac{\pi}{4}\,
\sum_{j}
\,
\left(
1 - s_{j}\,s_{j+1}
\right)
\,
\end{equation}•
due to the fact that, with periodic boundary conditions,
$
\sum_{j}\,
(1 - s_{j}\,s_{j+1})/2
$
is an even number.
Moreover, comparing
Eqs.~(\ref{eq: Action of W for the Z2T state on the basis elements})
and~(\ref{eq: W_{G}(s) for the Z2T state in one dimension})
with the results obtained in Sec~(\ref{sec: Z2 times Z2 in one dimension}), we conclude that, for any partition $\Sigma_{i}$ between sites $i$ and $i+1$, there are $2$ degenerate (equal to $1/2$) Schmidt eigenvalues and that
the SPT parent Hamiltonian for the ground state
Eq.~(\ref{eq: Non-trivial Z2T ground state in one dimension})
reads
\begin{equation}
\label{eq: Non-trivial Z2T Hamiltonian in one dimension}
\begin{split}
\mathcal{H}_{\mathbb{Z}^{T}_{2}}
\equiv
\mathbb{W}_{\mathbb{Z}^{T}_{2}}
\,
H_{\mathbb{Z}^{T}_{2}}
\,
\mathbb{W}_{\mathbb{Z}^{T}_{2}}^{-1}
&\,=
\omega_{0}\,
\sum^{N_{\mathrm{s}}}_{j=1}
\,
\left(
1
+
Z_{j-1}
\,
X_{j}
\,
Z_{j+1}
\right)
\,.
\end{split}•
\end{equation}•
Since this is essentially the same model as the
$
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
$
spin chain discussed in Sec~(\ref{sec: Z2 times Z2 in one dimension}),
we conclude that the
$
\mathbb{Z}_{2}^{T}
$
symmetric SPT Hamiltonian has $2$-fold degeneracy of the ground state per edge.
With the unitary transformation
Eqs.~(\ref{eq: Action of W for the Z2T state on the basis elements})
and~(\ref{eq: W_{G}(s) for the Z2T state in one dimension})
we can obtain the most general form of the SPT ground state
Eq.~(\ref{eq: SPT ground state})
by allowing $\beta \neq 0$ without changing either the gapped nature of the many-body energy spectrum or the topological properties of the ground state.
In this regard, the Hamiltonian
Eq.~(\ref{eq: Non-trivial Z2T Hamiltonian in one dimension})
provides a particular example of a larger class of models described in
Eq.~(\ref{eq: Definition of the SPT Hamiltonian}), with the phase factors
$
e^
{
i\,
W_{\mathbb{Z}_{2}^{T}}(s)
}
$
given in
Eq.~(\ref{eq: W_{G}(s) for the Z2T state in one dimension}).
\section{SPT states in two dimensions}
\label{SPT states in two dimensions}
\subsection{$\mathbb{Z}_{2}$ SPT states in $d=2$}
\label{subsec:Z2 SPT states in d=2}
The classification of bosonic SPT states in two dimensions, protected by
$\mathbb{Z}_{n}$ symmetry, asserts the existence of $n$ gapped phases of matter.~\cite{Chen-2012-a,Chen-2013,Lu-2012}
From the point of view of the one dimensional edge, these $n$ phases can be distinguished by the way right- and left-moving degrees of freedom transform under the symmetry: while in the trivial phase right- and left-moving modes carry the same $\mathbb{Z}_{n}$ charges, their $\mathbb{Z}_{n}$ charges are different for each of the other $n-1$ non-trivial SPT states.~\cite{Lu-2012,Chen-2012-b,Santos-2014}
As a consequence, in each of these $n-1$ non-trivial SPT states
the edge states cannot be gapped and symmetry preserving at the same time.
Levin and Gu have constructed a model on a triangular lattice
that describes a $\mathbb{Z}_{2}$ SPT paramagnet in two dimensions.~\cite{Levin-2012}
The $\mathbb{Z}_{2}$ symmetry is generated by
\begin{equation}
\label{eq: Z2 symmetry in 2 dimensions}
\widehat{S}_{\mathbb{Z}_{2}} = \prod_{j}\,X_{j}
\,,
\end{equation}•
where
$
\{ X, Y, Z\}
$
denote the three Pauli matrices
and
$
Z_{j}\,\ket{s_{j}}
=
s_{j}\,\ket{s_{j}}
$,
for
$s_{j}=\pm 1$.
In Ref.~\onlinecite{Levin-2012} the two kinds of paramagnetic ground states are represented by
\begin{equation}
\label{eq: Trivial Z2 paramagnet}
\begin{split}
\ket{\Phi_{\mathbb{Z}_{2}}}
=
\frac{1}{2^{N_{\textrm{s}}/2}}\,
\sum_{ s }\,
\ket{s}
\,,
\end{split}•
\end{equation}•
for the trivial paramagnet, and
\begin{equation}
\label{eq: Non-trivial Z2 paramagnet}
\begin{split}
\ket{\Psi_{\mathbb{Z}_{2}}}
=
\frac{1}{2^{N_{\textrm{s}}/2}}\,
\sum_{ s }\,
e^
{
i\,\pi\,L(s)
}\,
\ket{s}
\,,
\end{split}•
\end{equation}•
for the non-trivial paramagnet.
$N_{\textrm{s}}$ denotes the number of sites of the triangular lattice and $L(s) \in \mathbb{N}$ counts the number of loops that separate
domain-wall regions
for each one of the
$
2^{N_{\textrm{s}}}
$
many-body spin configurations $\ket{s}$
(see Figs.~\ref{fig:1},~\ref{fig:2}, and~\ref{fig:3}).
Thus while the coefficients in the expansion Eq.~(\ref{eq: Trivial Z2 paramagnet}) all have the same sign, in the non-trivial SPT ground state
Eq.~(\ref{eq: Non-trivial Z2 paramagnet}), the sign of the coefficients depends on the number of loops $L(s)$. This non-trivial (real) phase factor was shown in
Ref.~\onlinecite{Levin-2012} to
be responsible for the universal properties of the $\mathbb{Z}_{2}$ non-trivial paramagnetic state.
From
Eqs.~(\ref{eq: Trivial Z2 paramagnet})
and~(\ref{eq: Non-trivial Z2 paramagnet}),
we identify the ground states
of Ref.~\onlinecite{Levin-2012}
with Eqs.~(\ref{eq: RK ground state})
and~(\ref{eq: SPT ground state})
in the limit $\beta=0$,
and the unitary mapping
$\mathbb{W}_{\mathbb{Z}_{2}}$ between the two classes of paramagnets,
defined by its action on the basis $\{ \ket{s} \}$, to be
\begin{equation}
\label{eq: Phase factor in the Z2 SPT state in 2 dimensions}
\mathbb{W}_{\mathbb{Z}_{2}}\,\ket{s}
=
e^
{
i\,W_{\mathbb{Z}_{2}}(s)
}
\,\ket{s}
=
e^
{
i\,\pi\,L(s)
}
\,\ket{s}
\,.
\end{equation}•
Given that the unitary transformation defined in
Eq.~(\ref{eq: Phase factor in the Z2 SPT state in 2 dimensions})
encodes the universal properties of the two dimensional
$\mathbb{Z}_{2}$ SPT paramagnet, we can
move away from the ``infinite temperature" limit $\beta = 0$
studied in Ref.~(\onlinecite{Levin-2012})
by allowing a larger class of
$
\mathbb{Z}_{2}
$
SPT paramagnets with ground state given
by
Eq.~(\ref{eq: SPT ground state})
and
microscopic Hamiltonians
given by
Eq.~(\ref{eq: Definition of the SPT Hamiltonian}).
\subsection{$\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ SPT states in $d=2$}
\label{subsec: Z2 times Z2 SPT states in d=2}
\begin{figure}[h!]
\includegraphics[width=0.3\textwidth]{triangular-lattice-1.pdf}
\caption
{
(Color online) A particular many-body configuration $\ket{s^{(1)}}$ of spins
$
Z^{(1)}_{j}
$
(red),
where the loops are defined along the domain walls and have support on the
dual lattice.
As in Ref.~\onlinecite{Levin-2012}, in order to determine the
projection of the $\mathbb{Z}_{2}$ symmetry on the edge,
we take the outer spins in a reference state where they are all pointing
up.
}
\label{fig:1}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.3\textwidth]{triangular-lattice-2.pdf}
\caption
{
(Color online)
A particular many-body configuration $\ket{s^{(2)}}$ of spins
$
Z^{(2)}_{j}
$
(blue).
}
\label{fig:2}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.3\textwidth]{triangular-lattice-3.pdf}
\caption
{
(Color online)
A particular many-body configuration $\ket{s^{(12)}}$ of the spins
$
Z^{(12)}_{j}
=
Z^{(1)}_{j}
\,
Z^{(2)}_{j}
$
(green).
Note how the direction of the arrows at every site is consistent
with
Figs.~\ref{fig:1} and~\ref{fig:2}.
}
\label{fig:3}
\end{figure}
In Ref.~\onlinecite{J-Wang-2014}, the classification of two dimensional bosonic SPT states with
$\mathbb{Z}_{n} \times \mathbb{Z}_{m}$
symmetry was studied from the point of view of the effective one dimensional edge degrees of freedom.
The non-trivial SPT edge states were there shown
to be associated with non-trivial $3$-cocycles in the group cohomology.
However, no microscopic theory for the two-dimensional bulk states has been
investigated.
In this section, we construct a microscopic theory for the
$\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ bosonic SPT states using the formalism of RK Hamiltonians and show how the topological properties of the ground state give rise to
the edge physics studied in
Ref.~\onlinecite{J-Wang-2014}.
In addressing SPT states with $\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetry,
we allow, at each site $j$ of the lattice, two spin species represented by Pauli operators
$
\{\,X^{(1)}_{j}, Y^{(1)}_{j}, Z^{(1)}_{j}\,\}
$
and
$
\{\,X^{(2)}_{j}, Y^{(2)}_{j}, Z^{(2)}_{j}\,\}
$,
where the
$\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetry is generated by
\begin{equation}
\label{eq: Z2 times Z2 symmetry}
\begin{split}
\widehat{S}^{(1)}_{\mathbb{Z}_{2}}
=
\prod_{j}\,X^{(1)}_{j}
\,,
\quad
\widehat{S}^{(2)}_{\mathbb{Z}_{2}}
=
\prod_{j}\,X^{(2)}_{j}
\,.
\end{split}•
\end{equation}•
We immediately realize that there should exist at least four kinds of paramagnets, which are obtained by stacking two decoupled $\mathbb{Z}_{2}$ symmetric systems.
These four states of matter are parametrized by two integer numbers,
$p_{1}, p_{2} \in \{ 0, 1 \}$,
which contribute to the expansion of the ground state with the (real) phase factor
$
\exp
{
\left\{
i\,\pi\,
\left[
p_{1}\,L(s^{(1)})
+
p_{2}\,L(s^{(2)})
\right]
\right\}
}
$,
where
$L(s^{(1)})$ and $L(s^{(2)})$ count, respectively, the number of loops
defined by domain wall configurations formed by spins of species $1$ and $2$,
as in Figs.~\ref{fig:1} and~\ref{fig:2}.
In order to obtain $\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetric states that go beyond the tensor product of two $\mathbb{Z}_{2}$ symmetric states, we realize that, at each site $j$, we can consider an independent Ising spin
$
Z^{(12)}_{j}
\equiv
Z^{(1)}_{j}\,Z^{(2)}_{j}
$
with eigenvalues
$
s^{(12)}_{j}
=
s^{(1)}_{j}
\,
s^{(2)}_{j}
$.
We then propose
\begin{widetext}
\begin{equation}
\label{eq: Z2 times Z2 paramagnet ground state}
\begin{split}
\ket
{
\Psi^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
}
=
\frac{1}{2^{N_{\mathrm{s}}}}
\,
\sum_{ s^{(1)} }\,
\sum_{ s^{(2)} }\,
e^
{
i\,\pi\,
\left[
p_{1}\,L(s^{(1)})
+
p_{2}\,L(s^{(2)})
+
p_{12}\,L(s^{(12)})
\right]
}
\,
\ket
{
s^{(1)}; s^{(2)}
}
\,
\end{split}•
\end{equation}•
\end{widetext}
as a description of the $8$ classes of
$
\mathbb{Z}_{2} \times \mathbb{Z}_{2}
$
symmetric ground states parametrized by $3$ binary indices
$
p_{1}, p_{2}, p_{12} \in \{ 0, 1 \}
$.
We adopt the notation
$
\ket{s^{(1)}}
\otimes
\ket{s^{(2)}}
\equiv
\ket{s^{(1)};s^{(2)}}
$.
When
$
p_{1} = p_{2} = p_{12} = 0
$,
Eq.~(\ref{eq: Z2 times Z2 paramagnet ground state}) describes the trivial ground state
\begin{equation}
\label{eq: Trivial Z2 times Z2 paramagnet ground state}
\begin{split}
\ket
{
\Phi^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
}
=
\frac{1}{2^{N_{\mathrm{s}}}}
\,
\sum_{ s^{(1)} }\,
\sum_{ s^{(2)} }\,
\ket
{
s^{(1)}; s^{(2)}
}
\,,
\end{split}•
\end{equation}•
which is the tensor product of two trivial $\mathbb{Z}_{2}$ paramagnets.
Eq.~(\ref{eq: Z2 times Z2 paramagnet ground state})
corresponds to the choice of ground state in the
``high temperature" limit $\beta = 0$ and the SPT ground states can be obtained
from the trivial one,
Eq.~(\ref{eq: Trivial Z2 times Z2 paramagnet ground state}),
via the unitary transformation
$
\mathbb{W}^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
$,
whose action on the many-body basis
$
\{\,
\ket
{
s^{(1)}; s^{(2)}
}
\,\}
$
reads
\begin{widetext}
\begin{equation}
\label{eq: Phase factor in the Z2 times Z2 SPT state in 2 dimensions}
\mathbb{W}^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
\,
\ket
{
s^{(1)}; s^{(2)}
}
\equiv
e^
{
i\,
W^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
(s^{(1)};s^{(2)})
}
\,
\ket
{
s^{(1)}; s^{(2)}
}
=
e^
{
i\,\pi\,
\left[
p_{1}\,L(s^{(1)})
+
p_{2}\,L(s^{(2)})
+
p_{12}\,L(s^{(12)})
\right]
}
\,
\ket
{
s^{(1)}; s^{(2)}
}
\,.
\end{equation}•
\end{widetext}
In Eqs.~(\ref{eq: Z2 times Z2 paramagnet ground state})
and~(\ref{eq: Phase factor in the Z2 times Z2 SPT state in 2 dimensions}), in order to simplify the notation, we omit the indices $p_{1}$, $p_{2}$ and $p_{12}$ on which the SPT ground states and the unitary transformation depend upon.
\subsubsection
{
Parent Hamiltonians of the
$\mathbb{Z}_{2}\times\mathbb{Z}_{2}$
state
in
$
\mathrm{d}
=
2
$
}
\label
{
sec:
Parent Hamiltonians of the
Z2 times Z2 SPT state in d=2
}
The Hamiltonian
\begin{equation}
\label{eq: Trivial Hamiltonian of the Z2 times Z2 SPT state}
\begin{split}
H^{2 \mathrm{d}}_{\mathbb{Z}_{2}\times\mathbb{Z}_{2}}
=
\omega_{0}
\,
\sum_{j}\,
\left[
\left(
1
-
X^{(1)}_{j}
\right)
+
\left(
1
-
X^{(2)}_{j}
\right)
\right]
\end{split}•
\end{equation}•
describes a trivial paramagnetic system and has
Eq.~(\ref{eq: Trivial Z2 times Z2 paramagnet ground state}) as its ground state.
It is the sum of two RK Hamiltonians of the form
Eq.~(\ref{eq: Definition of the RK Hamiltonian})
with $\beta = 0$,
$
\omega_{s,s'}
=
\omega_{0}
>
0
$,
where the sum in
Eq.~(\ref{eq: RK Hamiltonian})
extends over pairs of states $(s,s')$ which differ by a single spin flip.
Then applying the unitary transformation
$
\mathbb{W}^
{
2 \mathrm{d}
}
_
{
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
}
$
in
Eq.~(\ref{eq: Phase factor in the Z2 times Z2 SPT state in 2 dimensions})
to the Hamiltonian
Eq.~(\ref{eq: Trivial Hamiltonian of the Z2 times Z2 SPT state})
yields the microscopic model realizing the $8$ classes of
$
\mathbb{Z}_{2}\times\mathbb{Z}_{2}
$
SPT paramagnets:
\begin{widetext}
\begin{equation}
\label{eq: Z2 times Z2 SPT Hamiltonian}
\begin{split}
\mathcal{H}^{2 \mathrm{d}}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
&\,=
\mathbb{W}^{2 \mathrm{d}}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
\,
H^{2 \mathrm{d}}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
\,
\left(
\mathbb{W}^{2 \mathrm{d}}_{\mathbb{Z}_{2} \times \mathbb{Z}_{2}}
\right)^{-1}
\\
&\,=
\omega_{0}\,
\sum_{j}\,
\Big\{
1
-
X^{(1)}_{j}\,
e^
{
i\,\pi\,
\left(
p_{1} +p_{12}
\right)
+
i\,\frac{\pi}{4}\,
\sum_{< \ell \, \ell' ; \, j >}\,
\left[
p_{1}\,
\left(
1 - Z^{(1)}_{\ell}\,Z^{(1)}_{\ell'}
\right)
+
p_{12}\,
\left(
1 - Z^{(12)}_{\ell}\,Z^{(12)}_{\ell'}
\right)
\right]
}
\Big\}
\\
&\,
+
\omega_{0}\,
\sum_{j}\,
\Big\{
1
-
\,X^{(2)}_{j}\,
e^
{
i\,\pi\,
\left(
p_{2} +p_{12}
\right)
+
i\,\frac{\pi}{4}\,
\sum_{< \ell \, \ell' ; \, j >}\,
\left[
p_{2}\,
\left(
1 - Z^{(2)}_{\ell}\,Z^{(2)}_{\ell'}
\right)
+
p_{12}\,
\left(
1 - Z^{(12)}_{\ell}\,Z^{(12)}_{\ell'}
\right)
\right]
}
\Big\}
\,,
\end{split}•
\end{equation}•
\end{widetext}
where the notation $< \ell \, \ell' ; \, j >$ denotes that the summation in the exponent runs over pairs of nearest neighbors $\ell$ and $\ell'$, which themselves belong to the $6$ nearest sites around a given site $j$ of the triangular lattice.
When $p_{12} = 0$ and $p_{1}=p_{2} = 1$,
Eq.~(\ref{eq: Z2 times Z2 SPT Hamiltonian}) describes, up to an overall energy shift, two decoupled Levin-Gu Hamiltonians.~\cite{Levin-2012}
On the other hand, the case $p_{12} = 1$
encodes strong entanglement between the two Ising systems.
\subsubsection
{
Edge symmetry transformations of the
$\mathbb{Z}_{2}\times\mathbb{Z}_{2}$
state
in
$
\mathrm{d}
=
2
$
}
The physical properties of the edge of two dimensional bosonic SPT states can be understood from the chiral action of the symmetry on the edge modes~\cite{Lu-2012,Chen-2012-b,Levin-2012,Santos-2014,J-Wang-2014,Else-2014}
which, on a microscopic scale, originates from a non-onsite symmetry
realization on the edge degrees of freedom.~\cite{Chen-2012-b,Levin-2012,Santos-2014,J-Wang-2014}
For the
$\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ SPT states
considered here, the non-onsite symmetry transformations on the edge
were shown in Ref.~\onlinecite{J-Wang-2014} to be a manifestation
of the non-trivial type-II cocycles of group cohomology.
The ground state Eq.~(\ref{eq: Trivial Z2 times Z2 paramagnet ground state}) is a simple product state and, as such, the edge states are disentangled from the bulk.
On the other hand, in the other SPT ground states whose expansion coefficients depend upon the domain wall configurations, the effect of flipping an edge spin can change the number of loops $L$, implying an entanglement between edge and bulk degrees of freedom. Therefore by studying the properties of the edge states, we can gain information about the nature of the bulk states and vice versa.
We have found, working directly with the bulk SPT wave functions
Eq.~(\ref{eq: Z2 times Z2 paramagnet ground state}), that the projection of the symmetry transformations Eq.~(\ref{eq: Z2 times Z2 symmetry}) onto the
boundary spins is
\begin{widetext}
\begin{subequations}
\label{eq: Z2 times Z2 symmetry transformation projected on the edge}
\begin{equation}
\begin{split}
\widehat{S}^{(1)}_{\mathbb{Z}_{2},\textrm{edge}}
=
&\,
\prod_{j}\,
X^{(1)}_{j}\,
e^
{
i\,\pi\,
\left(
p_{1} + p_{12}
\right)
+
i\,\frac{\pi}{4}\,
\left[
p_{1}
\left(
1
-
Z^{(1)}_{j}\,Z^{(1)}_{j+1}
\right)
+
p_{12}
\left(
1
-
Z^{(12)}_{j}\,Z^{(12)}_{j+1}
\right)
\right]
}
\,,
\end{split}•
\end{equation}•
\begin{equation}
\begin{split}
\widehat{S}^{(2)}_{\mathbb{Z}_{2},\textrm{edge}}
=
&\,
\prod_{j}\,
X^{(2)}_{j}\,
e^
{
i\,\pi\,
\left(
p_{2} + p_{12}
\right)
+
i\,\frac{\pi}{4}\,
\left[
p_{2}
\left(
1
-
Z^{(2)}_{j}\,Z^{(2)}_{j+1}
\right)
+
p_{12}
\left(
1
-
Z^{(12)}_{j}\,Z^{(12)}_{j+1}
\right)
\right]
}
\,,
\end{split}•
\end{equation}•
\end{subequations}•
\end{widetext}
where the product in
Eq.~(\ref{eq: Z2 times Z2 symmetry transformation projected on the edge})
is taken over the spins on the edge with periodic boundary conditions.
Remarkably these are the same non-onsite symmetry transformations studied in Ref.~\onlinecite{J-Wang-2014}, which shows that
Eq.~(\ref{eq: Z2 times Z2 paramagnet ground state})
gives the correct representation of the ground state
of the two dimensional bosonic SPT states protected by $\mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetry.
As is clear from
Eq.~(\ref{eq: Z2 times Z2 symmetry transformation projected on the edge}),
the non-onsite form acquired by the edge symmetry in the non-trivial SPT phase
implies that domain wall configurations on the edge carry projective representation
of the $\mathbb{Z}_{2}$ symmetry, i.e.,
upon flipping any given spin configuration twice,
there is a factor of
$
(-1)^{p}
$
for every domain wall.
Since the total number of domain walls is even,
it follows that
Eq.~(\ref{eq: Z2 times Z2 symmetry transformation projected on the edge})
is a faithful representation of the $\mathbb{Z}_{2}$ symmetry on the edge.
\section{Conclusions and future directions}
\label{sec:Conclusions and future directions}
We have demonstrated that Rokhsar-Kivelson models offer a useful framework
for constructing microscopic models whose ground states encode the universal properties of bosonic symmetry-protected topological states.
Although we have illustrated our construction for one- and two-dimensional
models and for certain discrete symmetries (unitary and anti-unitary), we believe our construction should hold true generically for any number of dimensions and for continuous symmetries as well.
We close with a few observations:
\begin{enumerate}
\item
It will certainly be interesting to realize parent Hamiltonians for bosonic SPT states in
two dimensions with discrete symmetry $\mathbb{Z}_{n}$, for $n \geq 3$.
In contrast to the $\mathbb{Z}_{2}$ case, the expansion of the
paramagnetic state in the ordered basis contains more than one type of
domain walls,
suggesting that the coefficients in the expansion of the SPT ground state
in the $\ket{s}$ basis may depend on orientable loops.
It remains an open question how these domain wall fluctuations can
coherently give rise to SPT ground states.
\item
We expect that this formalism may offer a path to constructing microscopic models for other classes of bosonic SPT states in three dimensions, as well as interesting examples of bosonic SPT states in two dimensions protected by
$\mathbb{Z}_{2} \times \mathbb{Z}_{2} \times \mathbb{Z}_{2}$ symmetry, which
are classified by non-trivial Type-III cocycles.~\cite{J-Wang-2014}
\item
A refinement of the group cohomology classification in terms of cobordisms
has been recently proposed to describe three dimensional bosonic SPT states
with time-reversal symmetry.~\cite{Kapustin-2014}
It could be worthwhile to study these time-reversal symmetric phases of matter
using the formalism of RK Hamiltonians, in particular, to understand
if there is any special meaning attached to the global unitary transformation
that connects the trivial and the non-trivial SPT ground states,
which could shed new light on the reason behind the
inability of the group cohomology classification
to account for these phases of matter.
\end{enumerate}
\section*{Acknowledgments}
We thank Eduardo Fradkin for constructive comments on the manuscript
and we acknowledge a useful discussion with Zheng-Cheng Gu
about Ref.~\onlinecite{Levin-2012}.
Research at the Perimeter Institute is supported by the Government
of Canada through Industry Canada and by the Province of Ontario
through the Ministry of Economic Development and Innovation.
|
\section{INTRODUCTION}
Two mechanisms, random mutation and natural selection, are considered in the mutation-selection model to explain an evolutionary process. The mutation-selection model is generally classified into two types, the coupled model and the decoupled model, depending on the mutation mechanism considered. The coupled model considers mutations that are only allowed as replication errors during reproduction events. The decoupled model considers mutations that are allowed at any time during the life cycle, such as those due to the effects of radiation, free radicals or thermal fluctuations\cite{1,2,3,4,5}.
Hence, mutation and selection are assumed to be independent processes that occur in parallel.
From a biological point of view, the relative contributions of the underlying mutation mechanisms for the two versions of the mutation-selection model are controversial. A mutation mechanism described by the coupled version implies a relatively constant mutation rate per generation, whereas the mutation mechanism described by the decoupled version implies a relatively constant mutation rate in time.
Crow and Kimura\cite{1} and Akin\cite{2} studied the decoupled mutation-selection model.
Baake {\it et al.} reported that the decoupled mutation-selection model was equivalent to an Ising quantum spin and calculated the observable quantities in three fitness landscapes using the methods from statistical mechanics\cite{3,4,5}.
Recently, Saakian {\it et al.}\cite{6} and Park and Deem\cite{7} examined the decoupled continuous-time mutation-selection model using a functional integral representation.
Several studies have considered the evolutionary process transiting to a neighboring higher fitness peak through the low-fitness valley.
Iwasa et al. calculated the rate of stochastic tunneling in a large population when the intermediate mutant had a lower fitness than the wild type, but the second mutant was beneficial\cite{8}. Weissman et al. provided a complete description of the rate at which an asexual population traversed a specific path through a genotype space to a particular fitter genotype by neglecting backward mutations and genetic diversity\cite{9}. Biological examples of the low-fitness valley can be found in pathogens that escape the immune response of their hosts, such as Escherichia coli\cite{10,11}, Salmonella typhimurium\cite{12} and the human immunodeficiency virus\cite{13}. Another example might be the types of cancer that occur only after a series of mutations\cite{14}.
Previous studies calculated the crossing time and the growing probability in the coupled discrete-time mutation-selection model by switching on an asymmetric sharply-peaked landscape, such that the reversal allele of the optimal allele had higher fitness than the optimal allele, from an initial steady state in a sharply-peaked landscape\cite{15,16,17,18}.
The growing probability was defined as the probability that additional offspring with the reversal allele grow to a measuring criterion.
The present study calculated the growing probability in the decoupled continuous-time mutation-selection model, which was evaluated as the ratio of the first arrival time to the final arrival time. The first arrival time was defined as the time when the first additional offspring arrived at the reversal allele. The final arrival time was defined as the arrival time of a successful offspring with the reversal allele that could grow to a measuring criterion.
The growing probability of additional offspring with the reversal allele in the decoupled continuous-time mutation-selection model was calculated systematically using computer simulations for various population sizes, sequence lengths, selective advantages, fitness parameters, and measuring parameters, and was compared with the theoretical formula for the growing probability in the Moran two-allele model\cite{19,20}.
This paper is organized as follows: Section II briefly describes the decoupled continuous-time mutation-selection models and the computer simulation method.
Section III presents the computer simulation results for the growing probability in the decoupled continuous-time mutation-selection models. Section IV summarizes the results.
\section{Model and Method} \label{sec-2}
This study considered the decoupled continuous-time mutation-selection model in a self-reproducing haploid population of $N$ individuals, whose size remains constant over the generations. The multiple alleles, $S_i$, might be represented by a string of $L$ spins, which use two values, $+1$ and $-1$.
The time dependence of the number, $n_i$, of individuals with the allele, $S_i$, in the decoupled continuous-time mutation-selection model can be described by the following equation:
\begin{equation} \label{eq-1}
\dot{n}_i (t) = ( w_i - \bar{w} )\: n_i(t)\: + \: \sum_j Q_{ij} \: n_j (t) \ ,
\end{equation}
where $w_i$ is the fitness of the allele $S_i$, and $\bar{w} (= {1 \over N} \sum_j w_j n_j(t) )$ is the mean fitness of the population.
In Eq. (\ref{eq-1}), the mutation matrix, $Q_{ij}$, which gives the mutation rate from $S_j$ to $S_i$, excludes double mutations\cite{4,5}:
\begin{equation} \label{eq-2}
Q_{ij}=\begin{cases}
-L \mu \ , &\mathrm{if}\ i=j\ , \cr
\mu \ , &\mathrm{if}\ d_{ij} = 1\ , \cr
0 \ , &\mathrm{otherwise}\ ,
\end{cases}
\end{equation}
where $\mu$ denotes the mutation rate per sequence element per unit time, and $d_{ij}$ denotes the Hamming distance between $S_i$ and $S_j$.
Consider a sharply-peaked landscape such that the optimal allele $S_0=(+1,+1, \cdots, +1)$ has a maximum fitness 1, and the remaining alleles have a lower fitness, $1-kL$.
The parameter $k$ is called the fitness parameter.
Equation (\ref{eq-1}) was solved using a computer simulation with the assumption that all individuals had the optimal allele initially. The population then approached a stochastic steady state, which is a quasispecies near the optimal allele in a sharply-peaked landscape.
After arriving at a stochastic steady state, the sharply-peaked landscape was changed suddenly to the following asymmetric sharply-peaked landscape:
\begin{equation} \label{eq-3}
w=\begin{cases}
1 \ , &\mathrm{if} \ S=S_0 \ , \cr
1+s \ , & \mathrm{if} \ S=S_0^\ast \ , \cr
1-kL \ , &\mathrm{otherwise}\ ,
\end{cases}
\end{equation}
where the maximum fitness was assigned to the reversal allele, $S_0^\ast = (-1,\cdots, -1)$, and the fitness of the remaining alleles was unchanged. All other parameters were fixed in this procedure, and the parameter, $s$, in Eq. (\ref{eq-3}) is called the selective advantage of the reversal allele over the optimal allele. The reversal allele had higher fitness than the optimal allele, assuming that the selective advantage, $s$, is a positive value. The asymmetric sharply-peaked landscape had two fitness peaks that were separated by a low-fitness valley with a tunable depth and width.
The numbers, $n_i$, were grouped into $L+1$ distinct classes according to their Hamming distances from the optimal allele, $S_0$, to improve the speed of the computer simulation\cite{21}.
The number, $N_l(=\sum_{\{i\}_{d=l}} n_i)$, of individuals in class $l$ was defined by the sum of all numbers $n_i$ with a Hamming distance, $d=l$, from the optimal allele.
The kinetic equation, Eq. (\ref{eq-1}), for the decoupled model can be expressed using the newly defined $N_l$ as follows:
\begin{equation} \label{eq-4}
\dot{N}_l (t)= (w_l - \bar{w})\: N_l(t)\: +\: \sum^L_{m=0} T_{lm}\: N_m (t) \ ,
\end{equation}
\begin{equation} \label{eq-5}
T_{lm}=\begin{cases}
-L \mu \ , &\mathrm{if}\ l=m\ , \cr
(L-m) \mu \ , &\mathrm{if}\ l=m+1\ , \cr
m \mu \ , &\mathrm{if}\ l=m-1\ , \cr
0 \ , &\mathrm{otherwise}\ ,
\end{cases}
\end{equation}
where $T_{lm}$ gives the mutation rate from class $m$ to class $l$.
The relative density, $X_l$, of class $l$ was then calculated by dividing the number, $N_l$, by the population size.
The computer simulations used the logistic branching process with the replacement of $\dot{N}_l (t)$ in Eq. (\ref{eq-4}) by $(N_l(t+dt)-N_l(t))/dt$\cite{19,22,23,24}.
The simulated population was evolved using the time steps of $dt=0.1$ generations in this study.
The mean fitness, $\bar{w}$, of the population was calculated at the beginning of each time step.
The branching process assumes that the population in the next time step is determined by the following two events: each individual in a class can mutate into any class with a certain probability, and each individual in a class can produce an offspring in the same class with a certain probability. More details of the simulation for a finite population can be found in Refs. 19 and 24.
The time was set to $t=0$ when the fitness landscape changed suddenly to an asymmetric sharply-peaked landscape. The increase in the relative density $\Delta X_d (t) (=X_d (t) - X_d (t=0) )$ was calculated. The computer simulation was completed when the increase in relative density with the reversal allele, $\Delta X_L (t)$, achieved the measuring criterion $\Delta X_L (t)=C$. The parameter, $C$, is called the measuring parameter.
The computer simulation recorded all arrival times, $t$, satisfying two conditions, $\Delta X_L(t-1) \le 0$ and $\Delta X_L(t) > 0$. The first arrival time, $t_a$, was defined as the earliest among the recorded arrival times, and the final arrival time was defined as the latest after the computer simulation is complete. Therefore, the final arrival time means the arrival time of a successful offspring with the reversal allele that can grow to a size, $NC$. The growing probability of additional offspring with the reversal allele was evaluated as the ratio of the first arrival time to the final arrival time, assuming that the additional offspring with the reversal allele are supplied at average time intervals of $t_a$ by stochastic fluctuations. This assumption holds well in the stochastic region, where the arrival of additional offspring with the reversal allele is a rare event with a quite long interval. Fortunately, the main consequence of this study was obtained from the simulation results for the growing probability in the stochastic region.
\section{Results} \label{sec-3}
The computer simulation results for the growing probability were obtained by averaging $M$ independent runs. The mean and the standard error were calculated from $M$ data values transformed into a logarithmic scale because the data showed a Gaussian distribution on a logarithmic scale.
The number of runs was set to $M=3000$ throughout this study.
The computer simulation results for an infinite population were calculated using the same method reported elsewhere\cite{24}.
Hereafter, $\Delta X_L^\infty (t)$ denotes the computer simulation result in an infinite population and $\Delta X_L(t)$ denotes the computer simulation result in a finite population.
Figure \ref{fig-1} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the measuring parameter, $\log(C)$, for $L=10$, $\mu =0.001$, $k=0.003$ and $s=10^{-4}$.
In Fig. \ref{fig-1}, the simulation results for the growing probability in the decoupled continuous-time mutation-selection model are represented by the solid lines with symbols, which correspond, right to left, to the population sizes $N=10^5, 10^6, 10^7, 10^8$ and $10^9$.
A previous study showed that the boundary between the deterministic and stochastic regions in the decoupled continuous-time mutation-selection model could be determined using the critical population size, $N_c$, satisfying $N_c\Delta X^\infty_L(t=0)=1$\cite{24}.
The computer simulation for an infinite population showed that $\Delta X^\infty_L(t=0)\approx 4.73\times 10^{-9}$ and $N_c\approx 2.12\times 10^8$ for $L=10$, $\mu=0.001$ and $k=0.003$.
A previous study reported that the growing probability in the stochastic region in the coupled discrete-time mutation-selection model was inversely proportional to the measuring parameter when $C<1/Ns$, bent when $C\approx 1/Ns$, and saturated when $C>1/Ns$\cite{18}.
Figure \ref{fig-1} shows that the simulation results for the growing probability for $N=10^5, 10^6$ and $10^7$, which belong to the stochastic region because $N\ll N_c$, are inversely proportional to the measuring parameter when $C<1/Ns^*$, bent when $C\approx 1/Ns^*$, and saturated when $C>1/Ns^*$, where $s^*$ is the effective selective advantage.
The simulation suggested that the effective selective advantage, $s^*$, is a hundred times larger than the selective advantage, or $s^* \approx 100 s \approx 10^{-2}$.
Let the selective advantage ratio, $r$, be the ratio of the effective selective advantage to the selective advantage($r=s^*/s$).
The simulation results suggested $r \approx 100$ for the parameter region in Fig. \ref{fig-1}.
The simulation indicated that the number of additional offspring with the reversal allele, $n_L$, in the stochastic region drifts neutrally until $n_L$ reaches the size, $NC$, when $n_L<1/s^*$, and that it grows deterministically when $n_L>1/s^*$.
Figure \ref{fig-1} also shows that the saturated growing probabilities, $P^*_g$, for $N=10^5, 10^6$ and $10^7$ become approximately the effective selective advantage, $s^* \approx 10^{-2}$, when $C\gg 1/Ns^*$ and $s^*\ll 1$.
Figure \ref{fig-2} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the measuring parameter, $\log(C)$, for $L=10$, $\mu =0.001$ and $k=0.003$.
In Fig. \ref{fig-2}, the simulation results for the growing probability in a finite population with a size, $N=10^7$, in the decoupled continuous-time mutation-selection model are represented by the symbols, which correspond, top to bottom, to the selective advantages, $s=10^{-3}$, $10^{-4}$, $10^{-5}$ and $10^{-6}$.
The parameters in Fig. \ref{fig-2} belong to the stochastic region, $N\ll N_c$, because the critical population size, $N_c$, is independent of the selective advantage. Figure \ref{fig-2} clearly shows that the growing probabilities for various selective advantages are inversely proportional to the measuring parameter when $C<1/Ns^*$, bend when $C\approx 1/Ns^*$ and become saturated when $C>1/Ns^*$, where $s^* \approx 100 s$.
Figure \ref{fig-2} also shows that the saturated growing probabilities, $P^*_g$, for various selective advantages become approximately the effective selective advantage when $C\gg 1/Ns^*$ and $s^*\ll 1$.
The following theoretical formula for the growing probability, $P^t_g$, represents the probability for a beneficial allele to grow to a size, $NC$, in the Moran two-allele model without mutations, when the selective advantage between two alleles is assumed to be equal to the effective selective advantage, $s^*$\cite{19,20}:
\begin{eqnarray}
P_g^t &=&\ {s^* \: (1+s^*)^{NC-1} \over (1+s^*)^{NC} - 1} \ , \label{eq-6} \\
&\approx &\ \begin{cases}
{s^* \over 1+s^*} \ , &\mathrm{if} \ s^*>1,\; s^*NC \gg 1 \ , \cr
\ s^* \ , &\mathrm{if}\ s^*\ll 1,\; s^*NC \gg 1\ , \cr
{1 \over NC} \ , &\mathrm{if}\ s^*NC \ll 1\ . \nonumber
\end{cases}
\end{eqnarray}
In Fig. \ref{fig-2}, the theoretical formula in Eq. (\ref{eq-6}) with $N=10^7$ and $s^* =100s$ is represented by the short-dotted lines, which correspond, top to bottom, to the selective advantages $s=10^{-3}$, $10^{-4}$, $10^{-5}$ and $10^{-6}$. A comparison of the symbols and the short-dotted lines shows that the simulation results for the growing probability, $P_g$, can be fitted well using the theoretical formula for the growing probability, $P^t_g$, in the Moran two-allele model.
A previous study showed that the growing probability in the stochastic region in the coupled discrete-time mutation-selection model could be described using the theoretical formula for the growing probability in the Wright-Fisher two-allele model\cite{18}.
The present study suggests that the growing probability in the stochastic region in the decoupled continuous-time mutation-selection model can be described using the theoretical formula for the growing probability in the Moran two-allele model.
Figure \ref{fig-3} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the selective advantage, $\log(s)$, for $L=10$, $\mu =0.001$ and $k=0.003$. In Fig. \ref{fig-3}, the symbols represent the simulation results for the growing probability in the decoupled continuous-time mutation-selection model in a finite population with a size, $N=10^7$, for a range of measuring parameters, $C=10^{-3}$, $10^{-4}$, $10^{-5}$ and $10^{-6}$.
The parameters in Fig. \ref{fig-3} belong to the stochastic region, $N\ll N_c$, because the critical population size, $N_c$, is independent of the measuring parameter.
Figure \ref{fig-3} shows that the growing probabilities approach $1/NC$ when $s^*\ll 1/NC$, where $s^* \approx 100 s$.
In Fig. \ref{fig-3}, the theoretical formula in Eq. (\ref{eq-6}) with $N=10^7$ and $s^*=100s$ is represented by the short-dotted lines, which correspond, bottom to top, to the measuring parameters $C=10^{-3}$, $10^{-4}$, $10^{-5}$ and $10^{-6}$. A comparison of the symbols and the short-dotted lines shows that the simulation results for the growing probability, $P_g$, can be fitted well using the theoretical formula for the growing probability, $P^t_g$, in the Moran two-allele model.
Figure \ref{fig-4} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the selective advantage, $\log(s)$, for $L=10$, $\mu =0.001$, $k=0.003$, and $C=10^{-4}$. In Fig. \ref{fig-4}, the symbols represent the simulation results for the growing probability in the decoupled continuous-time mutation-selection model for a range of population sizes, $N=10^5$, $10^6$ and $10^7$.
The parameters in Fig. \ref{fig-4} belong to the stochastic region, $N\ll N_c$, as mentioned in the explanation of Fig. \ref{fig-1}.
Figure \ref{fig-4} shows that the growing probabilities approach $1/NC$ when $s^*\ll 1/NC$, where $s^* \approx 100 s$.
In Fig. \ref{fig-4}, the theoretical formula in Eq. (\ref{eq-6}) with $C=10^{-4}$ and $s^*=100s$ is represented by the short-dotted lines, which correspond, top to bottom, to the population sizes, $N=10^5$, $10^6$ and $10^7$.
The simulation results in Figs. \ref{fig-2}, \ref{fig-3} and \ref{fig-4} show that the growing probability, $P_g$, in the stochastic region in the decoupled continuous-time mutation-selection model for various population sizes, selective advantages, and measuring parameters can be fitted well using the theoretical formula for the growing probability, $P^t_g$, in the Moran two-allele model.
Figure \ref{fig-5} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the fitness parameter, $\log(k)$, for $L=10$, $\mu =10^{-3}$, $C=10^{-2}$ and $s=10^{-4}$.
In Fig. \ref{fig-5}, the simulation results for the growing probability in the decoupled continuous-time mutation-selection model are represented by the solid lines with the symbols, which correspond, left to right, to the population sizes $N=10^6$, $10^8$, $10^{10}$ and $10^{12}$.
The growing probability, $P_g$, in Fig. \ref{fig-5} represents the saturated growing probability, $P^*_g$, because $C\gg 1/Ns^*$ for all parameters considered, where $s^* \approx 100 s \approx 10^{-2}$.
The arrows in Fig. \ref{fig-5} indicate the critical fitness parameters, $k_c$, left to right, for $N=10^6$, $10^8$, $10^{10}$ and $10^{12}$, which represent the fitness parameters satisfying $N\Delta X_L^\infty (t=0)=1$.
Figure \ref{fig-5} shows that the saturated growing probability in the deterministic region, $k\ll k_c$, approaches unity and that the saturated growing probability in the stochastic region, $k\gg k_c$, becomes approximately the effective selective advantage, regardless of the fitness parameter and the population size.
The simulation results in Figs. \ref{fig-2}, \ref{fig-3}, \ref{fig-4} and \ref{fig-5} indicate that the selective advantage ratio, $r$, maintains a similar value for various population sizes, selective advantages, measuring parameters, and fitness parameters, if the sequence length $L$ is the same.
Figure \ref{fig-6} shows the logarithm of the growing probability, $\log(P_g)$, as a function of the logarithm of the selective advantage, $\log(s)$, for $\mu =0.001$ and $k=0.002$. In Fig. \ref{fig-6}, the simulation results for the growing probability in the decoupled continuous-time mutation-selection model are represented by the symbols, that correspond, top to bottom, to the sequence lengths, $L=10$, $14$ and $18$.
The computer simulation for an infinite population showed that the critical population sizes $N_c$ satisfying $N_c\Delta X^\infty_L(t=0)=1$ were approximately $3.98\times 10^6$, $2.48\times 10^9$ and $1.61\times 10^{12}$ for $L=10$, $14$ and $18$, respectively.
Hence, the population sizes were set to $N=10^5$, $10^8$ and $10^{11}$ for $L=10$, $14$ and $18$, respectively, to belong to the stochastic region, $N\ll N_c$.
The measuring parameters were set to $C=10^{-2}$, $10^{-5}$ and $10^{-8}$ for $L=10$, $14$ and $18$, respectively, to maintain the same product of the population size and the measuring parameter, or $NC=10^3$, because the growing probability depends on $NC$, as shown in Eq. (\ref{eq-6}).
Figure \ref{fig-6} shows that the simulation results for the growing probability for $L=10$, $14$ and $18$ have a similar dependence on the selective advantage.
In Fig. \ref{fig-6}, the theoretical formula in Eq. (\ref{eq-6}) with $NC=10^3$ and $s^*=rs$ is represented by the short-dotted lines, which correspond, top to bottom, to the selective advantage ratios, $r=100$, 75 and 55.
A comparison of the symbols and the short-dotted lines showed that the simulation results for the growing probability, $P_g$, could be fitted well using the theoretical formula for the growing probability in the Moran two-allele model with the adjusted selective advantage ratio, which decreases with increasing sequence length.
The simulation showed that the selective advantage ratio does not depend on the population size, selective advantage, measuring parameter, and fitness parameter; instead, that the selective advantage ratio decreases with increasing sequence length.
\section{Summary} \label{sec-4}
The growing probabilities of additional offspring with the beneficial reversal allele for growing to a size, $NC$, for various population sizes, $N$, sequence lengths, $L$, selective advantages, $s$, fitness parameters, $k$, and measuring parameters, $C$, were calculated for a haploid, asexual population in the decoupled continuous-time mutation-selection model using the logistic branching process.
The growing probability was measured from the initial steady state in a sharply-peaked landscape, by switching on the asymmetric sharply-peaked landscape with a positive selective advantage of the reversal allele over the optimal allele. The growing probability was evaluated as the ratio of the first arrival time to the final arrival time, where the first and the final arrival times are defined as the arrival time of the first additional offspring and that of a successful offspring with the reversal allele that can grow to a size, $NC$, respectively.
Computer simulations showed that the growing probability in an asymmetric sharply-peaked landscape in the decoupled continuous-time mutation-selection model in the stochastic region satisfying $N\Delta X^\infty_L(t=0) \ll 1$ was inversely proportional to the measuring parameter when $C<1/Ns^*$, bent when $C\approx 1/Ns^*$ and became saturated when $C>1/Ns^*$, where $s^*$ is the effective selective advantage.
This suggests that the number of additional offspring with the reversal allele, $n_L$, in the stochastic region drifts neutrally until $n_L$ reaches the size, $NC$, when $n_L<1/s^*$, and that it grows deterministically when $n_L>1/s^*$.
Computer simulations also showed that the saturated growing probability in the stochastic region was approximately the effective selective advantage when $C\gg 1/Ns^*$ and $s^*\ll 1$.
Computer simulations showed that the growing probability in the stochastic region in an asymmetric sharply-peaked landscape in the decoupled continuous-time mutation-selection model for various population sizes, selective advantages and measuring parameters could be fitted well using the theoretical formula for the growing probability in the Moran two-allele model.
A previous study showed that the growing probability in the stochastic region in the coupled discrete-time mutation-selection model could be described using the theoretical formula for the growing probability in the Wright-Fisher two-allele model\cite{18}.
The present study suggests that the growing probability in the stochastic region in an asymmetric sharply-peaked landscape in the decoupled continuous-time mutation-selection model could be described using the theoretical formula for the growing probability in the Moran two-allele model.
Computer simulations showed that the selective advantage ratio, which represents the ratio of the effective selective advantage to the selective advantage, does not depend on the population size, selective advantage, measuring parameter, and fitness parameter; instead, that the selective advantage ratio decreases with increasing sequence length.
The growth behavior of additional offspring with the reversal allele in an asymmetric sharply-peaked landscape in the decoupled continuous-time mutation-selection model was controlled by the effective selective advantage of the reversal allele over the optimal allele despite there being many other alleles with lower fitness.
The present study might improve the understanding of the evolutionary process in the decoupled continuous-time mutation-selection model.
\begin{acknowledgements}
This study was supported by the Research Fund Program of Research Institute for Basic Sciences, Pusan National University, Korea, 2012, Project No. RIBS-PNU-2012-108.
\end{acknowledgements}
|
\section{Introduction}
\label{sect:intro}
Our recent search for the presence of a magnetic field in
the visually brightest early A-type supergiant HD\,92207, using FORS\,2 in spectropolarimetric mode,
revealed the presence of a weak mean longitudinal magnetic field of the order of a few
hundred Gauss (Hubrig et al.\ \cite{Hubrig2012}).
This target has been monitored for several years in
the $uvby$-Str\"omgren system by Sterken (\cite{Sterken1983}) and spectroscopically by
Kaufer et al.\ (\cite{Kaufer1996,Kaufer1997}),
who found cyclical changes of the brightness and substantial profile changes for metal lines and
H$\alpha$, and suggested that the observed photometric and H$\alpha$ line variations are the result
of a corotating structure in the wind, which they considered to be in the star's equatorial plane.
Furthermore, their study of the line profile variations revealed clear pulsation-like structures,
indicating the presence of non-radial pulsations (NRPs) with a period of 27\,days, while the stellar
rotation period is of the order of several months.
Ignace et al.\ (\cite{Ignace2009}) measured linear polarization
in the spectra of this star on seven different nights,
spanning approximately three months in time. For the continuum polarization, the
spiral-shaped wind density enhancement in the equatorial plane of the star suggested by
Kaufer et al.\ (\cite{Kaufer1996}) was explored.
Importantly, the authors reported that the polarization across
the H$\alpha$ line on any given night is typically different from the degree and position angle
of the polarization in the continuum. These night-to-night variations in the H$\alpha$ polarization
are hard to understand in terms of the spiral structure that was considered for the continuum polarization.
The determination of the atmospheric parameters was carried out by Przybilla et al. (\cite{Prz2006}), who also
established elemental abundances for over 20 chemical species.
In the subsequent spectropolarimetric study on this target, Hubrig et al.\ (\cite{Hubrig2014})
presented a careful inspection of the FORS\,2 spectra used for the
magnetic field determination in HD\,92207 and reported the detection of short-term spectral variability
on a time-scale of minutes or fractions of minutes. In the framework of that study, the authors
convincingly showed the absence of
instrument instabilities or flexures, presenting significantly stable spectral line profiles for a number of
other stars observed quasi-simultaneously and nearby in the sky to HD\,92207.
However, with the available data, it was not possible to decide
whether the discovered variations in line profiles of HD\,92207 were of periodic or stochastic nature.
In any case, given the size of the supergiant,
it was clear that the variability could not be referred to coherent line variations across the
entire surface on such short time-scales. Hubrig et al.\ (\cite{Hubrig2014}) estimated the length of the rotation period
as $P_{\rm rot}\le235\pm36$\,d.
The question how pulsations affect the magnetic field measurements is not yet solved in spite of
the fact that the number of studies of pulsating $\beta$~Cephei and slowly pulsating B (SPB) stars is
gradually increasing.
Already in \cite{Schnerr2006}, Schnerr et al.\ discussed the
influence of pulsations on the analysis of the magnetic field strength in the $\beta$~Cephei star $\nu$~Eri
in MUSICOS spectra and tried to model the signatures found in Stokes~$V$ and $N$ spectra.
Although the authors wrote that using some modeling they were able to quantitatively establish
the influence of pulsations on the magnetic field determination, they still detected profiles in
Stokes~$N$ and $V$ that were the result of the combined effects of the pulsations and the inaccuracies
in wavelength calibration that were not removed by their imperfect modeling of these effects.
Clearly, it is not possible to use the low-resolution FORS\,2 spectra to model the effect of
pulsations on the magnetic field measurements, and the potential of high-resolution spectropolarimetric
observations should be used in the search of short-term variations (e.g.\ Hubrig et al.\ \cite{Hubrig2011a}).
On the other hand, it is possible that the degree with which pulsations impact the magnetic field
measurements is time-dependent, e.g., if the line variability on short time scales were more or less pronounced
at certain rotation phases.
To investigate this possibility, we re-observed HD\,92207 with FORS\,2
on three different epochs in December 2013, January 2014, and February 2014 in the
framework of the ESO Programme 092.D-0209(A).
\begin{figure*}
\centering
\includegraphics[width=0.32\textwidth]{XV_0.clip_histo_night1.eps}
\includegraphics[width=0.32\textwidth]{XV_0.clip_histo_night2.eps}
\includegraphics[width=0.32\textwidth]{XV_0.clip_histo_night3.eps}
\caption{
Distributions from our Monte Carlo bootstrapping tests for the data sets of
HD\,92207 from 2013 December (left panel), 2014 January (middle panel), and 2014 February (right panel),
using the entire spectra.
The widths of these profiles are used to determine the errors listed in Table~\ref{tab:log_measfors}.
}
\label{fig:boot}
\end{figure*}
In this study, we discuss the new magnetic field measurements of HD\,92207. In addition, we present
the results of our analysis of HARPS archival spectropolarimetric daat and of our search for
short-term variability
using time series with the CORALIE spectrograph installed at the Swiss 1.2-m Euler telescope
on La Silla Observatory, Chile.
\section{Measurements of the magnetic field using FORS\,2 observations}
\subsection{Observations and data reduction}
\label{sect:fors_obs}
Three new spectropolarimetric observations of HD\,92207 were carried
out in service mode at the European
Southern Observatory with FORS\,2 mounted on the 8-m Antu telescope of the
VLT from 2013 December 4 to 2014 February 14.
This multi-mode instrument is equipped with
polarization analyzing optics, comprising super-achromatic half-wave and quarter-wave
phase retarder plates, and a Wollaston prism with a beam divergence of 22$\arcsec$ in
standard resolution mode.
Polarimetric spectra were obtained with the GRISM~600B and
the narrowest slit width of 0$\farcs$4 to achieve
a spectral resolving power of $R\sim2000$. The use of the mosaic detector
with a pixel size of 15\,$\mu$m allowed us to cover a
spectral range from 3250 to 6215\,\AA{}, which includes all hydrogen Balmer lines
from H$\beta$ to the Balmer jump.
From the raw FORS\,2 data, the spectra recorded in left- and right-hand polarized light, i.e.\
the ordinary and extraordinary beams
are extracted using a pipeline written in the MIDAS environment
by T.~Szeifert, the very first FORS instrument scientist.
This pipeline reduction by default includes background subtraction.
A single wavelength calibration frame is used for each night.
A first description of the assessment of the longitudinal magnetic field
measurements using FORS\,1/2 spectropolarimetric observations was presented
in our previous work (e.g.\ Hubrig et al.\ \cite{Hubrig2004a,Hubrig2004b},
and references therein).
To minimize the cross-talk effect, a sequence of subexposures at the retarder
position angles $-$45$^{\circ}$+45$^{\circ}$, +45$^{\circ}$$-$45$^{\circ}$,
$-$45$^{\circ}$+45$^{\circ}$, etc.\ is usually executed during observations and
the $V/I$ spectrum is calculated using:
\begin{equation}
\frac{V}{I} = \frac{1}{2} \left\{
\left( \frac{f^{\rm o} - f^{\rm e}}{f^{\rm o} + f^{\rm e}} \right)_{-45^{\circ}} -
\left( \frac{f^{\rm o} - f^{\rm e}}{f^{\rm o} + f^{\rm e}} \right)_{+45^{\circ}} \right\}
\label{eq:1}
\end{equation}
where +45$^{\circ}$ and $-$45$^{\circ}$ indicate the position angle of the
retarder waveplate and $f^{\rm o}$ and $f^{\rm e}$ are the ordinary and
extraordinary beams, respectively. Rectification of the $V/I$ spectra was
performed in the way described by Hubrig et al.\ (\cite{Hubrig2014}).
Null profiles, $N$, are calculated as pairwise differences from all available
$V$ profiles. From these, 3$\sigma$-outliers are identified and used to clip
the $V$ profiles. This removes spurious signals, which mostly come from cosmic
rays, and also reduces the noise. A full description of the updated data
reduction and analysis will be presented in a separate paper (Sch\"oller et
al., in preparation).
The mean longitudinal magnetic field, $\left< B_{\rm z}\right>$, is
measured on the rectified and clipped spectra based on the relation
\begin{eqnarray}
\frac{V}{I} = -\frac{g_{\rm eff}\, e \,\lambda^2}{4\pi\,m_{\rm e}\,c^2}\,
\frac{1}{I}\,\frac{{\rm d}I}{{\rm d}\lambda} \left<B_{\rm z}\right>\, ,
\label{eqn:vi}
\end{eqnarray}
\noindent
where $V$ is the Stokes parameter that measures the circular polarization, $I$
is the intensity in the unpolarized spectrum, $g_{\rm eff}$ is the effective
Land\'e factor, $e$ is the electron charge, $\lambda$ is the wavelength,
$m_{\rm e}$ is the electron mass, $c$ is the speed of light,
${{\rm d}I/{\rm d}\lambda}$ is the wavelength derivative of Stokes~$I$, and
$\left<B_{\rm z}\right>$ is the mean longitudinal (line-of-sight) magnetic field.
The magnetic field is usually measured in two ways: using only the absorption hydrogen Balmer
lines or using the entire spectrum including all available absorption lines in the spectra.
The Land\'e factor $g_{\rm eff}$ is assumed to be 1.0 in the spectral regions containing hydrogen lines and
it is slightly larger (1.25) in the spectral regions outside the hydrogen lines.
The feasibility of longitudinal magnetic field measurements in massive stars
using low-resolution spectropolarimetric observations was demonstrated by previous studies of O and B-type stars
(e.g., Hubrig et al.\ \cite{Hubrig2006,Hubrig2008,Hubrig2009,Hubrig2011a,Hubrig2011b,Hubrig2013a}).
The mean longitudinal magnetic field $\left<B_{\rm z}\right>$ is defined by the slope of the
weighted linear regression line through the measured data points, where
the weight of each data point is given by the squared signal-to-noise ratio
of the Stokes~$V$ spectrum. The formal $1\,\sigma$ error of
$\left<B_{\rm z}\right>$ is obtained from the standard relations for weighted
linear regression. This error is inversely proportional to the rms
signal-to-noise ratio of Stokes~$V$. Finally, the factor
$\sqrt{\chi^2_{\rm min}/\nu}$ is applied to the error determined from the
linear regression, if larger than 1. No significant fields were detected in the null spectra
calculated by combining the subexposures
in such a way that polarization cancels out. This allows us to verify that no spurious signals are present in the data.
\begin{table}
\caption[]{
Magnetic field measurements of HD\,92207 using FORS\,2.
All quoted errors are 1$\sigma$ uncertainties.
}
\label{tab:log_measfors}
\centering
\begin{tabular}{cr@{$\pm$}lr@{$\pm$}lr}
\hline \hline\\[-7pt]
\multicolumn{1}{c}{MJD} &
\multicolumn{2}{c}{$\left<B_{\rm z}\right>_{\rm all}$ [G]} &
\multicolumn{2}{c}{$\left<B_{\rm z}\right>_{\rm hyd}$ [G]} &
\multicolumn{1}{c}{S/N}\\
\hline\\[-7pt]
56630.308& $-$64 & 93 &$-$132 & 238 & 680\\%(3670 -6215)
56672.104 & 104 & 34 & 231 & 148 & 2056\\
56702.159 & 152 & 61 & 201 & 120 & 1675\\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.30\textwidth]{Fig02_newAA.eps}
\caption{
Variability of the Stokes~$I$ line profiles of \ion{Fe}{ii} $\lambda$4924, \ion{Fe}{ii} $\lambda$5169,
and \ion{He}{i} $\lambda$5876 (from top to bottom)
observed on three epochs in the FORS\,2 spectra.
}
\label{fig:si}
\end{figure}
Furthermore, we have carried out Monte Carlo bootstrapping tests.
These are most often applied with the purpose of deriving robust estimates of standard errors.
In these tests, we randomly generate 250\,000 samples from the original data
that have the same size as the original data set and analyse
the distribution of the $\left<B_{\rm z}\right>$ determined from all these newly generated data sets.
The measurement uncertainties obtained before and after Monte Carlo bootstrapping tests were found to be
in close agreement, indicating the absence of reduction flaws.
Distributions from our Monte Carlo bootstrapping tests for the data sets of
HD\,92207 from the three different epochs are presented in Fig.~\ref{fig:boot}.
The results of our magnetic field measurements along with the information related to the modified Julian date
and the signal-to-noise ratio in the observed spectra are listed in
Table~\ref{tab:log_measfors}.
A magnetic field at a significance level of 3$\sigma$ was achieved
during observations on 2014 January 15. We also find a change of polarity of the field between the observations
obtained in 2013 December and 2014 January.
\subsection{Variability}
\label{sect:fors_var}
As has already been reported in the work by Hubrig et al.\ (\cite{Hubrig2014}), the previous analysis of FORS\,2 polarimetric
spectra clearly showed the presence of short-term spectral
variability on time scales of the order of minutes and less.
Along with radial velocity
shifts up to 30\,km\,s$^{-1}$ detected for lines belonging to different elements, the authors also
detected notable changes in line intensities up to 3\%. Other stars observed during the same night,
also directly after the observations
of HD\,92207 at the same air mass and with similar short exposure times did not exhibit a similar kind
of variability, indicating that spectral variability of HD\,92207 was not introduced by
imperfection of the instrument. Hubrig et al.\ (\cite{Hubrig2014}) concluded that the detected spectral variability is
intrinsic, but with the current data, it can not be decided, if the variations are of
periodic or stochastic nature.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Fig3_corrV3.eps}
\caption{The behaviour of the \ion{Fe}{ii} $\lambda$4924, \ion{Fe}{ii} $\lambda$5169, and
\ion{He}{i} $\lambda$5876 lines (from top to bottom) in the ordinary and
extraordinary beams of each individual subexposure recorded in 2013 December.
In the respective upper row, the line profiles are shifted in vertical
direction for best visibility, while overplotted line profiles are presented in the respective lower row.
Note that one \ion{Fe}{ii} $\lambda$4924 profile in the EPMOS sub-exposure is distorted by a cosmic.}
\label{fig:fe1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Fig4_corr.eps}
\caption{ The behaviour of the \ion{Fe}{ii} $\lambda$4924, \ion{Fe}{ii} $\lambda$5169, and
\ion{He}{i} $\lambda$5876 lines (from top to bottom) in the ordinary and
extraordinary beams of each individual subexposure recorded in 2014 January.
In the respective upper row, the line profiles are shifted in vertical
direction for best visibility, while overplotted line profiles are presented in the respective lower row.
}
\label{fig:fe2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{Fig5_corr.eps}
\caption{The behaviour of the \ion{Fe}{ii} $\lambda$4924, \ion{Fe}{ii} $\lambda$5169, and
\ion{He}{i} $\lambda$5876 lines (from top to bottom) in the ordinary and
extraordinary beams of each individual subexposure recorded in 2014 February.
In the respective upper row, the line profiles are shifted in vertical
direction for best visibility, while overplotted line profiles are presented in the respective lower row.}
\label{fig:fe3}
\end{figure}
In our recent spectropolarimetric observations of HD\,92207 with FORS\,2, the Stokes~$I$ profile shape of spectral lines
changes from one epoch to the next. The differences in the line behaviour are illustrated
in Fig.~\ref{fig:si}, where we display as an example the behaviour of the
\ion{Fe}{ii} $\lambda$4924, \ion{Fe}{ii} $\lambda$5169, and \ion{He}{i} $\lambda$5876 lines with Land\'e factors
1.7, 1.3, and 1.0, respectively. An interesting fact
observed in the spectra obtained in 2014 January, i.e.\ at the epoch when the longitudinal magnetic field is detected at a
significance level of 3$\sigma$, is that the lines \ion{Fe}{ii} $\lambda$5169 and
\ion{He}{i} $\lambda$5876 appear slightly split.
To try to understand the character of the spectral variability of HD\,92207 in FORS\,2 spectra
on a short time scale and its impact on our
measurements of the longitudinal magnetic field, we decided to investigate the behaviour of the line profile
in left- and right-hand circular polarized light, i.e.\ in ordinary and extraordinary beams obtained at
the retarder waveplate positions +45$^{\circ}$ and $-$45$^{\circ}$, respectively. Such a study appears reasonable
since in the presence of
a non-zero longitudinal magnetic field, due to the Zeeman effect,
the line profiles display a radial velocity shift or/and
a different profile shape.
As we already mentioned above, for spectropolarimetric observations with FORS\,2, a sequence of subexposures at the retarder
position angles $-$45$^{\circ}$+45$^{\circ}$, +45$^{\circ}$$-$45$^{\circ}$,
$-$45$^{\circ}$+45$^{\circ}$, etc., is usually executed during observations. Ten subexposures with an exposure
time of 2\,sec were obtained in 2013 December, whereas twelve subexposures with an exposure
time of 4\,sec were taken in 2014 January, and fourteen subexposures with an exposure
time of 10\,sec were taken in 2014 February.
In Figs.~\ref{fig:fe1}--\ref{fig:fe3}, we present the behaviour of line profiles in the ordinary and
extraordinary beams at each epoch.
In the presented figures, the ordinary and extraordinary beams are labeled as OPMOS and EPMOS, respectively.
Please note that a spectrum from an ordinary beam at an angle of $-$45$^{\circ}$ corresponds to a spectrum
from an extraordinary beam at an angle of +45$^{\circ}$, and vice versa.
The differences seen on Figs.~\ref{fig:fe1}--\ref{fig:fe3} between the corresponding beams are due
to an imperfect wavelength calibration and are cancelling out in the $V/I$ spectrum, after the
application of Eq.~\ref{eq:1}.
No significant variation between the line profile shapes in the ordinary and extraordinary beams are observed
in FORS\,2 polarimetric spectra obtained in 2013 December.
At this epoch, no significant longitudinal magnetic field was detected.
In contrast, in 2014 January, the line profiles
in opposite circularly polarized light exhibit remarkable differences:
the line profiles of all three lines in the ordinary beams
appear slightly split. The line \ion{Fe}{ii} $\lambda$5169 exhibits split profiles also in the extraordinary beam,
but the shape of the line profile is different compared to that observed in the ordinary beams.
At that epoch, we measure a longitudinal magnetic field at a significance level of 3$\sigma$.
In 2014 February, the profiles of the \ion{Fe}{ii} $\lambda$4924 and \ion{Fe}{ii} $\lambda$5169 lines in
both ordinary and extraordinary beams appear similar, and the \ion{Fe}{ii} $\lambda$5169 line profile
shows a small distortion in the red wing.
However, the behaviour of the \ion{He}{i} $\lambda$5876 line is different, indicating the presence of a small splitting
in the profile shape observed in the ordinary beams.
At this epoch, a longitudinal magnetic field is measured at a
significance level of 2.5$\sigma$.
The detected behaviour of the line profile in opposite circularly polarized light has not been reported before. The inspection of the
previous FORS\,2 spectropolarimetric observations discussed in the work by Hubrig et al.\ (\cite{Hubrig2014}) revealed that also
in the observations obtained at MJD 55936.341 the lines
\ion{Fe}{ii} $\lambda$4924 and \ion{He}{i} $\lambda$5876 show a similar splitting in both ordinary and extraordinary beams, while
the line \ion{Fe}{ii} $\lambda$5169 does not show any splitting. No splitting in any line is detected on the dates
MJD55688.168 and MJD56018.223.
Since apart from the three lines discussed above, only very few other
\ion{Fe}{ii} lines show similar split line profile shapes, we can conclude that the detected splitting is not a result of an imperfect data
reduction. The interesting fact is that the lines \ion{Fe}{ii} $\lambda$4924 and \ion{Fe}{ii} $\lambda$5169 belong to
multiplet 42.
Lines of this multiplet present the principle emission lines observed in individual classes of emission objects, among them
Oe-, Be-, and P\,Cygni-stars.
However, the third member of this multiplet, the line \ion{Fe}{ii} $\lambda$5018,
was not found split in any FORS\,2 spectropolarimetric observation.
We have currently no explanation for the appearance
of the line splitting observed in the polarized line profiles.
We note that since previously only very few searches for the
presence of magnetic fields in a small number of early A-type supergiants were conducted, mostly using
least-squares deconvolution i.e.\ without any presentation of
individual line profiles in opposite circular polarization (e.g., Verdugo et al.\ \cite{Verdugo2005}), the necessary expertise for any
well-founded explanation of the observed behaviour is missing.
On the other hand, given the size and the shape of the splitting, it is obvious that the splitting is not related to
the Zeeman splitting frequently observed in slowly rotating Ap and Bp stars
with strong magnetic fields.
It is also likely that such a splitting can not be related to the appearance
of emission in the line cores since the observed H$\beta$ line in the FORS\,2 spectra does not appear in emission
on any of the three epochs. Noteworthy, we are not able to inspect the behaviour of the H$\alpha$ line in our spectra,
as it is located outside of the spectral range available with our FORS\,2 setting. Further, it is also not clear whether any
kind of pulsation would affect polarized profiles causing the appearance of the line splitting in polarized light.
\begin{figure*}
\centering
\includegraphics[width=0.46\textwidth]{FeII4924_LPVsh.eps}
\includegraphics[width=0.46\textwidth]{FeII5169_LPVsh.eps}
\caption{
FORS\,2 Stokes~$I$ profiles of the \ion{Fe}{ii}\,$\lambda$4924 (left panel) and
\ion{Fe}{ii}\,$\lambda$5169 (right panel) lines from individual subexposures of all three epochs.
For each epoch, in the upper rows, we present the line profiles shifted in vertical direction for best visibility.
The lower rows show all profiles overplotted.}
\label{fig:stokesi}
\end{figure*}
Hubrig et al.\ (\cite{Hubrig2014}) reported the presence of radial velocity
shifts of up to 30\,km\,s$^{-1}$ and changes in line intensities of up to 3\%.
Variations of the line profiles in Stokes~$I$ spectra with radial velocities reaching 10--20\,km\,s$^{-1}$ are observed
in the line profiles
recorded in 2013 December and 2014 February, while during the observations in 2014 January the line profiles
appear significantly stable with measured radial velocity shifts below 3\,km\,s$^{-1}$.
In all observed spectra is
the intensity variation below $\sim$2\% and mostly caused by a stronger noise in the line cores.
These results indicate that the short-term spectral variability most likely does not have any significant impact
on our magnetic field measurements using the spectra obtained in 2014 January.
As an example, we present in Fig.~\ref{fig:stokesi} overplotted Stokes~$I$ profiles of the \ion{Fe}{ii}\,$\lambda$4924 and
\ion{Fe}{ii}\,$\lambda$5169 lines in each individual subexposure on all three epochs. The shifts of individual line
profiles appear in general substantially smaller in 2014 January than those presented by
Hubrig et al.\ (\cite{Hubrig2014}) in their Figs.~2--5.
\section{Measurements of the magnetic field using HARPS observations}
\label{sect:harps}
\begin{table}
\caption[]{
Magnetic field measurements of HD\,92207 using HARPS polarimetric spectra.
All quoted errors are 1$\sigma$ uncertainties.
}
\label{tab:log_meas}
\centering
\begin{tabular}{ccr@{$\pm$}lr@{$\pm$}l}
\hline \hline\\[-7pt]
\multicolumn{1}{c}{HJD} &
\multicolumn{1}{c}{S/N} &
\multicolumn{2}{c}{$\left<B_{\rm z}\right>_{\rm Fe}$ [G]} &
\multicolumn{2}{c}{$\left<B_{\rm z}\right>_{\rm Fe,n}$ [G]} \\
\hline\\[-7pt]
2456345.677 &345 &$-$36 & 21 & $-$12 &25\\
2456347.676 &510 & 11 & 11 & 5 & 15 \\
2456348.677 &450 &2 & 13 & $-$3 & 16 \\
2456350.669 &485 &14 & 11 & 6 & 14 \\
\hline\\
\end{tabular}
\end{table}
Part of the spectra analyzed in this work were obtained with the HARPS polarimeter
(Snik et al.\ \cite{snik2008}) feeding the HARPS spectrometer
at the ESO 3.6-m telescope on La Silla.
We downloaded from the ESO archive the publically available polarimetric spectra of HD\,92207, obtained on four
different epochs in 2013 February.
All spectra have a resolving power of $R = 115\,000$ and the signal-to-noise ratio $(S/N)$ in the Stokes~$I$
profiles is between 345 and 510.
The HARPS archival spectra cover the wavelength range 3780--6913\,\AA{}, with a small gap around 5300\,\AA{}.
Each observation of the star is split into four sub-exposures,
obtained with four different orientations of the quarter-wave retarder plate relative to the
beam splitter of the polarimeter.
The reduction was performed using the HARPS data reduction
software available at the ESO headquarters in Garching. The normalization procedure for the HARPS spectra to the
continuum level is described in detail in the work of Hubrig et al.\ (\cite{Hubrig2013b}).
The Stokes~$I$ and $V$ parameters were derived following the ratio method described by
Donati et al.\ (\cite{Donati1997}),
ensuring in particular that all spurious signatures are removed up to first order.
The information on the individual observations, including the dates of observation and the $S/N$ values, is
given in Table~\ref{tab:log_meas}.
\begin{figure*}
\centering
\includegraphics[angle=0,width=0.58\textwidth]{Fig07_newAA.eps}
\caption{
The line profiles of different spectral lines, averaged over the four sub-exposures, recorded with HARPS in
left- and right-hand polarized light (indicated by the letters $L$ and $R$) on 2013 February 25.
The spectrum recorded in right-hand polarized light is shifted in vertical
direction for better visibility.
}
\label{fig:hd94}
\end{figure*}
We used for the measurements of the longitudinal magnetic field
the moment technique developed by Mathys (e.g.\ \cite{Mathys1991}).
The diagnostic potential of high-resolution circularly polarized spectra using the moment technique has been
discussed at length in numerous papers by Mathys (e.g.,\ \cite{Mathys1993}, \cite{Mathys1994}).
Wavelength shifts between right- and left-hand side circularly polarized spectra are
interpreted in terms of a longitudinal magnetic field $\left<B_{\rm z}\right>$.
Furthermore, this technique allows us not only the determination of the mean longitudinal magnetic field, but
also to investigate the presence of crossover effect, and to measure the quadratic magnetic field.
For each line in the selected sample of 50 \ion{Fe}{ii} lines, the measured shifts between the line profiles in the
left- and right-hand circularly polarized HARPS spectra
are used in a linear regression analysis in the
$\Delta\lambda$ versus $\lambda^2 g_{\rm eff}$ diagram, following the formalism discussed by
Mathys (\cite{Mathys1991,Mathys1994}).
The measured values for the mean longitudinal magnetic field $\left< B_{\rm z} \right>$ are presented in
Table~\ref{tab:log_meas}. They range from $-$36\,G obtained on the first
observing night to $+$14\,G obtained on the last observing night.
We do not detect a longitudinal magnetic field at a significance level of 3$\sigma$ for any night,
and, as we show below, on each night the profile shapes of spectral lines recorded in each subexposure in
the left- and right-hand polarized light appear rather identical, indicating the absence of a significant
longitudinal magnetic field.
As an example, we present in Fig.~\ref{fig:hd94} the line profiles recorded with HARPS in left- and
right-hand polarized light on 2013 February 25.
As a crosscheck, we measure the mean longitudinal magnetic field
on null polarization spectra, which are calculated by combining the subexposures
in such a way that polarization cancels out. The results of our measurements are presented in
Table~\ref{tab:log_meas}. The measurements on the spectral lines of \ion{Fe}{ii}
using null spectra are labeled with $n$.
Since no significant fields could be determined from null spectra, we conclude that
any noticeable spurious polarization is absent.
No significant crossover or mean quadratic magnetic field has been detected on the four observing epochs.
\begin{figure}
\centering
\includegraphics[angle=0,width=0.39\textwidth]{Fig08_newAA.eps}
\caption{
{\sl Upper panel:} The variability of the \ion{He}{i}\,$\lambda$5876 line in the HARPS spectra
of HD\,92207 observed from 2013 February 22 to 27.
{\sl Lower panel:} The variability of the \ion{Fe}{ii}\,$\lambda$5169 line over the same nights.
}
\label{fig:harps4}
\end{figure}
Clear day-to-day spectral variations are discovered in the spectra observed from 2013 February 22 to 27.
We detect that the radial velocity for metallic lines is increasing from night to night at a mean rate of
1.7\,km\,s$^{-1}$, while for the \ion{He}{i} lines it increases by 0.8\,km\,s$^{-1}$. The intensities of metallic lines vary opposite
to the intensities of \ion{He}{i} lines. We observe over the five nights
the increase of the metallic line intensity by about 3\% while the intensity for the \ion{He}{i} lines decreases by about 1\%.
As an example of the day-to-day spectral variations, we present in Fig.~\ref{fig:harps4} the behaviour of the
\ion{Fe}{ii}\,$\lambda$5169 and \ion{He}{i}\,$\lambda$5876 lines.
\begin{figure*}
\centering
\includegraphics[angle=0,width=0.75\textwidth]{HD92207_HeI5876_HARPS_nightlyLPV.eps}
\includegraphics[angle=0,width=0.75\textwidth]{HD92207_FeII5169_HARPS_nightlyLPV.eps}
\caption{
{\sl Upper panel:} The behaviour of the \ion{He}{i}\,$\lambda$5876 line in the HARPS spectra
of HD\,92207 over four nights from 2013 February 22 to 27 (from left to right).
Each panel presents the overplotted four Stokes~$I$
spectra obtained over each night.
{\sl Lower panel:} The behaviour of the \ion{Fe}{ii}\,$\lambda$5169 line over the same nights.
}
\label{fig:hd93}
\end{figure*}
Since HARPS spectropolarimetric observations are usually split into four sub-exposures, it is also possible to test
the line profile variability not only on the day-to-day time scale but also on shorter time scales corresponding
to the duration of sub-exposures, which in the case of HD\,92207 accounts for 8.3 to 10\,min.
Admittedly, these exposure times are much longer than those used for FORS\,2 observations. On the other hand, the
analysis of HARPS spectra allows us to explore the presence of the variability on 8--10 minute time scales.
In Fig.~\ref{fig:hd93},
we present the profiles of the \ion{Fe}{ii}\,$\lambda$5169 and \ion{He}{i}\,$\lambda$5876 lines recorded
on each individual night. No obvious changes in radial velocity or intensity are detected at these time scales.
In the sub-exposures obtained during the same night, the flux differences
are of the order of 0.5\% of the continuum level, which is consistent with the noise estimate.
\section{Spectroscopic time-series obtained with the CORALIE spectrograph}
\label{sect:CORALIE}
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{Fig10_newAA.eps}
\caption{
Low variability of the line profile of \ion{Fe}{ii}\,$\lambda$5197.6 in the CORALIE spectra observed on 2014 January 8.
}
\label{fig:var}
\end{figure}
To search for the presence of variability in HD\,92207 in high-resolution spectra on even shorter time scales of
the order of a few minutes,
we used the CORALIE spectrograph installed at
the Swiss 1.2-m Leonard Euler telescope on La Silla. CORALIE is a fibre-fed echelle spectrograph with $R\sim60\,000$
and with a wavelength coverage from 3800 to 6900\,\AA{}.
The instrument is described by Queloz et al. (\cite{que2001}), and instrumental upgrades made in 2007 are
detailed by S\'egransan et al. (\cite{seg2010}). The raw data were reduced using the efficient automated reduction
pipeline, which follows standard procedures and performs pre- and overscan bias correction, flatfielding using
Halogen lamps, background modelization, as well as cosmic removal. ThAr lamps are used for wavelength calibration.
Two time-series over several hours were obtained on the nights in 2014 January 8 and 12.
During the first night, we were able to record 40 spectra with an exposure time of 120\,s and a signal-to-noise ratio
of about 80--100.
Taking into account the readout overheads, the individual exposures are separated by 3.28\,min.
In Fig.~\ref{fig:var}, we present the behaviour of the line profile of the \ion{Fe}{ii}\,$\lambda$5197.6 line in
all 40 spectra.
Due to the presence of the rather strong noise, before plotting the line profiles, we have slightly smoothed
the profiles
using a Gauss filter with a width of 0.05\,\AA{}. Although the line profiles appear to be slightly variable,
the level of variability is rather low, much less than that observed by
Hubrig et al.\ (\cite{Hubrig2014}) in FORS\,2 spectra.
During the second night on January 12, an additional time series of 29 and 28 spectra with an exposure time
between 120 and 150\,sec and similar S/N, separated by a wavelength calibration exposure, were recorded.
The total observing time accounted for 3\,h 40\,min.
In these spectra, one can find similar line profile variations as in the night of January 8.
To check that the observed variability is not just due to the impact of the noise in the observed spectra,
we constructed
images of stacked spectra in several spectral regions.
In Fig.~\ref{fig:stacked_rms} in the upper panels, we present the stacked CORALIE spectra for the first and the
second series in different spectral regions around the lines \ion{Fe}{ii}\,$\lambda$4924, \ion{Fe}{ii}\,$\lambda$5169,
\ion{He}{i}\,$\lambda$5876, and \ion{Fe}{i}\,$\lambda$6678.
In the lower panels we show the residual spectra with respect to the mean spectrum.
All six panels are 13\,\AA{} long, and in all presented images, the $y$-dimension corresponds to the observing time.
The inspection of these images seems to indicate that some structure is present in the cores of
the \ion{He}{i}\,$\lambda$5876 and \ion{Fe}{i}\,$\lambda$6677 lines.
However, as we show in the same figure in the lower panels, if we subtract the mean
spectrum calculated using all observed spectra, this structure disappears, indicating
that the amplitude of the structure is of the same order as the spectrum noise.
The presented results can be explained in two different ways: either short-term spectral variability
on a time scale of 3-4\,min does not exist, or the amplitude of the line profile variability is of the same order as
the spectrum noise and cannot be detected in spectra obtained with a signal-to-noise ratio of about 80--100.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{fig11.eps}
\caption{Stacked CORALIE spectra of HD\,92207 with subtracted mean spectrum for the first and the second series in the
spectral regions around
a) \ion{Fe}{ii}\,$\lambda$4924,
b) \ion{Fe}{ii}\,$\lambda$5169,
c) \ion{He}{i}\,$\lambda$5876, and
d) \ion{Fe}{i}\,$\lambda$6678.
The upper panels present the observed spectra and the lower panels the residual spectra
with respect to the mean spectrum.
}
\label{fig:stacked_rms}
\end{figure}
\section{Discussion}
\label{sect:disc}
We obtained new magnetic field measurements of HD\,92207 on three different epochs in 2013 and 2014 using FORS\,2 in
spectropolarimetric mode.
Previous FORS\,2 observations revealed the presence of a weak mean longitudinal
magnetic field, but the definite
confirmation of the magnetic nature of this object was pending due to the detection of short-term spectral
variability probably
affecting the position of line profiles in left- and right-hand polarized spectra.
In our recent FORS\,2 observations obtained in 2014 January, the position of spectral lines appeared stable, with
measured radial velocity shifts below 3\,km\,s$^{-1}$.
Exactly at this epoch, using the entire spectrum
including all hydrogen lines for the measurement of the magnetic field, we achieved a 3$\sigma$ detection
with $\left<B_{\rm z}\right>_{\rm all}=104\pm34$\,G. This result is fully in agreement with the previous detection
of a weak longitudinal magnetic field in this star.
However, our analysis of the polarimetric FORS\,2 spectra revealed a strange phenomenon that is currently difficult to explain
and that has never been mentioned by other polarimetric studies of blue supergiants before: we discover that the line profiles in the light
polarized in a certain direction appear slightly split.
Given the size and the shape of the splitting, it is obvious that the splitting is not related to
the Zeeman splitting frequently observed in slowly rotating Ap and Bp stars
with strong magnetic fields.
We also cannot ascribe this splitting to the appearance of an emission in the line core, since
the H$\beta$ line, which is sensitive to the stellar wind, does not display any core emission or splitting in
polarized light.
Further, we are also not aware of any kind of pulsations that would affect polarized profiles in such a way.
Clearly, future monitoring of the behaviour of the line profiles in polarimetric spectra will be very valuable
to identify the mechanism causing the line splitting in circularly polarized light.
The available high-resolution polarimetric HARPS spectra on four different epochs show clear day-to-day variations in
all spectral lines, but short-term variability on a time scale of 8--10\,min is not detected.
We do not find any distinct Zeeman features
in the Stokes~$V$ spectra and the determined value of the longitudinal magnetic field is always below the 3$\sigma$ level.
No significant crossover or mean quadratic magnetic field is detected in the four HARPS observations
To search for short-term variability on even shorter timescales down to 3--4\,min, we obtained time series at two
different epochs with the CORALIE spectrograph.
The structure visible in a few spectral lines in the images presenting stacked spectra disappears if the mean spectrum is
subtracted. Due to the rather low signal-to-noise ratio of CORALIE spectra, we can not rule out that
spectral variability is present, but its amplitude is of the same order as
the spectrum noise.
Despite several decades of observational
efforts with ground-based photometry and spectroscopy of bright blue supergiants,
it is not yet certain what portion of
their variability is periodic, or how far they deviate from strict periodicity.
As a matter of fact, short-term variability was already identified
on a time scale of 1--3\,hours (Lefever et al.\ \cite{Lefever2007}; Kraus et al.\ \cite{Kraus2012}).
However, a variability on time scales of the order of minutes and less has not been investigated so far.
To improve the sensitivity to such variability, it would be important to use for the time series a high-resolution
spectrograph installed
at a telescope with a larger collecting area, such as the UV-Visual Echelle Spectrograph attached to the VLT 8\,m
Kuyen telescope.
|
\section{Introduction}
Quantum computers have exhibited tremendous power in algorithmic, cryptographic, information theoretic, and many other information processing tasks, compared with their classical counterparts. Meanwhile, for a large number of problems, quantum computers are not able to offer much advantage over classical ones. When and why quantum computers are more powerful are always at a central position in studies on quantum computation and quantum information processing. A particularly interesting class of scenarios is when there are, implicitly or explicitly, at least two parties involved who share a state, the correlation in this state is the key factor. What accounts for the quantum advantage is often \emph{entanglement}, one of the most distinctive characters of quantum information.
Indeed, it has been showed that a quantum algorithm with only slight entanglement can be simulated efficiently by a classical computer \cite{Vidal03}. In certain potential applications of quantum algorithms, it is also shown that entangled measurement is necessary for the existence of efficient quantum algorithms \cite{HMR+10}.
Recently people started to realize that entanglement is not always a necessary resource needed for generating quantum correlations. It has been found that \emph{discord}, another unique character of quantum states, also plays an important role in quantum information processing \cite{OZ02}. Discord is a relaxed version of entanglement---states with positive entanglement must also have positive discord, but there are states with positive discord but zero entanglement. People has discovered cases where quantum speed-up exists without entanglement involved, and discord is considered to be responsible for the quantum advantage \cite{DSC08}. Till today, discord is widely considered as necessary for the existence of quantum advantages.
In this paper, we reexamine this notion from the perspective of game
theory \cite{OR94}. Game theory studies the situation in which there
are two or more players with possibly different goals. There are two
broad classes of games, one is strategic-form (or normal-form)
games, in which all players make their choice simultaneously; a
typical example is Rock-Paper-Scissors. The other class is
extensive-form games, in which players make their moves in turn; a
typical example is chess.
The research on quantum games began about one decade ago, starting with two pioneering papers.\footnote{Note that there is also a class of ``nonlocal games'', such as CHSH or GHZ games \cite{BCMdW10}, where all the players have the \emph{same} objective. But general game theory focuses more on situation that the players have \emph{different} objective functions, and the players are selfish, each aiming to optimize her own objective function only.}
The first one \cite{EWL99} aimed to quantize a specific
strategic-form game called \emph{Prisoners' Dilemma} \cite{EWL99},
and it unleashed a long sequence of follow-up works in the same
model. Despite the rapid growth of literature, controversy also
largely exists \cite{BH01,vEP02,CT06}, which questioned the meaning
of the claimed quantum solution, the ad hoc assumptions in the
model, and the inconsistency with standard settings of classical
strategic games. Recently a new model was proposed for quantizing
general strategic-form games \cite{Zha12}. Compared with
\cite{EWL99}, the new model corresponds to the classical games more
precisely, and has rich mathematical structures and game-theoretic
questions; also see later theoretical developments
\cite{KZ12,WZ13,JSWZ13,PKL+15}.
Back to the early stage of the development of quantum game theory,
the other pioneering paper was \cite{Mey99}, which demonstrated the
power of using quantum strategies in an extensive-form game. More
specifically, Meyer considered the quantum version of the classical
Penny Matching game. The basic setting is as follows. There are two
players, and each has two possible actions on one bit: Flip it or
not. Starting with the bit being 0, Player 1 first takes an action,
and then Player 2 takes an action, and finally Player 1 takes
another action, and the game is finished. If the bit is finally 0,
then Player 1 wins; otherwise Player 2 wins. It is not hard to see
that if Player 2 flips the bit with half probability, then no matter
what Player 1 does, each player wins the game with half probability.
Now consider the following change of setting: The bit becomes a
qubit; the first player uses a quantum computer in the sense that
she can perform any quantum admissible operation on the bit; the
second player uses a classical computer in the sense that she can
perform either Identity or the flip operation $\begin{bmatrix} 0 & 1
\\ 1 & 0\end{bmatrix}$. In this new setting, Player 1 can win the
game with certainty! Her winning strategy is simple: she first
applies a Hadamard gate to change the state to $\ket{+} =
(\ket{0}+\ket{1})/\sqrt{2}$, and then no matter whether Player 2
applies the flip operation or not, the state remains the same
$\ket{+}$, thus in the third step Player 1 can simply apply a
Hadamard gate again to rotate the state back to $\ket{0}$. This
shows that a player using a quantum computer can have big advantage
over one using a classical computer.
Despite a very interesting phenomena it exhibits, the quantum
advantage is not the most convincing due to a fairness issue. After
all, the quantum player takes two actions and the classical player
takes just one. And the order of ``Player 1 $\rightarrow$ Player 2
$\rightarrow$ Player 1'' is also crucial for the quantum advantage.
One remedy is to consider normal-form games, in which the players
give their strategies \emph{simultaneously}, thus there is no longer
the issue of the action order. Taking the model in \cite{Zha12}, two
players play a complete-information normal-form game, with a
starting state $\rho$ in systems $(A_1,A_2)$, and $A_i$ being given
to Player $i$. A classical player can only measure her part of the
state in the computational basis, followed by whatever classical
operation $C$ (on the computational basis). In previous works
\cite{EWL99,Mey99,ZWC+12} the classical player is usually assumed to
be able to apply any classical operation on computational basis
(such as $X$-gate), followed by a measurement in the computational
basis. A classical operations there is implicitly assumed to be
unitary, so the operation in the matrix form is a permutation
matrix. Here we allow classical player to measure first and then
perform any classical operation, which gives her more power since
the second-step classical operation does need to be unitary. Indeed,
in Meyer's Penny Matching game, in the second step Player 2 could
measure the state first and then randomly set it to be $\ket{0}$ or
$\ket{1}$ each with half probability. Then in the third step, Player
1's Hadamard gate will change the state to $\ket{+}$ or $\ket{-}$,
in either case, Player 1 could win with only half probability.
Even if we now enlarge the space of possible operations of the
classical player, we will show examples where the quantum player has
advantage of winning the game. Furthermore, the examples have the
following nice properties respecting the fairness of the game:
\begin{enumerate}
\item If both players are classical, then both get expected payoff 0, and $\rho$ is a correlated equilibrium in the sense that any classical operation $C$ by one player cannot increase her expected payoff.
\item Suppose that one player remains classical and the other player uses a quantum computer. To illustrate the power of using quantum strategies, we cut the classical player some slack as follows. The classical player can (1) pick one subsystem, $A_1$ or $A_2$, of $\rho$, leaving the other subsystem to the quantum player, and (2) ``take side" by picking one of the two payoff matrices, leaving the other to the quantum player.
\end{enumerate}
Examples were found that even with the advantage of taking side and taking part of the shared state, the classical player still has a disadvantage compared to the quantum player. Consider the canonical $2\times 2$ zero-sum game with the payoff matrices being
\begin{equation}\label{eq:Penny game}
U_1=
\begin{pmatrix}
1 & -1\\
-1 & 1
\end{pmatrix}
\quad \text{ and } \quad
U_2=
\begin{pmatrix}
-1 & 1\\
1 & -1
\end{pmatrix}.
\end{equation}
\paragraph{Quantum game with entanglement}
Each player $i$ owns a 2-dimensional Hilbert space, and they share the quantum state
\begin{equation}
\ket{\psi} = \frac{1}{\sqrt{2}}(\ket{+0} + \ket{-1}) = \frac{1}{\sqrt{2}}(\ket{0+}+\ket{1-}),
\end{equation}
where $\ket{+}=\frac{1}{\sqrt{2}}(\ket{0}+\ket{1})$, and $\ket{-}=\frac{1}{\sqrt{2}}(\ket{0}-\ket{1})$.
It is not difficult to verify that if both players measure their parts in the computational basis, then each gets payoff 1 and $-1$ with equal probability, resulting an average payoff of zero for both players. This is a correlated equilibrium for classical operations.
Now suppose that Player 1 employs a quantum computer. Since the state is symmetric, it does not matter which part Player 2, the classical player, chooses. Let us assume that Player 2 chooses part 2, and the payoff matrix $U_2$. Then Player 1 can apply the Hadamard transformation on her qubit, followed by the measurement in computational basis. The state immediately before the measurement is
$\ket{\psi'} = (\ket{00} + \ket{11})/\sqrt{2}$. Therefore the measurement in the computational basis gives Player 1 and Player 2 payoff 1 and $-1$, respectively, with certainty. In other words, Player 1 wins with certainty, whereas she could only win with half probability when using a classical computer.
In this example where the quantum player has an advantage, the state shared by players is highly entangled, which motivates the following natural question: Is entanglement necessary for quantum advantage in the game? It turns out that the answer is no. Consider the example below.
\paragraph{Quantum game with discord} The payoff matrices are the same as before, but the quantum state shared by players is the following.
\begin{equation}
\rho =
\frac{1}{4}(\ket{+}\bra{+}\otimes\ket{0}\bra{0} + \ket{0}\bra{0}\otimes\ket{+}\bra{+} + \ket{-}\bra{-}\otimes \ket{1}\bra{1} + \ket{1}\bra{1}\otimes \ket{-}\bra{-}).
\end{equation}
This state is separable and thus does not have any entanglement. It can be checked that if the players measure this state in computational basis, the probability of getting each of the four possible outcomes is 1/4. Thus the overall payoff of each player is zero, and it can be verified that it is a classical correlated equilibrium.
In the quantum setting, again without loss of generality assume that the classical computer picks the second part of $\rho$ and the second payoff matrix. The quantum player can again perform a Hadamard operation on her system, resulting in a new state
\begin{equation}\label{eq:noEntan}
\rho' =
\frac{1}{4}(|0\rangle\langle0|\otimes|0\rangle\langle0|+|+\rangle\langle+|\otimes|+\rangle\langle+|+|1\rangle\langle1|\otimes|1\rangle\langle1|+|-\rangle\langle-|\otimes|-\rangle\langle-|).
\end{equation}
Measuring the new state, the quantum player gets state $\ket{00}$, $\ket{01}$, $\ket{10}$, $\ket{11}$ with probability
$3/8$, $1/8$, $1/8$, $3/8$ respectively. As a result, her winning probability
increases from 1/2 to 3/4; in other words, she gets an expected payoff of 1/2.
Note that the quantum state in Eq.\eqref{eq:noEntan} is separable,
and there is no any entanglement, but the quantum player still gets
a quantum advantage. Thus, entanglement is not necessary for quantum
advantage to exist in this game. Note that, however, the state in
Eq.\eqref{eq:noEntan} has a positive discord. As we have mentioned, it was known that in some scenarios, it is discord, rather than
entanglement, that produces non-classical correlations. So the
above example confirms this traditional notion in the new
game-theoretic setting.
These two examples were also experimentally verified recently \cite{ZWC+12}. The present paper makes further studies on the foregoing notion by asking the following fundamental question.
\begin{quote}
\emph{Is discord necessary for quantum advantage to exist in games where players share a symmetric state?}
\end{quote}
It is tempting to conjecture that the answer is Yes. In the rest of the paper, we will show that, first, discord is indeed necessary for any quantum advantage to exist in a 2-player games where each player has $n=2$ strategies. We will then show that when $n\geq 3$, however, there are games where the quantum player has a positive advantage even when the shared symmetric state has zero discord.
\section{Preliminaries}
Suppose that in a classical game there are $k$ players, labeled by
$\{1,2,\ldots,k\}$. Each player $i$ has a set $S_i$ of strategies.
To play the game, each player $i$ selects a strategy $s_i$ from
$S_i$. We use $s=(s_1,\ldots, s_k)$ to denote the \emph{joint
strategy} selected by the players and $S= S_1 \times \ldots \times
S_k$ to denote the set of all possible joint strategies. Each player
$i$ has a utility function $u_i: S \rightarrow \mbR$, specifying the
\emph{payoff} or \emph{utility} $u_i(s)$ of Player $i$ on the joint
strategy $s$. For simplicity of notation, we use subscript $-i$ to
denote the set $[k]-\{i\}$, so $s_{-i}$ is $(s_1, \ldots, s_{i-1},
s_{i+1}, \ldots, s_k)$, and similarly for $S_{-i}$, $p_{-i}$, etc.
In this paper, we will mainly consider 2-player games.
Nash equilibrium is a fundamental solution concept in game theory. Roughly, it says that in a joint strategy, no player can gain more by changing her strategy, provided that all other players keep their current strategies unchanged. The precise definition is as follows.
\begin{Def}
A \emph{pure Nash equilibrium} is a joint strategy $s = (s_1, \ldots ,s_k) \in S$ satisfying that
\begin{align}
u_i(s_i,s_{-i}) \geq u_i(s_i',s_{-i})
\end{align}
for all $i\in [k]$ and all $s'_i\in S_i$.
\end{Def}
Pure Nash equilibria can be generalized by allowing each player to independently select her strategy according to some probability distribution, leading to the following concept of \emph{mixed Nash equilibrium}.
\begin{Def}
A \emph{(mixed) Nash equilibrium (NE)} is a product probability distribution $p = p_1 \times \ldots \times p_k$, where each $p_i$ is a probability distributions over $S_i$, satisfying that
\begin{align}
\sum_{s_{-i}} p_{-i}(s_{-i}) u_i(s_i,s_{-i}) \geq \sum_{s_{-i}} p_{-i}(s_{-i}) u_i(s_i',s_{-i}),
\end{align}
for all $i\in [k]$, and all $s_i, s'_i\in S_i$ with $p_i(s_i)>0$.
\end{Def}
A fundamental fact proved by Nash \cite{Nas51} is that every game with a finite number of players and a finite set of strategies for each player has at least one mixed Nash equilibrium.
There are various further extensions of mixed Nash equilibria. Aumann \cite{Aum74} introduced a relaxation called \emph{correlated equilibrium}. This notion assumes an external party, called Referee, to draw a joint strategy $s = (s_1, ..., s_k)$ from some probability distribution $p$ over $S$, possibly correlated in an arbitrary way, and to suggest $s_i$ to Player $i$. Note that Player $i$ only sees $s_i$, thus the rest strategy $s_{-i}$ is a random variable over $S_{-i}$ distributed according to the conditional distribution $p|_{s_i}$, the distribution $p$ conditioned on the $i$-th part being $s_i$. Now $p$ is a correlated equilibrium if any Player $i$, upon receiving a suggested strategy $s_i$, has no incentive to change her strategy to a different $s_i' \in S_i$, assuming that all other players stick to their received suggestion $s_{-i}$.
\begin{Def} \label{thm:CE}
A \emph{correlated equilibrium (CE)} is a probability distribution $p$ over $S$ satisfying that
\begin{align}
\sum_{s_{-i}} p(s_i,s_{-i}) u_i(s_i,s_{-i}) \geq \sum_{s_{-i}} p(s_i,s_{-i}) u_i(s_i',s_{-i}),
\end{align}
for all $i\in [k]$, and all $s_i, s'_i\in S_i$.
\end{Def}
The above statement can also be restated as
\begin{equation}
\av_{s_{-i} \gets \mu|s_i}[u_i(s_i,s_{-i})] \ge \av_{s_{-i} \gets \mu|s_i}[u_i(s_i',s_{-i})].
\end{equation}
where $\mu|s_i$ is the distribution $\mu$ conditioned on the $i$-th component being $s_i$.
Notice that a classical correlated equilibrium $p$ is a classical Nash equilibrium if $p$ is a product distribution.
Correlated equilibria captures natural games such as the Traffic Light and the Battle of the Sexes (\cite{VNRT07}, Chapter 1). The set of CE also has good mathematical properties such as being convex (with Nash equilibria being some of the vertices of the polytope). Algorithmically, it is computationally benign for finding the best CE, measured by any linear function of payoffs, simply by solving a linear program (of polynomial size for games of constant players). A natural learning dynamics also leads to an approximate CE (\hspace{-.08em}\cite{VNRT07}, Chapter 4) which we will define next, and all CE in a graphical game with $n$ players and with $\log(n)$ degree can be found in polynomial time (\hspace{-.08em}\cite{VNRT07}, Chapter 7).
\section{Quantum game without discord}
In this section, we will address the question proposed at the end of the first section.
Suppose that a game has two players and both of them have $n$
strategies. In other words, each player holds an $n$-dimensional
quantum system. Recall that we also require the shared quantum state $\rho\in H\otimes H$ be symmetric, so that swapping the two systems does not
change the state. It is not hard to derive from the general criteria
of zero-discord state \cite{DVB10} that these quantum states $\rho$ have the form of
\begin{equation}\label{eq:def of rho}
\rho=\sum_{i,j=0}^{n-1}p(i,j)|\psi_i\rangle\langle\psi_i|\otimes|\psi_j\rangle\langle\psi_j|,
\end{equation}
where $\{|\psi_i\rangle\}$ is a set of orthogonal basis of the $n$-dimensional Hilbert space $H$, and $P=[p(i,j)]_{ij}\in {\mbR}_+^{n\times n}$ is a
symmetric matrix with nonnegative entries satisfying that $\sum_{ij}
p(i,j) = 1$. (In general, we use the upper case letter $P$ to denote
the matrix and the lower case letter $p$ to denote the corresponding
two-variate distribution $p(i,j)$.) We sometimes also write the
state as
\begin{equation}
\rho=\sum_{i}p_1(i)\ket{\psi_i}\bra{\psi_i}\otimes \sigma_i
\end{equation}
where $p_1(i) = \sum_j p(i,j)$ is the marginal distribution on the
first system, and $\sigma_i = \sum_j \frac{p(i,j)}{p_1(i)}
\ket{\psi_j}\bra{\psi_j}$ (if $p_1(i) = 0$ then let $\sigma_i =
\ket{0}\bra{0}$).
Consider the following game as a natural extension of the Penny Matching game in Section 1. The payoff matrices are
\begin{equation}\label{eq:dice}
U_1 = nI - J \quad and \quad U_2 = -U_1,
\end{equation}
where $J$ is the all-one matrix. Intuitively, whoever takes the first matrix bets that the two $n$-sided dice give the same side, and the other player bets that the two dice give different sides.
We first show that there is a unique correlated equilibrium in the game.
\begin{Lem}\label{lem:ce}
The game given by Eq.\eqref{eq:dice} has only one classical correlated equilibrium $Q = J/n^2$.
\end{Lem}
\begin{proof}
According to the definition of correlated equilibrium, if a distribution $q$ on $[n]\times [n]$ is a
classical correlated equilibrium, then the following relationships hold:
\begin{equation}\label{eq:forCE1}
\sum_{j}q(i,j)U_1(i,j)\geq \sum_{j}q(i,j)U_1(i',j), \qquad \forall i,i'\in\{0,1,...,n-1\},
\end{equation}
and
\begin{equation}\label{eq:forCE2}
\sum_{i}q(i,j)U_2(i,j)\geq \sum_{i}q(i,j)U_2(i,j'), \qquad \forall j,j'\in\{0,1,...,n-1\}.
\end{equation}
Plugging the definition of $U_1$ and $U_2$ into the above inequalities, one can verify that $Q=J/n^2$ is the only solution.
\end{proof}
Recall that $\rho=\sum_{i}p_1(i)\ket{\psi_i}\bra{\psi_i}\otimes \sigma_i$. Since $\rho$ is symmetric, it does not matter which part the classical player, Player 2, chooses to hold. For the convenience of discussions, let us assume that the classical player takes the second part. We use $\supp(p)$ to denote the support of a distribution $p$, \ie the set of elements with non-zero probability.
The next lemma gives a sufficient and necessary condition for the existence of quantum advantage.
\begin{Lem}\label{lem:advan}
Suppose that measuring the state $\rho$ gives a classical correlated equilibrium for the game given in Eq.\eqref{eq:dice}. Then Player 1 (who is quantum) does not have any
advantage if and only if
\begin{equation}\label{eq:lem2}
\langle i|\sigma_j|i\rangle=1/n, \ \ \ \forall i\in \{0,1,...,n-1\}
\text{ and } j\in \text{\supp}(p_1).
\end{equation}
\end{Lem}
\begin{proof}
\emph{``Only if"}: Assume that Player 1 first measures her part in
the orthonormal basis $\{|\psi_i\rangle\}$. Note that this does not affect the state. If outcome $j$
occurs, then Player 1 knows that the state of Player 2 is
$\sigma_j$. We consider which utility matrix in Eq.\eqref{eq:dice} Player 1 has. In the first case, Player 1 takes the utility matrix $U_1$. It is not hard to see that her optimal strategy is to replace
her part $\ket{\psi_j}$ by $\ket{i}$, where $i$ is a maximizer of
$\max_i \langle i|\sigma_j|i\rangle$. Thus Player 1 has a strict
positive advantage if and only if there is some $i$ and $j$, where
$j\in \supp(p_1)$, with $\bra{i}\sigma_j \ket{i} > 1/n$, which is
equivalent to saying that there is some $i$ and $j\in \supp(p_1)$
with $\bra{i}\sigma_j \ket{i} \neq 1/n$.
Similarly, if Player 1 takes the utility matrix $U_2$, then her
optimal strategy is to replace $\ket{\psi_j}$ with $\ket{i}$, where
$i$ is a minimizer of $\min_i \langle i|\sigma_j|i\rangle$. Thus
Player 1 has a strict positive advantage if and only if there is
some $i$ and $j$ with $\bra{i}\sigma_j \ket{i} < 1/n$, which is
again equivalent to saying that there is some $i$ and $j$ with
$\bra{i}\sigma_j \ket{i} \neq 1/n$.
\emph{``If"}: Player 2 measures her part in the computational
basis, yielding the state
\begin{equation*}
\frac{1}{n}\sum_{i,j} p_1(j) \ket{\psi_j}\bra{\psi_j}\otimes \ket{i}\bra{i}.
\end{equation*}
Now whatever quantum operation Player 1 applies, the probability of
observing the same bits (\ie the state after the measurement is
$\ket{ii}$ for some $i$) is $1/n$, with the expected payoff of 0
for both players.
\end{proof}
\medskip
Though the above lemma gives a sufficient and necessary condition, it is still not always clear whether quantum advantage could exist for any symmetric state $\rho$ with zero discord. Next we will further the study by considering a related matrix $M \in \mbR_+^{n\times n}$, whose $(i,j)$-th entry is defined to be
\begin{equation}\label{eq:M}
M (i,j) = |\qip{i}{\psi_j}|^2.
\end{equation}
It turns out that the rank of $M$ is an important criteria to our question. In the rest of this section, we will consider two cases, depending on whether $M$ is full rank or not.
\subsection{Case 1: $M$ is full-rank}
We will first show that if $M$ is full-rank, then the quantum player cannot have any advantage.
\begin{Thm}
Suppose that the two players of the game Eq.\eqref{eq:dice} share a symmetric state $\rho$, measuring which gives a classical correlated equilibrium. Then Player 1 (who is quantum) does not have any advantage if $M$ in Eq.\eqref{eq:M} is
full-rank.
\end{Thm}
\vspace{0.1in}
\begin{proof}
By Lemma \ref{lem:ce}, for any $0\leq k,j\leq n-1$ we have
\[
\sum_{i=0}^{n-1}p_1(i)|\langle k|\psi_i\rangle|^2\cdot\langle
j|\sigma_i|j\rangle=\frac{1}{n^2}
\]
Summing over $j$, we obtain another equality
\[
\sum_{i=0}^{n-1}p_1(i)|\langle k|\psi_i\rangle|^2=\frac{1}{n}.
\]
Combining these two equalities, we have
\[
\sum_{i=0}^{n-1}|\langle k|\psi_i\rangle|^2\cdot p_1(i)\left(\langle
j|\sigma_i|j\rangle-\frac{1}{n}\right)=0.
\]
Define a matrix $A=[a(i,j)]_{ij}\in {\mbR}^{n\times n}$ by
$a(i,j)=p_1(i)\left(\langle j|\sigma_i|j\rangle-\frac{1}{n}\right)$.
Then the above equality is just $\sum_i M(k,i)a(i,j) = 0$ for all
$k,j$. In other words, we have
\begin{equation}
M\cdot A=0.
\end{equation}
Since the matrix $M$ is assumed to be full-rank, we have $A = M^{-1}
0 = 0$. The conclusion thus follows by Lemma \ref{lem:advan}.
\end{proof}
Two corollaries are in order. First, note that $M$ is full-rank for a generic orthogonal basis $\{|\psi_i\rangle\}$, it is generically true that no discord implies no quantum advantage.
\begin{Cor}
If a set of orthonormal basis $\{\ket{\psi_i}\}$ is picked uniformly at random, then with probability 1, the quantum player does not have any advantage.
\end{Cor}
The second corollary considers the case of $n=2$, which is settled
by the above theorem completely. Indeed, when $n=2$, the rank of $M$
is either 1 or 2. The rank-2 case is handled by the above theorem.
If the rank is 1, it is not hard to see that the only possible $M$
is $M = \begin{bmatrix} 1/2 & 1/2 \\ 1/2 & 1/2 \end{bmatrix}$. In
this case, for any $i$ and any $k$ it holds that
\[
\bra{k}\sigma_i \ket{k} = \bra{k} \Big(\sum_j p(j|i) \ket{\psi_j} \bra{\psi_j}\Big) \ket{k} = \sum_j p(j|i) |\qip{k}{\psi_j}|^2 = \frac{1}{2} \sum_j p(j|i) = \frac{1}{2}.
\]
Applying Lemma \ref{lem:advan}, we thus get the following corollary
\begin{Cor}
There is no quantum advantage for the game defined in Eq.\eqref{eq:Penny game} on any symmetric state $\rho$ with zero discord.
\end{Cor}
\subsection{Case 2: $M$ is not full rank}
Somewhat surprisingly, the quantum player \emph{can} have an advantage when $M$ is not full-rank. In this section we exhibit a counterexample for $n=3$. In this case, recall that the payoff matrices are
\begin{equation}
U_1=
\begin{pmatrix}
2 & -1 & -1\\
-1 & 2 & -1\\
-1 & -1 & 2
\end{pmatrix}
\quad \text{ and } \quad
U_2=
\begin{pmatrix}
-2 & 1 & 1\\
1 & -2 & 1\\
1 & 1 & -2
\end{pmatrix}.
\end{equation}
We consider the following quantum state,
\begin{equation}
\rho=\sum_{i,j=0}^{2}p(i,j)|\psi_i\rangle\langle\psi_i|\otimes|\psi_j\rangle\langle\psi_j|,
\end{equation}
where
\begin{align}
|\psi_0\rangle = \frac{1}{\sqrt{2}}\left(|0\rangle+|1\rangle\right), \quad |\psi_1\rangle = \frac{1}{\sqrt{2}}\left(|0\rangle-|1\rangle\right), \quad |\psi_2\rangle = |2\rangle.
\end{align}
It is not hard to calculate $M$:
\begin{equation}\label{eq:forExa_1}
M=
\begin{pmatrix}
1/2 & 1/2 & 0\\
1/2 & 1/2 & 0\\
0 & 0 & 1
\end{pmatrix}.
\end{equation}
which has rank 2. Define
\begin{equation}\label{eq:forExa_2}
P=
\begin{pmatrix}
4/9 & 0 & 0\\
0 & 0 & 2/9\\
0 & 2/9 & 1/9
\end{pmatrix}.
\end{equation}
It can be easily verified that if the two players measure the state in computational basis, the probability distribution yielded is uniform, which is a classical Nash equilibrium.
Now suppose that Player 1 uses a quantum computer. One can verify that the condition in Lemma
\ref{lem:advan} does not hold. For a concrete illustration, let us consider the protocol in Lemma
\ref{lem:advan} again. Player 1 first measures in the
basis $\{\ket{+}, \ket{-}, \ket{2}\}$. With probability 4/9, she observes $\ket{+}$, then changes it to $\ket{0}$. Player 2's state is also $\ket{+}$ in this case, thus a measurement in the computational basis gives the $\ket{00}$ and $\ket{01}$ each with half probability. Thus Player 1's payoff in this case is $2\cdot \frac{1}{2} - 1\cdot \frac{1}{2} = \frac{1}{2}$. The second case is that Player 1 observes $\ket{-}$, which happens with
probability 2/9, and Player 2's state is $\ket{2}$ for sure. Player 1 changes her part
to $\ket{2}$, and gets payoff $2$. The third case is that Player 1 observes $\ket{2}$, which happens with probability 1/3, leaving Player 2 $\sigma_3 = (2/3) \ket{1}\bra{1} + (1/3) \ket{2}\bra{2}$. Player 1 then changes her qubit to $\ket{1}$, collides with Player 2's outcome with probability 1/3, thus Player 1's payoff is $2\cdot \frac{1}{3} - 1\cdot \frac{2}{3} = 0$. On average, the quantum player has a payoff of $(4/9)(1/2) + (2/9)\cdot 2 + (1/3) \cdot 0 = 2/3.$
It should be pointed out that the matrix $P$ achieving the quantum
advantage of 2/3 is not unique. For example, the following matrix
also works with the same effect:
\begin{equation}\label{eq:forExa_3}
P=
\begin{pmatrix}
2/9 & 2/9 & 0\\
0 & 0 & 2/9\\
1/9 & 1/9 & 1/9
\end{pmatrix}.
\end{equation}
\subsection{Optimization}
In this subsection, we show that the 3-dimensional example in the
above subsection is actually optimal for $M$ defined in Eq.\eqref{eq:forExa_1}. Actually the theorem below shows more. Note that
if the rank of $M$ is 1, it is easy to prove that $M$ must be the
uniform matrix, and the quantum advantage must be zero, thus in the
following we suppose the rank of $M$ to be 2.
\begin{Thm}\label{thm:Opt}
Suppose that measuring the state $\rho$ gives a classical correlated
equilibrium. Suppose the columns of $M$ are $M^0,M^1$ and $M^2$.
Without loss of generality, suppose $M^0=xM^1+(1-x)M^2$, where
$0\leq x\leq1$. Then the quantum advantage
\begin{equation}
QA\leq\frac{1}{3}+\frac{1}{3x_b},
\end{equation}
where $x_b=\max\{x,1-x\}$.
\end{Thm}
\vspace{0.1in}
\begin{proof}
By Lemma \ref{lem:ce}, for any $0\leq k,l\leq 2$ we have
\begin{equation}\label{eq:forOpt_1}
\sum_{i,j=0}^{2}p(i,j)|\langle k|\psi_i\rangle|^2\cdot|\langle
l|\psi_j\rangle|^2=\frac{1}{9}.
\end{equation}
This turns out to be equivalent
to
\begin{equation}
M\cdot P\cdot M^T=\frac{J}{9}.
\end{equation}
Noting that $M\cdot (J/9)\cdot M^T=J/9$, we know that $P$ can be
expressed as
\begin{equation}\label{eq:forOpt_2}
P = \frac{J}{9}+\bar{P},
\end{equation}
where $M\cdot\bar{P}\cdot M^T=0$. By straightforward calculation,
one can show that Eq.\eqref{eq:forOpt_2} indicates
\begin{equation}
M\cdot\bar{P}=0.
\end{equation}
Considering the form of $M$, $\bar{P}$ can now be expressed as
\begin{equation}
\bar{P}=
\begin{pmatrix}
k_0 & k_1 & k_2\\
-k_0x & -k_1x & -k_2x\\
-k_0(1-x) & -k_1(1-x) & -k_2(1-x)
\end{pmatrix},
\end{equation}
where $k_0,k_1$ and $k_2$ are real numbers.
According to the discussion above, we know that the maximal quantum
advantage is
\begin{equation}
QA = \sum_{i=0}^{2}p_1(i)[2\cdot\langle
l_i|\sigma_i|l_i\rangle-1\cdot(1-\langle l_i|\sigma_i|l_i\rangle)],
\end{equation}
where $l_i=\max_{l}\langle l|\sigma_i|l\rangle$. Then it holds that
\begin{align*}
QA & = 3\sum_{i=0}^{2}p_1(i)\cdot\langle l_i|\sigma_i|l_i\rangle -1 \\
& = 3\sum_{i,j=0}^{2}p(i,j)|\langle l_i|\psi_j\rangle|^2 -1 \\
& = 3\sum_{i,j=0}^{2}\left(\frac{1}{9}+\bar{p}(i,j)\right)|\langle l_i|\psi_j\rangle|^2 -1 \\
& = 3\sum_{i=0}^{2}\left(\sum_{j=0}^{2}\bar{p}(i,j)|\langle
l_i|\psi_j\rangle|^2\right),
\end{align*}
where $\bar{p}(i,j)$ is the element of $\bar{P}$. At the same time,
it can be obtained that $l_i=\max_{l}\sum_j \bar{p}(i,j)|\langle
l|\psi_j\rangle|^2$. Besides, recall that the rank of $M$ is 2, then
there must be one row of $M$, say $M_2$, has the form of
$aM_0+(1-a)M_1$, where $M_0$ and $M_1$ are the other two rows of
$M$, and $0\leq a\leq1$. Then it can be known that every $l_i$ must
be $0$ or $1$. Based on the form of $\bar{P}$, we have that $l_0\neq
l_1=l_2$. Without loss of generality, we suppose $l_0=0$, and
$l_1=l_2=1$. Then
\begin{align*}
QA & = 3\sum_{j=0}^{2}\bar{p}(0,j)|\langle
0|\psi_j\rangle|^2+3\sum_{j=0}^{2}(\bar{p}(1,j)+\bar{p}(2,j))|\langle
1|\psi_j\rangle|^2 \\
& = 3\sum_{j=0}^{2}\bar{p}(0,j)|\langle
0|\psi_j\rangle|^2-3\sum_{j=0}^{2}\bar{p}(0,j)|\langle
1|\psi_j\rangle|^2.
\end{align*}
Note that $\bar{P}+J/9$ is a matrix with nonnegative elements.
Thus, for any $0\leq i\leq 2$, if $k_i\geq0$ we have
\begin{equation}\label{eq:forOpt_3}
-k_ix\geq-\frac{1}{9} \quad \text{and} \quad
-k_i(1-x)\geq-\frac{1}{9},
\end{equation}
and if $k_i<0$, we have $-k_i\leq\frac{1}{9}$. And
Eq.\eqref{eq:forOpt_3} indicates that if $0<x<1$,
\begin{equation}\label{eq:forOpt_4}
k_i\leq\frac{1}{9x} \quad \text{and} \quad k_i\leq\frac{1}{9(1-x)},
\end{equation}
which is equivalent to
\begin{equation}\label{eq:forOpt_5}
k_i\leq\frac{1}{9x_b}.
\end{equation}
Actually, Eq.\eqref{eq:forOpt_5} also holds when $x=0$ or $x=1$.
Therefore, we obtain that
\begin{align*}
QA & = 3\sum_{j=0}^{2}\bar{p}(0,j)|\langle
0|\psi_j\rangle|^2-3\sum_{j=0}^{2}\bar{p}(0,j)|\langle
1|\psi_j\rangle|^2 \\
& = 3\sum_{j=0}^{2}k_j|\langle
0|\psi_j\rangle|^2-3\sum_{j=0}^{2}k_j|\langle
1|\psi_j\rangle|^2 \\
& \leq 3\cdot\frac{1}{9x_b}+3\cdot\frac{1}{9} \\
& = \frac{1}{3}+\frac{1}{3x_b},
\end{align*}
where the relationship
$\sum_j|\langle0|\psi_j\rangle|^2=\sum_j|\langle1|\psi_j\rangle|^2=1$
is utilized.
\end{proof}
Go back to the example in the above subsection. Note that for $M$ in
Eq.\eqref{eq:forExa_1} we have $M^0=1\cdot M^1+0\cdot M^2$(thus in
order to utilize Theorem.\ref{thm:Opt}, we need to adjust the order
of the columns). Thus we can choose $x=0$, and then $x_b=1$. As a
result, the discussion above shows that $QA\leq2/3$, which means the
choice of $P$ in Eq.\eqref{eq:forExa_2} is optimal for $M$ in
Eq.\eqref{eq:forExa_1}.
\section{Conclusion}
In this paper, we have shown that in certain fair equilibria of some
natural games, the player with a quantum computer can be more
powerful than the classical opponent, even if the equilibria have
zero entanglement or discord. This indicates that at least in games,
the standard understanding that nonzero discord is necessary to
produce non-classical correlations needs to be adjusted. Our work
provides new visions for further studies on quantum information,
especially when at least two parties are involved with different
objectives.
From the mathematical perspective, some questions remain open. Two
of them are listed as below: (1) What is the maximum gain in a
zero-sum $[-1,1]$-normalized game\footnote{A game is
$[-1,1]$-normalized if all utility functions have ranges within
$[-1,1]$.} on a state in symmetric subspace without entanglement?
(2) What is the maximum gain in a zero-sum $[-1,1]$-normalized game
on a state in symmetric subspace without discord?
\section*{Acknowledgments} Z.W. thanks Leong Chuan
Kwek and Luming Duan for helpful comments. Z.W. was supported by the Singapore National Research Foundation under NRF RF Award
No.~NRF-NRFF2013-13 and the WBS grant under contract no.
R-710-000-007-271. S.Z. was supported by Research Grants Council of the Hong Kong S.A.R. (Project no. CUHK419011, CUHK419413), and this research benefited from a visit to Tsinghua University partially supported by China Basic
Research Grant 2011CBA00300 (sub-project 2011CBA00301).
\bibliographystyle{alpha}
|
\section{The stable Heisenberg-Poincar\'{e} algebra and noncommutative
space-time}
When models are constructed for the natural world, it is reasonable to
expect that only those properties of the models that are robust have a
chance to be observed. Models or theories being approximations to the
natural world, it is unlikely that properties that are too sensitive to
small changes (that is, that depend in a critical manner on particular
values of the parameters) will be well described in the model. If a fine
tuning of the parameters is needed to reproduce some natural phenomenon,
then the model is basically unsound and its other predictions expected to be
unreliable. For this reason it would seem that a good methodological point
of view, in the construction of physical theories, would be to focus on the
robust properties of the models or, equivalently, to consider only models
which are stable, in the sense that they do not change, in a qualitative
manner, when some parameter changes.
This point of view had a large impact in the field of non-linear dynamics,
where it led to the rigorous notion of \textit{structural stability }\cit
{Andronov} \cite{Smale}. As pointed out by Flato \cite{Flato} and Faddeev
\cite{Faddeev} the same pattern seems to occur in the fundamental theories
of Nature. In fact, the two physical revolutions of the last century, namely
the passage from non-relativistic to relativistic and from classical to
quantum mechanics are deformations of two unstable Lie algebras to two
stable ones. A mathematical structure is said to be \textit{stable} (or
\textit{rigid}) for a class of \textit{deformations,} if any deformation in
this class leads to an equivalent (isomorphic) structure. When going from
Galilean to relativistic mechanics, the Galilean algebra, an isolated point,
is deformed to the stable Lorentz algebra and on the transition from
classical to quantum mechanics the unstable Poisson algebra is deformed to
the stable Moyal algebra.
This situation motivated the question of whether the full algebra of
relativistic quantum mechanics, the Heisenberg-Poincar\'{e} algebra, would
itself be stable. The answer was that it is not and that one possible
deformation to a stable one is defined by the following commutators \cit
{Vilela1}
\begin{equation}
\begin{array}{rll}
\lbrack M_{\mu \nu },M_{\rho \sigma }] & = & i(M_{\mu \sigma }\eta _{\nu
\rho }+M_{\nu \rho }\eta _{\mu \sigma }-M_{\nu \sigma }\eta _{\mu \rho
}-M_{\mu \rho }\eta _{\nu \sigma }) \\
\lbrack M_{\mu \nu },P_{\lambda }] & = & i(P_{\mu }\eta _{\nu \lambda
}-P_{\nu }\eta _{\mu \lambda }) \\
\lbrack M_{\mu \nu },x_{\lambda }] & = & i(x_{\mu }\eta _{\nu \lambda
}-x_{\nu }\eta _{\mu \lambda }) \\
\lbrack P_{\mu },P_{\nu }] & = & -i\frac{\epsilon _{4}}{R^{2}}M_{\mu \nu }
\\
\lbrack x_{\mu },x_{\nu }] & = & -i\epsilon _{5}\ell ^{2}M_{\mu \nu } \\
\lbrack P_{\mu },x_{\nu }] & = & i\eta _{\mu \nu }\Im \\
\lbrack P_{\mu },\Im ] & = & -i\frac{\epsilon _{4}}{R^{2}}x_{\mu } \\
\lbrack x_{\mu },\Im ] & = & i\epsilon _{5}\ell ^{2}P_{\mu
\end{array}
\label{1.1}
\end{equation
In this algebra, that will be denoted $\Re _{\ell ,R}=\{M_{\mu \nu },P_{\mu
},x_{\mu },\Im \}$, $M_{\mu \nu }$ are the generators of the Lorentz group,
P_{\mu }$ and $x_{\mu }$ the momenta and coordinates, $\Im $ a non-trivial
operator that replaces the center of the Heisenberg algebra and $\epsilon
_{4},\epsilon _{5}$ are $\pm $ signs. Velocities and actions are in units of
$c$ and $\hbar $ (that is $c=\hbar =1$). The stable algebra $\Re _{\ell ,R}$
is isomorphic to the algebra of the $6-$dimensional pseudo-orthogonal group
with metric
\begin{equation}
\eta _{aa}=(1,-1,-1,-1,\epsilon _{4},\epsilon _{5}),\bigskip \ \epsilon
_{4},\epsilon _{5}=\pm 1 \label{1.2}
\end{equation
The nonvanishing right hand side of the $[P_{\mu },P_{\nu }]$ commutator
simply means that flat space is an isolated point in the set of arbitrarily
curved spaces. This is why Faddeev \cite{Faddeev} points out that Einstein's
theory of gravity may also be considered as a deformation in a stable
direction. Einstein's theory is based on curved pseudo Riemann manifolds.
Therefore, in the set of Riemann spaces, Minkowski space is a kind of
degeneracy whereas a generic Riemann manifold is stable in the sense that in
its neighborhood all spaces are curved. However, as long as one is concerned
with the kinematical group of the tangent space to the space-time manifold,
and not with the group of motions in the manifold itself, it is perfectly
consistent to take $R\rightarrow \infty $ and this deformation would be
removed. In particular, because the curvature is not a constant, $R$ cannot
have the status of a fundamental constant.
In contrast, for the other stabilizing deformation, associated to the $\ell $
constant, there is no obvious reason to remove it and $\ell $ might play the
role of a new fundamental constant. Taking the $\Re _{\ell ,\infty }$
algebra as the kinematical algebra of tangent space, the main features are
the non-commutativity of the $x_{\mu }$ coordinates and the fact that $\Im
, previously a trivial center of the Heisenberg algebra, becomes now a
non-trivial operator. These are however the minimal changes that seem to be
required if stability of the algebra of observables (in the tangent space)
is a good guiding principle. Two constants define this deformation. One is
\ell $, a \textit{fundamental length}, the other the sign $\epsilon _{5}$.
The algebra $\Re _{\ell ,\infty }$ is seen to be the algebra of the
pseudo-Euclidean groups $E(1,4)$ or $E(2,3)$, depending on whether $\epsilon
_{5}$ is $-1$ or $+1$. It has a simple representation by differential
operators in a five-dimensional space with coordinates $(\xi _{0},\xi
_{1},\xi _{2},\xi _{3},\xi _{4})
\begin{equation}
\begin{array}{rll}
P_{\mu } & = & i\frac{\partial }{\partial \xi ^{\mu }}+iD_{P_{\mu }} \\
M_{\mu \nu } & = & i(\xi _{\mu }\frac{\partial }{\partial \xi ^{\nu }}-\xi
_{\nu }\frac{\partial }{\partial \xi ^{\mu }})+\Sigma _{\mu \nu } \\
x_{\mu } & = & \xi _{\mu }+i\ell (\xi _{\mu }\frac{\partial }{\partial \xi
^{4}}-\epsilon _{5}\xi ^{4}\frac{\partial }{\partial \xi ^{\mu }})+\ell
\Sigma _{\mu 4} \\
\Im & = & 1+i\ell \frac{\partial }{\partial \xi ^{4}}+i\ell D_{\xi ^{4}
\end{array}
\label{1.3}
\end{equation
The set ($\Sigma _{\mu \nu },\Sigma _{\mu 4}$) is an internal spin operator
for the groups $O(4,1)$ (if $\epsilon _{5}=-1$) or $O(3,2)$ (if $\epsilon
_{5}=+1$) and $D_{P_{\mu }}$ and $D_{\xi ^{4}}$ are derivations operating in
the space where ($\Sigma _{\mu \nu },\Sigma _{\mu 4}$) acts. For practical
calculations, in particular for the construction of quantum fields, it may
be convenient to use this representation. Notice however that only the
Poincar\'{e} part of $E(1,4)$ or $E(2,3)$ corresponds to symmetry operations
and only this part has to be implemented by unitary operators. Also,
although an extra dimension is used in the representation space, the
space-time coordinates are still only four, noncommutative ones. Physical
consequences of the non-commutative space-time structure implied by the $\Re
_{\ell ,\infty }$ algebra have already been explored in a series of
publications \cite{Vilela2}-\cite{Vilela8}. Depending on the sign of
\epsilon _{5}$ the time ($\epsilon _{5}=+1$) or one space variable (
\epsilon _{5}=-1$) will have discrete spectrum. In any case $\ell $, a new
fundamental constant, sets a natural scale for time and length. If $\ell $
is of the order of Planck's length, observation of most of the effects
worked out in the cited references will be beyond present experimental
capabilities. However, if $\ell $ is much larger than Planc
\'{
s length (for example of order $10^{-27}-10^{-26}$ seconds) the effects
might already be observable in the laboratory or in astrophysical
observations. Some of the most noteworthy effects arise from the
modification of the phase space volume and from interference effects.
However, most of the consequences worked out in the references \cite{Vilela2
-\cite{Vilela6} are rather conservative, in the sense that they simply
explore the nonvanishing of the right-hand-side of the commutators of
previously commuting variables. Deeper consequences are to be expected from
the radical change from a commutative to a non-commutative space-time
geometry, in particular from the new differential algebra associated to the
geometry. One such consequence will be described in this paper. The new
geometry was studied in Ref.\cite{Vilela7} to which I will refer for details
and notation.
\section{The differential algebra and an extended Dirac equation}
In the framework of the non-commutative geometry implied by the deformed
algebra, the differential algebra may be constructed either by duality from
the derivations of the algebra or from the triple $\left( H,\pi (U_{\Re
}),D\right) $, where $U_{\Re }$ is the enveloping algebra of $\Re _{\ell
,\infty }$, to which a unit and, for later convenience, the inverse of $\Im
, are added.
\begin{equation}
U_{\Re }=\{x_{\mu },M_{\mu \nu },p_{\mu },\Im ,\Im ^{-1},1\} \label{3.13}
\end{equation
$\pi (U_{\Re })$ is a representation of the $U_{\Re }$ algebra in the
Hilbert space $H$ and $D$ is the Dirac operator, the commutator with the
Dirac operator being used to generate the one-forms. In a general
non-commutative framework \cite{Connes2} \cite{Connes-Marcolli} it is not
always possible to use the derivations of the algebra to construct by
duality the differential forms. In particular, many algebras have no
derivations at all. However when the algebra has enough derivations it is
useful to consider them \cite{Dubois1} \cite{Dubois2} because the
correspondence of the non-commutative geometry notions to the classical ones
becomes very clear. One considers here the set $V$ of derivations with basis
$\{\partial _{\mu },\partial _{4}\}$ defined as follows\footnote
Notice that the definition of $\partial _{4}$ here is slightly different
from the one in Ref.\cite{Vilela7}.
\begin{equation}
\begin{array}{lll}
\partial _{\mu }(x_{\nu }) & = & \eta _{\mu \nu }\Im \\
\partial _{4}(x_{\mu }) & = & -\epsilon _{5}\ell p_{\mu }\Im \\
\partial _{\sigma }(M_{\mu \nu }) & = & \eta _{\sigma \mu }p_{\nu }-\eta
_{\sigma \nu }p_{\mu } \\
\partial _{\mu }(p_{\nu }) & = & \partial _{\mu }(\Im )=\partial _{\mu }(1)=0
\\
\partial _{4}(M_{\mu \nu }) & = & \partial _{4}(p_{\mu })=\partial _{4}(\Im
)=\partial _{4}(1)=
\end{array}
\label{3.14}
\end{equation
In the commutative ($\ell =0$) case a basis for $1$-forms is obtained, by
duality, from the set $\{\partial _{\mu }\}$. In the $\ell \neq 0$ case the
set of derivations $\{\partial _{\mu },\partial _{4}\}$ is the minimal set
that contains the usual $\partial _{\mu }$'s, is maximal abelian and is
action closed on the coordinate operators, in the sense that the action of
\partial _{\mu }$ on $x_{\nu }$ leads to the operator $\Im $ associated to
\partial _{4}$ and conversely.
The operators that are associated to the physical coordinates are just the
four $x_{\mu }$, $\mu \in (0,1,2,3)$. However, an additional degree of
freedom appears in the set of derivations. This is not a conjectured extra
dimension but simply a mathematical consequence of the algebraic structure
of $\Re _{\ell ,\infty }$ which, in turn, was a consequence of the
stabilizing deformation of relativistic quantum mechanics. No extra
dimension appears in the set of physical coordinates, because it does not
correspond to any operator in $\Re _{\ell ,\infty }$. However the
derivations in $V$ introduce, by duality, an additional degree of freedom in
the exterior algebra. For example, all quantum fields that are Lie
algebra-valued connections will pick up additional components. These
additional components, in quantum fields that are connections, are a
consequence of the length parameter $\ell $ which does not depend on its
magnitude, but only on $\ell $ being $\neq 0$.
The Dirac operator \cite{Vilela7} i
\begin{equation}
D=i\gamma ^{a}\partial _{a} \label{3.15}
\end{equation
with $\partial _{a}=\left( \partial _{\mu },\partial _{4}\right) $, the
\gamma $'s being a basis for the Clifford algebras $C\left( 3,2\right) $ or
C\left( 4,1\right)
\begin{equation}
\begin{array}{ccc}
\left( \gamma ^{0},\gamma ^{1},\gamma ^{2},\gamma ^{3},\gamma ^{4}=\gamma
^{5}\right) & & \epsilon _{5}=+1 \\
\left( \gamma ^{0},\gamma ^{1},\gamma ^{2},\gamma ^{3},\gamma ^{4}=i\gamma
^{5}\right) & & \epsilon _{5}=-
\end{array}
\label{3.16}
\end{equation
How to construct quantum, scalar, spinor and gauge fields, as operators in
U_{\Re }$, has been described in \cite{Vilela7}. In particular the role of
the additional dimension in the exterior algebra, on gauge interactions, has
been emphasized (see also \cite{Vilela6}). Here another potentially
interesting consequence for spinor fields will be described. Because
\begin{equation}
\left[ p_{\mu },e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\}
_{+}}\right] =k_{\mu }e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im
^{-1}\right\} _{+}} \label{3.17}
\end{equation
a spinor field is writte
\begin{equation}
\Psi =\int d^{4}k\delta (k^{2}-m^{2})\left\{ b_{k}u_{k}e^{-\frac{i}{2}k_{\nu
}\left\{ x^{\nu },\Im ^{-1}\right\} _{+}}+d_{k}^{\ast }v_{k}e^{\frac{i}{2
k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\} _{+}}\right\} \label{3.18}
\end{equation
\begin{equation}
\Psi \in U_{\Re }:D\Psi -m\Psi =0 \label{3.19}
\end{equation
From the field a wave function is constructed operating on a vacuum stat
\begin{equation}
\psi =\Psi \left\vert 0\right\rangle \label{3.19a}
\end{equation
Notice that both $b_{k},d_{k}^{\ast }$ and the elements of $U_{\Re }$
operate on $\left\vert 0\right\rangle $, in particular $p_{\mu }\left\vert
0\right\rangle =0$. Now, for a massless field, the (extended) Dirac equation
becomes
\begin{equation}
D\psi =i\gamma ^{a}\partial _{a}\psi =\left( i\gamma ^{\mu }\partial _{\mu
}+i\gamma ^{4}\partial _{4}\right) \psi =0 \label{3.20}
\end{equation
Writ
\begin{equation*}
\psi =e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\} _{+}}u\left(
k\right)
\end{equation*
Fro
\begin{equation}
\begin{array}{lll}
\partial _{\mu }e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\}
_{+}} & = & -ik_{\mu }e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im
^{-1}\right\} _{+}} \\
\partial _{4}e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\} _{+}}
& = & -i\epsilon _{5}\ell \left( -k^{\mu }p_{\mu }+\frac{1}{2}k^{2}\right)
e^{-\frac{i}{2}k_{\nu }\left\{ x^{\nu },\Im ^{-1}\right\} _{+}
\end{array}
\label{3.21}
\end{equation
one obtains, using (\ref{3.21}), (\ref{3.17}) and (\ref{3.19a}
\begin{equation}
\begin{array}{ccc}
\left( \gamma ^{\mu }k_{\mu }-\gamma ^{5}\ell \frac{1}{2}k^{2}\right)
u\left( k\right) =0 & & \epsilon _{5}=+1 \\
\left( \gamma ^{\mu }k_{\mu }+i\gamma ^{5}\ell \frac{1}{2}k^{2}\right)
u\left( k\right) =0 & & \epsilon _{5}=-
\end{array}
\label{3.22}
\end{equation
Let $\epsilon _{5}=-1$. Iterating (\ref{3.22}
\begin{equation}
\left( k^{2}-\frac{\ell ^{2}}{4}\left( k^{2}\right) ^{2}\right) u\left(
k\right) =0 \label{3.23}
\end{equation
This equation has two solutions, the massless solution ($k^{2}=0$) and
another one, of large mass ($\ell $ being small
\begin{equation}
k^{2}=\frac{4}{\ell ^{2}} \label{3.24}
\end{equation
For $\epsilon _{5}=+1$ one would obtain $k^{2}=0$ an
\begin{equation}
k^{2}=-\frac{4}{\ell ^{2}} \label{3.24a}
\end{equation
a tachyonic large $\left\vert k^{2}\right\vert $ solution.
The solutions of the extended Dirac equation for $k^{2}=0$ are the usual
ones. To find the nature of the solutions for $\left\vert k^{2}\right\vert
\frac{4}{\ell ^{2}}$, $\epsilon _{5}=-1$ and $+1$, use a Majorana imaginary
representation for the gamma matrice
\begin{eqnarray}
\gamma ^{0} &=&\left(
\begin{array}{cc}
0 & \sigma _{2} \\
\sigma _{2} &
\end{array
\right) ;\gamma ^{1}=\left(
\begin{array}{cc}
i\sigma _{1} & 0 \\
0 & i\sigma _{1
\end{array
\right) ;\gamma ^{2}=\left(
\begin{array}{cc}
0 & \sigma _{2} \\
-\sigma _{2} &
\end{array
\right) \notag \\
\gamma ^{3} &=&\left(
\begin{array}{cc}
i\sigma _{3} & 0 \\
0 & i\sigma _{3
\end{array
\right) ;\gamma ^{5}=i\gamma ^{0}\gamma ^{1}\gamma ^{2}\gamma ^{3}=\left(
\begin{array}{cc}
-\sigma _{2} & 0 \\
0 & \sigma _{2
\end{array
\right) \label{3.25}
\end{eqnarray}
\subsection{$\protect\epsilon _{5}=-1,k^{2}=\frac{4}{\ell ^{2}}$}
In the rest frame $k=(m_{0}=\pm \frac{2}{\ell },0,0,0)$. The second equation
in (\ref{3.22}) leads t
\begin{equation*}
\left( \pm \gamma ^{0}+i\gamma ^{5}\right) u=0
\end{equation*
\begin{equation*}
\begin{array}{ccc}
u=\left(
\begin{array}{c}
a \\
i
\end{array
\right) & & \text{Positive energy }\left( m_{0}=\frac{2}{\ell }\right) \\
u=\left(
\begin{array}{c}
a \\
-i
\end{array
\right) & & \text{Negative energy }\left( m_{0}=-\frac{2}{\ell }\right)
\end{array
\end{equation*
where $a$ is an arbitrary two-vector. The solutions of non-zero momentum are
obtained by the application of a proper Lorentz transformation $\exp \left(
\frac{1}{2}\alpha _{\mu \nu }\left\{ \gamma ^{\mu },\gamma ^{\nu }\right\}
_{+}\right) $.
One has $u^{\ast }\neq u$, hence these solutions are Dirac spinors.
\subsection{$\protect\epsilon _{5}=+1,k^{2}=-\frac{4}{\ell ^{2}}$}
Here one makes $k=(0,0,0,\frac{2}{\ell })$, obtainin
\begin{equation*}
\left( \gamma ^{3}-\gamma ^{5}\right) u^{\prime }=0
\end{equation*
Making $u^{\prime }=$ $\left(
\begin{array}{c}
a \\
\end{array
\right) $, $a$ and $b$ being two-vectors, yield
\begin{eqnarray*}
\left( \sigma _{2}+i\sigma _{3}\right) a &=&0 \\
\left( \sigma _{2}-i\sigma _{3}\right) b &=&0
\end{eqnarray*
meaning that $a$ and $b$ are independent two vector
\begin{equation*}
a=\left(
\begin{array}{c}
a_{1} \\
a_{1
\end{array
\right) ;b=\left(
\begin{array}{c}
b_{1} \\
-b_{1
\end{array
\right)
\end{equation*
with $a_{1}$ and $b_{1}$ real numbers $u^{\prime \ast }=u$ and this
tachyonic, large $\left\vert k^{2}\right\vert $, solution is a Majorana
spinor.
As before, general solutions would be obtained by the action of a Lorentz
transformation.\
\section{Coupling the two solutions}
In the previous section it was seen how the extended Dirac equation,
following from the exterior algebra of noncommutative space-time, has both a
massless and a large $\left\vert k^{2}\right\vert $ solution, large if $\ell
$ is small. If, for example, $\ell $ is in the $10^{-27}-10^{-26}$ seconds
range, $M=\frac{2}{\ell }$ would be of the order of $1$TeV. If the two
solutions mix, one expects that the massless solution would acquire a small
mass as in the seesaw mechanism proposed for neutrinos. In the seesaw
mechanism the large mass (of right-handed neutrinos) is hypothesized to be
obtained from the lepton number violation scale at grand unification. Here
the large mass arises from an independent solution of the same extended
equation.
Let us call $u_{1}$ the zero mass solution and $u_{2}$ the large mass
solution. Then let, us assume that they are coupled by interaction with a
background scalar field that acquires a nonzero vacuum expectation value
\phi =\left\langle \phi \right\rangle +h$, with Lagrangia
\begin{equation}
\mathcal{L}=\overline{u_{1}}i\gamma ^{\mu }\partial _{\mu }u_{1}+\overline
u_{2}}\left( i\gamma ^{\mu }\partial _{\mu }+\gamma ^{4}\frac{2}{\ell
\right) u_{2}+\left( g\overline{u_{1}}\left( \left\langle \phi \right\rangle
+h\right) u_{2}+h.c.\right) \label{4.1}
\end{equation
leading to the equations of motion
\begin{equation}
\begin{array}{ccc}
i\gamma ^{\mu }\partial _{\mu }u_{1}+g\left( \left\langle \phi \right\rangle
+h\right) u_{2} & = & 0 \\
\left( i\gamma ^{\mu }\partial _{\mu }+\gamma ^{4}\frac{2}{\ell }\right)
u_{2}+g^{\ast }\left( \left\langle \phi \right\rangle +h^{\ast }\right) u_{1}
& = &
\end{array}
\label{4.2}
\end{equation
With $\frac{2}{\ell }$ large, one neglects the kinetic term in the last
equation and obtains, in leading order
\begin{equation}
\begin{array}{ccc}
u_{2}\simeq -\frac{\ell }{2}g^{\ast }\left\langle \phi \right\rangle \gamma
^{5}u_{1} & & \epsilon _{5}=+1 \\
u_{2}\simeq i\frac{\ell }{2}g^{\ast }\left\langle \phi \right\rangle \gamma
^{5}u_{1} & & \epsilon _{5}=-
\end{array}
\label{4.3}
\end{equation
Substitution in the first equation of (\ref{4.2}) yields
\subsection{$\protect\epsilon _{5}=-1$}
\begin{equation*}
\left( i\gamma ^{\mu }\partial _{\mu }+i\left\vert g\right\vert
^{2}\left\langle \phi \right\rangle ^{2}\frac{\ell }{2}\gamma ^{5}\right)
u_{1}\simeq 0
\end{equation*
which has a small mass solution wit
\begin{equation*}
k^{2}=\left( \left\vert g\right\vert ^{2}\left\langle \phi \right\rangle ^{2
\frac{\ell }{2}\right) ^{2}
\end{equation*
For $k=\left( \pm \left\vert g\right\vert ^{2}\left\langle \phi
\right\rangle ^{2}\frac{\ell }{2},0,0,0\right) $ the solutions ar
\begin{equation*}
u_{1}=\left(
\begin{array}{c}
a \\
\pm i
\end{array
\right)
\end{equation*
a small mass Dirac spinor. $a$ is an arbitrary two-vector.
\subsection{$\protect\epsilon _{5}=+1$}
\begin{equation*}
\left( i\gamma ^{\mu }\partial _{\mu }-\left\vert g\right\vert
^{2}\left\langle \phi \right\rangle ^{2}\frac{\ell }{2}\gamma ^{5}\right)
u_{1}^{\prime }\simeq 0
\end{equation*
which has a small $\left\vert k^{2}\right\vert $ solution wit
\begin{equation*}
k^{2}=-\left( \left\vert g\right\vert ^{2}\left\langle \phi \right\rangle
^{2}\frac{\ell }{2}\right) ^{2}
\end{equation*
For $k=\left( 0,0,0,\left\vert g\right\vert ^{2}\left\langle \phi
\right\rangle ^{2}\frac{\ell }{2}\right) $ the solution i
\begin{equation*}
u_{1}^{\prime }=\left(
\begin{array}{c}
a_{1} \\
a_{1} \\
b_{1} \\
-b_{1
\end{array
\right)
\end{equation*
a small $\left\vert k^{2}\right\vert $ tachyonic Majorana spinor.
|
\section{Introduction}
Starting from the pioneering works of mid Seventies of last century, laser cooling of atoms has become a vast research field with many applications in physics, chemistry and biology.
\par
When an atom absorbs a photon, its kinetic energy is changed according to the laws of energy and linear momentum conservation. In general, but not always (section \ref{linassappsec}), if an atom flying against a photon absorbs it, the kinetic energy of the atom decreases. Instead, if a photon is chasing the atom, the variation of the atom's kinetic energy due to the absorption of the photon is always positive.
\par
The proposal of cooling atoms with laser beams has been put forward by H\"{a}nsch and Schawlow in 1975 \cite{hs}: the idea was that of illuminating the atoms with six laser beams (two opposite beams for each spatial dimension) red detuned with respect to an absorption line.
\par
Apparently, there are two limits to the cooling process: the Doppler limit and the recoil limit. The Doppler limit is due to the natural width $\Gamma$ (Full Width at Half Maximum) of the atomic transition used: since the first order Doppler shift is $\pm \Delta E (v/c)$ (where $\Delta E$ is the transition energy between the two atomic levels), when $\Delta E (v/c)\approx \Gamma$, the photon may be absorbed (with significant probability) not only by atoms flying against the photon but also by atoms flying in the opposite direction, thus limiting the cooling process.
Instead, the recoil limit is due to the fact that when the kinetic energy of the absorbing atom is of the same order of magnitude of its variation due to the absorption or emission of a photon no further cooling seems to be possible.
\par
The theoretical treatment of laser cooling is not a simple one. In the semi--classical approach, the atom is considered as a localized two--levels quantum system and the light field is treated classically \cite{stenholm}.
According to this theory, the lowest attainable average kinetic energy $<E_K>$ of the atoms is given by $\Gamma/4$. This result can not be valid in all conditions since it implies that $<E_K>\rightarrow 0$ as $\Gamma\rightarrow 0$. This physically unsound result is due to the assumption that the photon momentum is negligible with respect to the atomic one. This approximation implies that the kinetic energy of the atom is much larger than the recoil energy $E_R$. As we shall see, the semiclassical theory is applicable only when $\Gamma\gg E_R$.
\par
A quantum mechanical treatment of the motion of a two levels atom under laser light, applicable for any value of the ratio between the natural linewidth and the recoil energy, has been developed, among others, by Winelnd and Itano \cite{wineland} and by Castin, Wallis and Dalibard \cite{castin}.
\par
The discovery by Lett et al. that temperatures well below the Doppler limit can be achieved \cite{lett} inspired a re--formulation of the theoretical description: it was found that the multilevel nature of the atoms and the spatial variation of the light field polarization can be exploited for attaining temperatures close to the recoil limit \cite{dalibard}.
\par
Temperatures below the recoil limit can be achieved by sophisticated procedures like the Velocity Selective Coherent Population Trapping \cite{dark} and the Stimulated Raman Cooling \cite{raman}. The book by Metcalf and van der Straten may be taken as a guide for exploring all these issues and the vast available literature \cite{metstra}.
\par
Stimulated by these developments, the search for ever lower temperatures has attracted the attention of experimental and theoretical physicists. This notwithstanding, the theoretical reconsideration of laser cooling of two levels atoms maintains its importance, since it sheds light on essential features of the cooling process.
\par
This paper is based on the idea that the application of conservation laws to the absorption/emission of photons by atoms should yield the essential feature of laser cooling of two levels atoms at low laser intensities. We shall see that this is indeed the case,
\par
The conceptual framework of the following treatment is simple and the mathematics needed is limited to some algebraic manipulations. Therefore, this paper might be of some interest for university and high school teaching. It might also be of some value for researchers: being based on conservation principles, it sets down limiting conditions that should be met also by more sophisticated (and complicated) treatments like the quantum mechanical ones.
\par
As we shall see in due course, the present treatment begins with the use of relativistic dynamics; however, given the small atoms' velocities of interest in the final steps of the laser cooling process, the relativistic dynamics will be approximated for small velocities. Of course, it is possible to use directly the Newtonian mechanics (this may be a choice for high school teaching). However, it is worth stressing the conceptual differences between the two approaches. Within the relativistic approach, the rest energy $Mc^2$ appears naturally with its physical meaning and the transition energy $\Delta E$ between two atomic levels, being the difference of two rest energies, is a relativistic invariant; instead, in the Newtonian treatment, the physical meaning of $Mc^2$, which enters the expression of the recoil energy, remains obscure.
\section{Outline of the paper}
An outline of the main steps of the paper will help in understanding how the various calculations contribute to the overall description.
We shall deal only with energies: the transition energy $\Delta E_M$, the recoil energy $E_R=\Delta E_M^2/2Mc^2$, the natural linewidth $\Gamma$ and the energy $E_{ph}$ of the absorbed/emitted photon. The transition energy $\Delta E_M$ is the one corresponding to the most probable value given by the Lorentzian shape of the natural line.
\par
In section \ref{basic}, we shall assemble the basic formulas, giving them the most suitable form for the laser cooling process. Two dimensionless parameters are introduced:
\begin{equation}\label{BT}
B_T=\frac{E_R}{\Delta E_M}=\frac{\Delta E_M}{2Mc^2}
\end{equation}
and
\begin{equation}\label{BD}
B_D=\frac{\Gamma}{\Delta E_M}
\end{equation}
Their ratio $B_D/B_T$ is equal to $\Gamma/E_R$.
A third dimensionless parameter will enter the description: the parameter $B_1=v_1/c$, where $v_1$ is the norm of the atom's velocity vector $\vec v_1$ before the absorption/emission of a photon. The use of dimensionless parameters simplify the formulas and allows an easy comparison of the orders of magnitude; however, when needed, the basic physical quantities will be re--established.
\par
In section \ref{naturalsec}, the implications of a non zero linewidth are discussed. In particular, it is shown that when an atom absorbs a counter--propagating photon its transition energy is $\Delta E^A$, while when the atom absorbs a co--propagating photon of the same energy its transition energy is $\Delta E^P=\Delta E^A(1-B_1)/(1+B_1)$.
\par
In section \ref{linear}, the basic formulas are approximated in the limit of small atoms' velocities: in this limit, only the first order Doppler effect is taken into account and the relativistic dynamics can be replaced by the Newtonian one.
In order to identify this limit without ambiguity, we must take into account that the formulas contain terms of the first or higher order in the $B$'s ($B_1, B_T, B_D$): the linear approximation is valid insofar as only terms linear in the $B$'s can be safely retained. We shall distinguish between photon absorptions that decrease the atom's kinetic energy and photon absorptions that increase it; we shall also find that, on the average, the spontaneous emission of a photon increases the atom's kinetic energy by an amount equal to the recoil energy $E_R$ (in the linear approximation). The balance between these competing processes (cooling and heating) characterizes the steady state condition in laser cooling.
\par
In section \ref{laser}, the formulas of section \ref{linear} are applied to laser cooling, in the limit of low laser intensity, i.e. in the limit in which the stimulated emission is negligible. It is assumed that the laser photons are red detuned. The average variation of the kinetic energy $<\Delta E_K>$ of an atom, due to the absorption of a laser photon and the subsequent emission of a fluorescence one, is calculated by taking into account the different probability that the absorbed photon belongs to one of the two counter--propagating laser beams. The condition $<\Delta E_K>=0$ yields a unique value of the atom's velocity parameter $B_1$ that depends on the detuning parameter $\delta^*$. This value of $B_1$ can be minimized as a function of the detuning parameter $\delta^*$, thus obtaining the lowest value $B_{1min}$. The absorbed laser photons decrease the kinetic energy of the atoms with velocity parameter larger than $B_1$ (or $B_{1min}$) and increase the kinetic energy of the atoms with velocity parameter smaller than $B_1$ ($B_{1min}$).
By identifying $cB_1$ with $v_{rms}$ (or $cB_{1min}$ with $v_{rms}$),
we calculate the average kinetic energy $<E_K>=(1/2)Mc^2{B^2_1}$ in a steady state condition, or the lowest attainable average kinetic energy $<E_K>_{min}=(1/2)Mc^2{B^2_{1min}}$.
\par
Finally, in the Appendix, the absorption of a counter--propagation photon by an atom is dealt within Newtonian mechanics in order to highlights the basic conceptual differences with the relativistic treatment, in spite of the fact that the two approximated calculations yields the same equations.
\section{Basic formulas\label{basic}}
The first treatment of the absorption/emission of photons by atoms based on energy and linear momentum conservations is due to Schr\"odinger' \cite{erwin}. Schr\"odinger's paper has been ignored by his contemporaries and even nowadays it is not so popular: see, for instance, \cite{dragan, gg1, gg2}. Reasonably, this oblivion has been due to the deep rooting of the wave description of light in the background physical knowledge, in spite of the fact that, the absorption/emission of light by atoms is a discrete process.
In \cite{erwin}, Schr\"odinger's dealt only with the emission of photons and, in particular, he never introduced explicitly the energy difference $\Delta E$ between the two levels of the atomic transition. Therefore, in the following, the original treatment by Schr\"odinger is extended to the absorption case (as in \cite{gg1, gg2}) and the form of the equations is adapted to the problem under study. Schr\"odinger's approach, besides the conservations laws and special relativity, assumes that the absorption/emission process is instantaneous: more physically, that the duration of the absorption/emission process is much smaller than the lifetimes of the atomic energy levels.
\begin{figure}[htb]
\centerline{
\includegraphics[width=6cm]{fig_1.eps}
}
\caption{\label{erwinfig}
Emission of a photon by the atom $A$ in motion. The photon is emitted along the direction $A\rightarrow
O$. The subscript $1$ denotes the quantities before the emission; the subscript $2$ the quantities after the emission.
}
\end{figure}
\par\noindent
For the emission of a photon by an atom, the conservation laws are (see fig. \ref{erwinfig}):
\begin{equation}\label{energia2}
E_{ph}=\gamma_1 E_1-\gamma_2 E_2
\end{equation}
for energy, and
\begin{eqnarray}
\gamma_1 {{E_1}\over{c^2}}v_1 \cos\theta_1 &= & \gamma_2
{{E_2}\over{c^2}}v_2 \cos\theta_2 +{{E_{ph}}\over{c}}
\label{qmx}\\
\gamma_1 {{E_1}\over{c^2}}v_1 \sin\theta_1 &= & \gamma_2
{{E_2}\over{c^2}}v_2 \sin\theta_2\label{qmy}
\end{eqnarray}
for linear momentum.
$E_{ph}$ is the energy of the emitted photon; $E_1$ and $E_2$ are the rest energies of the atom before and after the emission; $\gamma_1,\,\gamma_2$ are the relativistic factors before and after the emission; $v_1$ and $v_2$ the atom's velocities before and after the emission; $\theta_1$ and $\theta_2$ the angles between $\vec v_1$ and $\vec v_2$ and the direction of the emitted photon. Notice that $E_1-E_2=\Delta E$, where $\Delta E$ is the energy difference between the two levels of the atomic transition: $\Delta E$ is a relativistic invariant, since it is given by the difference of two rest energies.
After some calculations, we get:
\begin{equation}\label{interab}
E_{ph}=\frac{1}{2}\left( \frac{E_1}{\varphi_1}-\frac{E_2}{\varphi_2} \right)
\end{equation}
where:
\begin{equation}\label{varphi}
\varphi_i= \gamma_i\left(1- \frac{v_i}{c} \cos\theta_i \right)=\gamma_i\left(1- B_i \cos\theta_i\right); \qquad i=1,2
\end{equation}
Equation (\ref{interab}), by taking into account that $E_1\varphi_1=E_2\varphi_2$, can be written as:
\begin{equation}\label{inter12}
E_{ph}=\frac{1}{2}\frac{E_1^2-E_2^2}{E_1\varphi_1}=\frac{1}{2}\frac{E_1^2-E_2^2}{E_2\varphi_2}
\end{equation}
Notice that $E_2=Mc^2$ and $E_1= Mc^2+\Delta E$ with $\Delta E\ll Mc^2$. From now on, we shall use the dimensionless parameter $B_T=\Delta E/(2Mc^2)$: as we shall see, $B_T$ is a threshold parameter and the velocity $v_T=cB_T$ a threshold velocity.
\par
It is easy to verify that :
\begin{equation}
\frac{E_1^2-E_2^2}{2E_2}= \frac{E_1^2-E_2^2}{2Mc^2}= \Delta E \left(1+\frac{\Delta E}{2Mc^2} \right)=\Delta E(1+B_T)
\end{equation}
\begin{equation}
\frac{E_1^2-E_2^2}{2E_1} = \Delta E \left(1-\frac{\Delta E}{2E_1} \right)\approx \Delta E(1-B_T)\label{bta}
\end{equation}
$\Delta E/(2E_1)$ differs from $B_T$ by a term of the order of $B_T^2$; therefore, we shall replace in (\ref{bta}) the $\approx$ sign with that of the equality.
Therefore, from equation (\ref{inter12}), we get:
\begin{equation}\label{uguale}
E_{ph} = \Delta E \left( 1 - B_T\right) {{\sqrt{1-B_1^2}}\over{1 -B_1\cos \theta_1 }}=\Delta E \left( 1 + B_T\right) {{\sqrt{1-B_2^2}}\over{1 -B_2\cos \theta_2 }}
\end{equation}
or, in compact form:
\begin{equation}\label{semplice}
E_{ph}=\frac{\Delta E}{\varphi_1}(1-B_T)=\frac{\Delta E}{\varphi_2}(1+B_T)
\end{equation}
The case of absorption can be treated in the same way, starting from adequately re-written conservation equations. It turns out that the energy of the absorbed photon is given by:
\begin{equation}\label{ugualeass}
E_{ph} = \Delta E \left( 1 + B_T\right) {{\sqrt{1-B_1^2}}\over{1 -B_1\cos \theta_1 }}=\Delta E \left( 1 - B_T\right) {{\sqrt{1-B_2^2}}\over{1 -B_2\cos \theta_2 }}
\end{equation}
or by the compact equation:
\begin{equation}\label{sempliceass}
E_{ph}=\frac{\Delta E}{\varphi_1}(1+B_T)=\frac{\Delta E}{\varphi_2}(1-B_T)
\end{equation}
Both equations (\ref{uguale}) yield the energy of the emitted photon, one in terms of the atom's velocity parameters before the emission ($B_1, \theta_1$), the other in terms of the atom's velocity parameters after the emission ($B_2, \theta_2$). This last equation, on the other hand, yields the energy of a photon absorbed by an atom with initial velocity parameters ($B_2, \theta_2$) (first equation of (\ref{ugualeass})). Therefore: if an excited atom with velocity parameters ($B_1, \theta_1$) emits a photon, the same atom, after the emission and, therefore, with velocity parameters ($B_2, \theta_2$), can absorb a photon of the same energy.
\par
The variation of the atom's kinetic energy due to the emission of a photon is given by:
\begin{equation}\label{emicine}
\Delta E_K^{emission}=(\gamma_2E_2-E_2)-(\gamma_1E_1-E_1)= \Delta E-E_{ph}
\end{equation}
When the atom is at rest before emission, $\Delta E_K=\Delta E^2/2Mc^2$: by definition, this is the recoil energy $E_R$.
The equation:
\begin{equation}\label{recvel}
E_R=\frac{1}{2}Mv_R^2
\end{equation}
defines the recoil velocity $v_R=\Delta E/Mc= 2v_T$.
Similarly, the variation of the atom's kinetic energy due to the absorption of a photon is given by:
\begin{equation}\label{asscine}
\Delta E_K^{absorption}=(\gamma_2E_2-E_2)-(\gamma_1E_1-E_1)= E_{ph}-\Delta E
\end{equation}
The recoil energy in the case of absorption is the same as that in the case of emission.
\section{The natural linewidth: implications\label{naturalsec}}
Each atomic transition has a natural linewidth $\Gamma$ defined as the Full Width at Half Maximum of a Lorentzian function centered at the value of the transition energy $\Delta E$ corresponding to the highest transition probability: we shall denote this value as $\Delta E_M$. The existence of a natural linewidth suggests to define another dimensionless parameter: $B_D= \Gamma/\Delta E_M$; $B_D$ defines the Doppler limit in laser cooling since $B_D\Delta E_M(1\pm B_T)\approx B_D\Delta E_M$ describes the first order Doppler effect.
\par
In laser cooling, we are particularly interested in the absorption of photons belonging to two counter--propagating laser beams. An atom with velocity parameter $B_1$ can absorb a counter--propagating photon if its transition energy $\Delta E^A$ satisfies the first of equations (\ref{ugualeass}) (the superscript $A$ stays for `anti--parallel'). The same atom can absorb a co--propagating photon if its transition energy $\Delta E^P$ satisfies the same equation (the superscript $P$ stays for `parallel').
Then:
\begin{equation}\label{viavai}
\Delta E^P= \Delta E^A\frac{1-B_1}{1+B_1}
\end{equation}
$\Delta E^P$ is always smaller than $\Delta E^A$ for $B_1\neq 0$ and equal to $\Delta E^A$ for $B_1=0$.
\section{The linear approximation\label{linear}}
In the linear approximation, the energy of the emitted photon is obtained from equations (\ref{uguale}) with the approximation $\sqrt{1-B_1^2}\approx 1-B_1^2/2$ and by keeping only terms linear in the $B$'s ($B_T$, $B_D$, $B_1$):
\begin{equation}\label{emiapp}
E_{ph}= \Delta E(1-B_T+B_1\cos\theta_1)=\Delta E(1+B_T+B_2\cos\theta_2)
\end{equation}
Similarly, the energy of the absorbed photon, in the linear approximation, is obtained from equations (\ref{ugualeass}):
\begin{equation}\label{assapp}
E_{ph}= \Delta E(1+B_T+B_1\cos \theta_1)= \Delta E(1-B_T+B_2\cos \theta_2)
\end{equation}
In the linear approximation, the energy of the emitted/absorbed photon depends only on the atom's velocity component along the direction of propagation of the emitted/absorbed photon.
\subsection{Variation of the atom's kinetic energy: absorption\label{linassappsec}}
As we have seen in equation (\ref{asscine}), the variation of the kinetic energy of the atom due to the absorption of a photon is given by:
\begin{equation}\label{var_kin}
\Delta E_{K}= E_{ph}-\Delta E
\end{equation}
Let us suppose that a photon with energy $E_{ph}$ is propagating along the negative direction of the $x$ axis. This photon can be absorbed by any atom whose velocity component $v_x=-v_1\cos\theta_1$ satisfies the first of equations (\ref{assapp}): the variation of the atom's kinetic energy is the same for all these atoms. This means that, in the calculations, we can consider only the cases $\theta_1=\pi$ and $\theta_1=0$.
Then:
\begin{equation}\label{cine+-}
\Delta E_{K}= \Delta E(B_T\pm B_1)
\end{equation}
where the minus sign corresponds to $\theta_1=\pi$ and the plus sign to $\theta_1=0$.
\par
We define a red detuned photon as a photon with energy
$E_{ph}=\Delta E_M(1+B_T)(1-\delta^*)$ where $\Delta E_M(1+B_T)$ is the maximum energy of an absorbed photon when the atom is at rest before absorption and $\delta^*$ is of the same order of magnitude of or smaller than $B_D$: then, $E_{ph}\approx\Delta E_M(1+B_T-\delta^*)$.
In the linear approximation, this photon can be absorbed in a head on collision if (first equation of (\ref{assapp})):
\begin{equation}\label{headonmp}
\Delta E^A= \Delta E_M(1-\delta^*+B_1 )
\end{equation}
The variation of the atom's kinetic energy is given by:
\begin{equation}\label{cinefine}
\Delta E_K^A= \Delta E_M(B_T-B_1)
\end{equation}
If $B_1>B_T$, the absorption decreases the atom kinetic energy: in laser cooling, this is the cooling mechanism. If $B_1<B_T$, the absorption increases the atom kinetic energy; in laser cooling, this is one of the heating processes at work.
$B_T$ is a threshold parameter.
The threshold velocity is:
\begin{equation}\label{soglia}
v_T=\frac{\Delta E_M}{2Mc}=\frac{v_R}{2}
\end{equation}
The same atom with the same velocity parameter can absorb a co--propagating red detuned photon of the sa can be absorbed by atoms flying in the same direction of the photon. In this case, the transition energy of the atom must satisfy the equation:
\begin{equation}\label{same}
\Delta E^P= \Delta E_M(1-\delta^*-B_1 )
\end{equation}
and the variation of the atom's kinetic energy due to the absorption is:
\begin{equation}\label{samecin}
\Delta E_K^P= \Delta E_M(B_T+B_1)
\end{equation}
When the atom is flying in the same direction of the photon, the variation of the kinetic energy due to the absorption of a photon is always positive. In laser cooling, this is a second heating process.
\subsection{Variation of the atom's kinetic energy: emission\label{emission}}
In the case of emission the energy conservation implies that:
\begin{equation}\label{ene_emi}
\Delta E_{K}= \Delta E- E_{ph}
\end{equation}
According to equation (\ref{emiapp}), the energy of the emitted photon, in the linear approximation, is given by:
\begin{equation}\label{emiMP}
E_{ph}= \Delta E(1-B_T+ B_1 \cos\theta_1)
\end{equation}
Correspondingly, the variation of the atom's kinetic energy is:
\begin{equation}\label{pinco}
\Delta E_K= \Delta E(B_T-B_1\cos\theta_1)
\end{equation}
We see that, also for emission, $B_T$ operates as a threshold parameter.
\par
Given an atom with a velocity parameter $B_1$, it is useful to consider the average energy of the emitted photon under the hypothesis that any direction of emission is equally probable. From (\ref{emiapp}) we get:
\begin{eqnarray}
<E_{ph}>&=& \Delta E (1-B_T) +\Delta E B_1\frac{1}{4\pi}\int_0^\pi \cos\theta_1 (2\pi\sin \theta_1d\theta_1)\nonumber\\
&=& \Delta E (1-B_T)+\frac{B_1\Delta E}{4}\left[\sin^2\theta_1\right]_0^\pi=\Delta E (1-B_T)\label{fotomedia}
\end{eqnarray}
i.e., in the linear approximation, the average energy of the emitted photon is equal to that of the photon emitted when the emitting atom is at rest before emission.
\par
The average variation of the atom's kinetic energy due to the emission of a photon with the same probability along any direction is given by:
\begin{equation}\label{emimedia}
<\Delta E_K>=\Delta E- <E_{ph}>=B_T\Delta E\approx B_T\Delta E_M=E_R
\end{equation}
i.e., the average variation of the atom's kinetic energy due to the emission of a photon along an arbitrary direction is positive, independent from $B_1$, and equal to the recoil energy (in the linear approximation). This is a third heating mechanism in laser cooling.
\par
We shall now consider a cycle composed by the absorption of a photon followed by a spontaneous emission along an arbitrary direction, taking into account the complications due to the linewidth. If an atom absorbs a photon in a head on collision, its transition energy is $\Delta E^A$ and the variation of its kinetic energy is $\Delta E_M(B_T-B_1)$ (equation (\ref{cinefine})). On the other hand, the average variation of its kinetic energy due to the emission of a fluorescence photon is simply $E_R$. Then, the average overall variation of its kinetic energy due to the cycle considered is given by:
\begin{equation}\label{somma}
<\Delta E_K^A>= \Delta E_M(B_T-B_1)+E_R=\Delta E_M(2B_T-B_1)
\end{equation}
This variation is negative for $B_1>2B_T$, null for $B_1=2B_T$ and positive for $B_1<2B_T$, where $B_1$ is the velocity parameter of the atom before absorption.
\par
Similarly, if an atom flying in the same direction of the photon absorbs it and subsequently undergoes a spontaneous emission, the average variation of its kinetic energy is given by (equations \ref{samecin}, \ref{emimedia}):
\begin{equation}\label{somma2}
<\Delta E_K^{P}>= \Delta E_M(2B_T+B_1)
\end{equation}
These two last equations contain all the information necessary for a quantitative treatment of laser cooling: taken together, along with the different transition probabilities for $\Delta E^A$ and $\Delta E^P$, they describe all the cooling and heating processes, under the conditions specified in next section.
\section{Laser cooling of two levels atoms\label{laser}}
Before proceeding, it is worth recallig what are the cooling and heating mechanisms at work. If we assume that the laser photons are propagating along the negative direction of the $x$ axis, then:
\begin{itemize}
\item if the atom's velocity component $v_x>v_R/2$, the absorption of a photon decreases the atom's kinetic energy (cooling mechanism);
\item if the atom's velocity component $v_x<v_R/2$, the absorption of a photon increases the atom's kinetic energy (heating mechanism);
\item the emission of a fluorescence photon increases, on the average, the atom's kinetic energy by an amount equal to the recoil energy $E_R$ (heating mechanism).
\end{itemize}
We shall assume that: the laser photons are mono--energetic; the laser intensity is low enough so that stimulated emission is negligible; only the first fluorescence cycle is relevant (this means that the probability that a photon emitted by an atom is absorbed by another atom is negligible).
\par
We suppose that the sample of atoms is illuminated by two opposite laser beams of red detuned photons for each direction axis. For symmetry reasons, we can deal only with the two beams propagating, say, along the $x$ direction.
As stated before, the energy of the red detuned photon is written as:
\begin{equation}\label{Eph}
E_{ph}=\Delta E_M(1+B_T)(1-\delta^*)\approx \Delta E_M(1+B_T -\delta^*)
\end{equation}
If an atom with velocity parameter $B_1$ absorbs a laser photon, this photon belongs to one of the two opposite beams. If the photon belongs to the beam flying against the atom, the transition energy $\Delta E^A$ satisfies
equation (\ref{headonmp}) (in the linear approximation):
\begin{equation}\label{headonmp2}
\Delta E^A=\Delta E_M(1-\delta^*) +B_1 \Delta E_M
\end{equation}
The smallest possible value of $\Delta E^A$ is $\Delta E^A=\Delta E_M(1-\delta^*)$, corresponding to $B_1=0$.
Instead, if the photon belongs to the beam flying in the same direction of the atom, the transition energy $\Delta E^P$ satisfies
the equation:
\begin{equation}\label{viavai2}
\Delta E^{P}=\Delta E_M(1-\delta^*)-B_1 \Delta E_M=\Delta E^A(1-2B_1)
\end{equation}
The maximum possible value of $\Delta E^P$ is $\Delta E^P=\Delta E_M(1-\delta^*)$, corresponding to $B_1=0$ and equal to the minimum value of $\Delta E^A$.
\par
The transition probabilities $P^A$ and $P^P$ for $\Delta E^A$ and $\Delta E^P$ are different, and given by the corresponding values of the normalized Lorentzian function describing the natural line.
The average variation of the kinetic energy of the atom due to the absorption of a photon and the subsequent emission of a fluorescence one, weighed by the relative probabilityted, is calculated by using the basic equations (\ref{somma}, \ref{somma2}) that we reproduce here for convenience:
\begin{equation}\label{s1}
<\Delta E_K^{A}> = \Delta E_M(2B_T-B_1)P^{A}
\end{equation}
for a photon flying against the atom, and
\begin{equation}\label{s2}
<\Delta E_K^{P}> = \Delta E_M(2B_T+B_1)P^P
\end{equation}
for a photon flying in the same direction of the atom.
Therefore, the average variation of the kinetic energy of the atom is obtained by summing the two equations (\ref{s1}, \ref{s2}) member by member:
\begin{equation}\label{sommatotale}
<\Delta E_K>=\Delta E_M[2B_T(P^{A}+P^{P})-B_1(P^{A}-P^{P})]
\end{equation}
Putting:
\begin{equation}\label{put}
Q(\delta^*)=2B_T\frac{P^A+P^P}{P^A-P^P}
\end{equation}
$<\Delta E_K>$ will be negative if $B_1>Q(\delta^*)$, null if $B_1=Q(\delta^*)$ and positive if $B_1<Q(\delta^*)$.
The condition $B_1=Q(\delta^*)$ yields:
\begin{equation}\label{bello}
B_1=2B_T\frac{P^A+P^P}{P^A-P^P}
\end{equation}
i.e.
\begin{equation}\label{first}
B_1=2B_T \frac{L[\Delta E_M(1-\delta^*+B_1)]+L[\Delta E_M(1-\delta^*-B_1)]}{L[\Delta E_M(1-\delta^*+B_1)]-L[\Delta E_M(1-\delta^*-B_1)]}
\end{equation}
where the $L$ is the normalized Lorentzian function describing the natural line shape. After some manipulations we get:
\begin{equation}\label{second}
B_1^2=\frac{B_T(B_D^2+4\delta^{*^2})}{4(\delta^*-B_T)}
\end{equation}
with $\delta^*>B_T$. This condition means that the energy of the red detuned photon $E_{ph}=\Delta E_M(1+B_T-\delta^*)$ must be smaller than $\Delta E_M$. Then:
\begin{equation}\label{detuning}
{B_1}=\frac{1}{2}\sqrt{B_T \frac{B_D^2+4\delta^{*^2}}{\delta^* -B_T}}
\end{equation}
The laser beams will reduce the average kinetic energy of the atoms with velocity parameter larger than $B_1$ and will increase the average kinetic energy of the atoms with velocity parameter smaller than $B_1$.
The smallest value $B_{1min}$ of $B_1$ is obtained for
\begin{equation}\label{delta*}
\delta^*=B_T+\frac{1}{2}\sqrt{4B_T^2+B_D^2}
\end{equation}
and
is given by:
\begin{equation}\label{bminimo}
{B_{1min}}={B_T}\sqrt{2+2\sqrt{1+\frac{B_D^2}{4B_T^2}}}
\end{equation}
These results are valid for any value of the ratio $B_D/B_T=\Gamma/E_R$.
It is interesting to consider three limits: $B_D\gg B_T$, $B_D=B_T$ and $B_D\ll B_T$. In the first case, $B_{1mim}$ is obtained for $\delta^*\approx B_D$; in the second, for $\delta^*=2.41\, B_T$; in the third for $\delta^* \approx 2B_T$.
\par
The absorption of a laser photon depends only on the atom's velocity component $v_x$.
Since the statistical distribution of $v_x$ is symmetric around $v_x=0$, the atoms' sample in the steady state can be described as if all the atoms have velocity component $v_x=\pm v_{x_{rms}}$. When this velocity is equal to $\pm cB_1$, the average kinetic energy of the atoms can not be reduced further. Then, by putting $cB_1=v_{x_{rms}}$, we get:
\begin{equation}\label{rms}
<E_K>=\frac{1}{2} M<v_x^2>=\frac{1}{2}Mc^2B_1^2=\frac{1}{16}E_R\frac{B_D^2+{4\delta^*}^2}{B_T(\delta^*-B_T)}
\end{equation}
The lowest attainable kinetic energy will be:
\begin{equation}\label{lowest}
<E_K>_{min}=\frac{1}{2}Mc^2B_{1min}^2=\frac{1}{2}E_R \left(1+\sqrt{1+\frac{B_D^2}{4B_T^2}} \right)
\end{equation}
This equation implies that $<E_K>_{min}\rightarrow E_R$ as $B_D/B_T\rightarrow 0$, i.e as the natural linewidth $\Gamma\rightarrow 0$ (Fig. \ref{minimo}).
\begin{figure}[h]
\centering{
\includegraphics[width=10cm]{fig_2.eps}\\
}
\caption{The lowest average kinetic energy goes to $E_R$ as the linewidth $\Gamma$ goes to zero.}\label{minimo}
\end{figure}
\par\noindent
When $\Gamma =0$, red detuned photons can be absorbed only by atoms flying against the photons and the minimum condition $<E_K>_{min}= E_R$ can be derived directly from equation (\ref{somma}). In fact, according to (\ref{somma}), the average variation of the atom's kinetic energy due to the absorption of a photon and the subsequent emission of a fluorescence one is zero for $B_1=2B_T$. Then, the average kinetic energy of an atom in the steady state condition in given by:
\begin{equation}\label{zerogamma}
<E_K>=\frac{1}{2}Mc^2B_1^2=\frac{1}{2}Mc^2(2B_T)^2=E_R
\end{equation}
\subsection{Comparison with the semiclassical theory}
From equations (\ref{rms}, \ref{lowest}), we get, dividing member by member:
\begin{equation}\label{ratio}
\frac{<E_K>}{<E_K>_{min}}=\frac{1}{8}\frac{B_D^2+{4\delta^*}^2}{(\delta^*-B_T)(B_T+\frac{1}{2}\sqrt{4B_T^2+B_D^2})}
\end{equation}
If $B_T=0$, this equation reduces to:
\begin{equation}\label{semi}
\frac{<E_K>}{<E_K>_{min}} =\frac{1}{2}\left(\frac{B_D}{2\delta^*}+\frac{2\delta^*}{B_D}\right)
\end{equation}
In literature, the detuning parameter $\delta$ is defined as $\delta=\omega_l-\omega_a$ where $\omega_l$ and $\omega_a$ are the laser and the atomic transition frequencies. Therefore: $\delta^*=\hbar|\delta|/\Delta E_M$.
Taking into account this relation and that $\Gamma=\hbar\gamma$ (where $\gamma$ is the natural width expressed in terms of angular frequency), equation (\ref{semi}) assumes the form:
\begin{equation}\label{classica}
\frac{<E_K>}{<E_K>_{min}}=\frac{1}{2}\left(\frac{\gamma}{2|\delta|} + \frac{2|\delta|}{\gamma}\right)
\end{equation}
This is the result of the semiclassical theory of two levels atoms (at low laser intensities): it is a limiting case of the present treatment based on conservation laws.
\par
As a matter of fact, if we approximate equations (\ref{rms}) and (\ref{lowest}) for $\delta^*\approx B_D\gg B_T$, we obtain:
\begin{equation}\label{cinappcl}
<E_k>=\frac{\Gamma}{8}\left( \frac{B_D}{2\delta^*} + \frac{2\delta^*}{B_D}\right)=\frac{\hbar\gamma}{8}\left( \frac{\gamma}{2|\delta|} +\frac{2|\delta|}{\gamma} \right)
\end{equation}
and
\begin{equation}\label{cinappclm}
<E_K>_{min}=\frac{\hbar\gamma}{4}
\end{equation}
respectively. We recognize in equation (\ref{cinappcl}) the average kinetic energy of the atom and in equation (\ref{cinappclm}) the lowest attainable average kinetic energy predicted by the semiclassical theory.
\par
The comparison with the semiclassical theory can be visualized by three figures.
Fig. \ref{sodio} shows that the two treatments are practically indistinguishable when $B_D\gg B_T$.
\begin{figure}[h]
\centering{
\includegraphics[width=10cm]{fig_3.eps}\\
}
\caption{Comparison between the semiclassical theory (dotted line) and the present treatment for the case of the $589.16$ nm transition of $^{23} Na$. }\label{sodio}
\end{figure}
\par\noindent
In the semiclassical theory, it is explicitly assumed that the photon momentum is negligible with respect to the atomic one. This approximation implies that the kinetic energy of the atom is much larger than the recoil energy. Consequently, in the semiclassical theory both the average kinetic energy and the lowest attainable average kinetic energy depend only on $\Gamma$ (on $B_D$ in the language of this paper): equations (\ref{cinappcl}) and (\ref{cinappclm}).
Instead, in the present treatment, they depend on both $B_D$ and $B_T$ (equations (\ref{rms}, \ref{lowest})).
However, when $B_D\approx B_T$ the differences between the two treatments are evident (fig. \ref{elio}).
\begin{figure}[h]
\centering{
\includegraphics[width=10cm]{fig_4.eps}\\
}
\caption{Comparison between the semiclassical theory (dotted line) and the present treatment for the case of the $388.98$ nm transition of $^{4} He^*$. The $^*$ remind us that the starting level of the transition is a metastable one. }\label{elio}
\end{figure}
\par\noindent
Finally, the lowest attainable kinetic energy is systematically lower in the semiclassical theory (Fig. \ref{tmtd}):
the ratio between the value predicted by the semiclassical theory and the one obtained by conservation laws tends to one for large values of the ratio $B_D/B_T$ but drops
dramatically as $B_D/B_T\rightarrow 0$. Of course, this behavior is due to the fact that $B_T$ does not enter into the semiclassical theory.
\begin{figure}[h]
\centering{
\includegraphics[width=10cm]{fig_5.eps}\\
}
\caption{Ratio between the lowest values of the average kinetic energy predicted by the semiclassical theory $<E_K>_{min}^{SC}$ and the conservation laws $<E_K>_{min}^{CL}$.} \label{tmtd}
\end{figure}
\section{Comparison with quantum mechanical treatments\label{qm}}
Quantum mechanical treatments of laser cooling of two levels atoms at low laser intensities have been carried out by several authors. Wineland and Itano \cite{wineland} deal with both free and bound atoms and begin with formulas based on conservation laws applied to the absorption/emission of photons by atoms. The difference with the present treatment lies in the fact that Wineland and Itano's calculation of the steady state average kinetic energy requires the knowledge of the atoms' velocity distribution, assumed to be always Gaussian. For free atoms, when the linewidth is much smaller than the recoil energy, they found that it should be possible to achieve average kinetic energies lower than the recoil energy. However, Wineland and Itano stress that it is difficult to obtain ``the proper conditions under
which these results hold \cite[p. 1525]{wineland}.''
Instead, in the present paper, the calculation of the average kinetic energy is based on the position $cB_1=v_{x_{rms}}$ where $cB_1$ is the atom's velocity which zeroes the average kinetic energy variation due to an absorption--emission cycle: no knowledge of the velocity distribution is required. As a consequence, as shown above, when $\Gamma\rightarrow 0$ $<E_K>\rightarrow E_R$: $E_R$ is the lowest kinetic energy allowed by conservation principles in laser cooling of two levels atoms at low laser intensities.
In the second part of their paper, Wineland and Itano show that the quantum mechanical treatment of free atoms does not alter the picture given in the first part of their paper.
\par
The quantum mechanical treatment by Castin, Wallis and Dalibard does not make any assumption on the atoms' velocity distribution and is valid for any value of the ratio between the natural linewidth and the recoil energy. As in the present paper, they show that the semiclassical theory is valid as long as the natural linewidth is much larger than the recoil energy; however, differently from the present paper, they found that, for very narrow transition lines, the lowest attainable average kinetic energy is about $0.5 E_R$ instead of $E_R$. It is not clear to me why a quantum mechanical treatment yields this limit value which seems to be incompatible with the conservation laws.
\section{Conclusions}
The straightforward application of energy and linear momentum conservation to the absorption/emission of photons by atoms allows to find out the essential features of laser cooling of two levels atoms at low laser intensities.
The lowest attainable average kinetic energy of the atoms depends on the ratio $\Gamma/E_R$ between the natural linewidth and the recoil energy and tends to $E_R$ as $\Gamma/E_R$ tends to zero. This treatment, like the quantum mechanical ones, is valid for any value of the ratio between the natural linewidth and the recoil energy and contains the results of the standard semiclassical theory of laser cooling as the limiting case in which the recoil energy is negligible with respect to the natural linewidth.
\renewcommand{\theequation}{A-\arabic{equation}}
\setcounter{equation}{0}
\section*{Appendix: Absorption of a photon in Newtonian mechanics \label{newton}}
Suppose that an atom, flying in the positive direction of the $x$ axis, absorbs a counter--propagating photon with energy $E_{ph}$.
The conservation of linear momentum reads:
\begin{equation}\label{momentum}
Mv_1-\frac{E_{ph}}{c}=Mv_2
\end{equation}
where $v_1$ and $v_2$ are the component of the atom velocity along the $x$ axis, before and after the absorption, respectively. Notice how the light speed $c$, extraneous to classical mechanics, enters this equation through the photon linear momentum.
The conservation of energy reads:
\begin{equation}\label{energy}
\frac{1}{2}Mv_1^2+ E_{ph}=\frac{1}{2}Mv_2^2+\Delta E
\end{equation}
where $\Delta E$ is the energy difference between the two levels of the atomic transition.
From (\ref{momentum}):
\begin{equation}\label{velocity}
v_2=v_1-\frac{E_{ph}}{Mc}
\end{equation}
Substituting this value in (\ref{energy}), we get:
\begin{equation}\label{energy2}
E_{ph}= \Delta E+\frac{1}{2Mc^2}E_{ph}^2-B_1E_{ph}
\end{equation}
where $B_1=v_1/c$.
If we write:
\begin{equation}\label{position}
E_{ph}=\Delta E(1+\alpha)
\end{equation}
equation (\ref{energy2}) becomes:
\begin{equation}\label{energy3}
\Delta E(1+\alpha)=\Delta E +\frac{\Delta E}{2Mc^2}\Delta E(1+\alpha)^2 -B_1\Delta E(1+\alpha)
\end{equation}
Putting $B_T=\Delta E/2Mc^2$ and retaining only terms of the first degree in $\alpha, B_T, B_1$ (linear approximation) we obtain:
\begin{equation}\label{finale}
\alpha= B_T-B_1
\end{equation}
Therefore:
\begin{equation}\label{finale2}
E_{ph}=\Delta E(1+B_T-B_1)
\end{equation}
which is identical to the first of equations (\ref{assapp}) in the case of $\theta_1=\pi$.
From (\ref{energy}) and (\ref{finale2}), the variation of the atom's kinetic energy is obtained:
\begin{equation}\label{cinetica}
\Delta E_K= E_{ph}-\Delta E=\Delta E(B_T-B_1)
\end{equation}
If $B_1=0$, $E_{ph}=\Delta E(1+B_T)$ and $\Delta E_K=\Delta E^2/2Mc^2=E_R$. $E_R$ is the recoil energy.
\par
The rest energy of special relativity $Mc^2$ appears in these equations as a consequence of having considered the absorption of a photon. Of course, in classical mechanics, the physical meaning of $Mc^2$ remains obscure.
\par
The energy of an emitted can be derived in a similar way, by writing down the adequately re--written conservation equations.
\vskip5mm\par\noindent
{\bf Acknowledgments.}
Thanks are due to Biagio Buonaura, for having sympathetically and carefully followed the development of this paper from the beginning; to Nicole Fabbri, for her critical reading of a first draft of the paper and for her encouragement; and to Ennio Arimondo, whose stringent comments have been a kind of driving force for the attainment of the final version of the paper.
\vskip5mm
|
\section{Introduction}
Many problems of classical Euclidean geometry explicitly or
implicitly consider objects up to similarity. Understanding and
using similarity is an important geometry competence feature for
schoolchildren.
Recall that two geometric figures $A$ and $B$ are similar if $B$
can be obtained from $A$ after a finite composition of
translations, rotations, reflections and dilations (homotheties).
Similarity is an equivalence relation and thus, for example, the
set of all triangles in a plane is partitioned into similarity
equivalence classes which can be identifies with \sl similarity
types\rm\ of triangles.
In many areas of mathematics objects are studied up to equivalence
relations. Depending on situation and traditions this is done
explicitly, implicitly or inadvertently. The problem of finding a
distinguished (\sl canonical, normal\rm) representatives of
equivalence classes of objects ir posed. Alternatively, it is the
problem of mapping the quotient set injectively back to the
original set. Let $X$ be a set with an equivalence relation $\sim$
or, equivalently, $R\subseteq X\times X$, denote the equivalence
class of $x\in X$ by $[x]$. Let $\pi:X\rightarrow X/R$ such that
$\pi(x)=[x]$ be the canonical projection map. We call a map
$\sigma: X/R \rightarrow X$ \sl normal object map\rm\ provided 1)
$\sigma$ is injective and 2) $\sigma\circ \pi=id_{X}$. For
example, there are various normal forms of matrices, such as the
Jordan normal form. See \cite{Sh} for examples of normal forms in
algebra and \cite{P} for a related recent work. Normal objects are
designed for educational, pure research (e.g. for classification)
and applied reasons. Normal objects are constructed as objects of
simple, minimalistic design, to show essential properties and
parameters of original objects. Often it is easier to solve a
problem for normal objects first and extend the solution to
atbitrary objects afterwards. Normal objects which are initially
designed for educational, pure research or problem solving
purposes are also used to optimize computations.
In elementary Euclidean geometry normal map approach does not seem
to be popular working with simple discrete objects such as
triangles. This may be related to the traditional dominance of the
synthetic geometry in school mathematics at the expense of the
coordinate/analytic approach. We can pose the problem of
introducing and using normal forms of triangles up to similarity.
This means to describe a set $S$ of mutually non-similar triangles
such that any triangle in the plane would be similar to a triangle
in $S$. We assume that Cartesian coordinates are introduced in the
plane, $S$ is designed using the Cartesian coordinates. For
triangle we offer three normal forms based on side lengths. Using
these normal forms the set of triangle similarity forms is
bijectively mapped to a fixed plane domain bounded by lines and
circles. For these forms two vertices are fixed and the third
vertex belongs to this finite domain, we call them \sl the one
vertex normal forms.\rm\ One vertex normal forms are also
generalized for quadrilaterals. Another normal form for triangles
is based on angles and circumscribed circles. For this form one
constant vertex is fixed on the unit circle and two other variable
vertices also belong to the unit circle, we call this form \sl the
circle normal form.\rm\
These normal forms may be useful in solving geometry problems
involving similarity and teaching geometry. The paper may be
useful for mathematics educators.
\section{Main results}
\subsection{Normal forms of triangles}
\subsubsection{Notations}
Consider $\mathbb{R}^{2}$ with a Cartesian system of coordinates
$(x,y)$ and center $O$. We think of classical triangles as being
encoded by their vertices. Strictly speaking by the triangle
$\triangle XYZ$ we mean the multiset $\{\{X,Y,Z\}\}$ of three
points in $\mathbb{R}^2$ each point having multiplicity at most
$2$. A triangle is called degenerate if points lie on a line.
Given $\triangle ABC$ we denote $\angle BAC=\alpha$, $\angle
ABC=\beta$, $\angle ACB=\gamma$, $|BC|=a$, $|AC|=b$, $|AB|=b$. We
exclude mutisets of type $\{\{XXX\}\}$.
We will use the following affine transformations of
$\mathbb{R}^{2}$: 1) translations, 2) rotations, 3) reflections
with respect to an axis, 4) dilations (given by the rule
$(x,y)\rightarrow (cx,cy)$ for some $c\in \mathbb{R}\backslash
\{0\}$). It is known that these transformations generate the \sl
dilation group\rm\ of $\mathbb{R}^{2}$, denoted by some authors as
$IG(2)$, see \cite{H}, \cite{P}. Two triangles $T_{1}$ and $T_{2}$
as similar if there exists $g\in IG(2)$, such that
$g(T_{1})=T_{2}$ (as multisets). If triangles $T_{1}$ and $T_{2}$
are similar, we write $T_{1}\sim T_{2}$ or $\triangle
X_{1}Y_{1}Z_{1}\sim \triangle X_{2}Y_{2}Z_{2}$.
The point $(x_{1},y_{1})$ is lexicographically smaller than the
point $(x_{2},y_{2})$ and denoted as $(x_{1},y_{1})\prec
(x_{2},y_{2})$ provided ($x_{1}<x_{2}$) or ($x_{1}=x_{2}$ and
$y_{1}<y_{2}$). The lexicographical order of points can be
extended to lexicographical ordering of sequences of points: the
sequence of points $[p_{1},p_{2}]$ is lexicographically smaller
than the sequence $[q_{1},q_{2}]$ denoted by $[p_{1},p_{2}]\prec
[q_{1},q_{2}]$ provided ($p_{1}\prec q_{1}$) or ($p_{1}=q_{1}$ and
$p_{2}\prec q_{2}$).
We use normal letters to denote fixed objects and $\backslash
mathcal$ letters to denote objects as function values.
\subsubsection{The $C$-vertex normal form}\label{1}
A normal form can be obtained by transforming the longest side of
the triangle into a unit interval of the $x$-axis. We call it \sl
the $C$-normal form.\rm\
In this subsection $A=(0,0)$ and $B=(1,0)$.
\begin{definition}
Let $S_{C}\subseteq \mathbb{R}^{2}$ be the domain in the first
quadrant bounded by the lines $y=0$, $x=\frac{1}{2}$ and the
circle $x^2+y^2=1$, see Figure 1.
\begin{center}
\epsfysize=70mm
\epsfbox{dom_C1_math.eps}
Fig.1. - the domain $S_{C}$.
\end{center}
In other terms, $S_{C}$ is the set of solutions of the system of
inequalities
$$
\left\
\begin{array}{ll}
y\ge 0 \\
x\ge \frac{1}{2}\\
x^2+y^2\le 1. \\
\end{array
\right.
$$
\end{definition}
\begin{theorem} Every triangle $UVW$ (including degenerate triangles)
in $\mathbb{R}^{2}$ is similar to a triangle $AB\mathcal{C}$,
where $A=(0,0)$, $B=(1,0)$ and $\mathcal{C}\in S_{C}$.
\end{theorem}
\begin{proof} Let $\triangle UVW$ has side lengths $a,b,c$
satisfying $a\le b\le c$. Perform the following sequence of
transformations:
\begin{enumerate}
\item translate and rotate the triangle so that the longest side
is on the $x$-axis, one vertex has coordinates $(0,0)$ and another
vertex has coordinates $(c,0)$, $c>0$;
\item if the third vertex has negative $y$-coordinate, reflect the
triangle with respect to the $x$-axis;
\item do the dilation with coefficient $\frac{1}{c}$, note that
the vertices on the $x$-axis have coordinates $(0,0)$ and $(1,0)$,
the third vertex has coordinates $(x'_{C},y'_{C})$, where
$x'^2_{C}+y'^2_{C}\le 1$ and $(x'_{C}-1)^2+y'^2_{C}\le 1$;
\item if $x'_{C}<\frac{1}{2}$, then reflect the triangle with
respect to the line $x=\frac{1}{2}$, denote the third vertex by
$\mathcal{C}=(x_{C},y_{C})$, by construction we have that
$\mathcal{C}\in S_{C}$.
\end{enumerate}
The image of the initial triangle $\triangle UVW$ is the triangle
$AB\mathcal{C}$, where $\mathcal{C}\in S_{C}$. All transformations
preserve similarity type therefore $\triangle UVW\sim \triangle
AB\mathcal{C}$.
\end{proof}
\begin{theorem} If $C_{1}\in S_{C}$, $C_{2}\in S_{C}$ and $C_{1}\neq C_{2}$, then $\triangle ABC_{1}\not\sim\triangle ABC_{2}$.
\end{theorem}
\begin{proof} If $\angle C_{1}AB=\angle C_{2}AB$ and $C_{1}\neq C_{2}$, then $\angle C_{1}BA\neq \angle
C_{2}BA$. By equality of angles for similar triangles it follows
that $\triangle ABC_{1}\not\simeq\triangle ABC_{2}$.
Let $\angle C_{1}AB\neq \angle C_{2}AB$. The angle $C_{i}AB$ is
the smallest angle in $\triangle ABC_{i}$. By equality of angles
for similar triangles it again follows that $\triangle
ABC_{1}\not\simeq\triangle ABC_{2}$.
\end{proof}
\begin{definition} A point $\mathcal{C}\in S_{C}$ such that $\triangle AB\mathcal{C}\sim \triangle
UVW$ is called \sl the $C$-normal point\rm\ of $\triangle UVW$.
\end{definition}
\begin{definition} The $C$-vertex normal form of $\triangle UVW$ is
$\triangle AB\mathcal{C}$, where $\mathcal{C}\in S_{C}$ is the
$C$-normal point of $\triangle UVW$.
\end{definition}
\begin{remark} Denote by $R_{C}$ the intersection of the circle
$(x-\frac{1}{2})^2+y^2=(\frac{1}{2})^2$ and $S_{C}$. Points of
$R_{C}$ correspond to right angle triangles. Points below and
above $R_{C}$ correspond to, respectively, obtuse and acute
triangles, see Figure 2.
Points on the intersection of the line $x=\frac{1}{2}$ and $S_{C}$
correspond to isosceles obtuse triangles. Points on the
intersection of the circle $x^2+y^2=1$ and $S_{C}$ correspond to
isosceles acute triangles. Points in the interior of $S_{C}$
correspond to scalene triangles. The point
$(\frac{1}{2},\frac{\sqrt{3}}{2})$ corresponds to the equilateral
triangle. Points on the intersection of the line $y=0$ and $S_{C}$
correspond to degenerate triangles. $\mathcal{C}=B$ for triangles
having side lengths $0,c,c$.
\end{remark}
\begin{remark} A similar normal form can be obtained reflecting
$S_{C}$ with respect to the line $x=\frac{1}{2}$.
\end{remark}
\begin{center}
\epsfysize=60mm
\epsfbox{dom_C2_math.eps}
Fig.2. - the subdomains of $S_{C}$ corresponding to obtuse and
acute triangles.
\end{center}
\subsubsection{The $B$-vertex normal form}
Another normal form can be obtained by transforming the median
length side (in the sense of ordering) of the triangle into a unit
interval of the $x$-axis. By analogy it is called \sl the
$B$-normal form.\rm\
In this subsection $A=(0,0)$ and $C=(1,0)$.
\begin{definition}
Let $S_{B}\subseteq \mathbb{R}^{2}$ be the domain in the first
quadrant bounded by the line $y=0$ and the circles $x^2+y^2=1$ and
$(x-1)^2+y^2=1$, see Figure 3.
\begin{center}
\epsfysize=60mm
\epsfbox{dom_B1_math.eps}
Fig.3. - the domain $S_{B}$.
\end{center}
In other terms, $S_{B}$ is the set of solutions of the system of
inequalities
$$
\left\
\begin{array}{ll}
y\ge 0 \\
x^2+y^2\ge 1\\
(x-1)^2+y^2\le 1.\\
\end{array
\right.
$$
\end{definition}
\begin{theorem} Every triangle $UVW$ (including degenerate triangles)
in $\mathbb{R}^{2}$ is similar to a triangle $A\mathcal{B}C$,
where $A=(0,0)$, $C=(1,0)$ and $\mathcal{B}\in S_{B}$.
\end{theorem}
\begin{proof} Let $\triangle UVW$ has side lengths $a,b,c$
satisfying $a\le b\le c$. Perform the following sequence of
transformations:
\begin{enumerate}
\item translate and rotate the triangle so that the side of length
$b$ is on the $x$-axis, one vertex has coordinates $(b,0)$ and
another vertex has coordinates $(b,0)$, $b>0$, the side of length
$c$ is incident to the vertex $(c,0)$;
\item if the third vertex has negative $y$-coordinate, reflect the
triangle with respect to $x$-axis;
\item do the dilation with coefficient $\frac{1}{b}$, note that
the vertices on the $x$-axis have coordinates $(0,0)$ and $(1,0)$,
at this point the third vertex $\mathcal{B}$ has coordinates
$(x'_{B},y'_{B})$, where $y'_{B}\ge 0$, $x'^2_{B}+y'^2_{B}\ge 1$
or $(x'_{B}-1)^2+y'^2_{B}\le 1$;
\item if $x'_{B}<\frac{1}{2}$, reflect the triangle with respect
to the line $x=\frac{1}{2}$, now the third vertex $\mathcal{B}$
has new coordinates $(x_{B},y_{B})$, where $x_{B}\ge \frac{1}{2}$,
$y_{B}\ge 0$, $(x_{B}-1)^2+y^2_{B}\le 1$.
\end{enumerate}
The image of the initial triangle $\triangle UVW$ is the triangle
$A\mathcal{B}C$, where $\mathcal{B}\in S_{B}$. All transformations
preserve similarity type therefore $\triangle UVW\sim \triangle
A\mathcal{B}C$.
\end{proof}
\begin{theorem} If $B_{1}=(x_{i},y_{i})\in S_{B}$, $B_{2}=(x_{2},y_{2})\in S_{B}$ and $B_{1}\neq B_{2}$, then $\triangle
AB_{1}C\not\sim\triangle AB_{2}C$.
\end{theorem}
\begin{proof} The angle $\angle B_{i}AC$ is
the smallest angle in the triangle $\triangle AB_{i}C$.
If $\angle B_{1}AC\neq \angle B_{2}AC$, then, since these are
smallest angles in the triangles, it follows that $\triangle
AB_{1}C\not\sim \triangle AB_{2}C$.
If $\angle B_{1}AC=\angle B_{2}AC$ and $B_{1}\neq B_{2}$, then
$\angle AB_{1}C\neq \angle AB_{2}C$. $\angle AB_{i}C$ is the
biggest angle in $\triangle AB_{i}C$, therefore $\angle
AB_{1}C\neq \angle AB_{2}C$ implies $\triangle AB_{1}C\not\sim
AB_{2}C$.
\end{proof}
\begin{definition} A point $\mathcal{B}\in S$ such that $\triangle A\mathcal{B}C\sim \triangle
UVW$ is called \sl the $B$-normal point\rm\ of $\triangle UVW$.
\end{definition}
\begin{definition} The $B$-vertex normal form of $\triangle UVW$ is
$\triangle A\mathcal{B}C$, where $\mathcal{B}\in S$ is the
$B$-normal point of $\triangle UVW$.
\end{definition}
\begin{remark} Denote by $R_{B}$ the intersection of the ray
$x=1$, $x\ge 0$ and $S_{B}$. Points of $R_{B}$ correspond to right
angle triangles. Points to the right and left of $R_{B}$
correspond to, respectively, obtuse and acute triangles, see
Figure 4.
\begin{center}
\epsfysize=60mm
\epsfbox{dom_B2_math.eps}
Fig.4. - the subdomains of $S_{B}$ corresponding to obtuse and
acute triangles.
\end{center}
Points on the intersection of the line $x^2+y^2=1$ and $S_{B}$
correspond to isosceles acute triangles. Points on the
intersection of the circle $(x-1)^2+y^2=1$ and $S_{B}$ correspond
to isosceles obtuse triangles. Points in the interior of $S_{B}$
correspond to scalene triangles. The point
$(\frac{1}{2},\frac{\sqrt{3}}{2})$ corresponds to the equilateral
triangle. Points on the intersection of the line $y=0$ and $S_{B}$
correspond to degenerate triangles. $\mathcal{B}=C$ for triangles
having side lengths $0,c,c$.
\end{remark}
\subsubsection{$A$-vertex normal form}
Finally a normal form can be obtained by transforming the shortest
side of the triangle into a unit interval of the $x$-axis. By
analogy it is called \sl the $A$-normal form.\rm\ In this case
again two vertices on the $x$-axis are $(0,0)$ and $(1,0)$, the
domain $S_{A}$ of possible positions of the third vertex is
unbounded.
In this subsection $B=(0,0)$ and $C=(1,0)$.
\begin{definition}
Let $S_{A}\subseteq \mathbb{R}^{2}$ be the unbounded domain in the
first quadrant bounded by the lines $y=0$, $x=\frac{1}{2}$ and the
circle $(x-1)^2+y^2=1$, see Figure 5.
In other terms, $S_{A}$ is the set of solutions of the system of
inequalities
$$
\left\
\begin{array}{ll}
y\ge 0 \\
x\ge \frac{1}{2}\\
(x-1)^2+y^2\ge 1.\\
\end{array
\right.
$$
\end{definition}
\begin{center}
\epsfysize=60mm
\epsfbox{dom_A2_math.eps}
Fig.5. - the domain $S_{A}$.
\end{center}
\begin{theorem} Every triangle $UVW$ (including degenerate
triangles but excluding the similarity type having side lengths
$0,c,c$) in $\mathbb{R}^{2}$ is similar to a triangle
$\mathcal{A}BC$, where $B=(0,0)$, $C=(1,0)$ and $\mathcal{A}\in
S_{A}$.
\end{theorem}
\begin{proof} Let $\triangle UVW$ has side lengths $a,b,c$
satisfying $a\le b\le c$. Perform the following sequence of
transformations:
\begin{enumerate}
\item translate and rotate the triangle so that the side of length
$a$ is on the $x$-axis, one vertex has coordinates $(0,0)$ and
another vertex has coordinates $(a,0)$, $a>0$, the side of length
$c$ is incident to the vertex $(0,0)$;
\item if the third vertex has negative $y$-coordinate, reflect the
triangle with respect to the $x$-axis;
\item do the dilation with coefficient $\frac{1}{a}$, note that
the vertices on the $x$-axis have coordinates $(0,0)$ and $(1,0)$;
\item if the third point has the $x$-coordinate less than
$\frac{1}{2}$, reflect the triangle with respect to the line
$x=\frac{1}{2}$, now the third vertex $\mathcal{A}$ has
coordinates $(x_{0},y_{0})$, where $x_{0}\ge \frac{1}{2}$,
$y_{0}\ge 0$, $(x_{0}-1)^2+y^2_{0}\le 1$.
\end{enumerate}
The image of the initial triangle $\triangle UVW$ is the triangle
$\mathcal{A}BC$, where $\mathcal{A}\in S_{A}$. All transformations
preserve similarity type therefore $\triangle UVW\sim \triangle
\mathcal{A}BC$.
\end{proof}
\begin{theorem} Let $B=(0,0)$, $C=(1,0)$. If $A_{1}=(x_{i},y_{i})\in S_{A}$, $A_{2}=(x_{2},y_{2})\in S_{A}$ and $A_{1}\neq A_{2}$, then $\triangle
A_{1}BC\not\sim\triangle A_{2}BC$.
\end{theorem}
\begin{proof} The angle $\angle BCA_{i}$ is
the largest angle in the triangle $\triangle A_{i}BC$.
If $\angle BCA_{1}\neq \angle BCA_{2}$, then since these are
largest angles in the triangles it follows that $\triangle
A_{1}BC\not\sim \triangle A_{2}BC$.
If $\angle BCA_{1}=\angle BCA_{2}$ and $A_{1}\neq A_{2}$, then
$\angle BA_{1}C\neq \angle BA_{2}C$. $BA_{i}C$ is the smallest
angle in $\triangle A_{i}BC$, therefore $A_{1}BC\neq A_{2}BC$
implies $\triangle AB_{1}C\not\sim AB_{2}C$.
\end{proof}
\begin{definition} A point $\mathcal{A}\in S_{A}$ such that $\triangle \mathcal{A}BC\sim \triangle
UVW$ is called \sl the $A$-normal point\rm\ of $\triangle UVW$.
\end{definition}
\begin{definition} The $A$-vertex normal form of $\triangle UVW$ is
$\triangle \mathcal{A}BC$, where $\mathcal{A}\in S_{A}$ is the
$A$-normal point of $\triangle UVW$.
\end{definition}
\begin{remark} Denote by $R_{A}$ the intersection of the ray
$x=1$, $x\ge 0$ and $S_{A}$. Points of $R_{A}$ correspond to right
angle triangles. Points to the right and left of $R_{A}$
correspond to, respectively, obtuse and acute triangles, see
Figure 6.
\begin{center}
\epsfysize=60mm
\epsfbox{dom_A3_math.eps}
Fig.6. - the subdomains of $S_{A}$ corresponding to obtuse and
acute triangles.
\end{center}
Points on the intersection of the line $x=\frac{1}{2}$ and $S_{A}$
correspond to isosceles acute triangles. Points on the
intersection of the circle $(x-1)^2+y^2=1$ and $S_{A}$ correspond
to isosceles obtuse triangles. Points in the interior of $S_{A}$
correspond to scalene triangles. The point
$(\frac{1}{2},\frac{\sqrt{3}}{2})$ corresponds to the equilateral
triangle. Points on the intersection of the line $y=0$ and $S_{A}$
correspond to degenerate triangles excluding the similarity type
with side lengths $0,c,c$, which corresponds to the point at
infinity. In contrast to the $C$-vertex and $B$-vertex normal
forms triangles for the $A$-vertex normal form are not bounded.
\end{remark}
\subsubsection{The circle normal form}
Consider $\mathbb{R}^{2}$ with a Cartesian system of coordinates
$(x,y)$ and center $O$. We also consider polar coordinates
$[r,\varphi]$ introduced in the standard way: the polar angle
$\varphi$ is measured from the positive $x$-axis going
counterclockwise.
Note that angles $\alpha,\beta,\gamma$ of a nondegenerate triangle
such that $\alpha\le \beta\le \gamma$ satisfy the system of
inequalities
$$
\left\
\begin{array}{ll}
0<\alpha\le \frac{\pi}{3},\\
\alpha\le \beta\le \frac{\pi-\alpha}{2}.\\
\end{array
\right.
$$
Similarity types of nondegenerate triangles are parametrized by
one point in the domain in $(\alpha,\beta)$-plane determined by
the system
$$
\left\
\begin{array}{ll}
\alpha>0,\\
\beta\ge\alpha,\\
\beta\le\frac{\pi}{2}-\frac{\alpha}{2}.\\
\end{array
\right.
$$
See Fig.7.
\begin{center}
\epsfysize=50mm
\epsfbox{alpha_beta_3.eps}
Fig.7. - parametrization of similarity types by $(\alpha,\beta)$.
\end{center}
For the normal form described in this subsection the vertex with
the biggest angle will be fixed at $(1,0)$, to be consistent with
previous notations we define $C=(1,0)$. For this normal form only
nondegenerate triangles are considered.
In this case normal form triangles are inscribed in the unit
circle $\mathbb{U}=\{x^2+y^2=1\}$ having $C$ as one of the
vertices.
\begin{definition} A triangle $\triangle ABC$ inscribed in $\mathbb{U}$ is called \sl normal circle
triangle\rm\ if
\begin{enumerate}
\item $0\le \alpha\le \frac{\pi}{3}$,
\item $\alpha\le \beta\le \frac{\pi}{2}-\frac{\alpha}{2}$,
\item $C=(1,0)$,
\item the point $A$ is above $x$-axis,
\item the point $B$ is below $x$-axis.
\end{enumerate}
\end{definition}
\begin{remark} For a normal triangle $\triangle ABC$ we have that
$\alpha\le \beta \le \gamma$.
\end{remark}
\begin{remark} A normal triangle with angles $\alpha\le \beta\le \gamma$ can be constructed in the
following way:
\begin{enumerate}
\item choose a point $B$ below the $y$-axis with the argument
equal to $2\alpha$, where $0\le 2\alpha\le \frac{2\pi}{3}$;
\item draw the bisector of $\angle BOC$, denote the intersection
of this bisector with the arc $BC$ having angle $2\pi-2\alpha$ by
$D$;
\item find the point $\widetilde{B}$ which is symmetric to $B$
with respect to the $x$-axis;
\item choose a point $A$ in the shorter arc $\widetilde{B}D$.
\end{enumerate}
See Figure 8.
\begin{center}
\epsfysize=70mm
\epsfbox{circle.eps}
Fig.8. - construction of a normal circle triangle.
\end{center}
\end{remark}
\begin{theorem} For every nondegenerate triangle $\triangle UVW$ there exists a
normal circle triangle $\triangle \mathcal{A}\mathcal{B}C$ such
that $\triangle UVW\sim \triangle \mathcal{AB}C$.
\end{theorem}
\begin{proof} Suppose $\triangle UVW$ has angles $\alpha\le \beta\le \gamma$.
Let $\mathcal{B}\in \mathbb{U}$ be the point with polar
coordinates $[1,-2\alpha]$. Let $\mathcal{A}\in \mathbb{U}$ be the
point with polar coordinates $[1,2\beta]$. Then since $\triangle
\mathcal{AB}C$ is inscribed in $\mathbb{U}$ we have that $\angle
\mathcal{BA}C=\alpha$, $\angle \mathcal{AB}C=\beta$ and thus
$\triangle \mathcal{AB}C\sim \triangle UVW$.
\end{proof}
\begin{theorem} Let $\triangle A_{1}BC_{1}$ and $A_{2}BC_{2}$ be
two distinct normal circle triangles: $A_{1}\neq A_{2}$ or
$B_{1}\neq B_{2}$. Then $\triangle A_{1}BC_{1}\not\sim\triangle
A_{2}BC_{2}$.
\end{theorem}
\begin{proof} If $A_{1}\neq A_{2}$, then $\angle B_{1}A_{1}C\neq \angle
B_{2}A_{2}C$. The angle $B_{i}A_{i}C$ is the smallest angle of
$\triangle A_{i}B_{i}C$. We have that $\angle B_{1}A_{1}C\neq
\angle B_{2}A_{2}C$ implies $\triangle
A_{1}B_{1}C\not\sim\triangle A_{2}B_{2}C$.
If $B_{1}\neq B_{2}$ and $A_{1}=A_{2}$, then $\angle
A_{1}CB_{1}\neq \angle A_{2}CB_{2}$. The angle $A_{i}CB_{i}$ is
the largest angle of $\triangle A_{i}B_{i}C$. We have that $\angle
A_{1}CB_{1}\neq \angle A_{2}CB_{2}$ in this case implies
$\triangle A_{1}B_{1}C\not\sim\triangle A_{2}B_{2}C$.
\end{proof}
\begin{remark} The only isosceles normal triangles are normal triangles of type $\triangle B\widetilde{B}C$ and $\triangle
BDC$. Right normal triangles are normal triangles with $AB$
passing through $O$. Acute/obtuse normal triangles as normal
triangles with $O$ inside/outside $\triangle ABC$. In contrast to
the one vertex normal forms triangles for the circle normal form
are unbounded from below.
\end{remark}
\begin{remark} Other normal forms of this type can be designed
choosing another point instead of $(1,0)$ and rearranging triangle
points.
\end{remark}
\subsubsection{Conversions}\
\begin{definition} Given a triangle with side lengths $a,b,c$
define $N_{X}(a,b,c)$ to be the Cartesian plane coordinates of the
$X$-normal point ($X\in \{A,B,C\}$) corresponding to this
triangle. Note that $N_{X}$ is a symmetric function. We can also
think of arguments of $N_{X}$ as multisets and think that
$N_{X}(a,b,c)=N_{X}(L)$, where $L$ is the multiset
$\{\{a,b,c\}\}$.
\end{definition}
\begin{proposition} Let $\triangle ABC$ has side lengths $a\le b\le c$.
Then
\begin{enumerate}
\item $N_{C}(a,b,c)=\Big(\frac{-a^2+b^2+c^2}{2c^2},
\frac{\sqrt{-a^4-b^4-c^4+2(a^2b^2+a^2c^2+b^2c^2)}}{2c^2}\Big)$;
\item $N_{B}(a,b,c)=\Big(\frac{-a^2+b^2+c^2}{2b^2},
\frac{\sqrt{-a^4-b^4-c^4+2(a^2b^2+a^2c^2+b^2c^2)}}{2b^2}\Big)$;
\item $N_{A}(a,b,c)=\Big(\frac{a^2-b^2+c^2}{2a^2},
\frac{\sqrt{-a^4-b^4-c^4+2(a^2b^2+a^2c^2+b^2c^2)}}{2a^2}\Big)$.
\end{enumerate}
\end{proposition}
\begin{proof}
1. Translate, rotate and reflect $\triangle ABC$ so that
$A=(0,0)$, $B=(c,0)$ and $C=(x,y)$ is in the first quadrant. For
$(x,y)$ we have the system $$\left\
\begin{array}{ll}
x^2+y^2=b^2\\
(c-x)^2+y^2=a^2\\
\end{array
\right. $$
and find $$\left\
\begin{array}{ll}
x=\frac{-a^2+b^2+c^2}{2c}\\
y=\frac{\sqrt{-a^4-b^4-c^4+2(a^2b^2+a^2c^2+b^2c^2)}}{2c}\\
\end{array
\right. $$
After the dilation by coefficient $\frac{1}{c}$ we get the given
formula.
2. and 3. proved are in a similar way.
\end{proof}
\begin{proposition}
Let a triangle $T$ have angles $\alpha\le\beta\le\gamma$.
Then
\begin{enumerate}
\item its $C$-normal point is $N_{C}(\frac{\sin\alpha}{\sin
\gamma},\frac{\sin\beta}{\sin\gamma},1)$;
\item if $T$ has the $C$-normal point $(x,y)$, then it has angles
$\alpha=\arctan{\frac{y}{x}}$, $\beta=\arctan{\frac{y}{1-x}}$,
$\gamma=\pi-\arctan{\frac{y}{x}}-\arctan{\frac{y}{1-x}}$.
\item its $B$-normal point is $N_{B}(\frac{\sin\alpha}{\sin
\beta},\frac{\sin\gamma}{\sin\beta},1)$;
\item if $T$ has the $B$-normal point $(x,y)$, then it has angles
$\alpha=\arctan\frac{y}{x}$,
$\beta=-\arctan\frac{y}{x}+\arctan\frac{y}{x-1}$,
$\gamma=\pi-\arctan\frac{y}{x-1}$;
\item its $A$-normal point is $N_{A}(\frac{\sin\beta}{\sin
\alpha},\frac{\sin\gamma}{\sin\alpha},1)$;
\item if $T$ has the $A$-normal point $(x,y)$, then it has angles
$\alpha=-\arctan\frac{y}{x}+\arctan\frac{y}{x-1}$,
$\beta=\arctan\frac{y}{x}$, $\gamma=\pi-\arctan\frac{y}{x-1}$.
\end{enumerate}
\end{proposition}
\begin{proof}
1. Let $\triangle ABC$ be the $C$-normal triangle with angles
$\alpha\le \beta \le \gamma$, i.e. $|AB|=1$. By the law of sines
we have $b=|AC|=\frac{\sin \beta}{\sin \gamma}$ and $a=\frac{\sin
\alpha}{\sin\gamma}$. By definition $C$ has coordinates
$N_{C}(\frac{\sin\alpha}{\sin\gamma},\frac{\sin\beta}{\sin\gamma},1)$.
2. Let $CD$ be a height of $\triangle ABC$. Formulas for angles
are obtained considering $\triangle ACD$ and $\triangle BCD$.
3.,4.,5.,6. ar proved similarly.
\end{proof}
\subsection{Normal forms of quadrilaterals}
In this subsection we consider mutisets of $4$ points in a plane.
A multiset of $4$ points can be interpreted as a quadrilateral. We
exclude the case of one point of multiplicity $4$. The multiset
$Q=\{\{X,Y,Z,T\}\}$ is also denoted as $\Box XYZT$. We define
$Q_{1}\sim Q_{2}$ provided there is an element of the dilation
group $g$ such that $g(Q_{1})=Q_{2}$.
A set of $4$ points defines a set of $6$ distances between these
points. Choosing any two points we can translate, rotate, reflect
and dilate the given $4$-point configuration so that the chosen
two points have coordinates $A=(0,0)$ and $B=(1,0)$. Different
normal forms can be obtained choosing pairs with different
relative metric properties. In this paper we consider only the
simplest case - two points having the maximal distance are mapped
to the $x$-axis.
\subsubsection{Longest distance normal form} Suppose we are given a quadrilateral $\Box XYZT$ such that $|XY|\ge |XZ|$,
$|XY|\ge |XT|$, $|XY|\ge |YZ|$, $|XY|\ge |YT|$, $|XY|\ge |ZT|$. We
map $X$ and $Y$ by a dilation to the $x$-axis (to $A=(0,0)$ and
$B=(1,0)$) and determine what are positions of the $2$ remaining
vertices $C$ and $D$ so that $\Box XYZT \sim \Box ABCD$.
\begin{definition} Let $p\in \mathbb{R}^2$. The mapping of $p$ by reflections of $p$ with respect to
the $x$-axis and the line $x=\frac{1}{2}$ to the domain $y\ge 0$,
$x\ge \frac{1}{2}$ is denoted by $p_{s}$.
\begin{definition} Let $p, p'\in \mathbb{R}^2$. We say that
$p$ is \sl quasilexicographically\rm\ smaller or equal to $p'$,
denoted by $p\vartriangleleft p'$, provided $p_{s}\prec p'_{s}$ or
$p_{s}=p'_{s}$. Given two pairs $[p,q]$ and $[p',q']$ we define
$[p,q]\vartriangleleft [p',q']$ provided ($p_{s}\prec p'_{s}$) or
($p_{s}=p'_{s}$ and $q\vartriangleleft q'$).
\end{definition}
\end{definition}
\begin{definition}
Let $S_{D}(x_{0},y_{0})\subseteq \mathbb{R}^2$ with
$(x_{0},y_{0})\in S_{C}$ (for the definition of $S_{C}$ see
section \ref{1}) be the set of solutions of the following system
of inequalities:
\begin{equation}\label{2}
\left\
\begin{array}{ll}
x^2+y^2\le 1\\
(x-1)^2+y^2\le 1\\
(x-x_{0})^2+(y-y_{0})^2\le 1\\
|x-\frac{1}{2}|\le |x_{0}-\frac{1}{2}|\\
if\ |x-\frac{1}{2}|=|x_{0}-\frac{1}{2}|,then\ |y|\le |y_{0}|\\
\end{array
\right.
\end{equation}
See Figure 9.
\begin{center}
\epsfysize=70mm
\epsfbox{quadr_6.eps}
Fig.9. - example of the domain $S_{D}$.
\end{center}
\end{definition}
\begin{remark} Conditions for $p\in S_{D}(x_{0},y_{0})$ consist of two
parts:
\begin{enumerate}
\item distance from $p$ to $A$, $B$ and $(x_{0},y_{0})$ is less
than or equal to $1$;
\item $p_{s}\vartriangleleft (x_{0},y_{0})$.
\end{enumerate}
\end{remark}
\begin{theorem}\label{3} Every $\Box UVWZ$ (including
multisets with multiplicities at most $3$) in $\mathbb{R}^{2}$ is
similar to $\Box AB\mathcal{CD}$, where $A=(0,0)$, $B=(1,0)$,
$\mathcal{C}\in S_{C}$ and $\mathcal{D}\in S_{D}$.
\end{theorem}
\begin{proof} Let $UVWZ$ be a multiset of points in $\mathbb{R}^{2}$ with at least two distinct elements.
Perform the following sequence of transformations:
\begin{enumerate}
\item translate and rotate the plane so that $2$ points with the
longest distance are on the $x$-axis, one vertex has coordinates
$(0,0)$ and another vertex has coordinates $(d,0)$, $d>0$; if
there is more than one possibility to choose two points with the
longest distance then choose this pair so that the remaining pair
is the largest in the quasilexicographic order;
\item do the dilation with coefficient $\frac{1}{d}$, note that
the vertices on the $x$-axis have coordinates $(0,0)$ and $(1,0)$,
suppose the the other two vertices have coordinates
$(x_{C},y_{C})$ and $(x_{D},y_{D})$;
\item if $|x_{C}-\frac{1}{2}|\ne |x_{D}-\frac{1}{2}|$, then put
the point with the maximal $|x-\frac{1}{2}|$ value into $S_{C}$ by
reflections with respect to the $x$-axis and the line
$x=\frac{1}{2}$;
\item if $|x_{C}-\frac{1}{2}|=|x_{D}-\frac{1}{2}|$, then put the
point with the maximal value of $|y|$ into $S_{C}$ by reflections
with respect to the $x$-axis and the line $x=\frac{1}{2}$;
\item if $|x_{C}-\frac{1}{2}|=|x_{D}-\frac{1}{2}|$ and
$|y_{C}|=|y_{D}|$, then map any of the points into $S_{C}$.
\end{enumerate}
Denote the point which is mapped to $S_{C}$ by this sequence of
transformations by $C=(x_{C},y_{C})$ and the fourth point by
$D=(x_{D},y_{D})$. For any $C=(x_{0},y_{0})\in S_{C}$ we have that
$S_{D}(x_{0},y_{0})\neq \emptyset$.
We check that $D\in S_{D}(x_{C},y_{C})$. From conditions $|AD|\le
1, |BD|\le 1$, $|CD|\le 1$ it follows that $D$ satisfies the first
three inequalities of the system \ref{2}. If $|y_{D}|>|y_{C}|$,
then $|x_{D}-\frac{1}{2}|<|x_{C}-\frac{1}{2}|$ due to the
quasilexicographic order condition.
\end{proof}
\begin{definition} \sl The longest distance normal form\rm\ of $\Box UVWZ$
is $\Box ABCD$ with $A=(0,0)$, $B=(1,0)$, $C=(x_{C},y_{C})\in
S_{C}$ and $D\in S_{D}(x_{C},y_{C})$ constructed according to the
algorithm given in the proof of \ref{3}.
\end{definition}
\begin{proposition} Let $\Box ABC_{1}D_{1}$ and $\Box ABC_{2}D_{2}$ be two
quadrilaterals constructed according to the longest distance
normal form algorithm.
If $C_{1}\neq C_{2}$ or $D_{1}\neq D_{2}$, then $\Box ABC_{1}D_{1}
\not\sim \Box ABC_{2}D_{2}$.
\end{proposition}
\begin{proof} $D_{i} \vartriangleleft C_{i}$, therefore if $C_{1}\neq C_{2}$, then
$\Box ABC_{1}D_{1}\not\sim\Box ABC_{2}D_{2}$.
Suppose $C_{1}=C_{2}$ and $D_{1}\neq D_{2}$. Under any similarity
mapping $C_{1}$ must be mapped to $C_{2}$. If a similarity mapping
fixes three noncollinear points $A$, $B$ and $C_{i}$, then it must
fix any other point of the plane. If $A$, $B$ and $C_{i}$ are on
the $x$-axis, then $D_{i}$ must also be on the $x$-axis and must
be fixed. Therefore $D_{1}\neq D_{2}$ implies $\Box
ABC_{1}D_{1}\not\sim \Box ABC_{2}D_{2}$ in this case.
\end{proof}
\begin{remark} If $C_{s}=D_{s}$, then there
are the following possibilities for the number of similarity types
of quadrilaterals with a given $C\in S_{C}$: 1) $1$ similarity
type if $C=D$, 2) $2$ similarity types if $C\neq D$ and
$|AC|=|BC|$, or $C\neq D$ and $C$ belongs to the $x$-axis or 3)
$4$ similarity types in other cases.
\end{remark}
\section{Possible uses of normal forms in education}
One vertex normal forms of triangles can be used to represent all
similarity types of triangles in a single picture with all
triangles having a fixed side, especially $C$-vertex and
$B$-vertex normal forms. It may be useful to have an example for
students showing that similarity type of triangle can be
parametrized by coordinates of a single point. One vertex normal
forms can also be used in considering quadrilaterals.
The circle normal form of triangles may be useful teaching
properties of circumscribed circles, e.g. inscribing triangles
with given angles in a circle.
Normal forms of triangles can also be used to teach the idea of
normal (canonical) objects using a case of simple and popular
geometric constructions.
\section{Conclusion and further development} It is relatively easy to define several normal forms of
triangles up to similarity. Since the main purpose of this work is
contribution to mathematics education only simplest approaches
which may be used in teaching are considered in this paper. One
approach is to map one side to the $x$-axis and use dilations and
reflections to position the third vertex in a unique way, in this
approach normal triangles are parametrized by one vertex. This
approach can be generalized for quadrilaterals. Another approach
considered in this paper is to design normal triangles as
triangles inscribed in a unit circle. Further development in this
direction may be related to using other figures related to a given
triangle, for example, the inscribed circle, medians, altitudes or
bisectors.
|
\section{Introduction}
Galactic star formation resides mostly within Giant Molecular Clouds
(GMCs), conventionally defined to have masses $\geq 10^4\:{M_\odot}$ and
observed to extend up to several~$\times 10^6\:{M_\odot}$
\citep[e.g.,][]{Blitz1993,Williams2000,McKeeOstriker2007}. With
typical mass surface densities of $\sim 100\:M_\odot\:{\rm pc}^{-2}$,
GMCs have mean radial sizes of $\sim 6 - 100$~pc, assuming simple
circular symmetry. However, GMCs are highly irregular and hierarchical
structures. Their dense clumps can spawn stellar clusters and
associations, creating the bulk of the Galactic field star
population. The efficiency and rate of star formation from these
clumps is relatively low, i.e., a few percent per local free-fall time
\citep[][]{ZuckermanEvans1974,KrumholzTan2007}. This appears to be
mostly because much of the GMC material is stable with respect to
gravitational collapse, especially material below a threshold
$A_V\sim10$~mag \citep[e.g.,][]{Lada2010}. Higher total star formation
efficiencies, $\sim10-50\%$, appear to be possible in the star-forming
clumps that form at least moderately bound clusters \citep{Lada2003}.
Different theoretical models of the processes that create star-forming
clumps within GMCs or prevent overdensities in the bulk of the cloud
are actively debated. These processes include the regulation of star
formation and stabilization of gas by magnetic fields
\citep{McKee1989,Mouschovias2001} or turbulence
\citep{KrumholzMcKee2005,PadoanNordlund2011}, and/or the initiation of
star formation by discrete triggering events, such as converging
atomic flows \citep[e.g.,][]{Heitsch2006}, cloud collisions
\citep[e.g.,][]{Tan2000} or stellar feedback
\citep[e.g.,][]{Samal2014}.
Infrared dark clouds (IRDCs) are likely to be examples of early stage
star-forming clumps
\citep[e.g.,][]{Perault1996,egan1998,Carey2000,Rathborne2006,ButlerTan2009,PerettoFuller2009,Battersby2010}. Thus
their study may help us understand the processes that initiate star
formation in GMCs. There have been many investigations of the internal
properties of IRDCs, including their temperatures
\citep[e.g.,][]{Pillai2006,Peretto2010,Ragan2011,Chira2013}, mass
surface density structure \citep[e.g.,
][]{ButlerTan2009,PerettoFuller2009,Ragan2011,ButlerTan2012,KainulainenTan2013,Butler2014},
kinematics \citep[e.g.,][]{Henshaw2013,JimenezSerra2014} and dynamics
\citep[e.g.,][]{HernandezTan2011,Hernandez2012}, CO depletion
\citep{Fontani2006,Hernandez2011}, chemistry \citep{Sanhueza2013}.
See \citet{Tan2014} for a review.
However, there have been fewer studies connecting IRDCs to their
larger-scale environments, such as the morphology, kinematics and
dynamics of their parent clouds. Theories involving production of
dense gas in shocks have been supported by
detection of large-scale SiO emission along IRDCs \citep{JimenezSerra2010,NguyenLuong2013}.
However, these studies focus only on a few individual clouds and are
still confined to a few-parsec scales in and around the
filamentary molecular clouds.
Here we study the $^{13}$CO(1-0)-emitting gas in and around 10 well-studied
IRDCs, utilizing data from the BU-FCRAO Galactic Ring Survey
\citep[GRS;][]{Jackson2006}. We consider a range of scales out to
30~pc projected radius, expected to encompass the potential GMC
environment of the IRDC. While our study connects to scales typical
of other large-sample GMC studies
\citep[e.g.,][]{Heyer2009,RomanDuval2009,RomanDuval2010}, by focusing
on just 10 regions
we are able to investigate their kinematic properties in much greater
detail. Other studies done on such a range of scales, from clump to
GMCs, have been performed on nearby GMCs, such as Orion A
\citep[e.g.,][]{Shimajiri2011}, Taurus \citep[e.g.,][]{Goldsmith2008},
and Perseus \citep{Ridge2006,Foster2009,Kirk2010}. However, these local
GMCs do not seem to give rise to the more extreme range of star-forming
clumps that is found in IRDCs.
The main focus of this paper is to ``bridge the gap'' between IRDC and
GMC studies. The questions we aim to address include: Are IRDCs
typically found within GMCs? Are IRDCs found within specific
locations with respect to GMCs? Are IRDCs and their surrounding GMCs
virialized, and does their degree of virialization vary as a function
of cloud physical scale?
The IRDC/GMC sample is presented in \S2. Methods for defining
cloud boundaries and estimating masses and kinematic properties are
described in \S3. Results are presented in \S4, including derived physical properties of
the clouds (\S\ref{S:physprop}), the location of IRDCs within their
respective GMCs (\S\ref{S:loc}), and the kinematic and dynamical
analysis of the IRDC and GMCs (\S\ref{S:kinematics}). We
conclude in \S5.
\section{The IRDC Sample and Molecular Line Data}
We utilize the $\thco (J=1\rightarrow0)$ data of the GRS survey
\citep[][]{Jackson2006}, which covers $18^{\circ}\le l \le
56^{\circ}$, $b\pm1^{\circ}$ and $-5 <v_{\rm LSR}<135~\kms$ (for $l
\le 40^{\circ}$) and $-5<v_{\rm LSR}<85~\kms$ (for $ l \ge
40^{\circ})$. The GRS had a spatial resolution of 46\arcsec, with
22\arcsec\ sampling, and a spectral resolution of $0.212~\kms$. The
typical rms noise is $\sigma \rm (T_A^*)=0.13$~K (or $\sigma \rm
(T_{B,\nu})\sim 0.26$~K with a main beam efficiency of 0.48).
Our selected IRDCs are the 10 clouds from \citet[][hereafter
BT09]{ButlerTan2009} \citep[see
also][]{ButlerTan2012,KainulainenTan2013}. This sample, a subset of
that of \citet{Rathborne2006}, was chosen while considering the
8$\mu$m (IRAC band 4) images from the \textit{Spitzer} Galactic Legacy
Mid-Plane Survey Extraordinaire
\citep[GLIMPSE;][]{Benjamin2003}. These IRDCs were selected for being
relatively nearby, massive, dark (i.e., relatively high contrast
against the surrounding diffuse emission), and surrounded by
relatively smooth diffuse emission within the 8$\mu$m GLIMPSE images.
Characteristic sizes and boundaries for each IRDC were taken from
\citet{Simon2006a}, where ellipses were fitted based on extinction in
MSX images. Although these fitted ellipses are not necessarily
accurate of IRDC shapes, they provide a convenient measure of the
approximate cloud structure. The catalog coordinates and sizes for the
10 IRDCs (including ``single component'' sub-classifications, labelled
``-s'', see below) are listed in Table \ref{tab1}.
\begin{table}[h]
\centering
\scriptsize
\caption{The IRDC Sample}
\begin{tabular}{lccccccccccccc}
\hline
IRDC & $l_{\rm IRDC}$\footnotemark[1]& $b_{\rm IRDC}$\footnotemark[1] & $r_{\rm maj}\footnotemark[1]$ & $r_{\rm min}\footnotemark[1]$ & PA$_e$\footnotemark[1] & $l_{0}$ & $b_{0}$ & $v_{0}$ & $l_{0,\tau}$ & $b_{0,\tau}$ & $v_{0,\tau}$ & $d$\footnotemark[2] & $\tex$\footnotemark[3] \\
& ($^{\circ}$)& ($^{\circ}$) & (\arcmin) & (\arcmin) & ($^{\circ}$) & ($^{\circ}$)& ($^{\circ}$) &$\kms$ & ($^{\circ}$)& ($^{\circ}$) &$\kms$ & $(\kpc)$ & (K)\\
\hline
\hline
A & 18.822 & -0.285 & 14.1 & 3.90 & 74.0 & 18.850 & -0.276 & 64.3 & 18.862 & -0.276 & 64.5 & 4.80 & 8.90 \\
B & 19.271 & 0.074 & 8.40 & 1.80 & 88.0 & 19.287 & 0.075 & 26.5 & 19.287 & 0.075 & 26.7 & 2.40 & 7.11 \\
C & 28.373 & 0.076 & 12.0 & 9.30 & 78.0 & 28.370 & 0.062 & 79.0 & 28.364 & 0.062 & 79.2 & 5.00 & 7.86 \\
D & 28.531 & -0.251 & 20.4 & 5.10 & 60.0 & 28.567 & -0.233 & 78.6 & 28.567 & -0.239 & 80.7 & 5.70 & 7.00 \\
D-s & & & & & & 28.555 & -0.251 & 87.9 & 28.549 & -0.257 & 87.7 & & \\
E & 28.677 & 0.132 & 14.6 & 4.10 & 103 & 28.672 & 0.130 & 80.7 & 28.684 & 0.130 & 81.3 & 5.10 & 7.00 \\
F & 34.437 & 0.245 & 5.30 & 2.00 & 79.0 & 34.434 & 0.240 & 57.3 & 34.428 & 0.240 & 57.5 & 3.70 & 6.48 \\
F-s & & & & & & 34.434 & 0.240 & 58.0 & 34.428 & 0.240 & 58.0 & & \\
G & 34.771 & -0.557 & 6.60 & 2.00 & 95.0 & 34.764 & -0.559 & 43.9 & 34.764 & -0.559 & 44.1 & 2.90 & 8.69 \\
G-s & & & & & & 34.770 & -0.559 & 44.4 & 34.764 & -0.559 & 44.4 & & \\
H & 35.395 & -0.336 & 20.6 & 6.40 & 59.0 & 35.422 & -0.319 & 47.5 & 35.416 & -0.319 & 48.0 & 2.90 & 6.61 \\
H-s & & & & & & 35.391 & -0.337 & 43.9 & 35.391 & -0.337 & 44.4 & & \\
I & 38.952 & -0.475 & 7.50 & 3.00 & 64.0 & 38.942 & -0.473 & 43.5 & 38.942 & -0.473 & 43.2 & 2.70 & 7.24 \\
J & 53.116 & 0.054 & 1.60 & 1.30 & 50.0 & 53.114 & 0.056 & 21.4 & 53.120 & 0.056 & 21.6 & 1.80 & 7.05 \\
\hline
\label{tab1}
\end{tabular}
\footnotetext[1]{IRDC coordinates and elliptical sizes adopted from \citet{Simon2006a}}
\footnotetext[2]{IRDC kinematic distances adopted from \citet{Rathborne2006}}
\footnotetext[3]{The $\thco$ excitation temperatures adopted from \cite{RomanDuval2010}}
\end{table}
\begin{figure}[ht!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=7.2in, angle=0, trim=0 0.8in 0 0.5in]{Fig1-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption{
The IRDC/GMC sample: each row presents a single cloud (A through E)
showing data out to a radius of 30 pc. \textit{Left to right:} GLIMPSE
$8\:{\rm \mu m}$ image with the IRDC boundary marked with a pink
ellipse (see Figs. \ref{mapA} and \ref{mapB}-\ref{mapE} for intensity scales);
$\thco$(1-0) GRS integrated intensity map (over the range $v_0\pm
15\:\kms$) in units of $\rm K ~\kms$; Full projections of the cloud in
position-velocity space: simple-extraction (SE) $\thco$(1-0)
$v_{\rm lsr}$ vs $l$ map, SE $\thco$(1-0) $v_{\rm lsr}$
vs $b$ map; connected extraction (CE) $\thco$(1-0) $v_{\rm lsr}$ vs
$l$ map; CE $\thco$(1-0) $v_{\rm lsr}$ vs $b$ map.
The IRDC is marked with a white box of width equal to the extent of
the elliptical boundary and height equal to the FWHM of the IRDC
co-added spectra. The solid black line shows the mass-weighted linear velocity gradient ($dv_0/ds$) across each cloud at $R=30$ pc.}
\label{PV1}
\end{figure}
Kinematic distances were adopted from \citet{Rathborne2006} and
\citet{Simon2006b}, where the IRDC central velocity was matched
morphologically between the mid-infrared (MIR) extinction estimated
from MSX and the $\thco$ emission from the GRS. These distances were
estimated assuming the rotation curve of \citet{Clemens1985}. We
assume uncertainties of 20\%, given the size of streaming motions of
$\sim 10 - 20\:\kms$ of clouds with trigonometric parallax
measurements \citep[e.g.,][]{Brunthaler2009,Reid2014}.
\section{Methods: Derivations of Molecular Cloud Properties}
\subsection{Definition of Cloud Boundaries}
We first determined the cloud center-of-mass in position-velocity
space. To do this, we co-added the $\thco$ spectra within an
elliptical boundary defined by twice the $r_{\rm maj}$ and $r_{\rm min}$
values, using the full GRS velocity range. Then, a velocity interval
of $30~\kms$ was considered, centered on the peak of the emission
profile. The cloud's center-of-mass was evaluated for both optically
thin and opacity corrected gas, $(l_0,b_0,v_0$) and
$(l_{0,\tau},b_{0,\tau},v_{0,\tau}$), respectively. We then use this
as the reference point around which to search for GMC-scale gas, out
to a projected radius of 30~pc and over a new velocity range of
$v_0\pm15\:\kms$. This range is wide enough to sample the
velocities of GMC gas, even those that are potentially undergoing
collisional interactions at velocities set by Galactic shear at the
tidal radii of the clouds \citep{Tan2000, TaskerTan2009}.
The GLIMPSE $8\mu$m and GRS $\thco$ integrated intensity maps centered
on these locations are presented in Figures \ref{PV1} and \ref{PV2}.
We examined $\thco$ emission within the $v_0\pm15$~km/s velocity range
and within apertures of varying radii from the cloud center-of-mass
coordinate. To explore how boundary definitions affect estimated
physical properties, we used two different methods to select $\thco$
line emission associated with the cloud. First, ``Simple Extraction
(SE)'' selected all the $\thco$ emission within radii $R=5, 10, 20$,
and 30~pc and $v_0\pm15$~km/s. Second, ``Connected Extraction (CE)''
defined a cloud as a connected structure in $l$-$b$-$v$-space with all
cloud voxels required to be above a given threshold intensity: we
defined a given voxel as ``molecular cloud gas'' if its $\thco$(1-0)
line intensity $T_{B,\nu}\geq 1.35$~K (i.e., the GRS $5\sigma_{\rm rms}$
noise level). The CE search was also limited to within a 30~pc radius
and $\pm15\:\kms$ of the IRDC center-of-mass. Position-velocity maps
for each cloud, defined by both SE and CE, are shown in Figures
\ref{PV1} and \ref{PV2}.
\begin{figure}[t!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=7.2in, angle=0, trim=0 0.8in 0 0.5in]{Fig2-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption{The same as Figure \ref{PV1}, but for Clouds F through J.}
\label{PV2}
\end{figure}
Clouds defined by CE do not have simple radial sizes. Therefore, we
estimated three circular boundaries, centered on the extracted cloud's
center-of-mass: (1) mass-weighted radius, $R_M$, defined as mean
projected radial distance of cloud mass from the center of mass; (2)
areal radius, $R_A$, defined by the total projected area $A = \pi R_A^2
= N_p A_p$, where $N_p$ is the total number of pixels subtended by the
cloud and $A_p$ is the area of one image pixel; (3) half-mass radius,
$R_{1/2}$, defined as the radius from cloud center that contains half
of the total mass.
For the IRDCs, we also selected ``cloud'' material via SE and CE using
the spatial coordinates and elliptical boundaries from
\citet{Simon2006a}, along with the $v_0\pm15~\kms$ velocity intervals.
\subsection{Column Densities and Masses from $\thco$ emission}\label{S:13COmass}
We estimated the $\thco$ column density of each molecular cloud voxel,
$dN_{\rm 13CO}$ from their $J=1\rightarrow 0$ line emission assuming a
partition function with a thermal distribution described by an
excitation temperature $T_{\rm ex}$ via:
\begin{equation} \frac{dN_{\rm 13CO}}{dv} = \frac{8\pi Q_{\rm
rot}}{A \lambda_0^3} \frac{g_l}{g_u} \frac{\tau_{\nu} }{1-{\rm
exp} (-h\nu /[kT_{\rm ex}])},
\label{eq:dN}
\end{equation}
where $Q_{\rm rot}$ is the partition function, $A = 6.294\times
10^{-8}{\rm s^{-1}}$ is the Einstein coefficient, $\lambda_0=
0.27204$~cm, $g_l = 1$ and $g_u = 3$ are the statistical weights of
the lower and upper levels and $\tau_\nu$ is the optical depth of the
line at frequency $\nu$, i.e., at velocity $v$.
Each GRS voxel has a velocity width of ${\rm
d}v=0.212~\kms$. For linear molecules, the partition function is
$Q_{\rm rot}=\sum_{J=0}^{\infty} (2J+1) {\rm exp}(-E_J/kT_{\rm ex})$
with $E_J =J(J+1)h B$ where $J$ is the rotational quantum number and
$B=5.5101\times 10^{10}\:{\rm s}^{-1}$ is the rotational constant. For
$\thco$(1-0) we have $E_J/k = 5.289$~K.
Many studies of the physical properties of GMCs and IRDCs have
accounted for line optical depth
when estimating their physical properties \citep[e.g.,][]{Heyer2009,
RomanDuval2010, HernandezTan2011, Hernandez2011, Hernandez2012}. In
\citet{HernandezTan2011}, we showed that optical depth correction
factors can increase the $\thco$ column density by a factor of $\sim2$
in the densest, sub-parsec scale clumps of IRDCs. However, for the
more diffuse GMCs the optical depth correction factors are expected to
be smaller. To gauge the importance of this effect, we carry out
column density estimates for both the optically thin assumption and
accounting for opacity corrections.
The optical depth is evaluated via
\begin{equation}
T_{B,\nu}=\frac{h\nu}{k}[f(T_{\rm ex})-f(T_{\rm bg})][1-e^{-\tau_{\nu}}],
\label{eq:tau}
\end{equation} where $T_{B,\nu}$ is the brightness temperature at frequency
$\nu$, $f(T)\equiv [{\rm exp}(h\nu/[kT])-1]^{-1}$, and $T_{\rm
bg}=2.725$~K is the background temperature. $T_{B,\nu}$ is derived
from the antenna temperature, $T_A$, via $T_A\equiv \eta f_{\rm
clump}T_{B,\nu}$, where $\eta$ is the main beam efficiency
($\eta=0.48$ for the GRS) and $f_{\rm clump}$ is the beam dilution
factor of the $\thco$ emitting gas, which we assume to be unity due to
the large scale extent of GMCs. Smaller scale structures are
undoubtedly present, e.g., as revealed in the BT09 MIR extinction
maps, but to gauge the effects of these on the CO emission requires
higher resolution molecular line maps of the clouds. For $\tau \ll 1$,
Equation (\ref{eq:tau}) can be simplified to express $\tau_{\nu}$ by:
$\tau_{\nu}=(T_{B,\nu}k / [h\nu] )[f(T_{\rm ex}) - f(T_{\rm
bg})]^{-1}$. With this simplification, for an observed voxel
$T_{B,\nu}$ and an assumed $T_{\rm ex}$, the optically thin $\thco$
column density per voxel is given by combining Equations (\ref{eq:dN})
and (\ref{eq:tau}):
\begin{equation} \frac{dN_{\rm 13CO}}{dv} = 1.251 \times 10^{14} \frac{Q_{\rm
rot}}{f(T_{\rm ex}) - f(T_{\rm bg})} \frac{T_{B,\nu}}{1-{\rm
exp}(-{h\nu/[kT_{\rm ex}])}} ~\frac{\rm{cm}^{-2}}{~\kms}.
\label{eq:dN2}
\end{equation}
Early studies of IRDCs estimated typical gas kinetic temperatures
$T_{\rm gas}\sim20$~K \citep{Carey1998, Carey2000, Pillai2006}. IRDC F
was estimated to have a temperature of 19~K based on $\rm NH_3$~(1,1)
and (2,2) VLA observations \citep{Devine2009}. However, as discussed
below, CO excitation temperatures appear to be significantly lower.
In our previous study of IRDC H, we used IRAM~30m observations of $\rm
C^{18}O$(2-1) and (1-0) emission from around the IRDC filament to
estimate a mean $T_{\rm ex}\sim7$~K \citep[][hereafter
H11]{Hernandez2011}. Here, for a uniform analysis of the 10 IRDCs,
we now use $\cotwo$ $T_{\rm ex}$ estimates from \citet[][hereafter
RD10]{RomanDuval2010}. In their study of 580 molecular clouds,
brightness temperatures from $\cotwo$(1-0) emission line data
(Univ. of Massachusetts-Stony Brook (UMSB) Galactic Plane Survey),
were used to derive proxy $\thco$ excitation temperatures, assuming
$\cotwo$ emission was optically thick and that
$\thco$ and $\cotwo$ excitation temperatures are
equal. Ultimately, RD10 cited a mean excitation temperature for all
their molecular clouds based on all cloud voxels above $4\sigma_{\rm rms}$. The
RD10 clouds were extracted from the GRS data using a modified version
of CLUMPFIND \citep{Williams1994}, which allowed for varying
thresholds (i.e., contour increment and minimum brightness, see
\citet{Rathborne2009} for details).
Eight of our IRDCs overlapped with at least one of their molecular
clouds, and in these cases we adopted $T_{\rm ex}$ from the RD10 value
from the overlapping cloud(s). For the remaining two IRDCs (D and E)
we set $T_{\rm ex}=7$~K, similar to the mean values of 7.2~K of H11
and 6.32~K of RD10. Our adopted values of $T_{\rm ex}$ are
listed in Table \ref{tab1}. For our column density estimates that
assume optically thin $^{13}$CO emission, these $T_{\rm ex}$ values
are assumed to be constant throughout the cloud. These
temperatures are slightly lower (by a few~K) than those used in
previous studies (e.g., \citealt{Simon2001,Simon2006b} who assumed a
fixed value of 10~K). However, the results from
\citet{Heyer2009} indicate that CO gas throughout GMCs is mostly sub-thermally excited.
Note that, in this optically thin limit, varying $T_{\rm ex}$
from 5~K to 10~K would change the derived column density of the cloud
by only $\sim 20\%$.
For the opacity-corrected case, the use of a single mean excitation
temperature of relatively low value ($\sim 7$~K) can lead to
non-physical results in certain regions of the cloud.
Equation \ref{eq:tau} implies
$\tau_{\nu}=-\ln{(1-[(T_{B,\nu}k / [h\nu] )[f(T_{\rm ex}) - f(T_{\rm bg})]^{-1}])}$.
Thus, for a given observed $T_{B,\nu}$, $\tau_{\nu}$ will become
undefined if $T_{\rm ex} \le T_{\rm ex,crit}$ (i.e., when
$[(T_{B,\nu}k / [h\nu] )[f(T_{\rm ex}) - f(T_{\rm bg})]^{-1}] >
1$). For example, a voxel with an observed brightness temperature of
$T_{B,\nu}=5$~K will have a numerically undefined opacity at $T_{\rm
ex,crit}<8.2$~K.
H11 showed modest temperature variations were present in IRDC H, with a peak
temperature of $T_{\rm ex} \sim10$~K within the densest clumps and $T_{\rm ex}
\sim 7$~K in the more diffuse gas within the filament envelope (see
H11, Fig. 1). Hence, we expect that the voxels containing the largest
brightness temperatures have excitation temperatures which are a few K
larger than the constant $T_{\rm ex}$ values adopted from RD10.
To estimate the opacity-corrected column density, we first apply our
adopted $T_{\rm ex}$ to estimate $\tau_{\nu}$ in each voxel. Then, for
each voxel with an undefined $\tau_{\nu}$, we specified a new
excitation temperature of $T_{\rm ex,crit}+1$~K, given the voxel's
observed $T_{B,\nu}$. This revised excitation temperature allows us to
estimate a real solution for $dN_{\rm 13CO}/dv$ throughout the
cloud. After considering a range of possible temperature offsets from
$T_{\rm ex,crit}$, we estimate that the uncertainty in $dN_{\rm
13CO}/dv$ in an individual voxel by this method is at a level of
$\sim 20\%$.
Ideally, excitation temperatures would be estimated locally from
$^{12}$CO $J=1-0$ observations of the clouds, assuming that the
$^{12}$CO line is optically thick. However, there are currently no
other $^{12}$CO surveys that match both the resolution and spatial
extent of the GRS. The widely used Columbia-CfA $^{12}$CO survey
covers the whole Galactic plane, but with a low angular resolution of
8\arcmin\ \citep{Dame2001}. The RD10 temperatures measurements are
based on the UMSB survey, which has a 44\arcsec\ spatial resolution
with 3\arcmin\ sampling.
For the clouds defined by SE extraction and with $R=30$ pc, we find
that on average $0.7\%$ of the cloud voxels require higher
temperatures than those listed in Table \ref{tab1}. Cloud D has the
highest percentage, 1.5\%, and for these voxels
$T_{\rm ex,crit}$ peaks at 16.4 K with a mean $T_{\rm ex,crit}=8.3$~K.
For the clouds defined by CE extraction and $R_A$, we find that on
average $10\%$ of voxels require higher excitation temperatures. Here,
Cloud J has the highest percentage, 27\%, with a peak $T_{\rm ex,crit}$ of
25.0~K and a mean $T_{\rm ex,crit}=9.2$~K.
The $\thco$-based mass per voxel,
$dM$ was then calculated by
assuming a $n_{\rm 12CO}/n_{\rm 13CO}=54$ \citep{Milam2005} and
$n_{\rm 12CO}/n_{\rm H2}=2.0 \times 10^{-4}$ \citep{Lacy1994}. Hence, the
assumed abundance of $\rm ^{13}CO$ to $\rm H_2$ is $3.70\times
10^{-6}$. The mass located at a given voxel is then given by
\begin{equation}
dM=1.45\times10^{-4} \frac{dN_{\rm
13CO}}{10^{13}~\rm{cm}^{-2}} \frac{dv}{0.212~\kms} \left(\frac{d}{\rm
kpc}\right)^2 \frac{\Delta l}{22\arcsec} \frac{\Delta b}{22\arcsec}
~\rm{M_{\odot}},
\label{eqn:M}
\end{equation}
where $d$ is the cloud distance, $\Delta l$ and $\Delta b$ are the
angular sizes of the GRS pixels, and assuming a mass per H nucleus of
$\mu_{\rm H}=2.34\times 10^{-24}\:{\rm g}$. The total $\thco$-derived
cloud mass, $M$, is simply the total mass of all cloud voxels within
radius $R$ and velocity range $v_0\pm15~\kms$. For clouds defined by
SE, any pixels with total integrated intensities below $5\sigma_{\rm
rms}$ were omitted from further analysis.
We estimate an uncertainty of $20\%$ in $T_{\rm ex}$ due to its
intrinsic variation, which in addition to uncertainties due to
intrinsic abundance variations, leads to $\sim30\%$ uncertainty in the mass
surface density ($\Sigma$). We thus estimate $\sim$50\% random errors in
$M$, after accounting for these uncertainties in
$\Sigma$ and the cloud kinematic distance estimates (assumed to be
$\sim 20\%$). However, we also anticipate that there could be
global systematic uncertainties in $M$ (of the whole cloud sample) of up to a factor of 2, given the
uncertainties in overall absolute $\thco$ abundance.
\subsection{Cloud Kinematics}
We used co-added $\thco$ column density velocity distributions (e.g.,
Fig. \ref{mapA}: \textit{Right}) to determine the mean velocity, $v_0$
of each extracted cloud.
The velocity dispersion was estimated using two standard methods: 1)
the rms 1D velocity dispersion, $\sigma$;
2) the width of a fitted Gaussian profile, $\sigma_G$. We estimate an mean uncertainty in $\sigma$
of $10\%$. Additionally, we used these Gaussian fitted profiles to estimate a Gaussian profile
mass, $M_G$, of each cloud. To visualize how $v_0$ and $\sigma$ vary
throughout the cloud, we show mass-weighted first and second moment
maps of each GMC (e.g., Fig. \ref{mapA}).
Using the first moment maps, we derive the velocity gradients in each
spatial direction. For example, the longitudinal velocity gradient,
$dv_0/d l$, was derived by first estimating the mean velocity at each
longitudinal position, via a mass-weighted sum along the perpendicular
($b$) direction, then finding the best (mass-weighted) linear fit to
these velocities. This method was repeated for $dv_0/db$. The
magnitude of the total linear velocity gradient across the cloud,
$dv_0/ds$, and its position angle direction, $\theta_v$, were then
calculated. Note, we choose to use mass-weighted velocity
gradients to prevent the results being unduly affected by tenuous
wisps of cloud material. Also, in the context of interpreting
velocity gradients as due to rotation and thus measuring rotational
energies of the cloud (below), this mass-weighted gradient is the
appropriate one to use.
Many studies of the dynamics of molecular clouds have interpreted
total GMC linear velocity gradients as due to solid body rotation
\citep[e.g.,][]{Phillips1999,Rosolowsky2003,ImaraBlitz2011}.
However, the possibility remains that the identified single ``cloud''
actually consists of spatially independent structures. For a cloud
undergoing solid body rotation, the line of sight velocity gradient in
the plane of the sky, $dv_{0}/ds$, is equal to the projected angular
velocity, i.e., $\Omega_0 = dv_{0}/ds$. The true angular velocity is
$\Omega = \Omega_0/{\rm sin} i$, where $i$ is the angle between the
rotation axis and the line of sight.
The position angle of the projected rotation axis of the cloud
contains information that may constrain theories of GMC formation and
evolution. For example, if a GMC forms rapidly from atomic gas in the
Galactic plane, then the GMC rotation is expected to be prograde with
respect to Galactic rotation \citep[e.g.,][]{TaskerTan2009}.
If strong gravitational encounters and collisions are frequent between
gravitationally bound GMCs (Tan 2000), then a more random set of
orientations of the positional angles of projected rotation axes are
expected, including both pro- and retrograde rotating clouds.
For clouds observed in the Galactic plane,
$-90^{\circ}<\theta_v<+90^{\circ}$ represents retrograde rotation and
$-90^{\circ}<\theta_v<-180^{\circ}$ and
$90^{\circ}<\theta_v<180^{\circ}$ represent prograde rotation.
We then also estimated the projected moment of inertia, $I_0 = \sum
dM_{i} s_{i}^2$, of each cloud using the sky projected rotation
axis, defined by $\theta_v$ and the cloud center-of-mass coordinate,
where $s_i$ is the shortest distance to the rotation axis. This allows
estimation of the projected rotational energy of the cloud, $E_{\rm
rot,0}=(1/2) I_0 \Omega_0^2$.
\begin{sidewaystable}[h]
\centering
\tiny
\caption{GMC Physical Parameters Estimates}
\begin{tabular}{lcccccccccccccc}
\hline
IRDC & & & & & & & & & & & & & & \\
Case & $l_{\rm CM}$,$l_{\rm CM,\tau}$ & $b_{\rm CM}$,$b_{\rm CM,\tau}$ & $ v_{\rm CM}$,$v_{\rm CM,\tau}$ & $\Delta s$ & $\Delta v$ &R &$M$,$M_{\tau}$,$M_{G}$ & $\sigma$,$\sigma_{\tau}$,$\sigma_{G}$ & $\frac{dv_0}{ds}$,$\frac{dv_{0,\tau}}{ds}$ & $\theta_v$,$\theta_{v,\tau}$ & $I_0$,$I_{0,\tau}$ & $E_{\rm rot,0}$,$E_{\rm rot,0,\tau}$ & $\beta,\beta_{\tau}$ & $\log(\alpha_{\rm vir})$,$\log(\alpha_{\rm vir,\tau})$, \\
& & & & & & & $\times10^4 $ & & $\times \rm 10^{-2}$ & & $\times \rm 10^{3}$ & $\times 10^{46}$ & ($10^{-2}$) & $\log(\alpha_{\rm vir,G})$ \\
(pc) & ( $^{\circ}$) & ($^{\circ}$) & ($\kms$)& (pc) & ($\kms$)& (pc) & $({M_\odot})$ & ($\kms$) & (km/s/pc) & ($\deg$) & ($M_{\odot}$~pc$^2$) & (erg) & & \\
(1) & (2) & (3) &(4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) & (12) &(13) & (14) &(15) \\
\hline
\hline
A& & & & & & & & & & & & & & \\
5&18.856,18.868&-0.282,-0.276&64.3,64.3&2.06,2.12&0.0,0.0&5.00&1.10,1.87,1.62&3.92,3.48,1.66&36.0,32.3&141,124&0.0814,0.131&1.37,4.01&7.62,3.39&0.912,0.575,-0.00704\\
10&18.862,18.875&-0.282,-0.282&63.9,64.1&2.58,2.57&-0.4,-0.2&10.0&4.83,7.20,5.73&5.12,4.53,1.94&15.2,16.1&124,123&1.36,1.97&13.3,29.6&2.34,1.72&0.802,0.521,-0.115\\
20&18.868,18.881&-0.301,-0.313&63.5,63.7&3.46,4.02&-0.9,-0.6&20.0&16.1,24.0,19.2&5.87,5.15,2.36&16.8,15.6&139,142&16.8,25.7&74.2,165&6.38,3.78&0.696,0.411,-0.170\\
30&18.868,18.881&-0.307,-0.319&62.6,63.1&3.72,4.37&-1.7,-1.3&30.0&31.0,44.1,34.9&6.58,6.01,3.02&12.7,12.1&96.8,96.3&68.2,96.7&183,370&5.97,3.80&0.688,0.457,-0.0389\\
$R_M$&18.881,18.893&-0.325,-0.337&63.3,63.5&5.10,6.60&-1.7,-1.3&17.4,17.0&9.63,16.7,15.7&3.38,2.88,2.00&17.5,14.0&151,153&7.63,12.9&30.5,94.2&7.67,2.69&0.380,-0.00984,-0.298\\
$R_A$&18.881,18.893&-0.325,-0.337&63.3,63.5&5.10,6.60&-1.7,-1.3&26.6,26.6&16.6,28.7,26.9&3.88,3.36,2.33&16.5,14.0&133,134&32.3,54.3&59.3,177&14.8,5.97&0.447,0.0862,-0.204\\
$R_{1/2}$&18.881,18.893&-0.325,-0.337&63.3,63.5&5.10,6.60&-1.7,-1.3&17.1,16.2&9.38,15.5,14.7&3.34,2.81,1.99&17.2,13.5&151,153&7.18,11.0&29.4,84.9&7.18,2.36&0.374,-0.0177,-0.291\\
\hline
B& & & & & & & & & & & & & & \\
5&19.299,19.299&0.050,0.050&27.4,26.7&1.65,1.46&0.9,0.6&5.00&0.799,1.38,1.19&5.27,4.39,1.59&34.1,34.3&105,97.6&0.0347,0.0547&0.729,2.18&5.50,2.94&1.31,0.909,0.0934\\
10&19.323,19.330&0.032,0.032&26.9,26.5&2.94,2.94&0.4,0.4&10.0&2.30,3.68,2.88&5.88,5.09,1.87&67.5,64.0&114,112&0.428,0.603&3.01,7.75&64.3,31.7&1.24,0.912,0.150\\
20&19.342,19.354&0.025,0.025&27.4,26.9&3.72,3.93&0.9,0.9&20.0&5.35,7.64,5.21&6.67,6.18,2.21&58.2,57.4&89.3,89.1&2.72,3.62&8.16,16.6&112,71.2&1.29,1.07,0.340\\
30&19.317,19.323&0.013,0.019&27.4,27.1&3.30,3.10&0.9,1.1&30.0&7.70,10.7,7.26&7.14,6.79,2.80&42.3,41.7&87.3,86.4&6.91,9.34&11.3,21.9&109,73.6&1.36,1.18,0.577\\
$R_M$&19.447,19.434&0.062,0.050&24.8,25.0&7.23,6.78&-1.1,-0.9&7.79,7.34&1.26,2.35,2.27&2.20,2.13,1.78&36.3,36.5&109,112&0.195,0.327&1.17,4.30&21.9,10.1&0.541,0.219,0.0764\\
$R_A$&19.447,19.434&0.062,0.050&24.8,25.0&7.23,6.78&-1.1,-0.9&12.2,12.2&1.87,3.42,3.26&2.19,2.08,1.62&31.5,31.1&100.,101&0.353,0.556&1.64,5.47&21.2,9.77&0.559,0.256,0.0577\\
$R_{1/2}$&19.447,19.434&0.062,0.050&24.8,25.0&7.23,6.78&-1.1,-0.9&7.20,6.60&1.12,2.02,1.96&2.25,2.19,1.84&40.8,41.5&110,113&0.159,0.253&0.987,3.54&26.6,12.2&0.581,0.259,0.124\\
\hline
C& & & & & & & & & & & & & & \\
5&28.370,28.364&0.062,0.062&78.4,78.4&0.54,0.76&-0.2,-0.2&5.00&1.75,3.23,3.10&4.18,3.49,2.76&11.6,11.5&-118,-114&0.110,0.191&3.51,11.9&0.423,0.212&0.763,0.341,0.154\\
10&28.370,28.364&0.055,0.055&78.6,78.6&1.07,1.20&0.0,0.0&10.0&5.77,9.31,8.82&4.67,4.09,3.08&6.40,6.77&-74.0,-86.7&1.31,1.86&19.0,49.5&0.282,0.171&0.643,0.320,0.0965\\
20&28.370,28.358&0.049,0.049&79.0,79.2&1.61,1.94&0.4,0.6&20.0&19.7,28.0,25.6&5.50,5.16,3.82&6.61,7.05&-121,-119&18.7,25.2&111,224&0.735,0.556&0.553,0.344,0.122\\
30&28.370,28.364&0.031,0.025&79.0,79.0&3.22,3.79&0.4,0.4&30.0&36.7,50.2,45.9&5.77,5.50,4.15&4.54,4.75&-116,-117&68.0,88.2&256,479&0.544,0.414&0.501,0.324,0.117\\
$R_M$&28.364,28.358&0.018,0.012&78.4,78.4&4.33,4.95&0.2,0.0&17.1,16.6&11.4,17.2,17.2&3.45,3.27,3.07&4.50,4.22&-75.6,-66.2&6.60,9.71&43.5,102&0.306,0.169&0.318,0.0792,0.0259\\
$R_A$&28.364,28.358&0.018,0.012&78.4,78.4&4.33,4.95&0.2,0.0&29.0,29.0&24.4,36.9,35.3&4.21,4.21,3.46&7.72,7.12&-91.4,-90.5&34.7,51.6&117,268&1.75,0.970&0.389,0.209,0.0588\\
$R_{1/2}$&28.364,28.358&0.018,0.012&78.4,78.4&4.33,4.95&0.2,0.0&17.8,17.2&12.1,18.2,18.2&3.48,3.27,3.09&5.09,4.43&-82.9,-75.5&7.38,10.6&47.3,110&0.402,0.189&0.314,0.0725,0.0235\\
\hline
D& & & & & & & & & & & & & & \\
5&28.573,28.567&-0.233,-0.239&80.7,82.0&0.61,0.86&0.2,0.4&5.00&1.95,3.23,2.09&7.59,7.03,1.83&20.2,19.9&-130,-126&0.113,0.176&4.32,11.9&1.06,0.583&1.24,0.949,-0.0280\\
10&28.579,28.573&-0.233,-0.239&79.6,80.9&1.22,0.61&-0.9,-0.6&10.0&5.96,9.23,4.66&7.61,7.31,1.88&18.4,20.1&118,112&1.42,2.15&20.3,48.6&2.35,1.78&1.05,0.828,-0.0534\\
20&28.586,28.592&-0.227,-0.233&79.4,80.3&1.93,1.84&-1.1,-1.3&20.0&18.5,27.6,27.9&7.45,7.47,11.7&17.9,19.3&62.4,59.4&19.6,28.8&97.6,218&6.37,4.92&0.845,0.673,1.06\\
30&28.573,28.567&-0.196,-0.202&79.6,80.3&3.72,3.12&-0.9,-1.3&30.0&36.1,54.8,55.1&7.07,7.11,9.24&8.59,10.4&50.0,47.3&81.5,123&249,571&2.41,2.34&0.684,0.509,0.733\\
$R_M$&28.567,28.567&-0.177,-0.184&79.6,80.5&5.05,4.45&-1.7,-1.9&17.6,18.0&10.5,18.3,10.7&6.57,6.74,2.46&19.9,23.7&67.6,63.1&9.31,16.8&35.8,106&10.3,8.86&0.926,0.716,0.0741\\
$R_A$&28.567,28.567&-0.177,-0.184&79.6,80.5&5.05,4.45&-1.7,-1.9&28.2,28.2&22.1,37.5,37.1&6.23,6.53,7.21&17.4,20.9&46.3,44.1&51.4,87.3&99.1,285&15.7,13.3&0.761,0.571,0.662\\
$R_{1/2}$&28.567,28.567&-0.177,-0.184&79.6,80.5&5.05,4.45&-1.7,-1.9&18.1,18.5&11.0,19.2,11.3&6.57,6.76,2.51&19.9,23.9&67.4,62.3&10.3,18.9&38.2,114&10.7,9.35&0.917,0.709,0.0795\\
\hline
D-s& & & & & & & & & & & & & & \\
5&28.567,28.567&-0.233,-0.239&87.1,86.9&0.61,0.86&-0.2,-0.2&5.00&0.991,1.94,1.96&1.97,1.66,1.66&4.22,5.23&172,-177&0.0874,0.355&1.12,4.31&0.138,0.224&0.358,-0.0818,-0.0853\\
10&28.573,28.573&-0.239,-0.245&87.5,87.3&0.61,1.22&0.2,0.2&10.0&2.44,4.45,4.41&2.06,1.78,1.72&11.2,10.8&124,121&0.547,0.968&3.41,11.3&2.02,0.992&0.305,-0.0813,-0.109\\
20&28.604,28.604&-0.257,-0.270&88.1,87.9&3.92,4.78&0.9,0.9&20.0&6.14,10.0,10.2&2.39,2.12,2.31&14.1,14.2&64.8,65.0&5.13,7.54&10.8,28.7&9.39,5.27&0.334,0.0174,0.0848\\
30&28.616,28.610&-0.245,-0.257&88.3,87.9&4.46,4.41&1.1,0.9&30.0&10.4,15.9,16.2&2.57,2.27,2.52&11.2,10.5&46.5,50.2&20.4,25.8&20.6,47.9&12.4,5.90&0.345,0.0551,0.137\\
$R_M$&28.635,28.629&-0.264,-0.270&87.9,87.7&6.29,6.12&1.1,0.9&13.5,13.4&3.58,7.14,7.08&1.96,1.90,1.88&22.3,22.1&74.4,78.3&1.28,2.56&5.43,21.7&11.6,5.72&0.224,-0.102,-0.107\\
$R_A$&28.635,28.629&-0.264,-0.270&87.9,87.7&6.29,6.12&1.1,0.9&20.5,20.5&4.91,9.35,9.32&2.09,2.03,2.06&21.5,20.6&51.6,60.8&3.15,5.32&6.71,24.4&21.7,9.20&0.328,0.0204,0.0361\\
$R_{1/2}$&28.635,28.629&-0.264,-0.270&87.9,87.7&6.29,6.12&1.1,0.9&11.7,11.1&3.00,5.76,5.73&1.86,1.75,1.72&21.3,21.4&69.5,75.4&0.890,1.64&4.40,17.1&9.11,4.37&0.194,-0.164,-0.176\\
\hline
E& & & & & & & & & & & & & & \\
5&28.672,28.684&0.130,0.130&79.4,79.4&0.55,0.54&-0.4,-0.6&5.00&0.907,1.27,0.966&6.67,6.16,2.71&31.3,11.8&144,128&0.0571,0.0742&0.939,1.84&5.92,0.559&1.46,1.24,0.646\\
10&28.672,28.690&0.118,0.118&80.7,81.1&1.23,1.55&0.9,1.1&10.0&3.52,5.03,4.20&6.45,6.28,3.66&18.3,23.5&74.8,79.6&0.747,1.01&7.08,14.5&3.52,3.87&1.14,0.960,0.568\\
20&28.653,28.666&0.105,0.099&80.3,80.5&3.10,2.95&0.4,0.4&20.0&15.0,21.1,18.6&6.52,6.40,4.32&5.69,6.81&-41.2,-28.0&15.6,22.7&63.8,127&0.787,0.822&0.820,0.655,0.369\\
30&28.641,28.647&0.081,0.075&80.3,80.5&5.47,5.64&0.4,0.4&30.0&32.4,46.4,40.4&6.93,6.87,4.86&6.34,7.14&-44.8,-41.1&68.8,101&199,409&1.38,1.25&0.714,0.551,0.310\\
$R_M$&28.616,28.622&0.056,0.050&79.8,80.3&8.21,8.21&0.0,0.4&17.7,17.6&9.01,14.8,13.7&5.47,5.76,3.86&14.3,15.9&-35.2,-35.6&7.29,12.1&26.2,71.3&5.66,4.26&0.834,0.662,0.349\\
$R_A$&28.616,28.622&0.056,0.050&79.8,80.3&8.21,8.21&0.0,0.4&27.5,27.5&20.2,34.2,31.4&5.44,5.62,4.02&10.5,11.3&-8.71,-3.01&40.1,71.4&84.7,242&5.16,3.77&0.671,0.472,0.216\\
$R_{1/2}$&28.616,28.622&0.056,0.050&79.8,80.3&8.21,8.21&0.0,0.4&18.1,17.9&9.41,15.3,14.1&5.47,5.76,3.86&14.2,15.6&-37.2,-36.3&7.93,12.8&27.9,74.2&5.67,4.18&0.826,0.655,0.344\\
\hline
F& & & & & & & & & & & & & & \\
5&34.428,34.422&0.240,0.240&56.9,56.9&0.89,1.25&-0.6,-0.6&5.00&0.897,1.51,1.30&4.09,3.55,2.09&20.2,25.8&75.7,64.8&0.0373,0.0535&0.919,2.60&1.65,1.36&1.03,0.688,0.289\\
10&34.422,34.409&0.234,0.228&56.0,56.3&1.43,2.31&-1.5,-1.3&10.0&2.87,4.32,3.68&4.99,4.50,2.93&25.6,24.1&109,103&0.642,0.861&4.70,10.7&8.92,4.66&1.00,0.737,0.434\\
20&34.415,34.397&0.215,0.209&54.6,55.0&2.54,3.66&-3.0,-2.6&20.0&10.6,16.2,14.8&5.64,5.25,4.18&18.0,18.6&161,159&12.6,19.5&32.0,74.9&12.8,8.91&0.845,0.597,0.438\\
30&34.465,34.458&0.197,0.197&53.7,53.9&3.55,3.39&-3.8,-3.6&30.0&21.3,31.7,30.6&6.07,5.79,5.46&10.3,11.2&162,160&58.7,88.6&86.1,191&7.21,5.73&0.782,0.568,0.531\\
$R_M$&34.471,34.458&0.215,0.215&53.5,53.7&2.81,2.32&-3.4,-3.4&18.9,18.6&5.03,9.23,7.44&4.76,4.45,2.44&25.5,25.3&-171,-172&5.81,10.3&7.66,26.1&49.1,25.2&0.996,0.667,0.240\\
$R_A$&34.471,34.458&0.215,0.215&53.5,53.7&2.81,2.32&-3.4,-3.4&26.2,26.1&8.91,16.2,14.8&4.95,4.84,3.85&15.7,16.7&-176,-178&18.8,33.9&17.3,57.5&26.7,16.4&0.923,0.643,0.482\\
$R_{1/2}$&34.471,34.458&0.215,0.215&53.5,53.7&2.81,2.32&-3.4,-3.4&19.6,19.2&5.39,9.79,7.91&4.79,4.51,2.54&24.5,24.9&-172,-173&6.67,11.7&8.48,28.5&47.2,25.4&0.987,0.667,0.261\\
\hline
\label{tab2}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}[h]
\centering
\tiny
\caption{Cont. Table \ref{tab2}}
\begin{tabular}{lcccccccccccccc}
\hline
IRDC & & & & & & & & & & & & & & \\
Case & $l_{\rm CM}$,$l_{\rm CM,\tau}$ & $b_{\rm CM}$,$b_{\rm CM,\tau}$ & $ v_{\rm CM}$,$v_{\rm CM,\tau}$ & $\Delta s$ & $\Delta v$ &R &$M$,$M_{\tau}$,$M_{G}$ & $\sigma$,$\sigma_{\tau}$,$\sigma_{G}$ & $\frac{dv_0}{ds}$,$\frac{dv_{0,\tau}}{ds}$ & $\theta_v$,$\theta_{v,\tau}$ & $I_0$,$I_{0,\tau}$ & $E_{\rm rot,0}$,$E_{\rm rot,0,\tau}$ & $\beta,\beta_{\tau}$ & $\log(\alpha_{\rm vir})$,$\log(\alpha_{\rm vir,\tau})$, \\
& & & & & & & $\times10^4 $ & & $\times \rm 10^{-2}$ & & $\times \rm 10^{3}$ & $\times 10^{46}$ & ($10^{-2}$) & $\log(\alpha_{\rm vir,G})$ \\
(pc) & ( $^{\circ}$) & ($^{\circ}$) & ($\kms$)& (pc) & ($\kms$)& (pc) & $({M_\odot})$ & ($\kms$) & (km/s/pc) & ($\deg$) & ($M_{\odot}$~pc$^2$) & (erg) & & \\
(1) & (2) & (3) &(4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) & (12) &(13) & (14) &(15) \\
\hline
\hline
F-s& & & & & & & & & & & & & & \\
5&34.428,34.422&0.240,0.240&57.7,57.5&0.89,1.25&0.0,0.0&5.00&0.664,1.21,1.18&2.51,2.21,1.77&8.85,9.76&-24.7,-21.8&0.0464,0.0838&0.503,1.68&0.718,0.472&0.742,0.368,0.192\\
10&34.415,34.403&0.221,0.221&58.2,58.0&2.25,2.86&0.4,0.4&10.0&1.89,3.13,2.79&2.89,2.67,1.88&1.87,2.17&9.94,-3.59&0.414,0.783&2.04,5.58&0.0706,0.0655&0.711,0.423,0.170\\
20&34.360,34.335&0.191,0.184&58.2,58.0&6.28,7.83&0.4,0.4&20.0&5.74,9.75,9.02&2.68,2.48,1.95&1.05,1.32&-105,-88.7&2.45,3.99&9.41,27.1&0.0287,0.0255&0.463,0.167,-0.00943\\
30&34.379,34.366&0.178,0.178&58.0,57.7&5.90,6.47&0.2,0.2&30.0&9.42,15.3,14.3&2.61,2.41,1.88&0.480,0.504&-129,-122&8.39,9.78&16.9,44.4&0.0114,0.00557&0.401,0.122,-0.0638\\
$R_M$&34.366,34.354&0.178,0.178&57.3,57.3&6.47,7.07&0.0,0.0&14.9,14.0&3.35,6.67,6.27&2.26,2.13,1.67&4.86,4.73&-84.3,-86.3&0.786,1.34&4.31,18.1&0.428,0.164&0.422,0.0453,-0.138\\
$R_A$&34.366,34.354&0.178,0.178&57.3,57.3&6.47,7.07&0.0,0.0&20.5,20.5&4.06,8.21,7.84&2.14,2.02,1.64&4.84,4.96&-80.3,-77.5&1.56,3.27&4.60,18.8&0.792,0.425&0.431,0.0733,-0.0877\\
$R_{1/2}$&34.366,34.354&0.178,0.178&57.3,57.3&6.47,7.07&0.0,0.0&12.3,10.9&2.84,5.44,5.04&2.33,2.20,1.65&2.60,2.20&-91.7,-104&0.458,0.602&3.73,15.5&0.0827,0.0187&0.437,0.0511,-0.162\\
\hline
G& & & & & & & & & & & & & & \\
5&34.758,34.758&-0.565,-0.571&43.9,44.1&0.44,0.88&0.4,0.9&5.00&1.05,1.61,1.58&3.58,3.23,2.85&47.1,46.0&47.0,50.8&0.0575,0.0804&1.25,2.96&10.2,5.71&0.853,0.577,0.475\\
10&34.764,34.764&-0.577,-0.590&44.6,44.6&0.93,1.59&1.1,1.3&10.0&3.16,4.47,4.14&4.57,4.14,3.08&12.1,12.7&-2.58,3.59&0.623,0.698&5.69,11.4&1.59,0.988&0.887,0.650,0.425\\
20&34.807,34.813&-0.614,-0.620&44.8,44.6&3.55,3.80&1.3,1.3&20.0&11.1,14.6,13.1&6.47,6.29,4.99&21.8,22.2&-161,-160&9.22,11.3&35.1,60.5&12.4,9.14&0.944,0.802,0.645\\
30&34.844,34.862&-0.571,-0.590&44.4,43.7&4.09,4.92&0.9,0.4&30.0&21.3,28.0,27.7&7.22,7.21,8.46&20.1,22.9&-161,-158&41.2,53.6&86.7,149&19.0,18.8&0.931,0.812,0.955\\
$R_M$&34.936,34.948&-0.602,-0.620&42.0,41.6&8.98,9.84&-1.5,-1.7&15.2,14.8&5.65,9.28,6.13&5.56,5.91,2.61&42.4,56.3&-128,-131&3.38,6.07&12.0,33.3&50.4,57.5&0.984,0.810,0.280\\
$R_A$&34.936,34.948&-0.602,-0.620&42.0,41.6&8.98,9.84&-1.5,-1.7&23.9,23.9&9.07,14.6,15.1&5.59,5.66,6.94&31.7,35.8&-134,-136&9.70,16.4&19.6,51.2&49.6,40.8&0.982,0.784,0.948\\
$R_{1/2}$&34.936,34.948&-0.602,-0.620&42.0,41.6&8.98,9.84&-1.5,-1.7&13.9,13.3&4.91,7.66,5.30&5.48,5.92,2.65&42.2,59.9&-124,-128&2.53,4.18&9.91,25.2&45.3,59.3&0.995,0.851,0.313\\
\hline
G-s& & & & & & & & & & & & & & \\
5&34.764,34.764&-0.571,-0.571&44.4,44.4&0.69,0.69&0.6,0.9&5.00&0.929,1.47,1.51&2.57,2.40,2.66&28.7,29.1&57.9,63.1&0.0487,0.0714&0.985,2.47&4.05,2.44&0.617,0.357,0.435\\
10&34.776,34.770&-0.583,-0.596&44.8,44.8&1.28,1.87&1.1,1.3&10.0&2.57,3.81,3.89&2.75,2.56,2.77&18.0,17.8&14.6,15.2&0.408,0.507&3.77,8.27&3.47,1.94&0.534,0.302,0.361\\
20&34.838,34.832&-0.590,-0.602&45.2,45.2&3.76,3.80&1.5,1.7&20.0&7.01,9.59,9.70&2.95,2.81,2.97&8.12,8.92&-18.9,-20.1&4.79,6.32&14.0,26.2&2.24,1.91&0.461,0.282,0.326\\
30&34.862,34.862&-0.528,-0.540&45.0,45.0&4.92,4.76&1.3,1.5&30.0&11.7,15.2,15.4&3.03,2.91,3.11&1.19,1.35&-26.1,-23.6&16.8,20.9&26.0,44.0&0.0909,0.0859&0.437,0.290,0.339\\
$R_M$&34.887,34.881&-0.547,-0.559&44.8,44.8&5.95,5.60&1.1,1.3&14.5,14.1&3.64,5.54,5.58&2.58,2.50,2.59&13.9,15.3&-7.89,-13.1&1.84,2.90&5.22,12.4&6.76,5.41&0.490,0.266,0.295\\
$R_A$&34.887,34.881&-0.547,-0.559&44.8,44.8&5.95,5.60&1.1,1.3&21.8,21.8&5.36,7.95,7.94&2.57,2.48,2.50&2.38,2.57&35.9,14.1&2.41,3.91&7.55,16.6&0.180,0.154&0.493,0.291,0.298\\
$R_{1/2}$&34.887,34.881&-0.547,-0.559&44.8,44.8&5.95,5.60&1.1,1.3&13.1,12.5&3.25,4.79,4.87&2.62,2.53,2.70&16.1,17.0&-4.82,-7.47&1.48,2.16&4.59,10.5&8.32,5.91&0.508,0.288,0.337\\
\hline
H& & & & & & & & & & & & & & \\
5&35.428,35.422&-0.313,-0.319&48.6,48.2&1.12,0.88&1.1,1.3&5.00&0.970,1.52,0.789&6.53,6.18,1.12&51.8,57.5&-87.4,-81.5&0.0500,0.0761&1.07,2.65&12.4,9.46&1.41,1.16,-0.0351\\
10&35.422,35.422&-0.288,-0.288&48.8,48.4&1.97,2.27&1.3,1.5&10.0&3.27,5.02,2.72&6.82,6.53,1.54&25.0,28.7&-68.9,-68.6&0.662,1.04&6.10,14.4&6.73,5.91&1.22,0.996,0.00640\\
20&35.403,35.397&-0.251,-0.257&49.2,49.0&3.75,3.79&1.7,2.1&20.0&9.90,14.5,6.60&7.16,6.95,2.12&31.6,33.1&-52.4,-53.3&9.80,14.7&27.9,59.6&34.8,26.8&1.08,0.890,0.198\\
30&35.367,35.354&-0.282,-0.307&47.8,46.9&3.08,3.06&0.2,0.0&30.0&20.9,31.0,31.0&7.92,8.05,13.2&32.7,35.7&-53.9,-54.6&53.7,84.0&82.8,183&69.0,58.3&1.02,0.863,1.29\\
$R_M$&35.348,35.336&-0.337,-0.362&46.5,45.6&3.48,4.34&-0.2,-0.6&18.6,18.6&4.52,8.30,4.91&5.95,5.90,1.60&40.4,40.5&-56.2,-60.0&4.55,8.61&6.27,21.1&118,66.4&1.23,0.958,0.0548\\
$R_A$&35.348,35.336&-0.337,-0.362&46.5,45.6&3.48,4.34&-0.2,-0.6&25.7,25.7&8.22,15.8,7.07&7.41,7.53,1.94&48.5,49.9&-52.9,-54.4&18.5,37.6&15.0,55.1&288,169&1.30,1.03,0.204\\
$R_{1/2}$&35.348,35.336&-0.337,-0.362&46.5,45.6&3.48,4.34&-0.2,-0.6&20.0,19.9&5.04,9.23,5.14&6.19,6.30,1.62&43.7,45.9&-56.1,-59.7&5.89,11.1&7.25,24.4&154,95.4&1.25,0.998,0.0716\\
\hline
H-s& & & & & & & & & & & & & & \\
5&35.428,35.422&-0.319,-0.325&43.9,43.9&1.56,1.12&0.2,0.0&5.00&0.478,0.848,0.781&1.50,1.37,1.10&13.9,15.5&-51.8,-46.3&0.0321,0.0616&0.260,0.820&2.38,1.80&0.436,0.108,-0.0466\\
10&35.416,35.410&-0.294,-0.301&43.9,43.9&2.27,1.89&0.2,0.0&10.0&1.50,2.64,2.62&1.57,1.47,1.46&16.2,16.6&-57.8,-56.0&0.397,0.740&1.29,3.99&8.03,5.09&0.283,-0.0205,-0.0250\\
20&35.330,35.330&-0.288,-0.294&43.7,43.7&4.49,4.32&0.0,-0.2&20.0&3.36,5.54,5.60&1.69,1.60,1.68&7.84,7.66&-47.5,-48.3&2.79,4.42&3.22,8.74&5.31,2.95&0.298,0.0293,0.0677\\
30&35.244,35.244&-0.276,-0.282&43.7,43.7&8.67,8.56&0.0,-0.2&30.0&5.99,9.34,9.53&1.81,1.72,1.86&2.84,2.66&-11.7,-14.5&9.29,14.4&6.82,16.6&1.09,0.612&0.279,0.0457,0.105\\
$R_M$&35.231,35.231&-0.294,-0.301&43.7,43.7&8.98,8.61&-0.2,-0.2&13.4,13.2&1.83,3.38,3.47&1.40,1.36,1.50&5.55,4.94&-14.3,-9.80&0.994,1.83&1.44,4.93&2.12,0.901&0.223,-0.0746,-0.00303\\
$R_A$&35.231,35.231&-0.294,-0.301&43.7,43.7&8.98,8.61&-0.2,-0.2&19.2,19.3&2.98,5.54,5.60&1.48,1.44,1.52&3.96,3.86&-8.14,-17.6&3.25,6.19&2.64,9.06&1.92,1.01&0.216,-0.0739,-0.0344\\
$R_{1/2}$&35.231,35.231&-0.294,-0.301&43.7,43.7&8.98,8.61&-0.2,-0.2&13.1,12.9&1.78,3.23,3.32&1.40,1.35,1.49&5.61,5.11&-15.5,-10.2&0.929,1.64&1.38,4.62&2.10,0.921&0.224,-0.0705,0.00264\\
\hline
I& & & & & & & & & & & & & & \\
5&38.936,38.936&-0.460,-0.460&44.3,42.6&1.04,1.04&0.6,0.4&5.00&0.814,1.47,1.27&5.92,4.47,1.21&1.76,14.9&-100.,130&0.0423,0.0671&0.756,2.47&0.0172,0.600&1.40,0.898,-0.173\\
10&38.930,38.930&-0.454,-0.460&44.9,43.2&1.45,1.30&1.3,1.1&10.0&2.25,3.58,2.71&6.33,5.26,1.42&11.9,6.46&-39.0,-10.1&0.452,0.588&2.90,7.32&2.21,0.333&1.32,0.955,-0.0629\\
20&38.942,38.942&-0.393,-0.411&46.9,45.4&3.81,2.96&3.2,3.2&20.0&5.15,6.96,3.86&6.87,6.42,1.61&24.3,22.6&-71.4,-67.8&2.69,3.54&7.57,13.8&20.8,13.1&1.33,1.14,0.194\\
30&38.961,38.961&-0.362,-0.380&47.1,45.6&5.22,4.36&3.4,3.4&30.0&6.81,8.96,4.35&7.29,6.94,1.62&13.5,15.1&-18.3,-10.2&9.39,11.1&8.83,15.3&19.4,16.6&1.43,1.27,0.326\\
$R_M$&38.936,38.936&-0.460,-0.467&41.3,41.1&1.04,0.91&0.0,-0.2&6.70,6.18&0.789,1.61,1.59&1.42,1.34,1.27&24.8,25.9&91.4,89.1&0.0849,0.152&0.531,2.40&9.80,4.24&0.297,-0.0932,-0.137\\
$R_A$&38.936,38.936&-0.460,-0.467&41.3,41.1&1.04,0.91&0.0,-0.2&11.6,11.6&1.35,2.71,2.58&1.79,1.67,1.32&3.43,5.61&139,112&0.269,0.696&0.891,3.61&0.353,0.604&0.506,0.145,-0.0381\\
$R_{1/2}$&38.936,38.936&-0.460,-0.467&41.3,41.1&1.04,0.91&0.0,-0.2&6.40,5.70&0.755,1.47,1.46&1.41,1.29,1.24&25.1,25.9&89.1,84.0&0.0754,0.120&0.509,2.17&9.33,3.66&0.289,-0.122,-0.158\\
\hline
J& & & & & & & & & & & & & & \\
5&53.139,53.145&0.056,0.050&21.6,21.6&0.77,0.98&0.4,0.4&5.00&0.434,0.794,0.679&5.44,4.61,1.11&48.3,50.5&44.6,41.5&0.0222,0.0403&0.215,0.720&23.9,14.2&1.60,1.19,0.0203\\
10&53.164,53.164&0.025,0.025&22.0,22.0&1.82,1.82&0.9,0.9&10.0&0.881,1.53,1.31&5.16,4.42,1.17&20.0,18.1&33.2,37.5&0.132,0.222&0.443,1.34&11.8,5.37&1.55,1.17,0.0849\\
20&53.250,53.280&0.062,0.050&22.7,22.7&4.25,5.22&1.5,1.5&20.0&1.65,3.05,2.64&4.87,4.17,1.22&12.2,10.9&-21.1,-19.0&1.21,2.26&0.775,2.65&23.2,10.1&1.52,1.12,0.119\\
30&53.293,53.324&0.124,0.105&22.7,22.7&5.99,6.75&1.5,1.5&30.0&2.31,4.09,3.35&5.28,4.59,1.24&6.84,6.59&10.8,9.23&2.83,4.71&1.01,3.18&13.0,6.40&1.62,1.25,0.203\\
$R_M$&53.317,53.342&0.013,0.019&22.7,22.7&6.52,7.24&1.5,1.5&9.89,9.34&0.697,1.47,1.50&1.10,1.06,1.12&12.4,13.1&-37.2,-44.5&0.189,0.353&0.280,1.33&10.3,4.54&0.301,-0.0818,-0.0402\\
$R_A$&53.317,53.342&0.013,0.019&22.7,22.7&6.52,7.24&1.5,1.5&13.1,13.1&0.939,2.00,2.04&1.11,1.08,1.15&8.78,8.75&-14.4,-14.9&0.483,1.03&0.385,1.74&9.62,4.53&0.301,-0.0541,-0.00341\\
$R_{1/2}$&53.317,53.342&0.013,0.019&22.7,22.7&6.52,7.24&1.5,1.5&9.20,8.60&0.629,1.28,1.30&1.08,1.06,1.12&13.3,15.1&-40.7,-50.1&0.137,0.227&0.246,1.09&9.85,4.75&0.301,-0.0549,-0.0138\\
\hline
\label{tab2c}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}[h]
\centering
\tiny
\caption{IRDC Physical Parameters Estimates}
\begin{tabular}{cccccccccccccc}
\hline
IRDC & Case & $l_{\rm CM}$,$l_{\rm CM,\tau}$ & $b_{\rm CM}$,$b_{\rm CM,\tau}$ & $ v_{\rm CM}$,$ v_{\rm CM,\tau}$ & $R$ &$M$,$M_{\tau}$,$M_{G}$ & $\sigma$,$\sigma_{\tau}$,$\sigma_{G}$ & $\frac{dv_0}{ds}$,$\frac{dv_{0,\tau}}{ds}$ & $\theta_v$,$\theta_{v,\tau}$ & $I_0$,$I_{0,\tau}$ & $E_{\rm rot,0}$,$E_{\rm rot,0,\tau}$ & $\beta,\beta_{\tau}$ & $\log(\alpha_{\rm vir})$,$\log(\alpha_{\rm vir,\tau})$, \\
& & & & & & $\times10^4 $ & & $\times \rm 10^{-2}$ & & $\times \rm 10^{3}$ & $\times 10^{46}$ & ($10^{-2}$) & $\log(\alpha_{\rm vir,G})$ \\
& & ( $^{\circ}$) & ($^{\circ}$) & ($\kms$)& (pc) & $({M_\odot})$ & ($\kms$) & (km/s/pc) & ($\deg$) & ($M_{\odot}$~pc$^2$) & (erg) & & \\
(1) & (2) & (3) &(4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) & (12) &(13) & (14)\\
\hline
\hline
A&SE&18.831,18.844&-0.282,-0.282&64.3,64.3&5.18&1.07,1.65,1.41&4.20,3.60,1.79&24.6,27.9&117,117&143,214&123,293&6.97,5.66&0.998,0.675,0.136\\
&CE&18.838,18.844&-0.282,-0.276&65.0,64.8&&0.802,1.36,1.30&2.18,1.97,1.63&27.1,28.7&122,121&114,187&69.4,199&12.0,7.73&0.553,0.236,0.0890\\
\hline
B&SE&19.268,19.274&0.075,0.075&26.5,26.1&1.36&0.0891,0.178,0.170&3.96,3.06,1.51&37.6,35.2&-93.6,-93.5&0.102,0.204&3.20,12.8&0.446,0.197&1.44,0.920,0.327\\
&CE&19.274,19.274&0.075,0.075&25.9,25.9&&0.0742,0.161,0.165&1.41,1.33,1.46&22.8,22.5&40.5,40.3&0.712,1.52&2.22,10.5&1.66,0.727&0.625,0.239,0.307\\
\hline
C&SE&28.370,28.370&0.068,0.068&78.6,78.6&7.68&3.58,6.19,5.96&4.09,3.52,2.83&10.5,10.6&-98.4,-102&371,599&947,2830&0.429,0.238&0.621,0.254,0.0788\\
&CE&28.370,28.370&0.068,0.068&78.1,78.4&&3.07,5.63,5.68&2.78,2.53,2.63&14.7,15.0&-92.7,-93.8&309,527&698,2340&0.956,0.506&0.351,0.00859,0.0370\\
\hline
D&SE&28.567,28.573&-0.233,-0.233&80.5,81.5&8.46&3.54,5.67,3.27&7.46,7.09,1.90&5.66,5.40&-4.15,37.0&1490,528&822,2110&0.577,0.0727&1.19,0.941,0.0380\\
&CE&28.579,28.579&-0.227,-0.227&81.3,82.4&&2.37,4.29,2.93&7.07,6.53,1.71&22.3,32.3&101,107&387,727&368,1210&5.21,6.25&1.32,0.990,-0.00652\\
\hline
D-s&SE&28.573,28.573&-0.233,-0.233&87.3,87.1&8.46&1.63,3.05,3.05&2.06,1.75,1.72&6.34,7.73&113,114&363,584&174,612&0.834,0.568&0.408,-0.00586,-0.0187\\
&CE&28.579,28.579&-0.233,-0.233&86.9,86.9&&1.35,2.77,2.84&1.58,1.49,1.63&24.3,24.5&121,120&259,481&120,502&12.7,5.71&0.260,-0.0996,-0.0343\\
\hline
E&SE&28.678,28.678&0.130,0.130&79.8,80.1&5.74&1.29,1.86,1.52&6.38,5.97,3.05&40.7,41.9&71.4,69.5&120,196&163,336&12.2,10.2&1.32,1.11,0.611\\
&CE&28.672,28.678&0.130,0.124&79.8,79.8&&0.770,1.26,1.21&3.58,3.60,2.41&20.0,18.5&57.4,53.4&136,251&57.8,155&9.32,5.53&1.05,0.838,0.506\\
\hline
F&SE&34.440,34.440&0.246,0.246&57.5,57.5&1.75&0.160,0.294,0.255&3.74,3.12,1.67&26.1,26.9&26.5,23.8&2.38,4.67&8.25,28.0&1.95,1.20&1.25,0.831,0.348\\
&CE&34.440,34.440&0.246,0.246&56.9,57.1&&0.122,0.250,0.243&1.82,1.69,1.55&46.8,46.2&-29.6,-31.9&2.29,4.58&4.82,20.3&10.4,4.79&0.742,0.367,0.305\\
\hline
F-s&SE&34.440,34.440&0.246,0.246&57.7,57.5&1.75&0.127,0.252,0.247&2.10,1.81,1.58&20.7,21.1&-23.4,-23.5&2.63,5.21&5.20,20.6&2.15,1.12&0.851,0.424,0.315\\
&CE&34.440,34.440&0.246,0.246&57.3,57.3&&0.111,0.235,0.239&1.44,1.40,1.50&32.7,34.6&-26.5,-25.9&2.18,4.66&4.01,17.9&5.80,3.10&0.580,0.230,0.286\\
\hline
G&SE&34.764,34.770&-0.559,-0.559&43.5,43.3&1.53&0.139,0.249,0.244&3.20,2.75,2.36&54.8,49.7&61.4,60.4&0.882,1.50&7.06,22.7&3.73,1.62&1.12,0.734,0.610\\
&CE&34.764,34.764&-0.559,-0.559&43.5,43.3&&0.127,0.236,0.239&2.41,2.17,2.28&70.5,67.1&67.0,66.5&0.628,1.09&5.94,20.4&5.22,2.40&0.912,0.550,0.590\\
\hline
G-s&SE&34.770,34.770&-0.559,-0.559&43.7,43.5&1.53&0.126,0.234,0.236&2.42,2.16,2.24&33.2,33.9&72.1,71.1&0.483,0.861&5.87,20.0&0.902,0.490&0.915,0.551,0.579\\
&CE&34.770,34.770&-0.559,-0.559&43.7,43.5&&0.122,0.228,0.233&2.23,2.03,2.20&36.0,36.0&69.9,69.5&0.514,0.902&5.42,19.1&1.22,0.610&0.865,0.507,0.569\\
\hline
H&SE&35.410,35.410&-0.325,-0.331&47.5,46.9&4.84&0.924,1.56,1.05&6.12,5.64,1.42&81.4,83.0&-60.2,-60.5&134,232&99.3,282&88.9,56.3&1.36,1.06,0.0358\\
&CE&35.416,35.416&-0.325,-0.331&46.7,46.3&&0.641,1.24,0.940&5.05,4.79,1.24&105,106&-61.6,-61.5&94.5,187&47.7,179&218,116&1.35,1.02,-0.0359\\
\hline
H-s&SE&35.403,35.403&-0.337,-0.337&43.7,43.9&4.84&0.560,1.06,1.04&1.54,1.43,1.40&30.6,29.4&-56.0,-56.2&95.8,183&36.5,130&24.6,12.2&0.380,0.0397,0.0277\\
&CE&35.403,35.397&-0.337,-0.337&43.9,43.9&&0.481,0.966,0.940&1.34,1.30,1.24&41.9,40.8&-56.2,-56.2&84.5,171&26.9,108&55.0,26.1&0.322,-0.00712,-0.0359\\
\hline
I&SE&38.954,38.954&-0.473,-0.473&43.7,42.2&1.86&0.187,0.396,0.361&5.33,3.56,1.06&17.2,28.6&150,137&3.17,5.27&10.6,47.7&0.871,0.896&1.52,0.842,-0.169\\
&CE&38.954,38.954&-0.473,-0.473&41.3,41.3&&0.143,0.353,0.356&1.10,1.01,1.04&27.6,30.0&-23.1,-23.3&2.20,4.93&6.24,38.1&2.68,1.16&0.260,-0.206,-0.181\\
\hline
J&SE&53.114,53.114&0.056,0.056&21.2,21.2&0.378&0.00333,0.00734,0.00718&2.30,1.75,0.728&217,208&5.45,5.68&0.0398,0.0780&0.0166,0.0809&1130,416&1.84,1.26,0.512\\
&CE&53.114,53.114&0.056,0.056&21.2,21.2&&0.00298,0.00693,0.00707&0.698,0.654,0.711&1.18,7.40&66.9,171&0.000768,0.0319&0.0133,0.0721&0.000795,0.241&0.857,0.434,0.497\\
\hline
\label{tabirdc}
\end{tabular}
\end{sidewaystable}
\section{Results}
\subsection{Molecular Cloud Physical Properties}
\label{S:physprop}
Table \ref{tab2} presents the derived physical properties for all 10
clouds. The global properties of each cloud were evaluated for data
extracted out to seven different radii, shown on different lines in
the table: 5, 10, 20, and 30 pc using SE; and out to mass weighted
($R_M$), areal ($R_A$), and half-mass ($R_{1/2}$) radii using CE. Each
cloud definition case is noted in column 1. The center-of-mass
coordinates of each cloud are listed in columns 2 and 3, evaluated for
the optically thin and opacity-corrected cases.
We note that the CE cloud definitions with $R_M$, $R_A$ and $R_{1/2}$
share the same center. Mean velocities are reported in column 4.
Columns 5 and 6 list the separations between the GMC centers and the
IRDC centers in plane of sky distance and velocity offset,
respectively, again for both optically thin and opacity-corrected methods.
Column 7 lists the cloud radii (with $R_M$, $R_A$ and $R_{1/2}$ cases
having two values for the optical thin and opacity-corrected
cases). Column 8 lists the three cloud mass estimates: optically thin
mass ($M$), opacity-corrected mass ($M_{\tau}$), and the mass
estimated from a gaussian fit of the co-added opacity corrected column
density spectra ($M_G$). Column 9 lists the three velocity dispersions
measured: the standard deviations of the cloud's optically thin and
opacity corrected column density spectra ($\sigma$, $\sigma_{\tau}$)
and the dispersion estimated from the gaussian fitted opacity
corrected column density spectra ($\sigma_{G}$). Column 10 presents
the two velocity gradient magnitudes, one estimated from the optically
thin column density weighted first-order moment map ($dv_0/ds$) and
one estimated from the opacity-corrected column density weighted
first-order moment map ($dv_{0,\tau}/ds$).
Column 11 presents the position angles of the projected cloud rotation
axes estimated from each cloud's angular momentum for both optically
thin and opacity corrected masses ($\theta_v$ and $\theta_{v,\tau}$). The
projected moments of inertia estimated using the optically thin mass
($I_0$) and opacity corrected mass ($I_{0,\tau}$) are presented in column
12. Column 13 shows projected rotational energies, $E_{\rm rot,0}$ and
$E_{\rm rot,0,\tau}$, and column 14 shows the ratios of these to the
gravitational energy, i.e., $\beta$ and $\beta_\tau$. Finally, column
15 shows log$_{10}$ of the virial parameters of the clouds using the three
measures of mass, i.e., $\alpha_{\rm vir}$, $\alpha_{\rm vir,\tau}$, $\alpha_{\rm vir,G}$.
A visual overview of each GMC-scale region is given in Figure
\ref{mapA} and Figures \ref{mapB}-\ref{mapJ}. The left side of the
figures display a series of images of the molecular gas out to a
radius of 30~pc. The first row displays the GLIMPSE $8\micron$ image
and the $\thco$ mass surface density map, $\Sigma_{\rm 13CO}$,
generated using the velocity interval given solely by the IRDC. The
IRDC elliptical boundary is also shown \citep[][]{Simon2006a}. The
subsequent rows display the $\Sigma_{\rm 13CO}$ maps assuming our
defined velocity interval of $v_0\pm15~\kms$, the column-density
weighted mean velocity (first moment), the linear momentum map derived
from each pixel's column density and velocity with respect to the
cloud center of mass, and the velocity dispersion map (second moment)
for the SE cloud (\textit{Left column}) and CE cloud (\textit{Right
column}). The right side of the figures display the optically thin
and opacity-corrected co-added column density velocity spectra at
various extraction radii for the SE and CE clouds and the SE IRDC.
Four of the clouds (D, F, G, and H) were found to contain multiple
velocity components, based on the co-added $\thco$ column density
spectra of their SE 60 pc diameter GMCs (e.g., Cloud D; Fig. \ref{mapD}). We
used a series of integrated intensity maps to isolate the velocity
range of the molecular gas associated with the IRDC
(Fig. \ref{chanmaps}). In general, these velocity ranges span
$\sim10~\kms$ (Cloud D: $84.1-93.7~\kms$; Cloud F: $54.3-65.5~\kms$;
Cloud G: $39.7-51.8~\kms$; Cloud H: $39.7-49.9~\kms$). To investigate
how the cloud's derived physical properties were affected by isolating
only these velocity ranges, we include these four cases as separate
entries in Table \ref{tab2} (noted with ``-s''). For all subsequent
analyses, the ``-s'' values for these four clouds are used in place of those
derived using the full velocity range of $v_0\pm15~\kms$.
\subsection{IRDC and GMC Morphologies}
\label{S:loc}
\subsubsection{Description of Individual IRDCs/GMCs}
\begin{figure}[ht!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=7.61in, angle=0]{Fig3-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption{
Cloud A: \textit{Left panel maps:} {\it Top row:} {\it Left:} GLIMPSE
8$\rm \mu$m image. {\it Right:} $\thco$ mass surface density
($\Sigma$) map over the velocity range of the IRDC emission profile
$v_{\rm lsr}=54.6-77.1~\kms$ (ellipse denotes IRDC boundary). {\it 2nd
row:} {\it Left:} SE $\thco$-derived $\Sigma$ map (circles denote 5,
10, 20~pc boundaries). {\it Right:} CE $\thco$-derived $\Sigma$
map. {\it 3rd row:} {\it Left:} SE $\thco$ mean velocity (first
moment) map. {\it Right:} CE $\thco$ mean velocity (first moment)
map
relative to $v_{0}$ defined from the IRDC.
The arrows show the direction of total (mass-weighted) line-of-sight
velocity gradients, $dv_0/ds$. An arrow is shown for each extraction
radius (defining the length of the arrow) and is centered on either
the extracted cloud's center of mass (for SE cases) or the $R=30$ pc cloud's
center of mass (for CE cases).
{\it 4th row:} {\it Left:} SE $\thco$-derived (optically thin) linear
momentum map. {\it Right:} CE $\thco$-derived (optically thin) linear
momentum map.
The arrows show the position angle, $\theta_v$, of the cloud's
rotation axis, i.e., orthogonal to the direction of the velocity
gradient, for each extracted cloud (with arrow length showing cloud
radius) {\it 5th row:} {\it Left:} SE $\thco$ velocity dispersion
(second moment; optically thin) map. {\it Right:} CE $\thco$ velocity
dispersion (second moment; optically thin) map. All maps in Rows 2-5
cover the velocity interval of $v_0\pm15~\kms$, with $v_0$ defined by
the emission of the IRDC. Omitted pixels, those that fall below the
5$\sigma_{\rm rms}$ level, are shown in gray.
\textit{Right panel velocity distributions:}
In each panel, the column density distributions with velocity are
derived assuming optically thin conditions (blue solid line) or with
opacity correction (red solid line) from the $\thco$(1-0) co-added
spectra. Gaussian fits are shown by the dotted blue and red lines. In all panels,
$\sigma$ and $\sigma_G$ are noted in the upper left corner of the
plot. {\it Left column:} {\it Top to bottom:} SE out to $R=5, 10, 20,
30$~pc. {\it Right column:} {\it Top to bottom:} SE of IRDC ellipse,
CE out to $R_M$, $R_A$, $R_{1/2}$.
}
\label{mapA}
\end{figure}
\underbar{\textit{Cloud A:}} We estimate a cloud areal radius of
$R_A\simeq 27$~pc and a mass of $M_{G,A}\simeq 2.7\times10^5~\rm
M_{\odot}$. The GRS $\thco$ GMC study by RD10 identified five molecular clouds within the extent of Cloud A, with a
total $\thco$ mass of $2.1\times10^5~\rm M_{\odot}$ (note, we scaled
all $\thco$ masses of RD10 by a factor 0.49 to
convert to our assumed $\thco$/\htwo ~abundance ratio; see
\S\ref{S:13COmass}). IRDC A is located about 20~pc north of its parent
GMC's highest $\Sigma$ clump, which appears bright at $\rm 8\:\mu m$
and is thus likely to be star-forming.
The WISE \ion{H}{2}~ region catalog \citep{Anderson2014} lists the W39 \ion{H}{2}~
region to be within $\sim3\arcmin$ of this clump's center, and with a
similar LSR velocity ($v_{\rm LSR}=65.5~\kms$). Another \ion{H}{2}~ region
lies in a smaller dense clump at $l=19.07^\circ$, $b=-0.28^\circ$ with
velocity and distance consistent with those of Cloud A. The isolated,
simple $\rm ^{13}CO$ emission profiles, along with the spatial
connection between the IRDC and the \ion{H}{2}~ regions, demonstrated
through CE, indicates that GMC A contains clumps at a variety of
stages of star formation activity.
\underbar{\textit{Cloud B:}} We estimate a radius of $R_{A}\simeq 12$~pc
and a mass of
$M_{G,A}\simeq3.3\times10^4\:M_{\odot}$. RD10
estimated a $\thco$ mass of $2.4\times10^4~\rm M_{\odot}$ and a
smaller projected areal radius of 6.3 pc.
The IRDC is located within the densest region of the GMC, which
extends $\sim10$ pc to the south-east of the IRDC and accounts
for most of the cloud's mass. Spatially, Cloud B is
well defined and completely isolated at an extraction radius of 20~pc
in CE. Although the cloud is isolated in velocity space, the
column density profile indicates a low level of $\thco$ emission at
neighboring velocities, both lower and higher.
Although the GLIMPSE $8\mu$m emission map indicates a large bright
emission structure to the west of the cloud, this feature is not at
velocities consistent with Cloud B ($v_{\rm LSR}\sim 50\kms$).
\begin{figure}[ht!]
\begin{center}$
\begin{array}{cc}
\includegraphics[width=2.51in, angle=-90]{Fig4a-eps-converted-to.pdf} & \includegraphics[width=2.51in, angle=-90]{Fig4c-eps-converted-to.pdf} \\
\includegraphics[width=2.51in, angle=-90]{Fig4b-eps-converted-to.pdf} & \includegraphics[width=2.51in, angle=-90]{Fig4d-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption{$\thco$ channel maps of the multi-component clouds. \textit{top left}: Cloud D; \textit{top right}: Cloud F; \textit{bottom left} Cloud G; \textit{bottom right}: Cloud H. The velocity interval of each panel is shown in the upper-right corner of each map. }
\label{chanmaps}
\end{figure}
\underbar{\textit{Cloud C:}} We estimate a GMC radius of $R_A\simeq
29$~pc and a mass of $M_{G,A}\simeq
3.5\times10^5\:M_{\odot}$. RD10 derived a radius of
20.9~pc and a mass of $\sim3\times10^5~\rm M_{\odot}$. IRDC C is
located within the dense central region of its parent GMC. The
integrated intensity maps indicate dense and clumpy structure
throughout the cloud (Fig. \ref{mapC}). The velocity distribution
shows a broad centrally-peaked profile, but with potential
substructure. Additionally, there is a low level of $\thco$ emission
at higher velocities. \citet{Butler2014} studied the IRDC C using
near- and mid-infrared extinction mapping to probe the high-dynamic
range of its mass surface density (down to $A_V\sim1$ mag), finding an
IRDC mass of $\sim7\times10^4\:M_{\odot}$. This is one of the most
massive, high-$\Sigma$ IRDCs known, with the potential to form a
massive star cluster. Using the same IRDC elliptical boundary from
\citet{Simon2006a}, we estimate a $\thco$ mass of
$5.7\times10^4\:M_{\odot}$. Our slightly lower mass estimate may be
the result of abundance uncertainties and/or CO depletion.
\underbar{\textit{Cloud D:}} For the GMC at $v_{\rm LSR}=84.1-93.7~\kms$,
we estimate a radius of $R_A\simeq 21$~pc and a mass of $M_{G,A} \simeq
9.3\times10^4\:M_{\odot}$. RD10 identified two
clouds within the Cloud D-s region with a total $\thco$ mass of
$\sim5.2\times10^4~\rm M_{\odot}$. The IRDC is located in the dense
central region of the GMC. However, the $\thco$ spectra show four
cloud components within $v_0\pm15~\kms$. Although the $\thco$
emission corresponding to the IRDC is evident within a radius of 5~pc,
a lower-velocity gas component begins to dominate on scales out to
about 20~pc. The position-velocity map (Fig. \ref{PV1}) shows there is a
significant amount of $\thco$ emission at lower velocities,
$v_{\rm LSR}=63.5$-$84.1~\kms$, that extends largely over the northern
region of the cloud. In the same velocity interval of the IRDC,
$v_{\rm LSR}=84.1-93.7~\kms$, a substantial molecular gas clump contains
the largest integrated peak intensity ($\sim50$~K~$\kms$) at the
western edge of the 30~pc extraction radius. Figure \ref{chanmaps}
suggests that this emission is only tenuously connected to the $\thco$
emission associated with the main GMC.
\underbar{\textit{Cloud E:}} We estimate a GMC radius of $R_A\simeq
28$~pc and a mass of $M_{G,A}\simeq 3.1\times10^5\:M_{\odot}$. No clouds
within the Cloud E spatial and velocity ranges were listed in the
RD10 catalog. The integrated emission map indicates
that the GMC is highly substructured. The IRDC is contained within a
small dense clump central to its parent GMC, but more massive clumps
are seen on scales beyond 5~pc. CE finds an extended region of clumpy
$\thco$ gas in the cloud's south-west region, beginning at 10~pc and
extending beyond the 30~pc boundary.
\underbar{\textit{Cloud F:}} For the cloud in the velocity range
$v_{\rm LSR}=54.3-65.6~\kms$, we estimate a radius of $R_A\simeq21$~pc
and a mass of $M_{G,A}\simeq7.8\times10^4\:M_{\odot}$, assuming a
distance of 3.7~kpc \citep{Rathborne2006}. Note, \citet{Kurayama2011}
estimated a parallax distance $1.56\pm0.12$~kpc, however this result
has been called into question by
\citet{Foster2012}. RD10 found two GMCs associated
with Cloud F: one cloud matches the location and velocity interval of
Cloud F's western massive clump, having a mass of
$\simeq6.9\times10^4\:M_{\odot}$ and radius of 12.3~pc. The other
molecular cloud is associated with the smaller eastern clump, for
which they estimate a mass of $\sim2.2\times10^4~\rm M_{\odot}$. Our
analysis recovers this larger total mass, $\sim 10^5\:M_\odot$, if we
include all the $\thco$ emission within $v_0\pm15\:\kms$.
The larger molecular cloud complex surrounding Cloud F was recently
cataloged within the Giant Molecular Filament (GMF) study of
\citet{Ragan2014}. They found that their filament GMF38.1-32.4 is
actually the projection of two overlapping filamentary structures.
Cloud F is within the further filament (3.0-3.7~kpc), which has a
total length of 232~pc. The GMF contains multiple infrared dark
regions (identified in both GLIMPSE $8\mu$m and HiGAL $250\mu$m
images) and several dense gas clumps (identified by the
\citet{Wienen2012} $\rm NH_3$ and Bolocam Galactic Plane surveys),
demonstrating that overall it is in an early phase of star
formation. However, in Cloud F alone, there are two \ion{H}{2}~regions
within the two densest parts of the cloud with consistent LSR
velocities and distances, at $l=34.256, b=0.136$ and $l=34.821,
b=0.351$ \citep{Anderson2014}. While IRDC~F, which is also highly
filamentary, contains some of the highest column densities in the GMC,
the highest $\Sigma$ clump is actually another feature extending west
of the IRDC between 5 and 25~pc (Fig. \ref{mapF}). The velocity
distribution of the GMC for material out to 30~pc indicates that there
are four cloud components within our defined velocity interval
$v_0\pm15~\kms$. Figure \ref{PV2} shows the PV maps of the cloud out
to 30 pc and indicates that the majority of the $\thco$ emission
resides within the velocity interval of the IRDC
($v_{\rm LSR}=54.3-65.6~\kms$) and is connected to the weaker molecular
emission at surrounding velocities. Additionally, the massive clump
and IRDC are kinematically associated with another dense clump on the
eastern edge of the cloud at $\sim25$ pc. The cloud defined by CE out
to a radius of 30 pc indicates that the regions above and below the
IRDCs parent cloud are filled with clumpy molecular material related
to the other three emission profiles.
\underbar{\textit{Cloud G:}} For the material in the velocity range of
$v_{\rm lsr}=39.7-51.8~\kms$ we estimate a cloud radius of $R_A\simeq
22$~pc and a mass of $M_{G,A}\simeq8.0\times10^4\:M_{\odot}$.
RD10 identified two clouds within the Cloud G
boundary with a total $\thco$ mass of
$\sim5.7\times10^4\:M_{\odot}$.
At small cloud radii, $<10$~pc, Cloud G's velocity distribution is
reasonably well described by a single peak. However, at larger radii
we clearly see three separate components within $v_0\pm15~\kms$. The
IRDC, and its parent GMC, are confined to the central velocity
interval of $v=39.7-51.8~\kms$. In the GMC defined by SE, the IRDC
lies at the northern edge of the cloud's densest region, which extends
south $\sim18$ pc. For the cloud defined by CE, this same cloud
region breaks up into three smaller clumps. The lower velocity
component of the cloud contains a dense molecular clump at
$l\simeq35.2,b\simeq-0.7$, a region which also contains a high level
of mid-infrared emission at $8\mu$m (Fig. \ref{mapG}). The WISE \ion{H}{2}~region
database lists a candidate \ion{H}{2}~region near the center of the
low-velocity clump with an estimated parallax distance of 2.2 kpc,
versus the Cloud G distance of 2.9~kpc.
\underbar{\textit{Cloud H:}} For gas in the velocity range of $v_{\rm
lsr}=39.7-47.1~\kms$, we estimate a cloud radius of $R_A\simeq19$~pc
and a mass of $M_{G,A}\simeq5.6\times10^4\:M_{\odot}$. RD10
identified three molecular clouds at the location of Cloud H, with a
total mass of $\simeq2.8\times10^4\:M_{\odot}$. The PV diagrams
indicate that there are three separate cloud components within
$v_0\pm15\:\kms$ (Fig. \ref{PV2}). The IRDC and its parent GMC are
confined to the velocity interval of $v=39.7-47.1~\kms$, in which the
IRDC is the main structure of a larger diffuse cloud extending to the
west. There are additional dense gas clumps at the neighboring
velocities: a large coherent clump about $25$~pc to the south at
$v=32.5-39.7~\kms$ (part of Cloud G), and a high concentration of
dense, clumpy gas along the entire northern region at velocities of
$v=47.1-62.6\:\kms$. These higher velocity clumps overlap with a
region of bright $8\mu$m dust emission (Fig. \ref{mapH}). The WISE
\ion{H}{2}~ database lists five \ion{H}{2}~regions overlapping these dense clumps
with similar LSR velocities \citep{Anderson2014}.
\underbar{\textit{Cloud I:}} We find that GMC I has a radius of $R_A\simeq
12$~pc and a mass of $M_{G,A}\simeq2.6\times10^4\:M_{\odot}$.
However, RD10 identified two clouds that match the location of Cloud I, and estimated the much
larger mass of $\sim2.4\times10^5~\rm M_{\odot}$. This difference is
due to the surrounding diffuse $\thco$ emission, which is included by
their cloud finding method that uses the CLUMPFIND algorithm
\citep{Rathborne2009}. IRDC I lies within the densest region of the
GMC, which extends roughly 7 pc north of the IRDC. The cloud's SE
column density spectra indicates a symmetric and well isolated
emission profile out to radii of $\sim10$ pc. However, at larger
cloud radii, diffuse $\thco$ gas at higher velocities is seen.
However, the spectra of the cloud defined by CE indicates that this
high-velocity emission is at intensities less than the $5\sigma$
threshold, suggesting that the GMCs is surrounded by a reservoir of
low-density molecular gas. For the cloud defined by CE, the central
dense region is broken down into three dense clumps and the GLIMPSE
$8\mu$m image indicates that there are multiple bright dust regions
along these dense $\thco$ clumps. However, only one \ion{H}{2}~ region is
identified within the WISE \ion{H}{2}~ catalog, with a slightly lower LSR
velocity of $\sim38~\kms$.
\underbar{\textit{Cloud J:}}
We find that Cloud J has a radius of $R_A\simeq 13$~pc and a mass of
$M_{G,A}\simeq2.0\times10^4\:M_{\odot}$.
However, Cloud J was included in the GMF catalog by
\citet{Ragan2014} in which they estimate a total cloud length of 79
pc. They estimated a total cloud mass of $6.8\times10^4\:M_{\odot}$
(part of this difference results from the larger considered area; part
from assuming a smaller $n_{\rm 12CO}/n_{\rm H2}$ ratio of
$1.1\times10^{-4}$). Our analysis also finds that at the largest extraction radii, Cloud J
is a filamentary cloud, extending $\sim60$~pc in the CE
case. Additionally, the co-added distribution of column density with
velocity at all cloud radii indicate that Cloud J is quite symmetric
and well isolated in velocity space. IRDC J, i.e., G053.11+00.05 from
\citet{Simon2006a}, defined the search region, but Cloud J also
includes the IRDC G053.31+00.00 within its 30~pc radius.
The $8\mu$m images of the Cloud J reveal a
significant amount of bright dust emission, which \citet{Ragan2014}
associated with a foreground high-mass star-forming region at
$v_{\rm LSR}\sim4~\kms$.
Our search through the WISE \ion{H}{2}~ database found one radio quiet
\ion{H}{2}~region at $l=53.09,b=0.12$ with an LSR velocity consistent with
Cloud J \citep{Anderson2014}.
\subsubsection{Spatial and Velocity Offsets between IRDCs and GMCs}
The spatial and velocity offsets between the IRDC and GMC center of
masses are listed in Table~\ref{tab2}, quoted to the nearest
resolution element of the GRS position-velocity grid. In general,
spatial offsets grow as one considers larger GMC scales. At the 30~pc
aperture scale applied to connected extraction, this offset is
typically $\sim 2-10$~pc. All but clouds C and I have this GMC center
located outside the \citet{Simon2006a} IRDC boundary.
Velocity offsets are typically relatively small, $\sim1$~km/s. This
suggests that IRDCs have been formed by contraction/compression of
structures that were already pre-existing within the GMC.
Clouds A, B, C, E, I, J have relatively simple CE CO morphologies in
position velocity space. The remaining clouds exhibit more complex,
multi-component structures.
\begin{figure}[t!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=7in, angle=0,trim=0 0.7in 0 1in]{Fig5-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Velocity gradient $d v_0/ ds$ versus radius $R$ of the clouds over
which the gradient is evaluated. Each panel represents a different
cloud definition: {\it (a) Top left:} simple extraction, optically
thin column density derivation; {\it (b) Bottom left:} simple
extraction, opacity-corrected column densities; {\it (c) Top right:}
connected extraction, optically thin column density derivation; {\it
(d) Bottom right:} connected extraction, opacity-corrected column
densities.
n each panel, the 10 IRDCs are shown by the colored filled triangles
as defined in the upper right-hand corner. For Clouds D, F, G, and H,
these estimates are based on the single component emission profile.
GMCs are shown with filled circles.
The best-fit power-law relation from \citet[][]{McKeeOstriker2007}
($dv_0 / d s \simeq \sigma_{\rm pc} (s/{\rm pc})^{\alpha_\sigma - 1}$,
with values of $\alpha_\sigma\simeq 0.5$ and $\sigma_{\rm pc}\simeq
0.72\:\kms$) is shown by the dotted lines.
The global (IRDCs and GMCs, excluding IRDC J) best-fit power-law relation
for each panel is shown by the solid lines.}
\label{Fig:omega}
\end{figure}
\subsection{Kinematics and Dynamics}\label{S:kinematics}
\subsubsection{Velocity Gradients and Implications for Turbulence and Rotation}
The GMCs defined by CE have a mean velocity gradient of
$dv_{0,\tau}/ds\simeq 0.14\:\kms\:\rm pc^{-1}$. A similar value of
$0.13~\kms~\rm pc^{-1}$ was found by \citet{ImaraBlitz2011} for five
GMCs. The $\sim100$-pc-scale filaments have global velocity gradients
that are smaller than these values. \citet{Jackson2010} find ${\rm d}
v/{\rm d} s <0.09\:{\rm km\:s^{-1}\:pc^{-1}}$ in ``Nessie'',
\citet{BattersbyBally2012} find ${\rm d} v/{\rm d} s <0.05\:{\rm
km\:s^{-1}\:pc^{-1}}$ in their 80-pc long cloud, and
\citet{Ragan2014} measure ${\rm d} v/{\rm d} s \simeq 0.06\:{\rm
km\:s^{-1}\:pc^{-1}}$ as an average of the 7 filaments in their
sample.
\begin{figure}[ht!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=6.5in, angle=0,trim=0 1in 0 0]{Fig6-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Velocity dispersion $\sigma$ versus radius $R$ of the clouds over
which the dispersion is evaluated. Each panel represents a different
cloud definition:
{\it (a) Top left:} simple extraction, optically thin column density
derivation; {\it (b) Middle left:} simple extraction,
opacity-corrected column densities; {\it (c) Bottom left:} simple
extraction, opacity-corrected column densities fitted by a Gaussian
profile; {\it (d) Top right:} connected extraction, optically thin
column density derivation; {\it (e) Middle right:} connected
extraction, opacity-corrected column densities. {\it (f) Bottom
right:} connected extraction, opacity-corrected column densities
fitted by a Gaussian profile.
In each panel, the 10 IRDCs are shown by
the colored filled triangles as defined in the upper right-hand
corner. For Clouds D, F, G, and H, these estimates are based on the
single component emission profile. GMCs are shown with filled circles.
The best-fit power-law relation from \citet[][]{McKeeOstriker2007}
($\sigma =\sigma_{\rm pc} (s/{\rm
1\:pc})^{\alpha_\sigma}$, with values of $\alpha_\sigma\simeq 0.5$
and $\sigma_{\rm pc}\simeq 0.72\:\kms$) is shown by the dotted
lines. The global (IRDCs and GMCs) best-fit power-law relation for
each panel is shown by the solid lines. In panel \textit{(f)}, the best-fit
power-law relations are also shown for the IRDCs (dot-dashed line) and the GMCs
(dashed line).}
\label{Fig:sigmaR}
\end{figure}
For the elliptical IRDC regions also defined by CE, we find a mean
velocity gradient of $dv_{0,\tau}/ds \simeq 0.26~\kms~\rm pc^{-1}$,
larger than that of the GMCs. For comparison, on smaller scales within
IRDC H, \citet{Henshaw2014} find global velocity gradients of $0.08,
0.07, 0.30\:{\rm km\:s^{-1}\:pc^{-1}}$ in several sub-filaments
measured over $\sim$2 parsec length scales, based on centroid
velocities of the dense gas tracer $\rm N_2H^+(1-0)$. Larger local
gradients $\sim 1.5-2.5\:{\rm km\:s^{-1}\:pc^{-1}}$ can be present on
sub-parsec scales.
\citet{Ragan2012} found velocity gradients of 2.4 and 2.1~$\rm km
s^{-1} pc^{-1}$ within sub-pc regions of IRDCs G5.85-0.23 and
G24.05-0.22, based on observations of $\rm NH_3~(1,1)$.
Observed velocity gradients may be a signature of turbulence
\citep[e.g.,][]{BurkertBodenheimer2000,McKeeOstriker2007}. For a
turbulent line-width-size (LWS) relation of the form $\sigma =
\sigma_{\rm pc} (s/{\rm 1\:pc})^{\alpha_\sigma}$, where values of
$\alpha_\sigma\simeq 0.5$ and $\sigma_{\rm pc}\simeq 0.72\:\kms$ have
been reported \citep[][]{McKeeOstriker2007}, we expect velocity
gradients measured over some scale $s$ to vary as $dv_0 / d s \simeq
\sigma_{\rm pc} (s/{\rm pc})^{\alpha_\sigma - 1}$.
Figure \ref{Fig:omega} displays our estimates of $dv_0/ds$ as a
function of cloud size for both IRDCs and GMCs
(we note that the small size of IRDC J makes it difficult for velocity
gradients to be properly resolved with GRS data, so we exclude it from
the following analysis). Fitting the data for clouds defined by CE and
including opacity-corrected column densities (lower-right panel) we
find $\alpha_\sigma= 0.39\pm0.14$ and $\sigma_{\rm pc}=0.54\pm0.19\kms$.
If each cloud is considered individually,
$\alpha_\sigma$ has mean and median values of $0.45$ and $0.51$,
respectively, and a standard deviation of 0.59. By inspection of the
other panels in Fig.~\ref{Fig:omega} we see that the results for the
slope and normalization of the power law relation are not too much
affected by the choice of cloud definition and mass measurement, but
the CE results show smaller dispersions.
We also examine directly the observed velocity dispersions as a
function of size scale (effective radius) of the clouds. Figure
\ref{Fig:sigmaR} displays estimates of $\sigma$ as a function of cloud
size for both IRDCs and GMCs. Now we see a larger effect on the
velocity dispersion versus size relation depending on how the cloud is
identified (i.e., SE or CE), how its mass is measured (i.e., whether
or not to carry out an opacity correction for the $^{13}$CO-derived
mass), and how to derive velocity dispersion (i.e., via a direct
estimate or via a fitted Gaussian). The CE, $\tau$-corrected,
Gaussian-fitted results yield a relatively tight power-law, with the
smallest dispersion.
Fitting the data for all clouds defined by CE and including
opacity-corrected column densities (lower-right panel) we find
$\alpha_\sigma=0.25\pm0.06$ and
$\sigma_{\rm pc}= 1.07\pm0.15\kms$. If each cloud is considered
individually, $\alpha_\sigma$ has a mean and median value of $0.17$
and $0.16$, respectively, and a standard deviation of 0.10.
If we fit IRDCs and GMCs separately, we find
$\alpha_\sigma=0.27\pm0.11$ and $\sigma_{\rm pc}= 1.17\pm0.16\kms$ for
IRDCs and $\alpha_\sigma=0.58\pm0.13$ and $\sigma_{\rm pc}=0.43\pm0.15\kms$ for GMCs.
For virialized, constant mass surface density clouds, where $\sigma
\propto (GM/R)^{1/2} \propto \Sigma^{1/2} R^{1/2}$ \citep{Larson1981,
McKeeOstriker2007, McKee2010} one expects $\alpha_\sigma=0.5$ and
this is consistent with the GMC-only results. However, for IRDCs alone
or for the full sample including IRDCs, we see a shallower relation,
which could still be consistent with expectations from virial
equilibrium since IRDCs have higher mass surface densities.
We will discuss the virialization of the clouds in more detail below
in \S\ref{S:virial}.
Observed velocity gradients within molecular clouds can also be
interpreted as due to cloud rotation
\citep[e.g.,][]{Goodman1993,ImaraBlitz2011}. With such an
interpretation, we assess the projected rotational energy of the
clouds $E_{\rm rot,0}=(1/2)I_0\Omega_0^2$, as discussed in \S3.3.
The derived values of $I_0$ and $E_{\rm rot,0}$ are listed in Table~2.
We next compare rotational to gravitational energies. Following
\cite[][, hereafter BM92]{Bertoldi1992}, the gravitational energy of
an ellipsoidal cloud is given by
\begin{equation}
W=\frac{3}{5}a_1 a_2 \frac{GM^2}{R},
\end{equation}
where parameter $a_1$ describes the effects of the internal density
distribution of the cloud, where for a power-law distribution
$\rho\propto r^{-k_\rho}$, $a_1=(1-k_{\rho}/3)/(1-2k_{\rho}/5)$.
Here, for both GMCs and IRDCs,
we adopt $k_{\rho}=1$, such that $a_1=10/9$, based on the IRDC density
profile study by BT12. Somewhat steeper density profiles with
$k_\rho\simeq 1.5$ are derived by BT12 if envelope subtraction is
carried out, but this would only change $a_1$ by $\sim10\%$. The
parameter $a_2$ accounts for the effect for the cloud's
ellipticity. Following BM92 and our ellipsoidal cloud virial analysis
from HT11, an ellipsoidal cloud with radius $R$ normal to the axis of
symmetry and size $2Z$ along the axis has an aspect ratio of $y=Z/R$,
and semi-major and semi-minor axes of $r_{\rm max}$ and $r_{\rm min}$,
respectively. Since we do not know the inclination angles of the IRDCs
or GMCs, we assume that $R=r_{\rm min}$, $Z=r_{\rm max}$, and that the
observed cloud radius is equivalent to the geometric mean radius,
$R_{\rm obs}=(R_{\rm min}R_{\rm max})^{1/2}$. These assumptions
differ from those of HT11, where we adopted fiducial inclination angle
values of $60^{\circ}$ for the filamentary IRDCs F and H. Now, with a
larger sample of both clumpy and filamentary IRDCs, we make no
assumption regarding their inclination angles. However, as noted in
HT11, an uncertainty of $15^{\circ}$ in inclination would lead to a
$\sim15\%$ uncertainty in the value of y. We follow the BM92
definition of $a_2$, where
\begin{equation}
a_2= \frac{R_m}{R} \frac{{\rm arcsinh} (y^2-1)^{1/2}}{(y^2-1)^{1/2}},
\end{equation}
where, $R_m$ is the mean value of $R_{\rm obs}$ averaged over all
viewing angles. For clouds with low aspect ratios ($y<10$), BM92 give
an approximation for the mean observed radius of $R_m/R\sim
y^{1/2}(1\pm0.27|1-y^{-2}|^{0.81})^{0.31}$, accurate to within
$2\%$. Since all of the IRDCs in this study have $y$ values below 10,
with IRDC B having the largest aspect ratio of $y=4.67$, the BM92
approximation of $a_2$ is valid for use in our gravitational energy
estimates.
We thus evaluate $W$, and then the ratio $\beta = E_{\rm rot}/|W|$,
which we also list in Tables \ref{tab2} and \ref{tabirdc}. We find
that the GMCs defined by CE and with opacity-corrected column density
estimates have mean $\beta_{\tau}\simeq 0.038$. For the elliptical
IRDCs also defined by CE, we find mean $\beta_{\tau}\simeq
0.051$. These values are larger than the $\beta \sim 3\times 10^{-4}$
values estimated by \citet{Ragan2012} for two IRDCs. On much smaller
scales, the study by \citet{Goodman1993} found that most dense cores
within dark clouds have values of $\beta\le0.18$. The increasing
values of $\beta$ that we see on going from GMC to IRDC scales may be
indicative of slightly more disturbed kinematics within
IRDCs. However, it could also be caused by systematic changes in mass
estimation from $\rm ^{13}CO$, e.g., due to increased CO freeze-out
within IRDCs.
\subsubsection{``Rotation'' Directions}
Figure~\ref{PA} displays the distribution of position angles of
cloud angular momentum vectors,
$\theta_{v,\tau}$, with respect to the
direction of Galactic rotation for the cloud definition cases assuming
opacity corrected column densities. Our results show that while a
cloud's position angle can vary depending on cloud definition, the
overall, global distribution of cloud position angles appears
consistent with near equal fractions of clouds with prograde and
retrograde rotation, although based on a relatively small sample of
only 10 analyzed clouds.
\begin{figure}[t!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=4.5in, angle=0, trim=0 0.6in 0.6in 0]{Fig7-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Distribution of relative position angle of the angular momentum
vectors of the 10 IRDCs/GMCs assuming opacity corrected column
densities with respect to the direction of Galactic rotation,
$\theta_{v,\tau}$. The effect of different cloud definitions is
illustrated in the different panels. In each panel, the black line
shows the results from the numerical simulation of
\citet{TaskerTan2009} for clouds at $\rm t=250~Myr$ and the
appropriate inclination of viewing angle of $0^{\circ}$ with respect
to the galactic plane, i.e., an in-plane view.}
\label{PA}
\end{figure}
Young GMCs, recently formed from the diffuse ISM should possess
position angles which are prograde with Galactic rotation ($\theta_v <
-90^{\circ}$; $\theta_v > 90^{\circ}$). For all of the clouds,
measured at various radii for the SE and CE definitions, we find that
there is a large fraction of clouds with retrograde rotation
($-90^\circ<\theta_v < +90^{\circ}$). We find that $\sim50\%$ of the
CE clouds analyzed with opacity corrected column densities have
retrograde motion.
In the Milky Way, \citet{Phillips1999} compiled a database of
rotational measurements for 156 individual clouds from previously
published studies, finding that there is a random distribution between
the direction of a cloud's angular velocity and the Galactic plane.
However, isolated clouds had orientations towards the Galactic
poles. They concluded that the rotation of large-scale structures was
due to Galactic shear, while the randomized orientation of cloud
position angles was due to dynamical and/or magnetic interactions. By
estimating the velocity gradients for more than 500 clouds within the
GRS, \citet{Koda2006} found that the populations of prograde and
retrograde rotation were nearly equal. Beyond the Milky Way,
\citet{Rosolowsky2003} found that 40\% of the GMC population in M33
are in retrograde rotation with respect to the galaxy. In a subsequent
study of the \ion{H}{1}~ envelopes of 45 GMCs from \citet{Rosolowsky2003},
\citet{Imara2011} found that 53\% of these GMCs are in retrograde
rotation with respect to the local \ion{H}{1}. Furthermore, they suggest
that if the linear gradients of the \ion{H}{1} ~envelopes are due to
rotation, then 62\% are in retrograde rotation with respect to M33.
In comparison to theoretical models, the disk galaxy simulations of
\citet{TaskerTan2009} found near equal amounts of apparent prograde
and retrograde clouds ($\sim50\%$) for simulated clouds at
$t=250$~Myr and a viewing angle of $0^{\circ}$ inclination to the
galactic plane, i.e., an in-plane view.
A substantial fraction of retrograde clouds results if GMCs live long
enough to undergo mergers and/or collisions with neighboring clouds.
\citet{Dobbs2013} performed 3D smooth particle hydrodynamic (SPH)
simulations of GMCs within a galactic disk, including self-gravity
and stellar feedback. They found a retrograde GMC population of
$\sim40\%$,
similar to their previous studies \citep[e.g.,
][]{Dobbs2011b}. However, the most massive GMCs tend to be prograde.
Studying the distribution of angular momentum vectors can test
scenarios of star formation that involve frequent galactic shear
driven cloud-cloud collisions \citep{Tan2000}. At the moment, with
only 10 clouds in our pilot sample, we are not yet in a position to
make detailed statistical comparison with numerical simulations. We
defer such an analysis, using larger samples of Galactic GMCs, to a
future paper.
\subsubsection{Cloud Virialization}\label{S:virial}
Various mechanisms, such as turbulence, magnetic fields, and thermal
gas pressure, may be responsible for supporting molecular clouds
against gravitational collapse \citep[e.g.,][]{McKeeOstriker2007}.
The dimensionless virial parameter, $\alpha_{\rm vir}$, describes a cloud's
dynamical state
via \citep{Bertoldi1992}:
\begin{equation}
\alpha_{\rm vir} \equiv \frac{5\sigma^2 R}{GM} = 2a \frac{E_K}{|E_G|}
\label{eqn:alpha}
\end{equation}
where $\sigma$ is the 1D mass-averaged velocity dispersion, $a\equiv
|E_G| / (3GM^2/[5R])$ is the ratio of gravitational energy, $E_G$
(assuming negligible external tides), to that of a uniform sphere, and
$E_K$ is the kinetic energy. As discussed above in
\S\ref{S:kinematics}, for spherical clouds with power-law density
distribution, $\rho \propto r^{-k_\rho}$, $a=
(1-k_\rho/3)/(1-2k_\rho/5)$, i.e, rising as 1, 25/24, 10/9, 5/4, 5/3
when $k_\rho$ rises as 0, 0.5, 1, 1.5, 2, respectively.
\begin{figure}[t!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=5in, angle=0, trim=0 1in 0 0.5in]{Fig8-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Comparison of cloud virial parameter, $\alpha_{\rm vir}$, with cloud
mass. Each panel represents a different cloud definition: {\it (a)
Top left:} simple extraction, optically thin column density
derivation; {\it (b) Middle left:} simple extraction,
opacity-corrected column densities; {\it (c) Bottom left:} simple
extraction, opacity-corrected column densities fitted by a Gaussian
profile; {\it (d) Top right:} connected extraction, optically thin
column density derivation; {\it (e) Middle right:} connected
extraction, opacity-corrected column densities. {\it (f) Bottom
right:} connected extraction, opacity-corrected column densities
fitted by a Gaussian profile. In each panel, the 10 IRDCs are shown by
the colored filled triangles (see legend). For Clouds D, F, G, and H,
these estimates are based on the isolated single fitted emission
profile. The error bar in the top-left of each panel represents the mean uncertainties
of $\sim40\%$ in $\alpha_{\rm vir}$ and $\sim50\%$ in $M$.
}
\label{alpha1}
\end{figure}
\begin{figure}[t!]
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.3in, angle=0, trim=0 1.2in 1in 0]{Fig9a-eps-converted-to.pdf} & \includegraphics[width=3.3in, angle=0, trim=0 1.2in 1in 0]{Fig9b-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Dependence of $\sigma/R^{1/2}$ with mass surface density,
$\Sigma$. Results for the IRDCs (\textit{Left}) and GMCs
(\textit{Right}) are shown separately. Each panel represents a
different cloud definition: {\it (a) Top left:} simple extraction,
optically thin column density derivation; {\it (b) Bottom left:}
simple extraction, opacity-corrected column densities; {\it (c) Middle
right:} connected extraction, optically thin column density
derivation; {\it (d) Middle right:} connected extraction,
opacity-corrected column densities. {\it (c) Bottom right:} connected
extraction, optically thin column density, Gaussian fitted profile
derivation; {\it (d) Bottom right:} connected extraction,
opacity-corrected column density, Gaussian fitted profile
derivation. In each panel, the IRDCs/GMCs are shown by the colored
filled triangles/circles as defined in the upper right-hand corners.
For Clouds D, F, G, and H, these estimates are based on the single
component emission profiles.
The error bar in the upper-left corner represents the mean
uncertainties of $\sim14\%$ in $\sigma/R^{1/2}$ and $\sim30\%$ in
$\Sigma$.
The results of RD10, a total of 329 clouds ($M>10^{4}~{M_\odot}$) scaled by
a factor of 0.49 to reflect our adopted $^{13}$CO abundance
\citep[see][]{Tan2013} are shown by the gray crosses. The dotted line
represents virialized conditions with $\alpha_{\rm vir}=1$, i.e.,
$\sigma/R^{1/2}=(\pi G/5)^{1/2} \Sigma^{1/2}$ (see text). The mean
results of \citet{Solomon1987} are also shown (black open square).
Best-fit power-law relations ($\sigma/R^{1/2} \propto \Sigma^n$) are
shown by the solid lines.}
\label{heyer}
\end{figure}
Effects of nonspherical geometry for ellipsoidal clumps have also been
considered by BM92, and found to be relatively small ($\lesssim 10\%$)
for aspect ratios $Z/R\lesssim6$, where $Z$ is the clump radius along
the axis of symmetry and $R$ the radius normal to this
axis. Therefore, here we will ignore these effects of clump
elongation.
Ignoring surface pressure and magnetic terms, a cloud in virial
equilibrium has $|E_G|=2 E_K$, so that $\alpha_{\rm vir}=a$. Such a
cloud in free-fall has $\alpha_{\rm vir}\rightarrow 2a$ from below as
time progresses. In principle one can imagine a cloud starting with
very small levels of internal motion, so that $\alpha_{\rm vir}\ll1$,
and then crossing through the condition $\alpha_{\rm vir}=a$ as part
of its free-fall collapse. However, such a tranquil and synchronized
initial condition would seem to be difficult to achieve in reality if
GMCs are accumulating from a turbulent ISM, including potential
agglomerating via GMC collisions.
Although it is difficult to differentiate infall motions from
virialized turbulent motions, in cases where infall rates have been
claimed towards dense clumps, including IRDCs,
\citep[e.g.,][]{Peretto2013}, the infall velocities are relatively
small compared to virial motions (equivalent to the mass doubling
times being long ($\sim 10\times$) compared to the free-fall time; see
\citet[][]{Tan2014}, for a review). On the larger scales of GMCs,
\citet{ZuckermanEvans1974} and \citet{KrumholzTan2007} have argued
against fast rates of global collapse and infall, since overall star
formation rates are much smaller. However, converging flows associated
with the collision of independent, self-gravitating clouds may be
relatively common \citep{Tan2000,TaskerTan2009}, and the kinematic
signatures of such flows would be difficult to distinguish from global
infall.
The effect of surface pressure on the cloud increases the velocity
dispersion of the virial equilibrium state and thus raises the value
of its virial parameter. For example, for a cloud with $k_\rho=1.5$
and negligible surface pressure we have $\alpha_{\rm
vir}=a=5/4$. However, if the cloud is embedded in an ambient medium
with the cloud's surface pressure equal to the ambient pressure, as in
the Turbulent Core model of \citet[][hereafter MT03]{McKeeTan2003},
then $\alpha_{\rm vir}= (5/2)(3-k_\rho)/[(5-2k_\rho)(k_\rho-1)] =
15/4$.
Note as $k_\rho\rightarrow 1$ from above,
$\alpha_{\rm vir}\rightarrow \infty$ (a self-similar singular
polytropic sphere with $k_\rho=1$ would have a spatially constant
pressure, i.e., zero pressure gradient, so there is no self-consistent
physical solution; MT03). These results show that virial equilibrium
pressure confined singular polytropic spherical clouds with $k_\rho
\lesssim 1.5$ have virial parameters $\gg 1$.
Support by large-scale magnetic fields acts to reduce the velocity
dispersion of the virial equilibrium state. For example, in the
fiducial MT03 Turbulent Core model, the mean velocity dispersion of
the virial equilibrium state is reduced by a factor
$\phi_B^{-1/2}\rightarrow 2.8^{-1/2}=0.598$ if allowing magnetic
fields strong enough that the mean Alfv\'en Mach number is unity. This
implies that the value of the virial parameter of the virial
equilibrium state would be reduced by a factor of 0.357.
Magnetic field observations in molecular clouds have been reviewed by
\citet{Crutcher2012} and \citet{Li2014}.
Field strengths increase with density approximately as $B\propto
\rho^{2/3}$ and are close to values expected for equipartition
magnetic and turbulent energy densities and also close (within a
factor of $\sim 2$) of the critical field strength that would prevent
collapse. It thus seems likely that magnetic fields can have
significant dynamical effects. Recently, \citet{Pillai2015} have
found evidence for strong, dynamically important magnetic fields in
two IRDCs, based on relatively ordered directions of polarized sub-millimeter
emission, thought to be caused by dust grains that have aligned with
the $B$-fields. Unfortunately for the IRDCs/GMCs in our sample, we do
not have detailed information on magnetic field strengths.
For the GMCs and IRDCs analyzed in this paper, for each cloud, three
virial parameters were estimated. First, we estimated $\alpha_{\rm
vir}$ assuming optically thin conditions to derive $\sigma$ and
$M$. Second, using the opacity corrected column densities,
$\alpha_{\rm vir,\tau}$ is estimated using $\sigma_{\tau}$ and
$M_{\tau}$. Finally, the virial parameter was estimated using the
Gaussian-fitted, opacity corrected column density spectra, i.e.,
$\alpha_{\rm vir,G}$ is estimated using $\sigma_G$ and $M_G$.
In the cases of clouds D, F, G, and H, the ``-s'' mass and velocity
dispersions were used to estimate the single component
virial parameters of the clouds.
If cloud properties are assessed using the full $v_0\pm15\kms$
velocity range (including multiple components), then larger values of
$M$ and especially $\sigma$ are inferred, leading to estimates of
$\alpha_{\rm vir}$ that are larger by factors of about 3.
All molecular cloud virial parameters estimates are listed in Tables
\ref{tab2} and \ref{tabirdc}. We note that for CE-defined GMCs, the virial analysis is carried out
using cloud centers that are the center-of-mass of all the CE gas
within the 30~pc radius region. The cloud masses are those that are
estimated from the material that is within the radii $R_M$, $R_A$ and
$R_{1/2}$. For SE-defined GMCs, the virial analysis uses the extracted
cloud radius as the characteristic size of the cloud.
For the same cloud definition cases, the virial parameter was also
estimated for each IRDC. Since these clouds are elliptical, we used
the geometric mean observed radius, $R=\sqrt{(R_{\rm maj}R_{\rm
min})}$, as an estimate of the cloud's equivalent spherical shape
(but note that, as discussed above, effects of asphericity are minor).
\begin{figure}[t!]
\begin{center}$
\begin{array}{c}
\includegraphics[width=6in, angle=0]{Fig10-eps-converted-to.pdf}
\end{array}$
\end{center}
\caption[]{
Dependence of $\sigma/R^{1/2}$ with mass surface density, $\Sigma$,
for the all clouds defined by CE, opacity corrected column densities,
and gaussian fitted velocity dispersions. As in Figure~\ref{heyer},
IRDCs are shown by the filled triangles and the GMCs by the filled
circles (see legend).
The error bar in
the upper-left corner represents the mean uncertainties of $\sim14\%$
in $\sigma/R^{1/2}$ and $\sim30\%$ in $\Sigma$. The results of RD10,
a total of 329 clouds with $M>10^{4}~{M_\odot}$, are shown by the gray
crosses. The mean results of
\citet{Solomon1987} are also shown (open black square).
The best-fit power-law relation ($\sigma/R^{1/2} \propto \Sigma^n$) is
shown for both the IRDCs
($n=1.27\pm0.52$; dot-dashed line) and the GMCs ($n=0.70\pm0.12$;
dashed line). Collectively, the IRDCs and GMCs have a best-fit
power-law relation of $n=0.81\pm0.10$ (solid line). The dotted lines
represents the expected scaling relation for virialized clouds with
different values of $\alpha_{\rm vir}=0.5$, 1, 2, 4, and 8.}
\label{CEtg}
\end{figure}
Figure \ref{alpha1} displays the virial parameters as a function of
mass for each cloud definition case.
The overall range of $\alpha_{\rm vir}$ is nearly the same for both
IRDCs and GMCs. We estimate a typical random uncertainty in
$\alpha_{\rm vir}$ of
$\sim40\%$ after accounting for uncertainties in the mass surface
densities, velocity dispersions, and the cloud kinematic distances.
However, systematic uncertainties in mass surface density of a factor
of $\sim2$ would lead to similar levels of uncertainty in $\alpha_{\rm
vir}$.
Considering the different cloud definition methods, the virial
parameters estimated using the CE are lower than those estimated using
SE. For both cloud extraction methods, $\alpha_{\rm vir}$ is generally
smaller for the clouds using opacity corrected column densities
compared to those assuming optically thin gas conditions. This trend
results from larger masses and slightly narrower line profiles in the
opacity-corrected cases. Similarly, the smallest $\alpha_{\rm vir}$
values were estimated in the cases of using the Gaussian fitted
spectra to measure both the velocity dispersion and mass, $\sigma_G$
and $M_G$.
Thus, the virial parameter is sensitive to the method used to extract
the $\thco$ emission associated with a GMC. Overall, the lowest values
of $\alpha_{\rm vir}$ were estimated for the clouds defined by CE,
opacity-corrected column densities, and Gaussian fitted masses and
velocity dispersions. The mean and median values of these estimates of
$\alpha_{\rm vir,G}$ are $1.12$ and $0.99$, respectively, with a
standard deviation of $0.51$ (similar to the expected size of the
random errors). Similarly, for the IRDCs under the same definitions,
the mean and median values of $\alpha_{\rm vir,G}$ are $1.88$ and
$1.93$, respectively, with a standard deviation of $1.11$.
We see that IRDCs have virial parameters that are about a factor of
two larger than GMCs. Again, this can be interpreted as evidence for
more disturbed kinematics. Alternatively, it could indicate that
$\thco$-derived masses for IRDCs have been underestimated, potentially
due to CO depletion.
A population of virialized clouds with similar degrees of pressure
confinement may be expected to have a relatively small dispersion in
their values of $\alpha_{\rm vir}$, although some variation is likely
due to variation in internal density structure and level of support by
large scale magnetic fields. If we denote the mean virial equilibrium
value of the virial parameters of such a cloud population as
$\bar{\alpha}_{\rm vir,eq}$, then from eq.~(\ref{eqn:alpha}) we can
derive
\begin{equation}
\frac{\sigma}{R^{1/2}} = \left(\frac{G\pi}{5}\right)^{1/2} \alpha_{\rm vir,eq}^{1/2} \Sigma^{1/2}
\label{eq:heyer}
\end{equation}
\citep[see also][]{Heyer2009,McKee2010}. By examining the observables
$\sigma/R^{1/2}$ and $\Sigma$ of our IRDC and GMC samples, we can see
if they follow this correlation that is expected if the clouds are
virialized. We estimate an uncertainty in $\sigma/R^{1/2}$ of
$\sim14\%$ based on $10\%$ uncertainties in both $\sigma$ and $R$.
Thus we considered the above scaling relationship for the results of
our ten IRDCs and GMCs (Fig. \ref{heyer}). For IRDCs, $\Sigma$ was
estimated for each ellipsoidal cloud using the geometric mean observed
radius $R$ and the total $\thco$-derived mass. For GMCs, $\Sigma$ was
estimated using the circular area defined by the extraction radius $R$
and its corresponding $\thco$-derived mass. We estimate a random
uncertainty in $\Sigma$ of $\sim30\%$ based on the uncertainty in
$T_{\rm ex}$ and potential variations in our assumed
abundances.
However, systematic uncertainties could be present at the
factor of $\sim2$ level.
For the ten GMCs, we find that the clouds analyzed using gaussian
profile fitting on the opacity corrected column density spectra to
estimate the cloud mass and velocity dispersion, yield results most
consistent with the scaling expected for gravitational virial
equilibrium. Specifically, we find the best-fit power-law relation
$\sigma/R^{1/2} \propto \Sigma^n$ has $n=0.54\pm0.05$ and
$n=0.70\pm0.12$, for the SE and CE extraction, respectively. We gauge
the significance of these fits by estimating the Pearson's correlation
coefficient, $\rho$, and the one-sided $p$-value. For all the GMCs
defined by CE and gaussian profile fitting on the opacity corrected
column density spectra, we estimate $\rho=0.71$ and a $p$-value of
$6.0\times 10^{-6}$, indicating a significant linearity in the
data.
On the other hand, for the ten IRDCs analyzed using gaussian
profile fitting on the opacity corrected column density spectra to
estimate the cloud mass and velocity dispersion, yield steeper
best-fit power-law relations, with $n=1.57\pm0.80$ and
$n=1.27\pm0.52$, for the SE and CE extraction, respectively. These
power-law fits are less significant than those seen for the GMCs, with
$\rho=0.56$ and a $p$-value of $0.047$ for the ``CE,$\tau$,G'' method.
Focusing on the ``CE,$\tau$,G'' extraction method,
these IRDC and GMC results are displayed together in Fig.~\ref{CEtg}.
Overall the IRDCs and GMCs together yield a best-fit power-law
relation of $n=0.81\pm0.10$, somewhat steeper than the $n=1/2$
expected for virialized clouds. If a power-law relation with $n=0.5$
is fit, then the normalization implies $\bar{\alpha}_{\rm vir}=1.60$
and 1.03 for IRDCs and GMCs, respectively, and $\bar{\alpha}_{\rm
vir}=1.15$ for the whole sample (with a Pearson's correlation
coefficient of $\rho=0.78$ with a $p$-value of $1.7\times 10^{-9}$).
We note that some of the dispersion in the data is due to
cloud-to-cloud variation, which might be expected if there is
systematic variation in degree of surface pressure confinement,
magnetization, distance uncertainty (affecting $R$) or $\thco$
abundance.
We conclude that the relatively tight correlation shown by the data in
Fig.~\ref{CEtg} is evidence of cloud virialization over a wide range
of scales within GMCs, perhaps representing a self-similar hierarchy
of self-gravitating virialized structures. There is tentative evidence
that IRDCs deviate from this virialized scaling relation in the sense
of being more kinematically disturbed.
However, the level of deviation of the IRDCs is quite modest and their
best-fit $\bar{\alpha}_{\rm vir}\simeq 1.6$ is a value that is quite
compatible with models of virialized singular polytropic
($k_\rho\simeq 1.5$) spherical clouds embedded in a pressure confining
ambient medium. Such a model would yield $\bar{\alpha}_{\rm vir}=3.75$
(but quite sensitive to $k_\rho$). Including support from large-scale
magnetic fields that lead to Alfv\'en Mach numbers of unity would
lower this to $\bar{\alpha}_{\rm vir}=1.34$. However, given these
uncertainties in $k_\rho$, together with those from $\thco$ abundance,
using the absolute value of $\bar{\alpha}_{\rm vir}$ to assess the
degree of virialization is problematic, and we argue it is better to
focus on the degree of correlation in the $\sigma/R^{1/2}$ versus
$\Sigma$ diagram.
For comparison, the results of RD10 for a total of 329 clouds with
$M>10^{4}~{M_\odot}$ and with their $\Sigma$ values scaled by a factor of
0.49 to reflect our adopted $^{13}$CO abundance \citep[see][]{Tan2013}
are also shown in Fig.~\ref{CEtg}. The RD10 GMCs also follow the
expected virialized scaling, but with larger scatter. This is to be
expected given that we also see larger scatter with our simpler cloud
extraction methods.
Consistent with our GMC results, the RD10 GMCs also have values of
$\bar{\alpha}_{\rm vir}\simeq 1.1$.
We also note here that the absolute values of $\Sigma$ that we derive
for GMCs are similar to those of H09 and RD10. H09 found an average
mass surface density of $\bar{\Sigma}\simeq 42\:{M_\odot}\:{\rm pc}^{-2}$
using 162 of the \citet{Solomon1987} $^{12}$CO-defined molecular
clouds (but they noted that the expected values are likely
underestimated by a factor of $\sim 2$ for these larger scale GMCs due
to sub-thermal excitation of $\thco$). Additionally, the 329 RD10
clouds with $M>10^{4}~{M_\odot}$ yield $\bar{\Sigma}\simeq 88\:{M_\odot}\:{\rm
pc}^{-2}$. For the 10
GMCs studied here, we find
$\bar{\Sigma}_{G}\simeq 80\:{M_\odot}\:{\rm pc}^{-2}$ assuming CE, opacity
corrected column densities, fitted gaussian line profile to derive
mass, and circular cloud radii defined by $R_A$. For the 10 IRDCs,
defined by the Simon et al. (2006) ellipses, we find
$\bar{\Sigma}_{G}\simeq 217\:{M_\odot}\:{\rm pc}^{-2}$.
However, it is important to note that there is a real spread in
$\Sigma$ values between and among the considered IRDC and GMC scales
of greater than a factor of 10.
\section{Conclusions}
We have presented a detailed study of the $\thco$ gas on various size
scales out to radii of 30~pc around 10 well-studied IRDCs (i.e., the
sample of BT09 and BT12). We find that all the IRDCs are embedded in a
$\thco$-defined GMC, i.e., a molecular cloud with $M\gtrsim
10^4\:M_\odot$. We have taken care to assess the effects of cloud
definition, specifically comparing simple extraction of a slab in
position-velocity space with extraction only of connected emission
features. We have also taken notice of the effects of multiple
Gaussian components when they are present. We have considered column
density estimates via assuming, first, optically thin $\thco$(1-0)
emission (with excitation temperatures estimated from associated
$^{12}$CO emission) and then correcting for line optical depth
effects. Our preferred method of ``CE,$\tau$,G'' cloud definition
involves connected extraction, opacity-corrected column densities and
fitting a single Gaussian profile to derive velocity dispersion and
mass.
We discussed the particular morphologies of the individual clouds in
both position-velocity space and in relation to plane of sky imaging
of the IRDCs. We find most IRDCs are offset from their 30-pc-scale GMC
center by $\sim 2-10$~pc, but have only small velocity offsets of
$\sim 1\:{\rm km\:s^{-1}}$. This suggests IRDCs have been formed by
contraction/compression of structures that were already pre-existing
with the GMCs rather than being formed from material swept-up by a shock. Six of
the ten clouds have relatively simple morphologies in
position-velocity space, while the other four exhibit more complex,
multi-component structures.
We then focussed on the kinematic and dynamical properties of the
clouds. We measured velocity gradients as a function of size scale,
$s$, finding they decrease approximately as $s^{-1/2}$, i.e., consistent
with observations that IRDCs show larger velocity gradients than
GMCs. Overall, this is also consistent with the gradients being caused
by turbulence with a velocity dispersion versus size relation of the
form $\sigma \propto s^{1/2}$. Such a scaling is seen in our sample of
GMCs, but a somewhat shallower relation is apparent when considering
IRDCs and GMCs together. Using velocity gradients to define projected
rotation axes we do not see correlations with Galactic plane
orientation and have approximately equal fractions of pro- and
retrograde rotators with respect to Galactic orbital rotation (though
this analysis is limited by the small sample of only 10
clouds). We measure rotational energy to gravitational binding energy ratios of
$\beta\simeq 0.045$, with a slight increase seen on going from GMC to
IRDC scales.
Assessing virial equilibrium, we find cloud definition has a
significant impact on the virial parameter, i.e., the ratio of
internal kinetic to gravitational energy. CE,$\tau$,G-defined GMCs
have $\bar{\alpha}_{\rm vir}\simeq 1.0$, while IRDCs have
$\bar{\alpha}_{\rm vir}\simeq 1.6$. We argue that it is difficult to
assess the dynamical state, i.e., how close to virial equilibrium,
from the absolute values of $\bar{\alpha}_{\rm vir}$, especially given
the uncertain effects of pressure confinement, large-scale magnetic
field support and mass surface density estimation from
$\thco$. However, these values $\bar{\alpha}_{\rm vir}\sim 1.0$ to 1.6
are quite compatible with Turbulent Core/Clump/Cloud models of
singular polytropic spherical clouds that are pressure-confined by an
ambient medium and have moderate large scale $B$-field support that
yields Alf\'enic Mach numbers of order unity. Note, that
$\thco$-defined clouds are likely to have significant surrounding
molecular material that could provide pressure confinement.
Rather than focus on absolute values of $\bar{\alpha}_{\rm vir}$, we
follow H09 in examining the predicted scaling in the $\sigma/R^{1/2}$
versus $\Sigma$ plane. The results can be sensitive to the method of cloud definition. For
our preferred method of CE,$\tau$,G-defined clouds, the GMCs, analyzed
over a range of scales, show a strong correlation in this diagram,
with a power law slope of $\sigma/R^{1/2} \propto \Sigma^n$ with
$n=0.70\pm0.12$, similar to the $n=1/2$ value expected under virial
equilibrium. Potentially this represents a self-similar hierarchy of
self-gravitating virialized structures.
However, the IRDCs exhibit a somewhat steeper scaling with
$n=1.27\pm0.52$ (or $n=0.81\pm0.10$ for the combined GMC and IRDC
sample). This may indicate that IRDCs have moderately more disturbed
kinematics than GMCs, which is qualitative expectation of models in
which star formation is triggered by GMC-GMC collisions
\citep[e.g.,][]{Scoville1986,Tan2000}. However, it is also possible that
systematic uncertainties affecting mass surface density estimation
from $\thco$, such as increased CO freeze-out in IRDCs \citep[e.g.,][]{Hernandez2011,
Hernandez2012}, are playing a role. Improved methods of
measuring mass and assessing degree of CO freeze-out in these clouds
are needed to help distinguish these possibilities and thus reliably
measure the dynamical state of gas that is likely to form star
clusters.
Additional future work will involve applying the techniques developed
here to larger samples of clouds to better study connections with
Galactic dynamics and Galactic environment, and potential systematic
differences of GMCs that are forming IRDCs or that have relatively
high dense gas mass fractions.
\acknowledgements A.K.H acknowledges support from grants awarded
to Bart Wakker at the University of Wisconsin-Madison.
We thank Brian Babler for contributing to the GLIMPSE data
processing. This publication makes use of molecular line data from the
Boston University-FCRAO Galactic Ring Survey (GRS). The GRS is a joint
project of Boston University and Five College Radio Astronomy
Observatory, funded by the National Science Foundation under grants
AST-9800334, AST-0098562, and AST-0100793.
J.C.T. acknowledges support from NASA grants ATP09-0094 and
ADAP10-0110.
\newpage
|
\section{Introduction and motivation}
Following \cite{ellis1, ellis2, n4}, the \textit{Schur multiplier of the pair} $(L,N)$, where $L$ is a Lie algebra with ideal $N$, is the abelian Lie algebra $M(L,N)$ which appears in the following natural exact sequence of Mayer--Vietoris type
\[H_3(L,\mathbb{Z}) \longrightarrow H_3(L/N,\mathbb{Z}) \longrightarrow M(L,N) \longrightarrow M(L) \longrightarrow M(L/N) \longrightarrow \]
\[ \longrightarrow \frac{L}{[L,N]} \longrightarrow \frac{L}{[L,L]} \ \longrightarrow \ \frac{L}{[L,L] + N} \ \longrightarrow 0, \]
in which we may note the third homology Lie algebras $H_3(L,\mathbb{Z})$ and $H_3(L/N,\mathbb{Z})$ with coefficients in the ring $\mathbb{Z}$ of the integers and the two homology Lie algebras $M(L)=H_2(L,\mathbb{Z})$ and $M(L/N)=H_2(L/N,\mathbb{Z})$, again over $\mathbb{Z}$. The notion of $M(L,N)$ is in fact categorical and can be formulated both for groups and Lie algebras (see \cite{ri1, ri2, sam, ellis1, ellis2, ellis3, ellis4, n4}).
On another hand, the \textit{nonabelian tensor product} $L \otimes N$ of $L$ and $N$ is defined by generators and relations via suitable actions of $L$ on $N$, and viceversa. Following \cite{ellis1, ellis2, guin1}, an action of $L$ on $N$ ($L$ and $N$ are supposed to be over the same field $F$) is an $F$-bilinear map $(l,n) \in L \times N \longmapsto ~^ln \in N$ satisfying $^{[l,l']}n={~^l}(^{l'}n)-{~^{l'}}(^{l}n)$ and $~^{l}[n,n']=[~^{l}n,n']+[n,~^{l}n']$ for all $l,l' \in L$ and $n, n' \in N$. Note that the Lie multiplication induces the action of $L$ on $N$, in fact, $L$ acts on $N$ via $~^ln=[l,n]$. Since $L$ and $N$ act on each other, and on themselves, by Lie multiplication, we say that their actions are \textit{compatible}, when $~^{~^nl}n'=~^{n'}(~^ln)$ and $~^{~^ln}l'=~^{l'}(~^nl)$ for all $l,l' \in L$ and $n,n' \in N$. Now $L \otimes N$ is the Lie algebra generated by the symbols $l\otimes n$ with defining
relations $c(l \otimes n)=cl \otimes n = l \otimes cn,$ $(l+l')
\otimes n = l \otimes n + l' \otimes n,$ $l \otimes (n+n') = l
\otimes n + l \otimes n',$ $~^ll' \otimes n = l \otimes ~^{l'}n -
l' \otimes ~^ln,$ $ l \otimes ~^nn'= ~^{n'}l \otimes n - ~^nl
\otimes n',$ $[l\otimes n, l' \otimes n']=- ~^nl \otimes
~^{l'}n'$, where $c \in F$, $l,l' \in L$ and $n,n' \in N$. In case $L=N$ and all actions are given by Lie multiplication, $L\otimes L$ is called \textit{nonabelian tensor
square} of $L$. Note that the nonabelian tensor product always exists and, in particular, we find the usual abelian tensor product $L \otimes_\mathbb{Z} N$, when $L$ and $N$
are abelian and the actions are all by Lie multiplication, compatible and trivial.
This construction plays a fundamental role in algebraic topology and homology; the reader may refer to \cite{n1, n2, n3, n4} for recent contributions in the theory of Lie algebras, but also to \cite{ellis3, ellis4, guin2, mas1, fra} for analogies with the context of groups. In particular, \cite{ellis4, guin1, n2} inspired most of the ideas that we are going to show in the present paper. We recall from \cite{ellis1, ellis2, n1, n4, simson} that it is possible to get the following commutative diagram:
\begin{equation}\label{diag}\begin{CD}
@. 0 @. 0\\
@. @VVV @VVV\\
\Gamma\left(\frac{N}{[N,L]} \right) @>>> J_2(L,N) @>>>M(L,N)@>>>0\\
@| @VVV @VVV @.\\
\Gamma\left(\frac{N}{[N,L]} \right) @>\psi>> L \otimes N @>\varepsilon_{L,N}>> L \wedge N @>>>0 @. \\
@. @V\kappa_{L,N} VV @V\kappa'_{L,N}VV\\
@. [L,N] @= [L,N]\\
@. @VVV @VVV\\
@. 0 @. 0\\
\end{CD}\end{equation}
where $\kappa_{L,N} : l \otimes n \in L \otimes N \longmapsto \kappa_{L,N}(l \otimes n)=[l,n] \in [L,N]$ is an epimorphism of Lie algebras such that $J_2(L,N)=\ker \kappa_{L,N} \subseteq Z(L \otimes N)$. Note that $J_2(L,L)=J_2(L)$ was described in \cite[pp.109--110]{ellis2} when $N=L$. On the other hand, \[L \square N=
\langle n \otimes n \ | \ n \in L \cap N\rangle\] is an ideal of
$L \otimes N$ contained in $Z(L \otimes N)$, called \textit{diagonal ideal} of $L \otimes N$. Its properties are discussed in \cite{ellis1, ellis2, n1, n2, n4} when $L=N$. The natural epimorphism $\varepsilon_{L,N} : l \otimes n \in L \otimes N \mapsto (l\otimes n)+ L \square N \in L \otimes N/L \square N$ allows us to form the Lie algebra quotient \[L\wedge N=\frac{L \otimes
N}{L \square N}=\langle l \otimes n + (L \square N) \ | \ l \in
L, n \in N \rangle =\langle l \wedge n \ | \ l \in L, n \in N\rangle,\]
called \textit{nonabelian exterior product} of $L$ and $N$. The natural epimorphism
$\kappa'_{L,N} : l \wedge n \in L \wedge N \longmapsto \kappa'_{L,N}(l \wedge n)=[l,n] \in [L,N]$ is
such that $M(L,N) \simeq \ker \kappa'_{L,N} \subseteq Z(L \wedge N)$: this is one of many connections between the theory of Schur multipliers of pairs and that of nonabelian tensor products. The reference give more details on this aspect. We also note that the columns of \eqref{diag} are central extensions, while the rows of \eqref{diag} form two long exact sequences. In fact $\Gamma$ denotes the \textit{quadratic Whitehead functor} and $\psi$ the \textit{quadratic Whitehead function}, properly defined in \cite[Definition, p.107]{ellis2} (see also \cite{simson} for a categorical definition of $\Gamma$ and $\psi$).
Since the notion of \textit{dimension} for a Lie algebra is in a certain sense ``more geometric than algebraic'',
we cannot expect full analogies with respect to the results in \cite{ellis3, guin2, mas1, fra}, when we replace this notion with that of order of a group. In fact the second and the fourth author have investigated the role of the homological invariants between Lie algebras and finite groups, making specific studies in \cite{n1, n2, n3, n4} during the last years. We continue the same line of research. In Section 2 we prove a theorem of splitting for $L \otimes N$ via $L \square N$ and $L \wedge N$. This generalizes results in \cite{n2} when $N=L$. Section 3 describes the role of $\square$ with respect to the operator $*$ of free product: another theorem of splitting is shown here.
\section{Splitting of $\otimes$ via $\square$ and $\wedge$}
We begin with elementary facts, which follow from the exactness of \eqref{diag}. Our first lemma generalizes \cite[Proposition 17]{ellis2} when $N \neq L$.
\begin{lem}\label{1}
Let $N$ be an ideal of a Lie algebra $L$. If $N/[N,L]$ is free, then
\[\begin{CD} 0 @>>>\Gamma\left(\frac{N}{[N,L]} \right) @>\psi>> L \otimes N @>\varepsilon_{L,N}>> L \wedge N @>>>0 \end{CD}\]
is an exact sequence.
\end{lem}
\begin{proof}
We just need to prove that
\[\psi: \gamma(n+[N,L]) \in \Gamma \left(\frac{N}{[N,L]} \right) \longmapsto n\otimes n \in L \otimes N \] is injective. Let
$\theta: L \otimes N \rightarrow L/[N,L] \otimes N/[N,L]$ be the homomorphism induced by the natural projection $l \in L \mapsto l+[N,L] \in L/[N,L]$. Then \[ \frac{L}{[N,L]} \otimes \frac{N}{[N,L]} \simeq \frac{L}{[N,L]} \otimes_{\mathbb{Z}} \frac{N}{[N,L]}\] is a free Lie algebra, and so is $\Gamma (N/[N,L])$. Moreover, the composition $\theta \circ \psi$ maps a basis of $\Gamma (N/[N,L])$
injectively into a part of a basis of $L/[L,N] \otimes N/[L, N]$. Thus $\theta \circ \psi$ is injective. This is enough to conclude that $\psi$ is injective.
\end{proof}
In general, the columns of \eqref{diag} are short exact sequences, but not the rows. However, if $N/[N,L]$ is free, then Lemma \ref{1} allows us to conclude that even the rows of \eqref{diag} become short exact sequences.
\begin{cor}\label{2}
Let $N$ be an ideal of a Lie algebra $L$ such that $N/[N,L]$ is free. Then the following
\[\begin{CD}
0 @>>> \Gamma\left(\frac{N}{[N,L]} \right) @>>> J_2(L,N) @>>>M(L,N)@>>>0\\
@. @| @VVV @VVV @.\\
0 @>>>\Gamma\left(\frac{N}{[N,L]}\right) @>>> L \otimes N @>>> L \wedge N @>>>0 @. \\
\end{CD}\]
is a commutative diagram with short exact sequences as rows.
\end{cor}
\begin{proof} Application of Lemma \ref{1} to \eqref{diag}.
\end{proof}
A crucial step, which is fundamental for our aims, deals with the description of the natural epimorphism
\begin{equation}\label{epi1}\pi: l \otimes n \in L \otimes N \mapsto (l + [N,L]) \otimes (n + [N,L]) \in \frac{L}{[N,L]} \otimes \frac{N}{[N,L]}
\end{equation}
and its restriction
\begin{equation}\label{epi2}\pi_|: n \ \square \ n \in L \ \square \ N \mapsto (n + [N,L]) \ \square \ (n + [N,L]) \in \frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}.\end{equation}
The kernel of \eqref{epi1} is studied in the next result. We inform the reader that we will use the notion of \textit{Lie pairing} in the next proof. This can be found in \cite[pp.101--102]{ellis2}, together with its fundamental properties.
\begin{lem}\label{5}
Let $N$ be an ideal of a Lie algebra $L$, $\pi$ as in \eqref{epi1} and \[M= (L \otimes [N,L]) + ([N,L] \otimes N).\] Then $\ker \pi= M.$
\end{lem}
\begin{proof}
Since $L \otimes N \supseteq M$ and $M$ is an ideal of $L\otimes N$, $M$ induces the homomorphism
\[\bar{\pi}: (l\otimes n) + M \in \frac{L \otimes N}{M} \mapsto (l + [N,L]) \otimes (n + [N,L]) \in \frac{L}{[N,L]} \otimes \frac{N}{[N,L]}.\]
On the other hand, define
\[ \alpha: (l+[N,L],n+[N,L]) \in \frac{L}{[N,L]} \times \frac{N}{[N,L]} \mapsto (l\otimes n) + M \in \frac{L \otimes N} {M}.\]
It is easy to check that $\alpha$ is well defined. Now for all $l_1,l_2 \in L$ and $n_1,n_2\in N$
\[\alpha( [l_1+[N,L],l_2+[N,L]] \ , \ n_1+[N,L] )\]
\[=([l_1,l_2] \otimes n_1) + M = (l_1\otimes [l_2,n_1]-l_2\otimes [l_1,n_1]) + M\]
\[=\alpha(l_1+[N,L],[l_2,n_1]+[N,L])-\alpha(l_2+[N,L],[l_1,n_1]+[N,L]).\]
Similarly, it is easy to see that
\[\alpha( ^{n_1+[N,L]} l_1+[N,L] \ , \ \ ^{l_1+[N,L]}n_2+[N,L] )\]
\[=-[\alpha(l_1+[N,L],n_1+[N,L]) \ , \ \alpha(l_2+[N,L],n_2+[N,L])]\]
and
\[\alpha(l_1+[N,L] \ , \ [n_1,n_2]+[N,L])\]
\[=\alpha(^{n_2+[N,L]} l_1+[N,L] \ , \ n_1+[N,L]) - \alpha(^{n_1+[N,L]}l_1+[N,L] \ , \ n_2+[N,L]).\]
Thus $\alpha$ is a Lie pairing and induces the homomorphism
\[\bar{\alpha}: \frac{L}{[N,L]} \otimes \frac{N}{[N,L]} \longrightarrow \frac{L\otimes N}{M}\] such that $\bar{\pi} \circ \bar{\alpha}=\bar{\alpha} \circ \bar{\pi}= 1$. Therefore \[\frac{L}{[N,L]} \otimes \frac{N}{[N,L]} \simeq \frac{L \otimes N}{M}\] and the result follows.
\end{proof}
The assumption that a Lie algebra must be of finite dimension appears now for the first time. The main results of the present section will deal with this kind of Lie algebras only.
\begin{rem}\label{extra1} Lemma \ref{1} and Corollary \ref{2} are true in particular when we replace the assumption that $N/[N,L]$ is free with $N/[N,L]$ of finite dimension. The above conditions are satisfied, for instance, when $L$ is of finite dimension.
\end{rem}
Firstly, we describe $\Gamma(N/[N,L])$ (in \eqref{diag}) via a suitable isomorphism with $L \square N$.
\begin{cor}\label{3}
Let $N$ be an ideal of a Lie algebra $L$ such that $N/[N,L]$ is of finite dimension. Then
\[\Gamma\left(\frac{N}{[N,L]}\right) \simeq L \ \square \ N \simeq \frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}.\]
\end{cor}
\begin{proof}
By Lemma \ref{1}, $\mathrm{Im} \ \psi = L \square N \simeq \Gamma(N/[N,L])$. On the other hand, \[\Gamma\left(\frac{N}{[N,L]}\right) \simeq \frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}.\]
Thus
$L\square N \simeq L/[N,L] \ \square \ N/[N,L]$, as required.
\end{proof}
Secondly, we describe better one of the isomorphisms in Corollary \ref{3}.
\begin{lem}\label{4}
Let $N$ be an ideal of a Lie algebra $L$ such that $N/[N,L]$ is of finite dimension. Then
the natural epimorphism \eqref{epi2} is injective and so it is an isomorphism.
\end{lem}
\begin{proof}
By Corollary $\ref{3}$, we have $\dim L \ \square \ N =\dim L/[N,L] \ \square \ N/[N,L]$ and so \eqref{epi2} is an isomorphism, because it is an epimorphism of abelian Lie algebras of same dimension.
\end{proof}
Now we begin to look for information on the bases of those Lie algebras, which will be involved in the main results of the present section. The abelian case plays a fundamental role and is discussed in the next result.
\begin{prop}\label{basis}Let $N= \langle x_1, x_2, \ldots , x_m \rangle$ be an ideal of dimension $m$ of an abelian Lie algebra $L = \langle x_1, x_2, \ldots, x_m, y_{m+1}, y_{m+2}, \ldots, y_n \rangle$ of dimension $n$. Then
\[L \otimes N \simeq (L \ \square \ N) \ \oplus \ \langle y_j\otimes x_t \ | \ 1\leq t\leq m, \ m+1 \leq j \leq n \rangle.\]
\end{prop}
\begin{proof}
From the decomposition of finite dimensional abelian Lie algebras (see \cite{knapp}) in one dimensional ideals, we have \[L \simeq \bigoplus^{n}_{i=m+1} \langle y_i \rangle \ \oplus \ \ \bigoplus^{m}_{j=1}\langle x_j \rangle, \ \mathrm{where} \ N \simeq \bigoplus^{m}_{j=1}\langle x_j \rangle\] and so
\[L \otimes N \simeq \langle x_i \otimes x_t + x_t \otimes x_i, x_i\otimes x_i \ | \ 1 \leq i < t \leq m \rangle\]
\[\oplus \ \langle y_j \otimes x_t \ | \ 1 \leq t \ \leq m, m+1 \leq j \leq n\rangle.\]
Now the result follows from the fact that
\[L \ \square \ N \simeq \langle x_i\otimes x_t + x_t \otimes x_i, x_i \otimes x_i \ | \ 1\leq i <t \leq m\rangle .\]
\end{proof}
We are ready to prove the first main result of the present section.
\begin{thm}\label{6}
Let $N$ be an ideal of a Lie algebra $L$ such that $L/[N,L]$ is of dimension $n$ with basis $\{ \bar{x}_{_1},\bar{x}_{_2}, \ldots,\bar {x}_{_m},\bar{y}_{_{m+1}}, \bar{y}_{_{m+2}}, \ldots,\bar{y}_{_n}\}$ and $N/[N,L]$ of dimension $m$ with basis $\{ \bar{x}_{_1},\bar{x}_{_2}, \ldots,\bar {x}_{_m}\}$. Then
\[\frac{L}{[N,L]} \ \otimes \ \frac{N}{[N,L]} \simeq \left(\frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}\right) \ \oplus \ \langle \bar{y_j} \otimes \bar{x_i} \ |\ 1\leq i \leq m, m+1\leq j \leq n \rangle.\]
\end{thm}
\begin{proof}
We do two observations. The first is that $[L/[N,L],N/[N,L]]=0$. The second is that the actions of $L/[N,L]$ on $N/[N,L]$, and viceversa of $N/[N,L]$ on $L/[N,L]$, are trivial, compatible and by conjugation. Then we may conclude that
\[\frac{L}{[N,L]} \ \otimes \ \frac{N}{[N,L]} \simeq {\left(\frac{L}{[N,L]}\right)}^{ab} \ \otimes_{\mathbb{Z}} \ \frac{N}{[N,L]}.\] From Corollary \ref{2}, we have
\[{\left(\frac{L}{[N,L]}\right)}^{ab} \ \otimes_{\mathbb{Z}} \ \frac{N}{[N,L]} \simeq \left({\left(\frac{L}{[N,L]}\right)}^{ab} \ \square \ \frac{N}{[N,L]}\right) \ \oplus \ \left({\left(\frac{L}{[N,L]}\right)}^{ab} \ \wedge \ \frac{N}{[N,L]}\right),\]
where the abelian Lie algebra factor $(L/[N,L])^{ab} \ \wedge \ N/[N,L]$ admits a basis exactly of the form $\{ \bar{y_j} \otimes \bar{x_i} \ |\ 1\leq i \leq m, m+1\leq j \leq n \}$ while the other abelian Lie algebra factor is $(L/[N,L])^{ab} \ \square \ N/[N,L] \simeq L/[N,L] \ \square \ N/[N,L]. $ Then the result follows from Proposition \ref{basis}.
\end{proof}
Of course, Theorem \ref{6} is true when the entire $L$ is of finite dimension, not only its factor $L/[N,L]$. The second main result of this section may be formulated again with the restriction of finite dimension on the factor $L/[N,L]$.
It describes a pure splitting of the operator $\otimes$ with respect to the operators $\square$ and $\wedge$.
\begin{thm}\label{7}
Let $N$ be an ideal of a Lie algebra $L$ such that $L/[N,L]$ is of finite dimension. Then
\[ L\otimes N \simeq (L \ \square \ N) \oplus (L \wedge N).\]
\end{thm}
\begin{proof}
Following the notations of Theorem \ref{6}, we note that
\[\frac{L}{[N,L]} \otimes \frac{N}{[N,L]} \simeq \left(\frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}\right) \oplus \langle\bar{y_j}\otimes \bar{x_i} \ | \ 1 \leq i \leq m, m+1\leq j \leq n\rangle.\]
On the other hand, $\pi_|$ in \eqref{epi2} is an isomorphism by Lemma \ref{4} and so
\[\left(\frac{L}{[N,L]} \ \square \ \frac{N}{[N,L]}\right) \oplus \langle \bar{y_j}\otimes \bar{x_i} \ | \ 1\leq i \leq m, m+1\leq j \leq n \rangle\]
\[=\pi_| ((L \square N) \ + \ \langle y_j \otimes x_i \ | \ 1 \leq i \leq m, m+1\leq j \leq n \rangle ),\]
where $\{\bar{x}_{_1},\bar{x}_{_2}, \ldots, \bar {x}_{_m} \}$ is a basis of $N/[N,L]$ and $\{\bar{x}_{_1},\bar{x}_{_2}, \ldots,\bar {x}_{_m}, \bar{y}_{_{m+1}}, \bar{y}_{_{m+2}}, \ldots, \bar{y}_{_n}\}$ of $L/[N,L]$. Following the notation adopted in Lemma \ref{5}, we find that
\[L\otimes N = (L\square N) + \langle y_j \otimes x_i \ | \ 1\leq i \leq m, m+1\leq j \leq n\rangle + M.\] Now \eqref{epi2} maps all the elements of $(L \ \square \ N) \cap (\langle y_j \otimes x_i \ | \ 1\leq i \leq m, m+1\leq j \leq n\rangle + M)$ to zero, and so
\[L\otimes N \simeq (L \ \square \ N) \ \oplus \ (\langle y_j \otimes x_i \ | \ 1\leq i \leq m, m+1\leq j \leq n\rangle + M).\]
Since we may easily check that $ L\wedge N \simeq \langle y_j\otimes x_i \ | \ 1\leq i \leq m,m+1\leq j \leq n\rangle +M $, the result follows.
\end{proof}
The role of the Lie algebra $J_2(L,N)$ has been investigated in \cite{ellis2} when $N=L$ and it is related to the homotopy theory in the sense of \cite[Theorems 27, 28]{ellis2}. There is not a version of \cite[Theorems 27, 28]{ellis2}, when $N \neq L$, even if some ideas can be found in \cite[Theorems 1, 2, 4, 5]{ellis4} for the case of groups. This emphasizes the following consequence of Theorem \ref{7}.
\begin{cor}\label{8}
Let $N$ be an ideal of a Lie algebra $L$ such that $L/[N,L]$ is of finite dimension. Then
\[J_2(L,N) \simeq (L \ \square \ N) \oplus M(L,N).\]
\end{cor}
\begin{proof}
By Theorem \ref{7},
\[J_2(L,N)=((L \ \square N) \oplus ( \langle y_j \otimes x_i \ | \ 1\leq i \leq m, m+1\leq j \leq n\rangle +M )) \cap J_2(L,N)\]
\[=(L \ \square \ N) \oplus (\langle y_j\otimes x_i \ | \ 1\leq i \leq m, m+1\leq j \leq n\rangle +M) \cap J_2(L,N).\]
The rest follows by Corollary \ref{2}; specifically by $J_2(L,N)/ (L \ \square \ N) \simeq M(L,N)$.
\end{proof}
\section{K\"unneth-type formulas}
Dealing with homology of algebraic structures, there are some standard results, which help to understand the splitting of the second homology Lie algebra (with integral coefficients) with respect to direct sums. Of course, one may ask whether similar formulas happen for other operators, and not necessarily for the sum with respect to the second homology functor.
We list some of these splittings in our context of investigation. They are often called \textit{K\"unneth type formulas}, due to the original version of K\"unneth (see \cite{knapp})
\begin{prop}[See Propositions 15, 21, Theorem 35 of \cite{ellis2}]\label{folklore} Two Lie algebras $H$ and $K$ satisfy the conditions:
\begin{itemize}\item[(i)] $\Gamma((H \oplus K)^{ab}) \simeq \Gamma(H^{ab}) \oplus \Gamma(K^{ab}) \oplus (H^{ab} \ \otimes_\mathbb{Z} \ K^{ab})$;
\item[(ii)]$M(H \oplus K) \simeq M(H) \oplus M(K) \oplus (H^{ab} \ \otimes_\mathbb{Z} \ K^{ab});$
\item[(iii)]$J_2(H \oplus K) \simeq J_2(H) \oplus J_2(K) \oplus (H^{ab} \ \otimes_\mathbb{Z} \ K^{ab})\oplus (K^{ab} \ \otimes_\mathbb{Z} \ H^{ab}).$
\end{itemize}
\end{prop}
Another version of Proposition \ref{folklore} can be formulated, when we consider the \textit{free product} $H * K$ of $H$ and $K$ in the usual sense (see \cite{knapp}). Roughly speaking, the ``nonlinear'' factor $H^{ab} \ \otimes_\mathbb{Z} \ K^{ab}$ in (ii) above vanishes in (ii) below, and so there is a perfect splitting. Similarly, $K^{ab} \ \otimes_\mathbb{Z} \ H^{ab}$ in (iii) above vanishes in (iii) below.
\begin{prop}[See Proposition 19 and its proof in \cite{ellis2}]\label{folklorebis} Two Lie algebras $H$ and $K$ satisfy the conditions:
\begin{itemize}\item[(i)] $\Gamma((H * K)^{ab}) \simeq \Gamma(H^{ab}) \oplus \Gamma(K^{ab}) \oplus (H^{ab} \ \otimes_\mathbb{Z} \ K^{ab})$;
\item[(ii)]$M(H * K) \simeq M(H) \oplus M(K);$
\item[(iii)]$J_2(H * K) \simeq J_2(H) \oplus J_2(K) \oplus (H^{ab} \ \otimes_\mathbb{Z} \ K^{ab}).$
\end{itemize}
\end{prop}
What happens to Propositions \ref{folklore} and \ref{folklorebis} for pairs of Lie algebras ?
Of course, we expect to have generalizations. The answer is positive and can be found in the context of groups in \cite{ellis3, ellis4, mas1}, even if some authors are beginning to investigate pairs of Lie algebras in this perspective (see \cite{ri1, ri2, sam, n2, n4}). This motivated us to prove the following result.
\begin{thm}\label{freeproducts}
Let $N_1$ and $N_2$ two ideals of two finite dimensional Lie algebras $L_1$ and $L_2$ such that $(L_1 * L_2, N_1 * N_2 )$ is a pair. Then
\begin{itemize}
\item[(i)]$J_2(L_1 * L_2, N_1 * N_2) \simeq J_2(L_1, N_1) \oplus J_2(L_2, N_2) \oplus \left(\frac{N_1}{[N_1,L_1]} \otimes \frac{N_2}{[N_2,L_2]}\right);$
\item[(ii)] $(L_1 * L_2) \square (N_1 * N_2) \simeq (L_1 \oplus N_1) \square (L_2 \oplus N_2);$
\item[(iii)]$M(L_1 * L_2, N_1 * N_2) \simeq M(L_1, N_1) \oplus M(L_2, N_2)$, whenever $N_1$ admits a complement in $L_1$ such that $[N_1,L_1]=[L_1,L_1]$.
\end{itemize}
\end{thm}
\begin{proof}
(i). The natural monomorphisms $l_1 \in L_1 \mapsto l_1 \in L_1 *L_2$,
$l_2 \in L_2 \mapsto l_2 \in L_1 *L_2$, $n_1 \in N_1 \mapsto n_1 \in N_1 * N_2$ and $n_2 \in N_2 \mapsto n_2 \in N_1 * N_2$ induce the natural monomorphisms
$\iota : l_1 \otimes n_1 \in L_1 \otimes N_1 \mapsto l_1 \otimes n_1 \in (L_1 * L_2) \otimes (N_1 * N_2)$
and
$j : l_2 \otimes n_2 \in L_2 \otimes N_2 \mapsto l_2 \otimes n_2 \in (L_1 * L_2) \otimes (N_1 * N_2)$ and so the following map
\[\mu : (l_1 \otimes n_1, l_2 \otimes n_2) \in (L_1 \otimes N_1) \oplus (L_2 \otimes N_2) \mapsto
\iota (l_1 \otimes n_1) + j (l_2 \otimes n_2) \in (L_1 * L_2) \otimes (N_1 * N_2)\]
restricts to an injective homomorphism
\[\zeta : J_2(L_1, N_1) \oplus J_2(L_2,N_2) \rightarrow J_2(L_1 * L_2, N_1 * N_2).\]
On the other hand,
\[\eta : \overline{n_1} \otimes \overline{n_2} \in \frac{N_1}{[N_1,L_1]} \otimes \frac{N_2}{[N_2,L_2]} \mapsto (n_1 \otimes n_2) + (n_2 \otimes n_1) \in J_2(L_1 * L_2, N_1 *N_2), \]
where $\overline{n_1}=n_1 + [N_1,L_1]$ and $\overline{n_2}=n_2 + [N_2,L_2]$, is a well defined epimorphism. Hence we consider $\xi = \zeta \oplus \eta$, where
\[\xi : (x,y,z) \in J_2(L_1, N_1) \oplus J_2(L_2, N_2) \oplus \left(\frac{N_1}{[L_1,N_1]} \otimes \frac{N_2}{[L_2,N_2]}\right) \longmapsto\]
\[\xi(x,y,z)=(\zeta \oplus \eta) (x,y,z)=\zeta(x,y) + \eta(z) \in J_2(L_1 * L_2, N_1 *N_2). \]
We claim that $\xi$ is bijective. Firstly, we note that
\[p : (L_1 * L_2) \otimes (N_1 * N_2) \rightarrow (L_1 \otimes N_1) \oplus (L_2 \otimes N_2) \oplus (L_1 \otimes N_2) \oplus (L_2 \otimes N_2)\]
is a canonical epimorphism and so
\[\xi (x,y,z) = 0 \Leftrightarrow (p \circ \xi) (x,y,0)= (p \circ \xi) (0,0, -z) \Leftrightarrow \xi (x,y,0)=0 \Leftrightarrow x=y=z=0. \]
This is enough to conclude that $\xi$ is injective. Now we show that $\xi$ is surjective. It is clear
that $(L_1 * L_2) \otimes (N_1 * N_2)$ is finitely
presented, because we have generators and relations by definition of $\otimes$. Using $i$, $j$ and $\mu$, $L_1 \otimes
N_1$, $L_2 \otimes N_2$, $L_1 \otimes N_2$, $L_2 \otimes
N_1$ (and their sums) are isomorphic to suitable ideals
of $(L_1 * L_2) \otimes (N_1 * N_2)$. Then we may consider $t \in (L_1 * L_2) \otimes (N_1 * N_2)$, $U=L_1 \otimes N_1$, $W=L_2 \otimes N_2$, $V= (L_1 \otimes N_2) + (L_2 \otimes N_1)$ and $t=u+v+w$ for some $u \in U$, $v \in V$, $w \in W$. Here the commutator map $\kappa_{L_1*L_2, N_1*N_2}=\kappa$ is defined by \[\kappa : t \in (L_1 * L_2) \otimes (N_1 * N_2) \mapsto \kappa(t)=a+b+c \in [N_1,L_1] \oplus [N_2,L_2] \oplus [L_1, L_2]\]
and is an epimorphism. Note that
$\kappa(t)=0 \Leftrightarrow a=b=0 $ and so $\kappa(t)=\kappa(v)=c=0$. This implies $u \in J_2(L_1,N_1)$ and $w \in J_2(L_2,N_2)$. We also note that $\kappa(v)$ is freely reduced in $[L_1,L_2]$, whenever $v$ does not contain words of the form $k_1=(n_1 \otimes n_2) + (n_2 \otimes n_1)$ or $k_2=(n_2 \otimes n_1) + (n_1 \otimes n_2)$ in it. This means that there is no loss of generality in assuming $v=h+k_i+l$ with $h,l \in V$, $i=1,2$ and $k_1, k_2$ as before. Note that $\kappa(h+l)=\kappa(v)=0$. Now one can do induction on the number of expressions of the form $k_1, k_2$ and note that $\eta$ is an epimorphism. This allows us to conclude that $\xi$ is surjective and so we have
\[J_2(L_1 * L_2, N_1 * N_2) \simeq J_2(L_1, N_1) \oplus J_2(L_2, N_2) \oplus \left(\frac{N_1}{[N_1,L_1]} \otimes \frac{N_2}{[N_2,L_2]}\right). \]
(ii). Having in mind $\iota$ and $j$ of the previous step (i), we may deduce from \eqref{diag} the commutativity of the following diagram
\[\begin{CD}
\Gamma\left(\frac{L_1 * L_2}{[N_1*N_2, L_1*L_2]} \right) {\overset{\lambda}\longrightarrow} J_2(L_1*L_2,N_1*N_2) \longrightarrow M(L_1 * L_2, N_1*N_2) \longrightarrow 0 \\
\uparrow \hspace{3.5cm} \uparrow \hspace{3cm} \uparrow \\
\Gamma\left(\frac{L_1 \oplus L_2}{[N_1 \oplus N_2, L_1 \oplus L_2]} \right) {\overset{\nu}\longrightarrow} J_2(L_1 \oplus L_2, N_1 \oplus N_2) \longrightarrow M(L_1 \oplus L_2, N_1 \oplus N_2) \longrightarrow 0 \end{CD}\]
which has exact sequences as rows.
Since
\[\mathrm{Im} \ \lambda= \langle x \otimes x \ | \ x \in (L_1*L_2) \cap (N_1 *N_2)\rangle= (L_1 * L_2) \square (N_1 * N_2) \] and
\[\mathrm{Im} \ \nu= \langle (a,b) \otimes (a,b) \ | \ (a,b) \in (L_1 \otimes L_2) \oplus (N_1 \otimes N_2) \rangle= (L_1 \oplus L_2) \ \square \ (N_1 \oplus N_2),\]
our scope is to prove that $\mathrm{Im} \ \lambda \simeq \mathrm{Im} \ \nu$.
We note that
the epimorphism $L_1*L_2 \longrightarrow L_1 \oplus L_2$ induces
an epimorphism \[\beta :(L_1*L_2) \otimes (L_1 * L_2)\longrightarrow (L_1 \oplus
L_2) \otimes (L_1 \oplus L_2)\] which may be restricted to an epimorphism $\beta_|$ from
$\mathrm{Im} \ \lambda$ onto $\mathrm{Im} \ \nu$. So, it is
enough to find a left inverse for $\beta_|$. We proceed to do this.
Keeping the notations of the above step (i), the canonical homomorphism
\[\varphi : (L_1 \oplus N_1) \otimes (L_2 \oplus N_1) \otimes (L_1 \oplus N_2) \otimes (L_2 \oplus N_2) \rightarrow\]
\[ (L_1 \otimes N_1) \oplus (L_2 \otimes N_1) \oplus (L_1 \otimes N_2) \oplus (L_2 \otimes N_2)\]
induces the following epimorphism
\[\sigma : (L_1 \oplus N_1) \otimes (L_2 \oplus N_1) \otimes (L_1 \oplus N_2) \otimes (L_2 \oplus N_2) \rightarrow
\left(\frac{L_1}{[N_1,L_1]} \otimes \frac{N_1}{[N_1,L_1]}\right) \oplus\]\[ \left(\frac{L_2}{[N_2,L_2]} \otimes \frac{N_1}{[N_1, L_1]}\right) \oplus \left(\frac{L_1}{[N_1,L_1]} \otimes \frac{N_2}{[N_2,L_2]}\right) \oplus \left(\frac{L_2}{[N_2,L_2]} \otimes \frac{N_2}{[N_2,L_2]}\right).\]
We denote by $\mathrm{proj}_2$ the projection of $\sigma$ onto the second factor. Now we have all that is necessary to define the left inverse of $\beta_|$, this is the map
\[\theta=\beta_| \circ \mathrm{proj}_2 \circ \sigma_| \circ \zeta \circ \varphi_|\]
where $\varphi_|$ denotes the restriction of $\varphi$ to $\mathrm{Im} \ \nu$ and $\sigma_|$ that of $\sigma$ again to $\mathrm{Im} \ \nu$. Now one can check that $\theta \circ \beta_| =1$ and the result follows.
(iii). Applying \cite[Lemma 7 (iii)]{ellis2} twice, that is, the distributive law
of the form $(L \oplus M) \otimes N \simeq (L \otimes
N) \oplus (M \otimes N)$ for three given Lie algebras $L$, $M$ and $N$, we get the isomorphism
\[(L_1 \oplus L_2) \otimes (N_1 \oplus N_2) \simeq (L_1 \otimes N_1) \oplus (L_2 \otimes N_1) \oplus (L_1 \otimes N_2) \oplus (L_2 \otimes N_2)\]
and this induces $(L_1 \oplus L_2) \ \square \ (N_1 \oplus N_2) \simeq (L_1 \ \square \ N_1) \oplus (L_2 \ \square \ N_2) \oplus V,$
where $V= \langle (n_2 \otimes n_1)+(n_1 \otimes n_2) \ | \ n_1 \in N_1, n_2 \in N_2\rangle \subseteq (L_2 \otimes N_1) + (L_1 \otimes N_2).$ Since $N_1$ admits a complement in $L_1$, the homomorphism $\eta$ in (i) above extends to an isomorphism onto $V$. Therefore the result follows from (ii) above and the commutativity of \eqref{diag}.
\end{proof}
It is possible to describe the Lie algebra quotient $J_2(L_1 \oplus L_2, N_1 \oplus N_2)/J_2(L_1 * L_2, N_1 * N_2) $ very well and this is done in the following result.
\begin{cor} Let $N_1$ and $N_2$ two ideals of two finite dimensional Lie algebras $L_1$ and $L_2$ such that $(L_1 * L_2, N_1 * N_2 )$ is a pair. Then
\[\frac{J_2(L_1 \oplus L_2, N_1 \oplus N_2)}{J_2(L_1 * L_2, N_1 * N_2)} \simeq \frac{N_1}{[N_1,L_1]} \otimes \frac{N_2}{[N_2,L_2]}. \]
\end{cor}
\begin{proof}
This follows from Theorem \ref{freeproducts} and Corollary \ref{8}.
\end{proof}
|
\section{Introduction}\label{s:intro}
In this work, we focus on the heterogeneous reaction-diffusion
equation
\begin{equation}\label{monostable}
\partial _t u= \mbox{div} (A (x) \nabla u) + q (x) \cdot \nabla u + f(
x,u), \quad t\in \mathbb R, x\in \mathbb R ^N.
\end{equation}
Here $A = (A_{i,j})_{1 \leq i,j \leq N}$ is a matrix field, and $q = (q_1 , ... ,q_N)$
is a vector
field, to be precised later. The nonlinearity $f$ is of either the
monostable type (not necessarily with the KPP assumption) or
ignition type, which we will define below. We would like to
understand, in the periodic framework, how the propagation
phenomena depend on the direction.
On the one hand, we prove that the minimal (and, in the ignition
case, unique) speed of the well known pulsating fronts depends
continuously on the direction of propagation. On the other hand,
we prove that the spreading properties are in some sense uniform
with respect to the direction, thus improving the seminal result
of Weinberger \cite{Wein02}. While in the ignition case, these
properties will mostly follow from the well known uniqueness of
the pulsating traveling wave, such uniqueness does not hold true
in the monostable case where the set of admissible speeds is
infinite. Our argument will be inspired
by~\cite{berestycki-nirenberg}, \cite{Ber-Ham-02}, and will rely
on an approximation of the monostable nonlinearity by some
well-chosen ignition nonlinearities.
\subsection{Main assumptions}
Let $L_1$,...,$L_N$ be given positive constants. A function $h:\mathbb R
^ N \to \mathbb R$ is said to be {\it periodic} if
$$
h(x_1,...,x_k+L_k,...,x_N)=h(x_1,...,x_N),
$$
for all $1\leq k \leq N$, all $(x_1,...,x_N)\in \mathbb R^N$. In such
case, $\mathcal{C} = (0,L_1)\times\cdots \times (0,L_N)$ is called
the cell of periodicity. Through this work, we put ourselves in
the spatially periodic framework and assume that
\begin{equation}\label{periodicity}
\begin{array}{l}
\mbox{ for all } 0 \leq i, j \leq N, \, \mbox{ the functions } A_{i,j}: \mathbb R^N \to \mathbb R \, , \ q_i : \mathbb R^N \to \mathbb R \quad \mbox{ are periodic},\vspace{3pt}\\
\text{ for all } u\in \mathbb R_+,\, \mbox{ the function } f(\cdot,u):
\mathbb R^N \to \mathbb R \text{ is periodic}.
\end{array}
\end{equation}
Moreover, we assume that $A = (A_{i,j})_{1 \leq i,j \leq N}$ is a $C^3$ matrix field which satisfies
\begin{equation}\label{ellipticity}
\begin{array}{l}
A(x) \mbox{ is a symmetric matrix for any } x \in \mathbb R^N,\vspace{3pt}\\
\exists 0 < a_1 \leq a_2 < \infty, \quad \forall (x, \xi) \in \mathbb R^N
\times \mathbb R^N, \quad a_1 |\xi|^2 \leq \sum_{i,j} A_{i,j} (x) \xi_i
\xi_j \leq a_2 |\xi |^2.
\end{array}
\end{equation}
Concerning the advection term, we assume that $q = (q_1 , ... ,q_N)$
is a $C^{1,\delta}$ vector
field, for some $\delta >0$, which satisfies
\begin{equation}
\begin{array}{l}
\mbox{div} \, q =0 \mbox{ in } \mathbb R^N \quad \mbox{ and } \quad
\forall 0 \leq i \leq N , \ \displaystyle \int_{\mathcal{C}} q_i =
0.
\end{array}
\end{equation}
The advection term in the equation is mostly motivated by
combustion
models where the dynamics of the medium also plays an essential role. In such
a context, the fact that the flow $q$ has zero divergence carries the physical
meaning that the medium is incompressible.
Furthermore, we will assume that $f$ satisfies either of the
following two assumptions.
\begin{assumption}[Monostable nonlinearity]\label{hyp:monostable} The function $f:\mathbb R ^N \times \mathbb R_+\to \mathbb R$ is of class $C^{1,\alpha}$ in $(x,u)$
and~$ C^2$ in $u$, and nonnegative on $\mathbb R^N\times [0,1]$. Concerning the steady states of the periodic
equation~\eqref{monostable}, we assume that
\begin{enumerate}[(i)]
\item the constants 0 and 1 are steady states (that is,
$f(\cdot,0)\equiv f (\cdot,1) \equiv
0$ in $\mathbb R^N$);
\item $\forall u \in (0,1), \ \exists x \in \mathbb R^N, \ \ f(x,u) >0$.
\item there exists some $\rho
>0$ such that $f (x,u)$ is nonincreasing with respect to $u$ in
the set $\mathbb R^N \times (1-\rho,1]$.
\end{enumerate}
\end{assumption}
Notice that, if $0\leq p(x)\leq 1$ is a periodic stationary state, then $p\equiv 0$ or $p\equiv 1$. Indeed, since $f(x,p)\geq 0$, the strong maximum principle enforces $p$ to be identically equal to its minimum, thus constant and, by $(ii)$, the constant has to be 0 or 1.
\begin{assumption}[Ignition nonlinearity]\label{hyp:ignition} The function $f:\mathbb R ^N \times \mathbb R_+\to \mathbb R$ is locally Lipschitz-continuous on $\mathbb R^N \times \mathbb R_+$.
Concerning the steady states of the periodic
equation~\eqref{monostable}, we assume that
\begin{enumerate}[(i)]
\item there exists $0<\theta<1 $ such that
$$\forall 0 \leq u \leq \theta, \ \ \forall x \in \mathbb R^N, \qquad f(x,u)=0,$$
as well as
$$\forall x \in \mathbb R^N , \qquad f(x,1)=0.$$
\item $\forall u \in (\theta,1), \ \exists x \in \mathbb R^N, \ \ f(x,u)
>0$.
\item there exists some
$\rho
>0$ such that $f (x,u)$ is nonincreasing with respect to $u$ in
the set $\mathbb R^N \times (1-\rho,1]$.
\end{enumerate}
\end{assumption}
Notice that, similarly as above, $(ii)$ implies that if $\theta\leq p(x)\leq 1$ is a periodic stationary state then $p\equiv \theta$ or $p\equiv 1$.
\subsection{Comments and related results}
Under Assumption~\ref{hyp:monostable},
Assumption~\ref{hyp:ignition}, equation~\eqref{monostable} is
referred to as the monostable equation, the ignition equation
respectively. Both sets of assumptions arise in various fields of
physics and the life sciences, and especially in combustion and
population dynamics models where propagation phenomena are
involved. Indeed, a particular feature of these equations is the
formation of traveling fronts, that is particular solutions
describing the transition at a constant speed from one stationary
solution to another. Such solutions have proved in numerous
situations their utility in describing the spatio-temporal
dynamics of a population, or the propagation of a flame modelled
by a reaction-diffusion equation.
Equation \eqref{monostable} is a heterogeneous version of the well
known reaction-diffusion equation
\begin{equation}\label{homogene}
\partial _t u=\Delta u +f(u),
\end{equation}
where typically $f$ belongs to one of the three following classes:
monostable, ignition and bistable. Homogeneous reaction-diffusion
equations have been extensively studied in the literature (see
\cite{Kan1}, \cite{Aro-Wei1, Aro-Wei2}, \cite{Fif-Mac},
\cite{Ber-Nic-Sch}, \cite{Volpert-Volpert-Volpert} among others)
and are known to support the existence of monotone traveling
fronts. In particular, for monostable nonlinearities, there exists
a critical speed $c^*$ such that all speeds $c\geq c^*$ are
admissible, while in the bistable and ignition cases, the
admissible speed $c= c^*$ is unique. Moreover, in both cases, the
speed $c^*$ corresponds to the so-called spreading speed of
propagation of compactly supported initial data.
Among monostable nonlinearities, one can distinguish the ones
satisfying the Fisher-KPP assumption, namely $u\mapsto
\frac{f(u)}u$ is maximal at 0 (meaning that the growth per capita
is maximal at small densities), the most famous example being
introduced by Fisher \cite{Fish} and Kolmogorov, Petrovsky and
Piskunov \cite{Kol-Pet-Pis} to model the spreading of advantageous
genetic features in a population:
\begin{equation*}
\partial_t u=\Delta u+u(1-u).
\end{equation*}
Let us notice that our work stands in the larger class of
monostable nonlinearities.
Nevertheless, much attention was more recently devoted to the
introduction of some heterogeneity, taking various forms such as
advection, spatially dependent diffusion or reaction term. Taking
such a matter into account is essential as far as models are
concerned, the environment being rarely homogeneous and may depend
in a non trivial way on the position in space (patches, periodic
media, or more general heterogeneity...). We refer to the seminal
book of Shigesada and Kawasaki \cite{Shi-Kaw}, and the
enlightening introduction in \cite{Ber-Ham-Roq1} where the reader
can find very precise and various references. As far as combustion models are concerned, one can consult \cite{Ber-Nic-Sch}, \cite{Xin} and the references therein.
Traveling front solutions in heterogeneous versions of
\eqref{homogene} with periodicity in space, in time, or more
general media are studied in \cite{Wein02},
\cite{berestycki-nirenberg}, \cite{HZ}, \cite{Xin3},
\cite{Ber-Ham-02}, \cite{Ber-Ham-Roq2}, \cite{Nad-09},
\cite{Nolen-Ryzhik} among others. For very general
reaction-diffusion equations, we refer to
\cite{berestycki-hamel-cpam2} for a definition of generalized
transition waves and their properties.
In this work, we restrict ourselves to the spatially periodic
case, which provides insightful information on the role and
influence of the heterogeneity on the propagation, as well as a
slightly more common mathematical framework. In this periodic
setting, let us mention the following keystone results for
ignition and monostable nonlinearities. Weinberger \cite{Wein02}
exhibited a direction dependent spreading speed for planar-shaped
initial data and proved, in the monostable case, that this
spreading speed is also the minimal speed of pulsating traveling
waves moving in the same direction. His approach relies on a
discrete formalism, in contrast with the construction of both
monostable and ignition pulsating traveling waves by Berestycki
and Hamel \cite{Ber-Ham-02}, via more flexible PDE technics. In
this PDE framework, note also the work of Berestycki, Hamel and
Roques \cite{Ber-Ham-Roq2} where KPP pulsating fronts are
constructed without assuming the nonnegativity of the
nonlinearity. Our main goal is to study how these results behave
when we vary the direction of propagation.
\medskip
Let us give another motivation for our analysis of the dependence
of the propagation on the direction. Our primary interest was
actually to study the {\it sharp interface limit} ${\varepsilon} \to 0$ of
\begin{equation}\label{papier2}
\partial _t u^{\varepsilon}= {\varepsilon} \Delta u^{\varepsilon}+\displaystyle \frac 1 {\varepsilon} f\left(\frac x
{\varepsilon},u^{\varepsilon}\right),
\end{equation}
arising from the hyperbolic space-time rescaling
$u^{\varepsilon}(t,x):=u\left(\frac t {\varepsilon}, \frac x {\varepsilon}\right)$ of
\eqref{monostable}, with $A\equiv Id$, $q\equiv 0$. The parameter
${\varepsilon}>0$ measures the thickness of the diffuse interfacial layer.
As this thickness tends to zero, \eqref{papier2} converges --- in
some sense--- to a limit interface, whose motion is governed by
the minimal speed (in each direction) of the underlying pulsating
fronts. This dependence of the speed on the (moving) normal
direction is in contrast with the homogeneous case and makes the
analysis quite involved. In particular, it turns out that we need
to improve (by studying the uniformity with respect to the
direction) the known spreading properties \cite{Wein02},
\cite{Ber-Ham-02}, for both ignition and monostable nonlinearities
in periodic media. We refer to \cite{A-Gil} for this singular
limit analysis, using some of the results of the present work.
\section{Main results}\label{s:results}
Before stating our main results in subsection~\ref{ss:varying},
let us recall the classical results on both pulsating fronts and
spreading properties in subsection~\ref{ss:known}.
\subsection{Pulsating fronts and spreading properties: known results}\label{ss:known}
The definition of the so-called pulsating traveling wave was
introduced by Xin \cite{Xin} in the framework of flame
propagation. It is the natural extension, in the periodic
framework, of classical traveling waves. Due to the interest of
taking into account the role of the heterogeneity of the medium on
the propagation of solutions, a lot of attention was later drawn
on this subject. As far as monostable and ignition pulsating
fronts are concerned, we refer to the seminal works of Weinberger
\cite{Wein02}, Berestycki and Hamel \cite{Ber-Ham-02}. Let us also
mention \cite{Ber-Ham-Roq2}, \cite{Ham}, \cite{Ham-Roq},
\cite{Nad-09} for related results.
\medskip
For the sake of completeness,
let us first recall the definition of a pulsating traveling wave
for the equation~\eqref{monostable}, as stated in
\cite{Ber-Ham-02}.
\begin{defi}[Pulsating traveling wave]\label{def:puls}
A pulsating traveling wave solution, with speed $c >0$ in the
direction $n \in \mathbb{S}^{N-1}$, is an entire solution $u(t,x)$
--- $t\in\mathbb R$, $x\in \mathbb R ^N$--- of \eqref{monostable} satisfying
$$
\forall k \in \prod_{i=1}^N L_i \mathbb{Z} , \qquad u(t
,x)=u\left(t + \frac{k\cdot n}{c} ,x+k\right),
$$
for any $t \in \mathbb R$ and $x\in \mathbb R^N$, along with the asymptotics
$$
u(-\infty,\cdot)=0 < u (\cdot , \cdot) < u(+\infty,\cdot) = 1,
$$
where the convergences in $\pm \infty$ are understood to hold
locally uniformly in the space variable.
\end{defi}
One can easily check that, for any $c>0$ and $n \in
\mathbb{S}^{N-1}$, $u(t,x)$ is a pulsating traveling wave with
speed $c$ in the direction $n$ if and only if it can be written in
the form
$$
u(t,x) = U (x\cdot n -ct,x),
$$
where~$U(z,x)$ --- $z\in \mathbb R$, $x\in \mathbb R^N$--- satisfies
$$
\text{ for all } z\in \mathbb R,\, U(z,\cdot): \mathbb R^N \to \mathbb R \text{ is
periodic},
$$
\begin{equation*
U (-\infty, \cdot) = 1 < U (\cdot , \cdot) < U (+\infty,\cdot)
=0 \quad \text{ uniformly w.r.t. the space variable},
\end{equation*}
along with the following equation
\begin{equation}\label{eq-tw}
\begin{array}{l}
\mbox{div}_x (A \nabla_x U) + (n \cdot An)\, \partial_{zz} U + \mbox{div}_x ( An \, \partial_z U) + \partial_z (n\cdot A \nabla_x U) \vspace{3pt}\\
\qquad \qquad + \, q \cdot \nabla_x U + (q \cdot n) \, \partial_z
U + c\partial _z U+f(x,U)=0 , \quad \text{ on } \mathbb R\times \mathbb R ^N.
\end{array}
\end{equation}
We can now recall the results of \cite{Ber-Ham-02} (see also
\cite{Wein02} for the monostable case), on the existence of
pulsating traveling waves for the spatially periodic monostable
and ignition equations. Precisely, the following holds.
\begin{theo}[Monostable and ignition pulsating fronts, \cite{Ber-Ham-02},\cite{Wein02}]
\begin{itemize}
\item Assume that $f$ is of the spatially periodic monostable
type, i.e. $f$ satisfies \eqref{periodicity} and
Assumption~\ref{hyp:monostable}. Then for any $n \in
\mathbb{S}^{N-1}$, there exists $c^* (n) >0$ such that pulsating
traveling waves with speed $c$ in the direction $n$ exist if and
only if $c \geq c^* (n).$ \item Assume that $f$ is of the
spatially periodic ignition type, i.e. $f$ satisfies
\eqref{periodicity} and Assumption~\ref{hyp:ignition}. Then for
any $n \in \mathbb{S}^{N-1}$, there exists a unique (up to time
shift) pulsating traveling wave, whose speed we denote by $c^* (n)
>0$.\medskip
\end{itemize}
Furthermore, in both cases, any pulsating traveling wave is
increasing in time.
\end{theo}
The introduction of these pulsating traveling waves was motivated
by their expected role in describing the large time behavior of
solutions of \eqref{monostable} for a large class of initial data.
In this context, let us state the result of \cite{Wein02} for
planar-shaped initial data.
\begin{theo}[Spreading properties,
\cite{Wein02}]\label{Wein02_cauchy} Assume that $f$ is of the
spatially periodic monostable or ignition type, i.e. $f$ satisfies
\eqref{periodicity} and either of the two
Assumptions~\ref{hyp:monostable} and~\ref{hyp:ignition}. Let $u_0$
be a nonnegative and bounded initial datum such that $\| u _0
\|_\infty < 1$ and$$ \exists C>0, \qquad x \cdot n \geq C \
\Rightarrow \ u_0 (x) = 0,
$$
$$
\liminf_{x \cdot n \to -\infty} u_0 (x) >0 \ \ (monostable \ case)
, \qquad \liminf_{x \cdot n \to -\infty} u_0 (x) >\theta \ \
(ignition \ case)
$$
for some $n\in \mathbb S ^{N-1}$.
Then the solution $u$ of \eqref{monostable}, with initial datum
$u_0$, spreads with speed $c^* (n)$ in the $n$-direction in the
sense that
\begin{equation}\label{conclusion_wein2}
\forall c <c^*(n), \qquad \displaystyle \lim_{t \to +\infty} \sup_{x \cdot n \leq ct} |
1 - u (t,x) | =0,
\end{equation}
\begin{equation}\label{conclusion_wein1}
\forall c > c^*(n) , \qquad \lim_{t \to +\infty} \sup_{x \cdot n \geq c t }
u (t,x) =0.
\end{equation}
\end{theo}
\begin{rem}[Link between spreading speed and wave speed]\label{rem_bug}
In \cite{Wein02}, Weinberger was actually concerned with a more
general discrete formalism
where pulsating waves are not always known to exist. Therefore, the fact that the
spreading speed and the minimal wave speed are one and the same was only explicitly
stated in the monostable case.
However, under the ignition Assumption~\ref{hyp:ignition}
and benefiting from the results in~\cite{Ber-Ham-02}, it is clear by a simple comparison
argument that the solution associated with any such initial datum spreads at most with the wave speed $c^* (n)$, namely \eqref{conclusion_wein1} holds true. Furthermore, one may check, using for instance $U^* (x \cdot n -
(c^*(n)-\alpha)t,x) - \delta$ as a subsolution of
\eqref{monostable}, where $U^*$ is the pulsating wave with speed
$c^* (n)$ and $\alpha >0$, $\delta >0$ are small enough, that
\eqref{conclusion_wein2} also holds true, at least for some large
enough initial data. Thus, the spreading speed exhibited by
Weinberger must be $c^* (n)$, as one would expect.
We will use a very similar argument in
Section~\ref{s:spread_ignition}, which is why we omit the details.
Moreover, it is a simplification of a
classical argument, which originates from \cite{Fif-Mac} in the homogeneous framework, and usually aims at proving the stronger
property that the profile of such a solution $u(t,x)$ of the Cauchy problem converges to that of the
ignition pulsating wave. We refer for instance to the work of Zlato{\v{s}}~\cite{Zlatos}, which
dealt with a fairly general multidimensional heterogeneous (not necessarily periodic) framework,
and covers the above result under the additional assumption that $f(x,u)$ is bounded from below by a standard homogeneous ignition nonlinearity.
\end{rem}
Various features of pulsating fronts and many generalizations of
spreading properties have been studied recently. Nevertheless, as
far as we know, nothing is known on the dependence of these
results on the direction of propagation. Our results stand in this
new framework and are stated in the next subsection.
\subsection{Pulsating fronts and spreading properties: varying the direction}\label{ss:varying}
As recalled above, the periodic ignition equation admits a unique
pulsating traveling wave in any direction $n \in
\mathbb{S}^{N-1}$, while the periodic monostable
equation~\eqref{monostable} admits pulsating traveling waves in
any direction $n \in \mathbb{S}^{N-1}$, for any speed larger than
some critical $c^* (n) >0$. The
latter is a consequence of the former, as was proved in \cite{Ber-Ham-02} by approximating the monostable equation with an ignition type equation. With some
modifications of their argument, we will prove the following continuity property.
\begin{theo}[Continuity of minimal
speeds]\label{th:continuity_speeds} Assume that $f$ is of the
spatially periodic monostable or ignition type, i.e. $f$ satisfies
\eqref{periodicity} and either of the two
Assumptions~\ref{hyp:monostable} or~\ref{hyp:ignition}.
Then the mapping $n \in \mathbb{S}^{N-1} \mapsto c^* (n)$ is
continuous.
\end{theo}
In the Fisher-KPP case the continuity of the velocity map
$n\mapsto c^*(n)$, even if not explicitly stated, seems to follow
from the characterization of $c^*(n)$ (see \cite{Wein02},
\cite{Ber-Ham-02}). However, for other types of nonlinearities (and in particular, in the more general monostable case), such a property seems to be far from obvious.
For the sake of completeness, let us state the continuity of the
profile of the ignition wave, which will be proved simultaneously.
\begin{theo}[Continuity of ignition waves]\label{th:continuity_igwaves}
If $f$ satisfies \eqref{periodicity} and the ignition
Assumption~\ref{hyp:ignition}, then the mapping
$$n \in \mathbb{S}^{N-1} \mapsto U^* (z,x;n)$$
is continuous with respect to the uniform topology, where
$$u^* (t,x;n) = U^* (x \cdot n - c^* (n) t ,x;n)$$
is the unique pulsating traveling wave in the $n$ direction,
normalized by $\min _{x\in \mathbb R^N}U^* (0,x;n)= \frac{1+ \theta
}{2}$.
\end{theo}
In Section~\ref{s:ignition}, we deal with the ignition case,
proving both the continuity of the speed (Theorem
\ref{th:continuity_speeds}) and that of the profile (Theorem
\ref{th:continuity_igwaves}). To do so we take advantage of the
uniqueness
of the pulsating wave in each direction.
Then, in Section \ref{s:monostable-speed}, we
approach our original monostable equation by some ignition type
problems, and prove that the associated ignition speeds converge
to $c^* (n)$ not only pointwise (as in \cite{Ber-Ham-02}), but
even uniformly with respect to $n \in \mathbb{S}^{N-1}$. The
continuity of the minimal speed (Theorem
\ref{th:continuity_speeds}) then immediately follows.
Unfortunately, the lack of a rigorous uniqueness result of the
monostable pulsating wave with minimal speed (at least up to our
knowledge) prevents us from stating
continuity of its profile with respect to the speed of propagation. We refer to~\cite{Ham-Roq} for uniqueness
results in the Fisher-KPP case and discussion on the general monostable framework.
\medskip
We also stated above the well known fact that for any planar-like
initial data in some direction~$n$, the associated solution of
\eqref{monostable} spreads in the $n$ direction with speed $c^*
(n)$. Our main result consists in improving (compare
Theorem~\ref{th:unif_spreading} with Theorem~\ref{Wein02_cauchy})
this property by adding some uniformity with respect to $n \in
\mathbb{S}^{N-1}$, as follows.
\begin{theo}[Uniform spreading]\label{th:unif_spreading}
Assume that $f$ is of the spatially periodic monostable or
ignition type, i.e. $f$ satisfies \eqref{periodicity} and either
Assumption~\ref{hyp:monostable} or Assumption~\ref{hyp:ignition}.
Let a family of nonnegative initial data $(u_{0,n})_{n \in
\mathbb{S}^{N-1}}$ be such that
\begin{equation}\label{hyp1}
\exists C >0, \qquad \forall n \in \mathbb{S}^{N-1}, \qquad x
\cdot n \geq C \ \Rightarrow \ u_{0,n} (x) = 0,
\end{equation}
\begin{equation}\label{hyp2}
\left.
\begin{array}{l}
\exists \mu > \theta \ \ (ignition \ case) \vspace{3pt}\\
\exists \mu > 0 \ \ (monostable \ case)
\end{array}
\right\} , \qquad \exists K >0, \qquad \inf_{n \in
\mathbb{S}^{N-1}, \; x \cdot n \leq -K } u_{0,n} (x) \geq \mu,
\end{equation}
\begin{equation}\label{hyp0}
\sup_{n \in \mathbb{S}^{N-1}} \sup_{x \in \mathbb R^N}
u_{0,n} (x)
< 1.
\end{equation}
We denote by $(u_n)_{n \in \mathbb{S}^{N-1}}$ the associated
family of solutions of \eqref{monostable}.
Then, for any $\alpha >0$ and $\delta >0$, there exists $\tau
>0$ such that for all $t \geq \tau$,
\begin{equation}\label{conclusion1}
\sup_{n \in \mathbb{S}^{N-1}} \ \sup_{x \cdot n \leq (c^*(n) -
\alpha)t} |1 -u_n (t,x)| \leq \delta,
\end{equation}
\begin{equation}\label{conclusion2}
\sup_{n \in \mathbb{S}^{N-1}} \ \sup_{x \cdot n \geq (c^*(n) +
\alpha)t} u_n (t,x) \leq \delta.
\end{equation}
\end{theo}
The difficult part is again to deal with the monostable case. The
proof of the lower spreading property \eqref{conclusion1} will
again rely on an ignition approximation of the monostable
equation, whose traveling waves will serve as nontrivial
subsolutions of \eqref{monostable}. This is performed in
Section~\ref{s:lower-spreading}. Then,
Section~\ref{s:upper-spreading} is devoted to the proof of the
upper spreading property: we prove \eqref{conclusion2} in
subsection~\ref{ss:with} and, for sake of completeness, relax
assumption~\eqref{hyp0} in subsection~\ref{ss:without}.
Last, in Section \ref{s:spread_ignition}, we prove Theorem
\ref{th:unif_spreading} in the ignition case.
\section{Continuity of ignition waves}\label{s:ignition}
Let us here consider a periodic nonlinearity $f$ of the ignition
type, namely satisfying Assumption~\ref{hyp:ignition}. As
announced, we will prove simultaneously the continuity of both
mappings $n \mapsto c^* (n)$ and $n \mapsto U^* (z,x;n)$, where
we recall that $c^* (n)$ and $U^* (x \cdot n - c^* (n)t,x ;n)$
denote respectively the unique admissible speed and the unique
pulsating wave in the direction $n$, normalized by
\begin{equation}\label{shift-min}
\min _{x\in \mathbb R ^N}U^* (0,x;n) = \frac{1+ \theta}{2}.
\end{equation}
\begin{proof} [Proofs of Theorem \ref{th:continuity_speeds} (ignition case) and Theorem \ref{th:continuity_igwaves}]
We first claim (we postpone the proof to the end of this section)
that
\begin{equation}\label{claim-kappa}
\kappa := \inf_{n \in \mathbb{S}^{N-1}} c^* (n) >0.
\end{equation}
Let us now prove that that $c^*(n)$ is also bounded from above,
using
$$(t,x)\mapsto v(t,x):= \min \{ 1 , \theta + C e^{-\lambda (x \cdot n - 2 a_1 \lambda t)} \}$$ as a
supersolution of \eqref{monostable}. Here $C$ and $\lambda$ are
positive constants to be chosen later, and $a_1$ comes from
hypothesis \eqref{ellipticity}. Indeed, when $v < 1$, it satisfies
\begin{eqnarray}
& & \partial_t v - \mbox{div} \, (A(x) \nabla v ) - q(x) \cdot \nabla v -f(x,v) \vspace{3pt}\nonumber\\
& = & \left[ 2 a_1 \lambda^2 - (n \cdot A n) \lambda^2 + \lambda \, \mbox{div} \, (An) + \lambda q \cdot n \right] \times C e^{- \lambda (x \cdot n - 2 a_1 \lambda t)} - f(x,v) \vspace{3pt}\nonumber\\
& \geq & \left[ a_1 \lambda^2 - \lambda \, |\mbox{div} \, (An)|
- \lambda |q \cdot n| - M \right] \times C e^{- \lambda (x \cdot
n - 2 a_1 \lambda t)} > 0,\label{gggg}
\end{eqnarray}
where
\begin{equation}\label{M}
M:= \sup_{x \in \mathbb R^N,u \in [0,1]} \frac{ f(x,u)}{|u - \theta|} < +
\infty
\end{equation}
comes from the Lipschitz continuity of $f$, and the last
inequality holds provided that $\lambda$ is large enough,
independently of $n \in \mathbb{S}^{N-1}$. As $1$ is a solution of
\eqref{monostable}, it is then clear that $v$ is a generalized
supersolution of \eqref{monostable}. Then, choosing $C>0$ so that
$v(t=0,x)$ lies above the traveling wave $u^* (t=0,x;n)=U^*(x\cdot
n,x;n)$ at time 0, we can apply the comparison principle
and obtain that $c^* (n) \leq 2 a_1 \lambda$. Putting this fact together with \eqref{claim-kappa}, we conclude that
\begin{equation}\label{speed_bounds0}
0< \kappa := \inf_{n \in \mathbb{S}^{N-1}} c^* (n) \leq \sup_{n
\in \mathbb{S}^{N-1}} c^* (n) =: K < +\infty.
\end{equation}
\medskip
We now let some sequence of directions $n_k \to n \in
\mathbb{S}^{N-1}$. As we have just shown, the sequence $c^* (n_k)$
is bounded and, up to extraction of a subsequence, $c^* (n_k) \to
c > 0$. We also choose the shifts $z_k$ so that, for all $k$,
$\max_{x \in \mathbb R^N} U^* (z_k, x;n_k) = \theta $. In particular,
recalling that $U^*$ is monotonically decreasing with respect to
its first variable, we have for all $k$ that
$$\forall z\geq z_k,
\forall x\in \mathbb R ^N , \quad 0 <U^* (z,x;n_k) \leq \theta.
$$
Then
$$u_{k} (t,x) := U^* (z_k + x \cdot n_k ,x + c^* (n_k) t n_k ;n_k)$$
satisfies
\begin{equation}\label{ignition_continuity1}
\partial_t u_{k}= \mbox{div} \, (A(x) \nabla u_{k}) + q(x) \cdot \nabla u_{k} + c^* (n_k) \nabla u_{k} \cdot
n_k,
\end{equation}
for all $t \in \mathbb R$ and all $x$ in the half-space $x \cdot n_k \geq
0$ (recall that $U^*$ solves \eqref{eq-tw} and that, in the
ignition case, $f (x,u)=0$ if $0 \leq u \leq \theta$).
\medskip
Let us now find a supersolution of \eqref{ignition_continuity1} of
the exponential type, namely
\begin{equation}\label{def-sur}\overline{u}_k (t,x):= \phi_k (t,x) \times e^{-\lambda_0 \, x \cdot n_k},
\end{equation}
where $\phi_k$ will be a well-chosen positive and bounded
function.
For any $n \in \mathbb{S}^{N-1}$, one may define (see
Proposition~5.7 in \cite{Ber-Ham-02}) the principal eigenvalue
problem
\begin{equation*}
\left\{
\begin{array}{l}
-L_{n , \lambda} \phi_{n,\lambda} = \mu (n,\lambda) \phi_{n,\lambda} \ \mbox{ in } \mathbb R^N , \vspace{3pt}\\
\phi_{n,\lambda}>0 \mbox{ is periodic},
\end{array}
\right.
\end{equation*}
where
$$
L_{n,\lambda} \phi := \mbox{div} \, (A \nabla \phi) + \lambda^2 (n
\cdot A n) \phi - \lambda (\mbox{div} (An \phi) + n\cdot A\nabla
\phi) + q \cdot \nabla \phi - \lambda (q \cdot n + \kappa ) \phi,
$$
with $\kappa >0$ given by \eqref{claim-kappa}. In the sequel, the
eigenfunction $\phi_{n,\lambda}$ is normalized so that
$$
\min_{x \in \mathcal{C}} \phi_{n,\lambda} (x) = \theta .
$$
As stated in Proposition~5.7 of \cite{Ber-Ham-02}, the function
$\lambda \mapsto \mu (n,\lambda)$ is concave and satisfies, for any~$n$, that $\mu (n, 0)= 0$ (any positive constant is clearly a principal eigenfunction of $-L_{n,0}$), and $\partial_\lambda \mu (n,0) = \kappa >0$.
It follows that one can find some small $\lambda_0 >0$ such that,
for any $n \in \mathbb{S}^{N-1}$,
$$\mu (n,\lambda_0) >0.$$ Indeed, proceed by contradiction and assume that for
any $j \in \mathbb{N}^*$, there exists $n_j$ such that $\mu (n_j,
1/j) \leq 0$. Then, by $\mu(n_j,0)=0$ and by concavity, one has
that $\mu (n_j,\lambda) \leq 0$ for all $\lambda > \frac{1}{j}$.
By uniqueness of the principal normalized eigenfunction, it is
straightforward to check that $\mu (n,\lambda)$ depends continuously on both $n$ and $\lambda$, as well as
$\phi_{n,\lambda}$ with respect to the
uniform topology. Thus, one can pass to the limit and conclude
that $\mu (n_\infty, \lambda) \leq 0$ for some $n_\infty = \lim
n_j$ (up to extraction of a subsequence) and all $\lambda \geq 0$.
This contradicts the fact that $\partial_\lambda \mu
(n_\infty,0)=\kappa >0$.
Notice that, by continuity of the eigenfunction with respect to
$n$ and $\lambda$ in the uniform topology, it is clear that for
any bounded set $\Lambda$,
\begin{equation}\label{eigenfunction_bound}
\max_{n \in \mathbb{S}^{N-1}} \max_{\lambda \in \Lambda} \ \max_{x
\in\mathcal{C}} \phi_{n,\lambda} (x)< +\infty.
\end{equation}
Choosing $\lambda_0$ as above and
$$
\phi_k (t,x) := \phi_{n_k,\lambda_0} (x + c^* (n_k) t n_k),
$$
in \eqref{def-sur}, one gets that
\begin{eqnarray*}
&& \partial_t \overline{u}_{k}- \mbox{div} \, (A(x) \nabla
\overline{u}_{k}) - q(x) \cdot \nabla \overline{u}_{k} - c^* (n_k)
\nabla \overline{u}_{k} \cdot
n_k \vspace{3pt} \\
& = & \left[ c^* (n_k) n_k \cdot \nabla \phi_{n_k ,\lambda_0}
-L_{n_k,\lambda_0} \phi_k + \lambda_0 ( - \kappa + c^* (n_k))
\phi_k - c^* (n_k) n_k \cdot \nabla \phi_{n_k ,\lambda_0}\right]
\times e^{-\lambda_0 (x \cdot n_k)} \vspace{3pt} \\
& = & \left[ \mu (n_k ,\lambda_0) + \lambda _0 (c^* (n_k) -
\kappa) \right] \overline{u}_k >0.
\end{eqnarray*}
In other words, as announced, $\overline{u}_k$ is a supersolution
of \eqref{ignition_continuity1}.
\medskip
Let us now prove that
\begin{equation}\label{ignition_continuity2}
\forall t\in \mathbb R,\; \forall x \cdot n_k \geq 0 , \quad u_{k} (t,
x) \leq \overline{u}_{k} (t,x) .
\end{equation}
Proceed by contradiction and define a sequence of points
$(t_j,x_j)_{j \in \mathbb{N}}$ such that
$$u_{k} (t_j,x_j) - \overline{u}_{k} (t_j,x_j)\to \sup_{t \in \mathbb R, x\cdot n_k \geq 0} ( u_{k} (t,x) - \overline{u}_{k}(t,x) ) >0.$$
Now write $x_j = (x_j \cdot n_k)n_k + y_j$ for any $j \geq 0$.
Note that, since $u_{k} (t,x) $ and $\overline{u}_k (t,x)$ both
tend to $0$ as $x \cdot n_k \to +\infty$ uniformly with respect to
$t$, then $(x_j \cdot n_k)_{j \in \mathbb{N}}$ must be bounded.
Thus, up to extraction of a subsequence, we can assume that $x_j
\cdot n_k \to a \geq 0$ as $j \to \infty$. Moreover, since $y_j$
is orthogonal to $n_k$, since $\phi_{n_k,\lambda _0}$ is periodic
and since $U^*$ is periodic with respect to its second variable,
we can assume without loss of generality that $y_j + c^* (n_k)t_j
n_k \in \mathcal{C}$ the cell of periodicity. As $y_j$ is
orthogonal to $n_k$ for all $j \in \mathbb{N}$, we can extract a
subsequence such that both $y_j \to y_\infty \in \mathbb R^N$ and $t_j
\to t_\infty \in \mathbb R$.
Finally, $u_k - \overline{u}_k$ reaches its positive maximum, over
$t \in \mathbb R$ and $x \cdot n_k \geq 0$, at $(t=t_\infty, x=an_k +
y_\infty)$. Moreover, as
$$\forall x \cdot n_k = 0 , \quad u_k (0,x) \leq \theta \leq \overline{u}_k (0,x) ,$$
the maximum is reached at an interior point, which contradicts the
parabolic maximum principle. Thus, \eqref{ignition_continuity2} is
proved.
\medskip
Now, by standard parabolic estimates and up to extraction of a
subsequence, we can assume that, as $k\to \infty$, the sequence
$u^* \left(t- \frac{z_k}{c^* (n_k)} ,x;n_k \right)=U^*(x\cdot
n_k-c^*(n_k)t+z_k,x;n_k)$ converges locally uniformly, along with
its derivatives, to a solution $u_\infty (t,x)$ of
\eqref{monostable}. Moreover, $u_\infty$ satisfies
$$\forall l \in \Pi_{i=1}^N L_i \mathbb{Z} , \quad u_\infty (t,x) = u_\infty \left( t+ \frac{l \cdot n}{c},x+l\right).$$
In a similar way than the discussion after
Definition~\ref{def:puls} of pulsating waves, this means that
$u_\infty (t,x) = U_\infty (x\cdot n- ct,x)$ where $U_\infty(z,x)$
is periodic with respect to its second variable and satisfies
\begin{equation*}
\begin{array}{l}\mbox{div}_x (A \nabla_x U) + (n \cdot An)\, \partial_{zz} U + \mbox{div}_x ( An \partial_z U) + \partial_z (n\cdot A \nabla_x U) \vspace{3pt}\\
\qquad \qquad + \, q \cdot \nabla_x U + (q \cdot n) \, \partial_z
U + c\partial _z U+f(x,U)=0\quad \text{ on } \mathbb R\times \mathbb R ^N.
\end{array}
\end{equation*}
It is then straightforward to retrieve that the sequence $U^*
(z+z_k,x;n_k)$ also converges, along with its derivatives, to this
function $U_\infty(z,x)$. In particular, $U_\infty$ is
nonincreasing with respect to its first variable, and satisfies
the inequalities
$$0 \leq U_\infty (z,x) \leq 1 .$$
Furthermore, noticing that $u^* \left(t- \frac{z_k}{c^* (n_k)}
,x;n_k \right)= u_{k} (t,x- c^* (n_k) tn_k )$,
it follows from passing to the limit in \eqref{ignition_continuity2}, and thanks to \eqref{eigenfunction_bound}, that
$$u_\infty (t,x) \leq A e^{-\lambda_0 (x\cdot n - c t)},$$
for some $A >0$ and all $x \cdot n \geq c t$.
Thus, $U_\infty (x \cdot n - ct,x) \leq A e^{-\lambda_0 ( x \cdot
n - c
t)}$, for all $t\in \mathbb R$ and $x \cdot n \geq ct$. This means that $U_\infty (z,x)$ converges exponentially to 0 as $z \to +\infty$, uniformly with respect to its second variable:
\begin{equation}\label{ignition_continuity3}
\forall z\geq 0,\; \forall x\in \mathbb R ^N, \quad U_\infty (z,x) \leq A
e^{-\lambda_0 z}.
\end{equation}
By monotonicity with respect to its first variable,
$U_\infty(z,x)$ converges as $z \to -\infty$ to some periodic
function $p (x)$. Or, equivalently, $u_\infty (t,x)$ converges as
$t \to +\infty$ to the same function $p (x)$. By standard
parabolic estimates, we get that $p (x)$ is a periodic and
stationary solution of \eqref{monostable}. Let us show that $p
\equiv 1$. From our choice of the shifts $z_k$ and up to
extraction of another subsequence, there exists some $x_\infty$
such that $U_\infty (0,x_\infty) =\theta$, hence $\max p \geq
\theta $. Assume first that $\max p = \theta$. Then $u_\infty
(t,x) \leq \theta$ for all $t \in \mathbb R$ and
$x \in\mathbb R^N$ and, by the strong maximum principle, $u_\infty \equiv \theta$.
This contradicts the inequality~\eqref{ignition_continuity3} above.
Therefore, $\max p > \theta$. Using again the strong maximum principle and
the fact that $f (x,u)=0$ for all $u \leq \theta$ and $x \in \mathbb R^N$, we reach
another contradiction if $\min p \leq \theta$. Therefore $\min p > \theta$ and, thanks to part $(ii)$ of our ignition Assumption~\ref{hyp:ignition}, $p \equiv 1 $
the unique periodic stationary solution of
\eqref{monostable} above $\theta$.
From the above analysis, we conclude that $U_\infty (\cdot ,
\cdot) = U^* (\cdot + Z, \cdot ;n)$ the unique pulsating traveling
wave in the $n$ direction with speed $c=c^* (n)$, where $Z$ is the
unique shift such that $ \max_{x \in \mathbb R^N} U^* (Z, x;n) = \theta$.
This in fact proves, by uniqueness of the limit, that the whole
sequence $c^*(n_k)$ converges to $c^*(n)$, and that the whole
sequence $U^* (\cdot +z_k ,\cdot;n_k)$ converges locally uniformly
to $U^*(\cdot + Z,\cdot;n)$. This in particular shows the
continuity of the map $n\mapsto c^*(n)$, that is
Theorem~\ref{th:continuity_speeds} in the ignition case.
\medskip
Let us now conclude the proof of
Theorem~\ref{th:continuity_igwaves}. Let us first prove that the
sequence of shifts $z_k$ is bounded. The normalization
\eqref{shift-min} implies that $U^*(0,y_k;n_k)=\frac{1+\theta}2$,
for some $y_k \in \mathcal C$ that (up to some subsequence)
converges to some $y\in \mathcal C$. Since $U ^*(z_k,y_k;n_k)\to U^*(Z,y;n)\leq \theta$ and $U^*
(0,y;n) = \frac{1+\theta}{2}$, the monotonicity of traveling waves
enforces $z_k \geq 0$ for $k$ large enough. Now proceed by
contradiction and assume that (up to some subsequence) $z_k \to
+\infty$. Then, for all $-z_k \leq z \leq 0$,
$$
U^* (z + z_k,y_k;n_k) \leq U^*(0,y_k;n_k)= \frac{1+
\theta}{2}.
$$
Passing to the limit as $k \to +\infty$, we get that
$$
U_\infty (z,y) \leq \frac{1 +\theta }{2}, $$ for all $z \leq 0$.
This contradicts the fact that $U_\infty$ is a pulsating traveling
wave and converges to $1 $ as $z \to -\infty$.
From the boundedness of the sequence $z_k$, we can now rewrite the
convergence as follows: the sequence $U^* (\cdot ,\cdot;n_k)$
converges locally uniformly to $U^* (\cdot,\cdot;n)$. It now
remains to prove that this
convergence is in fact uniform with respect to both variables. Note first that uniformity with respect to the second
variable immediately follows from the periodicity. Furthermore,
for a given $\delta >0$, let $K >0$ be such that, for any $x \in
\mathbb R^N$,
\begin{equation}\label{loin}
0 \leq U^* (z, x; n ) \leq \frac{\delta}{2} \ \mbox{ and } \ 1 -
\frac \delta 2 \leq U^* ( -z, x; n ) \leq 1, \textrm{ for all } z
\geq K.
\end{equation}
From the locally uniform convergence with respect to the first
variable, we have, for any $k$ large enough,
$$\| U^* (\cdot,\cdot;n_k) - U^* ( \cdot, \cdot; n ) \|_{L^\infty([-K,K]\times \mathbb R)} \leq \frac \delta 2 .$$
In particular, $U^* ( K, x; n _k) \leq \delta$ and $1 -\delta \leq
\ U^* ( -K, x; n_k )$, so that, by monotonicity with respect to
the first variable, for any $x \in \mathbb R^N$ and $k$ large enough,
\begin{equation}\label{loin2}
0 \leq U^* (z, x; n_k ) \leq \delta\ \mbox{ and } \ 1- \delta
\leq U^* ( -z, x; n_k ) \leq 1 , \textrm{ for all } z \geq K.
\end{equation}
Combining \eqref{loin} and \eqref{loin2}, we get
$$
\| U^* (\cdot,\cdot;n_k) - U^* ( \cdot, \cdot; n )
\|_{L^\infty((-\infty,-K)\cup (K, \infty)\times \mathbb R)} \leq \delta,
$$
for any $k$ large enough. As a result the convergence of $U^*
(\cdot,\cdot;n_k)$ to $U^* (\cdot,\cdot;n)$ is uniform in $\mathbb R
\times \mathbb R^N$. This ends the proof of the continuity of ignition
waves, that is Theorem~\ref{th:continuity_igwaves}.
\end{proof}
\begin{proof}[Proof of claim \eqref{claim-kappa}]
Proceed by contradiction and assume that there exists a sequence
$n_k \in \mathbb{S}^{N-1}$ such that $c^* (n_k) \to 0$.
Now for any $k$,
recall that the pulsating wave is normalized by
\begin{equation}\label{shift-min2}
\min _{x\in \mathbb R^N} U^* (0,x;n_k) = \frac{1+\theta}{2}.
\end{equation} Up to extraction of a subsequence, we can assume as
above that $n_k \to n$ and
$$u^* \left(t,x;n_k \right) \to u_\infty (t,x),$$
where the convergence is understood to hold locally uniformly, and
$u_\infty (t,x)$ is a solution of~\eqref{ignition_continuity1}. By
the strong maximum principle, it is clear that $0 < u_\infty < 1$.
We also know, by the monotonicity of $U^* (\cdot ,\cdot;n_k)$ with
respect to its first variable, by \eqref{shift-min2} and by
passing to the limit, that
$$
u_\infty (t,x) \geq \frac{1+\theta}{2}, \quad \forall x\cdot n
\leq 0.
$$
Applying Weinberger's result (see
Theorem \ref{Wein02_cauchy} as well as Remark~\ref{rem_bug}), we
get that the solution spreads at least at speed $c^* (n)$. In
particular, as $t\to+\infty$, $u_\infty(t,x)$ converges locally
uniformly to~1.
On the other hand, we fix $x \in \mathbb R^N$ and $s \geq 0$, then we let
some vector $l \in \Pi_{i=1}^N L_i \mathbb{Z}$ be such that $l
\cdot n > 0$. In particular, for any large $k$, one also has that
$l \cdot n_k \geq \frac{l \cdot n}{2} >0.$ Then, for all large
$k$, using the fact that $c^* (n_k) \to 0$ and the monotonicity of
$u^* (\cdot , \cdot ; n_k)$ with respect to its first variable, we
have that
$$u^* \left( s ,x ;n_k \right) \leq u^* \left( \frac{l \cdot n_k}{c^* (n_k)} , x ;n_k \right) = u^* \left(0 ,x - l ;n_k \right).$$
By passing to the limit as $k \to +\infty$, we obtain that
$$
u_\infty (s,x) \leq u_\infty (0,x-l)<1,
$$
for all $x \in \mathbb R^N$ and $s \geq 0$. This contradicts the locally
uniform convergence of $u_\infty(t,x)$ to $1$ as $t \to + \infty$.
The claim is proved.\end{proof}
\section{Continuity of the monostable minimal speed}\label{s:monostable-speed}
Let us here consider a periodic nonlinearity $f$ of the monostable
type, namely satisfying Assumption~\ref{hyp:monostable}. We will
prove the continuity of the mapping $n \mapsto c^* (n)$, that is
Theorem~\ref{th:continuity_speeds}. To do so, we introduce a
family $f_\varepsilon (x,u)$, for small $\varepsilon
>0$, of ignition nonlinearities which serve as approximations from
below of the monostable nonlinearity $f(x,u)$. Our aim is to
prove that, by passing to the limit as $\varepsilon \to 0$, we
indeed retrieve the dynamics of the monostable equation. This will
be enough to prove Theorem~\ref{th:continuity_speeds}.
\medskip
The family $(f_\varepsilon )_\varepsilon$, for small enough
$\varepsilon >0$, is chosen as follows:
\begin{equation*}
\forall x \in \mathbb R^N, \ \left\{\begin{array}{rl}
\forall u \in [-\varepsilon, 0], & f_\varepsilon (x,u) = 0 \vspace{3pt}\\
\forall u \in \left[0 , 1 - \varepsilon \right], & f_\varepsilon (x,u) = f (x,u) \vspace{3pt}\\
\forall u \in \left[1 - \varepsilon, 1 - \frac{\varepsilon}{2} \right], & f_\varepsilon (x,u) = f \left(x,1 - \varepsilon + 2 (u-(1-\varepsilon))
\right).
\end{array}\right.
\end{equation*}
Notice that $\Vert f_{\varepsilon} -f\Vert _{L^\infty(-{\varepsilon},1)}\to 0$ as ${\varepsilon}
\to 0$, and that, thanks to Assumption \ref{hyp:monostable}
$(iii)$,
$f_{\varepsilon}$ lies below $f$ and $0<{\varepsilon}<{\varepsilon}'$ implies $f_{\varepsilon} \geq
f _{{\varepsilon} '}$. Also, the equation
\begin{equation}\label{ignition_approx}
\partial _t u= \mbox{div} (A (x) \nabla u) + q (x) \cdot \nabla u + f_\varepsilon ( x,u),
\end{equation}
where $u$ is to take values between $-\varepsilon$ and $1-\frac
\varepsilon 2$, is of the ignition type in the sense of
Assumption~\ref{hyp:ignition} (where 0, $\theta$, 1 are replaced
by $-\varepsilon$, 0 and $1 - \frac{\varepsilon}{2}$
respectively). In particular, for each $n \in \mathbb{S}^{N-1}$,
there exists a unique ignition pulsating traveling wave
$$u_\varepsilon^* (t,x;n) = U_\varepsilon^* (x \cdot n -
c^*_\varepsilon (n) t,x;n)$$ of \eqref{ignition_approx} in the $n$
direction with speed $c^*_\varepsilon (n) >0$, normalized by
$$
\min _{x\in \mathbb R ^N}U_{\varepsilon} ^*(0,x;n)=\frac 12.
$$
Furthermore, we
have already proved in the previous section that the mappings $n
\mapsto c^*_\varepsilon (n)$ and
$n\mapsto U_\varepsilon^* (\cdot, \cdot;n)$ are continuous (with respect to the uniform topology).
\begin{theo}[Convergence of speeds] \label{th:uc_speeds} Assume that $f$ is of the spatially periodic monostable type, i.e.
$f$ satisfies \eqref{periodicity} and
Assumption~\ref{hyp:monostable}. Let $f_\varepsilon(x,u)$ be
defined as above.
Then, as $\varepsilon \to 0$,
$c^*_\varepsilon (n) \nearrow c^* (n)$ uniformly with respect to $n \in
\mathbb{S}^{N-1}$.
\end{theo}
As mentioned before, pointwise convergence was shown
in~\cite{Ber-Ham-02}, where the goal was to prove existence of
monostable traveling waves for
the range of speeds $[c^* (n) , +\infty)$. Here we prove that the convergence is actually uniform,
which together with the continuity of speeds in the ignition case, immediately insures the continuity of $n \mapsto c^* (n)$, that is Theorem~\ref{th:continuity_speeds} in the monostable case.
\begin{proof} First note that, for any fixed $n\in \mathbb{S}^{N-1}$ and $\varepsilon >0$, $c^*_{\varepsilon} (n) \leq c^* (n)$. Indeed, recalling that $U_\varepsilon^* (z,x;n)$
connects $1 - \frac{\varepsilon}{2}$ to $-\varepsilon$, one can
find some shift $Z\in \mathbb R$ such that $U^*_\varepsilon (z+Z,x;n)
\leq U^* (z,x;n)$, where $U^*$ denotes a monostable pulsating
traveling wave --- connecting 1 to 0--- with the minimal speed
$c^* (n)$. By a comparison argument, it follows that
$c^*_\varepsilon (n) \leq c^* (n)$. It is also very similar to
check that, for any $n \in \mathbb{S}^{N-1}$,
$0<\varepsilon<\varepsilon '$ implies $c_\varepsilon ^*(n)\geq
c_{\varepsilon '}^*(n)$.
Let us now consider some sequences $\varepsilon_k \to 0$ and $n_k
\to n$. Consider the estimate \eqref{speed_bounds0} where $\kappa$ and $K$ should {\it a priori} depend on $\varepsilon$. First, it is clear from the above that $\kappa (\varepsilon) := \inf_n c^*_\varepsilon (n)$ is
nonincreasing with respect to $\varepsilon$. Also, since
$$
\sup _{0<{\varepsilon}\leq {\varepsilon} _0} M_{\varepsilon}:=\sup _{0<{\varepsilon}\leq {\varepsilon} _0}\quad \sup_{x \in \mathbb R^N,u \in [-{\varepsilon},1-\frac {\varepsilon} 2]} \frac{ f_{\varepsilon} (x,u)}{|u|} < +
\infty
$$
(compare with \eqref{M}), arguing as we did to derive
\eqref{speed_bounds0}, we see that $K (\varepsilon) := \sup_n
c^*_\varepsilon (n)$ is uniformly
bounded from above. As a result, we have
\begin{equation}\label{speed_bounds0bis}
0< \kappa := \inf_{0<{\varepsilon}\leq {\varepsilon} _0} \inf_{n \in \mathbb{S}^{N-1}} c^* (n) \leq\sup _{0<{\varepsilon} \leq {\varepsilon} _0} \sup_{n
\in \mathbb{S}^{N-1}} c^* (n) =: K < +\infty.
\end{equation}
Hence, we can assume,
up to extraction of a subsequence, that $c^*_{\varepsilon_k} (n_k)\to c_\infty >0 $ as $k \to \infty$.
In order to prove Theorem~\ref{th:uc_speeds}, we have to prove that $c_\infty = c^* (n)$.
\medskip
We begin by showing that $ U^*_{\varepsilon_k} (z,x;n_k)$
converges as $k \to \infty$ to a monostable pulsating traveling
wave of \eqref{monostable},
up to extraction of a subsequence. Indeed, proceeding as before, one can use standard parabolic estimates to
extract a converging subsequence of pulsating ignition traveling
waves, such that
$$
U^*_{\varepsilon_k} (z,x;n_k) \to U_\infty (z,x),
$$
as $k \to +\infty$ locally uniformly with respect to $(z,x) \in \mathbb R
\times \mathbb R^N$. Furthermore, $0 \leq U_\infty (z,x ) \leq 1$ solves
\eqref{eq-tw} with $c = c_\infty$, is nonincreasing with respect
to $z$, periodic with respect to $x$, and satisfies $\min _{x\in
\mathbb R ^N}U_\infty (0, x) = \frac{1}{2}$. In particular, $U_\infty$
converges as $z \to \pm \infty$ to two periodic stationary
solutions of \eqref{monostable}, which under the monostable Assumption~\ref{hyp:monostable} can only be 0 and 1. We can conclude that $U_\infty$
is a monostable pulsating traveling wave with speed $c_\infty$, hence $c_\infty \geq c^* (n)$.
We now prove that $c_\infty = c^* (n)$. Notice that
$f_\varepsilon$ lies below $f$ but, since the direction varies, we
cannot use a simple comparison argument to conclude that $c_\infty
\leq c^* (n)$. Instead, we will use a sliding method as
in~\cite{Ber-Ham-02}. To do so, we shall need the following lemma.
\begin{lem}[Some uniform estimates]
There exists $C>0$ such that, for any small $\varepsilon >0$ and
$n \in \mathbb{S}^{N-1}$, the ignition pulsating traveling wave
$U_\varepsilon^* (z,x;n)$ satisfies
$$| \partial_{zz} U^*_\varepsilon (\cdot, \cdot;n) | \leq - C \partial_{z} U^*_\varepsilon (\cdot ,\cdot ,n),
\quad| \nabla_x \partial_z U^*_\varepsilon(\cdot ,\cdot ;n) | \leq
- C
\partial_z U^*_\varepsilon (\cdot,\cdot,n).$$
\end{lem}
\begin{proof}
Let us define $u^*_\varepsilon (t,x):= U^*_\varepsilon (x \cdot n
- c^*_\varepsilon (n) t,x;n)$. Then $v(t,x): =
\partial_t u^* _\varepsilon(t,x)>0$ satisfies
$$
\partial_t v = \mbox{div} (A (x) \nabla v) + q (x) \cdot \nabla v + v\, \partial_u f_\varepsilon
(x,u^*_\varepsilon), \quad \text{ a.e. in $\mathbb R \times \mathbb R ^N$.}
$$
From our definition of the ignition approximation
$f_\varepsilon(x,u)$, we see that $\Vert \partial_u f_\varepsilon
\Vert _{L^\infty( \mathbb R^N \times (-{\varepsilon},1-\frac {\varepsilon} 2))}$ is uniformly
bounded, independently on small $\varepsilon >0$ and $n\in \mathbb
S ^{N-1}$. Therefore, from the interior parabolic $L^p$-estimates
(see \cite[Theorem 48.1]{Qui-Sou} for instance) and Sobolev
embedding theorem, one gets
\begin{equation}\label{schauder}
\forall (t_0,x_0) \in \mathbb R \times \mathbb R^N \, , \ | \partial_t v
(t_0,x_0) | + | \nabla_x v (t_0,x_0) | \leq C_1 \max_{t_0 -1 \leq
t \leq t_0 , |x - x_0| \leq 1 } v (t,x),
\end{equation}
for some $C_1 >0$ which is independent on $t_0$, $x_0$, small
$\varepsilon>0$ and $n\in \mathbb S ^{N-1}$.
Furthermore, for any $n \in \mathbb{S}^{N-1}$, choose $N$ integers
$k_i (n) \in \{ -1,0,1 \}$ such that
$$k (n)L \cdot n = \max_{k_1 , ..., k_N \in \{-1,0,1\} } (k_1 L_1 , ... k_N L_N) \cdot n ,$$
where $k (n)L :=(k_1 (n) L_1 , ... k_N (n) L_N)$. Then
$$
0 < \inf_{n \in \mathbb{S}^{N-1}} k(n)L \cdot n \leq \sup_{n \in
\mathbb{S}^{N-1}} k(n)L \cdot n < +\infty,
$$ and hence, thanks to
\eqref{speed_bounds0bis},
$$
0 < \inf_{0<\varepsilon \leq \varepsilon_0, n \in \mathbb{S}^{N-1}} \frac{k(n) L \cdot n}{c^*_\varepsilon (n)} \leq
\sup_{0<\varepsilon \leq \varepsilon_0 , n \in \mathbb{S}^{N-1}} \frac{k(n) L \cdot n}{c^*_\varepsilon (n)} < + \infty.
$$
By the parabolic Harnack inequality for strong solutions (see \cite[Chapter VII]{Lie}
for instance), we get
\begin{equation}\label{harnack}\forall (t_0,x_0) \in \mathbb R \times \mathbb R^N \, , \
\max_{t_0 -1 \leq t \leq t_0 , |x - x_0| \leq 1 } v (t,x) \leq C_2
v \left( t_0 + \frac{ k(n) L \cdot n}{c^*_\varepsilon (n)} , x_0 +
k(n) L \right), \end{equation} for some $C_2
>0$ which is also independent on $t_0$, $x_0$, small
$\varepsilon>0$ and $n\in \mathbb S ^{N-1}$.
Combining \eqref{schauder}, \eqref{harnack} and the space-time
periodicity of the traveling wave, we get
$$
\forall (t_0,x_0) \in \mathbb R \times \mathbb R^N \, , \ |\partial_t v (t_0,x_0
) | + | \nabla_x v (t_0,x_0) | \leq C_3v(t_0,x_0),
$$
with $C_3=C_1C_2$. Now recall that $U^*_\varepsilon (z,x;n) =
u^*_\varepsilon \left( \frac{x\cdot n-z }{c^*_\varepsilon (n)} , x
\right).$ Thus
$$| \partial_{zz} U^*_\varepsilon | = \frac{1}{c^*_\varepsilon (n)^2} |\partial_t v | \leq \frac{C_3}{c^*_\varepsilon (n)^2} v =-\frac{C_3}{c_{\varepsilon}^*(n)} \partial_{z} U^*_\varepsilon,$$
$$| \nabla_x \partial_z U^*_\varepsilon | \leq \left| \frac{-1}{c^*_\varepsilon (n)^2} \, \partial_t v \, n - \frac{1}{c^*_\varepsilon (n)} \nabla_x v \right| \leq
\left(\frac 1{c^*_{\varepsilon}(n)^2}+\frac 1{c^*_{\varepsilon}(n)}\right)C_3v=-\left(\frac 1
{c^*_{\varepsilon}(n)}+1\right)C_3 \partial_z U^*_\varepsilon.
$$
Since $\kappa=\inf_{0 < \varepsilon \leq \varepsilon_0} \inf_{n\in
\mathbb{S}^{N-1}} c^*_\varepsilon (n)
>0$, this proves the lemma.
\end{proof}
Let us now go back to the proof of Theorem~\ref{th:uc_speeds}.
Proceed by contradiction and assume that
$c_\infty \geq c^* (n) + \delta$ for some $\delta >0$. We
plug $U^*_{\varepsilon_k}(\cdot,\cdot;n_k)$ into equation~\eqref{eq-tw} satisfied by
$U^*(\cdot,\cdot;n)$ and, thanks to the above lemma, get
\begin{eqnarray}
& & \mbox{div}_x (A \nabla_x U^*_{\varepsilon_k}) + (n \cdot
An)\, \partial_{zz} U^*_{\varepsilon_k} +
\mbox{div}_x ( An \, \partial_z U^*_{\varepsilon_k}) + \partial_z (n\cdot A \nabla_x U^*_{\varepsilon_k}) \vspace{3pt} \nonumber \\
&& \qquad \quad + \, q \cdot \nabla_x U^*_{\varepsilon_k} + (q \cdot n) \, \partial_z U^*_{\varepsilon_k} + c^*(n)\partial _z U^*_{\varepsilon_k} +f(x,U^*_{\varepsilon_k}) \vspace{3pt}\nonumber \\
& = & (n \cdot An - n_k \cdot A n_k) \partial_{zz} U^*_{\varepsilon_k} + \mbox{div}_x ((An - An_k) \times \partial_z U^*_{\varepsilon_k}) \nonumber \\
& & \qquad \quad + \partial_z ((n - n_k)\cdot A \nabla_x U^*_{\varepsilon_k}) + (q \cdot (n-n_k))\partial_z U^*_{\varepsilon_k} \nonumber \\
& & \qquad \quad + (c^* (n) -c^*_{\varepsilon_k} (n_k)) \partial_z U^*_{\varepsilon_k} + f (x,U^*_{\varepsilon_k}) - f_{\varepsilon_k} (x,U^*_{\varepsilon_k}) \vspace{3pt} \nonumber \\
& \geq & \left[ 4 a_2 C | n- n_k| + \mbox{div}_x \, A (n-n_k) +
|q| | n- n_k | - \frac \delta 2\right] \partial_{z}
U^*_{\varepsilon_k}
\vspace{3pt}\nonumber \\
& \geq & -\frac{ \delta}{3}
\partial_z U^*_{\varepsilon_k} >0,\label{ineg}
\end{eqnarray}
provided $k$ is large enough, and where $a_2>0$ comes from
\eqref{ellipticity}. We now use the sliding method. From the
asymptotics
$$
U^*_{\varepsilon_k} (+\infty, \cdot ;n_k) = - \varepsilon _k < 0 =
U^* (+\infty, \cdot ;n),
$$
$$
U^*_{\varepsilon_k} (-\infty,\cdot ;n_k) = 1 - \frac{ \varepsilon
_k}{2} < 1 = U^* (+\infty, \cdot;n),
$$
one can define
$$\tau_0 := \inf \{ \tau:\, U^*_{\varepsilon_k} (z + \tau,x ;n_k) < U^* (z,x;n), \forall z \in \mathbb R, \forall x \in \mathbb R ^N \} \in \mathbb R.$$
Then, using again the asymptotics as well as the periodicity with
respect to $x$ of any pulsating wave, there is some first touching
point $(z_0,x_0)\in\mathbb R \times \mathbb R^N$ such that
$$
U^*_{\varepsilon_k} (z_0 + \tau _0,x_0;n_k) = U^* (z_0,x_0;n),
\quad \text{ and }\quad U^*_{\varepsilon_k} (\cdot + \tau
_0,\cdot;n_k) \leq U^* (\cdot,\cdot;n).$$ Substracting the
equation~\eqref{eq-tw} satisfied by $U^*(z,x;n)$ to the
inequality~\eqref{ineg} satisfied by $U^*_{\varepsilon_k}(z+\tau
_0,x;n_k)$ above, and estimating it at point $(z_0,x_0)$, we get
that
$$
0 \geq - \frac{\delta}{3} \partial_z U^*_{\varepsilon_k}(z_0+\tau
_0,x_0;n_k)
>0,
$$
a contradiction. Hence, $c_\infty = c^* (n)$, and the convergence
of $c^*_\varepsilon(n)$ to $c^*(n)$ is uniform.
\end{proof}
\begin{rem}[On the convergence of profiles]
The argument above also shows that the ignition traveling waves
converge locally uniformly,
up to a subsequence, to a traveling wave with minimal speed of the monostable equation.
Proceeding as in Section~\ref{s:ignition} and thanks to
the monotonicity of traveling waves, one can check that this convergence is actually uniform in time
and space. In particular, they do not flatten as the parameter $\varepsilon \to 0$. However, as the uniqueness
of the monostable traveling wave with minimal speed is not known \cite{Ham-Roq}, we cannot conclude on the convergence of the whole sequence.
\end{rem}
\section{The uniform lower spreading}\label{s:lower-spreading}
In this section and the next, we will prove
Theorem~\ref{th:unif_spreading} under the monostable assumption.
The easier ignition case will be dealt with in the last section.
We begin here with the uniform lower spreading
property~\eqref{conclusion1} of Theorem~\ref{th:unif_spreading}.
The argument again relies on the approximation from below by an
ignition type problem, and follow the footsteps of the proof of
Theorem~\ref{th:uc_speeds}.
\begin{proof}[Proof of \eqref{conclusion1}]
Recall that $f_\varepsilon(x,u)$ is an ignition type nonlinearity
which approximates $f(x,u)$ from below as $\varepsilon \to 0$. We
still
denote $u_\varepsilon^* (t,x) = U^*_\varepsilon (x\cdot n - c^*_\varepsilon (n) t,x;n)$ the unique ignition pulsating traveling wave
of \eqref{ignition_approx} in the direction $n$, normalized by $\min _{x\in \mathbb R ^N}U_{\varepsilon} ^*(0,x;n)=\frac 12$.
As $f_\varepsilon \leq f$, it is clear that $u^*_\varepsilon$ is a
subsolution of \eqref{monostable}, whose speed is arbitrary close
to $c^* (n)$ as $\varepsilon \to 0$ thanks to
Theorem~\ref{th:uc_speeds}. This leads back to Weinberger's result
\cite{Wein02}, namely the fact that for any planar-like initial
datum in the $n$ direction, the solution of \eqref{monostable}
spreads with speed \lq\lq at least'' $c^* (n)$ in the $n$
direction.
Let us now make this spreading property uniform with respect to
the family of solutions $(u_n)_{n \in \mathbb{S}^{N-1}}$, as
stated in Theorem~\ref{th:unif_spreading}. In the following $\mu$
and $K$ are as in assumption~\eqref{hyp2} (monostable case). Let
$\alpha
>0$ and $\delta >0$ be given. In view of assumption \eqref{hyp0} and the comparison principle we
have $u_n(t,x)\leq 1$. Hence to prove \eqref{conclusion1}, we need
to find $\tau
>0$ so that
\begin{equation}\label{conclusion1-dessous} \inf_{n \in
\mathbb{S}^{N-1}} \ \inf_{x \cdot n \leq (c^*(n) - \alpha)t} u_n
(t,x) \geq 1 - \delta,
\end{equation}
holds for all $t\geq \tau$.
\medskip
In view of Theorem~\ref{th:uc_speeds}, we can fix $\varepsilon>0 $
small enough so that, for all $n\in \mathbb S ^{N-1}$,
\begin{equation}\label{speed-prime}
c^*_{\varepsilon}(n) \geq c^* (n) - \frac \alpha 2 .
\end{equation}
We then claim that one can find some $t_\varepsilon>0$ such that
\begin{equation}\label{claim:proof_uniformspread1}
u_n (t_\varepsilon,x) \geq 1 - \frac{\varepsilon}{2},
\end{equation}
for all $n \in \mathbb{S}^{N-1}$ and all $x$ such that $x \cdot n
\leq -K$. We insist on the fact that $t_\varepsilon$ does not
depend on $n \in \mathbb{S}^{N-1}$. To prove
\eqref{claim:proof_uniformspread1}, let us define
$$
\mathcal{S}= \{x \in \mathbb R^N : \; x \cdot n \leq c^* (n) \mbox{ for all } n \in \mathbb{S}^{N-1} \}.
$$
We know from Theorem \ref{th:continuity_speeds} that the mapping
$n \mapsto c^* (n)$ is positive and continuous, hence
$\mathcal{S}$ has nonempty interior. It is then known (see
Theorem~2.3 in \cite{Wein02}, as well as Remark~\ref{rem_bug}
above) that for compactly supported initial data \lq\lq with large
enough support'', the associated solution of \eqref{monostable}
converges locally uniformly to 1 as $t\to +\infty$ (in fact, even
uniformly on the expanding sets $t \mathcal{S}'$ for any subset
$\mathcal{S'}$ of the interior of $\mathcal{S}$; also, under the
additional assumption that $0$ is linearly unstable with respect
to the periodic problem, this is even true for any non trivial
and compactly supported initial datum, regardless of its
size~\cite{Ber-Ham-Roq1}, \cite{Ber-Ham-Nad08}). More precisely,
let $u_R$ be the solution of \eqref{monostable} associated with
the initial datum $u_{0,R} (x) = \mu \times \chi_{B_R} (x)$, where
$R$ is a large but fixed positive constant (depending on $\mu$)
which we can assume to be larger than $2 \sqrt{N} \max_i L_i$.
Here $B_R$ denotes the ball of radius $R$ centered at the origin.
Then $u_R$ converges locally uniformly to 1 as $t \to +\infty$. In
particular, $$u_R (t_\varepsilon,x) \geq 1 -
\frac{\varepsilon}{2},$$ for some $t_\varepsilon >0$ and all $x
\in B_{2R}$. Besides, for $x_0 \in \Pi_{i =1}^{N} L_i \mathbb{Z}$
such that $x_0 \cdot n \leq -K-R$, we have --- thanks to
\eqref{hyp2}--- that $u_n(0,x+x_0)\geq u_R(0,x)$. Then, by the
comparison principle,
$$
\forall x \in B_{2R}, \quad u_n (t_\varepsilon,x+x_0) \geq
u_R(t_\varepsilon,x)\geq 1-\frac \varepsilon 2 .
$$
Since $R > 2 \sqrt{N} \max_i L_i$, for all $x \cdot n \leq -K$,
there exists $x_0 \in \Pi_{i =1}^{N} L_i \mathbb{Z}$ such that
$x_0 \cdot n \leq -K -R$ and $x \in B_{2R} (x_0)$. Thus, we obtain
$u_n (t_\varepsilon ,x) \geq 1 - \frac{\varepsilon}{2}$, for all
$n \in \mathbb{S}^{N-1}$ and $x \cdot n \leq -K$, that is
claim~\eqref{claim:proof_uniformspread1}.
\medskip
Now, recall that $U^*_{\varepsilon} (\cdot , \cdot ;n)$ is the
pulsating traveling wave of equation~\eqref{ignition_approx} in
the direction~$n$, connecting $1-\frac{\varepsilon}{2}$ to
$-\varepsilon$. Hence, it follows from
\eqref{claim:proof_uniformspread1} that, for any $n\in \mathbb S
^{N-1}$, one can find some shift $Z_n$ such that
\begin{equation}\label{ineq_1234}
u_n (t_\varepsilon,x) \geq U^*_{\varepsilon} ( x \cdot n -
c^*_{\varepsilon} (n) t_\varepsilon + Z_n,x;n).
\end{equation}
Actually, it suffices to select
$$
Z_n := \min \{ z \in \mathbb R:\; \ \min_{x \in \mathcal{C}} U^*_{\varepsilon} ( - K -
c^*_{\varepsilon} (n) t_\varepsilon + z,x;n) \leq 0 \} \in
(0,\infty).
$$ Moreover, from the
uniform continuity of ignition traveling waves with respect to the
direction, namely Theorem~\ref{th:continuity_igwaves}, it is
straightforward that the family $(U^*_{\varepsilon} (z,x;n))_{n
\in \mathbb{S}^{N-1}}$ converges to $-\varepsilon$ as $z \to
+\infty$ uniformly with respect to $n \in \mathbb{S}^{N-1}$.
Therefore, we can also define the bounded real number $ Z :=
\sup_{n \in \mathbb{S}^{N-1}} Z_n\in (0,\infty)$, so that
\eqref{ineq_1234} is improved to
$$
\forall n\in \mathbb S ^{N-1}, \quad u_n (t_\varepsilon,x) \geq
U^*_{\varepsilon} ( x \cdot n - c^*_{\varepsilon} (n)
t_\varepsilon + Z,x;n).
$$
Then we can apply the parabolic comparison principle to get
\begin{equation}\label{ineq_12345}
\forall t \geq t_\varepsilon, \forall x \in \mathbb R^N, \forall n \in
\mathbb{S}^{N-1}, \quad u_n (t,x) \geq U^*_{\varepsilon} (x \cdot
n - c^*_{\varepsilon} (n) t + Z , x;n).
\end{equation}
Therefore it follows from \eqref{speed-prime}, \eqref{ineq_12345}
and the monotonicity of the front that
\begin{equation}\label{low_uspread1}
u_n (t,x) \geq U^*_\varepsilon \left( -\frac{\alpha}{2} t + Z ,
x;n \right),
\end{equation}
for all $n \in \mathbb{S}^{N-1}$, all $t \geq t_\varepsilon$ and
all $x$ such that $x \cdot n \leq (c^* (n) - \alpha)t$. Using
again the uniform continuity of ignition traveling waves with
respect to the direction, namely
Theorem~\ref{th:continuity_igwaves}, one can find some shift
$Z'>0$ such that, for all $n \in \mathbb{S}^{N-1}$,
\begin{equation}\label{low_uspread2}
z\leq - Z' \ \Rightarrow \ U^*_\varepsilon (z,x;n) \geq 1 -
\varepsilon.
\end{equation}
Up to decreasing $\varepsilon$, we can assume that
$\varepsilon<\delta$ without loss of generality. Now choose $\tau
\geq t_\varepsilon$ such that $-\frac{\alpha}{2} \tau + Z \leq
Z'$. Then, we get from \eqref{low_uspread1} and
\eqref{low_uspread2} that
$$
u_n (t,x) \geq 1 - \delta,
$$
for all $n \in \mathbb{S}^{N-1}$, $t \geq \tau$ and $x$ such that
$x \cdot n \leq (c^* (n) - \alpha) t$. We have thus proved
\eqref{conclusion1-dessous}, and hence \eqref{conclusion1}.
\end{proof}
\section{The uniform upper spreading}\label{s:upper-spreading}
We conclude here the proof of Theorem \ref{th:unif_spreading} (monostable case), by
proving the uniform upper spreading~\eqref{conclusion2} in
subsection \ref{ss:with}. Then in subsection \ref{ss:without} we
again prove \eqref{conclusion2} --- together with the uniform
lower spreading property \eqref{conclusion1}--- when
assumption~\eqref{hyp0} is relaxed.
\subsection{Proof of \eqref{conclusion2}}\label{ss:with}
We begin by proving some kind of uniform steepness of the
monostable minimal waves, which in turn will easily imply
\eqref{conclusion2}.
\begin{prop}[Steepness of critical waves]\label{steep_min_wave}
Assume that $f$ is of the spatially periodic monostable type, i.e.
$f$ satisfies \eqref{periodicity} and
Assumption~\ref{hyp:monostable}.
Let $u^* (t,x;n) = U^* (x \cdot n - c^* (n) t , x ;n)$ be a family
of increasing in time pulsating
traveling waves of \eqref{monostable}, with minimal speed $c^*(n)$
in each direction $n\in \mathbb{S}^{N-1}$, normalized by $U^* (0,0;n)= \frac{1}{2}$.
Then, the asymptotics $U^*( - \infty, x;n) = 1$, $ U^*(\infty,x;n)
= 0$ are uniform with respect to $n \in \mathbb{S}^{N-1}$.
Moreover, for any $K >0$, we have $\inf_{n \in \mathbb{S}^{N-1}}
\inf_{|z| \leq K} \inf_{ x\in \mathbb R^N} -\partial_z U^* (z,x;n)
>0$ and $\inf_{n \in \mathbb{S}^{N-1}}
\inf_{|z| \leq K} \inf_{ x\in \mathbb R^N} U^* (z,x;n)
>0$.
\end{prop}
\begin{rem} [Lack of uniqueness]\label{rem-lack}
Such a family of traveling waves is always known to exist.
However, the uniqueness of the traveling wave with minimal speed
in each direction is not known. We shall prove that any sequence
of increasing in time traveling waves with minimal speed in the
directions $n_k \to n$ converges, up to extraction of a
subsequence, to an increasing in time
traveling wave with minimal speed in the direction $n$, as we did in the ignition case. The
proposition then easily follows, but the lack of uniqueness is the reason we state this result in a slightly different way.
\end{rem}
\begin{proof}
Proceeding as explained in the above remark, choose some sequence
$n_k \to n \in \mathbb{S}^{N-1}$. As before, one can extract a
subsequence such that $u^* (\cdot ,\cdot;n_k)$ converges locally
uniformly to a solution $u_\infty$ of \eqref{monostable}. By the
continuity of the speeds $c^* (n)$ with respect to $n$, as proved
in Theorem~\ref{th:continuity_speeds}, the function $u_\infty$
also satisfies
$$\forall l \in \prod_{i=1}^N L_i \mathbb{Z} , \quad u_\infty (t,x) = u _\infty\left( t + \frac{l \cdot n}{c^* (n)},x+l \right).$$
Moreover, it is nondecreasing in time, hence increasing in time by
applying the strong maximum principle to $\partial_t u_\infty$. In
particular, it converges to two spatially periodic stationary solutions as $t \to \pm
\infty$ which, as before and thanks to the monostable assumption,
must be 0 and $1$. As announced, $u_\infty$ is an increasing in
time traveling wave with minimal speed in the direction $n$.
Reasoning by contradiction, it is now straightforward to prove
Proposition~\ref{steep_min_wave}.
\end{proof}
\begin{proof}[Proof of \eqref{conclusion2}] First, from
Proposition~\ref{steep_min_wave} above, and
hypotheses~\eqref{hyp1}---\eqref{hyp0}, one can find some shift
$K_1>0$ large enough so that, for any $n \in \mathbb{S}^{N-1}$,
$u_{0,n} (x) \leq U^* (x \cdot n - K_1,x;n)$. Thus, by the
comparison principle,
$$u_n (t,x) \leq U^* ( x \cdot n- c^* (n) t - K_1 ,x ;n).$$
For any $\alpha >0$ and $\delta >0$, let $\tau$ be such that $U^*
(\alpha \tau - K_1, x;n) \leq \delta$, for all $n \in
\mathbb{S}^{N-1}$ and $x \in \mathbb R^N$, which is again made possible
by Proposition~\ref{steep_min_wave}. Then \eqref{conclusion2}
immediately follows.
\end{proof}
\subsection{Relaxing assumption \eqref{hyp0}}\label{ss:without}
We here consider the case when the family $(u_{0,n})_{n \in
\mathbb{S}^{N-1}}$ does not necessarily satisfy \eqref{hyp0}, but
is only uniformly bounded: there is $M>0$ such that
\begin{equation}\label{bounded}
\forall x\in \mathbb R ^N, \forall n\in \mathbb S ^{N-1}, \quad
u_{0,n}(x)\leq M .
\end{equation}
We prove that, in such a situation, the uniform lower and upper
spreading properties \eqref{conclusion1} and \eqref{conclusion2}
remain true if we make the following additional assumptions on the
behavior of $f$, and in particular on its behavior above the
stationary state $p$.
\begin{assumption}[Additional assumptions]\label{ass_add}
\begin{enumerate}[(i)]
\item There is $\phi (t,x)$ a solution of \eqref{monostable} such
that $\phi (0,\cdot) \geq M$, and $\phi (t,x)$ converges uniformly
to $1$ as $t \to +\infty$. \item The steady state 0 of
\eqref{monostable} is linearly unstable with respect to periodic
perturbations. \item There exists some $\rho
>0$ such that $f (x,u)$ is nonincreasing with respect to $u$ in
the set $\mathbb R^N \times (1-\rho,1+ \rho)$.
\end{enumerate}
\end{assumption}
The first part of this assumption holds true, for instance, if
$f(x,s) < 0$ for all
$x \in \mathbb R^N$ and $s >1$. As we will see below, the second part can be expressed in
terms of some principal eigenvalue problem, and holds true as soon as $\partial_u f (x,0)$ is
positive on a non empty set. The last part is a natural extension of $(iii)$ of Assumption~\ref{hyp:monostable}.\\
Combining \eqref{conclusion1-dessous}, whose proof does not
require assumption \eqref{hyp0}, and a comparison of the solutions $(u_n)_{n\in\mathbb S ^{N-1}}$
with $\phi$ given by the above assumption, it is
clear that the lower spreading property \eqref{conclusion1} still holds true. In
the sequel, we prove the upper spreading property \eqref{conclusion2}. We
start with the following proposition, whose proof is identical to
that of Proposition~\ref{steep_min_wave} and does not require Assumption \ref{ass_add}.
\begin{prop}[Steepness of noncritical waves]\label{prop}
Assume that $f$ is of the spatially periodic monostable type, i.e.
$f$ satisfies \eqref{periodicity} and
Assumption~\ref{hyp:monostable}.
For any $\alpha >0$, let $u_\alpha (t,x;n) = U_\alpha (x \cdot n -
(c^* (n) +\alpha) t,x ;n)$ be a family
of increasing in time pulsating
traveling waves of \eqref{monostable}, in
direction $n$, with speed $c^* (n)+\alpha$, normalized by
$U_{\alpha} (0,0;n)=\frac{1}{2}$.
Then, the asymptotics $U_\alpha( - \infty, x;n) = 1$, $ U_\alpha
(\infty,x;n) = 0$ are uniform with respect to $n \in
\mathbb{S}^{N-1}$. Moreover, for any $K >0$, we have $\inf_{n \in
\mathbb{S}^{N-1}} \inf_{|z| \leq K} \inf_{ x\in \mathbb R^N} -\partial_z
U_\alpha (z,x;n)
>0$ and $\inf_{n \in
\mathbb{S}^{N-1}} \inf_{|z| \leq K} \inf_{ x\in \mathbb R^N}
U_\alpha (z,x;n)
>0$
\end{prop}
We now turn to the proof of the upper spreading
property~\eqref{conclusion2}, which relies on the construction of
a suitable family of supersolutions that were already used in
\cite{Ham-Roq} (following an idea of \cite{Fif-Mac}).
\begin{proof}[Proof of \eqref{conclusion2}]Let $\alpha
>0$ and $\delta >0$ be given. We need to find $\tau
>0$ so that estimate~\eqref{conclusion2} holds for all $t\geq \tau$.
First, we need to introduce some notations, and some well known
results (see \cite{Xin}, \cite{Ber-Ham-02}, \cite{Ham-Roq} among
others). We begin with the principal eigenvalue problem
\begin{equation}\label{principal0}
\left\{
\begin{array}{l}
- L_{0, n , \lambda} \phi_{n,\lambda} = \mu_0 (n,\lambda) \phi_{n,\lambda} \ \mbox{ in } \mathbb R^N , \vspace{3pt}\\
\phi_{n,\lambda} \mbox{ is periodic}, \; \phi_{n,\lambda} >0,\;
\Vert \phi_{n,\lambda} \Vert
_\infty=1,
\end{array}
\right.
\end{equation}
where
$$L_{0,n,\lambda} \phi =
\mbox{div} \, (A \nabla \phi) + \lambda^2 (n \cdot A n) \phi -
\lambda (\mbox{div} (An \phi) + n\cdot A\nabla \phi) + q \cdot
\nabla \phi - \lambda (q \cdot n) \phi + \partial_u f (x,0) \phi
.$$ This arises, similarly as in Section~\ref{s:ignition}, when
looking for moving exponential solutions of the type $e^{-\lambda
(x \cdot n-ct)} \phi_{n,\lambda} (x)$ of the linearized problem
around 0. Such solutions exist if and only if
\begin{equation*}
c \geq c^*_{lin} (n) := \min_{\lambda >0} \frac{- \mu_0
(n,\lambda)}{\lambda},
\end{equation*}
which is well-defined thanks to the linear instability of 0,
which reads as $\mu_0 = \mu_0 (n,0) < 0$. Moreover, it is known that $c^*
(n) \geq c^*_{lin} (n)$ \cite{Ham}. We introduce $\lambda (n)$ the
smallest positive solution of $-\mu_0 (n,\lambda) = \left(c^* (n)
+ \frac{\alpha}{4} \right) \lambda$. It is standard that $\mu_0
(n,\lambda)$ is continuous with respect to $n$ and, as it is known
to be concave, $\lambda (n)$ is also continuous with respect to
$n$. In particular
$$ 0 < \min_{n \in \mathbb{S}^{N-1}} \lambda (n) \leq \max_{n \in \mathbb{S}^{N-1}} \lambda (n) < +\infty.$$
Let some smooth and nonincreasing $\chi (z)$ be such that
$$\chi (z) =
\left\{ \begin{array}{l}
1 \mbox{ if } z < -1,\\
0 \mbox{ if } z > 1,
\end{array}
\right.
$$
and define, for $s\geq 0$ (a shift to be fixed later),
\begin{equation*}
\Phi (t,x)=\Phi _s(t,x;n) := \chi (\xi _s ) + (1- \chi (\xi _s))
\phi_{n,\lambda (n)} (x) e^{-\lambda (n)\xi _s},
\end{equation*}
where
$$
\xi_s=\xi_s(t,x;n)=x\cdot n- \left(c^*(n)+
\frac{\alpha}{2}\right)t-s.
$$
Note that $\Phi$ is nonnegative and, along with its derivatives,
is bounded uniformly with respect to $n$ and $s$.
Let us now define various positive constants. Choose $0<\eta
<\delta$ small enough so that
\begin{equation}\label{eta-zero}\forall x\in
\mathbb R^N , \ \forall 0 \leq u \leq \eta , \quad | \partial_u f(x,u) -
\partial_u f (x,0) | \leq \frac{\alpha}{4} \min_{n \in
\mathbb{S}^{N-1}} \lambda (n),
\end{equation}
\begin{equation}\label{eta-p}
\forall x\in \mathbb R^N , \ \forall 1 - \eta \leq u \leq 1 + \eta ,
\quad \partial_u f(x,u) \leq 0.
\end{equation}
Now, by Proposition~\ref{prop}, there is $K>1$ large enough such
that, for all $n \in \mathbb S ^{N-1}$, $x\in \mathbb R ^N$,
\begin{equation}\label{K}
\xi > K \Rightarrow 0\leq U_{\frac \alpha 4}(\xi,x;n)\leq \frac
\eta 2,\quad \xi <-K \Rightarrow 1-\frac \eta 2 \leq U_{\frac
\alpha 4}(\xi,x;n)\leq 1.
\end{equation}
Then, by Proposition~\ref{prop} again, we have
\begin{equation}\label{gamma}
\gamma:= \inf_{n \in \mathbb{S}^{N-1}} \inf_{|z| \leq K , x \in
\mathbb R^N} -
\partial_z U_{\frac{\alpha}{4}} (z,x;n) >0.
\end{equation}
Last, we define
$$\epsilon_1 := \frac \eta{2\Vert \Phi \Vert
_\infty} \ , \ \epsilon_2 := \frac {\alpha \gamma}{4(\Vert
\partial _t\Phi \Vert _ \infty+\Vert \mbox{div} (A\nabla \Phi)
\Vert _\infty + \Vert q \cdot \nabla \Phi \Vert _\infty+\Vert
\Phi \Vert _\infty \Vert \partial _u f\Vert _{L^\infty(\mathbb R^N \times
(0,1 +\frac \eta 2))})}$$ and
\begin{equation}\label{epsilon}
\epsilon:=\min\left(\epsilon _1 , \epsilon _2 \right)>0.
\end{equation}
\medskip
Now, we are going to show that
$$
v(t,x)=v_s(t,x;n):=U_{\frac{\alpha}{4}} \left(x \cdot n -
\left(c^* (n) +\frac{\alpha}{2}\right) t-s,x;n \right) +
\epsilon\Phi (t,x)=U_{\frac{\alpha}{4}} (\xi _s,x;n) +
\epsilon\Phi (t,x)
$$
is a supersolution of the monostable equation~\eqref{monostable}.
Straightforward computations and the mean value Theorem yield
\begin{eqnarray*}\mathcal L[v](t,x)&:=& \partial _t v(t,x)- \mbox{div} (A (x) \nabla v (t,x)) - q (x) \cdot \nabla v(t,x) - f( x,v(t,x))\\
&=&\epsilon \left[\partial _t \Phi(t,x)- \mbox{div} (A (x) \nabla
\Phi(t,x)) - q(x) \cdot \nabla \Phi (t,x) - \Phi(t,x)\partial _u f(x, \theta (t,x)) \right] \\
&& \qquad -\frac \alpha 4
\partial _z U_{\frac \alpha 4}(\xi _s,x;n),
\end{eqnarray*}
for some
$$
U_{\frac \alpha 4}(\xi _s,x;n)\leq \theta (t,x) \leq U_{\frac
\alpha 4}(\xi _s,x;n)+\epsilon \Phi(t,x).
$$
We distinguish three regions, depending on the values of $\xi _s$.
First, if $|\xi _s|\leq K$, the nonnegativity of $\mathcal
L[v](t,x)$ is obtained thanks to $-\frac \alpha 4
\partial _z U_{\frac \alpha 4}(\xi _s,x;n)\geq \frac \alpha 4 \gamma$
by \eqref{gamma} and the definition of~$\epsilon$ in
\eqref{epsilon}.
Next, if $\xi _s
>K$, then $\Phi(t,x)$ reduces to $\phi_{n,\lambda (n)} (x) e^{-\lambda
(n) \xi _s}$ and, dropping $-\frac \alpha 4
\partial _z U_{\frac \alpha 4}(\xi _s,x;n)$ which is positive, we
arrive at
\begin{eqnarray*}\frac 1 \epsilon\mathcal L[v](t,x)&\geq& \left[
\lambda(n)\left( c^*(n)+\frac{\alpha}{2}\right)+\mu
_0(n,\lambda(n))+\partial _u f(x,0)-\partial _u f(x,\theta(t,x))
\right] \phi_{n,\lambda (n)}
(x)e^{-\lambda (n) \xi _s}\\
&\geq &\left(\frac \alpha 4 \lambda (n)+\partial _u
f(x,0)-\partial _u f(x,\theta(t,x)) \right)\phi_{n,\lambda (n)}
(x)e^{-\lambda (n) \xi _s}.
\end{eqnarray*}
But, when $\xi _s >K$, \eqref{K} and $\epsilon\leq \epsilon _1$
imply $0\leq \theta (t,x)\leq \eta $, and the nonnegativity of
$\mathcal L[v](t,x)$ is obtained thanks to \eqref{eta-zero}.
Last, we consider the case where $\xi _s<-K$, so that $\Phi(t,x)$
reduces to $1$. Hence
$$\frac{1}{\epsilon} \mathcal L[v](t,x)\geq -\partial _u f(x,\theta(t,x)).
$$
But, when $\xi _s<-K$, \eqref{K} and $\epsilon\leq \epsilon _1$
imply $1-\eta\leq\theta(t,x)\leq 1+\eta$, and the nonnegativity of
$\mathcal L[v](t,x)$ is obtained thanks to \eqref{eta-p}. Hence, $
v_s(t,x;n)$ is a supersolution of \eqref{monostable}.
\medskip
Thanks to \eqref{bounded}, we get by the comparison principle that, for all $ n \in \mathbb{S}^{N-1}$, all $t\geq 0$, all $x \in \mathbb R^N$, $ u_n (t,x) \leq \phi (t,x)$,
where $\phi$ is
given by Assumption~\ref{ass_add}.
Now choose $T >0$ such that $\phi (T,x) \leq 1 +
\frac{\epsilon}{2}$, and get that
\begin{equation}\label{end1}
\forall n \in \mathbb{S}^{N-1}, \ \forall x \in \mathbb R^N , \quad u_n (T,x) \leq 1 +
\frac{\epsilon}{2}.
\end{equation}
Using the comparison principle and a computation identical to that of \eqref{gggg}, we get that, for any large $\lambda >0$, there is $C>0$ --- independent on $n$ thanks to \eqref{hyp1} and
\eqref{bounded}--- such that
$$
\forall n \in \mathbb{S}^{N-1}, \ \forall t \geq 0 , \ \forall x \in \mathbb R^N , \quad u_n (t,x) \leq Ce^{-\lambda (x\cdot n-2a_1\lambda t)}.
$$
In particular $u_n (T,\cdot)$ decays faster than any
exponential as $x\cdot n \to +\infty$, namely
\begin{equation}\label{faster}
\forall \lambda >0 , \quad u_n (T,x) e^{\lambda x\cdot n}
\to 0 \ \mbox{ as } x \cdot n \to +\infty, \ \mbox{ uniformly w.r.t. } n \in \mathbb{S}^{N-1}.
\end{equation}
Observe that, for all $s\geq 0$,
\begin{equation}\label{droite-s}
\forall n\in \mathbb S ^{N-1},\ \forall x\cdot n\geq \left(c^{*}(n)+\frac{\alpha}{2}\right)T+s+1, \ v_s(T,x;n)\geq \epsilon \phi _{n,\lambda(n)}(x)e^{-\lambda(n)x\cdot n}\geq \epsilon \gamma e^{-\lambda_{max}x\cdot n},
\end{equation}
where $\gamma:=\min_{n \in
\mathbb{S}^{N-1}} \min_{ x\in \mathbb R^N} \phi_{n,\lambda (n)} (x)
>0$ and $\lambda _{max}:=\max_{n\in \mathbb S ^{N-1}} \lambda(n)<\infty$ (recall that $n\mapsto \lambda (n)$ is
continuous and so is $(n,\lambda)\mapsto \phi_{n,\lambda}$). Now, select $A>1$ large enough so that, for all $s\geq 0$,
\begin{equation}\label{gauche-s}
\forall n\in \mathbb S ^{N-1},\ \forall x\cdot n\leq \left(c^{*}(n)+\frac{\alpha}{2}\right)T+s-A, \ v_s(T,x;n)\geq 1+\frac \epsilon 2,
\end{equation}
which is possible thanks to Proposition~\ref{prop}, and more
precisely the uniform with respect to~$n$ asymptotics of
$U_{\frac{\alpha}{4}} (z,x;n)$ as $z \to -\infty$. Proposition~\ref{prop} also enables to define
$$
\kappa:=\inf_{n\in \mathbb S ^{N-1}} \inf _{-A\leq z\leq 1} \inf _{x\in \mathbb R ^{N}}U_{\frac{\alpha}{4}} (z,x;n)>0,
$$
so that, for all $s\geq 0$,
\begin{equation}\label{milieu-s}
\forall n\in \mathbb S ^{N-1},\ \forall \left(c^{*}(n)+\frac{\alpha}{2}\right)T+s-A\leq x\cdot n\leq \left(c^{*}(n)+\frac{\alpha}{2}\right)T+s+1, \ v_s(T,x;n)\geq \kappa.
\end{equation}
In view of \eqref{faster}, we can now select a large enough shift $s_0>A$ so that
$$
\forall n\in \mathbb S ^{N-1},\ \forall x\cdot n\geq \left(c^{*}(n)+\frac{\alpha}{2}\right)T+s_0-A, \ u_n(T,x)\leq \min \{ \epsilon \gamma , \kappa \} e^{-\lambda_{max}x\cdot n}.
$$
Combining this with \eqref{end1}, \eqref{droite-s}, \eqref{gauche-s}, \eqref{milieu-s}, we have that, for all $n \in \mathbb{S}^{N-1}$ and $x \in \mathbb R^N$,
\begin{equation*
u_n(T,x)\leq v_{s_0}(T,x;n).
\end{equation*}
\medskip
Then, by the comparison principle, for all $t\geq T$, $x\in \mathbb R
^N$, $n\in\mathbb S ^{N-1}$,
$$
0\leq u_n(t,x)\leq U_{\frac \alpha 4} \left( x\cdot n- \left(
c^*(n)+ \frac{\alpha}{2} \right)t-s_0,x;n \right)+\epsilon
\Phi(t,x).
$$
Hence, when $x\cdot n\geq (c^*(n)+\alpha )t$, we have, since
$\epsilon \Vert \Phi \Vert _\infty \leq \frac \eta 2 \leq \frac
\delta 2$, that
$$
0\leq u_n(t,x)\leq U_{\frac \alpha 4}\left( \frac{\alpha}{2}
t-s_0,x;n\right)+\frac \delta 2 \leq \delta,
$$
as soon as $t\geq \tau$, where $\tau>0$ is large enough (again
independently on $n$ by Proposition~\ref{prop}). This proves
\eqref{conclusion2}.\end{proof}
\section{The uniform spreading: the ignition case}\label{s:spread_ignition}
For the sake of completeness, we give here the main steps to prove
Theorem~\ref{th:unif_spreading} in the (simpler) ignition
case. We will see that it follows from the continuity of
ignition waves, Theorem~\ref{th:continuity_igwaves},
together with the standard idea explained in Remark~\ref{rem_bug}. We will briefly sketch at
the end of this section how the hypothesis \eqref{hyp0} can again be relaxed.
\begin{proof}[Proof of Theorem \ref{th:unif_spreading} in the
ignition case]
First, the proof of the uniform upper
spreading~\eqref{conclusion2} is the same as that of subsection
\ref{ss:with}
in the monostable case, using Theorem~\ref{th:continuity_igwaves} (continuity of ignition waves) instead of Proposition~\ref{steep_min_wave} (steepness of critical
waves).
\medskip
Let us now prove the uniform lower spreading~\eqref{conclusion1}.
Let $\alpha >0$ and $\delta >0$ be given.
We may reduce $\delta$ without loss of generality, and assume that $\delta < \rho$ where $\rho$ is given by part
$(iii)$ of Assumption~\ref{hyp:ignition}. Using the same
arguments as in the proof of \eqref{claim:proof_uniformspread1},
we get the existence of some time $t_\delta>0$ such that
$$
u_n (t_\delta,x) \geq 1 - \frac{\delta}{2},
$$
for all $n \in \mathbb{S}^{N-1}$ and all $x$ such that $x \cdot n
\leq - K$. Now let, as usual, $U^* (x\cdot n - c^* (n) t,x;n)$ be
the unique ignition pulsating wave in the direction~$n$,
normalized by $\min _{x\in \mathbb R ^N} U^* (0,x;n) =
\frac{1+\theta}{2}.$ Thanks to the inequality above and the
continuity of the mapping $n \mapsto U^* (\cdot , \cdot;n)$ with
respect to the uniform topology, it is clear that there exists
some shift $Z>0$ such that, for all $n\in \mathbb{S}^{N-1}$,
$$
U^* (x \cdot n +Z ,x ;n) - \frac{\delta}{2} \leq u_n (t_\delta
,x), \quad \forall x\in \mathbb R ^N.
$$
We then check that $\underline{u} (t,x) := U^* (x \cdot n + Z -
(c^* (n) - \frac{\alpha}{2}) t,x;n) - \frac{\delta}{2}$ is a
subsolution of \eqref{monostable}. Indeed,
\begin{equation*}
\partial_t \underline{u}- \mbox{div} \, (A(x) \nabla \underline{u}) - q(x) \cdot \nabla \underline{u} - f(x,\underline{u}) = \frac{\alpha}{2} \partial_z U^* + f(x,U^*) - f(x,\underline{u}).
\end{equation*}
Assume first that $\underline{u} \leq \theta - \frac{\delta}{2}$.
Then $f(x,\underline{u})= f(x,U^* (x\cdot n + Z - (c^* (n) -
\frac{\alpha}{2} )t,x;n)=0$, which together with the monotonicity
of $U^*$ with respect to its first variable, gives the wanted
inequality. Assume then that $\underline{u} \geq 1 - \rho$. Then,
by the monotonicity of $f$ with respect to $u$ in the range
$[1-\rho,1]$, we again obtain the wanted inequality.
It remains to
prove that $\underline{u}$ is a subsolution when $\theta -
\frac{\delta}{2} \leq \underline{u} \leq 1 - \rho$ or,
equivalently, when $\theta \leq U^* (x\cdot n - (c^* (n)- \frac{\alpha}{2})t+Z,x;n) \leq 1 - \rho + \frac{\delta}{2}$.
Recall first that $1 - \rho + \frac{\delta}{2}< 1 - \frac{\rho}{2}$. Using again the continuity of the ignition wave
with respect to the direction, we have that there exists some~$R>0$ such
that, for all $n\in \mathbb S ^{N-1}$,
$$z \geq R \ \Rightarrow \ U^* (z + Z ,x;n) < \theta, \quad z \leq -R \ \Rightarrow \ U^* (z+Z,x;n) > 1 - \frac{\rho}{2},$$
and, furthermore,
$$\max_{n \in \mathbb{S}^{N-1}} \max_{|z| \leq R} \ \max_{x \in \mathbb R^N} \ \partial_z U^* (z+Z,x;n) <0.$$
Up to reducing $\delta$ again, we may assume that
$$\max_{n \in \mathbb{S}^{N-1}} \max_{|z| \leq R} \ \max_{x \in \mathbb R^N} \ \partial_z U^* (z+Z,x;n) < - \frac{M\delta}{\alpha},$$
where $M$ is a Lipschitz constant of $f$. Therefore, when $\theta
- \frac{\delta}{2} \leq \underline{u} \leq 1 - \rho$, then $|x
\cdot n - (c^* (n) - \frac{\alpha}{2} )t | \leq R$ and
\begin{eqnarray*}
&& \partial_t \underline{u}- \mbox{div} \, (A(x) \nabla \underline{u}) - q(x) \cdot \nabla \underline{u} - f(x,\underline{u}) \vspace{3pt}\\
& = & \frac{\alpha}{2} \partial_z U^* + f(x,U^*) - f(x,\underline{u}) \vspace{3pt}\\
& \leq & \frac{\alpha}{2} \partial_z U^* + M \frac{\delta}{2}
\leq 0,
\end{eqnarray*}
that is the wanted inequality.
We can therefore apply the comparison principle and conclude that
$$U^* \left(x \cdot n + Z - (c^* (n) - \frac{\alpha}{2})t ,x;n \right) - \frac{\delta}{2}\leq u_n (t_\delta +t,x),$$
for all $x \in \mathbb R^N$ and $t \geq 0$.
Noting that there exists some other shift $Z'>0$ such that, for
all $ n \in \mathbb{S}^{N-1}$,
$$z \leq -Z ' \ \Rightarrow \ U^* (z,x;n) - \frac{\delta}{2} \geq 1 -\delta,$$
we get the uniform lower spreading \eqref{conclusion1} as in the
end of Section \ref{s:lower-spreading}.
\end{proof}
\begin{proof}[Relaxing hypothesis \eqref{hyp0}]
In order to relax \eqref{hyp0}, assume now that $f$ satisfies
Assumption~\ref{hyp:ignition} and parts $(i)$ and $(iii)$ of
Assumption~\ref{ass_add}. As above, one can then show that $U^* (x
\cdot n - (c^* (n) + \frac{\alpha}{2})t,x;n) + \frac{\delta}{2}$
is a supersolution of \eqref{monostable}. Then, as in
Section~\ref{ss:without}, one can find some time $T$ and some
shift $s_0$ such that, for all $n$, the solution $u_n (T,x)$ lies
below $U^* (x \cdot n + s_0 - ( c^* (n)+ \frac{\alpha}{2})T,x;n) +
\frac{\delta}{2}$ in the whole space. It is then straightforward
to obtain the wanted uniform upper spreading~\eqref{conclusion2}.
\end{proof}
\bigskip
\noindent \textbf{Acknowledgement.} M. A. was supported by the
French {\it Agence Nationale de la Recherche} within the project
IDEE (ANR-2010-0112-01). T. G. was supported by the French {\it Agence Nationale de la Recherche} within the project NONLOCAL (ANR-14-CE25-0013). Both authors are grateful to Professor
Hiroshi Matano for great hospitality in the University of Tokyo,
where this work was initiated.
|
\section{Introduction}
\noindent
Controllability is one of the elementary concepts in mathematical control theory. This is a qualitative property of dynamical control systems and it is of particular importance in control theory. Roughly speaking, controllability generally means, that it is possible to steer dynamical control system from an arbitrary initial state to an arbitrary final state using the set of admissible controls. Most of the criteria, which can be met in the literature, are formulated for finite dimensional systems. It should be pointed out that many unsolved problems still exist as far as controllability of infinite dimensional systems are concerned.
In the case of infinite dimensional systems two basic concepts of controllability can be distinguished which are exact and approximate controllability. This is strongly related to the fact that in infinite dimensional spaces there exist linear subspaces, which are not closed. Exact controllability enables to steer the system to arbitrary final state while approximate controllability means that system can be steered to arbitrary small neighborhood of final state. In other words approximate controllability gives the possibility of steering the system to states which form the dense subspace in the state space. In recent years, controllability problems for various types of nonlinear dynamical systems in infinite dimensional spaces by using different kinds of approaches have been considered in many publications, see \cite{na1, ga1, ga2, jpd1, tg1, hrh4, nim1, nim2, nim4, nim5, nim7, nim8, rs6, rs1, rs2, vv5, vv6, zy1, zy2} and the references therein.
Recently, in \cite{nim8}, Mahmudov et al. studied the approximate controllability of second-order neutral stochastic evolution equations using semi-group methods, together with the Banach fixed point theorem. In \cite{rs2} Sakthivel et al. studied the approximate controllability of second-order systems with state-dependent delay by using Schauder fixed point theorem. In \cite{hrh1}, Henr\'{i}quez studied the existence of solutions of non-autonomous second order functional differential equations with infinite delay by using Leray-Schauder's Alternative fixed point theorem. In \cite{zy1}, Yan studied the approximate controllability of fractional neutral integro-differential inclusions with state-dependent delay in Hilbert spaces by using Dhage's fixed point theorem. In \cite{vv6}, Vijayakumar et al. discussed the approximate controllability for a class of fractional neutral integro-differential inclusions with state-dependent delay by using Dhage's fixed point theorem. In \cite{rs6} Sakthivel et al. studied the approximate controllability of fractional nonlinear differential inclusions with initial and nonlocal conditions by using Bohnenblust-Karlin's fixed point theorem. Very recently, in \cite{ga1} Arthi et al. established the sufficient conditiond for controllability of second-order impulsive evolution systems with infinite delay by using Leray-Schauder's fixed point theorem and in \cite{ga2}, proved existence and controllability results for second-order impulsive stochastic evolution systems with state-dependent delay by using Leray-Schauder's fixed point theorem. In \cite{tg1}, Guendouzi investigated the approximate controllability for a class of fractional neutral stochastic functional integro-differential inclusions Bohnenblust-Karlin's fixed point theorem.
Inspired by the above works, in this paper, we establish sufficient conditions for the approximate controllability for a class of second-order evolution differential inclusions in Hilbert spaces of the form
\begin{align}
x''(t)\in&A(t)x(t)+F(t,x(t))+Bu(t), \quad t\in I=[0,b],\label{vi1}\\
x(0)=&x_0\in X,\quad x'(0)=y_0\in X.\label{vi2}
\end{align}
In this equation, $A(t):D(A(t))\subseteq X\rightarrow X$ is a closed linear operator on a Hilbert space $X$ and the control function $u(\cdot)\in L^2(I,U)$, a Hilbert space of admissible control functions. Further, $B$ is a bounded linear operator from $U$ to $X$, and $F:I\times X\rightarrow 2^X\setminus\{\emptyset\}$ is a nonempty, bounded, closed and convex multivalued map.
Converting a second-order system into a first-order system and studying its controllability may not yield desired results due to the behavior of the semigroup generated by the linear part of the converted first order system. So, in many cases, it is advantageous to treat the second-order abstract differential system directly rather than to convert that into first-order system. To the best of our knowledge, the study of the approximate controllability for a class of second-order evolution differential systems in Hilbert spaces treated in this paper, is an untreated topic in the literature, to fill the gap, we study this interesting paper.
This paper is organized as follows. In Section \ref{s3}, we establish a set of sufficient conditions for the approximate controllability for a class of second-order evolution differential inclusions in Hilbert spaces. In Section \ref{s4}, we establish a set of sufficient conditions for the approximate controllability for a class of second-order evolution differential inclusions with nonlocal conditions in Hilbert spaces. In Section \ref{s5}, we establish a set of sufficient conditions for the approximate controllability for a class of second-order impulsive evolution differential inclusions with nonlocal conditions in Hilbert spaces. An example is presented in Section \ref{s6} to illustrate the theory of the obtained results.
\section{Preliminaries}\label{s2}
\noindent
In this section, we mention a few results, notations and lemmas needed to establish our main results. We introduce certain notations which will be used throughout the article without any further mention. Let $(X,\|\cdot\|_X)$ and $(Y,\|\cdot\|_Y)$ be Hilbert spaces, and $\mathcal{L}(Y,X)$ be the Banach space of bounded linear operators from $Y$ into $X$ equipped with its natural topology; in particular, we use the notation $\mathcal{L}(X)$ when $Y=X$. By $\rho(A)$, we denote the resolvent set of a linear operator $A$. Throughout this paper, $B_r(x,X)$ will denote the closed ball with center at $x$ and radius $r>0$ in a Hilbert space $X$. We denote by $\mathcal{C}$, the Banach space $C(J,X)$ endowed with supnorm given by $\|x\|_\mathcal{C}\equiv \sup_{t\in I}\|x(t)\|$, for $x\in \mathcal{C}$.
In recent times there has been an increasing interest in studying the abstract non-autonomous second order initial value problem
\begin{align}
x''(t)=&A(t)x(t)+f(t), \quad 0\le s, t \le b,\label{vp1}\\
x(s)=&x_0,\quad x'(s)=y_0,\label{vp2}
\end{align}
where $A(t):D(A(t))\subseteq X\rightarrow X$, $t\in I=[0,b]$ is a closed densely defined operator and $f:I\rightarrow X$ is an appropriate function. Equations of this type have been considered in many papers. The reader is referred to \cite{cjk1, ff1, yp1, yl1, eo1} and the references mentioned in these works. In the most of works, the existence of solutions to the problem $\eqref{vp1}-\eqref{vp2}$ is related to the existence of an evolution operator $S(t,s)$ for the homogeneous equation
\begin{align}
x''(t)=&A(t)x(t), \quad 0\le s, t \le b,\label{vp3}
\end{align}
Let as assume that the domain of $A(t)$ is a subspace $D$ dense in $X$ and independent of $t$, and for each $x\in D$ the function $t\mapsto A(t)x$ is continuous.
Following Kozak \cite{mk1}, in this work we will use the following concept of evolution operator.
\begin{definition}\label{vd1}
A family $S$ of bounded linear operators $S(t,s):I\times I\rightarrow \mathcal{L}(X)$ is called an evolution operator for $\eqref{vp3}$ if the following conditions are satisfied:
\begin{enumerate}
\item[$(Z1)$] For each $x\in X$, the mapping $[0,b]\times[0,b]\ni(t,s)\rightarrow S(t,s)x\in X$ is of class $C^1$ and
\begin{itemize}
\item[$(i)$] for each $t\in [0,b]$, $S(t,t)=0$,
\item[$(ii)$] for all $t,s\in [0,b]$, and for each $x\in X$,
$$\frac{\partial}{\partial t}S(t,s)x\Big|_{t=s}=x,\qquad \frac{\partial}{\partial t}S(t,s)x\Big|_{t=s}=-x.$$
\end{itemize}
\item[$(Z2)$] For all $t,s\in [0,b]$, if $x\in D(A)$, then $S(t,s)x\in D(A)$, the mapping $[0,b]\times[0,b]\ni(t,s)\rightarrow S(t,s)x\in X$ is of class $C^2$ and
\begin{itemize}
\item[$(i)$] $\frac{\partial^2}{\partial t^2} S(t,x)x=A(t)S(t,s)x$,
\item[$(ii)$] $\frac{\partial^2}{\partial s^2} S(t,x)x=S(t,s)A(s)x$,
\item[$(iii)$] $\frac{\partial}{\partial s}\frac{\partial}{\partial t} S(t,x)x\Big|_{t=s}=0$.
\end{itemize}
\item[$(Z3)$] For all $t,s\in [0,b]$, if $x\in D(A)$, then $\frac{\partial}{\partial s}S(t,s)x\in D(A)$, then $\frac{\partial^2}{\partial t^2}\frac{\partial}{\partial s}S(t,s)x$, $\frac{\partial^2}{\partial s^2}\frac{\partial}{\partial t}S(t,s)x$ and
\begin{itemize}
\item[$(i)$] $\frac{\partial^2}{\partial t^2}\frac{\partial}{\partial s}S(t,s)x=A(t)\frac{\partial}{\partial s}S(t,s)x$,
\item[$(ii)$] $\frac{\partial^2}{\partial s^2}\frac{\partial}{\partial t}S(t,s)x=\frac{\partial}{\partial t}S(t,s)A(s)x$,
\end{itemize}
and the mapping $[0,b]\times[0,b]\ni(t,s)\rightarrow A(t)\frac{\partial}{\partial s}S(t,s)x$ is continuous.
\end{enumerate}
\end{definition}
Throughout this work we assume that there exists an evolution operator $S(t,s)$ associated to the operator $A(t)$. To abbreviate the text, we introduce the operator $C(t,s)=-\frac{\partial S(t,s)}{\partial s}$. In addition, we set $N$ and $\widetilde{N}$ for positive constants such that $\sup_{0\le t,s \le b}\|S(t,s)\|\le N$ and $\sup_{0\le t,s \le b}\|C(t,s)\|\le \widetilde{N}$. Furthermore, we denote by $N_1$ a positive constant such that $$\|S(t+h,s)-S(t,s)\|\le N_1|h|,$$
for all $s,t, t+h\in [0,b]$. Assuming that $f:I\rightarrow X$ is an integrable function, the mild solution $x:[0,b]\rightarrow X$ of the problem $\eqref{vp1}-\eqref{vp2}$ is given by
\begin{gather*}
x(t)=C(t,s)x_0+S(t,s)y_0+\int_0^t S(t,\tau)h(\tau)d\tau.
\end{gather*}
In the literature several techniques have been discussed to establish the existence of the evolution operator $S(\cdot,\cdot)$. In particular, a very studied situation is that $A(t)$ is the perturbation of an operator $A$ that generates a cosine operator function. For this reason, below we briefly review some essential properties of the theory of cosine functions. Let $A:D(A)\subseteq X\rightarrow X$ be the infinitesimal generator of a strongly continuous cosine family of bounded linear operators $(C_{0}(t))_{t\in \mathbb{R}}$ on Hilbert space $X$. We denote by $(S_0(t))_{t\in \mathbb{R}}$ the sine function associated with $(C_{0}(t))_{t\in \mathbb{R}}$ which is defined by
\begin{gather*}
S_0(t)x=\int_0^t C(s)xds, \quad x\in X, \quad t\in \mathbb{R}.
\end{gather*}
We refer the reader to \cite{hof1, cct1, cct2} for the necessary concepts about cosine functions. Next we only mention a few results and notations about this matter needed to establish our results. It is immediate that
\begin{gather*}
C_{0}(t)x-x=A\int_0^t S(s)xds,
\end{gather*}
for all $X$. The notation $[D(A)]$ stands for the domain of the operator $A$ endowed with the graph norm $\|x\|_A = \|x\| + \|Ax\|$, $x \in D(A)$. Moreover, in this paper the notation $E$ stands for the space formed by the vectors $x\in X$ for which the function $C(\cdot)x$ is a class $C^1$ on $\mathbb{R}$. It was proved by Kisy\'{n}ski \cite{jk1} that the space $E$ endowed with the norm
\begin{gather*}
\|x\|_E=\|x\|+\sup_{0\le t\le 1}\|AS(t,0)x\|, \quad x\in E,
\end{gather*}
is a Hilbert space. The operator valued function
\begin{gather*}
G(t)=\left[\begin{array}{lr} C_{0}(t)& S_0(t)\\ AS_0(t) & C_0(t) \end{array}\right]
\end{gather*}
is a strongly continuous group of linear operators on the space $E\times X$ generated by the operator $ \mathcal{A}=\left[\begin{array}{lr} 0& I\\ A & 0 \end{array}\right] $ defined on $D(A)\times E$. It follows from this that $AS_0(t):E\rightarrow X$ is a bounded linear operator such that $AS_0(t)x\rightarrow 0$ as $t\rightarrow 0$, for each $x\in E$. Furthermore, if $x:[0,\infty)\rightarrow X$ is a locally integrable function, then $z(t)=\int_0^t S(t,s)x(s)ds$ defines an $E$-valued continuous function.
The existence of solutions for the second order abstract Cauchy problem
\begin{align}
x''(t)=&Ax(t)+h(t), \quad 0\le t\le b,\label{vp4}\\
x(0)=&x_0, \quad x'(0)=y_0,\label{vp5}
\end{align}
where $h:[0,b]\rightarrow X$ is an integrable function, has been discussed in \cite{cct1}. Similarly, the existence of solutions of the semilinear second order Cauchy problem it has been treated in \cite{cct2}. We only mention here that the function $x(\cdot)$ given by
\begin{gather}
x(t)=C_0(t-s)x_0+S_0(t-s)y_0+\int_s^t S_0(t-\tau)h(\tau)d\tau, \quad 0\le t \le b, \label{vp6}
\end{gather}
is called the mild solution of $\eqref{vp4}-\eqref{vp5}$ and that when $x_0\in E$, $x(\cdot)$ is continuously differentiable and
\begin{gather*}
x'(t)=AS_0(t-s)x_0+C_0(t-s)y_0+\int_s^t C_0(t-\tau)h(\tau)d\tau, \quad 0\le t \le b.
\end{gather*}
In addition, if $x_0\in D(A)$, $y_0\in E$ and $f$ is a continuously differentiable function, then the function $x(\cdot)$ is a solution of the initial value problem $\eqref{vp4}-\eqref{vp5}$.
Assume now that $A(t)=A+\widetilde{B}(t)$ where $\widetilde{B}(\cdot):\mathbb{R}\rightarrow \mathcal{L}(E,X)$ is a map such that the function $t\rightarrow \widetilde{B}(t)x$ is continuously differentiable in $X$ for each $x\in E$. It has been established by Serizawa \cite{hs1} that for each $(x_0,y_0)\in D(A)\times E$ the nonautonomous abstract Cauchy problem
\begin{align}
x''(t)=&(A+\widetilde{B}(t))x(t), \quad t\in \mathbb{R},\label{vp7}\\
x(0)=&x_0, \quad x'(0)=y_0,\label{vp8}
\end{align}
has a unique solution $x(\cdot)$ such that the function $t\mapsto x(t)$ is continuously differentiable in $E$. It is clear that the same argument allows us to conclude that equation \eqref{vp7} with the initial condition \eqref{vp5} has a unique solution $x(\cdot,s)$ such that the function $t\mapsto x(t,s)$ is continuously differentiable in $E$. It follows from \eqref{vp6} that
\begin{gather*}
x(t,s)=C_0(t-s)x_0+S_0(t-s)y_0+\int_s^t S(t-\tau)\widetilde{B}(\tau)x(\tau,s)d\tau.
\end{gather*}
In particular, for $x_0=0$ we have
\begin{gather*}
x(t,s)=S_0(t-s)y_0+\int_s^t S(t-\tau)\widetilde{B}(\tau)x(\tau,s)d\tau.
\end{gather*}
Consequently,
\begin{gather*}
\|x(t,s)\|_1\le \|S_0(t-s)\|_{\mathcal{L}(X,E)}\|y_0\|+\int_s^t \|S_0(t-s)\|_{\mathcal{L}(X,E)}\|\widetilde{B}(\tau)\|_{\mathcal{L}(X,E)}\|x(s,\tau)\|_1d\tau.
\end{gather*}
and, applying the Gronwall-Bellman lemma we infer that
\begin{gather*}
\|x(t,s)\|_1\le \widetilde{M}\|y_0\|,\quad s, t\in I.
\end{gather*}
We define the operator $S(t,s)y_0=x(t,s)$. It follows from the previous estimate that $S(t,s)$ is a bounded linear map on $E$. Since $E$ is dense in $X$, we can extend $S(t,s)$ to $X$. We keep the notation $S(t,s)$ for this extension. It is well known that, except in the case $\dim(X)<\infty$, the cosine function $C_0(t)$ cannot be compact for all $t\in \mathbb{R}$. By contrast, for the cosine functions that arise in specific applications, the sine function $S_0(t)$ is very often a compact operator for all $t\in \mathbb{R}$.
\begin{theorem}\cite[Theorem 1.2]{hrh1}.
Under the preceding conditions, $S(\cdot,\cdot)$ is an evolution operator for $\eqref{vp7}-\eqref{vp8}$. Moreover, if $S_0(t)$ is compact for all $t\in \mathbb{R}$, then $S(t,s)$ is also compact for all $s\le t$.
\end{theorem}
We also introduce some basic definitions and results of multivalued maps. For more details on multivalued maps, see the books of Deimling \cite{kde1} and Hu and Papageorgious \cite{sh1}.
A multivalued map $G: X\rightarrow 2^X\setminus \{\emptyset\}$ is convex (closed) valued if $G(x)$ is convex (closed) for all $x\in X$. $G$ is bounded on bounded sets if $G(C)=\bigcup_{x\in C} G(x)$ is bounded in $X$ for any bounded set $C$ of $X$, i.e., $\sup_{x\in C}\Big\{\sup\{\|y\|:y\in G(x)\}\Big\}<\infty$.
\begin{definition}
$G$ is called upper semicontinuous (u.s.c. for short) on $X$ if for each $x_0\in X$, the set $G(x_0)$ is a nonempty closed subset of $X$, and if for each open set $C$ of $X$ containing $G(x_0)$, there exists an open neighborhood $V$ of $x_0$ such that $G(V)\subseteq C$.
\end{definition}
\begin{definition}
$G$ is called completely continuous if $G(C)$ is relatively compact for every bounded subset $C$ of $X$.
\end{definition}
If the multivalued map $G$ is completely continuous with nonempty values, then $G$ is u.s.c., if and only if $G$ has a closed graph, i.e., $x_n\rightarrow x_*$, $y_n\rightarrow y_*$, $y_n\in Gx_n$ imply $y_*\in Gx_*$. $G$ has a fixed point if there is a $x\in X$ such that $x\in G(x)$.
\begin{definition}
A function $x\in \mathcal{C}$ is said to be a mild solution of system \eqref{vi1}-\eqref{vi2} if $x(0)=x_0$, $x'(0)=y_0$ and there exists $f\in L^1(I,X)$ such that $f(t)\in F(t,x(t))$ on $t\in I$ and the integral equation
\begin{align*}
x(t)=C(t,0)x_0+S(t,0)y_0+\int_0^tS(t,s)f(s)ds+\int_0^tS(t,s)Bu(s)ds,\ t\in I.
\end{align*}
is satisfied.
\end{definition}
In order to address the problem, it is convenient at this point to introduce two relevant operators and basic assumptions on these operators:
\begin{align*}
\Upsilon_0^b&=\int_0^bS(b,s)BB^*S^*(b,s)ds: X\rightarrow X,\\
R(a,\Upsilon_0^b)&=(a I+\Upsilon_0^b)^{-1}: X\rightarrow X.
\end{align*}
It is straightforward that the operator $\Upsilon_0^b$ is a linear bounded operator.
To investigate the approximate controllability of system \eqref{vi1}-\eqref{vi2}, we impose the following condition:
\begin{enumerate}
\item [$\bf H_0$] $aR(a,\Upsilon_0^b)\rightarrow 0$ as $a\rightarrow 0^+$ in the strong operator topology.
\end{enumerate}
In view of \cite{nim1}, Hypothesis $\bf{H_0}$ holds if and only if the linear system
\begin{align}
x''(t)=&Ax(t)+(Bu)(t),\quad t\in [0,b], \label{s2e1} \\
x(0)=&x_0 \quad x'(0)=y_0,\label{s2e2}
\end{align}
is approximately controllable on $[0,b]$.
Some of our results are proved using the next well-known results.
\begin{lemma}\cite[Lasota and Opial]{al1}
Let $I$ be a compact real interval, $BCC(X)$ be the set of all nonempty, bounded, closed and convex subset of $X$ and $F$ be a multivalued map satisfying $F:I\times X\rightarrow BCC(X)$ is measurable to $t$ for each fixed $x\in X$, u.s.c. to $x$ for each $t\in I$, and for each $x\in \mathcal{C}$ the set
\begin{align*}
S_{F,x}=\{f\in L^1(I,X):f(t)\in F(t,x(t)),\ t\in I\}
\end{align*}
is nonempty. Let $\mathscr{F}$ be a linear continuous from $L^1(I,X)$ to $\mathcal{C}$, then the operator
\begin{gather*}
\mathscr{F} \circ S_F:\mathcal{C}\rightarrow BCC(\mathcal{C}),\ x\rightarrow (\mathscr{F} \circ S_F)(x)=\mathscr{F}(S_{F,x}),
\end{gather*}
is a closed graph operator in $\mathcal{C}\times \mathcal{C}$.
\end{lemma}
\begin{lemma} \cite[Bohnenblust and Karlin]{hfb1}. \label{29}
Let $\mathcal{D}$ be a nonempty subset of $X$, which is bounded, closed, and convex. Suppose $G:\mathcal{D}\rightarrow 2^X\setminus \{\emptyset\}$ is u.s.c. with closed, convex values, and such that $G(\mathcal{D})\subseteq \mathcal{D}$ and $G(\mathcal{D})$ is compact. Then $G$ has a fixed point.
\end{lemma}
\section{Approximate controllability results}\label{s3}
\noindent
In this section, first we establish a set of sufficient conditions for the approximate controllability for a class of second order evolution differential inclusions of the form \eqref{vi1}-\eqref{vi2} in Hilbert spaces by using Bohnenblust-Karlin's fixed point theorem. In order to establish the result, we need the following hypotheses:
\begin{enumerate}
\item [$\bf H_1$] $S_0(t)$, $t>0$ is compact.
\item[$\bf H_2$] For each positive number $r$ and $x\in \mathcal{C}$ with $\|x\|_\mathcal{C}\le r$, there exists $L_{f,r}(\cdot)\in L^1(I,\mathbb{R}^+)$ such
that
\begin{gather*}
\sup\{\|f\|:f(t)\in F(t,x(t))\}\le L_{f,r}(t),
\end{gather*}
for a.e. $t\in I$.
\item[$\bf H_3$] The function $s\rightarrow L_{f,r}(s)\in L^1([0,t],\mathbb{R}^+)$ and there exists a $\gamma>0$ such that
\begin{gather*}
\lim_{r\rightarrow \infty} \frac{\int_0^t L_{f,r}(s)ds}{r}=\gamma <+\infty.
\end{gather*}
\end{enumerate}
It will be shown that the system \eqref{vi1}-\eqref{vi2} is approximately controllable, if for all $a>0$, there exists a continuous function $x(\cdot)$ such that
\begin{align}
x(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)f(s)ds+\int_0^t S(t,s)Bu(s,x)ds,\quad f\in S_{F,x},\label{c7m1}\\
u(t,x)&=B^*S(b,t)R(a,\Upsilon_0^b)p(x(\cdot)),\label{c7m2}
\end{align}
where $$ p(x(\cdot))=x_b-C(b,0)x_0-S(b,0)y_0-\int_0^t S(b,s)f(s)ds.$$
\begin{theorem}\label{34}
Suppose that the hypotheses $\bf{H_0}$-$\bf{H_3}$ are satisfied. Assume also
\begin{align}
\widetilde{N}\gamma\Big[1+\frac{1}{\alpha}\widetilde{N}^2M^2_{B}b\Big]<1,\label{31}
\end{align}
where $M_B=\|B\|$, Then system \eqref{vi1}-\eqref{vi2} has a solution on $I$.
\end{theorem}
\begin{proof}
The main aim in this section is to find conditions for solvability of system \eqref{vi1}-\eqref{vi2} for $a>0$. We show that, using the control $u(x,t)$, the operator ${\Gamma}:\mathcal{C}\rightarrow 2^{\mathcal{C}}$, defined by
\begin{align*}
{\Gamma}(x)=&\Big\{\varphi\in \mathcal{C}:\varphi (t)=C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)[f(s)+Bu(s,x)]ds,\ f\in S_{F,x}\Big\},
\end{align*}
has a fixed point $x$, which is a mild solution of system \eqref{vi1}-\eqref{vi2}.
We now show that $\Gamma$ satisfies all the conditions of Lemma \ref{29}. For the sake of convenience, we subdivide the proof into five steps.
\noindent{\bf Step 1.} $\Gamma$ is convex for each $x\in \mathcal{C}$.
In fact, if $\varphi_1$, $\varphi_2$ belong to $\Gamma(x)$, then there exist $f_1$, $f_2\in S_{F,x}$ such that for each $t\in I$, we have
\begin{align*}
\varphi_i(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)f_i(s)ds+\int_0^t S(t,s)BB^{*}S^{*}(b,t) R(a,\Upsilon_0^b)\Big[x_b-C(b,0)x_0\\&-S(b,0)y_0-\int_0^b S(b,\eta)f_i(\eta)d\eta\Big](s) ds,\quad i=1,2.
\end{align*}
Let $\lambda\in [0,1]$. Then for each $t\in I$, we get
\begin{align*}
\lambda\varphi_1(t)+(1-\lambda)\varphi_2(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)[\lambda f_1(s) +(1-\lambda)f_2(s)]ds\\&+\int_0^t S(t,s)BB^{*}S^{*}(b,t) R(a,\Upsilon_0^b)\Bigg[x_b-C(b,0)x_0-S(b,0)y_0\\&-\int_0^b S(b,s)[\lambda f_1(s) +(1-\lambda)f_2(s)]ds\Bigg](s)ds.
\end{align*}
It is easy to see that $S_{F,x}$ is convex since $F$ has convex values. So, $\lambda f_1+(1-\lambda)f_2\in S_{F,x}$. Thus,
$$\lambda\varphi_1+(1-\lambda)\varphi_2\in \Gamma(x).$$
\noindent{\bf Step 2.} For each positive number $r>0$, let $\mathfrak{B}_r=\{x\in \mathcal{C}:\|x\|_\mathcal{C}\le r\}$. Obviously, $\mathfrak{B}_r$ is a bounded, closed and convex set of $\mathcal{C}$. We claim that there exists a positive number $r$ such that $\Gamma(\mathfrak{B}_r )\subseteq \mathfrak{B}_r$.
If this is not true, then for each positive number $r$, there exists a function $x^r\in \mathfrak{B}_r$, but $\Gamma(x^r)$ does not belong to $\mathfrak{B}_r$, i.e.,
$$\|\Gamma(x^r)\|_\mathcal{C}\equiv\sup\Big\{\|\varphi^r\|_\mathcal{C}:\varphi^r\in (\Gamma x^r)\Big\}>r$$
and
\begin{align*}
\varphi^r(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)f^r(s)ds+\int_0^t S(t,s)Bu^r(s,x) ds,
\end{align*}
for some $f^r\in S_{F,x^r}$. Using $\bf H_1$-$\bf H_3$, we have
\begin{align*}
r<&\|\Gamma(x^r)(t)\|\\
\le&\|C(t,0)x_0\|+\|S(t,0)y_0\|+\int_0^t \|S(t,s)f^r(s)\|ds+\int_0^t \|S(t,s)Bu^r(s,x)\|ds\\
\le& \Big[N\|x_0\|+\widetilde{N}\|y_0\|+\widetilde{N}\int_0^t L_{f,r}(s)ds\Big]\\& +\frac{1}{\alpha}\widetilde{N}^2M^2_{B}b\Bigg[\|x_b\|+N\|x_0\|+\widetilde{N}\|y_0\|+\widetilde{N}\int_0^b L_{f,r}(s)ds\Bigg].
\end{align*}
Dividing both sides of the above inequality by $r$ and taking the limit as $r\rightarrow \infty$, using $\bf H_3$, we get
\begin{gather*}
\widetilde{N}\gamma\Big[1+\frac{1}{\alpha}\widetilde{N}^2M^2_{B}b\Big]\ge 1.
\end{gather*}
This contradicts with the condition \eqref{31}. Hence, for some $r>0$, $\Gamma(\mathfrak{B}_r )\subseteq \mathfrak{B}_r$.
\noindent{\bf Step 3.} $\Gamma$ sends bounded sets into equicontinuous sets of $\mathcal{C}$. For each $x\in \mathfrak{B}_r$, $\varphi\in\Gamma(x)$, there exists a $f\in S_{F,x}$ such that
\begin{align*}
\varphi(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)f(s)ds+\int_0^t S(t,s)Bu(s,x)ds.
\end{align*}
Let $0<\varepsilon<0$ and $0<t_1<t_2\le b$, then
\begin{align*}
|\varphi(t_1)-\varphi(t_2)|=&|C(t_1,0)-C(t_2,0)||x_0|+|S(t_1,0)-S(t_2,0)||\eta|\\&+\Big|\int_{0}^{t_1-\varepsilon}[S(t_1,s)-S(t_2,s)]f(s) ds\Big| +\Big|\int_{t-\varepsilon}^{t_1}[S(t_1,s)-S(t_2,s)]f(s) ds\Big|\\&+\Big|\int_{t_1}^{t_2}S(t_2,s)f(s)ds\Big| +\Big|\int_{0}^{t_1-\varepsilon}[S(t_1,\eta)-S(t_2,\eta)]Bu(\eta,x) d\eta\Big|\\& +\Big|\int_{t-\varepsilon}^{t_1}[S(t_1,\eta)-S(t_2,\eta)]Bu(\eta,x) d\eta\Big|+\Big|\int_{t_1}^{t_2}S(t_2,\eta)Bu(\eta,x) d\eta\Big|\\&+\Big|\int_{0}^{t_1-\varepsilon}[S(t_1,\eta)-S(t_2,\eta)]Bu(\eta,x) d\eta\Big|\\
\le &|C(t_1,0)-C(t_2,0)||x_0|+|S(t_1,0)-S(t_2,0)||\eta|\\&+\int_{0}^{t_1-\varepsilon}|S(t_1,s)-S(t_2,s)|L_{f,r}(s) ds\\& +\int_{t-\varepsilon}^{t_1}|S(t_1,s)-S(t_2,s)|L_{f,r}(s) ds+\widetilde{N}\int_{t_1}^{t_2}L_{f,r}(s) ds\\& +M_B\int_{0}^{t_1-\varepsilon}|S(t_1,\eta)-S(t_2,\eta)|\|u(\eta,x)\| d\eta\\& +M_B\int_{t-\varepsilon}^{t_1}|S(t_1,\eta)-S(t_2,\eta)\|u(\eta,x)\| d\eta+\widetilde{N}M_B\int_{t_1}^{t_2}\|u(\eta,x)\| d\eta.
\end{align*}
The right-hand side of the above inequality tends to zero independently of $x\in B_r$ as $(t_1-t_2)\rightarrow 0$ and $\varepsilon$ sufficiently small, since the compactness of the evolution operator $S(t,s)$ implies the continuity in the uniform operator topology. Thus $\Gamma(x^r)$ sends $B_r$ into equicontinuous family of functions.
\noindent{\bf Step 4.} The set $\Pi(t)=\big\{\varphi(t):\varphi\in \Gamma(\mathfrak{B}_r)\big\}$ is relatively compact in $X$.
Let $t\in (0,b]$ be fixed and $\varepsilon$ a real number satisfying $0< \varepsilon<t$. For $x\in B_r$, we define
\begin{align*}
\varphi_\varepsilon(t)=C(t,0)x_0+S(t,0)y_0+\int_0^{t-\varepsilon}S(t,s)f(s)ds+\int_0^{t-\varepsilon}S(t,\eta)Bu(\eta,x)d\eta.
\end{align*}
Since $S_0(t)$ is a compact operator, the set $\Pi_\varepsilon(t)=\{\varphi_\varepsilon(t):\varphi_\varepsilon\in \Gamma(B_r)\}$ is relatively compact in $X$ for each $\varepsilon$, $0<\varepsilon<t$. Moreover, for each $0<\varepsilon<t$, we have
\begin{align*}
|\varphi(t)-\varphi_\varepsilon(t)|\le \widetilde{N}\int_{t-\varepsilon}^tL_{f,r}(s)ds+\widetilde{N}M_B\int_{t-\varepsilon}^t\|u(\eta,x)\|d\eta.
\end{align*}
Hence there exist relatively compact sets arbitrarily close to the set $\Pi(t)=\{\varphi(t):\varphi\in \Gamma(B_r)\}$, and the set $\widetilde{\Pi}(t)$ is relatively compact in $X$ for all $t\in [0,b]$. Since it is compact at $t=0$, hence $\Pi(t)$ is relatively compact in $X$ for all $t\in [0,b]$.
\noindent{\bf Step 5.} $\Gamma$ has a closed graph.
Let $x_n\rightarrow x_*$ as $n\rightarrow \infty$, $\varphi_n\in {\Gamma}(x_n)$, and $\varphi_n\rightarrow \varphi_*$ as $n\rightarrow \infty$. We will show that $\varphi_*\in {\Gamma}(x_*)$. Since $\varphi_n\in {\Gamma}(x_n)$, there exists a $f_n\in S_{F,x_n}$ such that
\begin{align*}
\varphi_n(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^tS(t,s)f_n(s)ds+\int_0^t S(t,s)BB^{*}S^{*}(b,t)
R(a,\Upsilon_0^b)\Big[x_b\\&-S(b,0)x_0-\int_0^b S(b,\eta)f_n(\eta)d\eta\Big](s)ds.
\end{align*}
We must prove that there exists a $f_*\in S_{F,x_*}$ such that
\begin{align*}
\varphi_*(t)=&C(t,0)x_0+S(t,0)y_0+\int_0^t S(t,s)f_*(s)ds+\int_0^t S(t,s)BB^{*}S^{*}(b,t)\Big[y_0-S(b,0)x_0\\&-\int_0^b S(b,\eta)f_*(\eta)ds\Big](s)ds.
\end{align*}
Set
\begin{gather*}
\overline{u}_x(t)=B^{*}S^{*}(b,t)[x_b-C(b,0)x_0-S(b,0)y_0](t).
\end{gather*}
Then
\begin{gather*}
\overline{u}_{x_n}(t)\rightarrow \overline{u}_{x_*}(t), \quad \mbox{for} \ t\in I, \ \mbox{as} \ n\rightarrow \infty.
\end{gather*}
Clearly, we have
\begin{align*}
\Big\|\Big(\varphi_n&-C(t,0)x_0+S(t,0)y_0-\int_0^t S(t,s)BB^{*}S^{*}(b,t)
R(a,\Upsilon_0^b)\Big[x_b-S(b,0)x_0 \\&-\int_0^b S(b,\eta)f_n(\eta)d\eta\Big](s)ds\Big)-\Big(\varphi_*-C(t,0)x_0+S(t,0)y_0-\int_0^t S(t,s)BB^{*}S^{*}(b,t)\Big[y_0\\&-C(b,0)x_0-S(b,0)y_0-\int_0^b S(b,\eta)f_*(\eta)ds\Big](s)ds\Big)\Big\|_\mathcal{C} \rightarrow 0 \ \mbox{as} \ n\rightarrow \infty.
\end{align*}
Consider the operator $\widetilde{\mathscr{F}}:L^1(I,X)\rightarrow \mathcal{C}$,
\begin{align*}
(\widetilde{\mathscr{F}}f)(t)=\int_0^t S(t,s)\Big[f(s)-BB^{*}S^{*}(b,t)\Big(\int_0^b S(b,\eta)f(\eta)d\eta\Big)(s)\Big]ds
\end{align*}
We can see that the operator $\widetilde{\mathscr{F}}$ is linear and continuous. From Lemma \ref{29} again, it follows that $\widetilde{\mathscr{F}}\circ S_F$ is a closed graph operator. Moreover,
\begin{align*}
\Big(\varphi_n&-C(t,0)x_0-S(t,0)y_0-\int_0^t S(t,s)BB^{*}S^{*}(b,t)
R(a,\Upsilon_0^b)\Big[x_b-S(b,0)x_0 \\&-\int_0^b S(b,\eta)f_n(\eta)d\eta\Big](s)ds\Big)\in \mathscr{F}(S_{F,x_n}).
\end{align*}
In view of $x_n\rightarrow x_*$ as $n\rightarrow \infty$, it follows again from Lemma \ref{29} that
\begin{align*}
\Big(\varphi_*&-C(t,0)x_0+S(t,0)y_0-\int_0^t S(t,s)BB^{*}S^{*}(b,t)
R(a,\Upsilon_0^b)\Big[x_b-S(b,0)x_0 \\&-\int_0^b S(b,\eta)f_*(\eta)d\eta\Big](s)ds\Big)\in \mathscr{F}(S_{F,x_*}).
\end{align*}
Therefore $\Gamma$ has a closed graph.
As a consequence of {\bf Steps 1-5} together with the Arzela-Ascoli theorem, we conclude that $\Gamma$ is a compact multivalued map, u.s.c. with convex closed values. As a consequence of Lemma \ref{29}, we can deduce that $\Gamma$ has a fixed point $x$ which is a mild solution of system \eqref{vi1}-\eqref{vi2}.
\end{proof}
\begin{definition}
The control system \eqref{vi1}-\eqref{vi2} is said to be approximately controllable on $I$ if $\overline{R(b)}=X$, where $R(b)=\{x(b;u):u^2L(I,U)\}$ is a mild solution of the system \eqref{vi1}-\eqref{vi2}.
\end{definition}
\begin{theorem}\label{40}
Suppose that the assumptions $\bf {H_0}$-$\bf {H_3}$ hold. Assume additionally that there exists $N\in L^1(I,[0,\infty))$ such that $\sup_{x\in X}\|F(t,x)\|\le N(t)$ for a.e. $t\in I$, then the nonlinear fractional differential inclusion \eqref{vi1}-\eqref{vi2} is approximately controllable on $I$.
\end{theorem}
\begin{proof}
Let $\widehat{x}^a(\cdot)$ be a fixed point of $\Gamma$ in $\mathfrak{B}_r$. By Theorem \ref{34}, any fixed point of $\Gamma$ is a mild solution of \eqref{vi1}-\eqref{vi2} under the control
\begin{align*}
\widehat{u}^a(t)=B^*S^*(b,t)R(a,\Upsilon_0^b)p(\widehat{x}^a)
\end{align*}
and satisfies the following inequality
\begin{align*}
\widehat{x}^a(b)=x_b+aR(a,\Upsilon_0^b)p(\widehat{x}^a).\label{kd1}
\end{align*}
Moreover by assumption on $F$ and Dunford-Pettis Theorem, we have that the $\{f^a(s)\}$ is weakly compact in $L^1(I,X)$, so there is a subsequence, still denoted by $\{f^a(s)\}$, that converges weakly to say $f(s)$ in $L^1(I,X)$. Define
\begin{gather*}
w=x_b-C(b,0)x_0-S(b,0)y_0-\int_0^b S(b,s)f(s)ds.
\end{gather*}
Now, we have
\begin{align}
\nonumber \|p(\widehat{x}^a)-w\|=&\Big\|\int_0^b S(b,s)[f(s,\widehat{x}^a(s))-f(s)]ds\Big\|\\
\le&\sup_{t\in I}\Big\|\int_0^t S(t,s)[f(s,\widehat{x}^a(s))-f(s)]ds\Big\|.\label{kd2}
\end{align}
By using infinite-dimensional version of the Ascoli-Arzela theorem, one can show that an operator $l(\cdot)\rightarrow \int_0^{\cdot} S(\cdot,s)l(s)ds:L^1(I,X) \rightarrow \mathcal{C}$ is compact. Therefore, we obtain that $\|p(\widehat{x}^a)-w\|\rightarrow 0$ as $a\rightarrow 0^+$. Moreover, from \eqref{kd1} we get
\begin{align*}
\|\widehat{x}^a(b)-x_b\|\le & \|aR(a,\Upsilon_0^b)(w)\|+\|aR(a,\Upsilon_0^b)\|\|p(\widehat{x}^a)-w\|\\
\le&\|aR(a,\Upsilon_0^b)(w)\|+\|p(\widehat{x}^a)-w\|.
\end{align*}
It follows from assumption $\bf{H_0}$ and the estimation \eqref{kd2} that $\|\widehat{x}^a(b)-x_b\|\rightarrow 0$ as $a\rightarrow 0^+$. This proves the approximate controllability of differential inclusion \eqref{vi1}-\eqref{vi2}.
\end{proof}
\section{Second-order control systems with nonlocal conditions} \label{s4}
\noindent
There exist an extensive literature of differential equations with nonlocal conditions. Since it is demonstrated that the nonlocal problems have better effects in applications than the classical ones, differential equations with nonlocal problems have been studied extensively in the literature. The result concerning the existence and uniqueness of mild solutions to abstract Cauchy problems with nonlocal initial conditions was first formulated and proved by Byszewski, see \cite{lb1, lb2, lb3}. Since the appearance of these papers, several papers have addressed the issue of existence and uniqueness of nonlinear differential equations. Existence and controllability results of nonlinear differential equations and fractional differential equations with nonlocal conditions has been studied by several authors for different kind of problems \cite{sa1, sa2, hrh3, ke1, xf1, jl1, nim2, vv5, xx1}.
Inspired by the above works, in this section, we discuss the approximate controllability for a class of second-order evolution differential inclusions with nonlocal conditions in Hilbert spaces of the form
\begin{align}
x''(t)\in&A(t)x(t)+F(t,x(t))+Bu(t), \quad t\in I=[0,b],\label{vn1}\\
x(0)+&g(x)=x_0\in X,\quad x'(0)+h(x)=y_0\in X,\label{vn2}
\end{align}
where $g,h:\mathcal{C}\rightarrow X$ is a given function which satisfies the following condition:
\begin{enumerate}
\item[$\bf H_4$] There exists a constant $L_g>0$, $L_h>0$ such that
\begin{align*}
|g(x)-g(y)|\le& L_g\|x-y\|,\ \text{for}\ x,y\in \mathcal{C}\\
|h(x)-h(y)|\le& L_h\|x-y\|,\ \text{for}\ x,y\in \mathcal{C}.
\end{align*}
for all $x,y\in \mathcal{C}$.
\end{enumerate}
\begin{definition}
A function $x\in \mathcal{C}$ is said to be a mild solution of system \eqref{vn1}-\eqref{vn2} if $x(0)+g(x)=x_0$, $x'(0)+h(x)=y_0$ and there exists $f\in L^1(I,X)$ such that $f(t)\in F(t,x(t))$ on $t\in I$ and the integral equation
\begin{align*}
x(t)=C(t,0)(x_0-g(x))+S(t,0)(y_0-h(x))+\int_0^tS(t,s)f(s)ds+\int_0^tS(t,s)Bu(s)ds,\ t\in I.
\end{align*}
is satisfied.
\end{definition}
\begin{theorem}\label{3.4}
Assume that the assumptions of Theorem \ref{34} are satisfied. Further, if the hypothesis $\bf H_4$ is satisfied, then the system \eqref{vn1}-\eqref{vn2} is approximately controllable on $I$ provided that
\begin{align*}
\widetilde{N}\gamma\Big[1+\frac{1}{\alpha}\widetilde{N}^2M^2_{B}b\Big]<1
\end{align*}
where $M_B=\|B\|$.
\end{theorem}
\begin{proof}
For each $a>0$, we define the operator $\widehat{\Gamma}_a$ on $X$ by
\begin{align*}
(\widehat{\Gamma}_a x)=z,
\end{align*}
where
\begin{align*}
z(t)&=C(t,0)(x_0-g(x))+S(t,0)(y_0-h(x))+\int_0^t S(t,s)f(s)ds+\int_0^t S(t,s)Bu(s,x)ds,\\
v(t)&=B^*S^*(b,t)R(a,\Upsilon_0^b)p(x(\cdot)),\\
p(x(\cdot))&=x_b-C(t,0)(x_0-g(x))-S(t,0)(y_0-g(x))-\int_0^t S(t,s)f(s)ds
\end{align*}
where $ f\in S_{F,x}$.
It can be easily proved that if for all $a>0$, the operator $\widehat{\Gamma}_a$ has a fixed point by implementing the technique used in Theorem \ref{34}. Then, we can show that the second order control system \eqref{vn1}-\eqref{vn2} is approximately controllable. The proof of this theorem is similar to that of Theorem \ref{34} and Theorem \ref{40}, and hence it is omitted.
\end{proof}
\section{Second-order impulsive nonlocal control systems} \label{s5}
\noindent
Impulsive dynamical systems are characterized by the occurrence of an abrupt change in the state of the system, which occur at certain time instants over a period of negligible duration. The dynamical behavior of such systems is much more complex than the behavior of dynamical systems without impulse effects. The presence of impulse means that the state trajectory does not preserve the basic properties which are associated with non-impulsive dynamical systems. In fact, the
theory of impulsive differential equations has found extensive applications in realistic mathematical modeling of a wide variety of practical situations \cite{bmm1} and has emerged as an important area of investigation in recent years. It is known that many biological phenomena involving thresholds, bursting rhythm models in medicine and biology, optimal control models in economics, pharmacokinetics and frequency modulation systems, do exhibit impulse effects; see the monographs of Samoilenko and Perestyuk \cite{ams1}, Lakshmikantham et al. \cite{vl1}, Bainov et al. \cite{ddb1} and Benchohra et al. \cite{mb1} and the references cited therein. The literature on this type of problem is vast, and different topics on the existence and qualitative properties of solutions are considered. Concerning the general motivations, relevant developments and the current status of the theory, we refer the reader to \cite{na1, ga1, ga2, eh1, hrh2, hrh3, mk1, yp1, rs1, rs2, ss1, vv3, vv1}.
Inspired by the above works, in this section, we discuss the approximate controllability for a class of second-order impulsive evolution differential inclusions with nonlocal conditions in Hilbert spaces of the form
\begin{align}
x''(t)\in&A(t)x(t)+F(t,x(t))+Bu(t), \quad t\in I=[0,b],\label{vni1}\\
x(0)+&g(x)=x_0\in X,\quad x'(0)+h(x)=y_0\in X,\label{vni2}\\
\Delta x(t_i)=&I_i(x(t_i)),\label{vni3}\\
\Delta x'(t_i)=&J_i(x(t_i)),\quad i=1,2,\cdots,n,\label{vni4}
\end{align}
where $0<t_1<t_2<\cdots<t_n<b$ are fixed numbers and $I_i:X\rightarrow X$, $J_i:X\rightarrow X$, $i=1,...,n$, are suitable functions and the symbol $\Delta \xi(t)$ represents the jump of the function $\xi$ at $t$, which is defined by $\Delta \xi(t)=\xi(t^+)-\xi(t^-)$.
Consider the following assumption
\begin{enumerate}
\item[$\bf H_5$] The maps $I_i:X\rightarrow X$, $J_i:X\rightarrow X$ are completely continuous and uniformly bounded. In the sequel, we set $M_i=\sup\{\|I_i(x)\|:x\in X\}$ and $\widetilde{M}_i=\sup\{\|J_i(x)\|:x\in X\}$.
\end{enumerate}
\begin{definition}
A function $x\in \mathcal{C}$ is said to be a mild solution of system \eqref{vni1}-\eqref{vni4} if $x(0)+g(x)=x_0$, $x'(0)+h(x)=y_0$ and there exists $f\in L^1(I,X)$ such that $f(t)\in F(t,x(t))$ on $t\in I$ and the integral equation
\begin{align*}
x(t)=&C(t,0)(x_0-g(x))+S(t,0)(y_0-h(x))+\int_0^tS(t,s)f(s)ds+\int_0^tS(t,s)Bu(s)ds\\&+\sum_{0<t_i<t} C(t,t_i)I_i(u(t_i))+\sum_{0<t_i<t} S(t,t_i)J_i(u(t_i)),\quad t\in I.
\end{align*}
is satisfied.
\end{definition}
\begin{theorem}
Assume that the assumptions of Theorem \ref{34} and Theorem \ref{3.4} are satisfied. Further, if the hypothesis $\bf H_5$ is satisfied, then the system \eqref{vni1}-\eqref{vni4} is approximately controllable on $I$ provided that
\begin{align*}
\widetilde{N}\gamma\Big[1+\frac{1}{\alpha}\widetilde{N}^2M^2_{B}b\Big]<1
\end{align*}
where $M_B=\|B\|$.
\end{theorem}
\begin{proof}
For each $a>0$, we define the operator $\widehat{\Gamma}_a$ on $X$ by
\begin{align*}
(\widehat{\Gamma}_a x)=z,
\end{align*}
where
\begin{align*}
z(t)=&C(t,0)(x_0-g(x))+S(t,0)(y_0-h(x))+\int_0^t S(t,s)f(s)ds+\int_0^t S(t,s)Bu(s,x)ds\\&+\sum_{0<t_i<t} C(t,t_i)I_i(u(t_i))+\sum_{0<t_i<t} S(t,t_i)J_i(u(t_i)),\\
v(t)=&B^*S^*(b,t)R(a,\Upsilon_0^b)p(x(\cdot)),\\
p(x(\cdot))=&x_b-C(t,0)(x_0-g(x))-S(t,0)(y_0-g(x))-\int_0^t S(t,s)f(s)ds\\&+\sum_{i=1}^n C(t,t_i)I_i(u(t_i))+\sum_{i=1}^n S(t,t_i)J_i(u(t_i)),
\end{align*}
where $ f\in S_{F,x}$.
It can be easily proved that if for all $a>0$, the operator $\widehat{\Gamma}_a$ has a fixed point by implementing the technique used in Theorem \ref{34}. Then, we can show that the second order control system \eqref{vni1}-\eqref{vni4} is approximately controllable. The proof of this theorem is similar to that of Theorem \ref{34}, Theorem \ref{40} and Theorem \ref{3.4}, and hence it is omitted.
\end{proof}
\section{An application}\label{s6}
\noindent
In this section, we apply our abstract results to a concrete partial differential equation. In order to establish our results, we need to introduce the required technical tools. Following the equations $\eqref{vp7}-\eqref{vp8}$, here we consider $A(t)=A+\widetilde{B}(t)$ where $A$ is the infinitesimal generator of a cosine function $C_0(t)$ with associated sine function $S_0(t)$, and $\widetilde{B}(t):D(\widetilde{B}(t))\rightarrow X$ is a closed linear operator with $D\subseteq D(\widetilde{B}(t))$ for all $t\in I$.
We model this problem in the space $X=L^2(\mathbb{T},\mathbb{C})$, where the group $\mathbb{T}$ is defined as the quotient $\mathbb{R}/2\pi Z$. We will use the identification between functions on $\mathbb{T}$ and $2\pi$-periodic functions on $\mathbb{R}$. Specifically, in what follows we denote by $L^2(\mathbb{T},\mathbb{C})$ the space of $2\pi$-periodic 2-integrable functions from $\mathbb{R}$ into $\mathbb{C}$. Similarly, $H^2(\mathbb{T},\mathbb{C})$ denotes the Sobolev space of $2\pi$-periodic functions $x:\mathbb{R}\rightarrow \mathbb{C}$ such that $x''\in L^2(\mathbb{T},\mathbb{C})$.
We consider the operator $Ax(\xi)=x''(\xi)$ with domain $D(A)=H^2(\mathbb{T},\mathbb{C})$. It is well known that $A$ is the infinitesimal generator of a strongly continuous cosine function $C_0(t)$ on $X$. Moreover, $A$ has discrete spectrum, the spectrum of $A$ consists of eigenvalues $-n^2$ for $n\in \mathbb{Z}$, with associated eigenvectors $$w_n(\xi)=\frac{1}{\sqrt{2\pi}}e^{in\xi},\quad n\in \mathbb{Z},$$ the set $\{w_n:n\in Z\}$ is an orthonormal basis of $X$. In particular,
$$Ax=-\sum_{n=1}^\infty n^2\langle x, w_n\rangle w_n $$ for $x\in D(A)$. The cosine function $C_0(t)$ is given by $$C_0(t)x=\sum_{n=1}^\infty\cos(nt)\langle x, w_n\rangle w_n,\quad t\in \mathbb{R},$$ with associated sine function $$S_0(t)x=\sum_{n=1}^\infty \frac{\sin(nt)}{n}\langle x, w_n\rangle w_n,\quad t\in \mathbb{R}.$$
It is clear that $\|C_0(t)\|\le 1$ for all $t\in \mathbb{R}$. Thus, $C(\cdot)$ is uniformly bounded on $\mathbb{R}$.
Consider the second-order Cauchy problem with control
\begin{align}
\frac{\partial^2}{\partial t^2}z(t,\tau)\in \frac{\partial^2}{\partial \tau^2}z(t,\tau)+b(t)\frac{\partial}{\partial t}z(t,\tau)+\mu(t,\tau)+Q(t,z(t,\tau))\label{vex1}
\end{align}
for $t\in I$, $0\le\tau\le \pi$, subject to the initial conditions
\begin{align}
z(t,0)=&z(t,\pi)=0,\quad t\in I,\label{vex2}\\
z(0,\tau)=&z_0(\tau), \quad 0\le\tau\le \pi,\quad \frac{\partial}{\partial t}z(0,\tau)=z_1(\tau),\label{vex3}
\end{align}
where $b:\mathbb{R}\rightarrow \mathbb{R}$, $\mu:I\times [0,\pi]\rightarrow [0,\pi]$ are continuous functions. We fix $a>0$ and set $\beta=\sup_{0\le t\le a}|b(t)|$.
We take $\widetilde{B}(t)x(\tau)=b(t)x'(\tau)$ defined on $H^1(\mathbb{T},\mathbb{C})$. It is easy to see that $A(t)=A+\widetilde{B}(t)$ is a closed linear operator. Initially we will show that $A+\widetilde{B}(t)$ generates an evolution operator. It is well known that the solution of the scalar initial value problem
\begin{align*}
q''(t)=&-n^2q(t)+p(t),\\
q(s)=&0,\quad q'(s)=q_1,
\end{align*}
is given by $$q(t)=\frac{q_1}{n}\sin n(t-s)+\frac{1}{n}\int_s^t\sin n(t-\tau)p(\tau)d\tau.$$
Therefore, the solution of the scalar initial value problem
\begin{align}
q''(t)=&-n^2q(t)+in b(t)q(t),\label{vs5e6}\\
q(s)=&0,\quad q'(s)=q_1,\label{vs5e7}
\end{align}
satisfies the integral equation $$q(t)=\frac{q_1}{n}\sin n(t-s)+ i\int_s^t\sin n(t-\tau)b(\tau)q(\tau)d\tau.$$
Applying the Gronwall-Bellman lemma we can affirm that
\begin{align}
|q(t)|\le \frac{|q_1|}{n}e^{\beta(t-s)}\label{vs5e8}
\end{align}
for $s\le t$. We denote by $q_n(t,s)$ the solution of $\eqref{vs5e6}-\eqref{vs5e7}$. We define $$S(t,s)x=\sum_{n=1}^\infty q_n(t,s)\langle x, w_n\rangle w_n.$$
It follows from the estimate $\eqref{vs5e8}$ that $S(t,s):X\rightarrow X$ is well defined and satisfies the conditions of Definition \ref{vd1}.
Let us define$f:I\times X\rightarrow X$ be defined as $$F(t,z)(\tau)=Q(t,z(\tau)),\quad z\in X,\quad \tau\in [0,\pi].$$ and $u:I\rightarrow U$ be defined as $$B(u(t))(\tau)=\mu(t,\tau),\quad \tau\in [0,\pi],$$ where $\mu:I\times [0,\pi]\rightarrow [0,\pi]$ is continuous.
Assume these functions satisfy the requirement of hypotheses. From the above choices of the functions and evolution operator $A(t)$ with $B=I$ , the system \eqref{vex1}-\eqref{vex3} can be formulated as an abstract second order semilinear system \eqref{vi1}-\eqref{vi2} in $X$. Since all hypotheses of Theorem \ref{40} are satisfied, approximate controllability of system \eqref{vex1}-\eqref{vex3} on $I$ follows from Theorem \ref{40}.
|
\section{Foreword written by Yves Pomeau}
\label{foreword}
The present paper reports original work which seems to me
a very appropriate topic for this volume honoring Martine Ben Amar. Its matter is related to my first (very successful) collaboration with her. Our first joint paper \cite{joint}
studied the growth of needle crystals in an undercooled melt. At the time it was well understood that the classical Ivantsov solution for the needle crystal had to be supplemented by physical effects outside of the ones in Ivantsov theory to yield a solution which would be both unique and pertinent for the observations. It had long been suspected that surface tension, through the Gibbs-Thomson curvature-dependence of the equilibrium temperature, had to be taken into account. But, at the time, it was more or less an article of faith. With Martine, we did show that things did not work as believed by many in the field: no solution exists with isotropic solid-melt surface tension. Crystal anisotropy had to be taken into account \cite{joint}. This advance was followed by many others thanks to Martine and her collaborators so that this tricky problem can now be considered as solved.
The connection with the present work is that, as in the 1986 Europhys. Lett paper, we have to introduce
Stefan condition for melting to understand quantitatively how a supply of heat can explain the dynamics of the solid/liquid interface. Ice skating relies on the same dynamics, and as the literature shows too well, it is a quite non trivial matter to write the appropriate boundary conditions. It is quite amazing to come back to this "classical", but somewhat forgotten, physics after so many years.
I wish you Martine, many more years of active and fruitful research.
\section{Introduction}
\label{intro}
Skating is possible because of the very small friction felt by a thin blade sliding on ice, in sharp contrast with much larger solid-on-solid friction observed at temperatures well below the melting temperature. This has long been explained \cite{reynolds} by the existence of a thin lubricating layer of liquid water. Standard wisdom was that ice melts because of the local increase of pressure by the weight of the skater, a pressure leading to local melting because of the pressure dependence of equilibrium melting temperature. But this melting is an equilibrium phenomenon, then it is unclear how it can describe the continuous formation of liquid water from ice, in particular because this requires a supply of heat. According to another explanation \cite{A} the main phenomenon is melting by the heat generated by friction in-between ice and the sliding skate. This heat yields the energy needed to balance the latent heat, as given by Stefan condition on the ice surface. A condition of mechanical equilibrium expresses also how the weight of the skater is supported by the pressure generated inside the film, which avoids direct ice-skate contact.
We reconsider below this question in the light of a macroscopic approach, assuming that the liquid layer in-between ice and skate has macroscopic thickness. Even though this layer is often referred to as a lubricating layer, it acts differently of a regular lubricant: lubricants like oil are chemically different of the solids facing each other. Therefore a layer of molecular thickness can change the conditions of sliding. On the contrary, in the present case, the lubricating layer is made of the same molecules of water as solid ice, so that a layer of molecular thickness, which is about one angstrom for water, will be mixed with the molecules present at the surface of ice, and cannot change the ice-skate friction.
Our analysis is inspired by previous work on the Leidenfrost effect \cite{leiden} where a thin layer of vapor was also analyzed and where the balance of vertical forces (weight of the airborne Leidenfrost drop and pressure forces) plays a central role. A difference between the skating problem and the Leidenfrost effect is that in Leidenfrost the evaporation from the droplet (instead of the melting in skating) is due to the heating by the hot bottom plate although in skating the heat comes from viscous friction in the bulk of the lubricating layer. Here it is the longitudinal viscous friction (along the direction of motion) which is responsible for heating the film of water, although this longitudinal flow does not generate any vertical stress in the layer. The weight of the skater is actually supported by the transverse Stefan flow which undergoes a drop of pressure bigger than in the longitudinal direction in the skating case (this is not the case treated in \cite{A}, which considers a 2-D situation).
We answer a very natural question: when a sharp edge slides on ice, how deep is the furrow due to melting arising itself from the heat generated by viscous friction? This introduces a length scale (the depth of the furrow) which turns out to be much bigger than the thickness of the liquid layer due to melting, a layer in-between ice and skate. To estimate this contribution of melting one needs to find out how much water is made during the time a skate passes over a point on the ice surface, and so how much ice has melt to make the furrow. We show that this "lubricating" layer has macroscopic thickness in concrete situations, but that it is far thinner than the depth of the furrow, an important remark. Furthermore, as written above, the depth of the furrow is far smaller than its length which is far smaller than than the radius of profile of the blade. Hence one has to solve a problem with four different length scales, each one with a different order of magnitude than the the three others, see equation (\ref{eq:inegalfa}) referring to the lengths reported in Fig.\ref{fig:sec3}.
Here we present a theory using in a consistent way Stefan description of the melt/solid interface.
This is not a trivial endeavor, because one has to write the equations for fluid mechanics in the frame of reference where the boundaries of this film are fixed, namely in the frame of reference of the skater. Therefore when writing Stefan condition on the surface of ice, we must consider the interface as a moving surface with respect to ice \textit{and} the ice as moving with respect to the skater. It follows that the ice melting rate cannot be identified to the growth rate of the film thickness, as done in the recent model derived by Lozowski et al.\cite{LScana}-\cite{LScana2}. We show below that the missing term in their model is actually of prime importance.
By combining (i) Stephan condition with (ii) the balance between the weight of the skater and the viscous pressure force in the liquid layer, and (iii) the relation between the volume of the trough and the melting process during the passage of the skater, we derive two coupled integro-differential equations relating the film thickness to the length of contact. All physical lengths of the problem are then deduced from the data, in particular the length of contact between ice and skate, a parameter which cannot be set to its value at rest as done in \cite{LScana}-\cite{LScana2}.
The study of these coupled equations allows to define a small a-dimensional quantity depending on the longitudinal coordinate (in the direction of the motion) which can be neglected everywhere except close to the contact points at the front and the end of the blade. Help to this property we show that the length of contact can be deduced from the integral equation with a very good approximation. Then it remains to solve a single differential equation for the film thickness. From the analytical study of the two boundary layers at the front and at the end of the blade, we show that one can deduce an expression of the order of magnitude of the film thickness in these two regions in terms of the data. This provides without any calculation the relation between the unknown variables and the external parameters like the velocity and mass of the skater and the radius of profile and bite angle of the blade, in good agreement with the numerical study.
In section \ref{sec:hockey} we derive the model equations for standard ice-skating conditions, namely for V-shaped blade transverse profile, and study in details the solution analytically and numerically with applications to hockey and inclined speed- skating blades. Because of the existence of a small parameter we define a critical mass below which no skating is possible. This critical mass is so small that one may conclude that a macroscopical film of water is always formed in the skating case, a conclusion not valid for cavitation of water droplet in the Leidenfrost effect \cite {leiden}. The case of a vertical speed-skating or rectangular transverse blade profile is treated in section \ref{sec:rectangular} which gives results very similar to the ones for V-shaped blades, although with mote complex analytical expressions. Last section is devoted to conclusions.
\section{ Equations for V-shaped blades}
\label {sec:hockey}
To get close to ice-skating conditions, we consider first a skate having a smooth surface ending toward the ice with a sharp edge, its longer dimension being in the direction $y$ of the imposed speed, see fig.\ref{fig:sec3}. Such V-shaped blades are the ones sharpened for hockey skaters and also figure skatings. The bottom of each blade present actually a hollow separating two V-shaped edges but the skater spends most of his time on a single edge, therefore we focus here on the case of V-shape transverse profile on contact with the ice surface via a thin film of melted water. In this geometry the water layer has only a finite extend along the horizontal coordinate $x$ perpendicular to the direction of motion, a width much smaller than the length of the skate. Lastly the vertical direction is associated to coordinate $z$. The curved profile of the blade and the cross-section are schematized in figs.\ref{fig:sec3} where the length scales are not respected. The length of contact $\ell$ is defined as usual, as the longitudinal dimension of the wetted part of the blade which is situated between the first point of contact (where the thickness of the layer is zero) and the point of maximum penetration of the blade into the ice. Beyond this point the contact is assumed to stop because the bottom of the curved blade becomes above the surface of the through dug by the passage of the skater.
This geometry also describes the case of speed skating (having rectangular cross-section) with inclined blade making an angle $\pi/4$ with the vertical axis, considered in subsection \ref{sec:numerics}.
We neglect the fact that sharpening of the blade leaves a slightly rounded cylindrical wedge with a very small radius. This is a reasonable assumption because the water layer has a thickness of order of thirteen micrometers, although blades are sharpened with a radius of order of one micrometer. Therefore we consider a perfectly sharpened blade with V-shaped wedge. This hypothesis may be invalid in a very small domain (in front of the skate where the liquid layer is very shallow), that is not essential for our purpose which is to derive the geometry of the furrow whose main part has length scales much larger than the radius of curvature or the imperfections of sharpened wedges.
We derive below the equations for the flow created by the friction forces along $y$ which is partially evacuated laterally by a secondary flow flowing in the $x$ direction (perpendicular to the plane of Fig.(a)) generating a pressure supporting the weight of the skater. Note that there is also some flow ejecting fluid behind the blade which is generated by the pressure gradient along $y$. This gradient is negligible with respect to the lateral one because the length of contact along $y$ is much larger than the transverse dimension of the trough created by the passage of the blade.
\begin{figure}
{
(a) \includegraphics[height=1.25in]{schema1.eps}
(a) \includegraphics[height=1.25in]{schema2.eps}
}
\caption{ Schema of the blade digging into the ice surface. In (a) we set $y=0$ at the front contact point and $y=\ell$ at the back contact point where the letters $\epsilon$ and $\alpha$ are placed. (b) Schema of the cross-section at a given value of $y$. The geometry is deformed in order to see the various lengths.
}
\label{fig:sec3}
\end{figure}
\subsection{Derivation of the model with Stefan condition}
\label {sec:stefan sol}
Below we write Stefan equation which links the velocity of the flow (or melting rate, normal to the ice surface) and the film thickness via the viscosity, the surface in contact and the latent heat of the ice. This equation set the balance between the heat generated by viscous friction in the liquid layer of melt water separating solid ice from the skate and the latent heat melting ice. This viscous friction is due to the fact that the layer has a small thickness $ \alpha$ and that across this thickness the $y$-component of the velocity goes from $V$ on the surface of the skate to zero on solid ice. This yields a viscous stress and so a source of heat in the liquid layer. For the moment we assume that all this heat is used to melt ice.
Consider the slice $y,y+dy$. The film of water in this slice has a volume $\delta'(y) \alpha(y) dy$, where $\alpha(y)$ is the thickness of the film and $\delta'(y)$ is the length of the wetted part of the blade, represented in Fig.\ref{fig:sec3}-(b) as the length of the two sides of a triangle. The power dissipated by friction in the film of water of volume $\delta'(y) \alpha(y) dy$ is equal to
$dP= \eta \frac{V^2}{\alpha}\delta'(y)dy$, where $\eta$ is the shear viscosity of water.
In the cross section we define local coordinates $(s,n)$, not represented in the figure, with $-\frac{\delta'}{2} \le s \le +\frac{\delta'}{2}$ along the ice surface, and $ 0 \le n \le \alpha$ perpendicular to the surface, $n=\alpha$ on the skate. We assume that the thickness $\alpha(y)$ of the film is independent of $s$ in a given cross-section (an assumption consistent with the rest of the calculation) and that
\begin{equation}
\alpha(y) \ll \epsilon \ll \ell \ll r,
\label{eq:inegalfa}
\end{equation}
where $r$ is the radius of profile of the slightly curved blade which is of several meters or more (the curvature is intentionally exaggerated in the schema (b)).
This condition implies
$\alpha << \delta'$, because the bite angle $\theta$ is typically around $\frac{\pi}{2}$. This power $dP$ is responsible for the melting of a volume $ d\mathbb{V} =w_0(y)\delta'(y)dy$ , where $w_0(y)$ is the melt rate. We assume that $w_0(y)$ is independent of the local coordinate $s$.
The power $dP$ dissipated by the shear stress is equal to the power required to melt the volume $ d\mathbb{V}$, then one has $dP= L \rho_{s} d\mathbb{V}$ where $ \rho_{s}$ is the density of the ice (at the surface). It gives Stefan equation valid when the heat losses are neglected (for perfectly insulating blades)
\begin{equation}
\alpha(y) \vert w_0(y) \vert =\eta\frac{V^2}{ L \rho_s}.
\label{eq:wo1a}
\end{equation}
In the following we shall neglect the difference of density between ice and water to simplify the analysis, a minor approximation in regards with others (like the assumption of constant film thickness in the cross-section which will be assumed below), our aim being to get in fine an order of magnitude of the geometry of the trough and of the film, and ultimately to define the conditions for skating with small friction.
Note that the melt rate $w_{0}$ cannot be identified to $d\alpha /dt$, see the discussion below equation (\ref{eq:wosk}).
\subsubsection{Heat transfer in the film}
The equation (\ref{eq:wo1a}) is a possible writing of Stefan condition. It expresses the balance between the power input coming from the viscous stress and the latent heat needed to melt ice. This is an approximation only, because it short cuts the heat transfer problem in the liquid layer. If the matter of the skate had zero heat conduction, this formula were exact, because the only place the heat due to viscous friction could go is to ice in order to melt it. Consider the opposite limit where the matter of the skate is a perfect heat conductor. In this case one has also to assume something about the temperature of the skate. To simplify the matter let us take it as equal to the melting temperature. In this case, the temperature in the liquid layer will be symmetrical with respect to the mid plane of the layer, so that the heat flux toward the surface of the skate will be the same as the one toward the ice. In this case only half of the heat generated by viscous friction will be used to melt ice.
Moreover there is another possible effect, due to ice losses, because ice has to be heated to the melting temperature before to melt. As well known since Nansen \cite{Nansen}, this has a strong influence on sliding if ice is at very low temperature: then the viscous friction has first to heat ice and then to melt it. Because of the quite large latent heat of ice, this is a negligible effect in the condition of ice-skating in arenas, where the temperature of ice is just a little below melting. Therefore in this case, one can assume that all the heating is used to melt ice, except of course that it is also used to heat the skate itself. In the case of a skate conducting heat much better than water, as argued above only half of the heat is used to melt ice.
In the following we consider
a blade made of perfect heat conductor, that amounts to divide by $2$ the quantity of heat generated by the work of viscous forces and used for melting. With $ \rho_s=1$, Stefan relation (\ref{eq:wo1a}) becomes
\begin{equation}
\alpha(y) \vert w_0(y) \vert =\eta\frac{V^2}{ 2L }.
\label{eq:wo1}
\end{equation}
\subsubsection{Poiseuille flow description}
Neglecting the inertia of the flow (assumed to be at low Reynolds number), we can use the analytical solution of a Poiseuille flow. At given $y$, the gradient of the pressure $p$ in the film along the coordinate $s$ obeys the equation
\begin{equation}
\frac{\partial p}{\partial s}= \eta \frac{\partial^2 u}{\partial n^2},
\label{eq:dp1}
\end{equation}
where $u$ is the component of velocity in the film along $s$. Equation (\ref{eq:dp1}) has to be solved with the boundary conditions $\left(\frac{\partial u}{\partial n}\right)_{n=\alpha/2} =0$, $u(0)=u(\alpha)= u_{0} = -V\cos(\theta/2) \frac{\partial \epsilon_{sk}}{\partial y}$ which is due to the geometry of the skate sunk into the furrow, $\epsilon_{sk}$ being the penetration depth of the blade at given $y$. The solution is
\begin{equation}
u(s,n)= \frac{1}{2\eta}\frac{\partial p}{\partial s}(n^2-n\alpha) +u_{0}.
\label{eq:u1}
\end{equation}
The components $u$ and $w$ of the velocity in the cross-section are linked by the condition of incompressibility ($w$ is the component in the direction $n$, normal to the surface of the blade). Assuming that the depth of the trough (of order $\delta'$) and the thickness $\alpha$ of the film are much smaller than the length of contact $\ell$, we can neglect the gradient of the pressure in the direction $y$ of the skater motion as compared to the two other components of this gradient. The conservation of mass condition reduces to
\begin{equation}
\frac{\partial u}{\partial s}+ \frac{\partial w}{\partial n} =0.
\label{eq:u2}
\end{equation}
Inserting the solution (\ref{eq:u1}) in (\ref{eq:u2}) and integrating along $n$ in the film, we obtain
\begin{equation}
\frac{\partial^2 p}{\partial s^2}= \frac{12 \eta\Delta w(y)}{\alpha^3},
\label{eq:d2p}
\end{equation}
where
\begin{equation}
\Delta w(y)= w(y,n=\alpha) -w(y,n=0)
\label{eq:wosk}
\end{equation}
results from the integral over $n$ of the Stefan flow. The expression of $\Delta w(y)$ deserves some comments and requires to explain the meaning of the different velocities under consideration.
Generally Stefan condition expresses a constraint on the fluid velocity normal to an interface where there is a phase transformation. There is a subtle point as this condition depends if it is written in the frame of reference of the interface or in the frame of the solid that is melting. To simplify the discussion let us consider the case, approximately true for ice melting, where there is no change of density from solid to liquid, as assumed here. Let us put oneself in the frame of reference where neither liquid water nor solid ice move, but where ice melts so that the surface of separation moves at speed $w$ with respect to the frame of both ice and water (which is possible if one neglects their density difference). In this frame the fluid velocity normal to the surface of ice will be zero, because the phase change does not transport any matter. The consequence of that is that, in the frame of reference of the ice surface, which is moving at speed $-w$ with respect to the frame of reference just considered, the fluid velocity near the surface must be also $w$ because of the Galilean transform necessary to put the fluid in the moving frame of reference of the ice surface. Hence in this frame one must impose a Stefan condition to the normal velocity $w$ of the fluid on the ice surface, or with our notation $ w(y,0)= w_{0}(y)$.
On the contrary, on the surface of the skate, there is no phase transformation and the boundary condition in the frame of reference of the skate is the regular condition (for a viscous fluid) that all components of the fluid speed vanish on the surface of the skate in the frame of reference where it is at rest. Finally, in the skate frame, equation (\ref{eq:wosk}) writes
\begin{equation}
\vert \Delta w(y) \vert= w_{0}(y),
\label{eq:deltaw}
\end{equation}
a quantity which was defined as the rate of melting and cannot be identified to $\frac{d\alpha}{dt}$, see relation (\ref{eq:wo2}).
Assuming the right-hand side of (\ref{eq:d2p}) independent of $s$, integrating twice over $s$ and taking into account the boundary conditions $(\frac{\partial p}{\partial s})_0=0$ and $p(\delta'/2)=0$ we obtain
the drop of pressure along the blade in the cross section
\begin{equation}
p(s)= \frac{3}{2}\eta (\Delta w/\alpha^3)((\delta'_{sk})^{2} - 4s^2),
\label{eq:pr}
\end{equation}
where $\delta'_{sk}$ is the length of the wetted part of the blade in the cross-section, equal to
$\frac {2 \epsilon_{sk} }{\cos(\theta/2)} $,
$\epsilon_{sk}$ being given as a function of $y$ in equation (\ref{eq:epsk}).
Along the vertical direction $z$ the weight of the skater must be the opposite of the integral over $s$ of the vertical component of the pressure. This yields the condition
\begin{equation}
\int_{0}^{\ell}{ \eta \sin(\theta/2){ \Delta w(y)} \left(\frac{\delta'_{sk}(y)}{\alpha(y)}\right)^3}dy = M g ,
\label{eq:poids}
\end{equation}
where $\ell$ is the length of contact, $M$ the mass of the skater sliding on a single wedge of the blade (and assuming that its weight is supported by a single skate) and $g$ the acceleration of gravity.
Now we have to describe the formation of the trough, that provides a relation connecting $w_{0}$ and $ V \frac{ \partial {\epsilon}}{ \partial {y}}$ . The total area of the melted water in the V-shaped cross-section at $y$ is equal to $\frac{1}{2} \epsilon(y)\delta(y)$. This area results from the passage of the skater and the ensuing melting during the time $\delta t= y/V$. Therefore
\begin{equation}
\mathcal{A}(y)= \frac{1}{2} \epsilon(y) \delta(y)= \int_{0}^{y} {dy' w_{0}(y' ) \delta '(y')/V },
\label{eq:melt}
\end{equation}
or in a differential form, (after setting $\delta'(y)=\frac {2\epsilon(y)} {\cos (\theta/2)} $ and $\delta(y)=2\epsilon(y) \tan (\theta/2) $)
\begin{equation}
w_{0}(y)= \frac{\partial \epsilon}{\partial y} V\sin{\theta}/2,
\label{eq:meltdiff}
\end{equation}
where $\epsilon(y)$ may be be written as the sum of two terms as $\epsilon(y) =\epsilon_{sk}(y)+\frac{\alpha(y)}{\sin(\theta/2)}$. In this formula $\epsilon_{sk}$ is the penetration depth of the blade at a given $y$ in the frame of reference of the skate and $\frac{\alpha(y)}{\sin(\theta/2)}$ is for the vertical height of the liquid layer, a layer which originates also from the melting of ice.
\subsubsection{Length of contact and film thickness }
Using equations (\ref{eq:wo1})-(\ref{eq:poids})-(\ref{eq:meltdiff}) we can derive now the geometry and the components of the velocity in the film below the skate. Let us first calculate the length of contact $\ell$ using equation (\ref{eq:poids}), that requires to express $\Delta w(y)$ and $\frac{\delta' (y)}{\alpha(y)}$.
The blades of hockey and figure skates are curved along the propagation direction, in the plane $(z,x)$, with a radius of profile $r$.
This radius of profile being much larger than the length of contact (for hockey skating $r $ is generally between $2m$ and $4m$), the penetration depth of the blade writes
\begin{equation}
\epsilon_{sk}(y)= \frac{ y(2\ell-y)}{2r }.
\label{eq:epsk}
\end{equation}
In a given cross-section the wetted length of the blade is $\delta'_{sk}= \frac{ y(2\ell-y)}{r \cos(\theta /2)}$. Defining $\vert w_{sk}\vert =V \sin{(\theta/2)}\frac{\partial \epsilon_{sk}}{\partial y}$, which could be seen as the normal velocity of the blade with respect to an observer at rest (or the inverse), we have
\begin{equation}
\vert w_{sk}(y) \vert = V\sin(\theta/2)\frac{ l-y}{r },
\label{eq:wsk}
\end{equation}
and the flow velocity $w_{0} $ writes as
\begin{equation}
w_{0}(y)= \vert w_{sk}(y) \vert + V \frac{d\alpha}{dy},
\label{eq:wo2}
\end{equation}
Introducing equations (\ref{eq:wsk})-(\ref{eq:wo2}) in Stefan relation (\ref{eq:wo1}), gives the differential equation for the film thickness $\alpha$
\begin{equation}
\alpha \left( \sin(\theta/2 )\frac{\ell -y}{r } +\frac{d\alpha}{dy} \right )= k ,
\label{eq:wo3}
\end{equation}
which has to be solved with the initial condition $\alpha=0$ ahead of the blade (at $y=0$) because we have neglected the longitudinal squeezed flow. In (\ref{eq:wo3}) the length scale $k$ is given by the relation
\begin{equation}
k= \frac{\eta V}{2L},
\label{eq:k2}
\end{equation}
a very small length in the case of lubricating melted water film and realistic velocities (equal to $1.8 10^{-8} m$ for $V=12m/s$), therefore our study takes place in the limits
\begin{equation}
k \ll \alpha \ll \epsilon \ll \ell \ll r .
\label{eq:inegk2}
\end{equation}
As found below, the film thickness is of order $15\mu$, the penetration depth of order $1$mm, the length of contact of order $10 cm$, and the radius of profile of order of a few meters.
Equation (\ref{eq:wo3}) is to be considered jointly with the set of equations (\ref{eq:wo1}) and (\ref{eq:wsk})-(\ref{eq:wo2}), together with the condition (\ref{eq:poids}) balancing the skater weight with the beneath pressure. This later set can be reduced to the integral relation,
\begin{equation}
Mg=\frac{\eta V \sin(\theta/2)}{k^{3}} \int_{0}^{\ell}{ \left(\frac{2\ell y -y^{2}}{r \cos(\theta/2)} \right)^{3 }\left(\sin(\theta/2) \frac{\ell-y}{r } + \frac{d\alpha}{dy} \right)^{4} dy}.
\label{eq:poids2}
\end{equation}
Our mathematical analysis finally yields the coupled set of equations (\ref{eq:wo3})-(\ref{eq:poids2}). This makes our main result and it will be studied in the rest of the section.
\subsection{ Solution of coupled equations (\ref{eq:wo3}) and (\ref{eq:poids2}) }
Equation (\ref{eq:poids2}) fixes the value of $\ell$ once the first equation (\ref{eq:wo3}) is solved. In principle this makes a well defined problem. Since the equations are nonlinear, no analytical solution exists and a numerical solution must be looked at. This can be done by iterations by taking successive values of $\ell$ until the procedure converges: choose an arbitrary value $\ell^{(n)}$ for the length of contact, solve the differential equation (\ref{eq:wo3}) which gives a solution $\alpha^{(n)}$, insert this solution in the r.h.s of equation (\ref{eq:poids2}) and increment $ \ell$ using the same procedure as before until condition (\ref{eq:poids2}) is satisfied.
However in the situation of ice skating, because the length scales are widely different, it makes sense to try to find the solution in the frame of the inequality (\ref{eq:inegk2}).
\subsubsection{ Solution at leading order }
\label{sec:zero-order}
From inequality (\ref{eq:inegalfa}) one may conjecture that equation (\ref{eq:wo3}) has to be solved at leading order, by considering the derivative $\frac{d\alpha}{dy} $ as small in the term $\frac{w_{0}}{V}= \frac{\ell-y}{r }+\frac{d\alpha}{dy}$. Defining
\begin{equation}
\kappa(y) = \frac{d\alpha}{dt}/ w_{0}=V\frac{d\alpha}{dy}/ w_{0},
\label{eq:kappa}
\end{equation}
we get the simple solution
\begin{equation}
\alpha^{(0)} (y) = \frac{ r k}{ \sin(\theta/2)(\ell -y) )},
\label{eq:wo3.1}
\end{equation}
in the limit $\kappa \ll 1$ ,
where the superscript ${(0)}$ is to mean that this is the solution at leading order. Such an approximation requires to define a small parameter ($\gamma$, equation (\ref{eq:smallpara})) in terms of which the expansion of the solution is justified, at least in a certain domain.
This solution has two obvious defects. It diverges at $\ell = y$ which contradicts (\ref{eq:inegalfa}) and makes infinite the derivative $\frac{d\alpha}{dy}$ which was assumed to be much smaller than $\frac{\ell -y}{r }$ to derive equation
(\ref{eq:wo3.1}) from equation (\ref{eq:wo3}).Therefore there is a boundary layer near $\ell=y$ (physically where the blade becomes exactly horizontal). The second defect of the leading order solution (\ref{eq:wo3.1}) concerns the vicinity of $y=0$ where there is a second boundary layer, because at the front part of the ice-skate contact $\alpha(y)$ must tend to zero, a limit value obviously not satisfied by the solution (\ref{eq:wo3.1}).
However it turns out that, because
the two boundary layers have a small extension in the $y$ direction
one may neglect their contribution to the integral in equation (\ref{eq:poids2}), a result valid for all skating configuration as discussed in subsection \ref{sec:b.l.}. This integral which has to be calculated as well by canceling the derivative $\frac{d\alpha}{dy}$, becomes independent of $\alpha(y)$ and can be computed. Using the result $ \int_{0}^{\ell} { dy(\ell-y)^{4} (2\ell y -y^{2})^{3} }=\frac{16}{1155} \ell^{11}$,
we obtain
\begin{equation}
\ell^{(0)} = \left( \frac{1155}{16 } C Mg \right)^{1/11}
\label{eq:ll3.1}
\end{equation}
with
\begin{equation}
C= \frac{ (\cos(\theta/2))^{3} } { \sin(\theta/2))^{5} }\frac {(\eta V)^{2} r ^{7}} { (2L)^3}.
\label{eq:kl.1}
\end{equation}
This expression indicates that the length of contact is almost independent on the mass of the skater (as the power $M^{1/11}$), depends on the velocity as $V^{2/11}$ and mainly depends on the radius of profile (as $r^{7/11}$).
Note that ultimately the heat losses have a quite small effect on the water layer geometry, because it just amounts to reduce the length of contact by a factor $2^{3/11} \sim 1.2$.
To close the leading order solution of the coupled set (\ref{eq:wo3}) and (\ref{eq:poids2}), one has to replace of $\ell$ by $ \ell^{(0)}$ in (\ref{eq:wo3.1}).
Note that by plugging into this expression of $\ell ^{(0)}$ the numerical values for hochey skating, one finds that the inequalities (\ref{eq:inegk2})
are satisfied. In terms of the data, the approximation that the derivative $\frac{d\alpha}{dy}$ is negligible in equation (\ref{eq:wo3.1}), writes $\gamma \ll 1$ with
\begin{equation}
\gamma = \frac{ k r ^{2} }{ \ell ^{3}},
\label{eq:smallpara}
\end{equation}
because, by assuming (\ref{eq:wo3.1}), we have $(\frac{ d\alpha}{dy})^{(0)} = -\frac{ r k}{ \sin(\theta/2)(\ell -y) ^2)}$ which is of order $\frac{ r k}{\ell^{2}}$ in the major part of the domain and the term $ \sin(\theta/2 )\frac{\ell -y}{r } $ is of order $\ell/r $ (the bite angle $\theta$ is a parameter which can be taken away from the discussion of order of magnitude because it is about $\pi/2$).
If the inequality $ k r ^{2} \ll \ell^{3}$ is satisfied, the leading order solution (\ref{eq:wo3.1}) for $\alpha$ should be valid outside the boundary layers. This inequality is fulfilled for skating. This is not a consequence of $ k \ll \ell \ll r $, but follows from the order of magnitude typical of data, this ratio being of order $10^{-3}$ for figure, hockey and speed skating. However the condition $ \gamma \ll 1$ does not guarantee that the leading order solution (\ref{eq:ll3.1}) is a good approximation for $\ell$, this point will be discussed in subsection \ref{sec:b.l.} where another condition, more drastic, is derived.
To go further of the understanding of the role the $y$-dependent ratio $\kappa(y)$
we shall investigate the behavior of the solution close to the two boundary layers.
\subsubsection{Close to the boundary layers}
\label{sec:b.l.}
Close to $y=\ell$ and $y=0$ local forms of equation (\ref{eq:wo3}) can be analytically derived by using standard asymptotic expansion of a solution of a differential equation with a small parameter. This requires to define specific scalings valid in regions much smaller than $\ell$.
1) At the front of the blade, in the limit $y \to 0$, the solution of equation (\ref{eq:wo3}) with i.c. $\alpha(0)=0$ is $ \alpha(y) = (2k y)^{1/2},$ which has a square root singularity. In a certain domain one can neglect $y$ with respect to $\ell$ in (\ref{eq:wo3}). This leads to define, in this domain, the scaled variables $\tilde{y} =\frac{ y}{\ell_{0} }$ and $\tilde{\alpha} = \frac{\alpha}{\alpha_{0}}$ with the length scales $\ell_{0} =\frac{k r ^2}{\ell^2} $ and $\alpha_{0}=\frac{k r }{\ell} $, a scaling consistent with the use of the lubrication approximation for the liquid layer because $ \ell \ll r $. In the intermediate domain which makes the transition between the solution $ \alpha(y) = (2k y)^{1/2}$ and the exact one, equation (\ref{eq:wo3}) takes the local form
\begin{equation}
\tilde{ \alpha} \left( \sin(\theta/2 ) +\frac{d \tilde{\alpha}}{d \tilde{y}} \right )= 1.
\label{eq:cl2}
\end{equation}
Its solution
behaves asymptotically like
$$\tilde{ \alpha} \approx \frac{1}{\sin(\theta/2 )} $$ which matches the leading order solution (\ref{eq:wo3.1}) far away from the front part of the skate in these scaled variables. Restoring the original variables, the solution $\alpha^{b.l.}(y)$ of the local equation (\ref{eq:cl2}) is undistinguishable from the exact one for $0 \le y \le \ell_{0} $, as illustrated by the insert of Fig. \ref{fig:front}. Further the local solution tends to the constant value $ \alpha_{0} = \frac{k r }{\ell} $. The range $\ell_{0} << y << \ell$ defines the matching domain. At larger values of $y$, the local solution remains constant whereas the exact solution $ \alpha(y)$ remains close to $ \alpha^{(0)} (y) $ until the second boundary layer.
2) At the end of the contact, where $y$ tends to $\ell$, the leading order approximation is not valid, as written above, because the solution (\ref{eq:wo3.1}) is derived under the condition that the derivative $\frac{d\alpha}{dy} $ is small, whereas this derivative diverges at $ y = \ell$ (as well as $\alpha$).
Therefore we have to consider that all terms of equation (\ref{eq:wo3}) become of the same order of magnitude, that
leads to introduce the following length scales
\begin{equation}
\alpha_{*}= (k^{2} r )^{1/3}; \; \; \; \,\; \; \; \ \ell_{*}= (k r ^{2})^{1/3};
\label{eq:bl3}
\end{equation}
which gives in particular the order of magnitude of the thickness in this domain. Notice that because $ k\ll r $ the length scale $\alpha_{*}$ is much smaller than $\ell_{*}$ as necessary for maintaining the validity of the lubrication approximation for the flow in the layer between the skate and ice.
By introducing $\overline{y} =\frac{ \ell-y}{\ell_{*} }$ and $\overline{\alpha} = \frac{\alpha}{\alpha_{*}}$ equation (\ref{eq:wo3}) is transformed into
\begin{equation}
\overline{\alpha} \left( \sin(\theta/2 )\overline{y}- \frac{d\overline{\alpha} }{d\overline{y}} \right )= 1 ,
\label{eq:wo3.2}
\end{equation}
which is in fact identical to the original one, up to the rescaling defined in (\ref{eq:bl3}), and the change of variable from $\ell -y$
to $y$. Therefore in this case the local solution is the same as the global one and we cannot derive any analytical solution or numerical one
from initial conditions taken at $\overline{y}=0$ ($y=\ell$), because the forward solution $\overline{\alpha} $ of the differential equation (\ref{eq:wo3.2}) is unstable (as well as the backward solution $\alpha(y)$ of (\ref{eq:wo3})). This result could appear without any interest, but it is not, because the relation (\ref{eq:bl3}) is quite relevant, since it gives the order of magnitude of all the unknown lengths without solving any differential equation, in agreement with the exact numerical results, as reported next.
\subsection{Numerical results}
\label{sec:numerics}
In the numerical study we take the following data (in MKSA units) $\eta = 10^{-3}\; kg/(m s)$, $L = 3.3 \; 10^{5} \; W/m^3$, $g= 10 \;m/s^2$ and for the skater $V = 12 \;m/s$, $M = 75 \; kg$ (except in the subsection devoted to the role of these parameters). Moreover we set respectively {$r =3m$ and $r =25m$ depending on wether we describe figure/hockey skating or speed skating (the radius of profile being between about $2m$ for figure skating, between $2 m$ and $4 m$ for hockey, and may be longer than $25m$ for speed skating).
In all cases we show that
the assumption $\gamma \ll 1$, made to guarantee that the parameter $\kappa(y)$ is small in the major part of contact, is fulfilled. More importantly we are going to show that the leading order solution $\ell^{(0)}$ is a good approximation for $\ell$, that allows to shorten the numerical calculus of the solution of the two coupled equations by making a single iteration.
\subsubsection{Figure and Hockey skating}
The exact solution of the coupled equations requires to increment $\ell$ step by step, using the method outlined above: solve (\ref{eq:wo3}), calculate the r.h.s. of (\ref{eq:poids2}) until it is equal to its l.h.s. We find that the resulting length of contact is very close to $\ell^{(0)}$. Therefore a short cut to get a solution in good agreement with the exact one is to use this leading order value for $\ell$ and solve (\ref{eq:wo3}) only once. Doing this
the solution of equation (\ref{eq:wo3}) with initial condition $\alpha(0)=0$ is found to agree very well with the exact one (the error being less than one per thousand for the film thickness),
so that they are indistinguishable in fig. \ref{fig:final}, both merging in the red solid line. The numerical results obtained with the single step integration of (\ref{eq:wo3})) are
\begin{equation}
\alpha^{(1)}(\ell)= 14.83 \mu ; \; \; \; \, \epsilon^{(1)}(\ell)= 0.87mm \; \; \; \, \; \; \; \, \ell^{(0)}= 7.13 cm;
\label{eq:num2}
\end{equation}
in very good agreement with the exact values,
\begin{equation}
\alpha(\ell)= 14.83 \mu ; \; \; \; \ \epsilon(\ell) = 0.86mm; \; \; \; \, \; \; \; \, \ell= 7.10 cm,
\label{eq:num1}
\end{equation}
obtained by the multiple step method outlined above. These lengths are ordered as required by equation (\ref{eq:inegk2}), moreover we have $w_{0} \ll V$, as expected.
\begin{figure}
{
(a)\includegraphics[height=1.5 in]{bl-back.eps}
(b)\includegraphics[height=1.5 in]{eps-22janv.eps}
}
\caption{ (a) Film thickness $\alpha(y)$ and (b) penetration depth $\epsilon(y)$. The leading order solution (blue-dashed line) is compared to the exact solution of equations (\ref{eq:wo3}) -(\ref{eq:poids2}) with i.c. $\alpha(0)=0$, drawn in red-solid line which is indistinguishable from the first order one.}
\label{fig:final}
\end{figure}
For comparison the leading order solution is shown in blue dashed line. In (a) $\alpha^{(0)}(y)$ agrees with the exact one over the main part of the bottom blade, but differs in front close to $y=0$ (hardly visible in this figure but detailed in fig.\ref{fig:front}) and close to $y=\ell$, as stated previously (subsection \ref{sec:zero-order}).
To illustrate the validity of the conjecture made (at the beginning of the previous subsection) that the derivative $\frac{d\alpha}{dy} $ is small compared to $\frac{w_{0}}{V }$, we compare these two quantities in Fig. \ref{fig:wo-ap}. In this figure the derivative term is enhanced by a factor ten to be visible. The dashed curve, drawn for $10 \frac{d\alpha}{dy} $, shows that the derivative term is at least one order of magnitude lower than $\frac{w_{0}}{V }$ (solid line), except in the vicinity of $y=0$ and $y=\ell$, as expected. This result shows the importance of taking account of the relative velocity of the skater with respect to the ice surface, when writing the expression of the melting rate, as yet noted.
\begin{figure}
{\includegraphics[height=1.5 in]{wo-10.eps}}
\caption{ $\frac{ w_{0} }{V }$ (red solid line) and $\frac{d\alpha}{dy} $ versus $y$ (blue dashed line)}.
\label{fig:wo-ap}
\end{figure}
In the two boundary layers the numerics agree with the analytical study presented just before.
Close to the origin Fig. \ref{fig:front} shows the solution of the local equation (\ref{eq:cl2}) together with the exact one. They merge including in the matching domain which extends on a distance few times $ \ell_{0}$ where $\ell_{0}=\frac{k r ^2}{\ell^2}= 0.032mm$.
The insert displays the whole domain $0 \le y \le \ell$ to show that the local solution is no more valid beyond the matching domain where it remains constant.
The second boundary layer, close to $y=\ell$ , is clearly visible in Fig.\ref{fig:final}-(a) because the leading order solution $\alpha^{(0)}$ is well separated from the exact one over a noticeable distance. The extension of this layer, of order $\ell_{*}=0.5cm$ (the distance between the two vertical lines), is much larger than the extension of the boundary layer in front. We point out that the analytical study of the boundary layer presented above allows to derive straightforwardly
the order of magnitude of the thickness in terms of the data. Equation (\ref{eq:bl3}) gives $ \alpha_{*}=10\mu$ in qualitative agreement with
the exact result, a result obtained without solving the coupled set of equations (\ref{eq:wo3})-(\ref{eq:poids2}).
\begin{figure}
{\includegraphics[height=1.5in]{front-bl.eps}}
\caption{ Boundary layer description close to $y=0$: the exact solution (red solid line) is compared with the local solution of equation (\ref{eq:cl2}) $\alpha^{b.l.}$ (in restored variables, plotted in purple dashed line). They merge in the principal curve drawn in the domain $0 \le y \le20 \ell_{0}$.
$\ell_{0}=0.03 mm$ and $\alpha_{0}=0.77 \mu$.
The insert shows the solutions over the full domain to show that the matching domain (defined in the text) is much shorter than $\ell$. }
\label{fig:front}
\end{figure}
To compare our theory with the one presented in references \cite{LScana}-\cite{LScana2}, we note that they differ on several points. First the condition (\ref{eq:meltdiff}) necessary to describe the formation of the trough is lacking in this study, that leads them to estimate the length of contact from its value at rest, whereas the dynamical length of contact obtained here is five times larger. Secondly the authors of \cite{LScana}-\cite{LScana2} identify the melting rate $w_{0}$ with the thickness growth rate $\frac{d\alpha}{dt}$ although we show that it is more than one order of magnitude larger almost everywhere below the blade. This result shows the importance of taking into account the relative motion of the ice with respect the skater when writing the expression of the melting rate $w_{0}$ in the frame of the skater. Lastly the authors of \cite{LScana}-\cite{LScana2} add/subtract linearly the effects of the different contributions to the thickness growth rate, that leads to a \emph{single} differential equation, although our treatment leads to a set of \emph{two} coupled equations, one for the film thickness \textit{and} the other for length of contact.
\subsubsection{Role of the parameters}
\label{sec:data role}
It could be interesting to know the optimal conditions of skating. To give an idea of the role of the various parameters on the film thickness, let us investigate the role of the parameters which could be changed, like the mass $M$ and velocity $V$ of the skater, and the sharpening of the blade (its radius of profile $r $ and bite angle $\theta$). Here we investigate the role of these external parameters, keeping unchanged the other ones. The result is summarized in fig. \ref{fig:roleUr}.
\begin{figure}
{
(a)\includegraphics[height=1.2 in]{V-shape-U.eps}
(b)\includegraphics[height=1.2 in]{V-shape-r.eps}
(c) \includegraphics[height=1.2 in]{role-M.eps}
(c) \includegraphics[height=1.2 in]{role-theta.eps}
}
\caption{ Film thickness $\alpha(\ell)$ and length of contact $\ell$ versus the following data: (a) skater velocity $U$ (all other data unchanged), (b) radius of profile, (c) skater mass and (d) bite angle; in (b)-(d) $U$=8m/s. }
\label{fig:roleUr}
\end{figure}
Our curves show clearly that the film thickness and also the length of profile strongly depend on the velocity of the skater, moderately depends on the radius of profile and on the bite angle, although the geometry of the layer is almost insensitive to the skater mass.
Let us now compare these numerical results with the analytical study of solution of sec. \ref{sec:b.l.}. Introducing the bite angle dependence, the order of magnitude of the film thickness (at $y \backsimeq \ell$) given by (\ref{eq:bl3}) becomes,
\begin{equation}
\alpha_{*}= (\frac{ \eta }{ 2L })^{2/3} \left( \frac{ r V^{2} }{ \sin{(\theta /2)} } \right)^{1/3}.
\label{eq:num-para}
\end{equation}
This relation is in qualitative agreement with all numerical results displayed in fig.\ref{fig:roleUr}. It predicts a stronger dependence versus $V$ than versus $r $ and $\theta$, and independence with respect to $M$. Therefore we can claim that it contains the main physical result of the present study.
\subsection{Speed skating with an inclined blade}
\label {sec:speed}
The model derived above can be applied to the case of a speed skater strongly inclined with respect to the vertical axis, this angle being equal to $\pi/4$. A more general formulation can be derived in the general case of any angle value, but not presented here. For an inclined and symmetric rectangular blade , the geometry of the contact region is very close to the one considered for hockey or figure skating, but has to be calculated for larger radius of profile value and setting $\theta=\pi/2$. The radius of profile for speed skating is almost ten times larger than for figure/hocker skating, its typical value being $25m$, although the value of $\theta$ is close to the typical values of hockey-skating for which bite angles are between $70$ and $85$ deg.
Because the derivation of Stefan flow is the same for an inclined blade as for a V-shaped blade, it is enough to insert these two different parameters $ \theta, r $ into the equations obtained in section \ref{sec:hockey}. We get the following results ,
\begin{equation}
\alpha(\ell)= 29.13 \mu ; \,\,\, \epsilon(\ell) = 1.35mm; \, \, \, \ell= 25.6 cm,
\label{eq:numspeed-V}
\end{equation}
This shows the role of the radius of profile: increasing it by a factor $8$ increases by a factor $2$ the depth of the through and the thickness of the film. The geometry of the trough is detailed in Fig. \ref{fig:speed-V} where the exact solution of the coupled equations (\ref{eq:meltdiff})-(\ref{eq:poids2}) are drawn in solid lines, the leading order solution is drawn in dashed line for comparison. In summary we show that speed-skaters can plough a trough twice larger and deeper than hockey or figure skaters when skating with an inclined blade at the same velocity, and that the film of water is also twice thicker. These results are completely different from the study of \cite{LScana2} where the film thickness is of order $1 \mu$, whatever the angle of the blade.
\begin{figure}
{
(a)\includegraphics[height=1.2 in]{speed-V.eps}
(b)\includegraphics[height=1.2in]{eps-speed-V.eps}
(b)\includegraphics[height=1.2in]{wo-speed-V.eps}
}
\caption{ Solution for a speed skater making an angle $\pi/4$ with respect to the vertical (a) Film thickness $\alpha(y)$ , (b) penetration depth $\epsilon(y)$, (c) velocity of the melting flow normal to the ice surface in the through. The exact numerical solution of equations (\ref{eq:wo3}) -(\ref{eq:poids2}) are drawn in red solid lines and the leading order solution in blue-dashed lines .}
\label{fig:speed-V}
\end{figure}
\subsection{ Critical mass}
\label{sec:adim}
The analytical study of the coupled equations (\ref{eq:wo3}) -(\ref{eq:poids2}) has enlightened the role of the quantity $ \kappa(y)$ defined in (\ref{eq:kappa}). We have shown that this quantity can be neglected to get the solution of the integral equation (\ref{eq:poids2}): this gives $\ell^{(0)}$ in very good agreement with the exact length of contact, although $\kappa$ cannot be neglected to get the solution of the differential equation (\ref{eq:wo3}), because it becomes of order unity, as discussed in section \ref{sec:b.l.} and illustrated in Figs.\ref{fig:final}(a)-\ref{fig:wo-ap}-\ref{fig:front}.
There remains a question: What is the domain of validity of the approximation $\ell \backsimeq \ell^{0}$ ? In other words is there a range of parameters such that the leading order solution $\ell^{(0)}$ differs noticeably from the exact result? To answer this question we shall quantify the contribution of $\kappa$ in the integral equation (\ref{eq:poids2}) by using the results of subsection \ref{sec:b.l.}. Let us consider the integral in the r.h.s. of (\ref{eq:poids2})
\begin{equation}
\mathcal{I}=\int_{0}^{\ell}{ (\delta'_{sk})^{3} \left(\frac{w_{0}(y)}{V} \right) ^4}dy,
\label{eq:b1}
\end{equation}
and set $\Delta \mathcal{I}= \mathcal{I}-\mathcal{I}^{(0)} $ where $\mathcal{I}^{(0)}$ is for $w_{0}= w_{0}^{(0)}$. The condition $\frac{\Delta \mathcal{I} }{ \mathcal{I}} \ll 1$ is the one we are looking for, because it means that $\ell^{(0)}$ is a good approximation of $\ell$. The integrand in $\Delta \mathcal{I}$ is significant only in the region $y\sim \ell$ of extent $\ell_{*}$. Assuming that $\kappa \sim 1$ in this domain, and ignoring the role of the bite angle $\theta$ we obtain
$$\Delta \mathcal{I}=15(\frac{\ell_{*}}{r })^{4}\ell_{*}(\frac{\ell^{2}}{r })^{3}, $$ moreover we have $ \mathcal{I} \sim \frac{16}{1155}\frac{\ell^{11}} {r^{7} }$ at leading order. In terms of the length scales,
the condition $\frac{\Delta \mathcal{I} }{ \mathcal{I}} \ll 1$ can be written as
\begin{equation}
\frac{\ell_{*}}{\ell } \ll 0.25 ,
\label{a40}
\end{equation}
a condition fulfilled above (the ratio $\frac{\ell_{*}}{\ell }$ is equal to $0.03$ in the case of Fig.\ref{fig:final}). To compare with the inequality (\ref{eq:smallpara}) which is for the condition $\kappa \ll1$ in the central domain, we may notice that $\gamma= ( \frac{\ell_{*}}{\ell } )^{3}$, therefore the condition (\ref{a40}) is more drastic than (\ref{eq:smallpara}). This is because for deriving (\ref{a40}) we make one step more since we impose that the contribution of $\kappa$ in the boundary layer is negligible.
In terms of the data, using the order of magnitude (\ref{eq:bl3}) for $\ell_{*}$ and equation (\ref{eq:ll3.1}) for $\ell$, the condition (\ref{a40}) allows to define a set of critical parameters for the mass or velocity of the skater, and for the radius of profile of the blade. For example the inequality (\ref{a40}) writes $M \gg M_{*}(V,r)$ with
\begin{equation}
M_{*} \sim 2L \left( r k^{5} \right)^{1/3}
\label{eq:a4}
\end{equation}
where $M_{*}$ has the dimension of a mass and the length $k$ is proportional to the speed $V$. Using the data taken above for hockey skating with $V=12m/s$ , $r=3m$, this critical mass $M_{*}$ is very small ($0.2$g). To get critical mass values of order one kg, the velocity should be increased by a factor $100$ which would be inaccessible for skaters. This result shows that
the relation (\ref{eq:bl3}) covers all realistic skating situations.
From the previous estimates, one can also derive the friction force on the skate of length $l$. This is the friction on a surface $ l \delta'$ in a fluid where the velocity gradient is of order $V/ \alpha$. This total force (not per unit length of the skate) is $$ F_v = \frac{\eta V l \delta'}{\alpha} \mathrm{,}$$ where $\delta' $ and $\alpha$ have been estimated above. A dimensionless measure of this friction is the ratio of $ F_v$ to the weight $ M g $ of the skater
With the data taken above, this ratio is about $2 .10^{-4}$.
As a final remark, let us note that the geometry considered in this section, namely a narrow furrow made by melting ice, solves the problem of turning, because by tilting at the right angle, the skater rests on an inclined furrow which can stand a priori both his/her weight and the centrifugal force generated by the turn.
\section{Rectangular blade: Vertical speed skating}
\label{sec:rectangular}
We keep the same notations as before and assume that the skater moves on a vertical (not inclined) blade, with a rectangular cross-section of width $\delta$ independent of $y$. The blade is sunk into a furrow of total depth $\epsilon (y)= \epsilon_{sk}(y) + \alpha(y)$, where $\alpha$ is the thickness of the film of melted water. To simplify the calculations we assume that the thickness is constant all along the cross section, namely that $\alpha$ only depends on $y$ as above, whereas the value of $\alpha(s,y)$ could be larger vertically than horizontally where the film is squeezed by the weight of the skater. Then the furrow has a total depth $\epsilon (y)$, total width $\delta + 2 \alpha (y)$, and the length of the ice surface at the abscissa $y$ is $ \delta'(y)= \delta +2 \epsilon(y)$.
Equation (\ref{eq:wo1}) which describes the balance between the heat dissipated by viscous friction and absorbed to melt the ice surface is still valid, so are Poiseuille equations (\ref{eq:dp1})-(\ref{eq:u1}) and continuity condition (\ref{eq:u2}), together with equations (\ref{eq:d2p})-(\ref{eq:deltaw}). The relation between the melting rate $w_{0}$ and the slope of the skate blade is deduced as above by writing the condition for the formation of the trough,
$\mathcal{A}(y)= \epsilon(y) \delta= \int_{0}^{y} {dy' w_{0}(y' ) \delta '(y')/V }$, that gives after derivation with respect to $y$,
\begin{equation}
w_{0}(y)= V\frac{\delta}{\delta +2 \epsilon(y)} \frac{\partial \epsilon}{\partial y}.
\label{eq:melt2}
\end{equation}
The drop of pressure (\ref{eq:pr}) is also the solution of (\ref{eq:d2p}) (with the boundary conditions as in previous section), The integration of the pressure over the surface of contact, balanced with the weight of the skater writes here
\begin{equation}
Mg=\eta\delta \int_{0}^{\ell}{ w_{0}(y) \frac{ K_{1}(y)}{\alpha^{3}(y)} dy},
\label{eq:poids3}
\end{equation}
where $K_{1}(y)= \delta^{2} +6 \left(\epsilon^{2}(y) +\epsilon(y)\delta \right) $.
Finally, setting $\epsilon=\epsilon_{sk} +\alpha$ in (\ref{eq:poids3}) and (\ref{eq:melt2}), the two coupled equations relating the thickness $\alpha(y)$ to the length of contact $\ell$ are
\begin{equation}
\alpha \left( \frac{\ell -y}{r } +\frac{d\alpha}{dy } -2\frac{k}{\delta} \right )= k (1+ \frac{2\ell y-y^{2}}{\delta r })
\label{eq:wo4}
\end{equation}
which replaces (\ref{eq:wo3})
and
\begin{equation}
M g = \frac{V \eta\delta}{k^{3}} \int_{0}^{\ell} { dy \left(\frac{w_{0}}{V}\right)^{4} K_{1}(y)},
\label{eq:poids4}
\end{equation}
which replaces (\ref{eq:poids2}), with
\begin{equation}
\frac{w_{0}}{V}= \frac{\delta(\ell-y)+ \frac{d\alpha}{dy}r }{\delta r + (2\ell y-y^{2}) +2r \alpha(y) }
\label{eq:poids4b}
\end{equation}
and $$K_{1}(y)= \delta^{2} +6\left( (\frac{y\ell-y^{2}}{\ell})^{2} + \delta \frac{y\ell-y^{2}}{\ell} \right)$$
\subsubsection{Length of contact at leading order}
Neglecting the film thickness and its derivative in the above equations (\ref{eq:wo4})-(\ref{eq:poids4b}) gives the leading order solution $\frac{w_{0}^{(0)}}{V}= \frac{\delta}{\delta + \epsilon^{(0)}}\frac{d\epsilon^{(0)}}{ dy} $. With $\epsilon^{(0)}=(2\ell^{(0)} y -y^{2})/r $ we obtain $\frac{w_{0}^{(0)} }{V}=\frac{\ell-y}{r}\frac{1}{1+\frac{2\ell y-y^{2}} { r\delta }}$ and
\begin{equation}
\alpha^{(0)} (y)=k \frac{ \delta r -(2\ell^{(0)} y-y^{2}) }{ (\ell^{(0)} -y)\delta -2kr }
\label{eq:alfospeed}
\end{equation}
and a length of contact solution of the integral relation
\begin{equation}
\int_{0}^{\ell^{(0)}} { dy (\frac{w_{0}^{(0)}}{V})^{4} \left( \delta^{2} +6(\frac{ 2\ell^{(0)} y-y^{2} }{ r })^{2} +6\delta \frac{2\ell^{(0)}y-y^{2}}{r } \right) } = Mg\frac{k^{3}} {V \eta\delta}
\label{eq:mgospeed}
\end{equation}
which can be solved by iteration.
The numerical result is $\ell^{(0)}=22.2cm$, and a curve $\alpha^{(0)} (y)$ which diverges at $y=\alpha^{(0)} (y)$ as in the case of V-shaped blade studied above.
\subsubsection{Exact solution}
The exact solution may be obtained as above
by incrementing $\ell^{(n)}$ , solving (\ref{eq:wo4}) with $\ell=\ell^{(n)}$ to obtain $\alpha^{(n)}$, report this function in the r.h.s. of (\ref{eq:poids4}) until the right-hand side equals $M g$.
But we may also just solve equation (\ref{eq:wo4}) with initial condition $\alpha(y)=0 $ assuming that the length of contact is $ \ell=\ell^{(0)}$ because it gives quite the same result for typical data, as in the case of V-shaped blades. This procedure gives a film thickness and trough geometry of size,
\begin{equation}
\alpha(\ell)= 52 \mu ; \,\,\, \epsilon(\ell) = 1.0 mm; \, \, \; \, \, \ell= 22.2 cm,
\label{eq:numspeed}
\end{equation}
as illustrated in Fig. \ref{fig:speed-flat}.
\begin{figure}
{
(a)\includegraphics[height=1.5 in]{vert-speed-alf.eps}
(b)\includegraphics[height=1.5in]{vert-speed-eps.eps}
}
\caption{ Film thickness and penetration depth in the case of speed skate straight on ice (rectangular bottom blade) (a) Film thickness $\alpha(y)$ , (b) penetration depth $\epsilon(y)$,
where
the exact numerical solution of equations (\ref{eq:wo4}) -(\ref{eq:poids4}) are drawn in red solid lines and the leading order solution in blue-dashed lines .}
\label{fig:speed-flat}
\end{figure}
Note that these numerical results are similar to those of the inclined blade. Without making a detailed study of what happens in the general case of a rectangular blade wedged dissymmetrical into ice with any angle, we can make the conjecture that the numerical values of the lengths will be of the same order as those reported in (\ref{eq:numspeed-V}) and (\ref{eq:numspeed}).
\subsubsection{Estimation of $\alpha(\ell)$}
Because the calculations are heavier in the case of rectangular blade than in the case of V-shaped blade, it could be tempting to get the order of magnitude of the film thickness directly from the original parameters. This can be done as above by looking at the scalings close to $y=\ell$. There, defining the length scales $\ell_{*}$ for $\ell-y$ and $ \alpha_{*}$ for $\alpha$, we assume that all quantities in (\ref{eq:wo4}) have the same order of magnitude. Actually we may neglect the term $\frac{2k}{r}$ which is small with respect to $\ell_{*}/r$, (a condition that has to be verified a posteriori), that gives $ \ell_{*} =(\tilde{k} r^{2})^{1/3}$, and
\begin{equation}
\alpha_{*} =(r \tilde{k}^{2})^{1/3}
\label{eq:scalepeed}
\end{equation}
where $$\tilde{k}= k (1+ \frac{\ell^{2}}{r \delta }).$$ Taking the above data it gives $\ell_{*}=3cm$ and $ \alpha_{*} =39 \mu$ which still qualitatively agrees with the numerical value of the film thickness close to $y=\ell$. We must notice that $\alpha_{*}$ cannot be easily expressed in terms of the data, as it was in the case of V-shaped blade, because $\tilde{k}$ depends on the parameter $\ell$ whose expression at leading order, equation (\ref{eq:mgospeed}), is not simple.
equation (\ref{eq:mgospeed})
\section{Summary and conclusion}
The problem of understanding how and why it is possible to skate on ice almost without friction has remained imperfectly understood for a long time. In this work we introduced and discussed in depth all the necessary ingredient of a coherent theory. In particular we emphasized that, for a curved wedge-like blade, the one used in all skating sports, either speed, hockey or figure skating, the geometrical parameters and the velocity of the flow are linked together by three rather complex relations. One describes the melting of ice by viscous friction in the thin layer of melted ice (Stefan condition), another one expresses that direct contact between the skate and ice is avoided because the pressure of the Poiseuille flow in the layer is sufficient to lift the weight of the skater above the ice surface, the third one is the kinetic condition which connect the volume of melted ice to the volume of the furrow dug by the blade. Putting together those three relations we have shown that all the unknown physical quantities can be deduced from a set of two coupled equations, one differential the other integral, for the film thickness $\alpha$ and the length of contact $\ell$. The analytical study of both equations has revealed the existence of a parameter ($ \kappa = \frac{\frac{ d\alpha }{dt}}{ w_{0}} $) which plays a key role in the derivation of the solution because it is small all along the film, except at the ends. This property shows in particular that the melting rate $w_{0} $ definitely hasn't to be confused with the growth rate of the layer $\frac{d\alpha}{dt}$, an assumption found in the literature, because the two quantities differ by one order of magnitude in the correct theory. Besides this remark the fact that such a small parameter exists is of prime importance from a practical point of view, because it allows to get analytically the order of magnitude of the solution and inform about the effects of the input data.
A natural extension of this complete theory is to allow the design of improved skates by optimizing their shape to lower friction and increase the grip on the ice surface when the skater makes a turn.
\section{Acknowledgment}
One of us (YP) would like also to thank his grand son, Gaspard, for teaching him patiently the ways of hockey skating.
|
\section{Introduction}
The leading radiative corrections to scattering processes in closed string theories compactified on a $d$-dimensional torus are computed by modular integrals of the form
\begin{equation}
\mathcal{I} = \int_\mathcal{F} d\mu \, \varGamma_{d+k,d} \, \varPhi \,,
\label{intintro}
\end{equation}
where $\varGamma_{d+k,d}$ is the partition function of the Narain lattice with signature $(d+k,d)$,
$\varPhi$ is an appropriate modular form of negative
weight $(-\frac{k}{2},0)$ under the modular group
$\varGamma={\rm SL}(2;\mathbb{Z})$ of the worldsheet torus,
and $\mathcal{F}$ is a fundamental domain for the action
of $\varGamma$ on the Poincar\'e upper half plane $\mathbb{H}$. In order to extract quantitative predictions from string theory, it is therefore essential to develop efficient tools to compute such modular integrals.
In the traditional lattice unfolding method \cite{O'Brien:1987pn,McClain:1986id,Dixon:1990pc}, one decomposes the lattice sum $\varGamma_{d+k,d}$ into orbits under the modular group, and reduces each orbit to a single term while unfolding the integration domain. Although the resulting expression is very useful in extracting the asymptotic expansion of $\mathcal{I}$ in the limit where the volume of the torus $T^d$ becomes large, it depends on a choice of chamber in the Narain moduli space, and T-duality invariance is no longer manifest.
In a series of recent works \cite{Angelantonj:2011br,Angelantonj:2012gw,Angelantonj:2013eja}, we proposed a new method, which assumes that $\varPhi$ can be written as an absolutely convergent Poincar\'e series under the modular group. Upon unfolding the integration domain against the sum over images, and evaluating the integral, the result is expressed as a sum over charges satisfying a BPS-type condition. This can be recognised as a particular type of Poincar\'e series for the automorphism group $ {\rm O} (d+k,d;\mathbb{Z} )$ of the Narain lattice, which is thus manifestly invariant under T-duality. The analytic structure of the integral near points of enhanced gauge symmetry is also manifest, while the large volume asymptotics is no longer apparent. In fact, the method can be applied even when the integrand does not include the Narain lattice, {\it e.g.} in ten-dimensional string vacua.
If $\varPhi$ is a generic non-holomorphic modular form, there is no general known way to represent it as a Poincar\'e series. For BPS-saturated amplitudes, however, the modular form $\varPhi$ turns out to be weakly almost holomorphic, {\it i.e.} a polynomial in the ring $\mathbb{C}[\hat E_2,E_4,E_6,1/\Delta]$. In that case, it can actually be decomposed as a linear combination of so-called Niebur-Poincar\'e series $\mathcal{F}(s,\kappa,w;z)$ \cite{1004.11021, Angelantonj:2012gw}. These Poincar\'e series, introduced in the mathematics literature in \cite{0288.10010,0543.10020}, have the virtues of being eigenmodes of the Laplacian on $\mathbb{H}$ with a pole of order\footnote{While the integrands appearing in heterotic string theory only have a first order pole at $z=\infty$, it is useful to allow poles of arbitrary order $\kappa\geq 1$ as such poles typically occur in modular integrals for congruent subgroups, relevant for orbifold compactifications. Moreover, Niebur-Poincar\'e series $\mathcal{F}(s,\kappa',w)$ with $\kappa'>1$ occur in the Fourier expansion of modular integrals of the form \eqref{intintros}, even if $\kappa=1$.} $\kappa$ at the cusp $z=\infty$ and of being closed under modular derivatives and Hecke operators. In addition, they reduce to harmonic Maass forms when $s=1-\tfrac{w}{2}$, or to modular derivatives thereof when $s=1-\tfrac{w}{2}+n$ with $n$ a positive integer.
For suitable linear combinations, the non-analytic parts of $\mathcal{F}(s,\kappa,w;z)$ (encoded in the `shadow') cancel, producing the desired weakly almost holomorphic form $\varPhi$.
The integral \eqref{intintro}
is thus reduced to a sum of integrals of the form
\begin{equation}
\mathcal{I} (s,\kappa)= \int_\mathcal{F} d\mu \, \varGamma_{d+k,d} \, \mathcal{F}(s,\kappa,-k/2) \,,
\label{intintros}
\end{equation}
each of which can be computed by unfolding the fundamental domain against the Niebur-Poincar\'e series \cite{Angelantonj:2012gw}.
In this work, we compute the large volume expansion of the modular integral \eqref{intintros},
thus providing an alternative way to extract the asymptotics of \eqref{intintro}
within our unfolding method. The advantage of decomposing $\varPhi$ into a sum of
Niebur-Poincar\'e series is that the integral \eqref{intintro} is then decomposed
into a sum of terms which are separately eigenmodes of the Laplacian on
the Narain moduli space
\begin{equation}
\label{Narainmod}
G_{d+k,d} = {\rm O} (d+k,d;\mathbb{Z} ) \backslash {\rm SO} (d+k,d ) / {\rm SO} (d+k ) \times {\rm SO} (d) \, ,
\end{equation}
and which behave in a simple way under covariant derivatives with respect to the moduli.
For simplicity, we restrict our attention to integrals involving lattices of signature (2,2), where the moduli space \eqref{Narainmod} is parametrised by the standard variables $(T=T_1+\mathrm{i} T_2,U=U_1+\mathrm{i} U_2)\in \mathbb{H}_T \times \mathbb{H}_U$. They correspond to the K\"ahler modulus and complex structure of a two-torus, and are identified under separate actions of the modular group ${\rm SL}(2;\mathbb{Z})$ on $T$ and $U$, and under the involution $\sigma:T\leftrightarrow U$.
This case is relevant when the internal space includes a $T^2$ factor, {\it e.g.} in the context of heterotic strings compactified on $T^2$ (with 16 supercharges), on ${\rm K3}\times T^2$ (with 8 supercharges), or on $\mathcal{N} =1$ orbifolds with $\mathcal{N} =2$ sub-sectors, upon restriction to backgrounds with trivial or discrete Wilson lines\footnote{We defer the analysis of asymmetric lattices with non-trivial Wilson lines to a forthcoming publication.}. We consider both the limit where the volume $T_2$ is scaled to infinity keeping the complex structure modulus $U$ fixed, and the limit where both $T_2$ and $U_2$ are scaled to infinity (in which case the asymptotic expansion depends on a choice of chamber, $T_2>U_2$ {\it vs.} $T_2<U_2$). In the latter case, the result is expressed as an infinite sum of a special type of single-valued polylogarithms, known in the mathematics literature as Bloch-Wigner-Ramakrishnan polylogarithms \cite{zbMATH04144378}, which coincide with the real part of the `combined polylogarithms' introduced in \cite{Kiritsis:1997hf}.
In the former case, $T_2\to \infty$ with $U$ fixed, we show that the Fourier coefficients of the integral $\mathcal{I}(s,\kappa)$ with respect to $T_1$ are given by the same
Niebur-Poincar\'e series $\mathcal{F}(s,\kappa,0)$ present in the integrand, now as functions of the $U$ modulus. This is a manifestation of a phenomenon first observed in \cite{Bachas:1997mc} which, as we shall see, can be traced to the average value property of eigenmodes of the Laplacian on the Poincar\'e upper half-plane. An important consequence of this observation is that, using the closure of Niebur-Poincar\'e series under modular derivatives, one may express the integral \eqref{intintros} as an iterated derivative of a `harmonic prepotential' $h_n$\cite{Kiritsis:1997hf}
\begin{equation}
\label{iderpot0}
{\mathcal I} (n+1,\kappa) = 4 \, {\rm Re}\, \left[ \frac{(-D_T D_U)^n}{n!} h_n(T,U) \right]
- 24\, \sigma_1(\kappa)\, \delta_{n,0} \, \log T_2 U_2\ .
\end{equation}
The prepotential $h_n$ is by construction harmonic in $(T,U)$ and transforms covariantly under ${\rm SL}(2;\mathbb{Z})_T\times {\rm SL}(2;\mathbb{Z})_U$, while the last term originates from the infrared divergence of the integral when $n=0$. Crucially, it is also possible to construct a {\it holomorphic} prepotential
$f_n(T,U)$ which satisfies the same relation \eqref{iderpot0}. This holomorphic prepotential is
well-defined up to the addition of polynomials of degree $2n$ in $T$ and $U$ with real coefficients, which lie in the kernel of the operator ${\rm Re}(D_T D_U)^n$. As a result, it need not transform homogeneously under T-duality, but may pick up an additional degree $2n$ polynomial. This type of anomalous behaviour under modular transformations is characteristic of Eichler integrals. Indeed, we trace the modular anomaly of $f_n(T,U)$ to the presence of a term proportional to the Eichler integral of the holomorphic Eisenstein series $E_{2n+2}$ in the Fourier expansion. We use this fact to fully characterise the T-duality transformations of the generalised prepotential $f_n$.
From a physics viewpoint, the prepotentials $f_n$ for $n>1$ can be viewed as generalisations of the usual holomorphic prepotential $F$ that governs $\mathcal{N}=2$ supergravity in four dimensions -- or
more accurately, of the one-loop contribution to it \cite{Antoniadis:1995ct,deWit:1995zg,Harvey:1995fq,Forger:1997tu}. In fact, it was already recognised in \cite{Antoniadis:1995ct} that the quadratic terms appearing in the behaviour of the one-loop prepotential $f_1$ under monodromies around singularities in moduli space can be obtained from contour integrals of suitable modular forms, which are Eichler integrals in essence. Our results enable us to recover the monodromy matrices computed in \cite{Antoniadis:1995ct} in a streamlined fashion, and to extend the analysis to $n>1$. The case $n=2$ was studied in the context of $F^4$ couplings in heterotic strings compactified on $T^2$ \cite{Lerche:1998nx,Lerche:1999hg}, where a relation of the generalised prepotentials
$f_n$ to periods on symmetric products of $n$ K3's was suggested. However, the analogue of special geometry for $n>1$ is not understood at present. As a step in this direction, we construct
a class of `generalised Yukawa couplings', obtained from $f_n$ by iterated derivatives with respect to the $T$ and $U$ moduli, which compute modular integrals with insertions of momenta and are free of modular anomalies.
From a mathematics viewpoint, the prepotentials $f_n$ can also be viewed as interesting generalisations of Borcherds automorphic products \cite{zbMATH00220742} -- or rather, logarithms thereof. Indeed, in the special case where $\mathcal{F}(1+n,1,-2n)$
is a weakly holomorphic modular form, $f_n$ can be written schematically as a polylogarithm sum
\begin{equation}
f_n(T,U)=
\frac{1}{2^{2n}}
\sum_{N,M} F_n (NM) \, {\rm Li}_{2n+1} (q_T^M\, q_U^N) + P_{2n}(T,U)\ ,
\end{equation}
where $F_n(M)$ are the Fourier coefficients of
$\mathcal{F}(1+n,1,-2n)$ and $P_{2n}$ is the polynomial ambiguity mentioned above. For $n=0$, ${\rm Li}_1(z)=-\log(1-z)$, so $f_0$ is recognised as the logarithm of a
Borcherds automorphic product. For $n>0$, $f_n$ is no longer the logarithm of a product, but it provides an interesting example of an holomorphic `Mock' modular form on $(\mathbb{H}_T\times \mathbb{H}_U)/\sigma$ with known singular behaviour at all divisors.
Returning to physics, the integrals \eqref{intintro} do not exhaust the class of one-loop amplitudes of interest in string theory. In fact, in compactifications on non-factorisable tori \cite{Mayr:1993mq,Bailin:2014nna} or on freely-acting orbifolds \cite{Kiritsis:1998en, Kiritsis:1997ca, Angelantonj:2006ut, Angelantonj:2014dia, Faraggi:2014eoa} the integrand is typically expressed as a sum of various sectors $(h,g)$ involving shifted Narain lattices $\varGamma_{2,2}[{ h \atop g}]$ coupled to modular forms $\varPhi[{ h \atop g}]$.
Although the full integrand is, by construction, invariant under the full modular group ${\rm SL} (2;\mathbb{Z})$, each term is only invariant under a suitable congruence subgroup. These setups have not been extensively analysed so far in the literature and call for a systematic investigation.
Upon partial unfolding \cite{Angelantonj:2013eja}, one may compute the integral by focusing on a single sector now integrated over the fundamental domain of the congruence subgroup. In the case of freely-acting $\mathbb{Z}_N$ orbifolds of $\mathcal{N} =2$ compactifications, the relevant subgroup is the Hecke congruence subgroup $\varGamma_0 (N)$. In parallel
with the ${\rm SL} (2;\mathbb{Z})$ case, one may decompose $\varPhi[{ h \atop g}]$
as a linear combination of the $\varGamma_0 (N)$ Niebur-Poincar\'e series $\mathcal{F}_\mathfrak{a} (1+n,\kappa , 0)$ attached to the cusp $\mathfrak{a}$ \cite{Angelantonj:2013eja}. One may then evaluate the modular integral by unfolding the fundamental domain $\mathcal{F}_N$ of $\varGamma_0 (N)$ against the Poincar\'e series. This procedure was suggested in \cite{Angelantonj:2013eja} and applied to the study of radiative corrections to gauge couplings in classically stable non-supersymmetric heterotic vacua in \cite{Angelantonj:2014dia}. The result is a representation of the modular integral
as a Poincar\'e series under the T-duality group, in general a congruence subgroup
of ${\rm O}(2,2;\mathbb{Z})$.
In this work, we continue this analysis and extract the large-volume expansion of one-loop modular integrals for $\varGamma_0(N)$ with $N$ prime. We obtain expressions which extend the results for ${\rm SL}(2;\mathbb{Z})$ and are nicely covariant with respect to the choice of cusp $\mathfrak{a}$. The advantage of our approach, based on the special properties of $\mathcal{F}_\mathfrak{a} (1+n, \kappa , 0)$, is that we obtain a $T$-Fourier series expansion with $U$-dependent coefficients which are manifestly invariant under $\varGamma_0 (N)_U$, and which actually are given in terms of the same class of Niebur-Poincar\'e series. This property is instrumental in allowing us to derive new explicit results. Namely, in the $n=0$ case we re-sum the integral in terms of the $\varGamma_0 (N)$ Hauptmodul by generalising the celebrated ${\rm SL} (2;\mathbb{Z})$ Borcherds' product formula to the case of Hecke congruence subgroups. For $n\ge 1$ we identify generalised holomorphic prepotentials and fully characterise their transformation properties under T-duality. Finally, we express the generalised holomorphic Yukawa couplings in terms of the standard generators of the ring of $\varGamma_0 (N)$ modular forms.
The paper is organised as follows: in Section \ref{sec_intsl2z}, after recalling basic properties of Niebur-Poincar\'e and Eisenstein series, we compute the large volume asymptotics of the modular integral \eqref{intintros} or, equivalently, the Fourier expansion with respect to $T_1$, starting from its T-duality invariant representation. We express the results in terms of single-valued polylogarithms, and make contact with the literature. In Section
\ref{sec_prepot} we integrate this result in terms of generalised harmonic and holomorphic prepotentials, explain their relation to Eichler integrals, and determine their anomalous transformation properties under T-duality. We construct generalised Yukawa couplings by taking suitable modular
derivatives of the generalised prepotentials, and connect them to modular integrals with lattice-momentum insertions. Section \ref{HeckeIntegrals_sec} extends the previous analysis to modular integrals involving shifted
Narain lattices. We obtain new explicit results including `cusp-covariant' Fourier series expansions, closed expressions for special classes of integrals involving the $\varGamma_0 (N)$ Hauptmodul,
a full characterisation of the generalised holomorphic prepotentials, as well as concrete formul\ae\ for the generalised holomorphic Yukawa couplings. Appendix A collects our conventions for the Kloosterman sums and zeta functions, and Appendix B provides a detailed derivation of the modular properties of the relevant Eichler integrals.
\section{Modular integrals for ${\rm SL}(2;\mathbb{Z})$ \label{sec_intsl2z}}
In this section, we shall focus on modular integrals \eqref{intintro} involving the product of a signature (2,2) Narain lattice $\varGamma_{2,2}$ times a weakly almost holomorphic modular functions $\varPhi (\tau )$. As anticipated, the latter can be uniquely decomposed in terms of Niebur-Poincar\'e series of zero weight \cite{1004.11021, Angelantonj:2012gw}. Thus, we begin by collecting some useful properties of Eisenstein and Niebur-Poincar\'e series which are instrumental for the subsequent analysis.
\subsection{Niebur-Poincar\'e series for ${\rm SL}(2; \mathbb{Z})$}\label{secNPS}
The Niebur-Poincar\'e series \cite{ 0288.10010, 0543.10020,1004.11021}\footnote{We shall henceforth omit the explicit dependence of the Niebur-Poincar\'e series on the modular parameter $z$ whenever it is clear from the context.}
\begin{equation}
\begin{split}
\label{Fskw}
\mathcal{F}(s,\kappa,w ; z) =&\tfrac12 \sum_{\gamma\in \varGamma_\infty\backslash \varGamma} \,
\mathcal{M}_{s,w}(-\kappa y)\, e^{-2\pi\mathrm{i}\kappa x}\, \vert_w\, \gamma
\\
=& \tfrac{1}{2} \sum_{(c,d)=1} (cz + d)^{-w}\, {\mathcal M}_{s,w} \left(- \frac{\kappa\, y}{|cz + d|^2} \right) \, \exp \left\{ -2\mathrm{i}\pi \kappa \left( \frac{a}{c} -\frac{cx+d}{c|cz + d|^2}\right)\right\} \, .
\end{split}
\end{equation}
defines a one-parameter family of modular forms of non-positive weight $w$ and order $\kappa$ pole in the nome $q=e^{2i\pi z}$ at the cusp $z\equiv x+\mathrm{i} y =\infty$. Its seed is expressed in terms of the Whittaker
$M$-function
\begin{equation}
\mathcal{M}_{s,w}(t) = |4\pi t|^{-\frac{w}{2}}\, M_{\frac{w}{2}{\rm sgn}(t), s-\frac12} \left( 4\pi | t| \right) \,,
\label{curlyM}
\end{equation}
whose behaviour at $t\to 0$ guarantees absolute convergence when ${\rm Re}\, (s)>1$.
The Poincar\'e series can be shown to admit a meromorphic continuation to the complex $s$ plane.
Moreover, in the domain of analyticity, $\mathcal{F} (s,\kappa , w) $ is an eigenmode of the weight-$w$ hyperbolic Laplacian $\varDelta_{z}$,
\begin{equation}
\label{laplEskw}
\left[\varDelta_z + \tfrac{1}{2}\, s(1-s) +\tfrac{1}{8}\, w(w+2)\right] \, \mathcal{F}(s,\kappa,w) = 0\ ,
\end{equation}
where
\begin{equation}
\varDelta_{z} = 2\, \bar D_z \, D_{z}\,, \qquad {\rm with} \quad \bar D_z = -i \pi \, y^2 \partial_{\bar z} \,, \quad D_{z} = \frac{i}{\pi} \left( \partial_z - \frac{iw}{2y}\right)\,.
\end{equation}
Its Fourier expansion reads
\begin{equation}
\label{FskwF}
\mathcal{F}(s,\kappa,w)=\mathcal{M}_{s,w}(-\kappa y)\, e^{-2i \pi \kappa x}
+ \sum_{m\in\mathbb{Z}} \, \tilde\mathcal{F}_m(s,\kappa,w; y) \, e^{2 i \pi m x}\,,
\end{equation}
where, in the mode with frequency $m=-\kappa$, we have separated the contribution of the seed. Explicitly,
\begin{equation}
\label{NPFouriermodes}
\begin{split}
\tilde \mathcal{F}_0 (s,\kappa , w ; y) &= f_0 (s,\kappa , w) \, y^{1-s-\frac{w}{2}}\,,
\\
\tilde \mathcal{F}_m (s, \kappa , w ; y ) &= f_m (s,\kappa , w)\, \mathcal{W}_{s,w} (my)\,,
\end{split}
\end{equation}
with
\begin{equation}
\label{NPFouriermodes2}
\begin{split}
f_0 (s,\kappa,w) &=
\frac{2^{2-w}\, i^{-w}\, \pi^{1+s-\frac{w}{2}}\, |\kappa|^{s-\frac{w}{2}}\, \varGamma (2s-1)\, \sigma_{1-2s} (\kappa )}{\varGamma(s-\frac{w}{2}) \, \varGamma(s+\frac{w}{2})\, \zeta(2s)} \,,
\\
f_m (s,\kappa , w) &= \frac{4\pi\,i^{-w}\, |\kappa| \, \varGamma(2s)}{ \varGamma ( s+\frac{w}{2} \, {\rm sgn}(m))}\, \left| \frac{m}{\kappa}\right|^{\frac{w}{2}} \, \mathcal{Z} (m,-\kappa ; s) \,.
\end{split}
\end{equation}
In these expressions, $\mathcal{W}_{s,w} (t)$ is related to the Whittaker $W$-function
\begin{equation}
\mathcal{W}_{s,w} (t) = |4\pi t |^{-\frac{w}{2}}\, W_{\frac{w}{2}{\rm sgn}(t), s-\frac12}
\left(4\pi |t| \right) \,,
\end{equation}
$\sigma_\ell (n)$ is the divisor function, while $\mathcal{Z} (a,b; s)$ is the {\em associated} Kloosterman--Selberg zeta function defined in \eqref{KSzetadef}.
Niebur-Poincar\'e series are closed under the action of the modular derivatives,
\begin{equation}
\begin{split}
D_z\, \mathcal{F} (s,\kappa , w) &= 2 \kappa (s+\tfrac{w}{2}) \, \mathcal{F} (s,\kappa , w+2 )\,,
\\
\bar D_z \, \mathcal{F} (s,\kappa ,w) &= \frac{1}{8\kappa} (s- \tfrac{w}{2}) \, \mathcal{F} (s,\kappa , w-2)\,,
\end{split}
\label{NPderiv}
\end{equation}
and of the Hecke operator\footnote{The Hecke operator $H_n^{(z)}$ acts on the Fourier modes of a weight-$w$ modular form $\varPhi = \sum_m \varPhi (m,y) \,e^{2i \pi m x}$ as
\begin{equation}
H_n^{(z)} \, \varPhi (m,y) = n^{1-w} \sum_{d|(m,n)} d^{w-1}\, \varPhi (nm/d^2 , d^2 y/n)\,. \nonumber
\end{equation}
It satisfies the commutative
algebra $H_\kappa^{(z)} \, H_\lambda^{(z)} = \sum_{d|(\kappa , \lambda)} d^{1-w}\, H^{(z)} _{\kappa \lambda / d^2}$.}
\begin{equation}
\mathcal{F} (s,\kappa , w) = H_\kappa^{(z)} \, \mathcal{F}(s, 1 ,w)\, .
\label{NPHecke}
\end{equation}
For the special values $s = \frac{w}{2}$ and $s=1-\frac{w}{2}$, within the domain of analyticity,
$\mathcal{F} (s,\kappa ,w)$ becomes a harmonic Maa\ss\ form. For negative weight, the case of interest for most applications to string theory \cite{Angelantonj:2012gw}, the Niebur-Poincar\'e series
is absolutely convergent at the latter value $s=1-\frac{w}{2}$.
Moreover, since the first cusp form for ${\rm SL} (2;\mathbb{Z})$ appears at weight twelve, for $w \in \{ 0,-2,-4,-6,-8,-12\}$ the shadow of $\mathcal{F}(1-\frac{w}{2},\kappa,w)$ vanishes and the Niebur-Poincar\'e series is actually holomorphic, hence valued in the polynomial ring generated by $E_4$, $E_6$ and $\varDelta^{-1}$, where $E_{2n}$ are the holomorphic Eisenstein series and $\varDelta$ is the modular discriminant. Other Niebur-Poincar\'e series of interest are $\mathcal{F}(1-\frac{w}{2} +n,\kappa,w)$, with $n$ a positive integer, which can be reached from the harmonic Maa\ss\ form
$\mathcal{F}(1-\frac{w}{2} +n,\kappa,w-2n)$ by acting with $n$ modular derivatives. In fact, one can prove \cite{1004.11021, Angelantonj:2012gw, Angelantonj:2013eja} that any weak almost holomorphic modular form of negative weight in the polynomial ring generated by $\hat E_2 $, $E_4$, $E_6$ and $\varDelta^{-1}$, with $\hat E_2$ being the almost holomorphic Eisenstein series of weight two, can be uniquely decomposed into a linear combination of Niebur-Poincar\'e series,
\begin{equation}
\varPhi (z) = \sum_{0<m\le\kappa}\, \sum_{\ell =0}^n \, d_\ell (m) \, \mathcal{F} \left(1-\tfrac{w}{2}+\ell , m , w \right)\,.
\label{lineardec}
\end{equation}
The coefficients $d_\ell (m)$ are determined by the principal part of the $q$-Laurent expansion of $\varPhi$ itself \cite{Angelantonj:2013eja}.
The integer $n$ counts the maximal power of $\hat E_2$ in the ring decomposition of $\varPhi$, and is known as the {\em depth}. An important fact is that for each value of $\ell$, the shadows cancel in the linear combination \eqref{lineardec} .
It will also be useful to introduce the non-holomorphic weight-$w$ Eisenstein series
\begin{equation}
\begin{split}
E(s,w; z) &= \tfrac{1}{2} \sum_{\gamma\in\varGamma_{\infty}\backslash\varGamma} \, y^{s-\frac{w}{2}}\vert_w\, \gamma
\\
&= \tfrac{1}{2} \sum_{(c,d)=1} \frac{y^{s-\frac{w}{2}}} {(c\, z+d)^{w} \, |c\, z+d|^{2s-w}}
\end{split}
\label{defEsl2}
\end{equation}
which is absolutely convergent for ${\rm Re}\, (s)>1$, and admits the Fourier series expansion
\begin{equation}
\label{FourierEisw}
\begin{split}
E(s,w;z)=& y^{s-\frac{w}{2}}+
\frac{4^{1-s}\, i^{-w} \pi \varGamma(2s-1) \zeta(2s-1)}{\varGamma(s+\frac{w}{2})\, \varGamma(s-\frac{w}{2})\,\zeta(2s)} y^{1-s-\frac{w}{2}}
\\
&+ \sum_{m\neq 0}
\frac{i^{-w} (4\pi |m|)^{\frac{w}{2}} \pi^s}{\varGamma(s+\tfrac{w}{2} {\rm sgn} (m))\, \zeta(2s)}\,
|m|^{-s}\, \sigma_{2s-1}(|m|)\, \mathcal{W}_{s,w}(m y)\, e^{2i \pi mx}\,.
\end{split}
\end{equation}
The Eisenstein series $E(s,w)$ can be obtained formally as the limit $\kappa \to 0$ of $|4\pi \kappa |^{\frac{w}{2}-s}\, \mathcal{F} (s,\kappa , w) $, and thus shares several of the properties of the Niebur-Poincar\'e series. Namely, it is an eigenmode of the weight-$w$ hyperbolic Laplacian
\begin{equation}
\left[ \varDelta_w+\tfrac12 s(1-s)+\tfrac18 w(w+2) \right] \, E(s,w;z) =0\, ,
\end{equation}
is closed under the action of the modular derivatives
\begin{equation}
\begin{split}
D_z\, E(s,w;z) &= \frac{1}{2\pi} (s+\tfrac{w}{2})\, E(s,w+2;z)\,,
\\
\bar D_z\, E(s,w;z) &= \frac{\pi}{2} (s-\tfrac{w}{2})\, E(s,w-2;z)\,,
\end{split}
\label{EPderiv}
\end{equation}
and becomes harmonic in $z$ when $s=\tfrac{w}{2}$ or $s=1-\tfrac{w}{2}$. For $w>2$, $E (\tfrac{w}{2}, w)$ coincides with the holomorphic Eisenstein series $E_w $.
\subsection{Evaluation of the modular integral}
We can now proceed to the study of the integral \eqref{intintro}. Since any weak almost holomorphic modular form $\varPhi$ can be decomposed as in \eqref{lineardec} it is sufficient to focus our attention on the basic regularised modular integral
\begin{equation}
\mathcal{I} (s,\kappa ;T,U) = {\rm R.N.} \int_\mathcal{F} d\mu \, \varGamma_{2,2} (T,U; \tau ) \, \mathcal{F} (s,\kappa , 0; \tau)\,,
\label{modint}
\end{equation}
where $\varGamma_{2,2} (T,U;\tau)$ is the partition function of the even self-dual lattice of signature $(2,2)$, and ${\rm R.N.}$ stands for the modular invariant renormalisation prescription of
\cite{MR656029, Angelantonj:2011br, Angelantonj:2012gw}
for treating the infra-red divergences arising at the cusp. For $s=1$, the integral
\eqref{modint} develops a simple pole, which must be subtracted \cite{1004.11021, Angelantonj:2012gw}. By abuse of notation we denote by $\mathcal{I}(1,\kappa;T,U)$ the result of this subtraction.
The partition function
\begin{equation}
\label{G22ham}
\varGamma_{2,2} (T,U;\tau) =
\tau_2\, \sum_{m_1,m_2,n^1,n^2\in\mathbb{Z}} q^{\frac14 |p_L|^2} \bar q^{\frac14 |p_R|^2} \,,
\end{equation}
which we refer to as the Narain lattice, is a sum over integers $m_i$ and $n^i$ identified with the Kaluza-Klein momentum and winding numbers along a two-dimensional torus. The left-moving and right-moving momenta on the two-dimensional lattice
\begin{equation}
\begin{split}
p_{\rm R} &= \frac{m_2 - U\, m_1 + T n^1 + T U\, n^2 }{\sqrt{T_2\, U_2}} \,,
\\
p_{\rm L} &= \frac{m_2 - U\, m_1 +\bar T n^1 + \bar T U\, n^2 }{\sqrt{T_2\, U_2}} \,,
\end{split}
\end{equation}
depend on the K\"ahler modulus $T=T_1 + i T_2$ and the complex structure modulus $U=U_1 + i U_2$ of the two-torus.
$\varGamma_{2,2} (T,U;\tau)$ is modular invariant under ${\rm SL} (2;\mathbb{Z})$, which
requires the modular weight of the Niebur-Poincar\'e series in \eqref{modint} to vanish.
Eq. \eqref{G22ham} corresponds to the so-called Hamiltonian representation of the Narain lattice, and exhibits manifest invariance under the duality group ${\rm O} (2,2;\mathbb{Z} ) \sim {\rm SL} (2;\mathbb{Z})_T \times {\rm SL} (2;\mathbb{Z} )_U \ltimes \sigma_{T,U}$, where the $\mathbb{Z}_2$ generator $\sigma_{T,U}$ exchanges the $T$ and $U$ moduli.
A Poisson summation over the momenta $m_i$ yields the so-called Lagrangian representation
\begin{equation}
\label{G22lag}
\varGamma_{2,2}(T,U;\tau)=
T_2\, \sum_{A\in {\rm Mat}_{2\times 2} (\mathbb{Z})} e^{-2\pi\mathrm{i} T\,\det (A)}
\exp\left[ -\frac{\pi T_2}{\tau_2 U_2} \left| \begin{pmatrix}1 & U \end{pmatrix} \, A \, \begin{pmatrix}- \tau \\ 1 \end{pmatrix} \right|^2 \right]\,,
\end{equation}
which now exhibits manifest invariance under ${\rm SL} (2;\mathbb{Z})_\tau \times {\rm SL} (2;\mathbb{Z})_U \ltimes \sigma_{\tau , U}$. As a consequence, $\varGamma_{2,2}(T,U;\tau)$ is invariant under the action of ${\rm SL} (2;\mathbb{Z})_\tau \times {\rm SL} (2;\mathbb{Z})_T \times {\rm SL} (2;\mathbb{Z})_U \ltimes \sigma_{\tau , T, U}$, where $\sigma_{\tau , T, U}$ permutes the $\tau$, $T$ and $U$ moduli. This triality results in the identities
\begin{equation}
\label{DelTUtau}
\varDelta_T\, \varGamma_{2,2} \equiv \varDelta_U \varGamma_{2,2} \equiv \varDelta_{\tau} \varGamma_{2,2}\ ,\qquad
H_n^{(\tau)} \varGamma_{2,2}\equiv
H_n^{(T)} \varGamma_{2,2} \equiv
H_n^{(U)} \varGamma_{2,2} \, ,
\end{equation}
where $\varDelta_z$ is the weight-zero hyperbolic Laplacian in the variable $z$. Together with the differential equation \eqref{laplEskw}, the first equation in \eqref{DelTUtau} implies that the modular integral \eqref{modint} satisfy
\begin{equation}
\label{lapTU}
\left[\varDelta_T+\tfrac12 s(1-s) \right] \mathcal{I} (s,\kappa)=\left[\varDelta_U+\tfrac12 s(1-s) \right] \mathcal{I} (s,\kappa ) = 12\, \sigma_1(\kappa)\, \delta_{s,1}\,,
\end{equation}
where the source term is due to the subtraction of the simple pole in the integral \eqref{modint} at $s=1$.
Moreover, it suffices to consider the integral $\mathcal{I} (s) \equiv \mathcal{I} (s,1)$ since eqs. \eqref{NPHecke}, \eqref{DelTUtau} together with the self-adjointness of the Hecke operator allows one to express the integral \eqref{modint} for general positive integer $\kappa$ as
\begin{equation}
\mathcal{I} (s,\kappa ) = H^{(T)}_\kappa \, \mathcal{I} (s) = H^{(U)}_\kappa \, \mathcal{I} (s)\,.
\end{equation}
\subsubsection{The BPS state sum representation}\label{BPSPoinca}
A novel procedure for evaluating the integral \eqref{modint} has been discussed in \cite{Angelantonj:2012gw}. It amounts to unfolding the ${\rm SL} (2;\mathbb{Z})$ fundamental domain $\mathcal{F}$ against the absolutely convergent Niebur-Poincar\'e series. One thus arrives at the Schwinger-like representation
\begin{equation}
\mathcal{I} (s ) = \sum_{\rm BPS} \, \int_0^{\mathcal \infty} \frac{\mathrm{d}\tau_2}{\tau_2} \,
{\mathcal M}_{s,0} (- \tau_2 )\, e^{-\frac{\pi\tau_2}{2} (|p_{\rm L}|^2 + |p_{\rm R}|^2)} \,,
\label{BPSint}
\end{equation}
where $\sum_{\rm BPS}$ denotes a sum over integers $m_i, n^i$ subject to the `BPS constraint'\footnote{BPS states exist for any charges such that $m_1 n^1 + m_2 n^2\leq 1$,
so \eqref{BPSconstr} picks only a subset of the allowed BPS states associated to the $q$-pole, which corresponds to the ground state of the non-supersymmetric side of the heterotic string.}
\begin{equation}
m_1 n^1 + m_2 n^2 = 1\ .
\label{BPSconstr}
\end{equation}
Performing the $\tau_2$-integral then yields \cite{Angelantonj:2012gw}
\begin{equation}
\mathcal{I} (s) = \varGamma (s)\,
\sum_{\rm BPS}\,\, \left(\frac{ |p_{\rm L}|^2}{4} \right)^{-s} \,
{}_2 F_1 \left(s, s; 2s \,;\, \frac{4 }{ |p_{\rm L}|^2} \right)\,.
\label{int2F1}
\end{equation}
The sum is absolutely convergent for ${\rm Re} (s)> 1$ and can be analytically continued to a meromorphic function on the complex $s$ plane with a simple pole at $s=1$ \cite{1004.11021, Angelantonj:2012gw}. The renormalisation prescription implies that this pole should be subtracted in defining the value of \eqref{int2F1} at $s=1$. Note that the hypergeometric function reduces to simple elementary functions for integral positive values of $s$ \cite{Angelantonj:2012gw}.
Actually, eq. \eqref{int2F1} provides an explicit Poincar\'e series representation of the automorphic function $\mathcal{I}_{2,2} (s,1)$ of ${\rm O} (2,2;\mathbb{Z})$. To illustrate this point, we note that the BPS constraint \eqref{BPSconstr} allows one to write
\begin{equation}
|p_{\rm L}|^2 = \frac{| \tilde {\bar T} - \tilde U |^2}{\tilde T_2\, \tilde U_2}\,,\qquad {\rm with}\quad
\tilde T = \gamma_T \cdot T\,,
\qquad \tilde U = \gamma_U \cdot U \,,
\end{equation}
where $\gamma_T$ and $\gamma_U$ are both ${\rm SL} (2;\mathbb{Z})$ matrices. One may thus express the integral
\begin{equation}
\mathcal{I} (s) = \sum_{\gamma \in \varGamma_{\rm diag} \backslash (\varGamma_T \times \varGamma_U) } \varphi \left( \frac{|\bar T - U|^2}{T_2 U_2}\right) \Bigg|\, \gamma \,.
\end{equation}
as the ${\rm O} (2,2;\mathbb{Z})$ Poincar\'e series, with seed
\begin{equation}
\varphi (z) = \varGamma (s)\, \left( \frac{z}{4}\right)^{-s}\, {}_2 F_1 \left(s,s;2s; \frac{4}{z}\right) \,,
\label{seedBPS}
\end{equation}
invariant under the diagonal subgroup $\varGamma_{\rm diag}$ of $\varGamma_T \times \varGamma_U$.
It is straightforward to see that the action of the Hecke operator $H_\kappa \, \mathcal{I} (s,1) = \mathcal{I} (s,\kappa )$ amounts to rescaling $|p_L|^2 \to |p_L|^2 /\kappa$ and modifying the BPS constraint to $m_1 n^1 + m_2 n^2 = \kappa$.
The BPS-state-sum representation of the modular integral is thus manifestly invariant under the action of the duality group and is valid throughout the Narain moduli space. In particular, it makes manifest the singularities of the integral \eqref{modint} in the vicinity of points of symmetry enhancement where $|p_{\rm R}|^2\to 0$ for some choice of the integers $m_i,n^i$ \cite{Angelantonj:2012gw}.
If, however, one is interested in the asymptotic behaviour of \eqref{modint} at the boundary of the moduli space, one should instead resort to a Fourier-series representation of the integral.
This is best accomplished at the level of the Schwinger-like integral \eqref{BPSint} as we now show.
\subsubsection{The Fourier series representation}
The BPS constraint \eqref{BPSconstr}
admits an infinite number of solutions which are in one-to-one correspondence with elements of ${\rm SL} (2;\mathbb{Z})$. For any given co-prime pair $(n^1 , n^2)$ the Euclidean algorithm provides a pair of co-prime integers $(m_1 ^\star, m_2^\star)$ solving \eqref{BPSconstr}. The most general solution can thus be obtained as
\begin{equation}
\label{BPSsol}
m_1 = m_1^\star + \tilde M n^2\, ,
\quad
m_2 = m_2^\star - \tilde M n^1\, ,
\end{equation}
for any $\tilde M \in \mathbb{Z}$. Plugging this solution in \eqref{BPSint} and Poisson summing over $\tilde M$, provides the desired Fourier expansion of the integral,
\begin{equation}
\label{intNPunfold}
\begin{split}
{\mathcal I} (s) =& \sum_{M\in\mathbb{Z}}\, \sum_{(n^1,n^2)=1}\, \sqrt{T_2\,\tilde U_2}\ e^{2i \pi M (T_1 - \tilde U_1)}
\\
& \times \int_0^\infty \frac{\mathrm{d}\tau_2}{\tau_2^{3/2}} \,
{\mathcal M}_{s,0} (- \tau_2 )\, \exp\left[-\pi\tau_2 \left( \frac{T_2}{\tilde U_2} + \frac{\tilde U_2}{ T_2}\right)
-\frac{\pi M^2 T_2 \tilde U_2}{\tau_2} \right]
\end{split}
\end{equation}
where
\begin{equation}
\tilde U \equiv \tilde U_1 + i \tilde U_2 = \frac{m_1^\star U-m_2^\star}{n^2 U+n^1} \equiv \gamma\cdot U\,,
\quad {\rm with}\qquad \gamma=\begin{pmatrix} m_1^\star & -m_2^\star \\ n^2 & n^1 \end{pmatrix} \in {\rm SL} (2;\mathbb{Z})\,.
\end{equation}
The integral in \eqref{intNPunfold} can be readily evaluated for any $T_2$ and $\tilde U_2$, using the explicit expression
\begin{equation}
{\mathcal M}_{s,0} (t) = 2^{2s-1}\, \varGamma (s+\tfrac{1}{2})\, (4\pi |t|)^{\frac{1}{2}}\, I_{s-\frac{1}{2}} (2 \pi |t|) \,,\\
\end{equation}
and eq. (16) of Section 8.6 in \cite{erdelyi}\footnote{We reproduce this formula for convenience:
$$
\int_0^{\mathcal \infty} \frac{\mathrm{d} t}{t} e^{-\frac{1}{2t}-\beta t} I_\nu(\gamma t)
= 2\,
K_\nu\left( \sqrt{\beta+\sqrt{\beta^2-\gamma^2}}\right) \,
I_\nu\left( \sqrt{\beta-\sqrt{\beta^2-\gamma^2}}\right) \qquad (\beta >\gamma >0)\,.
$$
}, and reads
\begin{equation}
{\mathcal I} (s) = {\mathcal I}^{(0)} (s) + {\mathcal I}^{(+)} (s) + {\mathcal I}^{(-)} (s)\,,
\end{equation}
with the zero-frequency mode given by
\begin{equation}
\mathcal{I}^{(0)} (s) = 2^{4s-2}\, \sqrt{4\pi} \, \varGamma (s-\tfrac{1}{2}) \, \sum_{\gamma \in \varGamma_\infty \backslash \varGamma} (T_2 \tilde U_2)^s\, \left( T_2 +\tilde U_2 + |T_2 - \tilde U_2 |\right)^{1-2s} \,,
\end{equation}
the positive-frequency modes given by
\begin{equation}
\begin{split}
\mathcal{I}^{(+)} (s) &= \sum_{M>0}\sum_{\gamma \in \varGamma_\infty \backslash \varGamma} \frac{e^{2i \pi M (T_1 - \tilde U_1 )}}{M} \,
\\
&\qquad \times \mathcal{M}_{s,0} \left( \tfrac{1}{2} \, M \left[T_2 + \tilde U_2 - |T_2 - \tilde U_2 | \right] \right)\, \mathcal{W}_{s,0} \left( \tfrac{1}{2}\, M \left[T_2 + \tilde U_2 + |T_2 - \tilde U_2 | \right]\right)\,,
\end{split}
\end{equation}
and the negative-frequency modes simply obtained via complex conjugation, $\mathcal{I}^{(-)} (s) = [ \mathcal{I}^{(+)} (s^\ast)]^\ast$. Notice that we have recognised the sum over the co-prime integers $(n^1 , n^2)$ as the sum over cosets $\varGamma_\infty \backslash \varGamma$ where $\varGamma=\varGamma_U$.
Actually, the Fourier series expansion has different coefficients in the chambers $T_2>U_2$ and $T_2<U_2$. Without loss of generality we can always assume that both $T$ and $U$ lie in the fundamental domain $\mathcal{F}$. Then, whenever $U$ satisfies $U_2 < T_2 $, so do all its images $\tilde U = \gamma U$, for any $\gamma \in \varGamma_\infty \backslash \varGamma$.
On the contrary, when $T_2>1$ all points $U$ with $U_2 > T_2$ are mapped to points $\tilde U$ with $\tilde U_2 < T_2$, except when $\gamma$ is in the coset of the identity.
If $T_2 <1$, then both $U$ and $\tilde U = -1/U$ may satisfy the inequality $\tilde U_2 > T_2$.
As a result, the Fourier coefficients for $\mathcal{I} (s)$ read
\begin{equation}
\mathcal{I}^{(0)} (s) = 2^{2s}\, \sqrt{4\pi} \, \varGamma (s-\tfrac{1}{2}) \, \left[ T_2^{1-s}\, E (s,0;U) +\tfrac{1}{2} \sum_{\gamma = \mathbb{1}, S} \varTheta (\tilde U_2 - T_2 )\, \left( T_2^{s}\, \tilde U_2^{1-s} - T_2^{1-s}\, \tilde U_2 ^s \right) \right]\,,
\label{intzm}
\end{equation}
and
\begin{equation}
\begin{split}
\mathcal{I}^{(+)} (s) &= 2\, \sum_{M>0} \Biggl[ \frac{e^{2i\pi M T_1}}{M}\, \mathcal{W}_{s,0} (M T_2 )\, \mathcal{F} (s,M,0;U)
+ \sum_{\gamma = \mathbb{1}, S} \varTheta (\tilde U_2 - T_2 )
\\
&\quad \times \frac{e^{2 i \pi M (T_1 - \tilde U_1 )}}{M}\, \left( \mathcal{M}_{s,0} (MT_2 )\, \mathcal{W}_{s,0} (M\tilde U_2 ) - \mathcal{W}_{s,0} (M T_2) \, \mathcal{M}_{s,0} (M \tilde U_2 )\right) \Biggr]\, \,.
\end{split}
\label{intpf}
\end{equation}
where $\varTheta (x)$ is the Heaviside function. Note that, upon Fourier expanding the Niebur-Poincar\'e series, the expressions in the two chambers are related by $T\leftrightarrow U$ exchange. Indeed, one could have obtained the very same results by alternatively solving the BPS constraint in terms of pairs of co-prime momenta $(m_1 , m_2)$, which would result in a Fourier expansion with respect to $U$, convergent in the chamber $U_2>T_2$, with coefficients
manifestly invariant under ${\rm SL} (2;\mathbb{Z})_T$.
For convenience, we shall henceforth focus our attention to the fundamental chamber $T_2> U_2$, where the Heaviside function in \eqref{intzm} and \eqref{intpf} vanishes, and the above expressions simplify to
\begin{equation}
\begin{split}
\mathcal{I}^{(0)} (s) &= 2^{2s}\, \sqrt{4\pi} \, \varGamma (s-\tfrac{1}{2}) \, T_2^{1-s}\, E (s,0;U)\,,
\\
\mathcal{I}^{(+)} (s) &= 2\, \sum_{M>0} \frac{e^{2i\pi M T_1}}{M}\, \mathcal{W}_{s,0} (M T_2 )\, \mathcal{F} (s,M,0;U)\,.
\end{split}\label{intFourier}
\end{equation}
Remarkably, using \eqref{NPHecke}, the coefficient of the $M^{\rm th}$ Fourier mode can be seen to involve the Hecke transform of the same function
$\mathcal{F}(s,1,0)$ as the one appearing in the integrand of \eqref{modint}, except that the former
depends on $U$ while the latter depends on $\tau$. This is a manifestation of a fact first
noticed in \cite{Bachas:1997mc}, whose rationale will be explained in Section \ref{sec_unfoldlat}.
While the most relevant case for heterotic string theory is $\kappa =1$, on which we have concentrated so far, the results above may be extended to higher values of $\kappa$ by acting with the Hecke operator. Specifically, one obtains
\begin{equation}
\begin{split}
\mathcal{I}^{(0)} (s,\kappa ) &= 2^{2s} \, \sqrt{4\pi} \, \varGamma(s-\tfrac12) \, (\kappa T_2)^{1-s} \, \sigma_{2s-1}(\kappa)\, E(s,0;U) \,,
\\
\mathcal{I}^{(+)} (s,\kappa ) &= 2 \sum_{d|\kappa} \sum_{M>0}\, \frac{e^{2\pi\mathrm{i} d M T_1}}{M}\, \mathcal{W}_{s,0}(d M T_2) \, \mathcal{F} \left(s,\frac{\kappa M}{d},0;U \right) \,.
\end{split}
\end{equation}
\subsubsection{Unfolding against the Narain lattice \label{sec_unfoldlat}}
An alternative approach for obtaining the Fourier series representation of the modular integral, commonly used in the past string theory literature, is to unfold the fundamental domain against the Narain lattice \cite{O'Brien:1987pn,McClain:1986id, Dixon:1990pc}, rather than against the Niebur-Poincar\'e series. To this end, one resorts to the Lagrangian representation of the $(2,2)$ lattice and decomposes the sum of winding numbers $A$ into orbits of the ${\rm SL} (2;\mathbb{Z})_{\tau}$ action. The vanishing orbit contribution, corresponding to the trivial matrix $A=0$, is absent in this case since the average of any Niebur-Poincar\'e series over the fundamental domain vanishes. The degenerate orbits, corresponding to non-vanishing matrices $A$ with $\det A =0$, precisely reproduce $\mathcal{I}^{(0)}$ in \eqref{intFourier}.
The non-degenerate orbits, associated to the set of matrices
\begin{equation}
A = \begin{pmatrix} k & j \\ 0 & p \end{pmatrix}\,, \qquad 0\le j < k \,, \quad p\not= 0\,,
\end{equation}
can all be reconstructed from the contribution of the unit matrix $A=\mathbb{1}$, via the action of the Hecke operator $H_{\det (A)}$ on the complex structure $U$. In fact,
one can write
\begin{equation}
\mathcal{I}_{\rm non-deg} (s) = 4 \, T_2 \, {\rm Re}\, \sum_{M>0} e^{2i\pi M T} \, H_M^{(U)}\, \int_{\mathbb{C}_+} d\mu\, e^{- \frac{\pi M T_2}{\tau_2 U_2} |U-\tau |^2}\, \mathcal{F} (s,1,0; \tau)\,,
\label{nondegint}
\end{equation}
and upon evaluating the integral over the upper-complex plane one reproduces the non-vanishing frequency contributions to the Fourier expansion \eqref{intFourier}.
In the framework of the old unfolding method, the fact that the Fourier coefficients in \eqref{intFourier} involve the Hecke transform of the same function $\mathcal{F} (s, 1 , 0)$ appearing in the integrand is a consequence of Fay's theorem \cite{zbMATH03547622}, which states that any eigenfunction $G (z)$ of the hyperbolic Laplacian with eigenvalue $s (1-s)$, not necessarily automorphic, has the mean value property
\begin{equation}
\frac{1}{m (\rho ; s)} \int_{\mathbb{C}_+} d\mu \, \rho (r)\, G (z) = G (z_0 )\,.
\label{Fayth}
\end{equation}
Here $r$ is the geodesic distance between the points $z$ and $z_0$, and $\rho (r) $ is an arbitrary `measure' on the complex upper plane $\mathbb{C}_+$ that depends on $r$ but not on the polar angle between $z$ and $z_0$, and the weight factor $m (\rho ; s)$ is given by
\begin{equation}
\label{defmrho}
m (\rho ; s) = 2 \pi \int_0^\infty \, P_s ( \cosh r) \, \rho (r) \, \sinh r \, dr\,,
\end{equation}
with $P_s (t)$ the Legendre function. The integral in \eqref{nondegint} is seen to be a special
case of \eqref{Fayth} with $z=\tau, z_0=U$ and
\begin{equation}
\rho (r) = e^{- \frac{\pi M T_2}{\tau_2 U_2} |\tau - U|^2} = e^{-4\pi M T_2 \sinh^2 \frac{r}{2}} \,. \quad
\end{equation}
The integral \eqref{defmrho} can be computed with the help of 7.141.5 in \cite{GR},
\begin{equation}
m (\rho ; s) = \frac{e^{- 2 \pi M T_2}}{M T_2}\, \mathcal{W}_{s,0} (M T_2 )\ ,
\end{equation}
which thus yields the result \eqref{intFourier}.
\subsection{The special cases of integral $s$}
Let us now focus on the cases $s=1+n$, with $n$ a non-negative integer. For low enough values of $n$ the Niebur-Poincar\'e series $\mathcal{F}(1+n,1,0)$ reduce to weak almost holomorphic modular forms, and their modular integral \eqref{modint} admits alternative representations. The case $s=1$ is special in that $\mathcal{F} (1,1,0) = j (z )+24$ is holomorphic\footnote{Our convention for the Klein $j$-function does not involve a constant term in its Fourier expansion.}, and $\mathcal{I}(1)$ is related to Borcherds' automorphic products. For $n$ large the Niebur-Poincar\'e series are genuinely non-holomorphic. However, as discussed in Section \ref{secNPS}, the linear combinations \eqref{lineardec} still reduce to weak almost holomorphic modular forms, and are thus amenable to the same type of treatment.
\subsubsection{The case $s=1$}
As already noted, for $s=1$ the integral \eqref{modint} has a pole which must be
subtracted \cite{1004.11021, Angelantonj:2012gw}.
This pole originates from the non-holomorphic Eisenstein series $E(s,0;U)$ appearing in \eqref{intFourier}. Using the first Kronecker limit formula one thus obtains
\begin{equation}
\mathcal{I}^{(0)} (1) = - 24 \, \log \left(T_2 U_2\, |\eta (U) |^4 \right) + {\rm const}\, ,
\label{s1zeromode}
\end{equation}
where the additive constant is computable with the renormalisation scheme of \cite{MR656029, Angelantonj:2011br}, but inconsequential for our purposes.
The positive frequency modes are regular at $s=1$ and, using $\mathcal{W}_{1,0}(t)=e^{-2\pi| t|}$, they simplify to
\begin{equation}
\begin{split}
\mathcal{I}^{(+)} (1) &=2 \sum_{M>0} \frac{q_T^M}{M} \mathcal{F} (1,M,0; U )
\\
&= 2 \sum_{M>0} \frac{q_T^M}{M} \left[ \mathcal{F} (1,M,0; U ) - \tilde \mathcal{F}_0 (1,M,0; U) \right] -48\, \log \, \left( q_T^{-1/24}\, \eta (T)\right) \,,
\end{split}
\end{equation}
where we have introduced the nome $q_T$ associated to the K\"ahler modulus $T$, and $\tilde \mathcal{F}_0 (1,M,0) = 24\, \sigma_1 (M)$ is the constant zero-mode of Niebur-Poincar\'e series, see
\eqref{NPFouriermodes}.
Using holomorphy and modularity in the $T$ and $U$ variables, together with the behaviour $\mathcal{F} (1,M,0 ) - \tilde \mathcal{F}_0 (1,M,0) \sim q_U^{-M}$ at the cusp $\infty$, one may show
\begin{equation}
\sum_{M>0} \frac{q_T^M}{M} \left[ \mathcal{F} (1,M,0; U ) - \tilde \mathcal{F}_0 (1,M,0; U) \right] = - \log q_T \left( j (T) - j (U) \right)\,,
\end{equation}
which is simply a reformulation of the Borcherds' product formula \cite{zbMATH00220742}
\begin{equation}
q_T^{-1} \, \prod_{M>0 , N\in \mathbb{Z}} \left( 1 - q_T^M \, q_U^{N} \right)^{c (NM)} = j (T) - j (U) \,,
\end{equation}
where $c (M)$ are the Fourier coefficients of the Klein $j$-function.
Putting things together, the modular integral reads
\begin{equation}
\mathcal{I} (1) = - 24 \, \log (T_2 U_2\, |\eta (T) \, \eta (U) |^4 ) - \log | j (T) - j(U) |^4 + {\rm const}\, .
\label{modint2}
\end{equation}
The first term on the r.h.s. originates from the constant in the decomposition $\mathcal{F} (1,1,0) = j (z ) + 24$ and is the celebrated result of \cite{Dixon:1990pc}, whereas the second term originates from the Klein function and can be uniquely determined by harmonicity, automorphy and by the singularities of the modular integral \cite{LopesCardoso:1994ik, Harvey:1995fq}. These originate from additional massless states appearing at the special points $T = U$ and their ${\rm SL} (2;\mathbb{Z})$ images. Similar automorphic products arise for $\kappa>1$, and can be obtained by acting with the Hecke operators $H_{\kappa}$ in either $T$ or $U$ variables, but have more complicated singular divisors.
\subsubsection{The case $s=1+n$ with $n>0$}
For integer $s=1+n >1$ the Niebur-Poincar\'e series $\mathcal{F} (1+n, 1, 0)$ becomes weakly almost holomorphic. Using the fact
\begin{equation}
\mathcal{W}_{1+n,0} (t) = (-1)^n\, n!\, (4\pi |t|)^{-n}\, e^{- 2\pi | t| }L_n^{(-1-2n)} (4 \pi |t| ) \,,
\end{equation}
with $L_n^{(\alpha )}$ being the associated Laguerre polynomials, we can express
the integral in terms of the `combined polylogarithms'
introduced in \cite{Kiritsis:1997hf}
\begin{equation}
\textrm{L}_{(k)}(z)=\sum_{\ell=0}^{k}\frac{(k+ \ell )!}{\ell!(k-\ell )!(4\pi)^\ell } \, y^{k-\ell}\,\textrm{Li}_{k+\ell +1}(e^{2i \pi z})\, .
\label{SVPL}
\end{equation}
To this end, notice that the non-holomorphic Eisenstein series may be expressed as
\begin{equation}
E (1+n , 0; z) = y^{1+n} + \frac{ \sqrt{\pi} \, \varGamma (n+\frac{1}{2})\, \zeta (2n+1)}{n!\, \zeta (2n+2)}\, y^{-n} + 2 \, {\rm Re}\, \left[ \frac{\pi^{1+n}\, y^{-n}}{ n! \, \zeta (2n+2 )}\sum_{N>0}\, {\rm L}_{(n)} (Nz)\right]\,,
\label{EisenCombPolylog}
\end{equation}
so that $\mathcal{I}^{(0)}$ can be easily written in terms of the single-valued polylogarithms. As for the non-vanishing modes, we recall the Fourier series expansion of the Niebur-Poincar\'e series at the special value $s=1+n$,
\begin{equation}
\mathcal{F} (1+n,1,0 ) = \sum_{N\ge -1} F_n (N)\, (4\pi y)^{-n}\, L_n^{(-1-2n)} (4\pi N y)\, q^N\,,
\end{equation}
where
\begin{equation}
\begin{split}
F_n (-1) &= \varGamma (2n+2 ) \,,
\\
F_n (0) &= \frac{(2 \pi)^{2+2n}\, (-1)^n}{\zeta (2n+2)}\,,
\\
F_n (N>0) &= 4\pi (-1)^n\, \varGamma (2n+2 ) \, \frac{\mathcal{Z} (N,-1 ; 1+n)}{N^n}\,,
\end{split}\label{NPFcoeffs}
\end{equation}
as can be derived from eq. \eqref{NPFouriermodes2} and Appendix A.
Using \eqref{NPFouriermodes2}, one may check that the $F_n(N)$'s are in fact
the Fourier coefficients of
$\mathcal{F}(1+n,1,-2n)=\sum_{N\geq -1} F_n(N) q^N$.
Using the following expression for the product of two associated Laguerre polynomials
\begin{equation}
(n!)^2\, L_n^{(-1-2n)}(u )\, L_n^{(-1-2n)}(v)=
\sum_{k=0}^n \sum_{\ell=0}^k \frac{(n+k)!\, (k+\ell)!}{k!\, (n-k)!\, \ell ! \, (k-\ell)!} (uv)^{n-k}\, (u+v)^{k-\ell}\,,
\end{equation}
one finds
\begin{equation}
\mathcal{I}^{(+)} = \frac{2(-1)^n}{n!}\sum_{M>0}\sum_{N\in\mathbb{Z}}\, F_n (MN)\, \sum_{k=0}^n \frac{(n+k)!}{k!\,(n-k)!}\frac{(MN)^n}{(4\pi MN T_2 U_2 )^k}\, {\rm L}_{(k)} (MT + NU) \,,
\end{equation}
in agreement with \cite{Kiritsis:1997hf}.
Notice that the combination \eqref{SVPL} of the polylogarithms has the property that its real part is a single-valued function in the complex plane \cite{zbMATH04144378}. This ensures that the modular integral be a well-defined function in the fundamental Weyl chamber.
\section{Generalised prepotentials for ${\rm SL}(2;\mathbb{Z})$ \label{sec_prepot}}
Using the behaviour \eqref{NPderiv} of Niebur-Poincar\'e series under the action of modular derivatives,
we show in this section that for integer $s=1+n>1$, the modular integral \eqref{modint} can be expressed in terms of a holomorphic function $f_n(T,U)$, which we refer to as the
generalised prepotential\footnote{This equation holds for $n< 5$ and $n=6$, such that
$\mathcal{F}(1+n,1,-2n)$ is a weak holomorphic modular form. For $n=5$ and $n\geq 7$, $\mathcal{I}(1+n)$ can always
be integrated to a harmonic prepotential. For suitable linear combinations such that the
shadows of $\mathcal{F}(1+n,\kappa,-2n)$ cancel, there is no obstruction in taking the prepotential
to be holomorphic.},
\begin{equation}
\mathcal{I} (1+n) = 4\, {\rm Re}\, \frac{(- D_T D_U )^n}{n! }\, f_n (T,U) \ .
\label{ppotdef}
\end{equation}
The case $n=1$ arises in computations of one-loop threshold corrections to gauge
couplings in heterotic string theory compactified on ${\rm K3}\times T^2$, where $f_1(T,U)$ is identified as the one-loop correction to the usual holomorphic prepotential $F(X)$ encoding the vector-multiplet self-couplings in $\mathcal{N}=2$ supergravity theories in four dimensions \cite{Antoniadis:1995ct,deWit:1995zg,Forger:1997tu}. The case $n=2$ arose in studies of $F^4$ and $R^4$ couplings in heterotic strings on $T^2$ \cite{Lerche:1998nx}, although the physical significance of the prepotential
$f_2$ in the context of eight-dimensional supergravity remains to be elucidated. The generalised
prepotentials $f_n$ were first introduced in \cite{Kiritsis:1997hf}, where integrals of the type \eqref{intintro} were considered with $\varPhi$ being an arbitrary weakly holomorphic modular form. Since $\varPhi$ is in general a mixture of Niebur-Poincar\'e series $\mathcal{F}(1+n,1,0)$ with different values of $n$, the integral \eqref{intintro} becomes a sum of iterated derivatives of various prepotentials $f_n$. One advantage of decomposing $\varPhi$ into a sum of Niebur-Poincar\'e series $\mathcal{F}(1+n,1,0)$ is that it allows one to disentangle these contributions.
While it is interesting and useful that the integral \eqref{modint} can be integrated to a holomorphic function,
the drawback is that this function is ambiguous up to polynomials of degree $2n$ in $T$ and $U$ with real coefficients, as these polynomials lie in the kernel of ${\rm Re}(D_T D_U)^n$. This implies that under T-duality, the prepotential $f_n$ need not be a standard modular form of weight $-2n$, but rather can transform inhomogenously, picking up extra polynomials under the action of
the modular group. This behaviour is characteristic of Eichler integrals. Indeed, we shall see that
$f_n$ contains a $T$-independent term proportional to the Eichler integral of the holomorphic Eisenstein series $E_{2n+2}(U)$, which is responsible for this modular anomaly. Alternatively, one may express the integral as an iterated derivative of a {\it harmonic} function $h_n(T,U)$, which is modular covariant with weight $-2n$ but not holomorphic. This is in fact our first step in establishing
\eqref{ppotdef}.
\subsection{Generalised harmonic prepotentials}
To derive the harmonic prepotentials associated to the integral $\mathcal{I} (1+n)$, we recall the action \eqref{NPderiv} and \eqref{EPderiv} of the modular derivatives on the Niebur and Eisenstein Poincar\'e series, and observe that
\begin{equation}
\begin{split}
D^n \, q^{N} &= (-2N)^n\, \mathcal{W}_{n+1,0}( N y)\, e^{2\pi\mathrm{i} N x} \,,
\\
D^n \, 1 &= \frac{2^{2n}}{\sqrt\pi}\,\varGamma(n+\tfrac12) \, ( - 2\pi y)^{-n} \,,
\end{split}
\label{obscura}
\end{equation}
provided $q^M$ is treated as a generic mode in the Fourier-series expansion of a modular form of weight $-2n$ in $z$. Using \eqref{NPderiv} and \eqref{EPderiv} we may then cast our result \eqref{intFourier} for the integral as the action of the iterated derivatives
\begin{equation}
\mathcal{I} (1+n) = 4\, {\rm Re}\, \frac{(- D_T D_U )^n}{n! }\, h_n (T,U )\,,
\end{equation}
with
\begin{equation}
h_n (T,U) = (2\pi)^{2n+1}\, E(1+n,-2n;U) +
\sum_{N>0} \frac{2}{(2N)^{2n+1}} \, q_T^{N}\, \mathcal{F}(1+n,N ,-2n;U)\,.
\label{fnmaass}
\end{equation}
Since the Eisenstein and Niebur-Poincar\'e series appearing in \eqref{fnmaass} have $s=1-\frac{w}{2}$, the function $h_n(T,U)$ is harmonic in $U$, and holomorphic in $T$, and we shall refer to it as the generalised harmonic prepotential. For low values of $n$,
the non-vanishing frequency modes are in fact also holomorphic in $U$, but the zero-mode
part never is. Importantly, $h_n$ is manifestly modular under the action of ${\rm SL} (2;\mathbb{Z} )_U$, with weight $-2n$. Yet, it is not modular under the full T-duality group,
in particular it is not invariant under $T\leftrightarrow U$. Note that $h_n$ is a priori
defined up to terms in the kernel of ${\rm Re}(D_T D_U)^n$. Our choice ensures that
$h_n$ is modular under ${\rm SL} (2;\mathbb{Z} )_U$.
As before, this result generalises to the case of $\kappa > 1$ by the action of the Hecke operator on either $T$ or $U$. Specifically,
\begin{equation}
h_n (T,U;\kappa ) = \kappa^{-n} \, H_\kappa \cdot h_n (T,U)\,.
\label{prepHecke}
\end{equation}
The overall factor is a consequence of the identity
\begin{equation}
\kappa^n \, H_\kappa \, D^n \, \mathcal{F} (s,1,w) = D^n \, \mathcal{F} (s,\kappa , w )\,.
\end{equation}
Notice, that generalised {\em harmonic} prepotentials can always be introduced independently of the value of $n$, even when the Niebur-Poincar\'e series in the integrand is no longer weak almost holomorphic but is rather genuinely non-holomorphic, which occurs when $n=5$ and $n>6$.
\subsection{Generalised holomorphic prepotentials and Eichler integrals\label{sec_eichler}}
As we have seen, the harmonic prepotential $h_n$ emerges rather naturally and has the advantage of being modular covariant, at the cost of sacrificing holomorphy. Although this is legitimate from a mathematical viewpoint, it is however difficult to reconcile with the holomorphic structure of $\mathcal{N}=2$ supergravity, since $h_1$ should compute the radiative correction to the holomorphic prepotential $F(X)$. As we shall see, this is not a problem and holomorphy may always be restored, whenever the Niebur-Poincar\'e series in the integrand {\em is} weakly almost holomorphic.
For $n=1,2,3,4,6$ the only source of non-holomorphy in the harmonic prepotential is due to the zero-frequency mode $\mathcal{I}^{(0)}$. As a result, we shall first restrict our analysis to this contribution. Using \eqref{EPderiv} one has
\begin{equation}
E (1+n,0) = \frac{(2\pi)^n}{n!}\, D^n \, E (1+n , -2n) = \frac{(2\pi)^n}{n!}\, {\rm Re}\, D^n \, E (1+n , -2n) \,,
\end{equation}
where $E (1+n ; -2n)$ is a harmonic Maa\ss\ form. As a result, it can be decomposed
\begin{equation}
E (1+n ; -2n ) = \varPhi_{\rm h} + \varPhi_{\rm nh}\,,
\end{equation}
into its holomorphic (Mock modular) part $\varPhi_{\rm h} $ and the non-holomorphic complement $ \varPhi_{\rm nh}$, with
\begin{equation}
\varPhi_{\rm h} (z) = \left( \frac{z}{2i}\right)^{2n+1} + \frac{(-1)^n\, \pi \, \zeta (2n+1)}{2^{2n+1}\, \zeta (2n+2)}
+\frac{ (-1)^n\, \pi}{2^{2n}\, \zeta (2n+2)} \sum_{N>0} \sigma_{-2n-1} (N) \, q^N\,,
\end{equation}
and
\begin{equation}
\begin{split}
\varPhi_{\rm nh} (x,y) &= y^{2n+1} - \left( \frac{z}{2i}\right)^{2n+1} + \frac{(-1)^n\, \pi \, \zeta (2n+1)}{2^{2n+1}\, \zeta (2n+2)}
\\
&\quad +\frac{ (-1)^n\, \pi}{2^{2n}\, \zeta (2n+2)} \sum_{N>0} \sigma_{-2n-1} (N) \, \frac{\varGamma (2n+1 ; 4\pi N y )}{\varGamma (2n+1)} q^{-N}\,.
\end{split}
\end{equation}
The remarkable identities
\begin{equation}
\begin{split}
{\rm Re} \, \left[ D^n \left( \frac{z}{2i}\right)^{2n+1} \right] &= D^n \, y^{2n+1} \,,
\\
D^n \, \left[ \varGamma (2n+1;4 \pi N y)\, q^{-N} \right] &= \frac{(2n)!}{(4\pi N)^n}\, \left[ D^n \, q^N \right]^* \,,
\end{split}
\end{equation}
together with equations \eqref{obscura} yield the non-trivial result
\begin{equation}
D^n \, \varPhi_{\rm h} (z) = \left[ D^n \, \varPhi_{\rm nh} (x,y) \right]^* \, ,
\end{equation}
which originates from the fact that both the shadow and the ghost \cite{Angelantonj:2012gw} of the
harmonic Maa\ss\ form $E(n+1,-2n)$ are proportional to the holomorphic Eisenstein series $E_{2n+2} (z)$. In fact, one can prove that
\begin{equation}
D^{2n+1}\, \varPhi_{\rm h} = \left[ (D^*)^{n+1} \, D^n \, \varPhi_{\rm nh} \right]^* = y^{-2-2n}\, \left[ \bar D \, \varPhi_{\rm nh}\right]^*\,,
\end{equation}
where the left-hand side defines the ghost and the right-hand side defines the shadow.
The above considerations allow one to write
\begin{equation}
\mathcal{I}^{(0)} (n+1) = \frac{4}{n!} \, {\rm Re} \, (- D_T D_U )^n \, f_n^{(0)} \,,
\end{equation}
where
\begin{equation}
\label{fn0eich}
f_n^{(0)} (U) = \alpha\, \tilde E_{-2n} (U)\,,
\end{equation}
is the zero-frequency mode of the desired generalised {\rm holomorphic} prepotential, written in terms of
\begin{equation}
\tilde E_{-2n} (z) = \frac{\zeta (2n+2 )}{2\pi i }\, z^{2n+1} + \frac{\zeta (2n+1)}{2} + \sum_{N>0} \sigma_{-1-2n} (N) q^N\,,
\end{equation}
and
\begin{equation}
\alpha = \frac{(-1)^n \, (2\pi )^{2n+2} }{2^{2n}\, \zeta (2n+2)}\,.
\label{alphaConst}
\end{equation}
One may recognise $\tilde E_{-2n}(z)$ as the Eichler integral of the holomorphic Eisenstein series $E_{2n+2} (z)$ \cite{0990.11041}, defined by
\begin{equation}
E_{2n+2} (z) = \frac{2\pi\mathrm{i}}{(2n+1)! \, \zeta(2n+2)} \partial_z^{2n+1} \tilde E_{-2n}(z)\ .
\end{equation}
As anticipated, this holomorphic contribution does not transform covariantly under the action of the modular group, but rather involves an inhomogeneous term, as we shall discuss in the following Section. Altogether, the generalised {\em holomorphic} prepotential reads
\begin{equation}
f_n (T,U) = \alpha\, \tilde E_{-2n} (U) + \sum_{N>0} \frac{2}{(2N)^{2n+1}} q_T^{N}\, \mathcal{F}(1+n,N ,-2n;U)\,.
\label{holpp}
\end{equation}
It is important to stress that this choice of holomorphic prepotential is not unique and one has the freedom of adding a polynomial of degree $2n$ in $T$ and $U$ with real coefficients. In \eqref{holpp} we have made a convenient choice for this polynomial.
Upon Fourier-expanding $\mathcal{F}(1+n,1 ,-2n)=\sum_{N\geq -1} F_n(N) q^N$, the generalised holomorphic prepotential
admits the alternative representation
\begin{equation}
f_n (T,U) = \alpha\, \tilde E_{-2n} (U) +2^{-2n}\, \sum_{N>0 \atop M\in \mathbb{Z}} F_n (NM)\, {\rm Li}_{2n+1} (q_T^N \, q_U^M )
\label{holppLi}
\end{equation}
in terms of polylogarithms. By further Fourier expanding the Eichler integral and rearranging terms, one arrives at
\begin{equation}
\begin{split}
f_n(T,U) =&\alpha\,\zeta(2n+2)\left[\frac{U^{2n+1}}{2\pi i}-\tfrac{1}{2}\mathcal{Z} (0,0; 1+n)\right]+\frac{F_n (-1)}{2^{2n}}\,{\rm Li}_{2n+1}\left(\frac{q_T}{q_U}\right) \\
& +\frac{1}{2^{2n}} \sum_{N,M\geq 0} F_n(NM)\,{\rm Li}_{2n+1}(q_T^N \,q_U^M) \,,
\end{split}
\label{SL2ZprepPoly}
\end{equation}
which exposes the asymmetry of the generalised holomorphic prepotentials under $T\leftrightarrow U$ exchange, originating from the singularity at $T=U$ in the term proportional to $F_n (-1)$.
Although the zero-frquency part $\mathcal{I}^{(0)}$ may always be expressed in terms of a holomorphic function $f^{(0)}_n$
for any $n$, this is no longer true for the non-vanishing frequency modes when $n=5$ and $n>6$, since the Niebur-Poincar\'e series $\mathcal{F} (1+n,N , -2n)$ are genuine harmonic Maa\ss\ forms \cite{Angelantonj:2012gw}. This is related to the fact that, for these values of $n$, $\mathcal{F} (1+n,1,0)$ is a genuine non-holomorphic function. However, one can always associate a generalised holomorphic prepotential to the linear combination \eqref{lineardec}, representing weak almost holomorphic modular forms, since the shadows will cancel also in the prepotential.
Similarly to the harmonic case, the generalised holomorphic prepotential $f_n (T,U;\kappa )$ may be straightforwardly obtained via the action of the Hecke operator $H_\kappa$ as in eq. \eqref{prepHecke}.
\subsection{Modular properties of generalised holomorphic prepotentials\label{sec_modeich}}
As anticipated, the generalised holomorphic prepotential of eq. \eqref{holpp} does not transform homogeneously with weight $-2n$ under the action of the T-duality group. Rather, owing to the inherent polynomial ambiguity in $f_n$, one has
\begin{equation}
\begin{split}
(c T+d)^{2n}\, (c' U+d')^{2n}\,
f_n\left(\frac{a T+b}{c T+d},
\frac{a' U+b'}{c' U+d'} \right) =& f_n(T,U) + P_{\gamma,\gamma'}(T,U) \, ,
\\
f_n(T,U) =& f_n (U,T) + P_\sigma(T,U)\,,
\end{split}
\end{equation}
where $P_{\gamma,\gamma'}(T,U)$ and $P_\sigma(T,U)$ are polynomials of
degree $2n$ in $(T,U)$ with real coefficients. These polynomials themselves depend on the choice of
polynomial ambiguity in $f_n$, which we have fixed in the definition \eqref{holpp}.
It is important to note that there is no choice of this ambiguity that can trivialise all polynomials
$P$'s.
As far as
the action of ${\rm SL}(2;\mathbb{Z})_U$ is concerned, the anomaly can be traced to the presence of the
Eichler integral $\tilde E_{-2n}(U)$ in the zero-th Fourier mode \eqref{fn0eich} with respect to $T$. Indeed,
under a generic element $\gamma \in {\rm SL} (2; \mathbb{Z})$,
the Eichler integral transforms as
\begin{equation}
(c z + d)^{2n}\, \tilde E_{-2n} \left( \frac{az+b}{cz+d}\right) = \tilde E_{-2n} (z) + P_\gamma (z)\,,
\label{eichler1}
\end{equation}
with $P_\gamma (z)$ being a degree $2n$ polynomial in the $z$ variable, that depends on the choice of $\gamma$.
The polynomials $P_\gamma (z)$ are uniquely determined by $P_T (z)$ and $P_S (z)$, which are associated to the standard generators $T$ and $S$ of the modular group given in \eqref{defTS}. It is straightforward to derive $P_T$ since the nome $q$ is trivially invariant and the only source of inhomogeneity is the monomial $z^{2n+1}$. As a result,
\begin{equation}
P_T (z) = \frac{\zeta (2n+2 )}{2\pi i }\, \sum_{k=0}^{2n} {2n+1 \choose k} z^k \,.
\label{PT}
\end{equation}
The transformation under $S$ is more involved and can be determined by the analytic properties of the $L$-series
\begin{equation}
L^\star (s) = \int_0^\infty d y\, y^{s-1} \sum_{N>0} \sigma_{-1-2n} (N) e^{-2\pi N y} = (2 \pi )^{-s} \, \varGamma (s)\, \zeta (s) \, \zeta (s+2n+1)\,,
\end{equation}
associated to $\tilde E_{-2n}$, as we derive in Appendix \ref{appEichler}. Notice that although the $L$-series satisfies the functional equation $s \to w-s$, with $w=-2n$, and has simple poles at $s=0$ and $s=-2n$
as in the case for modular forms of weight $-2n$, the additional poles at $s=1-2p$ with $p=0, 1 , \ldots , n+1$ are actually responsible for the inhomogeneous transformation \eqref{eichler1}, and the associated residues uniquely fix
\begin{equation}
P_S (z) = - \frac{(2\pi i )^{2n+1}}{2} \sum_{k=1}^n \frac{B_{2k}\, B_{2 n - 2k +2}}{(2k)!\, (2n-2k+2)!}\, z^{2k-1}\,,
\label{PS}
\end{equation}
thus generalising the result of \cite{0990.11041} to generic $n$.
Clearly, $P_T$ and $P_S$ determine the behaviour of the generalised holomorphic prepotential $f_n (T,U)$ under the action of ${\rm SL} (2; \mathbb{Z})_U$. Moreover, using the property of the polylogarithms
\begin{equation}
{\rm Li}_k (e^{2 i \pi z}) + (-1)^k \, {\rm Li}_k (e^{-2i\pi z} ) = - \frac{ (2\pi i )^k}{k!} B_k (z) \,, \qquad 0 < {\rm Re}\, (z) <1\,,
\label{analyticContPoly}
\end{equation}
with $B_k (z)$ the Bernoulli polynomials, one obtains
\begin{equation}
P_\sigma (T,U) = - \frac{(2\pi i )^{2n+1}}{2^{2n}}\, \left[ B_{2n+1} (T-U) - T^{2n+1}+ U^{2n+1} \right]
\,,
\label{Psigma}
\end{equation}
when $T$ and $U$ lie in the fundamental chamber, with real parts in the interval $(0,1)$.
This expression generalises the result of \cite{Harvey:1995fq} to $n>1$.
The knowledge of the polynomials $P_S$, $P_T$ and $P_\sigma$ fully determines the transformation of the prepotential under the duality group, because any transformation $g' \in {\rm SL} (2;\mathbb{Z})_T$ can be obtained by conjugating the associated transformation $g\in {\rm SL} (2;\mathbb{Z})_U$ by $\sigma$, {\em i.e.} $g' = \sigma^{-1} g \sigma$. However, it is important to note that, since $P_{\sigma^2} \not=0 $, the generator $\sigma$ is no longer an involution at the quantum level, and as a result, modular transformations $g$ and $g'$ no longer commute. The group generated by these transformations, which we call the monodromy group, is an extension of the classical T-duality
group ${\rm SL}(2;\mathbb{Z})_T\times {\rm SL}(2;\mathbb{Z})_U\times \sigma$ given by the braid group of $(X\times X)\backslash \varDelta$, where $X$ is the Riemann sphere with three punctures located at images of $e^{i\pi /2}$, $e^{i\pi/3}$ and $i\infty$ under the action of the Klein $j$ function, and $\varDelta$ is the diagonal divisor in $X\times X$. The braid group is generated by the elements $g_1 :\ U \to -1/U$, $g_2:\ U\to -1/(U+1)$ and $\sigma$, obeying the relations
\begin{equation}
g_1 ^2 = g_2^3 = \mathbb{1}\,, \qquad g_i \sigma g_j \sigma^{-1} = \sigma g_j \sigma^{-1} g_i\,.
\end{equation}
Using the results above, we find that the generators act on the generalised holomorphic prepotential as
\begin{equation}
\begin{split}
f_n \left(T, -\frac{1}{U} \right) &= U^{-2n}\, \left[ f_n (T,U) + \alpha P_S (U) \right]\,,
\\
f_n \left(T,-\frac{1}{U+1}\right) &=(U+1)^{-2n} \, \left[ f_n (T,U) + \alpha \left(P_{T} (U) + P_S (U+1) \right) \right]\,,
\\
f_n (T,U) &= f_n (U,T ) + P_\sigma (T,U)\,,
\\
f_n \left(-\frac{1}{T},U\right) &= T^{-2n} \, \left[ f_n (T,U) + \alpha P_S (T) - P_\sigma (T,U) +T^{2n} \, P_\sigma \left( -\frac{1}{T}, U\right) \right]\,,
\\
f_n \left(-\frac{1}{T+1},U\right) &= (T+1)^{-2n} \, \biggl[ f_n (T,U) + \alpha \left( P_S (T+1) + P_T (T) \right)
\\
&\qquad\qquad\qquad\qquad - P_\sigma (T,U) +(T+1)^{2n} \, P_\sigma \left( -\frac{1}{T+1}, U\right) \biggr]\,,
\end{split}
\end{equation}
with $P_T$, $P_S$ and $P_\sigma$ the polynomials given in eqs. \eqref{PT}, \eqref{PS} and \eqref{Psigma}.
For $n=1$, our analysis is in full agreement with the results of \cite{Harvey:1995fq, Antoniadis:1995ct}
and extends it to higher values of $n$. The precise matching with \cite{Antoniadis:1995ct} at the level of the holomorphic prepotential is obtained by adding to $f_1 (T,U)$ the polynomial $3 + 5 U + 3 U^2$ that lies in the kernel of ${\rm Re} \, (-D_T D_U)$.
\subsection{Generalised Yukawa couplings \label{sec_yuk}}
In $\mathcal{N}=2$ supergravity in four dimensions, the Yukawa couplings $Y_{ijk}$ are obtained from the third derivatives of the prepotential $F$. They are modular covariant, and, in the case of $i=j=k=T$ or $U$, holomorphic in $(T,U)$. Similarly, for heterotic strings compactified on $T^2$, the one-loop contributions to $F^4_{\mu\nu}$ couplings are fifth derivatives of the generalised prepotential $f_2$. In this subsection, we construct a similar set of functions $Y_{T^p U^q}$ with $p+q=2n+1$, which we
call generalised Yukawa couplings, which are proportional to $(2n+1)^{\rm th}$ derivatives of the generalised prepotential
$f_n$, independent of the polynomial ambiguity, holomorphic when $p=0$ or $q=0$, modular covariant for any $p,q$, and related to one-loop modular integrals with momentum insertions. The physical significance of these functions is not clear at present, but we hope that they may provide hints towards a generalisation of the notion of special geometry to $n>1$.
We shall begin by studying the simplest instance of holomorphic generalised Yukawa couplings
\begin{equation}
Y_{T^{2n+1}} (T,U) = - 2^{-2n-1}\, D_T^{2n+1} \, f_n (T,U) =\left( \frac{\partial_T}{2\pi i}\right)^{2n+1}\, f_n (T,U) \,,
\end{equation}
where in the last equality we used Bol's identity \cite{Bol}, which equates the $(1-w)$-th modular derivative of a modular function of negative integer weight $w<0$ to the ordinary $(1-w)$-th derivative. Notice that, since the degree $2n$ polynomial in the inhomogeneous transformation of $f_n$ is annihilated by the derivatives, $Y_{T^{2n+1}}$ indeed transforms as a weight $(2n+2,-2n)$ holomorphic modular form with respect to ${\rm SL} (2;\mathbb{Z})_T \times {\rm SL} (2;\mathbb{Z})_U$, for $n$ sufficiently small. Inserting the expressions \eqref{holpp} and \eqref{holppLi}, one obtains
\begin{equation}
\begin{split}
Y_{T^{2n+1}} (T,U) &= 2^{-2n} \sum_{N>0} q_T^N\, \mathcal{F} (1+n , N, -2n ; U)
\\
&= 2^{-2n} \sum_{N>0 \atop M\in \mathbb{Z}} N^{2n+1} \, F_n (MN)\, \frac{q_T^N \, q_U^M}{1-q_T^N \, q_U^M}\,,
\end{split}
\end{equation}
thus generalising the `multi-cover' formula for the standard Yukawa coupling $Y_{T^3}$ \cite{Candelas:1990rm}. This expression clearly exhibits a simple pole divergence at $q_T = q_U$,
\begin{equation}
Y_{T^{2n+1}} \sim \frac{(2n+1)!}{2^{2n}}\, \frac{q_T}{q_U - q_T}\,.
\end{equation}
This behaviour, together with holomorphy and modularity, uniquely fixes the expression of the Yukawa couplings,
\begin{equation}
\begin{split}
Y_{T^3} &= \frac{3}{2} \, \frac{E_4 (T) \, E_4 (U) \, E_6 (U)}{\varDelta (U)\, [ j(T) - j(U)]} \,,
\\
Y_{T^7} &= \frac{315}{4} \, \frac{E_4^2 (T) \, E_6 (U)}{\varDelta (U)\, [ j(T) - j(U)]}\,,
\end{split}
\qquad
\begin{split}
Y_{T^5} &= \frac{15}{2} \, \frac{E_6 (T) \, E_4^2 (U) }{\varDelta (U)\, [ j(T) - j(U)]} \,,
\\
Y_{T^9} &= \frac{2835}{2} \, \frac{E_4 (T) \, E_6 (T) \, E_4 (U)}{\varDelta (U)\, [ j(T) - j(U)]} \,,
\end{split}
\end{equation}
which may be written compactly as
\begin{equation}
Y_{T^{2n+1}} (T,U) = 2^{-2n}\, \frac{E_{2n+2} (T) \, \mathcal{F} (1+n,1,-2n;U)}{j (T) - j (U)}\,,
\label{YukawasSL2Z}
\end{equation}
in terms of the holomorphic Eisenstein and Niebur-Poincar\'e series with $s=1-\frac{w}{2}$.
Note that the $Y_{T^{2n+1}}$ have already been discussed in the literature in the context of heterotic string compactifications on ${\rm K3} \times T^2$ and on $T^2$ \cite{ Lerche:1998nx,
Lerche:1999hg, Antoniadis:1995zn}, respectively. Similar results can be obtained for $Y_{U^{2n+1}}$ upon exchanging $T$ and $U$ in the above expressions.
The holomorphic generalised Yukawa couplings can be related to modular integrals with lattice-momentum insertions. In fact,
\begin{equation}
D_T^{n+1} \, \bar D _U^n \, \mathcal{I} (1+n) = (-1)^{n+1} \frac{4 (2n)!}{n!}\, Y_{T^{2n+1}} (T,U) + \frac{2}{n!}\, D_T^{n+1} \, \bar D_U^n \,\left[ (- D_{ T} D_{ U} )^n f_n ( T , U )\right]^* \,,
\end{equation}
where we have used again Bol's identity and the fact that $\bar D^n D^n $ is proportional to the identity operator when acting on Niebur-Poincar\'e series. The second term on the r.h.s. actually vanishes due to holomorphy of $f_n$ and of non-trivial combinatorial identities. Moreover, one can trade the covariant derivatives for insertions of lattice momenta to obtain
\begin{equation}
Y_{T^{2n+1}} = \frac{(-1)^{n+1}}{2^{4n+3}}\, \frac{U_2^n}{T_2^{n+1}}\, \int_\mathcal{F} d\mu \, \tau_2 \, \langle p_{\rm L}^{2n+1}\, \bar p_{\rm R} \rangle \, \mathcal{F} (1+n , 1 , -2n )\,.
\end{equation}
Here we have employed the notation
\begin{equation}
\langle f(p) \rangle = \tau_2 \, \sum_{m_i,n^i} f (p) \, q^{\frac{1}{4} |p_{\rm L}|^2 }\, \bar q^{\frac{1}{4} |p_{\rm R}|^2 }
\end{equation}
for the Narain lattice with generic momentum insertion $f (p)$.
Also in this case, the generalisation to higher values of $\kappa$ is straightforward and can be obtained via the action of the Hecke operator
\begin{equation}
Y_{T^{2n+1}} (T,U;\kappa ) = \kappa^{-n} \, H_\kappa \cdot Y_{T^{2n+1}} (T,U)\,.
\end{equation}
We can now move to discuss more general Yukawa couplings, which can be obtained as mixed derivatives of the generalised holomorphic prepotential. Aside from sporadic cases \cite{Lerche:1998nx, Foerger:1998kw, Lerche:1999hg}, these have not been extensively studied in the literature, and indeed there are ambiguities in their proper definition. In these cases Bol's identity is no longer applicable and modularity dictates a non holomorphic completion. For instance, in the combination
\begin{equation}
Y_{T^p U^q} (T,U;\kappa )= - \frac{D_U^q}{ 2^{2n+1}} \left[ D_T^p f_n (T,U;\kappa )- \frac{(2\pi)^{2q-2}\, p!}{(2n)!\, (q-1)!} \frac{ U_2^{2n}}{T_2^{2n-2q+2}} \left( D_T^{q-1} D_U^{2n}\, f_n (T,U;\kappa )\right)^* \right] ,
\label{mixedY}
\end{equation}
with $p+q=2n+1$ and $1 \le q \le n <p$, the degree $2n$ polynomial in the inhomogeneous transformation of $f_n$ is exactly cancelled. After some tedious algebra, one may cast them as modular integrals with lattice momentum insertions
\begin{equation}
Y_{T^p U^q} = (-1)^{n+1} \frac{\kappa^q\, q!}{2^{4n -2p +1}}\, \frac{U_2^{n-q}}{T_2^{p-n}}\, \int_\mathcal{F} d\mu \, \tau_2 \, \langle p_{\rm L}^{p-q}\, \bar p_{\rm R}\rangle \, \mathcal{F} (1+n , \kappa , 2q-2n )\,.
\end{equation}
These mixed Yukawa couplings are clearly non-holomorphic functions, but have holomorphic weights $2p-2n$ and $2q-2n$ with respect to ${\rm SL} (2;\mathbb{Z})_T$ and ${\rm SL} (2;\mathbb{Z})_U$, respectively.
Furthermore, observe that the quantity inside the brackets in eq. \eqref{mixedY} is itself modular covariant and may be expressed as the modular integral
\begin{equation}
(-1)^n \frac{(2\pi)^{2q} \,\kappa^{q-n-1}\, (2p)!}{2^{4n-q+4}\, n!}\, \frac{U_2^n}{T_2^{p-n}} \int_\mathcal{F} d\mu\, \tau_2^{q-1}\, \langle p_{\rm L}^p\, p_{\rm R}^{q-1}\rangle\, \mathcal{F} (1+n, \kappa, q-p-1) \,.
\end{equation}
This is a non-trivial result, and allows one to straightforwardly compute the Fourier series expansions of such integrals by expressing them as derivative of the known generalised prepotential.
\section{Modular integrals for $\varGamma_0 (N)$\label{HeckeIntegrals_sec}}
In perturbative computations in many string theory models, one encounters one-loop integrals over the fundamental domain of various congruence subgroups rather than of the full modular group. This happens {\em e.g.} in orbifold compactifications involving non-factorisable tori \cite{Mayr:1993mq,Bailin:2014nna} or when supersymmetry is spontaneously broken, either partially or completely \cite{Kiritsis:1998en, Kiritsis:1997ca, Angelantonj:2006ut, Angelantonj:2014dia, Faraggi:2014eoa}. In these cases, the one-loop amplitudes can be expressed as
\begin{equation}
\mathcal{I} = \int_\mathcal{F} d\mu \, \sum_{h,g} \mathcal{A} \left[{\textstyle{h\atop g}} \right] \,,
\label{OrbiInt}
\end{equation}
where the sum over $h$ runs over the untwisted and twisted orbifold sectors and the sum over $g$ imposes the orbifold projection. Clearly, the integrand is invariant under the full modular group, but each independent term in the sum is typically invariant only under some congruence subgroup of ${\rm SL} (2;\mathbb{Z})$, except for the untwisted unprojected sector $(h,g)=(0,0)$ which is invariant under the full modular group.
This suggests that an efficient way of evaluating \eqref{OrbiInt} is to partially unfold the modular integral \cite{Angelantonj:2013eja}, {\it i.e.} collect the $(h,g)$ sectors into orbits of ${\rm SL}(2;\mathbb{Z})$ and keep one element in each orbit, now integrated over the fundamental domain of the corresponding congruence subgroup. Schematically,
\begin{equation}
\sum_{h,g} \mathcal{A} \left[{\textstyle{h\atop g}} \right] = \sum_{a} \sum_{\gamma \in \varGamma_a \backslash \varGamma } \mathcal{A}_a \, |\, \gamma
\end{equation}
where the amplitudes $\mathcal{A}_a $ are now invariant under the congruence subgroup $\varGamma_a \subset \varGamma$, and the sum over $a$ runs over the possible subgroups. As a result,
\begin{equation}
\mathcal{I} = \sum_a \mathcal{I}_a = \sum_a \int_{\mathcal{F}_a} d\mu \, \mathcal{A}_a\,,
\end{equation}
$\mathcal{F}_a$ being the fundamental domain of $\varGamma_a$. The integrals $\mathcal{I}_a$ can then be computed by employing similar methods as in the case of the full modular group \cite{Angelantonj:2013eja}.
For simplicity, we shall restrict our analysis to the case of abelian $\mathbb{Z}_2$ and $\mathbb{Z}_3$ orbifolds, which play an important role in string vacuum constructions.
These cases are special in that they involve a single Hecke congruence subgroup, $\varGamma_0 (2)$ and $\varGamma_0 (3)$, respectively. Other $\mathbb{Z}_M$ and non-abelian orbifolds can be dealt with similar methods, though they typically involve several orbits.
We shall further focus on BPS-saturated amplitudes, where the orbifold
blocks are of the form
\begin{equation}
\mathcal{A} \left[{\textstyle{h\atop g}}\right]=\frac{1}{N} \varGamma_{2,2} \left[{}^h_g \right] (T,U)
\varPhi \left[{}^h_g\right] \,,
\end{equation}
where $\varPhi \left[{}^h_g\right] $ is a weakly almost holomorphic modular form under $\varGamma_0(N)$ and
\begin{equation}
\varGamma_{2,2} \left[ {}^h_g\right] (T,U,\tau) = \tau_2 \sum_{m_i , n^i \in \mathbb{Z}} e^{\frac{2\pi i g}{N} (\lambda_i n^i + \mu^i m_i )}\, q^{\frac{1}{4} |p_{\rm L} (h)|^2}\, \bar q^{\frac{1}{4} |p_{\rm R} (h)|^2} \,,
\end{equation}
is the shifted Narain lattice partition function with momenta
\begin{equation}
\begin{split}
p_{\rm L} (h) &= \frac{1}{\sqrt{T_2 U_2 }} \left[ m_2 (h) - U m_1 (h) + \bar T (n^1 (h) + U n^2 (h)) \right]\,,
\\
p_{\rm R} (h) &= \frac{1}{\sqrt{T_2 U_2 }} \left[ m_2 (h) - U m_1 (h) + T (n^1 (h) + U n^2 (h)) \right]\,,
\end{split}
\label{shiftedmomenta}
\end{equation}
where $m_i (h) = m_i + \frac{h}{N} \lambda_i$, $n^i (h) = n^i + \frac{h}{N} \mu^i$. The integral shift vector $(\lambda_i , \mu^i)$ is constrained by modular invariance to satisfy $ \lambda_i \mu^i = 0 \ \mod N$.
For $\mathbb{Z}_N$ orbifolds with $N=2,3$, partial unfolding leads to
\begin{equation}
\mathcal{I} = \frac{1}{N} \int_\mathcal{F} d\mu \, \varGamma_{2,2}\left[ {}^0_0 \right] (T,U)\, \varPhi \left[{}^0_0\right] + \frac{1}{N} \sum_{g=1}^{N-1} \int_{\mathcal{F}_N}d\mu \, \varGamma_{2,2} \left[{}^0_g \right] (T,U) \varPhi \left[{}^0_g\right] \,,
\end{equation}
where $\mathcal{F}_N$ is the fundamental domain of $\varGamma_0 (N)$. The relation $\varPhi [{}^h_g] = \varPhi [{}^{N-h}_{N-g}]$, valid for conjugate orbifold sectors, allows one to write
\begin{equation}
\mathcal{I} = \frac{1}{N} \int_\mathcal{F} d\mu \, \varGamma_{2,2}\left[ {}^0_0 \right]\, \left[ \varPhi \left[{}^0_0\right] - \sum_{\gamma \in \varGamma_0 (N)\backslash \varGamma} \varPhi [{}^0_1 ]\, \Big| \gamma \right] +
\int_{\mathcal{F}_N}d\mu \, \varPhi \left[{}^0_1\right]\,
\frac{1}{N} \sum_{g=0}^{N-1} \varGamma_{2,2} \left[{}^0_g \right] \,.
\label{integralFullCongr}
\end{equation}
Although this way of writing may appear contrived, it is actually instrumental in the subsequent analysis, since the sum over $g$ enforces a $\mathbb{Z}_N$ projection on the momenta and/or windings.
The term in the square bracket is ${\rm SL} (2;\mathbb{Z})$ invariant, and therefore the first integral can be straightforwardly evaluated using the results of \cite{Angelantonj:2011br,Angelantonj:2012gw} and of Sections \ref{sec_intsl2z}, \ref{sec_prepot}. The second integral instead can be treated using the techniques in \cite{Angelantonj:2013eja}. Since any weak almost holomorphic modular form of $\varGamma_0 (N)$ can be uniquely decomposed in terms of a linear combination of Niebur-Poincar\'e series \cite{Angelantonj:2013eja}, we shall restrict our analysis to the prototype integral involving the shifted lattice times the Niebur-Poincar\'e series of $\varGamma_0 (N)$.
\subsection{Symmetries of the shifted Narain lattice}
To study the symmetries of the shifted Narain lattice, it is convenient to introduce the matrices
\begin{equation}
M = \begin{pmatrix} n^1 & m_2 \\ n^2 & - m_1
\end{pmatrix}
\,,
\qquad
\varLambda = \begin{pmatrix} \mu_1 & \mu_2 \\ \lambda^2 & - \lambda^1
\end{pmatrix}\,,
\end{equation}
that encode the momentum and winding numbers, and the shift vectors. The shifted Narain lattice then reads
\begin{equation}
\begin{split}
\varGamma_{2,2} (T, U,\tau; \varLambda ) &\equiv \frac{1}{N}\sum_{g\in \mathbb{Z}_N} \varGamma_{2,2} \left[ {}^0_g\right] (T,U,\tau )
\\
&=\sum_{g \in \mathbb{Z}_N}\sum_{M\in {\rm Mat}_{2\times 2} (\mathbb{Z})}\frac{ e^{\frac{2 i \pi g}{N}\, {\rm tr}\, (M \, \varLambda )}}{N} \,
e^{-2 i \pi \tau \, {\rm det} (M)} \, \exp \left[ -\frac{\pi \tau_2}{T_2 U_2} \left| \begin{pmatrix} 1 & U \end{pmatrix} M \begin{pmatrix} T \\ 1 \end{pmatrix} \right|^2 \right]\,.
\end{split}
\end{equation}
Its duality symmetry is clearly a subgroup of ${\rm SL} (2;\mathbb{Z})_T \times {\rm SL}(2;\mathbb{Z} )_U \times \sigma$, which is the duality group for $\varLambda =0$, and can be easily determined as follows. A generic transformation in ${\rm SL} (2;\mathbb{Z})_T \times {\rm SL}(2;\mathbb{Z} )_U$ can be reabsorbed by the matrix redefinition
\begin{equation}
M \to \tilde M = \sigma_1 \gamma_U^t \sigma_1 M \gamma_T\,,
\end{equation}
where $\sigma_1$ is the Pauli matrix, and $\gamma_\alpha \in {\rm SL} (2;\mathbb{Z})_\alpha$, at the cost of changing the shift matrix
\begin{equation}
\varLambda \to \tilde \varLambda = \gamma_T^{-1} \, \varLambda \, \sigma_1 \gamma_U^{-t} \,\sigma_1\,.
\end{equation}
This transformation acts on the lattice as
\begin{equation}
\varGamma_{2,2} ( \gamma_T T , \gamma_U U ; \varLambda ) = \varGamma_{2,2} (T,U; \tilde\varLambda )
\end{equation}
and is a symmetry if and only if
\begin{equation}
\varLambda = \gamma_T^{-1} \, \varLambda \, \sigma_1 \gamma_U^{-t}\, \sigma_1 \qquad \mod\, N\,.
\label{dualityUT}
\end{equation}
Similarly, one can show that the involution $\sigma\, :\ T\leftrightarrow U$ is a symmetry if and only if
\begin{equation}
\varLambda = \sigma_1 \varLambda^t \sigma_1 \qquad \mod \, N\,.
\label{involTU}
\end{equation}
Actually, one has the option of combining $\sigma$ with any other involution which is in ${\rm SL} (2;\mathbb{Z})_T \times {\rm SL} (2;\mathbb{Z})_U$ but not in the perturbative duality group of $\varGamma_{2,2} (T,U;\varLambda )$.
As an example, let us consider the case of a momentum shift along the first direction, $\mu_i = 0 = \lambda^2$, $\lambda^1 =1$. Solving the eqs. \eqref{dualityUT} and \eqref{involTU} yields the duality group $\varGamma^1 (N) _T \times \varGamma_1 (N)_U \ltimes \sigma_{T\leftrightarrow -1/U}$, where the involution is now given by $\sigma S$, where $S : z \to - 1/z$ is an order-two element which is in ${\rm SL } (2;\mathbb{Z})$ but not in $\varGamma_1 (N)$ or $\varGamma^1 (N)$. Notice, however, that any even-weight modular form of the congruence subgroup $\varGamma_1 (N)$ is automatically modular under the larger Hecke subgroup $\varGamma_0 (N)$. As a result, since the lattice has zero-weight with respect to the $T$ and $U$ variables, the effective duality group is always enhanced and in this case reads $\varGamma^0 (N) _T \times \varGamma_0 (N)_U \ltimes \sigma_{T\leftrightarrow -1/U}$. Similarly, for the case of a winding shift along the second cycle of the $T^2$, with $\mu_2 =1$ and $\mu_1 = 0 = \lambda^i$, the duality group is $\varGamma_0 (N) _T \times \varGamma_0 (N)_U \ltimes \sigma_{T\leftrightarrow U}$. Actually, the lattice partition function is now invariant under the permutations of $T$, $U$, $\tau$
\begin{equation}
\varGamma_{2,2}(T,U,\tau;\varLambda) = \varGamma_{2,2}(U,T,\tau;\varLambda) =\varGamma_{2,2}(T,\tau,U;\varLambda)
\,,
\end{equation}
just like the unshifted lattice $\varGamma_{2,2} (T,U,\tau )$.
An important by-product is the relation between the Atkin-Lehner \cite{0177.34901} involution $z'=-1/(Nz)$ in the $T$, $U$ and $\tau$ variables
\begin{align}
\varGamma_{2,2} \left(T,-\tfrac{1}{NU},\tau;\varLambda \right) = \varGamma_{2,2} \left(-\tfrac{1}{NT}, U ,\tau;\varLambda \right)
=\varGamma_{2,2} \left(T,U,-\tfrac{1}{N\tau};\varLambda\right) \,.
\end{align}
Before we conclude, we observe that different shift vectors can actually be mapped into each other by a suitable redefinition of the $T$ and $U$ moduli \cite{Gregori:1997hi}.
\subsection{Niebur-Poincar\'e series for $\varGamma_0 (N)$}
In this subsection we recall basic facts about modular forms and Poincar\'e series with respect to the Hecke congruence subgroups $\varGamma_0 (N)$ with $N$ prime, referring the reader to the existing literature for more details \cite{iwaniec2, shimura, 0804.11039, Angelantonj:2013eja}. The fundamental domain $\mathcal{F}_N = \varGamma_0 (N) \backslash \mathbb{H}$ possesses two cusps, at $z=0$ and $z=\infty$. These special points are characterised by their {\em width} $m_\mathfrak{a}$, which essentially counts the number of copies of the ${\rm SL} (2;\mathbb{Z})$ fundamental domain in $\mathcal{F}_N$ that end at the cusp $z_\mathfrak{a}$. As a result, one finds $m_\infty =1$ while $m_0 = N$. The two cusps are mapped to each other via the Atkin-Lehner involution $z\to -1/(Nz)$.
For a given seed $f(z)$, invariant under $z\to z+1$, one defines the Poincar\'e series
attached to the cusp $\mathfrak{a}$ via
\begin{equation}
F_\mathfrak{a} (z) = \sum_{\gamma \in \varGamma_\mathfrak{a} \backslash \varGamma_0 (N) } f( z) \, \Big|_w \, \sigma_\mathfrak{a}^{-1}\, \gamma\,,
\label{HeckPS}
\end{equation}
where $\big|_w$ denotes the Petersson slash operator, $\varGamma_\mathfrak{a}$ is the stabiliser of the cusp $z_\mathfrak{a}$, and $\sigma_\mathfrak{a}$ are ${\rm SL} (2;\mathbb{R})$ matrices which relate $\varGamma_\mathfrak{a}$ to $\varGamma_\infty$, with
\begin{equation}
\sigma_\infty = \begin{pmatrix} 1 &0 \\ 0 & 1 \end{pmatrix} \,, \qquad
\sigma_0 = \begin{pmatrix} 0 & 1/\sqrt{N} \\ -\sqrt{N} & 0 \end{pmatrix} .
\end{equation}
A convenient choice of Poincar\'e series of $\varGamma_0 (N)$ consists in
choosing the same seed $f(z)=\mathcal{M}_{s,w} (- \kappa y)e^{-2 i \pi \kappa x}$ as for ${\rm SL} (2;\mathbb{Z})$ Niebur-Poincar\'e series. We shall denote them by $\mathcal{F}_\mathfrak{a} (s,\kappa , w)$, and we shall refer to them as the $\varGamma_0 (N)$ Niebur-Poincar\'e series attached to the cusp $\mathfrak{a}$. They are absolutely convergent for ${\rm Re}\, (s) >1$, and can thus be used for unfolding the fundamental domain. The functions $\mathcal{F}_\infty (s,\kappa , w) $ and $\mathcal{F}_0 (s,\kappa , w)$ are actually related via the action of the Atkin-Lehner involution, namely
\begin{equation}
\mathcal{F}_0 (s,\kappa , w; z) = N^{-w/2}\, z^{-w}\, \mathcal{F}_\infty (s,\kappa , w ; -1/Nz )\,.
\end{equation}
By construction, they are eigenmodes of the weight-$w$ hyperbolic Laplacian with the same eigenvalue as in \eqref{laplEskw} and satisfy the same closure properties under modular differentiations as in \eqref{NPderiv}. As a result, they are harmonic Maa\ss\ forms for $s= 1- \frac{w}{2}$, and turn out to be weakly holomorphic for low enough weight, when the shadow vanishes identically. The Hecke operators $H_n^{(N)}$ of $\varGamma_0 (N)$ \footnote{
The Hecke operator $H_n^{(N)}$ of $\varGamma_0 (N)$ acts on the Fourier modes of a weight-$w$ modular form $\varPhi = \sum_m \varPhi (m,y) \,e^{2i \pi m x}$ as
\begin{equation}
H^{(N)}_n \, \varPhi (m,y) = n^{1-w} \sum_{d|(m,n)} d^{w-1}\, \chi (d )\, \varPhi (nm/d^2 , d^2 y/n)\,, \nonumber
\end{equation}
with $\chi (d)$ the trivial Dirichlet character $\mod\ N$, such that $\chi (d) = 1$ if $(N,d)=1$ and $\chi (d) =0$ otherwise.
It satisfies the commutative algebra $H^{(N)}_\kappa \, H^{(N)}_\lambda = \sum_{d|(\kappa , \lambda)} d^{1-w}\, \chi (d)\, H^{(N)}_{\kappa \lambda / d^2}$.
}
acts on the Niebur-Poincar\'e series as
\begin{equation}
H^{(N)}_\kappa \, \mathcal{F}_{\mathfrak{a}} (s,1,w;z) = \mathcal{F}_{\mathfrak{a}} (s,\kappa , w;z) - \delta_{\mathfrak{a}\infty}\,\mathcal{F} (s,\kappa /N,w;Nz )\,.
\end{equation}
The second term in the r.h.s. is understood to be present only when $\mathfrak{a}=\infty$ and $N|\kappa$. It is important to note that $H^{(N)}_\kappa$ is self-adjoint with respect to the Petersson inner product only when $(N,\kappa)=1$ \cite{shimura, 0804.11039}.
The Fourier expansion of the Niebur-Poincar\'e series $\mathcal{F}_\mathfrak{a} (s,\kappa , w)$ around the two cusps can be found in \cite{Angelantonj:2013eja}. For small enough (negative) weight, in the holomorphic case $s=1-\frac{w}{2}$, one has
\begin{equation}
\begin{split}
\mathcal{F}_\mathfrak{a} \left(1-\tfrac{w}{2} , \kappa , w \right)\, |_w \, \sigma_\mathfrak{b} =& \delta_{\mathfrak{ab}} \, \frac{\varGamma (2-w)}{q^\kappa} + \frac{4\pi^2 \kappa}{(2\pi i \kappa)^w}\, Z_{\mathfrak{ab}} \left(0,\kappa ; 1-\tfrac{w}{2}\right)
\\
& + 4\pi \kappa \varGamma (2-w) \sum_{m\ge 1 }^\infty \left( \frac{-m}{\kappa}\right)^{\frac{w}{2}}\, \mathcal{Z}_{\mathfrak{ab}} \left( m , -\kappa ; 1 - \tfrac{w}{2}\right) \, q^m\,,
\end{split}
\end{equation}
where $Z_{\mathfrak{ab}} (a,b;s)$ and $\mathcal{Z}_{\mathfrak{ab}} (a,b;s)$ are the Kloosterman-Selberg zeta functions relating the two cusps $\mathfrak{a}$ and $\mathfrak{b}$ defined in appendix \ref{kloosterman}\footnote{Note that in the present work we employ a different notation for the Kloosterman-Selberg zeta function than the one of \cite{Angelantonj:2013eja}.}.
As for the case of the modular group, the Niebur-Poincar\'e series $\mathcal{F}_\mathfrak{a} (1-\frac{w}{2}+n,\kappa , w)$ can be used to decompose any generic $\varGamma_0 (N)$ modular form $\varPhi $ of negative weight as a linear combination
\begin{equation}
\varPhi (z) = \sum_{\mathfrak{a}=0,\infty} \sum_{-\kappa_\mathfrak{a} \le m < 0} \sum_{p=0}^n d_{\mathfrak{a} , p } (m)\, \mathcal{F}_\mathfrak{a} \left(1 - \tfrac{w}{2}+p, m , w \right)\,.
\label{HCdecomp}
\end{equation}
The coefficients $d_{\mathfrak{a} , p } (m)$ are uniquely determined by matching the singular behaviour of both sides of the equation around the two cusps. Their explicit expression can be found in \cite{Angelantonj:2013eja}.
We shall also need the non-holomorphic weight-$w$ Eisenstein series $E_\mathfrak{a} (s,w)$ obtained by choosing the seed $f(z) = y^s$ in \eqref{HeckPS}. As in the ${\rm SL}(2;\mathbb{Z})$ case, they are eigenmodes of the hyperbolic Laplacian with the same eigenvalue as the Niebur-Poincar\'e series, absolutely convergent for ${\rm Re} (s) >1$, and
satisfy the same closure properties under modular differentiation as in \eqref{EPderiv}. For $s=1-\frac{w}{2}$ and $s=\frac{w}{2}$ the Eisenstein series become harmonic Maa\ss\ forms. For even weight $w>2$, $E_{\mathfrak{a},w} \equiv E_\mathfrak{a} (\frac{w}{2},w )$ reduces to the conventional holomorphic Eisenstein series $E_w$ of $\varGamma_0 (N)$ attached to the cusp $\mathfrak{a}$. The Eisenstein series $E_\mathfrak{a} (s,w)$ can actually be expressed as linear combinations of the ${\rm SL} (2;\mathbb{Z})$ ones with rescaled arguments,
\begin{equation}
E_\infty(s,w;z)=\frac{N^{s+\frac{w}{2}}\,E(s,w;Nz)-E(s,w;z)}{N^{2s}-1}\,, \qquad
E_0(s,w;z)=\frac{N^s\,E(s,w;z)-N^{\frac{w}{2}}\,E(s,w;Nz)}{N^{2s}-1}\,.
\label{FuncRelatEisen}
\end{equation}
From these relations, the Fourier series expansions around the cusp $\mathfrak{b}$ of the $\varGamma_0 (N)$ Eisenstein series attached to the cusp $\mathfrak{a}$ can be straightforwardly extracted and one obtains
\begin{equation}
\begin{split}
E_{\mathfrak a}(s,w;z) |_{\mathfrak{b}} &= \delta_{\mathfrak{ab}}\,y^{s-\frac{w}{2}}+i^{-w}\sqrt{\pi}\,\frac{\varGamma(s)\,\varGamma(s-\frac{1}{2})}{\varGamma(s+\frac{w}{2})\,\varGamma(s-\frac{w}{2})}\, Z_{\mathfrak{ab}}(0,0;s)\,y^{1-s-\frac{w}{2}}
\\
&+ i^{-w}\pi^s (4\pi)^{\frac{w}{2}}\sum_{m\neq 0}\frac{1}{\varGamma(s+\frac{w}{2}{\rm sgn}(m))}\,|m|^{s+\frac{w}{2}-1}\, Z_{\mathfrak{ab}}(0,m;s)\,\mathcal W_{s,w}(my)\,e^{2\pi i m x} \,.
\end{split}
\end{equation}
\subsection{Properties of the $\varGamma_0 (N)$ modular integrals}
We can now proceed to the computation of the prototype integral
\begin{equation}
\mathcal{I}_N (s,\kappa ; \mathfrak{a} ) = {\rm R.N.} \int_{\mathcal{F}_N} d\mu \, \mathcal{F}_\mathfrak{a} (s,\kappa , 0 )\, \varGamma_{2,2} (T,U ; \varLambda )
\label{HCprot}
\end{equation}
by unfolding $\mathcal{F}_N$ against the Niebur-Poincar\'e series. Any $\varGamma_0 (N)$ integral involving the shifted Narain lattice times a generic weakly almost holomorphic modular form can be straightforwardly derived from \eqref{HCprot} by virtue of the decomposition \eqref{HCdecomp}. The symbol R.N. stands for the $\varGamma_0 (N)$ invariant renormalisation prescription of \cite{Angelantonj:2013eja} for treating the infra-red divergences that may arise at the two cusps. Special care is needed for $s=1$ where the integral develops a simple pole, in which case, the renormalisation prescription amounts to its proper subtraction.
From the differential properties of the Niebur-Poincar\'e series and from the identities
\begin{equation}
\varDelta_T \, \varGamma_{2,2} (T,U ; \varLambda) =
\varDelta_U \, \varGamma_{2,2} (T,U ; \varLambda) =
\varDelta_\tau \, \varGamma_{2,2} (T,U ; \varLambda) \,,
\end{equation}
obeyed by the shifted Narain lattice, for any choice of the shift $\varLambda$, it turns out that the $\varGamma_0 (N)$ modular integrals $\mathcal{I}_N (s,\kappa ; \mathfrak{a})$ are themselves eigenmodes of both the (weight-zero) hyperbolic Laplacians $\varDelta_T$ and $\varDelta_U$ with eigenvalue $\frac{1}{2} s (s-1)$. For $s=1$ a constant source-term appears, similarly to the ${\rm SL} (2;\mathbb{Z})$ case \eqref{lapTU}.
Whenever the shifted lattice $\varGamma_{2,2} (T,U ; \varLambda )$ is symmetric under the exchange of the modular parameter $\tau$ with either the $T$ and $U$ moduli,
and provided that $(\kappa , N) =1$ so that the $\varGamma_0 (N)$ Hecke operator is self-adjoint,
one can restrict the analysis to the case $\kappa =1$ and $\mathfrak{a} = \infty$, since all other cases can be recovered via the action of the Hecke and Atkin-Lehner operators.
Indeed, following the steps outlined in Section \ref{sec_intsl2z} one finds
\begin{equation}
\mathcal{I}_N (s , \kappa ; \mathfrak{a}) = H^{(N)}_\kappa\cdot \mathcal{I}_N (s ; \mathfrak{a}) \,,
\end{equation}
where the Hecke operator acts on either $T$ or $U$, or on both, depending on the available symmetries and $\mathcal{I}_N (s;\mathfrak{a}) \equiv \mathcal{I}_N (s,1;\mathfrak{a})$.
Moreover, the integrals associated to the two cusps are not independent, but are related by the action of the Atkin-Lehner involution $\sigma_0$. In fact, since
\begin{equation}
\mathcal{F}_0 (s,\kappa , w) = \mathcal{F}_\infty (s,\kappa , w) |\sigma_0\,,
\end{equation}
and since $\sigma_0$ is self-adjoint on the fundamental domain of $\varGamma_0 (N)$, one finds
\begin{equation}
\mathcal{I}_N (s; \mathfrak{a}) = \mathcal{I}_N (s ; \infty ) |\sigma_\mathfrak{a}\,,
\end{equation}
where the Petersson slash operator can act on either variable, depending on the available symmetries. For $s=1+n$, this relation can be extended to the level of both the harmonic and holomorphic prepotential
\begin{equation}
f_n (T,U ;\mathfrak{a} ) = f_n ( T,U ; \infty ) |\sigma_\mathfrak{a} +\, {\rm polynomial}\,,
\label{SlashPrepotCongr}
\end{equation}
due to the commutation property of the involution $\sigma_\mathfrak{a}$ with the modular derivative. Similarly, for $(\kappa,N)=1$, one may extend the relation \eqref{prepHecke} to both the harmonic and holomorphic prepotentials of $\varGamma_0(N)$
\begin{equation}
f_n(T,U,\kappa;\mathfrak{a}) =\kappa^{-n}\,H^{(N)}_\kappa\cdot f_n(T,U;\mathfrak{a}) \,.
\end{equation}
\subsection{Evaluation of the modular integral $\mathcal{I}_N (s; \mathfrak{a} )$}
For concreteness, we shall evaluate the integral \eqref{HCprot} for $\kappa =1$, and we shall consider explicitly only the case of a winding shift along the second cycle of the two-torus, with $\mu_2 =1$ and $\mu_1 = \lambda^i =0$, since the other cases can be treated similarly. In addition to the duality symmetry $\varGamma_0 (N)_T \times \varGamma_0 (N)_U \times \sigma_{T\leftrightarrow U}$, inherited by the integral, the shifted lattice is also invariant under the triality symmetry $T\leftrightarrow U \leftrightarrow \tau$, which implies that the results for the cusp at zero, or for higher values of $\kappa$, can be obtained by acting with the Atkin-Lehner involution or Hecke operator on $\mathcal{I} (s;\infty)$.
For the cusp at $\infty$ the shifted lattice $\varGamma_{2,2} (T,U; \varLambda )$ reduces to the conventional unshifted Narain lattice sum merely subjected to the constraint $n^2 = 0 \ \mod \, N$. One may thus readily obtain the result of the integral $\mathcal{I} (s;\infty )$ from its ${\rm SL} (2;\mathbb{Z})$ cousin $\mathcal{I} (s)$, both in the BPS-sum representation and in the Fourier-series one.
Unfolding the fundamental domain against the Niebur-Poincar\'e series yields the BPS-state sum representation \cite{Angelantonj:2013eja}
\begin{equation}
\mathcal{I}_{N} (s; \mathfrak{a}) = \frac{m_\mathfrak{a}}{N}\, \varGamma (s) \, \sum_{g=0}^{N-1} \sum_{{\rm BPS} (v_\mathfrak{a} g)} \, e^{2 i \pi u_\mathfrak{a} g n_2/N}\, {}_2 F_1 \left( s, s; 2s; \frac{4}{m_\mathfrak{a}\, |p_{\rm L} (v_\mathfrak{a} g)|^2}\right) \, \left( \frac{m_\mathfrak{a}\, |p_{\rm L} (v_\mathfrak{a} g)|^2}{4} \right)^{-s}\,,
\end{equation}
where the BPS-sum extends over integers $m_i, n^i$ subject to the `BPS-constraint'
\begin{equation}
m_\mathfrak{a}\, \left( |p_{\rm L} (v_\mathfrak{a} g)|^2 - |p_{\rm R} (v_\mathfrak{a} g )|^2\right) = 4\,,
\end{equation}
with the shifted left-moving and right-moving momenta given in eq. \eqref{shiftedmomenta}. The ratio $u_\mathfrak{a}/v_\mathfrak{a}$ identifies the coordinate of the cusp $\mathfrak{a}$ on the complex upper plane, {\em i.e.} $(u_0,v_0) = (0,1)$ and $(u_\infty,v_\infty) = (1,0)$.
Following the same arguments as in Section \ref{BPSPoinca} , this expression can be cast as the Poincar\'e series
\begin{equation}
\mathcal{I}_{N} (s;\mathfrak{a} ) = \sum_{\gamma \in \varGamma_{\rm diag} \backslash (\varGamma_T \times \varGamma_U)} \left. \varphi \left( \frac{|\bar T - U |^2}{T_2 U_2}\right) \right| \, \sigma_{\mathfrak{a},i}^{-1} \, \gamma\,,
\end{equation}
where the seed $\varphi (z)$ is the same as in eq. \eqref{seedBPS}, and the involution $\sigma_{\mathfrak{a},i}$ acts on either of the $T$ and $U$ variables. For generic $N$ including $N=1$, $\varGamma_T \times \varGamma_U = \varGamma_0 (N)_T \times \varGamma_0 (N)_U$ and $\varGamma_{\rm diag} = {\rm diag} (\varGamma_0 (N)_T \times \varGamma_0 (N)_U)$.
This representation is valid at any point in the Narain moduli space. It is tailored to exploit the analytic structure of the integral in the vicinity of points of gauge symmetry enhancement and provides a Poincar\'e series representation, thus making the duality symmetry manifest. As in the ${\rm SL} (2;\mathbb{Z})$ case, points of gauge symmetry enhancement arise at $p_{\rm R} (v_\mathfrak{a} g) =0$, {\rm i.e.} $T = \sigma_\mathfrak{a} \, U$ together with their $\varGamma_0 (N)$ images \cite{Angelantonj:2012gw}. Note that the singular locus depends on a choice of cusp.
The Fourier series representation depends on a choice of Weyl chamber. Although the structure of Weyl chambers depends on the value of $N$, in the chamber $T_2 > U_2$ one finds
\begin{equation}
\mathcal{I}_{N} (s ; \mathfrak{a} ) = \mathcal{I}_{N}^{(0)} (s ; \mathfrak{a} ) +\mathcal{I}_{N}^{(+)} (s ; \mathfrak{a} ) + \mathcal{I}_{N}^{(-)} (s ; \mathfrak{a} ) \,,
\end{equation}
with
\begin{equation}
\begin{split}
\mathcal{I}_{N}^{(0)} (s ; \mathfrak{a} ) &=2^{2s}\, \sqrt{4\pi} \, \varGamma (s-\tfrac{1}{2}) \, T_2^{1-s}\, E_\mathfrak{a} (s,0;U)\,,
\\
\mathcal{I}_{N}^{(+)} (s ; \mathfrak{a} ) &= 2\, \sum_{M>0} \frac{e^{2i\pi M T_1}}{M}\, \mathcal{W}_{s,0} (M T_2 )\, \mathcal{F}_\mathfrak{a} (s,M,0;U)\,,
\end{split}
\end{equation}
and $\mathcal{I}^{(-)}_N (s;\mathfrak{a} ) = [ \mathcal{I}^{(+)}_N (s ;\mathfrak{a})]^*$.
Notice that although these expressions are valid for $N$ prime, the choice $N=1$ reduces to eq. \eqref{modint}, if we assume that, in this case, $\mathfrak{a}$ labels the unique cusp $\infty$ of the modular group. In fact, in this case $\mathcal{F}^{(1)}_\infty (s,1,0) \equiv \mathcal{F} (s,1,0)$, $\mathcal{F}_1 \equiv \mathcal{F}$, while for the Narain lattice, any shift vector $\varLambda$ acts trivially by simply relabelling the integral momenta and windings, and thus $\varGamma_{2,2}^{(1)} (T,U;\varLambda ) \equiv \varGamma_{2,2} (T,U)$.
Of particular interest in string theory are the cases where $s=1+n$ is a positive integer. The Fourier series expansion then admits alternative representations in terms of polylogarithms or, for $n=0$, can be summed in terms of combinations of the $\varGamma_0 (N)$ Hauptmodul, thus generalising the Borcherd's product formul\ae\ valid for ${\rm SL} (2;\mathbb{Z})$.
\subsubsection{The case $s=1$}
The case $s=1$ is special since the ${\rm R.N.}$ prescription requires the subtraction of a simple pole from the non-holomorphic Eisenstein series entering the Fourier mode with vanishing frequency. Using the first Kronecker limit formula one obtains
\begin{equation}
\mathcal{I}_N^{(0)}(1;\mathfrak{a})= -\tfrac{24}{(N-1)\nu_N}\,\log\left( T_2^{N-1} U_2^{N-1}\left|\left[ \frac{\eta^N(N U)}{\eta(U)} \Big| \sigma_{\mathfrak a} \right]\right|^4\right)+{\rm const}\,,
\end{equation}
where $\nu_N = {\rm vol} (\mathcal{F}_N)/{\rm vol} (\mathcal{F})$ is the index of $\varGamma_0 (N)$ inside ${\rm SL} (2;\mathbb{Z})$.
Notice that this expression is valid also for $N=1$, if we allow the formal limit $N\to 1$. In fact, $\eta (N U) \sim \eta (U) + \mathcal{O} (N-1)$, and thus
\begin{equation}
\lim_{N\to 1} -\frac{24}{(N-1)\nu_N} \log \, \left( T_2^{N-1} U_2^{N-1}\left| \frac{\eta^N(N U)}{\eta(U)} \right|^4\right) = -24 \, \log \left(T_2 U_2 \, |\eta (U) |^4\right)\,,
\end{equation}
in agreement with eq. \eqref{s1zeromode}.
The positive-frequency modes are regular at $s=1$ and read
\begin{equation}
\begin{split}
\mathcal{I}_N^{(+)}(1;\mathfrak{a}) &= 2\sum_{M>0}\frac{q_T^M}{M}\,\mathcal{F}_{\mathfrak{a}}(1,M,0)
\\
&=2\sum_{M>0}\frac{q_T^M}{M}\,\left[\mathcal{F}_{\mathfrak{a}}(1,M,0)-\tilde{\mathcal{F}}_{\mathfrak{a}}^{(0)}(1,M,0)\right]
+4\pi i T\,\delta_{\mathfrak{a}\infty}
-\tfrac{24}{(N-1)\nu_N}\,\log\left( \left[ \frac{\eta^N(NT)}{\eta(T)} \Big| \sigma_{\mathfrak{a}}\right]\right)^2 \,,
\end{split}
\end{equation}
where $\tilde{\mathcal{F}}_{\mathfrak{a}}^{(0)}(1,M,0)=4\pi^2 M\, Z_{\mathfrak{a}\infty}(0,M;1)$ is the constant zero-mode of the Niebur-Poincar\'e series of $\varGamma_0(N)$ attached to the cusp $\mathfrak a$.
Using the properties of holomorphy and modularity in the $T$ and $U$ variables, together with the singularity structure of $\mathcal{F}_{\mathfrak{a}}(1,M,0)-\tilde{\mathcal{F}}_{\mathfrak{a}}^{(0)}(1,M,0)$ around the two cusps, one obtains the identity
\begin{equation}
\sum_{M>0}\frac{q_T^M}{M}\left[\mathcal{F}_{\mathfrak{a}}(1,M,0)-\tilde{\mathcal{F}}_{\mathfrak{a}}^{(0)}(1,M,0)\right]+2\pi iT\,\delta_{\mathfrak{a}\infty} = -\log\frac{\left( j_{\infty}(T)-j_{\mathfrak{a}}(U)\right)}{\left( j_\infty(T)-\frac{24}{N-1}\right)^{\delta_{\mathfrak{a}0}}} \,,
\label{Borch1congr}
\end{equation}
where $j_{\mathfrak{a}}(z)=\sum_{M}c_{\mathfrak{a}}(M)\,q^M$. In fact, $j_\infty(z)$ is the Hauptmodul for $\varGamma_0(N)$, while
$j_0(z)=j_\infty(-1/Nz)$ is its image under the Atkin-Lehner involution.
Eq. \eqref{Borch1congr} generalises the Borcherds' product formula \cite{zbMATH00220742} to the case of $\varGamma_0(N)$ congruence subgroups
\begin{equation}
\prod_{K>0,\,L\in\mathbb Z}\left(\frac{(1-q_T^K \,q_U^L)^N}{1-q_T^{NK}\,q_U^{NL}}\right)^{c_\mathfrak{a}(KL)}= \frac{\left( j_\infty(T)-j_{\mathfrak{a}}(U)\right)^N}{\left(j(NT)-j(NU)\right)^{\delta_{\mathfrak{a}\infty}}\left(j_\infty(T)-\frac{24}{N-1}\right)^{N\delta_{\mathfrak{a}0}}} \,.
\end{equation}
Putting things together, the modular integral reads
\begin{equation}
\begin{split}
\mathcal{I}_N (1; \mathfrak{a} ) =& -\tfrac{24}{(N-1)\nu_N}\,\log\left( T_2^{N-1} U_2^{N-1}\left|
\left[\frac{\eta^N(NU)}{\eta(U)}\, \Big| \, \sigma_{\mathfrak a}\right]
\,
\left[\frac{\eta^N(NT)}{\eta(T)} \, \Big| \sigma_{\mathfrak a}\,
\right]
\right|^4\right)
\\
& -\log\frac{\left| j_{\infty}(T) - j_{\mathfrak{a}}(U)\right|^4}{\left|j_\infty(T)-\frac{24}{N-1}\right|^{4\delta_{\mathfrak{a}0}}} +{\rm const}\,.
\end{split}
\label{congrResult1}
\end{equation}
The first term on the r.h.s. originates from the constant term in the decomposition $\mathcal{F}_\mathfrak{a} (1,1,0) = j_\mathfrak{a} (z) + \tilde{\mathcal F}_{\mathfrak{a}}^{(0)}(1,1,0)$, whereas the second term originates from the $\varGamma_0 (N)$ Hauptmodul. By employing the identity
\begin{equation}
\left( j_0(T)-j_\infty(U)\right)\left(j_\infty(T)-\tfrac{24}{N-1}\right) = \left( j_0(U)-j_\infty(T)\right)\left(j_\infty(U)-\tfrac{24}{N-1}\right) \,,
\end{equation}
it is straightforward to verify that the expression \eqref{congrResult1} is indeed invariant under the duality symmetry $\varGamma_0(N)_T\times\varGamma_0(N)_U \ltimes \sigma_{T\leftrightarrow U}$. The modular integral \eqref{congrResult1} develops logarithmic singularities at the point $T=U$ and its $\varGamma_0(N)$ images, ascribed to the presence of extra massless states that would arise at symmetry enhancement points in the Narain moduli space.
Before closing this Section, we shall also give the expressions for the harmonic integrals involving the $\varGamma_0(N)$ Hauptmodul for $N=2,3$. Using the result
\begin{equation}
{\rm R.N.}\int_{\mathcal F_N} d\mu\,\varGamma_{2,2}[^0_1] = -\frac{1}{N-1}\,\log (T_2 U_2)^{N-1}\,\left| \frac{\eta^N(NT)\,\eta^N(NU)}{\eta(T)\,\eta(U)}\right|^4 +{\rm const}\,,
\end{equation}
for integrals of the shifted Narain lattice established in \cite{Angelantonj:2013eja}, together with eq. \eqref{integralFullCongr}, we obtain
\begin{equation}
\begin{split}
{\rm R.N.} \int_{\mathcal F_N} d\mu\,\varGamma_{2,2}[^0_1]\,j_\infty = & -\frac{N}{N-1}\,\log\left| \frac{j_\infty(T)-j_\infty(U)}{\left(j_\infty(T)-\frac{24}{N-1}\right)\left(j_\infty(U)-\frac{24}{N-1}\right)}\right|^4 \\
&+ \frac{1}{N-1}\,\log\left|j(T)-j(U)\right|^4 +{\rm const} \,,
\end{split}
\end{equation}
and
\begin{equation}
{\rm R.N.} \int_{\mathcal F_N} d\mu\,\varGamma_{2,2}[^0_1]\,(j_0-\tfrac{24}{N-1}) = -\frac{N}{N-1}\,\log\left| j_0(T)-j_\infty(U)\right|^4 +{\rm const} \,,
\end{equation}
which generalise the ${\rm SL}(2;\mathbb Z)$ result \eqref{modint2} first obtained in \cite{Harvey:1995fq} to the $\varGamma_0(N)$ integrals. These results have been used in \cite{Angelantonj:2014dia} to derive the universal behaviour of gauge thresholds in stable non-supersymmetric heterotic vacua.
\subsubsection{The case $s=1+n$ with $n>0$}
For integer $s =1+n>1$, and for small enough $n$, the Niebur-Poincar\'e series $\mathcal{F}_\mathfrak{a}^{(N)} (1+n,1,0)$ becomes almost holomorphic and admits the Fourier series expansion at the cusp $\infty$
\begin{equation}
\mathcal{F}_\mathfrak{a}^{(N)} (1+n,1,0) = \sum_m F_{\mathfrak{a},n} (m) \, (4\pi y)^{-n}\, L_n^{(-2n-1)} (4 \pi m y )\, q^m\,,
\end{equation}
in terms of the associated Laguerre polynomials, where $F_{\mathfrak{a},n}$ are the Fourier coefficients of $\mathcal{F}_\mathfrak{a}(1+n,1,-2n)$ at the cusp at infinity,
\begin{equation}
\begin{split}
F_{\mathfrak{a},n}(-1) &= \varGamma(2n+2)\,\delta_{\mathfrak{a}\infty} \,,
\\
F_{\mathfrak{a},n}(0) &= (2\pi)^{2n+2}\, (-1)^n\, Z _{\mathfrak{a}\infty}(0,1;n+1) \,,
\\
F_{\mathfrak{a},n}(L>0) &= 4\pi\,(-1)^n\,\varGamma(2n+2)\,\frac{\mathcal Z_{\mathfrak{a}\infty}(L,-1;n+1)}{L^n} \,.
\end{split}
\end{equation}
Similarly to the ${\rm SL}(2;\mathbb Z)$ case, the integral may be expressed in terms of the single-valued polylogarithms of eq. \eqref{SVPL}. One finds
\begin{equation}
\begin{split}
\mathcal{I}_N^{(+)} (1+n;\mathfrak{a}) &= \frac{\delta_{\mathfrak{a\infty}}}{N} \, \mathcal{I}^{(+)} (NT, NU)
\\
&\quad + \frac{2\, (-1)^n}{n!} \sum_{M>0\atop L\in \mathbb{Z}} F_{\mathfrak{a},n} (LM) \, \sum_{k =0}^n \frac{ (n+k)!}{k!\,(n-k)!} \, \frac{(ML)^n}{(4 \pi ML T_2 U_2)^k}\, \hat {\rm L}^{(N)}_{(k)} (MT + LU) \,,
\end{split}
\end{equation}
where we have defined the `combined $\varGamma_0(N)$ polylogarithms'
\begin{equation}
\hat{\rm L}^{(N)}_{(k)}(z) \equiv {\rm L}_{(k)}(z) - \frac{1}{N^{2k+1}}\, {\rm L}_{(k)}(Nz) \,.
\end{equation}
The zero mode $\mathcal{I}_N^{(0)}$ similarly admits a representation in terms of single-valued polylogarithms, by using \eqref{EisenCombPolylog} together with the functional relations \eqref{FuncRelatEisen} in order to express the non-holomorphic Eisenstein series $E_\mathfrak{a}(1+n,0;z)$ of $\varGamma_0(N)$ in terms of the ${\rm SL}(2;\mathbb Z)$ ones, $E(1+n,0;z)$ and $E(1+n,0;Nz)$.
\subsection{Generalised holomorphic prepotentials for $\varGamma_0(N)$}
As in the case of the full modular group, for integer $s=1+n>1$ one may define generalised harmonic and holomorphic prepotentials also for the $\varGamma_0(N)$ integral $\mathcal{I}_N$ via
\begin{equation}
\begin{split}
\mathcal{I}_N(1+n;\mathfrak{a}) &= 4\,{\rm Re} \frac{(-D_T D_U)^n}{n!}\, h_{n}(T,U;\mathfrak{a}) \,,
\\
&= 4\,{\rm Re} \frac{(-D_T D_U)^n}{n!}\,f_{n}(T,U;\mathfrak{a})
\,.
\end{split}
\end{equation}
The expression of the integral in terms of iterated derivatives of the prepotential is again possible due to the closure property of Niebur-Poincar\'e series under modular differentiation.
The prepotentials for $\varGamma_0(N)$ can be obtained by an analysis parallel to the ${\rm SL}(2;\mathbb Z)$ case and one obtains
\begin{equation}
h_{n}(T,U;\mathfrak{a}) = (2\pi)^{2n+1}\,E_\mathfrak{a}(n+1,-2n;U) + \sum_{M>0}\frac{2}{(2M)^{2n+1}}\,q_T^M\,\mathcal{F}_{\mathfrak{a}}(1+n,M,-2n;U) \,,
\end{equation}
for the generalised harmonic and
\begin{equation}
f_{n}(T,U;\mathfrak{a}) = \frac{(-1)^n\,(2\pi)^{2n+2}}{2^{2n}\,\zeta(2n+2)}\,\tilde{E}_{-2n,\mathfrak{a}}(U) +\sum_{M>0}\frac{2}{(2M)^{2n+1}}\,q_T^M\,\mathcal{F}_{\mathfrak{a}}(1+n,M,-2n;U) \,,
\label{HolPrepotCongr1}
\end{equation}
for the generalised holomorphic prepotentials. Here, $\tilde E_{-2n,\mathfrak{a}}(U)$ is the $\varGamma_0(N)$ Eichler integral of weight $-2n$ associated to the holomorphic Eisenstein series $E_{2n+2,\mathfrak{a}} $, attached to the cusp $\mathfrak{a}$. Its Fourier expansion about a cusp $\mathfrak{b}$ reads
\begin{equation}
\tilde{E}_{-2n,\mathfrak{a}}(z) \big| \sigma_{\mathfrak{b}} = \zeta(2n+2) \left[ \frac{z^{2n+1}}{2\pi i}\,\delta_{\mathfrak{ab}}+\tfrac{1}{2} Z_{\mathfrak{ab}}(0,0;n+1)+\sum_{M>0} Z_{\mathfrak{ab}}(0,M;n+1)\,q^M \right]\,.
\end{equation}
The $\varGamma_0(N)$ Eichler integral $\tilde{E}_{-2n,\mathfrak{a}}$ inherits the functional relation properties of the Eisenstein series $E_{\mathfrak{a}}(n+1,-2n;z)$ and may be expressed in terms of ${\rm SL}(2;\mathbb Z)$ Eichler integrals via
\begin{equation}
\tilde{E}_{-2n,\infty}(z) = \frac{ N\,\tilde{E}_{-2n}(Nz)-\tilde{E}_{-2n}(z)}{N^{2n+2}-1} \,,\qquad \tilde{E}_{-2n,0}(z) = \frac{ N^{2n+1}\,\tilde{E}_{-2n}(z)-\tilde{E}_{-2n}(Nz)}{N^n(N^{2n+2}-1)} \,.
\label{EichlerCongrFuncRel}
\end{equation}
Of course, since holomorphic prepotentials are only defined up to terms in the kernel of ${\rm Re}(-D_T D_U)^n$, the expression \eqref{HolPrepotCongr1} for $f_n(T,U;\mathfrak{a})$ corresponds to a specific choice for the polynomial ambiguity. Notice that, due to the inhomogeneous transformation of the Eichler integral, the action of $\sigma_{0}$ now relates the holomorphic prepotential $f_n(T,U;0)$ attached to the cusp $\mathfrak{a}=0$ to the prepotential $f_n(T,U;\infty)$ attached to the cusp at $\infty$ with an additional polynomial dependence in the $U$ variable, so that eq. \eqref{SlashPrepotCongr} becomes
\begin{equation}
f_n(T,U;0) = f_n(T,U;\infty)|\sigma_0-\alpha\,\frac{N^{2n+1}\,P_S(U)-P_S(NU)}{N^n\,(N^{2n+2}-1)} \,,
\end{equation}
where $P_S(z)$ is the polynomial transformation of the ${\rm SL}(2;\mathbb Z)$ Eichler integral given in \eqref{PS} and $\alpha$ is the constant defined in \eqref{alphaConst}.
In terms of the Fourier coefficients $F_{\mathfrak{a},n}(M)$ of the Niebur-Poincar\'e series $\mathcal F_{\mathfrak{a}}(1+n,1,-2n)$, we may cast the generalised holomorphic prepotential in the alternative representation
\begin{equation}
\begin{split}
f_n(T,U;\mathfrak{a}) = & \frac{\delta_{\mathfrak{a}\infty}}{N^{2n+1}}\,f_n(NT,NU) + \frac{F_{\mathfrak{a},n}(-1)}{2^{2n}} \hat{{\rm Li}}_{2n+1}^{(N)}\left(\frac{q_T}{q_U}\right)
\\
&+ \frac{1}{2^{2n}}\sum_{M,L\geq 0} F_{\mathfrak{a},n}(ML)\,\hat{\rm Li}_{2n+1}^{(N)}(q_T^M\,q_U^L)
\\
&+ \tfrac{1}{2}\alpha\, \zeta(2n+2)\Bigr[ Z_{\mathfrak{a}\infty}(0,0;n+1)-\delta_{\mathfrak{a}\infty} Z(0,0;n+1)
\\
& -2(1-N^{-2n-1})\zeta(2n+1) Z_{\mathfrak{a}\infty}(0,1;n+1)\Bigr] \,,
\end{split}
\label{HolPrepotCongr2}
\end{equation}
where $f_n(T,U)$ is the ${\rm SL}(2;\mathbb Z)$ prepotential \eqref{SL2ZprepPoly} and, for convenience, we have introduced the `$\varGamma_0(N)$ polylogarithms'
\begin{equation}
\hat{\rm Li}_{k}^{(N)}(z) \equiv {\rm Li}_{k}(z)-\frac{1}{N^k}\,{\rm Li}_{k}(z^N) \,.
\end{equation}
Note that eq. \eqref{HolPrepotCongr2} is cusp-covariant and reduces to the ${\rm SL}(2;\mathbb Z)$ prepotential by setting $N=1$ and $\mathfrak{a}=\infty$, whereby the Kloosterman-Selberg zeta function $ Z_{\infty\infty}(a,b;s)$ simply reduces to its ${\rm SL}(2;\mathbb Z)$ counterpart, $Z(a,b;s)$.
Although the positive frequency modes in $T$ of the generalised holomorphic prepotential do transform covariantly as weight $-2n$ modular forms in the $U$ variable, the zero-frequency mode does not. Clearly, the inhomogeneous polynomial term is completely determined by the transformation properties of the Eichler integral $\tilde{E}_{-2n,\mathfrak{a}}(z)$ under the generators of $\varGamma_0(N)$. Since it is always possible to decompose the latter into repeated applications of the generators $S$ and $T$ of the full modular group, one may uniquely fix the polynomial transformation of $\tilde{E}_{-2n,\mathfrak{a}}(z)$ under any element of $\varGamma_0(N)$ by using eq. \eqref{EichlerCongrFuncRel} and the known transformation properties of the Eichler integral of ${\rm SL}(2;\mathbb Z)$.
Moreover, the behaviour of the generalised holomorphic prepotentials under $\sigma_{T\leftrightarrow U}$ can be straightforwardly determined from eq. \eqref{HolPrepotCongr2} and from the property \eqref{analyticContPoly} of polylogarithms to be
\begin{equation}
f_n(T,U;\mathfrak{a})-f_n(U,T;\mathfrak{a}) = \delta_{\mathfrak{a}\infty}\,P_{\sigma}(T,U) \,,
\end{equation}
with $T$ and $U$ lying in the fundamental chamber, with real parts in the interval $(0,1)$ and with $P_\sigma(T,U)$ being the polynomial \eqref{Psigma}.
\subsection{Generalised Yukawa couplings for $\varGamma_0(N)$}
In complete analogy to the ${\rm SL}(2;\mathbb Z)$ case, one may compute interactions in the low energy $\mathcal N=2$ supergravity by acting with suitable derivatives on the $\varGamma_0(N)$ generalised holomorphic prepotential. Here we shall restrict our attention to the simplest case of holomorphic generalised Yukawa couplings attached to the cusp $\mathfrak{a}$
\begin{equation}
Y_{T^{2n+1}}(T,U;\mathfrak{a}) = - 2^{-2n-1}\,D_T^{2n+1}\,f_n(T,U;\mathfrak{a}) = \left(\frac{\partial_T}{2\pi i}\right)^{2n+1}\,f_n(T,U;\mathfrak{a}) \,.
\end{equation}
The degree $2n$ polynomial in the inhomogeneous transformation of $f_n$ is again annihilated by the derivatives and the Yukawa coupling $Y_{T^{2n+1}}$ transforms as a weight $2n+2$ and $-2n$ holomorphic modular form under $\varGamma_0(N)_T$ and $\varGamma_0(N)_U$, respectively. In terms of the Fourier coefficients $F_{\mathfrak{a},n}(M)$ of the Niebur-Poincar\'e series $\mathcal F_{\mathfrak{a}}(1+n,1,-2n)$ expanded around the cusp $\infty$, one finds explicitly
\begin{equation}
\begin{split}
Y_{T^{2n+1}}(T,U;\mathfrak{a}) &= 2^{-2n}\sum_{M>0} q_T^M\,\mathcal F_{\mathfrak{a}}(1+n,M,-2n;U) \\
&= \delta_{\mathfrak{a}\infty}\,Y_{T^{2n+1}}(NT,NU)+2^{-2n} \sum_{M>0\atop L\in\mathbb{Z}} M^{2n+1}\, F_{\mathfrak{a},n}(ML) \mathcal{R}_N(q_T^M\,q_U^L) \,,
\end{split}
\label{YukawasCongr}
\end{equation}
which generalises the ``multi cover'' formula for the standard Yukawa couplings $Y_{T^3}$ to the case of freely-acting $\mathbb Z_N$ orbifolds. To lighten the notation, we have introduced the rational function
\begin{equation}
\mathcal{R}_N(z) \equiv \frac{z}{1-z}-\frac{z^N}{1-z^N}\,.
\end{equation}
From the above expression, one may readily extract the simple pole divergence at $q_T=q_U$,
\begin{equation}
Y_{T^{2n+1}}(T,U;\mathfrak{a}) \sim \delta_{\mathfrak{a}\infty}\,\frac{(2n+1)!}{2^{2n}}\,\frac{q_T}{q_U-q_T} \,.
\end{equation}
This singularity, together with holomorphy, modularity and eq. \eqref{YukawasCongr} uniquely fixes the expression of the Yukawa couplings
\begin{equation}
Y_{T^{2n+1}}(T,U;\mathfrak{a}) =2^{-2n} \, \frac{ E_{2n+2,\infty} (T)\,\mathcal{F}_{\mathfrak{a}}(1+n,1,-2n;U) }{j_\infty(T)-j_{\mathfrak{a}}(U)} \,,
\end{equation}
where it is understood that the holomorphic Eisenstein series, the Niebur-Poincar\'e series and the Hauptmodul pertain to the Hecke congruence subgroup $\varGamma _0 (N)$. For $N=1$ one recovers the ${\rm SL}(2;\mathbb Z)$ result \eqref{YukawasSL2Z}. Moreover, using the results of \cite{Angelantonj:2013eja}, it is straightforward to express the generalised Yukawa couplings $Y_{T^{2n+1}}(T,U;\mathfrak{a})$ in terms of the holomorphic modular forms of $\varGamma_0(N)$. For instance, one gets
\begin{equation}
\begin{split}
Y_{T^3} (T,U;\infty ) &= \frac{1}{6} \, \frac{ \left[ 4 X^2 (T) - E_4 (T) \right] \left[ E_6 (U) - 2 X (U) E_4 (U) \right]}{\varDelta_8 (U) \, \left[j_\infty (T) - j_\infty (U)\right]} \,,
\\
Y_{T^3} (T,U;0 ) &=\frac{1}{24} \, \frac{ \left[ 4 X^2 (T) - E_4 (T) \right] \left[ E_6 (U) + X (U) E_4 (U) \right]}{\varDelta_8 (U)\, \left[ j_\infty (T) - j_0 (U)\right] } \,,
\end{split}
\end{equation}
for $N=2$, and
\begin{equation}
\begin{split}
Y_{T^3} (T,U;\infty ) &=\frac{3}{2048} \, \frac{\left[ 9 X^2 (T) - 4 E_4 (T)\right]\, \left[ 9 X^2 (U) - 4 E_4 (U)\right]}{\varDelta_6 (U) \, \left[ j_\infty (T) - j_\infty (U)\right]}\,,
\\
Y_{T^3} (T,U;0 ) &=\frac{3}{2048} \, \frac{\left[ 9 X^2 (T) - 4 E_4 (T)\right]\, \left[ X^2 (U) - 4 E_4 (U)\right]}{\varDelta_6 (U) \, \left[ j_\infty (T) - j_0 (U)\right]}\,,
\end{split}
\end{equation}
for $N=3$. In these expressions $E_w (z)$ are the holomorphic Eisenstein series of ${\rm SL} (2;\mathbb{Z} )$, $\varDelta_w$ is the weight-$w$ cusp form of $\varGamma_0 (N)$,
and $X (z ) = E_2 (z) - N E_2 (Nz)$ is the weight-two holomorphic modular form of $\varGamma_0 (N)$ \cite{Angelantonj:2013eja}.
The generalisation to higher values of $\kappa$, with $(\kappa,N)=1$, is also straightforward and may be obtained by acting with the Hecke operator on the $\kappa=1$ result,
\begin{equation}
Y_{T^{2n+1}}(T,U,\kappa;\mathfrak{a}) = \kappa^{-n}\, H^{(N)}_\kappa\cdot Y_{T^{2n+1}}(T,U;\mathfrak{a}) \,.
\end{equation}
\section*{Acknowledgements} We are grateful to Per Berglund and Kathrin Bringmann for discussions. C.A. would like to thank the TH Unit at CERN and the Arnold Sommerfeld Centre at the Ludwig-Maximilians-Universit\"at M\"unchen and the Max-Planck-Institut f\"ur Physik in M\"unchen for hospitality during different stages of this project.
I.F. would like to thank to thank the Physics Department of Torino University for hospitality.
This work was partially supported by the European ERC Advanced Grant no. 226455 ``Supersymmetry, Quantum Gravity and Gauge Fields'' (SUPERFIELDS) and by the Compagnia di San Paolo contract ``Modern Application in String Theory'' (MAST) TO-Call3-2012-0088.
|
\section{Introduction}
The random sampling of combinatorial structures is an active topic, with many available techniques.
Several general methods have emerged which supply efficient algorithms in many cases of interest, e.g., rejection sampling~\cite{Rejection}, Boltzmann sampling~\cite{Boltzmann}, Markov chains~\cite{huber2015perfect, JerrumSinclair, MarkovChainBook, ProppWilson}.
Each method presents difficulties: rejection sampling provides exact samples in finite time, but that finite time is often too large to be useful in practice; Boltzmann sampling produces an object which does not always satisfy the restrictions; rapidly mixing Markov chains are often easily fashioned to a problem, along with bounds on the mixing time (though not always), but they are notably inexact in finite time;
and while Markov chain coupling from the past is an exact sampling method, it requires constructing a coupling over the complete set of objects, which can be practical when there is some type of monotonicity.
There are many effective approaches to random sampling from the set of $9\times 9$ Sudoku matrices, which is defined as the collection of $9 \times 9$ integer-valued matrices such that each row and column is a permutation of the set $\{1,\ldots, 9\}$, and certain $3\times 3$ sub-blocks (see~\eqref{subblocks}) are also permutations of $\{1,\ldots, 9\}$.
An efficient backtracking algorithm is utilized in \cite{DeSalvoSudoku}, though no quantitative bounds are given for the bias.
Another approach is to fashion a Markov chain with uniform stationary distribution; this was described in~\cite{SudokuMC}, although it was not proved to be rapidly mixing.
Importance sampling delivers a collection of Sudoku matrices as well as weights which allows one to calculate unbiased estimates of statistical parameters; this was done in \cite{RidderIS}.
We present several algorithms which are an application of probabilistic divide-and-conquer (PDC) \cite{PDC}.
The central idea is the division of the sample space into two smaller parts, each of which can be sampled separately, and then appropriately pieced back together to form an exact sample.
The method is versatile enough to be adapted to other types of numerical tables, and we have chosen the set of Latin squares as a natural complementary structure; a Latin square of order~$n$ is an $n \times n$ integer-valued matrix such that each row and column is a permutation of $\{1,\ldots, n\}$.
There is also a generalization of Sudoku matrices which can be sampled using PDC, presented in Section~\ref{Sudoku:RC}.
There are many ways to implement PDC, and in this paper we implement a version called \emph{almost deterministic second half}, which divides the sample space into two pieces, one of which can be sampled efficiently using a brute force approach. There are other PDC algorithms: self-similar PDC (see~\cite[Section~3.5]{PDC}) often provides an asymptotically efficient algorithm, but requires detailed information about the sample space; PDC with deterministic second half (see~\cite[Section~3.3]{PDC},~\cite{PDCDSH}) requires almost no information about the sample space, and is such that given one of the pieces, the second piece is uniquely determined; PDC with the recursive method (see~\cite{DeSalvoImprovements}) is such that given one of the pieces, the second piece can be efficiently sampled using the recursive method~\cite{NW, NWBook}; the advantage of our current approach is that it can be customized to suit the available knowledge of the sample space.
When we do have detailed knowledge about the sample space, e.g., the number of ways of completing a partially generated object, as is the case for Sudoku matrices, we show how one can more optimally design a PDC algorithm.
The random sampling of Latin squares and Sudoku matrices is a notoriously difficult problem, and we can only claim in the present work to have made modest improvements to existing algorithms.
PDC offers a versatile method for approaching the problem, and allows the fashioner to design an algorithm tailored to the particular computational aspects of the problem available.
In addition, our algorithms do not require extensive auxiliary constructions or complicated transformations.
The paper is organized as follows.
In Section~\ref{definitions} we give the definitions of a Latin Square of order $n$ and a Sudoku matrix.
In Section~\ref{PDC}, we introduce PDC and present some simple rejection sampling algorithms.
In Section~\ref{section:Sudoku} we present a PDC algorithm for the random sampling of Sudoku matrices, an analysis of the cost, and some alternative PDC parameterizations.
We do the same for Latin squares of order~$n$ in Section~\ref{section:Latin:squares}.
In Section~\ref{section:other} we review other approaches, and in Section~\ref{final:remarks} we explain our initial motivation and provide a link to a code implementation in C++.
\section{Definitions}\label{definitions}
For any $n \geq 1$, a Latin square of order~$n$ is an $n \times n$ matrix which satisfies the following row and column constraints, heretofore referred to as the \emph{Latin square conditions}:
\begin{itemize}
\item each row, labelled $R_1, \ldots, R_n$ is a permutation of $\{1,2,\ldots, n\}$;
\item each column, labelled $C_1, \ldots, C_n$ is a permutation of $\{1,2,\ldots, n\}$.
\end{itemize}
We let $LS_n$ denote the set of all Latin squares of order $n$.
By a \emph{partially completed} matrix, e.g., partially completed Latin square, we mean a matrix that has some subset of entries filled in which do not a priori violate the conditions.
No assumption is implied in general as to whether such a matrix can be completed.
We shall primarily be interested in partially completed matrices whose first $k$ rows are filled in, although there are several possible applications of PDC which could be exploited in more general circumstances, see Section~\ref{other:PDC:algorithms:LS}.
A partially completed Latin square of order~$n$ with the first $k$ rows filled in is called a \emph{$k \times n$ Latin rectangle}, for which much is already known, see for example~\cite{godsil1990asymptotic}.
A Sudoku matrix is a Latin square of order $9$ which also satisfies the additional block constraints:
there are nine $3 \times 3$ sub-blocks, labelled $B_1, B_2, \ldots, B_9$, each of which is a permutation of $\{1,2,\ldots, 9\}$; these blocks are indicated below.
\begin{equation}\label{subblocks}
\begin{array}{|ccc|ccc|ccc|}\hline
& & & & & & & & \\
& B_1 & & & B_2 & & &B_3 & \\
& & & & & & & & \\ \hline
& & & & & & & & \\
& B_4 & & & B_5 & & &B_6 & \\
& & & & & & & & \\ \hline
& & & & & & & & \\
& B_7 & & & B_8 & & &B_9 & \\
& & & & & & & & \\ \hline
\end{array}
\end{equation}
We refer to these three constraints -- row, column and block -- as the \emph{Sudoku conditions}, and we let~$S$ denote the set of all Sudoku matrices.
In the calculations that follow, $a \approx b$ is simply the elementary meaning where $a$ and/or $b$ have been rounded to some nearest value.
For a finite set $A$, the notation $|A|$ will denote its cardinality, i.e., the number of elements in the set~$A$.
In addition, we introduce notation for asymptotic analysis: suppose $f(n)$ and $g(n)$ are positive functions of a positive integer $n$, then
\begin{itemize}
\item we say $f(n) = O(g(n))$ if there exists a $C>0$ such that $\frac{f(n)}{g(n)} < C$ for all $n$ larger than some $n_0$;
\item we say $f(n) = o(g(n))$ if $\frac{f(n)}{g(n)} \to 0$ as $n$ tends to infinity;
\item we say $f(n) \sim g(n)$ if $\lim_{n\to\infty} \frac{f(n)}{g(n)} = 1$.
\end{itemize}
For real-valued $x$, let
\begin{itemize}
\item $[x]$ denote the integer closest to $x$; i.e., $x$ rounded to the nearest integer;
\item $\lfloor x \rfloor$ denote the largest integer smaller or equal to $x$, also known as the \emph{floor} of $x$.
\end{itemize}
\section{Probabilistic divide-and-conquer}
\label{PDC}
We begin by highlighting various simple algorithms for obtaining an element of $S$, a Sudoku matrix, uniformly at random (u.a.r.) over all possibilities.
Each of the approaches can be adapted to sampling an element of $LS_n$ in a straightforward manner, and so we specialize to $S$ for concreteness.
\begin{enumerate}
\item Sample each entry independent, and identically distributed (i.i.d.)~uniform over the set $\{1,2,\ldots,9\}$; restart if any of the Sudoku conditions are not satisfied.
\item Sample each row (or column or block) as i.i.d.~uniform over the set of all permutations of $\{1,2,\ldots, 9\}$; restart if any of the Sudoku conditions are not satisfied.
\item Let the first row be $(1,2,\ldots, 9)$. Sample rows 2 through 8 as i.i.d.~uniform over the set of all fixed--point free permutations of $\{1,2,\ldots, 9\}$; if row 9 can be completed, with all Sudoku conditions satisfied, then complete it, and finally apply a random permutation of $\{1,\ldots, 9\}$ to all entries and return the matrix; otherwise, restart.
\end{enumerate}
These algorithms are generally known as \emph{rejection algorithms}, since the target set of objects $S$, also referred to as the target region, lies within the complete sample space $\Omega$.
Since the algorithm generates a sample uniformly over all elements of $\Omega$, any Sudoku matrix generated by such a hard rejection algorithm is also uniform over $S$ (see, e.g.,~\cite{Yordzhev2} for a more detailed analysis in a more general setting).
The expected number of times one must sample elements of $\Omega$ before obtaining a sample from $S$ is simply the quotient $|\Omega| / |S|$.
It was shown in \cite{FelgenJarvis} that the number of Sudoku matrices is exactly
\[ |S| = 6670903752021072936960 \approx 6.67 \times 10^{21}.
\]
This number is too large to simply list all of the elements and select one at random.
There are $9^{81} \approx 2.0 \times 10^{77}$ matrices in $\Omega_1 := \{1,\ldots,9\}^{9\times 9}$, so for the first, most na\"ive approach of i.i.d.~entry sampling, the expected number of times one must sample from $\Omega_1$ is $|\Omega_1|/|S| \approx 3.3\times 10^{56}$.
There are $|\Omega_2| := \binom{9!}{9}\approx 3.0\times 10^{44}$ different ways of selecting $9$ distinct permutations of $\{1,2,\ldots, 9\}$, whence, $|\Omega_2|/|S| \approx 4.5\times 10^{22}$.
The third algorithm is the most respectable of the three, as it eliminates two of the rows and reduces the number of possible permutations from $9!$ to $[9!/e]$.
However, since we are fixing the first row, we can only sample from those elements of $S$ which also have the same first row; thus, we must normalize by $|S|/9!$ instead of $|S|$.
We have $\Omega_3 := \{$the set of all collections of seven fixed--point free permutations of $\{1,2,\ldots,9\}\}$, hence,
\[ \frac{|\Omega_3|}{|S|/9!} \approx \frac{ (9!/e)^7 }{1.8\times 10^{16}} \approx 4.2\times 10^{19}. \]
None of these rejection rates is practical.
At this stage, there are many different ways to proceed.
One approach is to relax the demand that the distribution on elements of $S$ is \emph{exactly} uniform, and instead adopt algorithms which are faster but not guaranteed to be uniform.
Another approach is to take advantage of more detailed properties of the elements in $S,$ which we now explore using PDC.
Rejection sampling algorithms can very often be improved, with very little extra information known about the target region, using PDC~\cite{PDC}.
We assume a decomposition of the sample space $\Omega$ of the form $\Omega = \mathcal{A} \times \mathcal{B}$, where elements in $\Omega$ can be sampled using a distribution on $\mathcal{A}$, and a distribution on $\mathcal{B}$, such that random variables $A \in \mathcal{A}$ and $B \in \mathcal{B}$ are mutually independent; we denote their distributions by $\cL(A)$ and $\cL(B)$, respectively.
In addition, we assume the target distribution on $S \subset \Omega$, denoted by $\cL(S)$, is of the form
\[ \cL(S) = \cL( (A,B) | h(A,B)=1 ),
\]
where $h: \mathcal{A}\times \mathcal{B} \to \{0,1\}$ is some (measurable) functional which determines whether or not the pair $(A,B)$ lies in the target set.
The PDC Lemma \cite[Lemma 2.1]{PDC} states that, assuming $\mathbb{P}(h(A,B)=1) > 0$, we have $\cL(S) = \cL(X,Y)$, where
\[ \cL(X) = \cL( A | h(A,B) = 1), \qquad \cL(Y | X = x) = \cL(B | h(x,B) = 1). \]
An algorithm to sample from $\cL(S)$ is then as follows (see~\cite[Algorithm~2]{PDC}):
\begin{enumerate}
\item Generate sample from $\cL(A\, |\, h(A,B) = 1),$ call it $x$.
\item Generate sample from $\cL(B\, |\, h(x,B) = 1),$ call it $y$.
\item Return $(x,y)$.
\end{enumerate}
Often, however, the conditional distributions are not known, and so the more practical PDC algorithm utilizes von Neumann's rejection sampling approach \cite{Rejection}, which allows us to sample from the conditional distribution $\cL(A | h(A,B)=1)$ using $\cL(A)$ and a biased coin.
\begin{algorithm}[H]{\rm
\begin{algorithmic}[1]
\STATE Fix any $\alpha$ such that $\displaystyle \max_{a\in \mathcal{A}} \P(h(a,B)=0) \leq \alpha \leq 1$. \\
\STATE For each $a \in \mathcal{A}$, define $\displaystyle t(a) := \frac{\P(h(a,B)=1)}{\alpha}$. \\
\STATE\label{line:A} Generate sample from $\cL(A),$ call it $a$.
\STATE Reject $a$ with probability $1-t(a)$; \ otherwise, restart from Line~\ref{line:A}.
\STATE Generate sample from $\cL(B\, |\, h(a,B) = 1),$ call it $y$.
\STATE Return $(a,y)$.
\end{algorithmic}}
\caption{\cite{PDC} Probabilistic Divide-and-Conquer via von Neumann}
\label{PDC procedure von Neumann}
\end{algorithm}
Presently, in order to sample directly from the conditional distribution $\cL(B\, |\, h(a,B) = 1),$ we choose $\mathcal{A}$ such that $\cL(B\, |\, h(a,B) = 1)$ can be sampled effectively using a brute force approach; we call this \emph{PDC with almost deterministic second half}.
Also, since the rejection probability $t(a)$ must satisfy, for some $0<\alpha\leq 1$,
\[ t(a) = \frac{\P( h(a,B) = 1)}{\alpha} \leq 1, \qquad \text{for all $a\in \mathcal{A}$,}\]
the optimal choice of $\alpha$ is $\max_a \P(h(a,B)=1)$, \emph{but the algorithm is still valid and unbiased for any $\alpha$ larger than this value.}
Thus, there are many ways to apply PDC, starting with an appropriate selection of sets $\mathcal{A}$ and $\mathcal{B}$, and it suffices to choose any universal upper bound of $\P(h(a,B)=1)$, where greater efficiency is achieved by selecting the optimal value of $\alpha$.
\section{Sudoku matrices}
\label{section:Sudoku}
\subsection{A PDC algorithm}
\label{section:Sudoku:algorithm}
We now describe an algorithm for Sudoku matrices which takes advantage of Algorithm~\ref{PDC procedure von Neumann}.
Given the first three rows $R_1, R_2, R_3$ of a partially completed Sudoku matrix, let $P_9' \equiv P_9'(R_1, R_2, R_3) \equiv P_9'(B_1, B_2, B_3)$ denote the set of all possible permutations of $\{1,2,\ldots, 9\}$ which would not violate the Sudoku conditions if placed in row four.
Also, we denote by $U$ a random variable with the uniform distribution over the interval $(0,1)$, independent of all other random variables, and each occurrence of $u$ in an algorithm means to generate a random variate from the distribution of $U$.
\begin{algorithm}[H]
\caption{PDC Sudoku algorithm.}
\begin{algorithmic}[1]
\STATE Let $B_1 = (1,2,\ldots,9)$.
\STATE Select $B_2, B_3$ in proportion to the number of completable Sudoku matrices.\label{Sudoku:special}
\STATE Generate $(R_4, R_5, R_6, R_7),$ each an i.i.d.~uniformly random element from $P_9'$. \label{line:generate}
\STATE Let $d$ denote the number of possible completions given $(R_1, R_2, R_3, R_4, R_5, R_6, R_7)$.
\IF {$u < \frac{d}{16}$,}\label{line:rejection}
\STATE\label{completion} Select a completion uniformly at random from the $d$ possible completions.
\ELSE \STATE Goto 3.
\ENDIF
\STATE Apply a random permutation to $(C_4, C_5, C_6)$.
\STATE Apply a random permutation to $(C_7, C_8, C_9)$.
\STATE Apply a random permutation to $( (C_4, C_5, C_6), (C_7, C_8, C_9) )$.
\STATE Apply a random permutation of $\{1,2,\ldots, 9\}$ to the entries and return.
\end{algorithmic}
\label{Sudoku}
\end{algorithm}
Line~2 demands some further explanation.
By considering various symmetries in the permutations of columns, it was shown in \cite{FelgenJarvis} that there are only 36288 essentially unique completions to blocks $B_2$ and $B_3$;\footnote{This number can be simplified to 71 after taking more symmetries into account~\cite{SudokuWebsite}.} and, for each configuration of blocks $B_1$, $B_2$, $B_3$, there is a certain number of completable Sudoku matrices, which was calculated via brute force.
A downloadable file containing these enumerations is available online, see~\cite{SudokuWebsite};
the first few lines look like the following:
\begin{verbatim}
[456789,789123,123456] => 108374976
[456789,789123,123465] => 102543168
[456789,789123,123546] => 102543168
[456789,789123,123564] => 100231616
[456789,789123,123645] => 100231616
...
\end{verbatim}
The list of three 6--tuples represent columns 3 through 9 of the first three rows, and the integer value at the end of the arrow counts the number of completable Sudoku matrices given these first three rows. For example, the text file says that the matrix with top three rows given by
\[
\begin{Young}
1 & 2 &3 & 4& 5& 6& 7& 8&9 \cr
4 & 5 &6 & 7& $8$&9 &1 &2 &3 \cr
7 & 8 & 9 & 1 & 2 & 3 & 6 & 4 & 5 \cr
\end{Young}
\]
has precisely 100231616 completions of the bottom 6 rows which are valid Sudoku matrices.
Line~\ref{Sudoku:special} is thus a straightforward sampling of the entries in $B_2$ and $B_3,$ \emph{in proportion to the number of completable Sudoku matrices}.
There are two reasonable ways to perform this sampling.
Denote by $x_i$ the number of possible completions given outcome $i$, $i=1,\ldots, 36288$.
One can normalize by the sum of all the completions and generate a value using the discrete probability distribution $\{x_i / \sum_{j} x_j\}_{i=1,\ldots,36288}$; this is not so ideal in general due to floating point considerations.
Instead, we apply PDC in our implementation by sampling uniformly from among the 36288 possibilities first, say we select outcome $i$, and we accept this sample with probability $x_i / \max_j x_j$.
Line~\ref{completion} is then the last task of the algorithm, and for this we have implemented a brute force approach which enumerates through all possibilities and picks one uniformly at random.
Specifically, each column has a set of two elements $\{a_i, b_i\}$, $i=1,\ldots, 9$, which must be used in that column in the final two rows, in some order, and we are able to greatly reduce the enumeration of all a priori $2^9 = 512$ possibilities by taking various symmetries into account; see for example the proof of Lemma~\ref{upper bound}.
\subsection{Cost of Algorithm~\ref{Sudoku}}
\begin{lemma}\label{upper bound}
Assume the first seven rows of a Sudoku matrix have been filled in, and none of the Sudoku conditions are violated based on these first seven rows. Then the total number of possible completions of this matrix which also do not violate the Sudoku conditions is at most 16. This bound cannot be improved.
\end{lemma}
\begin{proof}
WLOG, assume the last two rows of column $j$ can be completed by $\{2j-1,2j\}$ for $j=1,2,3$. We shall denote this by the following table
\[
\begin{Young}
1 & 3 &5 & & & & & & \cr
2 & 4 &6 & & & & & & \cr
\end{Young}
\]
By interchanging $2j-1$ and $2j$ for $j=1,2,3$, we obtain $2^3$ possible combinations.
For the elements $7,8,9$, there are two possibilities: either $7$ and $8$ share the same column, or they do not.
\begin{itemize}
\item In the case where they share the same column, we have WLOG
\[
\begin{Young}
1 & 3 &5 & 7& 9& & & & \cr
2 & 4 &6 & 8& $a$& & & & \cr
\end{Young}
\]
which implies another factor of 2 by exchanging the role of $7$ and $8$; however, whichever element is paired with the 9, say $a=5$, determines which row the 9 lies in, and so there are at most 16 possible completions.
\item In the case where $7$ and $8$ do not share the same column, whichever element the $7, 8,$ and $9$ are paired with uniquely determine their row, and so there are at most $8$ possible completions in this case.
\end{itemize}
Thus, there are at most $2^4=16$ possible completions given the first seven rows of a partially completed Sudoku matrix.
The following matrix (in reduced form) was generated by Algorithm~\ref{Sudoku}.
The first seven rows of this matrix allow for 16 potential completions, thus, 16 is a tight upper bound.
\[\begin{Young}
1 & 2 & 3 & 4 & 5 & 9 & 6 & 7 & 8 \cr
4 & 5 & 6 & 7 & 8 & 2 & 3 & 1 & 9 \cr
7 & 8 & 9 & 1 & 6 & 3 & 2 & 5 & 4 \cr
9 & 4 & 8 & 2 & 1 & 7 & 5 & 6 & 3 \cr
5 & 7 & 2 & 3 & 4 & 6 & 8 & 9 & 1 \cr
6 & 3 & 1 & 5 & 9 & 8 & 4 & 2 & 7 \cr
3 & 6 & 7 & 9 & 2 & 4 & 1 & 8 & 5 \cr
8 & 9 & 5 & 6 & 3 & 1 & 7 & 4 & 2 \cr
2 & 1 & 4 & 8 & 7 & 5 & 9 & 3 & 6 \cr
\end{Young}
\qedhere\]
\end{proof}
We now justify Line~2 of Algorithm~\ref{Sudoku}.
In \cite{FelgenJarvis}, the total number of Sudoku matrices is found by first reducing by symmetry the total number of possible completed first three rows; the number given is 36288.
For each of these 36288 configurations of the first three rows, a brute force enumeration is performed for the number of possible Sudoku matrices that could be completed.
\begin{lemma}[\cite{FelgenJarvis}]\label{table}
The table of 36288 possible configurations of $B_2$ and $B_3$ contains, for each configuration, the number of possible Sudoku matrices that can be completed given the first three rows. These numbers all lie between 94888576 and 108374976.
\end{lemma}
The fact that these values are relatively constant has a number of quantitative and qualitative interpretations.
From a rejection sampling perspective, it means that we could sample a configuration uniformly at random and then reject each generated configuration in proportion to the number of completions, \emph{normalized by the maximum possible}, i.e., 108374976.
Thus, the probability of rejecting a sample is at worst $1 - 94888576 / 108374976 \approx 0.12444$.
Practically, this means that sampling of the first three rows uniformly from among the 36288 possible choices and applying a rejection is efficient.
Also, it says that each configuration is approximately as completable as any other configuration; i.e., accepting any completable configuration of the first three rows introduces a small bias.
Thus, the main cost of the algorithm is the rejection sampling of rows 4 through 7.
Through a complete enumeration of all 36288 cases, the number of permutations in $P_9'$ is always between 12000 and 12096.
There are thus at most $12096^4 \approx 2\times 10^{16}$ different combinations of permutations that can be placed in rows 4 through 7, but only about $10^8$ valid completions to a Sudoku matrix are possible.
This means that on average we must sample about $2\times 10^8$ 4--tuples before we get lucky and obtain a quadruplet that yields a completable Sudoku matrix.
Even after we obtain a completable quadruplet, however, we must still survive the final rejection in Line~\ref{line:rejection}, which comes from the fact that different quadruplets yield a different number of completable Sudoku matrices.
\begin{lemma}[\cite{devroye}]\label{rejection lemma}
In a rejection sampling algorithm, let $f$ denote the target distribution, and $g$ denote the sampling distribution;
then the rejection probability, see for example Line~\ref{line:A} in Algorithm~\ref{PDC procedure von Neumann}, for exact sampling from distribution $f$ is given by
\[ t(a) = C\, \frac{f(a)}{g(a)}, \qquad a \in \mathcal{A},\]
where the optimal value for $C$ is given by
\[ C^\ast := \left(\max_a \frac{f(a)}{g(a)}\right)^{-1}. \]
The algorithm is unbiased for any $C \leq C^\ast$.
Algorithms which utilize rejection sampling have a geometrically distributed number of generations before acceptance, with expected value given by $\frac{1}{C}$.
\end{lemma}
\begin{theorem}\label{Sudoku theorem}
Algorithm~\ref{Sudoku} samples uniformly over the set of all Sudoku matrices.
The number of times to sample blocks $B_2$ and $B_3$ by choosing one of the 36288 possible choices u.a.r.~and using rejection sampling is geometrically distributed with expected value
\begin{equation}\label{rejection 1} \frac{36288\cdot 108374976}{3546146300288} \approx 1.1. \end{equation}
The number of times to sample rows $R_4, R_5, R_6, R_7$ is geometrically distributed with expected value at least $3.0 \times 10^9$ and at most $3.6 \times 10^9$.
\end{theorem}
\begin{proof}
The uniformity of the algorithm follows by two applications of probabilistic divide-and-conquer given by Algorithm~\ref{PDC procedure von Neumann}. We now calculate the precise rejection probabilities.
The first application of PDC is given by the division
\[ \text{$A = (B_1, B_2, B_3)$, \quad $B = (R_4, R_5, R_6, R_7, R_8, R_9)$.} \]
The rejection probabilities $t(a)$ are given by the number of possible completions normalized by the maximum number of possible completions; these values are from Lemma~\ref{table}.
We index each of the possible configurations of $B_2$ and $B_3$ by $i=1,2,\ldots, 36288$.
Then we select one uniformly at random, hence $g(i) = 1/36288$, but the distribution desired is the one that weights each configuration in proportion to the number of possible completions, denoted by $x_i$ previously, which is contained in the table of values computed in \cite{FelgenJarvis}. That is,
\[ f(i) = \frac{ x_i }{3546146300288}, \qquad i=1,\ldots, 36288, \]
where $3546146300288 = \sum_j x_j$ is the total number of possible completions by all elements in the table, which makes $f(i)$ a probability distribution.
The expected number of times to sample from blocks $B_2$ and $B_3$ is thus
\begin{equation*} \frac{1}{C} = \left(\max_i \frac{f(i)}{g(i)}\right) = \frac{36288\cdot (\max_i x_i)}{3546146300288}, \end{equation*}
where
\[ \max_i x_i = 108374976 \]
was computed by brute force over the set of all possibilities.
We have just proved~\eqref{rejection 1}.
The second application of PDC is the division
\[ \text{$A = (R_4, R_5, R_6, R_7)$, \quad $B = (R_8, R_9)$.}\]
The rejection probability $t(a)$ is given by the number of possible completions (calculated via brute force) given $a = (r_4, \ldots, r_7)$, normalized by 16, the optimal upper bound provided by Lemma~\ref{upper bound}.
Rather than sample from the set of all permutations of $\{1,\ldots,9\}$, since we have already accepted the first three rows of the matrix, we can automatically discard any permutations which would violate the Sudoku conditions. Thus, we sample these four rows from the set $P_9'$, which depends on the particular configuration of $B_2$ and $B_3$ accepted in the first part.
By brute force calculation over all 36288 possible first three rows, we have found that $12000 \leq |P_9'| \leq 12096.$
Thus,
\[ \frac{1}{12096^4} \leq g(j) = \frac{1}{|P_9'|^4} \leq \frac{1}{12000^4}, \qquad \text{for all $j$}. \]
Similarly, we have
\[ \frac{8}{108374976} \leq f(j) = \frac{\text{\# completions given first seven rows}}{\text{\# completions given first three rows}} \leq \frac{16}{94888576}. \]
Whence,
\[ 3.0 \times 10^9 \approx \frac{12000^4\times 16}{108374976} \leq \frac 1C = \left( \max_j \frac{f(j)}{g(j)} \right) \leq \frac{12096^4\times 16}{94888576} \approx 3.6 \times 10^9. \qedhere\]
\end{proof}
\subsection{Alternative PDC parameterizations for Sudoku matrices}
\label{other:PDC:algorithms}
There are two simple alternative approaches to Algorithm~\ref{Sudoku} that require minimal modification of the original algorithm; we indicate the replacement for lines~\ref{line:generate}--\ref{line:rejection} below in each case.
The first approach is to randomly sample $R_4$ through $R_8$.
3: Generate $(R_4, R_5, R_6, R_7, R_8),$ each an i.i.d.~uniformly random element of $P_9'$. \\
4: Let $d$ denote the number of possible completions given $(R_1, R_2, R_3, R_4, R_5, R_6, R_7, R_8)$. \\
5: {\bf if} {$d=1$,} {\bf then}
That is, conditional on the existence of a completion to $R_9$, we may simply accept the sample $R_4$ through $R_8$ and fill in the \emph{unique} completion to $R_9$; this variation is called PDC deterministic second half \cite{PDC}, see also \cite{PDCDSH}.
(In general, PDC deterministic second half also requires a rejection step, but in the case when $\cL(B\,|\,h(a,B)=1)$ is a uniform distribution over a singleton set for each $a \in \mathcal{A}$ which can be used to complete a sample, all samples are rejected with probability 0.)
This approach increases the probability of rejection in Line~\ref{line:rejection} considerably, see Proposition~\ref{prop:R8} below, but eliminates the task in Line~\ref{completion}.
\begin{proposition}\label{prop:R8}
Suppose Line~\ref{line:generate} in Algorithm~\ref{Sudoku} instead samples rows~$R_4$ through $R_8$.
Then the expected number of rejections until the first eight rows are a partially completed Sudoku matrix is within the interval
\[ \left[\frac{12000^5}{108374976}, \frac{12096^5}{94888576} \right] \approx [2.3, 2.73] \times 10^{12}. \]
Then, once these first eight rows are a partially completed Sudoku matrix, the probability of rejection is 0, and the last row is uniquely determined.
\end{proposition}
The proof of Proposition~\ref{prop:R8} is straightforward, with upper bounds on the number of possible completions provided by Lemma~\ref{completion:lemma}, and the rejection probability provided by~\cite[Theorem~7.1]{PDCDSH}.
Alternatively, we can instead randomly sample $R_4$ through $R_6$, and reject these three rows in proportion to the number of possible completions of the last three rows, indicated below:
3: Generate $(R_4, R_5, R_6),$ each an i.i.d.~uniformly random element from $P_9'$. \\
4: Let $d$ denote the number of possible completions given $(R_1, R_2, R_3, R_4, R_5, R_6)$. \\
5: {\bf if} {$u < \frac{d}{\lfloor 2^{4.5}6^3\rfloor}$,} {\bf then}
From a theoretical point of view, this approach is more ideal, since the probability of rejection in Line~\ref{line:rejection} is smaller (since $d$ is much less likely to be 0).
However, completing the sample as in Line~\ref{completion} can also be a bottleneck, since now there are a priori $3^9 \approx 2\times 10^4$ possible completions that must be checked, even though we can greatly reduce this number by considering various symmetries.
\begin{proposition}\label{prop:R6}
Suppose Line~\ref{line:generate} in Algorithm~\ref{Sudoku} instead samples rows~$R_4$ through $R_6$.
Then the expected number of rejections is within the interval
\[ \left[\frac{\lfloor 2^{4.5}\,6^{3}\rfloor \, 12000^3}{108374976}, \frac{\lfloor 2^{4.5}\,6^3\rfloor \, 12096^3}{94888576} \right] \approx [7.8, 9.1] \times 10^{7}. \]
Then, once these first six rows are determined, the last three rows can be completed in at most $\lfloor 2^{4.5} 6^3\rfloor$ ways.
\end{proposition}
The upper bound on the number of possible completions in Proposition~\ref{prop:R6} is borrowed from the theory of Latin squares, see Section~\ref{sect:Latin:analysis}, specifically Lemma~\ref{completion:lemma} using the values $n=9$ and $k=6$.
Despite the much smaller expected number of rejections in Proposition~\ref{prop:R6}, the reason why we champion Algorithm~\ref{Sudoku} presently is due to our brute force implementation of sampling from $\cL(B\, |\, h(a,B)=1)$, which is optimized to enumerate between all potential completions using fast bit--wise operations in C++, whereas in the more general case we do not have this option encoded.
We emphasize that while we have not yet efficiently coded this algorithm for more general cases, should such a module be coded efficiently it could very easily be faster in practice than Algorithm~\ref{Sudoku}.
In addition, it may be possible to improve upon the upper bound of $\lfloor 2^{4.5} 6^3\rfloor$ in this special case.
To help facilitate such an endeavor, we give a motivating example.
The upper bound in Lemma~\ref{completion:lemma} for the number of completable $(n-3) \times n$ Latin rectangles is $2^{n/2}\,6^{n/3}$.
With $n=9$, we have $\lfloor 2^{n/2}\,6^{n/3}\rfloor = 4887.$ However, the largest observed value for partially completed Sudoku matrices in a sample of size 1000 was 288.
If, in fact, 288 is the smallest upper bound, and was used instead of 4887, then the run--time of the algorithm which samples all but the final three rows would be reduced by a factor of about 17.
\subsection{Further reduction by symmetries}
When various symmetries are taken into account, one can reduce the total number of Sudoku matrices to $5472730538 \approx 5.4\times 10^9$ essentially different Sudoku matrices, see~\cite{RusselJarvis}.
This number is certainly more practical, and a comprehensive list of such matrices could be stored for random access if the memory was available, thus offering an algorithm which is $O(1)$ in terms of our costing of random generation; of course, one would also have to implement the various transformations, but this cost is certainly not prohibitive.
The advantage with our approach is that its memory requirements are entirely practical, even for a computer of modest means, and can be generalized in a straightforward manner.
Nevertheless, should Algorithm~\ref{Sudoku} be deemed not efficient enough, a more efficient random sampling algorithm may be achievable by taking these symmetries into account.
\section{Latin squares}
\label{section:Latin:squares}
\subsection{A PDC algorithm}
The upper bound in Lemma~\ref{upper bound} is specialized to the set of Sudoku matrices, and in general such tight bounds are not available.
Nevertheless, we present a similar PDC algorithm in this section for Latin squares of order~$n$ for any $n \geq 5$.
\begin{algorithm}[H]
\caption{PDC Uniform sampling from $LS_n$}
\begin{algorithmic}[1]
\STATE Let $R_1 = (1,\ldots,n)$.
\FOR {$i=2,\ldots, n-3$}
\STATE Generate $R_i$ uniformly from the set of fixed--point free permutations of $\{1,\ldots, n\}$.
\ENDFOR
\STATE Let $d$ denote the number of possible completions given $(R_1, \ldots, R_{n-3})$.\label{line:d:LS}
\IF {$U < \frac{d}{\lfloor 6^{\frac{n}{3}}2^{ \frac{n}{2}}\rfloor}$}\label{line:rejection:LS}
\STATE Select a completion uniformly at random from the $d$ possible completions.\label{line:complete}
\ELSE
\STATE Goto 2.
\ENDIF
\STATE Apply a random permutation to the rows of the matrix.
\STATE Apply a random permutation of $\{1,\ldots,n\}$ to the entries and return.
\end{algorithmic}
\label{Sudoku:n}
\end{algorithm}
As in Algorithm~\ref{Sudoku}, the selection of a completion in Line~\ref{line:complete} once the first set of rows has been accepted can be performed using brute force enumeration, or any alternative method which uniformly samples from the set of possible completions.
\subsection{Cost of Algorithm~\ref{Sudoku:n}}
\label{sect:Latin:analysis}
We now discuss the parameterized algorithm for Latin squares of order $n$.
There are two competing styles of analysis: one champions an analysis for a particular, finite value of $n$ chosen in advance, e.g., $n=9$ as in the previous section; the other is an asymptotic analysis of how the algorithm scales as $n$ tends to infinity.
Whereas our previous analysis for Sudoku matrices was specific to a fixed size~$9$, we shall now be primarily interested in how Algorithm~\ref{Sudoku:n} scales as $n$ tends to infinity.
For a partially completed Latin square of order $n$ with the first $k$ rows filled in, a more general bound for all $n$ and $k$ can be obtained using upper bounds on the permanent of an $n\times n$ matrix taking values in the set $\{0,1\}$, as was done in~\cite{vanLintWilson}.
We now adapt the argument in our setting.
\begin{lemma}\label{completion:lemma}
Suppose the first $k$ rows of any Latin square of order $n$ have been filled in, $k < n$, without violating any of the Latin square conditions.
\begin{enumerate}
\item There always exists at least one completion to a full Latin square of order~$n$, and when $k=n-1$ there is exactly one completion.
\item Let $C_{n,k}$ denote the maximum possible number of completions to a full Latin square of order~$n$ given these first $k$ rows. Then we have
\begin{equation}\label{Cnk} C_{n,k} \leq \prod_{\ell=1}^{n-k} (\ell!)^{n/\ell}. \end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
First, the existence of at least one completion of a partially completed Latin square is given by~\cite[Theorem~17.1]{vanLintWilson}, and the uniqueness of the completion when $k=n-1$ follows by taking $k=n-1$ in inequality~\eqref{Cnk}.
The remaining argument is the same as~\cite[Proof of Theorem~17.3]{vanLintWilson}.
That is, we let $B$ denote the $n\times n$ matrix where the $(i,j)$th entry is $1$ if element $i$ does not appear in the first $k$ rows of column $j$, and 0 otherwise.
Thus, the row sums of $B$ are all $n-k$, and so by~\cite[Theorem~11.5]{vanLintWilson}, we have
\[ per\, B \leq (n-k)!^{n/(n-k)}, \]
where $per\, B$ denotes the permanent of the matrix $B$, which is also the number of ways of completing the next row.
Thus, by applying the above inequality $n-k-1$ times, for completing rows $k+1, k+2, \ldots n-1$, we obtain
\[ C_{n,k} \leq \prod_{\ell=k}^{n-1} (n-\ell)!^{n/(n-\ell)}. \qedhere\]
\end{proof}
We shall also need for our analysis of asymptotic runtimes to estimate the magnitude of $|LS_n|$, which is still not precisely known asymptotically as $n$ tends to infinity, although the following asymptotic result is contained in~\cite[Theorem~17.3]{vanLintWilson}.
\begin{theorem}[\cite{vanLintWilson}]\label{Ln:asymptotic}
We have
\[ |LS_n|^{1/n^2} \sim e^{-2}\, n. \]
\end{theorem}
There are, however, well-known upper and lower bounds which shall at least give us the ability to estimate the efficiency of our algorithms, which we give below. See~\cite[Theorem~17.2]{vanLintWilson} for the lower bound and see~\cite[the proof of Theorem~17.3]{vanLintWilson} for the upper bound.
\begin{theorem}[\cite{vanLintWilson}]\label{LS:bounds}
We have
\[ \frac{(n!)^{2n}}{n^{n^2}} \leq |LS_n| \leq \prod_{k=1}^n (k!)^{n/k}. \]
\end{theorem}
\begin{theorem}\label{Latin square theorem}
Algorithm~\ref{Sudoku:n} samples uniformly over the set of all Latin squares of order $n$.
The number of times to sample rows $R_2,\ldots, R_{n-3}$ is geometrically distributed with expected value
\[E_n := \frac{\lfloor 6^{n/3}2^{n/2}\rfloor \left([n!/e]\right)^{n-4}}{|LS_n|/n!}, \qquad n \geq 5.\]
\end{theorem}
\begin{proof}
We note that $|LS_n|$ is normalized by $n!$ since we are taking the first row to be the permutation $(1,\ldots, n)$.
Each of the rows $R_i$, $i=2,\ldots, n-3$, is sampled from the set of all fixed-point-free permutations, of which there are $[n!/e]$ such permutations, and our sample space is the set of all $(n-4)$--tuples of fixed-point-free permutations.
The proof then proceeds in the same fashion as the proof of Theorem~\ref{Sudoku theorem}.
\end{proof}
A table of values is below, calculated from the exact quantities and rounded to a few significant digits; the values for $|LS_n|$ can be found in the recent survey~\cite{Stones}.
In addition, we calculate the asymptotic rate of increase of $E_n^{1/n^2}$ in Proposition~\ref{E:growth}.
\[
\begin{array}{|c|c|c|}\hline
n & |LS_n| & E_n \\ \hline\hline
5 & 1.61\times 10^5 & 8.15 \\
6 & 8.13\times 10^8 & 1.32 \times 10^2 \\
7 & 6.15\times 10^{13} & 1.10 \times 10^4 \\
8 & 1.09\times 10^{20} & 4.26\times 10^6 \\
9 & 5.52\times 10^{27} & 8.41\times 10^9 \\
10 & 9.98\times 10^{36} & 8.79\times 10^{13} \\
11 & 7.77\times 10^{47} & 5.03\times 10^{18} \\ \hline
\end{array}
\]
\begin{proposition}\label{E:growth}
We have
\[ E_n^{1/n^2} \sim e. \]
\end{proposition}
\begin{proof}
Recall Stirling's formula, which states that as $n$ tends to infinity, we have
\[ n! \sim \frac{n^n}{e^n} \sqrt{2\pi\, n}\, , \]
so
\[ n!^{1/n} \sim \frac{n}{e}\, . \]
We then have
\[ E_n^{1/n^2} = \left(\frac{\lfloor 6^{n/3}2^{n/2}\rfloor \left([n!/e]\right)^{n-3}}{|LS_n|}\right)^{1/n^2} \sim \frac{6^{1/(3n)}2^{1/(2n)} \left(n!/e\right)^{1/n-3/n^2}}{|LS_n|^{1/n^2}} \sim \frac{n/e}{n/e^2} \sim e.
\]
\end{proof}
These values indicate that the rejection probability is the dominating aspect of the computation.
See Section~\ref{other:PDC:algorithms:LS} for other ways to improve upon this rejection cost.
\subsection{Alternative PDC parameterizations for Latin squares}
\label{other:PDC:algorithms:LS}
Let $U_{n,k} := \prod_{\ell=1}^{n-k} (\ell!)^{n/\ell}$ denote the right-hand side of inequality~\eqref{Cnk}, i.e., an upper bound for the number of completions of a $k \times n$ Latin rectangle.
Algorithm~\ref{Sudoku:n} can be generalized with a parameter $k \geq 2$ so that rows $R_2, \ldots, R_{k}$ are sampled in Line~\ref{line:d:LS} for the first step in PDC.
In this case, the rejection probability in Line~\ref{line:rejection:LS} may be replaced with $d / U_{n,k}$, and the number of times to resample has expected value
\[ E_{n,k} := \frac{U_{n,k}\, [n!/e]^{n-k-1}}{L_n/n!}, \qquad n \geq 5, 1 \leq k \leq n-4.\]
We estimate the log of this rejection cost asymptotically as $n$ tends to infinity.
\begin{proposition}\label{asymptotic:Enk}
For
\begin{itemize}
\item $k = o(n)$, we have
\[ \log E_{n,k} \sim n^2 \log(n);\]
\item $k \sim (1-t)\, n$, for any $0<t<1$, we have
\[ \log E_{n,k} \sim (2t-1) n^2 \log(n);\]
\item $k = n - r\, n^\alpha$, for any $0<r<1$, and $0<\alpha<1$, we have
\[ \log E_{n,k} \sim r(1+\alpha)n^{1+\alpha} \log(n) - n^2 \log(n). \]
\end{itemize}
\end{proposition}
\begin{proof}
To estimate $\log E_{n,k}$, we note that the proof of~\cite[Theorem~17.3]{vanLintWilson} contains an approach to estimate both $\log |LS_n|$ and $\log|U_{n,k}|$; that is, we have
\[ \log |LS_n| \sim n^2 \log(n), \]
and in a similar fashion we obtain
\[ \log U_{n,k} \sim n(n-k) \log(n-k). \]
Thus, asymptotically as $n$ and $n-k$ tend to infinity, we have
\begin{align*}
\log E_{n,k} & = \log U_{n,k} + (n-k)\log(n!) - (n-k-1) - \log |LS_n| \\
& \sim n\, (n-k) \log(n-k) + (n-k) (n \log(n)) - n^2 \log(n).
\end{align*}
The proposition now follows by a straightforward calculation.
\end{proof}
There are several noteworthy aspects of the asymptotic analysis in Proposition~\ref{asymptotic:Enk}.
First, when $k = o(n)$, there is asymptotically as much uncertainty in the first $k$ rows as there are in the entire ensemble of Latin squares of order $n$.
Next, taking $k = n/2$, i.e., $t=2$, is a natural cutoff for when the uncertainty is no longer on the same exponential order as the entire ensemble, and indeed it is plausible that a self-similar PDC algorithm using this division may achieve an asymptotically optimal rejection rate.
Finally, the final item in Proposition~\ref{asymptotic:Enk} is an example of a low-rank parameterization (see for example~\cite{PIPARCS}), and yields the greatest immediate potential for generalizing the presented algorithms in a way which is still analytically tractable.
Of course, even with these divisions, one still has to sample from the remaining set of rows $R_{k+1}, \ldots, R_n$, which for $k$ moderately large is still a daunting task.
One can also apply PDC more generally, not just to $k \times n$ Latin rectangles, by filling in any amount of partial information.
There are many enticing partial results in this area, see~\cite[Theorem~17.4]{vanLintWilson} for an example which fills in an upper left rectangle, and~\cite[Theorem~17.5]{vanLintWilson} for an example which fills in an upper triangular section of entries.
There are also many cautionary theoretical results as well, see~\cite[Figure~17.4]{vanLintWilson} and~\cite{colbourn1984complexity}.
As a final note for this section, it is typical to reduce a Latin square by assuming the first row \emph{and} the first column are the identity permutation. The number of such reduced Latin squares is then $|LS_n|/(n! (n-1)!)$, and it is often easier to work with this reduced set.
This reduction could also be exploited for PDC.
It is valid to replace Lines~1--3 in Algorithm~\ref{Sudoku:n} with the following:
1: Let $R_1 = (1,\ldots,n)$. \\
2: {\bf for} $i=2,\ldots, n-3$ \\
3: Generate $R_i$ uniformly from the set of fixed--point free permutations starting with $i$. \\
This restriction forces the first $n-3$ entries of column 1 to be the identity permutation $(1,2,\ldots,n-3)$ (and one could easily force the remaining entries in the column to be $(n-2, n-1, n)$ ).
However, for $2 \leq i \leq n-3$, the set of all fixed--point free permutations of $n$ starting with an $i$ is given by the OEIS sequence $A000255$~\cite{OEIS}.
In~\cite[Page~373]{Flajolet}, it is shown that the cardinality of this set is asymptotically $n!/e$, which is asymptotically the same as the number of fixed--point free permutations of $n$ that we have used in our current analysis.
\section{Other sampling algorithms}\label{sect alt}
\label{section:other}
\subsection{The $R \times C$ generalization of a Sudoku matrix}
\label{Sudoku:RC}
There is a generalization to a Sudoku matrix where the block constraint is modified to be a block with dimensions $R \times C$.
The set of $(R,C)$-Sudoku matrices is then defined as the set of $R\,C \times R\, C$ matrices tiled with the $R \times C$ blocks, where each row, column, and $R\times C$ block is a permutation of $\{1, 2, \ldots, R\,C\}$.
Sudoku matrices are the special case with $R=C=3$.
In fact, Algorithm~\ref{Sudoku:n} is worded so that it can be applied to $(R,C)$-Sudoku matrices \emph{varbatim} using $n=R\,C$: one simply interprets Line~\ref{line:d:LS} and Line~\ref{line:complete} analogously.
Even the denominator of the rejection probability, i.e., $\lfloor 6^{n/3} 2^{n/2}\rfloor$, can be used as an upper bound on the number of completable $(R,C)$-Sudoku matrices.
Of course, there is some efficiency lost in the algorithm by using this generic upper bound, however, as was pointed out earlier, it does not affect the unbiased nature of the algorithm, only the efficiency, and should more efficient bounds become available, one could (and should!) substitute in those bounds instead.
\subsection{$S$-permutation matrices}
\label{S:permutation}
A recent approach to random sampling of Latin squares and Sudoku matrices is via permutation matrices, see~\cite{Dahl}; see also \cite{FontanaFractions, Fontana, YordzhevNumber}.
A permutation matrix of order~$N$ is an $N \times N$ matrix with a exactly one 1 appearing in each row and column, and the rest of the entries are 0.
An \emph{$S$-permutation matrix of order $N = n^2$} is an $N \times N$ permutation matrix with the additional constraint that only one~$1$ appears in each of the $n \times n$ sub-blocks indicated by the generalization in Section~\ref{Sudoku:RC} for $(n,n)$-Sudoku matrices.
Central to this idea is the fact that each $(n,n)$-Sudoku matrices $A$ can be written uniquely as
\begin{equation}\label{Sudoku:decomposition} A = 1\cdot P_1 + 2\cdot P_2 + \cdots + N\cdot P_N, \end{equation}
where $P_1, \ldots, P_N$ are each permutation matrices of order~$N$, with the additional property of being \emph{disjoint}, i.e., the supports of each matrix are pairwise disjoint.
It was shown in~\cite[Proposition~1]{Dahl} that the number of $S$-permutation matrices is precisely $n!^{2n}$.
In~\cite[Theorem~1 and Theorem~2]{YordzhevNumber}, an exact formula is given using inclusion-exclusion for the number of \emph{pairs} of disjoint $S$-permutation matrices.
Permutation matrices were exploited in~\cite{Fontana} for the random sampling of Latin squares of order~$N$, using the same decomposition in~\eqref{Sudoku:decomposition} but using just permutation matrices of order $N$ without the additional block constraint.
The sampling algorithm in~\cite[Section~3]{Fontana} generates a certain undirected graph whose largest cliques correspond to permutation matrices in~\eqref{Sudoku:decomposition} used to construct a Latin square.
It is the creation of this undirected graph along with the finding of all cliques which currently dominates the computation.
For Latin squares of order $n$, the procedure was shown to be an effective method of uniform sampling for $n \leq 7$.
\subsection{Importance sampling}
\label{section:importance}
In~\cite{DeSalvoSudoku}, a backtracking algorithm for Sudoku matrices is presented as follows:
\begin{itemize}
\item Generate the first row as a random permutation of $\{1,\ldots, 9\}$.
\item Simulate each subsequent row one at a time, and left to right each entry at a time, with $R_{i,j}$ denoting the set of elements in row $i$ and column $j$ which can be placed without violating the Sudoku conditions, by selecting an element from $R_{i,j}$ uniformly at random at each stage.
\item If for any $2 \leq i \leq 9$ and $1 \leq j \leq 9$ we have $R_{i,j} = \emptyset$, then delete the current row and the preceding row, and continue.
\end{itemize}
This algorithm was used to generate a sample size of $10^8$ in~\cite{DeSalvoSudoku}.
While no quantitative bounds on the bias were proved, the algorithm completed quickly, always producing a valid Sudoku matrix.
In~\cite{RidderIS}, a similar algorithm is championed, except if the algorithm encounters $R_{i,j} = \emptyset$ for any $i$ and $j$, the algorithm halts.
The advantage this algorithm has is that one can more easily keep track of the importance sampling weight, and hence establish an approximation for the number of Sudoku matrices as follows.
If we assume that each row is a uniform permutation of the set $\{1, \ldots, 9\}$, then each entry in column $j$ of a Sudoku matrix can be one of $10-j$ possible values, depending on which of the first $j$ elements were placed first.
Whereas in the algorithm we select an element uniformly from the set $R_{i,j}$ of size $|R_{i,j}| \leq 10-j$, we correct for this via the likelihood ratio
\[ \ell_{i,j} = \frac{|R_{i,j}|}{10-j}, \qquad R_{i,j} \neq 0,\]
with $\ell_{i,j} = 0$ if $R_{i,j} = \emptyset$.
The importance sampling weight is then the product of these ratios, i.e.,
\[ \prod_{i=2}^9 \prod_{j=1}^9 \ell_{i,j}. \]
This was used to estimate the approximate number of Sudoku matrices, obtaining $6.662 \cdot 10^{21}$ as a point estimate.
This algorithm can also be adapted to larger numerical tables, e.g., $(n,n)$-Sudoku matrices, but as the author points out, the probability of generating a feasible matrix, i.e., the probability that the algorithm does not halt before completion, becomes increasingly small.
\subsection{Markov chain approaches}
In~\cite{JacobsonMatthews}, two Markov chains on the set of Latin squares of order~$n$ are presented which involve rather intricate sets of transformations to transition from one state to the next, always yielding a valid Latin square of order~$n$.
It is proved in~\cite[Theorem~7]{JacobsonMatthews} that the stationary distribution is indeed the uniform distribution, and that transitioning from one state to another requires $O(n)$ arithmetic operations.
However, as indicated in~\cite[Section~6]{JacobsonMatthews}, there is no proof that the chain is rapidly mixing.
A Markov chain for $(n,n)$-Sudoku matrices was recently introduced in~\cite{SudokuMC}, via a connection with the set of binary contingency tables (the set of matrices with entries in $\{0,1\}$ with row sums and column sums specified), and then by finding a Markov basis, see~\cite{DiaconisSturmfels}.
The Markov basis ensures that all states are reachable, however, the technique is admittedly not feasible for the standard $(3,3)$-Sudoku matrices, and instead it is explored using $(2,2)$-Sudoku matrices.
While an intriguing and unique approach to random sampling, there is also no proof that the chain is rapidly mixing.
\section{Final Remarks}\label{final:remarks}
Our motivation for these algorithms is to verify a claim from \cite{DeSalvoSudoku}, in which the authors calculated and compared the Shannon entropy of a random sample of Sudoku matrices and Latin squares of order $9$ using a backtracking algorithm, see Section~\ref{section:importance}.
This algorithm is fast, though not necessarily unbiased, and allowed the authors to generate a sample of size $10^8$ that was used in the estimation of Shannon's entropy.
For Sudoku matrices, the entropy was estimated as $1.73312 \pm 0.000173$, and for Latin squares of order $9$ the entropy was estimated as $1.73544\pm 0.0001735$.
Assuming the bias in the backtracking algorithm is relatively small, this calculation suggests that a statistical error of at most $10^{-3}$ is required in order to definitively distinguish a difference in entropies between the set of Sudoku matrices and the set of Latin squares of order $9$, and so we anticipate needing an \emph{unbiased} sample of size at least $10^6$ to effectively distinguish between the entropies of these two sets of matrices.
While Algorithm~\ref{Sudoku} and Algorithm~\ref{Sudoku:n} are provably unbiased, they each took approximately 24 hours to generate a sample of size 1000 on a personal computer\footnote{1.8 GHz Intel Core i7.} (using $n=9$ in Algorithm~\ref{Sudoku:n}).
Using this i.i.d. sample of size 1000, an unbiased estimate for the entropy of Sudoku matrices is $1.73356 \pm 0.0383708$, and for Latin squares of order 9 our estimate for the entropy is $1.73335 \pm 0.0389771$, which is consistent with the backtracking algorithm.
We have supplied C++ source code used to generate the samples and posted it at \\
\url{https://github.com/stephendesalvo}, along with the files containing the random samples of size 1000, and scripts to load in the matrices into an $n\times n \times 1000$ array in {\tt Matlab} and a $1000 \times n \times n$ array in {\tt Mathematica}.
\section{Acknowledgements}
The author would like to acknowledge helpful conversations with James Zhao and Edo Liberty, and also helpful comments from an anonymous reviewer.
\bibliographystyle{plain}
|
\section{Introduction}
Let $S = \Bbbk[x_1,\dots,x_n] = \bigoplus_{i \geq 0} [S]_i$, where $\Bbbk$ is an algebraically closed field of characteristic zero. In 1954 Mammana \cite{mammana} introduced the variety of reducible plane curves. He was seeking to generalize work of C. Segre \cite{segre} (for conics), N. Spampinato \cite{spampianto} (for plane cubics) and G. Bordiga \cite{bordiga} (for plane quartics) as well as other works mentioned in his ample bibliography.
Here we generalize the idea further. Let $\lambda = [d_1,\dots,d_r]$ be a partition of $d = \sum_{i=1}^r d_i$, which we will write as $\lambda \vdash d$, where $d_1 \ge d_2 \ge \cdots \ge d_r \ge 1$ and $r \ge 2$.
Consider the variety $\mathbb X_{n-1,\lambda} \subset \mathbb P([S]_d) = \mathbb P^{N-1}$ of $\lambda$-reducible forms, where $N = \binom{d+n-1}{n-1}$. That is,
\[
\mathbb X_{n-1,\lambda} = \{ [F] \in \mathbb P^{N-1} \ | \ F = F_1 \cdots F_r \hbox{ for some }0\neq F_i \in [S]_{d_i} \}.
\]
The object of this paper is to study the dimension of the $(\ell-1)$-secant variety of $\mathbb X_{n-1,\lambda}$, which we denote by $\sigma_\ell (\mathbb X_{n-1,\lambda})$.
So $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is the closure of the union of the linear spans of all subsets of $\ell$ distinct points of $\mathbb X_{n-1,\lambda}$.
We will give a new approach to this problem.
We have
\[
\dim \mathbb X_{n-1, \lambda} = \sum_{i=1}^r \binom{d_i + n-1}{n-1} - r .
\]
Since all forms in two variables are products of linear forms, we always assume $n \ge 3$, $d \ge r \ge 2$, and $\ell \ge 2$.
We can (and will) view a general point of $\mathbb X_{n-1,\lambda}$ as the product of general forms in $S$ of degrees $d_1,\dots,d_r$ respectively.
Since, as it is easy to see, no hyperplane contains $\mathbb X_{n-1,\lambda}$,
$\ell\leq N$ general points of $\mathbb X_{n-1,\lambda}$ span a linear space of dimension $\ell -1$ (i.e., an
$(\ell -1)$-secant plane), so
by a simple parameter count we have $\dim \sigma_\ell (\mathbb X_{n-1,\lambda}) \leq \ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1$. But
it is possible that $\sigma_\ell (\mathbb X_{n-1,\lambda})$ fills its ambient space $\mathbb P^{N-1}$; this
clearly happens, for instance, if $\ell\geq N$.
We combine the two possibilities to obtain an upper bound for the actual dimension of
$\sigma_\ell(\mathbb X_{n-1\lambda})$, typically referred to as the \emph{expected dimension}:
\[
\hbox{exp.dim } \sigma_{\ell} (\mathbb X_{n-1,\lambda}) = \min \left \{ \binom{d+n-1}{n-1} -1, \; \ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1 \right \} .
\]
The {\em defect}, $\delta_\ell$, is the expected dimension minus the actual dimension. When this is positive, we say that $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is {\em defective}. An important part of our work will be to identify when $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is defective and to compute the defect.
Secant and join varieties of the Veronese, Segre and Grassmann varieties have been extensively studied. The recent intense activity in studying these varieties has certainly benefited from the numerous fascinating applications in Communication Theory, Complexity Theory and Algebraic Statistics as well as from the connections to classical problems in Projective Geometry and Commutative Algebra. (For a partial view of these applications consider the following references and their bibliographies: \cite{AOP}, \cite{AH:1}, \cite{BCS}, \cite{CGG:x}, \cite{CGG:6}, \cite{CS}, \cite{landsberg}, \cite{ravi}, \cite{SS}, \cite{vermeire}.)
However, very little is known about the secant varieties of the varieties of reducible hypersurfaces. Here also there are useful applications in the study of vector bundles on surfaces and connections to the classical Noether-Severi-Lefschetz Theorem for general hypersurfaces in projective space (see \cite{CCG:1}
\cite{CF}, \cite{Patn}).
The first significant results about the secant varieties of $\lambda$-reducible forms were obtained by Arrondo and Bernardi in \cite{AB} for the special partition $\lambda =[1, \dots ,1]$ (they refer to $\mathbb X_{n-1,\lambda}$ for this particular $\lambda$ as the variety of {\it split} or
{\it completely decomposable}
forms). They find the dimensions of secant varieties in this case for a very restricted, but infinite, family of examples. This was followed by work of Shin \cite{S:1} who found the dimension of the secant line variety to the varieties of split plane curves of every degree.
This latter result was further generalized by Abo \cite{A}, again for split curves, to a determination of the dimensions of all the higher secant varieties. Abo also dealt with some cases of split surfaces in $\P^4$ and split cubic hypersurfaces in $\P^n$, for any $n$.
In all the cases considered, the secant varieties have the expected dimension. Arrondo and Bernardi have speculated that the secant varieties for split hypersurfaces always have the expected dimension. We verify this
for $\sigma_{\ell}(\mathbb X_{n-1,\lambda}$) ($\lambda = [1,\dots , 1]\vdash d$) as long as $2\ell \leq n$. We also note that their speculation is a special case of our Conjecture \ref{conj:defectivity conj} (a).
The parameters for this work are $n\geq 3$, $\ell\geq2$ and any partition $\lambda = [d_1,\dots,d_r]$ with $r\geq2$ positive parts. All previous results assume $d_1=1$ (i.e., the split variety case \cite{A, AB, S:1}) or $n=3$ \cite{CGGS}, or $r=2$ \cite{BCGI, CCG:1}. We extend all of this previous work significantly. See, for example, Theorem \ref{thm: main thm - intro}, as an immediate consequence of which we obtain complete answers in many new cases, including each of the following:
for all $n\gg0$ (in fact $n\geq 2\ell$), fixing any $\ell$ and an arbitrary partition $\lambda$;
for all $\ell\gg0$ (in fact $\ell\geq \binom{s+n-1}{n-1}$), fixing any $n$ and $d_2\ge\cdots\ge d_r \ge 1$, where $s=d_2+\cdots+d_r$; and
for all $d_1\gg0$ (in fact $d_1\geq (s-1) (n-1)$), fixing any $n$ and $d_2\ge\cdots\ge d_r \ge 1$, and any $\ell$ {\em not} in the interval $(\frac{n}{2},n)$.
We also propose a conjecture (see Conjecture \ref{conj:main conj}), which, if true, gives a complete answer in all remaining cases and which has led us to many of our results.
All approaches to finding the dimension of the secant varieties to a given variety $\mathbb X \subset \P^{n-1}$ begin with Terracini's Lemma, including ours.
These all require a good understanding of the tangent space to $\mathbb X$ at a general point. Successful applications of Terracini's Lemma begin by identifying this tangent space as a graded piece of some relatively nice ideal (which we will call the {\it tangent space ideal}). To apply Terracini's Lemma, one then needs a way to deal with the sum of tangent space ideals at a finite set of general points of $\mathbb X$. For the Veronese, Segre and Grassmann varieties, the tangent space ideals were all artinian ideals in the appropriate polynomial ring. The standard method for dealing with the sum of such ideals (which, per se, have no geometric content) is to pass, using Macaulay duality, to a consideration of the intersection of the {\it perps} of the tangent space ideals (see. e.g. the discussion in \cite{Ger}). In the classical cases considered above, one obtains a union of special 1-dimensional ideals corresponding to zero dimensional projective schemes. One then uses geometric methods to get information about the union of the schemes defined by the perps of the tangent space ideals.
This clever use of Macaulay duality had its first notable success with the work of Alexander and Hirschowitz, who completed (after almost one hundred years) the solution of Waring's Problem for general forms (see \cite{AH:1}). Other work in this direction for these classical varieties can be found in \cite{CCG:1}, \cite{CGG:6}, \cite{CGG:x},
\cite{AOP}, \cite{LP}.
In the case of the varieties of reducible hypersurfaces, the method described above no longer works. In this case, the tangent space ideals already define very nice schemes of dimension $\geq 0$, namely arithmetically Cohen-Macaulay codimension 2 subschemes of $\mathbb P^n$, and their Macaulay duals are artinian! Thus, one is forced to deal with the sum of the tangent space ideals, i.e., with the {\it intersection} of the codimension 2 schemes defined by the tangent space ideals at general points.
This is the novelty of our approach: to deal with this intersection we use a version of the so-called diagonal trick from intersection theory, and we show how the so-called Lefschetz properties come in to play in order to study improper intersections.
As we describe in detail, finding this dimension amounts to viewing the intersection of the aforementioned codimension two subschemes in $\mathbb P^n$ as the result of consecutive hyperplane sections of their join in $\mathbb P^{n\ell -1}$, where the hyperplanes cut out the diagonal.
The dimension of the secant variety can then be read off from the Hilbert function in degree $d$ of the coordinate ring of the intersection of such schemes, although ``intersection" must be suitably interpreted in the artinian situation. Algebraically, we are interested in the Hilbert function in degree $d$ of $S/(I_{P_1} + \dots + I_{P_\ell})$ (the $I_{P_i}$ being the tangent space ideals at general points; see Proposition \ref{TangIdeal}), but the geometric notions from intersection theory and hyperplane sections guide our approach.
A key to our method is the observation that we can replace the hyperplanes cutting out the diagonal by truly general hyperplanes. This allows us to compute the dimension of the secant variety $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ in the case where the subschemes meet properly, which occurs precisely when $2 \ell \leq n$.
In the case of an improper intersection of the tangent spaces, i.e. $2 \ell > n$, we conjecture that the general hyperplanes induce multiplication maps that all have maximal rank. For a single hyperplane such behavior has been dubbed the Weak Lefschetz Property in \cite{HMNW}. Assuming this conjectured property of the hyperplane sections, we obtain a formula for the dimension of the secant variety, which is surprisingly uniform.
This single formula proposes the dimension for {\em any} choice of $n, \ell$ and $\lambda$. We will establish it in some cases. It is a conjecture in the rest, but we know of no cases of known results with which it does not agree.
To be more precise, for $0 \le j \le d$ and every given $\ell, n$ and partition $\lambda$ of $d$, we define integers $a_j (\ell, n , \lambda)$ by an explicit formula (see Definition \ref{def:integers a-j}).
Our formula for the dimension of the secant varieties is the following, which we state as a conjecture so that it can be applied for all $n, \ell$ and $\lambda$,
in addition to the many cases which we prove below.
\begin{conj}
\label{conj:main conj}
Let $\lambda = [d_1,\ldots,d_r]$ be a partition of $d$ with $r \ge 2$ parts. Then:
\begin{itemize}
\item[(a)] The secant variety $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ fills its ambient space if and only if there is some integer $j$ with $s = d_2 + \cdots + d_r \le j \le d$ such that $a_j (\ell, n , \lambda) \le 0$.
\item[(b)] If $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ does not fill its ambient space, then it has dimension
\begin{align*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = & \ \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \\[.5ex]
& \hspace*{0.3cm} {\displaystyle - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} } \\[.5ex]
& \hspace*{0.3cm} {\displaystyle - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} }.
\end{align*}
\end{itemize}
\end{conj}
\noindent The fact that this conjecture is a consequence of the indicated Lefschetz property is shown in Theorem \ref{thm: gen dim formula}.
Throughout this paper we use the convention that a binomial coefficient $\binom{a}{b}$ is zero if $a < 0$. Thus, for example, the last and the penultimate term in the above dimension formula are zero if $r \ge 3$. (A heuristic approach to the formula in Conjecture \ref{conj:main conj}(b) can be found in Remark \ref{sandro rmk}.)
Although stated differently, previous results imply that Conjecture \ref{conj:main conj} is true if $n = 3$ and $\lambda = [1,\ldots,1]$ (see \cite{A}), or if $n=3$, $\lambda$ is arbitrary, and $\ell = 2$ (see \cite{CGGS}). Here, we prove this conjecture in further cases, most of which are summarized in the following theorem. Note that part (b)(i) of the following result was proved in \cite[Theorem 5.1]{CCG:1} using different language.
\begin{thm}\label{thm: main thm - intro}
Let $\lambda=[d_1,\ldots,d_r]$ be a partition of $d=d_1+s$ into $r\geq 2$ parts,
where $s=d_2+\cdots+d_r$.
Then Conjecture \ref{conj:main conj} is true in the following cases:
\begin{itemize}
\item[(a)] $\ell \le \frac{n}{2}$ or $\ell\geq\binom{s+n-1}{n-1}$;
\item[(b)] $r = 2$ and either
\begin{itemize}
\item[(i)] $\ell \le \frac{n+1}{2}$, or
\item[(ii)] $\lambda = [d-1, 1]$, or
\item[(iii)] $n=3$; and
\end{itemize}
\item[(c)]
$r\geq3$ and $n\leq \ell\leq 1+\frac{d_1+n-1}{s}$.
\end{itemize}
\end{thm}
We prove Theorem \ref{thm: main thm - intro}(a) in Remark \ref{rem:comp with {CGGS}}(ii) and Proposition \ref{prop:non-defective}(c).
See Theorem \ref{thm:dim when froeberg} for parts (b)(i, iii), Theorem \ref{thm:case d-1, 1} for part (b)(ii), and
Corollary \ref{Thm1.2c} for part (c).
We also show that if $\ell = \frac{n+1}{2}$ (Proposition \ref{prop:upper bound})
or $r = 2$ (Proposition \ref{prop:Froeberg gives main conj}), then the number predicted by Conjecture \ref{conj:main conj} is at least an upper bound for $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})$.
Notice that the dimension formula in Conjecture \ref{conj:main conj} involves a series of comparisons, checking whether $a_j (\ell, n , \lambda) > 0$ for all $j = s, s+1,\ldots,d$.
Accordingly, it is worthwhile to point out more explicitly some of the consequences it suggests. Again, the following is stated as a conjecture even though in the different settings of the above theorem these results are proven.
\begin{conj}\label{conj:defectivity conj}
Let $\lambda = [d_1,\ldots,d_r]$ be a partition of $d$ with $r \ge 2$ parts. Then:
\begin{itemize}
\item[(a)] If $d_1 < d_2 + \cdots + d_r$ (and thus $r \ge 3$), then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective.
\item[(b)] If $d_1 \ge d_2 + \cdots + d_r$, then the secant variety $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is defective if and only if it does not fill its ambient space.
\end{itemize}
\end{conj}
This conjecture
highlights the role of the ``partition dividing hyperplane'' $d_1 = d_2 + \cdots + d_r$
in the space of partitions with $r$ positive parts,
as introduced in \cite{CGGS} and discussed here in Remark \ref{PDH}.
Of course, Conjecture \ref{conj:defectivity conj} might be true even if the more
specific formulation given in Conjecture \ref{conj:main conj} is not.
Moreover, while Conjecture \ref{conj:main conj} implies most of Conjecture \ref{conj:defectivity conj},
it is not yet clear that Conjecture \ref{conj:main conj} implies all of Conjecture \ref{conj:defectivity conj};
see Proposition \ref{prop:non-defective} and Remark \ref{NumericalTestingRemark}.
In particular, notice that Conjecture \ref{conj:defectivity conj}(a) is an immediate consequence of
Conjecture \ref{conj:main conj}, but
we can show that Conjecture \ref{conj:defectivity conj}(b) follows from
Conjecture \ref{conj:main conj} only in certain cases (see Proposition \ref{prop:non-defective}).
Our results on defectiveness show unconditionally:
\begin{thm}\label{thm:defectivity - intro}
Let $\lambda = [d_1,\ldots,d_r]$ be a partition of $d$ with $r \ge 2$ parts and let $s=d_2+\cdots+d_r$. Then:
\begin{itemize}
\item[(a)] If $d_1 < s$ (and hence $r\geq3$) and $2\ell \le n$, then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective.
\item[(b)] If $d_1 \ge s$ and $2\ell \le n$,
then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is defective if and only if it does not fill its ambient space.
\item[(c)] If $\ell\geq n$ and $d_1\geq(n-1)(s-1)$, then $\sigma_\ell(\mathbb X_{n-1,\lambda})$ always fills its ambient space,
while if $2\ell \le n$, then $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills its ambient space if and only if one of the following conditions is satisfied:
\begin{itemize}
\item[(i)] $n = 4$, $\ell = 2$, and $\lambda \in \{[1,1], [2,1], [1,1,1]\}$ \quad or
\item[(ii)] $n = 2 \ell \ge 6$ and $\lambda = [1,1]$.
\end{itemize}
\end{itemize}
\end{thm}
\noindent We give the proof near the end of Section \ref{sec:improper inters}.
We use our results in the case $r=2$ to study the {\em variety of reducible forms of degree $d$ in $n$ variables}
\[
\mathbb X_{n-1, d} = \bigcup_{k=1}^{\lfloor \frac{d}{2} \rfloor} \mathbb X_{n-1, [d-k, k]}.
\]
We show that
\[
\dim \mathbb X_{n-1, d} = \dim \mathbb X_{n-1, [d-1, 1]}
\]
and that all other irreducible components of $\mathbb X_{n-1, d}$ have dimension that is smaller than the dimension of $\mathbb X_{n-1, [d-1, 1]}$. Thus, one can hope that $\mathbb X_{n-1, [d-1, 1]}$ determines the dimension of the secant variety of $\mathbb X_{n-1, d}$. Indeed, we establish:
\begin{thm}\label{thm:intro red forms}
If $2\ell \le n$, then
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, d}) = \dim \sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]}).
\]
Moreover, $\sigma_{\ell} (\mathbb X_{n-1, d})$ is defective if and only it it does not fill its ambient space.
\end{thm}
\noindent We prove this result in Section \ref{sec:variety red forms} as a consequence of Theorem \ref{thm:red forms, small ell}.
Note that the dimension of $\sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]})$ is known for all $\ell, n$ and $d$ (see
Theorem \ref{thm:case d-1, 1} or \cite[Proposition 4.4]{BCGI}). Thus, we know exactly when $\sigma_{\ell} (\mathbb X_{n-1, [d-1,1]})$ is defective (see Theorem \ref{thm:red forms, small ell}). We suspect that $\dim \sigma_{\ell} (\mathbb X_{n-1, d}) = \dim \sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]})$ is true for all $n$ and $\ell$.
In Section \ref{sec:inters, diag} we recall the basic facts about the variety of reducible hypersurfaces and how Terracini's lemma is applied. We consider the coordinate ring of the join, which is arithmetically Cohen-Macaulay of codimension $2 \ell$ in $\mathbb P^{n\ell -1}$, with known minimal free resolution and Hilbert function. We discuss how the algebra that determines the dimension of $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is obtained by successive hyperplane sections (i.e. reduction by general linear forms). As long as these linear forms are regular elements, the intersection is proper and in Section \ref{sec:proper inters} we give formulas for the dimension and defect.
(As an aside, we point out that the varieties $\mathbb X_{n-1,\lambda}$ are not generally
arithmetically Cohen-Macaulay. For example, by direct computation
they are not for $n=3$ and $\lambda$ either $[1,1,1]$ or $[1,2]$, but we do not know about their secant varieties.)
In Section \ref{sec:secant lines} we summarize our results in the case $\ell = 2$, i.e. for the secant line variety. Proper intersection corresponds to $n \geq 4$. For the remaining case, $n=3$, we recall the results of \cite{CGGS}. This gives us a bridge from the proper intersections to the improper intersections and gives the idea of how the Lefschetz property is applied in general.
In Section \ref{sec:improper inters} we work out, in general, the connection between the computation of the dimension for arbitrarily large $\ell$, corresponding to improper intersections, and the study of Lefschetz Properties.
Indeed, based on experiments, we conjecture that the coordinate ring of a certain join variety has enough Lefschetz elements if $2\ell > n$ (see Conjecture \ref{conj:WLP}). If this conjecture is true, then Conjecture \ref{conj:main conj} follows (see Theorem \ref{thm: gen dim formula}). However, Conjecture \ref{conj:main conj} is weaker than the conjecture on the existence
of enough Lefschetz elements.
In Section \ref{sec:r is 2} we focus on the case $r =2$. We show that Conjecture \ref{conj:main conj} is a consequence of Fr\"oberg's Conjecture on the Hilbert function of ideals generated by generic forms.
In Section~\ref{sec:variety red forms} we study the variety of reducible forms.
We conclude in Section \ref{AOPstuff} by showing how our results imply cases of conjectures raised in \cite{AOP} about defectivity of Segre Varieties.
\section{Intersections and the Dimension of Secant Varieties}
\label{sec:inters, diag}
After recalling some background and introducing our notation, we lay out our method for computing the desired dimension of a secant variety. It is inspired by a technique from intersection theory. The method will be applied in later sections, where we treat the case of proper and improper intersections separately and carry out the needed computations.
\begin{nota}
Let $S = \Bbbk[x_1,\dots,x_n] = \bigoplus_{i \geq 0} [S]_i$ be the standard graded polynomial ring, where $\Bbbk$ is an algebraically closed field of characteristic zero. Let $\lambda = [d_1,d_2,\dots,d_r]$ be a {\it partition of $d$ into $r \ge 2$ parts}, i.e., $\lambda \vdash d$, $d_i \in \mathbb N$, $d_1 \geq d_2 \geq \dots \geq d_r >0$ and $\sum_{i=1}^r d_i = d$.
\medskip\medskip
If we set $N = \binom{d+n-1}{n-1}$ then the {\it variety of reducible forms in $[S]_d$ of type $\lambda$} (or {\it the variety of reducible hypersurfaces in $\P^{n-1}$ of type $\lambda$}) is, as noted above:
$$
\mathbb X_{n-1,\lambda}:= \{ [F] \in \P([S]_d) = \P^{N-1} \mid F = F_1 \cdots F_r,\ \ \deg F_i = d_i \} \ .
$$
The map $([F_1],\ldots,[F_r])\mapsto [F]=[F_1\cdots F_r]$ induces a finite morphism
\begin{equation}
\P([S]_{d_1}) \times \cdots \times \P([S]_{d_r}) \longrightarrow \mathbb X_{n-1,\lambda},
\end{equation}\label{SegreToRedFormsMorphism}
and so we have
\begin{equation}\label{dim X}
\dim \mathbb X_{n-1,\lambda} = \Big[ \sum_{i=1}^r \binom{d_i + n-1}{n-1} \Big] - r .
\end{equation}
As discussed above, given a positive integer $\ell\leq N$, the
variety $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is the subvariety of $\P^{N-1}$ consisting of the
closure of the union of secant $\P^{\ell-1}$'s to $\mathbb X_{n-1,\lambda}$; for $\ell\geq N$,
$\sigma_\ell (\mathbb X_{n-1,\lambda})$ is simply $\P^{N-1}$.
Following the classical terminology, $\sigma_2(\mathbb X_{n-1,\lambda})$ is called the {\it secant line variety}
of $\mathbb X_{n-1,\lambda}$ and $\sigma_3(\mathbb X_{n-1,\lambda})$ the {\it secant plane variety} of $\mathbb X_{n-1,\lambda}$.
\end{nota}
Our main interest in this paper is the calculation of the dimensions of the varieties $\sigma_{\ell}(\mathbb X_{n-1,\lambda})$.
\begin{rem} Notice that for $n = 2$ the question is a triviality since {\em every} hypersurface of degree $d$ in $\P^1$ is reducible of type $\lambda = [1,1,\dots,1] \vdash d$. Thus, we assume throughout $n \ge 3$.
\end{rem}
The fundamental tool for the calculation of dimensions of secant varieties is the following celebrated result \cite{T:1}.
\begin{prop}[Terracini's Lemma] \label{Terracini} Let $P_1, \ldots , P_\ell$ be general points on $\mathbb X_{n-1,\lambda}$ and let $T_{P_i}$ be the (projectivized) tangent space to $\mathbb X_{n-1,\lambda}$ at the point $P_i$.
The dimension of $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is the dimension of the linear span of $\bigcup_{i=1}^\ell T_{P_i}$.
\end{prop}
As mentioned above, we have the following definitions.
\begin{defn} \label{exp dim}
The
{\it{expected dimension}} of $\sigma_\ell(\mathbb X_{n-1,\lambda})$, written $\expd (\sigma_\ell (\mathbb X_{n-1,\lambda}))$, is
\[
\min\{ N-1, \ell \cdot \dim\mathbb X_{n-1, \ \lambda} + (\ell-1) \} .
\]
The {\em defect} of $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is
\[
\delta_\ell = \expd (\sigma_\ell (\mathbb X_{n-1,\lambda})) - \dim(\sigma_\ell(\mathbb X_{n-1,\lambda})) \geq 0.
\]
We say that $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is {\em defective} if $\delta_\ell > 0$.
\end{defn}
\begin{rem}\label{defective}
Thus $\sigma_ \ell (\mathbb X_{n-1, \lambda})$ is defective if and only if $\dim \sigma_\ell (\mathbb X_{n-1, \lambda}) < N-1$ and $\dim \sigma_\ell (\mathbb X_{n-1, \lambda}) < \ell (\dim(\mathbb X_{n-1, \lambda})) + (\ell - 1)$. We will see that in some cases it will be easier to write an expression for $\delta_\ell$ than it will be to show that it is positive.
\end{rem}
\medskip
Clearly, to be able to effectively use Terracini's Lemma it is essential to have a good understanding of the tangent spaces to $\mathbb X_{n-1,\lambda}$ at general (hence smooth) points. We do that now.
Let $P = [F] \in \mathbb X_{n-1,\lambda}$ be a general point. Then $F = F_1\cdots F_r$ where the $F_i$ are irreducible forms of degree $d_i $ in $S$. Let $G_i = F/F_i$, and so $\deg G_i = d-d_i$. Consider the ideal $I_P \subset S$, where $I_P = (G_1, \ldots , G_r)$.
\begin{prop}\label{TangIdeal}
$$
T_P = \P\Big( [I_P]_d \Big)
$$
\end{prop}
\begin{proof}This Proposition is well known and proofs can be found in several places ( see e.g. \cite{CCG:1} Prop. 3.2).
\end{proof}
We refer to the variety in $\mathbb{P}^{n-1}$
defined by $I_P$ as the {\em variety determining the (general) tangent space to $\mathbb X_{n-1, \lambda}$}.
As an immediate corollary of Propositions \ref{Terracini} and \ref{TangIdeal} we have the following:
\begin{cor}\label{dim sigma ell}
Let $P_1, \ldots , P_\ell$ be $\ell$ general points of $\mathbb X_{n-1,\lambda}$ and $I=I_{P_1} + \cdots + I_{P_\ell}$. Then
$$
\dim (\sigma_\ell(\mathbb X_{n-1,\lambda} )) = \dim_{\Bbbk} \left [ I \right ]_d -1 \ .
$$
\end{cor}
This means that $\dim (\sigma_\ell(\mathbb X_{n-1,\lambda} ))$ is determined by the Hilbert function in degree $d$ of the ring $S/(I_{P_1}+ \cdots +I_{P_\ell})$, which, when $\sigma_\ell$ does not fill its ambient space, is the coordinate ring of the intersection of $\ell$ varieties, in $\P^{n-1}$, determining tangent spaces to $\mathbb X_{n-1, \lambda}$.
\begin{rem}\label{N-1 less}
Let $P_1, \ldots , P_\ell$ be general points on $\mathbb X_{n-1,\lambda}$. Suppose that
$$
N-1\leq \ell \dim \mathbb X_{n-1,\lambda} + (\ell -1);
$$
i.e., the expected dimension of $\mathbb X_{n-1,\lambda}$ is $N-1$. Then
$$
\sigma_\ell\big( \mathbb X_{n-1,\lambda} \big) \hbox{ is defective } \Leftrightarrow \dim_{\Bbbk} [S/(I_{P_1} + \cdots + I_{P_\ell})]_d > 0.
$$
In this case, $\delta_\ell = \dim_{\Bbbk} [S/(I_{P_1} + \cdots + I_{P_\ell})]_d$.
\end{rem}
\medskip\medskip Now that we have seen the ideal that enters into the use of Terracini's Lemma, it remains to give a nicer description of the ideal $I_P$ determining the tangent space at the point $P$.
\begin{prop}\label{nicer}
Let $P$ be a general point of $\mathbb X_{n-1,\lambda}$, $P = [F_1\cdots F_r]$ where $\deg F_i = d_i$. Put $F = F_1\cdots F_r$. Then we have
$$
I_P = (F/F_1, \ldots ,F/F_r) = \bigcap _{1 \leq i < j \leq r} (F_i, F_j) .
$$
\end{prop}
\begin{proof} The first equality is given in Proposition \ref{TangIdeal}.
The second equality is well known (see for example \cite{PS:1}, Thm. 2.3).
\end{proof}
Thus the ideal $I_P$, for a general point $P \in \mathbb X_{n-1,\lambda}$, is of codimension 2 in $S$ and is a finite intersection of interrelated codimension 2 complete intersection ideals. Such ideals (and their generalization to the situation where the complete intersection ideals have higher codimension) have been studied in several papers for many different reasons (see e.g. \cite{GHM:1}, \cite{GHMN:1}, \cite{GMS:1} and \cite{CGVT:1}).
We now derive the graded minimal free resolution of the ideal $I_P$.
\begin{lem}\label{lem3.1}
Let $R=\Bbbk[Y_1,\dots,Y_r]$, $M=Y_1\cdots Y_r$, $M_i=M/Y_i$, where $r \ge 2$. If $I=(M_1,\dots,M_r)$ then $I ={\ds\bigcap_{1\le i<j\le r}}(Y_i,Y_j)$ and the minimal graded free resolution of $I$ is
$$
0 \to R^{r-1}(-r) \overset{A}{\to} R^r(-(r-1)) \to R \to R/I \to 0.
$$
\end{lem}
\begin{proof} Consider the matrix $A$, defined by
$$
A^t=\begin{bmatrix}
Y_1 & -Y_2 & 0 & \cdots & 0 \\
Y_1 & 0 & -Y_3 & \cdots & 0 \\
\vdots & \vdots &\vdots & \vdots & \vdots \\
Y_1 & 0 & 0 & \cdots & -Y_r
\end{bmatrix}_{(r-1)\times r\ .}
$$
The ideal generated by the maximal minors of $A$ is $I$. It has codimension 2 in $R$. Thus, the claim follows from the Hilbert-Burch theorem.
\end{proof}
\begin{rem} \label{info about S/J}
Let $R=\Bbbk[Y_1,\dots,Y_r]$ be as above and let $S=\Bbbk[x_1,\dots,x_n]$. Let $F_1,\dots,F_r$ be general homogeneous polynomials in $S$ of degrees $d_1,\dots,d_r$ respectively. Let $F=\prod^r_{i=1}F_i$, $\deg F=d=\sum^r_{i=1}d_i$ and $G_i=F/F_i$, $\deg G_i = d-d_i$, for $1\le i\le r$. Let
$$
\varphi : R \to S
$$
be defined by $\varphi(Y_i)=F_i$. Then, with $I$ as in Lemma \ref{lem3.1},
$\varphi(I)=(G_1,\dots,G_r)=\bigcap_{1\le i<j\le r} (F_i,F_j)=J$.
By the generality in the choice of the $F_i$, $J$ is Cohen-Macaulay of codimension $2$ (see Proposition \ref{nicer}) and its Hilbert-Burch matrix has transpose
$$
\begin{bmatrix}
F_1 & -F_2 & 0 & \cdots & 0 \\
F_1 & 0 & -F_3 & \cdots & 0 \\
\vdots & \vdots &\vdots & \vdots & \vdots \\
F_1 & 0 & 0 & \cdots & -F_r
\end{bmatrix} _{(r-1)\times r \ .}
$$
The minimal free graded resolution of $S/J$ is
\begin{equation}
\label{MFR}
0 \to S^{r-1}(-d) {\to} \bigoplus_{i=1}^r S(-(d-d_i)) \to S \to S/J \to 0.
\end{equation}
It is a simple consequence of this resolution that the artinian reduction of $S/J$ is level with socle degree $d-2$ and Cohen-Macaulay type $(r-1)$. \qed
\end{rem}
\begin{rem}
Observe that $J$ is the ideal $I_P$ for the point $P = [F]$ on $\mathbb X_{n-1, \lambda}$ if we assume $d_1 \ge d_2 \ge \cdots \ge d_r$.
In this paper the main goal is to consider sums of such ideals, i.e. ideals of the form $(G_1,\dots,G_r)$ arising from general forms of prescribed degrees as described above, and to compute the dimension of the component in degree $d$. The fact that they correspond to general points on $\mathbb X_{n-1,\lambda}$ is not needed for most of our computations. Thus, to emphasize the focus on the ideals rather than the points, we will write $I_{(1)} + \dots + I_{(\ell)}$ in place of $I_{P_1} + \dots + I_{P_\ell}$ when the geometric context is not needed, and retain the latter only when the geometry is important (e.g., Remark~\ref{sandro rmk}).
\end{rem}
\begin{rem}
\label{HS}
Let us recall a few results about the Hilbert series of a standard graded ring.
Let $A= \oplus_{i=0}^{\infty} [A]_i$. The {\it Hilbert series of $A$} is the formal power series
$$
\mathbf{HS}(A) = \sum_{i=0}^{\infty} (\dim [A]_i)t^i .
$$
It is a simple matter to show the following two facts, which we will use often in what follows:
$(a)$ If $L$ is a linear non-zerodivisor in $A$ then
$$
\mathbf{HS}(A/LA) = (1-t)\mathbf{HS}(A).
$$
$(b)$ ${\ds \mathbf{HS}(\Bbbk[x_1, \ldots ,x_n]) = \frac{1}{(1-t)^n}}$ and ${\ds \mathbf{HS}(\Bbbk[x_1, \ldots ,x_n](-a)) = \frac{t^a}{(1-t)^n}}$.
\medskip
Of course $(b)$ is a simple consequence of $(a)$.
$(c)$ We can apply these observations to the minimal free resolution \eqref{MFR} in Remark \ref{info about S/J} in order to conclude that
\begin{equation}
\label{eq:hilbSeries tangent}
\mathbf{HS}(S/J) = \frac{1}{(1-t)^n}\Big[ 1 - \sum_{i=1}^r t^{d-d_i} + (r-1)t^d\Big] \ .
\end{equation}
$(d)$ If $A$ and $B$ are graded $\Bbbk$-algebras, then
\[
\mathbf{HS} (A \otimes_{\Bbbk} B) = \mathbf{HS} (A) \cdot \mathbf{HS} (B).
\]
\end{rem}
\medskip
Consider a partition $\lambda=[d_1,\dots,d_r]$, $\lambda \vdash d$. In the polynomial ring $\Bbbk[x_1,\dots,x_n]$ choose general homogeneous forms
$F_1,\dots,F_r$ of degrees $d_1,\dots,d_r$ and, as in Remark \ref{info about S/J}, let $F=\prod^r_{i=1} F_i$, $G_i=F/F_i$ and $I=(G_1,\dots,G_r)=\bigcap_{1\le i<j\le r} (F_i,F_j)$.
\medskip
Inasmuch as we are interested in the secant variety $\sigma_\ell(\mathbb X_{n-1,\lambda})$ we form $\ell$ sets of general polynomials as above in $\Bbbk[x_1,\dots,x_n]$. Call the elements of the $j$-th set
$$
\{F_{j,1},\dots,F_{j,r}\},
$$
where $\deg F_{j,k}=d_k$. As in Remark \ref{info about S/J}, for $1\leq j \leq \ell$ form $M_j = \prod^r_{i=1}F_{j,i} $, and $G_{j,1},\dots,G_{j,r}$ where $G_{j,k}= M_j/F_{j,k}$.
Set
$$
I_{(j)}=(G_{j,1},\dots,G_{j,r})=\bigcap_{1\le i<k\le r} (F_{j,i}.F_{j,k}), \ \ \ \ 1\leq j \leq\ell .
$$
Notice that each of the quotients $S/I_{(j)}$ has the same Hilbert function and minimal free resolution as that of $S/J$ given in Remark~\ref{info about S/J}. Furthermore, each ideal $I_{(j)}$ defines a variety determining the tangent space to $\mathbb X_{n-1, \lambda}$ at the point
$P_j = [F_{j1}F_{j2}\cdots F_{jr}]$.
We can perform the same construction as above, but this time choosing each set of $r$ general polynomials in different polynomial rings, i.e., consider
$\{F_{j,1},\dots,F_{j,r}\}$ as polynomials in the ring $\Bbbk[x_{j,1},\dots,x_{j,n}]$.
We can form the sum of these ideals (extended) in
$$
T=\Bbbk[x_{1,1},\dots,x_{1,n},\dots,x_{\ell,1},\dots,x_{\ell,n}] \ ,
$$
setting $\tilde I=I_{(1)}^{(e)}+\cdots+I_{(\ell)}^{(e)}$ (i.e., the sum of the extended ideals).
\begin{thm}
\label{thm:join as tensor product}
The ring
$$
B = T/\tilde I \cong S/I_{(1)}\otimes_{\Bbbk} \cdots \otimes_{\Bbbk} S/I_{(\ell)}
$$
is Cohen-Macaulay of dimension $\ell(n-2)$. Its minimal graded
free resolution over $T$ is the tensor product (over $\Bbbk$) of the minimal graded free resolutions of $S/I_{(j)}$ over $S$ for $1\le j\le \ell$.
\end{thm}
\begin{proof}
This is a consequence of the K\"unneth formulas. See \cite[Lemma 3.5]{MNP} and its proof.
\end{proof}
Note that $B$ is the coordinate ring of the join of $\ell$ varieties, each of which has codimension 2 in $\P^{n-1}$, so their join is in $\P^{n\ell -1}$. The so-called diagonal trick gives
\[
S/(I_{(1)} + \cdots + I_{(\ell)}) \cong B/\Delta B,
\]
where the diagonal $\Delta$ is generated by the $(\ell - 1) n$ linear forms $x_{1,j} - x_{i, j}$ with $1 < i \le \ell$ and $1 \le j \le n$. Observe that the saturation of $I_{(1)} + \cdots + I_{(\ell)}$ defines the intersection of the indicated varieties in $\P^{n-1}$, provided this intersection is not empty.
A key to our approach is the fact that replacing the linear forms generating the diagonal by truly general linear forms gives a quotient ring with the same Hilbert function as \linebreak $S/(I_{(1)} + \cdots + I_{(\ell)})$. To illustrate the idea, fix a polynomial ring $R$ in $m$ variables, and let $L \in R$ be a general linear form. Since we have a surjection $R \rightarrow R/( L )$, if $\{ F_1,\dots,F_t \}$ is a set of general forms in $R$ of degrees $d_1,\dots,d_t$, then the restriction, $\{ \bar F_1 \dots, \bar F_t \}$, to $R/( L )$ can be viewed again as a set of general forms of degrees $d_1,\dots,d_t$ in $m-1$ variables. Furthermore, given a prescribed construction of an ideal in $m$ variables using general forms of prescribed degrees, the restriction to $R/( L )$ of this ideal can be viewed as an application of the same construction to an ideal of general forms of the same degrees but in $m-1$ variables.
In our setting, if $[I_P]_d = [(G_1,\dots,G_r)]_d$ is the vector space determining the tangent space to $\mathbb X_{n-1,\lambda}$ at a general point $P$ (see Proposition \ref{TangIdeal}), then $[\bar I_P]_d = [(\bar G_1,\dots,\bar G_r)]_d$ is the degree $d$ component of an ideal that determines the tangent space at a general point of the variety $\mathbb X_{n-2,\lambda}$. The analogous statement also holds for an ideal of the form $I_{P_1}+\dots + I_{P_\ell}$.
Returning to the above notation, let $\mathcal{L}$ be a set of $(\ell-1)n$ general linear forms in $T$. Then we have the following useful observation.
\begin{lem}
\label{sop}
The algebras $S/(I_{(1)} + \cdots + I_{(\ell)})$ and $B/{\mathcal{L}}B \cong T/({\mathcal{L}}, {\tilde I})$ have the same Hilbert series. I.e.
$$
\mathbf{HS} (S/(I_{(1)} + \cdots + I_{(\ell)})) = \mathbf{HS} (B/{\mathcal{L}}B) \ .
$$
\end{lem}
\begin{proof} Each ideal $I_{(j)} \subset S$ corresponds to a choice of a general point on $\mathbb X_{n-1,\lambda}$. Thus, it is generated by the $r$ products of $r-1$ distinct forms that are created using $r$ general forms of degrees $d_1, d_2, \ldots , d_r$ in variables $x_1, x_2, \ldots , x_n$ (see Proposition \ref{nicer}). The same is true for the summand $I^{(e)}_{(j)}$ of $\tilde I$, although these forms are in a new set of variables. Since the linear forms in $\mathcal{L}$ are general, the residue classes of the forms defining $I^{(e)}_{(j)}$ modulo $\mathcal{L}$ are again general forms in $S$. It follows that the image $\overline{I^{(e)}_{(j)}}$ of $I^{(e)}_{(j)}$ in $T/{\mathcal{L}}T \simeq S$ also corresponds to a general point on $X_{n-1,\lambda}$. Thus, the ideals $I_{(1)} + \cdots + I_{(\ell)}$ and $\overline{I_{(1)}^{(e)} } + \cdots + \overline{I_{(\ell)}^{(e)} }$ have the same Hilbert function and hence the same Hilbert series.
\end{proof}
\begin{rem}
Lemma \ref{sop} is, in a sense, the key to the results in this paper. In combination with Corollary \ref{dim sigma ell} it shows that computing the dimension of $\sigma_\ell (\mathbb X_{n-1,\lambda})$, for {\em arbitrary} $n, \ell$ and $\lambda$, is equivalent to finding the coefficient of $t^d$ in the Hilbert series of $B/\mathcal L B$. We emphasize here that the only requirement for the linear forms in $\mathcal L$ is that they be general. We do not need them to be regular elements. The next section will handle the case where they are regular elements, and subsequent sections deal with the case where some of the linear forms are not regular elements.
\end{rem}
For our discussion of $\dim_{\Bbbk} [B/{\mathcal{L}}B]_d$ and more generally the Hilbert series of $B/{\mathcal{L}}B$, it is helpful to consider two cases. We refer to them as {\it proper} and {\it improper} intersections.
Consider varieties $V_1,\ldots,V_s \subset \mathbb{P}^{n-1}$. Then their intersection
is defined by the saturation of $I_{V_1} + \cdots + I_{V_s}$ and satisfies
\[
\codim (I_{V_1} + \cdots + I_{V_s}) \le \codim I_{V_1} + \cdots + \codim I_{V_s}.
\]
Abusing notation slightly (in the case where $\codim (I_{V_1} + \cdots + I_{V_s}) = n$, i.e., the intersection is the empty set), we say
that the varieties $V_1,\ldots,V_s \subset \mathbb{P}^{n-1}$ {\em intersect properly} if
\[
\codim (I_{V_1} + \cdots + I_{V_s}) = \codim I_{V_1} + \cdots + \codim I_{V_s}.
\]
Otherwise, they {\em intersect improperly}.
In particular, this means that, fixing $n$ and the partition $\lambda$, if the intersection of the varieties $V(I_{(1)}),\ldots,V(I_{(\ell)})$ is the empty set for some $\ell = \ell_0$, then these varieties intersect improperly for all $\ell > \ell_0$.
\smallskip
We close this section with a fact we will have opportunities to apply later.
We can partially order partitions of an integer $d>0$ as follows. Given partitions $\lambda_1=[d_1,\ldots,d_p]$
and $\lambda_2=[e_1,\dots, e_q]$ of the same integer $d>0$,
write $\lambda_1\geq\lambda_2$ if for each $i\geq0$ we have $\sum_{j\leq i}d_j\geq \sum_{j\leq i}e_j$
(where we regard $d_j$ and $e_j$ as being 0 if $j$ is out of range).
Write $\lambda_1>\lambda_2$ if $\lambda_1\geq\lambda_2$ but $\sum_{j\leq i}d_j> \sum_{j\leq i}e_j$
for some $i$. So, for example, if $q > p$, then either $\lambda_1$ and $\lambda_2$ are incomparable (as happens with $\lambda_1 = [4,3,1]$ and $\lambda_2 = [5,1,1,1]$) or $\lambda_1 > \lambda_2$ (as happens with $\lambda_1 = [5,2,1]$ and $\lambda_2 = [5,1,1,1]$).
\begin{lem} \label{lem:compare binomials}
Let $\lambda_1=[d_1,\ldots,d_p]$ and $\lambda_2=[e_1,\dots, e_q]$ be partitions of the same integer $d$
with $\lambda_1>\lambda_2$. If $n\geq 3$, then $\dim \mathbb X_{n-1,\lambda_1}>\dim \mathbb X_{n-1,\lambda_2}$.
\end{lem}
\begin{proof}
Let $u$ be the least $i$ such that $d_i>e_i$. (There must be such an $i$ since
$d_i\leq e_i$ for all $i$ implies $\sum_{j\leq i}d_j\leq \sum_{j\leq i}e_j$ for all $i$.)
Note that if $u>1$, then
$e_{u-1}=d_{u-1}\geq d_u>e_u$
and $\sum_{j\leq u}d_j> \sum_{j\leq u}e_j$.
Next, let $v$ be the least $i>u$ such that $\sum_{j\leq i}d_j= \sum_{j\leq i}e_j$.
(There must be such an $i$ since both sums eventually are equal to $d$.)
Note that if $v<q$, then $e_v>e_{v+1}$. This is because if $v<q$, then
(by definition of $v$ and the fact that $\sum_{j\leq u}d_j> \sum_{j\leq u}e_j$) we have
$\sum_{j\leq v-1}d_j > \sum_{j\leq v-1}e_j$, but $\sum_{j\leq v}d_j = \sum_{j\leq v}e_j$,
so $e_v>d_v$, and $\sum_{j\leq v+1}d_j \geq \sum_{j\leq v+1}e_j$,
so $d_v\geq d_{v+1}\geq e_{v+1}$.
Now let $\lambda_3=[f_1,\ldots,f_r]$ where $f_u=e_u+1$, $f_v=e_v-1$, and
otherwise $f_j=e_j$. Then $f_j$ is nondecreasing since $e_j$ is and
either $u=1$ or $f_{u-1}=e_{u-1}\geq e_u+1=f_u$, and
either $v=q$ or $f_v=e_v-1\geq e_{v+1}=f_{v+1}$.
Moreover, $\sum_{j\leq i}d_j \geq\sum_{j\leq i} f_j\geq\sum_{j\leq i} e_j$ is true for all $i$.
It holds for $i<u$ since $f_j=e_j=d_j$ in this range.
It holds for $u\leq i<v$ since $\sum_{j\leq i}d_j >\sum_{j\leq i} e_j$
but $1+\sum_{j\leq i}e_j =\sum_{j\leq i} f_j$ in this range.
And it holds for $i\geq v$, since $\sum_{j\leq i}e_j =\sum_{j\leq i} f_j$
in this range.
Thus $\lambda_1\geq\lambda_3>\lambda_2$, so it suffices by induction to show
$\dim \mathbb X_{n-1,\lambda_3}>\dim \mathbb X_{n-1,\lambda_2}$. Writing each $f_j$ in terms of $e_j$, this is equivalent to showing
$\binom{e_u+1+n-1}{n-1}+\binom{e_v-1+n-1}{n-1}-2>\binom{e_u+n-1}{n-1}+\binom{e_v+n-1}{n-1}-2$.
This in turn is equivalent to
\begin{align*}
\textstyle{\binom{e_u+1+n-2}{n-2}} & = \textstyle{\binom{e_u+n-1}{n-2}} \\
& = \textstyle{\binom{e_u+n}{n-1}-\binom{e_u+n-1}{n-1}>
\binom{e_v+n-1}{n-1}-\binom{e_v-1+n-1}{n-1}=\binom{e_v-1+n-1}{n-2}=\binom{e_v+n-2}{n-2},}
\end{align*}
which is true because $\binom{j + n-2}{n-2}$ is a strictly increasing function of $j \ge 0$ if $n-2 \ge 1$.
\end{proof}
We now have:
\begin{cor}\label{cor:dim comparison}
Let $\lambda=[d_1,\ldots,d_r]$ be any partition of $d$ with $r\geq 2$,
and let $\lambda_2=[d_1,1,\ldots,1]$ also be a partition of $d$. Assume $n \geq 3$.
If $\lambda_2\neq\lambda\neq [d-1, 1]$, then
\[
\dim \mathbb X_{n-1, \lambda_2} <\dim \mathbb X_{n-1, \lambda} < \dim \mathbb X_{n-1, [d-1, 1]}.
\]
\end{cor}
\begin{proof}
This follows from Lemma \ref{lem:compare binomials} and the fact that $\lambda_2<\lambda<[d-1, 1]$.
\end{proof}
Notice that a result analogous to Lemma \ref{lem:compare binomials} is not true for the lexicographic order. For example, $\lambda_1 = [5,4,1,1,1,1] > [5,3,3,2] = \lambda_2$ in the lexicographic order, but $\dim \mathbb X_{2, \lambda_1} = 42 < 43 = \dim \mathbb X_{2, \lambda_2}$. Observe that $\lambda_1$ and $\lambda_2$ are not comparable in the partial order used in Lemma \ref{lem:compare binomials}.
\section{Proper Intersections}
\label{sec:proper inters}
In this section we focus on the case where the varieties determining tangent spaces to $\mathbb X_{n-1, \lambda}$ at $\ell$ general points meet properly. Our main result is Theorem \ref{thm:dim formula}. The case of improper intersections is the subject of a later section.
We first show that the $\ell$ varieties determining tangent spaces intersect properly if $\ell$ is small enough.
\begin{prop}
\label{prop:proper intersection}
Assume $2\ell \le n$. Then:
\begin{itemize}
\item[(a)] The $(\ell-1)n$ general linear forms in $\mathcal{L}$ are a $B$-regular sequence.
\item[(b)] The varieties defined by $I_{(1)},\ldots,I_{(\ell)}$ intersect properly, that is,
\[
\codim (I_{(1)}+\cdots+I_{(\ell)})= 2 \ell.
\]
\end{itemize}
\end{prop}
\begin{proof} By Theorem \ref{thm:join as tensor product}, the algebra $B$ is Cohen-Macaulay of dimension $\ell (n-2)$. The assumption on $\ell$ guarantees $(\ell - 1) n \le \ell (n-2)$. Hence $\mathcal{L}$ is a regular sequence and $B/\mathcal{L} B$ has dimension $n - 2 \ell$. Now Lemma \ref{sop} gives $
\codim (I_{(1)}+\cdots+I_{(\ell)} )= 2 \ell$.
\end{proof}
\begin{rem} If $\ell \le \frac{n}{2}$, then the minimal free graded resolution of $S/(I_{(1)}+\cdots+I_{(\ell)})$
has the same graded Betti numbers as the minimal free graded resolution of $T/\tilde I$ since forming a quotient by factoring with a regular sequence does not change the graded Betti numbers of the resolution modules.
\end{rem}
\begin{rem}
\label{formula}
Using the isomorphism of graded modules (see Theorem \ref{thm:join as tensor product})
\[
T/\tilde I \cong S/I_{(1)} \otimes_{\Bbbk} \cdots \otimes_{\Bbbk} S/I_{(\ell)},
\]
it follows that $\mathbf{HS}(T/\tilde I)=(\mathbf{HS}(S/J))^\ell$, where $J$ is as given in Remark \ref{lem3.1}.
Furthermore, if $2\ell\le n$, Proposition \ref{prop:proper intersection}(a), Remark \ref{HS}(d), and Lemma \ref{sop} give
\[
\begin{array}{lllllllllllllllll}
\mathbf{HS}(S/(I_{(1)}+\cdots +I_{(\ell)}))
& = &(1-t)^{n(\ell-1)} \cdot \mathbf{HS}(T/\tilde I) \\
& = & (1-t)^{n(\ell-1)} \cdot [\mathbf{HS}( S/J)]^\ell.
\end{array}
\]
Putting this together with Equation \eqref{eq:hilbSeries tangent} in Remark \ref{HS} we obtain
\[
\mathbf{HS}(\Bbbk[x_1,\dots,x_n]/(I_{(1)}+\cdots +I_{(\ell)})) = \frac{1}{(1-t)^n} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell .
\]
Rewriting this last expression we have, if $2\ell \leq n$, that
\begin{equation*} \label{key eqn}
\begin{array}{llllllllllllllllllllllll}
\mathbf{HS} (I_{(1)}+\cdots +I_{(\ell)})
= \ds \frac{1}{(1-t)^n} - \frac{1}{(1-t)^n}\left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell.
\end{array}
\end{equation*}
\end{rem}
\medskip\medskip
If we now put together Corollary \ref{dim sigma ell} and Remark \ref{formula} we obtain:
\begin{thm}
\label{HS and dim sigma ell}
Let $\lambda\vdash d$, $\lambda = [d_1, \ldots , d_r]$ and suppose that $2\ell \leq n$. Put
\[
A = \Bbbk[x_1,\dots,x_n]/(I_{(1)}+\cdots +I_{(\ell)}).
\]
Then
\[
\mathbf{HS} (A) = \frac{1}{(1-t)^n}\left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell.
\]
Moreover, if $a_d$ denotes the coefficient of $t^d$ in $\mathbf{HS} (A)$, then
$$
\dim \sigma_\ell (\mathbb X_{n-1,\lambda}) = \binom{d+n-1}{n-1} - a_d - 1 .
$$
\end{thm}
We now compute the coefficient $a_d$, which gives the main result of this section.
\begin{thm}\label{thm:dim formula}
Let $\lambda\vdash d$, $\lambda = [d_1,\ldots,d_r]$ with $r \ge 2$. If $2\ell \le n$ then:
\begin{eqnarray*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) & = & \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \\
&& \ - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} \\
& & - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1}.
\end{eqnarray*}
Moreover, $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills its ambient space if and only if one of the following conditions is satisfied:
\begin{itemize}
\item[(i)] $n = 4$, $\ell = 2$, and $\lambda \in \{[1,1], [2,1], [1,1,1]\}$ \quad or
\item[(ii)] $n = 2 \ell \ge 6$ and $\lambda = [1,1]$.
\end{itemize}
\end{thm}
\begin{proof} Let $P_1, \ldots , P_\ell$ be general points on $\mathbb X_{n-1, \lambda}$ and set
\[
A = \Bbbk[x_1,\dots,x_n]/(I_{P_1}+\cdots +I_{P_\ell}).
\]
Then we have seen in Theorem \ref{HS and dim sigma ell} that
\begin{equation}
\label{eq:hilb series}
\mathbf{HS}(A) = \frac{1}{(1-t)^n} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell .
\end{equation}
Observing that
\[
2 (d-d_1) + (d-d_2) \ge 2 d_2 + (d_1 + d_3 + \cdots + d_r) > d,
\]
we get
\begin{eqnarray*}
\left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell & = &
\left [ 1-\sum^r_{i=1} t^{d-d_i} \right] ^\ell + (r-1) \ell \cdot t^d + \cdots \\[1ex]
& = & 1 - \ell \sum^r_{i=1} t^{d-d_i} + \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \cdot t^{k (d-d_1)} \\
& & \ + \ell (\ell -1) \cdot t^{d-d_1 + d-d_2} + \binom{\ell}{2} \cdot t^{2 (d - d_2)} + (r-1) \ell \cdot t^d + \cdots
\end{eqnarray*}
where only the terms whose exponent of $t$ are potentially at most $d$ have been written out.
Using also
\[
\frac{1}{(1-t)^n} = \sum_{j \ge 0} \binom{j+n-1}{j} \cdot t^j ,
\]
we get from Equation \eqref{eq:hilb series}
\[
\mathbf{HS}(A) = \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell \cdot \left [ \sum_{j \ge 0} \binom{j+n-1}{j} \cdot t^j \right ] .
\]
The coefficient of $t^d$ in $\mathbf{HS} (A)$ is
\begin{eqnarray*}
\dim_{\Bbbk} [A]_d & = & \binom{d+n-1}{n-1} - \ell \sum^r_{i=1} \binom{d_i + n-1}{n-1} + (r-1) \ell \\
& & + \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} \\[1ex]
& & + \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} + \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1}.
\end{eqnarray*}
This gives
\begin{eqnarray*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) & = & -1 + \dim_{\Bbbk} [I_{(1)}+\cdots +I_{(\ell)})]_d \\
& & \\
& = & -1 + \ell \sum^r_{i=1} \binom{d_i + n-1}{n-1} - (r-1) \ell \\
& & \ - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} \\
& & - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} .
\end{eqnarray*}
Then using Formula (\ref{dim X}), we get
\begin{eqnarray*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) & = & \ell \cdot \dim \mathbb X_{n-1, \lambda} + (\ell -1) \\
& & \ - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} \\
& & - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1},
\end{eqnarray*}
as claimed.
It remains to show the characterization, for $2\ell \leq n$, of when $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space.
This clearly occurs if and only if $[A]_d = 0$. If $2\ell < n$, then $[A]_d$ cannot be zero because $A$ is not artinian as $\dim A = n - 2 \ell$.
Let $n = 2 \ell$. In this case we can write
$$
\mathbf{HS}(A) = \left[ \frac{1}{(1-t)^{2 \ell}} \left (1-\sum_{i=1}^r t^{d-d_i} + (r-1)t^d \right ) \right] ^\ell .
$$
Remark \ref{info about S/J} implies that the artinian reduction of $S/J$ is level of socle degree $d-2$. Hence, each factor
\[
\frac{1}{(1-t)^{2}} \left (1-\sum_{i=1}^r t^{d-d_i} + (r-1)t^d \right )
\]
is a polynomial of degree $d-2$. It follows that $\mathbf{HS} (A)$ is a polynomial of degree $\ell (d-2)$. Since $A$ is artinian, this shows that $[A]_d = 0$ if and only if $\ell (d-2) < d$. This is equivalent to
\[
d < 2 + \frac{2}{\ell-1}.
\]
If $\ell = 2$ (hence $n=4$), then we can have $d=2$ or 3 and so $\lambda$ can be $[1,1], [2,1]$ or $[1,1,1]$. If $\ell > 2$ then we must have $d = 2$ and $\lambda = [1,1]$.
\end{proof}
\begin{rem}\label{correctionExample}
Note that $\dim \sigma_\ell(\mathbb X_{n-1, \lambda})<N-1<\ell(\dim_\Bbbk(\mathbb X_{n-1, \lambda}))+\ell-1$ can occur (where $N=\binom{d+n-1}{n-1}$),
as happens, for example, when $n=4$, $\lambda=[2,2]$ and $\ell=2$.
Thus, when $\sigma_\ell(\mathbb X_{n-1, \lambda})$ does not fill its ambient space, the defect in Theorem \ref{thm:dim formula}
need not be given by
$$\textstyle\sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} +
\binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} + \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1}.$$
In general one must use
$$\textstyle\sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} +
\binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} + \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1}-\epsilon,$$
where $\epsilon=\max(0,\ell(\dim_\Bbbk(\mathbb X_{n-1, \lambda}))+\ell-1-(N-1))$.
\end{rem}
\begin{rem}\label{is zero}
Observe that the last term in the formula of Theorem \ref{thm:dim formula} is zero if and only if $r \ge 3$ and that the penultimate term is zero unless $ r = 2$ and $d_1 = d_2$.
\end{rem}
\begin{rem}\label{rem:syzygy interpretation}
There is an interesting way to interpret the formula in Theorem \ref{thm:dim formula}.
Let $\mathcal I \subset S$ be an ideal generated by $\ell$ general forms of degree $s = d_2 + \dots + d_r$, where $\ell \leq n$. Let $\Syz$ be the module of first syzygies of $\mathcal I$, i.e., the sequence
$$
0 \longrightarrow \Syz \longrightarrow S(-s)^\ell \longrightarrow {\mathcal I} \longrightarrow 0
$$
is exact.
From the Koszul complex we obtain the following resolution of $\Syz$,
\begin{equation}\label{KoszulES}
0\rightarrow S(-\ell s )^{\binom{\ell}{\ell}} \rightarrow \cdots \rightarrow S(-3 s)^{\binom{\ell}{3}}\rightarrow S(-2 s)^{\binom{\ell}{2}} \rightarrow \Syz\rightarrow 0
\end{equation}\label{SyzFormula}
and so
\begin{equation}\label{SyzFormula2}
\dim_{\Bbbk} [\Syz]_d = \sum_{k=2}^{\ell} (-1)^k\binom{\ell}{k} \binom{d-k s +n-1}{n-1} .
\end{equation}
\end{rem}
Using Remark \ref{rem:syzygy interpretation} and Theorem \ref{thm:dim formula} we get
\begin{cor}\label{syzygycor}
Let $2\ell \leq n$ and $\lambda = [d_1, \ldots , d_r] \vdash d$. Let $\mathcal I \subset S$ be an ideal generated by $\ell$ general forms of degree $s=d_2 + \cdots + d_r$ and let $\Syz$ be the first syzygy module of $\mathcal I$.
Then
\begin{align*}
\dim \sigma_\ell(\mathbb X_{n-1,\lambda}) = & \ \ell \dim \mathbb X_{n-1,\lambda} + (\ell - 1) - \dim_{\Bbbk} [\Syz]_d \\[.5ex]
& - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} \ .
\end{align*}
\end{cor}
\begin{rem} \label{defect with syz} It is immediate from Corollary \ref{syzygycor} that if $r\geq 3$ then
$$
\dim \sigma_\ell(\mathbb X_{n-1,\lambda})= \ell \dim \mathbb X_{n-1,\lambda} + (\ell - 1) - \dim_{\Bbbk} [\Syz]_d
$$
and if $r=2$ and $d_1> d_2$ then
$$
\dim \sigma_\ell(\mathbb X_{n-1,\lambda})= \ell \dim \mathbb X_{n-1,\lambda} + (\ell - 1) - \dim_{\Bbbk} [\Syz]_d - \ell(\ell -1) .
$$
\end{rem}
We now discuss the defectivity of $\sigma_\ell(\mathbb X_{n-1, \lambda})$ if $2\ell \leq n$. We note that we
state additional results in the case $\ell = 2$ without this restriction on $n$ in Section \ref{sec:secant lines}.
For convenience, we consider the case $r = 2$ separately.
\begin{comment}
\begin{thm}\label{general r=2}
Let $r=2$ with $2 \le \ell \le \frac{n}{2}$ and $\lambda = [d_1,d_2]$.
Then $\sigma_\ell (\mathbb X_{n-1,\lambda})$
fills its ambient space if and only $n = 2 \ell$ and $\lambda = [1,1]$.
In all other cases, $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is defective and the defect is
\begin{equation*}
\delta_\ell = \begin{cases}
2 \ell (\ell -1) & \text{if } d_1 = d_2 \\
\ell (\ell -1) + \dim_{\Bbbk} [\Syz]_d & \text{if } d_1 >d_2,
\end{cases}
\end{equation*}
where an explicit formula for $\dim_{\Bbbk} [\Syz]_d$ is given in \eqref{SyzFormula2}.
\end{thm}
\end{comment}
\begin{thm}\label{general r=2}
Let $r=2$ with $\ell \le \frac{n}{2}$ and $\lambda = [d_1,d_2]$.
Then $\sigma_\ell (\mathbb X_{n-1,\lambda})$
fills its ambient space if and only if $n = 2 \ell$ and
$\lambda = [1,1]$, or $n=4$, $\ell=2$ and $\lambda = [2,1]$.
In all other cases, $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is defective and the defect is
\begin{equation*}
\delta_\ell = \begin{cases}
2 \ell (\ell -1) - \epsilon & \text{if } d_1 = d_2, \text{ and}\\
\ell (\ell -1) - \epsilon+\dim_{\Bbbk} [\Syz]_d & \text{if } d_1 >d_2,
\end{cases}
\end{equation*}
where an explicit formula for $\dim_{\Bbbk} [\Syz]_d$ is given in \eqref{SyzFormula2},
and
$$\epsilon= \max\{ 0,\ \ell\cdot \dim_\Bbbk(\mathbb X_{n-1,\lambda}) + \ell -1-(N-1)\}. $$
\end{thm}
\begin{comment}
\begin{proof}
Theorem \ref{thm:dim formula} implies the first claim, and also that $\sigma_\ell (\mathbb X_{n-1,\lambda})$ does not fill its ambient space
in all other cases, in which case
\[
\delta_{\ell} = \dim_{\Bbbk} [\Syz]_d + \binom{\ell}{2} \binom{d_2 - d_1 + n-1}{n-1} + \ell (\ell -1) .
\]
Now a short computation gives the claim.
\end{proof}
\end{comment}
\begin{proof}
Theorem \ref{thm:dim formula} implies the first claim, and that $\sigma_\ell (\mathbb X_{n-1,\lambda})$ does not fill its ambient space
in all other cases. Keeping Remark \ref{correctionExample} in mind, we get $\delta_\ell$ for the case that $d_1 = d_2$ from Theorem \ref{thm:dim formula},
and from Remark \ref{defect with syz} for the other case.
\end{proof}
\begin{rem}
The paper \cite{CCG:1} solves the problem of determining when $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills its ambient space in the case when $r=2$ and $2 \ell \leq n+1$, in quite different language. We should point out, first, that their $r$ is our $\ell$ and their $n$ is our $n-1$. Most of the cases that they need to consider satisfy $2 \ell = n+1$, and so do not overlap with Theorem \ref{general r=2} (but see Theorem \ref{thm:dim when froeberg} (a)). Nevertheless, the second sentence of Theorem \ref{general r=2} follows from Theorem 5.1 of \cite{CCG:1}.
We included it above for reference and because it is such an easy consequence of our current approach.
\end{rem}
\begin{comment}
\begin{thm} \label{general r geq 3}
Let $\lambda \vdash d$, $\lambda =[d_1, \ldots , d_r]$, where $r \ge 3$. Assume $3 \le \ell \le \frac{n}{2}$.
\begin{enumerate}
\item[(a)] If $d_1 < d_2 + \dots + d_r$ then
$$
\dim \sigma_\ell( \mathbb X_{n-1,\lambda} ) = \ell \cdot \dim \mathbb X_{n-1,\lambda}+ (\ell - 1) \leq N-1
$$
and $\sigma_\ell( \mathbb X_{n-1,\lambda} )$ is not defective.
\item[(b)] If $d_1 \geq d_2 + \dots + d_r$ then $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is defective, with defect
$$
\delta_\ell = \\dim_{\Bbbk} [\Syz]_d > 0,
$$
where an explicit formula for $\dim_{\Bbbk} [\Syz]_d$ is given in \eqref{SyzFormula}.
\end{enumerate}
\end{thm}
\begin{proof}
We begin with (a).
The proof is immediate from Theorem \ref{thm:dim formula} since, as partly noted
in Remark \ref{is zero}, all the summands in that formula vanish except for $\ell \cdot \dim \mathbb X_{n-1,\lambda}+ (\ell - 1)$.
As for (b), Theorem \ref{thm:dim formula} shows that $\sigma_\ell( \mathbb X_{n-1,\lambda} ) $ does not fill its ambient space. Hence, Corollary \ref{syzygycor} gives that the defect is
\[
\delta_\ell = \dim_{\Bbbk} [\Syz]_d.
\]
It remains to show that it is positive. However, $\Syz$ is a submodule of a free module, and the generators of $\Syz$ have degree $2(d_2+\cdots+d_r) \le d$ (see \eqref{KoszulES}). We conclude that $\dim_{\Bbbk} [\Syz]_d >0$, and so $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is defective.
\end{proof}
\end{comment}
\begin{thm} \label{general r geq 3}
Let $\lambda \vdash d$, $\lambda =[d_1, \ldots , d_r]$, where $r \ge 3$. Assume $2\ell \le n$.
\begin{enumerate}
\item[(a)] If $d_1 < d_2 + \dots + d_r$, then
$$
\dim \sigma_\ell( \mathbb X_{n-1,\lambda} ) = \ell \cdot \dim \mathbb X_{n-1,\lambda}+ (\ell - 1) \leq N-1
$$
and $\sigma_\ell( \mathbb X_{n-1,\lambda})$ is not defective.
\item[(b)] If $d_1 \geq d_2 + \dots + d_r$, then $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is defective, with defect
$$
\delta_\ell = \dim_{\Bbbk} [\Syz]_d -\epsilon,
$$
where $\dim_{\Bbbk} [\Syz]_d$ is given explicitly in \eqref{SyzFormula}
and $\epsilon$ is as given in Remark \ref{correctionExample}.
\end{enumerate}
\end{thm}
\begin{proof}
We begin with (a).
The proof is immediate from Theorem \ref{thm:dim formula} since, as partly noted
in Remark \ref{is zero}, all the summands in that formula vanish except for $\ell \cdot \dim \mathbb X_{n-1,\lambda}+ (\ell - 1)$.
As for (b), Theorem \ref{thm:dim formula} shows that $\sigma_\ell( \mathbb X_{n-1,\lambda})$ does not fill its ambient space.
Thus, if $N-1\leq\ell(\dim \mathbb X_{n-1,\lambda})+\ell-1$, then $\sigma_\ell( \mathbb X_{n-1,\lambda})$ is defective, and using
Remarks \ref{correctionExample} and \ref{defect with syz} we see the defect is
$\delta_\ell=\dim_{\Bbbk} [\Syz]_d-\epsilon$.
Suppose now that
$N-1>\ell(\dim \mathbb X_{n-1,\lambda})+\ell-1$, so $\epsilon=0$
Hence, Remark \ref{defect with syz} gives that the defect is
\[
\delta_\ell = \dim_{\Bbbk} [\Syz]_d=\dim_{\Bbbk} [\Syz]_d-\epsilon.
\]
It remains to show that it is positive. However, $\Syz$ is a submodule of a free module, and the generators of $\Syz$ have degree $2(d_2+\cdots+d_r) \le d$ (see \eqref{KoszulES}). We conclude that $\dim_{\Bbbk} [\Syz]_d >0$, and so $\sigma_\ell(\mathbb X_{n-1,\lambda})$ again is defective.
\end{proof}
\begin{rem} \label{PDH}
Theorems \ref{general r=2} and \ref{general r geq 3} give
us our first view of the fact that the hyperplane $d_1 = d_2 + \dots + d_r$ in $\mathbb N^r$ separates two very different kinds of behaviors with respect to defectivity, when $[d_1, \ldots , d_r] = \lambda\vdash d$ is a partition of $d$ into $r\geq 2$ parts. This was observed for $n=2$ in \cite{CGGS}, and it will recur frequently in this paper. As a result, we will follow \cite{CGGS} in referring to this as the {\em partition dividing hyperplane in $\mathbb N^r$}.
\end{rem}
\begin{rem} \label{sandro rmk}
The formula for $\dim \sigma_\ell (\mathbb X_{n-1,\lambda})$ given in Conjecture \ref{conj:main conj}(b) and Theorem \ref{thm:dim formula}
comes from Corollary \ref{dim sigma ell} after interpreting
$\dim_{\Bbbk} [I_{P_1}+ \cdots +I_{P_\ell}]_d = \dim_{\Bbbk}([I_{P_1}]_d+ \cdots +[I_{P_\ell}]_d)$.
The simplest case occurs when the spaces $[I_{P_i}]_d$ meet pair-wise in just $0$. In that case
$\dim_{\Bbbk}([I_{P_1}]_d+ \cdots +[I_{P_\ell}]_d)=\dim_{\Bbbk}[I_{P_1}]_d+ \cdots +\dim_{\Bbbk} [I_{P_\ell}]_d
=\ell(\dim_{\Bbbk} [I_{P_1}])=\ell(1+\dim \mathbb X_{n-1,\lambda})$ so
$\dim \sigma_\ell (\mathbb X_{n-1,\lambda})=\ell(\dim \mathbb X_{n-1,\lambda})+\ell-1$.
Often the pairs will not meet only in 0. To consider that case,
we set $\underline{i}=\{i_1,\ldots,i_j\}$ for $1\leq i_1<\cdots< i_j\leq \ell$ and say $|\underline{i}|=j$.
We then define $V_{\underline{i}}=\cap_{t\in\underline{i}} [I_{P_t}]$.
For $1\leq u\leq\ell$, let $v_u=\sum_{|\underline{i}|=u}\dim V_{\underline{i}}$,
so $v_1=\dim_{\Bbbk}[I_{P_1}]_d+ \cdots +\dim_{\Bbbk} [I_{P_\ell}]_d$,
$v_2$ is the sum of the dimensions of the pair-wise intersections of the $[I_{P_i}]_d$,
$v_3$ is the sum of the triple intersections, and so on.
Inclusion-exclusion now gives
$\dim_{\Bbbk}([I_{P_1}]_d+ \cdots +[I_{P_\ell}]_d)=\sum_{1\leq u\leq\ell}(-1)^{u+1}v_u$.
Let's look at this in the case that $\lambda$ is such that $d_1\geq s = d_2+d_3+ \cdots +d_r$; i.e. $\lambda$ is
above the ``partition dividing hyperplane" in $\mathbb{N}^r$.
This is an example for which it is not possible that
$[I_{P_{i_1}}]_d\cap[I_{P_{i_2}}]_d=0$ for $i_1\neq i_2$.
To see this, say each $P_j$ corresponds to the form
$F_{j,1}F_{j,2} \cdots F_{j,r}$.
Then $I_{P_i} \cap I_{P_j}$ will, for every $i,j\in \{1, \dots ,\ell\}$, $i\neq j$, contain all the products of the type:
$F_{i,2} \cdots F_{i,r}F_{j,2} \cdots F_{j,r}G$,
where $G$ is any form of degree $d-2s=d_1-s$.
Thus $\dim_\Bbbk [I_{P_{i_1}}]_d\cap[I_{P_{i_2}}]_d\geq \binom{d-2s+n-1}{n-1}$ for each
pair $i_1\neq i_2$, so $v_2\geq \binom{\ell}{2}\binom{d-2s+n-1}{n-1}$.
Moreover, when $r=2$, $I_{P_i}\cap I_{P_j}$ will also contain the forms
$F_{1,i}F_{2,j}$,
for all $i\neq j\in \{1, \ldots ,\ell \}$ so in this case $v_2\geq \binom{\ell}{2}\binom{d-2s+n-1}{n-1}+\ell(\ell-1)$.
If $r=2$ and $d_1=d_2$, we also have
$F_{1,i}F_{1,j}$ in the intersection, so $v_2\geq \binom{\ell}{2}\binom{d-2s+n-1}{n-1}+\ell(\ell-1)+\binom{\ell}{2}$.
Similarly, if $d_1\geq 2s$, then $\dim_\Bbbk [I_{P_{i_1}}\cap I_{P_{i_2}}\cap I_{P_{i_3}}]_d\geq \binom{d-3s+n-1}{n-1}$,
since
$$F_{i_1,2} \cdots F_{i_1,r}F_{i_2,2} \cdots F_{i_2,r}F_{i_3,2} \cdots F_{i_3,r}H\in [I_{P_{i_1}}\cap I_{P_{i_2}}\cap I_{P_{i_3}}]_d$$
for any $H\in S_{d-3s}$. Hence $v_3\geq\binom{\ell}{3}\binom{d-3s+n-1}{n-1}$.
In the same way, if $d_1\geq ks$, then $v_k\geq\binom{\ell}{k}\binom{d-ks+n-1}{n-1}$.
If the lower bounds on each $v_u$ above were to equal the corresponding $v_u$, then
$\dim_{\Bbbk}([I_{P_1}]_d+ \cdots +[I_{P_\ell}]_d)=\sum_{1\leq u\leq\ell}(-1)^{u+1}v_u$
becomes precisely the formula given in
Conjecture \ref{conj:main conj}(b) and proved in a special case in Theorem \ref{thm:dim formula}.
Of course, because cancellations may occur in an alternating sum, it is possible
for the formula in Conjecture \ref{conj:main conj}(b) to hold even if some of the lower bounds
were to be strictly less than their corresponding $v_u$.
\end{rem}
\section{The Secant Line Variety and Passage to Improper Intersections} \label{secant line section}
\label{sec:secant lines}
A consequence of the work done up to this point is that if $n \geq 4$ we have the following result for the secant line variety of the variety of reducible forms of prescribed type:
\begin{thm}
\label{thm: l = 2, proper intersection}
Assume $\ell = 2$ and $n \geq 4$. Then, for all partitions $\lambda = [d_1,\dots,d_r]$ of $d$ with $r \geq 2$:
\begin{itemize}
\item[(a)] \begin{eqnarray*}
\dim \sigma_{2} (\mathbb X_{n-1, \lambda}) & = & 2 \dim \mathbb X_{n-1, \lambda} + 1 - \binom{d_1 - (d_2 + \cdots + d_r) + n-1}{n-1} \\[.5ex]
&& \ - \binom{2 d_2 - d + n-1}{n-1} - 2 \binom{d_1 + d_2 - d+ n-1}{n-1};
\end{eqnarray*}
\item[(b)] $\sigma_2( \mathbb X_{n-1,\lambda})$ fills its ambient space if and only if $\lambda \in \{[1,1], [2, 1], [1,1,1]\}$ and $n = 4$;
\item[(c)] if $\sigma_2( \mathbb X_{n-1,\lambda})$ does not fill its ambient space, then
$\sigma_{2} (\mathbb X_{n-1, \lambda})$ is not defective if and only if $d_1 < d_2 + \cdots + d_r$; and
\item[(d)] if $\sigma_2( \mathbb X_{n-1,\lambda})$ is defective, then the defect is
\[
\binom{d_1 - (d_2 + \cdots + d_r) + n-1}{n-1} + \binom{2 d_2 - d + n-1}{n-1} + 2 \binom{d_1 + d_2 - d+ n-1}{n-1}-\epsilon,
\]
where $\epsilon$ is as given in Remark \ref{correctionExample}.
\end{itemize}
\end{thm}
\begin{proof}
Parts (a) and (b) are immediate from Theorem \ref{thm:dim formula}.
Theorem \ref{thm:defectivity - intro} gives part (c). Claim (d) is a consequence of (a).
\end{proof}
Notice that Theorem \ref{thm: l = 2, proper intersection} assumes $n\geq 4$, since $\ell=2$ but it relies on results that assume $n\geq 2\ell$. The main purpose of this section is to understand what is needed to pass beyond the condition $2 \ell \leq n$ with our approach. We begin with a review of the main results of \cite{CGGS}, since that paper gives a careful analysis of the case $\ell = 2$, $n = 3$ using entirely different methods. Then we will see what would be needed in order to pass from the results of the previous section to this case. Having the essential idea in hand, subsequent sections of this paper will carry out the calculations. The main result is a single, explicit (albeit complicated) conjectured formula for the dimension of the secant variety for {\em any} choice of $n$, $\lambda$ and $\ell$ (see Conjecture \ref{conj:main conj} and Theorem \ref{thm: gen dim formula}). In particular, the results of the previous sections agree with this conjecture. We are then able to give many consequences, some proven unconditionally, some conjectural.
So first we recall the results of \cite{CGGS}, which in particular confirm the importance of the ``partition dividing hyperplane" mentioned in Remark \ref{PDH}.
We will see that, in general, the partitions \lq\lq below" the hyperplane, i.e.\ those partitions for which $d_1 < d_2+\cdots+d_r$, behave quite differently from those \lq\lq above" the hyperplane, i.e. those partitions for which $d_1 \geq d_2+\cdots+d_r$. In fact, we expect this to be true in all cases (see Conjecture \ref{conj:defectivity conj} and Proposition \ref{prop:non-defective}).
Propositions \ref{r>6} and \ref{r less than 6} can be deduced easily from the results of \cite{CGGS}.
We see that the condition for defectivity when $r\geq6$ is straightforward, but there are exceptions
when $2\leq r<6$.
\begin{prop}\label{r>6} Let $n=3$ with $\lambda= [d_1, \ldots , d_r]\vdash d$ a partition of $d$ into $r \geq 6$ parts
and $s=d_2+\cdots+d_r$. Set
$$
p= \sum _{2\leq i<j\leq r}d_id_j .
$$
\begin{itemize}
\item[(a)] If $d_1 \geq s$ , then $\sigma_2( \mathbb X_{2,\lambda})$ is always defective, and the defect is
\begin{equation}
\label{eq:defect}
\min\left\{ \binom{d_1-s+2}{2}, 2p-3s \right\}.
\end{equation}
\item[(b)] If $d_1 < s$, then
$$
\dim \left( \sigma_2( \mathbb X_{2,\lambda})\right) = 2\dim ( \mathbb X_{2,\lambda}) + 1,
$$
and hence $\sigma_2( \mathbb X_{2,\lambda})$ is not defective.
\end{itemize}
\end{prop}
\begin{prop} \label{r less than 6}
Let $n=3$ with $\lambda= [d_1, \ldots , d_r]\vdash d$ a partition of $d$ into $2\leq r<6$ parts and let $s=d_2+\cdots+d_r$.
\begin{itemize}
\item[(a)] If $r=2$, then the secant line variety fills its ambient space, and so it is never defective.
\item[(b)] For the following partitions the secant variety $\sigma_2( \mathbb X_{2,\lambda})$ fills its ambient space and so the defect is zero:
\begin{itemize}
\item[$\bullet$] $r=3$ and $\lambda \in \{ [d_1,d_2,1], [d_1,2,2], [d_1,3,2], [d_1,4,2], [d_1,5,2],[d_1,6,2], [d_1,3,3] \}$;
\item[$\bullet$] $r=4$ and $\lambda \in \{ [d_1,1,1,1],[d_1,2,1,1],[d_1,3,1,1],[d_1,4,1,1] \}$;
\item[$\bullet$] $r = 5$ and $\lambda = [d_1,1,1,1,1]$.
\end{itemize}
\item[(c)] Apart from the partitions described above, if $d_1\geq s$ and $r \ge 3$, then $\sigma_2( \mathbb X_{2,\lambda})$ is always defective. In this case the defect is equal to \eqref{eq:defect} above.
\item [(d)] If $d_1<s$ then $\sigma_2( \mathbb X_{2,\lambda})$ is never defective. Apart from the partitions described in (b), the secant line variety has dimension $2 \dim( \mathbb X_{2,\lambda}) + 1$.
\end{itemize}
\end{prop}
\begin{ex} Consider $\lambda = [2,2,2,1]$. This partition has $d_1<s$ and so we are below the partition dividing hyperplane.
By Proposition \ref{r less than 6}(d), $\dim \sigma_2( \mathbb X_{2,\lambda})=2\dim \mathbb X_{2,\lambda}+ 1$ and a computation shows that
$2\dim \mathbb X_{2,\lambda}+ 1=N-1$, hence $\sigma_2( \mathbb X_{2,\lambda})$ fills its ambient space for this example.
\end{ex}
\begin{rem} \label{hyper surf} In case $\ell = 2$, $\lambda\vdash d, \ \lambda = [d_1, \ldots , d_r]$, with $d_1 = s$ we can show that the only time that $\sigma_2 ( \mathbb X_{2,\lambda})$ is a hypersurface in its ambient space is when $\lambda = [9,7,2], [5,2,2,1]$ and $[7,5,1,1]$. It would be interesting to find the equations of these hypersurfaces. These are all defective secant line varieties, and up to this point whenever we have been able to find equations for such defective secant varieties they have been determinants. Is that the situation in this case as well?
\end{rem}
Now fix $n = 3$ and $\ell = 2$. Then the codimension of $\sigma_{2} (\mathbb X_{3, \lambda})$ is
given by \linebreak $\dim_{\Bbbk} [S/(I_{P_1}+I_{P_2})]_d$ (see
Corollary \ref{dim sigma ell}). By Lemma \ref{sop} and Theorem \ref{thm:join as tensor product}, the latter is equal to $\dim_{\Bbbk} [B/ \mathcal L B]_d$, where $B = T/(I_{(1)}^{(e)}+I_{(2)}^{(e)})$ and $\mathcal L \subset T$ is a regular sequence comprised of $(\ell-1)n = 3$ general linear forms. By Remark \ref{formula}, we know the Hilbert function of $B$.
Thus, the important question now is to determine the relation between the Hilbert function of $B$ and that of $B/ \mathcal L B$.
Since $\dim B = \ell(n-2) = 2$ (Theorem \ref{thm:join as tensor product}), the first two linear forms in $\mathcal L$ form a $B$-regular sequence by Proposition \ref{prop:proper intersection}. Say $\mathcal L = \{L_1,L_2,L_3 \}$ and $L_1,L_2$ form a regular sequence. Thus we have to find the Hilbert function after reducing by one more general linear form, $L_3$. Let $\mathcal L' = \{L_1,L_2 \}$ and let $ B' = B/\mathcal L' B$.
Note that there is an exact sequence
\begin{equation} \label{mult SES}
B'(-1) \stackrel{\times L_3}{\longrightarrow} B' \rightarrow B'/ L_3 B' \rightarrow 0.
\end{equation}
Thus, in order to determine the Hilbert function of $B/\mathcal L B \cong B'/ L_3 B'$ it is enough to know that $\times L_3$ has maximal rank at each degree. This is exactly what is provided by the Weak Lefschetz Property, as described in the next section. It is the basis for the remaining calculations and the general formulae that they give.
As an example, let us consider a case mentioned in Remark \ref{hyper surf} and verify that if $\lambda = [9,7,2]$, and if the maximal rank property holds for $ B'$, then our results of the previous sections imply that $\sigma_2 (\mathbb X_{2,\lambda})$ is a hypersurface. Using Remark \ref{formula}, we get for the Hilbert series of $ B'$
\begin{align*}
\mathbf{HS} ( B') & = (1 - t)^2 \cdot \mathbf{HS} (B) \\[.5ex]
& = \frac{(1-t)^6}{(1-t)^4} \cdot \mathbf{HS}(B) \\[.5ex]
& = \frac{1}{(1-t)^4}\left [1 - t^{16} - t^{11} - t^9 + t^{18} \right ]^2 \\[.5ex]
& = 1 + 4 t + \cdots + 634 t^{17} + 635 t^{18} + \cdots + 4 t^{32}.
\end{align*}
Thus,
\[
h_{17} = \dim_{\Bbbk} [ B']_{17} = 634 \quad \text{ and } \quad h_{18} = \dim_{\Bbbk} [ B']_{18} = 635 .
\]
Note that $d = 18$ in our example. If multiplication by $L_3$ on $ B'$ has maximal rank, we obtain (see Sequence \eqref{mult SES}) that for $S = \Bbbk[x_1,x_2,x_3]$,
\[
\dim_{\Bbbk} [S/(I_{P_1}+I_{P_2})]_d = \dim_{\Bbbk} [ B'/L_3 B']_d = \max \{ h_d - h_{d-1}, 0 \} = 1.
\]
Since this is precisely the codimension of $\sigma_2 (\mathbb X_{2,\lambda})$, we see that under the hypothesis that multiplication by $L_3$ has maximal rank we obtain that indeed $\sigma_2 (\mathbb X_{2,\lambda})$ is a hypersurface in its ambient space.
For arbitrary $n$ and $\ell > \frac{n}{2}$, we will need that the maximal rank property holds sequentially, using the right number of linear forms, until we arrive at $n$ variables. In the following sections we make use of this idea to give a general formula for the dimensions of the secant varieties (see Theorem \ref{thm: gen dim formula} and Conjecture \ref{conj:main conj}), and as a consequence we describe the defective cases, assuming that suitable maximal rank properties hold. In many cases we know for different reasons that these maximal rank properties do hold, and in those cases we obtain unconditional (not conjectural) formulas.
Given the many special cases and the seemingly disparate results covering them that have been found up to now, we were astonished to find a simple unifying principle that produces a single conjectural formula for the exact dimension of the secant variety for any given $n$, $\lambda$ and $\ell$.
\section{Improper Intersections and Lefschetz properties}
\label{sec:improper inters}
We now consider the case in which the $\ell$ varieties $V(I_{(j)})$ ($1 \le j \le \ell$)
that determine tangent spaces at $\ell$ general points to $\mathbb X_{n-1, \lambda}$ intersect improperly. By Proposition \ref{prop:proper intersection}(b), their intersection is proper if $2\ell \le n$. However, if $2\ell > n$ then we will see that the intersection is improper, so in particular the diagonal trick from intersection theory becomes more difficult to apply. To make up for this, we will use Lefschetz properties as formally introduced in \cite{HMNW} in order to determine $\dim_{\Bbbk} [S/(I_{(1)} + \cdots + I_{(\ell)})]_d$. Because we are dealing with general forms, these properties are known in some cases, and we conjecture them in the remaining cases.
More precisely, we conjecture that a general artinian reduction of the coordinate ring of the join of $V(I_{(1)}),\ldots,V(I_{(\ell)})$ has enough Lefschetz elements (see Conjecture \ref{conj:WLP} below).
We continue to use the notation introduced above. In particular,
\[
B = S/I_{(1)} \otimes_{\Bbbk} \cdots \otimes_{\Bbbk} S/I_{(\ell)} \cong T/\tilde{I}
\]
is the coordinate ring of the join of $V(I_{(1)}^{(e)}),\ldots,V(I_{(\ell)}^{(e)})$ in $\mathbb{P}^{\ell n - 1}$.
As above, let $\mathcal{L}$ be a set of $(\ell -1) n$ general linear forms in $T$. Let $\mathcal{L}' \subset \mathcal{L}$ be a subset consisting of $\min \{\ell (n-2), (\ell - 1) n\}$ such forms. Thus,
\[
\mathcal{L}' = \begin{cases}
\mathcal{L} & \text{if } \ell \le \frac{n}{2}\\
\subsetneqq \mathcal{L} & \text{if } \ell > \frac{n}{2}
\end{cases}.
\]
Then there is the following useful observation.
\begin{prop}
\label{prop:improper inters}
If $2\ell \ge n$, then:
\begin{itemize}
\item[(a)] The linear forms in $\mathcal{L}'$ form a $B$-regular sequence and $\dim B/\mathcal{L}' B = 0$.
\item[(b)] $\codim (I_{(1)}+\cdots+I_{(\ell)})= n$, and hence the varieties $V(I_{(1)}),\ldots,V(I_{(\ell)})$ intersect improperly if $2\ell > n$.
\end{itemize}
\end{prop}
\begin{proof}
By Theorem \ref{thm:join as tensor product}, $B$ is Cohen-Macaulay of dimension $\ell (n-2)$. Hence, Claim (a) follows by the generality of the linear forms in $\mathcal{L}'$.
Part (a) shows in particular that $\dim B/\mathcal{L} B = 0$. Hence Lemma \ref{sop} gives
\[
\codim (I_{(1)}+\cdots+I_{(\ell)})= n,
\]
and we are done.
\end{proof}
By Lemma \ref{sop}, we are interested in the Hilbert function of $B/\mathcal{L} B$. We know the Hilbert function of $B$, and thus the Hilbert function of its general artinian reduction $B/\mathcal{L}' B$ by Proposition \ref{prop:improper inters}(a). However, if $\ell > \frac{n}{2}$, then $\mathcal{L}' \neq \mathcal{L}$.
Recall that a linear form $L$ is a non-zerodivisor of a graded algebra $A$ if the multiplication by $L$ on $A$ is injective. If $A \neq 0$ is artinian, this cannot be true. However, one may hope that this multiplication still has maximal rank. This has been codified in \cite{HMNW} as follows:
\begin{defn}
\label{def:WLP}
Let $A = S/J$ be an artinian graded $\Bbbk$-algebra. Then $A$ is said to have the {\em Weak Lefschetz Property} if, for each integer $i$, multiplication
by a general linear form $L \in A$ from $[A]_i$ to $[A]_{i+1}$ has maximal
rank. In this case, the form $L$ is called a {\em Lefschetz element} of $A$.
We say $A$ has the {\em Strong Lefschetz Property} if, for all $i$ and $e$, multiplication by $L^e$ from $[A]_i$ to $[A]_{i+e}$ has maximal rank.
\end{defn}
The first systematic study of these properties in this generality was carried out in \cite{HMNW}. In particular, the Hilbert functions of algebras with either
the Weak Lefschetz Property\ or Strong Lefschetz Property\ were classified there, and a sharp bound was given on the possible graded Betti numbers. Thus, the presence of these
properties leads to strong restrictions on the possible invariants. Many natural families of algebras are expected to have a Lefschetz property. However, it is typically very difficult to establish this.
We refer to \cite{HMMNWW} and \cite{wlp-survey} for further information and results.
There is a useful numerical characterization of Lefschetz elements. To state it we need some notation.
\begin{defn}
\label{def:positive part series}
Let $\sum_{i \ge 0} a_i t^i$ be a formal power series, where $a_i \in \mathbb{Z}$. Then we define an associated power series with non-negative coefficients by
\[
\left | \sum_{i \ge 0} a_i t^i \right |^+ = \sum_{i \ge 0} b_i t^i,
\]
where
\[
b_i = \begin{cases}
a_i & \text{if } a_j > 0 \text{ for all } j \le i \\
0 & \text{otherwise}.
\end{cases}
\]
\end{defn}
\begin{lem}
\label{lem:char Lefschetz element}
Let $A$ be a standard artinian graded algebra, and let $L \in A$ be a linear form. Then the following conditions are equivalent:
\begin{itemize}
\item[(a)] $L$ is a Lefschetz element of $A$.
\item[(b)] The Hilbert function of $A/LA$ is given by
\[
\dim_{\Bbbk} [A/L A]_i = \max \{ 0,\; \dim_{\Bbbk} [A]_i - \dim_{\Bbbk} [A]_{i-1} \} \quad \text{ for all integers } i.
\]
\item[(c)] The Hilbert series of $A/LA$ is
\[
\mathbf{HS} (A/LA) = \left | (1-t) \cdot \mathbf{HS} (A) \right |^+.
\]
\end{itemize}
\end{lem}
This is immediate from the definitions.
It is natural to ask if an algebra has the Weak Lefschetz Property repeatedly, using more than one linear form. This notion was first introduced by Iarrobino (according to the introduction of \cite{HW}) and formalized in \cite{HW} and \cite{Constantinescu}.
\begin{defn}
\label{def:k-WLP}
An artinian standard graded $\Bbbk$-algebra $A $ is said to have the {\em $k$-Weak Lefschetz Property} (denoted $k$-WLP) if either $k=0$, or $k>0$ and there are linear forms $L_1,\ldots,L_k \in A$ such that $L_i$ is a Lefschetz element of $A/(L_1,\ldots,L_{i-1}) A$ for all $i = 1,\ldots,k$. In this case, $\{L_1,\ldots,L_k\}$ is called a {\em $k$-Lefschetz set} of $A$.
\end{defn}
By definition, every artinian algebra has the $0$-WLP. Moreover, if an algebra $A$ has the $k$-WLP, then a set of $k$ general linear forms is a $k$-Lefschetz set of $A$. Since all quotients of polynomial rings of at most two variables have the Weak Lefschetz Property\ by \cite{HMNW}, the conditions $(n-2)$-WLP and $n$-WLP are equivalent for quotients of $S$.
Using Lemma \ref{lem:char Lefschetz element}, the above property can be restated as follows.
\begin{lem}
\label{lem:char k-WLP}
An artinian standard graded algebra $A$ has the $k$-WLP if and only if there are linear forms $L_1,\ldots,L_k \in A$ such that
\[
\mathbf{HS} (A/(L_1,\ldots,L_k) A) = \left | (1-t)^i \cdot \mathbf{HS} (A) \right |^+ \quad \text{ for all } i \le k.
\]
\end{lem}
\begin{proof}
This follows by combining Lemma \ref{lem:char Lefschetz element} and \cite[Lemma 4]{froeberg}.
\end{proof}
We now want to use these concepts to determine the dimension of secant varieties $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$, using $B/\mathcal{L} B$. Since the linear forms in $\mathcal{L}$ are general, it is reasonable to ask if, in the case where $\ell > \frac{n}{2}$, the linear forms in $\mathcal{L} \setminus \mathcal{L}'$ form a Lefschetz set of the general artinian reduction $B/\mathcal{L}' B$. We illustrate the usefulness of this property.
\begin{ex}
\label{exa:WLP}
Consider the case $n = 4$, $\ell = 3$, and $\lambda = [3,2,2]$, so $d = 7$.
Then Remark \ref{formula} gives
\[
\mathbf{HS} (B) = \frac{(1 - t^4 - 2 t^5 + 2t^7)^3}{(1-t)^{12}}.
\]
Since $|\mathcal{L}'| = 6$, Proposition \ref{prop:proper intersection} implies
\begin{align*}
\mathbf{HS} (B/\mathcal{L}' B) & = \frac{(1 - t^4 - 2 t^5 + 2t^7)^3}{(1-t)^6} \\
& = 8 t^{15}+48 t^{14}+144 t^{13}+292 t^{12}+456 t^{11}+588 t^{10}+646 t^9+612 t^8 \\
& \quad +504 t^7+363 t^6+228 t^5+123 t^4+56 t^3+21 t^2+6 t+1\\
& = \sum_{i = 0}^{15} c_i t^i
\end{align*}
Note that the passage from $B$ to $C= B/\mathcal{L}' B$ corresponds to intersecting the join of $V(I_{P_1}), V(I_{P_2})$, and $V(I_{P_3})$ properly with a linear subspace of codimension three. However, $C$ is artinian, and thus any further hyperplane sections correspond to improper intersections.
We have checked by computer that $C$ has the $2$-WLP, so let $L_1, L_2$ be a $2$-Lefschetz set. Then, we may assume that $\mathcal{L} = \mathcal{L}' \cup \{L_1, L_2\}$, and Lemma \ref{sop} gives
\[
\mathbf{HS} (A) = \mathbf{HS} (B/\mathcal{L} B) = \mathbf{HS} (C/(L_1, L_2) C) = \left | (1-t)^2 \cdot \mathbf{HS} (C) \right |^+.
\]
We compute this in two steps. First, we get
\begin{align*}
\mathbf{HS} (C/L_1 C) & = \sum_{i \ge 0} \max \{0, c_i - c_{i-1} \} t^i \\
& = 34 t^9+108 t^8+141 t^7+135 t^6+105 t^5+67 t^4+35 t^3+15 t^2+5 t+1 \\
& = \sum_{i = 0}^{9} b_i t^i.
\end{align*}
Thus, we obtain
\begin{align*}
\mathbf{HS} (A) = \mathbf{HS} (C/(L_1, L_2) C) & = \sum_{i \ge 0} \max \{0, b_i - b_{i-1} \} t^i \\
& = 6 t^7+30 t^6+38 t^5+32 t^4+20 t^3+10 t^2+4 t+1.
\end{align*}
In particular, the secant variety $\sigma_{3} (\mathbb X_{3, \lambda})$ has codimension $\dim_{\Bbbk} [A]_d = 6$ in its ambient space. Hence, $\sigma_{3} (\mathbb X_{3, \lambda})$ is non-defective of dimension 113 and does not fill its ambient space.
\end{ex}
Computer experiments suggest that a similar analysis can always be carried out. Thus, we conjecture:
\begin{conj}[WLP-Conjecture]
\label{conj:WLP}
The algebra $B/\mathcal{L}' B$ has the $k$-WLP for $k = \max \{0, \, 2 \ell - n\}$.
\end{conj}
Notice that, by definition, the WLP-Conjecture is true if $\ell \le \frac{n}{2}$.
In Section \ref{sec:r is 2} we will see that in the case $r = 2$ this conjecture is closely related to a well-known conjecture by Fr\"oberg, lending additional evidence to the WLP-Conjecture.
Here we show that if true, the WLP-Conjecture allows us to extend Theorem \ref{thm:dim formula} to $\ell$ with $2\ell > n$. In order to express this we need more notation.
\begin{defn}
\label{def:integers a-j}
Let $\lambda = [d_1,\ldots,d_r]\vdash d$ be a partition with $r \ge 2$ , and let $\ell$ and $n$ be positive integers. For $j = 0,\ldots,d$, define integers $a_j = a_j (\ell, n, \lambda)$ by
\begin{align*}
a_j = & \binom{j+n-1}{n-1} - \ell \sum^r_{i=1} \binom{j + d_i - d+ n-1}{n-1} + (r-1) \ell \binom{j}{d} \\
& + \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{j - k (d - d_1)+ n-1}{n-1} \\[1ex]
& + \binom{\ell}{2} \binom{j + 2 d_2 - 2 d + n-1}{n-1} + \ell (\ell -1) \binom{j + d_1 + d_2 - 2 d+ n-1}{n-1}.
\end{align*}
\end{defn}
Observe that $a_j (\ell, n, \lambda) > 0$ if $0 \le j < s = d_2 + \cdots + d_r$ as, for example, $\binom{j + d_i - d+ n-1}{n-1} = 0$ in this case.
Now we explain the meaning of the numbers $a_j(\ell,n,\lambda)$.
\begin{thm}
\label{thm:explain a-j}
Assume that the WLP-Conjecture is true for some $\ell, n$, and $\lambda$. Let $P_1, \ldots , P_\ell$ be general points on $\mathbb X_{n-1, \lambda}$, and set $A = S/(I_{P_1}+\cdots +I_{P_\ell})$.
If $i \le d$ is a non-negative integer
then
\[
\dim_{\Bbbk} [A]_i = \begin{cases}
0 & \text{if $a_j \le 0$ for some $j$ with $0 \le j \le i$} \\
a_i > 0 & \text{otherwise}.
\end{cases}
\]
In particular, if $a_j (\ell, n , \lambda) > 0$ for all $j = 0,\ldots,i-1$ and $a_i(\ell,n,\lambda) \geq 0$, then
\[
a_i (\ell, n , \lambda) = \dim_{\Bbbk} [A]_i.
\]
\end{thm}
\begin{proof}
To simplify notation set $a_j = a_j (\ell, n , \lambda)$.
First consider the case where $2\ell \le n$. Then the proof of Theorem \ref{thm:dim formula} gives, for all $i \le d$,
\[
\dim_{\Bbbk} [A]_i = a_i,
\]
and so the conclusion holds without the hypothesis that $a_j>0$ for $j < i$. Note that, in this case, $a_j \geq 0$ for all $j\leq i$.
Now assume that $2\ell > n$. We have seen in Remark \ref{formula} that
\begin{align*}
\mathbf{HS}(B) & = \left [ \mathbf{HS}(S/I_{P_1}) \right]^\ell \\[1ex]
& = \frac{1}{(1-t)^{\ell n}} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell .
\end{align*}
Since the elements of $\mathcal{L}'$ provide a regular sequence in $B$ of length $\ell (n-2)$ by Proposition \ref{prop:improper inters}, we get
\[
\mathbf{HS} (B/ \mathcal{L}' B) = \frac{1}{(1-t)^{2 \ell}} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell .
\]
Hence Lemmas \ref{sop} and \ref{lem:char k-WLP} together with the WLP-Conjecture give
\begin{equation}
\label{eq:Hilb series predicted by WLP}
\mathbf{HS} (A) = \left | \frac{1}{(1-t)^{n}} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell \; \right |^+.
\end{equation}
Define integers $b_j$ by
\[
\sum_{j \ge 0} b_j t^j = \frac{1}{(1-t)^{n}} \left [ 1-\sum^r_{i=1} t^{d-d_i} +(r-1)t^d \right] ^\ell.
\]
Then computations as in the proof of Theorem \ref{thm:dim formula} provide for $j \le d$,
\begin{align*}
b_j = \ & \binom{j+n-1}{n-1} - \ell \sum^r_{i=1} \binom{j + d_i - d+ n-1}{n-1} + (r-1) \ell \binom{j}{d} \\
& \ + \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{j - k (d - d_1)+ n-1}{n-1} \\[1ex]
& \ + \binom{\ell}{2} \binom{j + 2 d_2 - 2 d + n-1}{n-1} + \ell (\ell -1) \binom{j + d_1 + d_2 - 2 d+ n-1}{n-1} \\
= & \ a_j.
\end{align*}
Thus, we conclude for all non-negative integers $i \le d$
\[
\dim_{\Bbbk} [A]_i = \begin{cases}
0 & \text{if $a_j \le 0$ for some $j$ with $0 \le j \le i$} \\
a_i > 0 & \text{otherwise}.
\end{cases}
\]
\end{proof}
Notice that when $ 2\ell < n$ the intersection of the varieties determining tangent spaces to $\mathbb X_{n-1, \lambda}$ is non-empty, and the $a_j(\ell,n,\lambda)$ give its Hilbert function. When $ 2\ell = n$, the intersection of the varieties determining tangent spaces to $\mathbb X_{n-1, \lambda}$ becomes empty, but this is still a proper intersection and so the methods of Section 3 continue to apply and result in the values given by the $a_j(\ell,n, \lambda)$.
As soon as $2\ell > n$, however, this intersection remains empty but becomes improper. Nevertheless, the $a_j(\ell,n, \lambda)\geq 0$ essentially provide the Hilbert function of the ``algebraic intersection," i.e. give the Hilbert function of $S/(I_{P_1}+\cdots +I_{P_\ell})$, as formalized in the previous result.
We now have the following extension of Theorem \ref{thm:dim formula}:
\begin{thm}
\label{thm: gen dim formula}
Let $\lambda = [d_1,\ldots,d_r]\vdash d$ be a partition with $r \ge 2$. Assume that the WLP-Conjecture is true for some $\ell$ and $n$. Then:
\begin{itemize}
\item[(a)] The secant variety $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ does not fill its ambient space if and only if
\[
a_j (\ell, n , \lambda) > 0 \quad \text{ for all } j = 0,\ldots,d.
\]
\item[(b)] If $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ does not fill its ambient space, then it has dimension
\begin{align*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = & \ \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \\[1ex]
& \hspace*{0.3cm} {\displaystyle - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1) (d_2 + \cdots + d_r) + n-1}{n-1} } \\[1ex]
& \hspace*{0.3cm} {\displaystyle - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} }
\end{align*}
\end{itemize}
\end{thm}
\begin{proof}
Using Theorem \ref{thm:explain a-j} and
\[
\dim \sigma_{\ell} (\mathbb X_{n-1,\lambda}) = \binom{d+n-1}{n-1} - 1 - \dim_{\Bbbk} [A]_d,
\]
this follows from a computation as in the end of the proof of Theorem \ref{thm:dim formula}.
\end{proof}
\begin{rem}
\label{rem:comp with {CGGS}}
(i) The argument in the proof of Theorem \ref{thm:explain a-j} shows more generally that the WLP-Conjecture allows us to determine the Hilbert function of the ring $A$ in every degree. However, for finding the dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$, it is enough to know this Hilbert function in degree $d$ only. Thus, even if the WLP-conjecture is not true, it is possible that Conjecture \ref{conj:main conj} is correct.
(ii) As noted above, the WLP-Conjecture is true if $2\ell \leq n$. Hence, Theorem \ref{thm: gen dim formula} shows Conjecture \ref{conj:main conj} is true if $2\ell \le n$, thus proving Theorem \ref{thm: main thm - intro}(a). We will establish further instances of Conjecture \ref{conj:main conj} in Section \ref{sec:r is 2}.
(iii) Complete results for the dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ have been previously obtained only if $n=3$, $\lambda = [1,\ldots,1]$ in \cite{A}, or $n=3$, $\ell = 2$ in \cite{CGGS}. Both results confirm Conjecture \ref{conj:main conj}. Moreover, if $\lambda = [1,\ldots,1]$ and $d \ge 3$, then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective, as predicted in Conjecture \ref{conj:defectivity conj}. The case $\lambda = [1,1]$ is covered by Theorem \ref{thm: main thm - intro}. The case $\ell = 2$ was discussed in Section \ref{sec:secant lines}.
Thus, Conjecture \ref{conj:main conj} presents a unified formula for $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})$ in all cases. It is consistent with all the known results that we have checked.
\end{rem}
We explore some consequences of our main conjecture, Conjecture \ref{conj:main conj}.
As we show in our next result, Conjecture \ref{conj:defectivity conj}(a)
is an immediate consequence of Conjecture \ref{conj:main conj}
and thus holds in the many cases for which we establish Conjecture \ref{conj:main conj}.
The situation for Conjecture \ref{conj:defectivity conj}(b) is more complicated.
For certain choices of the parameters $n$, $\ell$ and $\lambda$, our next result shows
that Conjecture \ref{conj:defectivity conj}(b) is true while for some others
it shows that Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
In the remaining cases, for each $d_2 \geq \cdots\geq d_r>0$ and $\ell \geq n$,
it shows that there are at most finitely many cases,
namely $s =d_2 + \cdots + d_r \leq d_1 < (n-1)(s-1)$, for which
we do not know either that
Conjecture \ref{conj:defectivity conj}(b) is true or that
Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
For these cases we have run numerical tests based on Proposition \ref{prop:non-defective}, as discussed in more detail below,
which support our expectation that Conjecture \ref{conj:main conj} implies
Conjecture \ref{conj:defectivity conj}(b) in these cases also.
\begin{comment}
\begin{prop}
\label{prop:non-defective} Consider $d, \lambda, \ell$ such that the conclusion of Conjecture \ref{conj:main conj} is true.
Then the following statements hold:
\begin{itemize}
\item[(a)] If $d_1 < d_2 + \cdots + d_r$ (and thus $r \ge 3$), then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective.
\item[(b)] If $d_1 \ge d_2 + \cdots + d_r$ and $\ell \le n$, then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is defective if and only if it does not fill its ambient space.
\end{itemize}
\end{prop}
\begin{proof}
(a) The assumptions imply that the second and third lines of the formula in Conjecture \ref{conj:main conj}(b) are zero, so
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 =
\min \left \{ \binom{d+n-1}{n-1} - 1, \; \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \right \},
\]
which is the expected dimension.
(b) By definition, $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective if it fills its ambient space.
Assume $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ does not fill its ambient space. Then Conjecture \ref{conj:main conj} gives
\begin{eqnarray*}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})
& = & \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \\\
&& - \sum_{j=2}^\ell (-1)^j \binom{\ell}{j} \binom{d_1 - (j-1) (d_2 + \cdots + d_r) + n-1}{n-1} \\
& & - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} .
\end{eqnarray*}
Let $\Syz$ be the first syzygy module of a complete intersection in $S$ that is generated by $\ell \le n$ forms of
degree $d-d_1 = d_2 +\ldots + d_r$. We observed in Remark \ref{rem:syzygy interpretation} that
\[
\sum_{j=2}^\ell (-1)^j \binom{\ell}{j} \binom{d_1 - (j-1) (d_2 + \cdots + d_r) + n-1}{n-1} = \dim_{\Bbbk} [\Syz]_d.
\]
Hence we get
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) \le \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 - \dim_{\Bbbk} [\Syz]_d.
\]
The Koszul resolution shows that the initial degree of $\Syz$ is $2 (d-d_1)$. Since $\Syz$ is torsion free, it
follows that $[\Syz]_d \neq 0$ if and only if $d \ge 2 (d - d_1)$, which is equivalent to
\[
d_1 \ge d - d_1 = d_2 +\ldots + d_r.
\]
Therefore our assumption gives
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) < \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1,
\]
and we are done.
\end{proof}
\end{comment}
\begin{prop}\label{prop:non-defective}
As usual, let $n\geq 3$, $r\geq 2$, $N=\binom{n+d-1}{n-1}$ and
$\lambda=[d_1,\ldots,d_r]$, where $d = d_1 + s$ and $s = d_2 + \dots + d_r$.
\begin{itemize}
\item[(a)] Assume $d_1 < s$ (and thus $r \ge 3$)). Then
Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(a)
(i.e., that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective).
\item[(b)] Now assume $d_1 \geq s$.
\begin{itemize}
\item[(i)] If $2\ell \leq n$, then Conjecture \ref{conj:defectivity conj}(b) is true (i.e., either $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$
fills its ambient space, $\P^{N-1}$, or it is defective).
\item[(ii)] If $\frac{n}{2} < \ell \leq n$, then Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
\item[(iii)] If $d_1 < 2s$, then Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
\item[(iv)] If $n\leq\ell$ and $(n-1)(s-1)\leq d_1$, then $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space
(and hence Conjecture \ref{conj:defectivity conj}(b) is true).
\end{itemize}
\item[(c)] If $\ell\geq \binom{s+n-1}{n-1}$, then Conjecture \ref{conj:main conj} is true and $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space
(hence Conjecture \ref{conj:defectivity conj} is true).
\item[(d)]
If $\ell\geq \binom{s+n-1}{n-1}/t$, where $t$ is the number of occurrences of $d_1$ in $\lambda$,
then Conjecture~\ref{conj:main conj} implies that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space (and
hence, for such $\ell$, if Conjecture~\ref{conj:main conj} is true, then so is
Conjecture \ref{conj:defectivity conj}.)
\end{itemize}
\end{prop}
\begin{proof}
(a) The assumptions imply that the second and third lines of the formula in Conjecture \ref{conj:main conj}(b) are zero, so
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 =
\min \left \{ \binom{d+n-1}{n-1} - 1, \; \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \right \},
\]
which is the expected dimension.
(b)(i) This follows from Theorem \ref{thm:dim formula}.
(b)(ii) By definition, $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is not defective if it fills its ambient space, $\P^{N-1}$,
so assume $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ does not fill its ambient space. Then
$\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})<N-1$ and Conjecture \ref{conj:main conj}(b) gives
\begin{equation}\label{Conj1.1b formula}
\begin{split}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) =\ & \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 \\
& - \sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1)s + n-1}{n-1} \\
& - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} .
\end{split}
\end{equation}
Let $\Syz$ be the first syzygy module of a complete intersection in $S$ that is generated by $\ell \le n$ forms of
degree $d-d_1 = d_2 +\ldots + d_r$. We observed in Remark \ref{rem:syzygy interpretation} that
\[
\sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d_1 - (k-1)s + n-1}{n-1} = \dim_{\Bbbk} [\Syz]_d.
\]
Hence we get
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) \le \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 - \dim_{\Bbbk} [\Syz]_d.
\]
The Koszul resolution shows that the initial degree of $\Syz$ is $2 (d-d_1)$. Since $\Syz$ is torsion free, it
follows that $[\Syz]_d \neq 0$ if and only if $d \ge 2 (d - d_1)$, which is equivalent to
\[
d_1 \ge d - d_1 = s.
\]
Therefore our assumption gives
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) < \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1,
\]
so $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})< \hbox{exp.dim } \sigma_{\ell} (\mathbb X_{n-1,\lambda})$, and we are done.
(b)(iii) As in the proof of (b)(ii), if $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ does not fill its ambient space
(and so $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})<N-1$) we must show that
it is defective. Put
\[
g=\sum_{j=2}^\ell (-1)^j \binom{\ell}{j} \binom{d_1 - (j-1)s + n-1}{n-1}.
\]
If $s \le d_1 < 2s$, then $g >0$. Thus, the summation in display \eqref{Conj1.1b formula} is positive, but it is subtracted so we have
$\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) < \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1$, and hence
$\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is defective. Thus in the presence of the restriction on $d_1$, Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
(b)(iv) To see that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space for $d_1\geq (n-1)(s-1)$ and
$\ell\geq n$, it is enough to do so for $\ell=n$. Take $\ell$ general points $P_j$ on $\mathbb X_{n-1, \lambda}$. Each ideal $I_{P_j}$
contains a minimal generator of degree $d - d_1 = s$. Thus, by genericity, the ideal $I=I_{P_1}+\cdots+I_{P_\ell}$
contains a complete intersection generated by $n$ forms of degree $s$. The socle degree of this complete intersection is $n(s-1)$. Thus $[R/I]_d=0$ if
$d_1+s=d>n(s-1)$; i.e., if $d_1\geq (n-1)(s-1)$. It follows that for these $\ell$ and $d_1$ we get that Conjecture \ref{conj:defectivity conj}(b) is true.
(c) Note that $a_s = a_s(\ell, n, \lambda) =\binom{s+n-1}{n-1} -t\ell$, hence $a_s\leq0$ for $\ell\geq \binom{s+n-1}{n-1}/t$, and so also for $\ell \geq \binom{s+n-1}{n-1}$. So with the latter hypothesis, to prove Conjecture \ref{conj:main conj} we must show that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space. Notice that $[S]_s$ has a basis consisting of monomials, hence a basis of forms each of which factors as a product of forms of degrees $d_2,\dots,d_r$. Thus we can find points $P_1,\dots,P_\ell$ for which $I=I_{P_1}+\cdots+I_{P_\ell}$ spans $[S]_s$, so also in degree $d$ we have $[I]_d = [S]_d$. The same is then true for a general choice of $\ell$ points, and so $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ indeed fills its ambient space. But then for these $\ell$, all parts of Conjecture \ref{conj:defectivity conj} are automatically true as well.
(d) Finally, assume that $\ell\geq \binom{s+n-1}{n-1}/t$. Since $a_s \leq 0$, Conjecture \ref{conj:main conj} would imply that
$\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space and so
Conjecture \ref{conj:defectivity conj} would hold.
(We also note that the bound $\ell\geq \binom{s+n-1}{n-1}/t$ is sharp,
since $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$
does not always fill its ambient space for $\ell<\binom{s+n-1}{n-1}/t$, as we see for
$\sigma_{\ell} (\mathbb X_{n-1, [1,1]})$ by Theorem \ref{thm:defectivity - intro}(c).)
\end{proof}
\begin{rem}\label{NumericalTestingRemark}
Proposition \ref{prop:non-defective}
is the basis for numerical tests that support our belief that all of the cases left open
do in fact follow from Conjecture \ref{conj:main conj}. In these cases, we have $n < \ell$ and $2s\leq d_1<(n-1)(s-1)$. As noted above, Conjecture \ref{conj:main conj} predicts: If $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ does not fill its ambient space, then
\begin{equation*}\begin{split}
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) =\ & \ell \cdot \dim \mathbb X_{n-1, \lambda} + \ell -1 - g \\
& - \binom{\ell}{2} \binom{2 d_2 - d + n-1}{n-1} - \ell (\ell -1) \binom{d_1 + d_2 - d+ n-1}{n-1} .
\end{split}
\end{equation*}
Thus, Conjecture \ref{conj:defectivity conj}(b) follows if $g > 0$.
We have used Macaulay2 \cite{Macaulay2}
to check for all cases satisfying the above restrictions with
$n,\ell,s\leq 60$ that Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b).
There were 57,345,933 such cases. Note that typically for each $s$ and $d_1$, there are many possible
partitions of $d=d_1+s$; thus for each of the 57,345,933 cases for which $g\leq0$, we merely checked that
$\ell\dim(\mathbb X_{n-1,[d_1,1,\ldots,1]})+\ell\geq N$, and hence by Proposition \ref{prop:non-defective}(b)(iii)
we see that Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b)
for each case when $\lambda=[d_1,1,\ldots,1]$. But by
Corollary \ref{cor:dim comparison}
this means
Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(b)
also for all other partitions $\lambda$ of $d=d_1+s$ for each of these 57,345,933 cases.
\end{rem}
\begin{comment}
\begin{lem}\label{littleLemma}
Let $n\geq 3$ and let $\lambda=[d_1,d_2,\ldots,d_r]$ and $[d_1,1,\cdots,1]$ be partitions of
$d=d_1+\cdots+d_r$ with $r\geq 2$ and $d_1\geq\cdots\geq d_r>0$.
Then $\dim \mathbb X_{n-1,\lambda}\geq\dim \mathbb X_{n-1,[d_1,1,\cdots,1]}$.
\end{lem}
\begin{proof} Let $s=d_2+\cdots+d_r$. Let $\nu=n-1$.
What we want to show is equivalent to
$\binom{d_2+\nu}{\nu}+\cdots+\binom{d_r+\nu}{\nu}-(r-1)\geq s\binom{\nu+1}{\nu}-(s-1)$.
Thus it suffices to show $\binom{d_2+\nu}{\nu}+\cdots+\binom{d_r+\nu}{\nu}\geq s\binom{\nu+1}{\nu}$ for $\nu\geq2$, and this
is an easy recursive consequence of the straightforward fact that
$\binom{a+1+\nu}{\nu}\geq \binom{a+\nu}{\nu}+\binom{\nu+1}{\nu}$ holds for $a\geq1$ and $\nu\geq2$.
(To justify $\binom{a+1+\nu}{\nu}\geq \binom{a+\nu}{\nu}+\binom{\nu+1}{\nu}$, use $\binom{a+1+\nu}{\nu}=\binom{a+\nu}{\nu}+\binom{a+\nu}{\nu-1}$
to rewrite it as $\binom{a+\nu}{\nu-1}\geq \nu+1$ or as
$\nu\cdots (a+\nu)\geq (a+1)!(\nu+1)$ hence as
$\nu(\nu+2)\cdots (a+\nu)\geq 2\cdot3\cdots (a+1)$
which is clearly true for $\nu\geq2$.)
\end{proof}
\end{comment}
\begin{comment}
\begin{rem}
\label{rem:cons small ell}
Proposition \ref{prop:non-defective} shows that Conjecture \ref{conj:main conj} implies Conjecture \ref{conj:defectivity conj}(a) and that it also gives Conjecture \ref{conj:defectivity conj}(b), provided $\ell \le n$.
\end{rem}
\end{comment}
\begin{cor}\label{Thm1.2c}
Let $3\leq n\leq \ell\leq 1+\frac{d_1+n-1}{s}$ with $\lambda=[d_1,\ldots,d_r] \vdash d$, $r \geq 3$, and $s=d_2+\cdots+d_r$.
Then Conjecture \ref{conj:main conj} is true for such $n$, $\ell$ and $\lambda$.
\end{cor}
\begin{proof}
Conjecture \ref{conj:main conj}(a) asserts that $\sigma_\ell(\mathbb{X}_{n-1,\lambda})$ fills its ambient space
if and only if $a_j(\ell,n,\lambda)$
is not positive for some integer $j$ with $s\leq j\leq d$,
while Conjecture \ref{conj:main conj}(b) applies only when $\sigma_\ell(\mathbb{X}_{n-1,\lambda})$ does not fill its ambient space.
Since $d_1 \geq (n - 1)(s-1)$ is equivalent to $1 + \frac{d_1+n-1}{s} \geq n$, Proposition \ref{prop:non-defective} implies $\sigma_\ell(\mathbb{X}_{n-1,\lambda})$ fills its ambient space. Thus
Conjecture \ref{conj:main conj} is true if we show $a_d(\ell,n,\lambda)\leq 0$.
Recall the identity
\[
\sum_{k=2}^\ell (-1)^k \binom{\ell}{k} \binom{d - k s + n-1}{n-1}=0.
\]
(See formula 10.13 of \url{http://www.math.wvu.edu/~gould/Vol.4.PDF},
where $n,k,r,y,x$ in 10.13 become, respectively, our
$\ell$, $j$, $n-1$, $d+n-1$ and $-s$, so the assumption $n>r$ in 10.13
becomes $\ell > n-1$ and is thus satisfied. We also note that 10.13 does not assume
that $\binom{a}{b}=0$ when $a<0$, but our assumption $\ell\leq 1+\frac{d_1+n-1}{s}$
is equivalent to $d-\ell s + n-1\geq 0$. This ensures that $d-js+n-1\geq0$ hence
the convention used in 10.13 agrees with our convention that
$\binom{d-js+n-1}{n-1}=0$ when $d-js+n-1<n-1$.)
Using the identity above, we have
$$\sum_{j=2}^\ell(-1)^j\binom{\ell}{j}\binom{d-js+n-1}{n-1}=-\binom{d+n-1}{n-1}+\ell\binom{d_1+n-1}{n-1},$$
and substituting this into the expression for $a_d(\ell,n,\lambda)$
given in Definition \ref{def:integers a-j} we obtain
$$a_d(\ell,n,\lambda)=-\ell\sum_{i>1}\Bigg(\binom{d_i+n-1}{n-1}-1\Bigg)<0.$$
\end{proof}
We are now ready to prove one of the main results of the paper, as mentioned in the introduction.
\begin{proof}[Proof of Theorem \ref{thm:defectivity - intro}]
Part (a) follows from Theorem \ref{general r geq 3}(a).
Part (b) follows from Theorem \ref{general r=2} and Theorem \ref{general r geq 3}(b).
Part (c) follows from Proposition \ref{prop:non-defective}(iv) and Theorem \ref{thm:dim formula}.
\end{proof}
We conclude this section with a case where we can show that the prediction of Conjecture~\ref{conj:main conj} is at least an upper bound.
\begin{prop}
\label{prop:upper bound}
If $2 \ell = n+1$, then the dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is at most the number predicted in Conjecture \ref{conj:main conj}.
\end{prop}
\begin{proof}
Using the above notation, put $A' = B/\mathcal{L}'$. Let $L \in A'$ be a general linear element. Notice that the assumption on $\ell$ gives $|\mathcal{L} | = |\mathcal{L}' | + 1$. Thus, we get
\begin{equation}
\label{eq:upper bound}
\mathbf{HS} (A) = \mathbf{HS} (B/\mathcal{L} B) = \mathbf{HS} (A'/L A') \ge \left | (1-t) \mathbf{HS} (A') \right |^+,
\end{equation}
which means that the comparison is true coefficientwise.
If $a_j (\ell, n, \lambda) \le 0$ for some non-negative $j \le d$, then Conjecture \ref{conj:main conj} says that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space, and so the estimate follows.
Otherwise, Equation \eqref{eq:upper bound} implies
\[
\dim_{\Bbbk} [A]_d \ge a_d (\ell, n, \lambda),
\]
which gives
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = \binom{d + n-1}{n-1} - 1 - \dim_{\Bbbk} [A]_d \le \binom{d + n-1}{n-1} - 1 - a_d (\ell, n, \lambda).
\]
Since, the right-hand side is the formula for $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})$ that is predicted by Conjecture \ref{conj:main conj}, this completes the argument.
\end{proof}
\begin{rem}
\label{rem:upper bound}
Consider an arbitrary graded $\Bbbk$-algebra, and let $L_1, L_2 \in [A]_1$ be two general elements. Then it is \emph{not} necessarily true that
\[
\mathbf{HS} (A/(L_1, L_2) \ge \left | (1-t)^2 \cdot \mathbf{HS} (A') \right |^+.
\]
Thus, the above argument cannot be easily extended to show that Conjecture \ref{conj:main conj} gives an upper bound for $\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})$ for all $\ell \ge \frac{n+1}{2}$. Note however that in the following section we will prove that Conjecture \ref{conj:main conj} does give an upper bound if $r = 2$ by using a different approach.
\end{rem}
\section{Forms with two factors and Fr\"oberg's Conjecture}
\label{sec:r is 2}
In this section we focus on the case $r = 2$, that is, we consider secant varieties to the varieties whose general point corresponds to a product of two irreducible polynomials.
We begin by recalling that the dimension of secant varieties, in case $r = 2$, is related to a famous conjecture of Fr\"oberg (see \cite{CCG:1}). We systematically relate this conjecture to our approach in the previous section. In particular, we will see that Fr\"oberg's Conjecture and the WLP-Conjecture lead to the same prediction for the dimension of the secant variety in the case $r = 2$. This allows us to establish further instances of Conjecture~\ref{conj:main conj}.
Fr\"oberg's Conjecture concerns the Hilbert function of an ideal generated by generic forms. More precisely, it says:
\begin{conj}[Fr\"oberg's Conjecture \cite{froeberg}]
\label{conj:Froeberg}
Let $J \subset S = \Bbbk[x_1,\ldots,x_n]$ be an ideal generated by $s$ generic forms of degrees $e_1,\ldots,e_s$ in $S$. Then the Hilbert series of $S/J$ is
\[
\mathbf{HS} (S/J) = \left | \frac{\prod_{i=1}^s (1 - t^{e_i})}{(1-t)^n} \right |^+ .
\]
\end{conj}
There is an equivalent version of Fr\"oberg's Conjecture that gives a recursion to predict the Hilbert function of such an algebra.
\begin{conj}[Fr\"oberg's Conjecture, recursive version]
\label{conj:Froeberg recursive}
Let $J \subset S = \Bbbk[x_1,\ldots,x_n]$ be an ideal that is generated by generic forms, and let $f \in S$ be a generic form of degree $e$. Then, for all integers $j$,
\[
\dim_{\Bbbk} [S/(J, f)]_j = \max \{0, \; \dim_{\Bbbk} [S/J]_j - \dim_{\Bbbk} [S/J]_{j - e} \}.
\]
\end{conj}
Comparing the latter version with Definition \ref{def:WLP} shows that $S/J$ has the Weak Lefschetz Property\ if the above is true and $e=1$.
We refer to \cite[Proposition 2.1]{MMN3} for further results on the relation between Fr\"oberg's
conjecture and the Lefschetz properties. Here we need only the following observation.
\begin{prop}
\label{prop:Froeberg gives WLP}
Assume Fr\"oberg's Conjecture is true for polynomial rings in up to $n$ variables. If $J \subset S = \Bbbk[x_1,\ldots,x_n]$ is
an ideal that is generated by at least $n$ generic forms, then $S/J$ has the $n$-WLP.
\end{prop}
\begin{proof}
Let $L \in S$ be a generic linear form. Then, as noted above, $L$ is a Lefschetz element for $S/J$.
Since $S/(J, L)$ is isomorphic to a quotient of a polynomial ring in $n-1$ variables modulo an ideal
generated by generic forms in these $(n-1)$ variables, we can use this argument $n$ times.
\end{proof}
Since each quotient of a polynomial ring in at most two variables has the Weak Lefschetz Property\ by \cite{HMNW},
the properties $n$-WLP and $(n-2)$-WLP are equivalent if $n \ge 2$.
We now relate this to the secant varieties of $\mathbb X_{n-1, \lambda}$, where $\lambda$ is a partition with two parts.
In this case, we simplify notation and write $\lambda = [d- k,k]$, where $1 \le k \le \frac{d}{2}$.
Our starting point for the case $r = 2$ is the following observation.
\begin{lem}
\label{lem:dim for r=2}
If $\lambda = [d- k,k]$, then
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, \lambda}) = -1 + \dim_{\Bbbk} [I]_d,
\]
where $I \subset S$ is an ideal generated by $\ell$ generic forms of degree $d-k$ and $\ell$ generic forms of degree $k$.
\end{lem}
\begin{proof}
This follows from Corollary \ref{dim sigma ell} and Proposition \ref{nicer}.
\end{proof}
In the case $r=2$, the definition of the integers $a_j (\ell, n, \lambda)$ becomes somewhat simpler.
\begin{rem}
\label{rem:simples a-j for r is 2}
Assume $\lambda = [d- k,k]$, where $1 \le k \le \frac{d}{2}$. Then
\begin{align*}
a_j (\ell, n, \lambda) = & \binom{j+n-1}{n-1} - \ell \left [ \binom{j + k - d+ n-1}{n-1} + \binom{j - k + n-1}{n-1} \right ] \\
& + \sum_{i=2}^\ell (-1)^i \binom{\ell}{i} \binom{j - i k + n-1}{n-1}
\end{align*}
if $0 \le j < d$, and
\begin{align*}
a_d (\ell, n, \lambda) = & \binom{d+n-1}{n-1} - \ell \left [ \binom{k + n-1}{n-1} + \binom{d - k + n-1}{n-1} \right ] \\
& + \sum_{i=2}^\ell (-1)^i \binom{\ell}{i} \binom{d - i k + n-1}{n-1} + \binom{\ell}{2} \binom{2 k - d +n -1}{n-1} + \ell^2.
\end{align*}
Observe that the penultimate term is zero, unless $k = \frac{d}{2}$.
\end{rem}
Lemma \ref{lem:dim for r=2} allows us to relate Fr\"oberg's Conjecture to our work in the previous sections.
\begin{prop}
\label{prop:Froeberg gives main conj}
Let $\lambda = [d-k, k]$ be a partition with two parts. Then, for each $\ell \ge 2$ and each $n \ge 3$, the value for the dimension of the secant variety in Conjecture \ref{conj:main conj} gives an upper bound for the dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$.
Moreover, if Fr\"oberg's Conjecture is true for
$S$, then, for all $\ell \ge 2$, the variety $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ has the dimension predicted in Conjecture \ref{conj:main conj}.
\end{prop}
\begin{proof}
If $\ell \le \frac{n}{2}$, Conjecture \ref{conj:main conj} is true by Theorem \ref{thm:dim formula}. Thus, we may assume $2 \ell > n$.
With $I$ as in Lemma \ref{lem:dim for r=2}, we have
\[
\dim \sigma_{\ell} (\mathbb X_{n-1,\lambda}) = \binom{d+n-1}{n-1} - 1 - \dim_{\Bbbk} [S/I]_d.
\]
The value predicted for $\dim \sigma_{\ell} (\mathbb X_{n-1,\lambda})$ by Conjecture \ref{conj:main conj}
comes, by Theorem \ref{thm:explain a-j}, from the value of $\dim_{\Bbbk} [S/I]_d$ predicted by the
WLP-Conjecture \ref{conj:WLP}. Thus it is enough to derive from Fr\"oberg's Conjecture \ref{conj:Froeberg}
the same value for $\dim_{\Bbbk} [S/I]_d$ as given by Conjecture \ref{conj:WLP}, and to show that this
value is a lower bound for the actual value.
Using our earlier notation, observe that $B/\mathcal{L}' B \cong U/J$, where $U$ is a polynomial ring in $2 \ell$
variables and $J$ is a complete intersection generated by $\ell$ general forms of degree $k$ and $\ell$
general forms of degree $d-k$. Hence the Hilbert series of $U/J$ is
\[
\mathbf{HS} (U/J) = \frac{(1 - t^{d-k})^\ell (1-t^k)^\ell}{(1-t)^{2 \ell}} .
\]
Lemma \ref{sop} shows that $S/I$ is isomorphic to $B/\mathcal{L} B$, which is obtained from $U/J$ by quotienting by $(2 \ell - n)$
general linear forms. Hence \cite[Theorem]{froeberg} gives
\[
\mathbf{HS} (S/I) \geq \left | \frac{(1 - t^{d-k})^\ell (1-t^k)^\ell(1-t)^{2\ell-n}}{(1-t)^{2\ell}} \right |^+=\left | \frac{(1 - t^{d-k})^\ell (1-t^k)^\ell}{(1-t)^n} \right |^+,
\]
the right hand side of which is exactly the Hilbert series of $S/I$ as predicted by Fr\"oberg's Conjecture \ref{conj:Froeberg}.
Moreover, the WLP-Conjecture \ref{conj:WLP} predicts this same Hilbert series for $S/I$ (see Equation \eqref{eq:Hilb series predicted by WLP}),
and hence equality holds if Fr\"oberg's Conjecture does.
\end{proof}
\begin{rem}
Notice that in Proposition \ref{prop:Froeberg gives main conj} we assumed the correctness of Fr\"oberg's Conjecture only for ideals in $S$. If we assume more, namely that this conjecture is true for all ideals in $\max \{n, \ 2 \ell \}$ variables, then Proposition \ref{prop:Froeberg gives WLP} shows that $B/\mathcal{L}' B$ has the $(2 \ell-n)$-Weak Lefschetz Property. Hence, in this case Theorem \ref{thm: gen dim formula} immediately gives the conclusion of the above proposition.
\end{rem}
We are ready to establish new instances where Conjecture \ref{conj:main conj} holds.
\begin{thm}
\label{thm:dim when froeberg}
Conjecture \ref{conj:main conj} is true if $r=2$ and
\begin{itemize}
\item[(a)] $2\ell \le n+1$ \quad or
\item[(b)] $n = 3$ \quad or
\item[(c)] $\lambda = [1, 1]$, that is, $d = 2$.
\end{itemize}
\end{thm}
\begin{proof}
We use Proposition \ref{prop:Froeberg gives main conj}.
Fr\"oberg's Conjecture is true for forms in at most three variables by a result of Anick \cite{anick}. This gives (b). The conjecture also holds for ideals generated by general linear forms, and thus (c) follows.
Turning to (a), by Theorem \ref{thm:dim formula} it suffices to consider the case where $2 \ell = n + 1$. Then $I$ is generated by $n+1$ general forms in $n$ variables. For such ideals Fr\"oberg's Conjecture is true because complete intersections of general forms have the Strong Lefschetz Property\ (see \cite{stanley, watanabe, RRR, sekiguchi} or \cite{HP}).
\end{proof}
\begin{rem}
With some work, part (a) of Theorem \ref{thm:dim when froeberg} can be shown to be equivalent to Theorem 5.1 of \cite{CCG:1}. There, however, all the cases where $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills its ambient space are enumerated.
\end{rem}
We now begin working out more explicit formulas for some particular partitions as consequences of Proposition \ref{prop:Froeberg gives main conj}. First we consider balanced partitions.
\begin{thm}
\label{thm:balanced}
Consider a balanced partition $\lambda = [\frac{d}{2},\frac{d}{2}]$ and fix $n \geq 3$. Assume that Fr\"oberg's Conjecture holds for the polynomial ring $S = \Bbbk[x_1,\ldots,x_n]$.
Put
\[
\ell_0 = \frac{1}{2} \left [ \binom{\frac{d}{2} + n-1}{n-1} + \frac{1}{2} - \sqrt{\left [ \binom{\frac{d}{2} + n-1}{n-1} + \frac{1}{2} \right ]^2 - 2 \binom{d+n-1}{n-1}} \ \right ].
\]
Then
\begin{multline*}
\dim \sigma_\ell(\mathbb X_{n-1, [\frac{d}{2},\frac{d}{2}]})
= \begin{cases}
\ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1 - 2 \ell (\ell - 1) < \binom{d+n-1}{n-1} - 1 & \text{ if } 2 \le \ell < \ell_0 \\[2ex]
\binom{d+n-1}{n-1} - 1 & \text{ if } \ell_0 \le \ell.
\end{cases}
\end{multline*}
In particular, the secant variety $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is defective if and only if it does not fill its ambient space.
Furthermore, the defect is
\begin{align*}
\delta_\ell & = \begin{cases}
2 \ell (\ell - 1) & \text{ if } 2 \le \ell \le {\displaystyle \frac{\binom{d + n-1}{n-1}}{2 \binom{\frac{d}{2} + n-1}{n-1} - 1}} \\[4ex]
\binom{d+n-1}{n-1} - 1 - \dim \sigma_\ell(\mathbb X_{n-1, [\frac{d}{2},\frac{d}{2}]}) & \text{ if } {\displaystyle \frac{\binom{d + n-1}{n-1}}{2 \binom{\frac{d}{2} + n-1}{n-1} - 1}} < \ell < \ell_0.
\end{cases}
\end{align*}
\end{thm}
\begin{proof}
Put $N = \binom{d+n-1}{n-1}$ and recall that
\[
\dim \mathbb X_{n-1,\lambda} = 2 \cdot \binom{\frac{d}{2}+n-1}{n-1} -2 .
\]
Thus, the expected dimension of $\sigma_\ell (\mathbb X_{n-1,\lambda})$ is
\begin{align*}
\hbox{exp.dim } \sigma_\ell (\mathbb X_{n-1,\lambda}) & = \min \{ N-1, \ell \cdot \dim \mathbb X_{n-1,\lambda} + (\ell -1) \} \\[1ex]
& = \min \left \{ N-1\ , \ 2 \ell \cdot \binom{\frac{d}{2}+n-1}{n-1} - \ell -1 \right \}.
\end{align*}
In particular,
\begin{equation}
\label{eq:est exp dim}
\hbox{exp.dim } \sigma_\ell (\mathbb X_{n-1,\lambda}) = \ell \cdot \dim \mathbb X_{n-1,\lambda} + (\ell -1) \text{ if and only if } 2 \le \ell \le \frac{N}{2 \binom{\frac{d}{2} + n-1}{n-1} - 1}.
\end{equation}
We now will consider various ranges for the value of $\ell$ and use Lemma \ref{lem:dim for r=2}. For the partition $\lambda$, the ideal $I$ is generated by $2 \ell$ general forms of degree $\frac{d}{2}$. Instead of applying Proposition \ref{prop:Froeberg gives main conj} directly, it is more convenient to use the recursive approach (see Conjecture \ref{conj:Froeberg recursive}).
Assume first $2 \le \ell < \frac{n}{2}$. Then Theorem \ref{thm:dim formula} gives that $\sigma_\ell(\mathbb X_{n-1,\lambda})$ does not fill its ambient space and has dimension
\begin{align*}
\dim \sigma_\ell(\mathbb X_{n-1,\lambda}) & = \ell \cdot \dim \mathbb X_{n-1,\lambda} + (\ell-1) - 2 \binom{\ell}{2} - \ell (\ell -1) \\[1ex]
& = \ell \cdot \dim \mathbb X_{n-1,\lambda} + (\ell-1) - 2 \ell (\ell -1).
\end{align*}
This proves the statement if $\ell < \frac{n}{2}$.
Assume now $\frac{n}{2} \le \ell$. In order to simplify notation, set $k = \frac{d}{2}$ and $t = 2 \ell - n \ge 0$.
In this range of $\ell$, $S/I$ is artinian and
\[
\dim_{\Bbbk} [S/I]_k = \max \left \{0, \; \binom{k+n-1}{n-1} - 2 \ell \right \}.
\]
Hence $[S/I]_k = 0$ if $2 \ell \ge \binom{k+n-1}{n-1}$, which implies $[S/I]_d = 0$. It follows that $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills $\mathbb{P}^{N-1}$ for such $\ell$.
We are left to consider $\ell$ such that $\frac{n}{2} \le \ell < \frac{1}{2} \binom{k+n-1}{n-1}$. Notice that this forces $k \ge 2$, that is, $d \ge 4$.
For $i = 0,1,\ldots,t = 2 \ell -n$, let
\[
\begin{array}{rcl}
\mathfrak a_i & = & \hbox{ideal generated by $n+i$ general forms of degree $k$}.
\end{array}
\]
Observe that $\mathfrak a_0$ is a complete intersection and $I = \mathfrak a_{t}$.
Notice that, for all $i$,
\[
\dim_{\Bbbk} [S/\mathfrak a_i]_k = \binom{k+n-1}{n-1} - n-i = \dim_{\Bbbk} [S/\mathfrak a_0]_k - i.
\]
The minimal free resolution of $S/\mathfrak a_0$ has the form
\[
\dots
\rightarrow
S(-d)^{\binom{n}{2}} \\
\rightarrow
S(-k)^n \\
\rightarrow S \rightarrow S/\mathfrak a_0 \rightarrow 0,
\]
where we only display the terms that are non-trivial in degree $d$. This shows
\[
\dim_{\Bbbk} [S/\mathfrak a_0]_d = N - n \cdot \binom{k+n-1}{n-1} + \binom{n}{2}.
\]
Fr\"oberg's Conjecture \ref{conj:Froeberg recursive} predicts, for all $i$,
\[
\dim_{\Bbbk} [S/\mathfrak a_{i+1}]_d = \max \{ \dim_{\Bbbk} [S/\mathfrak a_i]_d - \dim_{\Bbbk} [S/\mathfrak a_i]_k, \ 0 \}.
\]
Hence, we get
\begin{align*}
\dim_{\Bbbk} [S/I]_d & = \max \left \{ 0, \ \dim_\Bbbk [S/\mathfrak a_0]_d - t \cdot \dim_\Bbbk [S/\mathfrak a_0]_k + \binom{t}{2} \right \} \\[1ex]
& = \max \left \{ 0\ , \ \left [ N - n \cdot \binom{k+n-1}{n-1} + \binom{n}{2} \right ] - t \left [ \binom{k+n-1}{n-1} - n \right ] + \binom{t}{2} \right \} \\[1ex]
& = \max \left \{ 0 \ , \ N - 2 \ell \cdot \binom{k+n-1}{n-1} + \binom{n}{2} + t n + \binom{t}{2} \right \} \\[1ex]
& = \max \left \{ 0 \ , \ N - 2 \ell \cdot \binom{k+n-1}{n-1} + 2 \ell^2 - \ell \right \}.
\end{align*}
It follows that
\begin{align*}
\dim \sigma_\ell(\mathbb X_{n-1,\lambda}) & = \min \left \{ N -1 \ , \ 2 \ell \binom{k+n-1}{n-1} + \ell-1 - 2 \ell^2 \right \}
\end{align*}
(note that we subtracted 1 from the dimension of the component of the ideal).
Therefore, for $\ell < \frac{1}{2} \binom{k+n-1}{n-1}$, the variety $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills $\mathbb{P}^{N-1}$ if and only if
\[
N \le 2 \ell \binom{k+n-1}{n-1} + \ell - 2 \ell^2,
\]
which means
\[
\ell \ge \frac{1}{2} \left [ \binom{k+ n-1}{n-1} + \frac{1}{2} - \frac{1}{2} \sqrt{\left [ \binom{k + n-1}{n-1} + \frac{1}{2} \right ]^2 - 2 N} \ \right ] = \ell_0.
\]
An induction on $n \ge 2$ shows that the radicand is at least $\frac{1}{4}$, which also implies
\[
\left \lceil \frac{1}{2} \binom{k+n-1}{n-1} \right \rceil \ge \ell_0.
\]
We conclude that $\sigma_\ell(\mathbb X_{n-1,\lambda})$ fills its ambient space if and only if $\ell \ge \ell_0$ and that
\begin{align*}
\dim \sigma_\ell(\mathbb X_{n-1,\lambda}) & = 2 \ell \binom{k+n-1}{n-1} + \ell-1 - 2 \ell^2 \\
& = \ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1 - 2 \ell (\ell - 1)
\end{align*}
if $\frac{n}{2} \le \ell \le \ell_0$. This concludes finding the dimension of $\sigma_\ell(\mathbb X_{n-1,\lambda})$. Combining the result with Observation \eqref{eq:est exp dim} proves the assertion on the defect.
\end{proof}
Second, we consider the most unbalanced partition of $d$ into two parts. Notice that the following result is true unconditionally. Since the partition $[1, 1]$ has been dealt with in the previous result (see Theorem \ref{thm:dim when froeberg}), there is no harm in assuming $d \ge 3$ in the next statement.
The fact that $\sigma_\ell (\mathbb X_{n-1,[d-1, 1]})$ fills its ambient space if and only if $\ell \geq \ell_0$ was shown in \cite[Proposition 5.6]{CCG:1}. The dimension of $\sigma_\ell (\mathbb X_{n-1,[d-1, 1]})$ can also be found in \cite[Proposition 4.4]{BCGI}. We give a new proof of these facts using our methods.
Note that, in the following theorem, the formula for $\dim(\mathbb X_{n-1,[d-1,1]})$ is simply the specialization of the formula of Conjecture \ref{conj:main conj} to the case at hand.
\begin{thm}
\label{thm:case d-1, 1}
Assume $\lambda = [d-1, 1]$, where $d \ge 3$. Put
\[
\ell_0 =
\min \left \{ \ell \ge \frac{n}{2} \; | \; \ell \in \mathbb{Z} \text{ and } \binom{d - \ell + n - 1}{d} \le \ell (n- \ell) \right \}.
\]
Then $\ell_0 \le n-1$ and
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]}) =
\begin{cases}
\binom{d+n-1}{n-1} - \binom{d+n- \ell - 1}{d} + \ell (n- \ell) - 1 < \binom{d+n-1}{n-1} - 1 & \text{ if } 2 \le \ell < \ell_0 \\[.5ex]
\binom{d+n-1}{n-1} - 1 & \text{ if } \ell_0 \le \ell.
\end{cases}
\]
Moreover, the secant variety $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is defective if and only if it does not fill its ambient space. In this case, the defect is
\begin{align*}
\delta_\ell & = \begin{cases}
\binom{d+n- \ell - 1}{d} + \ell \cdot \binom{d+n-2}{n-1} - \binom{d+n-1}{n-1}+ \ell^2 \ge \ell^2 & \text{ if } 2 \le \ell \le {\displaystyle \frac{\binom{d + n-1}{n-1}}{\binom{d + n-2}{n-1} + n}} \\[4ex]
\binom{d+n-1}{n-1} - 1 - \dim \sigma_\ell(\mathbb X_{n-1, [\frac{d}{2},\frac{d}{2}]}) & \text{ if } {\displaystyle \frac{\binom{d + n-1}{n-1}}{\binom{d + n-2}{n-1} + n}} < \ell < \ell_0.
\end{cases}
\end{align*}
\end{thm}
\begin{proof}
Again we use Lemma \ref{lem:dim for r=2}. This time
the ideal $I = I_{(1)} + \cdots + I_{(\ell)}$ contains $\ell$ generic linear forms. Thus, $I = (x_1,\ldots,x_n)$ if $\ell \ge n$, and we are done in this case. If $\ell < n$, then we get
\[
A = S/I \cong \Bbbk[x_1,\ldots,x_{n-\ell}]/(G_1,\ldots,G_{\ell}),
\]
where each $G_j$ is a generic form of degree $d-1$ in $T = \Bbbk[x_1,\ldots,x_{n-\ell}]$. It follows that
\begin{align*}
\hspace{2em}&\hspace{-2em} \dim \sigma_{\ell} (\mathbb X_{n-1, \lambda})\\
& = -1 + \dim_{\Bbbk} [I]_d \\
& = -1 + \dim_{\Bbbk} [S]_d - \dim_{\Bbbk} [A]_d \\
& = -1 + \dim_{\Bbbk} [S]_d - \dim_{\Bbbk} [T]_d + \dim_{\Bbbk} [(G_1,\ldots,G_{\ell})]_d \\
& = -1 + \binom{d+n-1}{n-1} - \binom{d+n- \ell - 1}{d} + \min \left \{\binom{d+n- \ell - 1}{d}, \ell (n- \ell) \right \} \\
& = -1 + \binom{d+n-1}{n-1} - \max \left \{0, \binom{d+n- \ell - 1}{d} - \ell (n- \ell) \right \}.
\end{align*}
In order to see the penultimate equality consider the graded minimal free resolution of $(G_1,\ldots,G_{\ell})$. Its beginning is of the form
\[
\cdots \to F \to T^{\ell} (-d+1) \to (G_1,\ldots,G_{\ell}) \to 0,
\]
where $F$ is a graded free $T$-module. It follows that to compute $\dim_{\Bbbk} [(G_1,\ldots,G_{\ell})]_d$ it is enough to know the number of linearly independent linear syzygies of the ideal $(G_1,\ldots,G_{\ell})$. Since the forms $G_j$ are generic this number is the least possible by the main result in \cite{HL}, and the dimension formula follows. It shows that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ fills its ambient space if and only if
\[
\binom{d+n - \ell - 1}{d} \le \ell (n - \ell).
\]
If $\ell = n-1$, this is true. Hence, the number $\ell_0$ is well defined and satisfies $\ell_0 \le n - 1$. Furthermore, if $\sigma_\ell(\mathbb X_{n-1,\lambda})$
fills its ambient space, then so does $\sigma_{\ell + 1} (\mathbb X_{n-1,\lambda})$.
This completes the argument for finding the dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$.
It remains to discuss the defect of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$.
Note that
\[
\dim \mathbb X_{n-1, \lambda} = \binom{d+n-2}{n-1} + n -1,
\]
and so the expected dimension of $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is
\[
\hbox{exp.dim } \sigma_\ell (\mathbb X_{n-1,\lambda}) = \min \left \{\binom{d+n-1}{n-1} - 1, \; \ell \cdot \binom{d+n-2}{d-1} + \ell n - 1. \right \}.
\]
We need to show that $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ is defective if and only if $2 \le \ell < \ell_0 < n$. Since for such $\ell$ the variety $\sigma_{\ell} (\mathbb X_{n-1, \lambda})$ does not fill its ambient space, this is equivalent to proving
\[
\ell \cdot \binom{d+n-2}{d-1} + \ell n - 1 > \binom{d+n-1}{n-1} - \binom{d+n- \ell - 1}{d} + \ell (n- \ell) - 1,
\]
that is,
\begin{equation}
\label{eq:defect lin factor}
\binom{d+n- \ell - 1}{d} - \binom{d+n-1}{n-1} + \ell \cdot \binom{d+n-2}{d-1} > - \ell^2.
\end{equation}
Notice that
\begin{align*}
\binom{d+n- \ell - 1}{d} - \binom{d+n-1}{n-1} + \ell \cdot \binom{d+n-2}{d-1} &
= \sum_{j=2}^{\ell} (-1)^j \binom{\ell}{j} \binom{d-j+n-1}{n-1} \\
& = \dim_{\Bbbk} [\Syz]_d,
\end{align*}
where $\Syz$ is the first syzygy module of a complete intersection in $S$ that is generated by $\ell < n$ linear forms (see Remark \ref{rem:syzygy interpretation}). This shows that the left-hand side in Inequality \eqref{eq:defect lin factor} is non-negative, and hence establishes that this inequality is true.
In order to determine the positive defect, it is enough to observe that
\[
\hbox{exp.dim } \sigma_\ell (\mathbb X_{n-1,\lambda}) \le \binom{d+n-1}{n-1} - 1
\]
if and only if
\[
\ell \le {\displaystyle \frac{\binom{d + n-1}{n-1}}{\binom{d + n-2}{n-1} + n}}.
\]
This concludes the calculation of the defect.
\end{proof}
Since we discussed the case $\ell = 2$ in Section \ref{sec:secant lines}, we illustrate the last result in the case $\ell = 3$.
\begin{cor}
\label{cor:l=3 r=2}
Consider the secant plane variety $\sigma_{3} (\mathbb X_{n-1, [d-1, 1]})$.
\begin{itemize}
\item[(a)] $\sigma_{3} (\mathbb X_{n-1, [d-1, 1]})$ fills its ambient space if and only if
\begin{itemize}
\item[(i)] $n \in \{3, 4\}$ and $d \ge 2$, \quad or
\item[(ii)] $n = 5$ and $d \in\{ 2, 3, 4, 5\}$, \quad or
\item[(iii)] $n = 6$ and $d = 2$.
\end{itemize}
\item[(b)] In all other cases $\sigma_{3} (\mathbb X_{n-1, [d-1, 1]})$ is defective with dimension
\[
\dim \sigma_{3} (\mathbb X_{n-1, [d-1, 1]}) = \begin{cases}
6 n - 16 & \text{ if } d = 2 \\[.5ex]
{\ds \binom{d+n-1}{n-1} - \binom{d+n-4}{d} + 3n -10 } & \text{ if } d \ge 3.
\end{cases}
\]
\end{itemize}
\end{cor}
\begin{proof}
Consider first $d = 2$. Then we can apply Theorem \ref{thm:balanced}. However, it is easier to argue directly. In this case, the ideal $I$ in Lemma \ref{lem:dim for r=2} is generated by $2 \ell = 6$ linear forms. Hence $[I]_2 = [S]_2$ if and only if $n \le 6$. If $n \ge 7$, then we get
\begin{eqnarray*}
\dim \sigma_{3} (\mathbb X_{n-1, ([1, 1]}) = \dim_{\Bbbk} [I]_2 - 1 & = & \binom{n+1}{2} - \binom{n-5}{2} - 1 \\[1ex]
& = & 6n - 16.
\end{eqnarray*}
Moreover, $\sigma_{2} (\mathbb X_{n-1, [1, 1]})$ is defective if it does not fill its ambient space by Theorem \ref{thm:balanced}.
Assume now $d \ge 3$. Then Theorem \ref{thm:case d-1, 1} shows that $\sigma_{3} (\mathbb X_{n-1, [d-1, 1]})$ fills its ambient space if and only if
\[
\frac{n}{2} \le \ell_0 \le 3.
\]
Furthermore, $\sigma_{3} (\mathbb X_{n-1, [d-1, 1]})$ does not fill its ambient space if $n \ge 6$ by Theorem \ref{thm:dim formula}. Hence, it remains to consider the cases $n \in \{3, 4, 5\}$.
If $n = 5$, this forces $\ell_0 = 3$, which means
\[
d+1 \le 6 \quad \text{ and } \quad \binom{d+2}{2} > 6,
\]
that is, $d \in \{3, 4, 5\}$.
Since $\ell_0 \leq n-1$, we get $\ell_0 \le 3$ if $n \le 4$, and thus $\sigma_{3} (\mathbb X_{n-1, ([1, 1]})$ fills its ambient space. This shows Part~(a). Claim (b) follows by Theorem \ref{thm:case d-1, 1}.
\end{proof}
\section{The variety of reducible forms}
\label{sec:variety red forms}
Every reducible form of degree $d$ in $n$ variables corresponds to a point of the variety
\[
\mathbb X_{n-1, d} = \bigcup_{k=1}^{\lfloor \frac{d}{2} \rfloor} \mathbb X_{n-1, [d-k, k]}.
\]
(Notice that this holds even for reducible forms with more than two factors.)
Thus, we call $\mathbb X_{n-1, d}$ the {\em variety of reducible forms of degree $d$ in $n$ variables}. In this section we study its secant varieties. This is based on the results on the secant varieties of the various $\mathbb X_{n-1, [d-k, k]}$, where $k$ varies between $1$ and $\lfloor \frac{d}{2} \rfloor$.
\begin{rem}
\label{rem:d is 2}
The variety $\mathbb X_{n-1, d}$ is irreducible if and only if $d = 2$. In this case $\mathbb X_{n-1, 2} = \mathbb X_{n-1, [1, 1]}$ and
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, 2}) = \begin{cases}
2 \ell (n - \ell) + \ell -1 & \text{if } 2\ell < n\\
\binom{n+1}{2} - 1 & \text{if } 2\ell \ge n.
\end{cases}
\]
Moreover, $\sigma_{\ell} (\mathbb X_{n-1, 2})$ is defective if and only if $2 \le \ell < \frac{n}{2}$. In this case, the defect is $2 \ell (\ell - 1)$ (see, e.g., Theorem \ref{thm:balanced}).
\end{rem}
We begin by determining the dimension of $\mathbb X_{n-1, d}$. The next result is an immediate consequence of Corollary \ref{cor:dim comparison}.
\begin{prop}
\label{prop:dim red var}
For all $d \ge 2$ and $n \ge 3$,
\[
\dim \mathbb X_{n-1, d} = \dim \mathbb X_{n-1, [d-1, 1]} = \binom{d + n-2}{n-1} + n -1.
\]
\end{prop}
Corollary \ref{cor:dim comparison} also gives that $\mathbb X_{n-1, [d - 1, 1]}$ is the unique irreducible component of $\mathbb X_{n-1, d}$ that has the same dimension as $\mathbb X_{n-1, d}$.
As we noted above, $\mathbb X_{n-1,d}$ is not irreducible as soon as $d > 2$. Thus, in order to calculate the dimension of $\sigma_\ell(\mathbb X_{n-1,d})$ one must also consider the (embedded) $\ell$-joins of the irreducible components of $\mathbb X_{n-1,d}$.
Recall that if $\mathbb X_1, \cdots, \mathbb X_\ell$ are irreducible varieties in $\P^m$ (not necessarily distinct), then the {\it embedded join} of $\mathbb X_1, \cdots , \mathbb X_\ell$, denoted
$$
J(\mathbb X_1, \cdots , \mathbb X_\ell)
$$
is the Zariski closure of the union of all the linear spaces $\langle P_1, \ldots , P_\ell \rangle \subset \P^m$ where $P_i \in \mathbb X_i$. If all the $\mathbb X_i = \mathbb X$ then this is nothing other than $\sigma_\ell(\mathbb X)$. Furthermore, for any (possibly reducible) variety $\mathbb X \subset \P^m$, the parameter count mentioned in the introduction gives
\[
\dim \sigma_\ell(\mathbb X) \le \min \{m, \ \ell \cdot \dim \mathbb X + \ell -1\},
\]
and the right-hand side is called the \emph{expected dimension} of $ \sigma_\ell(\mathbb X)$.
The following Lemma shows that if $\mathbb X_1, \cdots, \mathbb X_\ell$ are any $\ell$ irreducible components of $\mathbb X_{n-1,d}$ and $2\ell \leq n$, then
\[
\dim \sigma_\ell(\mathbb X_{n-1,[d-1,1]}) \geq \dim J(\mathbb X_1, \cdots , \mathbb X_\ell).
\]
\begin{lem}
\label{prop:compare Hilb c.i.}
Consider integers $k_1, k_2,\ldots, k_{\ell} \in \{1,\ldots, \lfloor \frac{d}{2} \rfloor \}$, where $d \ge 2$.
Let $I \subset S = \Bbbk[x_1,\ldots,x_n]$ be an ideal generated by $2 \ell$ general forms of degrees $k_1, k_2,\ldots, k_{\ell}, d- k_1, d - k_2,\ldots, d- k_{\ell}$. Let $J \subset S$ be an ideal generated by $\ell$ general linear forms
and $\ell$ general forms of degree $d - 1$. If $2\ell \le n$, then, for all integers $j$,
\[
\dim_{\Bbbk} [S/I]_j \ge \dim_{\Bbbk} [S/J]_j.
\]
Moreover, if $k_i > 1$ for some $i \in \{1,\ldots,\ell\}$, then there is some $j$ such that this is a strict inequality.
\end{lem}
\begin{proof}
Consider first the case, where $n = 2 \ell$. If $n = 2$, then
\[
\dim_{\Bbbk} [S/J]_j = \begin{cases}
1 & \text{if } 0 \le j \le d-2\\
0 & \text{otherwise},
\end{cases}
\]
and the claim follows in this case. Using Hilbert series, this observation can be expressed as
\begin{equation}
\label{eq:compare c.i.}
\frac{1}{(1 - t)^2} (1 - t^k) (1 - t^{d-k}) \ge \frac{1}{(1 - t)^2} (1 - t) (1 - t^{d-1}) \quad \text{whenever } 1 \le k \le \frac{d}{2}.
\end{equation}
Let now $n \ge 4$. Then the Hilbert series of $S/I$ and $S/J$ are
\[
\mathbf{HS} (S/I) = \prod_{i = 1}^{\ell} \frac{(1 - t^{k_i}) (1 - t^{d-k_i})}{(1 - t)^2}
\]
and
\[
\mathbf{HS} (S/J) = \prod_{i = 1}^{\ell} \frac{(1 - t) (1 - t^{d-1})}{(1 - t)^2}.
\]
Thus, Inequality \eqref{eq:compare c.i.} gives
\[
\mathbf{HS} (S/I) \ge \mathbf{HS} (S/J),
\]
as desired.
Finally, assume $n > 2 \ell$. Then
\begin{align*}
\mathbf{HS} (S/I) & = \frac{1}{(1 - t)^{n - 2 \ell}} \cdot \prod_{i = 1}^{\ell} \frac{(1 - t^{k_i}) (1 - t^{d-k_i})}{(1 - t)^2} \\
& \ge \frac{1}{(1 - t)^{n - 2 \ell}} \cdot \prod_{i = 1}^{\ell} \frac{(1 - t) (1 - t^{d-1})}{(1 - t)^2} \\
& = \mathbf{HS} (S/J),
\end{align*}
where the estimate follows from the case $n = 2 \ell$ and the fact that the coefficients in the power series expansion of $\frac{1}{(1 - t)^{n - 2 \ell}}$ are all non-negative.
\end{proof}
We are ready for the main result of this section.
\begin{thm}
\label{thm:red forms, small ell}
Assume $2\ell \le n$. Then
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, d}) = \dim \sigma_{\ell} (\mathbb X_{n-1, [d - 1, 1]}).
\]
Moreover:
\begin{itemize}
\item[(a)] The variety $\sigma_{\ell} (\mathbb X_{n-1, d})$ fills its ambient space if and only if
\begin{itemize}
\item[(i)] $2\ell = n$ and $d = 2$; \quad or
\item[(ii)] $\ell = 2, \ n = 4$, and $d = 3$.
\end{itemize}
\item[(b)] If $\sigma_{\ell} (\mathbb X_{n-1, d})$ does not fill its ambient space, then its dimension is
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, d}) =
\binom{d+n-1}{n-1} - \binom{d+n- \ell - 1}{d} + \ell (n- \ell) - 1,
\]
and $\sigma_{\ell} (\mathbb X_{n-1, d})$ is defective.
\end{itemize}
\end{thm}
\begin{proof}
Let $P_1,\ldots,P_{\ell} \in \mathbb X_{n-1, d}$ be points such that each $P_i$ is a general point on some component, say $\mathbb X_{n-1, [d - k_i, k_i]}$, of $\mathbb X_{n-1, d}$.
Then Terracini's Lemma gives (as in Corollary \ref{dim sigma ell})
\[
\dim (\sigma_\ell(\mathbb X_{n-1,d} )) = \max \left \{\dim_{\Bbbk} \left [ I_{P_1} + \cdots + I_{P_\ell} \right ]_d -1 \; | \; k_1, k_2,\ldots, k_{\ell} \in \left \{1,\ldots, \textstyle{\left \lfloor \frac{d}{2} \right \rfloor} \right \} \right \}.
\]
Using the notation of Lemma \ref{prop:compare Hilb c.i.}, this implies
\[
\dim (\sigma_\ell(\mathbb X_{n-1,d})) = \dim_{\Bbbk} [J]_d - 1 = \dim (\sigma_\ell(\mathbb X_{n-1,[d-1, 1]} )),
\]
as desired.
Part (a) is now a consequence of the second part of Theorem \ref{thm:dim formula}. We claim that part (b) follows from Theorem \ref{thm:case d-1, 1}. Indeed, if $\ell < \ell_0$ then Theorem \ref{thm:case d-1, 1} gives the desired dimension. If $\ell_0 \leq \ell$ then Theorem \ref{thm:case d-1, 1} yields that $\sigma_\ell(\mathbb X_{n-1,[d-1,1]})$ fills its ambient space.
\end{proof}
\begin{rem}
\label{rem:compare secant dim r is 2}
Lemma \ref{prop:compare Hilb c.i.} also implies that
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, [d-k, k]}) \le \dim \sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]}) \quad \text{ if } 2 \le k \le \frac{d}{2},
\]
provided $\ell \le \frac{n}{2}$. We conjecture that this bound is true without the latter restriction. Notice that we have an upper bound for $\dim \sigma_{\ell} (\mathbb X_{n-1, [d-k, k]})$ by Proposition \ref{prop:Froeberg gives main conj} and that we know $\dim \sigma_{\ell} (\mathbb X_{n-1, [d-1, 1]})$ by Theorem \ref{thm:case d-1, 1}. This reduces this conjecture to a comparison of two numbers. However, we have been unable to establish the needed estimate.
\end{rem}
By Theorem \ref{thm:case d-1, 1}, we know exactly when the secant variety $\sigma_{\ell} (\mathbb X_{n-1, [d - 1, 1]})$ fills its ambient space. This gives:
\begin{thm}
\label{thm:secants red forms fill}
The secant variety $\sigma_{\ell} (\mathbb X_{n-1, d})$ fills its ambient space if $\ell \ge \ell_0$,
where
\[
\ell_0 =
\min \left \{ \ell \ge \frac{n}{2} \; | \; \ell \in \mathbb{Z} \text{ and } \binom{d - \ell + n - 1}{d} \le \ell (n- \ell) \right \}.
\]
In particular, $\sigma_{\ell} (\mathbb X_{n-1, d})$ fills its ambient space if $\ell \ge n -1$.
\end{thm}
\begin{proof}
Notice that
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, d}) \ge \dim \sigma_{\ell} (\mathbb X_{n-1, [d - 1, 1]}).
\]
Thus, the claim follows from Theorem \ref{thm:case d-1, 1} if $d \ge 3$. If $d = 2$, then $\ell_0 = \lceil \frac{n}{2} \rceil$, and we conclude by Remark \ref{rem:d is 2}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:intro red forms}]
Combine Theorems \ref{thm:red forms, small ell} and \ref{thm:secants red forms fill}.
\end{proof}
\begin{rem}
\label{rem:conj for red forms}
If $2\ell > n$, then Theorems \ref{thm:red forms, small ell} and \ref{thm:secants red forms fill} do not rule out the possibility that $\sigma_\ell(\mathbb X_{n-1,d})$ fills its ambient space even when $\sigma_\ell(\mathbb X_{n-1,[d-1,1]})$ does not. However, we do not expect this ever happening. In fact, we suspect that the following extension of Theorem \ref{thm:red forms, small ell} is true:
\[
\dim \sigma_{\ell} (\mathbb X_{n-1, d}) = \dim \sigma_{\ell} (\mathbb X_{n-1, [d - 1, 1]}),
\]
for all $\ell, \ n$ and $d$.
If so, then, by Theorem \ref{thm:case d-1, 1}, the converse of Theorem \ref{thm:secants red forms fill} is true and $\sigma_{\ell} (\mathbb X_{n-1, d})$ is defective whenever it does not fill its ambient space.
\end{rem}
\section{Application to secant varieties of Segre varieties}\label{AOPstuff}
The paper \cite{AOP} by Abo, Ottaviani and Peterson
classifies all Segre varieties $X$ such that the $\ell$-secant variety is defective for some $\ell < 7$
and it raises some questions as to conjecturally what happens in general.
Our results verify certain cases of these conjectures.
We first recall some terminology from \cite{AOP} and a related result from \cite{CGG:0}.
Assume $2\leq n_r\leq\ldots\leq n_1$ (this is consistent with our ordering convention, but it is the reverse of what \cite{AOP} does). Say that
$(n_1,\ldots, n_r)$ is \emph{balanced} if $n_1-1\leq \Pi_{i=2}^rn_i-\sum_{i=2}^r(n_i-1)$ and
\emph{unbalanced} otherwise. Using this terminology,
\cite{CGG:0} proves that with respect to the Segre embedding of $X={\mathbb P}^{n_1-1}\times \cdots\times {\mathbb P}^{n_r-1}$
in projective space, $X$ has a defective $\ell$-secant
variety for some $\ell$ if $(n_1,\ldots, n_r)$ is unbalanced. This result, \cite[Proposition 3.3]{CGG:0},
was paraphrased in \cite{AOP} essentially as follows (see \cite[Lemma 4.1]{AOP}):
\vskip\baselineskip
\begin{prop} With respect to the Segre embedding of $X={\mathbb P}^{n_1-1}\times \cdots\times {\mathbb P}^{n_r-1}$
in projective space, $\sigma_\ell(X)$ is defective for $\ell$ satisfying
$$\Pi_{i=2}^rn_i-\sum_{i=2}^r(n_i-1)<\ell<\min\{\Pi_{i=2}^rn_i,n_1\}.$$
\end{prop}
By a Segre variety $X$ being \emph{unbalanced} \cite{AOP} means $X={\mathbb P}^{n_1-1}\times \cdots\times {\mathbb P}^{n_r-1}$ where
$(n_1,\ldots,n_r)$ is unbalanced. With these definitions, \cite[Question 6.6]{AOP} then asks:
\vskip\baselineskip
\begin{ques}\label{AOPquest}
Is it true for a Segre variety $X$ such that $\sigma_\ell(X)$ is defective for some $\ell$,
that either:
1. $X$ is unbalanced; or
2. $X={\mathbb P}^2\times{\mathbb P}^{n-1}\times{\mathbb P}^{n-1}$ with $n$ odd; or
3. $X={\mathbb P}^2\times{\mathbb P}^3\times{\mathbb P}^3$; or
4. $X={\mathbb P}^1\times{\mathbb P}^1\times{\mathbb P}^{n-1}\times{\mathbb P}^{n-1}$?
\end{ques}
The conjecture in \cite{AOP} is that the answer is yes; to rephrase:
\begin{conj}\label{AOPconj}
Let $X$ be a balanced Segre variety but not any of those listed in items 2, 3 or 4 of Question \ref{AOPquest}.
Then $\sigma_\ell(X)$ is not defective for all $\ell\geq 2$.
\end{conj}
Our results verify this in various cases, as we now explain.
It is useful to reconsider the morphism \eqref{SegreToRedFormsMorphism}.
Let $X={\mathbb P}^{n_1-1}\times \cdots\times {\mathbb P}^{n_r-1}$. The points of the Segre embedding of $X$
in ${\mathbb P}^{N-1}$, where $N=\Pi_i n_i$, are exactly those of the form
$[\ldots, a_{i_11}a_{i_22}\cdots a_{i_rr},\ldots]$, where $(a_{1j},\ldots,a_{n_jj})\in {\mathbb P}^{n_j-1}$.
We can regard $[\ldots, a_{i_11}a_{i_22}\cdots a_{i_rr},\ldots]$ as the multi-homogeneous polynomial
\[
\begin{split}
F(\ldots,x_{ij},\ldots)&=\sum a_{i_11}a_{i_22}\cdots a_{i_rr}x_{i_11}x_{i_22}\cdots x_{i_rr}\\
&=\Pi_{j=1}^r (a_{1j}x_{1j}+\cdots+a_{n_jj}x_{n_jj})
\in \Bbbk[\ldots,x_{ij},\ldots]
\end{split}
\]
of multi-degree $(1^r)=(1,\ldots,1)$, where the variables $x_{ij}$ are indexed by
$1\leq i\leq n_j$ with $1\leq j\leq r$ and $\Bbbk[{\mathbb P}^{n_j-1}]=\Bbbk[x_{1j},\ldots,x_{n_jj}]$.
Now suppose that $n_j=\binom{d_j+n-1}{n-1}$ for each $j$, where $d_1\geq d_2\geq \cdots\geq d_r$ and $d=d_1+\cdots+d_r$.
Regarding the coordinate variables $x_{ij}$ of ${\mathbb P}^{n_j-1}$
as an enumeration of the monomials $M_{ij}\in \Bbbk[x_1,\ldots,x_n]$ of degree $d_j$ for each $j$,
we get a map
$$F\mapsto \overline{F}\in \Bbbk[x_1,\ldots,x_n],$$
where $\overline{F}$ is the degree $d$ singly homogeneous polynomial
$\overline{F}(\ldots,M_{ij},\ldots)$ obtained by substituting $M_{ij}$ into $x_{ij}$.
Note that the map $\overline{\vbox to.08in{\hbox to.1in{\hfil}}}$ is actually a linear homomorphism
\[
\overline{\vbox to.08in{\hbox to.1in{\hfil}}}:(\Bbbk[\ldots,x_{ij},\ldots])_{(1,\ldots,1)}\to (\Bbbk[x_1,\ldots,x_n])_d.
\]
The projectivizations of $(\Bbbk[\ldots,x_{ij},\ldots])_{(1,\ldots,1)}$ and $(\Bbbk[x_1,\ldots,x_n])_d$
are ${\mathbb P}^{N-1}$ and ${\mathbb P}^{\binom{d+n-1}{n-1}-1}$, respectively. The restriction of $\overline{\vbox to.08in{\hbox to.1in{\hfil}}}$
to the affine cone corresponding to the Segre embedding of $X$ in ${\mathbb P}^{N-1}$ is just the surjective morphism induced
on affine cones by the morphism \eqref{SegreToRedFormsMorphism} of $X$ to $\mathbb X_{n-1,\lambda}$
for $\lambda=[d_1,\ldots,d_r]$.
Since $\overline{\vbox to.08in{\hbox to.1in{\hfil}}}:(\Bbbk[\ldots,x_{ij},\ldots])_{(1,\ldots,1)}\to (\Bbbk[x_1,\ldots,x_n])_d$
is linear, it maps the affine cone of the $\ell$-secant variety of $X$ to the affine cone of the $\ell$-secant variety of $\mathbb X_{n-1,\lambda}$.
If $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ is not defective,
let us say that the $\ell$-secant variety $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$
of $\mathbb X_{n-1,\lambda}$ does not \emph{overly fill} its ambient space if
$\dim \sigma_{\ell} (\mathbb X_{n-1,\lambda}) = \ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1$. Also, given the partition $\lambda=[d_1,\ldots,d_r]$,
let $X_{n-1,\lambda}$ denote the Segre embedding of ${\mathbb P}^{n_1-1}\times \cdots\times {\mathbb P}^{n_r-1}$, where
$n_i=\binom{d_i+n-1}{n-1}$. We thus have:
\begin{thm}\label{AOPholds}
If $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ is not defective and does not overly fill its ambient space, then
$\sigma_\ell(X_{n-1,\lambda})$ is not defective.
\end{thm}
\begin{proof}
Assume $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ does not overly fill its ambient space.
If $\sigma_\ell(X_{n-1,\lambda})$ were defective, then
$\dim \sigma_\ell(X_{n-1,\lambda})
< \ell \cdot \dim X_{n-1,\lambda} + \ell-1=\ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1$.
Since $\sigma_{\ell} (\mathbb X_{n-1,\lambda})$ does not overly fill its ambient space, we have
$\ell \cdot \dim \mathbb X_{n-1,\lambda} + \ell-1
=\dim \sigma_\ell(\mathbb X_{n-1,\lambda})$ and hence $\dim \sigma_\ell(X_{n-1,\lambda})
< \dim \sigma_\ell(\mathbb X_{n-1,\lambda})$.
But $\overline{\vbox to.08in{\hbox to.1in{\hfil}}}:(\Bbbk[\ldots,x_{ij},\ldots])_{(1,\ldots,1)}\to (\Bbbk[x_1,\ldots,x_n])_d$
maps the affine cone of $\sigma_\ell(X_{n-1,\lambda})$ onto the affine cone of
$\sigma_\ell(\mathbb X_{n-1,\lambda})$, so we must have
$\dim \sigma_\ell(X_{n-1,\lambda})
\geq\dim \sigma_\ell(\mathbb X_{n-1,\lambda})$, hence $\sigma_\ell(X_{n-1,\lambda})$ cannot be defective.
\end{proof}
As an example, we have the following corollary which verifies Conjecture \ref{AOPconj} in a range of cases.
Since the results proved in \cite{AOP} mainly have $\ell<7$ or $d_1=\cdots=d_r$, most of the cases
of the corollary seem to be new.
Of course, our results verify the conjecture in many more instances too.
\begin{comment}
\begin{cor}
Let $1\leq d_r\leq \cdots\leq d_1<d_2+\cdots+d_r$, $3\leq r$ and $4\leq 2\ell\leq n$ be integers with
$n_i=\binom{d_i+n-1}{n-1}$ and $a=n_1\cdots n_r$. If $X$ is the Segre embedding of
${\mathbb P}^{n_1-1}\times\cdots\times {\mathbb P}^{n_r-1}$ in ${\mathbb P}^{a-1}$, then
$\sigma_\ell(X)$ is not defective. If, moreover, $d_1=\cdots=d_r$, then $X$ is balanced.
\end{cor}
\begin{proof}
By Theorem \ref{general r geq 3}, $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is not defective $\lambda=[d_1,\ldots,d_r]$
and does not overly fill its ambient space, hence by Theorem \ref{AOPholds}, $\sigma_\ell(X)$ is not defective.
Now we just need to check that $X$ is balanced when $d_1=\cdots=d_r$; i.e., that $n_1+\cdots+n_r\leq n_2\cdots n_r+r$,
or, using $d_1=\cdots=d_r$, that $rn_1\leq n_1^{r-1}+r$. It is enough to show that $r\leq n_1^{r-2}$ under the assumption that
$r\geq 3$ and $n_1\geq 4$, which we have since $n_1=\binom{d_1+n-1}{n-1}$ and $n\geq 4$.
But $r\leq3^{r-2}$ is clear, so $r\leq3^{r-2}\leq n_1^{r-2}$.
\end{proof}
\end{comment}
\begin{cor}
Let $1\leq d_r\leq \cdots\leq d_1<d_2+\cdots+d_r$, $3\leq r$ and $4\leq 2\ell\leq n$ be integers with
$n_i=\binom{d_i+n-1}{n-1}$ and $a=n_1\cdots n_r-1$. If $X$ is the Segre embedding of
${\mathbb P}^{n_1-1}\times\cdots\times {\mathbb P}^{n_r-1}$ in ${\mathbb P}^{a-1}$, then
$\sigma_\ell(X)$ is not defective. Moreover, $X$ is balanced for $d_2=\cdots=d_r\gg0$.
\end{cor}
\begin{proof}
By Theorem \ref{general r geq 3}, $\sigma_\ell(\mathbb X_{n-1,\lambda})$ is not defective for $\lambda=[d_1,\ldots,d_r]$
and does not overly fill its ambient space, hence by Theorem \ref{AOPholds}, $\sigma_\ell(X)$ is not defective.
Now we just need to check that $X$ is balanced when $d_2=\cdots=d_r\gg0$. Since
$\binom{d_1+n-1}{n-1} < \binom{d_2(r-1)+n-1}{n-1}$, it suffices to show that
$\binom{d_2(r-1)+n-1}{n-1}\leq \binom{d_2+n-1}{n-1}^{r-1}-(r-1)\binom{d_2+n-1}{n-1}$.
But $\binom{d_2(r-1)+n-1}{n-1}$ is a polynomial in $d_2$ of degree $n-1$ while
$\binom{d_2+n-1}{n-1}^{r-1}-(r-1)\binom{d_2+n-1}{n-1}$ is a polynomial in $d_2$ of degree $(r-1)(n-1)$,
hence the inequality must hold for $d_2\gg0$.
\end{proof}
\section{Further questions and comments}
In this section we pose some natural questions arising from our work.
It is clear from the previous sections that the major algebraic question left open in the paper is the extent to which
sums of at least two generic tangent space ideals to the varieties of reducible hypersurfaces have enough Lefschetz elements (see Conjecture \ref{conj:WLP}). One should note that less information than the full Weak Lefschetz Property is needed to establish the conjectured dimension of secant varieties to varieties of reducible forms. This allowed us to use results by Hochster and Laksov \cite{HL}, Anick \cite{anick}, and a theorem on complete intersections (see
\cite{stanley, watanabe, RRR, sekiguchi} or \cite{HP}) in the proofs of Theorems \ref{thm:case d-1, 1} and \ref{thm:dim when froeberg}, respectively. It would be very interesting to have new instances where (partial) Weak Lefschetz Properties are established.
As we have observed earlier in the paper, each variety of reducible hypersurfaces is a finite projection of a Segre embedding of a product of projective spaces. Are there more geometric conclusions (than those we have found in Section \ref{AOPstuff}) that we can draw about the secant varieties of the Segre embedding from the more abundant information we have for the secant varieties of the varieties of reducible hypersurfaces?
Another question is if any of the secant varieties of the varieties of reducible forms are arithmetically Cohen-Macaulay, apart from the trivial cases in which those secant varieties are themselves hypersurfaces in their ambient space.
Again, apart from some very small examples, we do not have equations for the varieties of reducible hypersurfaces, much less for their secant varieties (even in cases where we know the latter are hypersurfaces in their ambient spaces (see Remark \ref{hyper surf})). It would be interesting to have some intrinsic equations for these varieties or a bound on the degrees of their equations.
Mammana \cite{mammana} gives a formula for the degree of the variety of reducible plane curves but we have no generalization of that formula for varieties of reducible hypersurfaces beyond the case of plane curves. More generally, one would like to have a formula for the degree of the secant varieties of these varieties. This is not known even for varieties of plane curves, except in the most trivial of cases. It would even be interesting to know the degree when the variety is a hypersurface in its ambient space.
\vskip .5cm
\thanks
\noindent{\bf Acknowledgements:}
The authors wish to thank Queen's University and NSERC (Canada), in the person of the second author,
for kind hospitality during the preparation of this work.
Catalisano and Gimigliano were partially supported by GNSAGA of INDAM and by MUIR funds (Italy). Geramita was partially supported by NSERC (CANADA) under grant No. 386080, while Harbourne was partially supported by NSA (US) under grant NO. H98230-13-1-0213. Both Migliore and Nagel were partially supported by the Simons Foundation under grants No. 309556 (Migliore) and 317096 (Nagel). Shin was supported by the Basic Science Research Program of the NRF (Korea) under grant No. 2013R1A1A2058240/2.
\bibliographystyle{alpha}
|
\section{Introduction}
This article deals with the integrability of invariant distributions arising in certain dynamical systems.
The integrability of tangent distributions of $k-$planes is an important problem
that has been studied for more than a century. Under certain hypotheses, there have been
quite satisfactory answers to the unique integrability problem. For instance, Picard's theorem in the one-dimensional case, and Frobenius' theory in higher dimensions. However, both results involve regularity of the distributions. \par
The problem is that, in general, distributions arising from a dynamical system are only H\"older-continuous, even if the system is smooth. Therefore, other elements of analysis are required in order to establish necessary or sufficient conditions for integrability. \par
Consider, for instance, the case of hyperbolic systems. Namely, diffeomorphims $f:M\to M$ for which there exists an invariant splitting of the tangent bundle into two invariant distributions, called {\em stable} and {\em unstable}: $TM=E^{s}\oplus E^{u}$ such that, for some Riemannian metric $\|.\|$ and all unit vectors $v_{s}\in E^{s}$ and $v_{u}\in E^{u}$ we have:
$$\|Df(x)v_{s}\|<1<\|Df(x)v_{u}\|.$$
In this case, both distributions, stable and unstable, are uniquely integrable. This can be achieved by applying essentially either one of the following two methods: Hadamard's method \cite{hadamard}, which consists in seeing each invariant integral manifold as a fixed point of a contracting operator acting in an appropriate space of functions, or Perron's method \cite{perron}, which consists in applying the Implicit Function Theorem to an operator acting in an appropriate function space.\par
Here we shall focus on the integrability of some distributions arising from a more general kind of systems, the {\em partially hyperbolic} ones. These systems involve not only a contracting and an expanding bundle as the ones mentioned above, but also a {\em center bundle}, which has an intermediate behavior. Namely; a diffeomorphism $f:M\to M$ is {\em partially hyperbolic} if the tangent bundle of $M$ admits an invariant splitting into 3 bundles, called respectively, {\em stable}, {\em center} and {\em unstable} bundles, $TM=E^{s}\oplus E^{c}\oplus E^{u}$, such that, for some Riemannian metric $\|.\|$, and all unit vectors $v_{s}\in E^{s}$, $v_{c}\in E^{c}$ and $v_{u}\in E^{u}$ we have
$$\|Df(x)v_{s}\|<\|Df(x)v_{c}\|<\|Df(x)v_{u}\|\quad\text{and}\quad \|Df(x)v_{s}\|<1<\|Df(x)v_{u}\|$$
Hadamard's and Perron's methods can be used to show that also in this setting, the stable and unstable bundles are uniquely integrable, see for instance \cite{bp,hps}.
However, it is a more delicate matter to determine wether $E^{c}$ or even $E^{cs}=E^{s}\oplus E^{c}$ or $E^{cu}=E^{c}\oplus E^{u}$ are integrable. \par
We will say that $f$ is {\em $cs$-dynamically coherent} if there exists an $f$-invariant foliation tangent to $E^{cs}$. The $cu$-dynamical coherence is defined analogously. A diffeomorphism $f$ is said to be {\em dynamically coherent} if it is both $cs$- and $cu$-dynamically coherent. This implies in particular that there exists an $f$-invariant foliation tangent to $E^{c}$.\par
Partially hyperbolic diffeomorphisms are not dynamically coherent in general. Indeed, as it was observed by A. Wilkinson in \cite{wilk}, a non-dynamically coherent example is given by an algebraic Anosov diffeomorphim in a six dimensional manifold, that was presented in the well-known survey by S. Smale \cite{sm} and is attributed to A. Borel. In this example, the sub-bundles are $C^{\infty}$, the center bundle is 4-dimensional and it cannot be integrable since the Frobenius condition is not satisfied. \newline\par
Is the lack of Frobenius condition the only reason for non-integrability of the center bundle? What about the case of one-dimensional $E^{c}$, where Frobenius condition is always trivially satisfied? The question of whether a partially hyperbolic diffeomorphism existed for a one-dimensional non-integrable center bundle remained open since the 70's. In \cite{HHU.nondinco} the authors answered the question negatively. They constructed examples of non-dynamically coherent
partially hyperbolic diffeomorphisms on $\mathbb{T}^3$. In \cite{HHU.nondinco}, it is also given (using the same methods) an example of a dynamically coherent diffeomorphism with non-locally uniquely integrable center foliation. The existence of such examples
contrasts with the result obtained by M. Brin, D. Burago and S.
Ivanov \cite{brin_burago_ivanov2} proving that diffeomorphisms on
${\mathbb T}^3$ satisfying a more restrictive definition of partial hyperbolicity (the {\em absolute partial hyperbolicity}) are dynamically coherent. \newline\par
A natural question is then: are there necessary or sufficient conditions for dynamical coherence? In this paper we provide some necessary conditions:
\begin{theorem}\label{uniqueint} Let $M$ be a closed 3-dimensional manifold and $f:M\rightarrow M$ be a $cu$-dynamically coherent partially hyperbolic diffeomorphism. Then, the center unstable foliation has no compact leaves.
\end{theorem}
Let us note that any compact leaf tangent to $E^{cu}$ must be a torus, by Poincar\'e-Hopf, due to the fact that is is foliated by lines. We shall call $cu$-torus any 2-torus tangent to $E^{cu}$.\par
As a matter of fact, the non-dynamically coherent example \cite{HHU.nondinco} was inspired in this result. Since any center-unstable foliation cannot have a torus leaf, we constructed a plane field that was uniquely integrable outside a $cu$-torus in such a way that if a foliation existed, it should contain the torus as a leaf. The fact that this is not possible provided the first example of a non dynamically coherent partially hyperbolic diffeomorphism with one-dimensional center bundle. \par
Observe that Theorem \ref{uniqueint} does not prevent the existence of tori, even invariant, tangent to $E^{cu}$. Theorem \ref{uniqueint} asserts the impossibility of the existence of such tori {\em as part of an invariant foliation} tangent to $E^{cu}$. Indeed, in \cite{HHU.nondinco} we also provide an example of a dynamically coherent partially hyperbolic diffeomorphism of ${\mathbb T}^{3}$ with an invariant $cu$-torus. \par
Theorem \ref{uniqueint} then states that if there is a center-unstable invariant foliation, no leaf can be a torus. We conjecture the converse is also true:
\begin{conjecture}\label{conj}
If a partially hyperbolic diffeomorphism $f:M^3\rightarrow M^3$ is not dynamically coherent, then it admits either a $cu$- or an $sc$-torus.
\end{conjecture}
In Section \ref{section.cu.tori} we prove that, in fact, not every manifold can contain a $cu$- or an $sc$-torus:
\begin{theorem}\label{invariant.tori} Let $f:M\to M$ be a partially hyperbolic diffeomorphism of an orientable 3-manifold. If there exists a torus tangent to either $E^{s}\oplus E^{u}$, $E^{c}\oplus E^{u}$ or $E^{s}\oplus E^{c}$, then the manifold $M$ is either:
\begin{enumerate}
\item the 3-torus ${\mathbb T}^{3}$
\item the mapping torus of $-id:{\mathbb T}^{2}\to{\mathbb T}^{2}$
\item the mapping torus of a hyperbolic automorphism $A:{\mathbb T}^{2}\to {\mathbb T}^{2}$
\end{enumerate}
\end{theorem}
This follows essentially from Proposition \ref{cu.tori.anosov}, where we prove that the dynamics on any invariant torus tangent to $E^{cu}$, or $E^{sc}$ is isotopic to the one generated by a hyperbolic linear automorphism. When a manifold $M$ admits an embedded 2-torus and a global diffeomorphism $g:M\to M$ such that $g(T)=T$ and $g|_{T}$ is isotopic to an Anosov diffeomorphism, then we say that $M$ admits an {\em Anosov torus}.\par
In \cite{anosov_tori}, we proved that the only irreducible manifolds admitting an Anosov torus are the ones listed in Theorem \ref{invariant.tori}. But on the other hand, if $M$ admits a partially hyperbolic diffeomorphism, then $M$ is an irreducible manifold. See Section \ref{section.cu.tori} for more details.\par
As a consequence of Theorem \ref{invariant.tori}, if Conjecture \ref{conj} were true, then the only manifolds supporting non-dynamically coherent diffeomorphisms would be the ones listed in Theorem \ref{invariant.tori}. In Proposition \ref{cu.tori.periodic} we show that the existence of a $cu$-torus implies the existence of a {\em periodic} $cu$-torus, which is attracting. An analogous statement holds for $sc$-tori. Therefore, were Conjecture \ref{conj} true, all partially hyperbolic diffeomorphisms $f$ with $\Omega (f)=M$ would be dynamically coherent. Hammerlindl and Potrie have proven that Conjecture \ref{conj} is true for manifolds with solvable fundamental group \cite{hp}; namely, for the manifolds that are finitely covered by the ones listed in Theorem \ref{invariant.tori}. This is the greatest advance in Conjecture \ref{conj} so far. Theorem \ref{uniqueint} is used in their proof. \newline \par
A foliation is {\em taut } if there is an embedded circle that transversely intersects each one of its leaves. Taut foliations play an important role in the description of 3-dimensional manifolds. In Section \ref{taut} we show:
\begin{theorem} \label{theorem.taut} Let $E^s$ be the strong stable bundle of a partially hyperbolic diffeomorphism of a closed orientable 3-dimensional manifold $M$. If $\mathcal{F}$ is a foliation transverse to $E^s$ that is not taut, then there exists a periodic $cu$-torus. In particular, $M$ admits an Anosov torus.
\end{theorem}
A foliation ${\mathcal F}$ like the one mentioned in Theorem \ref{theorem.taut} always exists, due to Burago-Ivanov \cite{burago_ivanov}, see more details in Theorem \ref{burago.ivanov}. As a consequence of Theorem \ref{theorem.taut}, all manifolds supporting partially hyperbolic diffeomorphisms are finitely covered by manifolds supporting taut foliations. Perhaps the theory of taut foliations could give some enlightening to the description of partially hyperbolic systems (see, for instance, \cite{calegari}).
\section{Dynamics on $cu$- and $sc$-tori}\label{section.cu.tori}
In this section, we shall prove Theorem \ref{invariant.tori}, which will follow from certain dynamical properties of $cu$- and $sc$-tori. \par
As we said in the Introduction, a manifold $M$ {\em admits an Anosov torus} if there exists a diffeomorphism $g:M\to M$ and an embedded $g$-invariant torus $T$ such that $(g|_{T})_*:\pi_1(\mathbb{T}^2)\rightarrow \pi_1(\mathbb{T}^2)$ is hyperbolic. Admitting an Anosov torus is a global property. In \cite{anosov_tori}, we prove that very few 3-manifolds have such a property.
\begin{theorem}\cite{anosov_tori} Let $M$ be an irreducible orientable 3-manifold admitting an Anosov torus, then
$M$ is either:
\begin{enumerate}
\item the 3-torus ${\mathbb T}^{3}$
\item the mapping torus of $-id:{\mathbb T}^{2}\to{\mathbb T}^{2}$
\item the mapping torus of a hyperbolic automorphism $A:{\mathbb T}^{2}\to {\mathbb T}^{2}$
\end{enumerate}
\end{theorem}
A 3-manifold admitting a partially hyperbolic diffeomorphism is always irreducible, see \cite[Lemma 6.3]{survey}. Theorem \ref{invariant.tori} then follows from the following propositions:
\begin{proposition}\label{cu.tori.periodic}
The existence of a $cu$-torus implies the existence of a periodic $cu$-torus.
\end{proposition}
\begin{proposition}\label{cu.tori.anosov}
The dynamics on an invariant $cu$-torus is isotopic to hyperbolic.
\end{proposition}
\begin{proposition}\label{su.torus.anosov}
A manifold admitting an $su$-torus, admits an Anosov torus.
\end{proposition}
Proposition \ref{cu.tori.anosov} is a direct corollary of Lemma \ref{hyperbolic} below. See also Proposition 2.1 of \cite{brin_burago_ivanov} for a similar result and proof.
\begin{lemma}\label{hyperbolic} Let $\mathcal{W}$ be a foliation of $\,\mathbb{T}^2$ with continuous tangent bundle $T\mathcal{W}$ and invariant by a diffeomorphism $g$. Suppose, in addition, that $||dg|_{T\mathcal{W}}||>1$. Then, $g_*:\pi_1(\mathbb{T}^2)\rightarrow \pi_1(\mathbb{T}^2)$ is hyperbolic.
\end{lemma}
\begin{proof}
By taking $g^2$ if necessary we can suppose that $g$ preserves the orientation of $T\mathcal{W}$. Since $g$ preserves a foliation without compact leaves, $g_*:\mathbb{Z}^2\rightarrow \mathbb{Z}^2$ (we identify $\pi_1(\mathbb{T}^2)$ with $\mathbb{Z}^2$) has an eingenspace of irrational slope. This implies that either $g_*$ is hyperbolic or $g_*=Id$. In the second case $g$ has a lift $\tilde g:\mathbb{R}^2\rightarrow \mathbb{R}^2$ such that $\tilde g=Id+\alpha$ where $\alpha$ is a periodic, and in particular bounded, function. As a consequence we obtain that there exists a constant $K>0$ such that given any subset of $\mathbb{R}^2$, $X$, $\operatorname{diam}(g^n(X))\leq \operatorname{diam}(X)+nK$. Let $\gamma$ be an arc contained in a leaf of $\mathcal{W}$. Then, the length of $\gamma$ grows exponentially while its diameter grows at most linearly. This implies that given a small $\varepsilon>0$ there exists an iterate of $\gamma$ that contains a curve of length arbitrarily large and with end points at distance less than $\varepsilon$. Using Poincar\'{e}-Bendixon we obtain a compact leaf. This is a contradiction and then, $g_*$ is hyperbolic.
\end{proof}
\begin{proof}[Proof of Proposition \ref{cu.tori.periodic}] Let $T$ be a $cu$-torus, and consider the sequence $f^{-n}(T)$. Since the family of all compact subsets of $M$, considered with the Hausdorff metric $d_{H}$, is compact, there is a subsequence $f^{-n_{k}}(T)$ converging to a compact set $K\subset M$. Therefore, for each $\eps>0$ there are arbitrarily large $N>>L>0$ such that $d_{H}(f^{-N}(T), f^{-L}(T))<\eps$. \par
Since $T$ is transverse to the stable foliation, the union of all local stable leaves of $T$ forms a small tubular neighborhood of $T$, $U(T)$. Since stable leaves grow exponentially under $f^{-1}$, if $N>>L$ as above are large enough, then $f^{-L}(\overline{U(T)})\subset f^{-N}(U(T))$. This implies that
$f^{N-L}(\overline{U(T)})\subset U(T)$.
The set $T_{0}=\cap_{k=0}^{\infty}f^{k(N-L)}(U(T))$ is a periodic $cu$-torus. Indeed, it is easy to see that $T_{0}$ is periodic and homeomorphic to a torus. On the other hand, for each point $x$ of $T_{0}$, its tangent space is limit of the tangent spaces of points $x_{k}$ in $f^{k(N-L)}(T)$, which are $cu$-tori. Hence $T_{x}T_{0}=E^{c}_{x}\oplus E^{u}_{x}$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{su.torus.anosov}] Assume $f$ admits an $su$-torus, and consider the lamination $\Lambda$ of all $su$-tori of $f$. This is a compact lamination \cite{Hae62}. Therefore, there is a recurrent leaf., that is, there is a torus $T$ and an iterate $n$, such that $d_{C^{1}}(f^{n}(T),T)<\eps$ for small $\eps$. There exists a dipheotopy $i_{t}$ on $M$, taking $f^{n}(T)$ into $T$. Then $\phi=f^{n}\circ i_{1}$ fixes $T$ and $\phi|T$ is isotopic to an Anosov diffeomorphism.
\end{proof}
\section{Weak foliations of partially hyperbolic diffeomorphisms}\label{taut}
In this section we prove Theorem \ref{theorem.taut}. For any partially hyperbolic diffeomorphism of a 3-manifold such that the invariant bundles are orientable, Burago and Ivanov \cite{burago_ivanov} have proved that there are (not necessarily invariant) foliations that ``almost" integrate $E^{c\sigma}=E^c\oplus E^\sigma, \,\,\, \sigma=s,\,u$.
\begin{theorem}[Key Lemma 2.2, \cite{burago_ivanov}]\label{burago.ivanov} Let $f$ be a partially hyperbolic diffeomorphism of a closed 3-manifold and let $E^*$ be orientable for $*=s,\, c, \, u$.
Then, for every $\varepsilon> 0$ there is a foliation $\mathcal{F}^{c\sigma}_\varepsilon$ such that $T\mathcal{F}^{c\sigma}_\varepsilon$
is a continuous bundle and the angles between $T\mathcal{F}^{c\sigma}_\varepsilon$ and $E^{c\sigma}$ are no greater than
$\varepsilon$, $\sigma=s,\, u$.
\end{theorem}
In this section we prove that these foliations, if the manifold is different from the ones listed in Theorem \ref{invariant.tori}, are taut. Recall that a codimension one foliation is {\em taut} if there exists an embedded $\mathbb{S}^1$ that intersects transversely, and nontrivially, every leaf of the foliation (see \cite{calegari}).\newline\par
Let ${\mathcal F}$ be a codimension-one foliation. A {\em dead-end component} is an open submanifold
$N \varsubsetneq M$ which is a union of leaves of $\mathcal{F}$, such that there is
no properly immersed line transverse to $\mathcal{F}$. That is, there is no $\alpha:[a,b]\to M$ transverse to ${\mathcal F}$ such that
$\alpha(a,b)\subset N$, $\alpha(a),\alpha(b)\in\partial N$.
A {\em Reeb
component} is a solid torus whose interior is foliated by planes transverse
to the core of the solid torus, such that each leaf limits on the
boundary torus, which is also a leaf. Observe that the interior of a Reeb component is a dead-end component.
Other examples of dead-end components are obtained by taking a Reeb foliation of the annulus, multiplying by the interval and gluing the two boundary annulus using a rotation that preserves the Reeb foliation.
Observe that in both cases the boundary of the dead-end component consists of tori. This is a general fact that is stated in the following lemma.
\begin{lemma}[Lemma 4.28, \cite{calegari}]\label{nodead} Let $M$ be a 3-dimensional orientable closed manifold. A foliation $\mathcal{F}$ of $M$ is taut if and only if it contains no dead-end components. If
$N$ is a dead-end component, then the restriction of $\mathcal{F}$ to $\overline N$ is transversely orientable and $\overline N\setminus N$ consists of a union of tori leaves of F. Moreover, boundary leaves of $N$ cannot be joined by an arc in $N$ transverse to $\mathcal{F}$.
\end{lemma}
Observe that the last assertion implies that the boundary leaves of any dead-end component of a foliation have half-neighborhoods in $N$ with size uniformly bounded by below.\par
It is an obvious corollary of Lemma \ref{nodead} that foliations without compact leaves are taut. For instance, the weak stable and weak unstable foliations of Anosov flows of 3-dimensional manifolds are taut.
\begin{proof}[Proof of Theorem \ref{theorem.taut}]
Partial hyperbolicity implies that the forward iterates of $T\mathcal{F}$ converge to $E^{cu}$. Let $N$ be a dead-end of $\mathcal{F}$ and let $T$ be a boundary component of $N$. Hence $T$ is a torus, transverse to $E^{s}$. All iterates $f^{-n}(T)$ are tori transverse to $E^{s}$. The proof follows now exactly as in Proposition \ref{cu.tori.periodic}.
\end{proof}
\section{Proof of Theorem \ref{uniqueint}}\label{s_teo2}
Let us suppose that there exists an invariant foliation $\mathcal{F}^{cu}$ tangent to $E^{cu}$, and that $\mathcal{F}^{cu}$ has a compact leaf, which must be a torus. (Then $M$ must be one of the manifolds listed in Theorem \ref{invariant.tori}).
By Proposition \ref{cu.tori.periodic}, there exists a periodic $cu$-torus $T$. By taking an iterate, we can assume that is fixed. The dynamics on $T$ is isotopic to hyperbolic, due to Proposition \ref{cu.tori.anosov}. In \cite{anosov_tori} it is shown that cutting $M$ along $T$ we obtain a manifold with boundary, that is diffeomorphic to $\mathbb{T}^2\times [0,1]$. Moreover, $f$ induces a diffeomorphism $g$ of $\mathbb{T}^2\times [0,1]$ isotopic to $A\times id$ where $A$ is a hyperbolic automorphism of $\mathbb{T}^2$ and $id$ is the identity map of the interval $[0,1]$. Then, \cite{fr} implies that there exists a semiconjugacy $h:\mathbb{T}^2\times [0,1] \rightarrow \mathbb{T}^2$, between $g $ and $A$, homotopic to the projection $p:\mathbb{T}^2\times [0,1] \rightarrow \mathbb{T}^2$. Observe also that $h(\mathbb{T}^2\times \{0\})=\mathbb{T}^2$.\newline\par
The torus $\mathbb{T}^2\times \{0\}$ is foliated by a foliation $\mathcal{S}^u$ by lines that are integral curves of the strong unstable foliation. Call $\tilde{\mathcal{S}}^u$ the lift of $\mathcal{S}^u$ to $\mathbb{R}^2$, the universal cover of $\mathbb{T}^2\times \{0\}$. It is not difficult to see that if $x, \,\, y$ are in the same leaf $S^u $ of $\tilde{\mathcal{S}}^u$ the fact that $\operatorname{dist}_u(x,y)$ goes to infinity implies that $dist(x,y)$ goes to infinity ($\operatorname{dist}_u(x,y)$ is the length of the arc of $S^u $ joining $x$ and $y$). Let $\bar h$ be a lift of $h|_{\mathbb{T}^2\times \{0\}}$ to $\mathbb{R}^2$. Since the diameters of the sets $\bar h^{-1}(y)$ are uniformly bounded and $f(\bar h^{-1}(y))=\bar h^{-1}(A y)$, we obtain that the map $\bar h$ is injective when restricted to strong unstable manifolds. Recall that the image of an unstable manifold by $h$ is an unstable manifold of $A$. \newline\par
Let us show that the image of a center curve by $h$ is contained in a stable manifold of $A$. For this, it is enough to show that the length of the forward iterates of the curve are bounded. Let $\gamma$ be a (small) center curve and let $\delta $ be such that $W^u_\delta(x)\cap W^u_\delta(y)=\emptyset$ for all $x\ne y \in \gamma$. Let $W^u_\delta(\gamma)=\bigcup\{W^u_\delta(x);\, x\in \gamma\}$. Since $f$ expands the unstable bundle we have that $W^u_\delta(f^n(\gamma))\subset f^n(W^u_\delta(\gamma))$.
But since the angle between the center bundle and the unstable bundle is bounded by below, we have that the area of $W^u_\delta(f^n(\gamma))$ goes to infinity as the length of $f^n(\gamma)$ goes to infinity which is a contradiction. Moreover, this implies that $E^c$ is uniquely integrable for $f|_T$. If this were not the case there would exist, in the universal cover, two center curves $\gamma_1,\,\,\gamma_2$ beginning at the same point $x $ and cutting a nearby unstable manifold at two different points $y,\, z$. By forward-iterating this ``triangle" we would obtain that the distance between $f^n(y)$ and $f^n(z)$ goes to infinity. Then, either the length of $\gamma_1 $ or the length of $\gamma_2$ goes to infinity, which contradicts what we have proved before. Summarizing, $f|_T$ has two invariant foliation by lines, one tangent to the unstable bundle and the other one tangent to the center bundle; the semiconjugacy sends the unstable leaves to unstable lines of $A$ and the center leaves to stable lines of $A$. Again the distance between two points in the same center leaf in the universal cover of $T$ goes to infinity as the length of the center curve goes to infinity. It is not difficult to see that this implies that, for any $y\in \mathbb{T}^2$, $h^{-1}(y)\cap T$ is a connected arc contained in a center leaf.\newline\par
Let $p\in T$ be a periodic point. We may assume that $J_p=h^{-1}(h(p))\cap T$ is a very small arc. This is easy to obtain since there are infinitely many periodic points in different center curves. Let $U\subset \mathbb{T}^2\times I$ be a small neighborhood of $J_p$, and let $y\in h^{-1}(h(p))\cap U$. We will prove that $y\in W^{cs}_{loc}(p)$. On one hand, if $y\notin W^{cs}_{loc}(p)$ then $z=W^s_{loc}(y)\cap T$ is not in the local center manifold of $p$. On the other hand, $h(z)\in h(W^s_{loc}(y))\subset W^s_{loc}(h(p))$ is in the local stable manifold of $h(p)$ for $A$, which contradicts the fact that it is not in the local center manifold (in $T$) of $p$. \par
Let us focus on $W^{cs}_{loc}(p)$. The intersection of the center unstable foliation $\mathcal{F}^{cu} $ with $W^{cs}_{loc}(p)$ foliates $W^{cs}_{loc}(p)$ by center arcs.
By continuity, any of these center arcs, close enough to $T$, has a point of $h^{-1}(h(p))$. Certainly, the same is true for $\bar p$, a lift of $p$ to universal cover. We choose $x_1,\dots, x_N\in W^{cs}_{loc}(\bar p)$ points that are in $\bar h^{-1}(\bar h(\bar p))$ and in different center curves. Let $C>0$ be such that $\operatorname{diam} (\bar h^{-1}(y))<C$ for every $y\in T$ and $\varepsilon>0$ so small that if two points are at a distance less than $\varepsilon$, then they are in a trivializing chart of $\mathcal{F}^{cu}$. Now, we choose $N$ in such a way that if $N$ points are contained in a set of diameter $C$ then, at least two of them are a distance less than $\varepsilon$. Since $\bar f^n(\bar h^{-1}(\bar h(\bar p)))=\bar h^{-1}(\bar h(\bar f^n(\bar p)))$ we have that $\operatorname{diam} \{\bar f^n(x_1),\dots ,\bar f^n (x_n)\}<C, \,\, \forall n\in \mathbb{Z}$. Then, there exists a subsequence $n_k\rightarrow -\infty$ and two different points $x_i$ and $x_j$ such that $\operatorname{dist}(\bar f^{n_k}(x_i),\bar f^{n_k}(x_j))<\varepsilon, \,\, \forall k>0$. Now, take an arc $\alpha$ joining $x_i$ and $x_j$ that consists of two sub-arcs, $\alpha_1$ beginning at $x_i $ and tangent to the center bundle and $\alpha_2$ ending at $x_j$ and tangent to stable one (see Figure \ref{figure3}).
\begin{figure}
\begin{tikzpicture}
\draw (-3,-3) -- (5,-3);
\draw (-3,-2.75) -- (5,-2.75);
\draw (-3,-2.5) -- (5,-2.5);
\draw (-3,-2.25) -- (5,-2.25);
\draw (-3,-2) -- (5,-2);
\draw (-3,-1.75) -- (5,-1.75);
\draw (-3,-1.5) -- (5,-1.5);
\draw (-3,-1.25) -- (5,-1.25);
\draw (-3,-1) -- (5,-1);
\draw (-3,-0.75) -- (5,-0.75);
\draw (-3,-0.5) -- (5,-0.5);
\draw (-3,-0.25) -- (5,-0.25);
\draw (-3,0) -- (5,0);
\draw (-3,-3.25) -- (5,-3.25);
\draw[line width=2pt] (-1,-3.25) -- (1,-3.25);
\draw[line width=2pt] (0,-2.5) -- (4,-2.5)-- (4,-0.5);
\fill (0,-2.5) circle (3pt);
\fill (4,-0.5) circle (3pt);
\node at (0,-2.80) {$x_i$};
\node at (2,-2.15) {$\alpha_1\subset W^c(x_i)$};
\node at (4,-0.2) {$x_j$};
\node at (5.5,-1.2) {$\alpha_2\subset W^s(x_j)$};
\node at (0,-3.75) {$J_{\bar p}$};
\end{tikzpicture}
\caption{\label{figure3}
$W^{cs}(\bar p)$}
\end{figure}
For $k$ large enough we obtain that $\bar f^{n_k}(\alpha_2)$ is a very long stable curve, $\bar f^{n_k}(\alpha_1)$ is contained in a leaf of $\tilde {\mathcal{F}}^{cu}$ and the extremes of $\bar f^{n_k}(\alpha)$ are a distance less than $\varepsilon$ (see Figure \ref{figure_reeb}).
\begin{figure}[h]
\vspace{-7cm}
\begin{tikzpicture}
\draw (1.3,-1) .. controls +(-110:7cm) and +(100:10cm) .. (3,0);
\node at (0,1) {$\bar f^n(\alpha_2)$};
\draw (3,0) -- (1.7,-0.1);
\node at (2.2,0.3) {$\bar f^n(\alpha_1)$};
\fill (1.3,-1) circle (2pt);
\node at (2,-1) {$\bar f^n(x_2)$};
\fill (1.7,-0.1) circle (2pt);
\node at (1,-0.1) {$\bar f^n(x_1)$};
\draw[thick,dotted] (0.8,-0.7) -- (4.3,-0.7) -- (5.8,1.5) -- (2.3,1.5) -- cycle;
\node at (4.2,1) {$\bar F^{cu}(\bar f^n(x_1))$};
\end{tikzpicture}
\vspace{-4.5cm}\caption{\label{figure_reeb}
$\bar f^n(\alpha)$}
\end{figure}
Standard arguments of foliation theory imply that there is a closed curve transverse to $\tilde {\mathcal{F}}^{cu}$ which implies the existence of a Reeb component which is, as it is well known (see for instance \cite{burago_ivanov,rru_nilman}), impossible. This finishes the proof of Theorem \ref{uniqueint}.
|
\section{Introduction}
Recall that a cosine function taking values in a unital normed algebra $A$ with unit element $1_A$ is a family $C=(C(t))_{t\in \R}$ of elements of $A$ satisfying the so-called d'Alembert equation
\begin{equation}C(0)=1_A, C(s+t)+C(s-t)=2C(s)C(t) \ \ (s\in \R, t\in \R).\end{equation}
Equality (1) is also used to define $G$-cosine families $C=(C(g))_{g\in G}$ over an abelian group $G,$ and in particular cosines sequences $(C_n)_{n\in \Z}.$
A cosine function $C= (C(t))_{t\in \R}$ is said to be bounded if there exists $M>0$ such that $\Vert C(t)\Vert \le M$ for every $t \in \R.$ In this case we set
$$\Vert C \Vert_{\infty} =sup_{t \in R}\Vert C(t)\Vert, \ dist(C_1,C_2)=\Vert C_1-C_2\Vert_{\infty}.$$
A cosine function is said to be scalar if $C(t)\in \C.1_A$ for $t \in \R.$ It is well-known and easy to see that a bounded complex-valued cosine function $c$ takes values in $[-1,1]$ and that a bounded continuous complex-valued cosine function satisfies $c(t)=cos(at)$ where
$a=2lim_{t\to 0}{1-c(t)\over t^2} \in \R.$ More generally if $X$ is a Banach space, and if $C=(C(t))_{t\in \R}$ is a strongly continuous ${\mathcal B}(X)$-valued cosine function, then the generator $a$ of $C$ is bounded if and only if the cosine function $C$ is continuous at $0$ with respect to the operator norm on ${{\mathcal B}(X)},$ and in this situation $C(t)=cos(ta),$ where $cos(ta)$ is defined by the usual series, see for example \cite{so}.
Strongly continuous operator valued cosine functions are a classical tool in the study of differential equations, see for example \cite{abhn}, \cite{beh}, \cite{n}, \cite{tw}, and a functional calculus approach to these objects was developped recently in \cite{h1}.
Bobrowski and Chojnacki proved recently in \cite{bc} that if the set of all bounded strongly continuous cosine
functions on a Banach space $X$ is treated as a metric space under the operator norm on ${\mathcal B}(X),$
then the isolated points of this set are precisely the scalar cosine
functions.
They also gave a quantitative version of this result, by showing that if a strongly continuous operator valued cosine function on a Banach space $C(t)$ satisfies $sup_{t\ge 0}\Vert C(t) -c(t)\Vert <1/2$ for some scalar bounded continuous cosine function $c(t)$ then $C(t)=c(t)$ for $t \in \R,$ and Schwenninger and Zwart showed in \cite{zs} that this result remains valid under the condition $sup_{t\ge 0}\Vert C(t) -c(t)\Vert <1.$ An elementary proof of this result is given by Chojnacki in \cite{c}.
The purpose of this paper is to show that this result holds when $sup_{t\ge 0}\Vert C(t) -c(t)\Vert <{8\over 3\sqrt 3},$ which is the optimal constant since $sup_{t\in \R}\vert cos(at)-cos(3at)\vert ={8\over 3\sqrt 3}$ for every $a\neq0,$ and that no continuity condition on $C$ is needed if the scalar bounded cosine function $c$ is assumed to be continuous, see theorem 3.6 (ii). In fact, this $"0-{8\over 3\sqrt 3}$ law" was already proved in a very recent paper by Bobrowski, Chojnacki and Gregoriewicz \cite{bcg}, which appeared after the present paper was submitted. The methods developed here vary in various aspects from those used by
Bobrowski et al.
We also point out that the 'generic' distance between two continuous real-valued bounded cosine functions is 2, a consequence of the fact that finite independent subsets of the torus are "Kronecker sets": if $a\neq 0,$ then the set $\Omega(a,m):=\{ b\ge 0 \ \ sup_{t\in \R} \vert cos(at)-cos(bt)\vert \le m\}$ is finite, and every $b\in \Omega(a,m)$ has the form $b={pa\over q},$ where $p,q$ are odd, $gcd(p,q)=1$, $1\le p \le {\pi\over arccos(m-1)}, 1\le q \le {\pi\over arccos(m-1)}.$
This description of the set $\Omega(a,m)$ leads to a description of cosine functions $C=(C(t))_{t\in \R}$ in a Banach algebra $A$ satisfying $sup_{t\in \R}\Vert C(t) -c(t)\Vert =m$ when $m<2.$ In this case we show in theorem 3.6 (i) that there exists $k \le card \left ( \Omega(a,m)\right )$ such that the closed subalgebra $A_1$ generated by $C$ is isomorphic to $\C^k,$ and we also show that there exists a family $p_1,\dots, p_k$ of pairwise orthogonal idempotents of $A_1$ and a family $(b_1,\dots,b_k)$ of distinct elements of $\Omega(a,m)$ such that we have
$$C(t)=\sum \limits_{j=1}^kcos(b_jt)p_j.$$
This implies in particular that if a cosine family $(C(t))_{t\in \R}$ satisfies $sup_{t\in \R}\Vert C(t)-1_A\Vert <2,$ then $C(t)=1_A$ for $t \in \R.$ This result was proved very recently
by Schwenninger and Zwart in \cite{zs1} for strongly continuous cosine families of bounded operators, but the general case seems new.
The description of the set $\Omega(a,m)$ pertains to folklore, but the operator theoretical part of the proofs seems new. It is based on the fact that every bounded cosine function $(C(t))_{t\in \R}$ taking values in a commutative unital Banach algebra having a unique maximal ideal is scalar, see
theorem 2.3 and corollary 2.4 in section 2.
Since $sup_{n\ge 1}\left \vert 1 -cos\left ({2n\pi\over 3}\right )\right \vert={3\over 2},$ the constant ${8\over 3\sqrt 3}$ does not work for cosine sequences. It is nevertheless possible to show that if a $G$-cosine family $C$ in a unital Banach algebra $A$ satisfies $sup_{g\in G}\left \Vert C(g)-c(g)\right \Vert <{\sqrt 5\over 2}$ for some bounded scalar $G$-cosine family $c,$ then $C(g)=c(g)$ for $g\in G.$ and the constant ${\sqrt 5\over 2}= max_{n\ge 1}\left | cos \left ( {n\pi \over 5} \right )-cos\left ( {3n\pi \over 5} \right )\right |$ is obviously optimal. Details will be given elsewhere.
The author would like to give his very warm thanks to the referee for his insightful comments and corrections.
\section{Cosine sequences in commutative local Banach algebras}
Set $f(x)=arccos(x).$ Then $f'(x)=-\frac{1}{\sqrt{1-x^2}},$ and we have, for $x\in ]-1,1[$
$$\frac{1}{\sqrt{1-x^2}}=1+\sum\limits_{n=1}^{+\infty}\frac{(-1)^n}{n!}\left(-\frac{1}{2}\right )\left (- \frac{1}{2}-1\right )\dots \left (- \frac{1}{2}-n+1\right )x^{2n}$$ $$=
1+\sum \limits_{n=1}^{+\infty}\frac{(2n)!}{2^{2n}n!^2}x^{2n}.$$
Hence we have, for $x\in (-1,1),$ with the convention $0!=1,$
$$arccos(x)=\frac{\pi}{2} -\sum\limits_{n=0}^{+\infty}\frac{(2n)!}{2^{2n}(2n+1)n!^2}x^{2n+1}.$$
It follows from example from Stirling's formula and Riemann's criterion that the series $\sum\limits_{n=0}^{+\infty}\frac{(2n)!}{2^{2n}(2n+1)n!^2}$ is convergent, and it follows from Abel's lemma that the power series expansion of $arccos(x)$ remains valid for $x=1$ and $x=-1.$ Since $arccos(-1)=\pi,$ we have
$$\sum\limits_{n=0}^{+\infty}\frac{(2n)!}{2^{2n}(2n+1)n!^2}=\frac{\pi}{2}.$$
Now let $A$ be a unital Banach algebra of unit element $1_A.$ We will write $\lambda=\lambda.1_A$ when $\lambda$ is scalar if there is no risk of confusion. We define $e^x$ by the usual series and set, for $x\in A,$
$$cos(x)=\frac{e^{ix}+e^{-ix}}{2}=\sum_{n=0}^{+\infty}(-1)^n\frac{x^{2n}}{(2n)!},$$
so that $cos(x+i\pi n)=(-1)^ncos(x)$ for $n\in \Z.$ When $z\in \C,$ this gives the usual cosine function of a complex variable $z.$
Let $\widehat A$ be the Gelfand space of $A,$ i.e. the space of all algebra homomorphisms from $A$ onto $\C.$ Notice that if $\chi \in \widehat A$ we have
$$\chi(cos(x))=cos(\chi(x)).$$
If sup$_{n\ge 1}\Vert x^n\Vert \le M <+\infty$, then the series $\frac{\pi}{2} -\sum\limits_{n=0}^{+\infty}\frac{(2n)!}{2^{2n}(2n+1)n!^2}x^{2n+1}$ is convergent,
we can set
$$arccos(x)=\frac{\pi}{2} -\sum\limits_{n=0}^{+\infty}\frac{(2n)!}{2^{2n}(2n+1)n!^2}x^{2n+1},$$
and we have
\begin{equation} \Vert arccos(x)\Vert \le \frac{M+1}{2}\pi, \chi(arccos(x))=arccos(\chi(x)) \ \ (\chi \in \widehat A).\end{equation}
Also it follows from standard properties of the holomorphic functional calculus that $cos(arccos(\lambda x))=\lambda x$ for $\vert \lambda \vert <1.$ By continuity, we obtain the tautological formula
\begin{equation}cos(arccos(x))=x.\end{equation}
\begin{proposition} Let $A$ be a unital Banach algebra, and let $(c_n)_{n\in \Z}\subset A$ be a cosine sequence. If sup$_{n\ge 1}\Vert c_n\Vert \le M <+\infty,$ then sup$_{p\ge 1}\Vert c_n^p\Vert\le M$ for $n\ge 1,$
and we have
\begin{equation}\Vert arccos(c_n)\Vert \le \frac{M+1}{2}\pi.\end{equation}
Moreover $Spec(c_n)\subset [-1,1]$ for every $n \in \N,$ and for every character $\chi$ on $A$ we have
$$\chi(c_n)=cos(n\beta_{\chi}) \ (n\ge 1),$$
where $\beta_{\chi}=\chi(arccos(c_1))=arccos(\chi(c_1))\in [0,\pi].$
\end{proposition}
Proof: Let $p\ge 1,$ and assume that we have
\begin{equation}c_1^p=\sum\limits _{k=0}^p\alpha_{k,p}c_k,\end{equation}
where $\alpha_{k,p} \ge 0,$ $\sum \limits_{k=0}^p\alpha_{k,p}=1,$ which is trivially true for $p=1.$ Using (1), we obtain
$$c_1^{p+1}=\sum_{k=0}^p\alpha_{k,p}c_1c_k = \alpha_{0,p}c_1 +\sum\limits_{k=1}^p{\alpha_{k,p}\over 2}c_{k-1}+\sum\limits _{k=1}^p{\alpha_{k,p}\over 2}c_{k+1}$$ $$\alpha_{0,p}c_1 +\sum\limits_{k=0}^{p-1}{\alpha_{k+1,p}\over 2}c_{k}+\sum\limits _{k=2}^{p+1}{\alpha_{k-1,p}\over 2}c_{k}=\sum \limits_{k=0}^{p+1}\alpha_{k,p+1}c_k,$$
where $$\alpha_{0,p+1}={\alpha_{1,p}\over 2}, \alpha_{1,p+1}=\alpha_{0,p}+{\alpha_{2,p}\over 2}, \alpha_{k,p+1}={\alpha_{k-1,p}+\alpha_{k+1,p}\over 2} \ \mbox{for}\ 2\le k\le p-1,$$
$$\alpha_{p,p+1}={\alpha_{p-1,p}\over 2}, \alpha_{p+1,p+1}={\alpha_{p,p}\over 2}.$$
Clearly, $\alpha_{k,p+1}\ge 0,$ and $\sum\limits _{k=0}^{p+1}\alpha_{k,p}=1.$ We thus see that (5) holds for every $p\ge 1.$
Hence $sup_{p\ge 1}\Vert c_1^p\Vert \le m.$ Applying this result to the cosine sequence $(c_{nm})_{m\in \Z},$ we see that $sup_{p\ge 1}\Vert c_n^p\Vert \le M$ for every $n\ge 1.$ Inequality (4) follows then from (2).
If $\chi$ is a character on $A,$ then the sequence $(\chi(c_n))_{n\ge 1}$ is a bounded complex-valued cosine sequence. Hence $\chi(c_n) \in [-1,1]$ for $n\ge 1,$ and we have $c_n=cos(n\beta)$ for $n\ge 1,$ where $\beta$ is any real number satisfying $c_1=cos(\beta).$ This holds in particular when $\beta=\beta_{\chi}=\chi(arccos(c_1))=arccos(\chi(c_1))\in [0,\pi].$
$\square$
Since $cos(arccos(c_1))=c_1,$ we have $cos(narccos(c_1))=c_n,$ and we obtain an alternative approach to the group decomposition given in \cite{c} by setting $\nu=e^{iarccos(c_1)}.$
Now define the sine function and the "cardinal sine" function on a unital Banach algebra $A$ by the usual formulae
$$sin(x)=\sum_{n=0}^{\infty}(-1)^n\frac{x^{2n+1}}{(2n+1)!}, sinc(x)=\sum_{n=0}^{\infty}(-1)^n\frac{x^{2n}}{(2n+1)!},$$
so that $sinc(0)=1,$ $xsinc(x)=sin(x).$ We have again $\chi(sin(x))=sin(\chi(x))$ and $\chi(sinc(x))=sinc(\chi(x))$ for $\chi \in \widehat A.$ For $x,y \in A$ such that $yx=xy,$ we have the usual formula
$$cos(x)-cos(y)=2sin\left ( \frac{y-x}{2}\right )sin\left ( \frac{x+y}{2}\right ).$$
Recall $x \in A$ is said to be quasinilpotent if lim$_{n\to +\infty}\Vert x^n\Vert=0,$ which is equivalent to the fact that $\chi(x)=0$ for every $\chi \in \widehat A$ if $A$ is commutative.
\begin{lemma} Let $A$ be a commutative Banach algebra, let $x\in A$ and $y\in A$ be two quasinilpotent elements of $A$ and let $\lambda \in \C.$
(i) If $\lambda \notin \pi\Z,$ and if $cos(\lambda.1_A +x)=cos(\lambda.1_A +y),$ then $x=y.$
(ii) If $\lambda \in \pi \Z,$ and if $cos(\lambda.1_A +x)=cos(\lambda.1_A +y),$ then $x^2=y^2.$
\end{lemma}
Proof: If $cos(\lambda.1_A +x)=cos(\lambda.1_A +y),$ we have
$$sin\left ( \frac{y-x}{2}\right) sin\left (\lambda +\frac{x+y}{2}\right )=0.$$
If $\lambda \notin \pi\Z,$ we have, for $\chi \in \widehat A,$
$$\chi \left (sin\left (\lambda.1_A +\frac{x+y}{2}\right )\right) =sin(\lambda)\neq 0,$$
and so $sin\left (\lambda.1_A +\frac{x+y}{2}\right )$ is invertible and further
$$(x-y)sinc\left ( \frac{y-x}{2}\right)=2sin\left ( \frac{y-x}{2}\right)=0.$$
Since $\chi \left (sinc\left ( \frac{y-x}{2}\right)\right )=sinc(0)=1$ for every $\chi \in \widehat A,$ $sinc\left ( \frac{y-x}{2}\right)$ is invertible and $x=y.$
If $\lambda \in \pi\Z,$ then we have $cos(x)=cos(y),$ which gives
$$(y^2-x^2)sinc\left ( \frac{y-x}{2}\right) sinc\left (\frac{x+y}{2}\right)=4sin\left ( \frac{y-x}{2}\right) sin\left (\frac{x+y}{2}\right )=0.$$
We see again that $sinc\left ( \frac{y-x}{2}\right)$ and $sinc\left ( \frac{x+y}{2}\right)$ are invertible, which shows that $x^2=y^2.$
\begin{theorem} Let $(c_n)_{n\in \Z}$ be a bounded cosine sequence in a Banach algebra $A,$ and assume that $spec(c_1)$ is a singleton. Then $(c_n)_{n\ge 1}$ is a scalar sequence, and there exists
$b\in \R$ such that $c_n=cos(nb).1_A$ for $n\ge 1.$
\end{theorem}
Proof: Set $M=sup_{n\ge 1}\Vert c_n\Vert.$ We can assume that $A$ is a commutative Banach algebra generated by $c_1,$ so that $\widehat A$ consists of a single character $\chi.$ Let $\lambda_n=\chi(c_n)$ be the unique element of $spec(c_n),$ and set $\beta=\beta_{\chi}=arccos(\chi(c_1))=\chi(arccos(c_1)),$ so that $\lambda_n=cos(n\beta)$ for $n\ge 1.$
Set $x_n=arccos(c_n).$ Since $c_n=cos(x_n),$ we have
$cos(n\beta)=cos(\chi(x_n)).$ It follows then from standard properties of the cosine function on $\C$ that there exists $k_n\in \Z$ such that $\chi(x_n)=\pm n\beta +2k_n\pi,$
and we have $\chi(x_1)=\beta.$
If $\chi(x_n)=n\beta +2k_n\pi,$ set $y_n=x_n-2k_n\pi.1_A,$ and if $\chi(x_n)=-n\beta +2k_n\pi$ set $y_n=-x_n +2k_n\pi.1_A,$ with the convention $y_n=x_n-2k_n\pi.1_A$ when $\beta=0.$ Then $\chi(y_n)=n\beta=\chi(nx_1),$ and $cos(y_n)=cos(x_n)=c_n
=cos(nx_1).$ Since $\widehat A$ is a singleton, $y_n-n\beta.1_A$ and $nx_1-n\beta.1_A$ are quasinilpotent. When $\beta \neq 0$ and $\beta \neq \pi,$ it follows from item (i) of Lemma 2.2 that $y_n-n\beta.1_A=nx_1-n\beta.1_A,$ and hence $y_n=nx_1.$
But $\Vert y_n-n\beta.1_A\Vert =\Vert y_n-\chi(y_n).1_A\Vert =\Vert \pm(x_n-\chi(\pm x_n).1_A)\Vert \le (1+\pi)M.$ Hence, given that $\Vert y_n-n\beta.1_A\Vert =n\Vert x_1-\beta.1_A\Vert$, we see that $x_1=\beta.1_A,$ $(c_n)_{n\ge 1}$ is a scalar sequence, and $c_n=cos(n\beta).1_A$ for $n \ge 1.$
If $\beta=0,$ then $\chi(x_n)\in [0,\pi]\cap2\pi\Z=\{0\},$ $y_n=x_n$ and $nx_1$ are quasinilpotent, and it follows from item (ii) of Lemma 2.2 that $x_n^2=n^2x_1^2.$ Since the sequence $(x_n^2)_{n\ge 1}$ is bounded, we have $x_1=0,$ $c_1=1_A$ and so $c_n=1_A$ for every $n \ge 1.$ If $\beta =\pi,$ set $c'_n=c_{2n}$ for $n\ge 1,$ and set $\beta'=arccos(\chi(c'_1)).$ Then $(c'_n)_{n\ge 1}$ is a cosine sequence. Since $\beta'=0,$ we have $2c_1^2-1_A=c_2=c'_1=1_A,$ and $(c_1-1_A)(c_1+1_A)=0.$ Since $scpec(c_1)=\{-1\},$ $c_1-1_A$ is invertible, $c_1=-1_A,$ and $c_n=(-1)^n.1_A=cos(n\pi).1_A$ for $n \ge 1.$ $\square$
Recall that a commutative unital Banach algebra $A$ is said to be local if it possesses a unique maximal ideal. We obtain the following corollary.
\begin{corollary} Let $G$ be an abelian group, and let $C= (C(g))_{g\in G}$ be a bounded cosine family in a commutative unital local Banach algebra. Then $C$ is scalar, and so there exists a bounded cosine family $(c(g))_{g\in G}$ with values in $[-1,1]$ such that $C(g)=c(g).1_A$ for $g\in G.$
\end{corollary}
\section{When the distance to a bounded cosine function is strictly less than 2}
A standard result shows that every bounded complex cosine function $c$ takes values in $[-1,1]$. The following observation, which is the cosine counterpart of a standard result for discontinuous one-parameter unimodular groups, see \cite{hp}, section 4.17, is certainly well-known.
\begin{proposition} Let $(c(t))_{t\in \R}$ be a discontinuous bounded complex cosine function. Then for every $\alpha \in [-1,1]$ there exists a sequence $(t_n)_{n\ge 1}$ of positive real numbers such that $lim_{n\to +\infty}t_n=0$ and $lim_{n\to+\infty}c(t_n)=\alpha.$
\end{proposition}
Proof: The identity
$$(1-c(s-t))(1-c(s+t))=(c(s)-c(t))^2$$
shows as is well-known that this bounded cosine function with values in $[-1,1]$ is discontinous at $0.$ Denote by $G$ the set of all real numbers $x$ for which there exists a sequence $(t_n)_{n\ge 1}$ of positive reals such that $lim_{n\to +\infty}t_n=0$ and $lim_{n\to +\infty}c(t_n)=cos(x).$ Then $nG +2\pi\Z\subset G$ for every $n\in Z,$ $G$ is closed, and $G\neq 2\pi \Z.$
Let $x \in G\cap (0,\pi].$ If $x\over \pi$ is irrational, then the sequence $(e^{inx})_{n\ge 1}$ is dense in $\T,$ and so $G=\R.$ If $x\over \pi$ is rational let $u$ be the smallest positive integer such that $e^{iux}=1.$ Then $e^{2i\pi\over u}=e^{ipx}$ for some $p\ge 1,$ and so ${2\pi\over u}\in G.$ Let $(t_n)_{n\ge 1}$ be a sequence of positive reals converging to 0 such that $lim_{n\to +\infty}c(t_n)=cos\left ({2\pi\over u}\right ),$ let $q\ge 2,$ and let $\alpha$ be a limit point of the sequence $c\left ({t_n\over u^{q-1}}\right )_{n\ge 1}.$ There exists $y \in \R$ such that $cos(y)=\alpha,$ so that $y \in G,$ and such that $y= {2\pi\over u^q} +{2k\pi\over u^{q-1}}=(1+ku){2\pi\over u^q}.$ Then $gcd(1+ku,u^q)=1,$ and there exist a positive integer $r$ such that $ry - {2\pi\over u^q} \in 2\pi\Z,$ so that ${2\pi\over u^q}\in G.$ This implies that $G=\R.$ $\square$
\begin{corollary} Let $a\in \R,$ and let $(c(t))_{t\in \R}$ be a discontinuous bounded scalar cosine function. Then $sup_{t\in \R}\vert cos(at)-c(t)\vert=lim sup_{t\to 0}\vert cos(at)-c(t)\vert
=2.$
\end{corollary}
\begin{corollary} Let $X$ be a Banach space, let $(c(t))_{t\in R}$ be a scalar cosine function, and let $(C(t))_{t\in \R}$ be a bounded strongly continuous cosine family of bounded operators
on $X$ such that $sup_{t\in \R}\Vert C(t)-c(t)I_X\Vert <2.$ Then $c(t)$ is continuous, and so there exists $a\in \R$ such that $c(t)=cos(at)$ for $t\in \R.$
\end{corollary}
Proof: Let $x \in X$ such that $\Vert x \Vert=1.$ If $c(t)$ were discontinuous, there would exist a sequence $(t_n)_{n\ge 1}$ of positive real numbers such that $$lim_{n\to +\infty}t_n=lim_{n\to+\infty}c(t_n)+1=0,$$ which gives
$$sup_{t\in \R}\Vert C(t)-c(t)I_X\Vert \ge lim_{n\to +\infty}\Vert C(t_n)x -c(t_n)x\Vert=lim_{n\to +\infty}(1-c(t_n))\Vert x \Vert=2.$$
$\square$
The following observation is an easy consequence of Kronecker's theorem on independent finite subsets of the unit circle.
\begin{lemma} Let $a,b$ be two real numbers. If $pa-qb\neq 0$ for $(p,q) \in \Z^2\setminus \{0,0\},$ then $sup_{t\in \R}\vert cos(ta)-cos(tb)\vert =2.$
\end{lemma}
Since $sup_{t \in \R}\vert 1-cos(bt)\vert=2$ for every $b\neq 0,$ we can restrict attention to the case where $a\neq 0$ and $b\neq 0.$ Denote by $\T$ the unit circle. If $pa-qb \notin 2\pi \Z$ for $(p, q)\in \Z^2\setminus \{(0,0)\},$ then the set $\{(e^{ia},e^{ib})\}$ is independent, and it follows from Kronecker's theorem, see for example \cite{ks}, p. 21 that the set $\left \{(e^{ina},e^{inb})\right \}_{n\in \Z}$ is dense in $\T^2,$ which implies that $sup_{n\ge 1}\left \vert cos(na)-cos(nb)\right \vert =2.$
Now assume that $pa -qb =2k\pi$ for some $(p,q)\in \Z^2\setminus \{(0,0)\}$ and some $k\in \Z.$
If $p=0,$ then we have $b={2k\pi \over q},$ with $k\ge 1,$ $q\ge 1.$ If $a\over \pi$ is irrational, then $\{e^{inqa}\}_{n\ge 1}$ is dense in $\T,$ and we have
$$sup_{n\ge 1}\left | cos(naq) -cos(nqb)\right | = sup_{n\ge 1}\left\vert cos(naq)-1\right \vert=2.$$
Otherwise $a \over \pi$ is rational, and so is ${a\over b}.$
Now assume that $p\neq 0, q\neq 0,$ and $k\neq 0.$ If $a\over \pi$ is rational, then $b\over \pi$ is rational, and ${a\over b}$ is rational. Otherwise we have, since the set $\{e^{ian}\}_{n\ge 1}$ is dense in the unit circle,
$$sup_{t\in \R}\left | cos(bt) -cos(at)\right | \ge sup_{n\ge 1} \left | cos\left ( b(2kn+1){q\over 2k} \right ) -cos \left( a(2kn+1){q\over 2k} \right)\right |$$
$$=\left | cos\left (a (2kn+1){p\over 2k} -(2kn+1)\pi \right )- cos \left( a(2kn+1){q\over 2k} \right)\right |$$ $$=\left | cos\left ( \left (an+{a\over 2k}\right ){p} \right )+ cos \left( \left (an+{a\over 2k}\right ){q} \right)\right | =2.$$
Hence $k=0$ if $sup_{t\in \R}\vert cos(ta)-cos(tb)\vert <2.$ $\square$
\begin{lemma} Let $a \ge 0,$ and set \ $\Omega(a,m):=\left \{b \ge 0 \ | \ sup_{t\in \R} \left | cos(ta)-cos(tb) \right \vert \le m\right\}$ for $m\ge 0.$
(i) If $m<2,$ then $\Omega(a,m)$ is finite, and every $b\in \Omega(a,m)$ has the form $b={pa\over q},$ where $p,q$ are odd, $gcd(p,q)=1$, $1\le p \le {\pi\over arccos(m-1)}, 1\le q \le {\pi\over arccos(m-1)}.$
(ii) If $m<{8\over 3\sqrt 3},$ then $\Omega(a,m)=\{a\}.$
\end{lemma}
Proof: Assume that $m<2.$ Since $\Omega(0,m)=\{0\},$ we can assume that $a > 0.$ Let $b\in \Omega(a,m).$ Then $b \neq 0.$ We may restrict attention to the case where $a=1$, and there exists positive integers $p$ and $q$ such that $b={p\over q},$ with $gcd(p,q)=1.$ If $p$ or $q$ were even, we would have $sup_{t\in \R}\vert cos(t)-cos(bt)\vert \ge sup_{n\ge 1}\vert cos(np\pi)-cos(nq\pi)\vert =2,$ and so $p$ and $q$ are odd.
It follows from Bezout's theorem that there exists $(u,v) \in \Z$ such that $2up-(2v+1)q=1.$ We have
$$sup_{t\in \R}\vert cos(t)-cos(bt)\vert \ge cos(2u\pi)-cos\left ({2up\pi\over q}\right )=1- cos\left ((2v+1)\pi +{\pi\over q} \right )$$ $$=1+cos\left ({\pi\over q}\right ).$$
Hence $q \le {\pi\over arccos(m-1)}.$ Since $sup_{t\in \R} \vert cos(t)-cos(bt)\vert =sup_{t\in \R} \vert cos(t)-cos({t\over b})\vert,$ the same argument shows that $p\le {\pi\over arccos(m-1)}.$ This proves (i)
\smallskip
(ii) If $m< 1+cos\left ({\pi\over 5}\right )\approx 1,8090,$ then $arccos(m-1)>{\pi\over 5}$ and so every $b\in \Omega(a,m)$ can be written under the form $b={ap\over q},$ where $1\le p <5$ and $1\le q <5,$ with $p$ and $q$ odd, $gcd(p,q)=1$ and so $\Omega(a,m)\subset \left \{{a\over 3}, a, 3a \right \}.$
An elementary computation shows that $\vert cos(t) -cos(3t)\vert $ attains its maximum when $cos(t) =\pm {1\over \sqrt 3}$, which gives
$$max_{t \in \R}=\vert cos(t) -cos\left ({t/ 3}\right )\vert =max_{t \in \R}\vert cos(t) -cos(3t)\vert = {8\over 3\sqrt 3}\approx 1.5396.$$ Hence
$\Omega(a,m)=\left \{{a\over 3}, a, 3a \right \}$ if ${8\over 3\sqrt 3}\le m < 1+cos\left ({\pi\over 5}\right ),$ and $\Omega(a,m)=\{ a\}$ if $m<{8\over 3\sqrt 3}.$
$\square$
We obtain the following theorem.
\begin{theorem} Let $(C(t))_{t\in \R}$ be a cosine function in a unital Banach algebra, and let $A_1$ be the closed subalgebra of $A$ generated by $(C(t))_{t\in \R}.$
(i) If $m:=sup_{t\in \R}\left \Vert C(t)-cos(at).1_A\right\Vert <2$ for some $a\in \R,$ then $A_1$ is isomorphic to $\C^k$ for some $k\le card\left ( \Omega(a,m)\right ),$ and there exists a family $p_1,\dots,p_k$ of pairwise orthogonal idempotents of $A_1$ and a family $b_1,\dots, b_k$ of distinct elements of $\Omega(a,m)$ such that $C(t)=\sum \limits_{j=1}^kcos(b_jt)p_j$ for $t \in \R.$
(ii) If $sup_{t\in \R}\left \Vert C(t)-cos(at).1\right\Vert <{8\over 3\sqrt 3}$ for some $a\in \R,$ then $C(t)=cos(at).1_A$ for $t\in \R.$
\end{theorem}
Proof: (i) Let $\chi \in \widehat{A_1}.$ Then it follows from Corollary 3.2 that there exists $b_{\chi} \in \Omega(a,m)$ such that $\chi(C(t))=cos\left (tb_{\chi}\right )$ for $t\in \R.$
Since $b^2_{\chi}=2 lim_{t\to 0}{1-\chi(C(t))\over 2},$ the map $\chi \to b_{\chi}$ is one-to-one, and it follows from Lemma 3.5 that $\widehat A_1$ is finite.
Let $\chi_1, \dots, \chi_k$ be the elements of $\widehat{A_1}.$ It follows from the standard one-variable holomorphic functional calculus, see for example \cite{d}, that there exists for every $j\le k$ an idempotent $p_j$ of $A_1$ such that $\chi_j(p_j)=1$ and $\chi_i(p_j)=0$ for $i\neq j.$ Hence $p_jp_i=0$ for $j \neq i,$ and $1_A=1_{A_1}=\sum \limits _{j=1}^kp_j.$
Let $x\in A_1.$ Then $(p_jC(t))_{t\in \R}$ is a cosine function in the commutative unital Banach algebra $p_jA_1,$ $spec_{p_jA_1}(p_jC(1))=\{\chi_j(C(1))\},$ and $(cos(at)p_j)_{t\in \R}$ is a scalar cosine function in $p_jA_1.$ Since $sup_{t\in \R}\left \Vert cos(at) p_j-p_jC(t) \right \Vert \le 2 \Vert p_j\Vert,$ the cosine function $(p_jC(t))_{t\in \R}$ is bounded, and it follows from theorem 2.3 that $(p_jC(nt))_{n\in \Z}$ is a scalar cosine sequence for every $t\in \R.$ So $(p_jC(t))_{t\in \R}$ is a scalar cosine function, and $p_jC(t)=\chi_j(C(t)p_j=cos(b_jt)p_j,$ where $b_j=b_{\chi_j}\in \Omega(a,m).$
We obtain
$$C(t)=\sum \limits_{j=1}^kC(t)p_j=\sum \limits_{j=1}^kcos(tb_j)p_j \ (t\in \R).$$
Since the algebras $p_jA_1$ are one-dimensional, $A_1$ is isomorphic to $\C^k.$
(ii) If $m<{8\over 3\sqrt 3},$ then $\Omega(a,m)=\{a\},$ $k=1,$ $p_1=1_A$ and $C(t)=cos(at).1_A.$ $\square$
Since $\Omega(0,m)=\{0\}$ for every $m<2,$ we obtain the following result, which was obtained recently by Schwenninger and Zwart in \cite{zs1} for strongly continuous operator valued cosine functions. A very short argument to prove a weaker result with the constant $3/2$ instead of $2$ is given by Arendt in \cite{a}.
\begin{corollary} Let $(C(t))_{t\in \R}$ be a cosine function in a unital Banach algebra $A.$ If $sup_{t\in \R}\Vert C(t)-1_A\Vert <2,$ then $C(t)=1_A$ for $t\in \R.$
\end{corollary}
Let $C=(C(t))_{t\in \R}$ be a bounded strongly continuous cosine family of bounded operators
on a Banach space $X.$ Bobrowski and Chojnacki observed in \cite{bc}, lemma 3 that if $sup_{t\in \R}\Vert C(t)-c(t)I_X\Vert <1$ then the generator of $C$ is bounded, assuming that the scalar cosine function $c$ is continuous. The following corollary, which is an immediate consequence of collorary 3.3 and theorem 3.6, shows that the continuity condition on $c(t)$ is redundant, and that the generator of $C(t)$ is bounded whenever $sup_{t\in \R}\Vert C(t)-c(t)I_X\Vert <2.$
\begin{corollary} Let $X$ be a Banach space, let $(c(t))_{t\in R}$ be a bounded scalar cosine function, and let $C=(C(t))_{t\in \R}$ be a bounded strongly continuous cosine family of bounded operators
on $X$ such that $sup_{t\in \R}\Vert C(t)-c(t)I_X\Vert <2.$ Then there exists $a\in \R$ such that $c(t)=cos(at)$ for $t\in \R,$ and the conclusions of theorem 3.6 hold. In particular the generator of the cosine function $C$ is bounded.
\end{corollary}
|
\section{Introduction}
The detection of B-mode polarization on large angular scales by the Background Imaging of Cosmic Extragalactic Polarization II \citep[BICEP2]{BICEP2} collaboration, and its possible primordial origin as the result of relic gravitational waves has invigorated the cosmological community. There has been much debate about the interpretation of the BICEP2 data, and this is an important issue that will be resolved with independent measurements and multi-frequency observations. The recent BICEP2/KECK and \textit{Planck} 353GHz polarized map joint analysis has shown that at least a part of the BICEP2 and KECK signal is due to polarized dust \cite{BICEPKECK2015}. However, more importantly the observations have yielded renewed interest in constraining models of the early universe through their predictions of the spectrum of primordial gravitational waves and the testability of the inflationary scenario.
Fortunately, at least for a simple class inflationary models in which a single scalar field drove a period of inflation more or less directly preceding the radiation dominated era of hot big bang expansion of our universe, positive predictions can be used to test the inflationary paradigm. The primordial power spectrum of tensor modes is typically parameterized as a power law \citep{Liddle1993}:
\begin{equation}
P_t(k)=A_t(k_*)\left(\frac{k}{k_*}\right)^{n_t} \, , \label{eq:tensor_pk}
\end{equation}
where $A_t$ is the tensor amplitude, $k$ is the wavenumber, $k_*$ is some reference wavenumber, and $n_t$ is the tensor spectral index, or tilt. While there are, in general, deviations from this simple expression, it serves to parameterize a host of scenarios. For the simplest single field models of inflation there is a relation between the tensor spectral tilt and the tensor-to-scalar ratio $r$, sometimes referred to as the inflationary consistency condition \citep{liddle_lyth}.
At leading order in slow-roll, this relation is given by \cite{1992PhLB..291..391L,1993PhRvL..71..219C}
\begin{equation}
r\equiv \frac{A_t}{A_s}=-8n_t \, ,
\end{equation}
where $A_s$ is the amplitude of the scalar power spectrum. In more complicated models with, for example, a non-trivial sound speed or multiple dynamical fields, this relation is modified to be $r\leq-8n_t$ (see e.g. \cite{2014arXiv1409.2498P,2014arXiv1407.2621B,palma/soto:2014} for recent discussions of some of these issues). In all models of inflation where the tensor perturbations originate as vacuum fluctuations, the tensor amplitude $A_t$ is set only by the expansion rate which is nearly constant during inflation and always decreases when the null energy condition holds: inflation always predicts a small negative tensor tilt. The specific value of the tilt $n_t$ may depend on the details of the model, but a negative tensor tilt is a generic prediction of inflation when the null energy condition holds. There are however, other early universe models which make different predictions for the tensor spectrum \cite{2008PhR...465..223L,2004PhRvD..69l7302B,2007PhRvL..98w1302B}. Hence we take a phenomenological point of view and do not restrict our analysis to negative values of $n_t$. If the data provide evidence for a positive tensor tilt, we should further explore alternatives to inflation. A detection of a positive tensor tilt would suggest at the very least that the primordial gravitational waves were not produced as vacuum fluctuations during inflation\footnote{That being said, counter examples do exist even in the context of inflation \cite{2014JCAP...08..016S,2014arXiv1412.0665M,2012PhRvD..85b3525B}, but fortunately these models tend to make additional predictions for other observables such as non-Gaussianity which would allow additional checks of these scenarios.}.
The aim of this paper is to perform the first {\it joint analysis} of cosmic microwave background (CMB), Laser Interferometer Gravitational wave Observatory \citep[LIGO]{abbott/etal:2007}, and Parkes Pulsar Timing Array \citep[PPTA]{manchester/etal:2013} data to put a constraint on both the tilt $n_t$ and {\it and} the running $n_{t,\rm run}$ of the primordial tensor power spectrum. Previous studies have shown that such combinations of data would improve the available lever arm of the data \citep{smith/peiris/cooray:2006}. BICEP2 data alone yields only a weak constraint on the tensor tilt $n_t$ as the polarization signal of gravitational waves drops rapidly as a function of multipole; hence the CMB alone provides a limited lever arm to constrain the tensor power spectrum. We will show that the constraint on $n_t$ is dominated by the \textit{Planck} temperature power spectrum data, which prefers lower power on large angular scales\footnote{To put any reasonable constraints one has to take a pivot scale that is within the observable window. In this paper we chose the slightly unconventional $k_* = 0.01$ Mpc$^{-1}$. See e.g. Ref.~\cite{2014arXiv1409.6530C} for a discussion of pivot scale in the context of constraining gravitational waves with B-mode polarization.}, resulting in a preference for positive tensor tilt. The additional data extend our lever arm significantly by increasing the scales over which we can measure the gravitational wave energy density by 18-20 orders of magnitude. Although the constraints at the small scales probed by LIGO and PPTA are only upper limits on the gravitational wave energy density, earlier estimates have shown \citep{Caligiuri2014,Boyle2007} that these experiments are particularly sensitive to the tilt of the primordial tensor mode power spectrum, assuming a detection of a tensor-to-scalar ratio $r \gtrsim 0.01$. The results presented in this paper are relevant for constraining the details of the gravitational wave power spectrum, independent of the current interpretation of the BICEP2 data. We will perform each analysis with and without BICEP2 data, and show that one can get meaningful results even without constraints from BICEP2 on $r$. In addition, we will include the recent released cross analysis data of BICEP2/KECK and the 353 GHz \textit{Planck} polarization data \cite{BICEPKECK2015}.
The paper is organized as follows. In Sec.~\ref{sec:review} we briefly review the relevant physics. In Sec.~\ref{sec:results} we present our results on the bounds on $n_t$ and $n_{t,\rm run}$ from a variety of data combinations, including in Sec.~\ref{sec:discussion} a discussion of the observational constraints derived from the impact of the primordial gravitational wave energy density on the background expansion of the universe not included in the initial analysis.
We summarize some future prospects and conclude in Sec.~\ref{sec:conclusion}. A summary of the derived parameter constraints using various combination of data sets can be found in Tab.~\ref{table:n_t values}. All bounds quoted in the paper are $95\%$ bounds. A discussion of the definition of gravitational wave energy density is included in Appendix \ref{appendixA}.
\section{Gravitational wave spectrum and energy density}\label{sec:review}
While measurements of the B-mode polarization of the CMB constrain the tensor mode power spectrum directly, observations from gravitational wave observatories and pulsar timing arrays instead place limits on the energy density of gravitational waves at specific wavenumbers \citep{mtw1973,krauss/white:1992}
\begin{eqnarray}
\Omega_\mathrm{GW}(k)&=&\frac{P_t(k)}{12H_0^2}\left[\dot{\mathcal{T}}(\eta_0,k)\right]^2 \, , \label{eq:omgw}
\end{eqnarray}
where $\eta_0$ is the conformal time in the current epoch, $\mathcal{T}(\eta,k)$ is the tensor transfer function, and an overdot refers to a derivative with respect to cosmic time $t$. The transfer function projects the primordial signal from early to late times and is obtained by solving the Klein-Gordon equation for the tensor fluctuations in an expanding universe\footnote{In general a non-standard thermal history will affect the form of the transfer function. See for example \cite{Boyle2007,2014arXiv1407.4785K}.}. See Appendix~\ref{appendixA} for a derivation of this formula.
In the case of single field slow-roll inflation, the power spectrum of the tensor modes, $P_t(k),$ is well described by Eq.~\eqref{eq:tensor_pk} with $n_t \simeq 0$.
The leading deviation from the simple power law Eq.~\eqref{eq:tensor_pk} expected in some models is parameterized by allowing the tilt to vary with scale:
\begin{equation}
P_t(k)=A_t(k_*)\left(\frac{k}{k_*}\right)^{n_t(k_*)+\frac{1}{2}n_{t,\mathrm{run}}\ln(k/k_*)},
\end{equation}
where $n_{t,\mathrm{run}} = dn_t/d\ln k$ is the running of the tensor spectral index and again $A_t$ is the amplitude of the tensor power spectrum.
Single field slow-roll inflation makes a model-dependent prediction for the tensor-to-scalar ratio, i.e. $r\equiv A_t/A_s=16\epsilon$ \citep{Liddle1993}. The (single field slow-roll) inflationary consistency condition also tells us that $n_t = -r/8$, and similar relations can be derived for the running and higher order terms which are generally slow-roll suppressed \citep{Liddle1994}. As a result, the scale dependence of tensor fluctuations is described by a limited number of parameters, and should be uniquely determined by a limited number of measurements. Despite being slow-roll suppressed, higher order terms can become quite significant in determining the spectrum far from the pivot scale $k_*$ \citep{boyle/etal:2014}. We are not yet in the data-driven regime, however, and with only upper limits from PPTA and LIGO on the gravitational wave energy density, the data is not yet capable of falsifying the consistency conditions with high significance. Hence, we include the running but leave a treatment of higher order corrections to future work.
In this paper we assume that the power law form of the tensor power spectrum holds all the way up to some fixed ultraviolet cutoff, which we will discuss in more detail below. Physically, this means that we are assuming an instantaneous transition from the phase of the early universe responsible for producing the primordial spectrum to the phase of radiation domination. This instant reheating assumption is likely to be violated in a realistic model of the early universe where one would expect modifications to the power spectrum at wavenumbers near the cutoff (see \cite{2014arXiv1407.4785K} for an excellent review). Other effects, such as the late production of entropy can also modify the spectrum and weaken the constraints derived below. Hence our constraints apply to a phenomenological model with the simplest assumptions: a simple power law primordial spectrum (including the possibility of running), instantaneous reheating, and a standard thermal history. Any specific model which violates one or more of these features is likely to be subject to different observational constraints.
We use a pivot scale of $k_*=0.01$ Mpc$^{-1}$, particularly because the BICEP2 experiment is insensitive to smaller scales \cite{2014arXiv1409.6530C}. We use an analytic solution for the gravitational wave transfer function valid at late times, which has an accuracy of about 1\% in a flat universe and is relevant for very small scales $k\gg k_{eq}$ and late times $\eta \gg \eta_{eq}$ \citep{WMAP}.
\begin{figure}[t!]
\centering
\includegraphics[width=3.5in]{nt_rbb.pdf}
\caption{The marginalized 2D posterior constraints on $n_t$ and $r$. The large grey contour shows the constraints of using BICEP2 in conjunction with CMB data from \textit{Planck} \citep{Planck}, the HST prior on the Hubble constant \citep{HST}, and the constraints on the acoustic scale from BAO \citep{BAO} measurements (the latter three combined in the name `All'). Adding in data from the PPTA constraints on the gravitational wave energy density improves the constraints three-fold on $n_t$, while the constraint on $r$ is largely unchanged. Using data from LIGO instead tightens this even further, highlighting the constraining power of LIGO on the primordial tensor power spectrum. All contours agree on the value of $r$ announced by the BICEP2 team, confirming that the LIGO and PPTA data are not particularly sensitive to the value of $r$ directly, but they are sensitive to the tilt $n_t$, as shown in Eq.~\eqref{eq:omgw}.}
\label{fig:nt_r}
\end{figure}
We modified the publicly available Markov Chain Monte Carlo sampler {\tt cosmomc} \citep{cosmomc} and the Boltzmann code {\tt CAMB} \citep{Lewis:1999bs}, by adding a module\footnote{Code will be made publicly available through \url{http://www.astro.princeton.edu/~meerburg/coding/}. } that computes the gravitational wave energy density as a function of wavenumber $k$, which can be converted to frequency using the relationship
\begin{equation}
f = \frac{k}{2\pi a(\eta_0)} \, .
\end{equation}
The observations of pulsar timing arrays correspond to frequencies of $f \in$($10^{-9}$~Hz, $10^{-7}$~Hz), while gravitational wave observatories probe $f\in$(0.1~mHz, 5000~Hz) \citep{Moore2014}. The constraint from LIGO\footnote{We have approximated the constraints from LIGO by taking a constant upper limit over the frequency range considered. For the purpose of this analysis, this approximation is sufficient.} is $\Omega_\mathrm{GW}<\rm{5.6 \times 10^{-6}}\,(68/H_0)^2$ \citep{LIGO} at $2\sigma$, while the upper limit from PPTA is $\Omega_\mathrm{GW}<\rm{1.5 \times 10^{-9}}\,(68/H_0)^2$ \citep{PPTA} at $2\sigma$ where $H_0=100\,\,h\, \mathrm{km}\mathrm{s}^{-1}\mathrm{Mpc}^{-1}$.
\begin{figure}[t!]
\centering
\includegraphics[width=3.35in]{nt_rbb_BKP.pdf}
\caption{The marginalized contours in the $r-n_t$ plane comparing BICEP2 to the BKP data. If all the signal in BICEP2 is primordial, we obtain strong bounds on $n_t$, with a $2\sigma$ preference for $n_t>0$. The upper limit is driven by the LIGO constraint. From the BICEP2/KECK cross \textit{Planck} analysis $r$ is lowered and therefore in combination with the constraints from \textit{Planck} TT power spectrum leads to a shift in the peak value of $n_t$. With a small value of $r$, constraints on $n_t$ are weakened. However, including LIGO puts a strong upper bound on $n_t$ as before. }
\label{fig:BKP_compare}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=3.5in]{nt_rbb_nobicep.pdf}
\caption{The marginalized contours in the $r-n_t$ plane without using BICEP2 data. When $r_{0.01} = 0$ there is no constraint on $n_t$ in principle, however \textit{Planck} TT can be best fit with $r_{0.01} = 0.24$, hence there exists some constraint on $n_t$. \textit{Planck} prefers a very blue tensor spectrum, in order to create an artificial running of the temperature data on large scales. From this analysis it is also apparent that \textit{Planck} drives the value of $n_t$, not BICEP2 data. Interestingly, by adding LIGO, such large values of $n_t$ are ruled out, which has the effect that $r$ is significantly constrained $r_{0.01} < 0.2$ at $95\%$. }
\label{fig:noBICEP}
\end{figure}
For each value of the gravitational wave energy density, a likelihood is computed based on the upper limits from PPTA and LIGO ($\chi^2=4.0$, corresponding to 2$\sigma$ at 95\% confidence level) from a truncated, logarithmic distribution. We use an analytic solution to obtain the conformal time $\eta$ as a function of scale factor \cite{Sazhin2011}.
\begin{table*}[htbp!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Dataset &\multicolumn{5}{c|}{Parameter} \\ \hline
& \multicolumn{2}{c|}{$n_t$} &\multicolumn{2}{c|}{$r_{0.01}$}& \multicolumn{1}{c|}{$n_{t,\mathrm{run}}$} \\ \hline
& Best fit & Mean, 95\% limits& Best fit & 95\% limits & Upper, 95\% limits\\
\hline
All & 2.66 &$2.08^{+1.79}_{-1.95}$&0.24 &$<1.0$ & $-$\\
All+LIGO & 0.51 &$0.13^{+0.54}_{-0.75}$ &0.06 & $< 0.2$ &$-$\\
All+PPTA+LIGO (running) & 2.99 &$2.21^{+2.02}_{-1.85}$&1.06 &$ <1.7$ &$< -0.32$\\
All \,$~(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{20})$& 0.52&$0.12^{+0.61}_{-0.97}$& 0.01&$ < 0.21$ & $-$\\
All\,$~(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{24})$&0.44 &$0.05^{+0.44}_{-0.9}$& 0.01 &$ <0.17$ & $-$\\
\hline
All+BKxP+PPTA &0.65 &$0.32^{+0.71}_{-0.92}$ & 0.04 &$<0.1$ & $-$\\
All+BKxP+LIGO &0.26 &$0.06^{+0.63}_{-0.89}$&0.02 &$<0.1$ & $-$\\
All+BKxP+PPTA+LIGO (running) &3.96 &$1.67^{+2.94}_{-2.77}$&0.02 &$<0.11$ &$ < -0.22$\\
All+BKxP\,$~(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{20})$& 0.49&$0.15^{+0.55}_{-0.80}$&0.05 & $<0.09$ & $-$\\
All+BKxP\,$(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{24})$&0.39 &$0.07^{+0.52}_{-0.80}$&0.04 & $<0.09$ & $-$\\
\hline
All+BICEP2 & 1.42& $1.30^{+1.36}_{-1.03}$& 0.18 &$0.18^{+0.1}_{-0.09}$ & $ -$ \\
All+BICEP2+PPTA &0.66 &$0.55^{+0.29}_{-0.49}$&0.22 &$0.19^{+0.09}_{-0.08}$ &$-$\\
All+BICEP2+LIGO&0.56 &$0.35^{+0.20}_{-0.39}$ & 0.18 &$0.18^{+0.07}_{-0.06}$ &$-$\\
All+BICEP2+PPTA+LIGO (running) &0.56 &$1.15^{+1.36}_{-1.21}$&0.19 &$0.19^{+0.1}_{-0.1}$ &$ < -0.05$\\
All+BICEP2\,$~(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{20})$& 0.47&$0.4^{+0.09}_{-0.35}$&0.187 & $0.17^{+0.07}_{-0.06}$ & $-$\\
All+BICEP2\,$(N_\mathrm{eff}, k_\mathrm{UV}/k_*=10^{24})$&0.41 &$0.33^{+0.12}_{-0.25}$&0.16 & $0.17^{+0.07}_{-0.06}$ & $-$\\
\hline
\hline
\end{tabular}
\caption{Summary of parameter constraints using various combinations of data sets. ``All'' refers to \textit{Planck} (2013), WMAP low $\ell$ polarization, HST and BAO data. In all studied cases a positive $n_t$ is preferred over a negative $n_t$, but the significance is small when you remove any BICEP2/KECK data. Recall that we use a pivot scale $k_* = 0.01$ Mpc$^{-1}$. }
\label{table:n_t values}
\end{center}
\end{table*}
\section{Results}\label{sec:results}
\subsection{Combining CMB measurements with low-redshift constraints}
In all runs we include temperature data from the \textit{Planck} satellite \citep{Planck} with the polarization prior on the optical depth from WMAP \citep{WMAP}, WMAP low $\ell$ polarization, a prior on the Hubble parameter from the HST key project \citep{HST} and BAO measurements from the Sloan Digital Sky Survey \citep[SDSS]{BAO}. We then include a variety of other probes, including the recent measurements of large-scale polarization from the BICEP2 team \cite{BICEP2}, and constraints on $\Omega_\mathrm{GW}$ from the PPTA or from LIGO data. In order to be insensitive to lensing effects in polarization data, we restrict the BICEP2 data to the four lowest multipole bins. Last, we include an analysis combining the previous data sets with the BICEP2/KECK and \textit{Planck} cross correlation likelihood \cite{BICEPKECK2015}. Here we fit for the lensing amplitude and the dust contribution.
Our results are presented in Fig.~\ref{fig:nt_r} and Table~\ref{table:n_t values}. As expected, we get a weak constraint on the 95\% limit value of $n_t$ of $\rm{1.64}<n_t<\rm{2.63}$ when using only BICEP2 data. The constraint on $n_t$ is improved significantly by adding LIGO, lowering the 95\% upper limit by roughly a factor of 5 from 2.63 to 0.59. The mean value we get for the tensor tilt at CL 95\% is $n_t=0.394^{+0.209}_{-0.288}$. The upper limit is a very hard cut, because of the exponential drop off. Our results reflect the knowledge that it is very difficult to measure $n_t$ from CMB data alone (see e.g. Refs.~\cite{Dodelson,boyle/etal:2014}) and small scale GW experiments are necessary to test the inflationary consistency condition. We are also aware that a detection of gravitational waves by LIGO will not be particularly useful for determining the primordial tensor spectrum as the signal on scales to which LIGO is sensitive is dominated by astrophysical sources of GW, such as binary mergers.
Since current constraints are mainly upper limits on the gravitational wave energy density and any signal which is detected at LIGO is expected to be of non-primordial nature, what we present here are conservative constraints on $n_t$.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=5.9in]{TensorTTspectra}
\caption{The contribution from the tensors to the $TT$ power spectrum. A very large positive tilt leads to a rescaling of the spectrum, except on very large scales, where the spectrum is suppressed. This suppression is what \textit{Planck} prefers (though not significantly) over no tensor modes at all. }
\label{fig:TensorTTspectra}
\end{figure*}
When we replace the BICEP2 data with the cross data \cite{BICEPKECK2015}, there is a very strong upper bound on the tensor-to-scalar ratio from BICEP2/KECK data, as shown in Fig~\ref{fig:BKP_compare}. We find $n_t = 0.06^{+0.63}_{-0.89}$ and $r_{0.01}<0.1$.
Even after cross checking \textit{Planck} with BICEP2/KECK data, there is still significant uncertainty in the total fraction due to dust. For completeness, we also considered constraints using only \textit{Planck} temperature data and WMAP polarization, as shown in Fig.~\ref{fig:noBICEP}. In principle, no bound can be put on $n_t$ when $r=0$, but $r=0$ is as likely as any other small value\footnote{We do not encounter serious issues with convergence. An alternative procedure is to constrain $r$ at two different scales and to treat the tilt as a derived parameter. Note that additional scales would be required for constraints on the running and higher order scale dependence.}. It is interesting to note that \textit{Planck} drives the value of $n_t$, not the BICEP2 data. In fact, allowing $n_t$ to vary freely, leads to the best-fit value $r_{0.01} = 0.24$. We show the contributions from the best-fit spectra to the $TT$ power spectrum at low multipoles in Fig.~\ref{fig:TensorTTspectra}. For very positive tensor tilt, the $TT$ data can be made more consistent with the theory by providing a possibility to lower the power on large scales. This is also the reason why \textit{Planck} alone prefers a large value of $r$ (to increase the overall contribution from the tensors to the $TT$ spectrum) and with a large positive tensor tilt (to mimic suppression of the power on large scales). When LIGO data is added to the likelihood, such preference completely disappears, since $n_t$ can not be too large for most values of $r$. As a result, a very large value of $r$ is no longer allowed and we actually get a constraint on $r$ that is comparable to the constraint from $r$ without adding the tilt and the constraints after adding the BICEP2/KECK and \textit{Planck} cross data. As such, in the scenario that even a larger fraction in the BICEP2 patch is due to dust, simply adding LIGO puts a very strong bound on $r$. That being said, obviously if all measured signal in BICEP2/KECK is due to dust, this puts a very strong bound on $r$. It is interesting to see that even without BICEP2/KECK data one could have put a constraint on $r$ when allowing $n_t$ to vary; adding LIGO tells us that $r < 0.2$, which is consistent with the BICEP2/KECK cross \textit{Planck} data.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=5in]{ntrun_nt_rbb_4panel_BKPLANCK.pdf}
\caption
Constraints allowing for more freedom in the GW power spectrum. \textit{Left panel -} the marginalized 2D constraints in the $n_t-r$ for the cases including BICEP2 (\textit{bottom}) and BICEP2/KECK cross \textit{Planck} (\textit{top}) data. \textit{Right panel - } constraints on the $n_t - n_{t,\mathrm{run}}$ for the same cases. Allowing for running in the tensor spectral index does not change the constraint on the tilt significantly, as the errors on the running are large and there is no significant correlation between the two parameters. BICEP2 data improves constraints on $n_t$ as expected. The addition of $n_{t,\mathrm{run}}$ generally weakens the constraint on $r$. Note the different ticks in the figures on the right. }
\label{fig:DDMLIGO}
\end{figure*}
Next, we consider the running of the tensor tilt. Because the PPTA and LIGO measurements are at different scales we can use both PPTA and LIGO in our analysis. The results are presented in the right-hand panel of Figure~\ref{fig:DDMLIGO}. Without small scale constraints on the gravitational wave energy density data are incapable of putting any bound on the $n_{t,{\rm run}}$. As is shown in the right hand panel of Figure~\ref{fig:DDMLIGO}, the running is exclusively confined to negative values and this constraint comes solely from PPTA and LIGO. Naturally, as $n_t$ increases, $n_{t,{\rm run}}$ decreases in order to compensate for the large values of the tilt. This causes the constraint on $n_t$ to worsen as one moves to lower values of the running. For values above $n_t \sim 3$ (with BICEP2) or $n_t \sim 6 $ (with BICEP2/KECK cross \textit{Planck}), the CMB measurements control the constraint, and the joint data does not allow for larger values of the tilt. Since LIGO and PPTA do not provide lower limits, there is no useful constraints on the lower limit of the running. The constraints on the tensor-to-scalar ratio $r$ are weakened, since running adds another free parameter that allows the fit to improve on large scales in the TT power spectrum.
\subsection{Gravitational wave contribution to the massless degrees of freedom}\label{sec:discussion}
Our analysis has shown that the inclusion of small scale gravitational wave constraints results in a bound on $n_t \lesssim 0.5$ (at $1\sigma$). Since gravitational waves also contribute to the total radiation energy density of the universe, a consistent analysis must take into account their effect on the background expansion, as was considered in \cite{Boyle2007} and \cite{NeffSmith2006}. The increased radiation energy density alters the peak structure of the CMB as explained in e.g. \cite{2004PhRvD..69h3002B} and \cite{2013PhRvD..87h3008H} and also affects the prediction for the primordial abundance of light elements, see for example \cite{Weinberg:2008zzc}. Rossi et al. \citep{rossi/etal:2014} recently presented constraints from Lyman-$\alpha$ forest measurements in conjunction with a variety of probes.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=6.8in]{Chang_aeq_DeltaNeff}
\caption{Left: The relative change in the scale of matter-radiation equality as a function of the tensor tilt $n_t$, assuming $r=0.2$. The simple analytical estimate suggests that $n_t\sim 0.4$ would be modifying the scale of equality at the $5\%$ level, which is close to current bounds from WMAP and \textit{Planck}. Obviously, for smaller $r$, this constraint weakens accordingly. Right: The change in $N_{\rm eff}$ as a function of $n_t$ for $r=0.2$ using the two approximations given in Eqs.~(\ref{eq:Neff}) and (\ref{eq:Neff2}). The difference is very small up until the energy density in gravitational waves becomes of the order of the total energy density in the Universe. Our constraints are derived using Eq.~(\ref{eq:Neff}). A detailed explanation of both curves and how these are derived can be found in the text. }
\label{fig:aeq}
\end{figure*}
The energy density of gravitational waves is given by (see Appendix~\ref{appendixA} for details)
\begin{eqnarray}
\rho_{\rm GW} &=& \int_{k_{\rm IR}}^{k_{\rm UV}} d \log k \frac{P_t(k)}{32\pi Ga^2} \left[\mathcal{T}'(k,\eta)\right]^2 \nonumber \\
&=& \frac{A_s r }{32 \pi G a^2} \int_{k_{\rm IR}}^{k_{\rm UV}} k dk \left(\frac{k}{k_*}\right)^{n_T} j^2_1(k\eta).
\end{eqnarray}
Here we used that the conformal time derivative of the transfer function during radiation domination is given by $\mathcal{T}'(\eta < \eta_{\rm eq}, k > k_{\rm eq}) = - k j_1(k \eta)$ \cite{Komatsu2006}. We cannot simply integrate over all wavenumbers, since this integral diverges in the IR for $n_t\leq -4$ and in the UV for $n_t\geq -2$. Physically, one expects \cite{2005PhRvD..72h3501H} that superhorizon modes should not contribute to the local energy density (note that even this itself is an ambiguous statement in the sense that the total energy density could only be measured after averaging over several wavelengths \cite{mtw1973}) and this behavior is captured by the fact that small wavenumbers make a negligible contribution to the integral. Furthermore, we do not expect that the physics of the early universe generated primordial gravitational waves at arbitrarily small scales. The largest conceivable wavenumber at which primordial gravitational waves could have been produced in the early universe is given by the Planck scale at the beginning of the hot big bang expansion of the universe, since modes with larger wavenumbers would have been super-Planckian at some point in the history of the universe. If we assume that the mechanism which is responsible for solving the flatness and horizon problems is also responsible for generating the primordial spectrum of gravitational waves, then we would expect that the spectrum should extend over a range of scales equivalent to the amount of expansion required to solve those problems. Below we give results for the case where the spectrum extends only to LIGO scales and also for the case where the UV scale is 60 $e$-folds smaller than the present horizon size. It is also worth mentioning that although the scaling of the energy for each mode $k$ is equivalent to that of radiation, if there is some comoving IR bound, this scaling is (weakly) broken by the large scale modes which enter the horizon and begin to oscillate. For our purposes, we are only interested in positive $n_t$, for which the UV modes dominate the total energy density. It is natural then to introduce a new variable $q = k/k_{\rm UV}$, $dk = k_{\rm UV} dq$ and $k_{\rm IR}/k_{\rm UV} = \varepsilon$. In the cases we consider, we can safely take the limit $\varepsilon\rightarrow 0$, but our results will still depend on the choice of $k_\mathrm{UV}$, which we leave unspecified for the moment.
We can then perform the integral for positive $n_t$
\begin{eqnarray}
&k_{\rm UV}^2 \left(\frac{k_{\rm UV }}{k_*}\right)^{n_t} \int_{0}^{1} q^{n_t +1} dq j^2_1(q k_{\rm UV} \eta) \nonumber \\
&\qquad=\left(\frac{k_{\rm UV }}{k_*}\right)^{n_t} \frac{1}{2 n_t} \frac{1}{\eta^2} + \mathcal{O}(1/(k_{\rm UV}\eta)) \, .
\end{eqnarray}
Hence for the gravitational wave energy density we find (up to corrections of order $\varepsilon$)
\begin{equation}\label{rhoGW}
\rho_{\rm GW} (\eta) \simeq \frac{A_s r}{32 \pi G} \left(\frac{k_{\rm UV }}{k_*}\right)^{n_t} \frac{1}{2 n_t} \frac{1}{(a\eta)^2}.
\end{equation}
It is useful to relate this energy density to quantities which may be more familiar (for a similar analysis, see \cite{Maggiore:1999vm}). Deep in the radiation-dominated era, after big bang nucleosynthesis (BBN) but well before matter-radiation equality, total energy density of the universe is given by
\begin{align}
\rho_{\mathrm{tot}}&=\rho_\gamma+\rho_\nu+\rho_{\mathrm{GW}} \nonumber \\
& \equiv \rho_\gamma\left(1+\frac{7}{8}\left(\frac{4}{11}\right)^{4/3}N_\mathrm{eff}\right) \, ,
\end{align}
where $N_{\mathrm{eff}}=3.046$ in the standard model in the absence of gravitational waves \cite{Mangano:2005cc}. Using the fact that during this period, $1/(a\eta)^2=H^2=8\pi G \rho_\mathrm{tot}/3$, and also assuming that the gravitational waves are not the dominant source of energy density, we can write the fraction of total energy density in gravitational waves as
\begin{align}
\frac{\rho_{\mathrm{GW}}}{\rho_\mathrm{tot}}&=\frac{A_s r}{24n_t}\left(\frac{k_\mathrm{UV}}{k_*}\right)^{n_t} \nonumber \\
&=\frac{\frac{7}{8}\left(\frac{4}{11}\right)^{4/3}\left(N_\mathrm{eff}-3.046\right)\rho_\gamma}{ \left(1+\frac{7}{8}\left(\frac{4}{11}\right)^{4/3}N_\mathrm{eff}\right)\rho_\gamma} \,
\end{align}
which can be solved for $N_{\mathrm{eff}}$
\begin{equation}\label{eq:Neff}
N_\mathrm{eff}=\frac{\frac{8}{7}\left(\frac{11}{4}\right)^{4/3}\left[\frac{A_s r}{24n_t}\left(\frac{k_\mathrm{UV}}{k_*}\right)^{n_t}\right]+3.046 }{1-\left[\frac{A_s r}{24n_t}\left(\frac{k_\mathrm{UV}}{k_*}\right)^{n_t}\right]} \, .
\end{equation}
This expression breaks down when the gravitational wave energy density becomes comparable to the total energy density of the universe. A fully consistent analysis for larger values of the gravitational wave energy density would require deriving a new transfer function which does not assume radiation domination, but instead incorporates the backreaction of the gravitational waves on the expansion history. Luckily, the data is sufficient to constrain the gravitational wave energy density to a level significantly below the regime where this presents a problem. An alternative approximation for $N_\mathrm{eff}$ which does not have a singularity can be found by neglecting the contribution of $\rho_\mathrm{GW}$ to $H$ in Eq.~(\ref{rhoGW}) which gives the same result as the first order Taylor expansion of Eq.~(\ref{eq:Neff}) about $\rho_{GW}/\rho_{\mathrm{tot}}=0$,
\begin{equation}\label{eq:Neff2}
N_\mathrm{eff}\approx3.046+\left(3.046+\frac{8}{7}\left(\frac{11}{4}\right)^{4/3}\right)\frac{A_s r}{24n_t}\left(\frac{k_\mathrm{UV}}{k_*}\right)^{n_t} \, .
\end{equation}
Both Eq.~(\ref{eq:Neff}) and Eq.~(\ref{eq:Neff2}) should only be trusted in the regime $\rho_\mathrm{GW}/\rho_\mathrm{tot}\ll 1$. We use Eq.~(\ref{eq:Neff}) in our analysis since it gives more conservative constraints on $n_t$, though the difference is small with current data (see right panel of Fig.~\ref{fig:aeq} for a comparison of these approximations in the parameter range of interest).
Since the energy density of gravitational waves redshifts as radiation (so long as we can neglect the contribution of the long wavelength modes) the relative energy density of gravitational waves compared to that of photons, and thus the gravitational wave contribution to the value of $N_\mathrm{eff}$, remains constant for all times of interest.
In order to estimate the size of the effect of the additional radiation energy density, we will calculate the change in the redshift of matter-radiation equality. At the instant of matter-radiation equality, we have by definition
\begin{eqnarray}
\rho_m(a_\mathrm{eq}) = \rho_r(a_\mathrm{eq}) \, .
\end{eqnarray}
This can be written as an equation for $a_\mathrm{eq}$ as follows
\begin{align}
\frac{3H_0^2\Omega_m}{8\pi G}&\left(\frac{1}{a_\mathrm{eq}}\right)^3 \nonumber \\
&= a_\mathcal{B}T_{\gamma,0}^4\left(1+\frac{7}{8}\left(\frac{4}{11}\right)^{4/3}N_\mathrm{eff}\right)\left(\frac{1}{a_\mathrm{eq}}\right)^4 \, ,
\end{align}
where $a_\mathcal{B}$ is the radiation constant, $T_{\gamma,0}$ is the CMB temperature today, $N_\mathrm{eff}$ is given by Eq.~(\ref{eq:Neff}), and we have fixed $a_0=1$.
In order to place constraints on $r$ and $n_t$ we have to make an assumption about the ratio $k_{\rm UV}/k_*$. In our analysis of the data we chose a pivot scale $k_* = 0.01$ Mpc$^{-1}$. For LIGO we do not have to make an assumption on the UV cutoff, however we do assume the scaling is power law all the way up to $k_{\rm LIGO}$. The minimum for this is ratio would then be $k_{\rm UV}/k_* \simeq 10^{20}$, while if we assume that the power law spectrum extends over $\sim60$ $e$-folds\footnote{This value is chosen not due to some particular mechanism for generating the gravitational waves in the early universe, but instead to roughly correspond to the maximum amount of hot big bang expansion.} the ratio is given by $k_{\rm UV}/k_* \simeq 10^{24}$. We estimate the contribution of primordial gravitational waves to the total energy density at the time of equality between matter and the relativistic components of the universe. Shifts in the redshift of matter-radiation equality by roughly a percent will be constrained by the CMB power spectrum, since this the accuracy with which the redshift of equality is measured. In Fig.~\ref{fig:aeq} we plot the relative change of the $a_{\rm eq}$ as a function of $n_t$ for these two numbers. It can be seen from this estimate that even for the lower value of $k_\mathrm{UV}$, the observational constraint from the modified expansion history will be at a level similar the best constraint from the CMB+LIGO data.
For consistency we include this effect in our CMB analysis. We add the gravitational wave contribution to the total the massless degrees of freedom in $N_{\rm eff}$ as computed above. The way the sampler usually scans the likelihood is to sample the densities and the optical depth independently (slow parameters) of the primordial parameters (fast parameters); this scheme obviously does not work in case one of the densities depends explicitly on the primordial parameters. For that reason we changed the sampling. Unfortunately this makes the scanning of the likelihood slower, as one adds three additional parameters to the slow part of the code (namely the computation of the radiation transfer functions). In order to speed up the convergence, we therefore fix the foregrounds to their best fit values.
The approximations we have used are valid for a positive tensor tilt, and only for small values or negative values of the tilt would we have to worry about IR corrections. However, from the analytical analysis, the background will not be affected in that limit and hence the total energy density from gravitational waves will be negligible. The code does not compute a correction to $N_{\rm eff}$ when $n_t<0$. For very large values of $n_t$, Eq.~\ref{eq:Neff} breaks down predicting negative values of $N_{\rm eff}$, and so in that case, we replace $N_{\rm eff}$ with a very large number. The results are shown in Fig.~\ref{fig:nt_r_neff} for two assumptions about the UV cutoff; one in which the cutoff is set at the LIGO scale and one in which we assume the maximal cutoff $k_{\rm UV}/k_* \simeq 10^{24}$. The constraints are summarized in Tab.~\ref{table:n_t values}. Compared to our analytical estimate, we find the constraints to be slightly weaker. Although $z_{\rm eq}$ is constrained quite strongly, $H_0$ and the matter density tend to vary as $n_t$ increases in order to keep $z_{\rm eq}$ fixed, which therefore gives a weaker constraint than our naive prediction. This is shown in Fig.~\ref{fig:scatter_zeq}. The strongest bound, as expected, is derived when we include BICEP2 data (which drives $r$ to large values) and $k_{\rm UV}/k_* \simeq 10^{24}$. We find $n_t = 0.33^{+0.12}_{-0.25}$. This stronger bound is driven by a non-zero $r$. Without the addition of BICEP2 data constraints are weaker and are found to be remarkably close to LIGO bounds.
The change in the expansion history due to the gravitational wave background energy density also affects BBN. The presence of extra massless degrees of freedom during BBN increases the primordial Helium fraction by increasing the expansion rate and thus decreasing the time during which free neutrons can decay before becoming bound into light nuclei \cite{Weinberg:2008zzc}. In \cite{Brandenberger2007} and recently in \cite{Gerbino2014} it was claimed that BBN puts an upper limit (with unknown confidence) on $n_t \leq 0.15$. However, this constraint is derived under the assumption that $k_{\rm UV}/k_* \simeq 10^{61}$, which would suggest the power law spectrum of gravitational waves extends over more than 120 $e$-folds and that many modes would have certainly been super-Planckian during the hot big bang expansion of the universe. Without a consistent theory of quantum gravity, it is unclear how to treat the evolution of such high energy gravitons. If one repeats the analysis with more sensible numbers, we find $n_t \leq 0.4$ for $r = 0.2$, in line with \cite{Boyle2007} and \cite{2014arXiv1407.4785K}. This bound again is very close to the bound from LIGO and from the CMB, and for completeness it should be taken into account. However, we expect based on a very similar analysis in \cite{NeffSmith2006} that the modified expansion history leads to effects for which the CMB provides a stronger constraint, given the same $\rm UV$ cutoff. That being said, the BBN constraint depends on the value of both $Y_p$ and $\Omega_b$ and it would be possible to include this fully self consistently inside CAMB. In this analysis we have fixed the value of $Y_p = 0.24$, and we leave a full treatment including these effects to future work.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=6.5in]{Neff_bicep_combined.pdf}
\caption{The constraints on the gravitational wave spectrum when adding in the change in $N_{\rm eff}$ from gravitational waves. Allowing for the $N_{\rm eff}$ contribution cuts off the distribution on $n_t$ to more positive values. Adding in BICEP2 data again moves the constraints to more positive values of $r$, as expected. When considering the BKxP data in addition to \textit{Planck}, the upper limits on $n_t$ are hardly affected as can been see in the figure on the right. We also show the effect of changing the cut-off scale from $k_*/k_{\mathrm{UV}}=10^{20}$ to $k_*/k_{\mathrm{UV}}=10^{24}.$ The overall effect is to limit the upper bound on $n_t$. }
\label{fig:nt_r_neff}
\end{figure*}
When considering the constraints derived from the contribution of gravitational waves to the massless degrees of freedom, there are two important things to remember. Firstly, the total energy density in gravitational waves depends on the entire primordial spectrum, which we have assumed here to be a power law all the way up to some UV cutoff. A more gradual transition to the reheating phase, or production of additional gravitational waves during reheating could alter the constraints derived here.
In addition, since the constraints on modified expansion history are from `indirect' measures, the effect of gravitational waves would be degenerate with modifications to the neutrino sector or the addition of dark radiation which would have the same phenomenological effect on the measured CMB \cite{NeffSmith2006}. As such, this constraint is less robust than actually putting upper bounds on direct gravitational wave detection. In our analysis we have assumed that beyond photons and three families of standard model neutrinos, all of the radiation energy density is made up of gravitational waves, with no additional massless degrees of freedom making a significant contribution to the energy density of the universe.
\begin{figure*}[htbp!]
\centering
\includegraphics[width=6.5in]{scatter_BKP.pdf}
\caption{A scatter plot showing the relation between the tensor tilt $n_t$, the Hubble rate $H_0$ and the redshift of equality $z_{\rm eq},$ including the contribution from gravitational waves to $N_{\rm eff}$. The \textit{top} panel shows the case where the data are combined with BICEP2, while the \textit{middle} and \textit{lower} panels replace the BICEP2 data with the BICEP2/KECK-\textit{Planck} cross spectrum for two different values of the cutoff. In all cases, $z_{\rm eq}$ is fairly well constrained, hence in order to accommodate large $n_t$ (and hence a large $N_{\rm eff}$, $H_0$ is increased by adding more matter to the universe. Because of the steep increase of $N_{\rm eff}$ as a function of $n_t$ there is a very strong turn in degeneracy between $n_t$ and $H_0$. There is a shift in the upper bound of $n_t$ allowed by BICEP2 compared to the cross spectrum, and this limit also shifts with different values of the cut off, as can be seen by comparing the middle to lower panels of the figure.
\label{fig:scatter_zeq}}
\end{figure*}
\section{Conclusion}\label{sec:conclusion}
In this paper we performed the first joint analysis of CMB data and late time measures of gravitational waves. This study is interesting for the following reasons. First, a combined analysis allowed us to put a bound on both the tilt {\it and} the running of the primordial tensor power spectrum. Secondly, even without BICEP2 data or the BICEP2/KECK and \textit{Planck} cross data, we can put a bound on the tensor-to-scalar ratio $r$ that is close to the constraint without allowing for the tensor tilt to vary. This is especially important given the recent conclusion that dust could be responsible for all of the detected B-mode power. While additional sources of gravitational waves could explain the detection of $r$ (even without considering dust), this gives interesting bounds on the energy scale of inflation \citep{ozoy/etal:2014}, of the detailed perturbation physics in this early epoch \citep{palma/soto:2014}, and of other early universe physics \citep{2014arXiv1412.0407L}. Thirdly, it has shown that all possible constraints on the tilt (and the running) are of the same order, and a full analysis should consider all these effects at the same time.
Current data prefers the tilt to be positive and the running to be negative. The posterior parameters ranges in the combined analysis are reduced by factor 5 for the tilt and infinitely for the running (there is no constraint from the CMB). As expected, the data is not yet capable of putting the inflation consistency condition to the test, but this work suffices as a confirmation that such an analysis is possible and that the combined data is already much more constraining than the CMB alone. A true test of the inflation consistency condition requires improved CMB data and measurement of gravitational waves at frequencies that are not dominated by foregrounds.
We also showed that for large positive values of the tensor tilt, one expects the background evolution to be modified, altering the peak structure of the CMB. There are some caveats in this analysis, since it requires an explicit assumption about the UV cutoff of the tensor spectrum and one could in principle introduce other ingredients with degenerate cosmological effects, i.e. this measure is not as `clean' as direct constraints obtained with gravitational wave detectors. For that reason, one should be careful interpreting these results. Although analytical results indicated that the constraints would be better, performing the actual analysis showed that parameter degeneracies lead to bounds on the tilt that are slightly worse ($k_{\rm UV}/k_* = 10^{20}$) or only marginally better ($k_{\rm UV}/k_* = 10^{24}$) than constraints from direct gravitational wave detectors.
Constraints on $n_t$ from direct CMB measurements are obviously dependent on a non-zero detection of $r$, and are unlikely to greatly improve with future measurements given the large foregrounds and the limited lever arm for B-modes. On the other hand, direct measures place useful upper bounds on $n_t$ which are not particularly sensitive to the value of $r$. In the near future we expect LIGO to improve its bound on gravitational waves and as such on $n_t$. At the same time, the background expansion will also be measured more accurately, with an expected error on the number of massless degrees of freedom of order a percent (see for example \cite{Benson:2014qhw}). Hence, we expect both measures to be equally important. In the long term however, testing the consistency condition would require a more ambitious direct detection experiment with a much higher level of precision. For that reason alone, we anticipate the analysis presented in this paper to be of interest in the future, especially if the value of $r$ is constrained at the percent level.
\section*{Acknowledgments}
BH was supported by the Astrophysics summer school program. RH would like to thank David Marsh and Kendrick Smith for useful discussions. JM would like to thank Aaron Zimmerman for helpful conversations. PDM would like to thank Dan Grin, Marilena Loverde, Dan Green, Thorsten Battefeld, Marco Peloso, Vitaly Vanchurin, Eiichiro Komatsu and Kipp Cannon for insightful discussions on the gravitational wave energy density. We also like to thank Paul Steinhardt for a very interesting and insightful discussion on inflation.
|
\section{Introduction}
In the study of complex systems -- such as the brain, cells or our economies -- we face conceptual issues of a novel type, because the systems studied involve many variables, many of which are unknown. In addition, their behaviour is not constrained by well established laws, as in physics.
In such high dimensional inference problems one is hardly ever sampling correctly an underlying probability distribution, even with huge data sets. In order to evade the deep under-sampling domain, we implicitly or explicitly resort to dimensionality reduction schemes, where the data is projected into a low-dimensional space where statistics can provide accurate conclusions. Yet, in this process, the data processing inequality \cite{cover91} tells us that we inevitably loose relevant information on the system's ``laws of motion''. So understanding which are the relevant variables is crucial in order to limit information losses. This requires guiding principles for the choice of dimensional reduction schemes, or for measuring the relevance of a given set of variables.
Recently Ref. \cite{marsili13} suggested that the entropy of the frequency of observations (see later) can be used as a measure of relevance of a given representation of the data. This allows one to characterise {\em most informative samples} as those that maximise this measure, at a given resolution and for a given sample size. Remarkably, one finds that most informative samples, in the under-sampling regime, have a power law frequency distribution \cite{marsili13}.
This finding sheds light on the widespread observation of ``criticality'' (i.e. power law frequency/size distributions) in empirical data \cite{clauset09} ranging from language \cite{zipf32}, statistics of natural images \cite{ruderman94}, neural activity \cite{eguiluz05,schneidman06}, city size distribution \cite{gabaix99}, to name just a few cases. In brief, this strongly suggests that the observed power laws usually associated with ``criticality'' arise as a consequence of our choice of relevant variables and that they do not necessarily require hidden mechanisms of self-organisation to a critical point \cite{bak96}. Besides the academic interest of such an interpretation of ``criticality'', its implication for data analysis are far reaching because the proposed measure of relevance can be used as a universal guiding principle in the search of optimal dimensional reduction schemes (e.g. data clustering) or for the identification of relevant variables (e.g. keywords in texts, relevant amino acids in proteins) \cite{marsili13}.
The purpose of this paper is to ground the finding of Ref. \cite{marsili13} in a model selection Bayesian framework, thereby clarifying its information theoretic basis. In brief, within this approach, we shall see statistical models of the data emerge from a process of symmetry breaking between data points in the sample\footnote{In what follows, a sample is a sequence of data points, each of which belong to a set of possible outcomes, which are defined {\em a priori}.}, acquiring more and more details as the size of the sample increases. In this way, model selection informs us on what resolution in the space of outcomes is justified by the data. In order for different outcomes to be assigned different probabilities, the frequency with which they occur in the sample must be sufficiently different. Formally, this identifies an {\em optimal partition} which distinguishes outcomes that occur with different probabilities. The entropy of the size of the partitions provides a measure of the number of outcomes that can be distinguished in the sample (or of the number of parameters that can be estimated from the samples) and hence a measure of relevance. In what follows, for the sake of simplicity, we shall define and refer to this measure as \emph{relevance}.
The next section introduces the generic problem we deal with and discusses model selection. Simple examples are presented to provide the main intuition.
We shall first show that, barring atypical cases, an upper bound to the relevance is given by partitions in frequency classes. Next we shall see that most informative samples are characterised by power law frequency distributions.
This will be followed by an application to two different examples of real data sets. The results suggest that the entropy of the frequency as suggested in Ref. \cite{marsili13}, can be used in place of the entropy of the optimal partition, which is computationally more demanding, as a measure of relevance. A final discussion will close the paper.
\section{The problem}
Let $\hat s=(s^{(1)},\ldots,s^{(M)})$ be a dataset of $M$ observations of the state $s$ of a system.
Here $s^{(i)}$ is a discrete variable, that we can think of as the label of the cluster to which the $i^{\rm th}$ observation belongs, or the configuration $s=(s_1,\ldots,s_n)$ of a system of $n$ discrete degrees of freedom (e.g. the amino acid sequence of a protein domain). The number of possible different states $s$ is much larger than $M$ and it may even be unknown. We restrict attention to the case where $s^{(i)}$ can be thought of as outcomes of independent experiments, carried out in the same conditions.
The general question of interest is to infer the laws governing the system, from the data. This can be formalised by assuming that the data can be thought of as $M$ i.i.d. draws from a generative model $P\{s^{(i)}=s\}=p_s$, where the function $p_s$ should encode the property of the system and the functions it performs. The basic problem then becomes that of inferring the generative model.
\subsection{Resolution and relevance}
Reference \cite{marsili13} has shown that, if we think of each sample $s^{(i)}$ as a realisation of an optimisation problem of a function $U(s,\bar s)$ over an enlarged set of variables that includes also unknown variables ($\bar s$), then the frequency
\[
k_s =\sum_{i=1}^M \delta_{s^{(i)},s}
\]
with which a given observation $s$ occurs in the sample provides a noisy estimate of that part $u_s=E_{\bar s}[U(s,\bar s)]$ of the function that is being optimised. Hence the relevance of the particular choice of the variables $s$, among all those that enter $U$, is reflected in the statistics of the frequency $k_s$ of states $s$. Ref. \cite{marsili13} argues that a quantitative measure of \emph{relevance}, in information theoretic terms, is given by
\begin{equation}
\label{eqHk}
\hat H[K]=-\sum_{k} \frac{k m_k}{M}\log \frac{km_k}{M},
\end{equation}
where
\[
m_k=\sum_s\delta_{k_s,k}
\]
is the number of states that occur $k$ times in the sample $\hat s$. Notice that $\hat H[K]$ is the entropy of the random variable $K_i=k_{s^{(i)}}$ for a randomly chosen point $s^{(i)}$ of the sample. This is different from the entropy of the state $s$ itself \footnote{Again we use uppercase for random variables defined on the space of the points in the sample $\hat s$. Also we assume maximum likelihood estimates of the probability $P\{S=s\}=k_s/M$.}
\begin{equation}
\label{eqHs}
\hat H[S]=-\sum_s\frac{k_s}{M}\log\frac{k_s}{M}=-\sum_{k} \frac{k m_k}{M}\log \frac{k}{M}.
\end{equation}
Intuitively, this measures the {\em resolution} of the description based on the variable $s$. Indeed a more detailed definition of the state $s$ of the system likely results in a higher {\em resolution} (i.e. a larger value of) $\hat H[S]$ but not necessarily in a higher {\em relevance} $\hat H[K]$.
\subsection{Learning the generative model}
Given a generative model $\mathcal{M}=p_s$, the likelihood of $\hat s$ is defined as:
\begin{equation}
\label{ }
P(\hat s|\mathcal{M})=\prod_{i=1}^M p_{s^{(i)}}=\prod_s p_s^{k_s},~~~~k_s=\sum_{i=1}^M \delta_{s,s^{(i)}}.
\end{equation}
The frequentist approach estimates the best model as the one that maximises the likelihood. This results in equating probabilities with frequencies: $\hat p_s=k_s/M$. The Bayesian approach, instead, invokes Bayes rule to turn the likelihood into a (posterior) distribution over the parameters $\vec p$ of the model. This requires identifying a {\em prior} distribution $P_0(\vec{p})$ that reflects our ignorance on $\vec p$ before seeing the data. A minimal requirement is that $P_0(\vec p)$ should be a symmetric function of the $p_s$'s. Dirichelet priors
\begin{equation}
\label{ }
P_0(\vec p)=\Gamma\left(\sum_{s} a_{s}\right)\prod_{s}\frac{p_{s}^{a_{s}-1}}{\Gamma(a_{s})}\delta\left(\sum_{s} p_{s}-1\right)
\end{equation}
are a mathematically convenient choice, and ignorance requires by symmetry that $a_s=a$ is independent of $s$. The posterior is easily computed:
\begin{equation}
\label{ }
P_1(\vec p)=\Gamma\left[\sum_{s} (k_{s}+a)\right]\prod_{s}\frac{p_{s}^{k_{s}+a-1}}{\Gamma(k_{s}+a)}\delta\left(\sum_{s} p_{s}-1\right)
\end{equation}
This allows us to give a Bayesian estimate of the probabilities
\begin{equation}
\label{ }
\langle p_s\rangle_1=\int\!d\vec p p_s P_1(\vec p)=\frac{k_s+a}{M+aS}
\end{equation}
where $S$ is the number of states. When $M\gg aS$ this converges to the frequentist estimate $k_s/M$, reminding us that in the presence of a large enough data set, the choice of the prior does not matter.
There are a number of problematic issues with this procedure:
\begin{enumerate}
\item The set of possible states and their number $S$ should be known in advance. This is not always the case.
\item The estimate of the entropy $H[S]=-\sum_s \langle p_s\log p_s\rangle_1$ is strongly affected by the prior and it converges slowly to its true value, as shown in Ref. \cite{nemenman01}.
\item The model assumes a different parameter for each state that occurs in the sample. A posteriori, this assumption is not justified as there is nothing that can be learned from the data on how the probabilities of two states that are seen the same number of times differ. Indeed, the posterior estimate of these probability depends on the frequency $k_s$ and is exactly the same for two states $s,s'$ that occur the same number of times $k_s=k_{s'}$.
\end{enumerate}
In particular, the last point suggests that we are in a clear case of over-fitting and indeed this model does not survive a model selection test, as we shall see in what follows.
\section{Model selection}
The key issue is that the definition of states $s$ is made by the observer, not by the system.
If the distinction between $s$ and $s'$ is totally spurious, we expect that the data will not distinguish between the two states, i.e. $k_s \approx k_{s'}$. Conversely, if two states are seen the same number of times, there is no reason to assume that they have a different probability. In terms of inference, we are not allowed to think that $p_s \neq p_{s'}$ unless we have sufficient evidence.
\subsection{An illustrative case: two states}
Let there be only two states $s=0,1$ and assume there are $M$ observations, $k=k_1$ with $s=1$ and $M-k$ with $s=0$. There are two possibilities: one is that the two states are actually the same, i.e. that the underlying distribution has $p_0=p_1=1/2$, the other that the states are different, i.e. $p_1=p=1-p_0$.
These correspond to different models that we can identify with different partitions of states and the associated probabilities. So the first case corresponds to a model $\mathcal{M}_0=[(\{0,1\},1/2)]$ where the two states are symmetric, whereas the second to a model $\mathcal{M}_1=[(\{0\},1-p),(\{1\},p)]$.
Clearly $P\{\hat s|\mathcal{M}_0\}=2^{-M}$ whereas for $\mathcal{M}_1$ the likelihood $P\{\hat s|\mathcal{M}_1\}$ can be obtained by integrating the likelihood over the prior distribution of the parameter $p$, for which again we take a Dirichelet form. Hence
\begin{equation}
\label{ }
P\{\hat s|\mathcal{M}_1\}=\frac{\Gamma(2a)\Gamma(k+a)\Gamma(M-k+a)}{\Gamma(a)^2\Gamma(M+2a)}.
\end{equation}
In order to compare the two models, we invoke Bayes rule and compute the posterior probability
\[
P(\mathcal{M}_i|\hat{s})=\frac{P(\hat{s}|\mathcal{M}_i)P_0(\mathcal{M}_i)}
{\sum_j P(\hat{s}|\mathcal{M}_j)P_0(\mathcal{M}_j)}=\frac{P(\hat{s}|\mathcal{M}_i)P_0(\mathcal{M}_i)}
{P(\hat{s})}
\]
where $P_0(\mathcal{M}_i)$ is the prior probability of model $i$. For the sake of simplicity, we're going to assume that all models are {\em a priori} equally likely\footnote{By Occam's razor, one would be tempted to prefer simpler models, i.e. those with fewer parameters. Yet Occam's razor already arises from the integration over the parameters implied by Bayes rule, without the need to introduce it {\em ad hoc}.}.
So the most probable model is the one with the highest likelihood $P\{\hat s|\mathcal{M}\}$. In the present case, it is easy to check that, for $M\gg 1$, in the representative case of a uniform prior ($a=1$) we have that as long as
\[
\left|\frac{k}{M}-\frac{1}{2}\right|<\sqrt{\frac{\log(2M/\pi)}{2M}}
\]
the symmetric model $\mathcal{M}_0$ should be preferred.
Figure \ref{three_states} shows an extension for the 3-states case. Here the possible models are $\mathcal{M}_0$ with no parameters (each state has $p=1/3$), $\mathcal{M}_{1,i}$ where two out of the three states have the same probability ($p_i=p$ and $p_s=(1-p)/2$ for $i=1,2$ or $3$), and $\mathcal{M}_2$ where all states have a different probability. If the frequencies are close enough, the states should not be distinguished and the model with no parameters should be preferred (blue surface in \ref{three_states}). Conversely the red surface reflect the cases where two states should not be distinguished from each other, and the green shows the case were the three states should be distinguished.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{Data_figures/three_states.pdf}
\caption{Model selection in a three state system with $M=150$ observations. $k_1$, $k_2$ and $k_3$ are the number of observations of each state. The coloured surface shows the preferred model in terms of the likelihood $P(\hat{s}|\mathcal{M}_i)$. $\mathcal{M}_0$ is the model with no parameters ($p_i=1/3, \forall i$), $\mathcal{M}_1$ is the one with one parameter ($p_i=p, ~p_{j}=p_k=(1-p)/2$), and $\mathcal{M}_2$ is the one with two parameters ($p_i=p,~ p_j=q,p_k=(1-(p+q))$).}
\label{three_states}
\end{figure}
\subsection{The general case}
The argument above suggests that, in the general case, for each pair of states $s$ and $s'$ their probability should be the same, unless they occur in the data a sufficiently different number of times. If $k_s\approx k_{s'}$ instead, they should be assigned the same probability, i.e. the symmetry between states $s$ and $s'$ should not be broken.
Conversely, imagine the situation where the distinction between states $s$ and $s'$ is completely arbitrary, with no relation with the internal states of the system under study. Complete ignorance of the system about the distinction between states $s$ and $s'$ means that the probability distribution restricted to only these two states must be the one of maximal entropy, i.e. that $p_s=p_{s'}$.
We remind again that the definition of states $s$ is made by the observer, not by the system. If it distinguishes effectively different internal states of the system, then this definition is relevant and meaningful, otherwise it is not. One way to turn this observation into a quantitative criterium is to extend the model selection argument above.
Given the set $\mathcal{S}$ of states $s$ that are seen (with multiplicity $k_{s}>0$), then a generic model $\mathcal{M}=[\mathcal{Q},\vec \mu]$ is one where different states are divided into a partition
\[
\mathcal{Q}=(Q_1,Q_2,\ldots,Q_N),\qquad \bigcup_{q=1}^N Q_q=\mathcal{S}
\]
of a number $N$ of disjoint sets, and each state in the $q^{\rm th}$ subset of the partition ($s\in Q_q$) has the same probability\footnote{All quantities $N$, $m_q$, $Q_q$ $\mu_q$ depend on the model $\mathcal{M}$. We omit this dependence for the sake of simplifying formulas.} $\mu_q$. If $m_q=|Q_q|$ is the number of states in subset $Q_q$, then $\mu_q$ satisfies the normalisation
\begin{equation}
\label{normal}
\sum_{q} m_q\mu_q=1.
\end{equation}
Any possible partition corresponds to a different model, going from the one where each state is in the same subset ($s\in Q_1,\forall s$), to the one where each state is in a different subset ($s\in Q_{s}, \forall s$).
It is possible to consider more general structure that also includes yet not seen states (i.e. states with $k_s=0$).
We shall see below that these are less likely than those considered here.
Each partition $\mathcal{Q}$ identifies a different model $\mathcal{M}$. This is why we shall use the partition $\mathcal{Q}$ to refer to the model that is based on that partition.
It is straightforward to compute the likelihood of each model:
\begin{equation}
\label{ }
P\{\hat s|\mathcal{Q}\}=\int\!d\vec \mu \prod_{q}\mu_q^{K_q}P_0^{(\mathcal{Q})}(\vec \mu),\qquad K_q=\sum_{s\in Q_q}k_{s}
\end{equation}
where the prior $P_0^{(\mathcal{Q})}$ contains the constraint Eq. (\ref{normal}).
We take again conjugate (Dirichelet) priors
\begin{equation}
\label{ }
P_0^{(\mathcal{Q})}(\vec \mu)=\Gamma(aN)\prod_{q}\frac{m_q^a}{\Gamma(a)}\mu_q^{a-1}\delta\left(\sum_{q\in\mathcal{Q}} m_q\mu_q-1\right)
\end{equation}
where $N$ is the number of partitions in $\mathcal{Q}$, i.e of parameters in $\mathcal{Q}$. Then
\begin{equation}\label{likelihood}
\log P\{\hat s|\mathcal{Q}\}=\sum_{q}\left[\log\frac{\Gamma(K_q+a)}{\Gamma(a)}-K_q\log m_q\right]-\log\frac{\Gamma(M+aN)}{\Gamma(aN)}
\end{equation}
The posterior distribution, under model $\mathcal{Q}$ is
\begin{equation}
\label{P1}
P_1^{(\mathcal{Q})}(\vec \mu|\hat s)=\Gamma(M+aN)\prod_{q}\frac{m_q^{K_q+a}}{\Gamma(K_q+a)}\mu_q^{K_q+a-1}\delta\left(\sum_{q} m_q\mu_q-1\right)
\end{equation}
The expected value of $p_{s}$ for $s\in Q_q$ is
\begin{equation}
\label{ps_post}
\left\langle p_{s}|\mathcal{Q}\right\rangle_{1}=\frac{1}{m_q}\frac{K_q+a}{M+aN}, \qquad \forall s\in Q_q
\end{equation}
where $\langle \ldots|\mathcal{Q}\rangle_{1}$ indicates expected values over the posterior distribution Eq. (\ref{P1}).
The expected value of the entropy $H[S]=-\sum_{s}p_{s}\log p_{s}$ is given by
\begin{equation}
\label{Hs}
\left\langle H[S]|\mathcal{Q}\right\rangle_{1}=-\sum_{q}\frac{K_q+a}{M+aN}\left[\psi(K_q+a+1)-\log m_q-\psi(M+aN+1)\right]
\end{equation}
where $\psi(z)=\frac{d\log \Gamma(z)}{dz}$ is the digamma function.
Assuming that all models are {\em a priori} equally likely, $P\{\hat s|\mathcal{Q}\}$ is also proportional to the posterior probability $P\{\mathcal{Q}|\hat s\}$ of model $\mathcal{Q}$ given the data. Therefore the optimal model is given by\footnote{A fully Bayesian approach would entail considering all possible partitions $\mathcal{Q}$ with their probability $P\{\mathcal{Q}|\hat s\}$. Here we depart from this approach and focus on the most likely partition.}
\begin{equation}
\label{ }
\mathcal{Q}^*={\rm arg}\max_{\mathcal{Q}} P\{\hat s|\mathcal{Q}\}.
\end{equation}
The partition $\mathcal{Q}^*$ identifies an emergent description of the system in terms of effective states $q$, that we shall call $q$-states. This is the statistical description that can be resolved on the basis of the dataset $\hat s$. The states $s\in Q_q^*$ in the same partition $q$ cannot be distinguished one from the other, hence they all correspond to the same $q$-state. The variable $q$ is associated to a distribution $p_q=m_q\mu_q$, which is the probability to observe the $q$-state. The entropy of this distribution $H[Q]=-\sum_q p_q\log p_q$ provides a quantitative measure of the amount of information that the data provides on the generative model. It's expected value on the posterior distribution Eq. (\ref{P1})
\begin{eqnarray}
\left\langle H[Q]|\mathcal{Q}\right\rangle_{1} & = & -\sum_{q}\frac{K_q+a}{M+aN}\left[\psi(K_q+a+1)-\psi(M+aN+1)\right] \\
& = & \left\langle H[S]|\mathcal{Q}\right\rangle_{1}-\sum_{q}\frac{K_q+a}{M+aN}\log m_q \label{HQdiff}
\end{eqnarray}
is what we shall call {\em relevance}. Indeed, this is a measure of the relevance of the original description based on the states $s$. Eq. (\ref{HQdiff}) shows that $\left\langle H[Q]|\mathcal{Q}\right\rangle_{1} \le \left\langle H[S]|\mathcal{Q}\right\rangle_{1}$ with equality if and only if all partitions $Q_q$ contain only one state ($m_q=1~\forall q$).
The next section illustrates the behaviour of this measure in some specific examples.
Before doing that, it is instructive to discuss the issue of unsampled states and two special cases, to make contact with the results of Ref. \cite{marsili13}.
\subsection{Unsampled states}
\label{sec:unsampled}
In many instances, the sample contains only a partial coverage of the set of possible states. There are two ways in which not yet sampled states could be included in one of the partitions $\mathcal{Q}$ discussed above. Either adding them to one or more of the sets $Q_q$ or augmenting the partition with a set $Q_0$ that includes all states with $k_s=0$. In the first case, the partition $\mathcal{Q}$ changes into one which is identical on all sets $Q_{q'}$ with $q'\neq q$ and with $Q_q\to Q_q'=Q_q\bigcup Q_0$, where $Q_0$ is the set of unseen states. Since $k_s=0$ for $s\in Q_0$, the count $K_q$ does not change, and the change in the likelihood is given by $-K_q\log(1+m_0/m_q)$, where $m_0=|Q_0|$ is the number of states $s\in Q_0$. Since the change in the likelihood is negative, the optimal partition $\mathcal{Q}^*$ does not include not yet sampled states.
The change in the likelihood when the unseen states are added to the partition in a new set, $\mathcal{Q}\to \mathcal{Q}_{+0}=(\mathcal{Q},Q_0)$ can also be easily computed. The first two terms in Eq. (\ref{likelihood}) do not change, as $K_0=0$, so the only difference is due to the fact that the number of sets increases by one: $N\to N+1$. Hence the change in the likelihood
\begin{equation}\label{dlikelihood}
\log \frac{P\{\hat s|\mathcal{Q}_{+0}\}}{P\{\hat s|\mathcal{Q}\}}=
-\log\frac{\Gamma(M+aN+a)\Gamma(aN)}{\Gamma(M+aN)\Gamma(aN+a)}
\end{equation}
is again negative.
Hence models based on partitions that include unseen states are dominated by those discussed above, if they are considered equally likely {\em a priori}.
Yet, if one expects that the sample contains only a partial coverage of the set of possible states, the uniform prior hypothesis needs to be revised. Therefore
\begin{equation}
\label{dpost}
\log \frac{P\{\mathcal{Q}_{+0}|\hat s\}}{P\{\mathcal{Q}|\hat s\}}=\Delta_0
-\sum_{k=0}^{M-1}\log\left(1+\frac{a}{k+aN}\right)
\end{equation}
where $\Delta_0=\log\frac{P_0\{\mathcal{Q}_{+0}\}}{P_0\{\mathcal{Q}\}}$ encodes the {\em a priori} likelihood that states $s$ that are not present in the sample $\hat s$ exist. Notice that the second term in Eq. (\ref{dpost}) increases with $M$ (as $a\log (1+M/(aN))$ for $M,N\gg 1$). Hence for a given $\Delta_0$, we expect the model $\mathcal{Q}$ to become preferable to $\mathcal{Q}_{+0}$ as $M$ grows large. When instead the model $\mathcal{Q}_{+0}$ is the optimal, this approach also gives an estimate of the discovery probability
\begin{equation}
\label{ }
p_0=\frac{a}{M+aN+a}
\end{equation}
which is an intense subject of research in statistical learning\footnote{This discussion relates to the wider field of non-parametric Bayesian statistics which discusses models that reproduce sampling processes. For a general introduction, the reader is referred to \cite{OrbanzTeh2010}. A model of the sampling process based on our approach departs from this literature in that non-parametric Bayesian models such as the Dirichelet's process are based on a single partition ($\mathcal{S}$ in this case) whereas we consider selecting the optimal partition for each $M$. Further discussion of this issue would bring us too far from the main aim of the present paper and will be dealt with elsewhere.}, since the work of Good and Turing \cite{goodturing}.
\subsection{Special cases}
For the model based on the atomic partition $\mathcal{S}$, where each subset contains one state $Q_s=\{s\}$
\begin{eqnarray}
\log P\{\hat s|\mathcal{S}\} & = & \sum_{s}\log\frac{\Gamma(k_{s}+a)}{\Gamma(a)}-\log\frac{\Gamma(M+aN_s)}{\Gamma(aN_s)}
\\
& = & \sum_k m_k\log\frac{\Gamma(k+a)}{\Gamma(a)}-\log\frac{\Gamma(M+aN_s)}{\Gamma(aN_s)}
\end{eqnarray}
where $N_s=|\mathcal{S}|$ is the number of different states $s$ that occur in the sample $\hat s$. Note that $m_s=1$ and $K_s=k_s$ is simply the frequency of state $s$.
For the model based on the frequency partition $\mathcal{K}$, where subset $Q_k=\{s:~k_{s}=k\}$ for $k=1,2,\ldots$, we have $K_k=km_k$ and
\begin{equation}
\log P\{\hat s|\mathcal{K}\} = \sum_k \log\frac{\Gamma(km_k+a)}{\Gamma(a)m_k^{km_k}}-\log\frac{\Gamma(M+aN_k)}{\Gamma(aN_k)}
\end{equation}
where $N_k=|\mathcal{K}|$ is the number of different values of $k_s$ that appear in the sample.
Na\"ively one would expect that $P\{\hat s|\mathcal{K}\}>P\{\hat s|\mathcal{S}\}$, i.e. that the $\mathcal{K}$ partition should always be preferred to the atomic partition $\mathcal{S}$. \ref{proof_KS} proofs that this is indeed the case for $a=1$ and for $a\to 0$. But it also exhibit counterexamples where this is not so, in the limit of large $a$. These however correspond to rather atypical samples and no counterexample to the rule $P\{\hat s|\mathcal{K}\}>P\{\hat s|\mathcal{S}\}$ has been found in the data we have analysed. This strongly suggests that, in practical terms, the $\mathcal{K}$ partition should always be preferred to the $\mathcal{S}$ partition.
\noindent
\section{Properties of the optimal partition $Q^*$}
\label{Qalgorithm}
Finding the optimal partition $\mathcal{Q}^*$ for a given sample $\hat s$ is a non-trivial task. It is reasonable to
assume that partitions that merge states with adjacent frequencies are more likely than those that merge states with non-adjacent frequencies\footnote{If $k_{s_1}>k_{s_2}>k_{s_3}$ then a partition $\mathcal{Q}$ where $s_1,s_3\in Q_{q_1}$ and $s_2\in Q_{q_2}$ will be dominated by partitions where either all three states are in different sets, or $s_1,s_2\in Q_{q_1'}'$ and $s_3\in Q_{q_2'}'$, or $s_1\in Q_{q_1'}'$ and $s_2, s_3\in Q_{q_2'}'$, or they are all in the same set.}. Therefore, it is enough to consider partitions where all states $s\in Q_q$ have frequency $k_s$ which is larger than that of all states $s'\in Q_{q'}$ with $q>q'$. This leads us to the following heuristics to derive the optimal partition $\mathcal{Q}^*$ of a finite sample:
\begin{enumerate}
\item Starting from $\mathcal{Q}=\mathcal{K}$:
\item For every $q=1,\ldots,N_{\mathcal{Q}}-1$, define a new partition $\mathcal{Q}^{(q)}$ by merging the subsets $Q_q$ and $Q_{q+1}$ of the current partition $\mathcal{Q}$ and compute the change in the log-likelihood.
\item If the largest increase in the log likelihood over all possible values of $q$ is positive, then merge the corresponding subsets, update the partition $\mathcal{Q}$ accordingly and repeat the previous step.
\item If the largest increase in the log likelihood over all possible values of $q$ is negative, then return $\mathcal{Q}^*$ as the optimal partition.
\end{enumerate}
In order to explore the properties of $\mathcal{Q^*}$ we study ensembles where the states $s$ are drawn from power law distributions $P(s)\sim s^{-\alpha}$. This choice serves for generating data with a broad distribution of frequencies, such as those that are often observed in empirical studies. Varying $\alpha$ allows us to probe the merging algorithm proposed over a broad range of underlying distributions.
Figure \ref{barcode} gives a pictorial representation of the merging process during a typical run. Interestingly, visual inspection suggests that the frequencies of the optimal model $\mathcal{Q}^*$ are evenly spaced in a logarithmic scale.
One can think as well of variations to the algorithm such as selecting a favourable move at random in step (ii) instead of choosing the one that maximizes the likelihood, or merging triplets of subsets instead of pairs. We have seen that the overlaps in the final representations obtained using these variations in the algorithm are always larger than $90\%$. Moreover we see that for large samples the probability of finding a representation with $H[Q]$ greater than $H[Q^*]$ goes to zero, meaning that the later yields a more relevant description of the data. This issue is discussed in \ref{algorithms}.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{Data_figures/barcode.pdf}
\caption{Illustration of the Merging Process. $M=10^5$ data points were drawn from a distribution $P(s)\sim s^{-\alpha}$, with $\alpha=1.2$. The x-axis shows the estimated probability (\ref{ps_post})
for states in each subset $Q\in\mathcal{Q}$. The y-axis stands for the different partitions $\mathcal{Q}$ in the merging process from $\mathcal{K}$ to $\mathcal{Q}^*$.}
\label{barcode}
\end{figure}
\subsection{Distance between $\mathcal{Q^*}$ and $\mathcal{K}$ and scaling with the sample size}
Figure \ref{N_M_scaling} shows the difference between $\mathcal{Q^*}$ and $\mathcal{K}$ as a function of the sample size $M$. Panel A shows the estimated parameters $\left\langle p_s\right\rangle$ (Eq. \ref{ps_post}) for both models and two sample sizes $M_1=10^3$ and $M_2=10^6$. The states with higher frequency $k_s$ are not merged, so the partitions $\mathcal{S}$, $\mathcal{K}$ and $\mathcal{Q}^*$ overlap on the left tail of the curve on a number $\xi$ of identical subsets of states. We estimated the $\mathcal{Q^*}$ partition and the parameters $p_s$ using priors with $a$ ranging from 0.01 to 10. The different overlapping curves in panel A stand for the different values of $a$. Clearly neither the number of subsets in $\mathcal{Q^*}$ ($N_{\mathcal{Q}^*}$) nor the estimated parameters $\langle p_s\rangle$ vary strongly with $a$. In the following analysis we set $a=1$.
Panel B shows that the overlap $\xi$ between the two partitions scales with $M$ with a non-trivial exponent ($\gamma$) which depends on the underlying distribution parametrized by $\alpha$ (panel C).
The number of parameters ($N$) in each partition gives a measure of the overfitting done in $\mathcal{K}$ with respect to $\mathcal{Q^*}$. Panel D shows that $N\sim M^\delta$ has a power law dependence on $M$ with an exponent $\delta$ that depends on $\alpha$\footnote{For the $\mathcal{K}$ partition it is possible to show that $\delta=1/(1+\alpha)$. The argument relies on the fact that the frequency of state $s$ approximates the probability $k_s/M\simeq p_s\sim s^{-\alpha}$ as long as $k_s\gg 1$ is large enough.
We note that $m_k\simeq ds/dk$ is the number of states $s$ in an interval of frequency $dk=1$, hence $m_k\sim s^{\alpha+1}/M\sim k^{-1/\alpha-1}M^{1/\alpha}$. The number $N_k$ of states corresponds to the value of $k$ such that $m_k$ becomes of order one. Therefore $N_k\sim M^{1/(\alpha+1)}$. Interestingly, we also find that $\gamma=\delta/2$ for the $\mathcal{K}$ partition, to numerical precision. These relations do not hold for the $\mathcal{Q}$ partition.}
(panel E). The exponent $\delta$ for the $\mathcal{Q}^*$ partition is smaller than that of the $\mathcal{K}$ partition implying that the difference between $N_{\mathcal{Q}^*}$ and $N_{\mathcal{K}}$ increases with $M$.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{Data_figures/Nparam_scaling.pdf}
\caption{Scaling of the optimal partition with the sample size $M$. \textbf{Panel A} shows the estimated parameters $<p_s>$ for each subset in partitions $\mathcal{K}$ and $\mathcal{Q^*}$. The data was drawn from a power law distribution $P(s)\sim s^{-\alpha}$, with $\alpha=1.2$. For both partitions we show the estimated parameters for a sample of size $M_1=10^3$ and $M_2=10^6$. $\xi$ denotes the number of parameters which are identical under both models $\mathcal{K}$ and $\mathcal{Q^*}$. $N_{\mathcal{K}}$ and $N_{\mathcal{Q^*}}$ are the number of parameters (subsets) in each model. The different overlapping red (black) curves correspond to estimations using different values for the prior parameter $a$, ranging from 0.01 to 10. Panels B-E show analysis using $a=1$. \textbf{Panel B} shows the scaling of $\xi$ with the sample size, for $\alpha=1.2$, which follows a power law $\xi\sim M^{\gamma (\alpha)}$. \textbf{Panel C} shows the dependence of $\gamma$ with $\alpha$. \textbf{Panel D} shows the scaling of the number of parameters in each model with the sample size, which follows a power law $N\sim M^{\delta(\alpha)}$. \textbf{Panel E} shows that the number of parameters in $\mathcal{Q^*}$ grows slower with $M$ than in $\mathcal{K}$ for a wide range of systems ($\alpha$).}
\label{N_M_scaling}
\end{figure}
An interesting observation is that the optimal partition $\mathcal{Q}^*$ provides an estimate of the entropy of the underlying distribution that converges faster than that based on the $\mathcal{S}$ partition. The slow convergence of the entropy based on the $\mathcal{S}$ partition and its strong dependence on the prior where noticed in Ref. \cite{nemenman01}, that also proposed a remedy based on treating $a$ as a hyper-parameter in Bayesian inference. Figure \ref{Hs_convergence} shows that the estimate based on the optimal partition $\mathcal{Q}^*$ converges faster than finer representations, and that Bayesian inference and model selection are enough to have a reliable estimate of the entropy. This also suggests that the information kept in the coarser representation is truly relevant for characterising the sample, while the discarded information is noise associated with the finite number of data points.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Data_figures/Hs_convergence_a1_alpha3.pdf}
\caption{Convergence of the estimated entropy under different models. Labels $\mathcal{S}$, $\mathcal{K}$, $\mathcal{Q^*}$ stand for the bayesian estimates of the entropy (\ref{Hs}) in the respective model using a flat prior ($a=1$). $ML$ stand for the maximum likelihood estimate (\ref{eqHs}). The dashed line is the true entropy of the underlying distribution $P(s)\sim s^{-3}$. Error bars denote standard errors over 1000 samples of each size $M$. The inset shows the difference between the bayesian estimates based on the posterior distribution (\ref{Hs}) and the likelihood of the model (\ref{likelihood}) $\Delta H[\mathcal{S}]=\frac{1}{H^*}\left[ \left<H[\mathcal{S}]|\mathcal{Q}\right>_1- (-\frac{1}{M}\log(P\{\hat s|\mathcal{Q}\}))\right] $ where $H^*$ is the true entropy of the distribution.}
\label{Hs_convergence}
\end{figure}
\section{Criticality of maximally informative partitions}
Having provided a measure for the relevance of a given sample, allows one to characterise the typical properties of most relevant samples, i.e. of samples that are maximally informative. This question was partly addressed in Ref. \cite{marsili13}, where an upper bound to the entropy $\hat H[K]$, for a given sample size $M$ and at a given resolution $\hat H[S]$, was derived. Interestingly, this exercise shows that the distributions that achieve the upper bound in the under-sampling regime, are power laws, i.e. $m_k\sim k^{-\mu-1}$. This suggests that ``criticality'', i.e. the observation of scale-free frequency distribution, may be a consequence of choosing the most informative variables, and need not necessarily imply underlying mechanisms of self-organisation to a critical point.
In \ref{poissonisation} we revisit the argument leading to the upper bound and also derive a lower bound for $\hat H[K]$, showing that this is also achieved when the distribution of frequencies has a power law behaviour $m_k\sim k^{-\mu-1}$.
The observation (see Fig. \ref{barcode}) that model selection identifies partitions $\mathcal{Q}^*$ with posterior probabilities $\langle p_s|\mathcal{Q}^*\rangle_1$ that are evenly spaced on a logarithmic scale, suggests that the same may be true for samples of a given size $M$, with a maximal $\langle H[Q]|\mathcal{Q}^*\rangle_1$ at a given resolution $\langle H[S]|\mathcal{Q}^*\rangle_1$.
Yet, in order to further corroborate this conclusion, one needs to resort to numerical simulations. To this end, we generated samples from Montecarlo simulations maximising the measures of relevance proposed above.
The simulations consisted in the following steps:
\begin{enumerate}
\item Start with an arbitrary sample defined by the frequencies $\hat k=(k_1,...,k_N)$, with $\sum_s^N k_s=M$, and $N$ the initial number of states. Without loss of generality sort the frequencies in decreasing order $k_1\ge ...\ge k_N$.
\item Consider every possible move of $n$ samples from state $i$ into state $j\ne i$ for all $i\in [1,N]$,$j\in [1,N+1]$ and $n$, under the constraint that $k_i-n\ge k_{i+1}$ and $k_{j-1}\ge k_j+n$. Notice that $j=N+1$ implies defining a new state with frequency $n$. Conversely if $k_i=1$ this state will disappear after moving the one sample to state $j$.
\item Choose the move which maximizes the Lagrange function $\mathcal{L}=H[Q]+\mu H[S]$, with $Q=K$ and $Q=Q^*$ independently, for a fixed value of $\mu$.
\item Repeat (ii)-(iii) until there is no favourable move.
\end{enumerate}
Keeping $\mu$ fixed in step (iii) and allowing H[S] to fluctuate accordingly during the simulation favoured the ergodicity of the dynamics with respect to fixing H[S] and maximizing H[Q]. We repeated the simulations for different values of $\mu \in [-1,5]$. For each value of $\mu$ we repeated the simulations with different initial conditions. This was not essential for the maximization of H[K] but for the maximization $H[Q^*]$ the process converged to local maxima strongly dependent on the initial conditions. We therefore varied the initial number of states $N$ from 25 to 950, for a sample of size $M=1000$, and performed 20 independent realizations for each initial resolution. The absolute maximum of $H[Q^*]+\mu H[S]$ across realizations was kept for each $\mu$. Panel A in figure \ref{MC_samples} shows the results for both relevance measures $H[K]$ and $H[Q*]$. Values of $\mu<-1$ yield the trivial result of $H[Q]=0$ and $H[S]=0$ which corresponds to the solution $k_1=M$, $k_{j>1}=0$. $\mu=-1$ yields solutions with $H[Q]=H[S]$ in the well sampled regime (left part of the diagram). In the case of $Q=K$, the solutions are of the form $m_k\sim k^{-\mu-1}$ (see \ref{poissonisation}). Panel B shows the solutions obtained for $\mu=1$ which match the expected Zipf law. The dashed curves in panel A refer to theoretical upper and lower bounds for the value of $H[K]$ (\ref{poissonisation}).
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Data_figures/MC_samples_new.pdf}
\caption{Maximally relevant samples. Samples of size M=1000 were generated via Monte Carlo simulations maximizing $H[K]+\mu H[S]$ and $H[Q^*]+\mu H[S]$ (see main text), for fixed values of $\mu \in [-1,5]$ which determine the resolution $H[S]$ of the solutions. All entropies are normalised by $\log(M)$. Panel A shows the maximized $H[K]$(red) and $H[Q^*]$(blue) together with theoretical bounds for $H[K]$ (dashed). The solutions expected are of the form $m_k \sim k^{-\mu-1}$. Panel B shows the solutions obtained for $\mu=1$, which are power laws of exponent 2 (Zipf's law). The dashed line is an approximate solution to the theoretical lower bound for $H[K]$ where $m_k$ is assumed to be a poisson variable (see \ref{poissonisation}).}
\label{MC_samples}
\end{figure}
\section{Application to real data}
In this section we compare the models based on the $\mathcal{K}$ and $\mathcal{Q^*}$ partitions in two applications to real data. The $\mathcal{K}$ partition is derived directly and exactly from the data whereas the $\mathcal{Q^*}$ partition requires a calculation that may be heavy and approximate. The scope of this section is to show that in practical cases, the $\mathcal{K}$ partition is a very good approximation to the optimal one $\mathcal{Q^*}$. Intuitively, the reason why this is so relies on the fact that informative samples (those with a large $H[Q^*]$ or $H[K]$) have broad frequency distributions, and as we have seen, the $\mathcal{Q}^*$ and $\mathcal{K}$ partitions have a sizeable overlap in these cases.
In the first example, we analyse a financial market data set of stock returns. The data set (used previously in \cite{marsili02b,marsili02a}) span a period from 1st January 1990 to 30th April 1999 (2249 time points) and it covers the $M=2000$ most frequently traded stocks in the New York Stock Exchange in that period. Assuming that returns are gaussian with a block diagonal correlation matrix allows one to group stocks in clusters of ``sectors'', by maximum likelihood (see \cite{marsili02a,marsili02b} for details).
The cluster label $s_i$ of each stock $i=1,\ldots,M$ identifies the $\mathcal{S}$ partition in this context. As the number $N_s$ of clusters varies from $1$ to $M$, the algorithm produces partitions $\mathcal{S}$ with a different resolution $H[S]$. We compare the relevance of different levels of description by computing $H[K]$ and $H[Q^*]$. Here the optimal partition $\mathcal{Q}^*$ is obtained with the algorithm defined in Section \ref{Qalgorithm} starting from $\mathcal{K}$.
Panel A in figure \ref{Financial} shows both measures of relevance as a function of the resolution $H[S]$. The dashed curves are theoretical upper and lower bounds to the estimate of the maximal value of $H[K]$, given $H[S]$ and $M$ (see \ref{poissonisation}). Panel B illustrates the relation between the $\mathcal{K}$ and $\mathcal{Q}^*$ representations, at the resolution marked by the vertical dashed line in panel A. Bars in panel B denote the $\mathcal{K}$ partition, whereas the colours indicate which frequencies were merged together to form the coarser optimal model $\mathcal{Q}^*$.
Panels C and D provide a closer look at the distance between partitions $\mathcal{K}$ and $\mathcal{Q}^*$. Panel C shows the overlap between the $\mathcal{K}$ and $\mathcal{Q}^*$ partitions at each resolution. The overlap was computed by the Adjusted Rand Index (\cite{hubert85}), which is bounded above by 1 and yields 0 when the overlap matches the one expected by chance. For illustrative purposes we show the overlap between shuffled versions of the partitions (red curve), which indeed yield constant zero for all resolutions. We point out that in the strongly under-sampled domain, both models are practically the same.
Panel D shows the estimated parameters (Eq. \ref{ps_post}) in the $\mathcal{K}$ and $\mathcal{Q}^*$ models, at the resolution marked by the vertical line in A. The dashed line is a Zipf law for comparison.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{Data_figures/Financial_3.pdf}
\caption{$\mathcal{K}$ and $\mathcal{Q}^*$ representations of a financial market data set of stock returns. Stocks where clustered at different resolutions ($H[S]$) by means of the algorithm developed in Ref. \cite{marsili02b}. The relevance of each level of description is quantified by $H[K]$ and $H[Q^*]$ in Panel A (all entropies are normalized by $\log(M)$). Panel C shows the overlap between partitions $\mathcal{K}$ and $\mathcal{Q}^*$ at each resolution. The overlap index is the Adjusted Rand Index (ARI, Ref. \cite{hubert85}) which yields 1 for identical partitions and 0 when the overlap is expected by chance. The partition labels where shuffled before computing the ARI (red curve) to illustrate this case. Panels B and D refer to partitions $\mathcal{K}$ and $\mathcal{Q}^*$ of the data, at the resolution marked by the vertical dashed line in A ($H[S]\sim 0.75$). Bars in Panel B show the $\mathcal{K}$ partition. The x-axis are the frequencies and the y-axis are the number of states seen with each frequency. The colours show the coarser $\mathcal{Q}^*$ partition, obtained by the algorithm of section \ref{Qalgorithm}. Panel D shows the estimated parameters with the model based on $\mathcal{K}$ and $\mathcal{Q}^*$. The dashed line is a Zipf law for comparison.}
\label{Financial}
\end{figure}
As a second example, following Ref. \cite{marsili13}, we consider the problem of identifying relevant positions in the sequence of amino acid that correspond to a particular protein domain. In brief, the data consists of Multiple Sequence Alignments (MSA) of $M$ sequences of the same protein domain, across different species. We refer to \cite{grigolon15} for a detailed description, for our purposes here, suffice it to say that a protein domain can be identified by a sequence $\vec a=(a_1,\ldots,a_L)$ of $L$ amino acids, each being of one of 21 possible types (e.g. $a_i=V$ for valine, $a_i=A$ for alanine, etc) and that an MSA is a collection of $M$ such sequences across different organisms or species. The key point is that, while the whole sequence is subject to a random process of mutations, there are features which need to be conserved in order to perform the function the protein is supposed to do. In order to understand which positions along the sequence are relevant for the biological function, we observe that each subset $I\subseteq \{1,2,\ldots,L\}$ of the positions identifies a partition $\mathcal{S}_i$ of the MSA data, whose elements $s=(a_{i},i\in I)$ are the subsequences of the domain on the positions $i\in I$.
From this we can define the $\mathcal{K}$ and the optimal $\mathcal{Q}^*$ partitions, and compute both the resolution $H_I[S]$ and the relevance $H_I[Q^*]$ or $H_I[K]$ corresponding to this subset of positions.
This makes it possible to look for the most relevant subset of positions $I^*$, as the one that maximises $H_I[Q^*]$.
This program is carried out in Ref. \cite{grigolon15} to which we refer the interested reader. Here we confine the discussion to the comparison of the $\mathcal{K}$ and $\mathcal{Q}^*$ partitions. In brief, the maximisation (of either $\mathcal{K}$ or $\mathcal{Q}^*$) is done using a Montecarlo algorithm for subsequences of a fixed number $n$ of amino acids. We applied the algorithm to the Voltage Sensor Domain of ion channels (Pfam code PF000520). The data was the same as that used in Klein {\em et al.} \cite{MSApaper}. In brief, the algorithm of Ref. \cite{grigolon15} produces a distribution over the subsets $I^*$ of $n$ relevant sites that allows one to compute the probability that a site is either selected or not in two different realisations of $I^*$. Fig. \ref{Protein} shows that optimising the relevance of the $\mathcal{K}$ or the $\mathcal{Q}$ partitions provides a sharp separation between relevant and irrelevant positions, which is sharper for the $\mathcal{K}$ partition. In addition, the selected subsets of sites $I^*_{\mathcal{K}}$ and $I^*_{\mathcal{Q}^*}$ under the optimisation of $H[K]$ or $H[Q^*]$ have a large relative overlap: in 90\% of the cases, the two optimisation schemes yield the same prediction on whether a site that is relevant or not (see Fig. \ref{Protein} and the caption for details).
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{Data_figures/modselect_protein.pdf}
\caption{Self-overlap $O_{\mathcal{K}}$ and $O_{\mathcal{Q}^*}$ (filled circles) of the subset of site $I^*_{\mathcal{K}}$ and $I^*_{\mathcal{Q}^*}$ produced by the optimisation of $H[K]$ and $H[Q^*]$, respectively, on the MSA for the voltage sensor protein domain. Self-overlaps are computed as the probability that a randomly chosen site is either selected or not in two different runs of the algorithm of Ref. \cite{grigolon15} and are shown, in the plot, as a function of the number $n$ of sites in $I^*$. The overlap $O(K,Q)$ between the subsets $I^*_{\mathcal{K}}$ and $I^*_{\mathcal{Q}^*}$ is also computed in the same way, and it is shown with filled squares. The dashed line corresponds to the overlap between subsets of $n$ sites chosen randomly among the $L=114$ possible sites.}
\label{Protein}
\end{figure}
These two examples suggest that, in practical applications, $H[K]$ can be used as a proxy for $H[Q^*]$ in measuring the relevance. This is particularly useful to avoid the optimisation leading from $\mathcal{K}$ to $\mathcal{Q}^*$ and speed up numerical calculations.
\section{Conclusion}
The Big Data revolution has made available data of unprecedented detail on the working of complex systems, such as cells, networks of neurons and the brain, ecologies, social networks, economies and financial markets. This, in particular, indicates that quantitative approaches typical of hard sciences can be extended to life sciences as well. Yet, the fact that such phenomena are not constrained by well known laws, as in physics, makes inference of behaviour a daunting task. Indeed, one is rarely in the circumstance where behaviour depends on only few variables that can be controlled. In such cases, the resolution of high dimensional data, is not given by the number of variables that one can measure simultaneously, but rather is limited by statistical errors induced by finite sample size. Dimensionality reduction schemes have to be invoked to adjust the resolution so that reliable statistical information can be extracted from the data. This inevitably introduces a tradeoff between relevance and resolution, which is addressed in this paper.
The main contribution of this paper is to make this tradeoff explicit and quantitative in information theoretic terms, on the basis of a Bayesian model selection approach. We focus on the limiting case where the system under study is severely under-sampled and no other information apart the frequency of observations is available. There, models are in one to one correspondence with partition of the set of observed states. So while resolution is a measure of the number of different states, relevance can be defined in terms of the number of different elements in the partition, i.e. the number of different states that the data allows one to distinguish. We find that, as resolution increases from the coarser possible level, relevance increases up to a maximum, beyond which it starts decreasing. In the extreme limit where each observation is seen only once, relevance vanishes, signalling that data contains no relevant information on the system.
The resolution (i.e. the number of sets in the partition) also provides a natural cutoff in the number of parameters that the data allows us to infer, beyond which inference would result in overfitting. The number of parameters (and of partitions) increases with the sample size $M$. Loosely speaking, as $M$ increases, the model passes through a sequence of symmetry breaking transitions where more and more distinctions between states can be made. This process, indeed, bears well known formal analogies with the symmetry breaking process in physical systems when the temperature (here proportional to $1/M$) decreases.
There are several interesting directions for further research along these lines. One is to extend the approach in Section \ref{sec:unsampled} to explore sampling processes \cite{goodturing} that are consistent with the Bayesian model selection scheme. The second is to exploit these results for inference of graphical models in cases where states can be considered as a configuration $s=(\sigma_1,\ldots,\sigma_n)$ of an extended system. There, a well established technique is Boltzmann learning (see e.g. \cite{schneidman06}) which, given a set of relevant observables, invokes maximum entropy principle and predicts a distribution $P(\sigma_1,\ldots,\sigma_n)$. The set of relevant observables determines the model. Yet, no general criterium exists for dictating what relevant observables should be and it seems natural to invoke model selection schemes to address the issue.
Finally, the present approach also suggests a new perspective on the widespread occurrence of criticality. It suggests that the occurrence of broad frequency distributions is a consequence of sampling relevant variables in the under sampling regime. In this spirit the interesting question is not whether or why "biological systems are poised at criticality" \cite{MoraBialek} but rather how to use the "apparent criticality" of frequency distributions to select relevant variables.
\section{Acknowledgements}
We thank W. Bialek, A. Celani, P. Latham, Y. Roudi and M. Vergassola for interesting discussions.
This work was supported by the Marie Curie Training Network NETADIS (FP7, 290038).
|
\section{Introduction}
Statistics provided by EPTC \cite{EPTC:Statistics} -- the traffic managing agency in Porto Alegre/RS (Brazil) --
shows that in 2013, approximately 22.447 traffic accidents took place in the city of Porto Alegre,
an average of 1.870 traffic accidents per month. According to Saunier et al. \cite{saunier}, the social cost of road
collisions is the largest side effect
of road transportation . The costs of fatalities, injuries and property damage, as well as medical
care and traffic delays accounts for a significant impact on the finance of the people involved, cities and the
government. According to Brazilian National Traffic Department (DENATRAN) \cite{DENATAN:CustosAcidentes}, the
average cost for an accident (in federal highways) without victims is R\$ 16.840,00, for accidents with victims
this cost increases to R\$ 86.032,00 and for accidents with fatalities, the cost is R\$ 418.341,00.
Descriptive analysis of the situation in road safety and road accidents are important, but understanding
the factors related with dangerous situations and patterns in data is of even greater importance \cite{Beshah}.
Being able to predict when an traffic accident will result in an injury, can help traffic agencies to
provide faster medical care.
Another example of the benefits of understanding the factors behind the
injury risk is to guide traffic agencies to improve the road safety by means of infrastructure design (which
includes road signs and speed control devices) or even through the pedestrian/driver behavior improvements that could be obtained
with targeted marketing campaigns.
Data-driven decisions can also help the traffic agencies to reduce
the costs involved in a traffic accident.
This paper describes the efforts and experimental results obtained through the application of Machine Learning
techniques in order to provide a better understanding of the data that is being collected today by the traffic
agency. The main purpose of this work, is to get an overall understanding of the accident data as well as
build a predictor for the injury risk related to the traffic accidents in the city of Porto Alegre/RS (Brazil).
This study is organized as follows: in Section \ref{related_work}, an overview of the related work is provided, followed
by the Section \ref{proposed_approach}, where an overview of our proposed approach
to the problem is described. In Section \ref{experimental_methodology}, the experimental methodology is provided,
including which algorithms were used an how they were used, while in Section \ref{experimental_analysis} the
results of the experimental analysis are detailed followed by the Section \ref{conclusion}, that describes the
conclusions and remarks related to future work.
\section{Related Work}
\label{related_work}
An attempt has been made to search for existing accident analysis practice in the city of Porto Alegre, however
no published works were found related to the application of Machine Learning techniques by the traffic managing
agency (EPTC) in Porto Alegre. Only limited descriptive analysis were published
on the site of the traffic agency \cite{EPTC:Statistics}, but no deeper analysis of the factors or
injury risk assessment were published by the traffic managing agency.
The lack of standardization for the data collection process and for the data itself between different
traffic managing agencies worldwide makes the experimental results comparison very limited.
Beshah et al. \cite{Beshah} explored a rich data set, comprising of 14.254 accident cases described with
48 attributes containing information related to road users (drivers, pedestrians and passengers), vehicles
and road. In their study, two predictive modeling methods were used: CART and Random Forests. The
experimental results done using CART analysis to assess the injury risk, scored with respect to the
area under the ROC curve (AUC) a result of 0.8827. While running Random Forests, the authors also found
that the age of the victim, victim occupation, among others, were the attributes with the most predictive power.
Saunier et al. \cite{saunier} investigated the collision factors and processes (i.e. the chain of events
that lead to collisions) through the collection and analysis of microscopic data (road users trajectories).
Saunier et al. \cite{saunier} avoided the use of algorithms with a ``black box'' nature like ANNs (Artificial
Neural Networks) or SVMs (Support Vector Machines) and used C4.5 (Decision Trees) instead and clustering
analysis to investigate the collision factors. In their work, they found an strong relationship between the
evasive actions and the interaction outcome: in most collisions (62 out of 82), no evasive action was
attempted \cite{saunier}.
\section{Proposed Approach}
\label{proposed_approach}
The work described in this paper aims to evaluate different Machine Learning techniques in order to build
a predictive model for injury risk assessment of traffic accident events based on data that was collected
by a traffic managing agency. Since the injury risk has a dichotomous nature in relation to the dependent variable,
this facilitates the use of binary classifiers used in this study like Logistic Regression or Support Vector Machines.
This paper also evaluates the association between the traffic accident injury outcome and the possible
contributory factors, an effort to understand which are the most important factors in an accident with
an injury outcome.
\section{Experimental Methodology}
\label{experimental_methodology}
This section describes the data set used, as well as the tools and algorithms used to perform the analysis.
\subsection{Traffic Accident Data Set}
The data set used in this study was obtained through Datapoa \cite{Datapoa}, an initiative from the city hall
of Porto Alegre to provide open data access to many data sets related with the city itself. The traffic accident
data set available at the Datapoa is licensed under the Open Database License (ODbL) \cite{ODBL}, which is an Attribution
and Share-Alike license for databases.
Although the time span of the available traffic accident data sets ranges from the year 2000 up to 2013, only the data from
the most recent data set was used (the data set related to the accidents that happened in the year of 2013). The data set
is comprised of 20.798 accident records described using 44 attributes. Some attributes of the data set are irrelevant
for the purpose of this study and many attributes also presented duplicated data or invalid records, thus a step of data
cleansing was required before using the data set. The data set also lacks detailed information about vehicles (i.e. age, movement),
drivers (i.e. age of the driver, driver license level, driving experience, sex, etc.) and victims (i.e. age).
\subsection{Tools}
To plot heat maps with the geospatial distribution of the accidents, this study used the framework
Django GIS Brasil \cite{DjangoGISBrasil}, an open source project from the same author of this study,
that aggregates geospatial information related with the Brazilian territory. To provide data analysis,
the author used Pandas \cite{Pandas}, an open source library providing high-performance, data structures
and data analysis tools for the Python programming language. This study also used
scikit-learn \cite{scikit-learn} -- an open source Machine Learning framework for the Python language --
to perform data pre-processing and to build the predictive models.
\subsection{Machine Learning Techniques}
This study employed the following algorithms as classifiers for the injury risk assessment: Logistic Regression,
Support Vector Machines, Naive Bayes, K-nearest neighbors and Random Forests. The details about the use,
parametrization and model evaluation techniques used to assess the predictive models are described
in the next sections.
\subsection{Logistic Regression}
The Logistic Regression used in this study is the Logistic Regression present in the scikit-learn framework \cite{scikit-learn},
which in turn uses the LIBLINEAR \cite{Liblinear} implementation of the Logistic Regression. The LIBLINEAR implementation
solve the following optimization problem:
\begin{equation}
\min_{\substack{w}} \frac{1}{2} w^Tw + C \sum_{i=1}^{l} \xi (w; x_i, y_i)
\end{equation}
Given a set of instance-label pairs $(x_i, y_i), i = 1, \ldots, l$ where $C$ is the penalty parameter and $\xi (w; x_i, y_i)$
is the loss function, which for Logistic Regression is:
\begin{equation}
log(1 + e^{-y_i w^T x_i})
\end{equation}
In this study we used L2 regularized Logistic Regression with the penalty $C$ equal to 1.0.
\subsection{Support Vector Machines}
The Support Vector Machine (SVM) from scikit-learn \cite{scikit-learn} used in this work is based on the LIBSVM \cite{Libsvm} implementation,
which is a $C$-Support Vector Classification. For more details about the algorithm implementation, please refer to the elucidative LIBSVM \cite{Libsvm}
original paper. The SVM algorithm was parametrized with a linear kernel and with 9.0 as the error term, both parameters were chosen using
hyperparameter optimization through a non-exhaustive grid search between different kernel types (RBF, Polynomial and Linear) with different
error term and gamma values. It is also important to note that Support Vector Machine algorithms are not scale invariant, so the the author
applied a scaling function over the attributes before feeding attributes into the algorithm.
\subsection{Naive Bayes}
The Naive Bayes algorithm used in this study is also from scikit-learn \cite{scikit-learn}. The different
Naive Bayes classifiers implement in scikit-learn differ mainly by the assumptions they make regarding the
distribution of $P(x_i \mid y)$ \cite{scikit-learn}. The author decided to use the Gaussian Naive Bayes,
where the likelihood of the features is assumed to be Gaussian:
\begin{equation}
P(x_i \mid y) = \frac{1}{\sqrt{2\pi\sigma^2_y}} \exp\left(-\frac{(x_i - \mu_y)^2}{2\sigma^2_y}\right)
\end{equation}
And where the parameters $\sigma_y$ and $\mu_y$ are estimated using maximum likelihood \cite{scikit-learn}.
\subsection{K-nearest neighbors}
The K-nearest neighbors (kNN) algorithm used in this study is also from scikit-learn \cite{scikit-learn}, which
provides both unsupervised and supervised neighbors-based learning methods. Despite the simplicity of the
algorithm, kNN has been successful in a large number of classification and regression problems. The number
of neighbors used in this study (for the $k$ value) is 8. This value was also found using a non-exhaustive
hyperparameter optimization through the grid search technique. Attribute scaling was also performed before
using kNN, to ensure that the distance measure accords equal weight to each variable.
\subsection{Random Forests}
The author of the paper also used Random Forests from scikit-learn \cite{scikit-learn} as a binary classifier
and also to evaluate the feature importance in order to understand which are the most important factors while
predicting the injury risk. Random forests are a combination of tree predictors, where each tree
depends on the values of a random vector sampled independently and with the same distribution for
all trees in the forest \cite{RandomForest}.
When compared with the original publication from Brainman \cite{RandomForest}, the scikit-learn implementation
combines classifiers by averaging their probabilistic prediction, instead of letting each classifier vote
for a single class \cite{sklearn-randomforest}. Random Forests were also used to assess injury risk and the importance
of factors in the aforementioned study done by Beshah \cite{Beshah}.
The number of estimators (trees in the forest) used was 200, this number was chosen using non-exhaustive
hyperparameter optimization through grid search.
\subsection{Model Evaluation}
In order to evaluate the predictive models trained in this study, the author used a cross-validation with a
train data set with 60\% of the instances from the original data set and with a test data set with 40\%
of the original data set. Both the training and the testing data set were random sampled from the original
data set.
To evaluate the predictive models, the author used the Area Under the Curve (AUC) -- \textit{computed using the
trapezoidal rule} -- of the Receiver Operating Characteristic (ROC), which is a graphical plot that shows the performance of a
binary classifier varying the discrimination threshold. The curve in this study was plotted
using the true positive rate against the false positive rate at various threshold settings (one for
each different predictive outcome from each model).
Also, complementary to the ROC and AUC, the author of this study calculated tables presenting the Precision, Recall
and F1-Score for each class from each predictive model used.
Since some algorithms used in this study didn't have a natural probability estimate outcome like Logistic Regression has
for each class, the author hence used different estimating techniques
in order to be able to compare the ROC and AUC between different classifiers:
\subsubsection{Support Vector Machines} In this case, the probability estimates were calculated by LIBSVM using Platt scaling \cite{Platt}.
\subsubsection{K-nearest neighbors} For kNN, the predicted probability for each class is the ratio of neighbors voting for each label, i.e.
if $k=5$ and 4 neighbors predicted class 1 and only one neighbor predicted class 0, then the probabilities for that example is 0.2 and 0.8.
\subsubsection{Random Forest} The probabilities of a forest are the mean probabilities of the trees in the ensemble and the probabilities
returned by a single tree are the normalized class histograms of the leaf that a sample lands in.
\subsection{Data Cleansing}
Some accident records present in the data set had an invalid date/time format and since the amount of invalid records was
very low (less than 5) when compared with the amount of valid records (greater than 20.790), these records were just
removed from the data set without causing any significant loss in the experimental analysis.
\section{Experimental Analysis}
\label{experimental_analysis}
This section provides a brief descriptive analysis of the data set used as well as the experimental results
using different predictive models together with their model evaluations.
\subsection{Data set analysis}
The data set (after applying data cleansing), is comprised by 20.798 records of traffic accident events that took place in
the city of Porto Alegre/RS (Brazil). The attributes of the data set can be categorized in TODO different types:
\subsubsection{Geospatial Attributes}
These attributes, listed in Table \ref{table:geospatial_attributes} represents where the accident happened in space.
They weren't used in this study and were left for a further study.
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Geospatial Attributes}
\label{table:geospatial_attributes}
\centering
\begin{tabular}{ll}
\hline
Attribute Name & Description \\ \hline
LOG1 and LOG2 & Street names. \\
PREDIAL1 & Street numbers. \\
REGION & Region of the city. \\
LATITUDE and LONGITUDE & The geographical coordinates. \\
LOCAL\_VIA and REGION & Concatenation of LOG1, LOG2 and PREDIAL1.
\end{tabular}
\end{table}
\end{center}
\subsubsection{Irrelevant Attributes}
These attributes are irrelevant to the analysis of factors or injury risk assessment.
They are presented in Table \ref{table:irrelevant_attributes}.
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Irrelevant Attributes}
\label{table:irrelevant_attributes}
\centering
\begin{tabular}{ll}
\hline
Attribute Name & Description \\ \hline
ID & The unique ID of the accident. \\
BOLETIM & The ID of the traffic agency record.
\end{tabular}
\end{table}
\end{center}
\subsubsection{Attributes with data leakage}
Since the main goal of this study is to predict the risk of injury/non-injury, an extra care was taken to discover
attributes that could leak to the target class. The result of this evaluation is present in the Table \ref{table:leak_attributes}.
The attribute ``FONTE'' leaks information about the injury target class because the police is usually involved only when
there was someone injuried. The attribute ``UPS'' also leaks information about the target class because it assumes 3 different
values: 1 (accident only with property damage), 5 (accident with someone injuried) and 13 (accident with deaths), so when the
UPS is 5 or 13 it will perfect predict the injuried/non-injuried target classes.
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Attributes with data leakage}
\label{table:leak_attributes}
\centering
\begin{tabular}{ll}
\hline
Attribute Name & Description \\ \hline
FONTE & Whether the accident was registered by the traffic managing \\
& agency or by the police. \\
UPS & A severity measurement.
\end{tabular}
\end{table}
\end{center}
\subsubsection{Relevant attributes}
These are attributes that were used to train all the predictive models presented in this study. They are shown
in the Table \ref{table:attributes_relevant}. Except the counting attributes, all attributes were preprocessed
using one-hot encoding scheme (aka. one-of-K scheme).
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Relevant Attributes}
\label{table:attributes_relevant}
\centering
\begin{tabular}{ll}
\hline
Attribute Name & Description \\ \hline
LOCAL & Whether the accident happened on a street or in crossing streets. \\
TIPO\_ACID & The type of the accident (collision, fire, etc...). \\
DIA\_SEM & The day of the week. \\
CONSORCIO & If a bus were involved, the name of the company. \\
AUTO & The count of cars involved. \\
TAXI & The count of cabs involved. \\
LOTACAO & The count of small bus involved. \\
ONIBUS\_URB & The count of urban bus involved. \\
ONIBUS\_MET & The count of bus (others) involved. \\
CAMINHAO & The count of trucks involved. \\
MOTO & The count of motorcycles involved. \\
CARROCA & The count of carts involved. \\
BICICLETA & The count of bikes involved. \\
OUTRO & The count of vehicles (others) involved. \\
TEMPO & How was the weather (raining, clear, etc.). \\
NOITE\_DIA & If it was night or day. \\
MES & The month of the accident. \\
FX\_HORA & The hour that accident happened. \\
CORREDOR & Whether the accident happened in the bus lane road or not.
\end{tabular}
\end{table}
\end{center}
\subsubsection{Target attribute}
Since the aim of this work is to predict if the outcome of an traffic accident was an injury/non-injury,
the author merged (summed) the features shown in the Table \ref{table:target_attributes} and then
created a new attribute with this value that was later converted to 0 (non-injury) if sum was less or equal than
zero, or 1 (injury) if the sum was greater or equal to 1.
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Attributes merged to create the target attribute}
\label{table:target_attributes}
\centering
\begin{tabular}{ll}
\hline
Attribute Name & Description \\ \hline
FERIDOS & The count of injured people involved in the accident. \\
FERIDOS\_GR & The count of serious injured people involved in the accident. \\
MORTES & The count of deaths (local deaths) in the accident. \\
MORTES\_POST & The count of deaths (posterior deaths) happened after the accident. \\
FATAIS & The sum of MORTES and MORTES\_POST attributes.
\end{tabular}
\end{table}
\end{center}
It is also important to note that the data set is imbalanced and it has a ratio of records of at least 2:1
between the target classes (injury/non-injury), totaling 14.247 non-injury instances and 6.551 injury
records.
The geospatial information related to the accident events weren't used in this study, but the heat map shown in Figure
\ref{figure:heatmap} shows an important pattern that clearly confirms that the accidents density increases
on crossing streets. This information is represented not only in latitude/longitude attributes but also in
the ``LOCAL'' attribute used to train the predictive models in this study.
\begin{figure}[!t]
\centering
\includegraphics[width=3.5in]{heatmap.png}
\caption{Heat map of the traffic accidents.}
\label{figure:heatmap}
\end{figure}
\subsection{Model evaluation}
The model evaluation results (Precision, Recall, F1-score) using test data set for Logistic Regression and Support
Vector Machine (SVM) are shown in the Table \ref{table:svm_log}, the results for Naive Bayes
and K-nearest neighbors are shown in the Table \ref{table:bayes_knn} and the results for the
Random Forest model evaluation is shown in the Table \ref{table:random_forest}. The ROC curves and the AUC
value for each model for the positive class (injury) prediction with varying discriminative threshold is presented
in the Figure \ref{figure:roc}.
As we can see, Logistic Regression and Support Vector Machines models provided the best scores
in AUC and average Precision/Recall/F1-score, they also were very similar regarding the performance. Random Forest also performed
well with an AUC of 0.93 when compared with AUC of SVM and Logistic Regression that scored an AUC of 0.94 each.
K-nearest neighbor performed below the scores of SVM, Logistic Regression and Random Forest with an AUC of 0.90, but
it still performed better than the worst model which is Naive Bayes with an AUC of 0.83.
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.2}
\caption{Support Vector Machine and Logistic Regression evaluation}
\label{table:svm_log}
\centering
\begin{tabular}{llll|lll}
\hline
\multicolumn{4}{c}{Support Vector Machine} & \multicolumn{3}{c}{Logistic Regression} \\ \hline
& Precision & Recall & F1-score & Precision & Recall & F1-Score \\ \hline
Non-injury & 0.90 & 0.96 & 0.93 & 0.90 & 0.96 & 0.93 \\
Injury & 0.89 & 0.76 & 0.82 & 0.89 & 0.76 & 0.82 \\ \hline
Average & 0.90 & 0.90 & 0.89 & 0.89 & 0.90 & 0.89
\end{tabular}
\end{table}
\end{center}
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.2}
\caption{Naive Bayes and K-nearest neighbors evaluation}
\label{table:bayes_knn}
\centering
\begin{tabular}{llll|lll}
\hline
\multicolumn{4}{c}{Naive Bayes} & \multicolumn{3}{c}{K-nearest neighbors evaluation} \\ \hline
& Precision & Recall & F1-score & Precision & Recall & F1-Score \\ \hline
Non-injury & 0.96 & 0.23 & 0.38 & 0.85 & 0.96 & 0.90 \\
Injury & 0.37 & 0.98 & 0.54 & 0.88 & 0.63 & 0.73 \\ \hline
Average & 0.78 & 0.47 & 0.43 & 0.86 & 0.86 & 0.85
\end{tabular}
\end{table}
\end{center}
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.2}
\caption{Random Forest evaluation}
\label{table:random_forest}
\centering
\begin{tabular}{llll}
\hline
\multicolumn{4}{c}{Random Forest} \\ \hline
& Precision & Recall & F1-score \\ \hline
Non-injury & 0.90 & 0.94 & 0.92 \\
Injury & 0.85 & 0.76 & 0.80 \\ \hline
Average & 0.88 & 0.88 & 0.88
\end{tabular}
\end{table}
\end{center}
\begin{figure}[!t]
\centering
\includegraphics[width=3.6in]{roc.pdf}
\caption{The Receiver Operating Characteristic (ROC) and Area Under Curve (AUC) of each model for the positive class (injury).}
\label{figure:roc}
\end{figure}
\subsection{Variable Importance}
According to Strobl et al. \cite{random_bias}, Random Forests have been successfully applied to various problems and within a very
short period of time, random forests have become a major data analysis tool, that performs well in comparison with many standard methods.
One of the factors that greatly contributed to the popularity of Random Forests was that it produces variable
importance measures for each predictor variable.
The experimental results to evaluate the attributes that had more importance while predicting the injury risk were obtained using
the same trained model with the evaluation presented in the Table \ref{table:random_forest}. The first 10 most important attributes that
were described by this model, with their respective importances, are presented in the Table \ref{table:attribute_importance}.
while predicting the injury risk, followed by other attributes, like cars involved and if accident was a run over, among others.
As we can see in the Table \ref{table:attribute_importance}, the motorcycle count attribute had the largest importance
\begin{center}
\begin{table}[h]
\renewcommand{\arraystretch}{1.3}
\caption{Attribute Importances}
\label{table:attribute_importance}
\centering
\begin{tabular}{llll}
Importance & Attribute Name & Description \\ \hline
\hline
0.2108 & MOTO & The count of motorcycles involved. \\
0.0948 & AUTO & The count of cars involved. \\
0.0925 & TIPO\_ACID \\
& ATROPELAMENTO & If the type of the accident was a run over. \\
0.0391 & LOCAL \\
& LOGRADOURO & If the accident was on a normal street. \\
0.0368 & LOCAL \\
& CRUZAMENTO & If the accident was on crossing streets. \\
0.0267 & TIPO\_ACID \\
& COLISÃO & If the type of the accident was a collision. \\
0.0205 & TIPO\_ACID \\
& QUEDA & If the type of the accident was a fall. \\
0.0203 & CAMINHAO & The count of trucks involved. \\
0.0182 & TIPO\_ACID \\
& ABALROAMENTO & If the type of the accident \\
& & was a collision (on the side). \\
0.0181 & NOITE\_DIA \\
& DIA & If the accident happened during night time. \\
\end{tabular}
\end{table}
\end{center}
\section{Conclusions and Future Work}
\label{conclusion}
As we can see, the experimental results demonstrated that prediction models for
injury risk assessment can be created with
good precision, even with limited data sets, like the one used in this study that lacks
information about vehicle drivers, victims and vehicle movements.
These results, together with the variable importance analysis, can be used by traffic
managing agencies to understand the provided data sets with an even greater depth than the limited
descriptive analysis that is being carried today by these agencies.
This study didn't used the geospatial data, but the author believes that this information is also a
critical factor to the prediction of the injury risk associated with an traffic accident. The use of
the geospatial data was left to a future study due to the very specific nature of the geospatial
data format, which requires different preprocessing approach before being employed.
Future works can also include better hyperparameter optimization with a more intensive search
for better parameters such as kNN neighbor size, SVM error term, SVM kernel parameters, Random Forest
estimators count, among others.
This study also didn't applied feature selection techniques, but the author believes that a
future work could also improve the models performance by using feature selection methods.
|
\section{Introduction}
To mitigate the global climate change
it is commonly agreed that
greenhouse gas emissions, and in particular emissions of CO$_2$,
have to be reduced substantially
\cite
Foster-et-al-2010,IPCC-2007,IPCC-2013,IPCC-2014,Rockstroem-et-al-2009,Rockstroem-et-al-2009-Nature}.
Since 85\% of current primary energy driving global economies
are due to the combustion of fossil fuels,
and since consumption of fossil fuels accounts for 56.6\% of all anthropogenic
greenhouse gas emissions, introducing renewable energy sources
to support all areas of human life plays an essential role
in fighting global warming \cite{Edenhofer-et-al-2012}.
In particular, the generation of electricity by renewables will be
an important step towards this goal, requiring substantial changes
to current grid structures and power plant systems.
If generation and distribution of electricity is to be organized by market principles,
a preeminent challenge of a future electricity market mechanism design is to set effective price signals
to reward the introduction and the use of renewable energy sources
for the generation of electricity,
and simultaneously to penalize fossil fuel power plants.
However, the physical requirements of electricity grids
and the necessities of public life in present societies
impose special restrictions on
electricity markets.
In particular, a necessary condition for
grid stability is the reliability of electricity generation
and the immediate equality of supply and demand at any instant of time.
It is expected that the biggest contribution of renewable energy sources in electricity grids
will come from wind turbines and photovoltaic cells \cite{Agora-2013},
both producing electricity only with high volatility.
Their widespread installation therefore would challenge the reliability of electricity
supply and thus the stability of the grids.
Lacking sufficiently large storages for electricity,
to warrant reliability in grids with volatile energy sources
power plants with high operational flexibility are required
as a power reserve standing in
in cases of sudden scarcity of electricity supply or of blackouts.
\begin{figure*}[htp]
\centering
\begin{footnotesize}
\begin{tikzpicture}[>=stealth]
\draw[<->]
(0ex,20ex) -- node[near start,rotate=90,anchor=south west] {price}
(0ex, 0ex) -- node[very near end,anchor=north] {quantity}
(26ex,0ex);
\draw[dashed,red,thick] (-.5ex,7.75ex) -- (12.85ex,7.75ex) -- (12.85ex,-.5ex);
\draw[->,gray] (9.2ex,12.0ex) .. controls (10.2ex,10.8ex) .. (12.0ex,9.1ex);
\draw[->,blue] (15.2ex,12.0ex) .. controls (14.58ex,10.8ex) .. (13.3ex,9.1ex);
\draw[->,gray] (16.2ex,4.5ex) .. controls (14.6ex,5.6ex) .. (13.2ex,6.8ex);
\draw[->,blue] (10.2ex,4.5ex) .. controls (11.5ex,5.6ex) .. (12.6ex,6.8ex);
\node[anchor=east,red] at (0ex,7.25ex) {$p_*$};
\node[anchor=north,red] at (12.9ex,0ex) {$q_*$};
\draw[thick] (5ex,19ex) .. controls (8ex,9ex) and (19ex,3ex) .. node[very near start,anchor=east] {$D$} (24ex,1ex);
\draw[thick,blue] (18ex,19ex) .. controls (16ex,9ex) and (10ex,4ex) .. node[very near start,anchor=west] {$S$} (4ex,1ex);
\end{tikzpicture}
\hspace*{10ex}
\begin{tikzpicture}[>=stealth]
\draw[<->]
(0ex,20ex) -- node[near start,rotate=90,anchor=south west] {price}
(0ex,0ex) --
(36ex,0ex);
\node[anchor=north,blue] at (22.5ex,0ex) {$P_{\max}$};
\node[anchor=north east] at (36ex,0ex) {power};
\draw[thick, blue]
(0ex,3ex) -- (10ex,3ex) -- (10ex,7ex) -- (17ex,7ex) -- (17ex, 9ex) -- (22ex,9ex) -- (22ex,19ex);
\draw[dashed,blue,thick] (22ex,0ex) -- (22ex,20ex);
\draw[thick] (25.5ex,19ex) .. controls (26ex,9ex) and (27ex,4ex) .. (28ex,1ex);
\node[rotate=90, anchor=east] at (23.6ex,20ex) {\em blackout};
\end{tikzpicture}
\end{footnotesize}
\caption{\label{fig-blackout}\footnotesize
A perfect efficient market where any demand $D$ meets supply $S$
at a certain equilibrium price $p_*$ and quantity $q_*$ (left hand).
On an electricity spot market, a
blackout is a market failure due to inelastic demand and supply,
with the supply curve given by the merit order of the power plant
system (right hand).
Here a (“rolling”) blackout occurs if the demand is higher
than the total maximum power
$P_{\max}$ of all power plants \cite{Cramton-et-al-2013}.
Increasing the demand-side inelasticity, e.g., by smart grids,
could remove the problem on the long run, but in the short run
electricity markets require capacity reserves which are not demanded
for most of the time.
}
\end{figure*}
Cramton and Ockenfels \cite{Cramton-Ockenfels-2012}
proved the “missing money” theorem stating that,
in a competitive electricity market, prices are always too low to
pay for adequate capacity.
In fact, present electricity markets are not perfect efficient markets
since both supply and demand are price inelastic, see
Figure~\ref{fig-blackout}.
Future increase of demand elasticity, for instance by smart grids,
would relax the difficulties to a certain degree,
but inelasticity on the supply-side could only be removed by
capacity reserves or huge electricity storages.
The first option, however, requires long-term plannings at a magnitude
of decades, whereas the second option is technologically not
realizable to date. For more details see \cite{Cramton-et-al-2013}.
Besides these theoretical arguments there also exist empirical
clues to doubt that current electricity
markets encourage investments in operationally flexible power plants or
in the provision
of power reserves for cases of emergency or maintenance
\cite{Agora-2013-capacity-market
Ecke-et-al-2013-en,EWI-2012,Steuwer-2013}.
Several solutions to this problem have been
proposed recently to complement the present “energy-only” markets,
ranging from
separate capacity markets which trade backup capacity,
to strategic capacity reserves usually settled by long-term contracts
with national agencies
\cite{Agora-2013-capacity-market,Cramton-Ockenfels-2012,EWI-2012,Gawel-Purkus-2013,Growatsch-et-al-2013,Matthes-et-al-2012,Schoenberg-2015,Steuwer-2013}.
The main goal of this paper is to propose a solution for the economic
problem to finance the necessary capacity reserves guaranteeing grid
stability by market principles.
One necessary property of a power plant for being part of a capacity reserve is a fast guaranteed operational flexibility.
In our opinion the main problem of current market mechanism designs
is the fact that market prices do not regard operational flexibility,
being determined solely by the marginal costs of
electricity generation. Thus the costs of operational inflexibility
are market externalities
\cite[§14]{Bofinger-2007}, \cite[p 125]{Krugman-Wells-2006}
and reduce welfare.
By contrast, a sustainable electricity market mechanism design should induce market prices
which take into account both the direct variable production costs and the
external ecological costs of
electricity production, but also the costs caused by operational inflexibility of
each individual power plant.
Due to emissions trading \cite[§15.4]{Stroebele-et-al-2012},
the first two cost factors are
already priced in as marginal costs
on present electricity spot markets,
but operational flexibility does not play a role for the
determination of the spot market prices to date.
To internalize it into price calculation, at first we define
measure for the operational flexibility of a given power plant.
This measure then can be used
to compute an inflexibility fee for each power plant.
The total of these inflexibility fees then can serve to
pay power reserves provided by some given capacity mechanism.
This paper is organized as follows. In section
\ref{sec-definition-operational-flexibility}
a general class of functions measuring the operational flexibility
of a power pant
in dependence to its guaranteed start-up time is defined.
In section \ref{sec-inflexibility-fee} the effect of the inflexibility fee
on the offer price of a power plant on an electricity spot market
is calculated
and demonstrated by a prototypical exemplary “toy” grid in Example
\ref{beisp-toy-grid}.
A way how the accumulated inflexibility fees then can be used to finance
a capacity mechanism
is described in section \ref{sec-capacity-reserve},
before a short discussion concludes the paper.
\section{A measure of the operational flexibility of a power plant}
\label{sec-definition-operational-flexibility}%
We stipulate that the operational flexibility of a power plant
depends on its
\emph{guaranteed start-up time}\index{guaranteed start-up time}
$t_s \in [0,\infty)$
which is defined as the time that a power plant requires to supply a guaranteed
power of electricity.
Moreover, we claim that the measure should be a pure number expressing a degree
of flexibility ranging from $0$ to $1$,
with the property that the longer the guaranteed start-up time
the smaller the value of flexibility.
Consequently, we define a general
\emph{measure of operational flexibility}%
\index{measure of operational flexibility}\index{operational flexibility}\index{flexibility}
to be a strictly monotonically decreasing function $\varphi: [0,\infty) \to [0,1]$
of a single variable satisfying the limit behavior
\begin{equation}
\varphi(x) \to 1 \quad \mbox{as \ $x \to 0$,}
\quad \mbox{and} \quad
\varphi(x) \to 0 \quad \mbox{as \ $x \to \infty$}
.
\label{eq-operational-flexibility-norm-condition}
\end{equation}
Here the variable $x$ represents the starting time
of the power plant, measured in hours [h].
A simple example of such a measure is the differentiable function
\begin{equation}
\varphi(x)
= \frac{1}{x+1}
.
\quad
\begin{tabular}{c}
\begin{tikzpicture}[>=stealth, domain=0:3, yscale=1.2]
\draw[->,darkgray] (-0.2,0) -- (3.2,0) node[below] {$x$};
\draw[->,darkgray] (0,-0.2) -- (0,1.2) node[left] {$\varphi(x)$};
\draw
plot (\x,{1/(\x + 1)});
\end{tikzpicture}
\end{tabular}
\label{eq-operational-flexibility}
\end{equation}
In the sequel we will use this function to measure the operational flexibility of
a given power plant.
In Table \ref{tab-operational-flexibility-measures} there are listed
guaranteed start-up times $t_s$ and the respective flexibility measures
$\varphi(t_s)$ for some typical power plants.
\begin{table}[htp]
\centering
\begin{footnotesize}
\begin{tabular}{|l|c|c||c|}
\hline
& \textbf{Guaranteed} & & \textbf{Marginal} \\
\textbf{Power Plant} & \textbf{Start-Up} & $\mathbf{\varphi}$ &
\textbf{Costs $\fett{p}^{\mathbf{mc}}$}
\\
& \textbf{Time [h]} & & \textbf{[\euro/MWh]}
\\ \hline
wind turbine & $\infty$ & .000 & --- \\
hydroelectric power station & 0.02 & .979 & --- \\
gas turbine & 0.12 & .893 & 90 \\
cogeneration plant (CHP) & 0.17 & .855 & --- \\
combined cycle gas turbine & 5 & .167 & 50 \\
hard coal power plant & 6 & .143 & 60 \\
lignite power plant & 9 & .100 & 40 \\
nuclear power plant & 50 & .020 & \,\, 5 \\
\hline
\end{tabular}
\end{footnotesize}
\caption{\label{tab-operational-flexibility-measures}\footnotesize
Exemplary cold start-up times and their respective operational flexibility measures,
as well as exemplary marginal costs (without emissions trading).
The values are typical for current German electricity markets.
Data from \cite[p 71]{Grimm-2007-Diss} (start-up times)
and
\cite[pp 13, 19]{Heitmann-2010},
\cite[p 3]{von-Roon-Huck-2010} (marginal costs).
}
\end{table}
Note that a wind turbine is assigned a vanishing operational flexibility,
since due to the volatility of winds
a predetermined amount of energy by a wind turbine cannot be guaranteed
at a given future instant.
The highest operational flexibilities are exposed by hydroelectric power stations
and modern gas turbines.
\section{Fees on operational inflexibility}
\label{sec-inflexibility-fee}%
On an wholesale electricity market,
each participating power plant operator offers electric power with a sell bid
for each of its power plant.
The market maker collects all these sell bids and determines the market-clearing price
in accordance to the buy bids and the merit order
\cite{Das-et-al-2001,EPEX-2014,Wilson-1985,Satterthwaite-Williams-1989},
for a theoretical introduction see also \cite[§6.5, §7.4.5]{Fudenberg-Tirole-1991}.
Our main idea now is to rise a fee for operational inflexibility on each
power plant, its amount being calculated by the operational flexibility $\varphi$
as part of a factor to a given market-wide reference level.
In consequence, the offer price of each power plant must
take it into account its operational flexibility.
To be more precise, let
$p_i^{\mathrm{mc}}$ denote the marginal offer price per energy quantity of the power plant
regarding only the marginal costs,
including the variable costs of production and emissions trade certificates;
this is the price which would be offered for the power plant on a current wholesale spot market
\cite{von-Roon-Huck-2010}.
Assume moreover that all power plants participating at the spot market are uniquely numbered
by the indices $i$ $=$ $1$, $2$, \ldots, $n$.
The spot market offer price $p_i$ of plant $i$ taking into account its operational flexibility
$\varphi_i$ then is calculated by the formula
\begin{equation}
\fbox{$
p_i
= p_i^{\mathrm{mc}} + (1 - \varphi_i) \ p_0
.
$}
\label{eq-price-regarding-operational-flexibility}
\end{equation}
Here $p_0$ denotes a market-wide constant reference level price, set by the market
authority.
It therefore is a political or regulatory quantity,
not a market-inherent value or immediately economically deducible.
It is arbitrary in principle, but the higher
its amount the heavier the effect of operational flexibility on the final
spot market price.
\begin{table*}[htp]
\centering
\begin{footnotesize}
\begin{tabular}{|l|c|r||r|r|r|}
\hline
& & & \textbf{Marginal Price} & \multicolumn{2}{c|}{\textbf{Offer Price}} \\
\textbf{Power Plant}
& $\mathbf{\varphi}$
& $\mathbf{1 - \varphi}$
& \multicolumn{1}{c|}{$\fett{p}^{\mathbf{mc}}_i$}
& \multicolumn{2}{c|}{$\fett{p}_i$ \textbf{[\euro/MWh]}}
\\ \cline{5-6}
& &
& \multicolumn{1}{c|}{\textbf{[\euro/MWh]}}
& \multicolumn{1}{c|}{$\fett{p}_0$ = \textbf{10}}
& \multicolumn{1}{c|}{$\fett{p}_0$ = \textbf{70}}
\\ \hline
wind turbine
& .000 & $1.000$ & $1$ \qquad \ & $11$ \hspace*{1.2ex} & $71$ \hspace*{1.2ex}
\\
hydroelectric power station
& .980 & $.020$ & $1$ \qquad \ & $1$ \hspace*{1.2ex} & $2$ \hspace*{1.2ex}
\\
gas turbine
& .893 & $.107$ & $90$ \qquad \ & $91$ \hspace*{1.2ex} & $98$ \hspace*{1.2ex}
\\
cogeneration plant (CHP)
& .855 & $.145$ & $50$ \qquad \ & $51$ \hspace*{1.2ex} & $60$ \hspace*{1.2ex}
\\
combined cycle gas turbine
& .167 & $.833$ & $50$ \qquad \ & $58$ \hspace*{1.2ex} & $108$ \hspace*{1.2ex}
\\
hard coal power plant
& .143 & $.857$ & $60$ \qquad \ & $69$ \hspace*{1.2ex} & $120$ \hspace*{1.2ex}
\\
lignite power plant
& .100 & $.900$ & $40$ \qquad \ & $49$ \hspace*{1.2ex} & $103$ \hspace*{1.2ex}
\\
nuclear power plant
& .020 & $.980$ & $5$ \qquad \ & $15$ \hspace*{1.2ex} & $74$ \hspace*{1.2ex}
\\
\hline
\end{tabular}
\end{footnotesize}
\caption{\label{tab-prices-regarding-operational-flexibility}\footnotesize
Operational flexibilities, as given by the exemplary data
of Table \ref{tab-operational-flexibility-measures},
and the resulting offer price differences with respect to the reference level prices
$p_0 = 10$ \euro$/$MWh and $p_0 = 70$ \euro$/$MWh.
}
\end{table*}
It should be high enough to signal effective incentives to introduce and use
operationally flexible power plants for scarcity situations and black-outs,
but it must be low enough to avoid a too radical change of the merit order
such that too many flexible power plants are operational on the spot market
and thus unavailable for a capacity reserve
(see Figure \ref{fig-operational-flexibility-example}).
\begin{beispiel}
\label{beisp-toy-grid}%
Consider a small examplary grid
(called “toy grid” in the sequel)
consisting of the eight
power plants listed in
Table \ref{tab-operational-flexibility-measures}.
The prices resulting from the respective inflexibility fees
in dependence to different reference level prices
$p_0$ are listed in
Table \ref{tab-prices-regarding-operational-flexibility}.
If the reference level price is low (here $p_0$ $=$ $10$ \euro$/$MWh),
the modified offer prices do not change the merit order of the power plant system,
whereas a sufficiently high reference level price
(e.g., $p_0 = 70$ \euro$/$MWh)
changes it,
as is depicted in Figure \ref{fig-operational-flexibility-example}.
\begin{figure*}[htp]
\centering
\begin{footnotesize}
\begin{tikzpicture}[>=stealth,baseline]
\draw[fill=cyan] ( 0ex,40ex) rectangle node[rotate=90,anchor=west] {hydro}
( 5ex,40.25ex);
\draw[fill=green] ( 5ex,40ex) rectangle node[rotate=90,anchor=west] {wind}
(10ex,40.25ex);
\draw[fill=gold] (10ex,40ex) rectangle node[rotate=90,anchor=west] {nuclear}
(15ex,41.25ex);
\draw[fill=brown] (15ex,40ex) rectangle node[rotate=90] {lignite} (20ex,50.00ex);
\draw[fill=lightgray] (20ex,40ex) rectangle node[rotate=90] {CHP} (25ex,52.50ex);
\draw[gray,fill=cyan] (25ex,40ex) rectangle
(30ex,52.50ex);
\draw node[rotate=90] at (27.5ex,50ex) {combined cycle gas};
\draw[fill=lightgray] (30ex,40ex) rectangle node[rotate=90] {hard coal} (35ex,55.00ex);
\draw[fill=teal!50] (35ex,40ex) rectangle node[rotate=90] {gas turbine} (40ex,62.50ex);
\draw[dashed,red,thick] (-.5ex,52.50ex) node[anchor=south west] {$p_*$}
-- (23.80ex,52.50ex) -- (23.80ex,39.5ex) node[anchor=north] {$q_*$};
\draw[<->]
(0ex,64ex) -- node[very near start,rotate=0,anchor=south west] {\euro$/$MWh}
(0ex,40ex) -- node[very near end,anchor=north west] {MW}
(45ex,40ex);
\draw[<->]
(60ex,64ex) -- node[very near start,rotate=0,anchor=south west] {\euro$/$MWh}
(60ex,40ex) -- node[very near end,anchor=north west] {MW}
(105ex,40ex);
\draw[dashed,opacity=.5,fill=cyan] (60ex,40ex) rectangle (65ex,40.25ex);
\draw[dashed,opacity=.5,fill=green] (65ex,40ex) rectangle (70ex,40.25ex);
\draw[dashed,opacity=.5,fill=gold] (70ex,40ex) rectangle (75ex,41.25ex);
\draw[dashed,opacity=.5,fill=brown] (75ex,40ex) rectangle (80ex,50.00ex);
\draw[dashed,opacity=.5,fill=lightgray] (80ex,40ex) rectangle (85ex,52.50ex);
\draw[dashed,opacity=.5,fill=cyan] (85ex,40ex) rectangle (90ex,52.50ex);
\draw[dashed,opacity=.5,fill=lightgray] (90ex,40ex) rectangle (95ex,55.00ex);
\draw[dashed,opacity=.5,fill=teal] (95ex,40ex) rectangle (100ex,62.50ex);
\draw (60ex,40ex) rectangle node[rotate=90,anchor=west] {hydro} (65ex,40.30ex);
\draw (65ex,40ex) rectangle node[rotate=90,anchor=west] {\ wind} (70ex,42.75ex);
\draw (70ex,40ex) rectangle node[rotate=90,anchor=west] {\ \ nuclear} (75ex,43.70ex);
\draw (75ex,40ex) rectangle node[rotate=90] {lignite} (80ex,52.25ex);
\draw (80ex,40ex) rectangle node[rotate=90] {CHP} (85ex,52.86ex);
\draw (85ex,40ex) rectangle
(90ex,54.58ex);
\draw node[rotate=90] at (87.5ex,50ex) {combined cycle gas};
\draw (90ex,40ex) rectangle node[rotate=90] {hard coal} (95ex,57.14ex);
\draw (95ex,40ex) rectangle node[rotate=90] {gas turbine} (100ex,62.77ex);
\draw[dashed,red,thick] (59.5ex,52.86ex) node[anchor=south west] {$p_*$}
-- (83.80ex,52.86ex) -- (83.80ex,39.5ex) node[anchor=north] {$q_*$};
\draw[->] (45.0ex,52.0ex)
-- node[anchor=south] {$p_0 = 10$ \euro/MWh} (55.0ex,52.0ex);
\draw[fill=cyan] ( 0ex, 0ex) rectangle node[rotate=90,anchor=west] {hydro}
( 5ex, 0.25ex);
\draw[fill=green] ( 5ex, 0ex) rectangle node[rotate=90,anchor=west] {wind}
(10ex, 0.25ex);
\draw[fill=gold] (10ex, 0ex) rectangle node[rotate=90,anchor=west] {nuclear}
(15ex, 1.25ex);
\draw[fill=brown] (15ex, 0ex) rectangle node[rotate=90] {lignite} (20ex,10.00ex);
\draw[fill=lightgray] (20ex, 0ex) rectangle node[rotate=90] {CHP} (25ex,12.50ex);
\draw[gray,fill=cyan] (25ex, 0ex) rectangle
(30ex,12.50ex);
\draw node[rotate=90] at (27.5ex,10ex) {combined cycle gas};
\draw[fill=lightgray] (30ex, 0ex) rectangle node[rotate=90] {hard coal} (35ex,15.00ex);
\draw[fill=teal!50] (35ex, 0ex) rectangle node[rotate=90] {gas turbine} (40ex,22.50ex);
\draw[dashed,red,thick] (-.5ex,12.50ex) node[anchor=south west] {$p_*$}
-- (23.80ex,12.50ex) -- (23.80ex,-.5ex) node[anchor=north] {$q_*$};
\draw[<->]
(0ex,33ex) -- node[very near start,rotate=0,anchor=south west] {\euro$/$MWh}
(0ex, 0ex) -- node[very near end,anchor=north west] {MW}
(45ex, 0ex);
\draw[<->]
(60ex,33ex) -- node[very near start,rotate=0,anchor=south west] {\euro$/$MWh}
(60ex, 0ex) -- node[very near end,anchor=north west] {MW}
(105ex,0ex);
\draw[dashed,opacity=.5,fill=cyan] (60ex,0ex) rectangle (65ex, 0.25ex);
\draw[dashed,opacity=.5,fill=lightgray] (65ex,0ex) rectangle (70ex,12.50ex);
\draw[dashed,opacity=.5,fill=green] (70ex,0ex) rectangle (75ex, 0.25ex);
\draw[dashed,opacity=.5,fill=gold] (75ex,0ex) rectangle (80ex, 1.25ex);
\draw[dashed,opacity=.5,fill=teal] (80ex,0ex) rectangle (85ex,23.78ex);
\draw[dashed,opacity=.5,fill=brown] (85ex,0ex) rectangle (90ex,10.00ex);
\draw[dashed,opacity=.5,fill=cyan] (90ex,0ex) rectangle (95ex,12.50ex);
\draw[dashed,opacity=.5,fill=lightgray] (95ex,0ex) rectangle (100ex,15.00ex);
\draw (60ex,0ex) rectangle node[rotate=90,anchor=west] {hydro} (65ex, 0.59ex);
\draw (65ex,0ex) rectangle node[rotate=90] {CHP} (70ex,15.04ex);
\draw (70ex,0ex) rectangle node[rotate=90] {wind} (75ex,17.75ex);
\draw (75ex,0ex) rectangle node[rotate=90] {nuclear} (80ex,18.06ex);
\draw (80ex,0ex) rectangle node[rotate=90] {gas turbine} (85ex,24.38ex);
\draw (85ex,0ex) rectangle node[rotate=90] {\ lignite} (90ex,25.75ex);
\draw (90ex,0ex) rectangle node[rotate=90] {combined cycle gas} (95ex,27.08ex);
\draw (95ex,0ex) rectangle node[rotate=90] {hard coal} (100ex,30.00ex);
\draw[dashed,red,thick] (59.5ex,24.38ex) node[anchor=south west] {$p_*$}
-- (83.80ex,24.38ex) -- (83.80ex,-.5ex) node[anchor=north] {$q_*$};
\draw[->] (45.0ex,16.5ex)
-- node[anchor=south] {$p_0 = 70$ \euro/MWh} (55.0ex,16.5ex);
\end{tikzpicture}
\end{footnotesize}
\caption{\label{fig-operational-flexibility-example}\footnotesize
Effect of the operational inflexibility fee on the
price $p_*$ clearing the market consisting of the power plant system
in Table \ref{tab-prices-regarding-operational-flexibility},
neglecting operational flexibility (left) and regarding it (right).
The reference level price are assumed as $p_0 = 10$ \euro/MWh
and $p_0 = 70$ \euro/MWh, respectively.
For a given demand $q_*$ of electric power, the market-clearing spot price
increases more or less slightly, depending on $p_0$.
For a high operational inflexibility fee, as in the second case,
the merit order is changed.
}
\end{figure*}
In our toy grid we can recognize that,
if the amount of $p_0$ is too high,
the effect may be even counterproductive since the flexible gas turbine
is in the money and thus operating at a normal quantity demand,
leaving no power plant as a capacity reserve.
In case of a sudden scarcity or of a blackout,
the grid then would perform worse
than with the original merit order.
Moreover we observe that the higher the reference level price $p_0$,
the higher the spot market price.
The amounts, however, are not related to each other in a linear manner,
but depend discontinously on the changes of the merit order.
The total amount of inflexibility fees, at last, is directly calculated to
be either 48.4 \euro$/$MWh in case of $p_0$ $=$ 10 \euro$/$MWh,
or 339 \euro$/$MWh in case of $p_0$ $=$ 70 \euro$/$MWh.
We finally note that for the demanded quantity $q_*$ depicted in the scenarios
in Figure \ref{fig-operational-flexibility-example},
only five power plants are operational. Depending on the reference level price
the realized profit then is given by the following tables.
\noindent
\begin{footnotesize}
\[
\begin{tabular}{|l|c|}
\hline
\multicolumn{2}{|c|}{\textbf{$\mathbf{\fett{p}_0 = 10}$ \euro/MWh}}
\\ \hline
\textbf{Power plant} & \textbf{Profit for} $\mathbf{\fett{q}_*}$
\\ \hline
wind turbine & 40 \euro$/$MWh \\
hydroelectric & 50 \euro$/$MWh \\
CHP & \ \,0 \euro$/$MWh \\
lignite & \ \,2 \euro$/$MWh \\
nuclear & 37 \euro$/$MWh \\
\hline
\end{tabular}
\quad
\begin{tabular}{|l|c|}
\hline
\multicolumn{2}{|c|}{\textbf{$\mathbf{\fett{p}_0 = 70}$ \euro/MWh}}
\\ \hline
\textbf{Power plant} & \textbf{Profit for} $\mathbf{\fett{q}_*}$
\\ \hline
wind turbine & 27 \euro$/$MWh \\
hydroelectric & 95 \euro$/$MWh \\
CHP & 37 \euro$/$MWh \\
gas turbine & \ \,0 \euro$/$MWh \\
nuclear & 24 \euro$/$MWh \\
\hline
\end{tabular}
\label{eq-toy-grid-profits}
\]
\end{footnotesize}
Assume for simplicity that the demand remains constantly at $q_*$ during a certain
hour and that all power plants yield the same power of $5$ MW, say,
and let be $q_* = 25$ MW be the demanded electrical power for the hour considered
(such that the consumed electricity energy during this period is $E = 25$ MWh).
Then with Table \ref{tab-prices-regarding-operational-flexibility}
the total of the inflexibility fees for the five operational power plants
in the money
has the amount of
\begin{equation}
C_f
= (10 + 1 + 1 + 9 + 10) \cdot 5
= 205 \mbox{ \euro$/$h,}
\end{equation}
at a reference level price $p_0 = 10$ \euro$/$MWh,
and
\begin{equation}
C_f
= (70 + 1 + 8 + 10 + 69) \cdot 5
= 790 \mbox{ \euro$/$h,}
\end{equation}
at a reference level price $p_0 = 70$ \euro$/$MWh.
The total fee then can be distributed to the power plants participating
at a capacity mechanism, paying their time of reliability.
\end{beispiel}
The toy grid in Example \ref{beisp-toy-grid} demonstrates the possible
direct consequences of the inflexibility fee to the wholesale electricity market.
In essence, by Equation (\ref{eq-price-regarding-operational-flexibility})
a power plant with a low operational flexibility
is penalized more than one with a high operational flexibility.
In the limit case that all power plants participating on the spot market
are equally operationally flexible, i.e., $\varphi_i$ $=$ const,
all sell bids are raised by the same amount and the merit order cannot change.
On the other hand, if the power plants have different operational flexibilities
and the reference price level $p_0$ is chosen too high,
the merit order changes the merit order such that all flexible power plants
are operational on the spot market, such that no power plant is left for
the capacity reserve necessary to warrant grid reliability.
The total amount of inflexibility fees paid for each power plants
participating the spot market now is available for a capacity mechanism,
as described in the following section.
\section{Accumulated inflexibility fees paying capacity reserves}
\label{sec-capacity-reserve}%
A power plant serving as a power reserve for periods of scarcity or blackouts
should have fast and guaranteed start-up times, i.e.,
should be operationally flexible to a high degree.
There exist several proposed capacity mechanisms,
for instance capacity markets
or a strategic reserve determined by a grid agency.
In either of these approaches,
we therefore require a power plant to offer capacity reserves
to have a high operational flexibility $\varphi$, say
\begin{equation}
\varphi > \frac12
.
\label{eq-capacity-market-flexibility-requirement}
\end{equation}
This value means that the guaranteed start-up time of
a power plant participating the capacity mechanism
must be less than one hour.
A further natural requirement is that a power plant offering its reliability
on the capacity market cannot participate on the spot market.
Assume then that there are $k$ power plants participating on the capacity market,
each one established with a unique index $i$ $=$ $1$, \ldots, $k$.
Let $\varphi_i$ and $P_i$
denote the operational flexibility and the capacity (measured in MW)
of power plant $i$, respectively,
and let $C_f$ be the total of inflexibility fees accumulated
on the spot market in a certain past period, say, the day before.
It has the dimension currency per time, for instance \euro$/$h.
Then the \emph{reliability payment}\index{reliability payment} $\rho_i$
for power plant $i$ in that period is defined as
\begin{equation}
\rho_i
= \frac{\varphi_i \, P_i}{P_{\mathrm{flex}}}
\ C_f
\quad \left[\frac{\mbox{\euro}}{\mbox{h}}\right]
\qquad \mbox{where} \qquad
P_{\mathrm{flex}} = \sum_{j=1}^k \varphi_j \, P_j
.
\label{eq-reliablity-fee}
\end{equation}
Note that by construction $\sum_1^k \rho_i$ $=$ $C_f$, i.e., the sum over all
reliability payments equals the total amount of the inflexibility fees.
The quantity $P_{\mathrm{flex}}$ is the weighted sum of all available capacities,
where the weights are precisely the respective operational flexibilities.
\begin{beispiel}
\label{bsp-toy-grid-capacity-market}%
Assume the toy grid from Example \ref{beisp-toy-grid}.
Then by the requirement (\ref{eq-capacity-market-flexibility-requirement})
only three power plants can participate at the capacity market, namely
the hydroelectric power station, the CHP plant and the gas turbine.
\begin{table*}[htp]
\centering
\begin{footnotesize}
\begin{tabular}{|l|c|r||r|r||r|r|}
\hline
\textbf{Power Plant}
& $\mathbf{\varphi}$
& \textbf{Capacity}
& \multicolumn{4}{c|}{$\mathbf{\rho_i}$ \textbf{[\euro/h]}}
\\ \cline{4-7}
&
& \multicolumn{1}{c||}{\textbf{[MW]}}
& \multicolumn{2}{c||}{$\fett{C}_s$ = \textbf{205 \euro/h}}
& \multicolumn{2}{c|}{$\fett{C}_s$ = \textbf{790 \euro/h}}
\\ \hline
hydroelectric power station
& $.980$ & $5$ \quad \
& $74$ \hspace*{.2ex} & --- \hspace*{.5ex}
& $284$ \hspace*{.2ex} & --- \hspace*{.5ex}
\\
gas turbine
& $.893$ & $5$ \quad \
& $67$ \hspace*{.2ex} & $105$ \hspace*{.2ex}
& $259$ \hspace*{.2ex} & $404$ \hspace*{.2ex}
\\
cogeneration plant (CHP)
& $.855$ & $5$ \quad \
& $64$ \hspace*{.2ex} & $100$ \hspace*{.5ex}
& $247$ \hspace*{.2ex} & $386$ \hspace*{.2ex}
\\
\hline
\end{tabular}
\end{footnotesize}
\caption{\label{tab-toy-capacity-market}\footnotesize
The three power plants participating at the capacity market of our toy grid
in Example \ref{bsp-toy-grid-capacity-market}
and their reliability payments $\rho$ in dependence to the total inflexibility fee
coming from the spot market.
}
\end{table*}
In Table \ref{tab-toy-capacity-market} they are listed with their capacities
and the resulting reliability payments according to Equation (\ref{eq-reliablity-fee})
and depending on the amount of total inflexibility fee coming from the
spot market.
For calculational details refer to the Excel file
\url{http://math-it.org/climate/operational-flexibilities.xls}.
\end{beispiel}
\section{Discussion}
\label{sec-discussion}%
In this paper a proposal has been worked out to integrate
operational flexibility into
the sell bids of power plants participating wholesale electricity
spot markets.
The main idea is to calculate a fee for each power plant depending on its
operational flexibility.
For this purpose the concept of a
general measure of operational flexibility of a power plant is introduced
here as a strictly monotonically decreasing function $\varphi$
of the guaranteed start-up time, normed by
condition (\ref{eq-operational-flexibility-norm-condition}).
With such a measure, the inflexibility is priced in
by Equation (\ref{eq-price-regarding-operational-flexibility})
to the marginal price determining the sell bid
of each power plant at the spot market.
The amount depends on a market-wide reference level price $p_0$ which
is set by the market authority or the state.
The total operational inflexibility fee $C_f$ accumulated at the
spot markets then
is spread on the power plants participating in a given
capacity mechanism,
depending on their operational flexibilities according to Equation
(\ref{eq-reliablity-fee}).
Here the power plants forming a capacity reserve should
have a very high operational flexibility,
to guarantee reliability and stability of the grid.
A reasonable value is proposed by inequality
(\ref{eq-capacity-market-flexibility-requirement}).
A simple example of a measure for operational flexibility is given
by Equation (\ref{eq-operational-flexibility}).
Using this measure, the spot market and the corresponding
payments to power plants participating in a capacity mechanism
are applied to a simple but prototypical toy grid
in Examples \ref{beisp-toy-grid} and \ref{bsp-toy-grid-capacity-market}.
The most important consequence of our proposal,
as viewed from an economic perspective,
is the internalization of the negative externality of operational
inflexibility of power plants.
With the inflexibility fees determined as above, the currently
external costs would thus be paid by the spot markets
and could be used to pay capacity reserves,
be it on a separate capacity market or another capacity mechanism
such as a pool of power plants forming a strategic reserve.
The inflexibility fee therefore increases welfare without necessarily
decreasing dispatch efficiency.
A critical point of our approach, however, is the determination of
the reference level price
$p_0$. It is crucial since it can even change the
merit order of electricity markets if it is set very high.
Although a change of the merit order in itself does not
necessarily imply severe problems,
it could nonetheless lead to the paradox that
operationally flexible power plants participate in a short-term
spot market and therefore could not serve as a capacity reserve.
An amount $p_0$ too high would thus be adverse to the intention to
pay a capacity mechanism and thus would even diminish welfare.
We therefore are faced with the conflicting objectives of providing
enough means to fund the reserves of a capacity mechanism,
and of keeping suitable power plants with high
operational flexibility as capacity reserve.
Although this risk is calculable when choosing the amount
for a given grid cautiously such that experiences could be
gained over time,
a comprehensive theoretical framework to illuminate
effects and limits of inflexibility fees
on electricity markets should be accomplished.
Hints to tackle this problem may be indicated
by the optimal taxation due to Ramsey \cite{Ramsey-1927},
or by regulation theory \cite[§13]{Stroebele-et-al-2012}.
Further research in this direction appears worthwhile.
|
\section*{Introduction}
Let $X$ be a smooth projective rational surface
which admits a morphism $\pi\colon X\to\mathbb{P}^1$
whose general fiber is a smooth curve of genus
one. We say that $\pi$ is {\em minimal} if it does
not contract any $(-1)$-curve.
Denote by $K_X$ the canonical divisor class on $X$;
by~\cite{BHPV}*{12.1} the map $\pi$ is
induced by a complete linear system of the
form $|{-mK_X}|$, with $m>0$. As a consequence,
the anticanonical divisor on $X$ is semiample with $(-K_X)^2=0$.
It is possible to show~\cite{DoCo} that
$\pi$ is the unique elliptic fibration on $X$,
that is $\pi$ admits a unique multiple fiber and that
this multiplicity is $m$. The number $m$
is the {\em Halphen index} of $X$.
It is well known that
the Mori cone of $X$ is finitely generated if and only
if the Cox ring is finitely generated and these two conditions
are equivalent to requiring $X$ to have a finite number
of $(-2)$-curves and $(-1)$-curves~\cite{AL,To}.
While the former
are exactly the prime components of the reducible
fibers of $\pi$, thus easily enumerated, less
is known about the number of $(-1)$-curves of $X$.
The aim of this note is to provide a combinatorial
approach to determine the number of such curves.
More precisely, we
associate to any such surface $X$ a $\mathbb{Q}$-factorial
projective toric variety $Z_X$ in the following way.
Let $F$ be the free
abelian group with basis $\{f_C\}$ indexed by the
prime components $C$ of the reducible fibers of $\pi$.
By the genus formula it follows that such components
are exactly the $(-2)$-curves of $X$.
The linear map
\[
P_X
\colon
F\to \dfrac{ \hphantom{{}^\bot}
{K_X}^\bot} {\langle K_X\rangle} = {{\rm E}_8}\cong \mathbb{Z}^8
\qquad {\textrm{determined by}}\qquad
f_C\mapsto [C]
\]
defines the $1$-skeleton of the fan of a projective
toric variety $Z_X$. It is possible to show that there
is a unique complete toric variety $Z_X$ with this
skeleton (see Remark~\ref{fano}).
Recall that, given $E=\operatorname{Hom}(F,\mathbb{Z})$ the dual of $F$,
the cokernel $Q_X$ of the dual map $P_X^*$ is the
class map $Q_X\colon E\to \operatorname{Cl}(Z_X)$.
This allows us to define a homomorphism
\[
\alpha\colon\operatorname{Pic}(X)\to\operatorname{Cl}(Z_X)
\]
by mapping any divisor class $[D]$ to the element
$(\Gamma\mapsto D\cdot\Gamma)$ of $E$
and then applying $Q_X$.
In what follows we will denote by $=_{\rm free}$
equality up to torsion.
Our main result is the following theorem.
\begin{introtheorem}
\label{main}
Let $\pi \colon X \to \mathbb{P}^1$ be a minimal
elliptic fibration on a rational surface with Halphen
index $m$ and let $Z_X$ be the associated
toric variety.
\begin{enumerate}
\item
\label{one}
All the $(-1)$-curves of $X$ are mapped to the
same class $\delta_X\in\operatorname{Cl}(Z_X)$ by $\alpha$.
The set of $(-1)$-curves of $X$ is in bijection
with the set of integral points of the Riemann-Roch
polyhedron
\[
\Delta(\delta_X)
:=
Q_X^{-1}(\delta_X)\cap (E_\mathbb{Q})_{\geq 0}.
\]
Moreover the coordinates of an integral
point of the polyhedron $\Delta(\delta_X)$
are the intersection multiplicities of the
corresponding $(-1)$-curve with
the prime components of the fibers of $\pi$.
\item
\label{two}
For any reducible fiber $F_i$ of $\pi$ there exists
an element $w_i\in\operatorname{Cl}(Z_X)$ such that
for any prime component $C$ of $F_i$
\[
Q_X(C)=_{\rm free} n_C\, w_i
\]
where $n_C$ is the positive integer which
appear as a coefficient of the root $C$ in the unique
integer relation between the roots of the affine Dynkin
diagram corresponding to $F_i$.
\item
\label{three}
Let $F_1,\dots,F_r$ be the reducible fibers of $\pi$,
let $w_1,\dots,w_r\in\operatorname{Cl}(Z_X)$ be as in (2) and let
$F$ be the multiple fiber of $\pi$.
Then the following holds
\[
\delta_X
=_{\rm free}
\begin{cases}
m(w_1+\dots+w_r) & \text{ if $F$ is not in } \{ F_1 , \ldots , F_r\}, \\
m(w_1+\dots +w_r) - (m-1)w_i & \text{ if $F=F_i$}.
\end{cases}
\]
\end{enumerate}
\end{introtheorem}
The paper is organised as follows. In Section~\ref{sec:1}
we prove Theorem~\ref{main}. Section~\ref{sec:2}
is devoted to rational elliptic surfaces of Halphen
index two and the enumeration of their $(-1)$-curves
as an application of Theorem~\ref{main}.
Finally in the Appendix we list the degree matrices
of the thirteen $\mathbb{Q}$-factorial Fano toric varieties
related to root lattices and which appear in the
theory exposed in this note.
\section{$(-1)$-curves and lattice points}
\label{sec:1}
Let $\pi\colon X\to\mathbb{P}^1$ be a minimal elliptic fibration
on a smooth projective rational surface.
Denote by ${{\rm E}_8}$ the lattice ${K_X}^\perp / \langle K_X\rangle$ and by
$p \colon {K_X}^\perp \to {{\rm E}_8}$ the canonical quotient map; as the
notation suggests, the lattice ${{\rm E}_8}$ is isometric to the negative definite
${{\rm E}_8}$-lattice.
We let $F$ be the free abelian group with basis indexed
by the $(-2)$-curves on $X$ and we let $\operatorname{cl}_X \colon F \to {K_X}^\perp$
be the homomorphism assigning to each basis vector
the corresponding $(-2)$-class; we denote by $P_X$ the
composition $p \circ \operatorname{cl}_X$. The images of the rays
of the standard basis of $F$ under the homomorphism
$P_X$ define a one-dimensional fan in ${{\rm E}_8}$.
We will show in Remark~\ref{fano}
that there exists a unique projective $\mathbb{Q}$-factorial
toric variety $Z_X$ whose one-dimensional
cones are the cones of the fan defined by $P_X$.
We summarise the situation in the following
commutative diagram with exact rows
\begin{equation} \label{diag:fan}
\begin{array}{c}
\xymatrix{
F\ar@{=}[d]\ar[r]^-{\operatorname{cl}_X} & {K_X}^\perp \ar[d]^-p\ar[r]^-\tau & A\ar[r]\ar[d] & 0 \\
F \ar[r]^{P_X} & {{\rm E}_8}\ar[r]^-{\bar\tau} & A/\langle h\rangle\ar[r] & 0
}
\end{array}
\end{equation}
where $A$ is the cokernel of $\operatorname{cl}_X$ and
$h:=\tau(K_X)$.
\begin{remark}
The image of the homomorphism $P_X$ is
a root sublattice $\Lambda$ of ${{\rm E}_8}$.
According to~\cite{OgSh}*{Thm. 3.2 and Thm. 3.3}
there is a unique isometric embedding of $\Lambda$
into ${{\rm E}_8}$ modulo the action of the Weyl group
of ${{\rm E}_8}$ unless $\Lambda$ is one of the following
lattices: $A_7$, $2A_3$, $A_5+A_1$, $A_3+2A_1$, $4A_1$.
In particular, if the rank of $\Lambda$ is eight, then
the embedding is unique.
\end{remark}
Let $E=\operatorname{Hom}(F,\mathbb{Z})$ be the dual of $F$
and let $Q_X \colon E \to \operatorname{Cl}(Z_X)$ be the
cokernel of the dual map $P_X^* \colon {\rm E}_8^\ast \to E$.
Dualising the exact sequences in~\eqref{diag:fan}
we obtain the following commutative diagram with
exact rows
\begin{equation} \label{diag:cl}
\begin{array}{c}
\xymatrix{
({K_X}^\perp)^*\ar[r]^-{\operatorname{cl}_X^*} & E \ar[r]\ar@{=}[d]
& \dfrac{\operatorname{Cl}(Z_X)}{\langle\delta_X\rangle}\ar[r] & 0\\
({{\rm E}_8})^* \ar[r]^-{P_X^*}\ar[u]^-{p^*} & E\ar[r]^-{Q_X}
& \operatorname{Cl}(Z_X)\ar[r]\ar[u]^-\tau & 0
}
\end{array}
\end{equation}
where $\delta_X\in\operatorname{Cl}(Z_X)$ generates the kernel
of $\tau$ and is defined by
\[
\delta_X
=
(Q_X\circ\operatorname{cl}_X^*)(m),
\]
where $m\in ({K_X}^\perp)^*$ generates the cokernel
of $p^*$, that is $m$ is any projection of the
homomorphism $i\colon\mathbb{Z}\to{K_X}^\perp$ defined by
$1\mapsto -K_X$.
Observe that $\delta_X$ does not depend
on the choice of $m$.
Given a class $w\in\operatorname{Cl}(Z_X)$ we recall that the
{\em Riemann-Roch polytope} of $w$ is
the convex polytope
\[
\Delta(w) := Q_Z^{-1}(w)\cap (E_\mathbb{Q})_{\geq 0},
\]
where $(E_\mathbb{Q})_{\geq 0}$
is the positive orthant of the rational vector
space $E_\mathbb{Q} = E\otimes_\mathbb{Z}\mathbb{Q}$.
Assume now that $D$ is a $(-1)$-curve of
$X$ and denote by $e_D$ the linear function
$F\to\mathbb{Z}$ defined by $f\mapsto \operatorname{cl}_X(f)\cdot D$,
where the product is induced by the intersection
pairing of $X$.
We therefore obtain a function
\[
\varepsilon \colon \bigl\{ (-1){\textrm{-curves of }} X \bigr\} \longrightarrow E
\qquad
D\longmapsto e_D.
\]
The following lemma is needed in the proof of
Lemma~\ref{mlemma} (see also~\cite[Proposition~3.3]{LH}
for another proof).
\begin{lemma}
\label{lem:m1}
Let $\pi \colon X \to \mathbb{P}^1$ be a minimal
elliptic fibration on a rational surface and let $D$
be a divisor on $X$ such that $D^2=D\cdot K_X=-1$
and $D\cdot C\geq 0$ for any $(-2)$-curve of $X$.
Then $D$ is linearly equivalent to a $(-1)$-curve.
\end{lemma}
\begin{proof}
Since the elliptic fibration $\pi$ is minimal, the
anticanonical divisor $-K_X$ is linearly equivalent to a
multiple of a fiber and thus it is nef.
By the Riemann-Roch theorem either $D$ or ${K_X-D}$
is linearly equivalent to an effective divisor.
The equality $-K_X \cdot (K_X-D) = -1$ and the fact that
the divisor $-K_X$ is nef allow us to conclude that
the first possibility occurs. Thus, without loss
of generality, we can assume $D$ to be effective.
By the genus formula and the fact that $-K_X$ is nef,
the only curves of $X$ which are orthogonal
to $-K_X$ are $(-2)$-curves, while the only ones
which have intersection $1$ with $-K_X$ are the
$(-1)$-curves. Thus the equality $-K_X\cdot D=1$
implies $D = \Gamma + D_1$, where $\Gamma$
is a $(-1)$-curve and $D_1$ is a non-negative sum
of $(-2)$-curves. From $(\Gamma+D_1)^2=D^2=-1$
we deduce the following
\[
0\leq (D_1+\Gamma)\cdot D_1= -\Gamma\cdot D_1\leq 0,
\]
where the first inequality is due to the hypothesis
on $D=(D_1+\Gamma)$ and the definition of $D_1$,
while the second inequality is due to the fact that
$\Gamma$ is not a component of $D_1$.
Thus $(D_1)^2=\Gamma\cdot D_1=0$ and,
since $-K_X\cdot D_1=0$, by the Hodge index theorem
we conclude that $D_1$ is linearly equivalent to
$-\alpha K_X$ for some rational number $\alpha$.
Using the identities $-1=D^2=(\Gamma-\alpha K_X)^2=-1+2\alpha$
we deduce that $\alpha=0$ and thus $D=\Gamma$ is
a $(-1)$-curve.
\end{proof}
\begin{lemma}\label{mlemma}
The function $\varepsilon$ induces a bijection
between the set of $(-1)$-curves on $X$ and
the set of integral points of the polytope
$\Delta(\delta_X)$.
\end{lemma}
\begin{proof}
We begin by showing that the image of $\varepsilon$ is
contained in $\Delta(\delta_X)$. Indeed, if $D$ is a
$(-1)$-curve of $X$ the homomorphism
$e_D\colon F\to\mathbb{Z}$ factors through ${K_X}^\perp$
by definition and it defines a projection
$m_D\colon {K_X}^\perp\to\mathbb{Z}$ of $i\colon\mathbb{Z}\to{K_X}^\perp$
since $-K_X\cdot D = 1$.
Thus $Q_X(e_D)=\delta_X$ by our definition
of $\delta_X$. Moreover,
for each element $f$ of the standard basis
of $F$, we have
\[
e_D(f) = \operatorname{cl}_X(f) \cdot C \geq 0,
\]
where the last inequality is due to the fact that
$\operatorname{cl}_X(f)$ and $D$ correspond to distinct
irreducible curves on $X$. We deduce that the
image of $\varepsilon$ is contained in
$\Delta(\delta_X)$.
Now, we show that $\varepsilon$ is injective.
Let $D,D'$ be $(-1)$-curves on $X$ and suppose
that $e_D=e_{D'}$ holds.
It follows that the difference $D-D'$ is proportional
to $K_X$. We deduce that
$0 = (D-D')^2 = -2 - 2 D \cdot D'$, so that
$D$ and $D'$ coincide.
Finally, let $e$ be an integral point of $\Delta(\delta_X)$.
Therefore, we write $e = e_D + P_X^*(\gamma)$,
where $D$ is a $(-1)$-curve on $X$
and $\gamma$ is an element of ${\rm E}_8^\ast$.
We deduce that $e$ factors through $\operatorname{cl}_X$
inducing a retraction $m_e \colon {K_X}^\perp \to \mathbb{Z}$
of $i$. Using the perfect pairing
\[
\xymatrix{
{K_X}^\perp \times \dfrac{\operatorname{Pic}(X)}{\langle-K_X\rangle}
\ar[r] &
\mathbb{Z},
}
\]
we see that $m_e$ is multiplication by an element
$[D]+\mathbb{Z}(-K_X)$, where $D \cdot (-K_X)=1$
and for each $(-2)$-curve $\Gamma$ on $X$ the
inequality $D \cdot \Gamma \geq 0$ holds.
Let $r$ be an integer such that $(D+r K_X)^2 = D^2 - 2r = -1$;
such an integer $r$ exists since $D^2$ is odd by the
adjunction formula. By Lemma~\ref{lem:m1},
the class of $D+rK_X$ is the class of a
$(-1)$-curve and we are done.
\end{proof}
We now make explicit the homomorphism
$Q \colon E \to \operatorname{Cl}(Z)$.
We begin with a sublattice $\Lambda$ of ${{\rm E}_8}$
spanned by roots of ${{\rm E}_8}$.
Let $\Lambda = \Lambda_1 \oplus \cdots \oplus \Lambda_r$
be a decomposition of $\Lambda$ as a direct sum of irreducible
root lattices.
We fix a subset $\{\lambda_1 , \ldots , \lambda_s\}$
of the lattice $\Lambda$ consisting of simple roots
chosen as follows.
First of all we choose a basis of $\Lambda_1$
indecomposable roots together with the root corresponding
to the affine vertex in the extended Dynkin diagram of
$\Lambda_1$, equivalently, the last root is the opposite
of the longest root of $\Lambda_1$. Then we proceed
similarly with $\Lambda_2$ and so on.
Let $F$ be the free abelian group of rank $s$
with basis $f_1 , \ldots , f_{s}$.
Let $P \colon F \to {{\rm E}_8}$ be the homomorphism
defined by $f_1 \mapsto \lambda_1 , \ldots , f_{s} \mapsto \lambda_{s}$
and let $E=\operatorname{Hom}(F,\mathbb{Z})$ be the dual lattice of $F$.
The homomorphism $P$ defines a toric variety
$Z$ and an exact sequence
\[
\xymatrix@1{
{\rm E}_8^\ast\ar[r]^-{P*} & E\ar[r]^-Q & \operatorname{Cl}(Z)\ar[r] & 0.
}
\]
as in~\eqref{diag:cl}.
Let $\pi \colon X \to \mathbb{P}^1$ be a
minimal rational elliptic surface with singular fibers
of type $\widetilde{\Lambda}_1 , \ldots , \widetilde{\Lambda}_r$.
We denote the $(-2)$-curves of $X$ by
$C_1 , \ldots , C_{s}$ where the numbering is
compatible with the numbering of the roots
$\lambda_1 , \ldots , \lambda_{s}$ given above.
It follows from the definitions that the homomorphisms
$P$ and $P_X$ coincide, and thus that also
$Q$ and $Q_X$ coincide.
With the above notation the matrix $Q$
has the following form
\begin{equation}
\label{Q-mat}
Q
=
\left[
\begin{array}{cccc}
Q_1 & 0 & \cdots & 0\\
0 & Q_2 & \cdots & 0\\
\vdots & \ddots & \ddots & \vdots\\
0 & \cdots & 0 & Q_r\\
\hline
\multicolumn{4}{c}{\text{torsion}}
\end{array}
\right]
\end{equation}
where each $Q_i$ is the row vector of length
${\rm rk} \, \Lambda_i+1$ defined
by the unique relation between the simple roots
of the chosen basis of $\Lambda_i$
together with the root corresponding to its
affine vertex.
In these coordinates the matrix $Q$ defines
a map $E\to\mathbb{Z}^r\oplus ({{\rm E}_8}/\Lambda)$.
\begin{proof}[Proof of Theorem~\ref{main}]
The first part of~\eqref{one} follows from Lemma~\ref{mlemma}.
For the second part observe that a $(-1)$-curve $D$
gives an element $e_D\in E$ by taking intersection
with $D$. Thus the statement follows from
the fact that the canonical basis of $F$
consists of the prime components of the
fibers of $\pi$.
To prove items~\eqref{two} and~\eqref{three}
we begin by recalling that the class $\delta_X\in\operatorname{Cl}(Z_X)$
is defined by any
projection $m\colon{K_X}^\perp\to\mathbb{Z}$ of the inclusion
map $i\colon \mathbb{Z}\to{K_X}^\perp$ given by $1\mapsto -K_X$.
In particular such a projection is obtained by
intersecting with a $(-1)$-curve $D$ of $X$,
that is $m(C)=C\cdot D$ for any $C\in{K_X}^\perp$.
Due to the above presentation of the matrix $Q$
the integers in the row corresponding to the vector $Q_i$ are
exactly the multiplicities of the unique primitive
vector of self-intersection zero in the root
lattice $\widetilde\Lambda_i\subset{K_X}^\perp$
whose class is thus the class of $-mK_X$
whenever the corresponding $(-2)$-curves
form a fiber of $\pi$ or it is $-K_X$ whenever
these curves form the support of the non-reduced
fiber. Thus items~\eqref{two} and~\eqref{three} follow.
\end{proof}
\begin{remark}\label{fano}
Due to~\eqref{Q-mat} the free part of the
grading of any variable of the Cox ring
of the toric variety $Z$ is one of the canonical
vectors $w_1,\dots,w_r$ of $\mathbb{Z}^r$ and any
such vector occurs as a degree of at least
two variables. As a consequence, by
~\cite[Proposition~2.4.2.6]{ADHL}, there is
a unique projective variety $Z$ with the
given Cox ring and such that
\[
{\rm SAmple}(Z) = {\rm Mov}(Z) = {\rm Eff}(Z)
= {\rm cone}(w_1,\dots,w_r).
\]
It follows that any class in the relative interior
of the above cone defines the same $\mathbb{Q}$-factorial
projective toric variety $Z$. This is the case for
the anticanonical class $-K_Z$ so that
the variety $Z$ is Fano.
\end{remark}
\begin{remark}
Recall~\cite{MiSt}*{Lemma~8.16} that if $Z$
is a toric variety with Cox ring $\mathcal{R}(Z)$
and grading matrix $Q:=[w_1,\dots,w_n]$, then
the multigraded Hilbert series of
$\mathcal{R}(Z)$ is
\[
H(\mathcal{R}(Z),t)
=
\frac{1}{(1-t^{w_1})\cdots(1-t^{w_n})},
\]
where $t^w$ is the character of the quasi-torus
${\rm Spec}(\mathbb{C}[\operatorname{Cl}(Z)])$ corresponding to $w\in\operatorname{Cl}(Z)$.
Hence Table~\ref{gradings} and
Theorem~\ref{main}
allow us to determine the number of $(-1)$-curves
on any such surface.
\end{remark}
\begin{example}\label{ex:D8}
Let $X$ be a minimal rational elliptic surface of type
$D_8$ of Halphen index $2$ and whose multiple
fiber is irreducible. According to the table appearing in
the Appendix, the Hilbert series of the Cox ring
of the corresponding toric variety $Z$ is
\[
\frac{1}{(1-t)^2(1-tu)^2(1-t^2)^3(1-t^2u)^2}
\]
where the two variables are in $\mathbb{C}[t,u]/\langle u^2-1\rangle$.
Expanding the above series, a direct calculation
shows that the coefficient of the term $t^2$ is $9$,
while that of $t^{2}u$ is $6$.
Both cases are realized as shown in the next
paragraphs.
{\em Case 1}.
Choose nine points $p_1,\dots,p_9$ on a smooth
plane cubic curve $C$ with flex at $p_1$
such that the classes of $p_2-p_1,\dots,p_9-p_1$
are the following elements of $\mathbb{Z}/2\mathbb{Z}\oplus\mathbb{Z}/2\mathbb{Z}$:
$00,00,01,01,01,10,10,11,11$. For example we can take
$C:=V(x^3 - y^2z - xz^2)$,
$p_1=( 0, 1, 0 )$, $p_3=(-1, 0, 1)$, $p_6
= (0,0,1)$, $p_8=(1,0,1)$ and the remaining points
to be infinitely near to the given ones according
to the classes $p_i-p_1$ written above.
The corresponding pencil of sextics curves
is generated by $(x^3 - y^2z - xz^2)^2$ and
$x^2 y^2 (x - z)(x + z)$.
The classes of the $(-1)$-curves of $X$ are
the rows of the following matrix.
{\footnotesize
\[
\left[
\arraycolsep=3pt
\begin{array}{rrrrrrrrrr}
0&0&1&0&0&0&0&0&0&0\\
0&0&0&0&0&1&0&0&0&0\\
0&0&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&0&0&1\\
1&-1&-1&0&0&0&0&0&0&0\\
2&0&0&-1&-1&-1&-1&-1&0&0\\
2&0&0&-1&-1&-1&0&0&-1&-1\\
4&-2&-2&-1&-1&-1&-2&0&-1&-1\\
4&-2&-2&-1&-1&-1&-1&-1&-2&0
\end{array}
\right]
\]
}
The first column is the degree of the plane model,
while the remaining columns are the negative of the
multiplicities at the nine points. For example the first
row is an exceptional divisor while the last row is a
plane quartic with three double points and through
other five simple points. \\
{\em Case 2}.
Choose nine points $p_1,\dots,p_9$ on a smooth
plane cubic curve $C$ with flex at $p_1$
such that the classes of $p_2-p_1,\dots,p_9-p_1$
are the following elements of $\mathbb{Z}/4\mathbb{Z}$:
$0,0,2,2,2,1,1,3,3$. For example we can take
$C:=V(y^2z + xyz + yz^2 - x^3 - x^2z)$,
$p_1=( 0, 1, 0 )$, $p_3=(-1, 0, 1)$, $p_6
= (0,0,1)$, $p_8=(0,-1,1)$ and the remaining points
to be infinitely near to the given ones according
to the classes $p_i-p_1$ written above.
The corresponding pencil of sextics curves
is generated by $(y^2z + xyz + yz^2 - x^3 - x^2z)^2$
and $x^2(x+z)^2y(x+y+z)$. The classes of the
$(-1)$-curves of $X$ are the rows of the following matrix.
{\footnotesize
\[
\left[
\arraycolsep=3pt
\begin{array}{rrrrrrrrrr}
0&0&1&0&0&0&0&\hphantom{-}0&0&\hphantom{-}0\\
0&0&0&0&0&1&0&0&0&0\\
0&0&0&0&0&0&0&1&0&0\\
0&0&0&0&0&0&0&0&0&1\\
1&-1&-1&0&0&0&0&0&0&0\\
2&0&0&-1&-1&-1&-1&0&-1&0\\
\end{array}
\right]
\]
}
\end{example}
\section{Rational elliptic surfaces}
\label{sec:2}
Let $X$ be a smooth projective rational
surface with $-K_X$ nef and $(K_X)^2=0$.
Given a curve $C$ in the linear system $|{-K_X}|$
we can construct the homomorphism of groups
\[
\alpha \colon {K_X}^\perp \to {\rm Pic}^0(C)
\qquad
w\mapsto\imath^* w,
\]
where $\imath\colon C\to X$ is the inclusion.
We denote by $G$ the image of $\alpha$
and by $i\colon\mathbb{Z}\to{K_X}^\perp$ the
homomorphism defined by $1\mapsto -K_X$
and we let $H$ be the subgroup of $G$
generated by the images via $\alpha$ of the
classes of all the irreducible components of $C$.
Let $\kappa\colon\mathbb{Z}\to H$ be the homomorphism
defined by $1\mapsto h=\alpha(-K_X)$.
Using the isomorphism between the lattices
${{\rm E}_8}$ and ${K_X}^\perp / \mathbb{Z}(-K_X)$ we define
$p\colon{K_X}^\perp\to{{\rm E}_8}$ to be the corresponding projection.
We summarize these definitions in the
commutative diagram
\begin{equation}\label{main-diag}
\begin{array}{c}
\xymatrix{
0\ar[r] & \mathbb{Z} \ar[r]^-i \ar[d]^-\kappa
& {K_X}^\perp \ar[r]^p \ar[d]^\alpha
& {{\rm E}_8}
\ar[r] \ar[d]^\overline\gru & 0 \\
0\ar[r] & \langle h\rangle \ar[r]\ar@{^(->}[d] & G \ar[r]\ar@{=}[d]
& G/\langle h\rangle \ar[r]\ar@{->>}[d]^-\beta & 0 \\
0\ar[r] & H \ar[r] & G \ar[r] & G/H\ar[r] & 0 \\
}
\end{array}
\end{equation}
with exact rows.
We recall the proofs of the following well-known
lemmas for the sake of completeness.
See also~\cite{DoCo} and~\cite{Har}.
\begin{lemma} \label{res}
The order of the element $h$ is finite if and only if the
surface $X$ is a minimal rational elliptic surface.
Moreover, if $h$ has finite order $m$,
then the linear system $|{-mK_X}|$ is the
elliptic pencil.
\end{lemma}
\begin{proof}
Given $C\in |{-K_X}|$ there is an exact sequence
of sheaves
\begin{equation}
\label{seqK}
\xymatrix{
0\ar[r] & \mathcal{O}_X(K_X)\ar[r] & \mathcal{O}_X\ar[r] & \mathcal{O}_C\ar[r] & 0.
}
\end{equation}
Let $n>0$ be an integer.
By taking tensor product with $\mathcal{O}_X(-nK_X)$ and
passing to the long exact sequence in cohomology,
we see that the inequality $h^0(X,-nK_X)>1$ holds if and only
$\alpha(-nK_X)$ is trivial in ${\rm Pic}^0(C)$, that is
if and only if the class $h=\alpha(-K_X)$ is torsion. This is
equivalent to requiring that $h$ has finite order.
\end{proof}
\begin{lemma} \label{effective}
Let $R$ be a root of ${{\rm E}_8}$. Then one
of the following possibilities occurs:
\begin{enumerate}
\item
\label{case-1}
if $\overline\alpha(R)=0$ then $p^{-1}(R)$
contains an effective class;
\item
\label{case-2}
otherwise the only effective
classes in $p^{-1}(R)$ are components of $C$.
\end{enumerate}
\end{lemma}
\begin{proof}
Assume that we are in case~\eqref{case-1}.
Let $D'$ be a class in $p^{-1}(R)$, so that
$D'^2=-2$ and $D'\cdot K_X=0$,
and let $m$ be an integer such that $\alpha(D')=mh$.
We denote by $D$ the class $D'-mK_X$,
so that $\alpha(D)$ is trivial. By abuse
of notation, denote by $D$ also a divisor of $X$
whose class is $D$ and form the exact sequence
of sheaves
\[
\xymatrix{
0\ar[r] & \mathcal{O}_X(D+K_X)\ar[r] & \mathcal{O}_X(D)\ar[r] & \mathcal{O}_C(D)\ar[r] & 0.
}
\]
Since $\alpha(D)$ is trivial,
the first two cohomology groups of $\mathcal{O}_C(D)$
have dimension one.
The above exact sequence and
the vanishing of the Euler characteristic of
$\mathcal{O}_X(D)$ immediately imply that at least
one group $H^0(X,D)$, $H^2(X,D+K_X)^\vee\cong H^0(X,-D)$
is non-trivial. Thus either $D$ or $-D$ is linearly
equivalent to an effective divisor $\bar D$.
If $D$ is linearly equivalent to $\bar D$,
then we are done. Otherwise there exists
a positive integer $n$ such that $D-nK_X
\sim -\bar D-nK_X$
is linearly equivalent to an effective divisor
and again the statement follows.
To prove~\eqref{case-2} it suffices to observe
that any effective class of $p^{-1}(R)$
is the class of a non-negative sum of
$(-2)$-curves and that
if $\Gamma$ is a $(-2)$-curve of $X$ not
contained in $C$, then $\Gamma$ is disjoint
from $C$ and hence $\alpha([\Gamma])=0$.
\end{proof}
Given a smooth rational surface
$X$ with $-K_X$ nef and ${K_X}^2=0$, denote by
$\Lambda_X$
the sublattice of ${K_X}^\perp$ spanned
by the classes of $(-2)$-curves
of $X$:
\[
\Lambda_X
=
\langle
R \,:\, R \text{ is the class of a }(-2)\text{-curve}
\rangle.
\]
\begin{remark}
By the adjunction formula $\Lambda_X$ is a
negative semidefinite sublattice of ${K_X}^\perp$
and the quotient $\Lambda_X/\langle K_X\rangle$
is a negative definite lattice.
By Lemma~\ref{effective} each fiber of the
quotient homomorphism $\Lambda_X\to\Lambda_X/\langle K_X\rangle$
contains at most a finite number of
classes of $(-2)$-curves. Since the
image of the set of classes of $(-2)$-curves
of $X$ is a subset of the finite set of roots
of $\Lambda_X/\langle K_X\rangle\subseteq{{\rm E}_8}$
we conclude that
$X$ contains finitely many $(-2)$-curves.
\end{remark}
We assume now that $C$ is a smooth plane cubic,
$p_1,\ldots,p_9$ are nine points on $C$,
not necessarily distinct, and $p_1$ is a flex.
Denote by $X$ the blow up of $\mathbb{P}^2$ at the nine points
$p_1 , \ldots , p_9$, by $E_1,\ldots,E_9$ the corresponding
sequence of $(-1)$-classes in $\operatorname{Pic} (X)$ and by $L$
the class of the pull-back of a line.
Recall from~\eqref{main-diag} that
the homomorphism $\alpha$ is given by
\begin{equation}
\label{restr}
\alpha \colon {K_X}^\perp \to G
\qquad
\qquad
d L - (m_1 E_1 + \cdots + m_9 E_9) \mapsto \sum _{i = 2}^9 m_i p_i
\end{equation}
since $L|_C\sim 3p_1$ and $p_1$ is the origin
of the elliptic curve $(C,p_1)$.
The {\em characteristic sequence} of
the pair given by the points $p_1,\dots,p_9$
and the smooth plane cubic curve $C$ is the
sequence $[p_1,\dots,p_9]$ in $\operatorname{Pic}^9(C)$.
\begin{proposition}
Let $\Lambda$ be a sublattice of ${{\rm E}_8}$
of type ADE and of finite index.
Then there exists a minimal rational elliptic
surface $X$ such that $\Lambda_X/\langle K_X\rangle$ is
isomorphic to $\Lambda$.
An example of such a surface for
any such $\Lambda$ is given in Table~\eqref{tab}.
{\footnotesize
\begin{longtable}{l|l|l}
\label{tab}
Lattice & Group & Characteristic sequence\\
\midrule
$\operatorname{E}_8$ & $\langle 0\rangle$ & $[0,0,0,0,0,0,0,0,0]$ \\
$\operatorname{D}_8$ & $\mathbb{Z}/2$ & $[0,0,0,0,0,1,1,1,1]$ \\
$\operatorname{E}_7+\operatorname{A}_1$ & $\mathbb{Z}/2$ & $[0,0,0,0,0,0,0,1,1]$ \\
$\operatorname{A}_8$ & $\mathbb{Z}/3$ & $[0,0,0,0,1,1,1,1,2]$ \\
$\operatorname{E}_6+\operatorname{A}_2$ & $\mathbb{Z}/3$ & $[0,0,0,0,0,0,1,1,1]$ \\
$\operatorname{A}_7+\operatorname{A}_1$ & $\mathbb{Z}/4$ & $[0,0,0,0,1,1,2,2,2]$ \\
$\operatorname{D}_5+\operatorname{A}_3$ & $\mathbb{Z}/4$ & $[0,0,0,0,0,1,1,1,1]$ \\
$2\operatorname{A}_4$ & $\mathbb{Z}/5$ & $[0,0,0,0,1,1,1,1,1]$ \\
$\operatorname{A}_5+\operatorname{A}_2+\operatorname{A}_1$ & $\mathbb{Z}/6$& $[0,0,0,1,1,1,1,1,1]$ \\
$\operatorname{D}_6+2\operatorname{A}_1$ & $\mathbb{Z}/2\oplus\mathbb{Z}/2$ & $[0,0,0,0,0,10,10,01,01]$ \\
$2\operatorname{D}_4$ & $\mathbb{Z}/2\oplus\mathbb{Z}/2$ & $[0,0,0,10,10,01,01,11,11]$ \\
$2\operatorname{A}_3+2\operatorname{A}_1$ & $\mathbb{Z}/2\oplus\mathbb{Z}/4$ & $[0,0,0,10,10,01,01,01,01]$\\
$4\operatorname{A}_2$ & $\mathbb{Z}/3\oplus\mathbb{Z}/3$ & $[0,0,0,10,10,10,01,01,01]$\\[5pt]
\caption{Characteristic sequences for maximal rank lattices}\\
\end{longtable}
}
\end{proposition}
\begin{proof}
By hypothesis the group $G:={{\rm E}_8}/\Lambda$
is finite. Moreover, by the classification
of sublattices of ${{\rm E}_8}$ generated by roots, the
lattice $\Lambda$ is isomorphic to one of the
lattices of first column of the above table.
Assume now that there exists a characteristic
sequence $[p_1,\dots,p_9]$ of points on
a smooth plane cubic curve $C$ such that
$\Lambda$ is contained in the kernel of
the homomorphism $\alpha$ of~\eqref{restr}
and moreover $\alpha$ is surjective.
By Lemma~\ref{res} and the fact that
$G$ is a finite group we deduce that the
surface $X$ is elliptic and minimal.
Moreover the prime components of the unique
elliptic fibration $\pi\colon X\to\mathbb{P}^1$ form
a Coxeter-Dynkin diagram of type $\tilde\Lambda$
by Lemma~\ref{effective}, which gives
the statement.
Hence it suffices to find a characteristic
sequence realizing each of the thirteen
lattices; this is shown in the third
column of the table.
\end{proof}
As an application of our results we determine the
number of $(-1)$-curves on any minimal elliptic
surface of Halphen index $2$.
In what follows, given a root lattice $\Lambda$
we denote by $\tilde\Lambda$ the corresponding
affine root lattice and by $v_\Lambda$ a primitive
vector of null square of $\tilde\Lambda$.
\begin{lemma}
\label{lifting}
Let $\imath\colon \Lambda\to{{\rm E}_8}$ be an embedding of a
root lattice of rank eight into ${{\rm E}_8}$. Then
the embeddings $\tilde\imath\colon\tilde\Lambda\to{\rm \widetilde {E}}_8$,
which are compatible with $j$, are classified by the pairs $(m,\xi)$,
where $m$ is a positive integer and
$\xi\in {\rm Ext}^1_{\mathbb{Z}}({{\rm E}_8}/\Lambda,\mathbb{Z}/m\mathbb{Z})$.
\end{lemma}
\begin{proof}
Observe that there are two isometries
$\tilde\Lambda\cong\Lambda\oplus
\langle v_\Lambda\rangle$ and ${\rm \widetilde {E}}_8\cong{{\rm E}_8}\oplus
\langle v_{{\rm E}_8}\rangle$. Any isometric embedding
$\tilde\imath\colon\tilde\Lambda\to{\rm \widetilde {E}}_8$ must map $v_\Lambda$
to an integer multiple $mv_{{\rm E}_8}$ of $v_{{\rm E}_8}$ being
these the only vectors of null square of both
lattices. Moreover, up to sign change, we can
assume $m$ to be a positive integer. Then the
statement is consequence of the following
commutative diagram with exact rows and columns.
\[
\xymatrix@1{
& 0\ar[d] & 0\ar[d] & 0\ar[d] \\
0\ar[r] & \mathbb{Z}\ar[d]\ar[r]^-{\cdot m}& \mathbb{Z}\ar[r]\ar[d]
& \mathbb{Z}/m\mathbb{Z}\ar[d]\ar[r] & 0\\
0\ar[r] & \Lambda\oplus\mathbb{Z}\ar[r]^-{\tilde\imath}\ar[d]
& {{\rm E}_8}\oplus\mathbb{Z}\ar[d]\ar[r]^-{\tilde\beta}
& G\ar[d]^-\alpha\ar[r] & 0\\
0\ar[r] & \Lambda\ar[r]^-\imath\ar[d] & {{\rm E}_8}\ar[r]^-\beta\ar[d]
& {{\rm E}_8}/\Lambda\ar[r]\ar[d] & 0\\
& 0 & 0 & 0
}
\]
Indeed, given such an embedding $\tilde\imath$, its
cokernel $\tilde\beta$ gives an extension of ${{\rm E}_8}/\Lambda$
by $\mathbb{Z}/m\mathbb{Z}$ represented by the third column
of the above diagram. Observe that the cokernel
$\tilde\beta$ is unique since $\tilde\imath$ is a
full-rank homomorphism of free abelian groups.
Moreover equivalent embeddings, that is
embeddings which differ up to isometries
of the domain $\Lambda\oplus\mathbb{Z}$ and
of the codomain ${{\rm E}_8}\oplus\mathbb{Z}$, give
isomorphic extensions of ${{\rm E}_8}/\Lambda$
by $\mathbb{Z}/m\mathbb{Z}$.
On the other hand, given an extension of
${{\rm E}_8}/\Lambda$ by $\mathbb{Z}/m\mathbb{Z}$,
since ${{\rm E}_8}\oplus\mathbb{Z}$ is a free abelian group, one
one can lift the map $\beta$ to a map $\tilde\beta$
whose restriction to $\mathbb{Z}$ surjects onto $\mathbb{Z}/m\mathbb{Z}$.
The kernel $\tilde\imath$ of $\tilde\beta$ gives an
embedding of $\Lambda\oplus\mathbb{Z}$ into ${{\rm E}_8}\oplus\mathbb{Z}$.
\end{proof}
\begin{proposition}
Let $\pi\colon X\to\mathbb{P}^1$ be a minimal elliptic
fibration of Halphen index $2$ on a rational surface.
Then the number of $(-1)$-curves of $X$ is given
in Table~\eqref{halphen}.
{\footnotesize
\begin{table}[h]
\label{halphen}
\begin{tabular}{l|c}
Type & Number of $(-1)$-curves\\
\midrule
$E_8$ & $3$\\
$D_8$ & $6,9$\\
$E_7+A_1$ & $8,10$\\
$A_8$ & $15$\\
$E_6+A_2$ & $18$\\
$A_7+A_1$ & $24,30$\\
$D_5+A_3$ & $28,32$\\
$2A_4$ & $45$\\
$A_5+A_2+A_1$ & $60,66$\\
$D_6+2A_1$ & $26,28$\\
$2D_4$ & $28$\\
$2A_3+2A_1$ & $108,112$\\
$4A_2$ & $144$\\[5pt]
\end{tabular}
\caption{Number of $(-1)$-curves on Halphen elliptic surfaces of index $2$.}
\end{table}
}
\end{proposition}
\begin{proof}
Given a maximal rank sublattice $\Lambda$ of ${{\rm E}_8}$,
a positive integer $m$ and an element
$\xi\in {\rm Ext}^1_{\mathbb{Z}}({{\rm E}_8}/\Lambda,\mathbb{Z}/m\mathbb{Z})$,
in order to construct a Halphen
fibration of index $m$ whose configuration of
reducible fibers is of type $\tilde\Lambda$
we choose a characteristic
sequence $[p_1,\dots,p_9]$ of points on a smooth
cubic curve $C$ which satisfies the following conditions.
If $\xi$ is the following abelian extension
\[
\xymatrix@1{
0\ar[r] & \mathbb{Z}/m\mathbb{Z}\ar[r] & G\ar[r]^-\alpha & {{\rm E}_8}/\Lambda\ar[r] & 0
}
\]
we realize $G$ as a subgroup of the elliptic curve $(C,p_1)$
generated by nine points $p_1,\dots,p_9\in C$
such that the element $h=p_1+\dots+p_9$ of order $m$
generates the kernel of $\alpha$ and
$[p_1,\dots,p_9]\pmod{\langle h \rangle}\equiv
[\bar p_1,\dots,\bar p_9]$, where the second sequence
is the characteristic sequence
in Table~\ref{tab} associated to $\Lambda$.
Given such a characteristic sequence we reconstruct
the set of $(-2)$-curves of the blow-up $X$ of $\mathbb{P}^2$
at the nine points in the following way.
By Lemma~\ref{effective} each root $R$ of $\Lambda$
lifts to the class of an effective curve of $X$. The set $\mathcal S$
of such lifted classes spans a cone in $\operatorname{Pic}_\mathbb{Q}(X)$.
Fix an ample class $A$ of $\operatorname{Pic}(X)$ and form
the subset $\mathcal S_1$ of $\mathcal S$ consisting
of elements having minimal intersection with $A$.
Then construct the set $\mathcal S_n$ inductively
as follows.
Let $\mathcal S_n$ be the subset of
$\mathcal S \setminus \bigcup_{i \leq n-1}\mathcal S_i$
consisting of elements having minimal
intersection with $A$ and non-negative
intersection with any element of
$\bigcup_{i \leq n-1}\mathcal S_i$.
It follows that the set of $(-2)$-curves of $X$ is
\[
\mathcal C =\bigcup_{i}\mathcal S_i.
\]
According to Lemma~\ref{lifting} the classes
of these curves span a root sublattice of ${K_X}^\perp
\cong{\rm \widetilde {E}}_8$ isometric to $\tilde\Lambda$.
By means of these curves
we define the lifting $\operatorname{cl}_X
\colon F\to{K_X}^\perp$ of the map
$P\colon F\to{K_X}^\perp/\langle K_X\rangle={{\rm E}_8}$.
We thus get the class $\delta_X$ of the corresponding
toric variety, whose Riemann-Roch dimension
is the number of $(-1)$-curves of $X$.
Now assume $m = 2$.
Computing the multigraded Hilbert
series of the toric varieties given in the appendix
and excluding the cases where the trivial
extension of ${{\rm E}_8}/\Lambda$ by $\mathbb{Z}/2\mathbb{Z}$
needs more than two generators, one directly checks
that all the remaining cases are realized.
\end{proof}
\section*{Appendix}
We provide here the matrix $Q$ defining
the linear map $E\to\operatorname{Cl}(Z)$ for each type of
lattice analysed in this note.
{\footnotesize
\begin{longtable}{lll}
\label{gradings}
Type & Class group & Grading matrix\\
\\
\hline
\\
$E_8$
&
$\mathbb{Z}$
&
$\begin{bmatrix} 1&2&3&4&5&6&4&3&2 \end{bmatrix}$\\
\\
$D_8$
&
$\mathbb{Z}\oplus\mathbb{Z}/2\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&2&2&2&2&2\\
\bar 0&\bar 0&\bar 1&\bar 1&\bar 0&\bar 0&\bar 0&\bar 1&\bar 1
\end{bmatrix}
$\\
\\
$A_8$
&
$\mathbb{Z}\oplus\mathbb{Z}/3\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&1&1&1&1&1\\
\bar 0&\bar 1&\bar 2&\bar 0&\bar 1&\bar 2&\bar 0&\bar 1&\bar 2
\end{bmatrix}
$\\
\\
$E_7+A_1$
&
$\mathbb{Z}^2\oplus\mathbb{Z}/2\mathbb{Z}$
&
$
\begin{bmatrix}
1&2&3&4&3&2&1&2&0&0\\
0&0&0&0&0&0&0&0&1&1\\
\bar 0&\bar 0 &\bar 0&\bar 0 &\bar 1 &\bar 1 &\bar 1 &\bar 0 &\bar 0 &\bar 1
\end{bmatrix}
$\\
\\
$E_6+A_2$
&
$\mathbb{Z}^2\oplus\mathbb{Z}/3\mathbb{Z}$
&
$
\begin{bmatrix}
1&2&3&2&1&2&1&0&0&0\\
0&0&0&0&0&0&0&1&1&1\\
\bar 0& \bar 0& \bar 0& \bar 1& \bar 1& \bar 2& \bar 1& \bar 0& \bar 1& \bar 2
\end{bmatrix}
$\\
\\
$A_7+A_1$
&
$\mathbb{Z}^2\oplus\mathbb{Z}/4\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&1&1&1&1&0&0\\
0&0&0&0&0&0&0&0&1&1\\
\bar 0& \bar 1& \bar 2& \bar 3& \bar 0& \bar 1& \bar 2& \bar 3& \bar 0& \bar 2
\end{bmatrix}
$\\
\\
$D_5+A_3$
&
$\mathbb{Z}^2\oplus\mathbb{Z}/4\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&2&2&0&0&0&0\\
0&0&0&0&0&0&1&1&1&1\\
\bar 0& \bar 1& \bar 2& \bar 3& \bar 0& \bar 2& \bar 0& \bar 1& \bar 2& \bar 3
\end{bmatrix}
$\\
\\
$2A_4$
&
$\mathbb{Z}^2\oplus\mathbb{Z}/5\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&1&0&0&0&0&0\\
0&0&0&0&0&1&1&1&1&1\\
\bar 0& \bar 1& \bar 2& \bar 3& \bar 4& \bar 0& \bar 1& \bar 2& \bar 3& \bar 4
\end{bmatrix}
$\\
\\
$2D_4$
&
$\mathbb{Z}^2\oplus(\mathbb{Z}/2\mathbb{Z})^2$
&
$
\begin{bmatrix}
1&1&1&1&2&0&0&0&0&0\\
0&0&0&0&0&1&1&1&1&2\\
\bar 0& \bar 1& \bar 0& \bar 1& \bar 0& \bar 0& \bar 1& \bar 0& \bar 1& \bar 0\\
\bar 0& \bar 0& \bar 1& \bar 1& \bar 0& \bar 0& \bar 0& \bar 1& \bar 1& \bar 0
\end{bmatrix}
$\\
\\
$A_5+A_2+A_1$
&
$\mathbb{Z}^3\oplus\mathbb{Z}/6\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&1&1&0&0&0&0&0\\
0&0&0&0&0&0&1&1&1&0&0\\
0&0&0&0&0&0&0&0&0&1&1\\
\bar 0& \bar 1& \bar 2& \bar 3& \bar 4& \bar 5& \bar 0& \bar 2& \bar 4& \bar 0& \bar 3
\end{bmatrix}
$\\
\\
$D_6+2A_1$
&
$\mathbb{Z}^3\oplus(\mathbb{Z}/2\mathbb{Z})^2$
&
$
\begin{bmatrix}
1&1&1&1&2&2&2&0&0&0&0\\
0&0&0&0&0&0&0&1&1&0&0\\
0&0&0&0&0&0&0&0&0&1&1\\
\bar 0& \bar 1& \bar 0& \bar 1& \bar 0& \bar 0& \bar 1& \bar 0& \bar 1& \bar 0& \bar 0\\
\bar 0& \bar 0& \bar 1& \bar 1& \bar 0& \bar 0& \bar 0& \bar 0& \bar 1& \bar 0& \bar 1
\end{bmatrix}
$\\
\\
$2A_3+2A_1$
&
$\mathbb{Z}^4\oplus\mathbb{Z}/2\mathbb{Z}\oplus\mathbb{Z}/4\mathbb{Z}$
&
$
\begin{bmatrix}
1&1&1&1&0&0&0&0&0&0&0&0\\
0&0&0&0&1&1&1&1&0&0&0&0\\
0&0&0&0&0&0&0&0&1&1&0&0\\
0&0&0&0&0&0&0&0&0&0&1&1\\
\bar 0& \bar 1& \bar 2& \bar 3& \bar 0& \bar 1& \bar 2& \bar 3& \bar 0& \bar 2& \bar 0& \bar 0\\
\bar 0& \bar 1& \bar 0& \bar 1& \bar 0& \bar 0& \bar 0& \bar 0& \bar 0& \bar 1& \bar 0& \bar 1
\end{bmatrix}
$\\
\\
$4A_2$
&
$\mathbb{Z}^4\oplus(\mathbb{Z}/3\mathbb{Z})^2$
&
$
\begin{bmatrix}
1&1&1&0&0&0&0&0&0&0&0&0\\
0&0&0&1&1&1&0&0&0&0&0&0\\
0&0&0&1&0&0&1&1&1&0&0&0\\
0&0&0&0&0&0&0&0&0&1&1&1\\
\bar 0& \bar 1& \bar 2& \bar 0& \bar 0& \bar 0& \bar 0& \bar 1& \bar 2& \bar 0& \bar 1& \bar 2\\
\bar 0& \bar 0& \bar 0& \bar 0& \bar 1& \bar 2& \bar 0& \bar 1& \bar 2& \bar 0& \bar 2& \bar 1
\end{bmatrix}
$\\[30pt]
\caption{Grading matrices}
\end{longtable}
}
\begin{bibdiv}
\begin{biblist}
\bib{AL}{article} {
author={Artebani, Michela},
author={Laface, Antonio},
title={Cox rings of surfaces and the anticanonical Iitaka dimension},
journal = {Advances in Mathematics},
volume = {226},
number = {6},
pages = {5252 - 5267},
year = {2011}
}
\bib{ADHL}{book}{
AUTHOR = {Arzhantsev, Ivan},
AUTHOR = {Derenthal, Ulrich},
AUTHOR = {Hausen, J\"urgen},
AUTHOR = {Laface, Antonio},
TITLE = {Cox rings},
series={Cambridge Studies in Advanced Mathematics},
volume={144},
publisher={Cambridge University Press, Cambridge},
date={2014},
pages={530},
isbn={9781107024625},
}
\bib{BHPV}{book}{
author={Barth, Wolf P.},
author={Hulek, Klaus},
author={Peters, Chris A. M.},
author={Van de Ven, Antonius},
title={Compact complex surfaces},
series={Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics [Results in Mathematics and
Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]},
volume={4},
edition={2},
publisher={Springer-Verlag, Berlin},
date={2004},
pages={xii+436},
isbn={3-540-00832-2},
review={\MR{2030225 (2004m:14070)}},
}
\bib{DoCo}{book}{
author={Cossec, Fran{\c{c}}ois R.},
author={Dolgachev, Igor V.},
title={Enriques surfaces. I},
series={Progress in Mathematics},
volume={76},
publisher={Birkh\"auser Boston Inc.},
place={Boston, MA},
date={1989},
pages={x+397},
isbn={0-8176-3417-7},
review={\MR{986969 (90h:14052)}},
doi={10.1007/978-1-4612-3696-2},
}
\bib{Har}{article}{
author={Harbourne, Brian},
title={Anticanonical rational surfaces},
journal={Trans. Amer. Math. Soc.},
volume={349},
date={1997},
number={3},
pages={1191--1208},
issn={0002-9947},
review={\MR{1373636 (97f:14007)}},
doi={10.1090/S0002-9947-97-01722-4},
}
\bib{LH}{article}{
author={Lahyane, Mustapha},
author={Harbourne, Brian},
title={Irreducibility of $-1$-classes on anticanonical rational surfaces
and finite generation of the effective monoid},
journal={Pacific J. Math.},
volume={218},
date={2005},
number={1},
pages={101--114},
issn={0030-8730},
review={\MR{2224591 (2007b:14082)}},
doi={10.2140/pjm.2005.218.101},
}
\bib{MiPe}{article}{
author={Miranda, Rick},
author={Persson, Ulf},
title={On extremal rational elliptic surfaces},
journal={Math. Z.},
volume={193},
date={1986},
number={4},
pages={537--558},
issn={0025-5874},
review={\MR{867347 (88a:14044)}},
doi={10.1007/BF01160474},
}
\bib{MiSt}{book}{
author={Miller, Ezra},
author={Sturmfels, Bernd},
title={Combinatorial commutative algebra},
series={Graduate Texts in Mathematics},
volume={227},
publisher={Springer-Verlag},
place={New York},
date={2005},
pages={xiv+417},
isbn={0-387-22356-8},
review={\MR{2110098 (2006d:13001)}},
}
\bib{OgSh}{article}{
author={Oguiso, Keiji},
author={Shioda, Tetsuji},
title={The Mordell-Weil lattice of a rational elliptic surface},
journal={Comment. Math. Univ. St. Paul.},
volume={40},
date={1991},
number={1},
pages={83--99},
issn={0010-258X},
review={\MR{1104782 (92g:14036)}},
}
\bib{To}{article}{
AUTHOR = {Totaro, Burt},
TITLE = {The cone conjecture for {C}alabi-{Y}au pairs in dimension 2},
JOURNAL = {Duke Math. J.},
VOLUME = {154},
YEAR = {2010},
NUMBER = {2},
PAGES = {241--263}
}
\end{biblist}
\end{bibdiv}
\end{document} |
\subsection*{A1. Proof of Theorem~\ref{thm:exp_av_degree}}
\begin{proof} The proof is based on the fact that the total number of edges $|E_{\tau_n}|$ equals a total number of the $T2$ and $T3$ arrivals on $(0,\tau_n]$. By definition, $(0,\tau_n]$ contains exactly $(n-1)$ of $T1$ or $T2$ arrivals, hence, the number of $T2$ arrivals has a Binomial distribution with number of trials equal to $(n-1)$, and success probability $P(T2\,|\mbox{ $T1$ or $T2$})=\frac{p}{\lambda+p}$.
Next, the number of $T3$ arrivals on $[\tau_i,\tau_{i+1})$, where $i=1,\ldots,n-1$, has a shifted geometric distribution, namely, the probability of $k$ $T3$ arrivals on $[\tau_i,\tau_{i+1})$ is \[\left(1-\frac{1-p}{\lambda+1}\right)\left(\frac{1-p}{\lambda+1}\right)^k,\; k=0,1,\ldots.\] Observe that there have been $n-1$ of these transitions from 1 node to $n$. Hence,
the number of $T3$ arrivals on $(0,\tau_n]$ is the sum of $(n-1)$ i.i.d. Geometric random variables with mean $\frac{1-p}{\lambda+p}$. Summarizing the above, we obtain (\ref{eq:exp_av_degree}). For (\ref{eq:var_av_degree}) we also need to observe that the number of $T2$ and $T3$ arrivals on $[0,\tau_n]$ are independent.
\end{proof}
\subsection*{A2. Proof of Theorem~\ref{thm:exp_edges_discussion}}
\begin{proof}
Let $X_t$ be the number of $T2$ and $T3$ arrivals by time $t$. Note that $|E_t|=X_t$, and $|{\cal T}_t|=t-X_t+1$, which is the number of $T_1$ arrivals on $[0,t]$, since the first node at time $t=0$ is by definition a $T1$ arrival. Note that $X_t$ has a binomial distribution with parameters $t$ and $\prob{\mbox{$T2$ arrival}}+\prob{\mbox{$T3$ arrival}}=\frac{1}{\lambda+1}$. Furthermore, the number of $T1$ arrivals is $t-X_t+1$ since the first node at time $t=0$ is by definition a $T1$ arrival. Hence,
\begin{eqnarray*}
\E{\frac{|E_t|}{|{\cal T}_t|}} & = & \sum_{i=1}^t\frac{i}{t-i+1}{t\choose i}\left(\frac{1}{\lambda+1}\right)^i\left(\frac{\lambda}{\lambda+1}\right)^{t-i}\\
& = & \frac{1}{\lambda}\cdot\sum_{i=1}^t{t \choose i-1}\left(\frac{1}{\lambda+1}\right)^{i-1}\left(\frac{\lambda}{\lambda+1}\right)^{t-i+1},\\
\end{eqnarray*}
which proves (\ref{eq:exp_edges_discussion}). Next, we write
\begin{eqnarray}
&&\E{\left(\frac{|E_t|}{|{\cal T}_t|}\right)^2}
=\sum_{i=0}^t\left(\frac{i}{t-i+1}\right)^2{t\choose i}\left(\frac{1}{\lambda+1}\right)^i\left(\frac{\lambda}{\lambda+1}\right)^{t-i}\nonumber\\
&&\quad =\frac{1}{\lambda}\cdot\sum_{i=1}^t\frac{i}{t-i+1}{t\choose i-1}\left(\frac{1}{\lambda+1}\right)^{i-1}\left(\frac{\lambda}{\lambda+1}\right)^{t-i+1}\nonumber\\
&&=\frac{1}{\lambda}\E{\frac{t+1}{t-X_t}\mathbbm{1}_{\{X_t\le t-1\}}}-\frac{1}{\lambda}\left(1-\left(\frac{1}{1+\lambda}\right)^t\right),
\label{eq:exp_edges_discussion2}
\end{eqnarray}
where $\mathbbm{1}_{\{A\}}$ is an indicator of event $A$. Denoting
\begin{equation}
\label{eq:z0}Z_t=\frac{X_t-\E{X_t}}{\sqrt{\var{X_t}}}=\frac{(\lambda+1)X_t-t}{\sqrt{\lambda t}},\end{equation}
we further write
\begin{eqnarray}
\label{eq:z1}
&&\E{\frac{t+1}{t-{X_t}}\mathbbm{1}_{\{X_t\le t-1\}}}=\E{\frac{(t+1)(\lambda+1)}{\lambda t(1-\frac{Z_t}{\sqrt{\lambda t}})}\mathbbm{1}_{\{Z_t\le \sqrt{\lambda }t-\frac{\lambda+1}{\sqrt{\lambda t}}\}}}.
\end{eqnarray}
We now split the indicator above as follows:
\begin{equation}
\label{eq:z2}
\mathbbm{1}_{\{Z_t\le -\sqrt{\lambda }t\}}+ \mathbbm{1}_{\{-\sqrt{\lambda }t< Z_t<\sqrt{\lambda }t/2\}}+\mathbbm{1}_{\{\sqrt{\lambda }t/2\le Z_t\le \sqrt{\lambda }t-\frac{\lambda+1}{\sqrt{\lambda t}}\}}.
\end{equation}
For the first and the third term we use the Chernoff bound:
\begin{eqnarray}
\label{eq:z3} \E{\frac{1}{1-\frac{Z_t}{\sqrt{\lambda t}}}\mathbbm{1}_{\{Z_t\le -\sqrt{\lambda }t\}}}&\le& 2e^{-\lambda t/4},\\
\E{\frac{1}{1-\frac{Z_t}{\sqrt{\lambda t}}}\mathbbm{1}_{\{\sqrt{\lambda }t/2\le Z_t\le \sqrt{\lambda }t-\frac{\lambda+1}{\sqrt{\lambda t}}\}}}&\le& \frac{\sqrt{\lambda t}}{\lambda+1}2e^{-\lambda t/16},
\label{eq:z4}
\end{eqnarray}
and notice that both expressions above converge to zero faster than $1/t$. For the second case, note first that $\E{Z_t}=0$ and hence it follows from (\ref{eq:z0}) and (\ref{eq:z2})--(\ref{eq:z4}) that, as $t\to\infty$,
\[\E{Z_t\mathbbm{1}_{\{-\sqrt{\lambda }t< Z_t<\sqrt{\lambda }t/2\}}}=o\left(\frac{1}{t}\right).\]
Then we use the Taylor expansion to obtain:
\begin{eqnarray}
&&\left|\E{\frac{1}{1-\frac{Z_t}{\sqrt{\lambda t}}}\mathbbm{1}_{\{-\sqrt{\lambda }t< Z_t<\sqrt{\lambda }t/2\}}}-1\right|\nonumber\\
&&\hspace{2cm} \le \E{\frac{Z_t^2}{\lambda t}}+2\E{\frac{|Z_t|^3}{(\lambda t)^{3/2}}}+o\left(\frac{1}{t}\right),\label{eq:var_estimate}
\end{eqnarray}
as $t\to\infty$. By the central limit theorem, ${Z_t}\overset{D}{\longrightarrow}Z$ as $t\to\infty$.
Furthermore, for $r>0$, the convergence of moments holds~\cite{hall1983rate}:
$\lim_{t\to\infty}\E{|{Z_t}|^r}=\E{|Z|^r}$. In particular, in (\ref{eq:var_estimate}), $\E{|{Z_t}|^3}$ converges to a constant, and
$\E{{Z_t}^2}$ converges to 1 as $t\to\infty$. Thus, using (\ref{eq:exp_edges_discussion2})--(\ref{eq:z1}) and (\ref{eq:exp_edges_discussion}) we write
\begin{eqnarray*}
&&\var{\frac{|E_t|}{|{\cal T}_t|}}=\E{\left(\frac{|E_t|}{|{\cal T}_t|}\right)^2} -\left(\E{\frac{|E_t|}{|{\cal T}_t|}}\right)^2\\
&&=\E{\frac{(t+1)(\lambda+1)}{\lambda t(1-\frac{Z_t}{\sqrt{\lambda t}})}\mathbbm{1}_{\{Z_t\le \sqrt{\lambda }t-\frac{\lambda+1}{\sqrt{\lambda t}}\}}}-\frac{1}{\lambda}-\frac{1}{\lambda^2} +o\left(\frac{1}{t}\right).
\end{eqnarray*}
Now, subsequently using (\ref{eq:z2}) -- (\ref{eq:var_estimate}), we get
\begin{eqnarray*}
\var{\frac{|E_t|}{|{\cal T}_t|}}&=&
\frac{1}{\lambda}\,\frac{(t+1)(\lambda+1)}{\lambda t}\left(1+\frac{1}{\lambda t}+o\left(\frac{1}{t}\right)\right)\\
&&-\frac{1}{\lambda} - \frac{1}{\lambda^2} +o\left(\frac{1}{t}\right),
\end{eqnarray*}
which results in (\ref{eq:var_edges_discussion}). Statement (\ref{eq:distr_edges_discussion}) is proved along similar lines: we apply the expansion directly to the random variable
\[\frac{X_t}{t-X_t+1}=\frac{(t+1)(\lambda+1)}{(\lambda t+\lambda+1)(1-Z_t\frac{\sqrt{\lambda t}}{\lambda t+\lambda+1})}\mathbbm{1}_{\{Z_t\le \sqrt{\lambda t}\}}-1, \]
and then use the Chernoff bounds and the CLT to obtain the result.
\end{proof}
\subsection*{A3. Proof of Lemma~\ref{lem:distribution_component_sizes_one_step}}
\begin{proof}
Assume the arrival at time $t+1$ is of type $T1$. This occurs w.p. $\frac{\lambda}{\lambda+1}$, and then a new component consisting of size one is created in $G_{t+1}$, corresponding to the first case in (\ref{lem:distribution_component_sizes_one_step}).
Next, consider a $T2$ arrival, which occurs w.p. $\frac{p}{\lambda+1}$. We now add a node to an existing component $C_i$ w.p. $\frac{|C_i|}{|V|}$. Thus the probability that we add the new node to $C_i$ is $\frac{p}{\lambda+1}\cdot\frac{|C_i|}{|V|}$.
Last, we consider a $T3$ arrival. In this case we have two options. The new edge can either join two components, or join two nodes that are already in one component. For the first case, we derive the probability that $C_i$ and $C_j$ join as
\[\begin{aligned}
&\mathbb{P}\left(C_i\mbox{ and }C_j\mbox{ merge}\right)=\frac{1-p}{\lambda+1}\cdot\frac{2\cdot|C_i|\cdot|C_j|}{|V|^2-|V|}. \end{aligned}\]
Then for the second case, the number of ways a T3 arrival links two nodes that are already connected in a component, say $C_i$, is $|C_i|\left(|C_i|-1\right)$. Therefore with probability $\frac{\sum_{k=1}^m|C_k|^2-|C_k|}{|V|^2-|V|}$ the component size does not change.
\end{proof}
\section{Retweet graphs}
\label{sec:retweet graphs}
Our main object of study is the retweet graph $G=(V,E)$, which is a graph of users that have participated in the discussion on a specific topic. A directed edge $e=(u,v)$ indicates that user $v$ has retweeted a tweet of $u$. We observe the retweet graph at the time instances $t=0,1,2,\ldots$, where either a new node or a new edge was added to the graph, and we denote by $G_t=(V_t,E_t)$ the retweet graph at time $t$. As usual, the out- (in-) degree of node $u$ is the number of directed edges with source (destination) in $u$. In what follows, we model and analyse the properties of $G_t$. For every new message initiated by a new user $u$ a tree $T_{u}$ is formed. Then, ${\cal T}_t$ denotes the forest of message trees. Note that in our model a new message from an already existing user $u$ (that is, $u\in {\cal T}_t$) does not initiate a new message tree. We define $|{\cal T}_t|$ as the number of new users that have started a message tree up to time $t$.
After analyzing multiple characteristics of the retweet graphs for every hour of their progression, we found that the size of the largest (weakly) connected component (LCC) and its density are the most informative characteristics for predicting the peak in {\it Twitter}. In Figure~\ref{fig:PX} we show the development of these characteristics in the \emph{Project X} dataset. One day before the actual event, we observe a very interesting phenomenon in the development of the edge density of the LCC in Figure~\ref{fig:prog_edges_nodes}. Namely, at some point the edge density of the LCC exceeds $1$ (indicated by the dash-dotted gray lines), i.e. there is more than one retweet per user on average. We shall refer to this as the \emph{densification} (or dens.) of the LCC. Furthermore, the relative size of the LCC increases from 18\% to 25\% as well, see Figure \ref{fig:prog_gc}.
\begin{figure}[!ht]
\vspace{-25pt}
\centering
\subfloat[Edge density.\label{fig:prog_edges_nodes}]{
\includegraphics[height=0.5\textwidth,angle=270]{PX_progression_edges_nodes.pdf}}
\subfloat[Size of LCC.\label{fig:prog_gc}]{
\includegraphics[height=0.5\textwidth,angle=270]{PX_progression_gc.pdf}}
\caption{Progression for the edge density (a) and the size of the LCC (b) in the \emph{Project X} dataset.\label{fig:PX}}
\end{figure}
We have observed a densification of the LCC in each dataset that we have studied. Indeed, when the LCC grows its density must become at least one (each node is added to the LCC together with at least one edge). However, we have also observed that in each dataset the densification occurs before the main peak, but the scale of densification is different. For example, in the {\it Project X} dataset the densification already occurs one day before the peak activity. Plausibly, in this discussion, that ended up in riots, a group of people was actively participating before the event. On the other hand, in the {\it WCS} dataset, which tweets about an ongoing sport event, the densication of the LCC occurs during the largest peak. This is the third peak in the progression. Hence, our experiments suggest that the time of densification has predictive value for trend progression and classification. See Table~\ref{tab:param_values} for the density of the LCC in each dataset at the end of the progression.
\subsection{Growth of the graph}
The average degree, or edge density, is one of the aspects through which we give insight to the growth of the graph. The essential properties of this characteristic are presented in Theorem~\ref{thm:exp_av_degree}. The proof is given in the Appendix.
\begin{thm} Let $\tau_n$ be the time when node $n$ is added to the graph. Then
\begin{eqnarray}
\E{\frac{|E_{\tau_n}|}{|V_{\tau_n}|}} &=& \frac{1}{\lambda+p}-\frac{1}{n(\lambda+p)},\label{eq:exp_av_degree}
\\
\var{\frac{|E_{\tau_n}|}{|V_{\tau_n}|}} &=& \frac{(n-1)(\lambda +1- p)}{n^2(\lambda+p)^2}.
\label{eq:var_av_degree}
\end{eqnarray}
\label{thm:exp_av_degree}
\end{thm}
Note that the variance of the average degree in (\ref{eq:var_av_degree}) converges to zero as $n\to \infty$ at rate $\frac{1}{n}$.
The next theorem studies the observed ratio between $T2$ and $T3$ arrivals (new edges) and $T1$ arrivals (new nodes with a new message). As we see from the theorem, this ratio can be used for estimating the parameter $\lambda$. The proof is given in the Appendix.
\begin{thm} Let $G_t=(V_t,E_t)$ be the retweet graph at time $t$, let ${\cal T}_t$ be the set of all message trees in $G_t$. Then
\begin{eqnarray}
&&\E{\frac{|E_t|}{|{\cal T}_t|}} = \lambda^{-1}\cdot\left(1-\left(\frac{1}{\lambda+1}\right)^t\right),\label{eq:exp_edges_discussion}
\\
&&\lim_{t\to\infty}\frac{\lambda^3 t}{(1+\lambda)^2}\var{\frac{|E_t|}{|{\cal T}_t|}} =1,
\label{eq:var_edges_discussion}
\end{eqnarray}
Furthermore,
\begin{equation}
\label{eq:distr_edges_discussion}
\frac{\lambda^{3/2}\sqrt{t}}{\lambda+1}\left(\frac{|E_t|}{|{\cal T}_t|}-\frac{1}{\lambda}\right)\stackrel{D}{\to} Z,
\end{equation}
where $Z$ is a standard normal $ N(0,1)$ random variable, and $\stackrel{D}{\to}$ denotes convergence in distribution.
\label{thm:exp_edges_discussion}
\end{thm}
Note that, as expected from the definition of $\lambda$,
\begin{equation}
\lim_{t\rightarrow\infty}\E{\frac{|E_t|}{|{\cal T}_t|}} = \lambda^{-1}.\label{eq:lim_exp_edges_discussion}
\end{equation}
This will be used in Section \ref{sec:practice} for estimating $\lambda$.
\subsection{Component size distribution}
{\sloppy In the following, we assume that $G_t$ consists of $m$ connected components ($C_1,C_2,\ldots,C_m$) with known respective sizes ($|C_1|,\ldots,|C_m|$). We aim to derive expressions for the distribution of the component sizes in $G_{t+1}$.
}
\begin{lem} The distribution of the sizes of the components of $G_{t+1}$, given $G_t$ is as follows,
\begin{equation}
\begin{cases}
|C_1|,\ldots,|C_i|,|C_j|,\ldots,|C_m|,1 & \mbox{w.p. }\frac{\lambda}{\lambda+1}\\
|C_1|,\ldots,|C_i|+1,|C_j|,\ldots,|C_m| & \mbox{w.p. }\frac{p}{\lambda+1}\cdot\frac{|C_i|}{|V|}\\
|C_1|,\ldots,|C_i|+|C_j|,\ldots,|C_m| & \mbox{w.p. }\frac{1-p}{\lambda+1}\cdot\frac{2\cdot|C_i|\cdot|C_j|}{|V|^2-|V|}\\
|C_1|,\ldots,|C_i|,|C_j|,\ldots,|C_m| & \mbox{w.p. }\frac{1-p}{\lambda+1}\cdot\frac{\sum_{k=1}^m|C_k|^2-|C_k|}{|V|^2-|V|}
\end{cases}
\end{equation}
\label{lem:distribution_component_sizes_one_step}\end{lem}
The proof of Lemma~\ref{lem:distribution_component_sizes_one_step} is given in the Appendix. Lemma \ref{lem:distribution_component_sizes_one_step} provides a recursion for computing the distribution of component sizes. However, the computations are highly demanding if not infeasible. Also, deriving an exact expression of the distribution of the component sizes at time $t$ is very cumbersome because they are hard and they strongly depend on the events that occurred at $t=0,\ldots,t-1$. Note that if $p=1$, there is a direct correspondence between our model and the infinite generalized P\'{o}lya process~\cite{chung2003generalizations}. However, this case is uninformative as there are no $T3$ arrivals. Therefore, in the next section we resort to simulations to investigate the sensitivity of the graph characteristics to the model parameters.
\subsection{Influence of $q$, $p$ and $\lambda$}
We analyze the influence of the model parameters $\lambda$, $p$ and $q$ on the characteristics of the resulting graph numerically using simulations. To this end, we fix two out of three parameters and execute multiple simulation runs of the model, varying the values for the third parameter. We start simulations with graph $G_0$, consisting of one node. We perform $50$ simulation runs for every parameter setting and obtain the average values over the individual runs for given parameters.
Parameter $q$ affects the degree distribution~\cite{bhamidi2012twitter} and the overall structure of the graph. If $q=0$, then the graph contains less nodes that have many retweets. If $q=1$ each edge is connected to a superstar, and the graph consists of star-like sub graphs, some of which are connected to each other. In the {\it Project X} dataset, which is our main case study, $q\approx0.9$ results in a degree distribution that closely approximates the data. Since degree distributions are not in the scope of this paper, we omit these results for brevity.
We compare the results for two measures that produced especially important characteristics of the \emph{Project X} dataset: $\frac{|E_{\mbox{\tiny{LCC}}}|}{|V_{\mbox{\tiny{LCC}}}|}$ and $\frac{|V_{\mbox{\tiny{LCC}}}|}{|V|}$. These characteristics do not depend on $q$. In simulations, we set $t=1,{}000$, $q=0.9$ and vary the values for $p$ and $\lambda$. the results are give in Figure~\ref{fig:numerical}.
\begin{figure}[!ht]
\vspace{-25pt}
\subfloat[$\frac{|E_{\mbox{\tiny{LCC}}}|}{|V_{\mbox{\tiny{LCC}}}|}$\label{fig:E_over_V_in_GC_numerical}]{
\includegraphics[width=0.52\textwidth]{E_over_V_in_GC.pdf}}
\subfloat[$\frac{|V_{\mbox{\tiny{LCC}}}|}{|V|}$\label{fig:frac_in_GC_numerical}]{
\includegraphics[width=0.52\textwidth]{frac_in_GC.pdf}}
\caption{Numerical results for the model using $q=0.9$ and $t=1,{}000$\label{fig:numerical}.}
\vspace{-20pt}
\end{figure}
We see that the edge density in the LCC in Figure \ref{fig:E_over_V_in_GC_numerical} decreases with $\lambda$ and $p$. Note that according to (\ref{eq:exp_av_degree}), $|E|/|V|$ is well approximated by $1/(\lambda+p)$ when $\lambda$ or $p$ are large enough. The edge density in LCC shows a similar pattern, but it is slightly higher than in the whole graph. When $\lambda$ and $p$ are small, there are many $T3$ arrivals, and new nodes are not added frequently enough. This results in an unexpected non-monotonic behaviour of the edge density near the origin. For the fraction of nodes in the LCC, depicted in Figure \ref{fig:frac_in_GC_numerical}, we see that the parameter $\lambda$ is most influential. The parameter $p$ is of considerable influence only when it is large.
\section{Introduction}\label{sec:intro}
\input{intro}
\section{Related work}\label{sec:related}
\input{related}
\section{Datasets}\label{sec:data}
\input{data}
\section{Model}\label{sec:model}
\input{model_theory}
\section{The model in practice}\label{sec:practice}
\input{practice}
\section{Conclusion and Discussion}\label{sec:conclusion}
\input{conclusion_discussion}
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:intro}
Supermassive black holes (SMBHs) with mass $10^5{\rm{M}}_\odot\lesssim{M}\lesssim10^{10}{\rm{M}}_\odot$
are considered to reside at the center of most galaxies \citep{kr95}. Hitherto, SMBHs have been
found in 87 galaxies by observing the proper motion of stars bound by the SMBHs or by detecting
radiation emitted from gas pulled gravitationally by the SMBHs \citep{kho13}. $\rm{H}_2\rm{O}$
maser emission from active galactic nuclei (AGNs) in spiral galaxies provides a strong tool to measure
supermassive black hole (SMBH) masses, because it arises from a rotating disk on a subparsec scale
with a nearly Keplerian velocity distribution around a SMBH. Those maser disks have been observed
at the centers of NGC\,4258 \citep{miyoshi95}, NGC\,1068 \citep{green97}, NGC\,3079 \citep{yama04},
the Circinus galaxy \citep{green03}, UGC\,3789 \citep{reid09}, NGC\,6323 \citep{braatz07}, NGC\,2273,
NGC\,6264, and some more objects \citep{kuo11}. Several maser disks show warped structure at the radii
of $\sim0.1\,\rm{pc}$ \citep{green03,he05,kond08,kuo11}. From an observational point of view, maser spots
on the disk in NGC~4258 are spatially distributed along a line on each side of a central black hole. The SMBH
is then thought to be located at the center of a line connecting those two lines by a simple extrapolation, and
the disk starts to warp at the innermost maser spot. What mechanism makes the disk warped still remains an
open question.
Several scenarios have been proposed for explaining disk warping. \citet{pr96} showed that centrally illuminated
accretion disks are unstable to warping due to the reaction force of reradiated radiation. Such a radiation-driven
warping mechanism has also been applied to explain the disk warping in the context of X-ray binaries \citep{mb97,
wijers99,martin07,martin09}. If angular momentum vector of an accretion disk around a spinning black hole is
misaligned with the spin axis, differential Lense-Thirring torque due to the frame-dragging effect aligns the inner
part of the disk with the black-hole equatorial plane. Since the outer part of the disk retains its initial orientation,
the resultant disk is warped \citep{bp75}. This Bardeen-Petterson effect is also considered to be a plausible
mechanism for disk warping in maser disks \citep{caproni07}. Moreover, \citet{balex09} proposed that the warped
disk at the center of NGC 4258 is caused by the process of resonant relaxation, which is a rapid relaxation mechanism
to exchange angular momentum between the disk and the stars moving under the nearly spherical potential dominated
by the SMBH. These mechanisms have been mostly discussed based on the assumption that the central object
surrounded by the warped maser disk is a single SMBH.
The tight correlation between the mass of SMBHs and the mass or luminosity of the bulge of their host galaxies
(\citep{mtr98,geb00,fm00,mm13}; see also \citep{kho13} for a review) suggests that the SMBH at the center of
each galaxy should have evolved toward coalescence in a merged galaxy. If this is the case, a binary of SMBHs
on a parsec or subparsec scale should be formed in a merged galactic nucleus before two black holes finally
coalesce, yet no SMBH binaries have clearly been identified so far despite some claims (see \citep{ko06,po12} for
reviews and references therein).
A possible link between the presence of SMBH binaries and the warping of observed maser disks
has been studied by \citet{khetal14} (hereafter, H14a).
They have, for simplicity, assumed that the circumbinary disk is initially aligned with the binary orbital plane.
However, the angular momentum vector of the circumbinary disk does not always coincide with that of the binary orbital angular momentum, because the orientation of the circumbinary disk is primarily due to the angular momentum distribution of the gas supplied to the central region of AGNs. Therefore, the orientation of the circumbinary disk plane is expected to be distributed randomly with respect to the binary orbital plane.
The formation of such a misaligned circumbinary disk around SMBH binaries has been numerically examined by \citet{dun14}.
\citet{hsm13} have investigated the accretion process from misaligned circumbinary disks onto SMBHs in eccentric orbits.
Regarding the misaligned disk structure, the inner part of the circumbinary disk tends to align with the binary orbital plane,
because the tidal precession torque is stronger than the shear viscous torque in the vertical
direction, whereas the outer part tends to retain the original state because the tidal precession torque
is weaker than the vertical viscous torque. This is the origin of disk warping \citep{nix11,flp13,lf13}.
On the other hand, \citet{nix13} have recently proposed that the circumbinary disk is broken into mutually misaligned rings,
if the tidal precession torque is stronger than the local horizontal viscous torque in circular SMBH binaries.
However, little is known about the relationship between disk tearing and warping in eccentric SMBH binaries.
In this paper, we examine the tidally driven warping and tearing of misaligned
circumbinary disks around eccentric SMBH binaries.
In Section~\ref{sec:2}, we describe the tidal torques originating from a time-dependent binary
potential and derive the tidal warp radius of the misaligned circumbinary disk. In Section~\ref{sec:3},
we discuss a possibility that the observed warping of maser disks in several AGNs is caused by the
tidal effect of SMBH binaries. Finally, Section~\ref{sec:4} summarises our scenario.
\section{Tidally driven warping and tearing of a misaligned circumbinary disk}
\label{sec:2}
Let us consider the torques from the binary potential acting on the circumbinary disk,
which is misaligned with the binary orbital plane, surrounding two black holes
in a binary on a eccentric orbit. Figure~\ref{fig:schmatic} illustrates a schematic picture
of the setting of our model; binary black holes orbiting each other are surrounded
by a misaligned circumbinary disk.
The binary is put on the $x$-$y$ plane with its center of mass being at the origin
in the Cartesian coordinate.
The masses of the primary and secondary black holes are represented
by $M_1$ and $M_2$, respectively, and $M=M_1+M_2$. We put a circumbinary disk around the origin.
The unit vector of specific angular momentum of the disk is expressed
by (e.g. \citet{pr96})
\begin{eqnarray}
\mbox{\boldmath $l$}
=
\cos\gamma\sin\beta\hat{\boldmath{x}}
+
\sin\gamma\sin\beta
\hat{\boldmath{y}}
+
\cos\beta\hat{\boldmath{z}},
\label{eq:damvec}
\end{eqnarray}
where $\beta$ is the tilt angle between the circumbinary disk plane and the
binary orbital plane, and $\gamma$ is the azimuth of tilt.
Here, $\hat{\bmath{x}}$, $\hat{\bmath{y}}$, and $\hat{\bmath{z}}$ are unit
vectors in the $x$, $y$, and $z$, respectively.
The position vector of the disk can be expressed by
\begin{eqnarray}
\mbox{\boldmath $r$}
&=&
r(\cos\phi\sin\gamma+\sin\phi\cos\gamma\cos\beta)\hat{\bmath{x}}
\nonumber \\
&+&
r(\sin\phi\sin\gamma\cos\beta-\cos\phi\cos\gamma)\hat{\bmath{y}}
\nonumber \\
&-&
r\sin\phi\sin\beta\hat{\bmath{z}}
\label{eq:rin}
\end{eqnarray}
where the azimuthal angle $\phi$ is measured from the descending node.
The difference from equation~(3) of Paper~I is the position vector of each black hole,
which is given by
\begin{equation}
\mbox{\boldmath $r$}_{i}=r_{i}\cos{f_i}\hat{\bmath{x}}+r_{i}\sin{f_i}\hat{\bmath{y}} \hspace{2mm}(i=1,2),
\label{eq:ri}
\end{equation}
where $f_2=f_1+\pi$ is the true anomaly and $r_{i}$ is written as
\begin{equation}
r_{i}=\xi_{i}\frac{a(1-e^2)}{1+e\cos{f_i}}
\label{eq:rphi}
\end{equation}
with $\xi_1\equiv q/(1+q)$ and $\xi_2\equiv 1/(1+q)$.
Here, $e$ is the binary orbital eccentricity, $q=M_2/M_1$ is the binary mass
ratio, and $a=a_1 + a_2$ is the binary semi-major axis with $a_1\equiv\xi_1a$
and $a_2\equiv\xi_2a$. These and other model parameters are listed in Table~1.
We also assume that the disk total angular momentum is small compared with the
binary angular momentum.
\begin{table*}
\caption{Model parameters}
\begin{tabular}{ll}
\hline
Definition & Symbol\\
\hline
Total black hole mass & $M$ \\
Primary black hole mass & $M_1$ \\
Secondary black hole mass & $M_2$ \\
Schwarzschild radius & $r_{\rm{S}}=2GM/c^2$ \\
Binary mass ratio & $q=M_2/M_1$ \\
Mass ratio parameter{s} & $\xi_1=q/(1+q)$, $\xi_2=1/(1+q)$ \\
Binary semi-major axis & $a$ \\
Orbital eccentricity & $e$ \\
Orbital frequency & $\Omega_{\rm{orb}}=\sqrt{GM/a^3}$ \\
Orbital period & $P_{\rm{orb}}=2\pi/\Omega_{\rm{orb}}$ \\
True anomaly & $f_2=f_1+\pi$ \\
Tilt angle & $\beta$ \\
Azimuth of tilt & $\gamma$ \\
Azimuthal angle & $\phi$ \\
Shakura-Sunyaev viscosity parameter & $\alpha$ \\
Horizontal shear viscosity & $\nu_1$ \\
Vertical shear viscosity & $\nu_2$\\
Ratio of vertical to horizontal shear viscosities & $\eta=\nu_2/\nu_1$\\
Mass-to-energy conversion efficiency & $\epsilon$ \\
\hline
\end{tabular}
\label{tb:t1}
\end{table*}
\begin{figure}
\resizebox{\hsize}{!}{
\includegraphics{fig1.eps}
}
\caption{
Configuration of a SMBH binary on an eccentric orbit and a circumbinary disk surrounding it.
There are two angles ($\beta, \gamma$) which specify the orientation of the disk plane with
respect to the binary orbital plane ($x$-$y$ plane). The azimuthal angle ($\phi$) of an arbitrary
position on the disk is measured from the descending node.
}
\label{fig:schmatic}
\end{figure}
\subsection{Tidal torques acting on the misaligned circumbinary disk}
The gravitational force on the unit mass at position $\bmath{r}$
on the circumbinary disk can be written by
\begin{eqnarray}
\bmath{F}_{\rm{grav}}=-\sum_{i=1}^{2}\frac{GM_i}{|\bmath{r}-\bmath{r}_i|^3}(\bmath{r}-\bmath{r}_i)
\end{eqnarray}
The corresponding torque is given by
\begin{eqnarray}
\bmath{t}_{\rm{grav}}=\bmath{r}\times\bmath{F}_{\rm{grav}}=\sum_{i=1}^2\frac{GM_i}{|\bmath{r}-\bmath{r}_i|^3}(\bmath{r}\times\bmath{r}_i)
\end{eqnarray}
We consider the tidal warping/precession with timescales much longer than
local rotation period of the circumbinary disk. This allows us to use the torque
averaged in the azimuthal direction and over the orbital period:
\begin{eqnarray}
&&
\langle\bmath{T_{\rm{grav}}}\rangle
\approx
\frac{1}{4\pi^2}\int_0^{2\pi}
\int_0^{2\pi}
\bmath{t}_{\rm{grav}}
\,d\phi
d(\Omega_{\rm{orb}}t)
\nonumber \\
&&
=
\frac{1}{2\pi}\int_0^{2\pi}
\left[\frac{1}{2\pi}\int_0^{2\pi}
\sum_{i=1}^2\frac{GM_i(\bmath{r}\times\bmath{r}_i)}{|\bmath{r}-\bmath{r}_i|^3}
d\phi\right]
\frac{(1-e^2)^{3/2}}{(1+e\cos{f}_i)^2}df
\nonumber \\
&=&
\frac{3}{8}\xi_1\xi_2
\frac{GM}{r}
\left(\frac{a}{r}\right)^2
\Biggr[
(1-e^2)
\sin\gamma\sin2\beta\hat{\bmath{x}}
-
(1+4e^2)
\cos\gamma\sin2\beta\hat{\bmath{y}}
\nonumber \\
&+&
5e^2\sin2\gamma\sin^2\beta\hat{\bmath{z}}
\Biggr],
\label{eq:tgrav}
\end{eqnarray}
which is equivalent to equation~(7) of \citet{khetal14b},
where $\Omega_{\rm{orb}}=\sqrt{GM/a^3}$ is the angular frequency of the mean binary motion.
Here, we used for integration the following relationship:
\begin{eqnarray}
d(\Omega_{\rm{orb}}{t})=\frac{(1-e^2)^{3/2}}{(1+e\cos{f}_i)^{2}}df
\end{eqnarray}
and the approximations:
\begin{eqnarray}
|\bmath{r}-\bmath{r}_i|^{-3}&\approx&r^{-3}\left[1+3\frac{\bmath{r}\cdot\bmath{r}_i}{r^2}
+\mathcal{O}((r_i/r)^2)\right]
\label{eq:expan1}
\\
(1+e\cos{f}_i)^{-4}&\approx&1-4e\cos{f}_i+10e^2\cos^2{f}_i+\mathcal{O}(e^3),
\nonumber \\
\label{eq:expan2}
\end{eqnarray}
where $r\gg{a}$ is adopted for equations~(\ref{eq:expan1}).
The magnitude of the specific tidal torque is given by
\begin{eqnarray}
|\langle\bmath{T}_{\rm{grav}}\rangle|
&=&\frac{3}{8}\xi_1\xi_2
\frac{GM}{r}
\left(\frac{a}{r}\right)^2
\Theta(e,\beta,\gamma)
\label{eq:stgrav}
\end{eqnarray}
where $\Theta(e,\beta,\gamma)$ is a function given by
\begin{eqnarray}
\Theta(e,\beta,\gamma)
&
=
&
\Biggr
\{
4\biggr[
5e^2(2+3e^2)\cos^2\gamma
+
(e^2-1)^2
\biggr]
\cos^2\beta
\nonumber \\
&
+
&
25e^4\sin^22\gamma\sin^2\beta
\Biggr\}^{1/2}\sin\beta.
\label{eq:anglef}
\end{eqnarray}
Figure~2 shows dependence of $\Theta(e,\beta,\gamma)$ on
the three parameters: orbital eccentricity $e$, tilt angle $\beta$,
and azimuth of tilt $\gamma$.
It is noted from the figure that there are several zeros in
$\Theta(e,\beta,\gamma)$, e.g.,
for $\beta=0$ and $\pi$ with any $e$ and $\gamma$.
Except for these values, $\Theta(e,\beta,\gamma)$ is of the order of one.
\begin{figure}
\resizebox{\hsize}{!}{
\includegraphics{fig2a.eps}
\includegraphics{fig2b.eps}
\includegraphics{fig2c.eps}
}
\caption{
Dependence of $\Theta(e,\beta,\gamma)$ on orbital eccentricity
$e$, tilt angle $\beta$, and azimuth of tilt $\gamma$.
Panel (a); dependence on $e$ with fixed values of $\beta$ and
$\gamma$. The solid black, red dashed, and blue dotted lines indicate $\Theta(e,\pi/4,\pi/4)$,
$\Theta(e,\pi/2,\pi/4)$, and $\Theta(e,\pi/4,\pi/2)$, respectively.
Panel (b); dependence on $\beta$ with fixed values of $e$ and $\gamma$.
The solid black, red dashed, and blue dotted lines indicate $\Theta(0.0,\beta,\pi/4)$,
$\Theta(0.6,\beta,\pi/4)$, and $\Theta(0.9,\beta,\pi/4)$, respectively.
Panel (c); dependence on $\gamma$ with fixed values of $e$ and $\beta$.
The solid black, red dashed, and blue dotted lines indicate $\Theta(0.0,\pi/4,\gamma)$,
$\Theta(0.6,\pi/4,\gamma)$, and $\Theta(0.9,\pi/4,\gamma)$, respectively.
}
\label{fig:timescale}
\end{figure}
By going through the same procedure as for equation~(\ref{eq:tgrav}), we obtain
the azimuthally-averaged and orbit-averaged tidal potential acting on the misaligned
circumbinary disk around an eccentric binary:
\begin{eqnarray}
&&
\bar{\Phi}
=
\frac{1}{4\pi^2}\int_0^{2\pi}\int_0^{2\pi}
-\sum_{i=1}^2\frac{GM_i}{|\bmath{r}-\bmath{r}_i|}
d\phi{d(\Omega_{\rm{orb}}t)}
\nonumber \\
&\approx&
-\frac{GM}{r}\Biggr[
1
+
\frac{\xi_1\xi_2}{4}
\biggr[
1
+
\frac{3}{2}e^2
\biggr]
\left(\frac{a}{r}\right)^2
\nonumber \\
&-&
\frac{3\xi_1\xi_2}{8}
(
1-e^2+
5e^2\cos^2\gamma
)\sin^2\beta
\left(\frac{a}{r}\right)^2
\Biggr].
\label{eq:tpote}
\end{eqnarray}
Here, we used for integration equation~(\ref{eq:expan2})
and the following approximation:
\begin{eqnarray}
|\vec{r}-\vec{r}_i|^{-1}
\approx
r^{-1}\left[1+g_i\left(\frac{r_i}{r}\right)+\frac{1}{2}(3g_i^2-1)\left(\frac{r_i}{r}\right)^2+\mathcal{O}((r_i/r)^3)\right],
\label{eq:expan3}
\end{eqnarray}
where $g_i=\cos\phi\sin(\gamma-f_i)+\sin\phi\cos(\gamma-f_i)\cos\beta$.
Since the tidal potential has minima at $\beta=0$ (alignment) and $\beta=\pi$ (counter-alignment)
for arbitrary $e$ and $\gamma$,
the tidal torques tend to align or counter-align the tilted circumbinary disk with the orbital plane.
Note that equation (\ref{eq:tpote}) is reduced to equation (3) of \citet{khato09} if $\beta=0$.
The tidal precession timescale for eccentric binaries is given by
\begin{eqnarray}
\tau_{\rm{prec}}&=&
\frac{
|\bmath{J}|
}
{
|\langle
\bmath{T}_{\rm{grav}}
\rangle|
}
=\frac{8}{3}\left(\frac{1}{\xi_1\xi_2}\right)\left(\frac{r}{a}\right)^{2}
\frac{1}{\Theta(e,\gamma,\beta)}
\frac{1}{\Omega}
\label{eq:tprec}
\\
&=&
\frac{4}{3\pi}\left(\frac{1}{\xi_1\xi_2}\right)\left(\frac{r}{a}\right)^{7/2}
\frac{1}{\Theta(e,\gamma,\beta)}
P_{\rm{orb}}
\nonumber
\end{eqnarray}
where $\bmath{J}=r^2\Omega\bmath{l}$ with the disk angular frequency $\Omega=GM/r^3$
and $P_{\rm orb}=2\pi/\Omega_{\rm orb}$ are the specific angular momentum and binary orbital
period, respectively.
Note that the tidal precession timescale depends on the tilt angle and azimuth of tilt for $e\neq0$.
Since the inner edge of the disk is estimated to be $\sim2a$ \citep{al94},
the tidal precession timescale is longer than the binary orbital period.
\subsection{Tidal warping vs. tearing}
There are two types of important viscosities in the circumbinary disk.
The first type is the horizontal shear viscosity, $\nu_1$, which is the
viscosity normally associated with accretion disks. The second type
is the vertical shear viscosity, $\nu_2$, which tends to smooth out disk warping
when the disk is non-planar. \citet{og99} derived the relationship between $\nu_1$ and $\nu_2$:
\begin{eqnarray}
\eta=\frac{\nu_2}{\nu_1}=\frac{2(1+7\alpha^2)}{\alpha^2(4+\alpha^2)}
\label{eq:eta}
\end{eqnarray}
by taking a non-linear effect of the fluid on the warped disk,
where $\eta$ and $\alpha$ are the viscosity ratio parameter
and the Shakura-Sunyaev viscosity parameter, respectively.
For $\alpha\ll1$, the above equation is reduced to
\begin{eqnarray}
\eta\approx\frac{1}{2\alpha^2}.
\label{eq:eta2}
\end{eqnarray}
The global horizontal and vertical viscous timescales
for a geometrically thin disk are estimated to be
\begin{eqnarray}
\tau_{\nu_1}&=&\frac{2}{3}\frac{r^2}{\nu_1}\approx\frac{2}{3}\frac{1}{\alpha}\frac{1}{\Omega}
\left(\frac{H}{r}\right)^{-2},
\\
\tau_{\nu_2}&=&\frac{2}{3}\frac{r^2}{\nu_2}\approx\frac{2}{3}\frac{1}{\eta\alpha}\frac{1}{\Omega}
\left(\frac{H}{r}\right)^{-2},
\end{eqnarray}
where $\nu_1\approx\alpha\Omega{H^2}$ with the disk scale hight $H$ and
$\nu_2\approx\eta\alpha\Omega{H^2}$.
On the other hand, the local horizontal viscous timescale is estimated to be
\begin{eqnarray}
\Delta\tau_{\nu_1}&=&\frac{2}{3}\frac{r\Delta{r}}{\nu_1}\approx\frac{2}{3}\frac{1}{\alpha}\frac{1}{\Omega}
\left(\frac{H}{r}\right)^{-1},
\label{eq:localnu}
\end{eqnarray}
where $\Delta{r}\approx{H}$ is adopted.
The misaligned circumbinary disk is warped at a radius where the tidal precession timescale is longer than
the vertical viscous timescale. We call this radius the tidal warp radius, following \citet{martin07,martin09},
who used it in the context of X-ray binaries. The tidal warp radius is given by
\begin{eqnarray}
\frac{R_{\rm warp}}{r_{\rm{S}}}
&=&
\biggr[
\frac{3}{8}
\frac{\xi_1\xi_2}{\eta\alpha}
|\Theta(e,\beta,\gamma)|
\biggr]^{1/2}
\left(\frac{H}{r}\right)^{-1}
\left(\frac{a}{r_{\rm{S}}}\right),
\label{eq:rwarp}
\end{eqnarray}
Recently, \citet{nix13} have proposed that
the circumbinary disk is torn into mutually
misaligned gas rings if the
timescale of the tidal precession, which is strongly differential,
is shorter than the local horizontal viscous
timescale. The disk tearing occurs at the radius:
\begin{eqnarray}
\frac{R_{\rm tear}}{r_{\rm{S}}}
&=&
\biggr[
\frac{1}{4}
\frac{\xi_1\xi_2}{\alpha}
|\Theta(e,\beta,\gamma)|
\biggr]^{1/2}
\left(\frac{H}{r}\right)^{-1/2}
\left(\frac{a}{r_{\rm{S}}}\right).
\label{eq:rtear}
\end{eqnarray}
This reduces to equation (9) of \citet{nix13} when $e=0$.
The disk tearing is also confirmed by a three-dimensional
Smoothed Particle Hydrodynamics simulations \citep{nix13}.
Inside the tearing radius, the material rapidly accretes
onto the central binary, especially, in a retrograde, misaligned
circumbinary disk. Therefore, in order for the disk to have a warped
structure, $R_{\rm{warp}}$ must be larger than $R_{\rm{tear}}$.
This condition is equivalent to
\begin{eqnarray}
\frac{3}{2}\frac{1}{\eta}
&>&
\frac{H}{r}.
\label{eq:crit1}
\end{eqnarray}
This is then rewritten as
\begin{eqnarray}
\alpha&>&\alpha_{\rm{c}},
\label{eq:crit2}
\end{eqnarray}
where $\alpha_c$ represents the critical value for the disk viscosity parameter,
which is a function of $H/r$.
Figure~\ref{fig:alphac} shows the dependence of $\alpha_{\rm{c}}$ on $H/r$.
The red dashed line denotes the critical value of $\alpha$ for $\alpha\ll1$,
which is equal to $\sqrt{H/(3r)}$ obtained by substituting equation~(\ref{eq:eta2})
into equation~(\ref{eq:crit1}). From the figure, we note $\alpha_{\rm{c}}=\sqrt{H/(3r)}$ is
a good approximation for geometrically thin disks with $H/r<0.1$. For $\alpha>\alpha_{\rm{c}}$, the disk
warping occurs outside the tearing radius. Otherwise, no disk warping will occur
because there is little disk material at the tidal warp radius, which is inside the tearing radius.
Figure~\ref{fig:wtradii} shows the dependence of tidal-warp and
tearing radii on the semi-major axis. For a fixed value of $H/r=0.01$,
$\alpha_{\rm{c}}\sim0.06$. We can confirm that the tidal warp radius
is larger than the tearing radius for any semi-major axis for $\alpha=0.1$,
whereas it is smaller than the tearing radius for any semi-major axis for $\alpha=0.01$.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{fig3.eps}
}
\end{center}
\caption{
Critical value of viscosity parameter as a function of $H/r$.
The black solid line is the critical value, $\alpha_{\rm{c}}$, of $\alpha$ without approximation to $\eta$,
whereas the red dashed line is $\alpha_{\rm{c}}$ for $\eta\approx1/(2\alpha^2)$.
}
\label{fig:alphac}
\end{figure}
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{fig4.eps}
}
\end{center}
\caption{
Dependence of tidal-warp and tearing radii on the semi-major axis.
The black solid and dashed lines are the tidal-warp and tearing radii
for $\alpha>\alpha_{\rm{c}}$ case, respectively, whereas the red solid
and dashed lines are the tidal-warp and tearing radii for $\alpha<\alpha_{\rm{c}}$
case, respectively. Here, $H/r=0.01\,(\alpha_{\rm{c}}\sim0.06)$ is adopted
for each radius. The blue line shows the inner edge radius of the circumbinary disk.
The tidal-warp and tearing radii have other common parameters:
$M=10^7M_\odot$, $q=0.1\,(\xi_1\xi_2=10/121)$, $e=0.6$,
$\beta=\pi/4$, and $\gamma=\pi/4$, respectively.
}
\label{fig:wtradii}
\end{figure}
The disk evolves diffusively for $\alpha > H/r$, whereas it does with dispersive
wave propagation for $\alpha < H/r$ \citep{pp83}. For the latter case,
the communication time due to wave propagation is roughly given by
$2r/c_{\rm{s}}$, where $c_{\rm{s}}$ is the local sound speed.
The tearing radius is then given by
\begin{eqnarray}
\frac{R_{\rm tear,w}}{r_{\rm{S}}}
&=&
\biggr[
\frac{3}{4}
\xi_1\xi_2
|\Theta(e,\beta,\gamma)|
\biggr]^{1/2}
\left(\frac{H}{r}\right)^{-1/2}
\left(\frac{a}{r_{\rm{S}}}\right).
\label{eq:rtear2}
\end{eqnarray}
Comparing this equation with equation~(\ref{eq:rwarp}),
we obtain, for $\alpha\ll1$,
\begin{eqnarray}
\frac{R_{\rm{warp}}}{R_{\rm{tear,w}}}
\approx
\left(\alpha\frac{r}{H}\right)^{1/2}
<1.
\end{eqnarray}
Thus, no disk warping occurs
in a tilted circmbinary disk with $\alpha<{H/r}$.
\subsection{Warping and tearing of circumbinary disks in AGNs}
In this section, we apply our model to AGN disks.
A gaseous disk around a SMBH in an AGN is surrounded by a dusty torus (e.g. \citep{jaffe93}).
The grains in the dusty torus is evaporated above the dust sublimation temperature by the irradiation emitted from the central source. The inner radius of the dusty torus should therefore be determined by the dust sublimation radius: $r_{\rm{dust}}=3\,{\rm{pc}}\,(L/10^{46}\,{\rm{erg\,s^{-1}}})^{1/2}(T_{\rm{dust}}/1500\,\rm{K})^{-2.8}$, where $T_{\rm{dust}}$ is the dust sublimation temperature \citep{bar87}. Assuming that the AGN luminosity is the Eddington luminosity: $L_{\rm{Edd}}\simeq1.3\times10^{38}(M/{\rm M}_\odot)\,\rm{erg\,s^{-1}}$, the dust sublimation radius is rewritten as $r_{\rm{dust}}=4.7\times10^{-1}\,(M/10^7\,{\rm{M}}_\odot)^{1/2}\,{\rm{pc}}$ with the adoption of $T_{\rm{dust}}=1500\,{\rm{K}}$. Since the circumbinary disk should be also inside the dusty torus in our scenario,
we assume the disk outer radius and temperature at $R_{\rm{out}}$ to be given by
\begin{eqnarray}
\frac{R_{\rm{out}}}{r_{\rm{S}}}
\approx
4.8\times10^5
\left(
\frac{M}{10^7\,{\rm{M}}_\odot}
\right)^{-1/2}
\label{eq:rout}
\end{eqnarray}
and $T_{\rm{out}}=1500\,\rm{K}$, respectively.
Here, we assume that the radial profile of the disk temperature obeys
\begin{eqnarray}
T=T_{\rm{out}}\left(\frac{r}{R_{\rm{out}}}\right)^s \hspace{4mm}\,(s<0),
\end{eqnarray}
where $s$, $R_{\rm{out}}$, and $T_{\rm{out}}$ are
the power law index of the radial profile of the disk temperature, disk outer radius,
and temperature at $R_{\rm{out}}$, respectively.
With the thin disk approximation, the ratio of the disk scale hight to radius is then given by
\begin{eqnarray}
\frac{H}{r}=
\sqrt{2}
\left(\frac{c_{\rm{s,out}}}{c}\right)
\left(\frac{R_{\rm{out}}}{r_{\rm{S}}}\right)^{1/2}
\left(\frac{r}{R_{\rm{out}}}\right)^{(s+1)/2},
\label{eq:hor}
\end{eqnarray}
where $c_{\rm{s,out}}=\sqrt{(R_{\rm{g}}/\mu)T}$ is the isothermal sound
speed at $R_{\rm{out}}$ with the gas constant $R_{\rm{g}}$ and
molecular weight $\mu$.
Substituting equation~(\ref{eq:hor}) into equation~(\ref{eq:rwarp}),
the tidal warp radius is rewritten as
\begin{eqnarray}
\frac{R_{\rm warp,AGN}}{r_{\rm{S}}}
&=&
\biggr[
\frac{3}{16}
\frac{\xi_1\xi_2}{\eta\alpha}
\Theta(e,\beta,\gamma)
\biggr]^{1/(s+3)}
\left(\frac{c}{c_{\rm{s,out}}}\right)^{2/(s+3)}
\nonumber \\
&\times&
\left(\frac{R_{\rm{out}}}{r_{\rm{S}}}\right)^{s/(s+3)}\left(\frac{a}{r_{\rm{S}}}\right)^{2/(s+3)}.
\label{eq:rwagn}
\end{eqnarray}
Similarly, the tearing radius is rewritten as
\begin{eqnarray}
\frac{R_{\rm tear,AGN}}{r_{\rm{S}}}
&=&
\biggr[
\frac{1}{4\sqrt{2}}
\frac{\xi_1\xi_2}{\alpha}
\Theta(e,\beta,\gamma)
\biggr]^{2/(s+5)}
\left(\frac{c}{c_{\rm{s,out}}}\right)^{2/(s+5)}
\nonumber \\
&\times&
\left(\frac{R_{\rm{out}}}{r_{\rm{S}}}\right)^{s/(s+5)}
\left(\frac{a}{r_{\rm{S}}}\right)^{4/(s+5)}.
\label{eq:rtagn}
\end{eqnarray}
Figure~\ref{fig:cradii} shows characteristic radii of the misaligned circumbinary disk for $e=0.6$,
$q=0.1\,(\xi_1\xi_2=10/121)$, $M=10^7{\rm{M}}_\odot$, $\alpha=0.1\,(\eta\approx50)$,
$s=-3/4$, $\beta=\pi/4$, and $\gamma=\pi/4$.
The solid and dashed black lines show the tidal warp and tearing radii, respectively.
The red solid line shows the radius, which we call the orbital decay radius, where
the tidal precession timescale equals the timescale in which the binary orbit decays by
the gravitational wave (GW) emission. This orbital decay timescale for an
eccentric binary is given by \citep{p64}
\begin{eqnarray}
\tau_{\rm{gw}}=\frac{5}{8}\frac{1}{\xi_1\xi_2}\left(\frac{a}{r_{\rm{S}}}\right)^{4}\frac{r_{\rm{S}}}{c}\frac{(1-e^2)^{7/2}}{1+73e^2/24+37e^4/96}
\label{eq:tgw}
\end{eqnarray}
Equating equation~(\ref{eq:tgw}) with equation~(\ref{eq:tprec}), we obtain the orbital decay radius:
\begin{eqnarray}
\frac{R_{\rm{prec/gw}}}{r_{\rm{S}}}
&=&
\left[
\frac{15}{64\sqrt{2}}
\frac{\Theta(e,\beta,\gamma)(1-e^2)^{7/2}}{1+73e^2/24+37e^4/96}
\right]^{2/7}
\left(\frac{a}{r_{\rm{S}}}\right)^{12/7}.
\label{eq:rprecgw}
\end{eqnarray}
When it is larger than the tidal warp and tearing radii, the circumbinary
disk can be warped and torn before two SMBHs coalesce. Otherwise, two SMBHs coalesce prior
to disk warping and tearing.
The blue solid lower and upper lines show the inner and outer radii of the circumbinary disk, respectively.
The shaded area between the two blue lines shows the whole region of the circumbinary disk.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{fig5.eps}
}
\end{center}
\caption{
Characteristic radii of the warped and torn circumbinary disk around the SMBH binary
on an eccentric orbit for $e=0.6$, $\alpha=0.1\,(\eta\approx50)$, $s=-3/4$,
$q=0.1\,(\xi_1\xi_2=10/121)$, $M=10^7{\rm{M}}_\odot$, $\beta=\pi/4$, and
$\gamma=\pi/4$. The solid and dashed black lines show the tidal warp and tearing
radii, respectively. The red solid line shows the orbital decay radius where the tidal
precession timescale equals the orbital decay timescale due to the GW emission
(see equation~(\ref{eq:rprecgw})). While the blue lower line represents
the inner radius of the circumbinary disk $r_{\rm{in}}\sim2a$, the blue upper line represents
the outer radius of the circumbinary disk $R_{\rm{out}}/r_{\rm{S}}\approx4.8\times10^4\,(M/10^7\,{\rm{M}}_\odot)^{-1/2}$.
The shaded area between the two blue lines represents the whole region of the circumbinary disk.
}
\label{fig:cradii}
\end{figure}
\section{Discussion}
\label{sec:3}
In this section, we discuss a possible link between the observational signatures
of circumbinary disk warping and tearing and the presence of SMBH binaries.
If the warp and/or tearing radii are observationally determined,
we can estimate the semi-major axis of the SMBH binary from equations~(\ref{eq:rwagn}) and (\ref{eq:rtagn}) as
\begin{eqnarray}
\frac{a_{\rm{warp}}}{r_{\rm{S}}}
&=&
\Biggr[
\frac{16}{3}
\left(\frac{\eta\alpha}{\xi_1\xi_2}\right)
\frac{1}{\Theta(e,\beta,\gamma)}
\Biggr]^{1/2}
\left(\frac{c_{\rm{s,out}}}{c}\right)
\left(\frac{R_{\rm{out}}}{r_{\rm{S}}}\right)^{-s/2}
\nonumber \\
&\times&
\left(\frac{R_{\rm{obs}}}{r_{\rm{S}}}\right)^{(s+3)/2},
\label{eq:semiw}
\\
\frac{a_{\rm{tear}}}{r_{\rm{S}}}
&=&
\Biggr[
4\sqrt{2}
\left(\frac{\alpha}{\xi_1\xi_2}\right)
\frac{1}{\Theta(e,\beta,\gamma)}
\Biggr]^{1/2}
\left(\frac{c_{\rm{s,out}}}{c}\right)^{1/2}
\left(\frac{R_{\rm{out}}}{r_{\rm{S}}}\right)^{-s/4}
\nonumber \\
&\times&
\left(\frac{R_{\rm{obs}}}{r_{\rm{S}}}\right)^{(s+3)/4},
\label{eq:semitear}
\end{eqnarray}
where $R_{\rm{obs}}$ shows the tidal warp or tearing radius determined by observations.
Figure~\ref{fig:semi} shows the relationship between this observed radius
of disk warping or tearing and the binary semi-major axis.
Here, we adopt $\alpha=0.1\,(\eta\approx50)$, $s=-3/4$, $q=0.1\,(\xi_1\xi_2=10/121)$,
$\beta=\pi/4$, and $\gamma=\pi/4$.
The estimated semi-major axes are $a_{\rm{warp}}\sim0.01\,\rm{pc}$
and $a_{\rm{tear}}\sim0.02\,\rm{pc}$, although they mildly depend on the orbital eccentricity,
binary mass ratio, tilt angle, and azimuth of tilt, except for $\beta=\pi/2, 0$.
Because it is observationally more difficult to distinguish between
the disk warping and tearing, the actual
semi-major axis is likely to be $a_{\rm{warp}}\lesssim{a}\lesssim{a_{\rm{tear}}}$.
There are observational evidences for disk warping in the maser disks at the center of NGC 4258
\citep{he05}, Circinus \citep{green03}, NGC\,2273, UGC\,3789, NGC\,6264, and NGC\,6323 \citep{kuo11}.
The tidally-driven warping or tearing of misaligned circumbinary disks can be one of mechanisms
for explaining the warped structure of the maser disks in these systems. Since no observational
information is currently available about the warping or tearing radius, we simply assume that these
maser disks start to be warped or torn at the innermost maser spot radii. Here, we pick up NGC\,4258
case because of the most remarkable example of disk warping among the known maser disk systems.
The inner most maser spot radius is $0.17\,\rm{pc}$. Assuming that $R_{\rm{obs}}=0.17\,\rm{pc}$,
the semi-major axis is estimated to be $1.2\times10^{-2}\,\rm{pc}$.
However, it is difficult to distinguish, solely by the current analysis, whether the central object
is a single SMBH or such a small-scale SMBH binary. Other independent theoretical and observational approaches are needed.
Probing GWs from individual SMBH binaries with masses $\gtrsim10^7{\rm{M}}_\odot$
with pulser timing arrays (PTAs; \citep{lb01,svv09}) gives a powerful tool to
determine if the central object surrounded by the warped maser disk is a SMBH binary
or a single SMBH. For a typical PTA error box ($\approx40\,\rm{deg}^2$) in the sky, the
number of interloping AGNs are of the order of $10^2$ for more than $10^8\,{\rm M}_\odot$
black holes if the redshift range is between $0$ and $0.8$ (see Figure 1 of \citet{tk12} in detail).
As discussed in H14a, even if the observed warped maser disk systems host SMBH binaries,
the characteristic amplitudes of GWs from these SMBH binaries are three to four orders of magnitude
smaller than the current PTA sensitivity, so that it is unlikely for GWs to be detected from these systems.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{fig6.eps}
}
\end{center}
\caption{
Estimate of a semi-major axis of a SMBH binary
by the relationship between the tidal warp and tearing
radii and the observed radius, which is shown by the
purple solid line. The solid and dashed black lines are,
respectively, the tidal-warp and tearing radii of the misaligned
circumbinary disk around the eccentric SMBH binary with
$e=0.6$, $\alpha=0.1\,(\eta\approx50)$, $s=-3/4$, $q=0.1\,(\xi_1\xi_2=10/121)$,
$M=10^7{\rm{M}}_\odot$, $\beta=\pi/4$, and $\gamma=\pi/4$.
The shaded area between the two blue lines represents the
whole region of the circumbinary disk.
}
\label{fig:semi}
\end{figure}
\section{Conclusion}
\label{sec:4}
We have investigated the tidally driven warping and tearing of a geometrically thin, non-self-gravitating
circumbinary disk around two SMBHs in a binary on an eccentric orbit, where the original disk plane is misaligned with
the binary orbital plane. While tidal torques acting on the tilted circumbinary
disk tend to align it with the orbital plane, the viscous torque in the vertical direction tends
to retain its original orientation. The disk can then be warped at the tidal warp radius,
where the tidal precession timescale equals the vertical viscous timescale.
On the other hand, the disk is broken into mutually misaligned rings inside the tearing
radius, where the tidal precession timescale equals the local horizontal
viscous timescale.
There is a critical value of disk viscosity parameter.
If $\alpha>\sqrt{H/(3r)}$ with $H/r\lesssim0.1$,
the tidal warp radius is larger than the tearing radius.
Then, the disk tearing first occurs and subsequently disk warping
occurs outside the tearing radius. If $\alpha<\sqrt{H/(3r)}$ with $H/r\lesssim0.1$, only
the disk tearing occurs, because the disk material inside the
tearing radius is likely to accrete rapidly.
The tidal-warp and tearing radii most strongly depend on the binary
semi-major axis, although it mildly depends on the other orbital and disk parameters
except for $\beta=0\,,\pi$. This strong dependence enables us to estimate the
semi-major axis, once the tidal-warp or tearing radius is determined observationally.
For the tidal-warp or tearing radius of $0.1\,\rm{pc}$, for instance,
SMBH binaries with masses of $10^{7}{\rm{M}}_\odot$ and other typical
orbital and disk parameters are estimated to have a binary separation on a
$10^{-2}\,\rm{pc}$ scale.
\acknowledgments
They also thank Christopher~Nixon for helpful suggestions.
K.H. is grateful to Jongsoo~Kim for helpful discussions and his continuous encouragement.
K.H. would also like to thank the Kavli Institute for Theoretical Physics (KITP) for their hospitality
and support during the program on A Universe of Black Holes. B.W.S. and T.H.J. are grateful for
support from KASI-Yonsei DRC program of Korea Research Council of Fundamental Science
and Technology (DRC-12-2-KASI). This work was also supported in part by the Grants-in-Aid for
Scientific Research (C) of Japan Society for the Promotion of Science (23540271 T.N. and K.H.;
24540235 A.T.O. and K.H.).
|
\section{Introduction}
The availability of wireless communications has drastically increased in recent years and established new applications. Humans, agents, devices, robots, and applications interact together through more and more heterogeneous infrastructures, such as mobile \emph{ad hoc} networks (MANET), vehicular networks (VANET), (mobile) sensor and actuator networks (SAN), body area networks (BAN), as well as always evolving network infrastructures on the Internet. In such networks, items (users, links, equipments, {\em etc.}) may join, leave, or move inside the network at unforeseeable times. A common feature of these networks is their {\em high dynamic}, meaning that their topology keeps continuously changing over time. Dynamic, heterogeneity of devices, usages, and participants, and often the unprecedented scale to consider, make the design of such infrastructures extremely challenging. For a vast majority of them, the dynamics are also unpredictable. Classically, distributed systems are modeled by a static undirected connected graph where vertices are processes (nodes, servers, processors, etc.) and edges represent bidirectional communication links. Clearly, such modeling is not suitable for high dynamic networks.
Numerous models taking in account topological changes over time have have been proposed since several decades, {\em e.g.,}\xspace \cite{AKMUV12,AE84,CCF09,F03,F04,FGM07,SW09}. Some works aim at unifying most of the above approaches. For instance, in~\cite{XFJ03}, the authors introduced the {\em evolving graphs}. They proposed modeling the time as a sequence of discrete time instants and the system dynamic by a sequence of static graphs, one for each time instant. More recently, another graph formalism, called {\em Time-Varying Graphs} (TVG), has been provided in~\cite{CFQS12}. In contrast with evolving graphs, TVGs allow systems evolving within continuous time. Also in~\cite{CFQS12} and in companion papers~\cite{CFMS10,CFMS12}, TVGs are gathered and ordered into classes depending mainly on two main features: the quality of connectivity among the participating nodes and the possibility/impossibility to perform tasks.
In this paper, we focus on the {\em Minimal Dominating Set} (MDS) problem. A dominating set is a subset of vertices of a graph such as each vertex of this graph is either in the dominating set or neighbor of a vertex in the dominating set. A \emph{minimal} dominating set is such that none of its proper subsets is also a dominating set of the graph. Like many distributed covering structure (such as trees, coloring, matching, \emph{etc.}), Minimal Dominating Set is a key building block for numerous network protocols, {\em e.g.,}\xspace hierarchical routing and clustering, unicast, multicast, topology control, media access coordination, to name only a few.
Minimal Dominating Set and some of related problems (such as Maximal Independent Set and Connected Dominating Set) receive some attention in the context of dynamic networks, {\em e.g.,}\xspace \cite{BDTC05,WDCG12,SW10,CMM11}. The difficulty to define covering structures in dynamic networks (including MDS) is pointed out in~\cite{CF13r}. Indeed, the authors show that the definition of such structures may become ambiguous, incorrect, or even irrelevant when applied in dynamic systems. As an example, if the dynamicity of the graph is modeled as a sequence of static graphs and a new MDS is computed at each topological change as in \cite{WDCG12}, the stability of the MDS fully depends on the dynamic rate of the network ({\em i.e.,}\xspace the relative speed of appearance/disappearance of edges). This natural definition may hence lead to an high instability (or even impossibility of use) of the MDS. We discuss more precisely this issue in Section~\ref{sec:MDS}.
This paper aims at proposing a new approach suitable for Minimal Dominating Set construction in time-varying graph with weak connectivity, {\em i.e.,}\xspace the graph may be disconnected at each time instant and topological changes are unpredictable. The only assumption on communications is that every process is infinitely often able to communicate with other processes (not necessarily directly). In this context, our contribution is threefold. First, we propose a new definition of MDS for time-varying graphs that increases stability of this structure. More precisely, we require that each dominated node is infinitely often neighbor of at least one dominating node. Next, we provide a necessary and sufficient topological condition for the existence of a deterministic algorithm for MDS construction in our settings. Finally, we propose a new measure of time complexity in time-varying graph. This measure takes account of communication delays attributable to the dynamicity of the graph and not to the algorithm in order to allow fair comparison between algorithms.
The paper is organized as follows. Section \ref{sec:TVG} presents formally the time-varying graph model and our new measure of time complexity. We devote the Section \ref{sec:UG} to some preliminaries necessary to our main results on MDS presented in Section \ref{sec:MDS}. Finally, Section \ref{sec:conclu} concludes the paper.
\section{Time-Varying Graph: Model and Complexity}\label{sec:TVG}
This section aims to present formally the framework of our study of dynamic systems. In a first time, we recall in Section \ref{sub:model} the model of time-varying graphs (TVGs) introduced by \cite{CFQS12}. We present only definitions needed for the comprehension of our work and we refer the reader to \cite{CFQS12} for more details and an interesting taxonomy of TVGs.
Then, Section \ref{sub:complexity} focuses in complexity measures in this model. We think that a computational model without correct time complexity measure(s) is not complete. We are unable to find in previous works any such measure that is suitable for all TVGs. In consequence, we propose in this paper a new time complexity measure that captures the cost of the algorithm independently of delays introduced by topology changes and asynchronous communications.
\subsection{Model}\label{sub:model}
Let us first borrow the formalism introduced in~\cite{CFQS12} in order to describe the distributed systems prone to high dynamic. We consider {\em distributed systems} made of $n$ computing entities, henceforth indifferently referred to as {\em nodes}, {\em vertices}, or {\em processes}. A process has a local memory, a local sequential and deterministic algorithm, and input\slash output capabilities. We assume that each entity has a unique identifier. Moreover, given two distinct entities $p$ and $q$ identified respectively by $id_p$ and $id_q$, either $id_p<id_q$ or $id_q<id_p$. All these entities are gathered in a set $V$. Let $E$ be a set of edges (or relations) between pairwise entities, that describes interactions between processes, namely communication exchange. The presence of an edge between two vertices $p$ and $q$ at a given time $t$ means that each vertex among $\{p,q\}$ is able to send a message to the other at $t$.
The interactions between processes are assumed to take place over a time span $\mathcal{T} \subseteq \mathbb{T}$ called the {\em lifetime} of the system. The temporal domain $\mathbb{T}$ is generally assumed to be either $\mathbb{N}$ (discrete-time systems) or $\mathbb{R}^+$ (continuous-time systems).
\begin{definition}[Time-varying graph \cite{CFQS12}]
\label{def:TVG}
A time-varying graph (TVG for short) $g$ is a tuple $(V,E,\mathcal{T},\rho,$ $\zeta,\phi)$ where $V$ is a (static) set of vertices $\{v_1,\ldots,v_n\}$, $E$ a (static) set of edges between these vertices $E\subseteq V\times V$, $\rho:E\times\mathcal{T}\to\{0,1\}$ (called presence function) that indicates whether a given edge is available (\emph{i.e.} present) at a given time, $\zeta:E\times\mathcal{T}\rightarrow \mathbb{T}$ (called edge latency function) indicates the time it takes to cross a given edge if starting at a given date, and $\phi:V\times\mathcal{T}\rightarrow \mathbb{T}$ (called process latency function) indicates the time an internal action of a process takes at a given date.
\end{definition}
Given a TVG $g$, let $\mathcal{T}_g$ be the subset of $\mathcal{T}$ for which a topological event (appearance/disappearance of an edge) occurs in $g$. The evolution of $g$ during its lifetime $\mathcal{T}$ can be described as the sequence of graphs $\mathcal{S}_{g} = g_1, g_2,\ldots$, where $g_i=(V,E_i)$ corresponds to the static {\em snapshot} of $g$ at time $t_i \in \mathcal{T}_g$, \emph{i.e.} $e\in E_i$ if and only if $\forall t\in[t_i,t_{i+1}[,\rho(e,t) = 1$. Note that, by definition, $g_i \neq g_{i+1}$ for any $i$.
We consider {\em asynchronous} distributed systems, \emph{i.e.} no pair of processes has access to any kind of shared device that could allow to synchronize their execution rate. Furthermore, at any time, no process has access to the output of $\zeta$, \emph{i.e.} none of them can ({\em a priori}) predict a bound on the message delay. Note that the ability to send a message to another process at a given time does not mean that this message will be delivered. Indeed, the dynamicity of the communication graph implies that the edge between the two processes may disappear before the delivery of this message leading to the lost of messages in transit.
The presences and absences of an edge are instantly detected by its two adjacent nodes. We assume that our system provides to each process a non-blocking communication primitive named \textbf{Send\_retry} that ensures the following property. When a process $p$ invokes \textbf{Send\_retry}$(m,q)$ (where $m$ is an arbitrary message and $q$ another process of $V$) at time $t$, this primitive delivers $m$ to $q$ in a finite time provided that there exists a time $t'\geq t$ such that the edge $\{p,q\}$ is present at time $t'$ during at least $\zeta(\{p,q\},t')$ units of time. In other words, the delivery of the message is ensured if there is, after the invocation of the primitive, an availability of the edge that is sufficient to overcome the communication delay of the edge at this time. Note that this primitive may never deliver a message (\emph{e.g.} if the considered edge never appears after invocation). Details of the implementation of this primitive are not considered here but it typically consists in resending $m$ at each apparition of the edge $\{p,q\}$ until its reception by $q$. This primitive allows us to abstract from topology changes and asynchronous communication and to write high-level algorithms.
\paragraph{Configurations and executions} The \emph{state} of a process is defined by the values of its variables. Given a TGV $g$, a \emph{configuration} of $g$ is a vector of $n+2$ components $(g_i, M_i, p_1, p_2, \ldots, p_n)$ such that $g_i$ is a static snapshot of $g$ (\emph{i.e.} $g_i \in \mathcal{S}_{g}$), $M_i$ is the set of multi-sets of messages carried over $E_i$, and $p_1$ to $p_n$ represent the state of the $n$ processes in $V$. We say that a process $p$ outputs a value $v$ in a configuration $\gamma$ if one of its variable (called an output variable) has the value $v$ in $\gamma$.
An {\em execution} of the distributed system modeled by $g$ is a sequence of configurations $e=\gamma_{0}, \ldots ,\gamma_{k},$ $\gamma_{k+1}, \ldots$, such that for each $k\geq 0$, during an execution step $(\gamma_k,\gamma_{k+1})$, one of the following event occurs: $(i)$ $g_{k} \neq g_{k+1}$, or $(ii)$ at least one process receives a message, sends a message, or executes some internal actions changing its state. The \emph{algorithm} executed by $g$ describes the set of all allowed internal actions of processes (in function of their current state or external events as message receptions or time-out expirations) during an execution of $g$. We assume that during any configuration step $(\gamma_{k},\gamma_{k+1})$ of an execution, if $g_{k} \neq g_{k+1}$, then for each edge $e$ such that $e \in E_k$ and $e \notin E_{k+1}$ (\emph{i.e.} $e$ disappears during the step $(\gamma_k,\gamma_{k+1}$), none of the messages carried by $e$ belongs to $M_{k+1}$. Also, for each edge $e$ such that $e \in E_{k+1}$ and $e \notin E_{k}$ (\emph{i.e.} $e$ appears during the step $(\gamma_k,\gamma_{k+1})$), $e$ contains no message in configuration $\gamma_{k+1}$.
\paragraph{Connected over time TVGs} A key concept of time-varying graphs has been identified in~\cite{CFQS12}. The authors shows that the classical notion of path in static graphs in meaningless in TVGs. Indeed, some processes may communicate even if there is no (static) path between them at each time. To perform communication between two processes, the existence of a \emph{temporal path} (\emph{a.k.a.} \emph{journey}) between them is sufficient. They define such a temporal path as follows: a sequence of ordered pairs $\mathcal{J}=\{(e_1,t_1),(e_2,t_2),...,(e_k,t_k)\}$ such that $\{e_1,e_2,...,e_k\}$ is a path\footnote{A sequence of edges $\{v_1,v'_1\}, \{v_2,v'_2\}, \ldots, \{v_k,v'_k\}$ is a \emph{path} if $\forall i \in\{1,k-1\}, v_{i+1} = v'_{i}$.} if for every $i \in [1,k]$, $\rho(e_i,t_i)=1$ and $t_{i+1} \geq t_i + \zeta(e_i,t_i)$. In other words, a journey from process $p$ to process $q$ is a sequence of adjacent edges from $p$ to $q$ such that availability and latency of edges allow the sending of a message from $p$ to $q$ using the \textbf{Send\_retry} primitive at each intermediate process (refer to \cite{CFQS12} for a formal definition). Note that a journey is a non symmetric relation between two processes.
Based on various assumptions made about journeys (\emph{e.g.} recurrence, periodicity, symmetry, and so on), the authors propose in~\cite{CFQS12} proposes a relevant hierarchy of TVG classes. In this paper, we choose to make minimal assumptions on the dynamicity of our system since we restrict ourselves on \emph{connected-over-time} TVGs defined as follows:
\begin{definition}[Connected-over-time TVG \cite{CFQS12}]\label{def:COT}
A TVG $(V,E,\mathcal{T},\rho,\zeta,\phi)$ is connected-over-time if, for any time $t\in\mathcal{T}$ and for any pair of processes $p$ and $q$ of $V$, there exists a journey from $p$ to $q$ after time $t$. The class of connected-over-time TVGs is denoted by $\mathcal{COT}$\footnote{Authors of \cite{CFQS12} refer to this class as C5 in their hierarchy of TVG classes.}.
\end{definition}
Note that the lifetime of a connected-over-time TVG is necessarily infinite by definition. The class $\mathcal{COT}$
allows us to capture highly dynamic systems since we only require that any process will be always able to communicate
with any other one without any extra assumption on this communication (such as delay, periodicity, or used route). In particular, note that a connected-over-time TVG may be disconnected at each time and that the presence of an edge at a given time does not preclude that this edge will appear again after this time. Define an \emph{eventual missing edge} as en edge that appears only a finite number of time during the lifetime of the TVG. The main difficulty encountered in the design of distributed algorithms in $\mathcal{COT}$ is to deal with such eventual missing edges because no process is able to predict if a given adjacent edge is an eventual missing edge or not. Note that the time of the last presence of such an eventual missing edge cannot be even bounded.
\begin{definition}[(Eventual) Underlying Graph]\label{def:UG}
Given a TVG $g=(V,E,\mathcal{T},\rho,\zeta,\phi)$, the underlying graph of a $g$ is the (static) graph $U_g=(V,E)$. The eventual underlying graph of $g$ is the (static) subgraph $U^\omega_g=(V,E^\omega_g)$ with $E^\omega_g=E \setminus M_g$, where $M_g$ is the set of eventual missing edges of $g$.
\end{definition}
Intuitively, the underlying graph (sometimes referred to as {\em footprint}) of a TVG $g$ gathers all edges that appear at least once during the lifetime of $g$, whereas the eventual underlying graph of $g$ gathers all edges that are infinitely often present during the lifetime of $g$. Note that, for any TVG of $\mathcal{COT}$, both underlying graph and eventual underlying graph are connected by definition. Let us define the \emph{neighborhood} $\mathcal{N}_p$ of a process $p$ is the set of processes with which $p$ shares an edge in the underlying graph.
\paragraph{Induced subclasses}
In the following, we focus on specific subclasses of the class $\mathcal{COT}$ to establish our impossibility result. Informally, we focus on subclasses that gather all TVGs whose underlying graph belongs to a given set. The intuition behind this restriction is the following. In practice, some technical reasons may restrict or prevent the communication between some processes, that induces a given underlying graph for the TVG that models our system. In contrast, we cannot predict in general the availabilities and latencies of communication edges, that leads us to consider all TVGs sharing this underlying graph. More formally:
\begin{definition}[Induced subclass]\label{def:inducedsubclass}
Given a set of (static) graphs $\mathcal{F}$ and a class of TVGs $\mathcal{C}$, the subclass of $\mathcal{C}$ induced by $\mathcal{F}$ (denoted by $\mathcal{C}|_\mathcal{F}$) is the set of all TVGs of $\mathcal{C}$ whose underlying graph belongs to $\mathcal{F}$.
\end{definition}
\paragraph{Diameter}
For any given (static) graph $g$, we denote by $diam(g)$ the diameter of $g$ (that is, the longest distance between two processes of $g$).
\subsection{Complexity Measures}\label{sub:complexity}
At the best of our knowledge, there exists currently no time complexity measure that is suitable for any class of TVGs. Some previous works interested in complexity measure in the TVG model but restrict themselves to synchronous systems (see \emph{e.g.} \cite{KOM11c,KLO10c}), to message complexity (see \emph{e.g.} \cite{CFMS10c}), or to specific class of TVGs in which an existing notion of complexity naturally makes sense (see \emph{e.g.} \cite{IKW14c,CFMS10c}).
The first contribution of this paper is to propose a definition of a time complexity measure suitable for our model. We need a definition that captures the ``quality'' of an algorithm independently of delays introduced by asynchronous communications but also by topological changes. A typical example of such a delay is the waiting after the next apparition of an incident edge to a disconnected process that may introduce a long delay that is not imputable to the algorithm but only to the dynamicity of the system. To perform our goal, we propose to extend the classical notion of time complexity commonly adopted in asynchronous message passing (static) systems.
The classical way to deal with communication delays in time complexity measure in asynchronous message passing models is to consider as the unit of time of an execution the worst delay between the sending and the reception of a message during this execution (see \cite{AW04} for example). Using this time measure, we can bound the termination time of any execution of an algorithm independently of communication delays in this execution. This leads to a time complexity measure (the worst termination time over all possible executions of the algorithm) that induces a fair comparison between algorithms. Our proposal is to extend this idea to dynamic environments by including delays introduced by the dynamicity in this definition. In other words, we will consider as the unit of time of an execution the worst delay between the invocation of the \textbf{Send\_retry} primitive and the delivery of the message by this primitive during this execution.
This natural extension of the definition of time complexity measure of asynchronous message passing systems is not sufficient. Indeed, the dynamicity of the system may introduce another possibly arbitrarily long delay that we call initial delay. As an example, consider a problem that requires each process to propagate an initial value (think about consensus-like problems). An easy way to delay the termination of any algorithm for this problem is to disconnect one process for an arbitrary long (but bounded) time that leads all other processes to wait after its first apparition. Intuitively, this delay is not due to the algorithm but to the dynamicity of the system. Consequently, our complexity measure have to ignore such initial delay.
To deal with this issue, we propose to define for each problem a starting time as follows. It is the smallest time of an execution where the dynamicity of the system ``shows'' to processes the minimal topological information to solve the problem. Note that this starting time depends only of the problem (\emph{e.g.} first connexion of the last process for consensus-like problems) and that, in a static system, the starting time and the initial time are identical (since the system cannot delay apparition of any topological information).
Then, we propose to measure the complexity of an algorithm by the worst time (expressed in the time unit described above) between the starting time and the termination of the algorithm over all its possible executions. We believe that this time complexity measure allows us to fairly compare algorithms designed in our model based on TVGs since it exhibits their intrinsic communication costs and does not take in account delays introducing by asynchronous communications and topological changes.
We now state our complexity measure more formally. In the following, we first restrict to \emph{fixed point computation problems} on a TVG class $\mathcal{C}$, \emph{i.e.} problems that admit a specification of the following form: it is required that the execution on every TVG of $\mathcal{C}$ reaches in a finite time a suffix where each process outputs constantly a given value. The required value depends of the considered problem and is not necessarily the same at each process. Using this definition, leader election or spanning structure construction are fixed point computation problems whereas mutual exclusion or broadcast are not.
We consider now a (deterministic) distributed algorithm $\mathcal{A}$ that satisfies the specification of a fixed point computation problem $\mathcal{P}$ on a TVG class $\mathcal{C}$. Let $e$ be the execution of $\mathcal{A}$ on a given TVG of class $\mathcal{C}$. For any message $m$ sent during $e$, we call \emph{delay} (of $m$) the time between the invocation of the \textbf{Send\_retry} primitive by the sender of $m$ and the delivery of $m$ to its destination. Now, we call \emph{communication step} (or simply step) of $e$ the worst delay over the set of messages that are actually delivered during $e$ (note that we do not consider messages that are never delivered in $e$).
We associate to $\mathcal{P}$ a function $NPS_\mathcal{P}$, called the \emph{necessary presence sets function} of $\mathcal{P}$, that returns, for any TVG $(V,E,\mathcal{T},\rho,\zeta,\phi)$ of $\mathcal{C}$, a set of subsets of $E$. Note that the actual definition of this function depends of the problem itself and not of a TVG nor an execution. Each element of $NPS_\mathcal{P}(g)$ describes one of the set of edges whose apparition is necessary and sufficient to start the effective solving the problem (independently of the used algorithm). We give some examples in the following. For the underlying graph computation problem $\mathcal{UG}$, we have $NPS_\mathcal{UG}(g)=\{E\}$ since each edge of $E$ must appear in the output of any process. For a broadcast problem $\mathcal{B}$, we have $NPS_\mathcal{B}(g)=\{\{(p,q)\}|q\in \mathcal{N}_p\}$ (where process $p$ is the sender of the message) since the apparition of any edge adjacent to $p$ is necessary and sufficient to begin the broadcast of a message by $p$.
We define the \emph{starting time} of the execution $e$ of $\mathcal{A}$ over a TVG $g$ as the smallest time $t\in\mathcal{T}$ such that each edges of at least one element of $NPS_{\mathcal{P}}(g)$ are present at least once before $t$ in this execution. Note that, in a static distributed system, the initial time and the starting time are always identical since all edges of all elements of $NPS_\mathcal{P}(g)$ are present in the initial configuration whatever the definition of $NPS_\mathcal{P}$ is. Finally, the \emph{convergence time} of $\mathcal{A}$ on $g$ is the time (expressed in communication steps of $e$) between the starting time of $e$ and the smallest time in $e$ where the specification of $\mathcal{P}$ is satisfied.
\begin{definition}[Time complexity on a TVG class]
The time complexity of a distributed algorithm $\mathcal{A}$ that satisfies the specification of a fixed point computation problem $\mathcal{P}$ on a TVG class $\mathcal{C}$ is the worst convergence time of $\mathcal{A}$ on all TVGs of $\mathcal{C}$.
\end{definition}
Note that this definition may be naturally extended to so-called \emph{service problems} in the following way. First, we consider as starting time the maximum between the starting time defined above and the time of request of a service (\emph{e.g.} the sending of a message for a broadcast algorithm, the request of critical section for a mutual exclusion algorithm). Second, we substitute the convergence time of the algorithm by the time of achievement of the required service by the algorithm (\emph{e.g.} the delivery of a message to its destinations for a broadcast algorithm, the starting of critical section for a mutual exclusion algorithm).
\section{Underlying Graph Computation}\label{sec:UG}
In this section, we present an underlying graph computation algorithm (see Section \ref{sub:UGalgo}) and proves its time optimality with respect to our new measure (see Section \ref{sub:UGcomplexity}). This algorithm is used as a building block in the next section for our minimal dominating set construction algorithm. Before presenting our algorithm, we need to specify the underlying graph computation problem.
\begin{specification}[Underlying graph]
An algorithm $\mathcal{A}$ satisfies the underlying graph specification for a class of TVGs $\mathcal{C}$ if the execution $e=\gamma_0,\gamma_1,\ldots$ of $\mathcal{A}$ on every TVG $g$ of $\mathcal{C}$ has a suffix $e_i=\gamma_i,\gamma_{i+1},\ldots$ for a given $i\in\mathbb{N}$ such that each process outputs the underlying graph of $g$ in any configuration of $e_i$.
\end{specification}
\subsection{Algorithm}\label{sub:UGalgo}
Our underlying graph computation algorithm is presented in Algorithm \ref{algo:ug}. The intuition behind this algorithm is simple. Each process stores locally a graph, initially empty, that eventually gathers all edges of the underlying graph. At the first appearance of an edge, the two adjacent processes add this edge to their graph. Then, they try to propagate the last version of their graph to all processes that they have as neighbor at least once since the beginning of the execution. When a process receives such a message (that contains the current underlying graph of another process), it add to its own underlying graph every edge it does not already know. If its underlying graph grows during this operation, then the process propagates again its underlying graph to all processes that it has as neighbor at least once since the beginning of the execution.
\begin{algorithm}\caption{Underlying graph computation for process $p$.}\label{algo:ug}
\begin{description}
\item[Variables:]~\\
$g_p=(V_p,E_p)$: underlying graph built by $p$\\
$\mathcal{N}_p$: neighborhood of $p$
\item[Initialization:]~\\
$g_p:=(\{x\},\emptyset)$\\
$\mathcal{N}_p:=\emptyset$
\item[Upon appearance of an edge $\{p,q\}$:]~
\begin{tabbing}
xxx \= xxx \= \kill
\textbf{if} $\{p,q\}\notin E_p$ \textbf{then}\\
\> $\mathcal{N}_p:=\mathcal{N}_p\cup\{q\}$\\
\> $g_p:=(V_p\cup\{q\},E_p\cup\{\{p,q\}\})$\\
\> \textbf{foreach} $r\in\mathcal{N}_p$ \textbf{do}\\
\>\> \textbf{Send\_retry}$(add(g_p),r)$
\end{tabbing}
\item[On reception of $add(g_q)$ from $q$:]~
\begin{tabbing}
xxx \= xxx \= \kill
\textbf{if} $E_q\setminus E_p\neq\emptyset$ \textbf{then}\\
\> $g_p:=(V_p\cup V_q,E_p\cup E_q)$\\
\> \textbf{foreach} $r\in\mathcal{N}_p\setminus \{q\}$ \textbf{do}\\
\>\> \textbf{Send\_retry}$(add(g_p),r)$
\end{tabbing}
\end{description}
\end{algorithm}
This algorithm ensures that, upon the first apparition of the last edge of the underlying graph, this edge is added to the output of adjacent processes and then propagated (at least) to their neighbors in the eventual underlying graph in one step, and so on (note that we have no guarantees for neighbors in the underlying graph in general since it may exist some eventual missing edges). Hence, in any execution, after at most $diam(U^\omega_g)$ steps, this edge (and all others) appears in the output graph of any process. In other words, we have the following result:
\begin{theorem}
Algorithm \ref{algo:ug} satisfies the underlying graph specification for $\mathcal{COT}$. Moreover, its convergence time on any TVG $g$ of $\mathcal{COT}$ is $diam(U^\omega_g)$ steps.
\end{theorem}
\subsection{Time Optimality}\label{sub:UGcomplexity}
In this section, we interest in a lower bound result on the time complexity of underlying graph computation. We restrict ourselves to greedy algorithms that are the most natural ones for this problem. We define a {\em greedy algorithm} for the underlying graph computation as an algorithm that satisfies the following property. The initial output of any process is an empty graph and the graph outputted by a process can only grow (in the sense of inclusion) over time. In other words, such an algorithm ensures that, once a process start to output a given edge or process, this latter always appears in the output of this process afterwards. Note that Algorithm~\ref{algo:ug} falls in this category.
In the following, we prove that no greedy algorithm for underlying graph computation on $\mathcal{COT}$ can exhibit a better time complexity than our algorithm. Indeed, we prove that there exists, for any greedy algorithm, a TVG $g$ in $\mathcal{COT}$ such that this algorithm needs $diam(U^\omega_g)$ steps to compute the underlying graph of $g$. Note that the complexity of the underlying graph computation depends surprisingly of a parameter of the \emph{eventual} underlying graph.
We need a technical lemma for the proof of this optimality result.
\begin{lemma}\label{lem:optimality}
For any greedy algorithm $\mathcal{A}$ that satisfies the underlying computation graph, for any TVG $g=(V,E,\mathcal{T},\rho,\zeta,\phi)$ in $\mathcal{COT}$, for any edge $e\in E$ that is not a cut-edge of $U^\omega_g$, for any process $p\in V$, for any $t\in\mathcal{T}$, $e$ cannot belong to the graph outputted by $p$ in the execution of $\mathcal{A}$ on $g$ at time $t$ if there exists no temporal path from one extremity of $e$ to $p$ that starts after the first appearance of $e$ in $g$ and ends before $t$.
\end{lemma}
\begin{proof}
By contradiction, assume that there exists a greedy algorithm $\mathcal{A}$ that satisfies the underlying computation graph, a TVG $g=(V,E,\mathcal{T},\rho,\zeta,\phi)$ in $\mathcal{COT}$, an edge $\hat{e}\in E$ that is not a cut-edge of $U^\omega_g$, a process $\hat{p}\in V$, and a time $\hat{t}\in\mathcal{T}$ such that $\hat{e}$ appears in the graph outputted by $\hat{p}$ in the execution of $\mathcal{A}$ on $g$ at time $\hat{t}$ and that there exists no temporal path from one extremity of $\hat{e}$ to $\hat{p}$ that starts after the first appearance of $\hat{e}$ in $g$ and ends before $\hat{t}$.
Then, consider the TVG $g'=(V,E,\mathcal{T},\rho',\zeta,\phi)$ with:
\[
\forall e\in E,\forall t\in\mathcal{T},\rho'(e,t)=\begin{cases}
0 \text{ if } e=\hat{e}\\
\rho(e,t) \text{ otherwise}
\end{cases}
\]
Note that, according to the assumption that $\hat{e}$ is not a cut-edge of $U^\omega_g$, $g'$ belongs to $\mathcal{COT}$. Hence, due to the construction of $g$ and the determinism of $\mathcal{A}$, the process $\hat{p}$ receives exactly the same messages before time $\hat{t}$ in $g$ and $g'$ (the assumption on temporal paths after the first appearance of $\hat{e}$ in $g$ ensures us that the fact to remove $\hat{e}$ in $g'$ is not detectable by $\hat{p}$ before time $\hat{t}$). In other words, $\hat{p}$ cannot distinguish the executions of $\mathcal{A}$ on $g$ and $g'$ before $\hat{t}$. As a consequence, $\hat{e}$ appears in the graph outputted by $\hat{p}$ in the execution of $\mathcal{A}$ on $g'$ at time $\hat{t}$. As $\mathcal{A}$ is a greedy algorithm, this edge never disappear of the output of $\hat{p}$ in the execution of $\mathcal{A}$ on $g'$ after time $\hat{t}$. This is contradictory with the fact that $\mathcal{A}$ satisfies the underlying graph specification on $\mathcal{COT}$ since $\hat{e}$ does not belongs to $U_{g'}$ and proves the lemma.
\end{proof}
We are now ready to prove the following result.
\begin{theorem}\label{th:optimality}
For any greedy algorithm $\mathcal{A}$ that satisfies the underlying graph specification on $\mathcal{COT}$, there exists a TVG $g$ of $\mathcal{COT}$ such that the convergence time of $\mathcal{A}$ is at least $diam(U^\omega_g)$ steps.
\end{theorem}
\begin{proof}
Let $\mathcal{A}$ be a greedy algorithm that satisfies the underlying graph specification on $\mathcal{COT}$. Then, let us define the following family of TVGs. For any given $k\in\mathbb{N}^*$, let $g_k=(V_k,E_k,\mathbb{R}^+,\rho_k,\zeta_k,\phi_k)$ be the TVG defined by $V_k=\{p_0,\ldots,p_{3k}\}$, $E_k=\{\{p_i,p_{i+1}\}|i\in\{0,\ldots,3k-1\}\}\cup\{\{p_0,p_{2k}\},\{p_{2k},p_{3k}\}\}$, and
\[\forall e\in E_k,\forall t\in\mathbb{R}^+,\rho_k(e,t)=
\begin{cases}
1 \text{ if } e\in\{\{p_0,p_{2k}\},\{p_{2k},p_{3k}\}\} \text{ and } t< 1\\
0 \text{ if } e\notin\{\{p_0,p_{2k}\},\{p_{2k},p_{3k}\}\} \text{ and } t< 1\\
0 \text{ if } e\in\{\{p_0,p_{2k}\},\{p_{2k},p_{3k}\}\} \text{ and } t\geq 1\\
1 \text{ if } e\notin\{\{p_0,p_{2k}\},\{p_{2k},p_{3k}\}\} \text{ and } t\geq 1
\end{cases}
\]
\[\forall e\in E_k,\forall t\in\mathbb{R}^+,\zeta_k(e,t)=1\]
\[\forall p\in V_k,\forall t\in\mathbb{R}^+,\phi_k(p,t)=0\]
\begin{figure}
\centering
\includegraphics[scale=0.4]{optimality.pdf}
\caption{An illustration of the TVGs family in the proof of Theorem \ref{th:optimality}. \label{fig:optimality}}
\end{figure}
Refer to Figure \ref{fig:optimality} for an example of such a $g_k$. Note that, for any $k\in\mathbb{N}^*$, we have $diam(U^\omega_{g_k})=2k$ (and $diam(U_{g_k})<diam(U^\omega_{g_k})$ since $diam(U_{g_k})=k+1k$). As this graph is connected, $g_k$ belongs to $\mathcal{COT}$. By construction of $g_k$, the starting time of the execution of $\mathcal{A}$ on $g_k$ is $1$ for any $k\in\mathbb{N}^*$ (recall that $NPS_\mathcal{UG}(g)=\{E\}$). Note that, due to the choice of the latency function, any communication step of the execution of $\mathcal{A}$ on $g_k$ takes exactly one time unit.
Consider $e_k$ the execution of $\mathcal{A}$ on $g_k$ for any $k\in\mathbb{N}^*$. From Lemma \ref{lem:optimality}, we know that the edge $\{p_{k-1},p_k\}$ cannot appear in the graph outputted by $p_{3k}$ in $e_k$ before there exists at least one temporal path from $p_{k-1}$ or $p_k$ to $p_{3k}$. Note that the construction of $g_k$ implies that such a temporal path (after time $1$) needs at least $2k$ steps (the length of the path from $p_{k-1}$ or $p_k$ to $p_{3k}$ since $g_k$ is static after time $1$). As the edge $\{p_{k-1},p_k\}$ must eventually appear in the output of any process in $e_k$ by assumption on $\mathcal{A}$, we obtain that the convergence time of $\mathcal{A}$ is at least $diam(U^\omega_{g_k})$ steps, that ends the proof.
\end{proof}
\section{Minimal Dominating Set Construction}\label{sec:MDS}
Minimal dominating set construction is a classical problem in the context of distributed computing since this spanning structure have interesting properties for a lot of practical problems as clustering. Recall that, in a static distributed system, a dominating set $D$ is a subset of processes of the system such that each process that does not belong to $D$ have at least one neighbor in $D$. Such a dominating set is minimal when it has is no strict subset that is also a dominating set.
Regarding dynamic distributed systems, two different approaches have been proposed to handle minimal dominating set problem. We survey them quickly here and show that these definitions seem not relevant in our context, that motivates the need of our new definition presented in this section.
The most natural way to extend minimal dominating set definition in the context of dynamic systems is presented in \cite{WDCG12}. In this work, the dynamic graph is seen as a sequence of static graphs and a new minimal dominating set is computed at each topological change. This approach is not suitable in the case of highly dynamic systems since the system may be always in computation phase (the computation of the new dominating set at each topological change is not instantaneous). In this case, the dominating set may be never stable and is then useless for the application that required it.
The second approach, proposed by \cite{CF13r}, consists in computing a stable dominating set on the underlying graph of the TVG. This approach is interesting since the outputted dominating set is stable in spite of the dynamicity of the system but is still not suitable for our purpose. Indeed, as the dominating set is computed on the underlying graph that may contain eventual missing edges, it is possible for a process to be dominated only through such edges. In other words, a dominated process may have eventually only dominated neighbors, that is counter-intuitive for a minimal dominating set and makes sense only in TVGs where there is no eventual missing edges.
To overcome flaws of precedent definitions in our context of highly dynamic distributed systems (captured by the class of TVGs $\mathcal{COT}$), we propose a third definition. In this definition, we require the outputted minimal dominating set to be stable and each dominated process to be infinitely often neighbor of at least one dominating process. In other words, we want to compute a minimal dominating set on the \emph{eventual} underlying graph. Note that this definition is exactly the same as the one of \cite{CF13r} in TVGs where there is no eventual missing edges.
\begin{definition}[Minimal dominating set over time]
A set of processes $M$ is a minimal dominating set over time (MDST for short) of a TVG $g$ if $M$ is a minimal dominating set of $U^\omega_g$.
\end{definition}
We now specify the minimal dominating set construction problem over TVGs as follows.
\begin{specification}[Minimal dominating set]
An algorithm $\mathcal{A}$ satisfies the minimal dominating set specification for a class of TVGs $\mathcal{C}$ if the execution $e=\gamma_0,\gamma_1,\ldots$ of $\mathcal{A}$ on every TVG $g$ of $\mathcal{C}$ has a suffix $e_i=\gamma_i,\gamma_{i+1},\ldots$ for a given $i\in\mathbb{N}$ such that each process outputs constantly a boolean value in any configuration of $e_i$ and that the set of processes outputting true is a minimal dominating set overt time of $g$.
\end{specification}
\subsection{Preliminaries}\label{sub:MDSprelem}
In this section, we present some preliminary results that are needed in the following. First, we introduce the definition of a strong minimal dominating set of a graph as a dominated set of any connected spanning subgraph of this graph. In Section \ref{sub:MDSimp}, we prove that the existence of such a set in the underlying graph of a TVG is necessary to the existence of an algorithm to construct a minimal dominating set over time of this TVG. We claim in Section \ref{sub:MDSalgo} that this condition is also sufficient. To prove this result, we use a characterization of graphs that admit a strong minimal dominating set that we present in the end of this preliminary section.
\begin{definition}[Strong minimal dominating set]
A strong minimal dominating set (SMDS for short) of a (static) graph $g$ is a subset of processes of $g$ that is a minimal dominating set of every connected spanning subgraph of $g$.
\end{definition}
The following lemma follows directly from definitions and legitimates our interest for strong minimal dominating sets.
\begin{lemma}\label{lem:smdsismdst}
If the underlying graph of a TVG $g\in \mathcal{COT}$ admits a strong minimal dominating set $M$ then $M$ is a minimal dominating set over time of $g$.
\end{lemma}
The next result provides us a characterization of (static) graphs that admits a SMDS. We use this characterization in our minimal dominating set construction algorithm in the next section.
\begin{lemma}\label{lem:characterization}
For any (static) graph $g$ and any minimal dominating set $M$ of $g$, $M$ is a strong minimal dominating set of $g$ if and only if the set of edges $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is a cut-set in $g$ for every process $p\in V\setminus M$.
\end{lemma}
\begin{proof}
First, we prove that, for any SMDS $M$ of a graph $g$, the set of edges $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is a cut-set in $g$ for every process $p\in V\setminus M$. By contradiction, assume that there exists a SMDS $M$ of a graph $g=(V,E)$ such that the set of edges $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is not a cut-set in $g$ for a process $p\in V\setminus M$. Let $s_g=(V,E')$ the subgraph of $g$ defined by $E'=E\setminus\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$. By assumption, $s_g$ is a connected graph. Moreover, in $s_g$ the process $p$ has no neighbor in $M$, that means that $M$ is not a minimal dominating set of $s_g$. This contradicts the fact that $M$ is a SMDS of $g$ and proves the necessity of the condition.
Second, we prove that any minimal dominating set $M$ of a graph $g$ such that the set of edges $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is a cut-set in $g$ for every process $p\in V\setminus M$ is a SMDS of $g$. By contradiction again, assume that there exists a minimal dominating set $M$ of a graph $g=(V,E)$ such that the set of edges $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is a cut-set in $g$ for every process $p\in V\setminus M$ is not a SMDS of $g$. By definition of a SMDS, there exists a connected subgraph $s_g=(V,E')$ of $g$ such that $M$ is not a minimal dominating set of $s_g$. Let us study the two following cases.
\begin{enumerate}
\item $M$ is not a dominating set of $s_g$. Then, there exists a process $p$ such that no neighbors of $p$ in $s_g$ belongs to $M$. As $s_g$ is connected, that means that the set $\{\{p,q\}| q\in M\cap\mathcal{N}_p\}$ is not a cut-set in $g$, that is contradictory with the initial assumption on $M$.
\item $M$ is a dominating set of $s_g$ but is not minimal. We say that a process of a dominating set properly dominates one of its neighbor if it is the only dominating process in the neighborhood of this dominated process. Then, we know that there exists, in $s_g$, two neighbors $p\in M$ and $q\in M$ such that $p$ does not dominate properly any of its neighbors. As $M$ is a minimal dominating set in $g$, we deduce that $p$ dominates properly at least one of its neighbors $r\in V\setminus M$ (recall that $p$ and $q$ are neighbors in $g$ by construction). That means that the set of edges $\{\{r,s\}| s\in M\cap\mathcal{N}_r\}=\{\{p,r\}\}$ is not a cut-set of $g$ (since this edge does not belong to $s_g$ that is connected). This is contradictory with the initial assumption on $M$.
\end{enumerate}
These contradictions show us the sufficiency of the condition and ends the proof.
\end{proof}
\subsection{Impossibility Result}\label{sub:MDSimp}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.4]{mds_w.pdf}
\caption{An illustration of the sequence $(g_n)_{n\in\mathbb{N}}$ used in the proof of Theorem \ref{th:MDSimp}. \label{fig:impossibility}}
\end{figure*}
The proof of our impossibility result presented in Theorem \ref{th:MDSimp} makes use of a generic framework we proposed in another work. We recall here the minimal definitions and results to understand our proof. Due to the lack of space, the interested reader is referred to \cite{BDKP14r} for more details.
\paragraph{Summary of \cite{BDKP14r}} For a given time domain $\mathbb{T}$, a given static graph $(V,E)$ and a given latency function $\zeta$, let us consider the set $\mathcal{G}_{(V,E),\mathbb{T},\zeta}$ of all TVGs over $\mathbb{T}$ that admit $(V,E)$ as underlying graph and $\zeta$ as latency function. For the sake of clarity, we will omit the subscript $(V,E),\mathbb{T},\zeta$ and simply denote this set by $\mathcal{G}$. Remark that two distinct TVGs of $\mathcal{G}$ can be distinguished only by their presence function. For any TVG $g$ in $\mathcal{G}$, let us denote its presence function by $\rho_g$. We define now the following metric $d_\mathcal{G}$ over $\mathcal{G}$. If $g=g'$, then $d_\mathcal{G}(g,g')=0$. Otherwise, $d_\mathcal{G}(g,g')=2^{-\lambda}$ with $\lambda = \text{Sup }\{ t\in\mathbb{T}|\forall t'\leq t,\forall e\in E,\rho_g(e,t') = \rho_{g'}(e,t')\}$.
For a given algorithm $\mathcal{A}$ and a given TVG $g$, let us define the $(\mathcal{A},g)$-output as the function that associate to any time $t\in\mathbb{T}$ the state of $g$ at time $t$ when it executes $\mathcal{A}$. We say that $g$ is the supporting TVG of this output. Let us consider the set $\mathcal{O}_{\mathcal{A},\mathcal{G}}$ of all $(\mathcal{A},g)$-outputs over all TVGs $g$ of $\mathcal{G}$. For the sake of clarity, we will omit the subscript $\mathcal{A},\mathcal{G}$ and simply denote this set by $\mathcal{O}$. Remark that two distinct output of $\mathcal{O}$ can be distinguished only by their supporting TVG. For any output $o$ in $\mathcal{O}$, let us denote its supporting TVG by $g_o$. We define now the following metric $d_\mathcal{O}$ over $\mathcal{O}$. If $o=o'$, then $d_\mathcal{O}(o,o')=0$. Otherwise, $d_\mathcal{O}(o,o')=2^{-\lambda}$ with $\lambda = \text{Sup }\{ t\in\mathbb{T}|\forall t'\leq t,o(t') = o'(t')\}$.
Once we have observed that the metric spaces $(\mathcal{G},d_\mathcal{G})$ and $(\mathcal{O},d_\mathcal{O})$ are complete, we are now able to recall the main result of \cite{BDKP14r}. Intuitively, this theorem ensures us that, if we take a sequence of TVGs with ever-growing common prefixes, then the sequence of corresponding outputs also converges. Moreover, we are able to describe the output to which it converges as the output that corresponds to the TVG that shares all commons prefixes of our TVGs sequence. This result is useful since it allows us to construct counter-example in the context of impossibility results. Indeed, it is sufficient to construct a TVG sequence (with ever-growing common prefixes) and to prove that their corresponding outputs violates the specification of the problem for ever-growing time to exhibit an execution that violates infinitely often the specification of the problem. More formally, we have:
\begin{theorem}\label{th:convergence}
For any deterministic algorithm $\mathcal{A}$, if a sequence $(g_n)_{n\in\mathbb{N}}$ of $\mathcal{G}$ converges to a given $g_\omega\in\mathcal{G}$, then the sequence $(o_n)_{n\in\mathbb{N}}$ of the $(\mathcal{A},g_n)$-outputs converges to $o_\omega\in\mathcal{O}$. Moreover, $o_\omega$ is the $(\mathcal{A},g_\omega)$-output.
\end{theorem}
\paragraph{Application to minimal dominating set}
We are now in measure to prove our impossibility result. This result states that there exists no deterministic algorithm that satisfies the minimal dominating set specification on a TVG of $\mathcal{COT}$ as soon as the underlying graph of the considered TVG does not admit a strong minimal dominating set. Intuitively, this impossibility comes from the following fact. As no process is able to detect eventual missing edges, the minimal dominated set computed by any algorithm must be a minimal dominated set of any possible eventual underlying graph, that is of any connected subgraph of the underlying graph. In other words, the computed minimal dominated set is a strong minimal dominating set. The existence of such a set is then a necessary condition to the existence of an algorithm to compute a minimal dominating set over time. The main difficulty of the formal proof of this result lies in the construction of the TVGs sequence that allows us to apply Theorem \ref{th:convergence}.
\begin{theorem}\label{th:MDSimp}
For any set of (static) graphs $\mathcal{F}$ containing at least one graph that does not admit a strong minimal dominating set, there exists no deterministic algorithm that satisfies the minimal dominating set specification for $\mathcal{COT}|_\mathcal{F}$.
\end{theorem}
\begin{proof}
Let us introduce some notation first. We define, for any TVG $g=(V,E,\mathcal{T},\rho,\zeta,\phi)$, the TVG $g\odot\{(E_i,\mathcal{T}_i)|i\in I\}$ (with $I\subseteq \mathbb{N}$ and for any $i\in I$, $E_i\subseteq E$ and $\mathcal{T}_i\subseteq\mathcal{T}$) as the TVG $(V,E,\mathcal{T},\rho',\zeta,\phi)$ with:
\[\rho'(e,t)=\begin{cases}
0 \text{ if } \exists i\in I, e\in E_i \text{ and } t\in\mathcal{T}_i\\
1 \text{ if } \exists i\in I, e\in E\setminus E_i \text{ and } t\in\mathcal{T}_i\\
\rho(e,t) \text{ otherwise}
\end{cases}\]
By contradiction, assume that there exists a set of (static) graphs $\mathcal{F}$ containing at least one graph that does not admit a strong minimal dominating set and that there exists a deterministic algorithm $\mathcal{A}$ that satisfies the minimal dominating set specification for $\mathcal{COT}|_\mathcal{F}$. In consequence, any process that executes $\mathcal{A}$ outputs a boolean value at any time.
Let $g=(V,E,\mathcal{T},\rho,\zeta,\phi)$ be a TVG of $\mathcal{COT}|_\mathcal{F}$ such that $U_g$ does not admit a strong minimal dominating set and that all edges of $U_g$ are present during the first communication step of the execution of $\mathcal{A}$ on $g$ ($g$ exists by construction of $\mathcal{F}$ and by definition of $\mathcal{COT}|_\mathcal{F}$). Let $t_0$ be the time of completion of the first communication step of the execution of $\mathcal{A}$ on $g$. We construct then a sequence $(g_n)_{n\in\mathbb{N}}$ of TVGs as follows. We set $g_0=g$. Assume that we have already $g_i=(V,E,\mathcal{T},\rho',\zeta,\phi)$ for a given $i\in\mathbb{N}$ such that $g_i\in\mathcal{COT}|_\mathcal{F}$, $U_{g_i}=U_g$, and $\exists \alpha_i>t_0,\forall e\in E,\forall t\leq \alpha_i,\rho'(e,t)=\rho(e,t)$. Then, we define inductively $g_{i+1}$ as follows (refer to Figure \ref{fig:impossibility} for an illustration, gray boxes represent portions of executions where $\mathcal{A}$ outputs a stable minimal dominating set):
\begin{enumerate}
\item Consider the execution of $\mathcal{A}$ over $g_i$ and let $\eta_{i}\in\mathcal{T}$ be the smallest time strictly greater than $\alpha_i$ from which the set of processes that output true is constant ($\eta_i$ exists by assumption on $\mathcal{A}$ since $g_i\in\mathcal{COT}|_\mathcal{F}$);
\item Let $M_i$ be the minimal dominating set computed by $\mathcal{A}$ on $g_i$ (\emph{i.e.} the set of processes of $g_i$ outputting true after $\eta_i$). As $U_{g_i}=U_g$, we know by assumption on $U_g$ that $U_{g_i}$ does not admit a SMDS. In particular, $M_i$ is not a SMDS of $U_{g_i}$. Hence, there exists a process $p_i$ of $V\setminus M_i$ such that the set of edges $E_i=\{\{p_i,q\}| q\in M_i\cap\mathcal{N}_{p_i}\}$ is not a cut-set of $U_{g_i}$;
\item Let $g'_i=g_i\odot\{(E_i,\mathcal{T}\cap]\eta_i,+\infty[)\}$.
\item Remark that $U_{g'_i}=U_{g_i}=U_g$ (by construction of $g'_i$ since $\eta_i>t_0$) and that $U^\omega_{g'_i}$ is connected (since $E(U^\omega_{g'_i})=E(U_g)\setminus E_i$ by construction\footnote{where $E(g)$ denotes the set of edges of $g$.} and $E_i$ is not a cut-set of $U_{g}$). Hence, $g'_i\in\mathcal{COT}|_\mathcal{F}$ and we can consider the execution of $\mathcal{A}$ over $g'_i$. Let $\alpha_{i}\in\mathcal{T}$ be the smallest time strictly greater than $\eta_i$ from which the set of processes that output true is constant. Let $M'_i$ be the minimal dominating set computed by $\mathcal{A}$ on $g'_i$ (\emph{i.e.} the set of processes of $g'_i$ outputting true after $\alpha_i$). Note that $M'_i\neq M_i$ since $M_i$ is not a minimal dominating set of $U^\omega_{g'_i}$ (recall that, in $U^\omega_{g'_i}$, $p_i$ has no neighbor in $M_i$);
\item Let $g_{i+1}=g_i\odot\{(E_i,\mathcal{T}\cap]\eta_i,\alpha_i])\}$.
\end{enumerate}
It is straightforward to check that this construction ensures that, if there exists $g_i=(V,E,\mathcal{T},\rho',\zeta,\phi)$ for a given $i\in\mathbb{N}$ such that $g_i\in\mathcal{COT}|_\mathcal{F}$, $U_{g_i}=U_g$, and $\exists \alpha_i>t_0,\forall e\in E,\forall t\leq \alpha_i,\rho'(e,t)=\rho(e,t)$, then $g_{i+1}$ satisfies the same property. Moreover, as $g_0=g$, this property is naturally satisfied for $i=0$ with any $\alpha_0>t_0$. Hence, the sequence $(g_n)_{n\in\mathbb{N}}$ is well-defined. Note that, for any $i\in\mathbb{N}$, $\eta_i<\alpha_i$ and $\alpha_i<\eta_{i+1}$ (by construction).
That allows us to define the following TVG: $g_\omega=g\odot\{(E_i,\mathcal{T}\cap]\eta_i,\alpha_i])|i\in\mathbb{N}\}$. Note that $U_{g_\omega}=U_g$ and then that $g_\omega$ belongs to $\mathcal{COT}|_\mathcal{F}$. Observe that, for any $k\in\mathbb{N}^*$, we have $d_\mathcal{G}(g_k,g_\omega)=2^{-\eta_k}$ by construction of $(g_n)_{n\in\mathbb{N}}$ and $g_\omega$. Thus, $(g_n)_{n\in\mathbb{N}}$ converges in $\mathcal{COT}|_\mathcal{F}$ to $g_\omega$.
We are now in measure to apply the Theorem \ref{th:convergence} that states that the $(\mathcal{A},g_\omega)$-output is the limit of the sequence of the $(\mathcal{A},g_n)$-outputs. In other words, the $(\mathcal{A},g_\omega)$-output shares a prefix of length $\eta_i$ with the $(\mathcal{A},g_i)$-output for any $i\in\mathbb{N}$ (recall that the sequence of the $(\mathcal{A},g_n)$-outputs is Cauchy since it converges). That means that, for any $i\in\mathbb{N}^*$, the set of processes that output true in $g_\omega$ at $\eta_i$ is $M_i$ and the set of processes that output true in $g_\omega$ at $\alpha_i$ is $M'_i$. As we know that $M_i\neq M'_i$ for any $i\in\mathbb{N}$, we obtain that the set of processes that output true in $g_\omega$ never converges, that contradicts the fact that $\mathcal{A}$ satisfies the minimal dominating set specification for $\mathcal{COT}|_\mathcal{F}$ and ends the proof.
\end{proof}
\subsection{Algorithm}\label{sub:MDSalgo}
We are now able to prove the sufficiency of the existence of a strong minimal dominating set on the underlying graph for the construction of a minimal dominating set over time of any TVG of $\mathcal{COT}$. We prove this result simply by presenting an algorithm based on our underlying graph computation algorithm presented in Section \ref{sec:UG}.
This algorithm works as follows. Once a process has computed the underlying graph, it is easy to decide if this process belongs to the outputted minimal dominating set: the process enumerates (locally and in a deterministic order based \emph{e.g.} on process identities) all minimal dominating sets of the underlying graph and chooses the first one that satisfies Lemma \ref{lem:characterization}. This latter is then a strong minimal dominating set of the underlying graph and hence a minimal dominating set over time of the TVG by Lemma \ref{lem:smdsismdst}. In order to avoid the use of an algorithm of termination detection (for the underlying graph computation), each process repeats the local computation of its output at each update of its local copy of the underlying graph by the algorithm of Section \ref{sec:UG}. The existence of this simple algorithm allows us to state the following result:
\begin{theorem}
For any set of (static) graphs $\mathcal{F}$ containing only graphs that admit a strong minimal dominating set, there exists a deterministic algorithm that satisfies the minimal dominating set specification for $\mathcal{COT}|_\mathcal{F}$.
\end{theorem}
\section{Conclusion}\label{sec:conclu}
This paper addressed the construction of a minimal dominating set over time (MDST) in highly dynamic distributed systems. We considered the weakest connectivity assumption in the hierarchy of time-varying graphs: the graph may be disconnected at each time, topological changes are unpredictable but we know that any process is able to communicate with any other infinitely often using so-called temporal paths. In this context, we proposed a new definition of minimal dominating set increasing the stability of the computed MDST. Next, we provided a necessary and sufficient topological condition for the existence of a deterministic MDST algorithm. We then proposed a new measure of time complexity that takes in account the communication delays due to network dynamic.
The above results used the construction of an underlying graph. We showed the time optimality of our algorithm with respect to our measure. Note that our result (Theorem~\ref{th:optimality}) is valid for greedy algorithms only. We conjecture that all distributed underlying graph algorithms are greedy. This would lead to generalize our result of optimality. Also, we would like to extend our approach to other related overlay constructions.
\bibliographystyle{plain}
|
\section{Introduction}
Finite field transforms
link two sequences of data
according to a general pair of relationships
given by
\begin{align}
V_k
&=
\sum_{i=0}^{N-1}
v_i
K(i,k,\zeta),
\quad k = 0,1,\ldots,N-1, \\
v_i
&=
\sum_{k=0}^{N-1}
V_k
K^{-1}(i,k,\zeta),
\quad i = 0,1,\ldots,N-1,
\end{align}
where
$\mathbf{v}=(v_0\ v_1 \cdots v_{N-1})$ and
$\mathbf{V}=(V_0\ V_1 \cdots V_{N-1})$ are data vectors of length~$N$
with elements defined over a certain
Galois field;
$K(\cdot,\cdot,\cdot)$ and
$K^{-1}(\cdot,\cdot,\cdot)$
are the forward and inverse transformation kernels, respectively;
and
$\zeta$ is a fixed element of the given Galois field~\cite{thomas1983independent}.
In this framework, the concept of number theoretic transforms arises.
If, for each vector $\mathbf{v}$ with components defined over $\operatorname{GF}(p)$,
its associated transform pair $\mathbf{V}$ has also components defined over $\operatorname{GF}(p)$,
then the transformation is said to be a number theoretic transform~(NTT)~\cite{campello2000numericas}.
Number theoretic transforms have been subject of much interest,
because of their capability of exact calculation.
Being all operations required by an NTT performed
in a finite field,
there are no errors due to rounding or truncation.
Consequently,
in principle,
the possibility of error-propagation is eliminated.
The first proposed NTT, due to Pollard~\cite{pollard1971fast},
is a Fourier-like transform
often called finite field Fourier transform (FFFT).
Since its introduction,
the FFFT
was submitted to several
generalizations~\cite{dimitrov1994fermat,dimitrov1992generalized}
and
further methods
were devised~\cite{dimitrov1995golden,hong1993hartley,boussakta1992new,%
souza2004ffdct,souza1999ffht,campello2000numericas,campello1998trigonometry,campello2000hntt,silva2000transformadas}.
Standard number theoretic transforms
have been employed in
different frameworks,
such as:
(i)~fast calculation of convolutions and correlations~\cite{reed1975convolution,agarwal1975convolution,martens1984polynomial};
(ii)~algebraic coding theory~\cite{mcwilliams1988error};
and
(iii)~very large number multiplication~\cite{schonhage1971multiplikation}.
In the past years,
it was observed a broadening
of the NTT range of applications.
Other topics of research have also been benefited from the
utilization of NTT approaches.
Just to illustrate some of them,
one could cite:
(i)~solving Toeplitz system of equations~\cite{hsue1995toeplitz};
(ii)~speech coding problems~\cite{madre2003pitch};
and
(iii)~fast matrix multiplication~\cite{yagle1995matrix}.
Two-dimensional number theoretic transforms were subject to a comprehensive exposition in~\cite{julliendimitrov1996two}.
Additionally,
image processing problems,
involving motion estimation~\cite{toivonen2002motion}
and geometrical rotation~\cite{kriz1980rotation},
were also addressed.
In~\cite{tamori2002fragile}, Tamori~\emph{et al.}
suggested the use of the Fourier-based NTT
to provide a fragile watermarking scheme.
In the current work, this specific technique is denominated
Tamori-Aoki-Yamamoto scheme.
Generally, a watermarking operation consists of encapsulating a given information
(e.g., an image)
into raw image data.
The former is the watermark, the latter is the image to be watermarked.
Frequently, the watermark is required to be perceptually transparent.
Invisible watermarks are typically categorized in two different models:
robust or fragile watermarking.
Depending on the purpose of the watermarking,
robust or fragile schemes are chosen.
Robust watermarks are designed to endure
a variety of data manipulations,
such as adjustments of the compression ratio,
filtering, cropping, or scaling.
Therefore, robust watermarks are
suitable in situations involving
misappropriation of data,
such as
ownership assertion
and copyright enforcement~\cite{lin1999review}.
Despite of this, because of its robustness,
this class of watermarking methodology
is in part
ineffectual for recognizing tampering.
On the other hand,
fragile watermarking schemes
can furnish the necessary tools
for authentication
and integrity
attestation.
In fact, fragile watermarking techniques are supposed to detect tampering
and
determine the identity of the data originator
with high probability~\cite{lin1999review,george1999watermarking}.
However, fragile watermarks are not adequate in situations that require
copyright verification.
Since minimal image alterations are expected to promptly damage fragile watermarks,
fragile watermarking can not properly deal with misappropriation issues.
Several fragile watermarking techniques have been proposed,
usually, with the purpose of still image authentication.
A glimpse of the current investigations
embraces
works on
simultaneously robust and fragile systems~\cite{lu2001multipurpose};
authentication methods for {JPEG} images~\cite{li2004digital,lu2003fragile};
protection of video communications~\cite{chen2005video};
and
tamper detection with the aid of wavelet decompositions~\cite{hu2002multiresolution,kundur1999telltale}.
Some non-conventional applications have also
been reported, such as
blind estimation of the quality of communication links~\cite{campisi2003blind},
and
image quality assessment~\cite{zheng2003image}.
In account of the above,
the number theoretic transforms arise as a natural tool to provide fragile
watermarking methods.
Since
(i) the NTT domain has no physical meaning,
such as harmonic content,
and
(ii) the concept of energy over a finite field is not clear,
any perturbation on a data sequence
produces a dramatic alteration of
its associated number theoretic transformed sequence.
Moreover,
all computations required by an NTT
are integer modular arithmetic,
which can be efficiently implemented.
The aim of this paper is twofold.
First, an amplification of the Tamori-Aoki-Yamamoto fragile watermarking scheme is sought~\cite{tamori2002fragile}.
To explore this line, finite field trigonometry and trigonometrical transforms
are employed~\cite{souza2004ffdct,souza1999ffht,campello2000numericas,campello1998trigonometry,campello2000hntt,silva2000transformadas}.
Second,
a new mode of operation for the
discussed
watermarking scheme is suggested.
The original method by Tamori~\emph{et al.} is categorized as a private watermarking system.
In the present study, the proposal of a signature method with the ability of tamper detection and location
is made.
The rest of the paper is organized as follows.
Section~\ref{sectionfftt} is devoted
to the finite field trigonometry theory.
Central aspects of finite field transforms are also outlined.
Mainly, the focus is directed to the finite field cosine transform (FFCT)
and to the finite field Hartley transform (FFHT).
In Section~\ref{sectiontayschem},
the fragile watermarking technique proposed in~\cite{tamori2002fragile}
is described.
Then,
a new \emph{modus operandi} for the discussed watermarking scheme
is elaborated in Section~\ref{sectionsignature}.
Finally,
computational
results are presented in
Section~\ref{sectioncomputational}.
Concluding remarks are given in Section~\ref{sectionconclusion}.
\section{Finite Field Trigonometrical Transforms}
\label{sectionfftt}
In a series of papers
\cite{souza1999ffht,souza2004ffdct,campello2000hntt,campello1998trigonometry,silva2000transformadas}
by Campello \mbox{de Souza} and collaborators,
a trigonometry over finite fields was derived.
Equipped with such trigonometrical tools, it is possible to
define finite field transforms other than
Pollard's Fourier transform over finite fields~\cite{pollard1971fast}.
In particular,
the finite field trigonometry successfully offers
a formalism to encompass
the finite field cosine transform and
the finite field Hartley transform.
In this section, the theory of finite field trigonometry is briefly discussed and
its major properties are emphasized.
This is necessary to pave
the way for subsequent definitions of the
finite field trigonometrical transforms.
Let $\operatorname{GF}(p)$ be a Galois field
with odd characteristic~$p$.
If $p \equiv 3 \pmod{4}$,
then it is possible to construct an extension field $\operatorname{GF}(p^2)$
using the solution of the irreducible polynomial
$x^2+1=0$~\cite{dimitrov1995golden}.
In addition,
this extension field is isomorphic to the Gaussian integer field $\operatorname{GI}(p)
= \{ a + jb : a,b\in\operatorname{GF}(p) \}$, where
$j^2 \equiv -1 \pmod{p}$.
Furthermore,
compared to the usual complex numbers,
the Gaussian integers enjoy similar arithmetic operation rules.
For a given element $\zeta\in\operatorname{GI}(p)$ with multiplicative order $N$,
the trigonometrical cosine and sine functions are defined
by~\cite{souza1999ffht,campello1998trigonometry}
\begin{equation}
\label{eq.cosine.definition}
\cos(i)
\triangleq
\frac{\zeta^i + \zeta^{-i}}{2},
\quad
i = 0, 1, \ldots, N-1,
\end{equation}
\begin{equation}
\sin(i)
\triangleq
\frac{\zeta^i - \zeta^{-i}}{2j},
\quad
i = 0, 1, \ldots, N-1,
\end{equation}
respectively.
Remarkably,
for each $\zeta$ in $\operatorname{GI}(p)$,
the trigonometrical functions define a
different mapping.
Moreover,
unlike their real field counterparts,
the finite field cosine and sine
can assume complex values,
since $\zeta$ is a Gaussian integer.
The following definition and lemma provide
means to circumvent the need of complex arithmetic.
\begin{definition}[Unimodularity~\cite{silva2000transformadas}]
An element $a+jb\in \operatorname{GI}(p)$ is unimodular if
$a^2 + b^2 \equiv 1 \pmod{p}$.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
\end{definition}
In other words,
unimodular elements have unitary Gaussian integer norm,
denoted by $n(a+jb)\triangleq a^2 + b^2 \pmod{p}$~\cite{hardy1945introduction}.
Besides, if $\zeta$ has unitary norm,
then the computation of trigonometrical functions
are simplified,
as it is shown in the following lemma.
Let the real and the imaginary parts of a Gaussian integer $a+jb$,
where $a,b\in\operatorname{GF}(p)$,
be denoted by $\Re(a+jb)=a$ and $\Im(a+jb)=b$, respectively.
\begin{lemma}[Real Cosine and Sine~\cite{campello2000numericas}]
If
$\zeta\in \operatorname{GI}(p)$ is unimodular with multiplicative order~$N$,
then
$\cos(i) = \Re(\zeta^i)$ and
$\sin(i) = \Im(\zeta^i)$,
$i = 0, 1, \ldots, N-1$.
\end{lemma}
\noindent{\it Proof:~}
Manipulating the cosine function expression (Equation~\ref{eq.cosine.definition}) yields
\begin{align}
\cos(i) & = \frac{\zeta^i + \zeta^{-i}}{2} \\
&=\frac{\zeta^i + (\zeta^i)^{-1}}{2}.
\end{align}
Let $\zeta^i = a + j b$, where $a,b\in\operatorname{GF}(p)$.
Then, by usual arithmetical operations, the inverse of $\zeta^i$ is given by
\begin{align}
(\zeta^i)^{-1}
=
\frac{a - jb}{a^2 + b^2}
=
\frac{a - jb}{n(\zeta^i)}.
\end{align}
On account of the fact that
$n(\cdot)$ has a multiplicative property~\cite{hardy1945introduction},
$n(\zeta^i) = [n(\zeta)]^i = 1$.
Thus, the inverse of $\zeta^i$ is equal to its complex conjugate
$a-jb$.
Therefore, it yields
\begin{align}
\cos(i) & = \frac{(a+jb) + (a-jb)}{2} \\
&= a = \Re(\zeta^i).
\end{align}
A comparable derivation can be obtained for the sine function.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
This result is of paramount importance.
Once a unimodular element $\zeta$ is chosen,
any arithmetic manipulation involving finite field cosine or sine
functions results in quantities defined over the ground field $\operatorname{GF}(p)$.
Therefore,
the unimodularity of $\zeta$ is a sufficient
condition
for the definition of number theoretic transforms that utilize the finite field
cosine or sine.
Under this condition,
two consequences are accomplished:
(i) the finite field trigonometrical functions induce no complex arithmetic operations
and
(ii) all related computations are performed by means of modular integer arithmetic.
In terms of computational implementation,
these two properties are significant.
\subsection{Finite Field Cosine Transform}
Early investigations on the finite field discrete cosine transform were
reported in~\cite{julliendimitrov1996two}.
Independently,
an extended study of the FFCT,
based on the finite field trigonometry,
was introduced in~\cite{souza2004ffdct}.
A comprehensive derivation
and detailed proofs of related FFCT theorems are described in~\cite{souza2004ffdct}.
Generally, the FFCT is a transformation that
provides a spectrum that is not in $\operatorname{GF}(p)$,
thereby exhibiting a complex nature.
Nevertheless, for carefully selected unimodular $\zeta$ values,
it is possible to ensure that the FFCT spectrum
assumes purely real quantities.
Consequently,
the FFCT becomes an NTT.
In the present study,
only the NTT case is of interest.
The following simplified presentation of the FFCT is formulated to guarantee
a real spectrum.
\begin{definition}[FFCT~\cite{souza2004ffdct,julliendimitrov1996two}]
For a given unimodular element $\zeta\in\operatorname{GI}(p)$ with multiplicative order $4N$,
the
FFCT
maps an
$N$-dimensional vector
with elements $v_i \in\operatorname{GF}(p)$, $i = 0, 1, \ldots, N-1$,
into
another vector with components $V_k \in\operatorname{GF}(p)$, $k = 0, 1, \ldots, N-1$
according to
\begin{equation}
\label{ffct}
V_k \triangleq
\sum_{i=0}^{N-1}
2 v_i
\cos
\big(
(2i+1)k
\big),
\end{equation}
for $k = 0, 1, \ldots, N-1$.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
\end{definition}
An inversion formula can be derived and
is given by the following expression~\cite{souza2004ffdct,julliendimitrov1996two}:
\begin{equation}
\label{iffct}
v_i =
\frac{1}{N \pmod{p}}
\sum_{k=0}^{N-1}
a_k
V_k
\cos\big( (2i+1)k \big),
\end{equation}
for $i = 0, 1, \ldots, N-1$.
The auxiliary sequence $\mathbf{a}$ is defined by
\begin{equation}
\label{auxiliar}
a_i
=
\begin{cases}
2^{-1} \pmod{p}, & \text{if $i=0$}, \\
1, & \text{otherwise},
\end{cases}
\end{equation}
for $i = 0, 1, \ldots, N-1$.
Table~\ref{tab:zetas:ffct} brings a list of adequate unimodular $\zeta$ elements to be used
in the definition of cosine and sine functions
in order
to make the FFCT act as an NTT.
\input{table_ffct_zetas.tex}
\subsection{Two-dimensional FFCT}
As the kernel of FFCT is separable,
it follows that
the two-dimensional
FFCT can be performed by successive calls of the one-dimensional FFCT
applied to the rows of the image data; then to the columns of the resulting intermediate
calculation.
Invoking the Equation~\ref{ffct},
the two-dimensional finite field cosine transform can be synthesized in matrix form.
Let~$\mathbf{C}$ be the FFCT matrix,
whose elements are given by
\begin{equation}
c_{i,k} = 2\cos\big((2i+1)k\big),
\quad i,k = 0, 1, \ldots, N-1.
\end{equation}
The two-dimensional FFCT of an $N\times N$ image data $\mathbf{D}$ can be simply written as
\begin{equation}
\hat{\mathbf{D}} = \mathbf{C}\cdot\mathbf{D}\cdot\mathbf{C}^T,
\end{equation}
where the superscript~$T$ is the transposition operation.
The matrix formulation for the inverse transformation
can be obtained in a similar way:
\begin{equation}
\mathbf{D} = \mathbf{C}^{-1}\cdot\hat{\mathbf{D}}\cdot(\mathbf{C}^{-1})^T,
\end{equation}
where,
according to Equation~\ref{iffct},
the elements of $\mathbf{C}^{-1}$ are given by
\begin{equation}
c'_{i,k} =
\frac{1}{N \pmod{p}}
a_k
\cos\big( (2i+1)k \big),
\quad i,k = 0, 1, \ldots, N-1,
\end{equation}
and the quantities $a_k$ are defined in Equation~\ref{auxiliar}.
\subsection{Finite Field Hartley Transform}
Introduced independently in~\cite{hong1993hartley,campello1998trigonometry},
the definition of the FFHT resembles the formalism of
the conventional discrete Hartley transform (DHT)~\cite{bracewell1986hartleytransform}.
\begin{definition}[FFHT~\cite{campello1998trigonometry}]
For a given $\zeta\in\operatorname{GI}(p)$ with multiplicative order $N$,
the FFHT relates two $N$-dimensional vectors $\mathbf{v}$ and $\mathbf{V}$,
according to the following expression
\begin{equation}
V_k \triangleq \sum_{i=0}^{N-1} v_i \operatorname{cas}(ik), \quad k = 0, 1, \ldots, N-1,
\end{equation}
where
\begin{align}
\operatorname{cas}(i) &\triangleq \cos(i) + \sin(i) \\
&=\frac{(1-j)\zeta^i + (1+j)\zeta^{-i}}{2}
\end{align}
is the finite field version of the Hartley function~\cite{bracewell1986hartleytransform}.
\hspace*{\fill}~\QED\par\endtrivlist\unskip
\end{definition}
Consonant with the DHT theory,
apart from the scaling factor $N^{-1}\pmod{p}$,
the FFHT is a symmetrical
transformation
and its inversion formula is given by~\cite{campello2000hntt,campello1998trigonometry,campello2000numericas,souza1999ffht}
\begin{equation}
v_i =
\frac{1}{N\pmod{p}}
\sum_{k=0}^{N-1}
V_k
\operatorname{cas}(ik),
\quad
i = 0, 1, \ldots, N-1.
\end{equation}
Similar to the FFCT, the FFHT can exhibit a complex spectrum.
Nevertheless,
a judicious choice of $\zeta$ can effectively prevent this behavior,
making the components of the finite field Hartley spectrum to be real.
Again, unimodular elements are taken into consideration.
Table~\ref{tab:zetas:ffht} lists
some unimodular $\zeta$ elements for several fields,
and the associated FFHT blocklength.
The FFHT can be written in terms of matrices.
Let~$\mathbf{H}$ be the finite field Hartley matrix,
whose elements $h_{i,k}$ are
given by
\begin{equation}
h_{i,k}= \operatorname{cas}(ik),
\quad i,k = 0, 1, \ldots, N-1.
\end{equation}
Consequently, the forward and inverse FFHT are expressed by
\begin{align}
\mathbf{V} &= \mathbf{H}\cdot \mathbf{v}, \\
\mathbf{v} &= \frac{1}{N\pmod{p}}\mathbf{H}\cdot \mathbf{V},
\end{align}
respectively.
\input{table_ffht_zetas.tex}
\subsection{Two-dimensional FFHT}
Since
the FFHT lacks a separable kernel,
the two-dimensional FFHT cannot be performed by successive calls of the one-dimensional FFHT
applied to the rows of an image data; and then to its
columns~\cite{bracewell1986hartleytransform}.
In fact, the procedure for the calculation of the 2-D FFHT is analogous to the one utilized for
the computation of the 2-D DHT~\cite{bracewell1986hartleytransform}.
Initially, a temporary matrix $\mathbf{T}$ is computed,
which is given by
\begin{equation}
\mathbf{T} = \mathbf{H} \cdot \mathbf{D} \cdot \mathbf{H},
\end{equation}
where $\mathbf{D}$ is an $N\times N$ image and $\mathbf{H}$ is
the finite field Hartley transform matrix.
Subsequently, the two-dimensional FFHT of $\mathbf{D}$ is expressed by
\begin{equation}
\hat{\mathbf{D}} =
\frac{1}{2}\left( \mathbf{T} + \mathbf{T}^{\text{(c)}} + \mathbf{T}^{\text{(r)}} - \mathbf{T}^{\text{(c,r)}} \right),
\end{equation}
where
$\mathbf{T}^{\text{(c)}}$, $\mathbf{T}^{\text{(r)}}$, and $\mathbf{T}^{\text{(c,r)}}$
are built from the temporary matrix~$\mathbf{T}$.
Their elements are respectively given by
$t_{i,N-j\pmod{N}}$,
$t_{N-i\pmod{N},j}$, and
$t_{N-i\pmod{N},N-j\pmod{N}}$,
where $t_{i,j}$ are the elements of~$\mathbf{T}$,
for $i,j=0,\ldots,N-1$.
\subsection{Revisiting the FFFT}
In the light of the finite field trigonometry,
the classic finite field Fourier transform~\cite{pollard1971fast}
can be re-examined.
Mimicking the definition of the standard discrete Fourier transform,
one may consider a Fourier-like finite field transform
equipped with a kernel given by
$\cos(i) + j \sin(i)$, $i\in\operatorname{GF}(p)$.
This kernel suggests the following expression
\begin{align}
V_k &\triangleq
\sum_{i=0}^{N-1} v_i \big( \cos(ik) + j \sin(ik) \big) \\
&=
\sum_{i=0}^{N-1} v_i \left( \frac{\zeta^{ik}+\zeta^{-ik}}{2} + j \frac{\zeta^{ik}-\zeta^{-ik}}{2j} \right) \\
&=
\sum_{i=0}^{N-1} v_i \zeta^{ik},
\quad k = 0, 1, \ldots, N-1.
\end{align}
Because $\zeta$ is an element of $\operatorname{GI}(p)\cong\operatorname{GF}(p^2)$,
the above derivation is a simplified construction of Pollard's FFFT~\cite{pollard1971fast}.
The original FFFT definition considers
a more general field, $\operatorname{GF}(q^m)$,
where $q=p^r$ and $m,r$ are positive integers~\cite{blahut1979transform}.
Observe that the Fourier kernel degenerates
the
finite field trigonometrical functions into
the powers of $\zeta$.
Therefore, if $\zeta\in\operatorname{GF}(p)$, then the FFFT can always act as an NTT.
Otherwise, if $\zeta\in\operatorname{GI}(p)$, then a possibly complex spectrum can be obtained~\cite{reed1975convolution}.
The inversion formula is given by
\begin{equation}
v_i =
\frac{1}{N\pmod{p}}
\sum_{i=0}^{N-1} V_k \zeta^{-ik},
\quad i = 0, 1, \ldots, N-1.
\end{equation}
A proof of the inversion formula can be found in~\cite{blahut1979transform}.
\section{Fragile Watermarking over Finite Fields}
\label{sectiontayschem}
In this section,
a generalization of the
Tamori-Aoki-Yamamoto methodology
is presented.
In this derivation,
instead of using the Fourier-based NTT,
as suggested in~\cite{tamori2002fragile},
a general number theoretic transform is employed.
\subsection{Watermark Embedding}
Let $\mathbf{D}$ be an image data to be watermarked.
By means of modular arithmetic with respect to $p$,
the image data can
furnish its residue part~$\mathbf{D}_R$.
Therefore,
$\mathbf{D}$ can be written as
\begin{equation}
\mathbf{D} = \mathbf{D}_R + \mathbf{D}_M,
\end{equation}
where
$\mathbf{D}_R \equiv \mathbf{D} \pmod{p}$
and
$\mathbf{D}_M$ is an image containing elements that are multiples of~$p$.
The method consists of the insertion of a watermark image~$\mathbf{W}$
into the NTT spectrum of the residue of~$\mathbf{D}$.
Let $\hat{\mathbf{D}}_R$ be the \mbox{2-D}~NTT of~$\mathbf{D}_R$.
Consequently, it follows that
\begin{equation}
\hat{\mathbf{D}}_R' = \hat{\mathbf{D}}_R + \mathbf{W} \pmod{p},
\end{equation}
where $\hat{\mathbf{D}}_R'$ is the watermarked spectral contents of~$\mathbf{D}_R$.
Performing the inverse \mbox{2-D}~NTT
on $\hat{\mathbf{D}}_R'$
results in a watermarked spatial domain image~$\mathbf{D}_R'$
associated to the residue of the host image.
In view of that,
the resulting final watermarked image $\mathbf{D}'$ can be derived as follows
\begin{align}
\label{eq-watermarked}
\mathbf{D}' = \mathbf{D}_R' + \mathbf{D}_M.
\end{align}
\subsection{Watermark Extraction}
The inverse operation, watermarking extraction, can
be implemented in a similar way.
First, one may compute the modular residue
of the original and the watermarked images:
\begin{align}
\mathbf{D}_R &= \mathbf{D} \pmod{p},\\
\mathbf{D}_R' &= \mathbf{D}' \pmod{p}.
\end{align}
Both resulting residue images are submitted to
an NTT application.
Afterwards, the watermark can be recovered
from the
difference of the obtained NTT spectra:
\begin{align}
\mathbf{W} = \hat{\mathbf{D}}_R' - \hat{\mathbf{D}}_R \pmod{p},
\end{align}
where
$\hat{\mathbf{D}}_R'$ and $\hat{\mathbf{D}}_R$
are the number theoretic transform of
$\mathbf{D}_R'$ and $\mathbf{D}_R$, respectively.
\subsection{Private Watermarking Operation}
The above described
watermarking scheme
is intended to be
utilized as
a private watermarking technique for tampering detection and location,
in which the receiver must have the original image~\cite{tamori2002fragile}.
This is often the case in which the authority who marks the
data is also the interested party in verifying its integrity~\cite{lin1999review}.
Therefore, an original image $\mathbf{D}$ is watermarked to provide another image $\mathbf{D}'$,
which is transmitted.
The genuineness of $\mathbf{D}'$
is confirmed
when extracting an unaltered watermark.
Equation~\ref{eq-watermarked}
reveals that, to produce watermarked images that are perceptually
transparent,
the dynamic range of the elements of $\mathbf{D}_R'$
must not be excessively large.
Otherwise,
watermarked images can present significant distortions.
To illustrate this behavior,
the originally proposed scheme
was employed to embed a given watermark (Figure~\ref{fig:private:fourier}(b)) into standard Lena portrait
(Figure~\ref{fig:private:fourier}(a))
using two finite fields of different characteristics: $\operatorname{GF}(13)$ and $\operatorname{GF}(73)$.
The obtained watermarked images are shown in Figure~\ref{fig:private:fourier}(c--d),
respectively.
The calculation over $\operatorname{GF}(73)$ furnished an image
whose degradation is visually perceptible.
Indeed, the described experiment qualitatively demonstrates
a limitation in Tamori-Aoki-Yamamoto scheme.
Because visually transparent watermarked images are not always achievable,
the choice of finite fields becomes restricted to
small values of $p$.
\begin{figure}
\centering
\subfigure[$512\times 512$]{\includegraphics[width=0.4\linewidth,height=0.4\linewidth]{lena-original}}
\subfigure[$64\times64$]{
\hspace{0.1\linewidth}
\includegraphics[width=0.2\linewidth,height=0.2\linewidth]{small_watermark}
\hspace{0.1\linewidth}\strut
}
\\
\subfigure[$512\times 512$]{\includegraphics[width=0.4\linewidth,height=0.4\linewidth]{lena-watermarked-fourier-GF13} }
\subfigure[$512\times 512$]{\includegraphics[width=0.4\linewidth,height=0.4\linewidth]{lena-watermarked-fourier-GF73} }
\caption{Private Watermarking using the FFFT.
(a)~Original image,
(b)~watermark,
(c)~watermarked image in $\operatorname{GF}(13)$,
(d)~watermarked image in $\operatorname{GF}(73)$.}
\label{fig:private:fourier}
\end{figure}
\section{A Signature System}
\label{sectionsignature}
Apart from the private watermarking mode of operating,
the
described scheme can be interpreted from a different, new perspective.
The present study proposes
that the same system can be also considered as
a method for signature data generation.
The watermark embedding process is now
understood as a signature generation operation.
The output signature~$\mathbf{S}$ is simply defined as
\begin{equation}
\mathbf{S} \triangleq \mathbf{D}',
\end{equation}
where the image~$\mathbf{D}'$ is obtained according to
Equation~\ref{eq-watermarked}.
Once obtained, the signature data can be made publicly available.
In this context,
the original watermark extraction method becomes an authentication operation
performed on raw data~$\mathbf{D}$.
The input data to be verified is~$\mathbf{D}$,
which is owned by the user.
Therefore,
the associated signature data,
which can be retrieved from a trusted source,
and
the raw image can be submitted to
the discussed method.
An unaltered extracted watermark
is an evidence of
the integrity and authenticity of~$\mathbf{D}$.
In the case of a tampering attack,
either on~$\mathbf{D}$ or on~$\mathbf{S}$,
the extracted watermark
can indicate the data corruption
location.
It is worthwhile to emphasize that the aim is not to produce perceptually transparent
watermarked images.
Therefore,
this approach completely diminishes the sensitivity to the value of
the field characteristic,
as experienced in the original methodology.
Again because the
NTT transform domain lacks a physical meaning,
even an alteration in the least significant bit of a signal
can render a totally different NTT spectrum.
This is a desirable property.
No matter how subtle,
tampering
can
produce
strikingly
noticeable differences in the recovered watermark.
Consequently,
this approach can
verify the authenticity and integrity of raw digital data,
providing means to spatially locate alteration regions.
\section{Computational Results and Discussion}
\label{sectioncomputational}
In this section,
some computational experiments
are performed to
validate
the proposed
methods.
Selected standard images
available at
the University of Southern California
Signal and Image Processing Institute (USC-SIPI) Image Database~\cite{uscsipi}
were submitted to the
suggested watermarking scheme.
Private watermarking and signature generation
functionalities of the method
were explored,
with the application of
the two-dimensional finite field cosine and Hartley transforms.
Generally the dimensions of a practical number theoretic transformation matrix $N\times N$
are smaller than usual image sizes.
Typical values of the transform size $N$ are given in Table~\ref{tab:zetas:ffct} and~\ref{tab:zetas:ffht}.
Therefore,
the required two-dimensional transforms are obtained
conforming to a blockwise computation.
This approach simply consists of
decomposing a given host image into adjacent,
equal sized, nonoverlapping blocks.
These constituent blocks must match the dimensions
of the considered two-dimensional transformation matrix.
After that,
each block is
transformed
and
the resulting
\mbox{2-D}~spectral
blocks are reassembled,
which results in a
blockwise transformed image.
Figure~\ref{blockwise} illustrates the process.
Accordingly,
the discussed watermarking schemes are applied blockwise
as well.
\begin{figure}
\centering
\includegraphics{block_wise}
\caption{A blockwise two-dimensional transform application.
The size of the the dashed blocks matches the two-dimensional transform blocksize $N\times N$.}
\label{blockwise}
\end{figure}
Considering the FFCT,
a raw image (Figure~\ref{fig:private:cosine}(a))
was processed and furnished the watermarked data displayed in Figure~\ref{fig:private:cosine}(c).
Figure~\ref{fig:private:cosine}(d)
shows the original image after the introduction of random artifacts on its pixel values.
The modification consisted of
incrementing the image pixels by one
with probability $10^{-2}$.
A wavy pattern was selected as the watermark (Figure~\ref{fig:private:cosine}(b)) and
an FFCT over $\operatorname{GF}(7)$ with blocksize of two was chosen.
Over $\operatorname{GF}(7)$,
a possible choice is $\zeta=2+2j$,
which induces an NTT of length equal to 2.
This small blocksize was able to sharply indicate tampering locations (Figure~\ref{fig:private:cosine}(e)).
Similar results were obtained with the aid of the FFHT.
Using an 8-point \mbox{2-D}~FFHT over $\operatorname{GF}(31)$ ($\zeta = 4+4j$),
a modification introduced into a raw image (Figure~\ref{fig:private:hartley}(a))
could be detected and assessed.
Differently from the previous experiment,
a smaller watermark was utilized (Figure~\ref{fig:private:hartley}(b)).
After the calculations,
the recovered watermark image (Figure~\ref{fig:private:hartley}(e))
clearly discloses the tampering
and
shows the interfered regions.
Actually,
the use of a small transform blocklength made
possible to provide information
on the nature of the modification.
In this case, a small text had been written on
the original image.
If a larger transform were employed,
this capability would be reduced,
because the recovered watermark would present
the tampered areas in
larger blocks.
Ultimately the accuracy of
the tampering location would be mitigated.
Figures~\ref{fig:public:cosine} and~\ref{fig:public:hartley}
illustrate
the signature operation proposed in the current work.
An original image (Figure~\ref{fig:public:cosine}(a))
was processed with the inclusion of a watermark
(Figure~\ref{fig:public:cosine}(b)).
Consequently, the signature image data depicted
in Figure~\ref{fig:public:cosine}(c)
was obtained.
Being the calculations over $\operatorname{GF}(127)$,
a visually low-quality image was attained.
However, this is not important, because the
Figure~\ref{fig:public:cosine}(c) is
only intended to be
used as
signature data.
Only its mathematical properties are relevant.
A tampering consisting of incrementing a single pixel at position $(100,100)$
was applied to the original image,
and Figure~\ref{fig:public:cosine}(d) was derived.
The output of the authentication system is
shown in Figure~\ref{fig:public:cosine}(e).
As the chosen $\zeta$ for this calculations was
$2+39j$,
the transform blocksize was 32.
Therefore,
the precision of tampering positioning is restricted to $32 \times 32$ pixel regions.
For an improved tampering location accuracy,
one may consider shorter transforms,
as
in
the experiment depicted in Figure~\ref{fig:public:hartley}.
After signature generation in $\operatorname{GF}(251)$ using the FFHT
with~$\zeta=j$ (Figure~\ref{fig:public:hartley}(c)),
the host data (Figure~\ref{fig:public:hartley}(a)) was maliciously tampered.
Figure~\ref{fig:public:hartley}(d) shows that a forged image is introduced.
This specific~$\zeta$ implies a 4-point transform,
which provides a potentially adequate tampering location accuracy.
The recovered watermark clearly discriminates the modification site (Figure~\ref{fig:public:hartley}(e)).
Given a fixed subject image (Lena portrait),
watermarked/signature images were generated via the method discussed herein,
for all possible finite fields with characteristic ranging from 3 to 251.
Both FFCT and FFHT were considered and~$\zeta$ was chosen in such a way
that 4-point \mbox{2-D}~transformations were guaranteed.
Subsequently,
the peak signal-to-noise ratio (PSNR)
was
adopted as a
measure of the image deterioration after the watermarking process.
For small values of~$p$, visually imperceptible watermarks can be produced,
since the resulting images present high PSNR values.
However, for large values of~$p$, the image quality becomes unacceptable for invisible watermarking purposes.
Despite of that, for signature data generation,
obtaining low PSNR values becomes less relevant.
Quantitative data
concerning the image degradation
are summarized in Figure~\ref{fig:psnr}.
\begin{figure}
\centering
\subfigure[$512\times 512$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{elaine-original-cosine-private} }
\subfigure[$512\times 512$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{inverted-elaine-watermark-cosine-private} }
\subfigure[$512\times 512$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{elaine-watermarked-cosine-private} } \\
\subfigure[$512\times 512$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{elaine-corrupted-cosine-private} }
\subfigure[$512\times 512$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{inverted-elaine-recovered-cosine-private} }
\caption{Private watermarking using the FFCT.
(a)~Original image \texttt{elaine.512.tiff} (Elaine),
(b)~watermark (negative image for better visualization),
(c)~watermarked image,
(d)~tampered image with a Bernoulli distributed noise of $10^{-2}$ error probability,
(e)~extracted watermark with clear location of tampering (negative image for better visualization).
Calculations were performed over $\operatorname{GF}(7)$ with $\zeta = 2+2j$, transform size is 2.}
\label{fig:private:cosine}
\end{figure}
\begin{figure}
\centering
\subfigure[$256 \times 256$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5210-original-hartley-private} }
\subfigure[$64 \times 64$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5210-watermark-hartley-private} }
\subfigure[$256 \times 256$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5210-watermarked-hartley-private} } \\
\subfigure[$256 \times 256$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{5210-corrupted-hartley-private} }
\subfigure[$256 \times 256$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{5210-recovered-hartley-private} }
\caption{Private watermarking using the FFHT.
(a)~Original image \texttt{5210.tiff} (Stream and bridge),
(b)~watermark,
(c)~watermarked image,
(d)~tampered image with the additive insertion of another image,
(e)~extracted watermark with a hidden message (\textsc{send more money}).
Calculations were performed over $\operatorname{GF}(31)$ with $\zeta = 4+4j$, transform size is 8.}
\label{fig:private:hartley}
\end{figure}
\begin{figure}
\centering
\subfigure[$256 \times 256$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5112-original-cosine-public} }
\subfigure[$64 \times 64$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5112-watermark-cosine-public} }
\subfigure[$256 \times 256$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{5112-watermarked-cosine-public} } \\
\subfigure[$256 \times 256$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{5112-corrupted-cosine-public} }
\subfigure[$256 \times 256$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{5112-recovered-cosine-public} }
\caption{Signature data generation using the FFCT.
(a)~Original image \texttt{5112.tiff} (Clock),
(b)~watermark,
(c)~signature image (publicly available),
(d)~tampered image with the pixel at position $(100,100)$ incremented by one,
(e)~extracted watermark.
Observe the fragility of the signature embedding: a single bit alteration made a whole block
of recovered data to be distorted.
Calculations were performed over $\operatorname{GF}(127)$ with $\zeta = 2+39j$, transform size is 32.}
\label{fig:public:cosine}
\end{figure}
\begin{figure}
\centering
\subfigure[$512 \times 512$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{7108-original-hartley-public} }
\subfigure[$64 \times 64$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{7108-watermark-hartley-public} }
\subfigure[$512 \times 512$]{\includegraphics[width=0.3\linewidth,height=0.3\linewidth]{7108-watermarked-hartley-public-forgery} } \\
\subfigure[$512 \times 512$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{7108-corrupted-hartley-public-forgery} }
\subfigure[$512 \times 512$]{\includegraphics[width=0.45\linewidth,height=0.45\linewidth]{7108-recovered-hartley-public-forgery} }
\caption{Signature data generation using the FFHT.
(a)~Original image \texttt{7108.tiff} (APC),
(b)~watermark,
(c)~signature image (publicly available),
(d)~tampered image,
(e)~extracted watermark.
Calculations were performed over $\operatorname{GF}(251)$ with $\zeta = j$, transform size is 4.}
\label{fig:public:hartley}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{psnr_experiment}
\caption{Quality degradation of watermarked images generated by the discussed method as
a function of the finite field characteristic.
The PSNR was employed to assess the image degradation.
The Lena portrait was watermarked with 4-point \mbox{2-D}~FFCT (\mbox{---$\circ$---}) and FFHT (\mbox{---$\Box$---})
in all pertinent Galois fields $\operatorname{GF}(p)$, for $3\leq p \leq 251$.}
\label{fig:psnr}
\end{figure}
\section{Conclusion}
\label{sectionconclusion}
The utilization of finite fields as an avenue for fragile watermarking was
explored.
The watermarking scheme suggested in~\cite{tamori2002fragile}
was amplified and generalized by
the introduction of finite field trigonometrical transforms,
such as the finite field cosine and Hartley transforms.
The new proposed methodology generates signature data
that can provide
image authentication as well as tampering location capability.
\section*{Acknowledgments}
This work was partially supported by the
CNPq (Brazil); and DFAIT and NSERC (Canada).
{\small
\bibliographystyle{IEEEtran}
|
\section{Introduction}
How galaxies form and evolve remain open questions. Models
reproduce global properties of galaxies well, partly by construction,
but discriminative tests of plausible models are few and far
between. The greatest tests of theories and models often come from
considering the extremes of parameter space.
For example, some of the most demanding tests on the current
paradigm of structure formation come from the lowest mass galaxies, which appear to be
staggeringly underabundant \citep[the ``missing satellite problem";][]{moore,klypin}.
Here we focus
on the opposite extreme by examining some of the most massive galaxies. Rather than considering their numbers as a population, we will
examine their internal structure. While the
least massive galaxies test certain aspects of the standard paradigm,
namely the low mass end of the halo mass function, the nature of
parent vs. satellite galaxy, and the role of reionization, giant galaxies
help test the effects and results of a Hubble time of accretion, merging,
collisions, and cannibalisms of dwarf companions \citep[cf.][]{wang}.
In a theoretical framework of structure growth whose principal
characteristic is hierarchical accretion, it is natural to suspect that
the environment in which a galaxy resides plays a prominent role in
its evolution. However, the success of models in which observable
galaxy properties (luminosity, color, morphology) are based
primarily on the galaxy's stellar or dark matter mass \citep[see for examples the large literature on halo occupancy models,
some examples include][]{peacock,berlind,conroy} has been interpreted as supporting the view
that environment plays a subdominant
role in galaxy evolution \citep{kauffmann,bundy}.
That is not to say that environment plays no role in such models. The
parent vs. satellite nature of a galaxy \citep[cf.][]{berlind05,tinker12} is envisioned to be key
to explaining some critical observations \citep{peng,knobel}. However, disentangling the physics involved is
complicated greatly by the different degrees to which environment correlates with
various galaxy properties, such as morphology and
color \citep{blanton}. And so, efforts to include environmental
influences in models increase the complexity and complicate interpretations of the data \citep[cf.][]{wechsler,tinker12,woo}.
Empirically, the role of environment in galaxy evolution is often explored in the densest environments \citep[see][for a review]{boselli}, which
are necessarily the most complicated. Furthermore, the principal influence of the environment may not be direct, through phenomenon
such as ram pressure stripping or tidal interactions, but rather indirect through either ``assembly bias" \citep{sheth04, gao, wechsler}, which is purely a dark matter halo phenomenon in which close halo pairs tend to have earlier formation times than otherwise similar distant pairs, and ``history bias", in which the formation age of halos {\sl and} their
subsequent star formation history
varies across environments \citep{delucia}.
Identifying clean tests of the influence of mass and environment on galaxy evolution is imperative.
Defying the developed expectation that mass is universally the dominant
determinant of a galaxy's appearance and that massive galaxies should not contain cold gas
and not be actively star forming \citep[cf.][]{dekel,cattaneo}
is a class of massive, rotation dominated
galaxy that includes H{\small I}ZOA J0836-43, the most H{\small I}-massive
galaxy known
\cite[Sc, M$_{H{\small I}} = 7.5 \times 10^{10}$ M$_\odot$;][]{2011AJ....142..170H},
the fastest known rotator UGC 12591 \cite[S0/a, V$_{max} = 506$ km s$^{-1}$;][]{1986ApJ...301L...7G}, and a few other
giant disks found in the recent 40\% release of the ALFALFA survey \citep{2011AJ....142..170H}.
However, these are rare galaxies. Current determinations
of the local galaxy H{\small I}-mass function illustrate that
statistically, H{\small I}ZOA J0836-43 should not exist in the volume
explored \citep[probability of $3\times 10^{-8}$;][]{zwaan}. Nevertheless, such extreme
galaxies do exist, they are as massive as corresponding
massive early-type galaxies at low
redshift \citep{bernardi}, and they provide evidence by example against the
hegemony of mass in determining a galaxy's
properties.
Although such galaxies would clearly not be predicted in a model where mass alone determines a galaxy's properties,
we do not know whether more sophisticated models that are currently available successfully predict the properties of such extreme objects.
Because these objects are rare, global statistical tests of the models \citep[cf.][]{obreschkow} may not recognize the absence of such galaxies.
It is therefore critical to complement large-scale comparisons, which offer
the powerful advantage of statistics, with carefully crafted samples
that aim to challenge the models at the margins.
These unusual galaxies must have experienced a strikingly different
evolutionary history than that of massive early type galaxies, whether that means
fewer or no large accretion events, a different angular momentum
accretion history, and/or a distinctly different star formation history.
They are likely to be the nearest analogs of the ``classical" view of galaxy
formation \citep{eggen}, where galaxies form smoothly within an overdensity.
In fact, to produce disk-dominated galaxies numerical simulations
of spiral galaxies set in a cosmological context have had to
select to model systems in underdense environments, a reflection
on what is now understood to relate to the details of the
angular momentum accretion history \citep[cf.][]{governato,sales}. Therefore, the importance
of environment in the genesis of these galaxies is hardwired
into detailed physics-oriented models, in opposition to what is often done in statistically-oriented
theoretical treatments.
Complementing the morphological differences between early and
late type giant galaxies, there are differences in star
formation histories. Late type giant galaxies presumably provide a
measure of the quiescent mode of star formation and a contrast to what is
likely to have occurred in early type systems. Empirically, we need
to determine whether these differences are reflected in the overall
star formation efficiency, which can be defined either relative to the total ``condensed"
baryonic content (the mass in stars and cold gas within the disk) or to the total
mass. The former measures the fraction of all ``condensed" baryons that formed stars, while the latter
measures the fraction of the total mass converted to stars and so includes any variations
in the ``condensed" baryon fraction among galaxies.
At this point, we pause to consider the ill-defined, and often ill-measured, description of a galaxy's mass. Theoretically, the preference would be to describe galaxies by their total mass, where total refers to both dark and baryonic mass, or halo mass, where halo refers to the dominant dark matter component. Even there, these are moving targets because of the definition of an outer radius, such as a virial radius. However, such definitions are completely impractical in an empirical sense because we rarely have any direct measurements of an individuals galaxy's total or halo mass. Instead, investigators either default to estimating the baryonic mass, which can be problematic in its own ways (as discussed next), or to assuming that the ranking between baryonic (or, typically, more accurately stellar) mass and halo mass is nearly 1:1 \citep[see, for examples, halo abundance matching schemes;][]{berlind}. The latter approach has worked well in reproducing statistical measurements of galaxy populations from theoretical dark matter distribution models, lending support to the hypothesis that the rankings are not grossly different than 1:1. We will therefore measure baryonic masses, and presume that these are not only measurements of the baryons in these systems but also, in a ranked sense, measurements of the total mass.
The baryonic mass measurements come with their caveats as well. It is usually the case that only stellar mass, through luminosity and colors, is estimated and corrections for baryons in other phases are either ignored or treated in the mean. We will describe a somewhat more accurate process in which we use IR magnitudes that have been carefully calibrated to stellar masses and complement those with measurements of the gaseous baryons based on H{\small I} observations. This approach still ignores baryons in hot phases and there is a well-known baryon shortfall \citep{bregman}, so these masses are still not the full picture. However, to be specific, we will measure the stellar plus gaseous masses, and use these liberally as a proxy for the total galaxy mass. This assumption is broadly valid, not worse than a factor of two in total mass, as evidenced by the scatter about an apparently constant baryon fraction resulting from studies of the baryonic Tully-Fisher relation \citep{zbtf} for galaxies spanning many decades in baryon mass.
The Cosmic Flows project, which is consistently remeasuring
all available H{\small I} line widths, provides a uniform, unprecedented
dataset from which to draw these massive rotating and H{\small I}-massive
galaxies. In combination with the infrared imaging provided by the
{\sl Spitzer Space Telescope} and the calibrated conversion between 3.6
photometry and stellar mass \citep{eskew}, we are now in a position
to compare gaseous, stellar, and total masses for these rare
rotationally supported giant galaxies. In \S2 we present the data and present our two
subsamples of extreme galaxies in \S3. In \S4 we examine the properties of these galaxies
relative to the baryonic Tully-Fisher relation and draw inferences regarding the ``condensed" baryon fractions and star formation efficiencies. We summarize our conclusions in \S5. Where needed, we adopt H$_0 = 70$ km s$^{-1}$ Mpc$^{-1}$.
\section{The Data}
\subsection{The Cosmic Flows Program}
Since 2009, the Cosmic Flows project (CF) has gathered all the
digital H{\small I} spectra available from public archives of the largest
radio-telescopes worldwide and measured them in a consistent
way. Two sub-projects of CF, at Green Bank in the USA and at
Parkes in Australia \citep{2011MNRAS.414.2005C}, complete the archives for targets without
previous observations adequate for the Tully-Fisher method of
distance measurements \citep{1977A&A....54..661T}. The latest update of this catalog
is published in \cite{CT2015}.
The main goal of CF is to map the all-sky peculiar velocity
field locally and to uncover, at high spatial resolution, the
underlying dark matter distribution. In that cause, they have measured
tens of thousands of galaxy line widths with a new method
described by \cite{2009AJ....138.1938C} and \cite{2011MNRAS.414.2005C}. Briefly, for each galaxy, they evaluate the
line width parameter Wm50, which measures the H{\small I} profile width
at 50\% of the mean flux within the velocity range encompassing
90\% of the total H{\small I} flux. That quantity is transformed into the more
physically motivated parameter W$_{mx}$, which is the same width
corrected for the slight relativistic broadening and for broadening
due to finite spectral resolution, corrected for inclination and so is
effectively twice the maximum rotation velocity. Technical details
of this process are provided by \cite{2009AJ....138.1938C} and \cite{2011MNRAS.414.2005C}, and reviewed in \cite{2012ApJ...749...78T}.
The result is a catalog of H{\small I} measurements of unprecedented
size (14221 galaxies) and consistency. They use the ratio of the
signal level at 50\% of the mean flux to the noise measured beyond the
frequency extremities of the signal to parametrize the line width
uncertainty. Only profiles with uncertainty estimates smaller than
or equal to 20 km s$^{-1}$ are retained (more than 75\% of the 14221
galaxies, exactly 10733 galaxies satisfy this criteria) after a supplementary
visual inspection. This catalog is available for public use at
the Extragalactic Distance Database (EDD) website\footnote{http://edd.ifa.hawaii.edu; catalog ``All Digital H{\small I}"} where several
other parameters, such as the integrated H{\small I} line fluxes for those
lines with a flux calibration error better than
about 10 to 15\% and the average heliocentric velocities, are also
available and described \citep{2009AJ....138..323T}.
In addition to this ``all-digital" catalog, we also use the
``pre-digital" catalog of measurements at 20\% of the H{\small I} line peaks.
The relation
\begin{equation}
V_{m50} ({\rm all - digital}) = (1.015 \times W_{20}({\rm pre -digital}) - 11.25)
\end{equation}
connects the two catalogs \citep{2009AJ....138.1938C}. In total, we have 16,124 galaxies with
coherent H{\small I} measurements and 12,189 with sufficiently good quality
for distance measurements.
\subsection{The Infrared Data}
We use 3.6$\mu$m magnitudes measured from data obtained using the {\sl Spitzer Space Telescope} \citep{2004ApJS..154....1W} and the IRAC camera \citep{2004ApJS..154...10F}, from a number of different observing programs \citep[such as,][]{2009ApJ...703..517D,sheth,2012AJ....144..133S}. For Tully-Fisher studies, within the CF sample, there are 6,007 galaxies with sufficiently good H{\small I} linewidth measurements, $\sigma_{Wm50} < 20$ km s$^{-1}$ and inclinations greater than 45$^\circ$. Among these, 2,493 have been observed with {\sl Spitzer}. We adopt total magnitudes measured either using the Archangel pipeline \citep{2007astro.ph..3646S} or the S$^4$G pipeline \citep{munoz}.
Distances, needed to calculated absolute magnitudes, are presumed from the Hubble flow using H$_0 = $ 70 km s$^{-1}$ Mpc$^{-1}$ and the recessional velocities from the H{\small I} catalog.
We use the 3.6$\mu$m photometry because there now exist simple, and robust, conversions to stellar mass \citep{eskew,meidt14} that have been vetted with comparisons to SDSS SED-derived stellar masses \citep{cybulski}. \citet{eskew} give the conversion between stellar mass and infrared flux as M$_* = 10^{5.97} F_{3.6} ({D\over 0.05})^2$, where D is the distance to the galaxy in Mpc, $F_{3.6}$ is the flux in Jy, and M$_*$ is expressed in solar masses. A slightly more precise estimate can be obtained by including the 4.5$\mu$m flux, but both \citet{eskew} and \citet{querejeta} find only a weak dependence between M$_*$ and infrared color. The largest uncertainty in the derived stellar mass remains the overall normalization due to uncertainties in the adoption of a specific stellar initial mass function (the \cite{eskew} calculation adopts a Salpeter IMF).
\section{GIANT DISK GALAXIES}
To define our galaxy subsamples, we search both for those galaxies with
the largest H{\small I} masses and for those
with the largest rotational velocities. Due to our use of the H{\small I} database, we may miss gas-poor versions of the latter, although such galaxies are likely to be early type and therefore not the pristine test cases we are searching for.
Because of broad correlation between gas mass and total mass, these samples have some objects in common.
\subsection{The most H{\small I} massive galaxies}
We compute H{\small I} masses (M$_{\rm HI}$)
from the publicly available integral fluxes and average heliocentric
velocities \citep{2009AJ....138.1938C} using
\begin{equation}
{\rm M}_{\rm HI}= 2.36 \times 10^5 {\rm D}^2 {\rm FI}
\end{equation}
\noindent
where D is the distance in Mpc and FI is the integrated flux in the H{\small I}
line in Jy km s$^{-1}$.
The distribution in log M$_{\rm HI}$ for the 10733 galaxies for which we can calculate the necessary quantities covers the range of 5.53 to 10.72, with a median of
9.62, corresponding to a range of $3.4\times 10^5$ to $5.2\times 10^{10}$M$_\odot$. PGC8681/UGC01752 is the most H{\small I} massive
(Figure \ref{fig:PGC0008681}) galaxy in the full sample, but it is not in our subsample because we lack {\sl Spitzer} photometry for it. PGC17625, with only a slightly lower H{\small I} mass ($4.5\times10^{10} {\rm M}_\odot$), tops our list. In Figure \ref{fig:sample} we show the distribution of M$_{\rm HI}$ for the 100 most H{\small I} massive galaxies that also have existing {\sl Spitzer} photometry. Choosing the top 100, which was set only because it is a simple round number, corresponds roughly to selecting the top 5\% of the available sample. One galaxy (PGC90167) was removed from our subsample because even though it was observed with {\sl Spitzer}, the S$^4$G survey \citep{sheth,munoz} was unable to provide a reliable magnitude measurement.
We distinguish this sample from an H{\small I} rich sample, where in that case one would presumably select on either the ratio of gas to stellar mass, M$_{\rm HI}$/M$_*$, or to dark matter mass, M$_{\rm HI}$/M$_{\rm DM}$. Although such a sample is also of interest, that selection depends on the modeling of M$_*$ or M$_{\rm DM}$ and therefore complicates any interpretation.
\begin{figure}
\centering
\includegraphics[scale=0.4]{PGC0008681-UGC01752-shg2005-AOlf}
\includegraphics[scale=0.5]{PGC0008681}
\caption{H{\small I} profile of PGC0008681 as observed at Arecibo, and B band image from DSS2 survey at the size of Arecibo beam (3 arcmin). This is the largest H{\small I} mass galaxy recorded in the full available sample, but is not in our subsample due to lack of {\sl Spitzer} photometry.}
\label{fig:PGC0008681}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.5]{fig100.eps}
\caption{Distribution of galaxies in our subsamples selected by rotational velocity and H{\small I} mass. We highlight the H{\small I} mass selected sample using open blue circles and the rotational velocity selected sample using filled red circles. We will continue this color scheme throughout the paper to differentiate the samples. There are 26 galaxies in common between the two samples.}
\label{fig:sample}
\end{figure}
\subsection{The fastest rotators}
Our rotational velocities come from the $W_{mx}$ parameter:
\begin{equation}
{\rm v}_{rot}=\frac{W_{mx}/2}{\sin (i)}
\end{equation}
where v$_{rot}$ is the galaxy rotational velocity and $i$ the inclination to the line of sight.
Among the full sample of galaxies (not yet implementing the {\sl Spitzer} imaging requirement), the rotational velocities range from 11 to 514 km s$^{-1}$ for PGC71392/UGC12591 (see Figure \ref{fig:PGC0071392}) with a median of 145 km s$^{-1}$. Again the top ranked galaxy is not in our final sample due to the lack of infrared photometry, and so the top ranked galaxy in our sample is PGC17625 with a rotation velocity of 450 km s$^{-1}$. The distribution of rotational velocities of our sample is also illustrated in Figure \ref{fig:sample}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{PGC0071392-UGC12591-shg2005-AOlf}
\includegraphics[scale=0.5]{PGC0071392}
\caption{H{\small I} profile of PGC0071392 as observed at Arecibo, and B band image from DSS2 at the size of Arecibo beam (3 arcmin). This is the largest rotational velocity recorded in the full available sample, but is not in our subsample due to lack of {\sl Spitzer} photometry.}
\label{fig:PGC0071392}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig101.eps}
\caption{
Baryonic Tully Fisher. Comparison of the BTF for the S$^4$G sample and our subsamples of extreme disks.
The red filled circles represent the fast rotators, while the blue open circles represent the H{\small I}-massive galaxies. The red solid line is described in detail by \citet{zbtf} but in brief represents a model where a fixed fraction ($\sim$ 0.4) of the baryons in a halo of circular velocity of v$_{rot}$, which has the cosmological baryon fraction, are settled in the disk. As shown by those authors, this simple model is consistent with the empirical best fit line.
The dotted lines represent factors of two change in that fixed fraction.}
\label{fig:btf}
\end{figure}
\section{Discussion}
\subsection{The baryonic Tully-Fisher relationship}
Because in any sample there will always be some galaxies with the largest H{\small I} masses or rotation velocities, we require a fiducial against which to compare these galaxies to the average galaxies. Are they simply scaled up versions of lesser galaxies or are they otherwise distinct?
This class of question can be addressed using galaxy scaling relations. For rotationally supported galaxies the most commonly used such relation is that identified by \citet{1977A&A....54..661T}, which relates rotational velocity to optical luminosity. One can then use such a relation to ask such questions as whether these galaxies have unusual luminosites for their rotational velocities.
Since the original work, this scaling relation has been extended in one key way that has been shown to be particularly necessary for lower mass galaxies, namely the substitution of the baryonic mass in place of the luminosity \citep{freeman,walker,mcgaugh00,verheijen,geha}. For our purposes, we will use the baryonic Tully-Fisher (BTF) measured for S$^4$G galaxies \citep{zbtf} as our fiducial because the data used to establish that relation are most comparable to the data presented here.
To implement the BTF we require measurements of the stellar and gaseous masses. We described in \S2.2 how we obtain M$_*$. Obtaining a measurement of the gaseous mass, M$_{\rm GAS}$, involves correcting the H{\small I} mass for the missing He, other ``metals", and molecular gas. This is done here as described by \cite{zbtf} to enable a direct comparison to their results. The baryonic mass is then simply ${\rm M}_* + {\rm M}_{\rm GAS}$. As noted by those authors, this quantity is more accurately referred to as the ``condensed" baryonic mass because it only represents the baryons that have settled into the disk of the galaxy and is missing the possibly substantial baryonic components in the halos of these galaxies.
In Figure \ref{fig:btf} we present our first comparison of the properties of the giant galaxies to the more general population. As shown, both the fastest rotators and most H{\small I} massive galaxies lie generally on the BTF relation, although some differences exist. First, although slightly difficult to notice from this Figure but easier to spot in Figure \ref{fig:hist} where we plot the residuals from the mean BTF ($\Delta$TF$_{\rm Baryons}$), the fastest rotators fall somewhat below the mean trend as depicted by the red line in Figure \ref{fig:btf}. Second, and much clearer, the H{\small I}-massive sample lies systematically above the line. The one striking outlier,
PGC90167, has a T-Type $= -2$. Because it is such an early-type galaxy, its dynamical support is probably not dominated by rotational support and therefore the galaxy should not be expected to satisfy the BTF (it has a low v$_{rot}$ of 45 km s$^{-1}$).
We remove this one galaxy from further consideration, leaving both the fast rotator and H{\small I}-massive samples with 99 galaxies, and proceed to discuss the two subsamples and their location relative to the mean BTF relation.
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig103.eps}
\caption{Deviations from the mean BTF relation, $\Delta {\rm TF}_{Baryon}$, for the BTF shown in Figure \ref{fig:btf}, for galaxies with v$_{rot} > 2$. Upper panel shows the residual distribution for the S$^4$G sample from \citet{zbtf}, middle panel for our subsample of fast rotators, and lower panel for our subsample of H{\small I}-massive galaxies. Also shown in the lower panel, with the opposite hashing is the subpopulation of H{\small I}-massive galaxies that are also in the fast-rotator sample. The small horizontal bars placed above each histogram represent the mean of each distribution and the length represents the standard deviation of that mean.}
\label{fig:hist}
\end{figure}
There are a variety of potential explanations for why the fast rotators fall below the mean BTF:
\smallskip
\noindent
First, this result could be a manifestation of a Malmquist-like bias. Because we select this subsample using v$_{rot}$, errors that artificially inflate v$_{rot}$ will preferentially contaminate our sample with galaxies that appear to be ``baryon poor" for their rotation velocity. The mean of the distribution (Figure \ref{fig:hist}) is displaced by $\sim$ 0.15 dex, suggesting that we would need to have a similar magnitude error in v$_{rot}$ for this to be a plausible explanation. That level of error translates to an error of $\sim$ 40\% or over 100 km s$^{-1}$ for v$_{rot} = 300$ km s$^{-1}$. Given our demands on precisely determined H{\small I} widths and high inclinations, this level of error seems unlikely.
\smallskip
\noindent
Second, inclination errors, in the form of an underestimation of the true inclination, could lead us to preferentially place objects to the right of the mean BTF. We test for this effect by searching for a correlation between inclination and $\Delta$TF$_{\rm Baryons}$. There is only a very weak positive correlation (positive is in the necessary sense) that is entirely consistent with arising at random (can be rejected as arising randomly with only 37\% confidence).
We conclude there is no evidence that inclination is systematically affecting the BTF residuals of the high v$_{rot}$ galaxies.
\smallskip
\noindent
Third, at these high v$_{rot}$ values we may have a larger fraction of galaxies for which pressure supported components play a larger dynamical role. \citet{2007MNRAS.381.1463N}, when examining the high mass end of the Tully-Fisher (TF) relation, found that a number of their galaxies were S0/Sa types (as well as finding a similar displacement of these massive galaxies below the general TF). When dealing with such galaxies a number of investigators \citep{burstein,weiner,kassin,zall} have suggested combining the use of v$_{rot}$ and velocity dispersion, $\sigma$, into a single kinematic term that measures the full dynamical support. While this may be appropriate when observing stellar components, it is not for gaseous measurements (there is negligible pressure support of the gas component and, even if there were, such a correction would move these galaxies even further to the right in the Figure). If one does attribute the dip below the BTF for this subsample to the prevalence of early-type galaxies at these rotation velocities, the cause of the dip would have to be a general failure of (or departure from) the linear BTF, rather than from the omission of $\sigma$ in the evaluation of the dynamical support.
\smallskip
\noindent
Fourth, the intrinsic BTF may deviate from a straight line. Numerical simulations \citep{aumer} also find that the most massive galaxies tail off below the BTF \citep{zbtf} and so this phenomenon may be a natural result, for which an intuitive explanation might provide significant insight to the nature of galaxy formation. As a potentially interesting aside, massive early type galaxies also appear to fall below their analogous scaling relation, the Fundamental Plane \citep{bernardi}. As for the early types, for which curvature in the scaling relation is established \citep{zfm,bernardi}, a unified scaling relation \citep{zall,zrev} would imply the same phenomenon for late types.
\smallskip
\noindent
Finally, the position of these galaxies may indicate that the stellar mass has been underestimated in these systems relative to that of the lower v$_{rot}$ systems.
This error could arise because the stellar initial mass function (IMF) is more bottom heavy in the faster rotators. Although this hypothesis may seem farfetched,
there is empirical evidence for this type of behavior among early type galaxies \citep{vandokkum,cappellari}, where the IMF becomes progressively more bottom heavy as one considers galaxies with larger velocity dispersions. Evidence for IMF variations even extends to Local Group clusters \citep{z12,z14}, and so the possibility that such variations exist among disk galaxies as well should not be quickly dismissed.
As already mentioned, the more striking departure from the BTF is seen for the H{\small I}-massive subsample. Here the departure, shown clearly in Figures \ref{fig:btf} and \ref{fig:hist}, can be interpreted to mean that these galaxies have a larger fraction of their cosmologically apportionment of baryons within their disks, as stars or gas, than the typical disk galaxy. As with the fast rotator sample, there are a set of potential systematic errors that could also produce this result:
\smallskip
\noindent
First, one might wonder whether this could be the result of a Malmquist-like bias. We select systems with large H{\small I} masses, and therefore errors in the measured H{\small I} mass could both help populate the sample and make it appear as if these galaxies are H{\small I} rich for their rotation velocity. However, these galaxies also have larger stellar masses than typical, as shown in Figure \ref{fig:stars} where we compare to a stellar mass only version of the BTF (renormalized to produce a mean departure of zero for galaxies with log v$_{rot} > 2$ in S$^4$G). Because the stellar masses, and this stellar version of the BTF, are entirely independent of M$_{\rm HI}$, we conclude that the departure of these galaxies from the BTF is not a result of errors in the H{\small I} measurements themselves.
\smallskip
\noindent
Second, errors in the adopted distance would affect both the gaseous and stellar masses because distance plays a role in converting between observables and masses. The mean shift between either the gaseous or stellar masses and the fiducial of the S$^4$G sample is $\sim 0.22$ dex (see Figures \ref{fig:hist} and \ref{fig:stars}). Because both mass estimates are proportional to $D^2$, this offset corresponds to a systematic offset of 0.11 dex in $D$ or alternatively a 30\% error. We know such an error cannot arise from H$_0$ given the current degree of uncertainty in the Hubble parameter \citep{freedman}. Peculiar motions could be as large as several thousand km s$^{-1}$ in the richest galaxy clusters, but are typically a few hundred km s$^{-1}$ \citep{peebles,kaiser}, particularly in the low density environments that these galaxies reside in. One might dismiss peculiar motions on the basis that these should result in increased scatter rather than an offset, but recall that we select the most H{\small I} massive galaxies, and therefore are predisposed toward galaxies with overestimated distances. However, even with peculiar motions that are as large as 500 km s$^{-1}$, the induced error is $<$ 30\% in 97 out of the 99 galaxies in the H{\small I}-massive sample. We conclude that distance errors are not responsible for the systematic offset of H{\small I} massive galaxies from the BTF.
\smallskip
\noindent
Third, inclination errors will affect v$_{rot}$. However, unlike in the previous discussion, inclination does not play a role in the selection of these galaxies (other than the requirement that they be inclined by more 45$^\circ$). As before, we test for the influence of inclination by searching for a correlation between measured inclination and the offset from the BTF. Again we find no significant correlation (the probability of the observed correlation arising at random is 18\%). We conclude that inclination errors are not driving the BTF residuals.
\smallskip
\noindent
Finally, the adopted relation between v$_{rot}$ and halo mass may be incorrect. We have the taken a mean trend from \cite{bullock}, normalized using the Milky Way \citep[see][]{zbtf}, but different types of galaxies will have different degrees of adiabatic contraction, and therefore different relations between the inner rotation curve, v$_{rot}$, and the characteristic
halo circular velocity, v$_c$. A factor of two overestimation in the fraction of baryons within the halo that have settled into a disk, $f_C$, would arise if we had underestimated the halo mass of galaxies in our subsample by a factor of two, or alternatively underestimated v$_{c}$ by $\sqrt{2}$. Because it is difficult to imagine a scenario by which v$_{rot}$ is smaller than that due to the dark matter, we would posit that the problem, if it exists at all, is that the v$_c$ of the typical galaxies are overestimated by a factor of $\sqrt{2}$. For a typical galaxy in our sample with v$_{rot} \sim 180$ km s$^{-1}$, the necessary error would require the halo to have $v_{rot} \sim 127$. Such a steep drop between v$_{rot}$ and v$_c$ is not found in dynamical analyses \citep{zaritsky94} or gravitational lensing \citep{fischer}. While the conversion between v$_{rot}$ and v$_c$ has its associated systematic uncertainties, we conclude that it is unlikely to globally cause a factor of two increase in the apparent $f_C$ of the H{\small I} massive sample.
In the upper panel of Figure \ref{fig:efficiency}, we calculate how the BTF offset translates into differences in the level of baryon ``condensation" in galaxies \citep[see][for the original discussion of this term and the calculation relative to the BTF]{zbtf} as a function of v$_{rot}$. In other words, the Figure shows the fraction of baryons --- stars and gas --- apportioned to a given halo at the level of the universal baryon fraction that settle into the disks of these galaxies. In the lower panel of the same Figure, we plot the star formation efficiency, or alternatively the fraction of the ``condensed" mass that is turned into stars, $f_* \equiv {\rm M}_*/({\rm M}_*+{\rm M}_{GAS})$.
The H{\small I} massive galaxies are clearly distinct from the typical disk galaxy in that the fraction of condensed baryons, which in some cases reaches nearly 100\% of the cosmological baryon fraction (meaning that all of the baryons expected in a halo of the given mass are in gas and stars within the disks), is significantly larger over nearly the full range of v$_{rot}$.
We speculate that the reason that the difference becomes less striking at higher v$_{rot}$ is that the H{\small I} mass limit on which we select these galaxies is approaching the typical H{\small I} mass for these more massive galaxies.
In contrast, the star formation efficiency is remarkably similar for the H{\small I} massive galaxies and typical S$^4$G galaxies, even as that efficiency appears to depend on v$_{rot}$.
The rise in $f_*$ is interesting but difficult to interpret because it probably also reflects a change in the morphological mix of galaxies. It is nevertheless, particularly interesting that among the H{\small I} massive galaxies, the efficiency follows that of the S$^4$G galaxies despite variations of a factor of two in $f_C$. This result suggests that star formation is strongly governed by the available gas mass and internal regulation, a result that is at least broadly reminiscent of results obtained in detailed studies of star formation rates in galaxies \citep{kennicutt}, rather than external factors.
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig104.eps}
\caption{Deviations from the stellar mass Tully-Fisher relation, $\Delta {\rm TF}_{Stellar}$, where the stellar TF is in effect the classic TF because the stellar masses are recovered directly from the luminosities. Panels and lines indicating means of the distribution are as in Figure \ref{fig:hist}. The deviation of the H{\small I} selected galaxies is as strong as in Figure \ref{fig:hist} even though these residuals are independent of H{\small I} mass.}
\label{fig:stars}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.4]{fig105.eps}
\caption{The condensed baryon fraction, $f_C$, which is defined as the fraction of all the cosmologically expected baryons in a halo that are present in the disk as either cold gas or stars, as a function of v$_{rot}$ (upper panel) and the star formation efficiency, $f_*$, which is defined as the fraction of condensed baryons that are in stars ($M_*/(M_*+M_{GAS}))$. In this figure blue open circles represent the H{\small I} selected sample and black filled circles represent the S$^4$G sample. Points only plotted if bin has at least 3 galaxies in it. Error bars represent error in the mean.}
\label{fig:efficiency}
\end{figure}
\subsection{Environment}
Given our general understanding that massive galaxies tend to be early types, the
physically interesting question is why the massive galaxies we have selected are such strong examples of late type galaxies. If the two primary drivers of galaxy evolutionary differences among galaxies are mass and environment, and if mass is not the cause of the differences seen here, then
environment has somehow played a dominant role for these galaxies. Confirming this conclusion is, however, difficult. By the nature of the selection of the sample (isolated so that the H{\small I} beam is clean and the result is a well-defined double horned profile), we have already selected against galaxies in dense environments. Upon examination, we confirm that none of these galaxies is in a highly overdense environment, such as an Abell cluster.
Looking a bit further, we searched for differences in the environments of these galaxies relative to galaxies in the 2MASS redshift catalog (2MRS) \citep{2012ApJS..199...26H} by comparing the correlation function of our galaxies to that of similar galaxies in the redshift catalog. We found no convincing evidence of a difference.
We face two difficulties in exploring the role of environment further. First, environment means many different things. The local environment, whether a galaxy is a parent or satellite, appears to
play an important role in the observable characteristics of galaxies \citep{berlind05,tinker12}. On the other hand, the large scale (several Mpc) environment also appears to play a role \citep{gomez}. Therefore, any one measure of environment will be a far from perfect diagnostic of the role of environment in evolution. Second, galaxies, particularly most massive galaxies, have inhabited many different environments over their lifetime. The environment we measure today, especially for those in denser than average environments, is not a complete measure of the environmental history of a galaxy \citep{delucia}. Both of these issues suggest that standard analyses may at best yield only subtle differences, even if environment does play a role in a galaxy's evolution.
To pursue this topic further, we explore a novel characterization of environment using the local shear field, as used for the Laniakea discovery \citep{Tully2014Nature}.
A Wiener Filter reconstruction applied to the Cosmicflows-2 database \citep{2013AJ....146...86T} recovers the underlying 3D velocity field with an effective resolution of a few Megaparsecs \citep{2012ApJ...744...43C}. By taking spatial derivatives of the velocity field, we compute the dimensionless shear tensor. The tensor's eigenvalues ($\lambda_1$, $\lambda_2$, $\lambda_3$) describe the strength (relative to the Hubble expansion) of compression (positive values) or expansion (negative values) along the eigenvectors of the shear: $\hat{e}_1$, $\hat{e}_2$ , and $\hat{e}_3$.
At each position in space, we calculate the three eigenvalues of the velocity shear tensor. By ordering these eigenvalues from most positive to most negative, we set thresholds that capture four possibilities. Flows can be inward on all three axes, the condition of a cluster, inward on two axes and outward on the third, the condition of a filament, inward on one axis and outward on two, hence a sheet, or outward on all three axes, hence a void. We define boundaries around contiguous regions with the same shear properties and the contours outline the cosmic web as reconstructed by the V-web algorithm \citep{2012MNRAS.425.2049H}. The statistical uncertainty of the shear eigenvectors and eigenvalues are controlled by means of an ensemble of constrained realizations. The spatial cosmography and reconstructed velocity field of the giant galaxies' environment is shown in Figure~\ref{V-web}.\\
We compute the number of galaxies in cells that are classified as knots, filaments, sheets, or nothing (voids or no signal) for two redshift catalogs V8K \citep{Courtois2013} (the most complete redshift catalog within 8,000 km/s) and 2MRS \citep{2012ApJS..199...26H}.
Some of the studied volume is without signal (for example the zone of avoidance). It is classified as a void in the V-web, but we do not have galaxies there. By confining the analysis to within 8000 km/s, our samples of H{\small I} massive and fast rotators now contain 18 and 17 galaxies, respectively, see Tables 2, 3 and 4.
As seen in Table\ref{Vweb}, our H{\small I} giant galaxies live preferentially in filaments: 10 out of 18 (56\%) while only 21\% of all galaxies within the same volume located today in a filament. Filaments represent 10\% of the volume, contain 21\% of all galaxies, and 56\% of the giants. Adopting a binomial distribution with the probability of a ''true" output (filament) defined to be 0.21, we find that the probability of having 10 or more ''true" outcomes is 0.00025. so the signal is significant at greater than 3$\sigma$ confidence.
In contrast, although our fast rotators also show a hint of a preference for filaments, there the result is only statistically significant at $\sim 1\sigma$ confidence.
\begin{table*}
\caption{Environmental study using the Cosmic-V-web computed within a sphere of 8,000 km/s radius. Number and percentage (respectively)
of galaxies located in structures dynamically classified as knots, filaments, sheets and voids. The most massive H{\small I} galaxies are preferentially located in filaments,
with a frequency that is a factor of two larger than a random galaxy.}
\begin{tabular}{lrrrr}
\hline
Objects & in knots & in filaments & in sheets & in void cells \\
\hline
HI massive & 0 (0\%) & 10 (56\%)& 4 (22\%)& 4 (22\%)\\
Fast rotators & 1 (6\%) & 5 (29\%) & 8 (47\%) & 3 (18\%)\\
2MRS galaxies & 586 (2\%) & 8065 (21\%)& 18236 (47\%) & 11483 (30\%)\\
V8k galaxies & 500 (2\%) & 6408 (21\%)& 14468 (47\%) & 9324 (30\%)\\
Vweb cells & 2459 (0.2\%) & 112639 (10\%) & 451487 (42\%) & 518007 (48\%)\\
\hline
\end{tabular}
\label{Vweb}
\end{table*}
\begin{figure}
\centering
\hspace{-0.5cm}
\includegraphics[width=7cm]{SDvision_giants_vweb_paper_Fig1_v001}\\
\vspace{-4cm}
\includegraphics[width=7cm]{SDvision_giants_vweb_paper_Fig2_v003}
\vspace{-2cm}
\caption{The environment is studied using the Velocity-Cosmic-Web (V-web). Regions gravitationally collapsing along one direction are sheets, along two orthogonal directions are filaments, along three orthogonal directions are knots. Regions expanding in three directions are voids. Contours in grey, blue and red are showing regions classified respectively as sheets, filaments and knots. The top panel shows a 3D view. The bottom panel is a closer view of a slice of 1000 km/s width on Supergalactic Y axis. The galaxies in 2MASS redshift catalog (2MRS) are plotted with black dots. The cosmography is given with Laniakea on the left and Perseus-Pisces supercluster on the right. The most massive HI galaxies (blue spheres) are located preferentially in filaments with a $\sim$ 100\% higher probability than random galaxies in two different redshift catalogs, V8K or 2MRS. For the fast rotators (red spheres) we cannot confidently identify a preferred environment. }
\label{V-web}
\end{figure}
\section{Conclusions}
We investigated the properties of galaxies that potentially challenge the current paradigm of galaxy evolution. Albeit rare, these massive disk galaxies run counter to the idea that mass alone determines a galaxy's star formation rate or its morphology. Of course there are other previously known reasons to disfavor such an extreme view, but well-specified, carefully defined samples have the potential to highlight these discrepancies. We selected the 100 most H{\small I} massive galaxies and the 100 most rapidly rotating disk galaxies from H{\small I} catalogs available for public use at the Extragalactic Distance Database that contains thousands of coherently measured H{\small I} profiles.
To compare with the overall disk population, we use the baryonic Tully-Fisher (BTF) relation as a benchmark. That relation holds broadly because a nearly constant fraction of the available baryons in each galaxy's halo settle into the disk \citep{zbtf}. The classic Tully-Fisher (TF) relation then holds because for most galaxies a fixed fraction of those settled, or ``condensed", baryons turn into stars. Deviations from the BTF therefore indicate a variation in the fraction of condensed baryons, $f_C$, which could signal interesting evolutionary deviations from the typical galaxy. Deviations from the TF relation could similarly signal deviations $f_C$, or they could signal deviations in the star formation efficiency.
Using the H{\small I} data in combination with available 3.6$\mu$m photometry from {\sl Spitzer} observations, we place our two samples on the BTF. We find both samples deviate systematically from the existing BTF. The rapid rotators tend to lie below the BTF. We discuss a number of possible reasons for this deviation, including a suggestion that the scaling relation is not linear at the massive end. Such behavior is also seen in massive early-type galaxies \citep{bernardi}. The systematic deviation might also hint at bottom heavy IMFs in more massive galaxies, again as seen in massive early type galaxies \citep{vandokkum,cappellari}.
The H{\small I} massive galaxies lie systematically above the BTF. If these deviations are interpreted as variations in $f_C$, then we conclude that there are galaxies in which $f_C$ approaches 1, or in other words the entire allotment of baryons for that halo have found their way to the disk. In spite of this, or whatever other process may be driving the deviation from the BTF, the star formation efficiency, when quantified as the fraction of disk baryons that are in stars, is the same in these systems as in more typical galaxies. We conclude that the star formation efficiency is driven by internal self-regulation, while $f_C$ can be altered by external factors.
Although mass is clearly not the driver of the morphological or star formation history differences that exist between these galaxies and other similarly massive galaxies, we have not identified the driver. Our galaxies avoid the densest environments, but so do other galaxies that are not similarly extreme. When looking in detail for environmental differences, we found
that the most massive HI detected galaxies are located preferentially in filaments.
We present the first evidence of an environmental effect on galaxy evolution using a dynamical definition of a filament.
\section{Acknowledgments}
This research has made use of the NASA/IPAC Extragalactic
Database (NED) which is operated by the Jet Propulsion
Laboratory, California Institute of Technology, under
contract with the National Aeronautics and Space
Administration. We acknowledge the usage of the HyperLeda
database (http://leda.univ-lyon1.fr). HC and
JS acknowledge support from the Lyon Institute of Origins
under grant ANR-10-LABX-66 and from CNRS under
PICS-06233. DZ acknowledges support from a NASA ADAP award (NNX12AE27G) and
an NSF AAG grant (AST-1311326).
We thank Yehuda Hoffman from the Hebrew University in Jerusalem for providing a Cosmic-V-web computation of the cosmicflows-2 database.
\clearpage
\begin{table*}
\caption{The 20 largest H{\small I}-mass galaxies}
\begin{tabular}{ccccccccccc}
\hline
PGC & Name & J2000 coordinates & Vhel & $W_{mx}$ & error on $W_{mx}$ & inclination & $V_{max}$& $M_{H{\small I}}/M_{\odot}$ & morph. type & $log(d_{25})$ \\
& & & $km\,s^{-1}$& $km\,s^{-1}$& $km\,s^{-1}$ & degrees & $km\,s^{-1}$ & 10$^{10}$ solar mass units& & in log of 0.1 arcmin\\
\hline
17625 & NGC1961 & J054204.8+692243 & 3935 & 658 & 8 & 47.0 & 449.86 & 4.46 & 4.2 & 1.65 \\
9560 & NGC0958 & J023042.8-025621 & 5739 & 570 & 10 & 78.1 & 291.26 & 3.77 & 4.9 & 1.40 \\
14030 & UGC02885 & J035302.4+353522 & 5804 & 556 & 8 & 62.4 & 313.70 & 3.35 & 5.2 & 1.63 \\
8066 & NGC0812 & J020651.6+443418 & 5161 & 435 & 16 & 71.9 & 228.82 & 3.24 & 5.9 & 1.44 \\
64458 & UGC11537 & J201838.1-000902 & 5418 & 499 & 8 & 69.7 & 266.02 & 3.09 & 5.3 & 1.18 \\
34836 & NGC3646 & J112143.1+201011 & 4246 & 508 & 16 & 63.9 & 282.84 & 2.96 & 4.6 & 1.49 \\
52328 & NGC5720 & J143833.3+504855 & 7784 & 409 & 17 & 52.1 & 259.16 & 2.80 & 3.0 & 1.28 \\
56891 & ESO136-016 & J160349.3-605840 & 5426 & 521 & 12 & 90.0 & 260.50 & 2.76 & 5.2 & 1.52 \\
24830 & UGC04625 & J085017.8+032951 & 8476 & 422 & 18 & 90.0 & 211.00 & 2.74 & 6.1 & 1.23 \\
62178 & NGC6674 & J183833.9+252230 & 3429 & 431 & 16 & 62.9 & 242.08 & 2.63 & 3.0 & 1.60 \\
16537 & UGC03218 & J050043.7+621439 & 5228 & 495 & 15 & 56.4 & 297.15 & 2.62 & 3.1 & 1.24 \\
59884 & IC4633 & J171346.9-773210 & 2945 & 318 & 13 & 62.4 & 179.42 & 2.57 & 6.0 & 1.43 \\
61791 & ESO395-002 & J182226.4-354040 & 5607 & 538 & 13 & 90.0 & 269.00 & 2.56 & 3.4 & 1.32 \\
19531 & NGC2280 & J064449.1-273819 & 1897 & 384 & 7 & 66.2 & 209.84 & 2.49 & 5.9 & 1.81 \\
7387 & NGC0753 & J015742.2+355457 & 4902 & 314 & 8 & 52.5 & 197.89 & 2.41 & 4.9 & 1.14 \\
52361 & UGC09437 & J143910.6+184247 & 14401 & 404 & 16 & 65.6 & 221.81 & 2.38 & 3.3 & 0.90 \\
54250 & NGC5833 & J151153.6-725134 & 3031 & 407 & 14 & 76.7 & 209.11 & 2.36 & 4.2 & 1.49 \\
9399 & NGC0931 & J022814.5+311840 & 4999 & 422 & 9 & 81.3 & 213.45 & 2.30 & 3.6 & 1.39 \\
18089 & UGC03375 & J055525.3+515438 & 5791 & 490 & 11 & 64.2 & 272.13 & 2.29 & 5.2 & 1.18 \\
28196 & NGC2998 & J094843.6+440453 & 4772 & 374 & 7 & 61.8 & 212.18 & 2.27 & 5.2 & 1.39 \\
\hline
\end{tabular}
\label{tb1}
\end{table*}
\clearpage
\begin{table*}
\caption{The 20 fastest rotators}
\begin{tabular}{ccccccccccc}
\hline
PGC & Name & J2000 coord. & Vhel & $W_{mx}$ & error on $W_{mx}$ & inclination & $V_{max}$ & $M_{H{\small I}}/M_{\odot}$ & morph. type & $log(d_{25}) $ \\
& & & $km\,s^{-1}$ & $km\,s^{-1}$ & $km\,s^{-1}$ & degrees & $km\,s^{-1}$ & 10$^{10}$ solar mass units & & log of 0.1 arcmin\\
\hline
17625 & NGC1961 & J054204.8+692243 & 3935 & 658 & 8 & 47.0 & 449.86 & 4.46 & 4.2 & 1.65 \\
42407 & NGC4594 & J123959.4-113723 & 1095 & 733 & 17 & 59.4 & 425.80 & 0.0430 & 1.1 & 1.93 \\
63575 & NGC6824 & J194340.7+560634 & 3549 & 544 & 12 & 45.8 & 379.41 & 0.148 & 2.3 & 1.28 \\
9359 & NGC0936 & J022737.4-010921 & 1436 & 545 & 11 & 50.5 & 353.15 & 0.0907 & -1.2 & 1.65 \\
165398 & PGC165398 & J043157.1+592547 & 4626 & 489 & 20 & 45.5 & 342.80 & 0.440 & 0.0 & 0.94 \\
58596 & NGC6195 & J163632.6+390141 & 9021 & 517 & 18 & 50.2 & 336.47 & 0.221 & 3.1 & 1.16 \\
65375 & NGC6962 & J204719.1+001915 & 4222 & 475 & 13 & 45.0 & 335.88 & 1.12 & 1.7 & 1.43 \\
24685 & ESO563-021 & J084717.0-200208 & 4582 & 664 & 7 & 90.0 & 332.00 & 2.17 & 4.3 & 1.48 \\
26512 & NGC2841 & J092202.5+505837 & 633 & 592 & 12 & 65.2 & 326.07 & 0.236 & 3.0 & 1.84 \\
25161 & NGC2713 & J085720.5+025517 & 3916 & 629 & 18 & 77.7 & 321.89 & 0.503 & 2.5 & 1.53 \\
36706 & NGC3884 & J114612.2+202330 & 6947 & 462 & 20 & 45.9 & 321.67 & 0.660 & 0.6 & 1.28 \\
52665 & NGC5746 & J144456.0+015717 & 1723 & 634 & 11 & 90.0 & 317.00 & 0.456 & 3.0 & 1.86 \\
57173 & UGC10205 & J160640.2+300556 & 6562 & 541 & 20 & 59.0 & 315.58 & 0.725 & 1.0 & 1.16 \\
14030 & UGC02885 & J035302.4+353522 & 5804 & 556 & 8 & 62.4 & 313.70 & 3.35 & 5.2 & 1.63 \\
37617 & NGC3992 & J115735.9+532228 & 1048 & 459 & 8 & 47.4 & 311.78 & 0.218 & 4.0 & 1.91 \\
72233 & UGC12755 & J234349.7+282021 & 8794 & 503 & 19 & 54.7 & 308.16 & 0.648 & 3.1 & 1.07 \\
67966 & UGC11893 & J220406.7+355618 & 5589 & 606 & 20 & 80.2 & 307.48 & 0.924 & 6.3 & 0.81 \\
66880 & UGC11758 & J213057.6+135910 & 8635 & 612 & 19 & 90.0 & 306.00 & 0.942 & 4.2 & 1.12 \\
45947 & NGC5032 & J131327.0+274808 & 6413 & 524 & 19 & 59.5 & 304.08 & 0.212 & 3.0 & 1.29 \\
63286 & UGC11455 & J192956.3+720646 & 5392 & 608 & 15 & 90.0 & 304.00 & 1.29 & 5.8 & 1.39 \\
\hline
\end{tabular}
\label{tb2}
\end{table*}
\clearpage
\begin{table*}
\caption{HI-21cm spectrum and Spitzer $3.6 \mu$ images of the top 20 most heavy HI-mass galaxies, in decreasing mass order.}
\begin{tabular}{cccc}
\hline
PGC - Name &
HI mass ($log(M_{HI}$) in solar units and V$_{hel}$ &
PGC - Name &
HI mass ($log(M_{HI}$) in solar units and V$_{hel}$ \\
Telescope / HI profile &
Spitzer 3.6 $\mu$ IRAC 400X400 pixels &
Telescope / HI profile &
Spitzer 3.6 $\mu$ IRAC 400X400 pixels \\
\hline
\hline
PGC17625 & 4.46 / 3935~$km\,s^{-1}$ & PGC9560 & 3.77 / 5739~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0017625-NGC01961-ctf2009-GBT.png}&
\includegraphics[width=3cm]{PGC17625.jpeg}&
\includegraphics[width=4cm]{PGC0009560-NGC000958-archi12-PAKS.png}&
\includegraphics[width=3cm]{PGC9560.jpeg}\\
GBT & & Parkes 64m\\
\hline
PGC14030 & 3.35 / 5804~$km\,s^{-1}$ &
PGC8066 & 3.24 / 5161~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0014030-UGC02885-ctf2009-GBT.png}&
\includegraphics[width=3cm]{PGC14030}&
\includegraphics[width=4cm]{PGC0008066-UGC01598-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC8066}\\
GBT && GB300 \\
\hline
PGC64458 & 3.09 / 5418 ~$km\,s^{-1}$ &
PGC34836 & 2.96 / 4246 ~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0064458-UGC11537-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC64458}&
\includegraphics[width=4cm]{PGC0034836-UGC06376-shg2005-GB140.png}&
\includegraphics[width=3cm]{PGC34836}\\
GB300 && GB140 \\
\hline
PGC52328 & 2.80 / 7784 ~$km\,s^{-1}$ &
PGC56891 & 2.76 / 5426~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0052328-UGC09439-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC52328}&
\includegraphics[width=4cm]{PGC0056891-ES136-016-archi12-PAKS.png}&
\includegraphics[width=3cm]{PGC56891}\\
GB300 && Parkes 64 m\\
\hline
PGC24830 & 2.74 / 8476 ~$km\,s^{-1}$ &
PGC62178 & 2.63 / 3429~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0024830-UGC04625-shg2005-AOlf.png}&
\includegraphics[width=3cm]{PGC24830}&
\includegraphics[width=4cm]{PGC0062178-UGC11308-shg2005-GB140.png}&
\includegraphics[width=3cm]{PGC62178}\\
Arecibo && GB140 \\
\hline
\end{tabular}
\end{table*}
\clearpage
\newpage
\begin{table*}
\caption{Table4 .. continued ...HI-21cm spectrum and Spitzer $3.6 \mu$ images of the top 20 most heavy HI-mass galaxies, in decreasing mass order.}
\begin{tabular}{cccc}
\hline
PGC - Name & HI mass ($log(M_{HI}$) in solar units and V$_{hel}$ - & PGC - Name & HI mass ($log(M_{HI}$) in solar units and V$_{hel}$\\
Telescope / HI profile & Spitzer 3.6 $\mu$ IRAC 400X400 pixels & Telescope / HI profile & Spitzer 3.6 $\mu$ IRAC 400X400 pixels \\
\hline
PGC16537 & 2.62 / 5228 ~$km\,s^{-1}$ &
PGC59884 & 2.57 / 2945~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0016537-UGC03218-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC16537}&
\includegraphics[width=4cm]{PGC0059884-IC0004633-ksk2004-PAKS.png}&
\includegraphics[width=3cm]{PGC59884}\\
GB300 && Parkes 64m \\
\hline
\hline
PGC61791 & 2.56 / 5607~$km\,s^{-1}$ &
PGC19531 & 2.49 / 1897~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0061791-ESO395-002-ctm2010-GBT.png}&
\includegraphics[width=3cm]{PGC61791}&
\includegraphics[width=4cm]{PGC0019531-NGC002280-ksk2004-PAKS.png}&
\includegraphics[width=3cm]{PGC19531}\\
GBT && Parkes 64m\\
\hline
\hline
PGC7387 & 2.41 / 4902 ~$km\,s^{-1}$ &
PGC52361 & 2.38 / 14401~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0007387-UGC01437-shg2005-AOlf.png}&
\includegraphics[width=3cm]{PGC7387}&
\includegraphics[width=4cm]{PGC0052361-UGC09437-shg2005-AOlf.png}&
\includegraphics[width=3cm]{PGC52361}\\
Arecibo && Arecibo\\
\hline
\hline
PGC54250 & 2.36 / 3031~$km\,s^{-1}$ &
PGC9399 & 2.30 / 4999~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0054250-NGC005833-ksk2004-PAKS.png}&
\includegraphics[width=3cm]{PGC54250}&
\includegraphics[width=4cm]{PGC0009399-UGC01935-shg2005-AOlf.png}&
\includegraphics[width=3cm]{PGC9399}\\
Parkes 64m && Arecibo\\
\hline
\hline
PGC18089 & 2.29 / 5791 ~$km\,s^{-1}$ &
PGC28196 & 2.27 / 4772 ~$km\,s^{-1}$ \\
\includegraphics[width=4cm]{PGC0018089-UGC03375-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC18089}&
\includegraphics[width=4cm]{PGC0028196-UGC05250-shg2005-GB300.png}&
\includegraphics[width=3cm]{PGC28196}\\
GB300 && GB300 \\
\hline
\end{tabular}
\end{table*}
\clearpage
|
\section{Introduction}\label{s-intro}
Let $\mathcal{X}$ be a finite set and $K$ be the transition matrix of a discrete time Markov chain on $\mathcal{X}$. For $t\in[0,\infty)$, set
\[
H_t=e^{-t(I-K)}=e^{-t}\sum_{i=0}^\infty \frac{t^i}{i!}K^i.
\]
If $(X_m)_{m=0}^\infty$ is a Markov chain on $\mathcal{X}$ with transition matrix $K$ and $N_t$ is a Poisson process independent of $(X_m)_{m=0}^\infty$ with parameter $1$, then $H_t(x,\cdot)$ is the distribution of $X_{N_t}$ given $X_0=x$. It is well-known that if $K$ is irreducible with stationary distribution $\pi$, then
\[
\lim_{t\ra\infty}H_t(x,y)=\pi(y),\quad\forall x,y\in\mathcal{X}.
\]
If $K$ is assumed further aperiodic, then
\[
\lim_{m\ra\infty}K^m(x,y)=\pi(y),\quad\forall x,y\in\mathcal{X}.
\]
For simplicity, we use the triple $(\mathcal{X},K,\pi)$ to denote a discrete time irreducible Markov chain on $\mathcal{X}$ with transition matrix $K$ and stationary distribution $\pi$ and use $(\mathcal{X},H_t,\pi)$ to denote the associated continuous time chain introduced above.
In this paper, we consider the convergence of Markov chains in both total variation distance and separation. Let $\mu,\nu$ be two probabilities on $\mathcal{X}$. The total variation distance between $\mu,\nu$ and separation of $\mu$ w.r.t. $\nu$ are defined by
\[
\|\mu-\nu\|_{\text{\tiny TV}}:=\max_{A\subset\mathcal{X}}\{\mu(A)-\nu(A)\},\quad
\text{sep}(\mu,\nu):=\max_{x\in\mathcal{X}}\{1-\mu(x)/\nu(x)\}.
\]
With initial state $x$, the total variation distance and separation are defined by
\[
d_{\text{\tiny TV}}(x,m):=\|K^m(x,\cdot)-\pi\|_{\text{\tiny TV}},\quad d_{\text{\tiny sep}}(x,m):=\text{sep}(K^m(x,\cdot),\pi).
\]
As these quantities are non-increasing in $m$, it is reasonable to consider the corresponding mixing time, which are defined by
\[
T_{\text{\tiny TV}}(x,\epsilon):=\min\{m\ge 0|d_{\text{\tiny TV}}(x,m)\le\epsilon\}
\]
and
\[
T_{\text{\tiny sep}}(x,\epsilon):=\min\{m\ge 0|d_{\text{\tiny sep}}(x,m)\le\epsilon\},
\]
for any $\epsilon\in(0,1)$. We define the maximum total variation distance and maximum separation by
\[
d_{\text{\tiny TV}}(m):=\max_{x\in\mathcal{X}}d_{\text{\tiny TV}}(x,m),\quad d_{\text{\tiny sep}}(m):=\max_{x\in\mathcal{X}}d_{\text{\tiny sep}}(x,m).
\]
The corresponding mixing times are defined in a similar way and are denoted by $T_{\text{\tiny TV}}(\epsilon)$ and $T_{\text{\tiny sep}}(\epsilon)$. For the associated continuous time chains, we use $d_{\text{\tiny TV}}^{(c)}$, $d_{\text{\tiny sep}}^{(c)}$, $T_{\text{\tiny TV}}^{(c)}$ and $T_{\text{\tiny sep}}^{(c)}$. The inequalities,
\[
d_{\text{\tiny TV}}(m)\le d_{\text{\tiny sep}}(m)\le 1-(1-2d_{\text{\tiny TV}}(m))^2,
\]
provide comparisons between the maximum total variation distance and maximum separation. As a consequence, one has
\[
T_{\text{\tiny TV}}(\epsilon)\le T_{\text{\tiny sep}}(\epsilon)\le 2T_{\text{\tiny TV}}(\epsilon/4),\quad \forall \epsilon\in(0,1).
\]
Those results also apply for the continuous time chain and we refer the reader to \cite{AF} for detailed discussions and to \cite{LPW08} for various techniques in estimating the mixing times.
A birth and death chain on $\{0,1,...,n\}$ with transition rates $p_i,q_i,r_i$ is a Markov chain with transition matrix $K$ satisfying
\[
K(i,i+1)=p_i,\quad K(i,i-1)=q_i,\quad K(i,i)=r_i,\quad\forall 0\le i\le n,
\]
where $p_i+q_i+r_i=1$ and $p_n=q_0=0$. Conventionally, $p_i,q_i,r_i$ are called the birth, death and holding rates at $i$. In the above setting, it is easy to see that $K$ is irreducible if and only if $p_iq_{i+1}>0$ for $0\le i<n$ and the unique stationary distribution $\pi$ satisfies $\pi(i)=c(p_0\cdots p_{i-1})/(q_1\cdots q_i)$, where $c$ is a normalizing constant such that $\sum_i\pi(i)=1$. Ding {\it et al.} proved in \cite{DLP10} that, over all initial states, separation is maximized when the chain starts at $0$ or $n$ and Diaconis and Saloff-Coste provided a formula for maximum separation in \cite{DS06}. As a consequence, the mixing time for maximum separation (and then for the maximum total variation distance) is comparable with the sum of reciprocals of non-zero eigenvalues of $I-K$. In \cite{CSal13-2}, Chen and Saloff-Coste showed that both mixing times are of the same order as the maximum expected hitting time to the median of $\pi$ over all initial distributions concentrated on the boundary points.
The cutoff phenomenon was first observed by Aldous and Diaconis in 1980s. For a formal definition, if $d$ is the total variation distance or separation either in the maximum case or with a specified initial state, a family of irreducible Markov chains $(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ is said to present a cutoff in $d$, or a $d$-cutoff, if there is a sequence of positive integers $(t_n)_{n=1}^\infty$ such that
\[
\forall \epsilon\in(0,1),\quad\lim_{n\ra\infty}\frac{T_{n,d}(\epsilon)}{t_n}=1,
\]
where $T_{n,d}$ is the mixing time in $d$ of the $n$th chain. A family that presents a cutoff in $d$ is said to have a $(t_n,b_n)$ cutoff in $d$ or a $(t_n,b_n)$ $d$-cutoff if $t_n>0,b_n>0$, $b_n/t_n\ra 0$ and
\[
\forall \epsilon\in(0,1),\quad\limsup_{n\ra\infty}\frac{|T_{n,d}(\epsilon)-t_n|}{b_n}<\infty.
\]
In either case, the sequence $(t_n)_{n=1}^\infty$ is called a cutoff time and, in the latter case, the sequence $(b_n)_{n=1}^\infty$ is called the window with respect to $(t_n)_{n=1}^\infty$. The definition of cutoffs for families of continuous time chains is similar and we refer the reader to \cite{D96cutoff,CSal08} for an introduction and a detailed discussion of cutoffs.
Return to birth and death chains. To avoid the confusion of the total variation distances (resp. separation) in the maximum case and with a specified initial states, we use $\mathcal{F}$ and $\mathcal{F}_c$ for families of birth and death chains without starting states specified and write $\mathcal{F}^L,\mathcal{F}_c^L$ and $\mathcal{F}^R,\mathcal{F}_c^R$ respectively for families of chains started at the left and right boundary states. Diaconis and Saloff-Coste obtained in \cite{DS06} a spectral criterion for the existence of the separation cutoff and we cite part of their results in the following.
\begin{thm}\cite[Theorems 5.1-6.1]{DS06}\label{t-sep}
For $n=1,2,...$, let $K_n$ be the transition matrix of an irreducible birth and death chain on $\{0,1,...,n\}$ and $\lambda_{n,1},...,\lambda_{n,n}$ be the non-zero eigenvalues of $I-K_n$. Set
\[
t_n=\sum_{i=1}^n\frac{1}{\lambda_{n,i}},\quad \lambda_n=\min_{1\le i\le n}\lambda_{n,i},\quad\sigma_n^2=\sum_{i=1}^n\frac{1}{\lambda_{n,i}^2},\quad \rho_n^2=\sum_{i=1}^n\frac{1-\lambda_{n,i}}{\lambda_{n,i}^2}.
\]
Let $\mathcal{F}$ be the family $(K_n)_{n=1}^\infty$ and $\mathcal{F}_c$ be the family of associated continuous time chains.
\begin{itemize}
\item[(1)] $\mathcal{F}_c^L$ has a separation cutoff if and only if $t_n\lambda_n\ra\infty$.
\item[(2)] Suppose $K_n(i,i+1)+K_n(i+1,i)\le 1$ for all $i,n$. Then, $\mathcal{F}^L$ has a separation cutoff if and only if $t_n\lambda_n\ra\infty$.
\end{itemize}
Furthermore, if $t_n\lambda_n\ra\infty$, then $\mathcal{F}_c^L$ has a $(t_n,\sigma_n)$ separation cutoff and, under the assumption of \textnormal{(2)}, $\mathcal{F}^L$ have a $(t_n,\max\{\rho_n,1\})$ separation cutoff.
\end{thm}
\begin{rem}\label{r-t-sep}
In Theorem \ref{t-sep}, the $(t_n,\max\{\rho_n,1\})$ separation cutoff of $\mathcal{F}^L$ is not discussed in \cite{DS06} but is an implicit result of the techniques therein. We give a proof of this fact in the appendix for completion. In the proof that there is a $(t_n,\max\{\rho_n,1\})$ separation cutoff, we show that
\[
\mathcal{F}^L \text{ has a cutoff} \quad\Lra\quad \rho_n=o(t_n)\quad\Lra\quad
\max\{\rho_n,1/\lambda_n\}=o(t_n).
\]
\end{rem}
\begin{rem}\label{r-sep2}
For any irreducible birth and death chain, it was proved in \cite{DLP10} that the maximum separation of the associated continuous time chain is attained when the initial state is any of the boundary states. This is also true for the discrete time case if the transition matrix $K$ satisfies $\min_iK(i,i)\ge 1/2$. As a result, if $\mathcal{F},\mathcal{F}_c$ and $t_n,\lambda_n$ are as in Theorem \ref{t-sep}, then
\begin{itemize}
\item[(1)] $\mathcal{F}_c$ has a maximum separation cutoff if and only if $t_n\lambda_n\ra\infty$.
\item[(2)] Assuming that $\inf_{i,n}K_n(i,i)\ge 1/2$, $\mathcal{F}$ has a maximum separation cutoff if and only if $t_n\lambda_n\ra\infty$.
\end{itemize}
\end{rem}
For cutoffs in the maximum total variation, Ding, Lubetzky and Peres provide the following criterion in \cite{DLP10}.
\begin{thm}\cite[Corollary 2 and Theorem 3]{DLP10}\label{t-tv}
Let $\mathcal{F},\mathcal{F}_c,\lambda_n$ be as in Theorem \ref{t-sep} and let $T_{n,\textnormal{\tiny TV}},T_{n,\textnormal{\tiny TV}}^{(c)}$ be the maximum total variation mixing time of the $n$th chains.
\begin{itemize}
\item[(1)] $\mathcal{F}_c$ has a maximum total variation cutoff if and only if $T_{n,\textnormal{\tiny TV}}^{(c)}(\epsilon)\lambda_n\ra\infty$ for some $\epsilon\in(0,1)$.
\item[(2)] Assume that $\inf_{i,n}K_n(i,i)>0$. Then, $\mathcal{F}$ has a maximum total variation cutoff if and only if $T_{n,\textnormal{\tiny TV}}(\epsilon)\lambda_n\ra\infty$ for some $\epsilon\in(0,1)$.
\end{itemize}
\end{thm}
\begin{rem}\label{r-tv}
For any birth and death chain started at the left or right boundary state, the total variation distance can be different and the biased random walk with constant birth and death rates is a typical example. Further, the maximum total variation distance over all initial states is not necessarily attained at boundary states and a birth and death chain with valley stationary distribution, a distribution which is decreasing on $\{0,...,M\}$ and increasing on $\{M,...,n\}$ for some $0<M<n$, could illustrate this observation. This is very different from the case of separation and we refer the readers to Sections \ref{s-bs} and \ref{s-comp} for more discussions.
\end{rem}
To state our main results, we need the following notation. For $n\in\mathbb{N}$, let $\mathcal{X}_n=\{0,1,...,n\}$ and $(X_m^{(n)})_{m=0}^\infty$ be an irreducible birth and death chain on $\mathcal{X}_n$ with transition matrix $K_n$ and stationary distribution $\pi_n$. Let $N_t$ be a Poisson process independent of $(X_m^{(n)})$ with parameter $1$. For $i\in\mathcal{X}_n$, set
\begin{equation}\label{eq-tauc}
\tau^{(n)}_i=\inf\{m\ge 0|X_m^{(n)}=i\},\quad\widetilde{\tau}^{(n)}_i=\inf\{t\ge 0|X_{N_t}^{(n)}=i\}.
\end{equation}
For $j\in\mathcal{X}_n$, let $\mathbb{E}_j$ and $\text{Var}_j$ denote the conditional expectation and variance given $X^{(n)}_0=j$.
\begin{rem}
It follows from the definition of $\tau_i^{(n)},\widetilde{\tau}_i^{(n)}$ that $\mathbb{E}_j\tau^{(n)}_i=\mathbb{E}_j\widetilde{\tau}^{(n)}_i$ for all $i,j\in\mathcal{X}_n$. See \cite{AF} for more information of the hitting times $\tau_i^{(n)},\widetilde{\tau}_i^{(n)}$.
\end{rem}
\begin{thm}\label{t-sep2}
Let $\mathcal{F},\mathcal{F}_c,\lambda_n$ be as in Theorem \ref{t-sep} and $\tau^{(n)}_i,\widetilde{\tau}^{(n)}_i$ be the hitting times in \textnormal{(\ref{eq-tauc})}. For $n\ge 1$, let $M_n\in\{0,1,...,n\}$ and set
\[
s_n=\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}+\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}
=\mathbb{E}_0\tau^{(n)}_{M_n}+\mathbb{E}_n\tau^{(n)}_{M_n}
\]
and
\[
b_n^2=\textnormal{Var}_0\widetilde{\tau}^{(n)}_{M_n}
+\textnormal{Var}_n\widetilde{\tau}^{(n)}_{M_n},\quad
c_n^2=\textnormal{Var}_0\tau^{(n)}_{M_n}
+\textnormal{Var}_n\tau^{(n)}_{M_n}.
\]
Suppose that
\begin{equation}\label{eq-pin}
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad\inf_{n\ge 1}\pi_n([M_n,n])>0.
\end{equation}
Then, the following properties hold.
\begin{itemize}
\item[(1)] $\mathcal{F}_c^L$ has a separation cutoff if and only if $s_n\lambda_n\ra\infty$ if and only if $s_n/b_n\ra\infty$. Furthermore, if $s_n/b_n\ra\infty$, then $\mathcal{F}_c^L$ has a $(s_n,b_n)$ separation cutoff.
\item[(2)] Assume that $K_n(i,i+1)+K_n(i+1,i)\le 1$ for all $i,n$. Then, $\mathcal{F}^L$ has a separation cutoff if and only if $s_n\lambda_n\ra\infty$ if and only if $s_n/c_n\ra\infty$. Furthermore, if $s_n/c_n\ra\infty$, then $\mathcal{F}^L$ has a $(s_n,\max\{c_n,1/\lambda_n\})$ separation cutoff.
\end{itemize}
\end{thm}
\begin{rem}\label{r-sep}
Let $\sigma_n,\rho_n$ be the constants in Theorem \ref{t-sep}. Let $M_n,M_n'\in\{0,1,...,n\}$ and $b_n,c_n,b_n',c_n'$ be the constants in Theorem \ref{t-sep2} defined accordingly. Suppose $M_n,M_n'$ satisfy (\ref{eq-pin}). Then,
\[
b_n\asymp b_n'\asymp \sigma_n,\quad \max\{c_n,1/\lambda_n\}\asymp\max\{c_n',1/\lambda_n\}\asymp\max\{\rho_n,1/\lambda_n\},
\]
where $u_n\asymp v_n$ means that both sequences, $u_n/v_n$ and $v_n/u_n$, are bounded. See Corollary \ref{c-sep} for a proof. Comparing Theorems \ref{t-sep} and \ref{t-sep2}, one can see that the cutoff window for $\mathcal{F}_c^L$ is unchanged up to some universal multiples but the cutoff window for $\mathcal{F}^L$ can have a bigger order in Theorem \ref{t-sep2} due to the change of the cutoff time.
\end{rem}
In total variation, we have the following result.
\begin{thm}\label{t-tv2}
Let $\mathcal{F},\mathcal{F}_c,\lambda_n$ be as in Theorem \ref{t-sep} and $\tau^{(n)}_i,\widetilde{\tau}^{(n)}_i$ be the hitting times in \textnormal{(\ref{eq-tauc})}. Let $M_n\in\{0,1,...,n\}$ and set
\[
\theta_n=\max\left\{\mathbb{E}_0\tau^{(n)}_{M_n},\mathbb{E}_n\tau^{(n)}_{M_n}\right\}
=\max\left\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n},\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}\right\}
\]
and
\[
\alpha_n^2=\max\left\{\textnormal{Var}_0\widetilde{\tau}^{(n)}_{M_n},
\textnormal{Var}_n\widetilde{\tau}^{(n)}_{M_n}\right\}
\]
and
\[
\beta_n^2=\max\left\{\textnormal{Var}_0\tau^{(n)}_{M_n},
\textnormal{Var}_n\tau^{(n)}_{M_n}\right\}.
\]
Suppose
\begin{equation}\label{eq-pin2}
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad\inf_{n\ge 1}\pi_n([M_n,n])>0.
\end{equation}
In the maximum total variation distance:
\begin{itemize}
\item[(1)] $\mathcal{F}_c$ has a cutoff if and only if $\theta_n\lambda_n\ra\infty$ if and only if $\theta_n/\alpha_n\ra\infty$. Furthermore, if $\mathcal{F}_c$ has a cutoff, then $\mathcal{F}_c$ has a $(\theta_n,\alpha_n)$ cutoff.
\item[(2)] Assume that $\inf_{i,n}K_n(i,i)>0$. Then, $\mathcal{F}$ has a cutoff if and only if $\theta_n\lambda_n\ra\infty$ if and only if $\theta_n/\beta_n\ra\infty$. Furthermore, if $\mathcal{F}$ has a cutoff, then $\mathcal{F}$ has a $(\theta_n,\beta_n)$ cutoff.
\end{itemize}
\end{thm}
\begin{rem}\label{r-wcomp2}
In Theorem \ref{t-tv2}, if $\delta=\inf_{i,n}K_n(i,i)$, then $\delta\alpha_n^2\le \beta_n^2\le\alpha_n^2$. See Remark \ref{r-wcomp} for details.
\end{rem}
\begin{rem}\label{r-exp1}
Let $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of irreducible birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$. For $a\in(0,1)$, set $M_n(a)$ be a state in $\mathcal{X}_n$ satisfying
\[
\pi_n([0,M_n(a)])\ge a,\quad\pi_n([M_n(a),n])\ge 1-a.
\]
By Theorem \ref{t-sep} and Remark \ref{r-sep2}, if $\mathcal{F}_c$ has a cutoff in maximum separation, then
\begin{equation}\label{eq-limhit1}
\lim_{n\ra\infty}\frac{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}
+\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)}}
{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(b)}+\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(b)}}=1,
\quad\forall 0<a<b<1.
\end{equation}
From Theorem \ref{t-tv2}, if $\mathcal{F}_c$ has a cutoff in the maximum total variation, then
\begin{equation}\label{eq-limhit2}
\lim_{n\ra\infty}\frac{\max\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)},
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)}\}}
{\max\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(b)},\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(b)}\}}=1,
\quad\forall 0<a<b<1.
\end{equation}
But, the converse of these statements are not necessarily true. For example, let
\[
K_n(i,i+1)=K_n(i+1,i)=1/2,\quad\forall 0<i<n,\quad K_n(n,n)=1/2,
\]
and
\[
K_n(0,1)=K_n(1,0)=\xi_n,\quad K_n(0,0)=1-\xi_n,\quad K_n(1,1)=1/2-\xi_n,
\]
where $\xi_n\in(0,1/2)$. Note that $K_n$ can be regarded as the transition matrix of a simple random walk on $\mathcal{X}_n$ with specific transitions at the boundary states and a bottleneck between $0$ and $1$ when $\xi_n$ is small. It is clear that the stationary distribution satisfies $\pi_n(i)=1/(n+1)$ for all $0\le i\le n$. After some computations, one has, for $n$ large enough,
\[
M_n(a)\asymp n\asymp (n-M_n(a)).
\]
This implies
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}=\frac{1}{\xi_n}+M_n(a)(M_n(a)+1)-2\asymp \frac{1}{\xi_n}+n^2
\]
and
\[
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)}=(n-M_n(a))[n-M_n(a)+1]\asymp n^2.
\]
Let $p_{n,i},q_{n,i},r_{n,i}$ and $\lambda_n$ be the transition rates and the spectral gap of $K_n$. By Theorem 1.2 in \cite{CSal13-2}, we have
\begin{equation}\label{eq-gap}
\frac{1}{\lambda_n}\asymp\max\left\{\max_{j:j<M_n}\sum_{k=j}^{M_n-1}\frac{\pi_n([0,j])}{\pi_n(k)p_{n,k}},
\max_{j:j>M_n}\sum_{k=M_n+1}^j\frac{\pi_n([j,n])}{\pi_n(k)q_{n,k}}\right\},
\end{equation}
where $M_n=\lfloor n/2\rfloor$. This implies
\[
\frac{1}{\lambda_n}\asymp \frac{1}{\xi_n}+n^2.
\]
As a consequence of Theorems \ref{t-sep2} and \ref{t-tv2}, $\mathcal{F}_c$ has neither a maximum separation cutoff nor a maximum total variation cutoff. Let $s_n$ and $\theta_n$ be the constants in Theorems \ref{t-sep2} and \ref{t-tv2}. If $n^2\xi_n\ra 0$, then
\[
s_n\sim \theta_n\sim\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}\sim\frac{1}{\xi_n},\quad\forall a\in(0,1).
\]
The above example illustrates that (\ref{eq-limhit1}) and (\ref{eq-limhit2}) are necessary but not sufficient for the existence of the corresponding cutoffs.
\end{rem}
The following theorem describes one of the main applications of Theorems \ref{t-sep2}-\ref{t-tv2}.
\begin{thm}\label{t-randomm}
Consider a family $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ of irreducible birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$. For $n\ge 1$, let $(\Omega_n,\mathbb{P}^{(n)})$ be a probability space and $C_{n,1},...,C_{n,n}:\Omega_n\ra(0,1)$ be independent and identically distributed random variables. For $\omega_n\in\Omega_n$ and $0\le i\le n$, let $(\mathcal{X}_n,L_n^{(\omega_n)},\pi_n)$ be a Markov chain given by
\[
\begin{cases}
L^{(\omega_n)}_n(i,i+1)=K_n(i,i+1)C_{n,i+1}(\omega_n),\\
L^{(\omega_n)}_n(i+1,i)=K_n(i+1,i)C_{n,i+1}(\omega_n),\\
L^{(\omega_n)}_n(i,i)=1-L^{(\omega_n)}_n(i,i+1)-L^{(\omega_n)}_n(i,i-1),
\end{cases}
\]
and, for $\omega=(\omega_1,\omega_2,...)\in\prod_{n=1}^\infty\Omega_n$, let $\mathcal{F}^{(\omega)}=(\mathcal{X}_n,L_n^{(\omega_n)},\pi_n)_{n=1}^\infty$. Let $\mathcal{F}_c,\mathcal{F}_c^{(\omega)}$ be the continuous time families associated with $\mathcal{F},\mathcal{F}^{(\omega)}$. For $n\ge 1$, set $\mu_n=\mathbb{E}(1/C_{n,1})$, $\nu_n^2=\textnormal{Var}(1/C_{n,1})$ and let $\theta_n,\alpha_n,\beta_n$ be the constants in Theorem \ref{t-tv2}.
\begin{itemize}
\item[(1)] If $\mathcal{F}_c$ has a maximum total variation cutoff and $\nu_n\alpha_n=o(\mu_n\theta_n)$, then there is a sequence $E_n\subset\Omega_n$ such that $\mathbb{P}^{(n)}(E_n)\ra 1$ and, for any $\omega\in\prod_{n=1}^\infty E_n$, $\mathcal{F}^{(\omega)}_c$ has a maximum total variation cutoff with cutoff time $\mu_n\theta_n$.
\item[(2)] Assuming $\inf_{n,i}K_n(i,i)>0$ and replacing $\alpha_n$ by $\beta_n$, the statement in (1) also holds for the families $\mathcal{F},\mathcal{F}^{(\omega)}$.
\end{itemize}
\end{thm}
\begin{rem}
In Theorem \ref{t-randomm}, $L_n$ can be regarded as a random birth and death chain obtained by applying i.i.d. random slowdowns on $K_n$ without changing the stationary distribution.
\end{rem}
\begin{rem}
Theorem \ref{t-randomm} also holds in maximum separation.
\end{rem}
The remaining of this article is organized in the following way. Sections \ref{s-sep} and \ref{s-tv} contain the proofs of Theorems \ref{t-sep2} and \ref{t-tv2} respectively. The proof of Theorem \ref{t-randomm} is given in Section \ref{s-rbdc}. We also introduce another randomization of simple random walks on paths and discuss its cutoff and mixing time. In Section \ref{s-bs}, we consider families of chains started at one boundary states and provide criteria for the existence of a total variation cutoff and formulas for the cutoff time. We discuss the distinction between maximum total variation cutoffs and cutoffs from a boundary state and illustrate this with several examples in Section \ref{s-comp}. The main results of Section \ref{s-bs} are proved in Section \ref{s-proof}. In Section \ref{s-ex}, we apply the developed theory to compute the cutoff time of some classical examples. Some useful lemmas and auxiliary results are gathered in the appendix.
\section{Cutoff in separation}\label{s-sep}
This section is dedicated to the proof of Theorem \ref{t-sep2} and we need the following two lemmas. The first lemma concerns the mean and variance of hitting times and the second lemma provides a comparison of spectral gaps.
\begin{lem}\label{l-meanvar}
Let $K$ be the transition matrix of an irreducible birth and death chain on $\{0,1,...,n\}$. For $1\le i\le n$, let $\beta^{(i)}_1,...,\beta^{(i)}_i$ be the eigenvalues of the submatrix of $I-K$ indexed by $\{0,...,i-1\}$ and set
\begin{equation}\label{eq-taui}
\tau_i=\min\{m\ge 0|X_m=i\},\quad\widetilde{\tau}_i=\inf\{t\ge 0|X_{N_t}=i\},
\end{equation}
where $(X_m)_{m=0}^\infty$ is a Markov chain with transition matrix $K$ and $N_t$ is a Poisson process independent of $X_m$ with parameter $1$. Then, $\beta^{(i)}_j\in(0,2)$ for all $1\le j\le i$ and
\begin{equation}\label{eq-mean}
\mathbb{E}_0\tau_i=\mathbb{E}_0\widetilde{\tau}_i=\sum_{j=1}^i\frac{1}{\beta^{(i)}_j},
\end{equation}
and
\begin{equation}\label{eq-var}
\textnormal{Var}_0(\tau_i)=\sum_{j=1}^i\frac{1-\beta^{(i)}_j}{\left(\beta^{(i)}_j\right)^2},\quad
\begin{aligned}
\textnormal{Var}_0(\widetilde{\tau}_i)=\sum_{j=1}^i\frac{1}{\left(\beta^{(i)}_j\right)^2}.
\end{aligned}
\end{equation}
\end{lem}
\begin{proof}
Let $\widetilde{K}$ be the submatrix of $K$ indexed by $\{0,1,...,i-1\}$. Let $\beta$ be an eigenvalue of $\widetilde{K}$ and $x=(x_0,...,x_{i-1})$ be a left eigenvector associated with $\beta$. That is,
\[
\begin{cases}\beta x_j=K(j-1,j)x_{j-1}+K(j,j)x_j+K(j+1,j)x_{j+1},\quad\forall 0<j<i-1,\\
\beta x_0=K(0,0)x_0+K(1,0)x_1,\\
\beta x_{i-1}=K(i-2,i-1)x_{i-2}+K(i-1,i-1)x_{i-1}.\end{cases}
\]
By the irreducibility of $K$, if $x_{i-1}=0$, then $x_j=0$ for all $0\le j<i$. This implies $x_{i-1}\ne 0$ and then
\[
|\beta|\sum_{j=0}^{i-1}|x_j|\le\sum_{j=0}^{i-1}|x_j|-K(i-1,i)|x_{i-1}|<\sum_{j=0}^{i-1}|x_j|.
\]
Since $x$ is an eigenvector of $K$, $\sum_j|x_j|>0$ and thus $|\beta|<1$. This proves that $\beta^{(i)}_j\in(0,2)$ for all $1\le j\le i$. For (\ref{eq-mean}) and (\ref{eq-var}), note that the distribution of $\widetilde{\tau}_i$ was given by Brown and Shao in \cite{BS87} and the technique therein also applies for $\tau_i$. This leads to the desired identities, where we refer the reader to their work for details.
\end{proof}
\begin{rem}\label{r-var}
In Lemma \ref{l-meanvar}, the first equality of (\ref{eq-var}) implies
\[
\sum_{j=1}^i\frac{1}{(\beta^{(i)}_j)^2}\ge \sum_{j=1}^i\frac{1}{\beta^{(i)}_j},\quad\forall j\ge 1.
\]
\end{rem}
\begin{lem}\label{l-gap}
Let $K$ be the transition matrix of an irreducible birth and death chain on $\{0,1,...,n\}$ with stationary distribution $\pi$. For $0\le i\le n$, let $L_i$ be the sub-matrix of $K$ obtained by removing the row and column of $K$ indexed by state $i$. Let $\lambda_1<\cdots<\lambda_n$ be the non-zero eigenvalues of $I-K$ and $\lambda^{(i)}_1\le\cdots\le\lambda^{(i)}_n$ be the eigenvalues of $I-L_i$. Then,
\[
\lambda^{(i)}_j\le \lambda_j\le\lambda^{(i)}_{j+1}\le\lambda_{j+1},\quad\forall 1\le j<n,
\]
and
\[
\left(\frac{\min\{\pi([0,i]),\pi([i,n])\}}{4}\right)\lambda_1\le\lambda^{(i)}_1\le\lambda_1.
\]
In particular, if $M$ is a median of $\pi$, i.e. $\pi([0,M])\ge 1/2$ and $\pi([M,n])\ge 1/2$, then $\lambda_1/8\le \lambda^{(M)}_1\le\lambda_1$.
\end{lem}
The proof of Lemma \ref{l-gap} is based on a weighted Hardy inequality obtained in \cite{CSal13-2} and is discussed in the appendix. In what follows, for any two sequences of positive reals $a_n,b_n$, we write $a_n=o(b_n)$ if $a_n/b_n\ra 0$ and write $a_n=O(b_n)$ if $a_n/b_n$ is bounded. In the case that $a_n=O(b_n)$ and $b_n=O(a_n)$, we write $a_n\asymp b_n$ instead.
\begin{proof}[Proof of Theorem \ref{t-sep2}]
Let $\lambda_{n,i},\lambda_n,t_n,\sigma_n,\rho_n$ be constants in Theorem \ref{t-sep}. Note that, for $n\ge 2$,
\[
\max\{\rho_n^2,1/\lambda_n^2\}\le \sigma_n^2=\sum_{i=1}^n\frac{1}{\lambda_{n,i}^2}\le \frac{t_n}{\lambda_n}.
\]
This implies
\begin{equation}\label{eq-comp}
\sqrt{t_n\lambda_n}\le\frac{t_n}{\sigma_n}\le\frac{t_n}{\max\{\rho_n,1/\lambda_n\}}\le t_n\lambda_n.
\end{equation}
As a consequence, we have
\begin{equation}\label{eq-equi}
t_n\lambda_n\ra\infty\quad\Lra\quad \sigma_n=o(t_n)\quad\Lra\quad \max\{\rho_n,1/\lambda_n\}=o(t_n).
\end{equation}
Next, let $s_n,b_n,c_n$ be constants in Theorem \ref{t-sep2}. Observe that
\[
1/\lambda_n\le\max\{\rho_n,1/\lambda_n\}\le \sigma_n.
\]
Set $a_n=\min\{\pi_n([0,M_n]),\pi_n([M_n,n])\}$. By Lemmas \ref{l-meanvar} and \ref{l-gap}, one has
\[
t_n\le s_n\le t_n+\frac{4}{a_n\lambda_n}\le t_n+\frac{4\sigma_n}{a_n}
\]
and
\[
\sigma_n^2\le b_n^2\le \sigma_n^2+\left(\frac{4}{a_n\lambda_n}\right)^2
\le\frac{17\sigma_n^2}{a_n^2}.\notag
\]
According to the assumption of (\ref{eq-pin}), we have $a_n\asymp 1$ and this implies
\[
t_n\lambda_n\ra\infty\quad\Lra\quad s_n\lambda_n\ra\infty
\]
and
\begin{equation}\label{eq-equi2}
|t_n-s_n|=O(\sigma_n),\quad |t_n-s_n|=O(\max\{\rho_n,1/\lambda_n\}),\quad
b_n\asymp \sigma_n.
\end{equation}
As a consequence of (\ref{eq-equi}) and (\ref{eq-equi2}), we obtain
\begin{equation}\label{eq-equi3}
t_n\lambda_n\ra\infty\quad\Lra\quad b_n=o(s_n)\quad\Lra\quad \max\{c_n,1/\lambda_n\}=o(s_n).
\end{equation}
The first equivalence of (\ref{eq-equi3}) proves the criterion for cutoff in (1). For (2), if $\mathcal{F}^L$ has a separation cutoff, then Theorem \ref{t-sep} implies $t_n\lambda_n\ra\infty$. By the last identity in (\ref{eq-equi3}), we obtain $c_n=o(s_n)$. To see the inverse direction, observe that the mapping $u\mapsto (1-u)/u^2$ is decreasing on $(0,2]$ and $\lambda_{n,i}\in(0,2)$ for all $1\le i\le n$. In the same reasoning as before, Lemmas \ref{l-meanvar} and \ref{l-gap} yield
\begin{equation}\label{eq-equi4}
\rho_n^2\le c_n^2\le\rho_n^2+\frac{1-a_n\lambda_n/4}{(a_n\lambda_n/4)^2}
+\frac{\lambda_{n,n}-1}{\lambda_{n,n}^2}\le \rho_n^2+\frac{17}{a_n^2\lambda_n^2}.
\end{equation}
By the first inequality of (\ref{eq-equi4}), if $c_n=o(s_n)$, then $\rho_n=o(s_n)$. Accompanied with the facts,
\[
s_n=t_n+\frac{4}{a_n\lambda_n}\le\left(1+\frac{4}{a_n}\right)t_n,\quad a_n\asymp 1,
\]
we obtain $\rho_n=o(t_n)$. By Remark \ref{r-t-sep}, $\mathcal{F}^L$ has a separation cutoff.
To see a window, we recall Corollary 2.5(v) of \cite{CSal08}, which says that if a family has a $(t_n,\sigma_n)$ cutoff and
\[
b_n=o(t_n)\,\, (\text{or }b_n=o(s_n)),\quad |t_n-s_n|=O(b_n),\quad \sigma_n=O(b_n),
\]
then the family has a $(s_n,b_n)$ cutoff. By Theorem \ref{t-sep}, the desired cutoff for $\mathcal{F}^L_c$ is given by the first and third identities in (\ref{eq-equi2}), while the desired cutoff for $\mathcal{F}^L$ is provided by the second identity in (\ref{eq-equi2}), the third identity in (\ref{eq-equi3}) and the following observations
\[
\max\{\rho_n,1/\lambda_n\}\asymp \max\{c_n,1/\lambda_n\},\quad \max\{\rho_n,1\}=O(\max\{c_n,1/\lambda_n\}),
\]
which are implies by (\ref{eq-equi4}) and the fact $\lambda_n\le 2$.
\end{proof}
In the following corollary, we summarize some useful comparison between the variances of hitting times and the windows of cutoffs obtained in the proof of Theorem \ref{t-sep2}.
\begin{cor}\label{c-sep}
Let $K$ be the transition matrix of an irreducible birth and death chain on $\{0,1,...,n\}$ with stationary distribution $\pi$ and $\tau_i,\widetilde{\tau}_i$ be the hitting times in \textnormal{(\ref{eq-taui})}. Suppose $\lambda_1,...,\lambda_n$ be non-zero eigenvalues of $I-K$ and set
\[
t=\sum_{i=1}^n\frac{1}{\lambda_i},\quad\sigma^2=\sum_{i=1}^n\frac{1}{\lambda_i^2},\quad \rho^2=\sigma^2-t,
\quad \lambda=\min_{1\le i\le n}\lambda_i.
\]
Then, for $0\le i\le n$,
\[
t\le\mathbb{E}_0\widetilde{\tau}_i+\mathbb{E}_n\widetilde{\tau}_i=\mathbb{E}_0\tau_i+\mathbb{E}_n\tau_i
\le t+\frac{4}{a(i)\lambda}
\]
and
\[
\sigma^2\le\textnormal{Var}_0\widetilde{\tau}_i+\textnormal{Var}_n\widetilde{\tau}_i\le\frac{17\sigma^2}
{a(i)^2},\quad \rho^2\le \textnormal{Var}_0\tau_i+\textnormal{Var}_n\tau_i\le\rho^2+\frac{17}{a(i)^2\lambda^2},
\]
where $a(i)=\min\{\pi([0,i]),\pi([i,n])\}$.
\end{cor}
To determine a cutoff time and a window using Theorem \ref{t-sep2}, one needs to compute the mean and variance of the hitting time to some state given that the chain starts at one boundary state. Explicit formulas on both terms are available using the Markov property and we summarize them in Lemma \ref{l-meanvar2}.
The next proposition discusses the cutoff times obtained in Theorem \ref{t-sep2} and provides a universal lower bound on the corresponding windows using the transition rates and the stationary distribution.
\begin{prop}\label{p-hittime}
Let $K$ be the transition matrix of a birth and death chain on $\{0,1,...,n\}$ with transition rates $p_i,q_i,r_i$. Let $\tau_i,\widetilde{\tau}_i$ be the hitting times in \textnormal{(\ref{eq-taui})} and set
\[
s(i)=\mathbb{E}_0\widetilde{\tau}_i+\mathbb{E}_n\widetilde{\tau}_i,\quad b(i)^2=\textnormal{Var}_0(\widetilde{\tau}_i)
+\textnormal{Var}_n(\widetilde{\tau}_i).
\]
Suppose $K$ is irreducible with stationary distribution $\pi$ and spectral gap $\lambda$. Let $M\in\{0,1,...,n\}$ be a state satisfying $\pi([0,M])\ge 1/2$ and $\pi([M,n])\ge 1/2$. Then, for $0\le i\le j\le M$,
\begin{equation}\label{eq-hittime1}
s(i)-s(j)=\sum_{\ell=i}^{j-1}\frac{1-2\pi([0,\ell])}{p_\ell\pi(\ell)}\ge 0,
\end{equation}
and, for $0\le i\le n$,
\begin{equation}\label{eq-hittime2}
b(i)\ge\frac{1}{\lambda}\ge\frac{1}{2}\max_{0\le j\le M\le k\le n}\max\left\{\sum_{\ell=j}^{M-1}\frac{\pi([0,j])}{p_\ell\pi(\ell)},
\sum_{\ell=M+1}^k\frac{\pi([k,n])}{q_\ell\pi(\ell)}\right\}.
\end{equation}
\end{prop}
\begin{proof}
(\ref{eq-hittime1}) is given by Lemma \ref{l-meanvar2} and the first inequality of (\ref{eq-hittime2}) is obvious from Lemmas \ref{l-meanvar}-\ref{l-gap}, while the second inequality of (\ref{eq-hittime2}) is cited from Theorem A.1 of \cite{CSal13-2}.
\end{proof}
\begin{rem}
Let $s_n,t_n$ be the constants in Theorems \ref{t-sep}-\ref{t-sep2}. By Corollary \ref{c-sep}, one has $s_n-t_n\ge 0$ and, by (\ref{eq-hittime1}), the difference $s_n-t_n$ is minimized when $M_n$ satisfies
\[
\pi_n([0,M_n])\ge 1/2,\quad \pi_n([M_n,n])\ge 1/2.
\]
\end{rem}
\section{Cutoff in total variation}\label{s-tv}
This section is dedicated to the proof of Theorem \ref{t-tv2}. Throughout the rest of this article, we will write $\mathbb{P}_i$ to denote the probability given the initial state $i$. First, recall two useful bounds on the total variation.
\begin{lem}\cite[Proposition 3.8 and Equation (3.5)]{CSal13-2}\label{l-tv}
Consider a continuous time birth and death chain on $\{0,1,...,n\}$ with stationary distribution $\pi$. For $0\le i\le n$, let $\widetilde{\tau}_i$ be the first hitting time to state $i$ and $d^{(c)}_{\textnormal{\tiny TV}}(i,t)$ be the total variation distance at time $t$ with initial state $i$. Then, for $0\le i\le n$ and $0\le j\le k\le n$,
\[
d^{(c)}_{\textnormal{\tiny TV}}(i,t)\le \mathbb{P}_i(\max\{\widetilde{\tau}_j,\widetilde{\tau}_k\}>t)+1-\pi([j,k])
\]
and
\[
d^{(c)}_{\textnormal{\tiny TV}}(0,t)\ge\mathbb{P}_0(\widetilde{\tau}_i>t)-\pi([0,i-1]).
\]
\end{lem}
Based on the above lemma, we may bound the maximum total variation mixing time using the expected hitting times.
\begin{thm}\label{t-tvmix}
Let $\pi,\widetilde{\tau}_i$ be as in Lemma \ref{l-tv} and set
\[
\theta(i)=\max\{\mathbb{E}_0\widetilde{\tau}_i,\mathbb{E}_n\widetilde{\tau}_i\},\quad \alpha(i)^2=\max\{\textnormal{Var}_0\widetilde{\tau}_i,\textnormal{Var}_n\widetilde{\tau}_i\}.
\]
The maximum total variation mixing time satisfies
\[
T^{(c)}_{\textnormal{\tiny TV}}(\epsilon_1)\le \theta(j)+\mathbb{E}_j\widetilde{\tau}_k+\mathbb{E}_k\widetilde{\tau}_j
+\sqrt{\left(\tfrac{2}{\delta}-1\right)}\max\{\alpha(j),\alpha(k)\}
\]
and
\[
T^{(c)}_{\textnormal{\tiny TV}}(\epsilon_2)\ge\theta(j)-\mathbb{E}_k\widetilde{\tau}_j
-\sqrt{\left(\tfrac{1}{\delta}-1\right)}\max\{\alpha(j),\alpha(k)\},
\]
for any $0\le j\le k\le n$ and $\delta\in(0,1)$, where $\epsilon_1=1-\pi([j,k])+\delta$ and $\epsilon_2=\min\{\pi([j,n]),\pi([0,k])\}-\delta$.
\end{thm}
\begin{proof}
We first consider the upper bound. Set $\epsilon_1=1-\pi([j,k])+\delta$. By Lemma \ref{l-tv}, if $i\le j$, then
\[
d^{(c)}_{\text{\tiny TV}}(i,t)\le\mathbb{P}_0(\widetilde{\tau}_k>t)+1-\pi([j,k]).
\]
As a result of the one-sided Chebyshev inequality, this implies
\[
T^{(c)}_{\text{\tiny TV}}(i,\epsilon_1)\le \mathbb{E}_0\widetilde{\tau}_k+\sqrt{\left(\tfrac{1}{\delta}-1\right)}\alpha(k).
\]
Similarly, if $i\ge k$, then
\[
T^{(c)}_{\text{\tiny TV}}(i,\epsilon_1)\le \mathbb{E}_n\widetilde{\tau}_j+\sqrt{\left(\tfrac{1}{\delta}-1\right)}\alpha(j).
\]
Note that, in the case $j<i<k$,
\[
\mathbb{P}_i(\max\{\widetilde{\tau}_j,\widetilde{\tau}_k\}>t)
\le\mathbb{P}_i(\widetilde{\tau}_k>t)+\mathbb{P}_i(\widetilde{\tau}_j>t)
\le\mathbb{P}_j(\widetilde{\tau}_k>t)+\mathbb{P}_k(\widetilde{\tau}_j>t).
\]
This implies
\[
T^{(c)}_{\text{\tiny TV}}(i,\epsilon_1)\le \mathbb{E}_j\widetilde{\tau}_k+\mathbb{E}_k\widetilde{\tau}_j
+\sqrt{\left(\tfrac{2}{\delta}-1\right)}\max\{\alpha(j),\alpha(k)\}.
\]
Combining all above gives the desired upper bound.
For the lower bound, set $\epsilon_2=\min\{\pi([j,n]),\pi([0,k])\}-\delta$. By the second inequality of Lemma \ref{l-tv}, one has
\[
d^{(c)}_{\text{\tiny TV}}(0,t)\ge \pi([j,n])-\mathbb{P}_0(\widetilde{\tau}_j\le t).
\]
Setting $t=\mathbb{E}_0\widetilde{\tau}_j-\sqrt{(1/\delta-1)}\alpha(j)$ in the above inequality derives
\[
d^{(c)}_{\text{\tiny TV}}(0,t)\ge \pi([j,n])-\delta\ge\epsilon_2.
\]
This implies
\[
T^{(c)}_{\text{\tiny TV}}(\epsilon_2)\ge T^{(c)}_{\text{\tiny TV}}(0,\epsilon_2)\ge \mathbb{E}_0\widetilde{\tau}_j-\sqrt{(\tfrac{1}{\delta}-1)}\alpha(j).
\]
Similarly, for $k\ge j$, we have
\[
T^{(c)}_{\text{\tiny TV}}(\epsilon_2)\ge \mathbb{E}_n\widetilde{\tau}_k-\sqrt{(\tfrac{1}{\delta}-1)}\alpha(k)=
\mathbb{E}_n\widetilde{\tau}_j-\mathbb{E}_k\widetilde{\tau}_j-\sqrt{(\tfrac{1}{\delta}-1)}\alpha(k).
\]
Both inequalities combine to the desired lower bound.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-tv2}(Continuous time case)] It has been shown in \cite{DLP10} that separation is maximized when the chain started at any of the boundary states and the maximum total variation cutoff is equivalent to the maximum separation cutoff. It is clear that the constants, $s_n$ and $b_n$, in Theorem \ref{t-sep2} are respectively of the same order as the constants, $\theta_n$ and $\alpha_n$, in Theorem \ref{t-tv2}. As a consequence of Theorem \ref{t-sep2}, $\mathcal{F}_c$ has a cutoff in the maximum total variation if and only if $\theta_n\lambda_n\ra\infty$ if and only if $\theta_n/\alpha_n\ra\infty$.
To see a cutoff time and a window, we assume in the following that $\theta_n/\alpha_n\ra\infty$. Set
\[
\epsilon_0=\inf_n\min\{\pi_n([0,M_n]),\pi_n([M_n,n])\}.
\]
For $\epsilon\in(0,\epsilon_0)$, we may choose $x_n,y_n$ such that
\[
\pi_n([0,x_n])\ge \frac{\epsilon}{3},\quad\pi_n([x_n,n])\ge 1-\frac{\epsilon}{3},\quad\pi_n([0,y_n])\ge 1-\frac{\epsilon}{3},\quad\pi_n([y_n,n])\ge\frac{\epsilon}{3}.
\]
Clearly, $x_n\le y_n$. Replacing $j,k,\delta$ with $x_n,y_n,\epsilon/3$ in Theorem \ref{t-tvmix} yields
\[
T^{(c)}_{n,\textnormal{\tiny TV}}(\epsilon)\le\theta_n(x_n)+\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{y_n}
+\mathbb{E}_{y_n}\widetilde{\tau}^{(n)}_{x_n}+\sqrt{\frac{6}{\epsilon}}
\max\{\alpha_n(x_n),\alpha_n(y_n)\},
\]
where
\[
\theta_n(j):=\max\{\mathbb{E}_0\widetilde{\tau}^{(n)}_j,\mathbb{E}_n\widetilde{\tau}^{(n)}_j\},\quad \alpha_n^2(j)=\max\{\textnormal{Var}_0\widetilde{\tau}^{(n)}_j,\textnormal{Var}_n\widetilde{\tau}^{(n)}_j\}.
\]
In the above notations, $\theta_n=\theta_n(M_n)$ and $\alpha_n=\alpha_n(M_n)$. Since $x_n\le M_n\le y_n$, one has
\[
\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{x_n}=\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_n}
+\mathbb{E}_{M_n}\widetilde{\tau}^{(n)}_{x_n},\quad\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n}=
\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{x_n}+\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n}.
\]
Note that, for any positive reals $a,b,c,d$,
\[
|\max\{a+b,c\}-\max\{a,c+d\}|\le\max\{b,d\}.
\]
This implies
\[
|\theta_n(x_n)-\theta_n|\le\mathbb{E}\widetilde{\tau}^{(n)}_{M_n}
+\mathbb{E}_{M_n}\widetilde{\tau}^{(n)}_{x_n}\le\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{y_n}
+\mathbb{E}_{y_n}\widetilde{\tau}^{(n)}_{x_n}.
\]
According to the definition of $x_n,y_n,M_n$, Corollary \ref{c-sep} implies
\[
\quad\alpha_n(x_n)\asymp \alpha_n\asymp \alpha_n(y_n).
\]
Let $p_{n,\ell},q_{n,\ell}$ be the birth and death rates of the $n$th chain. The replacement of $j,M,k$ with $x_n,M_n,y_n$ in (\ref{eq-hittime2}) yields that, for any $0\le i\le n$,
\begin{align}
\alpha_n(i)&\ge\frac{1}{2\sqrt{2}}\max\left\{\sum_{\ell=x_n}^{M_n-1}\frac{\pi_n([0,x_n])}{p_{n,\ell}\pi_n(\ell)},
\sum_{\ell=M_n+1}^{y_n}\frac{\pi_n([y_n,n])}{q_{n,\ell}\pi_n(\ell)}\right\}\notag\\
&\ge\frac{\epsilon}{12\sqrt{2}}\sum_{\ell=x_n}^{y_n-1}\frac{1}{p_{n,\ell}\pi_n(\ell)}
=\frac{\epsilon}{12\sqrt{2}}\sum_{\ell=x_n+1}^{y_n}\frac{1}{q_{n,\ell}\pi_n(\ell)}\notag\\
&\ge\frac{\epsilon}{12\sqrt{2}}\max\{\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{y_n},
\mathbb{E}_{y_n}\widetilde{\tau}^{(n)}_{x_n}\},\notag
\end{align}
where the second inequality uses the fact $q_{n,\ell}\pi_n(\ell)=p_{n,\ell-1}\pi(\ell-1)$ and the last inequality applies the first identity in Lemma \ref{l-meanvar2}. As a consequence, we may conclude from the above discussions that
\[
T^{(c)}_{n,\textnormal{\tiny TV}}(\epsilon)-\theta_n\le\left(\frac{48\sqrt{2}}{\epsilon}+\sqrt{\frac{6}{\epsilon}}\right)
\max\{\alpha_n(x_n),\alpha_n(y_n)\}\asymp \alpha_n,
\]
for all $\epsilon\in(0,\epsilon_0)$. In a similar statement, one can show, by the second part of Theorem \ref{t-tvmix}, that
\[
\theta_n-T^{(c)}_{n,\textnormal{\tiny TV}}(1-\epsilon)\le\left(\frac{36\sqrt{2}}{\epsilon}+\sqrt{\frac{3}{\epsilon}}\right)
\max\{\alpha_n(x_n),\alpha_n(y_n)\}
=O(\alpha_n),
\]
for all $\epsilon\in(0,\epsilon_0)$. This proves the $(\theta_n,\alpha_n)$ cutoff for $\mathcal{F}_c$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-tv2}(Discrete time case)]
We will use the result in the continuous time case and \cite{CSal13-1} to deal with the discrete time case. Set
\[
\delta=\inf_{n,i}K_n(i,i),\quad K_n^{(\delta)}=(K_n-\delta I)/(1-\delta).
\]
In the assumption for discrete time case, we have $\delta\in(0,1)$. Let $\mathcal{X}_n=\{0,1,...,n\}$, $\mathcal{F}^{(\delta)}=(\mathcal{X}_n,K_n^{(\delta)},\pi_n)_{n=1}^\infty$ and $\mathcal{F}^{(\delta)}_c$ be the family of continuous time chains associated with $\mathcal{F}^{(\delta)}$. It was proved in \cite{CSal13-1} (See Theorems 3.1 and 3.3) that, in the maximum total variation,
\begin{equation}\label{eq-cut1}
\mathcal{F}\text{ has a cutoff}\quad\Lra\quad \mathcal{F}^{(\delta)}_c\text{ has a cutoff}
\end{equation}
and
\begin{equation}\label{eq-cut2}
\mathcal{F}\text{ has a $(t_n,b_n)$ cutoff}\quad\Lra\quad \mathcal{F}^{(\delta)}_c\text{ has a $((1-\delta)t_n,b_n)$ cutoff}.
\end{equation}
Let $\widetilde{\tau}_i^{(n,\delta)}$ be the hitting time to state $i$ of the continuous time chain associated with $K_n^{(\delta)}$ and $\mathbb{E}_i,\textnormal{Var}_i$ be the conditional expectation and variance given the initial state $i$. Set
\[
\theta_n^{(\delta)}=\max\left\{\mathbb{E}_0\widetilde{\tau}_{M_n}^{(n,\delta)},
\mathbb{E}_n\widetilde{\tau}_{M_n}^{(n,\delta)}\right\},\quad
\beta_n^{(\delta)}=\max\left\{\textnormal{Var}_0\widetilde{\tau}_{M_n}^{(n,\delta)},
\textnormal{Var}_n\widetilde{\tau}_{M_n}^{(n,\delta)}\right\}.
\]
For $\mathcal{F}^{(\delta)}_c$, it has been proved in the continuous time case that
\[
\mathcal{F}^{(\delta)}_c\text{ has a cutoff}\quad\Lra\quad \theta_n^{(\delta)}\lambda_n^{(\delta)}\ra\infty \quad\Lra\quad\theta_n^{(\delta)}/\beta_n^{(\delta)}\ra\infty,
\]
where $\lambda_n^{(\delta)}$ is the smallest non-zero eigenvalue of $I-K_n^{(\delta)}$. Furthermore, if it holds true that $\theta_n^{(\delta)}/\beta_n^{(\delta)}\ra\infty$, then $\mathcal{F}_c^{(\delta)}$ has a $(\theta_n^{(\delta)},\beta_n^{(\delta)})$ cutoff. As a result of (\ref{eq-cut1}) and (\ref{eq-cut2}), we have
\[
\mathcal{F}\text{ has a cutoff}\quad\Lra\quad \theta_n^{(\delta)}/\beta_n^{(\delta)}\ra\infty,
\]
and, further, if the right side holds, then $\mathcal{F}$ has a $(\theta_n^{(\delta)}/(1-\delta),\beta_n^{(\delta)})$ cutoff.
Let $\lambda_n,\theta_n,\beta_n$ be the constants in Theorem \ref{t-tv2}. Clearly, $\lambda_n=(1-\delta)\lambda_n^{(\delta)}$. To finish the proof, it suffices to show that
\begin{equation}\label{eq-cut3}
\theta_n^{(\delta)}=(1-\delta)\theta_n,\quad \beta_n^{(\delta)}\asymp\beta_n.
\end{equation}
Let $p_{n,i},q_{n,i},r_{n,i}$ be the transition rates of $K_n$ and $p_{n,i}^{(\delta)},q_{n,i}^{(\delta)},r_{n,i}^{(\delta)}$ be the transition rates of $K_n^{(\delta)}$. It is clear that
\[
p_{n,i}^{(\delta)}=p_{n,i}/(1-\delta),\quad q_{n,i}^{(\delta)}=q_{n,i}/(1-\delta),\quad r_{n,i}^{(\delta)}=(r_{n,i}-\delta)/(1-\delta).
\]
The first equality of (\ref{eq-cut3}) is an immediate result of the first identity of Lemma \ref{l-meanvar2}. To see the second part of (\ref{eq-cut3}), let $\lambda_{n,1},...,\lambda_{n,n}$ be eigenvalues of the submatrix of $I-K_n$ obtained by removing the $M_n$-th row and column. Clearly, $\lambda_{n,1}/(1-\delta),...,\lambda_{n,n}/(1-\delta)$ are eigenvalues of the submatrix of $I-K_n^{(\delta)}$ obtained by removing the $M_n$-th row and column. As a consequence of Lemma \ref{l-meanvar}, we have
\[
\beta_n^2\asymp\sum_{i=1}^n\frac{1-\lambda_{n,i}}{\lambda_{n,i}^2},\quad \left(\beta_n^{(\delta)}\right)^2\asymp\sum_{i=1}^n\frac{1}{\lambda_{n,i}^2}.
\]
Note that the application of Remark \ref{r-var} on the chain $(\mathcal{X}_n,K_n^{(\delta)},\pi_n)$ says
\[
(1-\delta)\sum_{i=1}^n\frac{1}{\lambda_{n,i}^2}\ge\sum_{i=1}^n\frac{1}{\lambda_{n,i}}.
\]
This implies $\beta_n\asymp\beta_n^{(\delta)}$.
\end{proof}
\section{A randomization of birth and death chains}\label{s-rbdc}
This section gives two nontrivial examples as applications of theorems in the introduction. The first example is stated in Theorem \ref{t-randomm} and we discuss its proof in the following.
\begin{proof}[Proof of Theorem \ref{t-randomm}]
The proofs for $\mathcal{F}_c$ and $\mathcal{F}$ are similar and we consider only the continuous time case. Let $M_n,\theta_n,\alpha_n$ be as in Theorem \ref{t-tv2}. For convenience, we let $(p_{n,i},q_{n,i},r_{n,i})$ be the transition rates of $K_n$. For $n\ge 1$, set
\[
\theta_{n,1}=\sum_{i=0}^{M_n-1}\frac{\pi_n([0,i])}{\pi_n(i)p_{n,i}},\quad \theta_{n,2}=\sum_{i=M_n+1}^{n}\frac{\pi_n([i,n])}{\pi_n(i)q_{n,i}}
\]
and
\[
\alpha_{n,1}^2=\sum_{i=0}^{M_n-1}\sum_{j=i}^{M_n-1}\frac{\pi_n([0,i])^2}{\pi_n(i)p_{n,i}\pi_n(j)p_{n,j}},\,
\alpha_{n,1}^2=\sum_{i=M_n+1}^{n}\sum_{j=M_n+1}^i\frac{\pi_n([i,n])^2}{\pi_n(i)q_{n,i}\pi_n(j)q_{n,j}}.
\]
It is clear from Lemma \ref{l-meanvar2} that
\[
\theta_n=\max\{\theta_{n,1},\theta_{n,2}\},\quad \alpha_n=\max\{\alpha_{n,1},\alpha_{n,2}\}.
\]
Without loss of generality, we may assume that $\theta_n=\theta_{n,1}$. For $n\ge 1$, let $U_{n,1},V_{n,1}$ be positive random variables defined by
\[
U_{n,1}=\sum_{i=0}^{M_n-1}\frac{\pi_n([0,i])}{\pi_n(i)p_{n,i}C_{n,i+1}},\,
V_{n,1}^2=\sum_{i=0}^{M_n-1}\sum_{j=i}^{M_n-1}\frac{\pi_n([0,i])^2}{\pi_n(i)p_{n,i}C_{n,i+1}\pi_n(j)p_{n,j}C_{n,j+1}}.
\]
By the independency of $C_{n,i}$, one may compute
\[
\mathbb{E}U_{n,1}=\mu_n\theta_{n,1}=\mu_n\theta_n,\quad \text{Var}(U_{n,1})=\nu_n^2\alpha_{n,1}^2\le\nu_n^2\alpha_n^2
\]
and
\begin{align}
\mathbb{E}V_{n,1}^2&=\sum_{0\le i<j\le M_n-1}\frac{\pi_n([0,i])^2}{\pi_n(i)p_{n,i}\pi_n(j)p_{n,j}}\mu_n^2
+\sum_{i=0}^{M_n-1}\frac{\pi_n([0,i])^2}{\pi_n(i)^2p_{n,i}^2}(\mu_n^2+\nu_n^2)\notag\\
&\le (\mu_n^2+\nu_n^2)\alpha_{n,1}^2\le[(\mu_n+\nu_n)\alpha_{n,1}]^2.\notag
\end{align}
The estimation for $\mathbb{E}V_n^2$ implies
\[
\mathbb{E}V_{n,1}\le\sqrt{\mathbb{E}V_{n,1}^2}\le(\mu_n+\nu_n)\alpha_{n,1}\le(\mu_n+\nu_n)\alpha_n.
\]
Set $a_n=\sqrt{(\mu_n\theta_n)/(\nu_n\alpha_n)}$, $b_n=\sqrt{(\mu_n\theta_n)/[(\mu_n+\nu_n)\alpha_n]}$ and
\[
E_{n,1}=\{\omega_n\in\Omega_n:|U_{n,1}(\omega_n)-\mu_n\theta_n|<a_n\nu_n\alpha_n,\,
V_{n,1}(\omega_n)<b_n(\mu_n+\nu_n)\alpha_n\}.
\]
Since $\mathcal{F}_c$ has a maximum total variation cutoff, Theorem \ref{t-tv2} implies $\alpha_n=o(\theta_n)$. In the assumption of $(\nu_n\alpha_n)=o(\mu_n\theta_n)$, it is easy to see that, for $\omega_n\in E_{n,1}$,
\[
U_{n,1}(\omega_n)\sim\mu_n\theta_n,\quad V_{n,1}(\omega_n)=o(\mu_n\theta_n).
\]
By the Chebyshev and Markov inequalities, the fact that $a_n,b_n\ra\infty$ yields $\mathbb{P}^{(n)}(E_{n,1})\ra 1$.
In the same way, we set
\[
U_{n,2}=\sum_{i=M_n+1}^{n}\frac{\pi_n([i,n])}{\pi_n(i)q_{n,i}C_{n,i}},\,
V_{n,2}^2=\sum_{i=M_n+1}^{n}\sum_{j=M_n+1}^{i}\frac{\pi_n([i,n])^2}{\pi_n(i)q_{n,i}C_{n,i}\pi_n(j)q_{n,j}C_{n,j}},
\]
and
\[
E_{n,2}=\{\omega_n\in\Omega_n:U_{n,2}(\omega_n)<\mu_n\theta_n+a_n\nu_n\alpha_n,\,
V_{n,2}(\omega_n)<b_n(\mu_n+\nu_n)\alpha_n\}.
\]
A similar reasoning as before yields that $\mathbb{P}^{(n)}(E_{n,2})\ra 1$ and, for $\omega_n\in E_{n,2}$,
\[
U_{n,2}(\omega_n)\le \mu_n\theta_n(1+o(1)),\quad V_{n,2}(\omega_n)=o(\mu_n\theta_n).
\]
As consequence, if we set $E_n=E_{n,1}\cap E_{n,2}$, then $\mathbb{P}^{(n)}(E_n)\ra\infty$ and, for $\omega_n\in E_n$,
\[
\max\{U_{n,1},U_{n,2}\}\sim \mu_n\theta_n,\quad \max\{V_{n,1},V_{n,2}\}=o(\mu_n\theta_n).
\]
The maximum total variation cutoff for $\mathcal{F}_c^{(\omega)}$ and the cutoff time $\mu_n\theta_n$ are immediate from Theorem \ref{t-tv2}.
\end{proof}
\begin{rem}
From the proof given above, one can derive a variation of Theorem \ref{t-randomm}. Namely, under the assumption of $\nu_n\alpha_n=o(\mu_n\theta_n)$, if $\mathcal{F}_c$ has no maximum total variation cutoff (resp. maximum separation cutoff), then there is a sequence $E_n\subset\Omega_n$ satisfying $\mathbb{P}^{(n)}(E_n)\ra 1$ such that $\mathcal{F}_c^{(\omega)}$ has no maximum total variation cutoff (resp. maximum separation cutoff) for $\omega\in\prod_{n=1}^\infty E_n$. Note that, the requirement $\nu_n\alpha_n=o(\mu_n\theta_n)$ and the assumption of no cutoff will imply the existence of a subsequence, say $i_n$, such that $\nu_{i_n}=o(\mu_{i_n})$. As a result of the Chebyshev inequality, $1/C_{i_n,1}-\mathbb{E}(1/C_{i_n,1})$ converges in probability to $0$. This turns $\mathcal{F}_c^{(\omega)}$ into a lazy version of $\mathcal{F}_c$ with high probability.
\end{rem}
Note that the hypothesis of $\nu_n\alpha_n=o(\mu_n\theta_n)$ requires the existence of a second moment of $1/C_{n,1}$. Next, we give an example where $1/C_{n,1}$ does not have a finite first moment.
\begin{thm}\label{t-randomsrw}
For $n\ge 1$, let $C_{n,1},...,C_{n,n}$ be i.i.d. uniform random variables over $(0,1)$ defined on $(\Omega_n,\mathbb{P}^{(n)})$. For $\omega=(\omega_1,\omega_2,...)\in\prod_n\Omega_n$, let $\mathcal{F}^{(\omega)}=(\mathcal{X}_n,K_n^{(\omega_n)},\pi_n)_{n=1}^\infty$ be a family of birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$ and
\[
\begin{cases}K_n^{(\omega_n)}(i,i+1)=K(i+1,i)=C_{n,i+1}/2,\quad\forall 0\le i<n,\\ K_n^{(\omega_n)}(i,i)=1-K_n^{(\omega_n)}(i,i+1)-K_n^{(\omega_n)}(i,i-1),\quad\forall i.\end{cases}
\]
Let $\mathcal{F}_c^{(\omega)}$ be the family of continuous time chains associated with $\mathcal{F}^{(\omega)}$ and, for $\omega_n\in\Omega_n$, let $T_{n,\textnormal{\tiny TV}}^{c}(\omega_n,\cdot)$ be the maximum total variation mixing time for $(\mathcal{X}_n,K_n^{(\omega_n)},\pi_n)$. Then, there is a sequence $E_n\subset\Omega_n$ satisfying $\mathbb{P}^{(n)}(E_n)\ra 1$ such that, for any $\omega=(\omega_1,\omega_2,...)\in\prod_{n=1}^\infty E_n$, the family $\mathcal{F}_c^{(\omega)}$ has no maximum total variation cutoff and $T_{n,\textnormal{\tiny TV}}^c(\omega_n,\epsilon)\asymp n^2\log n$ for $\epsilon\in(0,1/10)$.
\end{thm}
\begin{proof}
Let $M_n\in\mathcal{X}_n$ and $U_{n,1},U_{n,2}$ be as in the proof of Theorem \ref{t-randomm}. For $n\ge 1$, set
\[
\overline{\Omega}_n=\left\{C_{n,i}>\frac{1}{n\log n},\forall 1\le i\le n\right\},\quad \overline{\mathbb{P}}^{(n)}(\cdot)=\mathbb{P}^{(n)}(\cdot|\overline{\Omega}_n),
\]
where $\overline{\mathbb{P}}^{(n)}$ is the conditional probability of $\mathbb{P}^{(n)}$ given $\overline{\Omega}_n$. Clearly, $\mathbb{P}^{(n)}(\overline{\Omega}_n)=(1-1/n\log n)^n\ra 1$ and, in $\overline{\mathbb{P}}^{(n)}$, $C_{n,1},...,C_{n,n}$ are i.i.d. uniformly distributed over $(1/n\log n,1)$. Let $\overline{\mathbb{E}}$ and $\overline{\text{Var}}$ be the expectation and variance taken in $\overline{\mathbb{P}}^{(n)}$. It is an easy exercise to compute
\[
\overline{\mathbb{E}}(1/C_{n,1})=\frac{\log n+\log\log n}{1-1/n\log n}\sim \log n
\]
and
\[
\overline{\text{Var}}(1/C_{n,1})=n\log n-(\overline{\mathbb{E}}(1/C_{n,1}))^2\sim n\log n,
\]
This implies that, if $M_n\ra\infty$ and $n-M_n\ra\infty$, then
\[
\overline{\mathbb{E}}U_{n,1}\sim M_n^2\log n,\quad\overline{\mathbb{E}}U_{n,2}\sim(n-M_n)^2\log n,
\]
and
\[
\overline{\text{Var}}(U_{n,1})\sim M_n^2n\log n,\quad \overline{\text{Var}}(U_{n,2})\sim(n-M_n)^2n\log n.
\]
For $a\in (0,1)$, if $M_n=\lfloor an\rfloor$, we write $U_{n,i}^{(a)}$ for $U_{n,i}$. As a result of the above computation, we obtain
\[
\overline{\mathbb{E}}U_{n,1}^{(a)}\sim a^2n^2\log n,\quad \overline{\mathbb{E}}U_{n,2}^{(a)}\sim(1-a)^2n^2\log n,
\]
and
\[
\overline{\text{Var}}(U_{n,1}^{(a)})\sim a^2n^3\log n,\quad \overline{\text{Var}}(U_{n,2}^{(a)})\sim(1-a)^2n^3\log n.
\]
For $n\ge 1$, let
\[
E_n=\left\{\omega_n\in A_n:|U_{n,1}^{(a)}-a^2n^2\log n|<n^{3/2}\log n,\text{ for }a=1/4,1/2\right\}.
\]
It is easy to show that $\overline{\mathbb{P}}^{(n)}(E_n)\ra 1$ and, hence, $\mathbb{P}^{(n)}(E_n)\ge \mathbb{P}^{(n)}(A_n)\overline{\mathbb{P}}^{(n)}(E_n)\ra 1$. Furthermore, for $\omega_n\in E_n$,
\[
\max\{U_{n,1}^{(1/2)}(\omega_n),U_{n,2}^{(1/2)}(\omega_n)\}\sim\frac{n^2\log n}{4}
\]
and
\[
\max\{U_{n,1}^{(1/4)}(\omega_n),U_{n,2}^{(1/4)}(\omega_n)\}\sim\frac{9 n^2\log n}{16}.
\]
By Remark \ref{r-exp1}, $\mathcal{F}_c^{(\omega)}$ has no maximum total variation cutoff for $\omega\in\prod_nE_n$. The order of the mixing time is given by Theorems 3.1 and 3.9 of \cite{CSal13-2}.
\end{proof}
\begin{rem}
We refer the reader to \cite{DW13} for another randomization of birth and death chains, which is different from the one considered in Theorem \ref{t-randomsrw}.
\end{rem}
\section{Chains started at boundary states}\label{s-bs}
For continuous time birth and death chains, \cite{DLP10} shows that separation reaches its maximum when the initial state is any of the boundary states. This is not true in the case of total variation and it is easy to construct counterexamples. In this section, we discuss the total variation cutoff for families of birth and death chains started at a boundary state. As before, we use $\mathcal{F}$ and $\mathcal{F}_c$ for families of birth and death chains without starting states specified and write $\mathcal{F}^L,\mathcal{F}_c^L$ and $\mathcal{F}^R,\mathcal{F}_c^R$ respectively for families of chains started at the left and right boundary states.
The following theorem displays a list of equivalent conditions for the total variation cutoff. It is worthwhile to note that some of these conditions are very similar to the conditions in Theorem \ref{t-tv2}.
\begin{thm}\label{t-tv3}
Let $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of irreducible birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$ and $\mathcal{F}_c$ be the family of associated continuous time chains in $\mathcal{F}$. For $n\ge 1$, let $\widetilde{\tau}^{(n)}_i$ be the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}_c$ and, for $a\in(0,1)$, let $M_n(a)$ be a state in $\mathcal{X}_n$ satisfying
\[
\pi_n([0,M_n(a)])\ge a,\quad \pi_n([M_n(a),n])\ge 1-a,
\]
and let $\lambda_n(a)$ be the smallest eigenvalue of the submatrix of $I-K_n$ indexed by states $0,...,M_n(a)-1$. Set
\[
u_n(a)=\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)},\quad v_n^2(a)=\textnormal{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}.
\]
Assume that $\pi_n(0)\ra 0$. Then, the following are equivalent.
\begin{itemize}
\item[(1)] $\mathcal{F}_c^L$ has a total variation cutoff.
\item[(2)] $u_n(a)/v_n(a)\ra\infty$ for all $a\in(0,1)$.
\item[(3)] $u_n(a)\lambda_n(a)\ra\infty$ for all $a\in(0,1)$.
\item[(4)] There are $a\in(0,1)$ and a positive sequence $(t_n)_{n=1}^\infty$ satisfying
\[
t_n=O(u_n(c)),\quad \forall c\in(0,1)
\]
and
\[
\lim_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(a)}>(1-\epsilon)t_n\right)=1,\quad\forall \epsilon\in(0,1),
\]
and, for any $b\in(a,1)$, there is $\alpha_b\in(0,1)$ such that
\[
\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(b)}>(1+\epsilon)t_n\right)\le\alpha_b,\quad\forall
\epsilon>0,
\]
where $\mathbb{P}_i$ denotes the probability given the initial state $i$.
\end{itemize}
Furthermore, if \textnormal{(2)} or \textnormal{(3)} holds, then $\mathcal{F}_c^L$ has a cutoff with cutoff time $(u_n(a))_{n=1}^\infty$ for any $a\in(0,1)$. If \textnormal{(4)} holds, then $\mathcal{F}_c^L$ has a cutoff with cutoff time $(t_n)_{n=1}^\infty$.
\end{thm}
The discrete time version of the previous theorem can be stated as follows.
\begin{thm}\label{t-tv4}
Let $\mathcal{F},M_n(a),\lambda_n(a)$ be as in Theorem \ref{t-tv3}. For $n\ge 1$, let $\tau^{(n)}_i$ be the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}$ and, for $a\in(0,1)$, set
\[
u_n(a)=\mathbb{E}_0\tau^{(n)}_{M_n(a)},\quad w_n^2(a)=\textnormal{Var}_0\tau^{(n)}_{M_n(a)}.
\]
Assume that $\pi_n(0)\ra 0$, $\inf_{i,n}K_n(i,i)>0$ and $u_n(a)\ra\infty$ for some $a\in(0,1)$. Then, the conclusion in Theorem \ref{t-tv3} remains true for the family $\mathcal{F}^L$.
\end{thm}
\begin{rem}
One can see from the proof of Theorem \ref{t-tv3} that the condition $\pi_n(0)\ra 0$ is necessary for the existence of cutoff of $\mathcal{F}_c^L$. This is also true for the family $\mathcal{F}^L$ in Theorem \ref{t-tv4}.
\end{rem}
\begin{rem}\label{r-pin->0}
Let $\mathcal{F},\mathcal{F}_c$ be as in Theorem \ref{t-tv3} and $(p_{n,i},q_{n,i},r_{n,i})$ be the transition rates of the $n$th chains in $\mathcal{F}$. Let $M_n\in\mathcal{X}_n$ be a sequence of states satisfying (\ref{eq-pin2}), that is,
\[
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad\inf_{n\ge 1}\pi_n([M_n,n])>0,
\]
and $x_n\in\{0,n\}$ be a boundary state fulfilling the following equation
\[
\max\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n},\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}\}
=\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n}.
\]
By Lemma \ref{l-meanvar2} and Theorem A.1 of \cite{CSal13-2}, if $x_n=0$, then
\[
\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n}=\sum_{i=0}^{M_n-1}\frac{\pi_n([0,i])}{\pi_n(i)p_{n,i}}
\le\sum_{i=0}^{M_n-1}\frac{1}{\pi_n(i)p_{n,i}}
\]
and
\begin{align}
\frac{1}{\lambda_n}&\ge\min\{\pi_n([0,M_n]),\pi_n([M_n,n])\}\times
\max_{j:j<M_n}\sum_{i=j}^{M_n-1}
\frac{\pi_n([0,j])}{\pi_n(i)p_{n,i}}\notag\\
&\ge\min\{\pi_n([0,M_n]),\pi_n([M_n,n])\}\times\pi_n(0)\sum_{i=0}^{M_n-1}
\frac{1}{\pi_n(i)p_{n,i}}\notag
\end{align}
This implies
\[
\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n}\lambda_n\le\frac{1}{\min\{\pi_n([0,M_n]),\pi_n([M_n,n])\}\pi_n(0)}.
\]
In a similar way, this inequality also holds in the case $x_n=n$. As a consequence of Theorem \ref{t-tv2}, if $\mathcal{F}_c$ has a maximum total variation cutoff, then $\pi_n(x_n)\ra 0$. The above discussion also holds for $\mathcal{F}$ with the assumption $\inf_{n,i}K_n(i,i)>0$.
\end{rem}
\begin{rem}\label{r-cutoffitme}
Let $\mathcal{F}_c^L$ and $\mathcal{F}^L$ be the families in Theorems \ref{t-tv3} and \ref{t-tv4}. If $\mathcal{F}_c^L$ (resp. $\mathcal{F}^L$) has a total variation cutoff with cutoff time $t_n$, then
\[
t_n\sim\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n},\quad(\text{resp. } t_n\sim\mathbb{E}_0\tau^{(n)}_{M_n},)
\]
where $M_n\in\mathcal{X}_n$ is any sequence satisfying
\begin{equation}\label{eq-mn2}
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad \inf_{n\ge 1}\pi_n([M_n,n])>0.
\end{equation}
In particular, if $\mathcal{F}_c^L$ (resp. $\mathcal{F}^L$) has a total variation cutoff with bounded cutoff time, then
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}=O(1),\quad (\text{resp. }\mathbb{E}_0\tau^{(n)}_{M_n}=O(1),)
\]
for any sequence $M_n\in\mathcal{X}_n$ satisfying (\ref{eq-mn2}).
\end{rem}
\begin{rem}\label{r-cutoff}
Let $\mathcal{F}_c^L$ be the family in Theorems \ref{t-tv3}. If $\mathcal{F}_c^L$ has a total variation cutoff, then $u_n(a)\sim u_n(b)$ for all $a,b\in(0,1)$, or equivalently
\[
\mathbb{E}_{M_n(a)}\widetilde{\tau}^{(n)}_{M_n(b)}=o\left(\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(c)}\right),\quad\forall
a,b,c\in(0,1).
\]
This is also true for $\mathcal{F}^L$ with the assumption in Theorem \ref{t-tv4}. But, the converse is not necessarily true. For an illustration, recall the example in Remark \ref{r-exp1}. It has been proved that
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}\asymp \frac{1}{\lambda_n}\asymp \frac{1}{\xi_n}+n^2,\quad
\forall a\in(0,1).
\]
By Lemma \ref{l-meanvar2}, one may compute
\[
\text{Var}_0\widetilde{\tau}^{(n)}_1=\frac{1}{\xi_n^2}
\]
and
\[
\text{Var}_1\widetilde{\tau}^{(n)}_{M_n(a)}\ge\sum_{i=1}^{M_n(a)-1}\frac{1}{K_n(i,i+1)\pi_n(i)}
\sum_{\ell=1}^{i}\pi_n(\ell)\mathbb{E}_\ell\widetilde{\tau}^{(n)}_{i+1}\asymp n^4.
\]
Along with the fact $\text{Var}_0\widetilde{\tau}^{(n)}_i\le(\mathbb{E}_0\widetilde{\tau}^{(n)}_i)^2$, we may conclude from the above computations that $\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}\asymp \xi_n^{-2}+n^4$ for all $a\in(0,1)$. By Theorem \ref{t-tv3}, this implies that the family $\mathcal{F}_c^L$ has no total variation cutoff. It has been shown in Remark \ref{r-exp1} that if $n^2\xi_n\ra 0$, then $\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}\sim\xi_n^{-1}$ for all $a\in(0,1)$.
\end{rem}
\begin{rem}\label{r-wcomp}
Let $v_n(a)$ and $w_n(a)$ be the constants in Theorems \ref{t-tv3} and \ref{t-tv4}. It is remarkable that if $\delta=\inf_{i,n}K_n(i,i)>0$, then $\delta v_n^2(a)\le w_n^2(a)\le v_n^2(a)$ for all $a\in(0,1)$. To see this, we let $\beta^{(n)}_1,...,\beta^{(n)}_{M_n}$ be the eigenvalues of the submatrix of $I-K_n$ indexed by $0,...,M_n(a)-1$. By Lemma \ref{l-meanvar}, $\beta^{(n)}_i>0$ for all $i$ and
\[
v_n^2(a)=\sum_{i=1}^{M_n(a)}\frac{1}{(\beta^{(n)}_i)^2},\quad w_n^2(a)=\sum_{i=1}^{M_n(a)}\frac{1-\beta^{(n)}_i}{(\beta^{(n)}_i)^2}.
\]
Clearly, $w_n^2(a)\le v_n^2(a)$. For the lower bound of $w_n^2(a)$, set $K_n^{(\delta)}=(K_n-\delta I)/(1-\delta)$. Note that $K_n^{(\delta)}$ is also a stochastic matrix and the submatrix of $I-K_n^{(\delta)}$ indexed by $0,...,M_n(a)-1$ has eigenvalues $\beta^{(n)}_1/(1-\delta),...,\beta^{(n)}_{M_n(a)}/(1-\delta)$. By Remark \ref{r-var}, we have
\[
(1-\delta)\sum_{i=1}^{M_n(a)}\frac{1}{(\beta^{(n)}_i)^2}\ge\sum_{i=1}^{M_n(a)}\frac{1}{\beta^{(n)}_i}
\]
and this implies $w_n^2(a)\ge \delta v_n^2(a)$.
\end{rem}
\begin{rem}
Note that, in Theorems \ref{t-tv3} and \ref{t-tv4}, if one chooses $\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ and $\mathbb{E}_0\tau^{(n)}_{M_n(a)}$ as the cutoff times, the square roots of $\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ and $\text{Var}_0\tau^{(n)}_{M_n(a)}$ are no longer suitable for the respective cutoff windows. This is very different from the conclusion in Theorem \ref{t-tv2} and we refer the reader to Example \ref{ex-cutoff3} for an illustration of this observation.
\end{rem}
The next corollary provides a way of selecting cutoff windows.
\begin{cor}\label{c-tv}
Let $\mathcal{F}_c,u_n(a),v_n(a)$ be as in Theorem \ref{t-tv3}. If $\mathcal{F}_c^L$ has a total variation cutoff and $b_n>0$ is a sequence satisfying
\[
b_n=o(u_n(a)),\quad v_n(a)=O(b_n),\quad\forall a\in(0,1),
\]
then $\mathcal{F}_c^L$ has a $(u_n(a),b_n)$ total variation cutoff. The above statement is also true for $\mathcal{F}^L$ under the assumption of $\inf_{n,i}K_n(i,i)>0$ and $\inf_nb_n>0$ and the replacement of $v_n(a)$ by $w_n(a)$ in Theorem \ref{t-tv4}.
\end{cor}
\begin{ex}\label{ex-cutoff3}
Let $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of birth and death chains for which $\mathcal{X}_n=\{0,1,...,n\}$, $\pi_n(i)=2^{-n}\binom{n}{i}$ and
\[
\begin{cases}
K_n(i,i+1)=1-\frac{i}{n},\quad K_n(i+1,i)=\frac{i+1}{n}\quad\text{for }i\ne M_n,\\
K_n(M_n,M_n+1)=c_n\left(1-\frac{M_n}{n}\right),\quad K_n(M_n,M_n)=(1-c_n)\left(1-\frac{M_n}{n}\right),\\
K_n(M_n+1,M_n)=\frac{c_n(M_n+1)}{n},\quad
K_n(M_n+1,M_n+1)=\frac{(1-c_n)(M_n+1)}{n},
\end{cases}
\]
where $c_n\in(0,1)$ and $M_n\in\mathcal{X}_n$ is a state satisfying $\pi_n([0,M_n])\ge 1/4$ and $\pi_n([M_n,n])\ge 3/4$. Let $\mathcal{F}_c$ be the family associated with $\mathcal{F}$ and $\widetilde{\tau}^{(n)}_i$ be the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}_c$. We will also use $M_n(a)$ with $a\in(0,1)$ to denote a state satisfying $\pi_n([0,M_n(a)])\ge a$ and $\pi_n([M_n(a),n])\ge 1-a$. When $c_n=1$, $(\mathcal{X}_n,K_n,\pi_n)$ is the Ehrenfest chain on $\{0,1,...,n\}$. The spectral information of the Ehrenfest chain is well-studied and it is easy to derive by Lemma \ref{l-gap} that
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{\lfloor n/2\rfloor}=\frac{1}{4}n\log n+O(n),\quad
\text{Var}_0\widetilde{\tau}^{(n)}_{\lfloor n/2\rfloor}\asymp n^2.
\]
One may use Stirling's formula to show that, for $0<a<b<1$,
\[
\left|\frac{n}{2}-M_n(a)\right|\asymp\sqrt{n},\quad \pi_n(i)\asymp\frac{1}{\sqrt{n}}\quad\text{uniformly for }M_n(a)\le i\le M_n(b).
\]
By Lemmas \ref{l-meanvar2}, \ref{l-gap} and \ref{l-hit2}, this implies that, for $a\in(0,1)$,
\begin{equation}\label{eq-ehren}
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}=\frac{1}{4}n\log n+O(n),\quad
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}\asymp n^2.
\end{equation}
When $c_n$ is small, $(\mathcal{X}_n,K_n,\pi_n)$ is the modification of the Ehrenfest chain with bottleneck between states $M_n$ and $M_n+1$. In the following, we will discuss the total variation cutoff and the cutoff window of $\mathcal{F}_c^L$ when $c_n$ is small.
First, we consider the total variation cutoff of $\mathcal{F}_c^L$. By Lemma \ref{l-meanvar2} and (\ref{eq-ehren}), one can show without difficulty that, for $a\in(0,1/2)$,
\begin{equation}\label{eq-ehren2}
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}=\frac{1}{4}n\log n+O(n),\quad
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}\asymp n^2,
\end{equation}
and, for $a\in(1/2,1)$,
\begin{equation}\label{eq-ehren3}
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}=\frac{1}{4}n\log n+O(n)+\frac{1+o(1)}{2c_n\pi_n(M_n)},\quad
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}\asymp n^2+\frac{n}{c_n^2},
\end{equation}
where $\pi_n(M_n)\asymp 1/\sqrt{n}$. By Theorem \ref{t-tv3}, $\mathcal{F}_c^L$ has a total variation cutoff if and only if $c_n\sqrt{n}\log n\ra\infty$.
Next, we discuss the cutoff window of $\mathcal{F}_c^L$. Assume that $c_n\sqrt{n}\log n\ra\infty$. By Corollary \ref{c-tv} and Equations (\ref{eq-ehren2}) and (\ref{eq-ehren3}), $\mathcal{F}_c^L$ has a $(\frac{1}{4}n\log n,\max\{\sqrt{n}/c_n,n\})$ total variation cutoff. We will prove that the window is optimal when $c_n\sqrt{n}\ra 0$. Suppose $c_n\sqrt{n}\ra 0$ and set
\[
s_n=\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n},\quad t_n=\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n+1},\quad a_n^2=\text{Var}^{(n)}_0\widetilde{\tau}^{(n)}_{M_n},\quad b_n^2=\text{Var}^{(n)}_0\widetilde{\tau}^{(n)}_{M_n+1}.
\]
Let $T_{n,\text{\tiny TV}}^c(0,\epsilon)$ be the total variation mixing time of the $n$th chain in $\mathcal{F}_c^L$ and recall (\ref{eq-mixing}) in the following
\[
T_{n,\text{\tiny TV}}^{(c)}(0,\epsilon)\begin{cases}\le\mathbb{E}_0\widetilde{\tau}^{(n)}_i
+\sqrt{(\tfrac{1-\delta}{\delta})\text{Var}_0(\widetilde{\tau}^{(n)}_i)}&\text{for } \epsilon=\delta+\pi_n([i+1,n])\\
\ge\mathbb{E}_0\widetilde{\tau}^{(n)}_i-\sqrt{(\tfrac{\delta}{1-\delta})\text{Var}_0(\widetilde{\tau}^{(n)}_i)}&
\text{for }\epsilon=\delta-\pi_n([0,i-1])\end{cases}.
\]
In the first inequality, the replacement of $i=M_n$ and $\delta=1/8$ implies
\[
T_{n,\text{\tiny TV}}^c(0,7/8)\le s_n+3a_n.
\]
In the second inequality, the replacement of $i=M_n+1$ and $\delta=3/8$ gives
\[
T_{n,\text{\tiny TV}}^c(0,1/8)\ge t_n-\frac{4}{5}b_n.
\]
These two inequalities yield
\[
T_{n,\text{\tiny TV}}^c(0,1/8)-T_{n,\text{\tiny TV}}^c(0,7/8)\ge \mathbb{E}_{M_n}\widetilde{\tau}^{(n)}_{M_n+1}-3a_n-\frac{4}{5}b_n.
\]
Under the assumption that $c_n\sqrt{n}\ra 0$, one may compute using Lemma \ref{l-meanvar2} that
\[
a_n\asymp n,\quad b_n\sim \mathbb{E}_{M_n}\widetilde{\tau}^{(n)}_{M_n+1}\asymp\frac{\sqrt{n}}{c_n}=\frac{n}{c_n\sqrt{n}}.
\]
Consequently, when $c_n\sqrt{n}\ra 0$, the cutoff window can be $\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ for any $a\in (1/4,1)$ but not for $a\in(0,1/4)$. Similar observation also happens in $\mathcal{F}_c^R$.
We would like to point out an interesting observation arising from the bottleneck effect in this example. Compared with the case $c_n=1$ for all $n$, when $c_n$ is of order bigger than $1/\sqrt{n}$, $\mathcal{F}_c^L$ has a cutoff with the same cutoff time and window. When $c_n$ is of order between $1/\sqrt{n}$ and $1/\sqrt{n}\log n$, $\mathcal{F}_c^L$ has a cutoff with the same cutoff time but different (larger) cutoff window. When $c_n$ is of order smaller than $1/\sqrt{n}\log n$, the cutoff of $\mathcal{F}_c^L$ disappears.
\end{ex}
The proofs of Theorems \ref{t-tv3} and \ref{t-tv4} and Corollary \ref{c-tv} are complicated and are given in Section \ref{s-proof}.
\section{Comparison of total variation cutoffs}\label{s-comp}
In this section, we make a comparison of cutoffs introduced in Sections \ref{s-tv} and \ref{s-bs}. To avoid confusion, we use $\mathcal{F},\mathcal{F}_c$ to denote families of birth and death chains without initial states specified and let $\mathcal{F}^L,\mathcal{F}_c^L$ and $\mathcal{F}^R,\mathcal{F}_c^R$ be families of chains started at respectively left and right boundary states. The following theorem is an immediate corollary of Theorems \ref{t-tv3} and \ref{t-tv4} and the proof is given in the end of this section.
\begin{thm}\label{t-tv5}
Let $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of irreducible birth and death chains with $\mathcal{X}_n=\{0,...,n\}$ and $\mathcal{F}_c$ be the family of continuous time chains associated with $\mathcal{F}$. For any sequence $S=(x_n)_{n=1}^\infty$ with $x_n\in\mathcal{X}_n$, let $\mathcal{F}^S,\mathcal{F}_c^S$ be the families of chains in $\mathcal{F},\mathcal{F}_c$ for which the $n$th chain started at $x_n$.
\begin{itemize}
\item[(1)] If $\mathcal{F}_c^L$ and $\mathcal{F}_c^R$ have a total variation cutoff with cutoff time $r_n$ and $s_n$, then $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$, where $t_n=\max\{r_n,s_n\}$.
\item[(2)] Let $M_n\in\mathcal{X}_n$ be a sequence of states satisfying
\[
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad\inf_{n\ge 1}\pi_n([M_n,n])>0
\]
and let $S=(x_n)_{n=1}^\infty$, where $x_n\in\{0,n\}$ is a state such that
\[
\max\left\{\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n},\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_n}\right\}
=\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n}
\]
and $\widetilde{\tau}^{(n)}_i$ is the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}_c$. If $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$, then $\mathcal{F}_c^S$ has a total variation cutoff with cutoff time $t_n$. In particular, $\mathcal{F}_c^S$ has a $(\mathbb{E}_{x_n}\widetilde{\tau}^{(n)}_{M_n},b_n)$ total variation cutoff with $b_n^2=\max\{\textnormal{Var}_0\widetilde{\tau}^{(n)}_{M_n},\textnormal{Var}_n\widetilde{\tau}^{(n)}_{M_n}\}$.
\end{itemize}
The above statements also apply for $\mathcal{F}$ under the assumption $\inf_{n,i}K_n(i,i)>0$.
\end{thm}
\begin{rem}
Let $\mathcal{F}_c,\widetilde{\tau}^{(n)}_i,M_n(a)$ be as in Theorem \ref{t-tv3}. By Theorem \ref{t-tv5}(2) and Remark \ref{r-cutoff}, if $\mathcal{F}_c$ has a maximum total variation cutoff, then
\[
\mathbb{E}_{M_n(a)}\widetilde{\tau}^{(n)}_{M_n(b)}
=o\left(\max\left\{\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n(c)},
\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n(c)}\right\}\right),\quad\forall a,b,c\in(0,1).
\]
\end{rem}
The following example gives counterexamples to the converse of (1) and (2) in Theorem \ref{t-tv5}.
\begin{ex}\label{ex-cutoff2}
Consider the family $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$, where $\mathcal{X}_n=\{0,1,...,n\}$ and
\[
\begin{cases}K_n(i,i+1)=1-\frac{i}{2n},\quad\forall 0\le i<n,\,i\ne i_n,\\
K_n(i+1,i)=\frac{i+1}{2n},\quad\forall 0\le i<n-1,\,i\ne i_n,\quad K_n(n,n-1)=1,\\
K_n(i_n,i_n+1)=c_n(1-\frac{i_n}{2n}),\quad K_n(i_n+1,i_n)=c_n\frac{i_n+1}{2n},\\
K_n(i_n,i_n)=(1-c_n)(1-\frac{i_n}{2n}),\quad K_n(i_n+1,i_n+1)=(1-c_n)\frac{i_n+1}{2n},\end{cases}
\]
with $0\le i_n<n$ and $c_n\in[0,1]$, and
\[
\pi_n(i)=2^{1-2n}\binom{2n}{i},\quad\forall 0\le i<n,\quad \pi_n(n)=2^{-2n}\binom{2n}{n}.
\]
As before, we use $M_n(a)$ to denote a state in $\mathcal{X}_n$ satisfying $\pi_n([0,M_n(a)])\ge a$ and $\pi_n([M_n(a),n])\ge 1-a$ and let $\widetilde{\tau}^{(n)}_i$ be the first hitting time to state $i$ of the continuous time chain associated with $(\mathcal{X}_n,K_n,\pi_n)$. Let $0<\lambda_{n,1}<\lambda_{n,2}<\cdots<\lambda_{n,n}$ be eigenvalues of $I-K_n$. It follows immediately from the central limit theorem that
\begin{equation}\label{eq-mna}
n-M_n(a)\asymp \sqrt{n},\quad\forall a\in(0,1).
\end{equation}
In what follows, we discuss the total variation cutoffs of $\mathcal{F}_c$, $\mathcal{F}_c^L$ and $\mathcal{F}_c^R$ with specific $c_n$ and $i_n$.
First, assume that $c_n=1$ for all $n$. In this setting, the chain $(\mathcal{X}_n,K_n,\pi_n)$ is exactly the collapsed chain of the Ehrenfest model on $\{0,1,...,2n\}$ obtained by combining states $\{i,2n-i\}$ into a new state for $0\le i<n$. The spectral information of the Ehrenfest model is well-studied and this implies
\[
\lambda_{n,i}=\frac{2i}{n},\quad\forall 1\le i\le n.
\]
By Theorem \ref{t-sep}, $\mathcal{F}_c$ has a maximum separation cutoff with cutoff time $\frac{1}{2}n\log n$ and, thus, has a maximum total variation cutoff. A simple computation with the Stirling formula gives
\[
\pi_n(i)\asymp \frac{1}{\sqrt{n}},\quad\text{uniformly for }M_n(a)\le i\le n.
\]
By Lemma \ref{l-meanvar2}, this implies that, for $a\in(0,1)$,
\[
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)}\asymp n,\quad \text{Var}_n\widetilde{\tau}_{M_n(a)}\asymp n^2,
\]
and, by Theorem \ref{t-sep2}, we have $\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}\sim \frac{1}{2}n\log n$ for any $a\in(0,1)$. As a consequence of Theorems \ref{t-tv3} and \ref{t-tv5}(2), $\mathcal{F}_c^R$ has no total variation cutoff, but $\mathcal{F}_c^L$ has with cutoff time $\frac{1}{2}n\log n$. Furthermore, by Theorem \ref{t-tv5}(1), the total variation cutoff time for $\mathcal{F}_c$ can be $\frac{1}{2}n\log n$. This gives a counterexample to the converse of Theorem \ref{t-tv5}(1).
Next, we consider the case $n-i_n=o(\sqrt{n})$ and $c_n$ is small. The second assumption means that a bottleneck arises between states $i_n$ and $i_n+1$. Under the first assumption, (\ref{eq-mna}) implies that, for $a\in(0,1)$, both $\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ and $\text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ remain the same as in the case $c_n=1$. This implies that $\mathcal{F}_c^L$ has a total variation cutoff with cutoff time $\frac{1}{2}n\log n$. For the cutoff of $\mathcal{F}_c^R$, one may compute using the formula in Lemma \ref{l-meanvar2} that, for any $a\in(0,1)$,
\[
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)}\asymp n+\frac{n-i_n}{c_n},\quad \text{Var}_n\widetilde{\tau}_{M_n(a)}\asymp
\left(n+\frac{n-i_n}{c_n}\right)^2.
\]
Consequently, Theorem \ref{t-tv3} implies that $\mathcal{F}_c^R$ has no cutoff in total variation. Moreover, Theorem \ref{t-tv2} implies that if $(n-i_n)/c_n=o(n\log n)$, then $\mathcal{F}_c$ has a maximum total variation cutoff. If $n\log n=O((n-i_n)/c_n)$, then $\mathcal{F}_c$ has no maximum total variation cutoff, which gives a counterexample to the converse of Theorem \ref{t-tv5}(2).
\end{ex}
The next theorem provides more information on the comparison of cutoffs and should be regarded as a complement to Theorem \ref{t-tv5}.
\begin{thm}\label{t-comp}
Let $\mathcal{F}=\{(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$ and $\mathcal{F}_c$ be the family of continuous time chains associated with $\mathcal{F}$. Suppose that, in total variation, $\mathcal{F}_c^L$ has a cutoff with cutoff time $t_n$ but no subsequence of $\mathcal{F}_c^R$ has a cutoff. Let $M_n$ be a state in $\mathcal{X}_n$ and set
\[
R=\limsup_{n\ra\infty}\frac{\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}}{t_n},\quad\forall a\in(0,1).
\]
Then, the following are equivalent.
\begin{itemize}
\item[(1)] $\mathcal{F}_c$ has a maximum total variation cutoff. In particular, $t_n$ is a cutoff time.
\item[(2)] $R=0$ for some sequence $(M_n)_{n=1}^\infty$ satisfying
\begin{equation}\label{eq-mn}
\inf_{n\ge 1}\pi_n([0,M_n])>0,\quad\inf_{n\ge 1}\pi_n([M_n,n])>0.
\end{equation}
\item[(3)] $R=0$ for any sequence $(M_n)_{n=1}^\infty$ satisfying \textnormal{(\ref{eq-mn})}.
\end{itemize}
The above statement also holds for $\mathcal{F}$ provided $\inf_{n,i}K_n(i,i)>0$.
\end{thm}
\begin{proof}
We first consider the continuous time case. Since $\mathcal{F}_c^L$ has a total variation cutoff with cutoff time $t_n$, Theorem \ref{t-tv3} implies
\begin{equation}\label{eq-tn}
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}\sim t_n,\quad \text{Var}_0\widetilde{\tau}^{(n)}_{M_n(a)}=o(t_n^2),\quad\forall a\in(0,1).
\end{equation}
Under the assumption of (\ref{eq-mn}), one may choose $0<a<b<1$ such that $M_n(a)\le M_n\le M_n(b)$. By (\ref{eq-tn}), this implies
\begin{equation}\label{eq-tn2}
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}\sim t_n,\quad \text{Var}_0\widetilde{\tau}^{(n)}_{M_n}=o(t_n^2).
\end{equation}
(3)$\Ra$(2) is obvious. Now, we prove (2)$\Ra$(1) and assume that (2) holds.
Note that $R=0$ is equivalent to $\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}=o(t_n)$. This implies $\text{Var}_n\widetilde{\tau}^{(n)}_{M_n}=o(t_n^2)$ using the fact $\text{Var}_n\widetilde{\tau}^{(n)}_i\le(\mathbb{E}_n\widetilde{\tau}^{(n)}_i)^2$. Along with (\ref{eq-tn2}), we may conclude
\begin{equation}\label{eq-tn3}
\sqrt{\max\left\{\text{Var}_0\widetilde{\tau}^{(n)}_{M_n},\text{Var}_n\widetilde{\tau}^{(n)}_{M_n}\right\}}
=o\left(\max\left\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n},\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}\right\}\right).
\end{equation}
By Theorem \ref{t-tv2}, $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$.
For (1)$\Ra$(3), we prove the equivalent implication by assuming that $R>0$ for some sequence $(M_n)_{n=1}^\infty$ satisfying (\ref{eq-mn}). Since $R>0$, we may choose a subsequence $(k_n)_{n=1}^\infty$ such that
\begin{equation}\label{eq-tn4}
t_{k_n}=O\left(\mathbb{E}_{k_n}\widetilde{\tau}^{(k_n)}_{M_{k_n}}\right).
\end{equation}
As the subfamily of $\mathcal{F}_c^R$ indexed by $(k_n)_{n=1}^\infty$ is assumed to have no total variation cutoff, we may refine, by Theorem \ref{t-tv3}, the selection of $k_n$ such that
\begin{equation}\label{eq-tn5}
\sqrt{\text{Var}_{k_n}\widetilde{\tau}^{(k_n)}_{M_{k_n}}}\asymp\mathbb{E}_{k_n}\widetilde{\tau}^{(k_n)}_{M_{k_n}}.
\end{equation}
Combining (\ref{eq-tn2}) with the above discussion leads to
\[
\sqrt{\max\left\{\text{Var}_0\widetilde{\tau}^{(k_n)}_{M_{k_n}},\text{Var}_{k_n}\widetilde{\tau}^{(k_n)}_{M_{k_n}}\right\}}
\asymp\max\left\{\mathbb{E}_0\widetilde{\tau}^{(k_n)}_{M_{k_n}},\mathbb{E}_{k_n}\widetilde{\tau}^{(k_n)}_{M_{k_n}}\right\}.
\]
By Theorem \ref{t-tv2}, the subfamily of $\mathcal{F}_c$ indexed by $(k_n)$ has no maximum total variation cutoff.
Next, we consider the discrete time case. (3)$\Ra$(2) is clear. For (2)$\Ra$(1), assume that $R=0$ for some sequence $M_n$ satisfying (\ref{eq-mn}). Observe that
\begin{equation}\label{eq->n}
\mathbb{E}_0\tau^{(n)}_{M_n}+\mathbb{E}_n\tau^{(n)}_{M_n}\ge n.
\end{equation}
By Remark \ref{r-cutoffitme}, (\ref{eq->n}) implies $t_n\ra\infty$. Using Theorem \ref{t-tv4}, one may derive a discrete time version of (\ref{eq-tn}), (\ref{eq-tn2}) and (\ref{eq-tn3}). As a consequence of Theorem \ref{t-tv2}, $\mathcal{F}$ has a maximum total variation cutoff with cutoff time $t_n$.
For (1)$\Ra$(3), we assume the inverse of (3) that $R>0$ for some sequence $M_n$ satisfying (\ref{eq-mn}). Consider the following two cases.
{\bf Case 1:} $t_{k_n}\ra\infty$ for some subsequence $k_n$.
{\bf Case 2:} $t_{k_n}=O(1)$ for some subsequence $k_n$.
The proof of Case 1 is the same as the continuous time case. For Case 2, since the subfamily of $\mathcal{F}^L$ indexed by $(k_n)$ has a cutoff with cutoff time $t_{k_n}$, Remark \ref{r-cutoffitme} implies that
\[
\mathbb{E}_0\tau^{(k_n)}_{M_{k_n}}=O(1),\quad \text{Var}_0\tau^{(k_n)}_{M_{k_n}}=O(1).
\]
By (\ref{eq->n}), we have $\mathbb{E}_{k_n}\tau^{(k_n)}_{M_{k_n}}\ra\infty$ and, by Theorem \ref{t-tv4}, we obtain a discrete version of (\ref{eq-tn4}) and then (\ref{eq-tn5}). Consequently, Theorem \ref{t-tv2} implies that $\mathcal{F}$ has no maximum total variation cutoff.
\end{proof}
The next theorem is a special version of Theorem \ref{t-tv5} which identifies two different cutoffs discussed in this section.
\begin{thm}\label{t-tv6}
Let $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ be a family of irreducible birth and death chains with $\mathcal{X}_n=\{0,...,n\}$ and $\mathcal{F}_c$ be the families of continuous time chains associated with $\mathcal{F}$. Assume that $K_n(i,j)=K_n(n-i,n-j)$ for all $i,j\in\mathcal{X}_n$ and $n\ge 1$.
\begin{itemize}
\item[(1)] $\mathcal{F}^L_c$ has a total variation cutoff with cutoff time $t_n$ if and only if $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$.
\item[(2)] Under the assumption that $\inf_{n,i}K_n(i,i)>0$, $\mathcal{F}^L$ has a total variation cutoff with cutoff time $t_n$ if and only if $\mathcal{F}$ has a maximum total variation cutoff with cutoff time $t_n$.
\end{itemize}
\end{thm}
\begin{proof}[Proof of Theorem \ref{t-tv5}(Continuous time case)]
As before, we use $\widetilde{\tau}^{(n)}_i$ to denote the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}_c$ and use the notation $M_n(a)$ with $a\in(0,1)$ to denote a state in $\mathcal{X}_n$ satisfying $\pi_n([0,M_n(a)])\ge a$ and $\pi_n([M_n(a),n])\ge 1-a$.
For (1), assume that $\mathcal{F}_c^L,\mathcal{F}_c^R$ have total variation cutoffs with cutoff times $r_n,s_n$. By Theorem \ref{t-tv3}, we have
\[
\sqrt{\text{Var}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)}}
=o\left(\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\right),
\quad \mathbb{E}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\sim r_{n},
\]
and
\[
\sqrt{\text{Var}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}}
=o\left(\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\right),
\quad \mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\sim s_{n}.
\]
Clearly, this implies
\[
\sqrt{\max\left\{\text{Var}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)},\text{Var}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\right\}}
=o\left(\max\left\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)},\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\right\}\right)
\]
and
\[
\max\left\{\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_{n}(1/2)},\mathbb{E}_{n}\widetilde{\tau}^{(n)}_{M_{n}(1/2)}\right\}
\sim \max\{r_{n},s_n\}=t_n.
\]
By Theorem \ref{t-tv2}, $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$.
For (2), let $\widehat{\mathcal{F}}=(\mathcal{X}_n,\widehat{K}_n,\widehat{\pi}_n)_{n=1}^\infty$ be a family given by
\[
\widehat{K}_n=K_n,\quad\widehat{\pi}_n=\pi_n\quad\text{if }x_n=0,
\]
and
\[
\widehat{K}_n(i,j)=K_n(n-i,n-j),\quad\widehat{\pi}_n(i)=\pi_n(n-i),\quad\forall i,j\in\mathcal{X}_n\quad\text{if }x_n=n.
\]
Let $\widehat{\mathcal{F}}_c$ be the family of continuous time chains associated with $\widehat{\mathcal{F}}$. Suppose that $\mathcal{F}_c$ has a maximum total variation cutoff with cutoff time $t_n$. It is obvious that $\widehat{\mathcal{F}}_c$ also has a maximum total variation cutoff with cutoff time $t_n$ and, to show that $\mathcal{F}_c^S$ has a total variation cutoff with cutoff time $t_n$, it is equivalent to prove that $\widehat{\mathcal{F}}_c^L$ has a total variation cutoff with cutoff time $t_n$.
Let $\widehat{\tau}^{(n)}_i$ be the first hitting time to state $i$ of the continuous time chain associated with $(\mathcal{X}_n,\widehat{K}_n,\widehat{\pi}_n)$ and set $\widehat{M}_n$ be a state defined by
\[
\widehat{M}_n=\begin{cases}M_n&\text{if }x_n=0\\n-M_n&\text{if }x_n=n\end{cases}.
\]
We use $\widehat{M}_n(a)$ to denote a state such that
\[
\widehat{\pi}_n([0,\widehat{M}_n(a)])\ge a,\quad \widehat{\pi}_n([\widehat{M}_n(a),n])\ge 1-a.
\]
By Theorem \ref{t-tv2}, the total variation cutoff of $\widehat{\mathcal{F}}_c$ with cutoff time $t_n$ implies
\[
t_n\sim\max\left\{\mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n},
\mathbb{E}_n\widehat{\tau}^{(n)}_{\widehat{M}_n}\right\}=\mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n}
\]
and, for any $a\in(0,1)$,
\begin{equation}\label{eq-cutoff}
\sqrt{\max\left\{\text{Var}_0\widehat{\tau}^{(n)}_{\widehat{M}_n(a)},
\text{Var}_{n}\widehat{\tau}^{(n)}_{\widehat{M}_{n}(a)}\right\}}=o(t_n)
=o\left(\mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n}\right).
\end{equation}
As a result of Lemma \ref{l-hit2} and (\ref{eq-cutoff}), we have, for $0<b<a<1$,
\[
\mathbb{E}_{\widehat{M}_n(b)}\widehat{\tau}^{(n)}_{\widehat{M}_n(a)}
=O\left(\sqrt{\text{Var}_{\widehat{M}_n(b)}\widehat{\tau}^{(n)}_{\widehat{M}_n(a)}}\right)
=o\left(\mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n}\right),
\]
which leads to
\[
\mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n(a)}\sim \mathbb{E}_0\widehat{\tau}^{(n)}_{\widehat{M}_n},\quad\forall a\in(0,1).
\]
Applying the last identity to (\ref{eq-cutoff}) yields
\[
\sqrt{\text{Var}_0\widehat{\tau}^{(n)}_{\widehat{M}_n(a)}}
=o\left(\mathbb{E}_{0}\widehat{\tau}^{(n)}_{\widehat{M}_n(a)}\right),\quad\forall a\in(0,1).
\]
By Theorem \ref{t-tv3}, $\widehat{\mathcal{F}}_c^L$ has a total variation cutoff with cutoff time $t_n$. The precise description of the cutoff time and window is given by Theorem \ref{t-tv2}, Corollary \ref{c-tv} and Remark \ref{r-sep}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-tv5}(Discrete time case)]
We use $\tau^{(n)}_i$ to denote the first hitting time to state $i$ of the $n$th chain in $\mathcal{F}$ and $M_n(a)$ for a state in $\mathcal{X}_n$ satisfying $\pi_n([0,M_n(a)])\ge a$ and $\pi_n([M_n(a),n])\ge 1-a$.
For (1), assume that $\mathcal{F}^L,\mathcal{F}^R$ have cutoffs with respective cutoff times $r_n,s_n$. Given an increasing sequence $\mathcal{K}=(k_n)_{n=1}^\infty$ in $\{1,2,...\}$, let $\mathcal{F}(\mathcal{K})$ be the family of chains in $\mathcal{F}$ indexed by the sequence $\mathcal{K}$. By Proposition 2.1 in \cite{CSal10}, to prove $\mathcal{F}$ has a maximum total variation cutoff, it suffices to show that, for any increasing sequence of positive integers, there is a subsequence, say $\mathcal{K}$, such that $\mathcal{F}(\mathcal{K})$ has a maximum total variation cutoff. Note that, by Remark \ref{r-cutoffitme}, $r_n+s_n$ must tend to infinity. This implies that $\mathcal{K}$ can be chosen to satisfy one of the following cases.
{\bf Case 1:} $r_{k_n}\ra\infty$ and $s_{k_n}\ra\infty$.
{\bf Case 2:} $r_{k_n}\ra\infty$ and $s_{k_n}=O(1)$.
{\bf Case 3:} $r_{k_n}=O(1)$ and $s_{k_n}\ra\infty$.
The proof for Case 1 is the same as the continuous time case. The proofs of Case 2 and Case 3 are similar and we discuss Case 2, here. By Theorem \ref{t-tv4} and Remark \ref{r-cutoffitme}, the cutoffs of $\mathcal{F}^L,\mathcal{F}^R$ imply that, for $a\in(0,1)$,
\[
\mathbb{E}_0\tau^{(k_n)}_{M_{k_n}(a)}\sim r_{k_n},\quad \sqrt{\text{Var}_0\tau^{(k_n)}_{M_{k_n}(a)}}=o(r_{k_n}),
\]
and
\[
\sqrt{\text{Var}_{k_n}\tau^{(k_n)}_{M_{k_n}(a)}}
\le\mathbb{E}_{k_n}\tau^{(k_n)}_{M_{k_n}(a)}=O(1).
\]
This implies, for $a\in(0,1)$,
\[
\sqrt{\max\left\{\text{Var}_0\tau^{(k_n)}_{M_{k_n}(a)},
\text{Var}_{k_n}\tau^{(k_n)}_{M_{k_n}(a)}\right\}}
=o\left(\max\left\{\mathbb{E}_0\tau^{(k_n)}_{M_{k_n}(a)},
\mathbb{E}_{k_n}\tau^{(k_n)}_{M_{k_n}(a)}\right\}\right)
\]
and
\[
\max\left\{\mathbb{E}_0\tau^{(k_n)}_{M_{k_n}(a)},
\mathbb{E}_{k_n}\tau^{(k_n)}_{M_{k_n}(a)}\right\}\sim \max\{r_{k_n},s_{k_n}\}=t_{k_n}.
\]
By Theorem \ref{t-tv2}, $\mathcal{F}(\mathcal{K})$ has a maximum total variation cutoff with cutoff time $t_{k_n}$.
For (2), based on the following observation
\[
n\le \mathbb{E}_0\tau^{(n)}_i+\mathbb{E}_n\tau^{(n)}_i,\quad\forall 0\le i\le n,
\]
we have $\mathbb{E}_{x_n}\tau^{(n)}_{M_n}\ra \infty$. The remaining proof is similar to the continuous time case and is skipped.
\end{proof}
\section{Proof of Theorems \ref{t-tv3}, \ref{t-tv4} and Corollary \ref{c-tv}}\label{s-proof}
This section is dedicated to the proof of Theorems \ref{t-tv3} and \ref{t-tv4} and we need the following lemmas.
\begin{lem}\label{l-hit2}
Let $(\mathcal{X},K,\pi)$ be an irreducible birth and death chain on $\{0,1,...,n\}$ and $\tau_i,\widetilde{\tau}_i$ be the first hitting times to state $i$ of the discrete time chain and the associated continuous time chain. Let $\lambda_i$ be the smallest eigenvalue of the submatrix of $I-K$ indexed by $0,...,i-1$. Then, for $i<j$,
\[
\frac{\pi([0,i])}{2\pi([0,j-1])}(\mathbb{E}_i\widetilde{\tau_j})^2\le
\textnormal{Var}_i(\widetilde{\tau}_j)\le\frac{2}{\lambda_j}\mathbb{E}_i\widetilde{\tau}_j
\]
and
\[
\frac{\delta\pi([0,i])}{2\pi([0,j-1])}(\mathbb{E}_i\tau_j)^2\le
\textnormal{Var}_i(\tau_j)\le\frac{2}{\lambda_j}\mathbb{E}_i\tau_j,
\]
where $\delta=\min_iK(i,i)$. In particular,
\[
\mathbb{E}_i\tau_j=\mathbb{E}_i\widetilde{\tau}_j\le\frac{4\pi([0,j-1])}{\pi([0,i])\lambda_j}.
\]
\end{lem}
\begin{lem}\label{l-hit1}
Let $K$ be the transition matrix of an irreducible birth and death chain on $\{0,1,...,n\}$ and $\widetilde{\tau}_i$ be the first hitting time to state $i$ for the continuous time chain associated with $K$. For $0<i\le n$ and $a\in(0,1)$,
\[
\mathbb{P}_0(\widetilde{\tau}_i>a\mathbb{E}_0\widetilde{\tau}_i)\ge \min\left\{e^{-\sqrt{a}},\frac{(1-a)^2}{\sqrt{a}+(1-a)^2}\right\}.
\]
\end{lem}
\begin{lem}\label{l-hit3}
Let $K$ be the transition matrix of an irreducible birth and death chain on $\mathcal{X}=\{0,1,...,n\}$ with transition rates $p_i,q_i,r_i$ and stationary distribution $\pi$. Let $\tau_i,\widetilde{\tau}_i$ be as in Lemma \ref{l-hit2}. Then, for $i<j<k$,
\[
\mathbb{E}_j\min\{\tau_i,\tau_k\}=\mathbb{E}_j\min\{\widetilde{\tau}_i,\widetilde{\tau}_k\}=A/B,
\]
where
\[
A=\sum_{\begin{subarray}{c}i+1\le \ell_1\le j\\ j\le\ell_2\le k-1 \end{subarray}}\frac{\pi([\ell_1,\ell_2])}
{\pi(\ell_1)q_{\ell_1}\pi(\ell_2)p_{\ell_2}},\quad B=\sum_{\ell=i}^{k-1}\frac{1}{\pi(\ell)p_{\ell}}.
\]
\end{lem}
\begin{lem}\label{l-unimodal}
Let $(\mathcal{X},K,\pi)$ be an irreducible birth and death chain on $\{0,1,...,n\}$ and $H_t=e^{-t(I-K)}$. Then,
\begin{itemize}
\item[(1)] $H_t(0,i)/\pi(i)\ge H_t(0,i+1)/\pi(i+1)$ for $0\le i<n$ and $t\ge 0$,
\item[(2)] Assume that $\min_iK(i,i)\ge 1/2$. Then, $K^m(0,i)/\pi(i)\ge K^m(0,i+1)/\pi(i+1)$ for $0\le i<n$ and $m\ge 0$.
\end{itemize}
\end{lem}
We relegate the proofs of Lemmas \ref{l-hit2}, \ref{l-hit1} and \ref{l-hit3} to the appendix and refer the reader to Lemma 4.1 in \cite{DLP10} for a proof of Lemma \ref{l-unimodal}.
\begin{proof}[Proof of Theorem \ref{t-tv3}]
We first prove the equivalence for cutoffs. Note that $\pi_n(0)\ra 0$ is necessary for the total variation cutoff since
\[
\liminf_{n\ra\infty}d_{n,\text{\tiny TV}}^{(c)}(0,t)\le \liminf_{n\ra\infty}d_{n,\text{\tiny TV}}^{(c)}(0,0)=1-\limsup_{n\ra\infty}\pi_n(0).
\]
Under the assumption that $\pi_n(0)\ra 0$, it is easy to see that, for any $a\in(0,1)$, $M_n(a)\ge 1$ if $n$ is large enough. For $a\in(0,1)$ and $n\ge 1$ such that $M_n(a)\ge 1$, we let
\[
\lambda_{n,1}(a)<\cdots<\lambda_{n,M_n(a)}(a)
\]
be the eigenvalues of the submatrix of $I-K_n$ indexed by $0,1,...,M_n(a)-1$. Clearly, $\lambda_n(a)=\lambda_{n,1}(a)$ and, by Lemma \ref{l-meanvar},
\[
u_n(a)=\sum_{i=1}^{M_n(a)}\frac{1}{\lambda_{n,i}(a)},\quad v_n^2(a)=\sum_{i=1}^{M_n(a)}\frac{1}{\lambda^2_{n,i}(a)}.
\]
As in the proof of (\ref{eq-comp}), we have
\[
\sqrt{u_n(a)\lambda_n(a)}\le\frac{u_n(a)}{v_n(a)}\le u_n(a)\lambda_n(a).
\]
This implies the equivalence of (2) and (3).
To prove the remaining equivalences, we let $d_{n,\text{\tiny TV}}^{(c)}$ be the total variation distance of the $n$th chains. By Lemma \ref{l-tv}, one has
\begin{equation}\label{eq-tv}
d_{n,\text{\tiny TV}}^{(c)}(0,t)\begin{cases}\le\mathbb{P}_0(\widetilde{\tau}^{(n)}_i>t)+\pi_n([i+1,n]),\\
\ge\mathbb{P}_0(\widetilde{\tau}^{(n)}_i>t)-\pi_n([0,i-1]).\end{cases}
\end{equation}
As a result of the one-sided Chebyshev inequality, this implies
\begin{equation}\label{eq-mixing}
T_{n,\text{\tiny TV}}^{(c)}(0,\epsilon)\begin{cases}\le\mathbb{E}_0\widetilde{\tau}^{(n)}_i
+\sqrt{(\tfrac{1-\delta}{\delta})\text{Var}_0(\widetilde{\tau}^{(n)}_i)}&\text{for } \epsilon=\delta+\pi_n([i+1,n]),\\
\ge\mathbb{E}_0\widetilde{\tau}^{(n)}_i-\sqrt{(\tfrac{\delta}{1-\delta})\text{Var}_0(\widetilde{\tau}^{(n)}_i)}&
\text{for }\epsilon=\delta-\pi_n([0,i-1]),\end{cases}
\end{equation}
where $\delta\in(0,1)$.
Now, we prove (2)$\Ra$(1) and assume that (2) holds. By the last inequality of Lemma \ref{l-hit2}, we have, for $0<\delta<\epsilon<1$,
\begin{equation}\label{eq-unvn}
0\le u_n(\epsilon)-u_n(\delta)\le\frac{4\epsilon}{\delta\lambda_n(\epsilon)}\le\frac{4\epsilon v_n(\epsilon)}{\delta}=o(u_n(\epsilon)).
\end{equation}
Fix $\epsilon\in(0,1)$ and let $0<\epsilon_1<\epsilon<\epsilon_2<1$. By (\ref{eq-mixing}), the replacement of $i=M_n(\epsilon_2)$, $\delta=\epsilon_2-\epsilon$ in the first inequality and the replacement of $i=M_n(\epsilon_1)$, $\delta=1-\epsilon+\epsilon_1$ in the second inequality yield
\[
\begin{cases}T_{n,\text{\tiny TV}}^{(c)}(0,1-\epsilon)\le u_n(\epsilon_2)+\sqrt{(\frac{1}{\epsilon_2-\epsilon}-1)}v_n(\epsilon_2)=(1+o(1))u_n(\epsilon_2),\\
T_{n,\text{\tiny TV}}^{(c)}(0,1-\epsilon)\ge u_n(\epsilon_1)-\sqrt{(\frac{1}{\epsilon-\epsilon_1}-1)}v_n(\epsilon_1)=(1+o(1))u_n(\epsilon_1).\end{cases}
\]
As a result of (\ref{eq-unvn}), we obtain that $T_{n,\text{\tiny TV}}^{(c)}(0,\epsilon)=(1+o(1))u_n(\eta)$ for any $\epsilon,\eta\in(0,1)$,
which proves (1).
Next, we prove (4)$\Ra$(3). Assume that $(t_n)_{n=0}^\infty$ is a positive sequence satisfying $t_n=O(u_n(c))$ for all $c\in(0,1)$ and $a\in(0,1)$ is a constant such that
\begin{equation}\label{eq-lim1}
\lim_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(b)}>(1-\epsilon)t_n\right)=1,\quad\forall b\in(a,1),
\end{equation}
and, for any $b\in(a,1)$, there corresponds a constant $\alpha_b\in(0,1)$ such that
\begin{equation}\label{eq-lim2} \limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(b)}>(1+\epsilon)t_n\right)\le\alpha_b,
\end{equation}
for all $\epsilon\in(0,1)$. Note that $\lambda_n(a_2)\le\lambda_n(a_1)$ for $0<a_1<a_2<1$. To prove (3), it suffices to show that $t_n\lambda_n(b)\ra\infty$ for all $b\in(a,1)$. Now, we fix $b\in(a,1)$. Since $\pi_n(0)\ra 0$, it is clear that $M_n(b)\ge 1$ for $n$ large enough. By \cite{BS87}, if $M_n(b)\ge 1$, we may write $\widetilde{\tau}^{(n)}_{M_n(b)}=T_n(b)+S_n(b)$, where $T_n(b)$ and $S_n(b)$ are independent, $T_n(b)$ is an exponential random variable with parameter $\lambda_n(b)$ and $S_n(b)$ is a sum of independent exponential random variables with parameters $\lambda_{n,2}(b),...,\lambda_{n,M_n(b)}(b)$. Note that
\begin{align}
\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(b)}>(1-\epsilon)t_n\right)&=\int_0^\infty
\lambda_n(b)e^{-\lambda_n(b)s}\mathbb{P}_0(S_n(b)>(1-\epsilon)t_n-s)ds\notag\\
&\le (1-e^{-\lambda_n(b)t})\mathbb{P}_0(S_n(b)>(1-\epsilon)t_n-t)+e^{-\lambda_n(b)t},\notag
\end{align}
where the inequality is obtained by separating the region of integration into $(0,t)$ and $[t,\infty)$,
and
\begin{align}
\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(b)}>(1+\epsilon)t_n\right)&=\int_0^\infty
\lambda_n(b)e^{-\lambda_n(b)s}\mathbb{P}_0(S_n(b)>(1-\epsilon)t_n-s)ds\notag\\
&\ge\mathbb{P}_0(S_n(b)>(1+\epsilon)t_n-r)e^{-\lambda_n(b)r}.\notag
\end{align}
By (\ref{eq-lim1}) and (\ref{eq-lim2}), the replacement of $t=C/\lambda_n(b)$ and $r=2C/\lambda_n(b)$ with $C=\frac{1}{4}\log\frac{1}{\alpha_b}$ in the above inequalities yields that, for all $\epsilon\in(0,1)$,
\[
\lim_{n\ra\infty}\mathbb{P}_0(S_n(b)>(1-\epsilon)t_n-C/\lambda_n(b))=1
\]
and
\[
\limsup_{n\ra\infty}\mathbb{P}_0(S_n(b)>(1+\epsilon)t_n-2C/\lambda_n(b))\le\sqrt{\alpha_b}<1.
\]
As a consequence, for $\epsilon\in(0,1)$, if $n$ is large enough, one has
\[
(1+\epsilon)t_n-2C/\lambda_n(b)\ge (1-\epsilon)t_n-C/\lambda_n(b),
\]
which implies $t_n\lambda_n(b)\ge C/(2\epsilon)$. This proves $t_n\lambda_n(b)\ra \infty$.
To finish the proof of those equivalences, it remains to show (1)$\Ra$(4). Assume that $\mathcal{F}_c$ has a cutoff with cutoff time $t_n$. The replacement of $i=M_n(a)$ in (\ref{eq-tv}) implies that, for all $\epsilon\in(0,1)$,
\begin{equation}\label{eq-liminf}
\liminf_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(a)}>(1-\epsilon)t_n\right)\ge a
\end{equation}
and
\begin{equation}\label{eq-limsup}
\limsup_{n\ra\infty}
\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(a)}>(1+\epsilon)t_n\right)\le a.
\end{equation}
By the Markov inequality, (\ref{eq-liminf}) implies that $t_n=O(u_n(a))$ for all $a\in(0,1)$. As a result of Lemma \ref{l-hit1}, (\ref{eq-limsup}) implies that $u_n(a)=O(t_n)$ for all $a\in(0,1)$, which leads to $t_n\asymp u_n(a)$ for all $a\in(0,1)$.
To fulfill the requirement in (4), one has to prove that there is $a\in(0,1)$ such that
\begin{equation}\label{eq-convto1}
\lim_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n)}_{M_n(a)}>(1-\epsilon)t_n\right)=1,\quad\forall \epsilon\in(0,1).
\end{equation}
To see the above limit, we fix $\epsilon\in(0,1)$ and show that, for any subsequence of positive integers, there is a further subsequence satisfying (\ref{eq-convto1}). Let $k_n$ be a subsequence of positive integers and set
\[
R(a):=\lim_{b\ra 1}\liminf_{n\ra\infty}\frac{\mathbb{E}_{M_{k_n}(a)}\widetilde{\tau}^{(k_n)}_{M_{k_n}(b)}}{t_{k_n}}.
\]
Clearly, $R(a)$ is nonnegative and non-increasing in $a$.
We consider the following two cases of $R(a)$. First, assume that $R(a)=0$ for some $a\in(0,1)$ and let $b_n$ be a sequence in $(a,1)$ that converges to $1$. Since $R(b_1)=0$, we may choose $\ell_1\in\{k_1,k_2,...\}$ such that $\mathbb{E}_{M_{\ell_1}(a)}\widetilde{\tau}^{(\ell_1)}_{M_{\ell_1}(b_1)}<t_{\ell_1}/2$. Inductively, for $n\ge 1$, we may select, according to the fact $R(b_{n+1})=0$, a constant $\ell_{n+1}\in\{k_1,k_2,...\}$ satisfying $\ell_{n+1}>\ell_n$ and
\[
\mathbb{E}_{M_{\ell_{n+1}}(a)}\widetilde{\tau}^{(\ell_{n+1})}_{M_{\ell_{n+1}}(b_{n+1})}<t_{\ell_{n+1}}/2^{n+1}.
\]
This implies
\[
\mathbb{E}_{M_{\ell_n}(a)}\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}=o(t_{\ell_n}),\quad\forall b\in(a,1).
\]
By Lemma \ref{l-meanvar}, $u_n(a)\asymp t_n$ implies $1/\lambda_n(a)=O(t_n)$ and, by Lemma \ref{l-hit2}, this yields $\text{Var}_{M_{\ell_n}(a)}\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}=o(t_{\ell_n}^2)$ for all $b\in(a,1)$. As a consequence of the one-sided Chebyshev inequality, we obtain
\[
\lim_{n\ra\infty}\mathbb{P}_{M_{\ell_n}(a)}\left(\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}\le\eta t_{\ell_n}\right)=1,\quad\forall b\in(a,1),\,\eta>0.
\]
This leads to
\begin{align}
&\liminf_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(a)}
>(1-\epsilon)t_{\ell_n}\right)\notag\\
\ge&\liminf_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}
>(1-\epsilon/2)t_{\ell_n},
\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}-\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(a)}
\le\epsilon t_{\ell_n}/2\right)\notag\\
=&\liminf_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(\ell_n)}_{M_{\ell_n}(b)}
>(1-\epsilon/2)t_{\ell_n}\right)\ge b,\notag
\end{align}
for all $b\in(a,1)$, where the last inequality uses (\ref{eq-liminf}). Letting $b$ tend to $1$ gives the desired limit.
Next, we assume that $R(a)>0$ for all $a\in(0,1)$. Along with this fact $u_n(a)\asymp t_n$ for all $a\in(0,1)$, it is easy to see that, for any $a\in(0,1)$, there is $b\in(a,1)$ such that $\mathbb{E}_{M_{k_n}(a)}\widetilde{\tau}^{(k_n)}_{M_{k_n}(b)}\asymp t_{k_n}$. To prove (\ref{eq-convto1}) for the subsequence $k_n$, we need the following discussion. For $n\ge 1$, set $H_{n,t}=e^{-t(I-K_n)}$ and let $(X_{n,t})_{t\ge 0}$ be a realization of the semigroup $H_{n,t}$ and, for $\eta\in(0,1)$, let
\[
N_n(\eta)=\max\{0\le i\le n|H_{n,(1-\eta)t_n}(0,i)>\pi_n(i)\}.
\]
By Lemma \ref{l-unimodal}, we have
\[
d_{n,\text{\tiny TV}}^{(c)}(0,(1-\eta)t_n)=H_{n,(1-\eta)t_n}(0,[0,N_n(\eta)])-\pi_n([0,N_n(\eta)]).
\]
Since $\mathcal{F}_c$ has a cutoff with cutoff time $(t_n)_{n=1}^\infty$, this implies
\[
\lim_{n\ra\infty}H_{n,(1-\eta)t_n}(0,[0,N_n(\eta)])=1,\quad \lim_{n\ra\infty}\pi_n([0,N_n(\eta)])=0.
\]
Obviously, this yields
\begin{equation}\label{eq-convto1-2}
\lim_{n\ra\infty}\mathbb{P}_0\left(X_{n,(1-\eta)t_n}\le M_n(a)\right)=1,\quad\forall a,\eta\in(0,1).
\end{equation}
Back to the case that $R(a)>0$ for all $a\in(0,1)$, one may choose $0<b<a^-<a<a^+<c<1$ such that
\begin{equation}\label{eq-sub}
\mathbb{E}_{M_{k_n}(b)}\widetilde{\tau}^{(k_n)}_{M_{k_n}(a^-)}\asymp t_{k_n}\asymp
\mathbb{E}_{M_{k_n}(a^+)}\widetilde{\tau}^{(k_n)}_{M_{k_n}(c)}.
\end{equation}
This implies that $M_{k_n}(b)<M_{k_n}(a^-)$ and $M_{k_n}(a^+)<M_{k_n}(c)$ for $n$ large enough. Next, let $L$ be a positive integer and set
\[
\Delta_n=\Delta_n(L):=\frac{(1-\epsilon)t_n}{L}.
\]
Note that, for $0\le j\le L-1$,
\begin{align}
&\mathbb{P}_0\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(a)}\in(j\Delta_{k_n},(j+1)\Delta_{k_n}],
X_{k_n,(j+1)\Delta_{k_n}}\le M_{k_n}(b)\right)\notag\\
\le&\mathbb{P}_0\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(a)}\in(j\Delta_{k_n},(j+1)\Delta_{k_n}]\right)
\mathbb{P}_{M_{k_n}(a)}\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(b)}\le\Delta_{k_n}\right).\notag
\end{align}
By (\ref{eq-convto1-2}), summing up the above inequalities over $j$ and then passing $n$ to the infinity yields
\begin{align}
&\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(a)}\le(1-\epsilon)t_{k_n}\right)\notag\\
\le&\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(a)}\le(1-\epsilon)t_{k_n}\right)
\times\limsup_{n\ra\infty}\mathbb{P}_{M_{k_n}(a)}\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(b)}\le\Delta_{k_n}\right).\notag
\end{align}
Observe that if there is $L>0$ such that
\begin{equation}\label{eq-convto<1}
\limsup_{n\ra\infty}\mathbb{P}_{M_{k_n}(a)}\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(b)}\le\Delta_{k_n}\right)<1,
\end{equation}
then
\[
\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(k_n)}_{M_{k_n}(a)}\le(1-\epsilon)t_{k_n}\right)=0,
\]
as desired. To get the limit in (\ref{eq-convto<1}), it suffices to show that there is $L>0$ such that
\[
\limsup_{n\ra\infty}\mathbb{P}_{M_{k_n}(a)}\left(T_{k_n}\le\Delta_{k_n}\right)<1,
\]
where $T_n=\min\{\widetilde{\tau}^{(n)}_{M_{n}(b)},\widetilde{\tau}^{(n)}_{M_{n}(c)}\}$. By Lemma \ref{l-hit3}, $\mathbb{E}_{M_{k_n}(a)}T_{k_n}=A_{k_n}/B_{k_n}$, where
\[
A_n=\sum_{\begin{subarray}{c}M_n(b)+1\le \ell_1\le M_n(a)\\ M_n(a)\le\ell_2\le M_n(c)-1 \end{subarray}}
\frac{\pi_n([\ell_1,\ell_2])}{\pi_n(\ell_1)q_{n,\ell_1}\pi_n(\ell_2)p_{n,\ell_2}},\quad B_n=\sum_{\ell=M_n(b)}^{M_n(c)-1}\frac{1}{\pi_n(\ell)p_{n,\ell}}.
\]
It is easy to see from the first identity in Lemma \ref{l-meanvar2} that
\[
A_
\ge(a^+-a^-)\mathbb{E}_{M_n(b)}\widetilde{\tau}^{(n)}_{M_n(a^-)}
\mathbb{E}_{M_n(a^+)}\widetilde{\tau}^{(n)}_{M_n(c)},\quad B_n\le\mathbb{E}_{M_n(b)}\widetilde{\tau}^{(n)}_{M_n(c)}/b.
\]
Along with the fact that $u_n(a)\asymp t_n$ for all $a\in(0,1)$, one may apply (\ref{eq-sub}) to the above inequalities to get $\mathbb{E}_{M_{k_n}(a)}T_{k_n}\asymp t_{k_n}$. Now, we choose $L>0$ such that
\[
0<\mathbb{E}_{M_{k_n}(a)}T_{k_n}-\Delta_{k_n}\asymp t_{k_n},
\]
where the first inequality holds for $n$ large enough. Since $T_n\le \widetilde{\tau}^{(n)}_{M_n(c)}$, one also has
\begin{align}
\text{Var}_{M_n(a)}T_n&\le\mathbb{E}_{M_n(a)}T_n^2\le\mathbb{E}_{M_n(a)}(\widetilde{\tau}^{(n)}_{M_n(c)})^2
=\text{Var}_{M_n(a)}\widetilde{\tau}^{(n)}_{M_n(c)}+(\mathbb{E}_{M_n(a)}(\widetilde{\tau}^{(n)}_{M_n(c)})^2\notag\\
&\le\text{Var}_{0}\widetilde{\tau}^{(n)}_{M_n(c)}+(\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n(c)})^2
\le 2(\mathbb{E}_{0}\widetilde{\tau}^{(n)}_{M_n(c)})^2=2u_n(c)^2\asymp t_n^2.\notag
\end{align}
As a result of the one-sided Chebyshev inequality, this implies
\[
\limsup_{n\ra\infty}\mathbb{P}_{M_{k_n}(a)}\left(T_{k_n}\le\Delta_{k_n}\right)\le\limsup_{n\ra\infty}
\left(1+\frac{(\mathbb{E}_{M_{k_n}(a)}T_{k_n}-\Delta_{k_n})^2}{\text{Var}_{M_{k_n}(a)}T_{k_n}}\right)^{-1}<1.
\]
In the assumption of (2) and (3), the proof for choosing $(u_n(a))_{n=1}^\infty$ as a cutoff time is given in the proof for (2)$\Ra$(1). In the assumption of (4), the equivalence of cutoffs implies that $v_n(a)=o(u_n(a))$ for all $a\in(0,1)$. As a consequence of the Chebyshev inequality, this yields that, for all $a\in(0,1)$ and $c>0$,
\[
\limsup_{n\ra\infty}\mathbb{P}_0\left(\left|\widetilde{\tau}^{(n)}_{M_n(a)}-u_n(a)\right|
>cu_n(a)\right)\le\limsup_{n\ra\infty}\frac{v_n(a)^2}{c^2u_n(a)^2}=0.
\]
Along with the assumption of (4), one has $t_n\sim\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n(a)}$ for all $a\in(0,1)$, as desired.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-tv4}]
Set
\[
\delta=\inf_{i,n}K_n(i,i),\quad K_n^{(\delta)}=(K_n-\delta I)/(1-\delta),\quad H^{(\delta)}_{n,t}=e^{t(K_n^{(\delta)}-I)}.
\]
It is easy to see that $(\mathcal{X}_n,K_n,\pi_n)$ and $(\mathcal{X}_n,H_{n,t}^{(\delta)},\pi_n)$ are respectively the $\delta$-lazy walk and the continuous time chain associated with $(\mathcal{X}_n,K_n^{(\delta)},\pi_n)$. Let $d_{n,\text{\tiny TV}},d_{n,\text{\tiny TV}}^{(c,\delta)}$ and $T_{n,\text{\tiny TV}},T_{n,\text{\tiny TV}}^{(c,\delta)}$ and $\tau_i^{(n)},\tau_i^{(n,\delta)}$ be respectively the total variation distances, the total variation mixing times and the first hitting times to state $i$ of chains $(\mathcal{X}_n,K_n,\pi_n)$ and $(\mathcal{X}_n,H_{n,t}^{(\delta)},\pi_n)$. As a result of the following observation
\begin{equation}\label{eq-kernelcomp}
H_{n,t}^{(\delta)}=e^{t(K_n^{(\delta)}-I)}=e^{t(K_n-I)/(1-\delta)},
\end{equation}
it is easy to see that the ratio of the spectral gaps of $(\mathcal{X}_n,K_n,\pi_n)$ and $(\mathcal{X}_n,H_{n,t}^{(\delta)},\pi_n)$ is constant in $n$ and, further,
\begin{equation}\label{eq-comp3}
\mathbb{E}_0\widetilde{\tau}^{(n,\delta)}_{M_n(a)}=(1-\delta)u_n(a),\quad
\text{Var}_0\widetilde{\tau}^{(n,\delta)}_{M_n(a)}\asymp w_n(a),
\end{equation}
where the latter also uses Remark \ref{r-wcomp}. This is consistent with (\ref{eq-cut3}).
Set $\mathcal{F}^{(\delta)}_c=(\mathcal{X}_n,H_{n,t}^{(\delta)},\pi_n)_{n=1}^\infty$ and let $\mathcal{F}_c^{(\delta,L)}$ denote the family of chains in $\mathcal{F}_c^{(\delta)}$ started at the left boundary points. The remaining proof for the equivalence of (1), (2) and (3) is very similar to the proof of the discrete time case in Theorem \ref{t-tv2} if (\ref{eq-cut1}) and (\ref{eq-cut2}) hold under the replacement of $\mathcal{F},\mathcal{F}_c^{(\delta)}$ by $\mathcal{F}^L,\mathcal{F}_c^{(\delta,L)}$. These two equivalences are given by Theorem 3.4 in \cite{CSal13-1} but the prerequisite of this theorem asks the existence of some $\epsilon\in(0,1)$ such that $T_{n,\text{\tiny TV}}(0,\epsilon)\ra\infty$ and $T_{n,\text{\tiny TV}}^{(c,\delta)}(0,\epsilon)\ra\infty$. (The authors of \cite{CSal13-1} point out the observation that such a requirement is missed in their article.)
First, consider the requirement $T_{n,\text{\tiny TV}}^{(c,\delta)}(0,\epsilon)\ra\infty$. Recall the second inequality in Lemma \ref{l-tv} in the following
\[
d_{n,\text{\tiny TV}}^{(c,\delta)}(0,t)\ge \mathbb{P}_0\left(\widetilde{\tau}_{M_n(a)}^{(n,\delta)}>t\right)-a.
\]
By Lemma \ref{l-hit1}, (\ref{eq-comp3}) and the fact $\text{Var}_0\widetilde{\tau}^{(n,\delta)}_{M_n(a)}
\le(\mathbb{E}_0\widetilde{\tau}^{(n,\delta)}_{M_n(a)})^2$, the above inequality implies
\[
d_{n,\text{\tiny TV}}^{(c,\delta)}(0,\alpha(1-\delta)u_n(a))
\ge\min\left\{e^{-\sqrt{\alpha}},\frac{(1-\alpha)^2}{\sqrt{\alpha}+(1-\alpha)^2}\right\}
-a,\quad\forall \alpha\in(0,1).
\]
This yields that
\begin{equation}\label{eq-liminf2}
\liminf_{n\ra\infty}\frac{T_{n,\text{\tiny TV}}^{(c,\delta)}(0,\epsilon)}{u_n(a)}>0\quad\text{for $\epsilon$ small enough}.
\end{equation}
Since $u_n(a)\ra\infty$, we have $T_{n,\text{\tiny TV}}^{(c,\delta)}(0,\epsilon)\ra\infty$ for $\epsilon$ small enough.
Next, we prove $T_{n,\text{\tiny TV}}(0,\epsilon)\ra\infty$. Note that one may use (\ref{eq-kernelcomp}) and the triangle inequality to derive
\begin{equation}\label{eq-comp2}
d_{n,\text{\tiny TV}}^{(c,\delta)}(0,t)\le\mathbb{P}(N_t\le m)+\mathbb{P}(N_t>m)d_{n,\text{\tiny TV}}(0,m),
\end{equation}
where $(N_t)_{t\ge 0}$ is a Poisson process with parameter $1/(1-\delta)$. A simple application of the weak law of large numbers says that $N_t/t$ converges to $1/(1-\delta)$ in probability as $t$ tends to infinity. By (\ref{eq-liminf2}) and the assumption $u_n(a)\ra\infty$, the replacement of $t=\beta u_n(a)$ and $m=\lceil \beta u_n(a)\rceil$ in (\ref{eq-comp2}) with small $\beta$ implies that
\begin{equation}\label{eq-liminf3}
\liminf_{n\ra\infty}\frac{T_{n,\text{\tiny TV}}(0,\epsilon)}{u_n(a)}>0\quad\text{for $\epsilon$ small enough}.
\end{equation}
This yields that $T_{n,\text{\tiny TV}}(0,\epsilon)\ra\infty$ for $\epsilon$ small enough.
To show (1)$\Lra$(4), let $(N_t)_{t\ge 0}$ be the Poisson process as before. It is easy to see from (\ref{eq-kernelcomp}) that if $(X_m^{(n)})_{m=0}^\infty$ is a realization of $(\mathcal{X}_n,K_n,\pi_n)$, then $(X_{N_t}^{(n)})_{t\ge 0}$ is a realization of $(\mathcal{X}_n,H_{n,t}^{(\delta)},\pi_n)$. This implies
\begin{equation}\label{eq-discts}
\begin{aligned}
\mathbb{P}_0\left(\widetilde{\tau}^{(n,\delta)}_i>s\right)&=\mathbb{P}_0\left(X_{N_r}^{(n)}<i,\,\forall
0\le r\le s\right)\\
&=\mathbb{P}_0(X_m^{(n)}<i,\,\forall m\le N_s)=\mathbb{P}_0\left(\tau^{(n)}_i>N_s\right).
\end{aligned}
\end{equation}
Since $u_n(a)\ra\infty$ for some $a\in(0,1)$, we obtain
\[
\mathcal{F}^L\text{ has a cutoff}\quad\Lra\quad\mathcal{F}_c^{(\delta,L)}\text{ has a cutoff}.
\]
By Theorem \ref{t-tv3}, the latter is equivalent to the existence of a sequence $t_n>0$ and a constant $a\in(0,1)$ satisfying
\begin{equation}\label{eq-2-4-1}
t_n=O\left(\mathbb{E}_0\widetilde{\tau}^{(n,\delta)}_{M_n(c)}\right),\quad\forall c\in(0,1)
\end{equation}
and
\begin{equation}\label{eq-2-4-2}
\lim_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n,\delta)}_{M_n(a)}>(1-\epsilon)t_n\right)=1,\quad\forall
\epsilon\in(0,1)
\end{equation}
and, for any $b\in(a,1)$, there is $\alpha_b\in(0,1)$ such that
\begin{equation}\label{eq-2-4-3}
\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n,\delta)}_{M_n(b)}>(1+\epsilon)t_n\right)\le\alpha_b,
\quad\epsilon\in(0,1).
\end{equation}
As a result of (\ref{eq-comp3}), one can see that (\ref{eq-2-4-1}) is equivalent to $t_n=O(u_n(a))$ and further, by (\ref{eq-discts}), (\ref{eq-2-4-2}) implies
\begin{align}
\liminf_{n\ra\infty}\mathbb{P}_0\left(\tau^{(n)}_{M_n(a)}>\frac{(1-\epsilon)t_n}{1-\delta}\right)
&\ge\liminf_{n\ra\infty}\mathbb{P}_0\left(\tau^{(n)}_{M_n(a)}>N_{(1-\epsilon/2)t_n}\right)\notag\\
&=\liminf_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n,\delta)}_{M_n(a)}>(1-\epsilon/2)t_n\right)=1.\notag
\end{align}
and (\ref{eq-2-4-3}) implies
\begin{align}
\limsup_{n\ra\infty}\mathbb{P}_0\left(\tau^{(n)}_{M_n(b)}>\frac{(1+\epsilon)t_n}{1-\delta}\right)&
\le\limsup_{n\ra\infty}\mathbb{P}_0\left(\tau^{(n)}_{M_n(b)}>N_{(1+\epsilon/2)t_n}\right)\notag\\
&=\limsup_{n\ra\infty}\mathbb{P}_0\left(\widetilde{\tau}^{(n,\delta)}_{M_n(b)}>(1+\epsilon/2)t_n\right)\le\alpha_b,
\notag
\end{align}
for all $\epsilon\in(0,1)$. This gives the desired properties in (4). Conversely, one may use a similar statement to prove (\ref{eq-2-4-2}) and (\ref{eq-2-4-3}) based on the observation of (4) and this part is omitted.
For a choice of the cutoff time, if (2) or (3) holds, the proof for the selected cutoff time is given by (\ref{eq-comp3}) and Theorem 3.4 in \cite{CSal13-1}. If (4) holds, the proof is exactly the same as that of Theorem \ref{t-tv3} and we skip it here.
\end{proof}
\begin{proof}[Proof of Corollary \ref{c-tv}]
The $(u_n(a),b_n)$ cutoff of $\mathcal{F}_c^L$ is immediately from (\ref{eq-mixing}) and Lemma \ref{l-hit2}. For the $(u_n(a),b_n)$ cutoff of $\mathcal{F}^L$, the assumption $\inf_nb_n>0$ and $b_n=o(u_n(a))$ implies that $u_n(a)\ra\infty$ for all $a\in(0,1)$, which means that the cutoff time tends to infinity. The remaining proof also uses Theorem 3.4 in \cite{CSal13-1} and is similar to the proof of the discrete time case in Theorem \ref{t-tv2}. We refer the reader to Section \ref{s-tv} for details.
\end{proof}
\section{Examples} \label{s-ex}
In this section, we consider some classical examples and use the developed theory to examine the existence of cutoff and, in particular, compute the cutoff time. First, we write $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$ for a family of irreducible birth and death chains with $\mathcal{X}_n=\{0,1,...,n\}$ and write $\mathcal{F}^L,\mathcal{F}^R$ for families of chains in $\mathcal{F}$ started at the left and right boundary states. For the continuous time case, those families are written as $\mathcal{F}_c,\mathcal{F}_c^L,\mathcal{F}_c^R$ instead. For $n\ge 1$, let $p_{n,i},q_{n,i},r_{n,i}$ be the birth, death and holding rates in $K_n$ and $\tau_i^{(n)},\widetilde{\tau}_i^{(n)}$ be the first hitting times to state $i$ of the $n$th chains in $\mathcal{F},\mathcal{F}_c$. For $a\in(0,1)$, $M_n(a)$ denotes a state in $\mathcal{X}_n$ satisfying $\pi_n([0,M_n(a)])\ge a$ and $\pi_n([M_n(a),n])\ge 1-a$.\vskip2mm
{\it (1) Biased random walk.} For $n\in\mathbb{N}$, let
\[
p_{n,i}=r_{n,n}=p,\quad q_{n,i+1}=r_{n,0}=q,\quad\forall 0\le i<n,\,n\ge 1,
\]
with $q=1-p\in(0,1/2)$. Note that the stationary distribution satisfies
\[
\pi_n(i)=\frac{p/q-1}{(p/q)^{n+1}-1}\left(\frac{p}{q}\right)^i,\quad\forall 0\le i\le n.
\]
This implies
\begin{equation}\label{eq-pq}
\frac{\pi_n([0,i])}{\pi_n(i)}=\frac{p/q-(p/q)^{-i}}{p/q-1},\quad\forall 0\le i\le n.
\end{equation}
By Lemma \ref{l-meanvar2}, one has
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{n}=\sum_{i=0}^{n-1}\frac{\pi_n([0,i])}{p\pi_n(i)},\quad
\zeta_{n,i}\le\text{Var}_i\widetilde{\tau}_{i+1}^{(n)}\le 2\zeta_{n,i},
\]
where
\[
\zeta_{n,i}=\frac{1}{p^2\pi_n(i)}\sum_{\ell=0}^i\left(\frac{\pi_n([0,\ell])}{\pi_n(\ell)}\right)^2
\pi_n(\ell).
\]
Applying (\ref{eq-pq}) to the computation of $\mathbb{E}_0\widetilde{\tau}^{(n)}_n$ and $\zeta_{n,i}$ yields
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{n}=\frac{n}{p-q}-\frac{p^2}{q(p-q)^2}
\left(1-\left(\frac{q}{p}\right)^n\right),\quad
\frac{1}{p^2}\le \zeta_{n,i}\le\frac{p}{(p-q)^3},\quad\forall i,
\]
where the bound of $\zeta_{n,i}$ leads to $\text{Var}_0\widetilde{\tau}^{(n)}_n\asymp n$.
Observe that $\pi_n([0,n])=1$ and $\pi_n(n)\ra 1-q/p$. As a consequence of Theorems \ref{t-sep2}, \ref{t-tv2} and \ref{t-tv5} with $M_n=n$, the families $\mathcal{F}_c,\mathcal{F}_c^L$ have a $(\frac{n}{p-q},\sqrt{n})$ cutoff in total variation and separation. To examine the existence of cutoff for $\mathcal{F}_c^R$, we fix $a\in(q/p^2-1,q/p)$. Based on the observation that $\pi_n(n-1)\ra(p-q)/p^2$, one has $M_n(a)=n-1$ for $n$ large enough and this implies $\text{Var}_n\widetilde{\tau}^{(n)}_{M_n(a)}=(\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n(a)})^2$. By Theorem \ref{t-tv3}, $\mathcal{F}_c^R$ has no cutoff in total variation.\vskip2mm
{\it (2) Metropolis chains for exponential distributions} Consider an increasing positive function $f$ on $(0,\infty)$. For $n\ge 1$, let $\pi_n(i)=\pi_n(0)f(i)$ and
\begin{equation}\label{eq-rates1}
p_{n,i}=r_{n,0}=1/2,\quad q_{n,i+1}=\frac{f(i)}{2f(i+1)},\quad\forall 0\le i<n,
\end{equation}
and
\begin{equation}\label{eq-rates2}
r_{n,i+1}=\frac{1}{2}-\frac{f(i)}{2f(i+1)},\quad\forall 0\le i<n-1,\quad r_{n,n}=1-\frac{f(n-1)}{f(n)}.
\end{equation}
One can check that the $n$th chain is the Metropolis chain for $\pi_n$ with base chain the simple random walk on $\mathcal{X}_n$ with holding probability $1/2$ at boundaries. We refer the reader to \cite{DS98} for details of Metropolis chains.
It is worthwhile to note that $K_n$ is monotonic, i.e. $p_{n,i}+q_{n,i+1}\le 1$ for all $0\le i<n$. By Corollary 4.2 in \cite{DLP10}, separation of the $n$th chain in $\mathcal{F},\mathcal{F}^L,\mathcal{F}^R$ (and respectively in $\mathcal{F}_c,\mathcal{F}_c^L,\mathcal{F}_c^R$) is the same. As a result of Theorem \ref{t-sep}, the existence of separation cutoff of $\mathcal{F}$ is equivalent to that of $\mathcal{F}_c$ and the cutoff time and window for $\mathcal{F}_c$ given by Theorem \ref{t-sep} is applicable to $\mathcal{F}$. For the total variation distance, if $\inf_{n,i}r_{n,i}>0$ is assumed, then Theorems \ref{t-tv2}, \ref{t-tv3} and \ref{t-tv4} and Remarks \ref{r-wcomp2} and \ref{r-wcomp} imply that the existence of cutoff of $\mathcal{F}$ (respectively $\mathcal{F}^L,\mathcal{F}^R$) is equivalent to that of $\mathcal{F}_c$ (respectively $\mathcal{F}_c^L,\mathcal{F}_c^R$). Furthermore, the cutoff times and windows for $\mathcal{F},\mathcal{F}_c$ given by Theorem \ref{t-tv2} (respectively for $\mathcal{F}^L,\mathcal{F}_c^L$ and for $\mathcal{F}^R,\mathcal{F}_c^R$ given by Theorems \ref{t-tv3} and \ref{t-tv4}) are consistent in the way that the cutoff times are equal and the cutoff windows are of the same order.
In this example, $f(x)=\exp\{\alpha x^\beta\}$ with $\alpha>0$ and $\beta>0$. Note that $\inf_{n,i}r_{n,i}>0$ if $\beta\ge 1$ and $\inf_{n,i}r_{n,i}=0$ if $\beta\in(0,1)$. In what follows, the cutoff phenomenon is discussed case by case according to $\beta$.
{\bf Case 1: $\beta>1$.} We first make some computations. Note that
\[
\frac{d}{dx}f(x)=\alpha\beta x^{\beta-1}f(x)\ge\alpha\beta f(x)\quad\forall x\ge 1.
\]
This implies
\[
\sum_{j=0}^if(j)\le 1+f(i)+\int_1^if(x)dx\le 1+f(i)+\frac{f(i)-f(1)}{\alpha\beta}\le\left(2+\frac{1}{\alpha\beta}\right)f(i).
\]
When $i$ tends to infinity, one has
\[
\left(1-\frac{1}{i}\right)^\beta=1-\frac{\beta}{i}+O\left(\frac{1}{i^2}\right).
\]
This leads to
\[
\frac{f(i-1)}{f(i)}=\exp\left\{-\alpha\beta i^{\beta-1}\left(1+O\left(\frac{1}{i}\right)\right)\right\}
=O\left(\frac{1}{i^2}\right).
\]
As a result, we obtain
\[
1\le \frac{\pi_n([0,i])}{\pi_n(i)}=1+\frac{\pi_n([0,i-1])}{\pi_n(i)}
\le 1+\left(2+\frac{1}{\alpha\beta}\right)\frac{f(i-1)}{f(i)}
=1+O\left(\frac{1}{i^2}\right).
\]
Replacing $i$ with $n$ gives $\pi_n(n)\ra 1$ and, by Lemma \ref{l-meanvar2}, one has
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_n=2\sum_{i=0}^{n-1}\frac{\pi_n([0,i])}{\pi_n(i)}
=2n+O(1)
\]
and
\[
\text{Var}_i\widetilde{\tau}^{(n)}_{i+1}\asymp\frac{1}{\pi_n(i)}
\sum_{\ell=0}^i\left(\frac{\pi_n([0,\ell])}{\pi_n(\ell)}\right)^2\pi_n(\ell)\asymp 1\quad\text{uniformly for }0\le i<n.
\]
The estimation of the variance implies $\text{Var}_0\widetilde{\tau}^{(n)}_n\asymp n$. By Theorem \ref{t-sep2}, \ref{t-tv2} and Theorem \ref{t-tv5}, both $\mathcal{F}_c$ and $\mathcal{F}_c^L$ have a $(2n,\sqrt{n})$ cutoff in total variation and separation. For the family $\mathcal{F}_c^R$, the observation, $\pi_n(n)\ra 1$, implies that the total variation mixing time of the $n$th chain is equal to $0$ when $n$ is large enough.
{\bf Case 2: $\beta=1$.} Set $\delta=(1-e^{-\alpha})/2$. Note that $(K_n-\delta I)/(1-\delta)$ is the biased random walk on $\mathcal{X}_n$ with $p=1/(1+e^{-\alpha})$. The result for biased random walks implies that $\mathcal{F}_c$ and $\mathcal{F}_c^L$ have a $(\frac{2n}{1-e^{-\alpha}},\sqrt{n})$ cutoff in total variation and separation but $\mathcal{F}_c^R$ has no total variation cutoff.
In Cases 1 and 2, one has $\inf_{n,i}r_{n,i}>0$. This implies that, in the total variation distance, the conclusion on the existence of cutoff, the cutoff time and the cutoff window also applies to $\mathcal{F}_c,\mathcal{F}_c^L,\mathcal{F}_c^R$.
{\bf Case 3: $0<\beta<1$.} First, observe that
\[
\frac{d}{dx}(x^{1-\beta}f(x))=\alpha\beta f(x)+(1-\beta)x^{-\beta}f(x).
\]
This implies
\[
\frac{i^{1-\beta}f(i)-j^{1-\beta}f(j)}{\alpha\beta+(1-\beta)j^{-\beta}}\le\int_j^if(x)dx\le\frac{i^{1-\beta}f(i)}{\alpha\beta},
\quad\forall 1\le j<i.
\]
and then
\begin{equation}\label{eq-alphabeta1}
\frac{1}{\alpha\beta+(1-\beta)j^{-\beta}}\left(1-\frac{j^{1-\beta} f(j)}{i^{1-\beta}f(i)}\right)\le\frac{f(1)+\cdots+f(i)}{i^{1-\beta}f(i)}
\le \frac{1}{\alpha\beta}+i^{\beta-1}
\end{equation}
When $i\ge 2j$ and $j\ra \infty$, one has
\[
\frac{f(j)}{f(i)}=\exp\left\{-\alpha i^\beta\left(1-\left(\frac{j}{i}\right)^\beta\right)\right\}=o\left(\frac{1}{i}\right).
\]
Consequently, we obtain, as $j\ra\infty$,
\begin{equation}\label{eq-alphabeta2}
f(0)+\cdots+f(i)=i^{1-\beta}f(i)\left(\frac{1}{\alpha\beta}+O(j^{-\beta}+i^{\beta-1})\right)\quad\text{uniformly for $i\ge 2j$}.
\end{equation}
Replacing $i,j$ with $n,\lfloor n/2\rfloor$ in (\ref{eq-alphabeta2}) gives
\[
\frac{1}{\pi_n(0)}=n^{1-\beta}f(n)\left(\frac{1}{\alpha\beta}+o(1)\right)\quad\text{as }n\ra\infty.
\]
Next, we fix $c>0$ and let $c_n$ be a sequence converging to $c$ such that $c_nn^{1-\beta}\in\mathcal{X}_n$. Set $M_n=n-c_nn^{1-\beta}$. Replacing $i,j$ with $M_n,\lfloor M_n/2\rfloor$ in (\ref{eq-alphabeta2}) yields
\[
\lim_{n\ra\infty}\pi_n([0,M_n])=\lim_{n\ra\infty}\pi_n(0)\sum_{\ell=0}^{M_n}f(\ell)=e^{-c\alpha\beta}\in(0,1).
\]
By Lemma \ref{l-meanvar2}, one has, when $2j_n\le i_n\le M_n$ and $j_n\ra\infty$,
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}=2\sum_{\ell=0}^{M_n-1}\frac{f(0)+\cdots+f(\ell)}{f(\ell)}
=\frac{2n^{2-\beta}}{\alpha\beta(2-\beta)}+O\left(n^{2-\beta} j_n^{-\beta}+i_n^{2-\beta}+n\right),
\]
where the second equality is given by separating $\sum_{\ell<M_n}$ into $\sum_{\ell<i_n}$ and $\sum_{i_n\le \ell<M_n}$ and then applying (\ref{eq-alphabeta1}) and (\ref{eq-alphabeta2}) respectively, and
\[
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n}\asymp\sum_{0\le\ell\le i<M_n}\frac{(f(0)+\cdots+f(\ell))^2}{f(i)f(\ell)}
\asymp\sum_{i=0}^{M_n-1}\frac{1}{f(i)}\sum_{\ell=0}^i\ell^{2-2\beta}f(\ell),
\]
where the computation uses (\ref{eq-alphabeta1}). Observe that $4-3\beta>2-\beta$. Setting $j_n=\lfloor n^{1/2}\rfloor$ and $i_n=\lfloor n^{\frac{4-3\beta}{4-2\beta}}\rfloor$. Clearly, $i_n\ge 2j_n$ for $n$ large enough and, in the computation of expectation, this leads to
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}=\frac{2n^{2-\beta}}{\alpha\beta(2-\beta)}+O\left(n^{2-\frac{3}{2}\beta}+n\right).
\]
Applying the following fact
\[
\frac{d}{dx}(x^{3-3\beta}f(x))=[(3-3\beta)x^{2-3\beta}+\alpha\beta x^{2-2\beta}]f(x)\asymp x^{2-2\beta}f(x),\quad\forall x\ge 1,
\]
to the computation of the variance yields
\[
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n}\asymp\sum_{i=0}^{M_n-1}i^{3-3\beta}\asymp n^{4-3\beta}.
\]
Similarly, one may use the observation that $\frac{f(M_n)}{f(n)}=\ra e^{-c\alpha\beta}$ to derive
\[
q_{n,i}\asymp 1\quad\text{uniformly for $M_n\le i\le n$}.
\]
By Lemma \ref{l-meanvar2}, this implies
\[
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}\asymp\sum_{i=M_n+1}^n\frac{f(i)+\cdots+f(n)}{f(i)}\asymp n^{2-2\beta}
\]
and
\[
\text{Var}_n\widetilde{\tau}^{(n)}_{M_n}\asymp\sum_{M_n<i\le \ell\le n}\frac{(f(\ell)+\cdots+f(n))^2}{f(i)f(\ell)}\asymp n^{4-4\beta}.
\]
As a consequence of Theorem \ref{t-sep2}, \ref{t-tv2} and \ref{t-tv5}, $\mathcal{F}_c$ and $\mathcal{F}_c^L$ have a $(\frac{2n^{2-\beta}}{\alpha\beta(2-\beta)},n^{2-\frac{3}{2}\beta}+n)$ cutoff in total variation and separation but, by Theorem \ref{t-tv3}, $\mathcal{F}_c^R$ has no total variation cutoff. Note that, when $\beta\in(2/3,1)$, a better choice of the cutoff window is $n^{2-\frac{3}{2}\beta}$. To have this cutoff window, a more subtle estimation of the cutoff time is required.
We summarize the above results in the following theorem.
\begin{thm}\label{t-metro}
Let $f(x)=\exp\{\alpha x^\beta\}$ with $\alpha>0,\beta>0$. Consider the family $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$, where $\mathcal{X}_n=\{0,1,...,n\}$, $\pi_n(i)=\pi(0)f(i)$ and $K_n$ is a birth and death chain with transition rates
\[
p_{n,i}=r_{n,0}=1/2,\quad q_{n,i+1}=\frac{f(i)}{2f(i+1)},\quad r_{n,i+1}=\frac{1}{2}-\frac{f(i)}{2f(i+1)},\quad\forall 0\le i<n.
\]
Then, $\mathcal{F}_c$ and $\mathcal{F}_c^L$ have a $(t_n,b_n)$ cutoff in total variation and separation but $\mathcal{F}_c^R$ has no total variation cutoff, where
\[
t_n=\begin{cases}2n&\text{for }\beta>1\\\frac{2n}{1-e^{-\alpha}}&\text{for }\beta=1\\\frac{2n^{2-\beta}}{\alpha\beta(2-\beta)}&\text{for }0<\beta<1\end{cases},\quad
b_n=\begin{cases}\sqrt{n}&\text{for }\beta\ge 1\\n^{2-\frac{3}{2}\beta}+n&\text{for }0<\beta<1\end{cases}.
\]
\end{thm}
\vskip2mm
{\it (3) Metropolis chains for polynomial distributions} In this example, we consider the family of Metropolis chains given by (\ref{eq-rates1}) and (\ref{eq-rates2}) with the replacement of $f(x)$ by $g(x)=\exp\{\alpha(\log(x+1))^\beta\}$, where $\alpha,\beta$ are positive. It has been shown in \cite{CSal13-2} that $\mathcal{F}_c$ has a cutoff in total variation and separation when $\beta>1$ but has no cutoff when $0<\beta\le 1$. The following theorem provides a cutoff time and a cutoff window when $\beta>1$.
\begin{thm}\label{t-metro2}
Let $g(x)=\exp\{\alpha(\log(x+1))^\beta\}$ with $\alpha>0$ and $\beta>1$. Consider the family $\mathcal{F}=(\mathcal{X}_n,K_n,\pi_n)_{n=1}^\infty$, where $\mathcal{X}_n=\{0,1,...,n\}$, $\pi_n(i)=\pi(0)g(i)$ and $K_n$ is a birth and death chain with transition rates
\[
p_{n,i}=r_{n,0}=1/2,\quad q_{n,i+1}=\frac{g(i)}{2g(i+1)},\quad r_{n,i+1}=\frac{1}{2}-\frac{g(i)}{2g(i+1)},\quad\forall 0\le i<n.
\]
Then, $\mathcal{F}_c$ and $\mathcal{F}_c^L$ have a $(t_n,b_n)$ cutoff in total variation and separation but $\mathcal{F}_c^R$ has no total variation cutoff, where
\[
t_n=\sum_{\ell=0}^N\frac{n^2}{\alpha\beta B_\ell(\log n)^{\beta+\ell-1}},\quad
b_n=\frac{n^2}{(\log n)^{\frac{3}{2}(\beta-1)}}
\]
and $B_0=1$, $B_\ell=2^\ell(\beta-1)\beta\cdots(\beta+\ell-2)$, $N=\lceil\frac{\beta-3}{2}\rceil\ge 0$.
\end{thm}
\begin{rem}
Note that, in Theorem \ref{t-metro2}, $\beta+N-1<\frac{3}{2}(\beta-1)\le\beta+N$.
\end{rem}
The proof of Theorem \ref{t-metro2} is similar to the proof of the case $\beta\in(0,1)$ in Theorem \ref{t-metro} and is placed in the appendix.\vskip2mm
{\it (4) Metropolis chains for binomial distributions} For $n\ge 1$, let $\pi_n(i)=2^{-n}\binom{n}{i}$ and
\[
p_{n,i}=q_{n,n-i}=\frac{1}{2},\quad q_{n,i+1}=p_{n,n-i-1}=\frac{i+1}{2(n-i)},\quad\forall 0\le i<n/2,
\]
and $r_{n,i}=1-p_{n,i}-q_{n,i}$ for $0\le i\le n$. It is easy to check that $K_n$ is the Metropolis chain for $\pi_n$ with base chain the simple random walk on $\mathcal{X}_n$ with holding probability $1/2$ at the boundary states. The separation cutoff of this family is proved in \cite{DS06} and we will discuss the cutoff time and the cutoff window in this example. First, one may use Lemma \ref{l-meanvar2} and (\ref{eq-ehren}) to derive
\begin{equation}\label{eq-ehrenmeanvar}
\sum_{i=0}^{M_n-1}\frac{\pi_n([0,i])}{\pi_n(i)(1-\frac{i}{n})}=\frac{n\log n }{4}+O(n),\quad
\sum_{0\le \ell\le i<M_n}\frac{\pi_n([0,\ell])^2}{\pi_n(\ell)\pi_n(i)}\asymp n^2,
\end{equation}
for any sequence $M_n\in\mathcal{X}_n$ satisfying $|M_n-\frac{n}{2}|=O(\sqrt{n})$. Note that
\[
\pi_n(i)\left(1-\frac{i}{n}\right)=\frac{\pi_{n-1}(i)}{2},\quad \pi_n([0,i])=\pi_{n-1}([0,i])-\frac{\pi_{n-1}(i)}{2}.
\]
This implies
\begin{equation}\label{eq-pin3}
\frac{\pi_n([0,i])}{\pi_n(i)(1-\frac{i}{n})}=\frac{2\pi_{n-1}([0,i])}{\pi_{n-1}(i)}-1.
\end{equation}
Set $M_n=\lfloor n/2\rfloor$. By Lemma \ref{l-meanvar2}, (\ref{eq-ehrenmeanvar}) and (\ref{eq-pin3}), we obtain
\[
\mathbb{E}_0\widetilde{\tau}^{(n)}_{M_n}=\sum_{i=0}^{M_n-1}\frac{2\pi_n([0,i])}{\pi_n(i)}=
\sum_{i=0}^{M_n-1}\left(\frac{\pi_{n+1}([0,i])}{\pi_{n+1}(i)(1-\frac{i}{n+1})}+1\right)=\frac{n\log n}{4}+O(n)
\]
and
\[
\text{Var}_0\widetilde{\tau}^{(n)}_{M_n}\asymp \sum_{0\le \ell\le i<M_n}\frac{\pi_n([0,\ell])^2}{\pi_n(\ell)\pi_n(i)}\asymp n^2.
\]
In a similar way, one has
\[
\mathbb{E}_n\widetilde{\tau}^{(n)}_{M_n}=\frac{n\log n}{4}+O(n),\quad \text{Var}_n\widetilde{\tau}^{(n)}_{M_n}\asymp n^2.
\]
As a consequence of Theorems \ref{t-sep2}, \ref{t-tv2} and \ref{t-tv5}, $\mathcal{F}_c$ has a $(\frac{1}{2}n\log n,n)$ separation cutoff and $\mathcal{F}_c,\mathcal{F}_c^L$ have a $(\frac{1}{4}n\log n,n)$ total variation cutoff.
|
\section{Introduction} \label{sec:Intro}
\vspace{-2mm}
\begin{figure}[ptb]
\vspace{-1mm}
\begin{center}
\includegraphics[width=3.2in]{Framework.pdf}
\end{center}
\vspace{-4mm}
\caption{An illustration of the proposed approach. A query individual is represented as a compositional part-based template, and part proposals are extracted from multiple instances at each parts. Human re-identification is thus posed as compositional template matching. Note certain parts are omitted for clear specification.}
\label{fig:framework}
\vspace{-4mm}
\end{figure}
Person re-identification at a distance increasingly receives attention in video surveillance, particularly for the applications restricting the use of face recognition. But this task is very challenging due to the following difficulties,
{\small \textbullet}\; Robust human representation (signature). There are large variations for human body in appearance, (e.g., different views, poses, lighting conditions). It is usually intractable to construct a template of the individual to be recognized by extracting only low-level image features.
{\small \textbullet}\; Effective human matching (localizing). Given the template, re-identifying targets with the global body information often suffers from high matching false positives, as the targets are possibly occluded or conjuncted with others and backgrounds in realistic surveillance applications. Furthermore, it is desired to accurately localize human body parts in general.
The objective of human re-identification in this work is to recognize an individual by employing body information to address the above difficulties. We study the problem with the following setting based on the application requirements in surveillance: (1) The clothing of individuals remain unchanged across different scenarios. (2) The individual to be re-identified should be in a moderate resolution, (e.g., $\ge\,120$ pixels in height). Our approach builds a compositional part-based template to represent the target individual and matches the template with input images by employing a stochastic cluster sampling algorithm, as illustrated in Fig.~\ref{fig:framework}.
We organize the template of a query individual with an expressive tree representation that can be produced in a very simple way. We perform the human body part detectors~\cite{PictorRevisit,ViewPictor} on several reference images of the individual, and the images of detected parts are grouped according to their semantics. That is, a human template is decomposed into body parts, e.g., head, torso, arms, each of which associates with a number of part instances. Note that we can prune the instances sharing very similar appearances with others. This expressive template fully exploit information from multiple reference images to capture well appearance variability, partially motivated by the recently proposed hierarchical and part-based models in object recognition~\cite{AOHumParse,AOTObjRec,GeneralSample}. Specifically, several possible instances (namely proposals), extracted from different references, exist at each part in the template, and we regard this representation as the multiple-instance-based compositional template (MICT). As a result, new appearance configurations can be composed by the part proposals in the MICT. One may question the scalability issue for building such a customized template. We argue that the critical concern is accurately identifying the target in realistic scenarios, e.g., searching for one suspect across scenes, rather than processing numbers of targets at the same time.
In the inference stage, the body part detectors are initially utilized to generate possible part locations in the scene shot, and human re-identification is then posed as the task of part-based template matching. Unlike traditional matching problems, the multiple part proposals in the MICT make the search space of matching combinatorially large, as the part proposals need to be activated alone with the matching process. Handling the false alarms and misdetections by the part detectors is also a non-trivial issue during matching. Inspired by recent studies in cluster sampling~\cite{SWCuts,LayerGM,C4}, we propose a stochastic algorithm to solve the compositional template matching.
The matching algorithm is designed based upon the candidacy graph, where each vertex denotes a pair of matching part proposals, and each edge link represents the contextual interaction (i.e. the compatible or the competitive relation) between two matching pairs. Compatible relations encourage vertices to activate together, while competitive relations depress conflicting vertices being activated at the same time. Specifically, two vertices are encouraged to be activated together, as they are kinematically or symmetrically related, whereas two vertices are constrained that only one of them can be activated, as they belong to the same part type or overlap. The algorithm iterates in two steps for optimal matching solution searching. (i) It forms several possible partial matches (clusters) by turning off the edge links probabilistically and deterministically. (ii) It activates clusters to confirm partial matches, leading to a new matching solution that will be accepted by the Markov Chain Monte Carlo (MCMC) mechanism~\cite{SWCuts}. Note that body parts are allowed to be unmatched to cope with occlusions.
The main contributions of this paper are two-fold. First, we propose a novel formulation to solve human re-identification by matching the composite template with cluster sampling. Second, we present a new database including realistic and general challenges for human re-identification, which is more complete than existing related databases.
\vspace{-2mm}
\section{Related Work} \label{sec:Literature}
\vspace{-1mm}
In literature, previous works of human re-identification mainly focus on constructing and selecting distinctive and stable human representation, and they can be roughly divided into the following two categories.
{\bf Global-based methods} define a global appearance human signature with rich image features and match given reference images with the observations~\cite{ELF,ETHZReid,SDALF}. For example, D. Gray et al. propose the feature ensemble to deal with viewpoint invariant recognition. Some methods improve the performance by extracting features with region segmentation~\cite{PrinAxis,SACM,DCDSign}. Recently, advanced learning techniques are employed for more reliable matching metrics~\cite{ReidRDCConfer}, more representative features~\cite{FeatureReid}, and more expressive multi-valued mapping function~\cite{ImplicitTransReid}. Despite acknowledged success, this category of methods often has problems to handle large pose/view variance and occlusions.
{\bf Compositional approaches} re-identify people by using part-based measures. They first localize salient body parts, and then search for part-to-part correspondence between reference samples and observations. These methods show promising results on very challenging scenarios~\cite{iLIDS}, benefiting from powerful part-based object detectors. For example, N. Gheissari et al.~\cite{STAppear} adopt a decomposable triangulated graph to represent person configuration, and the pictorial structures model for human re-identification is introduced~\cite{CPS}. Besides, modeling contextual correlation between body parts is discussed in~\cite{SCR}.
Many works~\cite{STAppear,SDALF,CPS} utilize multiple reference instances for individual, i.e. multi-shot approaches, but they omit occlusions and conjunctions in the target images and re-identify the target by computing a one-to-many distance, while we explicitly handle these problems by exploiting reconfigurable compositions and contextual interactions during inference.
The rest of this paper is organized as follows. We first introduce the representations in Section~\ref{sec:Represent}, and then discuss the inference algorithm in Section~\ref{sec:Algorithm}. The experimental results are shown in Section~\ref{sec:Experiments}, and finally comes the conclusion in Section~\ref{sec:Conclusion}.
\vspace{-2mm}
\section{Representation} \label{sec:Represent}
\vspace{-1mm}
In this section, we first introduce the definition of multiple-instance-based compositional template, and then present the problem formulation of human re-identification.
\vspace{-1mm}
\subsection{Compositional Template} \label{sec:Template}
\vspace{-1mm}
In this work, we present a compositional template to model human with huge variations.
A human body is decomposed into $N=6$ parts: head, torso, upper arms, forearms, thighs and calfs, and each limb is further decomposed into two symmetrical parts (i.e. left and right), as shown in Fig.~\ref{fig:relations}(a)). Each part $g$ is modeled as a rectangle and indicated by a $5$-tuple $(t,x,y,\theta,s)$, where $t$ denotes the part type, $x$ and $y$ the part center coordinates, $\theta$ the part orientation, $s$ the part relative scale, as widely employed in pictorial structures model~\cite{PictorOR,PictorRevisit}. The multiple-instance-based compositional template (MICT) $T$ is defined as
\vspace{-1mm} \begin{equation}
T = \{\,T_i :\; T_i = \{g\}\;\}_{i=1}^N,
\vspace{-1mm} \end{equation}
where $g$ denotes a part proposal and $T_i$ the set of proposals for the $i$th part in template.
Given reference images of an individual, the MICT is constructed as follows.
We first employ body part detectors to scan every reference image and obtain detection scores for all body parts. The training and detecting process of part detectors closely follows~\cite{ViewPictor}. Given detection scores, we further prune impossible part configurations by several strategies: (i) For all parts, the firing detection is pruned if the overlap rate of foreground mask (done by background subtraction) is less than $75\%$. (ii) The reference image is segmented into $4$ horizonal strips with equal height. Head is detected in the first strip (the first to fourth top to bottom), parts of upper body (i.e. torso, upper arms and forearms) in the second, and parts of lower body (i.e. thighs and calfs) the rest. Finally, we apply non-maximum suppression and collect the $K$ proposals with highest responses for each part from all reference images.
Given target images (scene shots) to be matched, we can obtain the target proposal set $G$ by a similar process as constructing the MICT, except the firing detection being pruned only by the foreground mask. Considering realistic complexities in surveillance, there probably exist large numbers of detection false alarms in the target proposal set $G$.
\vspace{-1mm}
\subsection{Candidacy Graph} \label{sec:CandiGraph}
\vspace{-1mm}
Given the template $T$ and the target proposal set $G$, the problem of human re-identification can be posed as the task of part-based template matching and solved by two steps: (i) activating one proposal for each part in $T$, (ii) finding the match in $G$.
We define the set of activated part proposals $\Psi$ from $T$, each of which corresponds to a certain part:
\vspace{-1mm} \begin{equation}
\Psi \,=\, \{\,\Psi_i:\; l(\Psi_i) = 1,\;\; \Psi_i \in T_i\,\}_{i=1}^{N}.
\vspace{-1mm} \end{equation}
The binary label $l(\cdot)$ indicates whether the proposal is activated or remains inactivated, i.e. $l(\cdot) = 1$ for activated and $l(\cdot) = 0$ for inactivated. The set of matched part proposals from $G$ can be defined as
\vspace{-1mm} \begin{equation}
\Phi \,=\, \{\,\Phi_i:\; l(\Phi_i) = 1,\;\; \Phi_i \in G_i \cup \{\emptyset\}\,\}_{i=1}^N,
\vspace{-1mm} \end{equation}
where $\Phi_i$ maps the activated proposal of the $i$th part in $T$ to a proposal in $G$. Note that $\Psi_i$ not necessarily has a match $\Phi_i$ (i.e. $l(\emptyset)=1$), in case the matched part is occluded or missed in $G$.
To solve these two steps simultaneously, we propose a candidacy graph representation and further formulate the problem by graph labeling. We define the candidacy graph $\mathbb{G} = {<} \mathbb{C},\mathbb{E} {>}$, where each vertex $c_i \in \mathbb{C}$ denotes a candidate matching pair $(\Psi_i,\Phi_i)$. A similar binary label $l(c_i)$ is employed to indicate whether a matching pair $c_i$ is activated or not. Solving the matching problem is equivalent to labeling vertices $\mathbb{C}$ in the candidacy graph $\mathbb{G}$. The label set $L$ is thus defined as
\vspace{-1mm} \begin{equation}
L \,=\,\{l(c_i) = l_i: l_i \in \{0,1\}, i=1,...,|\mathbb{C}|, c_i \in \mathbb{C}\}.
\vspace{-1mm} \end{equation}
Each edge $e_{ij} = {<} c_i,c_j {>}$ in $\mathbb{G}$ denotes the relation between two matching pairs $c_i$ and $c_j$. We incorporate two kinds of relations, i.e. compatible and competitive relations, to model the contextual interactions in scene shots. In the following discussion, we drop the notation of edge index $ij$ for notation simplicity.
\begin{figure}[ptb]
\begin{center}
\includegraphics[width=3.3in]{Relations.pdf}
\end{center}
\vspace{-4mm}
\caption{An illustration of compatible relations. (a) Kinematics (navy blue edges) and symmetry (brown edges) relations within the compositional template. (b) An example to show how target part proposals are coupled together by kinematics and symmetry relations.}
\label{fig:relations}
\vspace{-3mm}
\end{figure}
\begin{figure*}[ptb]
\begin{center}
\includegraphics[width=6in]{Example.pdf}
\end{center}
\vspace{-5mm}
\caption{A re-identification example used to illustrate our inference algorithm. Given (a) reference images and (b) a scene shot, proposals of four parts: head, torso, left thigh, left calf are drawn and numbered in the image. Note that we omit the other parts and only keep few proposals for clear specification.}
\label{fig:example}
\vspace{-4mm}
\end{figure*}
{\bf Compatible relations} encourage matching pairs to activate together in matching. We represent compatible relations as how two target part proposals are coupled together and mainly explore two cases: (i) kinematics relations for coupling kinematic dependent parts. (ii) symmetry relations for coupling symmetrical parts. That is,
\vspace{-1mm} \begin{equation} \label{equ:posedge}
p_e^+(c_i,c_j) \,=\, \begin{cases}
& p_k(\Phi_i,\Phi_j),\;\quad {<} t_i,t_j {>} \in \text{Kin},\\
& p_s(\Phi_i,\Phi_j),\;\quad {<} t_i,t_j {>} \in \text{Sym},
\end{cases}
\vspace{-1mm} \end{equation}
where $t_i$ and $t_j$ denotes the part type of $\Phi_i$ and $\Phi_j$, respectively.
(i) Kinematics relations describe spatial relationship between kinematic dependent parts (navy blue edges in Fig.~\ref{fig:relations}(a)). The spatial distribution $p_k(\Phi_i,\Phi_j)$ between two proposals $\Phi_i$ and $\Phi_j$ is modeled as a zero-mean Gaussian distribution under the coordinate system of their connected joint:
\vspace{-2mm} \begin{equation}
p_k(\Phi_i,\Phi_j) \,\varpropto\, N(F_{ji}(\Phi_i)-F_{ij}(\Phi_j),0,\Sigma_{ji}),
\vspace{-1mm} \end{equation}
where $F_{ji}(\cdot)$ and $F_{ij}(\cdot)$ are the transformations of $\Phi_i$ and $\Phi_j$ from image coordinate system to joint coordinate system. For detailed explanations, see~\cite{PictorOR,PictorRevisit}.
In the experiment, kinematics relations are learnt from reference images with body part annotations.
(ii) Symmetry relations measure the appearance similarity between symmetrical parts (brown edges in Fig.~\ref{fig:relations}(a)). We suppose symmetrical parts from the same individual tend to share similar appearance while those from different individuals don't. Therefore the symmetry relations are represented as
\vspace{-2mm} \begin{equation}
p_s(\Phi_i,\Phi_j) \,\varpropto\, e^{-D(\Phi_i,\Phi_j)},
\vspace{-1mm} \end{equation}
where $D(\cdot)$ measures the distance between two part proposals and is defined in Equ.(\ref{equ:distance}).
We give an example to illustrate how kinematics relations and symmetry relations work in scene shots, as shown in Fig.~\ref{fig:relations}(b). Note that we omit certain part proposals for clear specification.
{\bf Competitive relations} depress conflicting matching pairs being activated at the same time. We also develop two cases for competitive relations: (i) Two target proposals with the same part type cannot be activated simultaneously. (ii) The overlapped region between two target part proposals should only be compared once. That is,
\vspace{-1mm} \begin{equation} \label{equ:negedge}
p_e^-(c_i,c_j) \varpropto \begin{cases}
& \quad\quad\, 1, \quad\quad\quad t_i=t_j,\\
& e^{\lambda \cdot \mathbf{IoU}(\Phi_i,\Phi_j) },\; t_i\ne t_j, \Phi_i \cap \Phi_j \ne \emptyset,
\end{cases}
\vspace{-1mm} \end{equation}
where $\mathbf{IoU}(\Phi_i,\Phi_j)$ indicates the overlap intersection-over-union between $\Phi_i$ and $\Phi_j$, $\lambda$ is a scaling constant.
An illustration of the candidacy graph representation is shown in Fig.~\ref{fig:candigraph}, corresponding to the example in Fig.~\ref{fig:example}.
\begin{figure}[ptb]
\vspace{-3mm}
\begin{center}
\includegraphics[width=3.3in]{CandidacyGraph.pdf}
\end{center}
\vspace{-4mm}
\caption{An illustration of the candidacy graph representation. We visualize the candidacy graph of Fig.~\ref{fig:example}. In the graph, vertices denote candidate matches, blue and red edges indicate compatible and competitive edges between vertices, respectively.}
\vspace{-3mm}
\label{fig:candigraph}
\end{figure}
\begin{figure*}[ptb]
\begin{center}
\includegraphics[width=6.4in]{ClusterSampling.pdf}
\end{center}
\vspace{-4mm}
\caption{An illustration of one transition in composite cluster sampling. The first row and the second row denote labels of part proposals, labels of the composite cluster and matching configurations of two successive states (A and B) in one reversible transition, respectively.}
\label{fig:clustersample}
\vspace{-4mm}
\end{figure*}
In summary, the problem of matching the template $T$ to the target proposal set $G$ can be represented as
\vspace{-1mm} \begin{equation}\label{equ:model}
M \,=\,(N_u,N_s,L),
\vspace{-1mm} \end{equation}
where $N_u$ denotes the number of unmatched part pairs and $N_s$ the number of scales of the activated proposals, and they can be computed from the labeling set $L$. According to Bayes' Rule, $M$ can be solved by maximizing a posterior probability:
\vspace{-1mm} \begin{equation} \label{equ:posterior} \begin{aligned}
M^* &\,=\,\underset{M}{\arg\,\max}\;p(M|T,G)\\
&\,=\,\underset{M}{\arg\,\max}\;p(M|\mathbb{G})\\
&\,\varpropto\,\underset{M}{\arg\,\max}\;p(\mathbb{G}|M) \cdot p(M).
\end{aligned} \vspace{-1mm} \end{equation}
{\bf Likelihood} $p(\mathbb{G}|M)$ measures the appearance similarity between the template and the matching target. Assuming the appearance similarity of each matching pair is independent, then $p(\mathbb{G}|M)$ can be factorized into
\vspace{-1mm} \begin{equation}
p(\mathbb{G}|M) \,\varpropto\, \prod_{c_i \in \mathbb{C}}\, p(c_i|l_i) \,=\, \prod_{ l_i = 1, c_i \in \mathbb{C}}\, e^{-D(\Psi_i,\Phi_i)}\,,
\vspace{-1mm} \end{equation}
where $D(\cdot)$ denotes the distance between two proposals.
We adopt modified HSV color histogram~\cite{CPS} and MSCR descriptor~\cite{MSCR} to describe the visual statistics for each part proposal, which has been widely used in existing human re-identification studies~\cite{SDALF,CPS}. The distance $D(\cdot)$ between two arbitrary proposals $g_i$ and $g_j$ is defined as
\vspace{-1mm} \begin{equation} \label{equ:distance}
D(g_i,g_j) = D_{Bh}(H_i,H_j) + D_{\text{MSCR}}(Dc_i,Dc_j),
\vspace{-1mm} \end{equation}
where $H$ denotes the normalized HSV color histogram, $Dc$ the MSCR descriptor, $D_{Bh}(\cdot)$ and $D_{\text{MSCR}}(\cdot)$ the Bhattacharyya distance and the distance defined in~\cite{SDALF}, respectively.
{\bf Prior} $p(M)$ penalizes the undesired activation of matching pairs (e.g. missing parts) and matching inconsistency among the activated matching pairs. We define $p(M)$ as
\vspace{-1mm} \begin{equation} \begin{aligned}
p(M) \,&=\, p(N_u) \cdot p(N_s) \cdot p(L) \\
\,&\varpropto\, \exp\{-\alpha_u N_u-\alpha_s N_s\}\cdot p(L),
\end{aligned} \vspace{-1mm} \end{equation}
where $\alpha_u$ and $\alpha_s$ are corresponding parameters for $N_u$ and $N_s$, respectively.
$p(L)$ imposes constraints on the edge links among activated vertices, that is
\vspace{-1mm} \begin{equation}
p(L) \varpropto \underset{l_i=l_j=1, e \in \mathbb{E}^+}{\prod} p_e^+(c_i,c_j) \underset{l_i=l_j=1, e \in \mathbb{E}^-}{\prod} (1 {-} p_e^-(c_i,c_j)),
\vspace{-1mm} \end{equation}
where $\mathbb{E}^+$ and $\mathbb{E}^-$ indicate the compatible edges and competitive edges in the candidacy graph $\mathbb{G}$, respectively.
\vspace{-2mm}
\section{Inference Algorithm} \label{sec:Algorithm}
\vspace{-1mm}
In a scene shot containing multiple individuals, matching the template to the target becomes an extremely complicated problem. For example, in Fig.~\ref{fig:example}, the four individuals in the shot all share similar appearance with the template. As a result, solving Equ.(\ref{equ:posterior}) probably leads to a local optimal solution. In this case, popular inference algorithms, such as EM, Belief Propagation and Dynamic Programming, are easily struck and thus fail to re-identify the correct target (i.e. finding global optimal solution), while Composite Cluster Sampling, as introduced in~\cite{LayerGM,C4}, overcomes this problem by jumping from partial coupling matches in each MCMC step. Therefore, we employ Composite Cluster Sampling to search for optimal match between the template and the correct target.
Composite Cluster Sampling algorithm consists of the following two steps:
(I) {\bf Generating a composite cluster}. Given a candidacy graph $\mathbb{G} = {<} \mathbb{C},\mathbb{E} {>}$ and the current matching state $M$, we first separate graph edges $\mathbb{E}$ into two sets: set of inconsistent edges $\{e \in \mathbb{E}^+: l(c_i) \ne l(c_j)\} \cup \{e \in \mathbb{E}^-: l(c_i) = l(c_j)\}$ (i.e. edges violating current state) and set of consistent edges in the other two cases. Next we introduce a boolean variable $\omega_e \in \{1,0\}$ to indicate an edge is being turned on or turned off. We turn off inconsistent edges deterministically and turn on every consistent edge with its edge probability $p_e$. Afterwards, we regard candidates connected by "on" positive edges as a cluster $Cl$ and collect clusters connected by "on" negative edges to generate a composite cluster $V_{cc}$.
(II) {\bf Relabeling the composite cluster}. In this step, we randomly choose a cluster from the obtained composite cluster $V_{cc}$ and flip the labels of the selected cluster and its conflicting clusters (i.e. the clusters connected with the selected cluster), which generates a new state $M'$. To find a better state and achieve a reversible transition between two states $M$ and $M'$, the acceptance rate of the transition from state $M$ to state $M'$ is defined by a Metropolis-Hastings method~\cite{MHAlgorithm}:
\vspace{-2mm} \begin{equation}
\alpha(M {\to} M') = \min(1,\frac{q(M' {\to} M) \cdot p(M'|\mathbb{G})}{q(M {\to} M') \cdot p(M|\mathbb{G})}),
\vspace{-1mm} \end{equation}
where $q(M' {\to} M)$ and $q(M {\to} M')$ denote the state transition probability, $p(M'|\mathbb{G})$ and $p(M|\mathbb{G})$ the posterior defined in Equ.(\ref{equ:posterior}).
Following instructions in~\cite{SWCuts}, the state transition probability ratio is computed by
\vspace{-1mm} \begin{equation} \label{equ:transrate}\begin{aligned}
\frac{q(M' {\to} M)}{q(M {\to} M')} &\varpropto \frac{q(V_{cc}|M')}{q(V_{cc}|M)} \\
&\varpropto \frac{\prod_{e \in \mathcal{E}^+_{M'}}(1-\rho_e^+)\prod_{e \in \mathcal{E}^-_{M'}}(1-\rho_e^-)} {\prod_{e \in \mathcal{E}^+_M}(1-\rho_e^+)\prod_{e \in \mathcal{E}^-_M}(1-\rho_e^-)},
\end{aligned} \vspace{-1mm} \end{equation}
where $\mathcal{E}^+$ and $\mathcal{E}^-$ denote the sets of positive and negative edges being turned off around $V_{cc}$, respectively, that is,
\begin{equation} \begin{aligned}
& \mathcal{E}^+ = \{e \in\mathbb{E}^+:c_i\in V_{cc}, c_j \not\in V_{cc}, l(c_i)=l(c_j)\}, \\
& \mathcal{E}^- = \{e \in\mathbb{E}^-:c_i\in V_{cc}, c_j\not\in V_{cc}, l(c_i)\ne l(c_j)\}.
\end{aligned} \end{equation}
Note that the subscript of $\mathcal{E}^+$, $\mathcal{E}^-$ in Equ.(\ref{equ:transrate}) indicates the current state and is omitted for simplicity in the above definition.
We show an example of one transition in composite cluster sampling in Fig.~\ref{fig:clustersample}. In this figure, $V_{cc}$ contains two clusters $\{Cl_1,Cl_2\}$. In state A, $Cl_1$ is activated and the conflicting cluster $Cl_2$ is deactivated while in state B labels of $Cl_1$ and $Cl_2$ are flipped. The transition from state $A$ to state $B$ achieves a fast jump between two kinds of partial coupling matches and coincides with an individual-to-individual comparison in re-identification.
Applying the above mechanism, we summarize the inference algorithm in Algorithm 1.
\begin{figure}[htb]
\vspace{-2mm}
\begin{center}
\includegraphics[width=3.2in]{Algorithm.pdf}
\end{center}
\label{fig:algorithm}
\vspace{-6mm}
\end{figure}
\vspace{-2mm}
\section{Experiments} \label{sec:Experiments}
\vspace{-1mm}
In this section, we first introduce the datasets and the parameter settings, and then show our experimental results as well as component analysis of the proposed approach.
\vspace{-1mm}
\subsection{Datasets and Settings} \label{sec:Dataset}
\vspace{-1mm}
We validate our method on three public databases as follows.
(i) {\bf VIPeR dataset}\footnote{Available at \url{www.umiacs.umd.edu/~schwartz/datasets.html}}. It is commonly used for human re-identification, containing $632$ people in outdoor, and there are $2$ images for each individual.
(ii) {\bf EPFL dataset}\footnote{Available at \url{cvlab.epfl.ch/data/pom/}}. This database is very challenging, originally proposed for tracking in multi-views~\cite{EPFL}. It consists of $5$ different scenarios that are filmed by three or four cameras from different angles. For evaluating our method, we extract individuals from the original videos and annotate each of them with ID and location (bounding box). In total, there are $70$ reference images for $30$ different individuals, (normalized to $175$ pixels in height), and $80$ shots in $360 \times 288$, which contain $294$ targets to be re-identified.
(iii) {\bf CAMPUS-Human dataset}\footnote{Available at \url{http://vision.sysu.edu.cn/projects/human-reid/}}. We construct this database including general and realistic challenges for people re-identification in surveillance. There are $370$ reference images normalized to $175$ pixels in height, for $74$ individuals, with IDs and locations provided. We present $214$ shots containing $1519$ targets for evaluating methods, and the targets often appear with diverse poses/views, conjunctions and occlusions, see Fig.~\ref{fig:resulttable} (bottom row). Note all images in both EPFL dataset and CAMPUS-Human dataset are captured from the original videos with large time gap to guarantee appearance varieties (unlike ETHZ dataset~\cite{ETHZReid}).
{\em Experiment settings}. For VIPeR dataset, we adopt the common setting that running the algorithm on random partitions containing $316$ pairs. For EPFL and CAMPUS-Human dataset, we randomly select reference images for each individual, and all target images are tested to match. The results on all three datasets are computed by taking average over ten runs. Our approach is evaluated under cases of both single reference image (single-shot, SvsS) and multiple reference images (multi-shot, MvsS, $M = 2,3$).
All the parameters are fixed in the experiments, including $\lambda = 10$ for scaling the overlap $\mathbf{IoU}$, $\alpha_u = 12$ and $\alpha_s = 3$ for penalizing the activation of vertices. We construct the MICT for each individual with their selected reference images. In the re-identification, a number $K$ of body part proposals are generated. In practice, we set $K$ approximately $3$ times the number of individuals in the shot.
We implement our approach with C++ and run the program on a PC with I5 2.8GHZ CPU and 4GB memory. On average, the inference algorithm converges after around $500$ samplings, which costs $2s \sim 40s$. The time cost is related with the complexity of the candidacy graph.
\begin{figure*}[ptb]
\begin{center}
\includegraphics[width=1.69in]{CMC_VIPeR.pdf}
\includegraphics[width=1.69in]{CMC_EPFL.pdf}
\includegraphics[width=1.69in]{CMC_CAMPUS_single.pdf}
\includegraphics[width=1.69in]{CMC_CAMPUS_multi.pdf}
\end{center}
\vspace{-4mm}
\caption{Performance comparisons using the CMC curves on VIPeR (left1), EPFL (left2), CAMPUS-Human (right1 and right2) datasets. EPFL and CAMPUS-Human datasets are evaluated in both single-shot and multi-shot cases.}
\label{fig:CMCCurves}
\vspace{-4mm}
\end{figure*}
\vspace{-1mm}
\subsection{Experimental Results} \label{sec:ExpResults}
\vspace{-1mm}
We compare our approach with the state-of-the-arts methods: Pictorial Structures (PS)~\cite{PictorRevisit}, View-based Pictorial Structures (VPS)~\cite{ViewPictor}, Custom Pictorial Structures (CPS)~\cite{CPS}, Symmetry-driven Accumulation of Local Features (SDALF)~\cite{SDALF} and Ensemble of Localized Features (ELF)~\cite{ELF}. We adopt the provided code of PS and implement VPS and CPS according to their descriptions. For fair comparison, the same likelihood is employed for PS, VPS and CPS as the proposed method. The results are evaluated by two ways: (i) re-identifying individuals in segmented images, i.e. targets already localized, and (ii) re-identifying individuals from scene shots without provided segmentations.
For the first evaluation, we adopt the cumulative match characteristic (CMC) curve for quantitative analysis, as in previous works~\cite{VIPeRReid,ETHZReid}. The curve reflects the overall ranked matching rates; precisely, a rank $r$ matching rate indicates the percentage of correct matches found in top $r$ ranks. As Fig.~\ref{fig:CMCCurves} shows, we demonstrate the superior performance over the competing approaches in both single-shot case and multi-shot case. And our method yields the best rank $1$ matching rate on EPFL and CAMPUS-Human datasets. We observe that the performance of re-identification can be improved significantly by fully exploiting reconfigurable compositions and contextual interactions in inference. Our performance only improves slightly on VIPeR dataset, as most erroneous matchings are due to severe illumination changes, which has been approved in~\cite{SDALF}.
\begin{table}[pb]
\vspace{-5mm}
\centering
\caption{Matching rate of re-identifying targets in scene shots without provided segmentations.}
\begin{tabular}{ccc}
\addlinespace
\toprule
Dataset & EPFL & CAMPUS-Human \\
\midrule
\multicolumn{1}{l}{Our M=2} & {\bf 57/294} & {\bf 215/1519}\\
\multicolumn{1}{l}{VPS M=2} & 54/294 & 175/1519\\
\multicolumn{1}{l}{PS M=2} & 32/294 & 141/1519 \\
\midrule
\multicolumn{1}{l}{Our Single} & {\bf 50/294} & {\bf 173/1519}\\
\multicolumn{1}{l}{VPS Single} & 49/294 & 139/1519\\
\multicolumn{1}{l}{PS Single} & 24/294 & 118/1519 \\
\bottomrule
\end{tabular}%
\label{tab:batchresult}%
\end{table}%
The second test is stricter, since the algorithms should also localize the target during re-identification. We adopt the PASCAL Challenge criterion to evaluate the localization results: a match is counted as the correct match only if the intersection-over-union ratio ($\mathbf{IoU}$) with the groundtruth bounding box is greater than $50\%$. We compare our method with PS~\cite{PictorRevisit}, VPS~\cite{ViewPictor}, which can localize the body at the same time as localizing the parts. The quantitative results are reported in Table~\ref{tab:batchresult}. A number of representative results generated by our method are exhibited in Fig.~\ref{fig:resulttable}. From the results, existing methods perform poor when individuals are not well segmented and scaled to uniform size. In contrast, our method can re-identify challenging target individuals by searching and matching their salient parts and thus achieves better performance. Note the performance of our approach also drops significantly due to inaccurate part localizations and interference of other individuals.
\begin{figure*}[ptb]
\begin{center}
\includegraphics[width=\textwidth]{ResultTable.pdf}
\end{center}
\vspace{-4mm}
\caption{Results generated by our approach on EPFL and CAMPUS-Human datasets. In each result, the query individual is specified by the image beside the shot. Green boundings denote the target groundtruth location, while red boundings are generated by algorithm.}
\label{fig:resulttable}
\vspace{-4mm}
\end{figure*}
{\em Component Analysis}. We further analyze component benefits of our approach on CAMPUS-Human dataset under the setting: multi-shot $M=3$. Regarding feature effectiveness, we separately evaluate different image features, as shown in Fig.~\ref{fig:termanalysis}(left). It is apparent that the combined feature improves the result. We also demonstrate the effectiveness of the constraints employed, and Fig.~\ref{fig:termanalysis}(right) confirms that both kinematics and symmetry constraints help construct better matching solution.
\vspace{-3mm}
\begin{figure}[htb]
\begin{center}
\includegraphics[width=1.62in]{Features.pdf}
\includegraphics[width=1.62in]{Constraints.pdf}
\end{center}
\vspace{-4mm}
\caption{Empirical studies on different features (left) and constraints (right) used in our approach on CAMPUS-Human dataset.}
\label{fig:termanalysis}
\vspace{-4mm}
\end{figure}
\vspace{-2mm}
\section{Conclusion} \label{sec:Conclusion}
\vspace{-1mm}
This paper studies a novel compositional template for human re-identification, in the form of an expressive multiple-instance-based compositional representation of the query individual. By exploiting reconfigurable compositions and contextual interactions during inference, our method handles well challenges in human re-identification. Moreover, we will explore more robust and flexible part representations and better inter-part relations in future works.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
\label{intro}
The advent of large area surveys with high sensitivity and the cross-match of
radio and optical data opened the possibility to use large samples of
extragalactic sources to investigate the links between the radio properties,
the central engine, and the host galaxies. Based on the recent results of
\citet{baldi09} and \citet{baldi10a}, a new population of radio-emitting
sources turns out to dominate in the local Universe.
\begin{figure*}
\includegraphics[scale=1.0]{diag.ps}
\caption{Spectroscopic diagnostic diagrams for the galaxies of the
sample. The small dot points correspond to the SDSS/NVSS sample
studied by \citet{baldi10a}, while the symbols in color represent
galaxies with the new JVLA observations. The solid lines are
from \citet{kewley06} and separate star-forming galaxies (SFG),
LINER, and Seyfert. In the first panel the region between the two
curves is populated by the composite galaxies. The dashed lines
instead separate RL AGN into HEG from LEG \citep{buttiglione10}
which roughly correspond to Seyfert and LINER classification for
RL AGN. The red squares are the LEGs, while the blue circles are
the HEGs. The large filled symbols are associated with BH masses
$>$10$^{8}$ M$_{\odot}$, large empty with $\sim$10$^{7.7}$
M$_{\odot}$, and small symbols with $<$10$^{7.3}$ M$_{\odot}$. }
\label{diag}
\end{figure*}
\citet{best05a} built a sample of 2215 radio-galaxies (RGs) by
cross-correlating the SDSS (DR2), NVSS, and FIRST datasets (hereafter the
SDSS/NVSS sample). This sample is selected at $F_{\rm 1.4} > 5$ mJy and it
includes RGs up to $z\sim$ 0.3, covering the range of radio
luminosity $L_{\rm 1.4} \sim 10^{22} - 10^{26}$ W Hz$^{-1}$. All morphologies
are represented, including twin-jets and core-jet FR Is, narrow and wide angle
tails, and FR IIs. However, most of them ($\sim$ 80 \%) are unresolved or
barely resolved at the 5$\arcsec$ FIRST resolution.
\begin{figure*}
\includegraphics[scale=0.75]{lrlo3_all.ps}
\caption{FIRST vs [O~III] line luminosity (erg s$^{-1}$). The colored
points represent the 12 galaxies observed with the JVLA coded as in
Fig.~\ref{diag}. The small dot points correspond to the SDSS/NVSS sample
studied by \citet{baldi10a} . The solid line represents the correlation
between line and radio-luminosity derived for the 3CR/FR~I sample. The
dotted lines include the region where Seyfert galaxies are found. The empty
pink triangles are the CoreG and the green stars are the 3CR/FRIs.}
\label{lrlo3}
\end{figure*}
The SDSS/NVSS sample is mostly ($\sim$90\%) composed of RL AGN, representing
the bulk of the RGs population (with a space density $>$ 100 times higher than
3C sources) but it is virtually unexplored. A few partial results have been
already obtained, even though the radio data available are extremely limited,
generally only FIRST and NVSS images with angular resolution problems and the
missing radio spectral coverage (they are generally too faint to be detected
in the existing high frequency surveys). In a previous paper \citet{baldi10a}
found the following results on this sample:
1) the AGN power estimated from the optical line luminosity of the SDSS/NVSS
sample is at the same level of classical FR~I RGs. However, the
sample shows a deficit of a factor $\sim$100 of radio emission with respect to the
3C sources with the same optical power.
2) the SDSS/NVSS sample generally shows optical emission line ratios typical
of AGN: most of them ($\sim 70 \%$) can be classified as Low Excitation
Galaxies (LEG) similarly to the FR Is, but a substantial fraction of High
Excitation Galaxies (HEG) is also present\footnote{We bear in mind that HEG
galaxies in the 3C catalog are all FR II RGs \citep{buttiglione10}}.
3) Most of the hosts of the SDSS/NVSS are statistically indistinguishable from
those of 3C sources from the point of view of morphology, color, stellar and
black hole masses. They are generally associated with giant elliptical
galaxies, with no signs of star formation, located in dense environment. Thus
the deficit in radio emission cannot be ascribed to differences in their
hosts.
4) A small fraction ($\sim$10\%) of the sample shows rather different
photometric properties and it represents the RQ AGN contamination to the main
population of RL AGN.
\smallskip
These results have been recently confirmed by the analysis of a larger sample,
obtained from the SDSS DR7, by \citet{best12}.
Similar results to the analysis of the SDSS/NVSS sample come from a parallel
study of a sample of 14 nearby giant early-type galaxies (ETGs) of extremely
low radio luminosity (10$^{20-22}$ W Hz$^{-1}$ at 1.4 GHz)
\citep{baldi09}. They are named CoreG and are much fainter than FR~Is
\citep{balmaverde06core}. Despite the low radio power, the CoreG host genuine
``miniature'' RL nuclei, indistinguishable from the FR~Is. In fact they extend
the FR~I relations at different wavelengths
\citep{chiaberge99,balmaverde06a,balmaverde06core,baldi10b}, indicating a
common synchrotron origin of the nuclear emission. However, CoreG show a less
extended radio structure and a core dominance up to $\sim$100 times higher
than FR~Is, but similar to the SDSS/NVSS sample. A high core dominance is
generally interpreted as evidence of Doppler boosting in a radio source
oriented at a small angle with respect to the line of sight. The correlation
found with the core radio power and emission line luminosity (independent of
orientation) indicates that we are seeing a genuine deficit of extended radio
emission and that no geometric effect is present.
However, the differences between CoreG and FR~Is might be driven by their much
lower radio luminosity. For this reason a high resolution radio study of the
SDSS/NVSS sample, objects sharing the same range of power of FR~Is, is
required. In fact, the radio properties of the bulk of the RL population are
virtually unexplored and we have little or no information on their radio
morphology or on their spectral properties. This severely hampers our ability
to understand their nature and how they are related, on the one hand, to
classical RGs and, on the other hand, to low luminosity AGN. An
improvement of the radio data is needed to provide us with a more complete
view of the radio emission phenomenon.
We then undertook a project aimed at understanding the nature of the radio
emission of such objects and its connection with the host galaxies. Precisely,
we present new high-resolution observations with the JVLA of a pilot sample of
12 sources selected from the SDSS/NVSS sample. We obtained A array
observations in L and C bands, reaching a resolution of 0\farcs2. In
particular, the C band provides us with the possibility of measuring the high
frequency core emission, crucial for our purposes.
We aim at answering the following questions:
1) which is the nature of the radio emission: star formation, radio outflows as
in RQ AGN or genuine emission from a relativistic jet as in RL AGN?
2) once we isolate the RL AGN, which is their level of core dominance? Are
they highly core dominated similarly to the CoreG, but unlike the 3C sources?
Or, alternatively, they are just compact, but with well developed extended
radio structures which produce most of the radio emission?
3) do they follow the relation between radio core and line emission defined by
3C/FR~Is? The new observations aim at isolating their genuine radio core emission,
an essential information to test their nuclear similarity with the FR~Is.
4) is their radio structure characterized by jet(s), diffuse plumes or
a double-lobed morphology? Are these jets one or two sided? The
distribution of jet sidedness will constrain their speed.
5) which is the fraction of Compact Steep Spectrum in the sample?
6) is there a relationship between the radio properties and the
optical spectroscopic classification, i.e., HEG and LEG differ from the
point of view of their radio structure?
The paper is organized as follows. In Sect. \ref{sample} we define the
sample. Section~\ref{obs} presents the new JVLA observations for 12
sources. In Sect.~\ref{results} we analyze the radio and
spectro-photometric properties of the sample. We discuss the results
in Sect.~\ref{discussion}, presenting a new class of radio-sources,
FR~0s (Sect.~\ref{FR0}). The summary and conclusions to our findings
are given in Sect.~\ref{conclusion}. We provide the notes on the radio
properties of the sources which show extended morphology in the
Appendix A.
\section{The sample}
\label{sample}
The objects observed with the JVLA have been extracted from the
SDSS/NVSS \citeauthor{best05a} sample adopting the following criteria:
\begin{itemize}
\item redshift $z<0.1$,
\item the main optical emission lines detected at least 5 $\sigma$ significance
(for the \Hb\ we relaxed this requirement to 3 $\sigma$) to ensure a reliable
spectral classification,
\item equivalent width for the \oiii\ lines larger than
3 \AA\ and with a measurement error smaller than 1 \AA; this ensures that the
line emission is associated with an AGN and not with a stellar origin \citep{capetti10},
\item declination in the range $-10 < \delta < 10$ in order to optimize the
observing strategy.
\end{itemize}
This procedure returns 68 objects. Twelve of them have been then randomly
selected for the observations. In Tab.~\ref{table1} we give the main
properties of the observed objects.
We use the spectroscopic diagnostic diagrams
defined by \citet{buttiglione10} for the 3CR sample to recognize the nature of
their nuclear emission. These diagnostics are formed by pairs of nuclear
emission line ratios, to separate active nuclei from star-forming galaxies
(e.g. \citealt{baldwin81}), and, furthermore to differentiate AGN into
branches of different excitation level, i.e. low-excitation galaxies (LEG) and
high-excitation galaxies (HEG). This separation is similar to that introduced
by \citet{kewley06} as diagnostic to separate among the RQ AGN population,
LINER and Seyfert \citep{heckman80}.
From the point of view of their optical spectral classification, our sample
covers the whole range of ionization (Fig.~\ref{diag}), with eight objects
located into the LEG area and two objects falling into in the HEG region. Two
objects (ID~590 and 625) straddle the separation between the two
classes; we classify 590 (625) as LEG (HEG) since in 2 out of three
diagnostic ratios it is consistent with such a classification.
In Fig.~\ref{lrlo3} we show their location in a diagram comparing the FIRST
radio and optical [O~III]. All objects show a large deficit of radio emission
when compared to the FR~Is part of the 3C sample by a factor ranging from
$\sim$ 30 to $\sim$ 1000. They are all (but one) located in the region of
higher density of the SDSS/NVSS sample, just avoiding the lowest line
luminosity, due to the third requirement for their selection. ID~656 is on the
boundary of the region of RQ Seyferts. Only one object (ID~625) falls in a
poorly populated area, being one of the objects of higher radio and line
luminosity. In Fig.~\ref{lrlo3} we also stress the similarity of our
sample to the CoreG sample in terms of L$_{FIRST}$/\loiii\ ratio.
Fig.~\ref{SDSSimage} shows the SDSS optical images of the galaxies of
the sample. Most of the targets has an optical morphology akin of
elliptical galaxies. Some, namely ID~519, 537, 547 and 568, show
edge-on disks, while 567 shows spiral arms.
\begin{figure}
\includegraphics[scale=0.40]{sdss_image.eps}
\caption{SDSS optical images of the galaxies of our sample. The size of each
frame is 40\arcsec $\times$ 40\arcsec.}
\label{SDSSimage}
\end{figure}
\input{tab.tex}
\input{tab2new.tex}
\input{tab2bis.tex}
\section{The JVLA observations}
\label{obs}
We obtained 6 hours of observations with the JVLA of NRAO\footnote{The
National Radio Astronomy Observatory is a facility of the National
Science Foundation operated under cooperative agreement by
Associated Universities, Inc.} with A array configuration in 3 days
on December 2012 and January 2013. We observed 12 objects in scans of
2 hours for each group composed of 4 sources. Each source was observed
for $\sim$20 minutes spaced out by the pointing to the phase
calibrators (J0022+0014, J0059+0006,J0217+0144, and J2337-0230). The
flux calibrator was 3C~48 observed for $\sim$6-7 minutes. The
observations were performed in L and C bands; the exposure time was
split in the two band configurations. While the L band configuration
corresponds to the default 1GHz-wide band centered at 1.4 GHz, the C
band was modified based on our purpose. We divide the available 2-GHz
bandwidth into two sub-band of 1 GHz centered at 4.5 and 7.5 GHz. This
strategy allows to obtain images in 3 different radio frequencies in
two integration scan. Each of the three bands was configured in 7
sub-bands of 64 channels of 1 MHz. The object ID~656 has been observed
only in L-band due to a telescope failure.
The data reduction was performed with {\it AIPS} (Astronomical Image
Processing System) package according to standard procedures. The
final image was then produced from the calibrated data set by {\it
CLEAN}ing to convergence. The final map was obtained using the task
{\it IMAGR} with a beam size ranging between $\sim$0\farcs2 and
$\sim$2\farcs6, according to the band, and different weights to
increase the signal to noise ratio. We self-calibrated the maps
of the sources with the gain flux $\gtrsim$10 mJy. The core component
parameters have been measured with the task {\it JMFIT}.
The noise level was measured in background regions near the
target. The final noise is strongly affected (mainly in L band) by the
presence of interferences within the large bandwidth. At high
frequency some observations show a gain error in a few telescopes that
we were not able to recover in the calibration phase. Since in some
sources the total flux density is too low to properly self-calibrate
the data in gain ($\sim$10 mJy or less) and even in phase (a few mJy
or less), all these problems produced final images with a measured
noise level higher than the expected thermal noise.
In Tab.~\ref{table2} we collect the radio properties. In particular we give
the total and core (when visible in the radio images) flux
densities in the 3
observing bands.
\section{Results}
\label{results}
\subsection{Radio properties}
The FIRST radio maps of the selected objects show unresolved compact
radio sources on a scale of 5\arcsec, with the exception of ID~625
which is extended on a scale of $\sim$ 35\arcsec.
In the JVLA observations, most of the sources appears unresolved or
slightly resolved down to a resolution of $\sim$0\farcs2 which
corresponds to 0.2--0.4 kpc. Some sources show instead radio
structures on a scale of 1-3 kpc: ID~547 (a twin-jet source), ID~568
(a double source and lacking of the radio core), and ID~590 (with an
elongated radio morphology on a scale of 0\farcs8, possibly due to a
bent two-sided jet structure). Finally, ID~625 shows a hybrid
radio-structure \citep{gopal00}, being a FR~I on the North side and a
FR~II on the South side on a scale of $\sim$40 kpc. Tab.~\ref{table2}
collects all the radio properties derived from the new
observations. Appendix~\ref{notes} provides additional comments on the
radio maps for the sources which show extended structures (Figures
from \ref{547} to \ref{625}).
One of the main purposes of our program is to isolate the radio core
component and to measure the core dominance of the sources of our
sample. However, the high resolution JVLA observations required with
this aim and obtained with relatively short exposure time might be
missing faint extended emission. In order to explore this possibility
we compared the JVLA 1.4 GHz fluxes with those measured by FIRST and
NVSS (see Table \ref{table2bis}). The JVLA/FIRST flux ratios range
from 0.71 to 1.28, with an average value of 0.95 indicating that we
recovered most of the radio emission. The comparison with the NVSS
data is more complex due to the low resolution of this survey. Indeed,
in two cases (namely ID 567 and ID 656) two point-like FIRST sources
are blended into a single NVSS catalog entry. Leaving these two
sources aside, the JVLA/NVSS flux ratios cover a range between 0.36
and 1.14, with an average of 0.78. We conclude that our measurement
recover most of the total radio flux of our sources. In general, the
fraction of missing flux is sufficiently small and it does not to
alter our view of the overall radio structure.
We obtained matched-beam radio maps to derive the radio spectra at the three
frequencies (see Fig.~\ref{SED}). For the object ID~656 it is not possible to
derive its spectrum because of the lack of its C-band observation.
In Fig.~\ref{SED} we report the spectral slope $\alpha$ ($F_{\nu}
\propto \nu^{\alpha}$), obtained between 1.4 and 4.5 GHz. The
spectral indices range from -1.16 to -0.04, indicative of both steep
and flat spectra. We note that source 568 shows a strong steepening at
7.5 GHz. This might be due to some missing flux at the highest
frequency, although we note that its double structures at 4.5 and 7.5
GHz are very similar. We suggest that the radio lobes spectra might be
intrinsically steep.
One of the key point of this study is the measurement of the radio
core. The high resolution radio maps at 7.5-GHz reveals a radio core
component for five sources (ID~511, 519, 537, 547, and 625). For the
remaining sources, we can use the radio spectra to measure the core
emission. Two sources (ID 590 and 605) show an overall flat ($\alpha
>$ -0.5) radio spectrum, consistent with a self-absorbed synchrotron
emission; the flux at 7.5 GHz is dominated by the radio core
emission. For ID~535 and 567 a spectral flattening occurs at 7.5 GHz
where the core emerges over the optically-thin radio emission. For the
last three objects of the sample we use the total flux at 7.5 GHz as
upper limit for the radio core measurement.
\begin{figure*}
\includegraphics[scale=0.7]{allsed.eps}
\caption{Radio spectra of the 11 objects observed with JVLA observed at 1.4,
4.5 and 7.5 GHz. The dotted line is the extrapolation of the 1.4-4.5 GHz
slope which is reported in each panel.}
\label{SED}
\end{figure*}
Having measured the radio core component thanks to the new observations, we
can include our sources in the $L_{core}$ vs. \loiii\ plane
(Fig.~\ref{coreo3}), similarly to what done for CoreG in \citet{baldi09}. The
radio core power distribution span a range similar to the local
3CR/FR~Is. Several sources lie in the same region of the 3CR/FR~Is also
considering their line luminosities, but others show a radio core deficit (or
line excess) with respect to the relation defined by the FR~Is. The sources
classified as LEGs, on average, lie closer to the relation than the HEGs.
The core dominance of our sources, $R$, is defined as the ratio
between the source nuclear emission at 7.5 GHz and the total flux
density, for which we adopt the NVSS measurement (Table
\ref{table2bis}). We preferred to use the 7.5 GHz core measurements,
with respect to the usual definition based on 5 GHz data, to limit as
much as possible the contamination from the extended structures. The
core dominance ranges between 0.05 and 0.86. The smallest R belongs to
the most extended object, the FR~I/FR~II ID~625.
\begin{figure*}
\includegraphics[scale=0.7]{lcorelo3.ps}
\caption{Core radio power vs. [O~III] line luminosity (erg s$^{-1}$)
for CoreG (pink triangles), 3CR/FR~I radio-galaxies (green stars),
and our sample (colored symbols like in Fig.~\ref{diag}). The line
indicates the best linear the for 3CR/FR~Is.}
\label{coreo3}
\end{figure*}
\input{tab3.tex}
\subsection{Spectro-photometric properties}
We now explore the physical properties of the present objects
based on the optical spectroscopic
and photometric information available from the SDSS survey. We consider the
mass of their central black hole (BH), the host mass and morphology, and the
$D_n(4000)$ index (see Tab.~\ref{table3}), similarly to the study of SDSS/NVSS
sample by \citet{baldi10a}.
We estimate the BH mass ($M_{\rm{BH}}$) from the stellar velocity dispersion
adopting the relation of \citet{tremaine02}. They range from $\sim$10$^{7}$ to
$\sim$10$^{9}$ $M_{\odot}$. The sources are associated with galaxies with a
distribution of masses\footnote{$M_{*}$ is estimated from the $g$ and $z$ SDSS
magnitudes following following \citet{bell07} resulting in slightly
different masses from those reported by \citet{best05a}.} in the range
$10^{10}$--$10^{11.5}$ M$_{\odot}$ (and a median of $\sim 10^{11}$ M$_{\odot}$).
\begin{figure*}
\centerline{
\includegraphics[angle=90,scale=0.4,angle=0]{cr_dn4000_fr0.ps}
\includegraphics[angle=90,scale=0.4,angle=0]{LrsuMsDn4000_fr0.ps}}
\caption{Left panel: concentration index in the r band, $C_r$, versus the
$D_n(4000)$ index. The vertical lines correspond to the two
definition of ETGs ($C_{r}>$ 2.86 or 2.60, see text for details), while the
horizontal line separates the blue from red sources. Right panel:
$D_{n}(4000)$ versus $L_{\rm 1.4 \,\,GHz}/M_{*}$. The dashed
curve is the empirical separation between the star forming and AGN radio
emitting sources, performed by \citet{best05a} (read the text for
details). The dotted curve is shifted 0.1 above in $D_{n}(4000)$ of the
dashed curve, in order to separate the objects with a possible
star-formation contribution to their radio emission. The small dot points
correspond to the SDSS/NVSS sample studied by \citet{baldi10a}. The
different color/shape symbols correspond to the different classes like in
Fig.~\ref{diag}.}
\label{spdd}
\end{figure*}
We then study the host type by using the concentration index $C_{r}$ (defined
as the ratio of the radii including 90\% and 50\% of the light in the $r$ band
respectively). ETGs have $C_{r}\geq 2.86$
(e.g. \citealt{strateva01,kauffmann03b,bell03}). This cut basically
corresponds to E type with a small contamination of S0 and Sa
\citep{bernardi10}. A less conservative selection of ETGs uses $C_{r}\geq
2.60$ (e.g. \citealt{nakamura03,shen03}), which yields a larger contribution
from S0 and Sa.
The 4000\AA\ break strength, $D_n(4000)$ (defined as the ratio between
the fluxes in the wavelength range 3850-3950\AA\ and 4000--4100\AA) is
instead sensitive to the presence of a young stellar population
\citep{balogh99}. Massive galaxies (log $M_*/M_{\odot} \gtrsim 10.5$)
shows a bimodal distribution of $D_n(4000)$ \citep{kauffmann03b}. The
transition between the two groups is located approximately at
$D_n(4000) \sim 1.7$ with the red (blue) hosts having larger (smaller)
values.
A further diagnostic panel tool enables us to to qualitatively measure the
amount of contamination in radio emission in the galaxy due to star
formation. This method is based on the location of a galaxy in the
$D_{n}(4000)$ versus $L_{{\rm 1.4 \,\,GHz}}/M_{*}$ plane, where $M_{*}$ is the
galaxy's stellar mass. \citet{best05a} predict the radio emission per
unit of galaxy mass expected from a stellar population of a given
age: when a galaxy shows an excess larger than 0.225 in
$D_{n}(4000)$ above the curve corresponding to the prediction of a star
formation event lasting 3 Gyr and exponentially decaying, its radio emission
is associated with the AGN.
Figure~\ref{spdd}, in the left panel, shows the distribution of the objects in
the spectro-photometric diagram composed of $C_{r}$ vs. $D_n(4000)$,
while in the right panel we compare $L_{\rm 1.4 \,\,GHz}/M_{*}$ and
$D_{n}(4000)$. In order to explore the properties of the sample, we
separated the sources into three groups based on the BH masses: large
$>$10$^{8}$ M$_{\odot}$, intermediate $\sim$10$^{7.7}$ M$_{\odot}$ and
small $<$10$^{7.3}$ M$_{\odot}$.
We note a different behavior of the sources in our sample, related with their
BH mass. A first group consists in 4 sources with small (and one intermediate)
BH masses, namely ID~511, 567, 568, and 656. Three of them (see
Fig. ~\ref{spdd}, right panel), have a likely contribution of a star formation
to their radio emission; this is supported also by their large NVSS/FIRST
radio flux ratios, an indication of the presence extended diffuse radio
emission, missed by the FIRST images. The fourth object (ID~568) is consistent
with being a Seyfert 2 galaxy, based on its location in Fig.~\ref{lrlo3} and
on its spectrophotometric properties. They are all blue galaxies, having
$D_{n}(4000) < 1.6$. In the diagram comparing the core power, L$_{\rm core}$,
and emission line luminosity (Fig.~\ref{coreo3}) they all show a large deficit
of radio core emission, typically a factor of $\gtrsim$100 , with respect to the
relation defined by 3CR/FR~Is.
Conversely, the second group includes sources with large/intermediate
BH masses, whose radio emission is dominated by the AGN. With the sole
exception of ID~625 (the hybrid FR~I/FRII radio source) they are
associated with red massive ETGs belonging to the LEG spectroscopic
class. With respect to the first group, they follow more closely the
L$_{\rm core}$ vs \loiii\ relation.
Fig.~\ref{cd} shows that the core dominance for this sub-sample ranges
from $\log R \sim$ -1 to $\sim 0$ with a mean value of $\sim$-0.5.
This distribution is significantly different from from that of
3CR/FRIs (with a $>$99.9\% probability, according to a Kolmogorov
Smirnoff test) while it is not distinguishable from the $R$
distribution of CoreG. We remind that we estimated $R$ as the ratio
between the 7.5-GHz core emission and the total 1.4-GHz flux (NVSS),
while for the 3CR/FRIs and CoreG we used the 5 GHz core flux against
the total 1.4-GHz flux (NVSS). However, since the radio core emission
has generally a flat spectrum this quantity is only weakly dependent
on the frequency used for the core measurement and this comparison is
robust.
\begin{figure}
\includegraphics[scale=0.45]{cdisto_fr0.ps}
\caption{Histograms of core dominance Log R = log (L$_{\rm
core}$/L$_{\rm tot}$) for the 3CR/FR~Is (black solid line), CoreG
(black dashed line) and our sample of seven FR~0s (red solid line).}
\label{cd}
\end{figure}
\section{Discussion}
\label{discussion}
The results discussed in the previous sections indicate that the
twelve objects, belonging to the SDSS/NVSS sample and observed with
the JVLA, can be divided in two main groups, corresponding to specific
radio, spectroscopic, and photometric properties. Before discussing
these groups in more detail, we treat here the only object that shows
rather different properties. Not surprisingly, by exploring a sample
of radio-sources, we found a classical extended RG, namely ID~625. As
already mentioned it has a hybrid FR~I/FRII morphology extending over
$\sim$40 kpc. All its characteristics are typical of FR~II sources
considering its BH mass, the blue host color, its spectroscopic class
(HEG), and its location in both the L$_{\rm core}$ and L$_{\rm tot}$
vs \loiii\ diagram in the region populated by the FR~II part of 3C
sample \citep{chiaberge02,buttiglione10,chiaberge11}.
The first main group consists of four sources which are characterized
by their low BH masses, mostly $\sim$10$^{7}$ M$_{\odot}$ and their
blue color. Their radio and spectrophotometric properties indicate
they are RQ AGN (Tab~\ref{table3}). The presence of an active nucleus
is witnessed by their optical line ratios and equivalent widths,
characteristic of AGN. Nonetheless, three of them shows a substantial
contamination from star formation to their radio emission. The fourth
object (ID~568) shows a double radio morphology with a size of 0.8 kpc
and a steep spectrum, often observed in Seyfert galaxies
\citep{edelson87,nagar05}. In addition, its radio and line power
locate this source in the area typical of the Seyfert galaxies. The
contamination from RQ AGN ($\sim$33\% of the sample) is higher than
observed in the whole SDSS/NVSS sample \citep{baldi10a}. This is due
to the fact that the selection criterion of this pilot sample based on
the high [O~III] equivalent width favors the inclusion of bright
emission-line galaxies, which are preferentially RQ AGN.
The second group consists of seven sources, representing (together
with the powerful source ID~625 discussed above) the genuine RL AGN
population in the SDSS/NVSS sample. They live in red massive
($\sim$10$^{11}$ M$_{\odot}$) ETGs, have BH masses $\gtrsim$10$^{8}$
M$_{\odot}$ and are spectroscopically classified as LEGs (apart from
the HEG-spectrum FR~I/FR~II 625). All these properties are shared with
FR~I RGs (e.g.,
\citealt{zirbel96,cao04,floyd08,chiaberge99,baldi08,baldi09,buttiglione10}). Furthermore,
their radio core and [O~III] luminosities lie in the range typical of
FR~Is. Therefore, all the nuclear and host characteristics of this
group are indistinguishable from those of FR~Is. Nonetheless, the
JVLA observations presented here show compact radio structures (with a
limit to their size of $\sim$ 0.5 kpc or, in a few cases, extending by
at most 1-3 kpc) and lead to an estimate of their (average) core
dominance a factor of $\sim$30 higher than FR~Is. The only feature
distinguishing them from FR~Is is then the substantial lack of
extended radio emission. For such a paucity we define them ’{\it
FR~0}’.
\subsection{The FR~0 population}
\label{FR0}
The FR~0 classification corresponds to a combination of radio and
spectro-photometric properties. In the radio band, FR~0s are
characterized by a core responsible for a fraction larger than
$\gtrsim$10\% of the total emission with structures extending for, at
most, a few kpc. In classical FR~Is and FR~IIs RGs the core dominance
is typically $\lesssim$1\% and their sizes are of tens or hundreds of
kpc. (e.g, \citealt{morganti93,morganti97}). The high FR~0 core
dominance appears to be due to the paucity of extended radio emission,
rather than to an enhanced radio core, since FR~0s and FR~I show
similar ratios of radio-core to emission line luminosity. In addition,
the FR~0 definition involves the photometric/spectroscopic nuclear and
host properties (e.g., BH mass, color, and emission line ratios,
similar to FR~Is), which distinguish them from the radio-sources
associated with RQ AGN and star forming galaxies.
This new class of RGs is similar to the low-luminosity radio sources
hosted in ETGs, studied by \citet{slee94}, which contain parsec-scale
radio cores and do not produce extended radio emission. Furthermore,
recently, \citet{sadler14} found that the bulk of the 20-GHz RG
population consists of compact radio sources lacking of extended radio
emission, analogous to our FR~0s.
The CoreG fulfill the requirements for a FR~0 classification: they show kpc
scale radio structures and are of high core dominance, they are hosted in red
giant ellipticals and are characterized by LEG line ratios. They are
$\sim$100 times less luminous than the FR~0s of the SDSS/NVSS sample discussed
here. They smoothly extend the various nuclear multi-wavelengths relations
seen in FR~Is. In this sense
they represent the low luminosity end of the FR~0 population that therefore
extends at least down to a radio power of $\sim 10^{36}$ $\ergs$.
The small size and the steep spectra of FR~0s might suggest the
presence of Compact Steep Sources (CSS), i.e. of rapidly growing young
radio-sources (e.g. \citealt{odea98,snellen00}). CSS show a
correlation between the turnover frequency in their spectra and their
linear sizes (e.g., \citealt{fanti90,odea98}). The sizes of our FR~0
sources ($<$3 kpc) does not contrast with their steep spectra since
the turnover might occur at lower frequencies than the lowest value
sampled by our observations, 1.4 GHz. However, the core dominance of
CSS, although generally more difficult to measure and uncertain with
respect to more extended radio-sources, is much lower than in FR~0s
(typically ${\rm log} \,R \sim$-1 for quasars and $\log \,R \sim$-2.4
for galaxies \citealt{saikia95,saikia01}) ruling out a general
association of the FR~0s with the CSS.
As reported in the Introduction, the vast majority of the SDSS/NVSS
objects fulfill the FR~0 definition with the notable exception of the
lack of a direct measurement of the core dominance, impossible from
the FIRST images. Indeed, this was the main motivation to obtain the
higher resolution data presented in the previous
sections. Nonetheless, the possibility that they are generally high
core dominance objects, and thence bona-fide FR~0, appears to be
reasonably founded. In fact, most of them ($\sim$80 \%) are unresolved
or barely resolved in the FIRST maps, limiting their sizes to, at
most, a few kpc. This is an indication that, similarly to the 7
objects studied, they have a deficit of extended radio emission with
respect to classical extended RGs. It is clearly highly desirable to
obtain proper radio data for a larger sample to base this conclusion
on firmer ground. Therefore, if this result was confirmed for a larger
sub-sample of the SDSS/NSS RGs, the FR~0 population would turn out to
be the dominant class of RGs in the local Universe with a space
density $>$100 times higher than 3C sources, as also suggested by
\citet{sadler14}.
FR~0s apparently differ from FR~Is just for the lack of large scale radio
structures. In the following we present two scenarios that can account for
this specific property.
In the first scenario, the central engines of FR~0s and FR~Is are
indistinguishable and the paucity (and small size) of the extended
radio emission is ascribed to an evolutionary effect. FR~0s might be
young RGs that did not yet developed an extended radio
structure. However, their number density should be smaller than that
of FR~Is, the opposite of what is observed. This scenario can hold if
FR~0s are intermittent sources. Rapid AGN intermittency (see, e.g.,
\citealt{readhead94,reynolds97,czerny09}) would prevent FR~0s from
becoming well developed RGs. However, this does not explain why
intermittency should affect only FR~0s, particularly considering the
indications that accretion in FR~Is and CoreG is associated with a
long-lasting inflow of hot coronal gas \citep{allen06,balmaverde08},
related only to the host properties. Therefore, no differences would be
expected between FR~0s and FR~Is.
As a second scenario, the differences between FR~0s and FR~Is might be
due to a lower jet bulk speed $\Gamma$ in FR~0s than in FR~Is. In this
scheme, the innermost regions of FR~0 and FR~Is do not differ
significantly since they share a common range of accretion rates
(proved by the similarity in line emission luminosity and optical line
ratios) and of radio core powers (which represent the synchrotron
emission from the base of the jet). The different radio behavior
should arise on a larger scale. In the hypothesis that jets in FR~0s
are slower than FR~Is, they are more subject to instabilities and
entrainment \citep{bodo13} and this causes their premature
disruption. Indeed, the typical scale of the radio emission in FR~0s
is smaller than the core size of their hosts, a region characterized
by a dense interstellar medium that obstructs the passage of the
jet. This idea is supported, albeit with a small number statistics, by
the absence of one-sided kpc scale morphologies among the FR~0s
observed, the typical sign of relativistic jet boosting. The ultimate
origin of the lower $\Gamma$ factor in FR~0 is apparently not related
to any directly observable quantity. We speculate that this could be
due to a different spin of their central BH, by assuming a dependence
between the BH spin and $\Gamma$, as suggested by, e.g.,
\citet{mckinney05}, \citet{tchekhovskoy10}, \citet{chai12}, and
\citet{maraschi12}. The FR~I radio morphology is produced only when
the BH spin is close to its maximum value, while smaller spin values
could be associated with FR~0s. Recently, \citet{ghisellini14}
predicted a population of jetted sources which have smaller BH spin
than classical RL AGN: the FR~0s might represent or be part of this
population.
\section{Summary and Conclusions}
\label{conclusion}
We present the observations of 12 objects with the JVLA in A-array
configuration. These sources are selected from a large sample
(cross-matching NVSS/FIRST and SDSS) which represents the bulk of the
local radio-emitting AGN population \citep{baldi10a}. Based on the
FIRST radio maps, most of the sources are compact on an angular scale
of 5\arcsec.
The new high-resolution observations at 1.4, 4.5, and 7.5 GHz reveal
that they still show a compact morphology, extending at most over 1-3
kpc (with just one exception, a hybrid FR~I/FR~II which extends over
$\sim$40 kpc). We isolate and measure the radio core component thanks
to the new high-resolution maps or to the radio spectra. Furthermore,
the SDSS survey provides with the spectro-photometric properties of
the sample, such as, BH mass, host type, emission lines. Based on
these properties, we divide the sample into two groups.
The first group consists of four sources which represent the RQ AGN
contamination to the SDSS/NVSS sample. They have small BH masses (mostly
$\sim$10$^{7}$ M$_{\odot}$) associated with blue galaxies which show evidences
of star formation.
The second group consists of seven RL AGN in red massive ($M_{*}
\sim$10$^{11}$ M$_{\odot}$ and BH masses $\gtrsim$10$^{8}$
M$_{\odot}$) ETGs spectroscopically classified as LEG. All these
characteristics are shared by local FR~Is. In particular, the members
of this group have similar radio and [O~III] line luminosities of
FR~Is, lying on the same relation; however, they show a core dominance
higher than in FR~Is by a factor $\sim$30. What distinguishes this
group from FR~Is is one single aspect: the lack of substantial
extended emission. For this characteristic we define these objects
`FR~0'. This new RG class is consistent with a sub-class of
Gigahertz-Peaked Spectrum (GPS) radio sources, proposed by
\citet{tingay15}, with low luminosity and with jet-dominated compact
morphologies similar to FR~Is and lacking of extended radio
emission. Furthermore, recently, \citet{sadler14}, studying the
20-GHz radio-source population in the local Universe, suggest that
the FR~0s might be a mixed population of young CSS and GPS radio
galaxies.
What causes the deficit of extended radio emission in FR~0s? We
discuss two possible interpretations. The first scenario proposes that
the FR~0s consist in young intermittent radio sources that will
eventually evolve into extended RGs. However, this contrasts with the
picture in which low luminosity RGs are powered by continuous
accretion of hot gas from their X-ray coronae. Furthermore, it does
not explain why the intermittency should only occur in FR~0s and not
in FR~Is, considering the similarity of their nuclear and host
properties.
The second scenario suggests that FR~0s have smaller jet Lorentz factors
$\Gamma$ than FR~Is: their jets are less stable against entrainment and their
passage through the dense ISM region of the central regions of their hosts
causes their premature disruption. This low-$\Gamma$ scenario can be tested by
looking at the asymmetry between the opposite jets in FR~0s, a parameter
directly linked to the jet speed. Although twin symmetric jets are common in
FR~Is, they often have one jet substantially brighter than the other on kpc
scale (e.g., \citealt{parma87}). This indicates that the high Doppler factors
derived for FR~Is on pc scale (based on the detection of superluminal motions,
e.g., \citealt{giovannini01}, and, indirectly, on the association with BL~Lac
objects in the AGN unified model, e.g. \citealt{urry95}) are preserved to
larger scales. In the present sample we find that all three FR~0s with an
extended radio structure are two-sided and rather symmetric. The number is too
small to draw any firm conclusion (and two-sided jets are expected to be more
common than a one-sided morphology in a randomly oriented sample) but it is
consistent with this hypothesis. We speculate that the ultimate difference
between FR~0s and FR~Is is the spin of their BH, being smaller for FR~0s.
High resolution imaging of a larger sample of local low luminosity
radio sources is needed to put this result on firmer statistical
ground and confirm the presence of the FR~0 population. In order to
improve the efficiency of such a survey, the results obtained in this
study indicate that it should be focused on objects with a LEG
spectrum, hosted in red early type galaxies. If the FR~0 population
was confirmed, they would represent the bulk of the RL AGN in the
local Universe with a large impact on our view of the low-luminosity
radio-emitting population.
|
\section*{Acknowledgments}
This work was supported by
RIKEN Brain Science Institute and the Brain Mapping by Integrated Neurotechnologies for Disease
Studies (Brain/MINDS) by the Ministry of Education, Culture, Sports,
Science and Technology of Japan (MEXT).
|
\section{Introduction}
\label{Sec:Intro}
Harnessing strong and coherent interactions between quantum light
and matter is an important ability in quantum science. These
interactions can be used to build light-matter interfaces enabling
reversible quantum state transfer between photons and atoms. One
important application of these interfaces is the realization of
photonic quantum memories
\cite{Hammerer2010,Lvovsky2009,Simon2010,Bussieres2013} which
allow storage of quantum information carried by photons. Quantum
memories are important devices in quantum information science
because they can be used as synchronization devices when many
different probabilistic quantum processes are linked together.
They are therefore required for scalable protocols using photons,
with potential applications in optical quantum computing,
generation of multiphoton states from probabilistic pair sources
\cite{Nunn2013}, quantum information networks \cite{Kimble2008}
and long-distance quantum communication using quantum repeaters
\cite{Briegel1998,Duan2001,Sangouard2011}.
The realization of quantum memories for light requires efficient
and reversible mapping of photons onto long lived atomic
coherences. This in turn requires strong interactions between
light and matter. However, in free space the interaction between a
single photon and a single atom is usually weak. One way to
overcome the problem is to place the atom in a high finesse cavity
which strongly enhances the interaction \cite{Specht2011} (see
chapter (Kuhn)). Another way is to use a collection of atoms,
where the atom light coupling is enhanced by a factor of
$\sqrt{N}$ with $N$ being the number of atoms involved. Single
photons are stored in atomic ensembles as collective atomic
excitation, sometimes called superatoms. These superatoms have the
important property that they can be efficiently converted to
single photons in a well defined spatio-temporal mode thanks to a
collective interference between all the involved emitters
\cite{Duan2001}. This so-called collective enhancement is at the
heart of most quantum memory protocols in atomic ensembles. The
quantum control of collective atomic coherences is therefore a key
task in the field of quantum memories.
For applications involving transfer of quantum information over
large distances, remote quantum memories must be entangled
\cite{Duan2001,Simon2007a}. This requires that the remote quantum
memories must exchange quantum information using e.g. single
photons, or that photons emitted by the quantum memories interfere
at a central location between the two quantum memories. If optical
fibers are used, this means that quantum memories must be
connected to the optical fiber network, in particular to photons
at telecom wavelengths in order to minimize optical losses in the
fiber transmission.
Quantum memories for light were first demonstrated in atomic
gases, both room-temperature gases and ensembles of laser-cooled
atoms. Several review papers can also be found on that subject
\cite{Hammerer2010,Sangouard2011}. Some solid-state systems offer
interesting perspectives as quantum memories for light, such as
rare-earth doped crystals
\cite{Riedmatten2008,Hedges2010,Clausen2011,Saglamyurek2011,Rielander2014},
nitrogen-vacancy centers in diamond \cite{Togan2010,Bernien2013},
phonons in diamond \cite{Lee2012,England2013} and quantum dots
\cite{deGreve2012,Bao2012,Simon2010} (see also chapter McMahon). Realizing quantum memories
in solid-state systems would, in general, have several advantages,
such as the absence of atomic motion and the prospects for
integrated devices, which may facilitate large-scale deployment of
these techniques in future quantum networks. But, controlling
light-matter interactions in solid-state materials also poses
important challenges, such as preserving the quantum coherence in
a solid-state environment.
Here we will discuss rare-earth-ion doped crystals for quantum
memories. These crystals provide a large number of atoms naturally
trapped in a solid-state matrix, with spectrally narrow optical
and spin transitions. Due to their particular electron level
structure, they also provide exceptional coherence properties,
both for the optical and spin transitions, when cooled to
cryogenic temperatures.
This chapter will describe experiments and techniques developed to
store non-classical light in rare-earth-ion doped crystals. In
section \ref{Sec:RECrystals}, we describe in more details the
relevant properties of rare-earth-ion doped crystals and the
reasons why these are interesting materials for quantum light
storage. In section \ref{Sec:Protocols}, we describe quantum
memory protocols that have been proposed to store quantum
information in doped crystals. In section
\ref{Sec:State_of_the_art}, we review the experimental
state-of-the-art of storing non-classical states of light in
crystals. In the following two sections we describe the
development of two specific sources of non-classical light
(Section \ref{Sec:SPDCsource:General}) and their application to
non-classical light storage (Section \ref{Sec:QStorage}) in
Pr$^{3+}$ and Nd$^{3+}$ ions doped crystals. Finally, in section
\ref{Sec:prospects}, we comment on the prospects for extending
quantum light storage to longer storage times in these systems.
\section{Rare-earth-ion doped crystals}
\label{Sec:RECrystals}
The
energy structure of ions in solid-state materials is usually
strongly affected by the lattice of the host crystal, resulting in
broad optical transitions with very short optical coherence time.
Striking exceptions to this are rare-earth ion impurities in
crystals, whose 4f-4f transitions were found to be extremely
narrow when the first high-resolution resonance spectra were
obtained in the 1970s \cite{Macfarlane2002}. The sharp lines are
due to the shielding of the 4f electron shell from outermost 5s
and 5p electrons, which reduces the coupling of the 4f electrons
to the lattice. This explains the atomic-like properties of the
lanthanides in a crystal. There exist several excellent books
\cite{Macfarlane1987,Liu2005} and reviews of optical properties of
rare-earth-ion doped crystals
\cite{Macfarlane2002,Macfarlane1990,Macfarlane2007} and their
application in quantum information science
\cite{Tittel2010,Thiel2011}. Here we will summarize some
properties that are particularly relevant for quantum memory
applications.
The effective shielding of the 4f electron shell results in
extremely narrow homogeneous and inhomogeneous line widths of the
radiative 4f-4f transitions in rare-earth doped crystals. At
cryogenic temperatures, the inhomogeneous broadening is the
dominating broadening process, analogous to the Doppler broadening
of room temperature alkali gases. But in contrast to the dynamical
Doppler broadening, where atoms jump between velocity classes due
to velocity changing collisions, the inhomogeneous broadening of
rare-earth ion doped crystals is to a large degree static. This is
a result of the physical origin of the inhomogeneous broadening
\cite{STONEHAM1969}, which can be due to local crystal strain or
interactions between dopants \cite{Sellars2004}. In some cases one
can observe a time-dependent broadening of a spectral channel over
time, which is known as spectral diffusion \cite{Bottger2006a}.
However, this effect is usually weak and is taken into account by
a time-dependent homogeneous line width \cite{Bottger2006a}. The
static inhomogeneous broadening and the large number of spectral
channels that can be manipulated with precise lasers have
important consequences for quantum storage experiments. These
features are used in the quantum memory schemes specifically
developed for these materials, resulting in capabilities difficult
to obtain in gas phase experiments.
The optical inhomogeneous linewidth varies strongly between
crystal hosts. In some cases the degree of broadening can be
related to the ionic radius mismatch between the rare-earth dopant
ion and the lattice ion that it replaces. Usual inhomogeneous
broadenings range from a few hundreds of MHz to tens of GHz,
although extreme values of around 10 MHz \cite{Macfarlane1998} and
250 GHz have been observed \cite{Sun2002}.
The optical homogeneous linewidth can be extremely narrow (ranging
from $<$1 kHz to 1 MHz) if the sample is cooled to temperatures
below 5-10 K \cite{Macfarlane1987,Liu2005}. Above this
approximative temperature range, the homogeneous linewidth usually
increases rapidly, displaying a $T^7$ or $T^9$ dependence, due to
coupling to phonons (spin-lattice relaxation). Below this
temperature range, due to the shielding by the 5s and 5p
electrons, the homogeneous linewidth is often limited by magnetic
interactions with other rare-earth ion dopants or magnetic
constituents of the lattice. This has led to the general
understanding that crystal materials with low nuclear spin
concentration \cite{Macfarlane1981,Yano1991} and low rare-earth
ion dopant concentration \cite{Bottger2006a} provides a way of
obtaining long coherence times, both for optical and hyperfine
transitions. To increase the coherence times of hyperfine levels,
it was also realized that one can exploit the non-linear magnetic
Zeeman effect of the hyperfine levels in order to find sweet spots
where there is a zero first-order Zeeman effect (ZEFOZ) of the
hyperfine transition \cite{Fraval2004}. This effectively decouples
the hyperfine transition from the fluctuating magnetic
environment, which can increase the hyperfine coherence time with
orders of magnitude, akin to clock transitions used in alkali
atoms. In addition, one can apply dynamical decoupling schemes to
further increase the spin coherence time
\cite{Fraval2005,Pascual-Winter2012,Lovric2013,Heinze2013}. A recent experiment demonstrated a coherence time of 6 hours in
Eu$^{3+}$:Y$_2$SiO$_5$ \cite{Zhong2015}.
The parity-forbidden 4f-4f transitions are only weakly allowed in
crystals, and arise due to admixtures of excited configurations of
different parity to the 4f configuration. As a consequence
radiative lifetimes are long, usually in the range of 100 $\mu$s
to 10 ms. The oscillator strengths of transitions relevant to
quantum information applications are in the range of 10$^{-8}$ to
10$^{-6}$ \cite{McAuslan2009}. The weak absorption probability of
individual ions, as compared to the alkali D$_1$ and D$_2$ lines,
is compensated by the high density of ions typical for a crystal.
The doping concentrations are in the range of 10 to 1000 pm,
resulting in typical number densities in the range of 10$^{17}$ to
10$^{19}$ ions/cm$^{3}$. The absorption coefficients for typical
doping levels range from $\alpha$ = 1 to 50 cm$^{-1}$, which often
make the transitions close to opaque for a 1 cm long crystal.
We finally comment on the electronic ground-state substructure,
which is relevant for quantum information processing in general.
The RE$^{3+}$ ions can be divided into Kramers and non-Kramers
ions, which have odd or even number of electrons, respectively.
This is relevant since the two groups react differently to the
interaction with the surrounding lattice ions (crystal-field
interaction). In a low-symmetry RE$^{3+}$ doping site, the crystal
field interaction completely lifts the electronic ground-state
degeneracy for non-Kramers ions. For Kramers ions the
crystal-field interaction results gives rise to a series of
degenerate doublets, which is due to Kramers time-reversal
symmetry. As a consequence the Kramers ions have a strong magnetic
dipole moment (order of the Bohr magneton), while non-Kramers ions
usually have weak nuclear magnetic moments due to the quenching of
the electronic magnetic moment. In some cases, however, the
nuclear moment can be strongly enhanced in non-Kramers ions
through interactions induced by the crystal-field Hamiltonian. In
addition we also need to consider different types of hyperfine
interactions. As an example, in Eu$^{3+}$ and Pr$^{3+}$ doped
crystals, which are non-Kramers ions, these ions have three
hyperfine states, due to quadrupole-type interactions
\cite{Macfarlane1987,Liu2005}. The hyperfine level separations in
these ions are of the order of 10 to 100 MHz. In Kramers ions,
such as Er$^{3+}$ and Nd$^{3+}$, a spectrally resolved Zeeman
ground-state doublet can be formed by applying a rather weak field
($<$1 Tesla) \cite{Sun2008}. For isotopes with non-zero nuclear
spin, the interaction with unquenched electron spin results in a
strong hyperfine interaction, of the order of 1 GHz for Kramers
ions \cite{Guillot-Noel2006}. These different considerations are
important for different aspects of quantum memory applications,
since they affect the frequencies used for state preparation of
the ions and possible limitations of the memory bandwidth due to
interfering optical-hyperfine transitions.
\section{Quantum memory protocols}
\label{Sec:Protocols}
In this section we will shortly discuss quantum memory protocols
that can be used in rare-earth-ion doped crystals. We will
particularly discuss the atomic frequency comb protocol, which so
far is the only protocol that has been used in these systems for
storage of quantum states of light. For the reader who wants an
overview of different quantum memory schemes we refer to the many
excellent reviews that have been published
\cite{Hammerer2010,Lvovsky2009,Tittel2010,Simon2010,Bussieres2013}.
An important class of quantum memory schemes have been inspired by
the photon echo process \cite{Tittel2010}, which is the optical
analogue of the spin echo. In the conventional photon echo process
the optical pulse to be stored is absorbed by an inhomogeneous
ensemble of atoms, typically rare-earth impurity ions in a
crystal. The induced atomic coherence then undergo inhomogeneous
dephasing, but the dephasing can be reversed by applying an
optical $\pi$-pulse a time $\tau$ after the input pulse. After a
time 2$\tau$ the atomic coherences are back in phase, which
results in a strong collective emission known as a photon echo.
Photon echo processes have been investigated for storing and
processing coherent states of light since the 1980s
\cite{Tittel2010}. It was therefore natural to consider if the
same processes and materials could be used to store the quantum
state of a single photon. It was realized, however, that the
optical $\pi$-pulse used in conventional photon echoes would cause
too much spontaneous emission noise, due to the high degree of
atomic excitation induced by it \cite{Ruggiero2009}. In 2001
Moiseev and Kr\"oll \cite{Moiseev2001} made an initial proposal
for a noise free photon-echo quantum memory scheme. The scheme was
proposed for Doppler-broadened lines, but sparked the interest for
finding a similar scheme adapted to inhomogeneously broadened
solid-state ensembles. Around 2005-2006 Nilsson and Kr\"{o}ll
\cite{Nilsson2005}, Kraus et al. \cite{Kraus2006} and Alexander et
al. \cite{Alexander2006} proposed a quantum memory scheme where
the inhomogeneous dephasing was controlled not by a strong
$\pi$-pulse, but by an external electric field gradient. The basic
idea is to create a narrow spectral feature using optical pumping
techniques, which is then broadened using an external field. The
scheme was coined controlled reversible inhomogeneous broadening
(CRIB), while a later modified version of the scheme was named
gradient echo memory (GEM) \cite{Hetet2008}. Several experimental
realizations of the CRIB/GEM storage scheme in rare-earth-ion
doped crystals followed \cite{Alexander2006}, \cite{Hetet2008a},
including storage of weak coherent states at the single photon
level \cite{Hedges2010}, \cite{Lauritzen2010}. It is worth noting
that the GEM experiment reported by Hedges et al.
\cite{Hedges2010} demonstrated one of the highest efficiencies reached
in any quantum memory, 69\%, and the highest reached in a
solid-state memory. In rare-earth-ion doped crystals the CRIB/GEM
experiments were implemented using external electric fields, based
on the linear Stark shift. In some systems, however, the linear
Stark shift is zero due to symmetry considerations
\cite{Chaneliere2008}.
\begin{figure}[b]
\includegraphics[scale=.60]{deRiedmatten_Fig1.pdf}
\caption{(a) The input pulse is absorbed on a strongly absorbing
transition whose inhomogeneous spectrum is shaped into a comb with
periodicity $\Delta$. After the absorption of the input, a control
pulse converts the initial optical coherence into a spin
coherence, see panel (b) for the timing. Another control pulse
applied a time $T_S$ after the first one re-establishes the
optical coherence, which evolves towards an echo emission after a
total storage time $1/\Delta+T_S$. This scheme is the complete AFC
spin-wave memory. If the control pulses are not applied, then the
input will give rise to an output echo after a total storage time
$1/\Delta$, called the AFC echo scheme. We refer to the text for
more details on the processes and the required energy structure.}
\label{fig:AFCscheme}
\end{figure}
In 2008 the atomic frequency comb (AFC) quantum memory scheme was
proposed \cite{Riedmatten2008,Afzelius2009}. The motivation for
the scheme was the storage of trains of pulses, so-called temporal
multimode storage, which turned out to be difficult using the CRIB
scheme due to the scaling of the number of modes as a function of
the optical depth of the storage material. In CRIB memories the
number of modes scales linearly with the optical depth $d$ of the
transition \cite{Simon2007a}, while for AFC it is independent of
optical depth, although depending on other critical parameters, as
will be discussed later. For electro-magnetically induced
transparency (EIT) and Raman techniques, two common light-storage
techniques, the number of modes scales as $\sqrt{d}$
\cite{Nunn2008a}.
The AFC scheme is based on a spectral tailoring of the
inhomogeneous absorption spectrum of an optical transition
$|g\rangle - |e\rangle$, where one ideally wants to make a
periodic series of narrow, highly absorbing peaks. This forms the
atomic frequency comb, which can be characterized by its spacing
$\Delta$ and finesse $F$, in analogy with an optical cavity. A
single-photon state, with a bandwidth $\gamma_p$ larger than
$\Delta$, can be completely absorbed by the comb. Indeed, if the
input pulse is short enough (meaning $\gamma_p > \Delta$), the
effect of the short interaction in time is a spectral averaging of
the sharp AFC structure into a smooth distribution, due to Fourier
arguments, allowing for uniform absorption over the photons
bandwidth. The effective optical depth of the comb is roughly
$\tilde{d}=d/F$ \cite{Afzelius2009}, depending on the shape of the
peaks, where $d$ is the peak absorption depth. The absorption
probability is $1-\exp(-\tilde{d})$, showing that complete
absorption can be achieved for high enough $d$ for any finesse
$F$. For a more complete description, we refer the reader to
\cite{Afzelius2009}.
Conditioned on the absorption of a single photon, the atomic state
can be described by a collective Dicke state $\sum_k c_k
|g\cdot\cdot\cdot e_k \cdot\cdot\cdot g\rangle$ \cite{Dicke1954},
where the amplitudes $c_k$ depend on the detuning and spatial
position of the particular atom $k$. These modes are initially in
phase, but the collective state will rapidly dephase into a
non-collective state $\sum_k \exp(-i2\pi\delta_kt) c_k
|g\cdot\cdot\cdot e_k \cdot\cdot\cdot g \rangle$, since each term
acquires an individual phase depending on the detuning $\delta_k$
of each excited atom. If we consider an AFC having very sharp
peaks, then the detunings $\delta_k$ are approximately a discrete
set such that $\delta_k = m_k\Delta$, where $m_k$ are integers. It
follows that the collective state is re-established after a time
$1/\Delta$, which leads to a coherent photon-echo
\cite{Mossberg1979,Carlson1984,Mitsunaga1991} type re-emission in
the forward spatial mode defined by the absorbed photon. We note that the AFC echo can also be interpreted as a slow-light effect induced by the comb structure \cite{Bonarota2012}.
The scheme described here can only be used as a delay-line, with a
fixed storage time $1/\Delta$. In the following we will refer to
this scheme as an \textit{AFC echo scheme}. A recent quantum
repeater protocol is entirely based, however, on this scheme, but
which requires heavy frequency and time multiplexing to be
efficient \cite{Sinclair2014}. It should also be emphasized that
the AFC echo scheme provides a dynamical delay-line, which can be
re-programmed with a rate related to the comb creation time. The
temporal multimode capacity of the AFC scheme does not depend on
optical depth, as mentioned above. The number of modes that one
can store depends simply on the number of peaks in the comb, which
in turn depends on the ratio of the comb periodicity $\Delta$ to
the total comb bandwidth $\Gamma$.
To be able to read out a AFC memory on demand, the original
proposal was based on a conversion of the optical excitation into
a spin excitation \cite{Afzelius2009}. This can be done by
applying an optical control pulse that transfers the single
optical excitation to a spin state, for instance a $\pi$-pulse,
after the absorption of the single photon, but before the
appearance of the AFC echo. This requires an additional ground
state level $|s\rangle$, such that the states
$|g\rangle$,$|e\rangle$ and $|s\rangle$ form a so-called
$\Lambda$-system. To read out the memory a second control pulse is
applied after a spin-wave storage time $T_S$, after which the
collective Dicke state continues to evolve towards the AFC echo
emission, after a total storage time $T_S+1/\Delta$. In addition
to providing on-demand read out of the memory, it can also provide
a longer total storage time, since the spin coherence time can be
orders of magnitude longer than the optical coherence time. In the
following we will refer to this scheme as a \textit{AFC spin-wave
memory}. The spin-wave storage requires a coherent spin-transition
$|g\rangle$-$|s\rangle$ for storing the spin coherence, but also
an additional state $|aux\rangle$ with long population lifetime.
The auxiliary ground state $|aux\rangle$ is used for storing
population that has been pumped away optically during the AFC
creation process. The state $|aux\rangle$ is also needed in the
AFC echo scheme described above. The need for three ground-state
levels and a coherent spin transition $|g\rangle$-$|s\rangle$
limits the number of known materials that can be used for
spin-wave storage.
The efficiency of the complete AFC spin-wave memory depends on
several factors. The most important one is the efficiency of the
AFC echo, which in turn depends on the optical depth of the
material, the comb parameters and the direction of recall. But one
also needs to consider the efficiency of the optical control
pulses and spin dephasing during the spin-wave storage time $T_S$.
In most cases these factors act independently on the total
efficiency, leading to the simple efficiency formula
\begin{equation}
\eta=\eta_{AFC} \eta_{C}^2 \eta_{S}
\end{equation}
\noindent where $\eta_{AFC}$ is the AFC echo efficiency,
$\eta_{C}$ the efficiency of one optical control pulse and
$\eta_{S}$ accounts for loss of efficiency due to spin
decoherence. We here assume that the control pulses introduce no
decoherence, we only take into account a limited transfer
efficiency of population.
The AFC echo efficiency depends on the direction of recall. In
forward direction the re-absorption effect in a optically dense
medium limits the efficiency to 54\% \cite{Moiseev2004,Sangouard2007}. In
backward recall an interference effect makes it possible to reach
100\% in principle \cite{Sangouard2007,Afzelius2009}. Backward
recall can be achieved by using counter-propagating control
pulses, but then only in spin-wave storage. The AFC echo
efficiency formulas for both cases are given below
\begin{equation}
\eta^{fw}_{AFC}=\tilde{d}^2 \exp(-\tilde{d}) \eta_{deph},
\end{equation}
\begin{equation}
\eta^{bw}_{AFC}=(1-\exp(-\tilde{d}))^2 \eta_{deph},
\end{equation}
\noindent where $\eta_{deph}$ is a dephasing factor that accounts
for the finite width and shape of the AFC teeth. It should be
emphasized that these formulas also apply to the CRIB/GEM scheme.
As shown in Refs \cite{Sangouard2007,Afzelius2009}, $\eta_{deph}$
is simply the Fourier transform of a \textit{single tooth}
function in the comb, evaluated at the time of the AFC echo
$1/\Delta$. The effective absorption depth $\tilde{d}$ also
depends on the exact shape of the AFC teeth. In Ref.
\cite{Afzelius2009} Gaussian peaks where considered, while in Ref.
\cite{Chaneliere2010} formulas were given for Lorentzian shaped
teeth. Later Bonarota \textit{et al.} \cite{Bonarota2010} showed
that square peaks give the highest efficiency for a given peak
optical depth $d$. For square-shaped peaks $\tilde{d}=d/F$ exactly
and $\eta_{deph}=\rm{sinc}^2(\pi/F)$.
The optical depth is in practice the most crucial parameter, which
led to the proposal to
put the memory in an optical cavity to enhance the effective interaction
length \cite{Moiseev2010a},\cite{Afzelius2010a}. It was shown that
the cavity could be operated in an optimal regime, where the input
mirror reflectivity $R$ of an asymmetric cavity is tuned to the
effective optical depth of the memory $\tilde{d}$ such that
$R=\exp(-2 \tilde{d})$, assuming $\tilde{d}<1$, which is called an
\textit{impedance-matched} cavity. At the impedance-match point
complete absorption can in principle be achieved, if all other
losses are much smaller than $\tilde{d}$, and the efficiency is
then bounded only by the intrinsic dephasing $\eta_{deph}$. As a
consequence the cavity approach can lead to close to 100\%
efficiency, without resorting to the phase-matching operation
required for backward recall. Note also that the impedance-matched
cavity scheme can be applied to any memory scheme based on control
of the inhomogeneous dephasing, e.g. AFC, CRIB or GEM. Recent
experimental demonstrations of the cavity scheme reached AFC echo
efficiencies with bright pulses of 56\% \cite{Sabooni2013} in
Pr$^{3+}$:Y$_2$SiO$_5$ and 53\% \cite{Jobez2014} in Eu$^{3+}$:Y$_2$SiO$_5$, the highest reported
AFC efficiencies to date.
\section{State of the art}
\label{Sec:State_of_the_art}
Although the focus of this chapter is the interaction of quantum
light with rare-earth doped solids, we first review a series of
experiments that have been performed with weak coherent states at
the single photon level. This type of experiment allows the
testing of several aspects relevant for quantum light storage, in
particular the coherence preservation, the noise added in the
storage and retrieval processes as well as the waveform
preservation. In addition, although the light at the input is
classical, it has been shown that it is possible to infer the
quantum character of the storage under certain conditions
\cite{Hedges2010,Specht2011,Gundogan2012,Sinclair2014}.
The first demonstration of storage and retrieval of light at the
single photon level in a solid state device, which was also the
first demonstration of the AFC echo scheme, was done in 2008 at
the University of Geneva \cite{Riedmatten2008}. Weak coherent
light pulses were stored for up to 1 $\mu s$ using the atomic
frequency comb scheme in a Nd$^{3+}$:YVO$_4$ crystal. Single
photon level time-bin qubits were also stored and the coherence
was shown to be preserved to a high degree during the storage and
retrieval process. Finally, a proof of principle experiment of
temporal multimodality of the protocol was done, with the storage
and retrieval of 4 temporal modes. The storage and retrieval
efficiency was $< 1 \%$ in that initial demonstration. However,
several other single photon level experiments in other materials
have since then demonstrated the AFC echo scheme at the single
photon level with much higher efficiencies, reaching 9 $\%$
efficiency in Tm$^{3+}$:YAG \cite{Chaneliere2010} and 25 $\%$ in Pr$^{3+}$:Y$_2$SiO$_5$
\cite{Sabooni2010,Maring2014}.
Another
aspect that has been improved in recent experiments is the
multimode capacity. The reversible mapping of up to 64 weak pulse
temporal modes has been demonstrated in a Nd:YSO crystal
\cite{Usmani2010}. The coherence was verified by simultaneously
storing and analyzing multiple time-bin qubits. It has also been
shown that, combined with phase modulators, AFC can be used as a
programmable processor for spectral and temporal manipulation of
single qubits \cite{Saglamyurek2014}. Hong-Ou-Mandel interference
between two AFC echoes recalled from Tm doped waveguides has been
demonstrated \cite{Jin2013}. Finally, the AFC storage was also
extended to polarization qubits
\cite{Gundogan2012,Clausen2012,Zhou2012}, and more recently to
spectrally multiplexed time-bin qubits with selective readout in
frequency \cite{Sinclair2014}. Note that for all AFC echo
experiments mentioned above, the storage was done in the excited
state only, leading to short and pre-determined storage time. The
CRIB/GEM protocol allows on demand read-out even with storage in
the excited state, as has been demonstrated at the single photon
level in an Erbium doped crystal \cite{Lauritzen2010} and in a Pr
doped crystal \cite{Hedges2010}. The latter experiment reported
the highest efficiency for any solid state memory so far (69
$\%$). It also showed that the memory operated in the quantum
regime, meaning that the storage and retrieval fidelity was
measured to be higher than the one achievable with a classical
memory.
The first demonstration of quantum light storage in solid state
device was reported simultaneously in 2011 by two groups, one from
the University of Geneva \cite{Clausen2011} and one from the
University of Calgary \cite{Saglamyurek2011}. Both experiments
demonstrated the storage of entangled photons in a rare-earth
doped crystal, using the AFC echo scheme. Both experiments created
compatible photon pair sources with one photon matching the
storage device and the other photon at telecommunication
wavelength. The two teams used different storage media with
different properties and bandwidth. The Calgary experiment used a
broadband AFC in a Tm doped Lithium Niobate waveguide absorbing
light at 793 nm. One photon of the pair was stored for 7 ns in the
waveguide, with a storage and retrieval efficiency of 2 $\%$ (
excluding coupling losses in the waveguide). It was also shown
that time-bin entanglement was preserved during the storage and
retrieval, and a violation of a Bell inequality was demonstrated
between the telecom photon and the stored and retrieved photon.
The Geneva experiment used a Nd doped crystal absorbing at 883 nm,
with a storage bandwidth of 120 MHz and featured a maximal storage
time of 200 ns and a maximal efficiency of 20 $\%$. The
preservation of energy-time entanglement was demonstrated. These
experiments will be described in more detail in section
\ref{Sec:QStorage:Nd}.
These experiments demonstrated for the first time entanglement
between a telecom photon and a collective optical atomic
excitation in a solid state device. It should be noted however
that in both cases the photons were stored as
optical atomic excitations, leading to short and only
pre-determined storage times.
Following these initial experiments, further developments by the
same groups included the storage and retrieval of polarization
\cite{Clausen2012} and time-bin qubits \cite{Saglamyurek2012}
carried by heralded single photons. The Geneva group also reported
an experiment demonstrating entanglement between two crystals
\cite{Usmani2012} (it will be described in more detail in section
\ref{Sec:QStorage:Nd}), followed by a experiment demonstrating
quantum teleportation of a telecom wavelength photon onto a
collective atomic optical excitation \cite{Bussieres2014}. Furthermore, the quantum storage of a 3-dimensional orbital-angular-momentum entangled photon has been reported in a Nd$^{3+}$:YVO$_4$ crystal by a group in Hefei \cite{Zhou2014}.
In 2014 an experiment demonstrating quantum storage of heralded
single photons using the AFC echo scheme in a Praseodymium doped
crystal absorbing at 606 nm was carried out at ICFO. This material
has demonstrated promising properties in the storage of classical
light, including long storage times up to 1 minute
\cite{Longdell2005,Heinze2013} and high storage and retrieval
efficiencies as mentioned above \cite{Hedges2010}. Contrary to
materials used in previous demonstrations, it also possesses 3
ground state levels, such that spin-wave storage is in principle
possible. However this comes with the drawback that the spacing
between the hyperfine states is small, therefore limiting the
storage bandwidth to a few MHz. This poses strong challenges for
the realization of a suitable quantum light source. The
realization of such a source \cite{Fekete2013} and the storage
experiment \cite{Rielander2014} will be described in more detail
in sections \ref{Sec:SPDCsource:Pr} and \ref{Sec:QStorage:Pr},
respectively.
While the previous experiments have been performed in rare-earth
doped crystals, a recent experiment demonstrated that it is also
possible to store non classical light states in an amorphous
environment. An erbium-doped standard telecom glass fiber was used
as storage device, and photons at telecom wavelengths were stored
using the atomic frequency comb scheme, with an efficiency around
1 $\%$ and a storage time of 5 ns \cite{Saglamyurek2015}. The
experiment also showed that entanglement was preserved during the
storage in the fiber.
Finally, we shortly mention experiments aiming at using crystals
as a source of photon pairs with embedded memory. In these
protocols, the crystal is illuminated with classical pulses,
creating non classical correlations between an emitted photon and
a stored collective atomic excitation. The protocols include the
rephasing of amplified spontaneous emission (RASE)
\cite{Ledingham2010} and a combination of the DLCZ and AFC scheme
\cite{Sekatski2011}. First demonstrations of the RASE scheme have
been realized, with strong but still classical correlations
obtained in Pr$^{3+}$:Y$_2$SiO$_5$ \cite{Beavan2012} in the photon counting
regime, and evidence of non-classical correlations obtained with
homodyne detection in Tm$^{3+}$:YAG \cite{Ledingham2012}.
\section{Quantum Light sources compatible with solid state quantum memories}
\label{Sec:SPDCsource:General}
In order to achieve strong interactions between a single photon
and a crystal, and to achieve high efficiency storage, it is
crucial that the quantum light has spectral properties that match
those of the quantum memory. The bandwidth of AFC memories is
given by the width of the AFC that can be created in the crystal.
For obtaining the high-finesse combs necessary to achieve high
efficiency storage, the width of the comb is limited by the
spacing between the adjacent states in the ground or excited state
manifolds. Note that AFC broader than the spacing between ground
and excited states can be created, however with a low finesse
leading to limited storage efficiencies, see e.g.
\cite{Saglamyurek2011,Bonarota2011}. In the case where the spacing
becomes bigger than the inhomogeneous broadening of the optical
transition, the limit is then given by the latter. This situation
could be encountered with Kramers ions (e.g. Nd$^{3+}$, Er$^{3+}$), where
moderate magnetic fields could split the states by several GHz .
In principle, high efficiency storage using the AFC echo scheme
could therefore reach GHz bandwidth. In practice however, creating
high quality combs for long storage times over such a large
bandwidth is experimentally challenging. For achieving spin-wave
storage, non-Kramers ions such as Pr and Eu are good candidates.
In these materials, the spacing between hyperfine states are much
smaller, leading to much smaller bandwidths, which can be as low
as a few tens of MHz for Eu$^{3+}$:Y$_2$SiO$_5$ and 4.6 MHz for the excited state
of Pr$^{3+}$:Y$_2$SiO$_5$. The creation of single photons with such a narrow
linewidth is experimentally challenging. Designing and
implementing narrowband quantum light sources
\cite{Ou1999,Pomarico2009,Chuu2012,Foertsch2013,Kaiser2013} that
can be interfaced with atoms has been the subject of several
investigations in recent years
\cite{Chaneliere2005,Neergaard-Nielsen2007,Bao2008,Scholz2009,Haase2009,Wolfgramm2011}.
(see also chapters (Chih-Sung Chuu, Mitchell, Zhao)).
The quantum light sources compatible with solid-state quantum
memories that have been realized up to now make use of spontaneous
parametric down conversion (SPDC), where a pump photon is
probabilistically split into a photon pair, with energy and
momentum conservation. A great advantage of this solution is that
it is very flexible in terms of wavelengths of the created
photons, which allows the coupling to quantum memories operating
at any wavelengths. Moreover, it can create two photons with
different frequencies, which can be used for example to create non
classical correlations between a quantum memory operating in the
visible range and a photon at telecommunication wavelengths.
However, the spectrum of the photons emitted by spontaneous down
conversion typically goes from 100 GHz to THz, several orders of
magnitude larger than the bandwidth of quantum memories. Therefore
extensive filtering must be applied in order to generate quantum
memory
compatible quantum light using spontaneous down conversion. \\
Filtering can be applied after the source, using passive filters.
However this requires extremely bright sources \cite{Clausen2011},
e.g. waveguide sources. The waveguide increases the production
rate of photon pairs significantly \cite{Tanzilli2001}, allowing
for an efficient source while pumping it with a low peak-power cw
laser. This is particularly important for a strongly filtered
source, in order to have a sufficiently high probability of
creating a photon pair within the filtered spectral regions. It is
important to note that when passive filtering is used, the number
of photon pairs per coherence time (or the spectral brightness
expressed in pairs per second per mW of pump power and per MHz of
bandwidth) does not change with the filter width
\cite{Halder2008}. While the rate of created pairs per second
decreases with the filter width, the coherence time of the photon
increases, leading to a constant spectral brightness.
Another way of implementing a narrow-band quantum light source
from spontaneous down conversion is to insert the non linear
crystal in an optical cavity \cite{Ou1999} (see also chapters
(Zhao,Mitchell)). This not only has the advantage of providing
filtering, but also enhances the probability of generating a
photon within a cavity mode, with respect to the no cavity case.
In the ideal case, the enhancement is given by $Q = F^3/(\pi
F_0)$ \cite{Ou1999}, where $F$ is the finesse of the cavity and
$F_0$ is the finesse calculated only from the mirrors
reflectivity. In order to reach this enhancement, both signal and
idler fields need to be resonant with the cavity.
Beyond the use of SPDC, several other systems could be used as
quantum light sources. It could be possible to use the doped
crystals themselves as sources of photon pairs with the required
spectral properties, as shown with quantum memories based on
atomic gases (see. e.g. \cite{Sangouard2011} for a review).
However, this turns out to be much more difficult to implement in
rare-earth doped solids, due to the very small oscillator strength
of the optical transition. Several schemes have been proposed,
including the rephasing of amplified spontaneous emission
(RASE)\cite{Ledingham2010} and a combination of the DLCZ and AFC
scheme \cite{Sekatski2011}. Solid state single photon emitters may
also be used as compatible single photon sources, e.g. quantum
dots or single molecules. Quantum dots have typically GHz
spectral bandwidths (see chapters Lanco, Schneider,McMahon), which may be compatible with the AFC
echo scheme with broadband combs. The challenge is to tune the
quantum dot in resonance with the rare-earth ensemble, without
broadening the line. Single molecules in solid state matrices have
shown much narrower spectral bandwidth, down to tens of MHz
\cite{Lettow2010}, which could be used with spin-wave quantum
memories. Again, the challenge is to tune them near resonance. A
potential solution for the frequency mismatch is to implement
quantum frequency conversion \cite{Tanzilli2005} which has been
used recently to interface telecom photons to quantum memories
\cite{Albrecht2014,Maring2014}.
\subsection{Characterizing photon pair
sources}
In this section, we discuss various ways to characterize photons
pairs emitted by SPDC and to quantify the correlations between
signal and idler fields. The state created by single mode SPDC is
given by (for $p\ll 1$):
\begin{equation}
\label{Eq:TMSS}
|\Psi\rangle_{s,i}=\sqrt{1-p}\sum_{n=0}^{\infty}p^{\frac{n}{2}}|n\rangle_s|n\rangle_i
\end{equation}
where $p$ is the probability to create a photon pair and is
proportional to the pump power, and $|n\rangle_s (|n\rangle_i)$ is
a n photon Fock state in the signal (idler) mode.
This state is known as two-mode
squeezed state. It displays very strong quantum correlations
between the two modes, i.e. the signal and idler fields.
For a photon pair source, the quality of the correlations between
signal and idler fields is usually quantified measuring the second
order cross-correlation function $G_{s,i}^{(2)}(\tau)$ between the
two fields, by performing a coincidence measurement. The
normalized form of $G_{s,i}^{(2)}(\tau)$, denoted as
$g_{s,i}^{(2)}(\tau)$ can be expressed as:
\begin{equation}
g_{s,i}^{(2)}(\tau) \equiv \frac{\langle E_s^{\dagger}(t)
E_i^{\dagger}(t+\tau) E_i(t+\tau) E_s(t) \rangle}{\langle
E_i^{\dagger}(t+\tau) E_i(t+\tau) \rangle
\langle E_s^{\dagger}(t) E_s(t) \rangle}, \\
\end{equation}
where $E_{s,i}^{\dagger}$ ($E_{s,i}$) is the electric field
creation (annihilation) operator for the signal and idler fields.
For the ideal two mode squeezed state of Eq. \ref{Eq:TMSS}, the
cross-correlation function is given by :
\begin{equation}
g^{(2)}_{s,i}=1+\frac{1}{p}
\end{equation}
We see that the correlation decreases when increasing $p$. This is
the consequence of the creation of multiple pairs. We also see
that $g^{(2)}_{s,i}$ can become arbitrarily high for low
excitation probability. However, this comes at the expense of the
count rate, since the mean number of photons in the signal mode is
given by
\begin{equation}
\bar{n}_s=\frac{p}{1-p}
\end{equation}
This illustrates a fundamental limitation of SPDC: there is a
trade off between the degree of quantum correlation and the count
rate that can be obtained. In practice, $g^{(2)}_{s,i}$ is
determined over a detection
window
$\Delta \tau$ as :
\begin{equation}
g^{(2)}_{s,i}(\Delta \tau)=\frac{p_{s,i}}{p_sp_{i}}
\end{equation}
where $p_{s,i}$ is the probability to detect a coincidence between
signal and idler photons and $p_s$ ($p_{i}$) is the probability to
detect a signal (idler) photon in a time interval $\Delta \tau$.
The second order cross-correlation function can also be used to
gain information on the spectral content of the photon pairs. For
example, in the case of doubly-resonant cavity-enhanced
downconversion, following the theory used in
\cite{Scholz2009,Wolfgramm2011}, $g_{s,i}^{(2)}(\tau)$ takes the
form:
\begin{eqnarray}
\begin{split}
g_{s,i}^{(2)}(\tau) \propto {} & \Bigg| \sum_{m_s, m_i = 0}^\infty \frac{\sqrt{\gamma_s \, \gamma_i \, \omega_s \, \omega_i}}{\Gamma_s + \Gamma_i} \\
& \times
\begin{cases}
e^{-2 \pi \Gamma_s (\tau-(\tau_0/2))}{\rm sinc}{(i \pi \tau_0 \Gamma_s)} \hspace{3 mm} \hspace{0.5 mm} \tau \geq \frac{\tau_0}{2}\\
e^{+2 \pi \Gamma_i (\tau-(\tau_0/2))}{\rm sinc}{(i \pi \tau_0 \Gamma_i)} \hspace{4 mm} \hspace{0.5 mm} \tau < \frac{\tau_0}{2}
\end{cases} \hspace{-4 mm} \Bigg|^2,
\end{split}
\label{eq:G2}
\end{eqnarray}
where $\gamma_{s,i}$ are the cavity damping rates for signal and
idler, $\omega_{s,i}$ are the central frequencies,
$\Gamma_{s,i} = \gamma_{s,i}/2+i m_{s,i} \rm{FSR}_{s,i}$ with mode indices $m_{s,i}$ and
free spectral ranges $\rm{FSR}_{s,i}$, and $\tau_0$ is the transit
time difference between the signal and idler photons through the
SPDC crystal. Eq. \ref{eq:G2} shows that the second order
cross-correlation function for a multimode cavity output displays
an oscillatory behavior with peaks separated by the inverse of the
cavity free spectral range. The width of the peaks is directly
related to the number of spectral modes in the signal and idler
fields. For two modes, the oscillation will be sinusoidal, and the
width will then decrease with the number of modes. In practice,
the minimum width that can be detected is given by the detectors'
time resolution.
For the single mode case, we have :
\begin{equation}
g_{s,i}^{(2)}(\tau)=1+\frac{4}{p}\frac{\gamma_s\gamma_i}{(\gamma_s+\gamma_i)^2}f(\tau)
\end{equation}
where
\begin{eqnarray}
\begin{split}
f(\tau)=
\begin{cases}
e^{-\gamma_s \tau} \hspace{3 mm} \hspace{0.5 mm} \tau \geq 0\\
e^{\gamma_i \tau} \hspace{4 mm} \hspace{0.5 mm} \tau < 0
\end{cases}
\end{split}
\end{eqnarray}
This is the same expression obtained for the case without cavity,
but with single mode Lorentzian filters with width $\gamma_s$ and
$\gamma_{i}$ inserted in the signal and idler modes, respectively
\cite{Clausen2014a}.
Information about the spectral content of the created photons can
also be obtained by measuring the unconditional second order
autocorrelation function $g^{(2)}_{s,s}$ and $g^{(2)}_{i,i}$, for
the signal and idler fields, respectively. For an ideal two mode
squeezed states, the unconditional field exhibit thermal
statistics with $g^{(2)}_{s,s}(0)= g^{(2)}_{i,i}(0)=2$. However,
if several spectral modes are present, it has been shown
\cite{Christ2011} that the autocorrelations decrease as
$g^{(2)}_{s,s}(0)= g^{(2)}_{i,i}(0)=1+1/K$ where $K$ is the number
of modes. By measuring the unconditional second order
autocorrelation function, one can therefore bound the number of
spectral modes present in the fields. This is valid as long as the
measurement is not limited by noise.
The non-classical nature of the correlation between signal and
idler fields can be experimentally assessed with a Cauchy-Schwarz
inequality. For a pair of independent classical fields the
following inequality must be fulfilled:
\begin{equation}
R=\frac{(g^{(2)}_{s,i})^2}{g^{(2)}_{s,s}g^{(2)}_{i,i}}\leq1
\end{equation}
If the signal and idler fields exhibit thermal or sub-thermal
statistics ($g^{(2)}_{s,s}(0)= g^{(2)}_{i,i}(0)\leq 2$, the
measurement of $g^{(2)}_{s,i}>2$ is therefore a signature of
non-classical correlations. However, in order to prove
non-classicality without assumptions on the created state, the
unconditional autocorrelation functions should be measured as
well.
\subsection{A quantum light source compatible with Nd doped crystals}
\label{Sec:SPDCsource:Nd}
\begin{figure}[b]
\includegraphics[scale=.40]{deRiedmatten_Fig2.pdf}
\sidecaption[t]
\caption{Second-order cross-correlation $g^{(2)}_{s,i}(\Delta \tau)$
function of the filtered SPDC source developed for Nd doped
crystals, as a function of the power of the laser pumping the SPDC
source. The detection integration window was $\Delta \tau$ = 10
ns.}
\label{fig:source_Nd_corr_funcs}
\end{figure}
Here we discuss a filtered SPDC source producing photons at 883 nm
(signal) and 1338 nm (idler), which was developed to interface
with a Nd$^{3+}$:Y$_2$SiO$_5$ quantum memory operating at 883 nm,
having a memory bandwidth of 120 MHz. This particular source has
been described in Refs \cite{Clausen2011,Usmani2012}, while a
similar source, slightly more broadband, of polarization-entangled
photons for quantum storage was described in a more recent work
\cite{Clausen2014a}.
The SPDC source was based on a periodically-poled potassium
titanyl phosphate (PPKTP) crystal with an optical waveguide. The
PPKTP crystal was pumped by a continuous-wave (cw) 532 nm laser,
which is convenient since powerful and frequency stable
single-mode Nd:YAG lasers exist at this wavelength. Since one
photon should be resonant with the 883 nm transition in
Nd$^{3+}$:Y$_2$SiO$_5$, the choice of the pump laser imposed the
wavelength of the idler photon to 1338 nm, in the
telecommunication O-band. In principle one could use another pump
wavelength to produce an idler photon in the more conventional
telecommunication C-band around 1550 nm.
Without external frequency filtering the source produced photons
with a bandwidth of about 800 GHz, such that strong filtering of
the photon pairs was necessary to match the memory bandwidth of
120 MHz. To this end a combination of diffraction gratings,
optical cavities and a fibre-based filter was used. The gratings
provided full-width at half-maximum (FWHM) bandwidths of 90 and 60
GHz for the signal and idler photons, respectively. For the signal
photon two etalons placed in series resulted in a single
longitudinal mode with a bandwidth of 350 MHz. It should also be
emphasized that the 6 GHz wide inhomogeneous absorption profile of
the Nd$^{3+}$:Y$_2$SiO$_5$ crystal provided additional filtering,
since the 120 MHz comb was created within this absorption profile.
For the idler photon a home-made narrow-band cavity filtered down
the photons to a FWHM linewidth of 43 MHz. A single longitudinal
mode of this cavity was selected by a fiber Bragg grating (FBG).
The total transmission coefficients from the PPKTP waveguide to
the single-mode fibers were 22\% and 14\% for signal and idler
photons \cite{Usmani2012}, respectively, including fiber coupling.
The more recent version of this source \cite{Clausen2014a} reach
higher efficiencies, partly because high-efficiency volume Bragg
gratings (VBGs) replaced the diffraction gratings.
Strong non-classical correlations between the signal and idler photons can only be obtained if the central frequencies of the filters on each mode satisfy the energy conservation of the SPDC processes. This is a non-trivial task when dealing with highly non-degenerate SPDC sources, particularly when one mode must be resonant with an external quantum memory. A solution to this problem was introduced in the work discussed here. A reference laser that is resonant with the signal filtering system and the quantum memory is injected into the SPDC source. This will create light at the idler wavelength through difference frequency generation (DFG), which obey the necessary energy conservation. This can be used to adjust central frequency of the idler filtering system, alternatively one can also change the wavelength of the pump laser.
The filtered SPDC source can be characterized by the pair creation
rate within the filtered modes (spectral brightness) and the
second-order auto and cross correlation functions of the signal
and idler modes. The intrinsic spectral brightness of the PPKTP
waveguide was 6.3$\cdot$10$^3$ pairs/(mW $\cdot$ MHz $\cdot$ s)
\cite{Usmani2012}, which does not include the transmission through
the filtering elements. This means that the probability $p$ to
create a pair (in the limit $p \ll 1$) is then $p \approx 2.7
\cdot 10^{-3}$/mW per 10 ns within a 43 MHz wide spectral window.
The duration of the detection window should be set with respect to
the photon pair coherence time, which in our case is dominated by
the 43 MHz filter on the idler side (7 ns coherence time). The
spectral brightness of the source including the filtering elements
and fiber coupling was $\approx$ 200 pairs/(mW $\cdot$ MHz $\cdot$
s). If we again assume a 43 MHz filtered bandwidth we arrive at a
final pair creation rate of 8600 pairs/s per mW pump power.
The second-order auto-correlation functions of both the idler and
signal modes were characterized \cite{Usmani2012}, resulting in
$g_{i,i}^{(2)}(0)=1.9$ and $g_{s,s}^{(2)}(0)=1.8$. These values
give effective mode numbers $K=1.1$ and $K=1.25$ for the idler and
signal modes, which indicate single frequency modes on both the
signal and idler side.
In Figure \ref{fig:source_Nd_corr_funcs} we show the second-order
cross-correlation function $g_{s,i}^{(2)}$ of the source as a
function of pump power which exhibits the expected $1+1/p$
behaviour \cite{Clausen2011}. Based on the intrinsic spectral
brightness we expect a cross-correlation $g_{s,i}^{(2)} \approx
75$ for $P=5$ mW, which is very close to the measured value of
about 80. The cross-correlation function is larger than the
classical upper bound of 2, where one assumes
$g_{i,i}^{(2)}(0)=g_{s,s}^{(2)}(0)=2$, for all pump powers,
clearly demonstrating the strong non-classical correlations of the
source.
\subsection{A quantum light source compatible with Pr doped
crystals} \label{Sec:SPDCsource:Pr} In the context of generating
quantum light compatible with Pr$^{3+}$:Y$_2$SO$_5$, one of the
photons must be at 606 nm, with a bandwidth smaller than 4 MHz. In
addition, in order to use this source to generate entanglement
between remote crystals, it is desirable to have the second photon
of the pair at a telecom wavelength. An experiment showing these
properties was reported in \cite{Fekete2013}.
\begin{figure}[h]
\includegraphics[scale=.45]{deRiedmatten_Fig3.pdf}
\caption{Measured second-order cross-correlation $G^{(2)}_{s,i}(\tau)$
function (non normalized) of the cavity enhanced SPDC source
developed for Pr doped crystals \cite{Fekete2013}. The FWHM correlation time is 104
ns. The zoom displays a higher temporal resolution, allowing to
observe the oscillations in the $G^{(2)}_{s,i}(\tau)$,
characteristic of multi spectral mode output. From this data, a
number of four mode per cluster is inferred.}
\label{fig:source_Pr}
\end{figure}
In order to meet these requirements, the following source was
used. A pump laser at 426 nm was used to generate widely non
degenerate photon pairs at 606 nm and 1436 nm in a periodically
poled lithium niobate crystal. The crystal was placed in a bow
tie optical cavity with a free spectral range of 400 MHz and a
finesse of around 200 for the signal and idler modes. A special
coating was used to achieve high reflectivity (99.99 $\%$) for
three mirrors, and a specified reflectivity of 98.5 $\%$ for the
output coupler. The SPDC output is modified by the cavity modes.
In the case of degenerate photons with the same polarization, the
spectrum would be a convolution between the phase matching
bandwidth and the cavity modes. However, in the non degenerate
case, the signal and idler modes experience different dispersion
characteristics in the crystal, which leads to slightly different
free spectral ranges. Since both modes have to be resonant with
the cavity to enable enhancement, only a subset of modes will be
enhanced. These modes are grouped in so called clusters
\cite{Pomarico2009,Pomarico2012}. The width and spacing of the
clusters is determined by the dispersion properties of the
crystals and the cavity geometry.
In the experiment of Ref. \cite{Fekete2013}, the number of modes
per cluster was inferred to be around 4, by looking at the
oscillations in the $g_{s,i}^{(2)}(\tau)$ function and comparing
to Eq. \ref{eq:G2} (see Fig. \ref{fig:source_Pr}). This number was
also verified by direct measurement of the signal and idler fields
with a narrow band filter cavity, scanned over the spectrum. The
number of clusters was then inferred by measuring the first order
correlation function of the idler field. Within the phase-matching
bandwidth of the free space down conversion photons of around
$\Gamma_{pm}$= 80 GHz, the spectrum of the photons leaving the
cavity was finally inferred to be composed of one main cluster
containing around four longitudinal modes, and two smaller side
clusters separated by 45 GHz from the main one and suppressed by
around 80 $\%$. For degenerate photons, the number of longitudinal
modes would be given by $\Gamma_{pm}/FSR$ = 200 modes. The
clustering effect therefore leads to a suppression of the number
of modes of around 50. Within a longitudinal mode, the photons
pair created featured a correlation time of 104 ns (see Fig.
\ref{fig:source_Pr}), the longest demonstrated so far with SPDC
sources. From the decay of the $G^{(2)}_{s,i}(\tau)$ (fitted with
exp($-2\pi\Delta\nu\tau$)), spectral-linewidth of $\Delta \nu$ of
1.7 MHz and 2.9 MHz were inferred for the idler field at 1436 nm
and the signal field at 606 nm, respectively. The asymmetry can be
attributed to the different intra-cavity losses for the two
wavelengths. A intrinsic spectral brightness of 8$\cdot$10$^3$
pairs/(mW $\cdot$ MHz $\cdot$ s) was inferred, leading to $p
\approx 6.4 \cdot 10^{-3}$/mW per 400 ns within a 2 MHz wide
spectral window. However due to large optical losses, the spectral
brightness before the detectors (inside single mode fibers) was 11
pairs/(mW $\cdot$ MHz $\cdot$ s). In a more recent version of the
source, this number was increased to 190 pairs/(mW $\cdot$ MHz
$\cdot$ s).
This source still requires additional filtering for selecting a
single frequency mode. However, the filtering requirement are
considerably relaxed due to the low number of modes present in the
spectrum. Side clusters can easily be removed by placing etalons
with high transmission efficiency in the signal or idler modes. In
order to select a single mode, a narrow band filter cavity with
FSR = 16.8 GHz and linewidth of 80 MHz has been placed in the
signal arm. In this configuration, the measured
$G_{s,i}^{(2)}(\tau)$ was measured and no oscillation was
observed, confirming that only one mode per cluster was present.
The suppression of the side cluster could also be inferred first
order autocorrelation.
\section{Quantum Light Storage experiments}
\label{Sec:QStorage}
\subsection{Quantum entanglement storage in Nd:YSO crystals}
\label{Sec:QStorage:Nd} In 2008 the first AFC echo storage
experiment in a Nd$^{3+}$-doped YVO$_4$ crystal at the single
photon level was demonstrated \cite{Riedmatten2008}, which was the
starting point for considering storing true quantum states of
light. To this end the filtered SPDC source described in section
\ref{Sec:SPDCsource:Nd} was developed. The first Nd:YVO$_4$ memory
had a bandwidth of a few MHz, making the filtering of the SPDC
source difficult. A more wideband quantum memory was therefore
developed in a Nd$^{3+}$-doped crystal Y$_2$SiO$_5$, using the
$^4$I$_{9/2}-^4$F$_{3/2}$ transition at 883 nm
\cite{Usmani2010,Clausen2011}. This optical transition has an
inhomogeneous broadening of about 6 GHz. The ground state was
split into a Kramers spin doublet by applying a magnetic field of
around 300 mTesla, producing a Zeeman split of about 11 GHz. The
comb structure was created on one of the Zeeman transitions by
performing spectral hole burning, i.e. ions were optically pumped
into the other Zeeman spin state. In practice this was done by
scanning the laser frequency of a narrow-band external-cavity
diode laser (EDCL) with a acousto-optic modulator (AOM), while
periodically switching off and on the light. The total bandwidth
was 120 MHz, limited by the scan range of the AOM. The maximum
efficiency at low storage times $1/\Delta$=25 ns was 20\%, close
to the optimal value for the optical depth of the crystal. In
Figure \ref{fig_histNd} we show an example of storage of 883 nm
signal photons produced by the filtered SPDC source in the crystal
\cite{Clausen2011}. The cross correlation function between the
signal and idler modes after storing the signal mode clearly shows
quantum correlations, for all storage times. This can be
interpreted as storing a single photon at 883 nm, heralded by the
detection of an idler photon.
The second-order cross-correlation function $g^{(2)}_{s,i}(\Delta
\tau)$ after storage of the signal photon is lower than that of
the source, cf. Figure \ref{fig:source_Nd_corr_funcs}. In
\cite{Usmani2012} it was shown that it was due to the delay
introduced by the memory. Indeed, the signal photons released from
the memory are superimposed with uncorrelated signal photons
created later during the AFC echo emission, which are transmitted
through the memory with a certain probability. This noise source
depends on the ratio of the memory transmission to the memory
efficiency \cite{Usmani2012}. A solution to this problem is to
turn off the SPDC pump laser before the AFC echo, which was done
in the quantum storage experiment in Pr$^{3+}$:Y$_2$SiO$_5$ described below
\cite{Rielander2014}.
\begin{figure}[th]
\includegraphics[scale=.50]{deRiedmatten_Fig4.pdf}
\caption{Detection event histogram as a function
of time. The peak at around zero time was recorded by making a
large transparency hole in the Nd$^{3+}$:Y$_2$SiO$_5$ absorption,
therefore it represents the input mode. When creating an AFC with
periodicity $\Delta$ = 20 MHz, the expected AFC echo appears as a
peak at around 50 ns. The excellent signal-to-noise ratio
indicates a strong quantum correlation between the idler photon
(1338 nm) and the signal photon (883 nm), after the latter has
been stored in the memory. The inset quantifies this by showing
the second-order cross-correlation $g^{(2)}_{s,i}(\Delta \tau)$
function as a function of storage time $1\Delta$. The
$g^{(2)}_{s,i}(\Delta \tau)$ function is significantly larger than
the classical limit of 2, for all storage times. The detection
integration window was $\Delta \tau$ = 10 ns.}
\label{fig_histNd} %
\end{figure}
The photons produced by a SPDC source pumped by a CW pump laser
are entangled in energy-time \cite{Franson1989}, provided that the
coherence time of the pump is significantly longer than the
coherence time of the signal-idler pair. This condition is
naturally met when using a single-mode frequency CW pump laser, as
in this case. This energy-time entanglement can be revealed by
making projective measurements on different time basis states, on
both the signal and idler modes. Practically, one can place
Mach-Zehnder (MZ) interferometers in each mode, each MZ having the
same path difference $\Delta T$, which is known as a Franson-type
set-up \cite{Franson1989}. The two-photon interference fringes
obtained when varying the phases in each interferometer reveals
the energy-time entanglement. The coincidence rate varies like
$V\cos(\Delta \phi_s+\Delta \phi_s)$, where $\phi_s$ and $\phi_i$
are the phase settings on the signal and idler interferometers and
$V$ is the visibility. More formally the quantum entanglement can
be detected by a violation of the Clauser-Horne-Shimony-Holt
(CHSH) inequality \cite{Clauser1969}, where the CHSH parameter $S$
is larger than 2 for any entangled state. The presence of
entanglement can also be inferred from a fringe visibility larger
than $1/\sqrt{2} \approx 70.7\%$.
\begin{figure}[t]
\includegraphics[scale=.6]{deRiedmatten_Fig5.pdf}
\sidecaption[t]
\caption{Number of detected signal-idler coincidences as a function of the signal analyser setting $\Delta \phi_s$, for two settings $\Delta \phi_i$ of the idler analyser shown as square and circle symbols. The solid ($V=78 \pm 4 \%$) and dashed ($V=84 \pm 4 \%$) lines are fits to the circles and squares, respectively.}
\label{fig:energy-time-fringes}
\end{figure}
In the experiment we describe here \cite{Clausen2011}, a
particular twist was introduced. While the idler photon was
analyzed using a standard interferometer, the signal photon was
analyzed inside the Nd$^{3+}$:Y$_2$SiO$_5$ memory. We exploited
the fact that more complicated absorption features than periodic
combs can be created. Indeed, by creating an absorption structure
that was the sum of two combs with different periodicities
$\Delta_1$ and $\Delta_2$, we could create an effective unbalanced
interferometer within the crystal. By setting
$1/\Delta_1-1/\Delta_2 = \Delta T$ we could analyze the
entanglement \textit{in memory}.
In Figure \ref{fig:energy-time-fringes} we show examples of
interference fringes as a function of $\Delta \phi_s$ for two
values of $\Delta \phi_i$. Both visibilities are well above the
limit $\approx 70.7\%$, strongly indicating the presence of
entanglement. By explicitly measuring a CHSH parameter of $S =
2.64 \pm 0.23$ the entanglement between the idler photon and the
photon stored in the crystal was clearly demonstrated.
In parallel to the work described above, Saglamyurek \textit{et
al.} also demonstrated storage of a photon entangled with a
another photon stored in and released from a rare-earth doped
crystal \cite{Saglamyurek2011}. In their experiment the entangled
photons were produced by a bulk periodically poled lithium niobate
(PPLN) crystal, which was pumped using a pico-second laser with
high peak power. Time-bin entangled photons were produced by
splitting the 16 ps pump pulse into two coherent pulses separated
by 1.4 ns. The idler photon was at the telecom wavelength 1532 nm,
while the signal photon was at 795 nm. The quantum memory was
based on a atomic frequency comb created on the
$^3$H$_6$-$^3$H$_4$ transition at 795 nm in a Thulium-doped
lithium niobate crystal (Tm$^{3+}$:LiNbO$_3$). A particular
feature of this experiment was the optical waveguide on the
surface of the Tm$^{3+}$:LiNbO$_3$ chip. The waveguide technology
is widespread in integrated optics and opens up interesting
perspectives of combining the storage device with other optical
elements. The AFC memory in this experiment was also very
broadband, the AFC spanned a range of 5 GHz. This allowed storage
of the short 16 ps photons produced by the filtered SPDC source.
This also implies that the degree of filtering was less stringent
than in the experiment described previously. A drawback of this
material is, however, the short storage time and the low
efficiency. The entanglement experiment discussed here
demonstrated a storage time of 7 ns, with an AFC echo efficiency
of 2\%. The preservation of entanglement was proved by testing the
Clauser-Horne-Shimony-Holt(CHSH) inequality, resulting in a CHSH
parameter of S=2.25 $\pm$ 0.06.
In the next experiment that we will describe the goal was to
entangle two Nd$^{3+}$:Y$_2$SiO$_5$ crystals \cite{Usmani2012}.
This could be done by storing each photon out of an entangled
pair, but this requires that both photons are resonant with a
memory. With only one photon in resonance with the Nd doped
crystal at 883 nm, the choice was made instead to store a
path-entangled state of a single photon \cite{Enk2005}. More
specifically, this state can be created by sending a single photon
state $|1\rangle$ through a balanced beam-splitter, which creates
the state
$1/\sqrt{2}(|1\rangle_A|0\rangle_B+|0\rangle_A|1\rangle_B)$ of the
spatial output modes A and B of the beam-splitter. This state is
sometimes referred to as single-photon entanglement
\cite{Enk2005}. Now, entanglement between two memories can be
realized by placing one memory in each path. This approach was
first used by Choi \textit{et al.} \cite{Choi2008} to entangle two
spatial modes in the same cloud of laser-cooled caesium atoms,
where each mode was stored in the cloud using EIT. In the
experiment described here, this this approach was used to entangle
the modes of two physically distinct crystals and the modes where
stored using the AFC scheme \cite{Usmani2012}. The single photon
state was produced by detecting the idler photon from the filtered
SPDC source, which creates a state very close to a single photon
in the signal mode (a heralded single-photon source).
To characterize the single-photon entanglement one can read out
the memories and then performing measurements on the photonic
state, which provides a lower bound of the entanglement present
while storing the two modes. To detect the entanglement we used
the tomographic approach developed by Chou \textit{et al.}
\cite{Chou2005}, in which the entanglement is quantified through a
single parameter, the concurrence $C$. The concurrence is positive
$C>0$ for an entangled state, more specifically a separable state
gives $C=0$ and a maximally entangled state $C=1$. To compute the
concurrence one needs to measure the probability of finding
exactly one signal photon in any of the two modes, $p_{01}$ and
$p_{10}$, and the probability of finding one signal photon in each
path, $p_{11}$. This requires two- and three-fold coincidence
measurements, respectively, since one also needs to detect the
heralding idler photon. In addition one needs to measure the
one-photon visibility $V$ associated with the two paths A and B.
The concurrence can then be calculated with the formula:
$C=V(p_{10}+p_{01})-\sqrt{2p_{00}p_{11}}$ \cite{Chou2005}. By
performing all the necessary measurements for a relatively high
pump power of 16 mW, we reached a concurrence of $C=6.3 \pm 3.8
\cdot 10^{-5}$ indicating presence of entanglement
\cite{Usmani2012}. The low concurrence was essentially due to
losses, since the detection probabilities include all propagation
losses, memory efficiency and detection probabilities.
Alternatively one can use an approach where the $p_{11}$
probability is inferred from the cross-correlation function
$g^{(2)}_{s,i}$, which is a much more time efficient experiment as
it uses only two-fold detections. Using this approach the presence
of entanglement at the pump power of 16 mW could be confirmed, and
the concurrence was measured for a range of lower pump powers
where the direct method turned out to be too time consuming due to
the rarity of three-fold coincidences. The concurrence reached
about $C=1.1 \pm 0.1 \cdot 10^{-4}$ for the lowest pump power of 1
mW, showing the presence of entanglement of the stored modes in
the two Nd$^{3+}$:Y$_2$SiO$_5$ crystals.
\subsection{Quantum Storage of heralded single photon in a Pr$^{3+}$:Y$_2$SiO$_5$ crystal}
\label{Sec:QStorage:Pr} As mentioned previously, Pr doped solids
have demonstrated exceptional properties for light storage
experiments, including long storage times
\cite{Longdell2005,Lovric2013,Heinze2013} and high efficiencies
\cite{Hedges2010,Sabooni2013}.
Despite these very promising properties, there is currently only
one demonstration of storage of quantum light in this system. In
this section, we describe in more detail this demonstration,
initially reported in \cite{Rielander2014}.
\begin{figure}[t]
\includegraphics[scale=.45]{deRiedmatten_Fig6.pdf}
\caption{ Exemple of an atomic frequency comb in Pr$^{3+}$:Y$_2$SiO$_5$ . The optical depth ($d$) is plotted as a function of the relative frequency. The input and output photons are resonant with the $\pm 1/2_g$ $\rightarrow$ $\pm 3/2_e$ transition. Inset : Relevant energy level scheme of the Pr$^{3+}$:Y$_2$SiO$_5$ crystal, with 3 hyperfine ground states and 3 excited states. }
\label{AFC}
\end{figure}
The ultra-narrowband photon pair source described in section
\ref{Sec:SPDCsource:Pr} can be used to generate heralded single
photons compatible with the Praseodymium doped crystal. For
sufficiently high correlations between the two fields, the
detection of an idler telecom photon will indeed project the
signal mode in a single photon Fock state.
These heralded single photons have been stored in the Pr$^{3+}$:Y$_2$SiO$_5$ as
collective optical atomic excitations using the AFC echo scheme.
The storage device is a 3 mm thick Y$_2$SiO$_5$ sample doped with
a Pr$^{3+}$ concentration of 0.05 $\%$. The relevant optical
transition connects the $^3H_4$ ground state to the $^1D_2$
excited state at a wavelength of 605.977 nm and features a
measured absorption coefficient of 23 cm$^{-1}$, and an
inhomogeneous linewidth of 5 GHz. At zero magnetic field, the
ground state and excited states manifolds are split in 3
metastable states, denoted $\pm 1/2_k$, $\pm
3/2_k$ and $\pm 5/2_k$ (see inset of Fig. \ref{AFC}), where
$k={g,e}$ denotes the ground or the excited state. The spacing
between the hyperfine states is of the order of a few MHz. This
gives an upper limit for the bandwidth as already mentioned. But,
since the separation between ground states is much smaller than
the inhomogeneous broadening, it also creates a complication for
isolating a single class of atoms. A laser sent in the crystal can
indeed be resonant with up to 9 distinct classes of atoms, within
the inhomogeneous broadening. To select a single class of atoms,
the optical pumping scheme first demonstrated by Nilsson et al was
used \cite{Nilsson2004}. After selecting one class of atoms, an
atomic frequency comb is created on the $\pm 1/2_g$ $\rightarrow$
$\pm 3/2_e$ transition (see Fig. \ref{AFC} for an example of AFC).
One important issue to couple photons to the solid state atomic
ensemble is that the frequency of the photons must be stable
within a few hundred kHz. This can be insured by implementing a
feedback lock system using the laser used to prepare the atomic
frequency comb. One the one hand, the length of the cavity is
locked on the laser with a Pound-Drever-Hall scheme. This insures
that at least one spectral mode emitted by the cavity enhanced
source is resonant with the crystal absorption. On the other hand,
the required double resonance for signal and idler modes is
insured by adjusting the frequency of the pump laser using a
classical signal at the idler wavelength created by difference
frequency generation between the pump laser at 426 nm and the
laser at 606 nm.
In the experiment, a single longitudinal mode was selected for the
idler mode thanks to a Fabry Perot filter cavity. For the heralded
single photon, the crystal itself was used as a filter to prevent
the modes non resonant with the AFC to reach the detector. The probability to have
a single photon in the signal mode before the cryostat conditioned on a detection in
the idler mode (called the heralding efficiency $\eta_H$) was around a few percents in this experiment, limited by dark counts in the idler detector, noise in the idler mode, cavity escape efficiency and optical losses in the signal mode from the cavity to the crystal.
The heralded single photon at 606 nm was first characterized by
measuring the second order correlation $g^{(2)}_{s,i}(\Delta
\tau)$ function by sending it through the crystal where a 12 MHz
wide transparency window was created. In Fig. \ref{fig_g2Pr}b, the
values of $g^{(2)}_{s,i}(\Delta \tau)$ for the incoming photons
are plotted as a function of the pump power. The value of
$g^{(2)}_{s,i}(\Delta \tau)$ increases when the pump power
decreases, as expected for a two-mode squeezed state. The non
classicality of the input photon was also demonstrated by
violating a Cauchy-Schwarz inequality. The detected coincidence
count rate was around 0.8 Hz per mW of pump power. By correcting
for known optical losses and detection efficiencies, a creation
rate outside the cavity of 2.8 kHz /mW was inferred.
\begin{figure}
\includegraphics[scale=.45]{deRiedmatten_Fig7.pdf}
\caption{ Results of heralded single photon
storage in a Pr$^{3+}$:Y$_2$SiO$_5$ crystal \cite{Rielander2014}. (a) $G^{(2)}_{s,i} (t)$ histogram
without (blue) and with (red) AFC. The preprogrammed storage
time is $2\,\mathrm{\mu s }$ and the power of the
$426.2\,\mathrm{nm}$ pump is $2\,\mathrm{mW}$. (b) The $g^{(2)}_{s,i}$ values as a function of
the pump power for the AFC echo (plain squares) are compared to
those for the input photons (plain circles).
The dotted line corresponds to the
classical limit $g^{(2)}_{s,i} = 2$ for two-mode squeezed states. (c) Storage and retrieval efficiency (blue circle) and $g^{(2)}_{s,i}$ (green squares) as a function of the storage time. For (b) and (c) the data are evaluated for a detection window $\Delta \tau$ = 400 ns and the error bars are evaluated from the raw number of counts assuming Poissonian statistics.}
\label{fig_g2Pr}
\end{figure}
After having confirmed the non classicality of the input light,
the heralded single photon was stored in and retrieved from the
crystal using atomic frequency comb. In order to avoid the
spurious noise effect described in section \ref{Sec:QStorage:Nd},
the photon pair source pump light was switched off after detection
of an idler photon. The efficiency of the storage and retrieval
process was measured to be up to 10 $\%$ for short storage times.
The second order cross correlation is measured for the stored and
retrieved photon as can be seen in Fig. \ref{fig_g2Pr}a. The values
of $g^{(2)}_{s,i}$ are also plotted as a function of the pump
power in Fig. \ref{fig_g2Pr}b. Surprisingly, the values of $g^{(2)}_{s,i}$ after retrieval
are higher than the one for the input photons, for a big range of
pump powers. This effect has been attributed to the fact that the
memory act as a filter for broadband noise emitted by the photon
pair source. Since the pump light is switched off after the
detection of an idler photon, the atomic frequency comb delays the
signal photon in a noise free region, which therefore increases
the signal to noise ratio and the $g^{(2)}_{s,i}(\Delta \tau)$
\cite{Rielander2014}. This effect highlights that under certain
conditions, quantum memories can act as purifiers by storing only
the signal and not the noise. It has been observed also in other
experiments \cite{McAuslan2012,Maring2014}.
The storage time in the crystal can be chosen by tuning the comb
periodicity $\Delta$. In theory, the minimal $\Delta$ achievable
is given by $2\gamma_hF$, where $\gamma_h$ is the homogeneous
linewidth of the optical transition and $F$ the finesse of the
comb. However, in practice, several effects will limit the
achievable $\Delta$, including power broadening, finite laser
linewidth, spin inhomogeneous broadening, crystal vibrations in
closed loop cryostats, etc. In the present experiment, non
classical correlations between heralding photon and stored and
retrieved heralded photons have been observed until a storage time
of $4.5 \mu s$ (see Fig. \ref{fig_g2Pr}c). This is more than 20 times longer than previous
realizations \cite{Clausen2011,Saglamyurek2011} and would allow
entanglement between crystals separated by km long distance. The
storage and retrieval efficiency dropped by a factor around 10
between 1.5 and 4.5 us (see Fig. \ref{fig_g2Pr}c). This is mainly due to the fact that the
finesse could not be kept constant when decreasing $\Delta$,
because of the minimum achievable width of an absorption peak, due
to the limitations mentioned above. Longer storage times in the
excited state of up to 10 us have recently been obtained with weak
coherent states in Pr$^{3+}$:Y$_2$SiO$_5$ \cite{Maring2014}. Note
also that excited state storage times of up to 30 us have been
obtained using bright pulses storage in a Eu$^{3+}$:Y$_2$SiO$_5$
crystal \cite{Jobez2014}.
\section{Prospects for spin-wave storage with quantum light}
\label{Sec:prospects} In order to increase the storage time and to
achieve an AFC spin wave memory with on demand read-out, it has
been proposed to transfer collective optical atomic excitations to
collective spin excitations (or spin waves), using control fields
(see section \ref{Sec:Protocols}). Proof of principle experiments
have been realized in the classical regime for storage of strong
pulses \cite{Afzelius2010,Gundogan2013,Jobez2014} and we discuss
in this section the prospects to extend this experiments to
quantum light. As mentioned in section \ref{Sec:Protocols}, spin
wave storage requires materials with at least 3 long lived ground
state levels. The best known materials with the required
properties are Pr$^{3+}$ and Eu$^{3+}$ doped crystals.
The main experimental challenge for reaching the quantum regime
with spin-wave storage is to suppress the noise generated by
the control pulses, which are very close in frequency from the
single photon output (e.g. from 10 to 17 MHz in Pr$^{3+}$:Y$_2$SiO$_5$ and from 35
to 120 MHz in Eu$^{3+}$:Y$_2$SiO$_5$ ). In that respect, one advantage of the AFC
scheme is that the control pulses are temporally separated from
the single photon output, which allow the use of temporal
filtering. However, the control pulses also create transient
phenomena inside the crystal due to e.g. the interaction with
unwanted residual population in the storage state because of
imperfect optical pumping. This can give rise to free induction
decay and fluorescence, which lead to noise emission simultaneous
with the single photon emission. The suppression of this noise
requires narrow spectral filtering, which can be realized using
e.g. a narrow-band optical cavity or a crystal filter
\cite{Zhang2012,Beavan2013}. An interesting figure of merit in the
context of single-photon-level spin-wave storage, taking into
account the noise generated and the storage and retrieval
efficiency, is the mean number of input photons to achieve a
signal-to-noise ratio (SNR) of 1 in the output mode, denoted as
$\mu_1$. In 2013, an experiment in Eu$^{3+}$:Y$_2$SiO$_5$ achieved $\mu_1$=2.5,
using a Fabry-Perot narrowband filter cavity \cite{Timoney2013}. Very recently, a lower value of $\mu_1$=0.1 has been demonstrated in Eu$^{3+}$:Y$_2$SiO$_5$ \cite{Jobez2015}. A low value of $\mu_1$=0.07 has also been shown in Pr$^{3+}$:Y$_2$SiO$_5$ \cite{Gundogan2015}, using a reconfigurable transparency window in an another
Pr$^{3+}$:Y$_2$SiO$_5$ crystal as spectral filter. High fidelity storage of time-bin qubits encoded in weak-coherent states at the single photon level was also demonstrated \cite{Gundogan2015}, allowing the spin-wave memory to operate in the quantum regime.
In order to achieve high
fidelity storage of quantum light, it is important to obtain
$\mu_1\ll 1$. This is because in practice, because of various
optical losses, it is very difficult to obtain a single photon in
front of the memory, with efficiency approaching unity. To achieve
quantum storage with high SNR using the photon pair sources
presented in section \ref{Sec:SPDCsource:General}, the condition
$\mu_1\ll \eta_H$ must be fulfilled. Such an experiment has so far
not been demonstrated.
Beyond the challenge of storing a single photon as a single
spin-wave excitation, another important challenge is to increase
the spin-wave storage time in this single-excitation regime. The
spin-wave memory in a rare-earth-ion doped crystal is limited by
the inhomogeneous spin linewidth \cite{Afzelius2010}. This
limitation can be lifted by utilizing spin-echo techniques to
rephase the inhomogeneous spin dephasing, in which case the
storage time is limited by the spin coherence time. But one can
push the storage time still further by implementing dynamical
decoupling sequences to reach storage times beyond 1 s, a
technique that has been successfully implemented in EIT storage
experiments of strong optical pulses in rare-earth-ion doped
crystals \cite{Longdell2005,Heinze2013}. However, the use of spin echo tecyhniques for extending the storage time of ensemble-based quantum memories introduces a new potential source of noise. The challenge lies in
avoiding to populate the $|s\rangle$ state with too many atoms,
which would lead to spontaneous emission noise in the output when
reading out the single spin excitation. At first this might appear
to be almost impossible for a single excitation in $|s\rangle$
\cite{Johnsson2004}, but it was shown later that the strong
collective emission into a particular spatial mode of the stored
single excitation provides a very effective spatial filtering of
the spontaneous emission noise \cite{Heshami2011}. The spin-echo
technique must be very efficient, however, to avoid this noise,
and it remains to be seen if the storage fidelity of a single
photon can be high enough when applying spin echo techniques. In a very recent experiment \cite{Jobez2015} it was shown that spin echo techniques could indeed be used when manipulating an average spin excitation of around 1, in an ensemble composed of 10$^{10}$ Eu$^{3+}$ ions, without adding significant noise to the optical read out of the memory. This was made possible by using a robust and error compensating spin echo sequence which limited the population introduced to the $|s\rangle$ state to well below 1\%. In this way a $\mu_1$ parameter of around 0.3 could be maintained for up to 1 ms of spin-wave storage time, showing that the noise level should in principle allow quantum state storage on a milliseconds time scale. One future goal of this research is to implement dynamical decoupling sequences in order to further extend the spin-wave storage time.
\section{Outlook}Progress to harness the interaction between quantum light and atomic ensembles in a solid state environment has been fast in recent years. However, the full capabilities of these materials have not yet been exploited in the quantum regime and several challenges remain. For example, an experiment demonstrating simultaneously high efficiency and long storage time of quantum light has not been demonstrated yet. The quantum networking capabilities also need to be improved. Current research directions include work towards the realization of long lived heralded entanglement between remote solid state multimode quantum memories \cite{Simon2007a}. Such an experiment would pave the way to functional elementary segments of quantum repeaters with multiplexed entanglement generation. These applications would strongly benefit from the realization of a solid state photon pair source with embedded memory \cite{Sekatski2011}. Another research direction actively pursued is to increase the spectral multiplexing capabilities, together with selective frequency read-out \cite{Sinclair2014,Thiel2014}. Another promising research direction is the integration of these quantum memories with micro and nanostructures, which would open many interesting opportunities in terms of miniaturization, scalability and integration other optical elements such as quantum light sources and single photon detectors. The coupling of rare-earth ions with nanophotonic structures like photonic crystal waveguides or cavities would also provide increased light-matter interaction and potentially lead to cavity QED experiments with a low number of rare-earth ions \cite{Zhong2014}. This may also facilitate the detection and manipulation of single rare-earth ions that could be used as quantum bits \cite{Kolesov2012,Utikal2014}. Finally, rare-earth ion doped crystals can be used as spin ensemble that can be coupled to superconductive cavities \cite{Staudt2012,Probst2013}, with the long term goal of connecting superconducting qubits and optical photons.
\begin{acknowledgement}
The results described here has been obtained during several years
of work by the teams at the University of Geneva and ICFO. The
authors would like to particularly acknowledge the contributions
by their co-authors of the work reviewed here, Christoph Clausen,
Imam Usmani, F\'{e}lix Bussi\`{e}res, Nicolas Sangouard, Nicolas
Gisin, Daniel Rielander, Kutlu Kutluer, Mustafa Gundogan, Patrick
Ledingham, Margherita Mazzera, Julia Fekete and Matteo Cristiani.
\end{acknowledgement}
|
\section{Introduction}\label{1}
The forward error correction (FEC) scheme is one of the integral parts of the wireless systems. The turbo coding scheme [1] has been adopted for the air interface standard called Long Term Evolution (LTE), that has been defined by the 3rd Generation Partnership Project (3GPP) [2]. Low-density parity-check codes (LDPC) [3] are also gaining popularity as it has been chosen for IEEE 802.11n WLAN systems [4], IEEE 802.16e WiMAX systems [5] and DVB-S2 [6]. Due to their excellent performance, the turbo and LDPC codes are the primary candidates for FEC scheme for the next generation communication systems. The roaming between WLAN and LTE systems requires a multimode FEC support. Therefore, a decoder which is able to support LDPC and turbo would be beneficial.
The hardware designs of turbo and LDPC decoders are at a matured stage due to extensive efforts of researchers. Some of the efficient hardware implementations of turbo decoder can be found in [7] and [8] and some of the efficient hardware implementations of LDPC decoders can be found in [9] and [10] etc. Sun and Cavallaro [11] even designed multimode decoders as pure hardware designs. The hardware implementations provide high throughput, but the development time is not as rapid as processor based implementations. Besides, the hardware implementations suffers from inflexibility. Therefore, the hardware implementation of a multimode decoder might not be useful for other purposes. The design presented in this paper has the potential to be used as a detector or equalizer running on factor graphs, for example.
The software implementations provide the required flexibility to support a multimode decoder, but requires a careful design to achieve the target throughput. Programmable accelerators, which enable software-hardware co-design method might be an attractive solution to overcome these bottlenecks.
Several application-specific instruction-set processors (ASIP) for multimode decoders with high throughput have been designed in [12], [13] and [14]. However, all of the ASIPs have been programmed with low level language which does not meet the time-to-market requirement. Besides, the utilization of the same function units for both decoding algorithms has not been described explicitly.
The design of software and hardware together to grind out the best performance and to ensure programmability is not a straightforward task. The designer needs a very efficient tool, which can be used to design the processor easily for a particular application.
In this paper, we propose a design of a processor based on the transport triggered architecture (TTA) for turbo and LDPC decoder. TTA is a very good processor template for a programmable ASIP. The TTA based codesign environment (TCE) tool enables the designer to write an application with a high level language and design the target processor in a graphical user interface at the same time [15].
Up to our knowledge, this is the first attempt to design a TTA processor for the LDPC decoder. The turbo decoder with TTA has been designed by Salmela $et$ $al.$ [16] and Shahabuddin $et$ $al.$ [17]. As a TTA processor can be best utilized to support different algorithms, a unified processor for turbo and LDPC decoding is the natural research direction.
The max-log-MAP algorithm is used for the component decoders of turbo decoding. The parity-check matrix of size $(M,N)$ for LDPC decoding is decomposed into $M$ rows of two state trellises or supercodes. The trellis based sum-product algorithm is used on these supercodes for check node calculation. The processor achieves 22.64 Mbps throughput for turbo decoding with a single iteration and 10.12 Mbps throughput for LDPC decoding with five iterations.
The rest of the paper is organized in the following way: In Sections \ref{2} and \ref{ldpcblock}, an overview of the turbo decoding and LDPC decoding algorithms is presented. In Section \ref{unified}, the simplification techniques and the similarities between turbo and LDPC decoding algorithm are presented. The common special function unit design is presented in Section \ref{SFU}. The processor design has been presented in Section \ref{4}. In Section \ref{5}, the throughput results and comparison with other implementations are given. The conclusion is given in Section \ref{6}.
\section{Review of Turbo Decoding} \label{2}
\subsection{Turbo Decoding}
The turbo decoder consists of two soft-input soft-output (SISO) decoders, with interleavers and de-interleavers between them as shown in Fig. 1. The inputs of the turbo decoder come from the soft demodulator, which produces the log-likelihood ratios (LLR) for the systematic bits and the parity bits. The LLRs of the systematic bit, $\it{LuI}$ and first parity bits, $\it{LcI1}$ goes to the first SISO decoder. The SISO decoder produces soft outputs based on these LLRs. These soft outputs are used in the second SISO decoder as the additional information. The inputs of the second SISO decoder are the LLRs coming from the systematic bits, second parity bits denoted by $\it{LcI2}$ and output of the first SISO decoder. The LLRs of the systematic bits are scrambled this time with the same interleaving pattern used at the encoder. Similarly, the soft outputs coming from the first SISO decoder are scrambled also with the same interleaving pattern, which are used as $a$ $priori$ values for the second SISO decoder.
\begin{figure}[h]
\includegraphics[keepaspectratio,width=1\columnwidth]{./Drawing20.eps}
\caption{Block diagram of the turbo decoder.}
\label{fig:}
\end{figure}
The heart of the turbo coding is the iterative decoding procedure. The output of the second SISO decoder does not produce the hard outputs immediately, but the soft output is used again in the first SISO decoder for more accurate approximation. The process continues in a similar fashion in an iterative manner.
A single iteration by both the first and the second SISO decoder is referred to as a full iteration. On the other hand, the operation performed by a single SISO decoder can be referred to as a half iteration. At the beginning of the first iteration, the $a$ $priori$ values are set at zero. Six to eight full iterations are used to achieve sufficient performance [1].
\subsection{MAP Algorithm for Component Decoder}
The MAP algorithm for the component decoder applied here has been proposed by Benedetto $\it{et}$ $\it{al}$. [18]. The algorithm can be stated like:
1. Initialize the values of the forward state metric as $\alpha_{0}(s)=0$ if $s=S_{0}$ and $\alpha_{0}(s)=-\infty $ otherwise.
2. Calculate all the forward state metric of the same window through the forward recursion according to
\begin{equation}
\begin{split}
\alpha_{k}(s)=\hspace{.1cm}&\max_{e}{}^{*}(\alpha_{k-1}[s^{S}(e)]+u(e)LuI[k-1]\\
&+c_1(e)LcI\it{1}[k-1]+c_2(e)LcI\it{2}[k-1]).
\end{split}
\end{equation}
3. Initialize the values of the backward state metric as $\beta_{n}(s)=0$ if $s=S_{n}$ and $\beta_{n}(s)=-\infty $ otherwise.
4. Calculate all the backward state metric of the same window through the backward recursion as
\begin{equation}
\begin{split}
\beta_{k}(s)=\hspace{.1cm}&\max_{e}{}^{*}(\beta_{k+1}[s^{E}(e)]+u(e)LuI[k+1]\\
&+c_1(e)LcI\it{1}[k+1]+c_2(e)LcI\it{2}[k+1]).
\end{split}
\end{equation}
5. The LLR values for the information and both parity bits can be calculated as following:
\begin{equation}
\begin{split}
LLR(.;O)=\hspace{.1cm}&\max_{e}{}^{*}(\alpha_{k-1}[s^{S}(e)]+c_1(e)LcI\it{1}[k-1]\\
&+c_2(e)LcI\it{2}[k-1]+\beta_{k}[s^{E}(e)]).
\end{split}
\end{equation}
For max-log-MAP algorithm, $\text{max}^*(x,y)\approx \text{max}(x,y)$ [19].
The decoding is done in smaller windows so that the decoding process can be done in parallel and the decoder does not have to wait for the whole block to arrive before starting the decoding process. This windowing is sometimes referred to as a sliding window method.
\section{Review of LDPC Decoding} \label{ldpcblock}
\subsection{Quasi-Cyclic LDPC Codes}
LDPC codes are linear block codes which consist of codewords
satisfying the parity-check equation
\begin{equation}
\mathbf{Hx}^T = \mathbf{0} \textrm{,}
\end{equation}
where $\mathbf{H}$ is the parity-check matrix and $\mathbf{x}$ is a codeword. The parity-check matrix $\mathbf{H}$
is `sparse' or consists of a small number of non-zero entries in case of LDPC codes.
The non-zero entries of the parity check matrix $\mathbf{H}$ are usually distributed pseudo-randomly according
to some distribution. Although this pseudo-random distribution leads to very good FER performance, it makes
the encoding and decoding of LDPC codes difficult. Therefore, a structure is imposed on $\mathbf{H}$ to ease
the encoding and decoding by slightly sacrificing the performance. A good trade-off between
complexity and performance is provided by the quasi-cyclic (QC) LDPC codes [20].
The parity check matrix of the QC-LDPC consists of square blocks which are either all zero matrix or
cyclic shifts of the identity matrix. This structure of the parity check matrix leads to
efficient encoding and decoding architectures. Due to their architecture aware construction, QC-LDPC
codes have been adopted by several wireless standards such as IEEE 802.11n, IEEE 802.16e and DVB-S2.
\subsection{Decoding of LDPC Codes}
LDPC codes can be visualized by a bipartite graph consisting of check and variable nodes
which represent the rows and columns of the parity check matrix respectively.
The decoding algorithm of LDPC codes is described as a message passing algorithm running on this graph.
Alternative LDPC decoding algorithms differ basically in two aspects which are
message computation at the check nodes and message flow schedules. In LDPC decoding, messages
are almost always represented in LLR domain to make the decoding numerically stable.
\begin{figure}[h]
\centering
\includegraphics[keepaspectratio,width=1\columnwidth]{./ldpc.eps}
\caption{Super codes from the parity-check matrix of LDPC codes.}
\label{fig:}
\end{figure}
The sum-product algorithm is employed at the check nodes in exact or
approximate fashion. An approximation, which is proposed in [21],
computes outgoing messages by finding the minimum of the absolute values of a subset of the incoming
messages. Although this approximation is very popular in the LDPC decoding literature,
the hardware it imposes is not useful for turbo decoding. The hardware for computing the minimum and absolute values are redundant for turbo decoding. Therefore, using this
approximation reduces the hardware reusability. Hence, we use the forward-backward
algorithm running on a binary trellis similar to [9] and [11]. This algorithm is derived by decomposing $(M,N)$ parity check matrix into $M$ binary or two-state trellises, which can also be called supercodes. The supercodes are shown in Fig. 2. The algorithm
can be described for a check node of weight $Z$ as follows.
\textbf{1.} The forward and backward recursion metrics are initialized as
\begin{eqnarray}
\alpha_0(0)&=&\beta_{Z}(0)=0 \nonumber \text{.} \\
\alpha_0(1)&=&\beta_{Z}(1)=-\infty \nonumber \text{.}
\end{eqnarray}
\textbf{2.} Forward recursion metrics are computed for $k=1,2,\ldots,Z-1$ as
\begin{eqnarray}
\alpha_{k}(s)=\max_{b}{}^{*} \left\{ \alpha_{k-1}(s \oplus b) + (-1)^{b}Li_{k-1} \right\} \textrm{,}
\end{eqnarray}
where $s$ and $b$ are binary variables, $\oplus$ denotes the binary addition, and $Li_k$ denotes
the incoming message from the $k^{th}$ neighbor.\\
\textbf{3.} Backward recursion metrics are computed for \mbox{$k=Z-1,Z-2,\ldots,1$} as
\begin{eqnarray}
\beta_{k}(s)=\max_{b}{}^{*} \left\{ \beta_{k+1}(s \oplus b) + (-1)^{b}Li_{k+1} \right\} \textrm{.}
\end{eqnarray}
\textbf{4.} Finally the outgoing message to the $k^{th}$ neighbor is given by
\begin{eqnarray}
Lo_k&=&\frac{1}{2}\left(\max_{s}{}^{*} \left\{ \alpha_{k-1}(s)+\beta_{k}(s) \right\} \right. \nonumber \\
&& \left.- \max_{s}{}^{*} \left\{ \alpha_{k-1}(s)+\beta_{k}(s \oplus 1) \right\}\right)\textrm{.}
\end{eqnarray}
Notice that the steps 3 and 4 can be carried in the same run to get rid of storing $\beta_k(.)$.
We prefer layered schedule [22] as the message flow schedule. This schedule can be formally described
as follows.
\textbf{1.} Initialize $A(n)$ to $\lambda(n)$ for $n=1,2,\ldots,N$ where
$\lambda(n)$ denotes the LLR of the $n^{th}$ bit received from the channel and
$N$ is the block length of the code. \\
\textbf{2.} For each row repeat the following. \\
\textbf{2.a} Assign $Li_k=A(\pi_j(k))-Lp_{j,k}$ where $\pi_j(k)$ denotes location
of the $k^{th}$ $1$ on the $j^{th}$ row of $\mathbf{H}$ and $Lp_{j,k}$ is the outgoing
message computed by $j^{th}$ row in the previous iteration for the $\pi_{j}(k)^{th}$ bit. For the first iteration
take $Lp_{j,k}$ as $0$. \\
\textbf{2.b.} Compute the outgoing messages according to the algorithm above for the $j^{th}$
row.\\
\textbf{2.c.} Update $A(\pi_j(k))$ as $A(\pi_j(k)) \leftarrow A(\pi_j(k))+ Lo_k$ and assign $Lp_{j,k}=Lo_k$
to use in the next iteration. \\
\textbf{3.} Goto Step 2 until a certain number of iteration. \\
\textbf{4.} $A(n)$ holds the estimated LLR's from the LDPC decoding.
\section{Shared Calculations between Turbo and LDPC Decoding} \label{unified}
There are 16 branch metric computations between two states for forward metric, backward metric and LLR calculations in the trellis diagram of an eight state convolutional code.
\begin{figure}[h]
\centering
\includegraphics[keepaspectratio,width=.8\columnwidth]{./trellis.eps}
\caption{Trellis of 3GPP turbo code.}
\label{fig:}
\end{figure}
From the trellis structure of the 3GPP turbo code in Fig. 3, it can be seen that four calculations of branch metric are being repeated to result in total sixteen calculations. The four calculations can be expressed as
\begin{equation} \small
\begin{split}
\gamma_1 = LuI+LcI\it{1}+LcI2,\\
\gamma_2 = -LuI-LcI\it{1}+LcI2,\\
\gamma_3 = LuI+LcI\it{1}-LcI2,\\
\gamma_4 = -LuI-LcI\it{1}-LcI2,
\end{split}
\end{equation}
where $\gamma_4$ can be represented as $-\gamma_1$ and $\gamma_3$ can be represented as $-\gamma_2$. Therefore, it is sufficient to calculate $\gamma_1$ and $\gamma_2$ only.
\subsection{Forward Metric calculation}
The trellis of the 3GPP turbo code can be divided into four butterfly pairs. Using the branch metric values given above, the forward metric calculation of a butterfly pair can be given as
\begin{equation}
\begin{split}
\alpha_k(1) = \text{max}^*(\alpha_{k-1}(1)-\gamma_{k-1}(1),\alpha_{k-1}(2)+\gamma_{k-1}(1)) \text{.}\\
\alpha_k(5) = \text{max}^*(\alpha_{k-1}(1)+\gamma_{k-1}(1),\alpha_{k-1}(2)-\gamma_{k-1}(1)) \text{.}
\end{split}
\end{equation}
The forward metric calculation for a supercode in LDPC decoding is also similar.
\begin{equation}
\begin{split}
\alpha_k(1) = \text{max}^*(\alpha_{k-1}(0)-Li_{k-1},\alpha_{k-1}(1)+Li_{k-1}) \text{.}\\
\alpha_k(0) = \text{max}^*(\alpha_{k-1}(0)+Li_{k-1},\alpha_{k-1}(1)-Li_{k-1}) \text{.}
\end{split}
\end{equation}
The value of the branch metric in LDPC decoding equals to a single LLR value between two time instances. On the other hand, the branch metric value in turbo decoding is a combination of three LLR values.
\subsection{Backward Metric and LLR calculation}
The backward metric calculation is also similar for turbo and LDPC. For turbo decoding, the backward metric calculation of a butterfly pair can be presented as
\begin{equation}
\begin{split}
\beta_k(1) = \text{max}^*(\beta_{k+1}(1)-\gamma_{k+1}(1),\beta_{k+1}(5)+\gamma_{k+1}(1)) \text{.}\\
\beta_k(2) = \text{max}^*(\beta_{k+1}(1)+\gamma_{k+1}(1),\beta_{k+1}(5)-\gamma_{k+1}(1)).
\end{split}
\end{equation}
and for LDPC,
\begin{equation}
\begin{split}
\beta_k(1) = \text{max}^*(\beta_{k+1}(0)-Li_{k+1},\beta_{k+1}(1)+Li_{k+1}) \text{.}\\
\beta_k(0) = \text{max}^*(\beta_{k+1}(0)+Li_{k+1},\beta_{k+1}(1)-Li_{k+1})\text{.}
\end{split}
\end{equation}
The operations needed to calculate the forward and backward metric is similar. However, the output LLR computation is different for the algorithms.
The output LLR of turbo involves eight forward metric, eight backward metrics and all the branch metric in between. The calculation can be presented as
\begin{equation} \small
\begin{split}
LLR_k=\text{max}&(\alpha_{k-1}(1)+\beta_k(1)+\gamma_1(k),\alpha_{k-1}(2)+\beta_k(5)+\gamma_1(k),\\
&\alpha_{k-1}(7)+\beta_k(8)+\gamma_1(k),\alpha_{k-1}(8)+\beta_k(4)+\gamma_1(k),\\
&\alpha_{k-1}(3)+\beta_k(6)+\gamma_2(k),\alpha_{k-1}(4)+\beta_k(2)+\gamma_2(k),\\
&\alpha_{k-1}(5)+\beta_k(3)+\gamma_2(k),\alpha_{k-1}(6)+\beta_k(7)+\gamma_2(k))\\
-\text{max}&(\alpha_{k-1}(1)+\beta_k(5)-\gamma_1(k),\alpha_{k-1}(2)+\beta_k(1)-\gamma_1(k),\\
&\alpha_{k-1}(5)+\beta_k(7)-\gamma_1(k),\alpha_{k-1}(6)+\beta_k(3)-\gamma_1(k),\\
&\alpha_{k-1}(3)+\beta_k(2)-\gamma_2(k),\alpha_{k-1}(4)+\beta_k(6)+\gamma_2(k),\\
&\alpha_{k-1}(7)+\beta_k(4)-\gamma_2(k),\alpha_{k-1}(8)+\beta_k(8)+\gamma_2(k)) \text{.}
\end{split}
\end{equation}
On the other hand, the LLR calculation of the super code in LDPC is simple.
\begin{equation}
\begin{split}
LLR_k = \frac{1}{2}(\text{max}(\alpha_{k-1}(0)+\beta_{k}(0),\alpha_{k-1}(1)+\beta_{k}(1))\\
-\text{max}(\alpha_{k-1}(0)+\beta_{k}(1),\alpha_{k-1}(1)+\beta_{k}(0))) \text{.}
\end{split}
\end{equation}
The output LLR calculation of turbo decoding needs to be divided into four parts to make the calculations similar.
\section{Special Function Unit Design} \label{SFU}
\subsection{ALPHA Special Function Unit}
A function unit can be made with three inputs and two outputs to compute the forward and backward metric. In turbo case, the unit can use $\alpha_{k-1}(1)$, $\alpha_{k-1}(2)$ and $\gamma_{k-1}(1)$ as inputs and compute the outputs of $\alpha_k(1)$ and $\alpha_k(5)$. In case of LDPC, the same unit can use $\alpha_{k-1}(0)$, $\alpha_{k-1}(1)$ and LLR as inputs and compute $\alpha_k(0)$ and $\alpha_k(1)$ as outputs. The same function unit can be used for both of the cases because the operations are same as can be seen from (9) and (10).
\begin{figure}[h]
\centering
\includegraphics[keepaspectratio,width=.8\columnwidth]{./alpha.eps}
\caption{ALPHA unit for a single butterfly pair.}
\end{figure}
One of the ALPHA special function units calculates two forward metric values based on two earlier state forward metric in the same butterfly pair. Therefore four ALPHA unit can calculate all the necessary forward metric values for one time instant. A block diagram of the ALPHA unit used for LDPC decoding is presented in Fig. 4.
On the other hand, the LDPC can utilize these four units by processing four supercodes parallely.
\subsection{BetaLLR Special Function Unit}
The backward metric and the LLR is computed together to reduce memory requirement. For a single algorithm it is easier to design a unit for beta separately and LLR separately. However, the LLR calcualtion of LDPC and turbo is not the same.
It can be seen from (14) that we can calculate with a special function unit the two maximization properties as
\begin{equation}
\begin{split}
output\text{\it{1}} = \text{max}(\alpha_{k-1}(0)+\beta_{k}(0),\alpha_{k-1}(1)+\beta_{k}(1)) \text{.}\\
output\text{\it{2}} = \text{max}(\alpha_{k-1}(0)+\beta_{k}(1),\alpha_{k-1}(1)+\beta_{k}(0)) \text{.}
\end{split}
\end{equation}
The unit would calculate the earlier state backward metrics $\beta_{k}(0)$ and $\beta_{k}(1)$ based on $\beta_{k+1}(0)$ and $\beta_{k+1}(1)$.
Therefore, the BetaLLR unit takes five inputs and produce four outputs. For example, the unit takes $\beta_{k+1}(0)$, $\beta_{k+1}(1)$, $\alpha_{k-1}(0)$, $\alpha_{k-1}(1)$ and $Li_{k+1}$ as inputs and can produce the earlier state backward metrics $\beta_{k}(0)$, $\beta_{k}(1)$, $output\text{\it{1}}$ and $output\text{\it{2}}$ of (15).
A block diagram of the unit is given in Fig. 5.
\begin{figure}[h]
\centering
\includegraphics[keepaspectratio,width=1\columnwidth]{./BetaLLR.eps}
\caption{BetaLLR unit for a single butterfly pair.}
\end{figure}
Equation (13) has to be divided in some similar form to utilize the BetaLLR unit to compute the backward metric and output LLR for turbo decoding.
Equation (13) can be expressed as
\begin{equation} \small
\begin{split}
LLR_k=\text{max}(\text{max}(\alpha_{k-1}(1)+\beta_k(1),\alpha_{k-1}(2)+\beta_k(5))+\gamma_1(k),\\
\text{max}(\alpha_{k-1}(7)+\beta_k(8),\alpha_{k-1}(8)+\beta_k(4))+\gamma_1(k),\\
\text{max}(\alpha_{k-1}(3)+\beta_k(6),\alpha_{k-1}(4)+\beta_k(2))+\gamma_2(k),\\
\text{max}(\alpha_{k-1}(5)+\beta_k(3),\alpha_{k-1}(6)+\beta_k(7))+\gamma_2(k))\\
-\text{max}(\text{max}(\alpha_{k-1}(1)+\beta_k(5),\alpha_{k-1}(2)+\beta_k(1))-\gamma_1(k),\\
\text{max}(\alpha_{k-1}(5)+\beta_k(7),\alpha_{k-1}(6)+\beta_k(3))-\gamma_1(k),\\
\text{max}(\alpha_{k-1}(3)+\beta_k(2),\alpha_{k-1}(4)+\beta_k(6))+\gamma_2(k),\\
\text{max}(\alpha_{k-1}(7)+\beta_k(4),\alpha_{k-1}(8)+\beta_k(8)+\gamma_2(k)) \text{.}
\end{split}
\end{equation}
We can divide (16) in four parts to use the BetaLLR unit. For example, one of the parts is given here as
\begin{equation}
\begin{split}
output\text{\it{1}} = \text{max}(\alpha_{k-1}(1)+\beta_k(1),\alpha_{k-1}(2)+\beta_k(5)) \text{.} \\
output\text{\it{2}} = \text{max}(\alpha_{k- 1}(1)+\beta_k(5),\alpha_{k-1}(2)+\beta_k(1)) \text{.}
\end{split}
\end{equation}
The branch metric $\gamma$ needs to be added or subtracted with the left side of (17) and have to use maximization unit to get the final LLR output. A block diagram is given in Fig. 6 to calculate a LLR in turbo mode.
\begin{figure}[h]
\centering
\includegraphics[keepaspectratio,width=1\columnwidth]{./oneLLR.eps}
\caption{LLR computation with four BetaLLR units in turbo mode.}
\end{figure}
Four BetaLLR unit can be utilized in LDPC mode by processing four supercodes parallely.
\section{Transport Triggered Architecture Processor} \label{4}
\subsection{Top level architecture}
\begin{figure*}[!t]
\centering
\includegraphics[keepaspectratio,width=2\columnwidth]{./processorldpc.eps}
\caption{Implemented processor with reduced number of function units.}
\label{fig:}
\vspace*{4pt}
\end{figure*}
A part of the TTA processor designed for the LDPC and turbo decoding is illustrated in Fig. 7. For readability, the whole processor figure is not given. The blocks on the upper part of the figure represent the function units and register files of the processor. The black horizontal straight lines represent the buses of the processor. The vertical rectangular blocks represent the sockets. The connection between function units and buses is illustrated by black dots in the sockets.
The fixed point processor includes load/store unit (LSU), arithmetic logic unit (ALU), global control unit (GCU) and register files. Based on the resource requirements in high level language, function units and register files are added.
The ALU unit is used to perform the basic arithmetic operations like addition, subtraction etc. Operations like shifting right or left are also included in ALU.
For turbo decoding, the forward and backward metric computations between two states need at least four of the ALPHA and BetaLLR units. Therefore, four ALPHA and four BetaLLR units are used in the processor.
LSU units are used to support the memory accesses. The LSU units are used to read and write memory. The memory can be read in three clock cycles and can be written in a single cycle.
Several register files are used to save the intermediate results. In terms of the power consumption, registers can be more expensive than memory, but to meet the latency requirement register files are needed. A single Boolean register file has been included in the processor design.
Thirty buses have been used for the processor. The number of buses is crucial to ensure the parallel processing. However, the complexity also increases with the increased number of buses.
The LLR outputs have been written using a first-in-first-out (FIFO) buffer by using the function unit called STREAM. The STREAM units can write every output sample in three clock cycle.
\subsection{Programming the processor}
The processor is programmed with high level language C. Several macros have been used to call the function units and use part of that code with that specific function unit.
The turbo decoding algorithm use three blocks of 6,144 input LLRs. The blocks have been divided in smaller windows to save the memory requirements. Only the forward metrics have been saved in a window. The backward metric is calculated with the BetaLLR unit and used immediately to calculate the corresponding LLR.
The forward metric and backward metrics increase in each step and that is why the forward and backward metrics is normalized to avoid memory overflow.
Before processing the ALPHA and BetaLLR values, the $\gamma$ values need to be calculated. As shown in figure 7, The output of the BetaLLR also needs to maximize and added or subtracted with $\gamma$ values to find the output LLR.
In the LDPC mode, the processor is programmed for the LDPC code of IEEE WLAN 802.11n of code rate 1/2 and output block size of 648. Due to the data dependencies, a single special function unit is used several times to calculate the required forward and backward metric values of a supercode in serial fashion. For example, the first row of the $\mathbf{H}$ matrix of this particular code configuration has seven nonzero elements. The two-state trellis should be a matrix of $8 \times 2$ to calculate the forward and backward metrices for this row. The initilization values of the metrices are zero. Therefore, only one ALPHA and one BetaLLR units are used seven times each to get all the necessary output LLR from this row. Four of this rows can be processed in parallel as there are four ALPHA and BetaLLR units available.
The variable node update is done by simply adding the LLR outputs of the super codes with the corresponding original LLRs of the same position. Shifting operations are required to
\section{Results and Discussion} \label{5}
The designed processor takes 166,224 clock cycles to process three blocks of 6,144 samples for a full iteration for the turbo decoding. The processor takes 10,368 clock cycles to decode a LDPC code for IEEE WLAN 802.11n of block size 648 and code rate 1/2 after five iterations.
The throughput can be calculated using the following equation as
\begin{equation}
\text{Throughput} = \frac{\text{Size of the code block} \times \text{device clock frequency}}{\text{latency} \times \text{number of iterations}}\text{.}
\end{equation}
The throughput achieved for the turbo mode is 22.64 Mbps for a single iteration and for LDPC mode 10.12 Mbps for five iterations for a clock frequency of 200 MHz.
The buses of the processor are perfectly utilized to achieve the best possible result due to the perfect scheduling. The number of some of the operations during the algorithm execution has been summarized in Table \ref{tab1}.
\begin{table}[h]
\centering
\caption{Number of operations}
\label{tab1}
\begin{tabular}{|c|c|c|}
\hline
Operation & \# of OPS in turbo & \# of OPS in LDPC\\
\hline
ADD & 431,009 & 87,134\\
\hline
SUB & 96,354 & 14,231\\
\hline
MAX & 43,008 & 0\\
\hline
ALPHA & 24,576 & 2,376\\
\hline
BetaLLR & 24,576 & 2,376\\
\hline
STREAM & 6,144 & 648\\
\hline
\end{tabular}
\end{table}
The number of addition operations does not only represent the addition for the algorithm, but for several other purposes like loop indexing for the code. The maximization units are not used in case of LDPC decoding because the maximization operations are done inside the ALPHA and BetaLLR units.
A comparison with different other programmable implementations of turbo decoder has been presented in Table \ref{tab2}. The throughput results are normalized for a clock frequency of 200 MHz. Our proposed processor with turbo mode provides very good throughput compared to other programmable implementations. The TTA processors of [16] and [17] provide higher throughputs but the designs were dedicated for only turbo decoding.
\begin{table}[h]
\centering
\caption{Programmable Turbo Processors}
\label{tab2}
\begin{tabular}{|c|c|c|c|}
\hline
Reference & Architecture & Algorithm & Throughput\\
\hline
[23] & TMS320C6201 DSP & max-log-MAP & 2 Mbps\\
\hline
[24] & VLIW ASIP & max-log-MAP & 5 Mbps\\
\hline
proposed & TTA proc. in turbo mode & max-log-MAP & 22.64 Mbps\\
\hline
[17] & TTA proc. for LTE & max-log-MAP & 31.21 Mbps\\
\hline
[25] & Nvidia C1060 & max-log-MAP & 33.85 Mbps\\
\hline
[16] & TTA proc. & max-log-MAP & 98 Mbps\\
\hline
\end{tabular}
\end{table}
A comparison with different other programmable implementations of LDPC decoder has been presented in Table \ref{tab3}. The throughput results are normalized for a clock frequency of 200 MHz. Our proposed processor with LDPC mode provides moderate throughput compared to most of the programmable implementations.
\begin{table}[h]
\centering
\caption{Programmable LDPC processors}
\label{tab3}
\begin{tabular}{|p{1cm}|p{2cm}|p{2cm}|p{2cm}|}
\hline
Reference & Architecture & Algorithm & Throughput\\
\hline
[26] & TMS320C64xx DSP & min-sum & 1.8 Mbps @ 10 it.\\
\hline
proposed & TTA proc. for LDPC & supercode based sum-product & 10.12 Mbps @ 5 it.\\
\hline
[27] & SDR SODA & min-sum & 15.2 Mbps @ 10 it.\\
\hline
[14] & VLIW ASIP & offset min-sum & 53 Mbps @ 10 it.\\
\hline
[12] & VLIW ASIP & offset min-sum & 16.32 - 128.5 Mbps @ 10-20 it.\\
\hline
\end{tabular}
\end{table}
Alles $\it{et}$ $\it{al}$. presented an efficient implementation of multimode decoder in [12]. The ASIP achieved 34.5 Mbps to 257 Mbps for LDPC codes of different code configurations and block size of IEEE 802.11n when the clock frequency is 400 MHz. The lowest throughput of 34.5 Mbps at 400 MHz clock frequency after 20 iterations was achieved when the code rate and block size is the same as presented in this paper. The design of [14] achieved high throughput with a different code configuration. Besides, all the implementations presented here used the assembly language.
\section{Conclusion} \label{6}
The paper discussed the design issues of a turbo and LDPC decoder on a TTA processor. The design shows the promise of the possibility of designing several decoding techniques on a single TTA processor. The processor designed in this paper can be used for tasks beyond decoding, for instance, it can be programmed for detection and equalization algorithms running on factor graphs [28]. The target throughput of LTE can also be reached by multi-core TTA processor. The flexibility gained from that processor could provide very interesting results and would be a fruitful direction for future research.
\section{Acknowledgement} \label{con}
This research was supported by the Finnish Funding Agency for Technology and Innovation (Tekes), Renesas Mobile Europe, Nokia Siemens Networks, Elektrobit and Xilinx.
Special thanks are due to Dr. Perttu Salmela from Tampere University of Technology and Dr. Frederik Naessens from IMEC for sharing their insights on programmable turbo and LDPC decoder implementations.
|
\section{Introduction}
In this paper, an operator is always assumed to be linear but not necessarily
bounded.
Let $X$ be a Banach space. A one-parameter family $(T(t))_{t\geq 0}$ of bounded linear operators from $X$ to $X$ is a semigroup of bounded
linear operators on $X$ if
\begin{enumerate}[(i)]
\item $T(0)=\mathop{\rm Id}$, the identity operator on $X$;
\item $T(t+s)=T(t)T(s)$ for every $s,t\geq 0$.
\end{enumerate}
The linear operator $A$ defined by
\[D(A)=\left\{ x\in X: \lim_{t\downarrow 0}\frac{T(t)x-x}{t}\mbox{ exists}\right\}\]
and
\[Ax=\lim_{t\downarrow 0}\frac{T(t)x-x}{t}\mbox{ for }x\in D(A)\]
is the (infinitesimal) generator of the semigroup $T(t)$, $D(A)$ is the domain of $A$.
The semigroup $(T(t))_{t\geq 0}$ of bounded linear operators on $X$ is strongly continuous if
\[\lim_{t\downarrow 0}\|T(t)x-x\|_X=0 \mbox{ for every }x\in X.\]
Such semigroups are also called $C_0$-semigroups.
A straightforward consequence of the uniform boundedness theorem is that
given a $C_0$-semigroup $(T(t))_{t\geq 0}$ on the Banach space $X$,
there exist $w\geq 0$ and $M\geq 1$ such that
\[\|T(t)\|\leq Me^{wt}\mbox{ for all }0\leq t<\infty.\]
In the particular case where $M=1$, the semigroup is said to be
quasicontractive. For $M=1$ and $w=0$, $(T(t))_{t\geq 0}$ is a
semigroup of contractions. \\
In 1978, Berkson and Porta \cite{B_P} gave a complete description of
the generator $A$ of semigroups of composition operators on the Hardy
space $H^2(\DD)$ (see Section~\ref{sec:2.3}),
induced by a semigroup of analytic self-maps of $\DD$ (see
Section~\ref{sec:2.2} for the definition of such semigroups).
Abate \cite{Abate} rediscovered the main results of \cite{B_P}, using a
different approach, and considering higher dimensions of the scalar
space.
Berkson and Porta \cite{B_P} noticed that such semigroups are strongly continuous on $H^2(\DD)$ and
Siskakis
\cite{S, sisk96} noticed that they are
strongly continuous on the Dirichlet space $\D$.
Moreover, it is not difficult to see that the generator $A$ of
a semigroup of composition operators is of the form $Af=Gf'$. The
aim of this paper is to give a complete description of quasicontractive
$C_0$-semigroups of bounded operators on $H^2(\DD)$ and $\D$ whose
generator $A$ is of the form $Af=Gf'$; that is, unlike previous authors, we do not
assume {\em a priori\/} that we are working with semigroups of composition operators..\\
The paper is organized as follows. First, in Section~\ref{sec:2}, we
recall the Lumer--Phillips theorem in order to obtain a contractive or
quasicontractive
$C_0$-semigroup by means of the numerical range of its generator. We
also present the main result of \cite{B_P} concerning the semigroups
of holomorphic functions on $\DD$. Then, the weighted Hardy spaces are
defined, their main properties are recalled (in particular the fact
that some of them are reproducing kernel Hilbert spaces is emphasized) and finally we also study
the norm of composition operators induced by a univalent symbol
$\varphi$ on $H^2(\DD)$ and $\D$. This study is essential to check
that the semigroups of composition operators are indeed strongly
continous and quasicontractive. To that aim, on the Dirichlet space,
the optimal estimates proved in \cite{martin05} are crucial.
In Section~\ref{sec:3}, we present our main result on $H^2(\DD)$, Theorem~\ref{ThmB},
which asserts that the only quasicontractive
$C_0$-semigroups whose generator is of the form $Af=Gf'$ are the
semigroups of composition operators. We also prove necessary and
sufficient conditions on $G$, different from the one of Berkson and
Porta.
In Section~\ref{sec:4}, we prove in Theorem~\ref{thm:dir} that the assertions of
Theorem~\ref{ThmB} are equivalent to the fact that $A$ generates a
quasicontrative $C_0$-semigroup on the Dirichlet space, which is
itself equivalent to the fact that $A$ generates a semigroup of
composition operators on $\D$.
The last section contains comments on well-known algorithms to test
the conditions on $G$, and explicit examples of constructions of the
semigroup of composition operators for a class of analytic polynomial $G$.
\section{General background}\label{sec:2}
\subsection{Characterization of contractive $C_0$-semigroups of
bounded operators on a Hilbert space}
Besides the well-known Hille--Yosida theorem (see for example Thm.~3.1 in \cite{Pazy}) which characterizes $C_0$-semigroups in terms of the growth of the resolvent of their generators $A$, another
useful theorem is the Lumer--Phillips theorem (see for example Thm.~4.3 in \cite{Pazy}) which is well-adapted for the characterization of quasicontractive $C_0$-semigroups in terms of the numerical range of $A$.
From now on, we assume that the Banach space $X$ on which $T(t)$ is defined is a complex Hilbert space,
and we denote it by $H$. This hypothesis will simplify the definition of dissipative operators involved in the Lumer--Phillips theorem.
Let $A:D(A)\to H$ be a linear operator. Then $A$ is dissipative if
\[\RE \langle Ax,x\rangle \leq 0\mbox{ for all }x\in D(A), \|x\|=1.\]
In other words, $A:D(A)\to H$ is dissipative if the numerical range of $A$ is in the left half-plane.
\begin{thm}[Lumer--Phillips]
Let $A$ be a linear operator with dense domain $D(A)$ in $X$.
\begin{enumerate}[(i)]
\item If $A$ is dissipative and there exists $\lambda_0>0$ such that
$(\lambda_0\mathop{\rm Id}-A)D(A)=H$, then $A$ is the generator of
contractive $C_0$-semigroup
on $H$.
\item If $A$ is the generator of a contractive $C_0$-semigroup on $H$,
then $A$ is dissipative and for all $\lambda >0$,
$(\lambda\mathop{\rm Id}-A)D(A)=H$.
\end{enumerate}
\end{thm}
This theorem is also of great interest for the characterization of
quasicontractive $C_0$- semigroups observing that $\|T(t)\|\leq e^{wt}$ if and only if
$\|\widetilde{T}(t)\|\leq 1$, where $\widetilde{T}(t):=T(t)e^{-wt}$ is the semigroup whose generator is $A-w\mathop{\rm Id}$, if $A$ is the generator of
$(T(t))_{t\geq 0}$.
In particular we have then the following result.
\begin{cor}\label{cor:lm}
Let $A:D(A)\to H$ be a linear operator with a dense domain. Then $A$ generates a quasicontractive
$C_0$-semigroup if and only if $\sup\{\RE\left( \langle
Ax,x\rangle\right):x\in D(A), \|x\|=1\}<\infty$ and there exists
$\lambda>0$ such that $(A-\lambda\mathop{\rm Id})D(A)=H$.
\end{cor}
\subsection{Semigroups of analytic functions}\label{sec:2.2}
\begin{thm}[Denjoy-Wolff]
Let $\varphi:\DD\to\DD$ holomorphic such that $\varphi$ is not an elliptic automorphism. Then there is a point
$b\in \overline{\DD}$ such that $\varphi_n$(:=$\varphi\circ\varphi\circ\cdots\varphi$, $n$ times) converges to $b$ uniformly
on compact subsets of $\DD$. \\
If $|b|<1$, then $\varphi(b)=b$, while if $|b|=1$, then $b$ behaves as a fixed point in the sense that $\lim_{r\to 1^-}\varphi (rb)=b$.
This distinguished point is called the Denjoy--Wolff point of $\varphi$.
\end{thm}
In the exceptional case of elliptic automorphisms different from the identity map, the sequence of iterates move around
an interior fixed point without converging to it.
\begin{defn}
A one-parameter semigroup of analytic functions of $\DD$ into
itself is a family $\Phi=\{\varphi_t:t\geq 0\}$ of analytic self-maps of $\DD$
such that
\begin{enumerate}
\item $\varphi_0(z)=z$ for all $z\in \DD$;
\item $\varphi_{t+s}(z)=\varphi_t\circ \varphi_s(z)$ for all $t,s\geq
0$ and $z\in \DD$;
\item $(t,z)\mapsto \varphi_t(z)$ is continuous on $[0,\infty)\times
\DD$.
\end{enumerate}
\end{defn}
Using Vitali's Theorem on convergence of holomorphic functions, it
follows that the continuity of $(t,z)\mapsto \varphi_t(z)$ on $[0,\infty)\times
\DD$ is equivalent to the continuity of $t \mapsto \varphi_t(z)$
for each $z\in \DD$.
Such semigroups have extensively been studied by
Berkson and Porta \cite{B_P} (see also \cite{AS}), who proved the
following useful result.
\begin{prop}\label{Sisk}
Let $\Phi=(\varphi_t)_{t\in\RR_+}$ be a semigroup of analytic functions on $\DD$, then:
\begin{enumerate}[(i)]
\item For every $t\in\RR_+$, the function $\varphi_t$ is univalent.
\item There is a holomorphic mapping $G:\DD\to \CC$ called the
generator of $\Phi$ such that
\begin{equation}\label{eq:gene}
\frac{\partial \varphi_t(z)}{\partial t}=G(\varphi_t(z))
\end{equation}
for all $t\geq 0$ and all $z\in \DD$. The convergence
\[G(z)=\lim_{t\to 0^+} \frac{\partial \varphi_t(z)}{\partial t}\]
is uniform on every compact subsets of $\DD$.
\item Moreover, the infinitesimal generator $G$ of $\Phi$ has the unique representation
\[G(z) = F(z)(\overline{\alpha}z-1)(z-\alpha),\ \forall z\in\DD,\]
where $F:\DD\to\CC$ is analytic and satisfies $\RE(F)\geqslant 0$, and $\alpha$ is the Denjoy--Wolff point of one (and thus any) $\varphi_t$, $t>0$.
\item Conversely, let $F:\DD\to\CC$ be analytic with $\RE(F)\geqslant 0$ and $\alpha\in\overline{\DD}$, the function
$z \mapsto F(z)(\overline{\alpha}z-1)(z-\alpha)$, generates a semigroup of analytic function on $\DD$.
\end{enumerate}
\end{prop}
\subsection{Operators on weighted Hardy spaces}\label{sec:2.3}
Our aim is to study semigroups of bounded operators on classical spaces of
analytic functions such as the Hardy space $H^2(\DD)$ and the
Dirichlet space $\cal D$, which are particular cases of the so-called
``weighted Hardy spaces''.
\begin{defn}
Take $(\beta_n)_{n \ge 0}$ a sequence of positive real numbers.
Then $H^2(\beta)$ is the space of analytic functions
\[
f(z)=\sum_{n=0}^\infty c_n z^n
\]
in the unit disc $\DD$ that have finite norm
\[
\|f\|_\beta= \left( \sum_{n=0}^\infty |c_n|^2 \beta_n^2 \right)^{1/2}.
\]
The case $\beta_n=1$ gives the usual \textbf{Hardy} space $H^2 (\DD)$. \\
The case $\beta_0=1$ and $\beta_n=\sqrt{n}$ for $n\geq 1$ provides the \textbf{Dirichlet} space
$\cal D$, which is included in $H^2(\DD)$. \\
The case $\beta_n=1/\sqrt{n+1}$ produces the \textbf{Bergman} space,
which contains $H^2(\DD)$.
\end{defn}
Obviously the polynomials are in $H^2(\beta)$ and with an extra condition on $(\beta_n)_n$, the Hilbert space $H^2(\beta)$ is also a
{\em reproducing kernel space}, i.e. for all $w\in \DD$, there exists a function $k_w\in H^2(\beta)$ such that
\[\langle f,k_w\rangle =f(w),\]
for all $f\in H^2(\beta)$ (see p. 19 in \cite{Cowen-Mac-Cluer} and p. 146 in \cite{Par97}). More precisely, if $(\beta_n)_n$ is such that
\begin{equation}\label{eq:rks}
\sum_{n\geq 0}\frac{|w|^{2n}}{\beta_n^2} <\infty\mbox{ for all }w\in\DD,
\end{equation}
it follows that $H^2(\beta)$ is a reproducing kernel Hilbert space and
\[k_w(z)=\sum_{n\geq 0}\frac{\overline{w}^n}{\beta_n^2}z^n \qquad \mbox{with}
\qquad \|k_w\|^2_{H^2(\beta)}=\sum_{n\geq 0}\frac{|w|^{2n}}{\beta_n^2}.\]
In fact (\ref{eq:rks}) is also equivalent to the more explicit condition $\liminf (\beta_n)^{1/n}\geq 1$.
Given an operator $A$ (possibly unbounded) defined by $A f(z) = G(z) f'(z)$ on its domain $D(A)=\{f\in H^2(\beta),\ Gf'\in H^2(\beta)\}$ where $G\in H^2 (\beta)$, we would like to know if there exists a $C_0$-semigroup on $H^2(\beta)$ with generator $A$. The next proposition asserts that two necessary conditions for $A$ to be a $C_0$-semigroup generator are satisfied.
\begin{prop}
Let $(\beta_n)_{n \ge 0}$ a sequence of positive real numbers such that $zH^2(\beta)\subset H^2(\beta)$.
Any operator $A$ defined by $A f(z) = G(z) f'(z)$ on $D(A)$ where $G\in H^2(\beta)$ is densely defined on $H^2(\beta)$ and closed.
\end{prop}
\beginpf
The operator $A$ is defined on polynomials which form a dense family of function in $H^2(\beta)$.
We consider a sequence $(f_n: z \mapsto\sum_k a_k^nz^k)\in H^2(\beta)^\N$ and two functions $f: z \mapsto\sum_k a_kz^k,
g: z \mapsto\sum_k c_kz^k \in H^2(\beta)$ such that $f_n \rightarrow f$ and $Gf_n' \rightarrow g$ in $H^2(\beta)$. We denote $G(z)=\sum_k b_kz^k$.
We now consider the truncated sums, up to the $N$-th exponent:
\[\|(Gf_n' - Gf')_N \|_2^2 \leqslant \sum_{k=0}^N \beta_k\left|\sum_{i+j=k} b_{i-1}j(a_{j}^n -a_j)\right|^2.\]
As $f_n \rightarrow f$ in $H^2(\beta)$, one has
\[ \sum_{k=0}^\infty \beta_k|a_{k}^n -a_k|^2 \rightarrow 0\]
and thus $\forall k\in\N, |a_{k}^n -a_k| \rightarrow 0$.
Hence,
\[
\|(g-Gf')_N\|_2 \leqslant \|(g-Gf_n')_N\|_2+ \|(Gf_n' - Gf')_N \|_2\rightarrow 0.
\]
We have shown that $\forall k\leqslant N$, $c_k=\sum_{i+j=k}j
b_{i-1}a_j$ and since this can be done for each choice of $N$, we
conclude that $g=Gf'\in H^2(\beta)$, $Gf_n' \rightarrow Gf'$ in
$H^2(\beta)$ and $A$ is closed.
\endpf
\begin{prop}
If $A$ is defined by $A f(z) = G(z) f'(z)$ where $G\in H^2(\DD)$ and
$\frac 1G\in H^\infty(\DD)$, then
$A $ cannot be the generator of a one-parameter semigroup.
\end{prop}
\beginpf
Let $\lambda$ be a real number; then $\lambda\in\sigma(A)$ if there exists $f\in H^2(\DD)$ such that
\[ G(z)f'(z) =\lambda f(z).\]
Since $\frac 1G\in H^\infty(\DD)$ the function $u=\int \frac{\lambda}{G}dz$ lies in $H^\infty(\DD)$ and $f=e^{u}\in H^\infty \subset H^2(\DD)$ satisfies $G(z)f'(z) =\lambda f(z)$. Thus $\RR \subset \sigma(A)$. This cannot occur for $C_0$-semigroups, see e.g. \cite[Chap.II,1.13]{E_N}.
\endpf
\begin{cor}
If $A$ is defined on $H^2(\DD)$ by $A f(z) = p(z) f'(z)$ where $p$ is polynomial with no roots in the unit closed disc $\overline{\DD}$, then $A$ cannot be the generator of a one-parameter semigroup.
\end{cor}
\subsection{Bounded composition operators on Hardy and
Di\-ri\-chlet spaces}
Composition operators on the Hardy space $H^2(\DD)$ have a quite surprising
property, namely, provided that they are well-defined, they are
always continuous. This fact is not true on the Dirichlet
space. Moreover, we have the following upper bound for the norm (see
Thm.~3.8 in \cite{Cowen-Mac-Cluer}).
\begin{thm}\label{th:compoH2}
Let $\varphi:\DD\to \DD$ be an analytic function. Then
$C_\varphi$ maps $H^2(\DD)$ continuously into $H^2(\DD)$, and moreover
\[\| C_\varphi \|\leq \left(
\frac{1+|\varphi(0)|}{1-|\varphi(0)|}\right)^{1/2}.\]
\end{thm}
The previous result is very useful to estimate the growth of the norm
of semigroups of composition operators on the Hardy space. Indeed, a
first consequence of
Theorem~\ref{th:compoH2} is that each semigroup $\Phi$ of analytic functions
on $\DD$ induces a $C_0$-semigroup of bounded operators on $H^2(\DD)$.
\begin{cor}\label{cor:quasicont}
Let $\Phi=(\varphi_t)_{t\geq 0}$ be a semigroup of analytic functions
on $\DD$. Then $(C_{\varphi_t})_{t\geq 0}$ is a quasicontractive $C_0$-semigroup on $H^2(\DD)$.
\end{cor}
\beginpf
The continuity of $t\mapsto \varphi_t(0)$ implies that $K:=\{\varphi_t(0):0\leq t\leq 1\}$ is a compact subset of
$\DD$. Since $G$, the generator of $\Phi$, is holomorphic on $\DD$, we get
\[\sup_{0\leq t\leq 1}|G(\varphi_t(0))|<\infty.\]
By (\ref{eq:gene}),
it follows that there exists $M>0$ such that $|\varphi_t(0)|\leq Mt$, and then,
for $0\leq t\leq \frac{1}{2M}$, $|\varphi_t(0)|\leq \frac{1}{2}$.
Using Theorem~\ref{th:compoH2}, we also know that
\[ \|C_{\varphi_t}\|\leq \left( \frac{1+|\varphi_t(0)|}{1-|\varphi_t(0)|}\right)^{1/2},\]
which implies that $\|C_{\varphi_t}\|\leq 1+O(t)$, and thus there exists $w\geq 0$ such that
\[\| C_{\varphi_t} \|\leq e^{wt},\]
for all $t\geq 0$.
Moreover the hypotheses on $\Phi$ imply that $C_{\varphi_t}f(z)$ tends to $f(z)$ as $t$ tends to $0$, for all $z\in\DD$ and all
$f\in H^2(\DD)$. In other words the semigroup $(C_{\varphi_t})_t$ is weakly continuous. It follows that
$(C_{\varphi_t})_t$ is strongly continuous (see Thm. I.5.8 in \cite{E_N}) .
\endpf
On the Dirichlet space, it is not true that $C_\phi$ is well defined
whenever $\varphi$ is a self-map of $\DD$. For example, for $\phi$ an infinite Blaschke
product, $C_\varphi$ is not a bounded composition operator on $\cal
D$. Nevertheless, if $\varphi$ is univalent, $C_\varphi$ is bounded on
$\cal D$ (see Section~6.2 of \cite{Ransfordetal}). We have therefore
the following preliminary result.
\begin{prop}\label{prop:compodirichlet}
Let $\Phi=(\varphi_t)_{t\geq 0}$ be a semigroup of analytic functions
on $\DD$. Then $(C_{\varphi_t})_{t\geq 0}$ is a semigroup of bounded
operators on $\cal D$.
\end{prop}
Using \cite{CP14}, it is possible to prove that
Proposition~\ref{prop:compodirichlet} is still true for any space
$H^2(\beta)$ containing $\cal D$. But this is beyond the scope of this
paper.
\section{Quasicontractive semigroups on the Hardy space}\label{sec:3}
From now on, the function $G$ will lie in $H^2(\DD)$ and the operator $A$ will be defined by $A f=G f'$ on the domain $D(A)=\{f\in H^2(\DD),\ Gf'\in H^2(\DD)\}$.
\begin{thm}\label{ThmB}
The operator $A$ generates a $C_0$-semigroup of composition operators on $H^2(\DD)$ if and only if $\forall z\in\DD$,
\begin{equation}\label{B}
2\RE(\overline{z}G(z)) + \left( 1 - |z|^2\right)\RE(G'(z))\leqslant 0.
\end{equation}
\end{thm}
\beginpf
Suppose $A$ is such a generator, let $(\varphi_t)$ denote the corresponding semigroup. From analyticity, one has for small $t$ and fixed $z$:
\[ \varphi_t(z) = z+G(z)t+o(t),\]
\[ \varphi'_t(z)= 1 +G'(z)t+o(t).\]
From the Schwarz--Pick lemma (see \cite{A}),
\[ |\varphi_t'(z)| \leqslant \frac{ 1-|\varphi_t(z)|^2}{1-|z|^2},\]
and thus,
\[1+\RE(G'(z))t+o(t)\leqslant \frac{1-|z|^2}{1-|z|^2}-\frac{2\RE(\overline{z}G(z))}{1-|z|^2}t+o(t).\]
The condition (\ref{B}) appears as $t$ tends to $0^+$.\\
We now assume that the condition (\ref{B}) is valid. For $z_0\in\DD$, consider the initial value problem
\[\frac{dw}{dt}=G(w),\quad w(0)=z_0.\]
Since $G$ is analytic and thus locally Lipschitz, there exist local solutions $w(t)=\varphi_t(z_0)$ by the Cauchy--Peano theorem with values in $\DD$. Let
\[\rho(z_1,z_2)=\min_{\gamma(0)=z_1;\gamma(1)=z_2}\int_\gamma \frac{2}{1-|z|^2}|dz|.\]
So
\begin{eqnarray*}
\rho(z_0,\varphi_t(z_0))&\leqslant& \int_0^t \frac{2}{1-|\varphi_s(z_0)|^2}\left|\frac{\partial \varphi_s(z_0)}{\partial s}\right|ds \\
& = & \int_0^t \frac{2}{1-|\varphi_s(z_0)|^2}\left|G( \varphi_s(z_0))\right| ds.
\end{eqnarray*}
Write $\displaystyle f:t\mapsto \frac{2}{1-|\varphi_t(z_0)|^2}\left|G( \varphi_t(z_0))\right|$, so that
\begin{eqnarray*}
f'(t)&=& \frac{2}{(1-|\varphi_t(z_0)|^2)^2}\bigg[ \frac{\partial |G(\varphi_t(z_0))|}{\partial t} (1-|\varphi_t(z_0)|^2)\\
& & \qquad\qquad +2\RE\left( \overline{\varphi_t(z_0)}G(\varphi_t(z_0)) \right)|G(\varphi_t(z_0))|\bigg]\\
&=& \frac{2|G(\varphi_t(z_0))|}{(1-|\varphi_t(z_0)|^2)^2}\Big[ \RE(G'(\varphi_t(z_0))) (1-|\varphi_t(z_0)|^2)\\
& & \quad\qquad\qquad\qquad\qquad+2\RE\left( \overline{\varphi_t(z_0)}G(\varphi_t(z_0)) \right)\Big]\\
&\leqslant &0 \quad \text{by condition (\ref{B}) at $\varphi_t(z_0)$}.
\end{eqnarray*}
We conclude that $f$ is a decreasing function, and thus, for $0\leqslant t_1 <t_2<\eta$,
\[\rho(\varphi_{t_1}(z_0),\varphi_{t_2}(z_0))\leqslant (t_2-t_1) \frac{2|G(\varphi_{t_1}(z_0))|}{1-|\varphi_{t_1}(z_0)|^2}.\]
Therefore, on $[0, \eta)$, $\varphi_t$ remains in a compact subset of $\DD$, so
\[\rho(\varphi_{t_1}(z_0),\varphi_{t_2}(z_0))\leqslant K|t_2-t_1|,\]
where $K$ is a constant independent of $t_1,t_2$ for $0\leqslant t_1 <t_2<\eta$.
Thus, $\varphi_t(z_0)$ converges as $t$ tends to $\eta$. This proves that there exists a solution on $\RR_+$ of the initial value problem. Following \cite{B_P}, $A$ generates a $C_0$-semigroup of composition operators on $H^2(\DD)$.
\endpf
\begin{rem}{\rm
A condition similar to \eqref{B} appears in the paper \cite{irish}, expressed in the language of
semi-complete vector fields (semiflows); that is, solutions to the Cauchy problem
\begin{eqnarray*}
\frac{du}{dt}+f(u) &=& 0, \\
u(0) &=& x,
\end{eqnarray*}
together with the alternative condition
\[
\re f(z)\overline z \ge \re f(0)\overline z (1-|z|^2), \qquad z \in \DD,
\]
(see also \cite[Prop. 3.5.2]{shoikhet}). Thus, as in \cite{B_P}, they start with
a semigroup of functions under composition.
}
\end{rem}
\begin{Not}
For each $G(z)=\sum_{n=0}^\infty \alpha_n z^n \in H^2(\beta)$, we write $\tilde G(z) = \alpha_1 + (\alpha_2+\overline{\alpha_0})z+\sum_{n=3}^\infty \alpha_n z^{n-1}$.
\end{Not}
An easy test using numerical ranges gives the following necessary condition for the generation of a $C_0$ semigroup
of quasicontractions. A more general result (with a more complicated proof) appears in
Proposition~\ref{prop:tinto}.
\begin{prop}
If the operator $A$ generates a $C_0$-semigroup of quasicontractions on $H^2(\beta)$ with $\beta =\left(n^{-\alpha}\right)$ and $\alpha \ge 0$, then
\begin{equation}\label{C}
\mathop{\rm ess}\sup_{z\in\TT} \RE (\tilde G(z))=\mathop{\rm ess}\sup_{z\in \TT} \RE (\overline{z} G(z))\leqslant 0.
\end{equation}
\end{prop}
\beginpf Observing that
\[\sup_{z\in \TT} \RE (\overline{z} G(z)) =
\sup_{\theta\in\RR} \left\{\RE(\alpha_1) +\RE\left((\overline{\alpha_0} +\alpha_2)e^{i\theta} + \sum_{n=3}^{\infty}\alpha_ne^{i(n-1)\theta}\right)\right\},\]
we can compute the numerical range of $A$. Let $f$ be an analytic
function defined by $f(z)=\sum_{n=0}^\infty a_nz^n$ with
$\|f\|_{H^2(\beta)}=1$ and $f\in D(A)$. Then we have
\begin{eqnarray*}
\RE\left(\left\langle G(z)f'(z),f(z)\right\rangle\right)
&=&\RE\left(\left\langle \tilde G(z)zf'(z),f(z)\right\rangle
+\overline{ \alpha_0} \sum_{n=0}^\infty \beta_n\beta_{n+1}a_{n} \overline{a_{n+1}} \right)\\
& =& \RE(\alpha_1)\sum_{n=0}^\infty\beta_{n}^2 n |a_n|^2 \\
& & +\RE\left((\alpha_2 +\overline{\alpha_0})\sum_{n=1}^\infty\beta_n\beta_{n+1} n a_n \overline{a_{n+1}} \right)\\
& & + \RE\left(\overline{ \alpha_0} \sum_{n=0}^\infty \beta_n\beta_{n+1}a_{n} \overline{a_{n+1}} \right)\\
& & + \RE\left(\sum_{k=3}^{\infty}\alpha_k \sum_{n=0}^\infty \beta_n\beta_{n+k-1}n a_n \overline{a_{n+k-1}}\right).
\end{eqnarray*}
Consider
the polynomial functions
(obviously in $D(A)$) defined by
\[f_N(z)=c_N\sum_{n=1}^N \frac{\sqrt{6}e^{-in\theta}}{\pi
n^{1-\alpha}}z^n,\]
where $c_N$ is a positive real chosen so that $\|f_N\|_{H^2(\beta)}
=1$. It is clear that $c_N$ tends to $1$ as $N$ tends to $\infty$.
Note now that if (\ref{C}) is not satisfied, then a suitable choice of $\theta$ makes
$\RE\left( \langle Af_N,f_N\rangle\right)$
tend to $\infty$ as $N$ tends to $\infty$. It follows that if
(\ref{C}) is not satisfied, then $A$ cannot generate a $C_0$-semigroup, see
e.g. \cite[Chap.II, 3.23]{E_N}.
\endpf
\begin{rem}
It is easy to see that condition (\ref{B}) implies condition (\ref{C}).
\end{rem}
\begin{prop}\label{prop:CimpliesB}
The condition (\ref{C}) implies the condition (\ref{B}).
\end{prop}
\beginpf
Assume $G$ satisfies condition (\ref{C}) and let $H(z)=z\tilde G(z)$. Condition (\ref{C}) and the maximum principle implies that
$\sup_{z\in\DD}\tilde G(z)\leqslant 0$. Thus by \cite[Theorem 3.3]{B_P} and Corollary~\ref{cor:quasicont}, $f\mapsto Hf'$ generates a $C_0$-semigroup of composition operators on $H^2(\DD)$. Hence, by Theorem \ref{ThmB}, $\forall z\in\DD$,
\[ X:=2\RE(\overline z H(z))+(1-|z|^2)\RE H'(z)\leqslant 0.\]
Now
\begin{eqnarray*}
X &=&
\RE((1+|z|^2)a_1+2(\overline{a_0}+a_2)z + \sum_{k=3}^\infty a_k z^{k-1}(k-(k-2)|z|^2) \\
&=& \RE(2 a_0 \overline z+(1+|z|^2)a_1+2a_2z + \sum_{k=3}^\infty a_k z^{k-1}(k-(k-2)|z|^2) \\
&=& \RE\left(2\left(a_0 \overline z + \sum_{k=1}^\infty a_k z^{k-1}|z|^2\right)+(1-|z|^2)\sum_{k=1}^\infty ka_k z^{k-1}\right)\\
&=& 2\RE(\overline z G(z))+(1-|z|^2)\RE G'(z),
\end{eqnarray*}
giving condition (\ref{B}).
\endpf
\begin{prop}\label{prop:tinto}
Let $(\beta_n)_n$ be a decreasing sequence of positive reals such that $\liminf_{n\to\infty}|\beta_n|^{1/n}\geq 1$ and
let $G\in H^2(\beta)$ such that
\[\mathop{\rm ess}\sup_{w\in \TT}\RE (\overline{w}G(w))>0.\]
Then \[\sup\RE \{\langle Af,f\rangle :f\in D(A),\|f\|_{H^2(\beta)}=1\}=+\infty,\]
where $A$ is defined on $D(A)=\{f\in H^2(\beta):Gf'\in H^2(\beta)\}$ by $Af=Gf'$.
\end{prop}
Before proceeding to the proof, we state the following technical lemma which explains the hypothesis on monotonicity of $(\beta_n)_n$.
\begin{lem}\label{lem:decreasing}
Let $(\beta_n)_n$ be a decreasing sequence of positive reals. Then for all positive integer $N$, there exists $\eta=\eta(N)>0$ such that for all $z\in\DD$ with
$|w|>1-\delta$, we have
\[\sum_{n=0}^N\frac{|w|^{2n}}{\beta_n^2}< \sum_{n=N+1}^\infty \frac{|w|^{2n}}{\beta_n^2}.\]
\end{lem}
\beginpf
Since $(1/\beta_n)_n$ is increasing, we have
\[\sum_{n=0}^N\frac{|w|^{2n}}{\beta_n^2}\leq \frac{1}{\beta_N^2}(1+|w|^2+\cdots +|w|^{2N})=\frac{1}{\beta_N^2}\left( \frac{1-|w|^{2N+2}}{1-|w|^2}\right).\]
On the other hand, we have
\[\sum_{n=N+1}^\infty\frac{|w|^{2n}}{\beta_n^2}\geq \frac{1}{\beta_{N+1}^2}\sum_{n=N+1}^\infty |w|^{2n}=\frac{|w|^{2N+2}}{\beta_{N+1}^2(1-|w|^2)}\geq
\frac{|w|^{2N+2}}{\beta_N^2(1-|w|^2)}.\]
Since $1-|w|^{2N+2}<|w|^{2N+2}$ is equivalent to $|w|>(1/2)^{1/(2N+2)}$, for all $w\in\DD$ such that $|w|>\eta(N)$ with $\eta(N)=1-(1/2)^{1/(2N+2)}$, we have
\[\sum_{n=0}^N\frac{|w|^{2n}}{\beta_n^2}< \sum_{n=N+1}^\infty
\frac{|w|^{2n}}{\beta_n^2}.\]
\endpf
\noindent\textbf{Proof of Proposition~\ref{prop:tinto}}:
By hypothesis, there exists $\delta>0$ and a sequence
$(w_k)_k\subset \DD$ such that $|w_k|\to 1$ and $\RE (\overline{w_k}G(w_k))\geq \delta$.
Moreover the condition $\liminf_{n\to\infty}|\beta_n|^{1/n}\geq 1$ guarantees that the space $H^2(\beta)$ has reproducing kernels
$k_w$ for all $w\in\DD$. Now consider the sequence $(\widehat{k_{w_k}})_k$
of normalized reproducing kernels associated with $(w_k)_k$,
i.e. $\widehat{k_{w_k}}=\frac{k_{w_k}}{\|k_{w_k}\|_{H^2(\beta)}}$.
First assume that $k_{w_k}\in D(A)$. In this case,
the remainder of the proof consist in checking that
\[\lim_{k\to \infty}\RE\left( \langle A\widehat{k_{w_k}}, \widehat{k_{w_k}}\rangle_{H^2(\beta)} \right)=+\infty.\]
Note that
\[\langle A\widehat{k_{w_k}}, \widehat{k_{w_k}}\rangle_{H^2(\beta)} =\langle G(\widehat{k_{w_k}})', \widehat{k_{w_k}}\rangle_{H^2(\beta)} =
\frac{1}{\|k_{w_k}\|^2_{H^2(\beta)}}G(w_k)k'_{w_k}(w_k),\]
where $k'_{w_k}(z)=\sum_{n\geq 1}\frac{n\overline{w_k}^n}{\beta_n^2}z^{n-1}$. It follows that
\[\langle Ak_{w_k},k_{w_k}\rangle_{H^2(\beta)}=\sum_{n\geq 1}\frac{nG(w_k)\overline{w_k}|w_k|^{2(n-1)}}{\beta_n^2},\]
and thus
\[\langle A\widehat{k_{w_k}}, \widehat{k_{w_k}}\rangle_{H^2(\beta)} =\frac{\overline{w_k}G(w_k)}{|w_k|^2}
\frac{\sum_{n\geq 1}\frac{n|w_k|^{2n}}{\beta_n^2}}{\sum_{n\geq 0}\frac{|w_k|^{2n}}{\beta_n^2}}.\]
Now, for all positive integer $N$, take $\eta(N)$ as in Lemma~\ref{lem:decreasing}, and $k$ sufficiently large so that $|w_k|>1-\eta(N)$.
Then we have
\begin{eqnarray*}
\frac{\sum_{n\geq 1}\frac{n|w_k|^{2n}}{\beta_n^2}}{\sum_{n\geq 0}\frac{|w_k|^{2n}}{\beta_n^2}} & = &
\frac{\sum_{n=0}^N\frac{n|w_k|^{2n}}{\beta_n^2}+ \sum_{n=N+1}^\infty \frac{n|w_k|^{2n}}{\beta_n^2}}{\sum_{n\geq 0}^N
\frac{|w_k|^{2n}}{\beta_n^2} + \sum_{n=N+1}^\infty\frac{|w_k|^{2n}}{\beta_n^2}}\\
& \geq & \frac{(N+1)\sum_{n=N+1}^\infty \frac{|w_k|^{2n}}{\beta_n^2}}{2\sum_{N+1}^\infty\frac{|w_k|^{2n}}{\beta_n^2}}=\frac{N+1}{2}.
\end{eqnarray*}
Therefore, for $k$ sufficiently large (so that $|w_k|>1-\eta(N)$), we get
\[\RE\left( \langle A\widehat{k_{w_k}}, \widehat{k_{w_k}}\rangle_{H^2(\beta)} \right)\geq \frac{(N+1)}{2|w_k|^2}\RE(\overline{w_k}G(w_k)).\]
Since $\RE (\overline{w_k}G(w_k))\geq \delta$ and since $|w_k|$ tends to $1$, we get the desired conclusion.
If $k_{w_k}$ is not in $D(A)$, the conclusion follows from similar
calculation, considering the sequence of polynomials
$(k^M_{w_k})_{M\geq 0}$
defined by
\[ k^M_{w_k}=\sum_{n=0}^M\frac{\overline{w_k^n}}{\beta_n^2}z^n,\]
which belongs to $D(A)$ and tends to $k_{w_k}$ in $\D$.
\endpf
We are now ready for the main theorem of this section.
\begin{thm}\label{thm:hardy}
Let $G \in H^2(\DD)$ and $A$ the operator $f \mapsto Gf'$, defined on the
domain $D(A)=\{f \in H^2(\DD): Gf' \in H^2(\DD)\}$ which is dense in $H^2(\DD)$. Then the following conditions are equivalent:\\
(i) $A$ generates a $C_0$-semigroup of composition operators on $H^2(\DD)$;\\
(ii) $2\RE \overline z G(z) + (1-|z|^2) \RE G'(z) \le 0$ for $z \in \DD$;\\
(iii) $A$ generates a quasicontractive $C_0$-semigroup on $H^2(\DD)$;\\
(iv) $\esssup_{z \in \TT} \RE \overline z G(z) \le 0$.
\end{thm}
\beginpf
The equivalence between (i) and (ii) is Theorem~\ref{ThmB}.
The implication (i)$\Rightarrow$(iii) is Corollary~\ref{cor:quasicont}.
The implication (iii)$\Rightarrow$(iv) follows immediately from Proposition~\ref{prop:tinto} with $\beta_n=1$ for all $n$.
Finally, the implication (iv)$\Rightarrow$(ii) is Proposition~\ref{prop:CimpliesB}.
\endpf
Let $A$ be defined on $D(A):=\{f\in H^2(\DD): Gf'\in H^2(\DD)\}$ by $A f(z) = G(z) f'(z)$ where $G(z)=\sum_{n=0}^{\infty}\alpha_nz^n$.
An easier condition than Condition (\ref{B}) to test is the following:
\begin{equation}\label{A}
\RE(\alpha_1) +|\overline{\alpha_0} +\alpha_2| + \sum_{n=3}^{\infty}|\alpha_n|\leqslant 0.
\end{equation}
In the sequel we present the link between Condition (\ref{A}) and
Condition (\ref{B}).
\begin{prop}
\begin{enumerate}[(i)]
\item Condition (\ref{A}) implies Condition (\ref{B}).
\item If $G\in \CC_2[X]$ (i.e., a polynomial of degree at most 2), then conditions (\ref{B}) and (\ref{A}) are equivalent.
\item There exists a polynomial function of degree $3$ such that condition (\ref{B}) holds and condition (\ref{A}) does not.
\end{enumerate}
\end{prop}
\beginpf
\begin{enumerate}[(i)]
\item The condition (\ref{B}) is equivalent to
\[ (1+|z|^2)\RE(\alpha_1) +2\RE((\overline{\alpha_0}+\alpha_2)z) + \sum_{n=3}^\infty \RE(\alpha_n((2-n)|z|^2+n)z^{n-1})\leqslant 0.\]
The condition (\ref{A}) is
\[\RE(\alpha_1) +|\overline{\alpha_0} +\alpha_2| + \sum_{n=3}^{\infty}|\alpha_n|\leqslant 0,\]
that is,
\[(1+|z|^2)\RE(\alpha_1) +(1+|z|^2)|\overline{\alpha_0} +\alpha_2| + \sum_{n=3}^{\infty}(1+|z|^2)|\alpha_n|\leqslant 0.\]
Note that $2\RE((\overline{\alpha_0}+\alpha_2)z)\leqslant 2|\overline{\alpha_0} +\alpha_2||z|\leqslant(1+|z|^2)|\overline{\alpha_0} +\alpha_2|$. On the other hand, the arithmetico-geometric inequality gives, $\forall k\in\NN^*,\ \forall x\in\RR_+$,
\[ \frac{1+x^2+(k-1)x^{k+2}}{k+1} \geqslant \sqrt[k+1]{1\times x^2\times x^{(k+2)(k+1)}}=x^k,\]
i.e. \[ 1+x^2 \geqslant x^k((k+1)-(k-1)x^2).\]
We now observe that
\[\RE( \alpha_n((2-n)|z|^2+n)z^{n-1}) \leqslant |\alpha_n||z^{n-1}|((2-n)|z|^2+n) \leqslant (1+|z|^2)|\alpha_n|.\]
\item Let $G(z) = \alpha_0+\alpha_1z+\alpha_2z^2$.\\
If condition (\ref{B}) is true, we have in particular
\[\forall \theta\in\RR,\ 2\RE(e^{-i\theta}G(e^{i\theta}))\leqslant 0\quad \text{ i.e.}\quad \RE(\alpha_1+(\overline{\alpha_0}+\alpha_2)e^{i\theta})\leqslant 0.\]
For $\theta=-\operatorname{arg}(\overline{\alpha_0}+\alpha_2)$, we get \[\RE(\alpha_1)+|\overline{\alpha_0}+\alpha_2|\leqslant 0.\]
\item Take $G(z)=-z+\frac{z^2}{\sqrt{3}}-\frac{z^3}{\sqrt{3}}$. Note
that $G$ does not satisfy condition (\ref{A}). On the other hand,
note that
$G(z)=-z(1-\frac{z}{\sqrt{3}}+\frac{z^2}{\sqrt{3}})=-zF(z)$. In
\cite{B_P}, it is shown that, if $\RE(F)\geqslant 0$, then
$G$ generates a $C_0$-semigroup of composition operators on
$H^2(\beta)$, which is equivalent to condition (\ref{B}). Since
$\RE(F)$ satisfies the maximum principle, for
$h(\theta):=\RE(F(e^{i\theta})=1-\frac{1}{\sqrt{3}}\cos(\theta)+\frac{1}{\sqrt{3}}\cos(2\theta)$,
$F$ maps the disc into the right-half plane if $h$ is
nonnegative. For that purpose, note that
\[ h'(\theta)=\frac{1}{\sqrt{3}}\sin(\theta)(1-4\cos(\theta)).\]
It follows that $h'(\theta)=0 \Leftrightarrow \theta =0 \text{ or } \theta = \pi
\text{ or } \cos(\theta)=\frac 14$. A direct computation gives: $h(0)=1$, $h(1)=1+\frac{2}{\sqrt{3}}$ and if $\cos(\theta)=\frac 14$, $h(\theta)=1-\frac{9}{8\sqrt{3}}>0$. Therefore $G$ satisfies condition (\ref{B}).
\end{enumerate}
\endpf
\section{Quasicontractive semigroups on the Dirichlet space}\label{sec:4}
Recall that the Dirichlet space norm is defined by
\begin{equation}\label{eq:sep29}
\|f\|^2_\D = |a_0|^2 + \sum_{k=1}^\infty k|a_k|^2 = |f(0)|^2 + \int_\DD |f'(z)|^2 \, dA(z),
\end{equation}
for $f(z)=\sum_{k=0}^\infty a_k z^k$,
and it is induced by an inner product that may be written, at least formally, as
\[
\langle f,g \rangle_\D = \langle f,zg' \rangle_{H^2(\DD)} + f(0)\overline{g(0)}.
\]
\begin{prop}\label{prop:doc5}
For $G \in \D$ and $A$ the operator $f \mapsto Gf'$, defined on the
domain $D(A)=\{f \in \D: Gf' \in \D\}$, which in dense in $\D$, the following two conditions are equivalent:\\
(i) $\esssup_{z \in \TT} \RE \overline z G(z) \le 0$;\\
(ii) $\sup\{ \RE \langle Af,f \rangle_\D : f \in D(A), \|f\|_\D=1 \} < \infty$.
\end{prop}
\beginpf
Suppose that $\esssup_{z \in \TT} \RE \overline z G(z) \le 0$.
Then
\begin{eqnarray*}
\RE \langle Af,f\rangle_\D &=&
\RE \langle Gf',zf'\rangle_{H^2(\DD)} + \RE\left( G(0)f'(0)\overline{f(0)} \right)\\
&=&
\RE \left( \frac{1}{2\pi}\int_0^{2\pi} G(z) \overline z |f'(z)|^2 \, d\theta\right) + \RE\left( G(0)f'(0)\overline{f(0)}\right),
\end{eqnarray*}
with $ z=e^{i\theta}$,
and the supremum of this quantity over $\|f\|_\D=1$, $f\in D(A)$ is clearly finite.\\
Conversely, suppose that $\esssup \RE \overline z G(z) > 0$.
By considering an (outer) function $u$ with $|u|=1$ on a set of positive measure where $\RE \overline zG(z) > \delta > 0$ and
$|u|=1/2$ on its complement (see Thm. 4.6 in \cite{Garnett})
we see that $\liminf_{n \to \infty}\RE \langle Gu^n,zu^n \rangle_{H^2(\DD)} > 0$.
It now follows that there is a function $f \in D(A)$ with $\langle f,f \rangle_\D=1$
and $\RE \langle Gf',zf' \rangle_{H^2(\DD)} > 0$.
Now define a sequence $(f_k)_k$ in $\D$ by setting $f_k'=z^k f'$ and $f_k(0)=0$.
So if $f(z)=\sum_{n=0}^\infty a_n z^n$, then
\[
f_k(z)= \sum_{n=1}^\infty a_n \frac{n}{n+k} z^{n+k},
\]
and hence
\begin{eqnarray*}
\langle f_k , f_k \rangle_\D &=& \sum_{n=1}^\infty |a_n|^2 \left( \frac{n}{n+k} \right)^2 (n+k) \\
&=&
\sum_{n=1}^\infty |a_n|^2 \frac{n^2}{n+k} \le \|f\|^2_\D,
\end{eqnarray*}
and thus this tends to zero by the dominated convergence theorem.
Now
\[
\RE \langle Af_k,f_k \rangle_\D = \RE \langle \overline z G f'_k,f'_k \rangle_{H^2(\DD)} = \RE \langle G f',f' \rangle_{H^2(\DD)}
\]
On normalizing the functions $f_k$ we see that
\[
\sup\{ \RE \langle Af,f \rangle_\D : f \in \D, \|f\|_\D=1 \} = \infty.
\]
\endpf
\begin{prop}\label{prop:univalent}
Let $G \in\D$ and $A$ the operator $f \mapsto Gf'$, defined on the
domain $D(A)=\{f \in \D: Gf' \in \D\}$, which is dense in $\D$. If $A$ generates a $C_0$-semigroup of composition operators on $\D$, then
this semigroup is quasicontractive.
\end{prop}
\beginpf
Given a semigroup $(C_{\varphi_t})_{t\geq 0}$ acting on $\D$, we must
show that $\|C_{\varphi_t} f\|_\D = \|f\|_\D (1+ O(t))$ for small $t>0$.
First, since $\phi_t$ is injective, we have the well-known inequality
\begin{eqnarray*}
\int_\DD |(f \circ \phi_t)'(z)|^2 \, dA(z) &=& \int_\DD |(f' \circ \phi_t(z))|^2 |\phi'_t(z)|^2 \, dA(z) \\
&= & \int_{\phi(\DD)} |f'(w)|^2 \, dA(w) \le \int_\DD |f'(w)|^2 \, dA(w),
\end{eqnarray*}
taking $w=\phi(z)$. Therefore the composition operator $C_{\phi_t}$ is
bounded on $\D$. Moreover, by \cite[Thm.~2]{martin05},
\[\|C_{\varphi_t}\|\leq \sqrt{ \frac{L+2+\sqrt{L(4+L)}}{2}}
,\]
where $L=\log\left( \frac{1}{1-|\varphi_t(0)|^2} \right)$. This upper
bound is sharp since it is an equality whenever $\DD\setminus
\varphi_t(\DD)$ is of Lebesgue area measure equal to $0$.
Siskakis \cite{sisk96} proved that, as in the case of the Hardy space, $A$ is of the form $A(f)=G(z)f'(z)$,
where $G$ is an holomorphic function on $\DD$ and
$\varphi_t(z)=z+G(z)t+o(t)$. It follows that, for $t\to 0$,
\[\|C_{\varphi_t}\|\leq 1+O(t),\]
since $L=O(t^2).$
Therefore, there exists $w>0$ such that $\|C_{\varphi_t}\|\leq e^{wt}$
for all $t\geq 0$, and thus $(C_{\varphi_t})_{t\geq 0}$ is then a
quasicontractive $C_0$-semigroup.
\endpf
\begin{thm}\label{thm:dir}
Let $G \in\D$ and $A$ the operator $f \mapsto Gf'$, defined on the
domain $D(A)=\{f \in \D: Gf' \in \D\}$, which is dense in $\D$. Then the following conditions are equivalent:\\
(i) $A$ (extended to its natural domain in $H^2(\DD)$) generates a $C_0$-semigroup of composition operators on $H^2(\DD)$;\\
(ii) $2\RE \overline z G(z) + (1-|z|^2) \RE G'(z) \le 0$ for $z \in \DD$;\\
(iii) $A$ (extended to its natural domain in $H^2(\DD)$) generates a quasicontractive $C_0$-semigroup on $H^2(\DD)$;\\
(iv) $\esssup_{z \in \TT} \RE \overline z G(z) \le 0$;\\
(v) $\sup\{ \RE \langle Af,f \rangle_\D : f \in D(A), \|f\|_\D=1 \} < \infty$;\\
(vi) $A$ generates a quasicontractive $C_0$-semigroup on $\D$;\\
(vii) $A$ generates a $C_0$-semigroup of composition operators on $\D$.
\end{thm}
\beginpf
Conditions (i)--(iv) have already been shown to be equivalent in Theorem~\ref{thm:hardy}.
The equivalence of conditions (iv) and (v) is shown in Proposition~\ref{prop:doc5}.
For (i)$\Rightarrow$(vii) is detailed in \cite{sisk96}.
The fact that (vii)$\Rightarrow$(vi) is given in Proposition~\ref{prop:univalent}.
Finally, (vi)$\Rightarrow$(v) by Lumer--Phillips result (see
Corollary~\ref{cor:lm} ).
\endpf
\section{Comments}
In \cite{B_P}, as well as in Condition~\ref{A}, the description of the
generator of a $C_0$-semigroup of composition operators
relies on analytic
functions $F$ or $\tilde{G}$ which map $\DD$ into the right or left half-plane.
For that purpose, let us recall the Carath\'eodory--Toeplitz theorem \cite{B_C,S}.
\begin{thm}[Carath\'eodory--Toeplitz]
Let $f(z)=\sum_{n=0}^\infty \mu_nz^n$ and consider for $k\geqslant 1$ the matrices $M_k=(m_{i,j})_{1\leqslant i,j \leqslant k}$ where $m_{i,j}=\mu_{j-i}$ if $i\leqslant j$ and $m_{i,j}=0$ otherwise. Then $f$ maps the disc to the right half plane if and only if the Hermitian matrix $N_k=M_k+\overline{M_k'}$ is nonnegative definite for all $k\geqslant 1$.
\end{thm}
This theorem has to be considered with the Sylvester Criterion.
\begin{thm}[Sylvester Criterion]
Let $A=\big(a_{ij}\big)_{1\leqslant i,j\leqslant n}$ be Hermitian. Then $A$ is positive definite if and only if the $n$ matrices $A_p=\big(a_{ij}\big)_{1\leqslant i,j\leqslant p}$ with $1\leqslant p \leqslant n$ have positive determinant.
\end{thm}
Here is an example where we can use those tools.
\begin{ex}
Let $G(z)=a_0+a_1z+a_2z^2 \in\CC_2[X]$, thanks to condition (\ref{C}), we have that $G$ generates a $C_0$-semigroup of analytic functions on $\DD$ if and only if $\sup(\tilde{G}(\T))\leqslant 0$. Besides, $\sup(\tilde{G}(\TT))< 0$ if and only if
\[\det \begin{pmatrix} -\RE(a_1) \end{pmatrix} >0,\quad \det\begin{pmatrix}
-\RE(a_1) & -(\overline{a_0}+a_2) \\ -(\overline{a_2}+a_0)&-\RE(a_1)
\end{pmatrix} >0\]
i.e.
\[ \left\{ \begin{array}{l}
\RE(a_1)<0\\
\RE(a_1)^2-|\overline{a_0}+a_2|^2=(\RE(a_1)-|\overline{a_0}+a_2|)(\RE(a_1)+|\overline{a_0}+a_2|)>0
\end{array}
\right.\]
i.e.
\[ \RE(a_1)+|\overline{a_0}+a_2|<0.\]
We have recovered condition (\ref{A}).
\end{ex}
One may wonder if the quasicontractive $C_0$-semigroup whose generator
is given by $A$ can be determined on $H^2(\DD)$ or $\D$. We know that
it is a semigroup of composition operators $C_{\varphi_t}$, with
\[
\frac{\partial \varphi_t(z)}{\partial t}=G(\varphi_t(z)).
\]
This is an important and not so easy issue, which can be answered in
some particular cases, as follows. Those examples are slight
generalizations of the ones presented in \cite{AS}.
\begin{enumerate}[(i)]
\item If $G(z)=az+b$ with $a\neq 0$, we have
\[\varphi_t(z)= e^{at}z+\frac{b}{a}(e^{at}-1).\]
Furthermore, the Denjoy--Wolff point $\alpha$ of this holomorphic semigroup is $\alpha = -\frac ba\in\DD$.
\item When \emph{$G$ is a polynomial of degree $2$}, defined by $G(z)=c(z-a)(z-b)$:
\begin{itemize}
\item If $a\neq b$, we get
\[
\varphi_t(z) = \frac{z(ae^{bct}-be^{act})+ab(e^{act}-e^{bct})}{z(e^{bct}-e^{act})+(ae^{act}-be^{bct})},
\]
whose Denjoy--Wolff point is $\alpha = a\in\DD$ if $\RE a<\RE b$ and
$\alpha = b\in\DD$ if $\RE a>\RE b$.
In the case where $\RE a =\RE b$ it happens that $\varphi_{t_n}=Id$ for $t_n=\frac{2\pi n}{\IM a-\IM b}$, so $\varphi_t$ is an automorphism.
\item If $a=b$, we find another expression for $\varphi_t$:
\[
\varphi_t(z) = \frac{z(1-act)+a^2ct}{-zct+(1+act)},
\]
whose Denjoy--Wolff point is $\alpha=a$.
\end{itemize}
\item As \emph{$G$ is polynomial of higher degree}, we usually do not have explicit expression of the semigroup $(\varphi_t)$. Yet, some cases can be found:
\begin{itemize}
\item If $G(z)=c(z-a)^n$ then $\forall t\in\RR_+,\forall z\in\DD$,
\[\varphi_t(z)=a+\frac{z-a}{(1-nct(z-a)^{n-1})^{\frac{1}{n-1}}}.\]
Note that, if $c=1$, the only possible case is when $a=1$.
\item
If $G(z)=cz(z^n-a)$ then $\forall t\in\RR_+,\forall z\in\DD$,
\[\varphi_t(z)=\frac{ze^{-ct}}{\left(1-z^n\left(\frac{1-e^{-nct}}{a}\right)\right)^{\frac{1}{n}}}.\]
\end{itemize}
\end{enumerate}
\subsection*{Acknowledgement}
The authors are grateful to the referee for drawing their
attention to the book of Shoikhet \cite{shoikhet}, and hence
indirectly to \cite{irish}.
|
\section{Introduction}
The main innovation of~\cite{CG} was to develop a framework for modularity
lifting theorems in contexts in which the Taylor--Wiles method did not apply. One
of the main examples in~\cite{CG} (Theorem~1.4) was a
minimal modularity lifting theorem for odd two-dimensional Galois representations
which are unramified at $p$. This result was simultaneously a generalization and a
specialization of the main theorem of Buzzard--Taylor~\cite{BuzzT}; generalized because it
related Galois representations modulo~$\varpi^n$ to Katz modular forms of weight one modulo~$\varpi^n$
neither of which need lift to characteristic zero,
and specialized because it required a minimality hypothesis at primes away
from $p$.
One of the goals of the present paper is to provide a theorem which is a new proof of many cases
of~\cite{BuzzT} in the spirit of~\cite{CG}. Our methods could be viewed as hybrid of both~\cite{CG}
and~\cite{BuzzT} in the following sense: as in~\cite{CG}, we prove an integral~$R=\mathbf{T}$ theorem for torsion
representations by working directly in weight one, however, as a crucial input, we use
ordinary modularity lifting theorems in higher weight (as in~\cite{BuzzT}, although we only need to work
in weight~$p$) in order to show that the patched Hecke modules see every component of the generic
fibre of the global deformation ring.
In order to simplify some of our arguments, we do not
strive for maximal generality. The assumption that the representations are unramified at~$p$, however, seems essential for the method (if one does not use base change), in contrast to~\cite{wild}.
Let us fix a prime $p > 2$ and a local field $[E:\mathbf{Q}_p] < \infty$ with ring of integers $\mathcal{O}$ and
residue field~$k = \mathcal{O}/\varpi$.
\begin{theorem} \label{theorem:example}
Let $p > 2$, and let
$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\mathcal{O})$
be a continuous odd Galois representation ramified at finitely many primes and unramified at $p$.
Suppose that $\overline{\rho}$ is absolutely irreducible. If $\rho$ is ramified at a prime $\ell$,
assume that
$\rho|D_{\ell}$ is reducible.
Then $\rho$ is modular of weight one.
\end{theorem}
This result will be deduced from our main result, which is an integral
$R = \mathbf{T}$ theorem which we now describe.
Let
$$\overline{\rho}: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(k)$$ be a continuous absolutely irreducible
odd representation unramified at~$p$.
For each $\ell$, let $\epsilon$ denote the cyclotomic character. Let $\psi$ denote the Teichmuller lift
of $\det(\overline{\rho})$.
Let $N = S \cup P$ be a set of primes not containing~$p$ such
that $\psi$ is ramified exactly at the primes contained in $P$ and unramified at primes contained $S$.
By abuse of notation, we also let~$N$ denote the product of the conductor of~$\psi$ with the primes in~$S$.
We consider the functor~$D^1$ from complete local Noetherian
$\mathcal{O}$-algebras $(A,\mathfrak{m})$ with residue field $k$ defined (informally) as follows.
Fix a collection of elements~$a_{\ell} \in k$ for~$\ell$ dividing~$N$. Let
$D^1(A)$ consist of deformations $\rho$ to $A$ together with a collection of elements $\alpha_{\ell} \in A$
for $\ell \in N$ such that:
\begin{enumerate}
\item $\det(\rho) = \psi$.
\item $\rho$ is unramified outside $N = S \cup P$.
\item If $\ell \in P$, then
$\rho |D_{\ell} \simeq
\chi^{-1} \psi |_{D_{\ell}} \oplus \chi$ for some
unramified character $\chi$ with $\chi(\mathrm{Frob}_{\ell}) = \alpha_{\ell} \equiv a_{\ell} \mod \mathfrak{m}$.
\item If $\ell \in S$, then
$\rho | D_{\ell} \simeq \displaystyle{ \left( \begin{matrix} \chi^{-1} \psi|_{D_{\ell}} & * \\
0 & \chi \end{matrix} \right)}$
for some
unramified character $\chi$ with $\chi(\mathrm{Frob}_{\ell}) = \alpha_{\ell}
\equiv a_{\ell} \mod \mathfrak{m}$.
\end{enumerate}
In fact, the actual definition of~$D^1(A)$ needs to be somewhat modified
(see~\S\ref{section:definitions} for precise definitions),
but this description will be valid for rings of integers such as~$\mathcal{O}$.
Naturally enough, we also assume that $D^1(k)$ is non-empty, and
that $k$ is also
large enough to contain
the eigenvalues of every element in the image of $\overline{\rho}$.
The elements $a_{\ell} \in k$ are determined by $\overline{\rho}$ for primes
in $P$, but not necessarily for primes in~$S$, because when~$\overline{\rho}$
is unramified at~$\ell$, there is a choice of eigenvalue for the unramified line.
Hence $D^1$ will not strictly be a Galois deformation ring;
we refer to such rings (and we have several in this paper) as modified
deformation rings because they depend not only on~$\overline{\rho}$ but also
on some auxiliary data.
The functor $D^1$ is representable by a complete local $\mathcal{O}$-algebra $R^1$.
The ring $R^1$ comes with elements $a_{\ell} \in k$ and~$\alpha_{\ell} \in R^1$ for $\ell$ dividing~$N$. For all other
primes $\ell$, define $a_{\ell} \in k$ to be $\Tr(\overline{\rho}(\mathrm{Frob}_{\ell}))$, including when
$\ell = p$.
We are now ready to state out main theorem. By abuse of notation, let $N$ denote the conductor of $\psi$ times
the primes in $S$ --- it is divisible exactly by the primes in $N = S \cup P$.
Let $X_H(N)$ denote the quotient of $X_1(N)$ by the Sylow $p$-subgroup of $(\mathbf{Z}/N\mathbf{Z})^{\times}$.
After enlarging~$S$ if necessary, we may assume that the curve~$X_H(N)$ is a fine moduli space.
($X_H(N)$ will be automatically a fine moduli space if~$p > 3$, see~\S\ref{section:stacks}.)
\begin{theorem} \label{theorem:main} Let $p > 2$.
Let $\mathbf{T} \subset \mathrm{End}_{\mathcal{O}} H^0(X_H(N),\omega_{E/\mathcal{O}})$ be the $\mathcal{O}$-algebra
generated by Hecke endomorphisms.
Let $\mathfrak{m}$ be the maximal ideal of $\mathbf{T}$ generated by the elements
$\langle \ell \rangle - \psi(\ell)$ for $\ell \nmid N$ and
$T_{\ell} - a_{\ell}$ for all $\ell$. Then there is an isomorphism
$R^1 \simeq \mathbf{T}_{\mathfrak{m}}$.
\end{theorem}
\subsection{Theorem~\ref{theorem:main} implies Theorem~\ref{theorem:example}}
Suppose that $\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\mathcal{O})$ is a continuous
Galois representation ramified at $\ell \ne p$ satisfying the conditions of
Theorem~\ref{theorem:example}. Then after a global twist and enlarging $\mathcal{O}$ if necessary
to contain a choice of elements $a_{\ell}$ for $\ell$ dividing $N$, $\rho$ gives rise to an
element of $D^1(\mathcal{O})$.
The modularity of $\rho$ then follows.
\medskip
As an application of Theorem~\ref{theorem:main}, we prove the following:
\begin{theorem} \label{theorem:blog} Let $p$ be any prime. There exists
a Katz modular form $f \in H^0(X(\Gamma)_{\mathbf{F}_p},\omega)$ for some level $\Gamma$ prime to $p$ which
does not lift to characteristic zero.
\end{theorem}
The original argument of Wiles~\cite{W,TW} for modularity theorems at non-minimal level was to use an induction
argument and a certain numerical criterion involving complete intersections which were
finite over $\mathcal{O}$. This does not seem
to be obviously generalizable to weight one --- although one still has access to forms of Ihara's Lemma,
the Hecke rings $\T^1$ are no longer complete intersections in general, and are certainly
not flat over $\mathcal{O}$. It remains open as to whether one can proceed using such an argument.
Instead, we use modularity theorems in weight~$p$ in order to show $R^1_{Q}/\varpi \simeq
\T^1_{Q,\mathfrak{m}}/\varpi$ for various sets of auxiliary primes~$Q$, and we then use this information to show
that the patched Hecke modules in weight one are ``big enough.'' To pass between weight one and
weight~$p$ we \emph{crucially} rely on $q$-expansions.
For this reason, the methods of this paper will probably not be generalizable beyond $\mathrm{GL}(2)$
(although they may have implications for Hilbert modular forms of partial weight one).
Note that, in writing the paper~\cite{CG}, we tried to avoid the use of $q$-expansions as much as possible, whereas
the philosophy of this paper is quite the opposite.
\begin{remark} \emph{The methods of our paper may well be able to handle
more precise local deformation conditions than those considered above.
However,
these assumptions considerably simplify some aspects of the arguments. We particularly shun
Diamond's vexing primes, which did indeed cause considerable vexation in~\cite{CG}. In fact,
we try so hard to avoid them that we assume that $\overline{\rho} |D_{\ell}$ is reducible, when certainly
some such representations --- for example those with $\overline{\rho} | I_{\ell}$ irreducible --- may well
be amenable to our methods.}
\end{remark}
\subsection{Acknowledgements}
The debt this paper owes to~\cite{CG} is clear, and the author thanks David Geraghty for many conversations.
We thank
Mark Kisin for the explaining a proof of
Lemma~\ref{lemma:kisin}, and we also thank Brian Conrad for a related proof of
the same result
in the context of rigid analytic geometry.
We thank Toby Gee and Patrick Allen for several useful comments.
We also thank Gabor Wiese for the original idea of proving modularity theorems in weight~one by
working in weight~$p$.
\section{Preliminaries}
\label{section:prelim}
\subsection{Local modified deformation rings}
Let~$\overline{\rho}: G_{\mathbf{Q}_{\ell}} \rightarrow \mathrm{GL}_2(k)$ be a representation, and let~$R^{\univ}_{\ell}$
denote the universal framed local deformation ring, and~$\rho^{\mathrm{univ}}$ the universal local deformation.
We assume in this entire section that~$p > 2$. The calculations
in this section will mostly be concerned with the case that~$\ell \ne p$.
Fix a lift of Frobenius~$\phi \in G_{\mathbf{Q}_{\ell}}$, and choose an eigenvalue~$a_{\ell}$
of~$\overline{\rho}(\phi)$, which, after enlarging~$k$ if necessary, we may assume to lie in~$k$.
We define the universal modified framed local deformation ring
$\widetilde{R}^{\univ}_{\ell}$ to be the localization
of the ring
$$R^{\univ}_{\ell}[\alpha_{\ell}]/(\alpha^2_{\ell} - \alpha_{\ell} \mathrm{Tr}(\rho^{\mathrm{univ}}(\phi))
+ \det(\rho^{\mathrm{univ}}(\phi)))$$
at~$(\alpha_{\ell} - a_{\ell})$. The quadratic polynomial satisfied by~$\alpha_{\ell}$
is the characteristic polynomial of Frobenius.
\begin{lemma} \label{lemma:hensel} If~$\overline{\rho}(\phi)$ has distinct eigenvalues, then $\widetilde{R}^{\univ}_{\ell} \simeq R^{\univ}_{\ell}$.
If~$\overline{\rho}(\phi)$ does not have distinct eigenvalues, then $\widetilde{R}^{\univ}_{\ell}$ is a finite flat
extension of $ R^{\univ}_{\ell}$ of degree two.
\end{lemma}
\begin{proof} If~$\overline{\rho}(\phi)$ has distinct eigenvalues, then the characteristic polynomial
of Frobenius is separable over~$k$. Since~$R^{\univ}_{\ell}$ is complete, the
polynomial also splits over~$R^{\univ}_{\ell}$ by Hensel's Lemma, and the quadratic extension above is,
(before localization), isomorphic to $R^{\univ}_{\ell} \oplus R^{\univ}_{\ell}$. Localizing
at~$(\alpha_{\ell} - a_{\ell})$ picks out the factor on which we have the
congruence~$\alpha_{\ell} \equiv a_{\ell} \mod \mathfrak{m}$.
If the eigenvalues of~$\overline{\rho}(\phi)$ are both~$a_{\ell}$, then the quadratic extension is already local.
\end{proof}
A modified local deformation ring will simply be a quotient of~$\widetilde{R}^{\univ}_{\ell}$.
Proposition~3.1.2 of~\cite{gee} proves the
existence of quotients~$R^{\mathrm{univ},\psi,\tau}_{\ell}$
of~$R^{\univ}_{\ell}$
which are reduced, $\mathcal{O}$-flat, equidimension of dimension~$4$, and such that,
for any finite extension~$F/E$, a map
$$x: R^{\mathrm{univ}} \rightarrow F$$
factors through~$R^{\mathrm{univ},\psi,\tau}$ if and only if the
corresponding~$F$ representation~$V_x$
has determinant~$\psi$ and is of type~$\tau$. (For this section~$\psi$
may be any unramfied character.)
For our purposes, it will suffice to
consider the trivial type~$\tau$, which corresponds to representations on which
$$\rho_x: G_{\mathbf{Q}_{\ell}} \rightarrow \mathrm{GL}(V_x) = \mathrm{GL}_2(F)$$
restricted to the inertial subgroup~$I_{\ell} \subset G_{\mathbf{Q}_{\ell}}$ has unipotent (and so
possibly trivial) image.
\begin{lemma} \label{lemma:gee}
Suppose that~$p > 2$ and~$\ell \ne p$.
Let~$\tau$ denote the trivial type. There exists a quotient
$R^{\mathrm{mod},\psi}_{\ell} := \widetilde{R}^{\mathrm{univ},\psi,\tau}_{\ell}$ of~$\widetilde{R}^{\univ}_{\ell}$
which is reduced, $\mathcal{O}$-flat, equidimensional of dimension~$4$,
and such that, for any finite extension~$F/E$, a map
$$x: \widetilde{R}^{\univ}_{\ell} \rightarrow F$$
factors through~$R^{\mathrm{mod},\psi}_{\ell}$ if and only if the
corresponding~$F$ representation~$V_x$ has determinant~$\psi$,
is ordinary, and has an unramified quotient on which the action of~$\mathrm{Frob}_{\ell}$
is by the image of~$\alpha_{\ell}$.
\end{lemma}
The arguments are very similar to those already in the literature, but for want
of a reference which covers this case exactly, we give the details.
\begin{proof} Suppose that the eigenvalues of~$\overline{\rho}(\phi)$ are distinct.
Then, under the isomorphism~$R^{\univ}_{\ell} \simeq \widetilde{R}^{\univ}_{\ell}$,
we may take~$ \widetilde{R}^{\mathrm{univ},\psi,\tau}_{\ell} = R^{\mathrm{univ},\psi,\tau}_{\ell}$.
Hence we may assume that the eigenvalues are the same.
Any representation
$$\rho_x: G_{\mathbf{Q}_{\ell}} \rightarrow \mathrm{GL}_2(F)$$
for which the image of inertia has non-trivial unipotent image is, up to twist, and
enlarging~$F$ if necessary,
an extension of~$F$ by~$F(1)$. In particular, the ratios of the eigenvalues of~$\rho_x(\phi)$
must be equal to~$\ell$. Since we are assuming
the eigenvalues of~$\overline{\rho}(\phi)$ coincide,
then, if~$\ell \not \equiv 1 \mod p$,~$ R^{\mathrm{univ},\psi,\tau}_{\ell}$ has no such quotients, and
will consist precisely the unramified locus. In this case,
we may take $\widetilde{R}^{\mathrm{univ},\psi,\tau}_{\ell}$ to be the double cover corresponding
to the unramified locus with a choice of Frobenius eigenvalue.
Hence we may assume that~$\ell \equiv 1 \mod p$, and in
particular~$\ell \not\equiv -1 \mod p$.
Assume that~$\overline{\rho}$ is unramified.
The ring~$R^{\mathrm{univ},\psi,\tau}_{\ell}$
admits two natural quotients; a quotient~$R^{\mathrm{unr},\psi}_{\ell}$ corresponding to
representations which are unramified, and a quotient
corresponding to representations for which the ratios of the eigenvalues of Frobenius are equal to~$\ell$.
Because the determinant is fixed, this latter quotient is given by imposing the the equation
$$\Tr(\rho(\phi))^2 = \ell^{-1} (1 + \ell)^2 \psi(\ell).$$
Since~$\ell \not \equiv -1 \mod p$, the right hand side is a unit, and hence there is exactly
one square root of this equation which is compatible with the choice of~$a_{\ell}$, and so this is equivalent
to the equation
$$\Tr(\rho(\phi)) = \ell^{-1/2} (1 + \ell) \psi^{1/2}(\ell)$$
for the appropriate choice of square root.
The ring obtained by imposing this relation on~$R^{\mathrm{univ},\psi,\tau}_{\ell}$ may
or may not be either~$\mathcal{O}$-flat or reduced, but let~$R^{\mathrm{sp},\psi}_{\ell}$ denote
the largest quotient with this property ($\mathrm{sp}$ is for special). Its~$F$-points will still include all ramified
representations of type~$\tau$.
The pre-image of the corresponding affine scheme under
the projection $\mathrm{Spec}(\widetilde{R}^{\univ}_{\ell}) \rightarrow \mathrm{Spec}(R^{\univ}_{\ell})$ is equal to
$$R^{\mathrm{sp},\psi}_{\ell}[\alpha_{\ell}]/(\alpha^2_{\ell} - \ell^{-1/2} (1 + \ell) \psi^{1/2}(\ell) \cdot \alpha_{\ell} + \psi(\ell)).$$
The quadratic relation factors as
$$(\alpha_{\ell} - \psi^{1/2}(\ell) \cdot \ell^{1/2})(\alpha_{\ell} - \psi^{1/2}(\ell) \cdot \ell^{-1/2}).$$
Define~$\widetilde{R}^{\mathrm{sp},\psi}_{\ell}$ to be the quotient on which~$\alpha_{\ell} = \psi^{1/2}(\ell) \cdot \ell^{-1/2}$.
There is a corresponding isomorphism
$$R^{\mathrm{sp},\psi}_{\ell} \rightarrow \widetilde{R}^{\mathrm{sp},\psi}_{\ell}.$$
On the other hand, the quotient $R^{\mathrm{unr},\psi}_{\ell}$ is a formally smooth. In this case, we
let~$\widetilde{R}^{\mathrm{unr},\psi}_{\ell}$ be the finite flat degree two extension given by adjoining
an eigenvalue~$\alpha_{\ell}$ of the characteristic polynomial of Frobenius.
We now let~$R^{\mathrm{mod},\psi}_{\ell} = \widetilde{R}^{\mathrm{univ},\psi,\tau}$ be the image of~$\widetilde{R}^{\univ}_{\ell}$
under the map
$$\widetilde{R}^{\univ}_{\ell} \rightarrow \widetilde{R}^{\mathrm{unr},\psi}_{\ell} \oplus \widetilde{R}^{\mathrm{sp},\psi}_{\ell}.$$
It is~$\mathcal{O}$-flat and reduced because both~$ \widetilde{R}^{\mathrm{unr},\psi}_{\ell}$ and~$ \widetilde{R}^{\mathrm{sp},\psi}_{\ell}$
have this property. Moreover, the~$F$-points for finite extensions~$F/E$
correspond exactly
to either an unramified representation together with a choice of Frobenius, or a ramified ordinary
representation together with~$\alpha_{\ell}$ being sent to the action of Frobenius on the unramified
quotient. Geometrically, $R^{\mathrm{mod}}_{\ell}$ consists of the union of two components, one the special
component of~$R^{\mathrm{univ},\psi,\tau}_{\ell}$, and the other a double cover of the unramified
component of~$R^{\mathrm{univ},\psi,\tau}_{\ell}$.
We also remark that, by construction, the image
of the universal deformation ring~$R^{\mathrm{univ},\psi}$ in~$R^{\mathrm{mod}}_{\ell}$ will
be precisely~$R^{\mathrm{univ},\psi,\tau}$.
\end{proof}
We also note the following:
\begin{corr} \label{corr:non}
Suppose that~$p > 2$ and~$\ell \ne p$.
Let $x: \widetilde{R}^{\univ}_{\ell} \rightarrow E[\epsilon]/\epsilon^2$ be a surjective map so that the image
of~$R^{\univ}$ is~$E$, and the corresponding Galois representation~$\rho_x: G_{\mathbf{Q}_{\ell}}
\rightarrow \mathrm{GL}_2(E)$ is unramified. Then~$x$ factors through~$R^{\mathrm{mod},\psi}_{\ell}$.
\end{corr}
\begin{proof} Such a representation exists exactly when the eigenvalues of $\rho_x(\phi)$ are equal.
We see that~$x$ certainly factors through~$\widetilde{R}^{\mathrm{unr},\psi}$, which is a quotient
of~$R^{\mathrm{mod},\psi}_{\ell}$.
\end{proof}
If the determinant is explicit from the context, we write~$R^{\mathrm{mod}}_{\ell}$ rather than~$R^{\mathrm{mod},\psi}_{\ell}$.
We give a precise description of the special fibre of~$R^{\mathrm{mod}}_{\ell}$ when~$\ell \equiv 1 \mod p$,
~$\psi = 1$,
and~$\overline{\rho}$ is trivial (this result will only be used for the proof
of Theorem~\ref{theorem:blog} in section~\S\ref{section:blog}.) Note that,
since~$\overline{\rho}$ by assumption is tamely ramified at~$\ell$,
the image
of any deformation also factors through tame inertia, hence through the
group $\langle \tau,\phi \rangle$ with $\phi \tau \phi^{-1} = \tau^{\ell}$.
\begin{lemma} \label{lemma:special} Suppose that~$p > 2$, that~$\ell \equiv 1 \mod p$, that~$\psi | G_{\mathbf{Q}_{\ell}} \rightarrow
\mathrm{GL}_2(k)$ is trivial. Then~$R^{\mathrm{mod}}_{\ell}$ represents the functor of deformations of~$\overline{\rho}$
to~$A$ together with an~$\alpha_{\ell} \in A$ satisfying the following conditions:
\begin{itemize}
\item $\mathrm{Tr}(\rho(\tau)) = 2$,
\item $(\rho(\tau) - 1)^2 = 0$,
\item $(\rho(\tau) - 1)(\rho(\phi) - \alpha_{\ell}) = 0$,
\item $(\rho(\phi) -
\alpha^{-1}_{\ell})(\rho(\tau) - 1) = 0$,
\item $(\rho(\phi) - \alpha_{\ell})(\rho(\phi) - \alpha^{-1}_{\ell}
) = 0$.
\end{itemize}
\end{lemma}
The argument is similar (but easier) to the corresponding
arguments of Snowden (\cite{Snowden}~\S4.5). In fact, our argument amounts
to the case~$d = 0$ of a theorem proved by Snowden for all integers~$d > 0$.
The only reason that Snowden does not consider this case is that, in his context,
$d$ is the degree of a finite extension of~$\mathbf{Q}_p$.
\begin{proof}
The last equation says that~$\alpha_{\ell}$
satisfies the characteristic polynomial of Frobenius. Hence the functor is certainly represented
by a quotient~$R^{\dagger}_{\ell}$ of the universal such ring~$\widetilde{R}^{\univ}_{\ell}$. Let us show that~$\mathrm{MaxSpec}(R^{\dagger}_{\ell}) = \mathrm{MaxSpec}(R^{\mathrm{mod}}_{\ell})$
(inside~$\mathrm{MaxSpec}(\widetilde{R}^{\univ}_{\ell})$).
Let~$x: R^{\dagger}_{\ell} \rightarrow F$ be a point of~$\mathrm{MaxSpec}(R^{\dagger}_{\ell})$.
If~$\rho(\tau)$ is trivial, then the equations reduce to the statement that~$\alpha_{\ell}$
is an eigenvalue of Frobenius, and these correspond exactly to the unramified points of~$R^{\mathrm{mod}}_{\ell}$.
If~$\rho(\tau)$ is non-trivial, then, from the first equation, its minimal polynomial will be~$(X-1)^2$, and so, after conjugation,
has the shape
$$\rho(\tau) = \left( \begin{matrix} 1 & 1 \\ 0 & 1 \end{matrix} \right).$$
The other equations then imply that
$$\rho(\phi) = \left( \begin{matrix} \alpha^{-1}_{\ell} & * \\ 0 & \alpha_{\ell} \end{matrix} \right).$$
Finally, from the equation $\phi \tau \phi^{-1} = \tau^{\ell}$, we deduce
that~$\alpha^{-2}_{\ell} = \ell$. In particular, the representation is, up to twist,
an extension of~$F$ by~$F(1)$, which corresponds exactly to points on the
special component of~$R^{\mathrm{mod}}_{\ell}$. It follows from
Corollary~2.3 of~\cite{KhareW2} (see also
Lemma~\ref{lemma:kisin} below) that~$\mathcal{O}$-flat
reduced quotients of~$R^{\univ}_{\ell}$ are characterized by their~$F$ points
for finite extensions~$F/E$. Since we have shown that~$R^{\dagger}_{\ell}$ and~$R^{\mathrm{mod}}_{\ell}$ have
the same such quotients, and since~$R^{\mathrm{mod}}_{\ell}$
is~$\mathcal{O}$-flat and reduced, it suffices to show that~$R^{\dagger}_{\ell}$ is~$\mathcal{O}$-flat and reduced.
The special fibre~$R^{\dagger}_{\ell}/\varpi$
is exactly the completion
of $\mathcal{C}_0$ at $c = (1;1;0)$ in the notation of~\cite{Snowden} (\S3.5).
The proof of this is identical to the proof of Lemma~4.7.4 of~\cite{Snowden}.
On the special fibre, the equation~$(\rho(\tau) - 1)^2 = 0$ implies that~$(\rho(\tau) - 1)^p = 0$
and so~$\rho(\tau^p)$ is trivial, and~$\rho(\tau^{\ell}) = \rho(\tau)$. Hence
the action of conjugation by~$\rho(\phi)$ on~$\rho(\tau)$ is trivial.
In~\cite{Snowden}, the image of inertia factors through an exponent~$p$
commutative group which,
as a module for~$\mathbf{F}_p \llbracket T \rrbracket$ where~$1+T$ acts as conjugation by~$\sigma$,
is isomorphic to~$U = \mathbf{F}_p \oplus \mathbf{F}_p\llbracket T \rrbracket^{\oplus d}$. In our context, the action of inertia
commutes with~$\sigma$ and
factors through a group~$U = \mathbf{F}_p$.
In particular, letting~$m = \rho(\tau) - 1$ and~$\varphi$ be the image of~$\rho(\phi)$,
the tuple~$(\varphi,\alpha,m)$ is the corresponding point on~$\mathcal{C}_0$. The rest
of the argument follows the proof of Theorem~4.7.1 of~\cite{Snowden}.
The ring~$\mathcal{C}_{0}$ has two
minimal primes (corresponding to $(\alpha - 1)$ and~$m$, which
come from the components~$\mathcal{A}_2$ and~$\mathcal{B}_0$ respectively, in the
notation of~\cite{Snowden}). On the other hand, as we have shown,
$R^{\dagger}_{\ell}[1/\varpi]_{\mathrm{red}} = R^{\mathrm{mod}}[1/\varpi]$ has two minimal primes
corresponding to the unramified and ordinary locus, so~$R^{\dagger}_{\ell}[1/\varpi]$
has two minimal primes, and so, by Propositions~2.2.1 and~2.3.1 of~\cite{Snowden},
it follows that~$R^{\dagger}_{\ell}$ is~$\mathcal{O}$-flat and reduced, and we are done.
\end{proof}
\subsection{The functors~$D_Q$ and~$D^1_Q$}
\label{section:definitions}
In this section,~$\overline{\rho}$ will be a global Galois representation unramified at~$p$ with the primes~$N = S \cup P$
as in the introduction.
We now define modified deformation rings $R^1_{Q}$ and $R_{Q}$
for certain sets $Q$ of auxiliary primes distinct from $N$ and $p$.
Let~$D_{\ell}$ denote the decomposition group~$G_{\mathbf{Q}_{\ell}} \subset G_{\mathbf{Q}}$.
The superscript~$1$ refers to weight one,
and the lack of superscript will refer to weight~$p$.
Note that $R^1_{\emptyset} = R^1$.
Besides the representation~$\overline{\rho}$, part of the data required
to define~$D^1_Q$ and~$D_{Q}$ consists of a
fixed choice of elements~$a_{\ell} \in k$ for~$\ell$ dividing~$N$ and~$Q$.
Moreover,
for~$D_{Q}$, we also fix an~$a_p \in k$.
Let $D^1_Q(A)$ and $D_Q(A)$ consist of deformations $\rho$ to $A$ and a collection of elements $\alpha_{\ell} \in A$
for $\ell \in N$ (and $\alpha_p \in A$ for $D_Q(A)$) such that:
\begin{enumerate}
\item $\det(\rho) = \psi$, where $\psi$ is the Teichmuller
lift of~$\det(\overline{\rho})$ for $\rho \in D^1_{Q}(A)$,
and $\det(\rho) = \psi \epsilon^{p-1}$ for $\rho \in D_{Q}(A)$.
\item $\rho$ is unramified outside $N \cup Q = S \cup P \cup Q$ for $\rho \in D^1_{Q}(A)$, and unramified outside $N \cup S
\cup Q \cup \{p\}$ for $\rho \in D_{Q}(A)$.
\item \label{item:P} If $\ell \in P$, then
$\rho |D_{\ell} \simeq
\chi^{-1} \psi |_{D_{\ell}} \oplus \chi$ if $\rho \in D^1_{Q}(A)$ and $\rho |D_{\ell} \simeq
\chi^{-1} \psi \epsilon^{p-1} |_{D_{\ell}} \oplus \chi$ if $\rho \in D_Q(A)$, where
$\chi$ is an unramified character and
$\chi(\mathrm{Frob}_{\ell}) = \alpha_{\ell} \equiv a_{\ell} \mod \mathfrak{m}$.
\item \label{item:S}
If $\ell \in S$, then
$\rho | D_{\ell}$ corresponds to an~$A$-valued quotient of~$R^{\mathrm{mod}}_{\ell}$,
where we take the determinant to be~$\psi$ if~$\rho \in D^1_{Q}(A)$
and~$\psi \cdot \epsilon^{p-1}$ if~$\rho \in D_{Q}(A)$, and, in either case~$\alpha_{\ell}
\in R^{\mathrm{mod}}_{\ell}$ is $a_{\ell} \mod \mathfrak{m}$.
\item \label{item:Q} If $\ell \in Q$, then $\ell \equiv 1 \mod p$, and $\overline{\rho}(\mathrm{Frob}_{\ell})$ has distinct eigenvalues.
Then $\rho |D_{\ell} \simeq \phi^{-1} \psi |_{D_{\ell}} \oplus \phi$, where $\phi$ is a character of
$\mathbf{Q}^{\times}_{\ell} \subset G^{\mathrm{ab}}_{\mathbf{Q}_{\ell}}$ such that
$\phi(\ell) = \alpha_{\ell} \equiv a_{\ell} \mod \mathfrak{m}$.
\item If $\rho \in D_Q(A)$ and $\ell = p$, then $\rho|D_p$ is ordinary with eigenvalue $\alpha_p \equiv a_p \mod \mathfrak{m}$.
\end{enumerate}
In order for these functors to be
non-zero, the~$a_{\ell}$ for
$\ell \in N \cup Q$ must be chosen to be one of the eigenvalues of $\overline{\rho}(\mathrm{Frob}_{\ell})$, and $a_p$
must be one of the eigenvalues
of $\overline{\rho}(\mathrm{Frob}_p)$. As always, we may extend scalars from $k$ to a field which contains
all necessary eigenvalues.
For each $\ell \in N$, there exists a corresponding universal framed local deformation
ring associated to our deformation problem. There is no subtlety in defining
these rings outside the case of primes in~$S$, and at the prime~$p$.
The first case was addressed in the previous section.
For~$\ell = p$,
we use the modified deformation rings as constructed by Snowden (\cite{Snowden}, see
in particular \S4.6).
For each~$\ell$, we denote the corresponding modified local deformation ring (with
the appropriate determinant) by~$R^{\mathrm{mod}}_{\ell}$.
\begin{prop} \label{prop:CM} For all of the $\ell$ different from~$p$, the corresponding modified local deformation
ring $R^{\mathrm{mod}}_{\ell}$ is an $\mathcal{O}$-flat reduced equidimensional ring of relative dimension~$3$ over~$\mathcal{O}$. If $\ell = p$ and $D = D_{Q}$, then $R^{\mathrm{mod}}_{\ell}$ is an $\mathcal{O}$-flat reduced
equidimensional ring of dimension~$4$ over~$\mathcal{O}$.
\end{prop}
\begin{proof} We consider each deformation ring in turn.
\begin{enumerate}
\item Suppose that $\ell \in P$. By assumption, $\psi$ is ramified at~$\ell$ and hence $\alpha_{\ell}$
is uniquely determined by $\rho |D_{\ell}$. Hence we recover the framed local deformation ring, and
the result follows from Lemma~4.11 of~\cite{CG}.
\item Suppose that $\ell \in S$. Then the result follows from Lemma~\ref{lemma:gee}.
\item Suppose that~$\ell \in Q$. The assumption that~$\ell \equiv 1 \mod p$ and
that~$\overline{\rho}(\mathrm{Frob}_{\ell})$ has no distinct eigenvalues implies that there is no
distinction between~$R^{\univ}_{\ell}$ and~$\widetilde{R}^{\univ}_{\ell}$. Moreover,
all deformations of~$\overline{\rho}$ will be tamely ramified and split as a direct sum
of two characters, and so~$R^{\mathrm{mod}}_{\ell} = R^{\mathrm{univ},\psi}_{\ell}$ in this case.
The ring~$R^{\mathrm{univ},\psi}_{\ell}$ has the desired properties by a
direct computation, see for example Proposition~7 of~\cite{Shotton}: it
may be identified with $\mathcal{O} \llbracket X,Y,Z,P \rrbracket/((1+P)^{m} - 1)$, where~$m$ is
the largest power of~$\ell$ dividing~$p-1$.
\item If $\ell = p$, and $\overline{\rho}(\mathrm{Frob}_p)$ has distinct eigenvalues, then the usual definition of
an ordinary deformation ring $R_{p}$
requires a choice of eigenvalue of the unramified quotient, and hence $R^{\mathrm{mod}}_{p}$ is just the usual
Kisin ring $R_p$ in this case.
If $\overline{\rho}(\mathrm{Frob}_p)$ has the same
eigenvalues, then the local modified deformation ring is exactly the completion
of $\mathcal{B}_1$ at $b = (1;1;0)$ considered in~\cite{Snowden}~\S3.4 and denoted by~$\widetilde{R}^{\dagger}$
in~\cite{CG}~\S3.7. The case when $\overline{\rho}(\mathrm{Frob}_p)$
has the same eigenvalues but is non scalar corresponds to the localization of $\mathcal{B}_1$
at $\left(\left( \begin{matrix}1 & 1 \\ 0 & 1 \end{matrix} \right);1;0\right)$.
In either case, $R^{\mathrm{mod}}_{p}$ is~$\mathcal{O}$-flat, reduced, equidimensional of relative
dimension~$4$ (over~$\mathcal{O}$), and Cohen--Macaulay.
\end{enumerate}
\end{proof}
We also present here the following proposition which will be useful later.
(cf. Lemma~3.4.12 of~\cite{kisin-moduli}.)
\begin{lemma} \label{lemma:kisin} Let~$A$ and~$B$ be complete local Noetherian reduced~$\mathcal{O}$-flat algebras with residue field~$k$.
Then~$A \widehat{\otimes}_{\mathcal{O}} B$ is reduced and~$\mathcal{O}$-flat.
\end{lemma}
\begin{proof} The~$\mathcal{O}$-flatness follows from Lemma~19.7.1.2
of section~$0$ of~\cite{EGAfourone}.
Because~$B$ is reduced, it follows from Corollary~2.3 of~\cite{KhareW2} that
the intersection of the kernels of all
morphisms~$B \rightarrow \mathcal{O}'$ for the ring
of integers finite extensions~$E'/E$ is trivial.
Using this, we may write $B$ as an inverse limit
$B = \varprojlim B_i$,
where each $B_i$ is reduced and finite flat over~$\mathcal{O}$. Then~$C = \varprojlim A \otimes B_i$ (now we
can replace~$\widehat{\otimes}$ by~$\otimes$) and it suffices to
prove the claim for the usual tensor product when~$B$ is finite flat over~$\mathcal{O}$, which we now assume.
Since~$C$ is~$\mathcal{O}$-flat,
it suffices to show that~$C[1/\varpi] = A[1/\varpi] \otimes_E B[1/\varpi]$ is reduced. However,
this follows
from~\cite{bourbaki}, Chap.~V~\S15.5; Theorem~$3$(d)).
\end{proof}
\subsection{Modular Curves} \label{section:stacks}
Let $N \ge 5$, and
let $X_H(N) = X(\Gamma_H(N))$ denote the quotient of $X_1(N)$ by the Sylow $p$-subgroup of $(\mathbf{Z}/N\mathbf{Z})^{\times}$
considered as a smooth
proper scheme over $\mathrm{Spec}(\mathcal{O})$~\cite{deligne-rapoport}. To be precise, the curve~$X_H(N)$
is a fine moduli space providing that either~$p \ge 5$ or~$p = 3$ and~$N$ divisible by
a prime~$q \ge 5$ such that $q \equiv -1 \mod 3$.
This follows either from the computation of stabilizers at the CM points
(as in~\cite{MazurEisenstein},~\S2, p.64), or, in the second case, because~$X_H(N)$ is a cover of~$X_1(q)$.
If~$p = 3$ and~$X_H(N)$ is not a fine moduli space,
we simply add a prime~$q \ge 5$ and~$q \equiv -1 \mod 3$ to~$S$
such that~$\overline{\rho}$ is unramified at~$q$.
If~$Q$ is a collection of auxiliary primes disjoint from~$N$, let~$X_H(N\kern-0.05em{Q})$ denote a quotient
of~$X_1(N)$ by the~$p$-Sylow subgroup of~$(\mathbf{Z}/N\mathbf{Z})^{\times}$
and some subgroup of~$(\mathbf{Z}/Q \mathbf{Z})^{\times}$. (In practice, the cokernel of the corresponding
subgroup of~$(\mathbf{Z}/Q\mathbf{Z})^{\times}$ will be a~$p$-group.)
\subsection{Hecke algebras}
Let $\omega$ be
the usual pushforward $\pi_{*} \omega_{\mathcal{E}/X_H(N)}$ of the relative dualizing sheaf
along the universal generalized elliptic curve. If $A$ is an $\mathcal{O}$-module, then
let $\omega^n_A = \omega^n \otimes_{\mathcal{O}} A$.
The (Katz) space of modular
forms of weight $k$ and level $N$ is defined to be
$H^0(X_H(N),\omega^k_{A})$.
\medskip
We shall now consider a number of Hecke algebras, and discuss the relationship between them.
Our coefficient ring or module will either be $A = \mathcal{O}$, $A = E = \mathcal{O} \otimes \mathbf{Q}$, $A = \mathcal{O}/\varpi = k$, $A = \mathcal{O}/\varpi^n$,
or $A = E/\mathcal{O}$ unless otherwise specified.
\begin{df} The Hecke algebra $\mathbf{T}_A$ in weight~$k$ is the $A$ sub-algebra of
$$\mathrm{End}_A(H^0(X_H(N),\omega^k_{A}))$$
generated by the operators $T_n$ for $n$ prime to $p$ and diamond operators $\langle d \rangle$
for $d$ prime to~$N$.
\end{df}
Note that this definition includes the operators~$T_{\ell}$ for ${\ell}|N$.
These operators can also be denoted by~$U_{\ell}$ (which is what we shall do below).
We now define a variant of these Hecke algebras where we include the Hecke operator at~$p$.
\begin{df} Let $\widetilde{\T}_A$ denote the ring $\mathbf{T}_A$ together with the operator $T_p$.
\end{df}
Note that a maximal ideal $\mathfrak{m}$ of $\mathbf{T}$ need no longer a priori be maximal in $\widetilde{\T}$.
That is,~$\widetilde{\T}_{\mathfrak{m}}$ will not always be a local ring. However, it will always be a semi-local ring,
that is, a direct sum of finitely many local rings.
\subsection{Hecke algebras at auxiliary level~$Q$}
Let $Q$ be a finite collection of primes congruent to $1 \mod p$ and distinct from $N$.
Let $\mathbf{T}_{Q,A}$ denote the~$\mathcal{O}$-algebra generated by Hecke operators away
from~$p$
acting
at level $X_H(N\kern-0.05em{Q})$ with coefficients in~$A$ together with diamond operators~$\langle d \rangle$
for $d$ prime to~$N$,
and let~$\widetilde{\T}_{Q,A}$ denote~$\mathbf{T}_{Q,A}$ together with the operator~$T_p$.
\medskip
Suppose that $\overline{\rho}$ is a representation such that the modified deformation ring $R$ is non-zero -- equivalently, that $D(k) \ne 0$.
Recall that~$N$ is equal to the conductor of~$\psi$ times the primes in some auxiliary set~$S$ which includes (but may be
larger than)
the set of primes~$\ell \nmid p$ where~$\overline{\rho} | I_{\ell}$ is non-trivial and unipotent. In particular, if~$\ell$ divides~$S$,
then~$\ell$ divides~$N$ exactly once.
By Serre's Conjecture~\cite{KhareW}, any such $\overline{\rho}$ is modular of level $N(\overline{\rho}) | N$ and
weight~$p$, so we now specialize to the case of weight~$p$, and let~$\mathbf{T}_{Q} = \mathbf{T}_{Q,\mathcal{O}}$.
Let $\widetilde{\mathfrak{m}}$ be a maximal ideal of $\widetilde{\T}_{Q}$ corresponding to~$\overline{\rho}$ (and to a choice of~$a_{\ell}$
for all~$\ell$ dividing~$N$,~$Q$, and~$p$).
Let us also suppose that for every prime~$\ell$ dividing~$Q$, the matrix~$\overline{\rho}(\mathrm{Frob}_{\ell})$
has distinct eigenvalues (since this is an assumption in part~\ref{item:Q} of the definition of~$D_Q$).
\begin{prop} \label{prop:action}
There exists a deformation
$$\rho_Q: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}})$$
of $\overline{\rho}$ unramified outside $pN\kern-0.05em{Q}$ such that~$\rho(\mathrm{Frob}_{\ell}) = T_{\ell}$ for~$\ell$ prime
to~$p N\kern-0.05em{Q}$. Let~$\rho'_Q = \rho_{Q} \otimes \eta$,
where~$\eta^2 = \psi \cdot \epsilon^{p-1} \cdot \det(\rho_Q)^{-1}$. Then~$\rho'_{Q}$ is a deformation
of~$\overline{\rho}$ in~$D_Q(\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}})$. In particular, there is a corresponding map
$$R_Q \rightarrow \widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$$
sending $\Tr(\rho^{\mathrm{univ}}(\mathrm{Frob}_{\ell})) \in R_Q$ to $\eta(\ell) \cdot T_{\ell}$
for $\ell$ not dividing $pN\kern-0.05em{Q}$, sending $\alpha_{\ell}$ to $\eta(\ell) \cdot U_{\ell}$ for $\ell$ dividing $NQ$, and sending
$\alpha_p + p^{\kern+0.1em{p}-1} \psi(p) \alpha^{-1}_p$ to $\eta(p) \cdot T_p$, or equivalently, $\alpha_p$ to the unit root of
$$X^2 - \eta(p) \cdot T_p X + \psi(p) p^{\kern+0.1em{p}-1} = 0,$$
which lies in $\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$ by Hensel's Lemma.
\end{prop}
This proposition is (mostly) an exercise in Atkin--Lehner--Li theory.
Indeed, if one assumes that the action of~$U_{\ell}$ on forms of level~$\ell \| N$ is semi-simple
(which conjecturally is always the case), then the space of modular forms under consideration
will decompose into a direct sum of eigenforms for all the Hecke operators in~$\widetilde{\T}_{Q,\mathfrak{m}}$,
and then the claim follows immediately from known local-global compatibility
for classical modular forms. (The only local--global compatibility we require is given
by Theorem~3.1 of~\cite{Fermat}.)
In practice, we have to allow for the possibility that~$U_{\ell}$
may not act semi-simply, although this is not difficult.
\begin{proof}
The space of modular forms of weight~$p$ is torsion free, so the Hecke algebra
is determined by its action on
$$H^0(X_H(N\kern-0.05em{Q}),\omega^p)_{\widetilde{\mathfrak{m}}} \otimes E.$$
It suffices to prove the proposition after further decomposing this space into a direct sum of~$\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$-modules.
Enlarging~$E$ if necessary, we may assume that all the eigenvalues of
all Hecke operators at level dividing~$N\kern-0.05em{Q}$ are defined over~$E$.
Let~$\T^{\mathrm{an}}_Q$ denote the anaemic Hecke algebra consisting of
endomorphisms of~$H^0(X_H(N\kern-0.05em{Q}),\omega^p)$ generated by Hecke operators~$T_n$ for~$n$ prime
to~$pN\kern-0.05em{Q}$
and diamond operators~$\langle d \rangle$ for~$d$ prime to~$N\kern-0.05em{Q}$. There is a map~$\T^{\mathrm{an}}_Q \rightarrow \widetilde{\T}_Q$;
let~$\mathfrak{m} \subset \T^{\mathrm{an}}_Q$ denote the inverse of the maximal ideal~$\widetilde{\mathfrak{m}}$
(which is determined by~$\overline{\rho}$).
Note that~$\mathfrak{m}$ may correspond to several~$\widetilde{\mathfrak{m}}$ in~$\widetilde{\T}_{Q}$; the possible~$\widetilde{\mathfrak{m}}$ are indexed by the possible
choices of~$a_{\ell}$ for~$\ell$ dividing~$pN\kern-0.05em{Q}$. In any event, there will always be an inclusion:
$$H^0(X_H(N\kern-0.05em{Q}),\omega^p)_{\widetilde{\mathfrak{m}}} \otimes E \subset
H^0(X_H(N\kern-0.05em{Q}),\omega^p)_{\mathfrak{m}} \otimes E.$$
(This would be an equality if we replaced the left hand side with a direct sum over all~$\widetilde{\mathfrak{m}}$
which pull back to~$\mathfrak{m}$.)
The space~$H^0(X_H(N\kern-0.05em{Q}),\omega^p) \otimes E$ decomposes under~$\T^{\mathrm{an}}_{Q}$ into eigenspaces indexed by
newforms~$f$ of level dividing~$N\kern-0.05em{Q}$.
Associated to a cuspidal newform~$f$ is a Galois representation~$\rho_f$.
In particular, combining all these Galois representations over~$f$ with~$\overline{\rho}_f = \overline{\rho}$,
we obtain a Galois
representation
$$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\T^{\mathrm{an}}_{Q,\mathfrak{m}} \otimes E).$$
Because the traces of Frobenius elements lie in~$\T^{\mathrm{an}}_{Q,\mathfrak{m}}$, and
because~$\overline{\rho}$ is absolutely irreducible, we may take the image of this
Galois representation to land in~$\mathrm{GL}_2(\T^{\mathrm{an}}_{Q,\mathfrak{m}})$ by~$(2.6)$ of~\cite{Lenstra}.
To this point, we have simply reconstructed the usual construction of the Galois
representation into the (anaemic) Hecke algebra.
\medskip
Let~$\rho_Q$ denote the Galois representation induced by composing this
with the image of~$\T^{\mathrm{an}}_{Q,\mathfrak{m}}$ in~$\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$. This will be the~$\rho_{Q}$
of the proposition. The reason for the twist by~$\eta$
is to match the determinant with the required determinant for the functor~$D_Q$.
The main point of this proposition is to show that
that the extra old forms associated to~$f$ (with their concomitant actions of~$U_{\ell}$)
contain exactly the extra information needed to obtain a modified deformation of~$\overline{\rho}$
of type~$D_Q$.
The eigenspaces corresponding to~$f$ will contribute to the localization at~$\mathfrak{m}$ if
and only if~$\overline{\rho}_{f} = \overline{\rho}$.
In particular, the level of~$f$ must be divisible by the Serre conductor of~$\overline{\rho}$, and hence the level of~$f$ is
of the form~$N\kern-0.05em{Q}/D$, where~$D$ is only divisible by primes dividing either~$S$ or~$Q$.
Hence the integer~$D$ is square-free and prime to~$N\kern-0.05em{Q}/D$.
Suppose that~$D$ has~$d$ prime divisors.
The form~$f$ generates a space of~$2^d$ oldforms of level~$N\kern-0.05em{Q}$
consisting of~$f = f(q)$ together with the forms~$f(q^m)$ for~$m|D$.
By Atkin--Lehner-Li theory (Theorem~9.4 of~\cite{Stein}) this exhausts the entire space of
oldforms associated to~$f$
which appear in~$H^0(X_H(N\kern-0.05em{Q}),\omega^p)_{\mathfrak{m}} \otimes \mathbf{Q}$.
Let us now describe the action of~$U_{\ell}$ on these spaces for~$\ell$ dividing~$D$.
Again by Atkin--Lehner, this is given as the
tensor product over~$\ell | D$ of a two-dimensional space on which~$U_{\ell}$ acts by the matrix
$$\left( \begin{matrix} \Tr(\rho_f(\mathrm{Frob}_{\ell})) & \langle \ell \rangle \ell^{p-1} \\
-1 & 0 \end{matrix} \right).$$
Here~$\Tr(\rho_f(\mathrm{Frob}_{\ell}))$ may also be identified with the eigenvalue of~$f$ under the Hecke operator~$T_{\ell}$ acting at level~$N\kern-0.05em{Q}/D$.
The element~$ \Tr(\rho_f(\mathrm{Frob}_{\ell}))$ will lie in the image of~$\T^{\mathrm{an}}_{Q}$ by the Cebotarev density theorem.
Note that the eigenvalues of this matrix are precisely
the eigenvalues of~$\rho_f(\mathrm{Frob}_{\ell})$.
There are now two possibilities:
\begin{enumerate}
\item The eigenvalues of~$\rho_f(\mathrm{Frob}_{\ell})$ are distinct.
In this case, the space of oldforms over~$E$ decomposes further into eigenspaces under~$U_{\ell}$. The eigenvalues of~$U_{\ell}$ will correspond
precisely to Galois representations together with a choice of eigenvalue of~$\rho_f(\mathrm{Frob}_{\ell})$.
Each choice of eigenvalue will contribute to the localization at~$\widetilde{\mathfrak{m}}$ if and only if the corresponding eigenvalue
is~$a_{\ell} \mod \varpi$.
After the global twist to match determinants,
such representations will naturally be algebras over~$R^{\mathrm{mod}}_{\ell}$, where~$\alpha_{\ell}$ is sent
to~$\eta(\ell) \cdot U_{\ell}$ (this follows by the construction of the rings~$R^{\mathrm{mod}}_{\ell}$,
in particular Lemma~\ref{lemma:gee} for unramified primes of type~$S$.
\item The eigenvalues of~$\rho_f(\mathrm{Frob}_{\ell})$ are equal. Call the unique eigenvalue~$b_{\ell}$.
From the explicit matrix description of the action of~$U_{\ell}$
above, we see that~$U_{\ell}$ is not a multiple of the scalar matrix, and so
it is not diagonalizable. In particular, in the Hecke algebra, the operator~$U_{\ell}$ satisfies the relation~$(U_{\ell} - b_{\ell})^2 = 0$.
However, once again (after twisting), there will be
a map from~$R^{\mathrm{mod}}_{\ell}$ sending~$\alpha_{\ell}$ to~$\eta(\ell)
\cdot U_{\ell}$, by Corollary~\ref{corr:non} (the eigenvalues can
only be the same for primes of type~$S$).
\end{enumerate}
We remark that the second case above conjecturally never occurs in weight~$\ge 2$
(see~\cite{Coleman}).
Let us now consider the operators~$U_{\ell}$ for~$\ell$ not dividing~$D$. In this case,
the Galois representation~$\rho_f$ is ramified at~$\ell$, and local--global compatibility of Galois representations
implies that, after twisting by~$\eta$, the Galois representation has an unramified quotient on which~$\mathrm{Frob}_{\ell}$ acts via~$U_{\ell}$,
and hence we have a natural map from~$R^{\mathrm{mod}}_{\ell}$ to~$\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$ sending
(after twisting)~$\alpha_{\ell}$ to~$\eta(\ell) \cdot U_{\ell}$.
Finally, since (by definition)~$a_p \in k$ is a unit (it is an eigenvalue of an invertible matrix), the representation~$\rho_f$
is ordinary at~$p$, and action of Frobenius on the Galois
representations associated to any form~$f$ will admit an unramified at~$p$ quotient on which~$\mathrm{Frob}_p$
acts as~$U_p$. Hence there will be a natural map from~$R^{\mathrm{mod}}_p$ to~$\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$ sending~$\alpha_p$ to~$\eta(p) \cdot U_p$,
which is related to $T_p$ via the equation $U^2_p - T_p U_p +
\langle p \rangle p^{\kern+0.1em{p}-1} = 0$.
\end{proof}
\begin{remark} \emph{The theorem above is true in any weight~$k \ge 2$, providing that one modifies the
definition of~$D_Q$ to take into account the weight, and one still works in the ordinary context (so~$a_p \in k^{\times}$).}
\end{remark}
\subsection{Modularity lifting theorems in weight~\texorpdfstring{$p$}{p}}
\label{section:higher}
The main goal of this section is to prove the following:
\begin{theorem} \label{theorem:minimal} There is an isomorphism
$R_Q \rightarrow \widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$.
\end{theorem}
Before proving this theorem, we remark that the ``modularity'' theorem
one can deduce from this~$R = \mathbf{T}$ theorem is already well known. In particular,
one knows that~$(R_Q[1/\varpi])^{\mathrm{red}} = (\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}[1/\varpi])^{\mathrm{red}}$
(our Hecke algebras will not be reduced if the action
of~$U_{\ell}$ is not semi-simple). Hence the content of this theorem
is to upgrade this known result to an integral statement. In order to see how one might do this,
note that the modifications of Taylor--Wiles due to Diamond, Kisin, and others
(\cite{DiamondMult},\cite{kisin})
proceed
by constructing a patched module~$M_{\infty}$ over a patched deformation ring~$R_{\infty}$
and a ring of auxiliary diamond operators~$S_{\infty}$.
By hook or by crook, one tries to prove that~$M_{\infty}$ is faithful (or nearly faithful) as
an~$R_{\infty}$-module. To recover a classical statement, one takes the quotient
of~$R_{\infty}$ and~$M_{\infty}$ by
the augmentation ideal~$\mathfrak{a}$ of~$S_{\infty}$, and recovers the classical ring~$R$ and a module~$M$
of classical modular forms
on which~$R$ acts via the quotient~$\mathbf{T}$.
An essential difficulty, however, is that even if one knows that~$M_{\infty}$ is faithful as an~$R_{\infty}$-module, this does not imply that~$M_{\infty}/\mathfrak{a}$ is faithful as an~$R_{\infty}/\mathfrak{a}$-module; that is, faithfulness is
not preserved under quotients. Hence these methods often only allow one to deduce
weaker statements concerning reduced quotients.
In Wiles' original arguments, however, the auxiliary modules~$M_D$ are free over the corresponding
Hecke algebras, and one ultimately deduces that the patched module~$M_{\infty}$ is also
free over~$R_{\infty}$,
from which one can certainly conclude that~$M_{\infty}/\mathfrak{a}$ is free over~$R_{\infty}/\mathfrak{a}$,
and hence that~$R = \mathbf{T}$. In our argument, we exploit the fact that,
by using all the Hecke operators,
the multiplicity one theorem for~$q$-expansions allows us to also
show that the auxiliary modules~$M_D$ are free
over certain Hecke algebras, and hence we are able to deduce (as in Wiles)
an integral~$R = \mathbf{T}$ theorem.
\begin{remark} \emph{{\bf An apology concerning notation:\rm} The notation~$Q$ that we have
used is meant to suggest a collection of Taylor--Wiles primes. Indeed,
the primes denoted by~$Q$ will play the role of Taylor--Wiles primes in the modularity
proof of~\S\ref{section:one}. However, in the proof
of Theorem~\ref{theorem:minimal} below, the set of primes~$Q$ will be fixed,
and there will be an auxiliary choice~$T_D$ of Taylor--Wiles primes~$x \equiv 1 \mod p^D$. Explicitly,
we are proving an~$R = \mathbf{T}$ theorem at level~$p$ and a level which already
includes a fixed collection of Taylor--Wiles primes~$Q$. Hence we require a second auxiliary choice
of Taylor--Wiles primes for which we use the letter~$T$ rather than~$Q$.}
\end{remark}
\begin{proof}
We first define a classical unmodified (``natural'') deformation ring $R^{\natural}_{Q}$ which records deformations
which are of the same type as considered in $R_{Q}$, except now the extra choice of
eigenvalues is omitted,
as is the choice of eigenvalue at $\ell = p$.
There is a natural isomorphism
$$R_{Q} \simeq R^{\natural}_{Q} \otimes_{R^{\mathrm{loc}}} R^{\mathrm{mod}},$$
where
$\displaystyle{R^{\mathrm{loc}}:=\widehat{\bigotimes}_{\ell | p N\kern-0.05em{Q}} R_{\ell}}$
denotes the corresponding local deformation rings for $R^{\mathrm{loc}}$, and
$$R^{\mathrm{mod}}:=\widehat{\bigotimes}_{\ell | p N\kern-0.05em{Q}} R^{\mathrm{mod}}_{\ell}.$$
We remind the reader that one should think about the $R^{\mathrm{loc}}$-algebra~$R^{\mathrm{mod}}$ as follows: it
is the algebra obtained by including the extra information over~$R^{\mathrm{loc}}$ coming from
a choice of Frobenius eigenvalue, and taking a localization of this ring corresponding
to fixing (residually) a choice of such an eigenvalue. In particular, the set of
components of the generic fibre
of~$R^{\mathrm{mod}}$ is a subset of the components of the generic fibre of~$R^{\mathrm{loc}}$.
The ring~$R^{\mathrm{mod}}$ is also reduced by Lemmas~\ref{prop:CM} and~\ref{lemma:kisin}.
We now patch together
coherent cohomology modules and we also simultaneously
patch Betti cohomology.
Namely, we patch the pairs of modules
$$M^{C}_{D} =
H^0(X_{H_D}(N\kern-0.05em{Q} \cdot T_D),\omega^{\otimes p}_{E/\mathcal{O}})^{\vee}_{\widetilde{\mathfrak{m}}},$$
$$M^{B}_{D} = H^1(X_{H_D}(N\kern-0.05em{Q} \cdot T_D),\mathrm{Sym}^{p-2}((E/\mathcal{O})^2))^{\vee}_{\mathfrak{m}}.$$
The notation $C$ and $B$ refers
to coherent and Betti cohomology, respectively.
Here~$T_D$ is a collection of Taylor--Wiles primes~$x$ (distinct from primes
dividing~$N\kern-0.05em{Q}$) such that~$x \equiv 1 \mod p^D$,
and~$H_D$ is the subgroup of~$(\mathbf{Z}/N\kern-0.05em{Q} \cdot T_D \mathbf{Z})^{\times}$ generated
by the kernel of the
map~$(\mathbf{Z}/T_D\mathbf{Z})^{\times} \rightarrow (\mathbf{Z}/p^D \mathbf{Z})^{\# T_D }$ together with a fixed
subgroup of~$(\mathbf{Z}/N\kern-0.05em{Q} \mathbf{Z})^{\times}$.
The first module has a faithful action of~$\widetilde{\T}_{Q \cdot T_D,\widetilde{\mathfrak{m}}}$, and the second
has a faithful action of~$\mathbf{T}_{Q \cdot T_D,\mathfrak{m}}$. Moreover, the tensor product
$M^{B}_{D} \otimes_{R^{\mathrm{loc}}} R^{\mathrm{mod}}$ has a natural action of~$\widetilde{\T}_{Q \cdot T_D,\widetilde{\mathfrak{m}}}$.
We patch together both of these modules for the following reason. The patched Betti cohomology
module is known to be nearly faithful over the patched framed natural Galois deformation
rings $R^{\mathrm{loc}}\llbracket x_1,\ldots,x_{r+d-1}\rrbracket$ by a theorem of Kisin~\cite{kisin} --- this essentially
amounts to the fact that we already \emph{have} modularity lifting theorems in this context; the goal
is to upgrade these theorems to integral statements.
On the other hand, the coherent cohomology will be free over the corresponding modified
Hecke rings, which allows for an easier passage from patched objects back to finite level.
By Lemma~\ref{lemma:count}, the modules $M^{C}_{D}$ are free of rank one over~$\widetilde{\T}_{Q \cdot T_D,\widetilde{\mathfrak{m}}}$,
and so $M^{C}_{D}/\varpi^D$ is free of rank one over~$\widetilde{\T}_{Q \cdot T_D,\widetilde{\mathfrak{m}}}/\varpi^D$. On the other hand,
$M^{B}_{D}$ need not be free. However, the action of the Hecke on
$M^{B}_{D}/\varpi^D$ certainly factors through~$\mathbf{T}_{Q \cdot T_D,\mathfrak{m}}/\varpi^D$, and the action of
the full Hecke algebra (with Hecke operators for primes dividing~$N$) on $M^{B}_{D}/\varpi^D \otimes_{R^{\mathrm{loc}}} R^{\mathrm{mod}}$ factors
through~$\widetilde{\T}_{Q \cdot T_D,\widetilde{\mathfrak{m}}}/\varpi^D$.
If we patch together all this data simultaneously (together with framings), we deduce (as in the
pre-Diamond argument of Wiles) that the patched framed module $M^{\square,C}_{\infty}$
is free of rank one over~$\widetilde{\T}^{\square}_{\infty}$
(this is tautological, because each of the modules that is being patched will be free),
and that the action
of the patched modified Galois deformation ring on $M^{\square,B}_{\infty} \otimes_{R^{\mathrm{loc}}} R^{\mathrm{mod}}$ acts
through~$\widetilde{\T}^{\square}_{\infty}$.
If $R^{\square}_{Q}$ denotes the framed version of $R_{Q}$ over $d$ primes dividing $p N\kern-0.05em{Q}$, then the Taylor--Wiles method as modified by Kisin
gives presentations:
$$R_{Q}\llbracket T_1,\ldots,T_{4d-1}\rrbracket \simeq
R^{\square}_{Q} \simeq R^{\mathrm{mod}}\llbracket x_1,\ldots,x_{r+d-1}\rrbracket/(f_1,\ldots,f_{r+1}),$$
$$R^{\natural}_{Q}\llbracket T_1,\ldots,T_{4d-1}\rrbracket \simeq
R^{\square,\natural}_{Q} \simeq R^{\mathrm{loc}}\llbracket x_1,\ldots,x_{r+d-1}\rrbracket/(f_1,\ldots,f_{r+1}).$$
From~\cite{kisin} we know that~$M^{\square,B}_{\infty}$ is a nearly faithful
$R^{\mathrm{loc}}\llbracket x_1,\ldots,x_{r+d-1}\rrbracket$-module. From the freeness of~$M^{C}$
over the Hecke algebra, it follows that~$\widetilde{\T}^{\square}_{\infty}$ is a quotient
of the power series ring~$R^{\mathrm{mod}} \llbracket x_1,\ldots, x_{r+d-1} \rrbracket$.
If~$\widetilde{\T}^{\square}_{\infty}$
is actually isomorphic to this ring,
then by taking the quotient by diamond operators, we arrive at the required
isomorphism~$R_{Q} \simeq \widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}$.
On the other hand, the components of the generic fibre of $R^{\mathrm{mod}} \llbracket x_1,\ldots,
x_{r+d-1}\rrbracket$ are a subset of those for~$R^{\mathrm{loc}}$, so we deduce (cf. Lemma~2.3 of~\cite{Taylor}) that~$\widetilde{\T}^{\square}_{\infty}$
is a nearly faithful~$R^{\mathrm{mod}} \llbracket x_1,\ldots,
x_{r+d-1}\rrbracket$-module. Since~$R^{\mathrm{mod}}$ is reduced and Noetherian, the power series
ring has no nilpotent elements, and hence being nearly faithful over this ring is equivalent
to being faithful. Thus~$\widetilde{\T}^{\square}_{\infty}$ is a faithful module, and
hence isomorphic to~$R^{\mathrm{mod}} \llbracket x_1,\ldots, x_{r+d-1} \rrbracket$, and the
proof follows.
\end{proof}
\begin{corr} \label{corr:metro} There is an isomorphism $R_Q/\varpi \simeq \widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}/\varpi \simeq \widetilde{\T}_{Q,k,\widetilde{\mathfrak{m}}}$.
\end{corr}
\begin{proof} The first isomorphism is an immediate consequence
of the previous theorem, so it suffices to show that $\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}/\varpi \simeq \widetilde{\T}_{Q,k,\widetilde{\mathfrak{m}}}$, or equivalently,
that $\widetilde{\T}_{Q,\widetilde{\mathfrak{m}}}/\varpi$ acts faithfully on $H^0(X,\omega^{\kern+0.1em{p}}_k)$. For this it suffices to
note that
$$H^0(X,\omega^{\kern+0.1em{p}}_k)^{\vee} = (H^0(X,\omega^{\kern+0.1em{p}}_{E/\mathcal{O}})[\varpi])^{\vee}
= H^0(X,\omega^{\kern+0.1em{p}}_{E/\mathcal{O}})^{\vee}/\varpi$$
and that $H^0(X,\omega^{\kern+0.1em{p}}_{E/\mathcal{O}})^{\vee}$ is free of rank
one over~$\widetilde{\T}_{\mathfrak{m}}$ by Lemma~\ref{lemma:count} below.
\end{proof}
\section{Katz modular forms}
We now study more closely the action of Hecke operators in characteristic~$p$, especially in weights~$p$
and~$1$. In this section, we denote $\mathbf{T}_{Q,k}$ and $\widetilde{\T}_{Q,k}$ by $\mathbf{T}$ and $\widetilde{\T}$.
We use freely the $q$-expansion principle, namely, that a form
in $H^0(X,\omega^n_{k})$ is determined by its image in $k\llbracket q\rrbracket$.
Multiplication by the Hasse invariant induces a map:
$$A: H^0(X,\omega_k) \rightarrow H^0(X,\omega^{\kern+0.1em{p}}_k)$$
which is an injection and is the identity map on $q$-expansions. It follows that this map is $\mathbf{T}$-equivariant,
but it is not in general $\widetilde{\T}$-equivariant. There is another map between these spaces induced by the
map $q \mapsto q^{\kern+0.1em{p}}$:
$$V: H^0(X,\omega_k) \rightarrow H^0(X,\omega^{\kern+0.1em{p}}_k),$$
this map $V$ is also $\mathbf{T}$-equivariant. (Although the corresponding Hecke algebras $\mathbf{T}$ in weights one and $p$ are
not the same rings, the meaning of $\mathbf{T}$-equivariance should be clear.)
\begin{lemma} \label{lemma:count} Let $\widetilde{\mathfrak{m}}$ be a maximal ideal in $\widetilde{\T}$ in weight~$p$, and let $\mathfrak{m}$ be the corresponding
maximal ideal in $\mathbf{T}$. Then
$\dim H^0(X,\omega^{\kern+0.1em{p}}_k)[\widetilde{\mathfrak{m}}] = 1$ and
$$\dim H^0(X,\omega^{\kern+0.1em{p}}_k)[\mathfrak{m}] = 1 + \dim H^0(X,\omega_k)[\mathfrak{m}],$$
where
$\dim H^0(X,\omega_k)[\mathfrak{m}] \le 1$.
\end{lemma}
\begin{proof} The Hecke operators at all primes determine the $q$-expansion, which proves the first equality.
For the second, let $f \in H^0(X,\omega^p_k)[\widetilde{\mathfrak{m}}]$ denote the unique eigenform with leading coefficient~$1$.
There is a homomorphism
$$H^0(X,\omega^p_k)[\mathfrak{m}] \rightarrow k\llbracket q^{\kern+0.1em{p}}\rrbracket \cap H^0(X,\omega^{\kern+0.1em{p}}_k)[\mathfrak{m}]$$
given by $g \mapsto g - a_1(g) f$. By the $q$-expansion principle, the kernel of
this map is one-dimensional.
By the main theorem of Katz~\cite{Katz}, the
space $k\llbracket q^{\kern+0.1em{p}}\rrbracket \cap H^0(X,\omega^{\kern+0.1em{p}}_k)[\mathfrak{m}]$, which lies in the kernel $\ker(\theta)$ of the theta operator,
may be identified with the image of $H^0(X,\omega_k)[\mathfrak{m}]$ under $V$.
This gives the first
equality. To prove the inequality, we repeat
the same argument in weight one, except now~\cite{Katz} implies that $ k\llbracket q^{\kern+0.1em{p}}\rrbracket \cap H^0(X,\omega_k) = 0$.
\end{proof}
It follows that $\mathbf{T}^1 = \widetilde{\T}^1$ in weight~one because $T_p \in \mathbf{T}^1$ and $\mathfrak{m} \mathbf{T}^1 = {\widetilde{\mathfrak{m}}} \mathbf{T}^1$.
\subsection{Doubled modules}
We define the notion of a doubled module with respect to $\mathbf{T}$ and~$\widetilde{\T}$.
\begin{df} Let $N \subset H^0(X,\omega^{\kern+0.1em{p}}_k)$ be invariant under the action of $\widetilde{\T}$,
let~$\widetilde{I} = \mathrm{Ann}_{\widetilde{\T}}(N)$,
and let $I = \mathrm{Ann}_{\mathbf{T}}(N) = \widetilde{I} \cap \mathbf{T}$. We say that $N$ is doubled if the action of $\widetilde{\T}$ on $N$ acts
faithfully through a quotient $\widetilde{\T}/{\widetilde{I}}$ such that
$$\mathrm{length}(\widetilde{\T}/{\widetilde{I}}) = 2 \cdot \mathrm{length}(\mathbf{T}/I).$$
\end{df}
\begin{lemma} There exists a maximal doubled sub-module of $H^0(X,\omega^{\kern+0.1em{p}}_k)_{\mathfrak{m}}$.
\end{lemma}
\begin{proof}
If $\dim H^0(X,\omega^{\kern+0.1em{p}}_k)[\mathfrak{m}] = 1$, then $H^0(X,\omega^{\kern+0.1em{p}}_k)^{\vee}_{\mathfrak{m}}$
is free of rank one over $\mathbf{T}_{\mathfrak{m}}$ and $\widetilde{\T}_{\mathfrak{m}}$, so $\mathbf{T}_{\mathfrak{m}} \simeq \widetilde{\T}_{\mathfrak{m}}$ in that case,
and the maximal doubled quotient is trivial. Hence we assume that
$\dim H^0(X,\omega^{\kern+0.1em{p}}_k)[\mathfrak{m}] = 2$. By Nakayama's Lemma applied to $\mathbf{T}_{\mathfrak{m}}$, it follows that
$\widetilde{\T}_{\mathfrak{m}}$ has rank at most $2$ over $\mathbf{T}_{\mathfrak{m}}$, or equivalently that $T_p$ satisfies a quadratic
relation. If $N$ is doubled, however,
then~$\widetilde{\T}_{\mathfrak{m}}/\widetilde{I}$ must be free of rank two as a~$\mathbf{T}_{\mathfrak{m}}/I$-module.
In particular,
$I$ must act trivially on $\widetilde{\T}_{\mathfrak{m}}/\mathbf{T}_{\mathfrak{m}}$,
so it must contain the annihilator of this module. Let $J$ be the
annihilator of $\widetilde{\T}_{\mathfrak{m}}/\mathbf{T}_{\mathfrak{m}}$ as a~$\mathbf{T}_{\mathfrak{m}}$-module. This is an ideal of~$\mathbf{T}_{\mathfrak{m}}$;
we claim that it is actually an ideal of~$\widetilde{\T}_{\mathfrak{m}}$.
By definition, if~$a \in \mathbf{T}_{\mathfrak{m}}$ is any element, then~$a$ annihilates~$\widetilde{\T}_{\mathfrak{m}}/\mathbf{T}_{\mathfrak{m}}$
if and only if it lies in~$J$.
Equivalently, we have $a x \in \mathbf{T}_{\mathfrak{m}}$ for all~$x \in \widetilde{\T}_{\mathfrak{m}}$ if and only if~$a \in J$.
To show that~$J$ is an ideal of~$\widetilde{\T}_{\mathfrak{m}}$, it suffices to show
that $a T_p \in J$. By the previous equivalence, we have~$a T_p \in \mathbf{T}_{\mathfrak{m}}$.
Moreover, since~$a x \in \mathbf{T}_{\mathfrak{m}}$ for every~$x \in \widetilde{\T}_{\mathfrak{m}}$, we also have~$a T_p x \in \mathbf{T}_{\mathfrak{m}}$
for every~$x \in \widetilde{\T}_{\mathfrak{m}}$. Thus~$a T_p \in J$, and~$J$ is an ideal of~$\widetilde{\T}_{\mathfrak{m}}$.
We then observe that
$H^0(X,\omega^{\kern+0.1em{p}}_k)[J]$ is doubled, and is thus the maximal doubled
sub-module.
\end{proof}
The ideal~$J$ is the analogue in this context of the (global) doubling ideal
denoted$\JJ^{\mathrm{glob}}$ in~\cite{CG}.
Let $M \subset H^0(X_{k},\omega^{\kern+0.1em{p}})_{\mathfrak{m}}$ be a maximally doubled module.
Hence $M^{\vee}$ is free of rank $2$ over ${\T}/J$ and free of rank~$1$ over $\widetilde{\T}/J$,
where $J = \mathrm{Ann}_{{\T}}(M)$.
The only maximal ideal of ${\T}/J$ is $\mathfrak{m}$, so ${\T}/J$ is a finite local ring.
Let $\ker(\theta)$ denote the subset of elements annihilated by the $\theta$ operator,
and let $\ker_M(\theta) = \ker(\theta) \cap M$
\begin{lemma} \label{lemma:faithful} The module $M/\ker_M(\theta)$ is a faithful ${\T}/J$-module.
\end{lemma}
\begin{proof} We have a surjection:
$$({\T}/J)^2 \simeq M^{\vee} \rightarrow \ker_M(\theta)^{\vee}$$
The module $\ker_M(\theta)$ is isomorphic as a $\T$-module to $H^0(X,\omega_k)_{\mathfrak{m}}$. Hence,
by Lemma~\ref{lemma:count}, $\ker_M(\theta)^{\vee}$ is cyclic as a $\T/J$-module.
If $K$ denotes the kernel, it follows that $K/\mathfrak{m} \rightarrow ({\T}/\mathfrak{m})^2$ has non-trivial image.
Let $x \in K$ denote an element which maps to a non-zero element in $({\T}/\mathfrak{m})$. Then the cyclic
module in $({\T}/J)^2$ generated by $x$ is a faithful ${\T}/J$-module, and hence $K$ is also a faithful
${\T}/J$-module. We then have $M/\ker_M(\theta) = K^{\vee}$.
\end{proof}
\begin{df} There is a $\T$-equivariant pairing
$\T/J \times M \rightarrow k$ defined as follows:
$$\langle T_{n}, f \rangle = a_{1}(T_n f).$$
\end{df}
\begin{lemma} \label{lemma:perfect} The map $\langle *,* \rangle$
is a perfect pairing between $\T/J$ and $M/\ker_M(\theta)$.
\end{lemma}
\begin{proof} If $f = \sum a_n q^n$ and $\theta(f) = 0$, then $a_n = 0$ for all $(n,p) = 1$, so
$\langle T_n,f \rangle = 0$ for all $T_n \in \T$, and hence for all $T \in \T$. Conversely, if $\langle T_n ,f \rangle = 0$, then $a_n = 0$
for all $(n,p) = 1$ and $f$ lies in the kernel of $\theta$. Now suppose that $\langle T,f \rangle = 0$ for all $f \in M$.
Since the map is Hecke equivariant, it follows that
$$\langle T_n, T \kern-.02em{f} \rangle = \langle T \kern+0.1em{T_n}, f \rangle = \langle T_n T, f \rangle = \langle T, T_n f\rangle = 0$$
for all $n$, and hence $T \kern-.02em{f}$ is trivial in $M/\ker_M(\theta)$. But $\T/J$ acts faithfully on $M/\ker_M(\theta)$, so $T = 0$.
\end{proof}
\begin{lemma} $\ker_M(\theta)$ is a faithful $\T/J$-module, and $\ker_M(\theta)^{\vee}$ is free over $\T/J$ of rank one.
\end{lemma}
\begin{proof} By definition, $M$ is free of rank two over $\T/J$, and so it has the same
length as $(\T/J)^2$. However, by Lemma~\ref{lemma:perfect},
$M/\ker_M(\theta)$ and $\T/J$ have the same dimension over $k$, and hence the same length.
It follows that $\ker_M(\theta)$ has the same length as $\T/J$. Since $\dim \ker_M(\theta)[\mathfrak{m}] = 1$,
the module $\ker_M(\theta)^{\vee}$ is cyclic of the same length as $\T/J$, and thus free of rank one
over $\T/J$. Hence $\ker_M(\theta)$ is also faithful as a $\T/J$-module.
\end{proof}
Since, by Lemma~\ref{lemma:count}, $\ker(\theta)^{\vee}_{\mathfrak{m}}$ is also free of rank one over $\T^1 = \T^1_{Q,k}$,
the Hecke algebra in weight one, and
$\ker_M(\theta)^{\vee}$ is a quotient of this module, we deduce the
immediate corollary:
\begin{corr} \label{cor:pickle}There is a surjection $\T^1 \rightarrow \mathbf{T}/J$.
\end{corr}
Now let us fix~$X = X_H(N\kern-0.05em{Q})$, and suppose that~$\mathfrak{m}$ and~$\widetilde{\mathfrak{m}}$
correspond to our residual Galois representation~$\overline{\rho}$ together with a suitable choice of~$a_{\ell}$.
\begin{prop} \label{prop:porpoise} There exists a doubled submodule $M \subset H^0(X_H(N\kern-0.05em{Q}),\omega^{\kern+0.1em{p}}_k)_{\mathfrak{m}}$ such that the action of
$\mathbf{T}$ on $M$ acts faithfully through $R^1_Q/\varpi$. In particular, there is a surjection~$\mathbf{T}/J \rightarrow R^1_Q/\varpi$.
\end{prop}
\begin{proof}
Let~$\widetilde{R}^1_Q$ denote the modification of~$R_Q$ where one also takes into account an eigenvalue~$\alpha_p$
of~$\rho(\mathrm{Frob}_p)$ (recall that representations associated to~$D^1_Q$ are unramified at~$p$).
The ring~$\widetilde{R}^1_Q$ is a finite flat degree~$2$ extension of~$R^1_Q$ given as the quotient of~$R^1_Q[\alpha_p]$
by the monic quadratic polynomial~$\alpha_p$ corresponding to the characteristic polynomial
of Frobenius in the universal representation associated to~$R^1_Q$.
Let us distinguish two cases. The first is when the eigenvalues of~$\overline{\rho}(\mathrm{Frob}_p)$ are distinct, and the
second is when they are the same (in the latter case,~$\overline{\rho}(\mathrm{Frob}_p)$ may or may not be scalar).
Let~$\Sigma$ denote the set of eigenvalues, so~$|\Sigma| = 2$ or~$1$.
If~$|\Sigma| = 1$, then~$\widetilde{R}^1_Q$ is a local ring, and if~$|\Sigma| = 2$, it is a semi-local ring with two maximal
ideals; indeed, by Hensel's Lemma the quadratic relation satisfied by~$\alpha_p$ splits over~$R^1_Q$, and so there is an isomorphism~$\widetilde{R}^1_Q = R^1_Q \oplus R^1_Q$ in this case.
(This is essentially Lemma~\ref{lemma:hensel}.)
In particular, the quadratic polynomial has exactly two roots in~$\widetilde{R}^1_Q$.
There is a surjection
$$\bigoplus_{\Sigma} R_Q \rightarrow \bigoplus_{\Sigma} R_Q/\varpi \rightarrow \widetilde{R}^1_Q/\varpi.$$
Here the sum is over the rings~$R_Q$ corresponding to each choice of eigenvalue~$a_p \in \Sigma$.
The latter map sends~$\alpha_{\ell}$ to~$\alpha_{\ell}$ for all~$\ell$ dividing~$ N\kern-0.05em{Q}$. If~$|\Sigma| = 2$,
then each~$\alpha_p$ goes to the corresponding eigenvalue of Frobenius.
If~$|\Sigma| = 1$, then~$\alpha_p$ goes to~$\alpha_p$. These maps are well defined because,
after reduction modulo~$\varpi$, all the local conditions (including the determinant)
in the definition of~$D^1$ and~$D^1_Q$ coincide with the exception of~$\ell = p$. For~$\ell = p$,
the enriched ring~$\widetilde{R}^1_Q$ recieves a map from~$R^{\mathrm{mod}}_p$, because~$\alpha_p$ in~$R^{\mathrm{mod}}_p$
is exactly an eigenvalue of Frobenius.
Since $\widetilde{\T}_{Q,k,\widetilde{\mathfrak{m}}} \simeq R_Q/\varpi$ by Corollary~\ref{corr:metro}, there is a surjection
$$\bigoplus_{\Sigma} \widetilde{\T}_{Q,k,\widetilde{\mathfrak{m}}} \rightarrow \widetilde{R}^1_Q/\varpi.$$
Since $\bigoplus H^0(X_H(N\kern-0.05em{Q}),\omega^{\kern+0.1em{p}}_k)_{\widetilde{\mathfrak{m}}}$
is co-free over $\bigoplus \widetilde{\T}_{Q,k,\widetilde{\mathfrak{m}}}$, there certainly exists a module $M$
such that the action of~$\widetilde{\T}$ on~$M$ is precisely via this quotient~$\widetilde{R}^1_{Q}/\varpi$.
Yet this quotient is also finite flat of degree two over~$R^1_Q/\varpi$, which is precisely the image of~$\mathbf{T}$ under this map.
Hence the submodule~$M$ is doubled, and the corresponding action of~$\mathbf{T}$ is via~$R^1_Q/\varpi$.
The final claim follows from the fact that~$J$ is the ideal corresponding to the largest doubled submodule.
\end{proof}
\begin{remark} \emph{One alternative way to write this paper was to define the functors~$D_Q$, etc. \emph{without}
making a fixed choice of~$a_{\ell}$. This would have amounted to replacing the universal
local deformation rings~$R_Q$, etc. by universal semi-local deformation rings, which would be
isomorphic to a direct sum
over all the local rings in this paper and over all possible choices of~$a_{\ell}$. We have decided to work with
the version of these rings in which choices have been made, however, as evidenced by the proof
of the previous proposition, one
still has to deal with semi-local rings in some cases at~$\ell = p$, because when~$\overline{\rho}(\mathrm{Frob}_p)$ has distinct eigenvalues,
the corresponding maximal ideal in weight one is determined by the sum~$\alpha + \beta$ of these eigenvalues whereas
the local rings in higher weight require a choice of~$\alpha$ or~$\beta$. }
\end{remark}
\section{Passage from weight~$p$ to weight~one}
Let $\T^1_{Q} = \T^1_{Q,E/\mathcal{O}}$. The rings ~$\mathbf{T}^1$, $\widetilde{\T}^1$, $\mathbf{T}$, and~$\widetilde{\T}$ of the previous section were
abbreviations for
the rings~$\mathbf{T}^1_{Q,k}$, $\widetilde{\T}^1_{Q,k}$, $\mathbf{T}_{Q,k}$ and~$\widetilde{\T}_{Q,k}$ respectively; we return to this expanded notation now.
\begin{corr} \label{corr:nail} There is an isomorphism $R^1_{Q}/\varpi \rightarrow \T^1_{Q,k,\mathfrak{m}} \simeq \T^1_{Q,\mathfrak{m}}/\varpi$.
\end{corr}
\begin{proof} For the first isomorphism, it suffices to note that there exists a map
$$R^1_{Q}/\varpi \rightarrow \T^1_{Q,k,\mathfrak{m}} \rightarrow \mathbf{T}_{Q,k,\mathfrak{m}}/J \rightarrow R^1_{Q}/\varpi$$
whose composite is the identity.
The existence of the first map follows from the fact that Galois representations in weight~one are unramified
at~$p$, which follows from Theorem~3.11 of~\cite{CG} (together with the appropriate local--global
compatibility away from~$p$, which follows as in the proof of Theorem~3.11 of~\cite{CG} by reduction to characteristic zero in higher weight, together with the proof of Proposition~\ref{prop:action}). The second map comes from Corollary~\ref{cor:pickle}. The
existence of the
third map follows from Proposition~\ref{prop:porpoise}. The identification of $\T^1_{Q,k,\mathfrak{m}}$
with $\T^1_{Q,\mathfrak{m}}/\varpi$ follows from the fact that
$$H^0(X_H(N\kern-0.05em{Q}),\omega_k)_{\mathfrak{m}} = H^0(X_H(N\kern-0.05em{Q}),\omega_{E/\mathcal{O}})_{\mathfrak{m}}[\varpi]$$
and the fact that $H^0(X_H(N\kern-0.05em{Q}),\omega_{E/\mathcal{O}})^{\vee}$ is free over $\T^1_{Q,\mathfrak{m}}$ by Lemma~\ref{lemma:count}.
\end{proof}
\subsection{Modularity in weight one} \label{section:one}
Now that we have an isomorphism $R^1_{Q}/\varpi \simeq \mathbf{T}^1_{Q,\mathfrak{m}}/\varpi$ for all collections of Taylor--Wiles primes $Q = Q_D$,
and we apply the machinery of~\cite{CG}, in particular Prop.~2.3 as applied in~\S3.8 of~\emph{ibid}.
We patch the modules
$H^0(X_{H_D}(N\kern-0.05em{Q}_D),\omega_{E/\mathcal{O}})^{\vee}_{\mathfrak{m}}$,
where~$Q$ is a collection of Taylor--Wiles primes~$x \equiv 1 \mod p^D$, and~$H$
is the subgroup of~$(\mathbf{Z}/N\kern-0.05em{Q}_D \mathbf{Z})^{\times}$ which is generated by the kernel
of~$(\mathbf{Z}/ Q_D \mathbf{Z})^{\times} \rightarrow (\mathbf{Z}/p^D \mathbf{Z})^{\times}$ and the~$p$-Sylow of~$(\mathbf{Z}/N\mathbf{Z})^{\times}$.
We obtain
a module $M_{\infty}$ which is a module over
the framed ring
of diamond operators $S^{\square}_{\infty}$,
and a patched deformation ring $R^{1,\square}_{\infty}$ which is also an algebra
over this ring. In contrast to~\cite{CG}, the ring~$R^{1,\square}_{\infty}$ is a power series ring over a completed
tensor product
$$R^{\mathrm{mod}}:=\widehat{\bigotimes}_{\ell | N} R^{\mathrm{mod}}_{\ell},$$
instead of a power series ring over~$\mathcal{O}$.
Because
the modules $H^0(X_{Q},\omega_{E/\mathcal{O}})^{\vee}_{\mathfrak{m}}$ are free over $\mathbf{T}_{Q,\mathfrak{m}}$, the module $M_{\infty}$ is
cyclic as an $R^{1,\square}_{\infty}$-module.
Hence we know that:
\begin{enumerate}
\item $M_{\infty}/\varpi$ is free of rank one over $R^{1,\square}_{\infty}/\varpi$, as follows
from our mod-$p$ modularity results above, in particular Corollary~\ref{corr:nail}.
\item $M_{\infty}$ is pure of co-dimension one as an $S^{\square}_{\infty}$-module; that is,
$M_{\infty}$ is a torsion $S^{\square}_{\infty}$-module, and there
exists a presentation:
$$0 \rightarrow (S^{\square}_{\infty})^n \rightarrow (S^{\square}_{\infty})^n \rightarrow M_{\infty} \rightarrow 0;$$
this is exactly the output of the construction of~\cite{CG}.
\end{enumerate}
The second result is essentially a formal consequence of the method of~\cite{CG} rather than anything in this paper.
This on its own is enough to show that~$M_{\infty}$ will certainly be supported on~\emph{some} components
of the generic fibre of~$R^{1,\square}_{\infty}$. However, as soon as~$S$ contains primes for which~$\overline{\rho}$ is unramified
(that is, as soon as we are at non-minimal level),
the ring~$R^{1,\square}_{\infty}[1/\varpi]$ will have multiple components. The usual technique for showing that the support of~$M_{\infty}$
is spread over all components is to produce modular lifts with these properties. In our context this is not
possible: there are no weight one forms in characteristic zero which are Steinberg at a finite place~$q$
(see the proof of Prop.~\ref{prop:katzfail} below).
Our replacement for producing modular points in characteristic zero is to work
on the special fibre, and to show that~$M_{\infty}/\varpi$
is (in some sense) spread out as much as possible over~$R^{1,\square}_{\infty}/\varpi$. And we do this (and this is the main point
of everything so far) by working in weight~$p$ and then descending back to weight one using the doubling method.
In particular, we know that~$M_{\infty}/\varpi$ is free of rank one over $R^{1,\square}_{\infty}/\varpi$.
From these two properties, we will now deduce that $M_{\infty}$ is free of rank one over $R^{1,\square}_{\infty}$,
which will imply all our modularity results.
By Nakayama's Lemma, there is certainly an exact sequence of $S^{\square}_{\infty}$-modules.
$$0 \rightarrow K \rightarrow R^{1,\square}_{\infty} \rightarrow M_{\infty} \rightarrow 0$$
It suffices to show that~$K = 0$. By Nakayama's Lemma again, it suffices to show that $K/\varpi = 0$.
Tensoring with $S^{\square}_{\infty}/\varpi$ (that is, reducing modulo~$\varpi$), we get a long exact sequence
$$\mathrm{Tor}^1(S^{\square}_{\infty}/\varpi,M_{\infty}) \rightarrow K/\varpi
\rightarrow R^{1,\square}_{\infty}/\varpi \stackrel{\simeq}{\longrightarrow} M_{\infty}/\varpi \rightarrow 0.$$
Here the last map is an isomorphism by property~$(1)$ above. Hence, to prove that~$K/\varpi$ is trivial, it suffices
to show that
$$\mathrm{Tor}^1(S^{\square}_{\infty}/\varpi,M_{\infty}) = M_{\infty}[\varpi]$$
is trivial. If $M_{\infty}[\varpi]$ is non-trivial, then, from the purity of $M_{\infty}$, we claim that $M_{\infty}/\varpi$ will have positive rank over
$S^{\square}_{\infty}/\varpi$. To see this, simply tensor the presentation of $M_{\infty}$ with
$S^{\square}_{\infty}/\varpi$ to obtain the exact sequence:
$$0 \rightarrow \mathrm{Tor}^1(S^{\square}_{\infty}/\varpi,M_{\infty}) \rightarrow
(S^{\square}_{\infty}/\varpi)^n \rightarrow (S^{\square}_{\infty}/\varpi)^n \rightarrow M_{\infty}/\varpi \rightarrow 0,$$
from which it follows that
$$\mathrm{rank}_{S^{\square}_{\infty}/\varpi} M_{\infty}/\varpi =\mathrm{rank}_{S^{\square}_{\infty}/\varpi}
\mathrm{Tor}^1(S^{\square}_{\infty}/\varpi,M_{\infty}).$$
If~$\mathrm{Tor}^1(S^{\square}_{\infty}/\varpi,M_{\infty})$ is non-zero, then, as it is a submodule of a free module over
$S^{\square}_{\infty}/\varpi$ which has no annihilator, it also has no annihilator
as an~$S^{\square}_{\infty}/\varpi$ module. However,
a module with no annihilator over a power series ring over~$k$ certainly must have positive rank.
This implies that (if~$M_{\infty}[\varpi]$ is non-zero) that~$M_{\infty}/\varpi$ has positive rank.
Yet this
contradicts the fact that $R^{1,\square}_{\infty}/\varpi \simeq M_{\infty}/\varpi$ does not have positive rank, as $R^{1,\square}_{\infty}$
is flat over $\mathcal{O}$ (Lemma~\ref{lemma:kisin}) and of smaller dimension than the ring of diamond operators (by one).
Hence $M_{\infty}$ is free of rank one over $R^{1,\square}_{\infty}$. But now specializing down to finite level,
we deduce that $H^0(X,\omega_{E/\mathcal{O}})^{\vee}_{\mathfrak{m}}$ is free of rank one over $R^1$, which
proves Theorem~\ref{theorem:main}.
\subsection{Producing Torsion Classes} \label{section:blog}
Let
$$f = \sum a_n q^n \in S_1(\Gamma_H(N),\eta)$$
be a cuspidal eigenform of some level~$N$ and character~$\eta$. Let
$$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(E)$$
denote the corresponding Artin representation. Assume
that~$\rho|D_{\ell}$ is reducible for any prime~$\ell$ dividing~$N$.
\begin{prop} \label{prop:katzfail} Let~$f$ be as above. Let~$p>2$ be a prime such that~$\overline{\rho}_f$ is absolutely
irreducible and~$p$ is prime to the level~$N$ of~$f$ and the order of~$\eta$.
Then there exists a set of primes~$\ell$ of positive
density so that, for each such~$\ell$, the map:
$$H^0(X( \Gamma_H(N) \cap \Gamma_0(\ell))_{\mathbf{Z}_p},\omega)_{\mathfrak{m}} \otimes \mathbf{F}_p \rightarrow H^0(X( \Gamma_H(N) \cap \Gamma_0(\ell))_{\mathbf{F}_p},\omega)_{\mathfrak{m}}$$
is not surjective. \end{prop}
\begin{remark} \emph{This implies that Katz' base-change theorem
(Theorem~1.7.1 of~\cite{KatzBig}) fails as badly as possible in weight one.}
\end{remark}
\begin{proof} Suppose, to the contrary, that all
such forms of this level lift to characteristic zero.
There are no
forms in characteristic zero which are new at~$\ell$ of level $\Gamma_0(\ell)$, because any such form would have to be (up to unramified twist)
Steinberg at~$\ell$,
and no weight one form in characteristic zero can be Steinberg at any place.
The easiest way to see this is that the eigenvalue of~$U_{\ell}$ would have to be a root of unity times
${\ell}^{-1/2}$, but this is impossible because Hecke eigenvalues
of modular forms are algebraic integers.
Hence any Galois representation arising
from forms of this level must come from level $\Gamma_H(N)$, and so in particular be unramified at~${\ell}$.
Thus, by Theorem~\ref{theorem:main}, it suffices to show
that there is a non-trivial deformation of~$\overline{\rho}$ to
the dual numbers which is minimal at~$N$, corresponds to
a quotient of~$R^{\mathrm{mod}}_{{\ell}}$ at the new auxiliary prime~${\ell}$, and is
unramified everywhere else.
The reduced tangent space of the unramified
deformation ring is given by the
Selmer group $H^1_{\emptyset}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$.
Denote the dual Selmer group by $H^1_{\emptyset^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho})(1))$.
Since these groups are both finite, there exists a finite extension
$F/\mathbf{Q}$ which contains the fixed field of $\ker(\rho)$ and such that
all the classes in $H^1_{\emptyset^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho})(1))$ split
completely. Let~${\ell}$ be a prime which splits completely in~$F(\zeta_p)$.
Let $H^1_{\Sigma}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$ denote the modified Selmer group
where classes are allowed to be arbitrarily ramified at~${\ell}$. By construction of~${\ell}$, the dual Selmer group $H^1_{\Sigma^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$
consisting of all dual Selmer classes which split completely at~${\ell}$ is equal
to $H^1_{\emptyset^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))$, because the localization map
factors through the restriction to $G_F$, and by construction all classes in the latter group are assumed to
split completely over~$\mathbf{Q}_{\ell}$.
Hence, the Greenberg--Wiles Euler characteristic formula:
$$\begin{aligned}
\frac{|H^1_{\Sigma}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))|}{|H^1_{\emptyset}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))|} & \ =
\frac{|H^1_{\Sigma}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))|}{|H^1_{\Sigma^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho})(1))|} \cdot
\frac{|H^1_{\emptyset^*}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho})(1))|} {|H^1_{\emptyset}(\mathbf{Q},\mathrm{ad}^0(\overline{\rho}))|}\\
& \ =\frac{|H^1(\mathbf{Q}_{\ell},\mathrm{ad}^0(\overline{\rho}))|}{|H^1(\mathbf{F}_{\ell},
\mathrm{ad}^0(\overline{\rho}))|} = |H^0(\mathbf{Q}_{\ell},\mathrm{ad}^0(\overline{\rho})(1))| = |\mathrm{ad}^0(\overline{\rho})(1)|, \end{aligned}$$
the final equality coming from the assumption that~${\ell}$ splits completely in~$F(\zeta_p)$.
Note that $\dim \mathrm{ad}^0(\overline{\rho})(1) = 3 > 0$.
It follows
that for such choices of~$\ell$,
there exists a deformation
$$\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(k[\epsilon]/\epsilon^2)$$
which is minimal at all primes away from~$\ell$ and genuinely
ramified at~$\ell$.
Moreover, $\overline{\rho} |D_{\ell}$ is trivial and $\ell \equiv 1 \mod p$.
It suffices to show that the corresponding deformation
arises from a quotient of~$R^{\mathrm{mod}}_{\ell}$,
which was described explicitly in this case
by Lemma~\ref{lemma:special}.
Since we are considering
fixed determinant deformations, the
trace of the image~$\rho(g)$ of any element~$g$ is $1 + \det(g) = 2$.
However, it is apparent the description of~$R^{\mathrm{mod}}_{\ell}$ in
Lemma~\ref{lemma:special} that all the relations apart from $\mathrm{Tr}(\rho(\tau)) = 2$ lie in~$\mathfrak{m}^2$, and so
are automatically satisfied for any deformation to $k[\epsilon]/\epsilon^2$. (Note that, associated to~$\rho$,
there is a corresponding
surjection~$R^{\mathrm{mod}}_{\ell} \rightarrow k[\epsilon]/\epsilon^2$ for any choice of~$\alpha_{\ell} \in 1 + \epsilon \cdot k[\epsilon]/\epsilon^2$.)
\end{proof}
\subsection{Proof of Theorem~\ref{theorem:blog}}
To prove Theorem~\ref{theorem:blog}, it suffices to apply Proposition~\ref{prop:katzfail} to suitably
chosen~$f$. Note that the class numbers of the fields~$\mathbf{Q}(\sqrt{-23})$ and~$\mathbf{Q}(\sqrt{-47})$ are~$3$ and~$5$ respectively. This gives rise to suitable weight one forms~$f$ with image~$D_3$ and~$D_5$ and level~$\Gamma_1(23)$ and~$\Gamma_1(47)$ respectively (both with quadratic nebentypus). Applying Proposition~\ref{prop:katzfail}, we deduce the
existence of mod-$p$ Katz modular forms which fail to lift for all $p \ne 2,3,23$ in the first example and
$p\ne 2,5,47$ in the second. For~$p=2$, the theorem is known by an example of Mestre~\cite{Mestre},
completing the proof.
\begin{remark} \emph{The computations of~\cite{schaeffer} (see also~\cite{buzzcomp}) suggest that Theorem~\ref{theorem:blog}
should also be true if one insists that~$f$ is an \emph{eigenform}. By Serre's conjecture, this would follow
if for each~$p > 5$ there existed an odd Galois representation:
$\overline{\rho}: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(k)$ unramified at~$p$ with image containing~$\mathrm{SL}_2(\mathbf{F}_p)$,
although proving this appears difficult.}
\end{remark}
\begin{remark} \emph{If the cuspform~$f$ in Proposition~\ref{prop:katzfail} is exceptional
--- that is, the projective image of~$\rho_f$ is~$A_4$,~$S_4$, or~$A_5$ --- then
the resulting torsion class at level $\Gamma_H(N) \cap \Gamma_0(q)$ will
not lift to characteristic zero at any higher level. The reason is that any form~$g$ with
$\overline{\rho}_f = \overline{\rho}_g$ will have to satisfy $\rho_f = \rho_g$ up to a~$p$-power twist, and the resulting
Hecke algebra in weight one cannot give rise to the infinitesimal deformations~$\rho$ which
arise in the proof of Proposition~\ref{prop:katzfail}.
}
\end{remark}
\bibliographystyle{amsalpha}
|
\section{Introduction}\label{sec1}
The abundance of dark matter in galaxies and the structure of the central engines of quasars are two interesting astrophysical problems that can be probed by gravitationally lensed quasars. The presence of point-like stars and their remnants in the otherwise smooth distribution of matter of the lens galaxy induces very strong local changes in the gravitational field that give rise to large changes in the magnification
of the lensed source compared to a smooth model known as microlensing (see the review by Wambsganss 2006). The amplitude of these anomalies is sensitive to the local stellar surface mass density fraction as compared to that of the dark matter at the image location (Schechter \& Wambsganss 2002). It is also sensitive to the source size, because larger sources more heavily smooth the magnification patterns, and so are less magnified than smaller sources. Both effects are important for the amplitude and statistics of the microlensing we see in the images of multiply imaged quasars.
Two basic experimental approaches have been used to measure microlensing in lensed quasars: photometric monitoring and single epoch spectroscopy/photometry. Photometric monitoring measures the magnification changes (microlensing variability) induced by the relative motions of the quasar source, the lens galaxy stars and the observer (Chang \& Refsdal 1979; Gott 1981; Kayser, Refsdal \& Stabell 1986) by comparing the light curves of the lensed images. After correcting the curves for the time delay between the images,
the time varying microlensing signal can be analyzed (e.g. Kochanek 2004). Alternatively, if frequent monitoring of the source is not available, valuable information can still be extracted by using single epoch spectroscopy/photometry. In this case, emission lines
or a smooth macrolens model can be used
as a reference with respect to which the microlensing of the different images can be measured.
The advantage of using the emission lines as a reference is that they are much less sensitive to microlensing (e.g. Guerras et al. 2013)
and systematic errors in the macro lens model (see Mediavilla et al. 2009, hereafter MED09).
Single epoch microlensing magnification estimates are observationally much less expensive than photometric monitoring, and can be easily obtained for relatively large samples of lensed quasars. The challenge is to adequately control for the systematic uncertainties in the reference magnification created by time variability, lens substructures and absorption/extinction.
Gravitational microlensing in the optical has proven a very powerful tool in many studies of individual lenses to estimate the size
of the quasar accretion disks (see, e.g., Morgan et al. 2008, 2010, 2012, Mediavilla et al. 2011b, Mu\~noz et al. 2011, Motta et al. 2012, Mosquera et al. 2009, 2011, 2013, Rojas et al. 2014) and also the fraction of surface mass density in the form of stars (Schechter \& Wambsganss 2004, Kochanek et al. 2004, 2006, Bate et al. 2011).
In the optical, most of the works based on large lens samples are focused on either the fraction of mass in stars $\alpha$ (e.g., Mediavilla et al. 2009) or the quasar size $r_s$ (e.g. Blackburne et al. 2011; Jim\'enez-Vicente et al. 2012) separately. However, since microlensing is sensitive to both physical effects, some degeneracy was expected in the microlensing based estimates of these parameters (see MED09). This has been shown by Jim\'enez-Vicente et al. (2015) who, despite the strong covariance between stellar mass fraction $\alpha$ and source size $R_{1/2}$, found reasonably good estimates for both parameters of $\alpha= 0.21\pm0.14$ (at a radius of approximately $1.8 R_{e}$ where $R_e$ is the effective radius of the lens, and $R_{1/2}=2.0^{+1.0}_{-0.6}\times 10^{16}\sqrt{M/0.3M_\sun}$ cm at an average rest wavelength of $\lambda=1734$ \AA, where $M$ is the mean mass of the microlenses, .
Microlensing of multiply imaged quasars has also been observed in X-rays (e.g. Pooley et al. 2006, 2007, Blackburne et al. 2006, Morgan et al. 2008, Chartas et al. 2009, Dai et al. 2010, Morgan et al. 2012, Blackburne et al., 2014, 2015, Mosquera et al. 2013, MacLeod \& Morgan, 2014). Size estimates in these studies generally find that the soft X-rays are emitted from a region with a size of $R_{1/2}\sim 20r_g$ (where $r_g=GM_{BH}/c^2$ is the gravitational radius of the black hole), while hard X-rays may come from a slightly more compact region. The ratio of the sizes of the optical and the soft X-ray sources is of order $\sim$ 10.
Studies using samples of lenses with X-ray data have been used to estimate the stellar mass fraction $\alpha$ (Pooley et al. 2012, Schechter et al. 2014), and they have generally assumed that the size of the X-ray emitting region is small enough to have little impact on the estimate of $\alpha$. MED09 showed that there was a significant covariance between stellar mass fraction $\alpha$ and the source size $r_s$ in the optical, and Jim\'enez-Vicente et al. (2015) found that despite the strong covariance, a determination of both parameters was still possible.
As X-ray microlensing has proven to be sensitive to both size and fraction of mass in stellar mass objects, it is reasonable to wonder
whether there is also a covariance between these two parameters for X-rays, similar to what is found at optical wavelengths, and how it may affect the estimates of both parameters. A joint study of both parameters using X-ray microlensing estimates is therefore needed to clarify this point.
Beyond measuring a mean value for the stellar mass fraction at an average radius, $\alpha$, measuring the radial profile
$\alpha$ is a key ingredient in understanding how galaxies formed; in particular the interaction of dark and baryonic matter during the initial collapse (including processes like baryonic cooling, settling, star formation and feedback) and subsequent mergers (cf. Diemand \& Moore 2011). Most studies addressing this issue use other methods (X-ray emission, dynamics or strong lensing) to estimate total masses, and use the brightness/color distribution and a suitable IMF to estimate the stellar mass (see, for example, the review by Courteau et al. 2014). But measuring radial profiles using these procedures is model dependent, particularly through the IMF, which is itself a subject of study using lensing (e.g. Treu et al. 2010, Sonnenfeld et al. 2012).
Microlensing can provide an estimate of the local stellar mass fraction $\alpha$ at the location of the images without any strong dependence on the specific shape of the mass distribution.
In this respect, microlensing studies, either based on X-ray or optical data, have mainly focused on obtaining an estimate of the fraction of mass in stars at an average distance. Pooley et al. (2012) made a first attempt to measure a radial gradient in the stellar fraction and did not detect a significant gradient.
Here we make a new attempt to detect such a radial gradient in the stellar mass fraction.
We will combine the available optical and X-ray data to estimate the fraction of mass in the form of stars at different radii within lens galaxies. We will subsequently check the consistency of our estimates of the stellar/dark matter fraction by comparing them with previous results and models. In Section 2 we analyze the sensitivity of X-ray microlensing to the source size,
and we discuss the dependence of X-ray sizes on the black hole mass. In Section 3, we address the
joint estimate of the stellar mass fraction and the typical size of the X-ray emitting region.
Section 4 is devoted to study the radial profile of the stellar mass fraction in lens galaxies. Finally, the main results are summarized in Section 5.
\section{The Dependence of X-ray Sizes on Black-Hole Mass}\label{sec2}\
We start by estimating the size of the X-ray emitting region for 10 quadruple lens systems using the flux ratios from Schechter et al. (2014), who used the soft (0.5-8 keV) X-rays fluxes from Pooley et al. (2007) and Blackburne et al. (2011). To compare microlensing magnification estimates for different models to observations, we follow the procedures of Jim\'enez-Vicente et al. (2012).
We compute magnification maps for each of the four images in the 10 systems using the Inverse Polygon Mapping technique (Mediavilla et al. 2006, 2011a). We take the values for $\kappa$ and $\gamma$ provided by Schechter et al. (2014), and put 20\% of the surface mass density in the form of stars, as derived from microlensing in the optical by Jim\'enez-Vicente et al. (2015). We use stars of fixed mass, which we have chosen to be $M=0.3M_\sun$. The maps have $2000\times 2000$ pixels and span 100 light days with a fixed pixel size of 0.05 light days.
\begin{deluxetable}{llr}
\tabletypesize{\footnotesize}
\tablewidth{0pt}
\tablecolumns{3}
\tablecaption{Microlensing data.\label{tab1}}
\tablehead{\multicolumn{3}{c}{X-rays}\\ \hline
\colhead{Object} & \colhead{Pair} & \colhead{$\Delta m$} }
\startdata
HE0230$-$2130 & B$-$A & 0.90 \\
& C$-$A & -0.21 \\
& D$-$A & -0.76 \\
MGJ0414+0534 & A2$-$A1 & 0.56 \\
& B$-$A1 & -0.53 \\
& C$-$A1 & -0.32 \\
HE0435$-$1223 & B$-$A & 1.14 \\
& C$-$A & 1.12 \\
& D$-$A & 0.63 \\
RXJ0911+0551 & B$-$A & 1.74 \\
& C$-$A & 2.29 \\
& D$-$A & 0.16 \\
SDSSJ0924+0219 & B$-$A & 0.34 \\
& C$-$A & 1.27 \\
& D$-$A & 2.00 \\
PG1115+080 & A2$-$A1 & 1.94 \\
& B$-$A1 & -0.51 \\
& C$-$A1 & -0.01 \\
RXJ1131$-$1231 & B$-$A & -3.01 \\
& C$-$A & -2.24 \\
& D$-$A & -3.43 \\
SDSSJ1138+0314 & B$-$A & 0.53 \\
& C$-$A & 0.90 \\
& D$-$A & 0.90 \\
B1422+231 & B$-$A & 0.84 \\
& C$-$A & 0.08 \\
& D$-$A & -0.16 \\
WFIJ2033-4723 & A2$-$A1 & -0.68 \\
& B$-$A1 & -0.63 \\
& C$-$A1 & -0.67 \\
\hline
\hline
\multicolumn{3}{c}{Optical}\\
\hline
\tablehead{\multicolumn{3}{c}{Optical}\\}
HE0047$-$1756 & B$-$A & -0.19 \\
HE0435$-$1223 & B$-$A & -0.24 \\
& C$-$A & -0.30 \\
& D$-$A & 0.09 \\
SDSS0806+2006 & B$-$A & -0.47 \\
SBS0909+532 & B$-$A & -0.60 \\
SDSS0924+0219 & B$-$A & 0.00 \\
FBQ0951+2635 & B$-$A & -0.69 \\
QSO0957+561 & B$-$A & -0.30 \\
Q1017$-$20 & B$-$A & -0.26 \\
HE1104$-$1805 & B$-$A & 0.60 \\
PG1115+080 & A2$-$A1 & -0.65 \\
B1422+231 & A$-$B & 0.16 \\
& C$-$B & 0.02 \\
& D$-$B & -0.08 \\
SBS1520+530 & B$-$A & -0.39 \\
WFIJ2033-4723 & B$-$C & -0.50 \\
& A2$-$A1 & 0.00 \\
\enddata
\end{deluxetable}
The source brightness is modelled as a Gaussian $I(r)\propto \exp(-r^2/2r_s^2)$. Mortonson et al. (2005) show that
estimates of the half light radius $R_{1/2}$ depend little on
the specific shape of the radial profile. For a Gaussian, the half light radius is $R_{1/2}=1.18 r_s$. We convolve the
magnification maps with Gaussians of 16 different sizes over a logarithmic
grid $\ln(r_s/0.05\, \mbox{lt-days})=0.3\times k$ with $k=0,\cdots, 15$, which spans
$r_s\sim 0.05$ to $r_s\sim 4.5$ light-days.
We want to compare the observed X-ray fluxes of the images with the predictions of a microlensing model as a function of $r_s$.
We model the magnitude of image $i$ as
\begin{equation}
\label{eq1}
m_i=m_0+\mu_i+\Delta m_i
\end{equation}
where $\mu_i$ and $\Delta m_i$ are, respectively, the macro and micro magnifications of image $i$. As we do not know the intrinsic flux of the source $m_0$, we will use one of the other three images as a reference.
We can calculate the differential microlensing magnification between image $i$ and the reference $r$ as
\begin{equation}
\label{eq2}
\Delta m_{ir}=m_i-m_r-(\mu_i-\mu_r)=(\Delta m_i-\Delta m_r),
\end{equation}
where the difference in the macro magnifications for each image can be accounted for from the lens model. The X-ray microlensing magnifications are presented in Table \ref{tab1}. In principle, errors in the macro model or other secondary effects (e.g. millilensing, extinction, intrinsic variability) can introduce additional noise in our differential microlensing estimates. The possible influence of these effects has been thoroughly analyzed by Schechter et al. (2014), who found that they have a modest influence in their microlensing results, with the
largest uncertainty coming from possible errors in the macro model. We try to account for this through the assumed uncertainty $\sigma=0.2$ mags in the differential magnifications. We can compare the measured differential microlensing magnitude with the prediction of the model and calculate a likelihood
for parameter $r_s$ using the observed differential microlensing magnifications, $\Delta m_{ir}^{obs}$,
\begin{equation}
\label{likel1}
L(r_s|\Delta m_{ir})=p(\Delta m_{ir}|r_s)=\sum_i \sum_r e^{-\chi^2/2},
\label{lik1}
\end{equation}
where
\begin{equation}
\chi^2=\frac{(\Delta m_{ir}-\Delta m_{ir}^{obs})^2}{\sigma^2}
\end{equation}
and $\sigma$ is a typical error in the estimate of $\Delta m_{ir}^{obs}$ which we have taken as 0.2 mags.
The summations in Equation \ref{likel1} are over $10^4$ points in the convolved magnification maps of images $i$ and $r$ respectively. The likelihood in Equation \ref{lik1} is therefore calculated using $10^8$ pixel pairs.
We calculate the total likelihood for lens $k$ by combining the likelihoods of the three image pairs relative to the reference image,
\begin{equation} L_k(r_s)=\prod_{i=1,3} L_i(r_s|\Delta m_{ir}), \end{equation}
and the joint probability distribution
\begin{equation} L(r_s)=\prod_{k=1,10} L_k(r_s), \end{equation}
is the product of the individual probabilities for all the lenses in the sample.
The resulting likelihood function using all 10 lenses favours very small sizes, as shown by the dashed line of Figure \ref{fig1}. The
distribution is, however, dominated by a single object, RXJ~1131$-$2131. If we exclude RXJ~1131$-$2131 from the sample, the likelihood function (see Figure 1) is rather different, with a clear maximum at $\log(R_{1/2}/{\rm cm})=15.5^{+0.3}_{-0.7}$ (68\% confidence interval), indicating that most objects in the sample have sizes close to $\sim$1 light day. This average value may look large at first sight, but it is in agreement with previous measurements (e.g. Mosquera et al. 2013, Morgan et al. 2008). RXJ~1131$-$2131 is an unusual case.
It has both the largest microlensing magnifications and the smallest estimated black hole mass,
$M_{BH}=6\times10^7 M_\sun$ (Peng et al. 2006), in the sample.
\begin{figure}
\epsscale{0.85}
\plotone{Fig1.eps}
\caption{Likelihood function for the size of the X-ray emission region $R_{1/2}$ for microlenses of mass $M=0.3M_\sun$. The dashed line shows the joint likelihood for the whole sample. The continuous line shows the likelihood excluding RXJ~1131$-$2131. \label{fig1}}
\end{figure}
\begin{figure}
\epsscale{0.85}
\plotone{Fig2.eps}
\caption{Half-light radii, $R_{1/2}$, of the X-ray emission region as a function of the mass of the central black hole $M_{BH}$ (diamonds).
Black hole masses are taken from Peng et al. (2006)
and Assef et al. (2011). Previous estimates are indicated as triangles for HE0435$-$1223 (Blackburne et al. 2014), HE1104$-$1805 (Blackburne et al. 2015), RXJ~1131$-$2131 (Dai et al. 2010), Q2237+0305 (Mosquera et al. 2013), QJ0158$-$4325 (Morgan et al. 2012), SDSS0924+0219 (McLeod \& Morgan 2014) and PG1115+080 (Morgan et al. 2008). For Q2237+0305, the hard and soft X-ray band estimates are shown separately. Upper limits are indicated by the arrows. The solid line is a power law fit to the size estimates, $\log(R_{1/2}/{\rm cm})=\log(R_9/{\rm cm})+x\log(M_{BH}/10^9M_\sun)$. The dot-dashed line corresponds to 24 gravitational radii ($R_{1/2}=24r_g=24GM_{BH}/c^2$).\label{fig2}}
\end{figure}
\begin{figure}
\epsscale{0.85}
\plotone{Fig3.eps}
\caption{Probability distribution for the scaled size of the X-ray emission region size $R_9$ assuming that $R_{1/2}=R_9(M_{BH}/10^9 M_\sun)$.
The upper x-axis shows the size in units of the gravitational radius $r_g=M_{BH}G/c^2$. We assume microlens masses
of $M=0.3M_\sun$. \label{fig3}}
\end{figure}
This suggests that we should examine the scaling with mass even though single epoch microlensing estimates will have large uncertainties.
Figure \ref{fig2} show estimates of the half light radius, $R_{1/2}$, for the eight individual objects in our sample with black hole masses estimates, $M_{BH}$, from Peng et al. (2006) or Assef et al. (2011).
We also show the X-ray size estimates for HE0435$-$1223 (Blackburne et al. 2014), HE1104$-$1805 (Blackburne et al. 2015),
RXJ~1131$-$2131 (Dai et al. 2010), Q2237+0305 (Mosquera et al. 2013), QJ0158$-$4325 (Morgan et al. 2012), SDSS0924+0219 (McLeod \& Morgan 2014) and PG1115+080 (Morgan et al. 2008).
In spite of the large uncertainties (several objects have only upper size limits), there is a clear increase in the size with the mass of the black-hole, and very good agreement with previous estimates from other authors. A fit of our individual size estimates to a
power law, $\log(R_{1/2}/{\rm cm})=\log(R_9/{\rm cm})+x\log(M_{BH}/10^9M_\sun)$, gives $\log(R_9/{\rm cm})=15.6\pm0.3$, and an exponent $x=1.2\pm 0.5$. The fit can be slightly improved if we include previous size estimates from the literature. In this case, the resulting
paramaters are $\log(R_9/{\rm cm})=15.6\pm0.2$, and an exponent $x=1.1\pm 0.3$. This fit is shown in Figure \ref{fig2} as a continuous line. The exponent is very close to unity, as found by Mosquera et al. (2013), albeit with a large error.
In units of the gravitational radius $r_g=M_{BH}G/c^2$, the size of the X-ray emitting region of our sample is reasonably well fit by a line with $R_{1/2}=24r_g$, as also shown in Figure \ref{fig2}.
Taking into account this trend of the X-ray source size with black hole mass, we can now recalculate the joint probability distribution of sizes, but this time we scale the size of each individual object with the mass of its black hole as $R_{1/2}=R_9(M_{BH}/10^9 M_\sun)$. In this case, we will not need to exclude RXJ~1131$-$2131, as its contribution to the joint probability distribution is properly scaled to take account of the low mass of its black hole. In Figure 3 we show the probability distribution for $R_9=R_{1/2}/(M_{BH}/10^9 M_\sun)$. Despite including objects with a wide range in black hole masses, this likelihood is more peaked than Fig 1, which is an indication that scaling the size linearly with the black hole mass has reduced the scatter.
The maximum likelihood result for the scaled size is $\log(R_9/{\rm cm})=15.5^{+0.2}_{-0.3}$
, in very good agreement with the results of Mosquera et al. (2013) (their Fig. 4) and with the results of the linear scaling shown in Figure \ref{fig2}. Figure \ref{fig3} also shows
that the size of the X-ray emitting region is restricted to a rather narrow region of $R_{1/2}=(24\pm12)r_g$ in units of the gravitational radius $r_g$. Our present estimate, based on completely independent method and dataset, is in excellent agreement with previous estimates from microlensing (e.g. Mosquera et al., 2013, Blackburne et al., 2014, 2015, Morgan et al., 2012, Dai et al., 2010 ), but also in
good agreement with size estimates by other means such as absorption variability (cf. Agis-Gonz\'alez et al., 2014, Sanfrutos et al., 2013, Uttley et al., 2014). Estimates of the height of the corona above the accretion disk from reverberation lags are also in the range of a few $r_g$ (Reis \& Miller 2013, Emmanoulopoulos 2014, Cackett et al. 2014, Shappee et al. 2014)
\section{Joint Determination of the Stellar Mass Fraction and the X-ray Size}\label{sec3}\
Next, we allow the stellar mass fraction to vary as well. We simply repeat the calculations, but now include
a logarithmic grid for the stellar mass fraction $\alpha$ such that
$\alpha_j=0.025\times 2^{j/2}$ with $j=0,\cdots,10$, so that $\alpha$ ranges from 0.025 to 0.8.
We again linearly scale the source size $R_{1/2}$ with the black hole mass.
For every image pair, a likelihood is calculated for each of the 176 possible combinations of parameters $(\alpha, R_9)$
to compute
\begin{equation}
L(\alpha,R_9| \Delta m_{ir})=p(\Delta m_{ir}|\alpha,{R_9}).
\end{equation}
From these we can calculate likelihoods for each individual object $L_k(\alpha,R_9)$, and a joint likelihood as
the product of the eight individual likelihoods
\begin{equation}L(\alpha,R_9)=\prod_{k=1,8} L_k(\alpha,R_9). \end{equation}
The resulting 2D likelihood function is shown in Figure \ref{fig4}.
\begin{figure}
\epsscale{0.85}
\plotone{Fig4.eps}
\caption{Likelihood function for the stellar mass fraction $\alpha$ and the X-ray source size $R_9$ at $M_{BH}=10^9 M_\sun$ for microlenses of mass $M=0.3M_\sun$. The contours are drawn at likelihood intervals of 0.25$\sigma$ for one parameter from the maximum. The contours at 1$\sigma$ and 2$\sigma$ are heavier. Vertical lines indicate the estimate (solid line) and 68\% confidence interval (dashed lines) from a similar analysis of optical microlensing data by Jim\'enez-Vicente et al. (2015). \label{fig4}}
\end{figure}
The joint likelihood distribution shows much less covariance between $\alpha$ and $R_{1/2}$ than found in the optical (cf. Jim\'enez-Vicente et al. 2015).
In fact, $\alpha$ is very well constrained with little dependence on size, as expected for small sizes.
The maximum likelihood estimate is $\alpha=0.20\pm0.05$ (for an average distance of $\sim 1.9R_e$), which is in
excellent agreement with the optical result, as shown in Figure \ref{fig4}.
This agreement strengthens our confidence in the robustness of the method.
The maximum likelihood estimate for the average size of the emitting region in X-rays is $\log(R_9/{\rm cm})=15.5\pm0.2$,
(68\% confidence interval). This estimate is very similar to the obtained in the previous section in a single parameter analysis, which is not surprising, as we used a value of $\alpha=0.2$ which matches the best estimate of the new analysis. Again, the two parameter analysis confirms that the X-ray emitting region is roughly
$R_{1/2}=(24\pm12) r_g$ in size.
We also recomputed the results sequentially dropping each lens and found that the results are not dominated by any single system.
Taking into account the results of Jim\'enez-Vicente et al. (2015) for the average size of accretion disks in the optical (at an
average rest wavelength of 1736 \AA) of $R_{1/2}^{opt}=7.9^{+3.8}_{-2.6}$ light days, the typical ratio of the half light radius in the optical and X-rays is $R_{1/2}^{opt}/R_{1/2}^{X-ray}\sim 8$, which is in reasonable good agreement with previous results from studies of individual lenses (Morgan et al. 2008, Chartas et al. 2009, Dai et al. 2010, Morgan et al. 2012, Blackburne et al., 2015, 2014, Mosquera et al. 2013, MacLeod \& Morgan 2014). As pointed out by Mosquera et al. (2013), if $R_{1/2}^{X-ray}\propto M_{BH}$ and $R_{1/2}^{opt}\propto M_{BH}^{2/3}$ (cf. Morgan et al. 2010), then $R_{1/2}^{opt}/R_{1/2}^{X-ray}\propto M_{BH}^{-1/3}$, and this ratio should be
larger for smaller masses. Our ratio of $R_{1/2}^{opt}/R_{1/2}^{X-ray}\sim 8$ is a typical value for a mass of $M_{BH}\sim 10^9M_\sun$, which is roughly the average black hole mass in our sample.
\section{The Stellar/Dark Matter Surface Mass Density Profile}
\begin{figure}
\epsscale{0.85}
\plotone{Fig5.eps}
\caption{Probability distributions for the stellar mass fraction $\alpha$ at two different radii. The continuous (dashed) line is the likelihood function for the inner (outer) radius bin.
\label{fig5}}
\end{figure}
\begin{figure}
\epsscale{0.85}
\plotone{Fig6.eps}
\caption{Radial profile for the stellar mass fraction.
The star (triangle) use only the X-ray (optical) data and a single radial bin for all objects.
Circles (diamonds) are the estimates for the two radial bins using a logarithmic (linear) prior on size and both the X-ray and optical data.
The dashed line corresponds to a simple model with a de Vaucouleurs stellar component and a total mass corresponding to a SIS with a flat rotation curve equal to the maximum rotational velocity of the stellar component. The thick line is the fiducial galaxy model from Schechter et al. (2014), which is a rescaled version of the previous model. The grey band is the best fit profile for the sample of lenses analyzed by Oguri et al. (2014). The open diamonds and squares correspond to a model using a Hernquist component for the stars, embedded in an NFW halo with (open squares) and without (open diamonds) adiabatic contraction of the dark matter, also from Oguri et al. (2014) \label{fig6}}
\end{figure}
Microlensing measurements of the stellar mass fraction (including remnants) have the advantages of being local (not integrated within a certain radius) and insensitive to the stellar IMF.
On the other hand, the estimates can only be made at the location of the multiple images, which do not sample a broad range of radii (particularly in quadruply imaged systems).
Moreover, single epoch microlensing provides individual estimates with large uncertainties, making necessary the combination of the estimates from a large enough sample of systems to reduce the uncertainties.
Here, we combine the X-ray and optical microlensing estimates to examine $\alpha$ (cf. Table \ref{tab1}). With this enlarged sample, we split the data into two radial bins and derive independent estimates for the stellar mass fraction at two different radii $R_E/R_{e}$, where $R_E$ is the Einstein
radius and $R_{e}$ is the effective radius of the lens galaxy. For the X-ray sample we have taken the ratios from Schechter et al. (2014), and for the optical sample, we used the estimates from Jim\'enez-Vicente et al. (2015) for the 18 pairs in the 13 lenses with available estimates of $R_E/R_{e}$. This gives us a total of 48 independent microlensing estimates for 18 different lens systems (there are five lenses in common to both samples). The radial distribution of the observed pairs is bimodal in $R_E/R_{e}$ with a minimum around 1.7, so we split the sample into two radial bins containing objects with $R_E/R_{e}$ smaller and larger than 1.7 respectively. This results in
average distances of $(1.3\pm0.3) R_{e}$ and $(2.4\pm0.4) R_{e}$ for these two bins, where we are giving the dispersion about the mean
of each bin, not the uncertainty in the mean. We combine the marginalized (using a logarithmic prior on the size $r_s$) probability distributions for all the pairs in each of the two bins to produce a joint probability
distribution for $\alpha$ at these two radii. The result of this procedure is shown in Figure \ref{fig5}.
This is the first direct detection of a significant radial variation of the stellar mass fraction using microlensing measurements.
The resulting Bayesian estimates (using a logarithmic prior in $\alpha$) for the stellar mass fraction at these radii are $\alpha=0.31\pm0.15$ and $\alpha=0.14\pm0.05$.
We have also repeated the calculations using a linear prior on size. In this case, the estimates are $\alpha=0.40\pm0.18$ and $\alpha=0.18\pm0.05$ for the inner and outer radial bins respectively.
In Figure \ref{fig6} we also compare these new estimates to our earlier optical microlensing results (Jim\'enez-Vicente et al. 2015), a simple theoretical model and estimates for a galaxy sample based on strong lensing models (Oguri et al. 2014). The simple theoretical model is an early-type galaxy consisting of a de Vaucouleurs component for the stars and a singular isothermal sphere (SIS) for the total mass. Two different scalings of this model are shown. The dashed line is scaled so that the (flat) rotation curve of the SIS equals the maximum rotational velocity of the de Vaucouleurs stellar system. The continuous thick line based on the mass fundamental plane scaling from Schechter et al. (2014), which they take as their fiducial galaxy. Despite the relatively large errors, our results are in good agreement with simple theory and previous results. In particular, there is very good agreement with both the strong lensing results by Oguri et al (2014) and the fiducial galaxy based on Schechter et al. (2014).
\section{Conclusions}
We have performed a statistical analysis of the effect of source size and the fraction of surface mass densisty in stars
on the microlensing in X-rays for a sample of 10 lensed quasars taken from Schechter et al. (2014).
\begin{enumerate}
\item Pre-existing studies of X-ray microlensing have found that sizes increase roughly in proportion to the estimated black hole
mass (Mosquera et al. 2013). From a fit of our individual size estimates to a power law $\log(R_{1/2}/{\rm cm})=\log(R_9/{\rm cm})+x\log(M_{BH}/M_\sun)$, we find $\log(R_9/{\rm cm})=15.6\pm0.3$, and $x=1.2\pm 0.5$ (improving slightly when previous size estimates are included to $\log(R_9/{\rm cm})=15.6\pm0.2$, and $x=1.1\pm 0.3$).
Based on this, we assumed a linear scaling ($x=1$) of size $R_{1/2}$ with black hole mass $M_{BH}$, and we find an average size
for $M_{BH}=10^9M_\sun$ of $\log(R_9/{\rm cm})=15.5^{+0.2}_{-0.3}$.
This result is rougly consistent with the simple scaling $R_{1/2}\simeq(24\pm12) r_g$ (with $r_g=GM_{BH}/c^2$). This agrees well with previous determinations
for individual sources (e.g. Morgan et al. 2008, Blackburne et al. 2015, Mosquera et al. 2013).
\item Microlensing in X-rays produces an estimate for the local surface density in stars and stellar remnants of $\alpha=0.20\pm0.05$ at a typical radius of 1.9 effective radii, in excellent agreement with the independent result obtained using microlensing in the optical by Jim\'enez-Vicente et al. (2015).
\item By combining the microlensing estimates in X-rays and in the optical, we have been able to obtain the stellar mass fraction at two different radii. We find a drop in the stellar mass fraction from $\alpha=0.31\pm0.15$ at a radius of $(1.3\pm0.3)R_{e}$ to $\alpha=0.13\pm0.05$ at $(2.3\pm0.3)R_{e}$. This result is in very good agreement, given the uncertainties, with results from strong lensing analysis of a large sample by Oguri et al. (2014) and with the scaling of the mass fundamental plane found by Schechter et al. (2014).
\end{enumerate}
The application of the present method to a significantly larger sample of lens systems with measured microlensing (preferably, but not necessarily, in X-rays) should allow the determination of the shape of the dark matter density profile relative to the stars in the radial range from roughly 0.5 and 3 effective radii.
\
\noindent Acknowledgements:
The authors would like to thank M. Oguri for kindly providing the differential version of their results for comparison with the present work shown in Figure \ref{fig6}.
This research was supported by the Spanish Mi\-nis\-te\-rio de Educaci\'{o}n y Ciencia with the grants AYA2011-24728, AYA2010-21741-C03-01 and AYA2010-21741-C03-02. JJV is also supported by the Junta de Andaluc\'{\i}a
through the FQM-108 project. JAM is also supported by the Generalitat Valenciana with the grant PROMETEOII/2014/060.
CSK is supported by NSF grant AST-1009756.
|
\section{Introduction}
Filled skutterudites {\fs} ($R =$ rare-earth; $M =$ Fe, Ru, Os, etc.; $X =$ P, As, and Sb), which crystallize with the LaFe$_4$P$_{12}$-type structure (Space group: $Im\bar{3}$, $Z = 2$)~\cite{Jeit77}, exhibit a variety of physical properties, including superconductivity, semiconductivity, ferromagnetism and antiferromagnetism, depending on the combination of $R$, $M$ and $X$~\cite{Sale03,Sato09}.
The highest superconducting transition temperature {\tc} among {\fs} was 10.3~K for LaRu$_4$As$_{12}$~\cite{Shir97} before the discovery of superconductivity with $\tc = 17$~K for {\RhP}, which was synthesized using the high-pressure technique reported by Shirotani and colleagues in 2005~\cite{Shir05, Take07, Imai07, Imai07b}. The relatively high {\tc} in the latter filled skutterudite has encouraged materials scientists to search for new superconductors with similar compositions.
One feature of {\RhP} is that it includes a cobalt group element, $M =$ Rh; most of the existing filled skutterudites involve iron group elements.
If filled skutterudites with cobalt group elements are synthesized under ambient pressure, then the site occupancy of the $R$ atoms will become rather low, as observed for La$_{0.2}$Co$_4$P$_{12}$~\cite{Zemn86} and La$_{0.05}$Rh$_4$Sb$_{12}$~\cite{Zeng00}.
Thus, the high site occupancy of {\RhP} can be attributed to high-pressure synthesis, and is believed to be essential for superconductivity.
A new candidate superconductor is {\Sbx}, because {\As} synthesized at high pressure was reported not to be a superconductor, but rather a semiconductor with a narrow energy gap {\eg} of 0.03~eV~\cite{Arii09,Arii10}.
Therefore, with the aim of finding a new superconductor, {\Sbx} filled skutterudites were synthesized utilizing high-pressure techniques.
The synthesis conditions were optimized to reduce the impurity phases and obtain a sample with a high La-site occupancy.
Electrical resistivity and magnetic susceptibility measurements were then performed to verify whether {\Sbx} is a superconductor.
\section{Materials and methods}
To determine the optimal conditions for the solid-state synthesis of polycrystalline {\Sbx}, samples were synthesized using starting materials with various atomic ratios at different temperatures and pressures. The purity of the starting materials was 99.9\% for La, 99.9\% for Rh and 99.999\% for Sb.
Rh$_{3.7}$Sb$_{12}$, which consists of RhSb$_3$ and Sb, was prepared in advance by solid-state reaction at ambient pressure. These materials were weighed with molar ratios of La:Rh:Sb = 1:4:12 and 0.5:3.7:12, La:Rh$_{3.7}$Sb$_{12}$ = 0.5:1, and La:Sb = 1:3 in an argon-gas-filled glove box, and ground for 3 min using a vibrating mill. The following procedures were different between the high-pressure and ambient pressure syntheses. For high-pressure synthesis, the resulting powder was pressed into a pellet, placed in a hexagonal boron nitride capsule, sealed in a gold capsule under an argon gas atmosphere, and heat-treated at 1073 or 1173~K under pressures of 5.5 or 7.7~GPa for 2~h using a belt-type high-pressure apparatus~\cite{Yama92}. For ambient pressure synthesis, a pellet of the milled materials was placed in an alumina crucible, sealed in a quartz tube under an argon gas atmosphere, and heat-treated at 1073~K for 100~h in an electric furnace. The samples were characterized using powder X-ray diffraction (XRD; RINT TTR-III, Rigaku) with Cu K$\alpha$ radiation (40~kV/150~mA). A one-dimensional position-sensitive Si detector was used. The chemical composition was determined from wavelength-dispersive X-ray spectroscopy (WDS) measurements using an electron probe microanalyzer (JXA-8500F, Jeol) and by energy-dispersive X-ray spectroscopy (EDS) with a scanning electron microscope (SU70, Hitachi High-Technologies). For physical property measurements, the sample with the least impurity phases was selected. DC magnetization measurements were conducted using a commercial superconducting quantum interference device (SQUID) magnetometer (MPMS, Quantum Design) under an applied field of $H = 10000$~Oe from room temperature down to 2.0~K. Four-probe electrical resistivity measurements were performed using an in-house-built apparatus with a DC current density of 0.1~mA/mm$^2$. A $^4$He closed-cycle cryostat, of which the lowest temperature was 2.8~K, was used to cool the sample.
\section{Results}
\subsection{Optimization of synthesis conditions}
\begin{figure}
\begin{center}
\includegraphics[bb=0 0 288 403, width=0.8\hsize]{./fig1.eps}
\caption{\label{fig1} XRD patterns for the prepared samples. For samples (a)-(c), the starting materials and synthesis temperature were fixed at La:Rh:Sb = 1:4:12 and 1073~K, while the synthesis pressures were ambient pressure, 5.5 and 7.7~GPa, respectively. For sample (d), the synthesis conditions were La:Sb = 1:3, 1073~K and 7.7~GPa. For sample (e), the synthesis conditions were La:Rh:Sb = 1:4:12, 1173~K and 7.7~GPa. For samples (f)-(g), the synthesis temperature and pressure were fixed at 1073~K and 7.7~GPa, while the starting materials were La:Rh:Sb = 0.5:3.7:12 and La:Rh$_{3.7}$Sb$_{12}$ = 0.5:1, respectively. The solid bars represent the Bragg peak positions for RhSb$_3$, RhSb$_2$ and LaSb$_2$. For RhSb$_3$, the index numbers are noted beside the bars.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[bb=0 0 1202 1073, width=0.9\hsize]{./fig2.eps}
\caption{\label{fig2} BEC images of samples prepared at 7.7~GPa with a starting material ratio of La:Rh:Sb = 1:4:12 at (a) 1073~K and (b) 1173~K, and at 7.7~GPa and 1073~K for (c) La:Rh:Sb = 0.5:3.7:12 and (d) La:Rh$_{3.7}$Sb$_{12}$ = 0.5:1.}
\end{center}
\end{figure}
The synthesis pressure was initially optimized for a synthesis temperature of 1073 K with a stoichiometric 1:4:12 molar mixture of La, Rh and P as starting materials. Figs.~\ref{fig1}(a), (b) and (c) show XRD patterns of the three samples synthesized at ambient pressure, and at 5.5 and 7.7~GPa, respectively. All the strong peaks are indexed to RhSb$_3$~\cite{Kjek74}, which indicates that RhSb$_3$ or {\Sbx} was formed. It should be noted that filled skutterudites have similar diffraction patterns to unfilled skutterudites because they both have the same space group of $Im\bar{3}$~\cite{Take07}. The $a$ lattice constants were estimated to be 9.231(1), 9.285(1) and 9.285(1) {\AA} for the samples synthesized at ambient pressure, 5.5 and 7.7~GPa, respectively. The $a$ lattice constant for the sample synthesized at ambient pressure is the same as that for RhSb$_3$(9.232~{\AA})~\cite{Kjek74}, while those for the samples synthesized at 5.5 and 7.7~GPa are significantly larger than that for RhSb$_3$.
The weak extra XRD peaks for the samples produced at ambient pressure and at 5.5 and 7.7~GPa are attributed to LaSb$_2$ and RhSb$_2$, respectively.
There was little difference between the samples synthesized at 5.5 and 7.7~GPa; therefore, 7.7~GPa was selected as the synthesis pressure for further experiments.
A backscattered electron composition (BEC) image of the sample prepared at 7.7~GPa is shown in Fig.~\ref{fig2}(a). The composition of the gray area was determined by WDS to be La$_{0.50}$Rh$_{3.72}$Sb$_{12.0}$.
It should be noted that the compositional formula for the filled skutterudite is described with the molar ratio of Sb being 12 in this article.
Both the composition and XRD spectrum indicate that the filled skutterudite was successfully synthesized.
The La site occupancy $x = 0.50$ was higher than that for {\Sbx} ($x = 0.05$) synthesized at ambient pressure~\cite{Zeng00}, which indicates that high pressure is effective for La insertion into unfilled skutterudites.
In addition, these results suggest that the $a$ lattice constant increases with $x$.
A large amount of impurity phases was also observed in Fig.~\ref{fig2}(a). The white, dark gray and black areas were assigned to LaSb$_3$, RhSb and a mixture of Rh and Rh$_3$Sb, respectively. The XRD pattern of the largest impurity phase, LaSb$_3$, has not been reported; therefore LaSb$_3$ was synthesized at 7.7~GPa and 1073~K with the starting material ratio La:Sb = 1:3, and was confirmed by BEC imaging to consist of single phase LaSb$_3$ (not shown). The XRD pattern for LaSb$_3$ is presented in Fig.~\ref{fig1}(d). The peaks are broad, which indicates that the structural correlation length is short. These characteristics made it difficult to detect LaSb$_3$ impurity using XRD for the {\Sbx} sample (Fig.~\ref{fig1}(c)), although a large amount of LaSb$_3$ was present.
The synthesis temperature was subsequently optimized to 7.7~GPa. Figs.~{\ref{fig1}(e) and {\ref{fig2}}(b) present an XRD pattern and a BEC image, respectively, for the sample synthesized at 1173~K using a starting material ratio of La:Rh:Sb = 1:4:12. Besides peaks for {\Sbx}, strong RhSb$_2$ peaks are observed in Fig.~{\ref{fig1}}(e). La$_{0.47-0.58}$Rh$_{3.77}$Sb$_{12.0}$, for which the composition was determined by WDS, occupies the small area in Fig.~\ref{fig2}(b), and impurities of LaSb$_3$, Rh, and Rh-Sb phases such as RhSb$_2$, RhSb, Rh$_2$Sb, Rh$_3$Sb occupy a large area. The volume fraction of impurities is larger for the sample prepared at the higher temperature of 1173~K than that at 1073~K, which indicates incongruent melting of {\Sbx}. For these reasons, 1073~K was selected as the optimal synthesis temperature. The appearance of a large amount of LaSb$_3$ impurity phase in both samples was attributed to an excess supply of La. Rh and the surrounding Rh-Sb phase were deduced to be non-equilibrium phases that appear due to the high melting temperature of Rh and the finite reaction time.
Finally, the starting material ratio was optimized by fixing the synthesis pressure and temperature at 7.7~GPa and 1073~K. Two starting material mixtures were attempted: a 0.5:3.7:12 molar mixture of La, Rh, and Sb and a 0.5:1 molar mixture of La and Rh$_{3.7}$Sb$_{12}$. Figs.~{\ref{fig1}}(f) and (g) show XRD patterns for the samples synthesized from these mixtures at 7.7~GPa and 1073~K. The peak intensity for the impurity phases is much smaller in the XRD pattern for the 0.5:3.7:12 molar ratio sample than that for the 1:4:12 molar ratio. BEC images of the sample synthesized using La:Rh:Sb = 0.5:3.7:12 and that using La:Rh$_{3.7}$Sb$_{12}$ = 0.5:1 are presented in Figs.~\ref{fig2}(c) and (d), respectively. The chemical compositions were determined from EDS measurements. The sample synthesized with the 0.5:3.7:12 molar ratio for La, Rh, and Sb consists of almost single phase La$_{0.5}$Rh$_{4.0}$Sb$_{12}$, although a small amount of the Au$_{0.14}$Sb$_{0.86}$ impurity phase is evident~\cite{Gies68, Iyo14}, which could be due to contamination from the gold cell. In the sample started from Rh$_{3.7}$Sb$_{12}$, a large amount of impurity phase was observed as shown in Fig.~{\ref{fig2}}(d), which is consistent with the XRD pattern in Fig.~{\ref{fig1}}(g). The La content $x$ in {\Sbx} fluctuates between 0.1 and 0.6, which indicates that the phase has not reached the equilibrium state within the heating time of 2~h. This may be attributed to the slow reaction between La and RhSb$_3$ due to the high stability of RhSb$_3$ compared with Rh and Sb.} As a result, the appropriate conditions for the synthesis of {\Sbx} were determined to be 1073~K and 7.7~GPa with a starting material ratio of La:Rh:Sb = 0.5:3.7:12.
The sample synthesized under the optimized conditions had a lattice constant of 9.284(2) {\AA} and the chemical composition was 3.9(3) wt\% La, 21.1(2) wt\% Rh, and 74.9(2) wt\% Sb, which corresponds to the chemical formula La$_{0.55(5)}$Rh$_{3.99(4)}$Sb$_{12.0(2)}$.
\subsection{Physical property measurements}
\begin{figure}
\begin{center}
\includegraphics[bb= 0 0 288 360, width=0.8\hsize]{./fig3.eps}
\caption{\label{physProp} (a) Temperature $T$ dependence of the magnetic susceptibility ${\chi}$ of {\Sb} from 2~K to room temperature. (b) $T$ dependence of the electrical resistivity $\rho$ for {\Sb} from 2.8~K to room temperature. The inset shows the inverse temperature dependence of the conductivity, $1/\rho$, where the solid line represents a fit with the exponential function above 270~K.}
\end{center}
\end{figure}
The sample prepared under the optimized conditions was used for physical property measurements. Fig.~{\ref{physProp}}(a) shows the temperature $T$ dependence of the magnetic susceptibility $\chi$ for {\Sb} from 2~K to room temperature, which indicates weak diamagnetism that is independent of $T$. The absence of perfect diamagnetism indicates that {\Sb} is not a superconductor. In addition, the lack of paramagnetism derived from free electrons implies that this is not a metal, but a semiconductor or insulator. The diamagnetism was approximately $-3.1{\times}10^{-7}$~emu/(gOe), which corresponds to $-6{\times} 10^2$~cm$^3$/mol for {\Sb}. This value is comparable to Larmor diamagnetism. The weak increase at low temperatures may be due to magnetic impurities.
Fig.~{\ref{physProp}}(b) shows the $T$ dependence of the electrical resistivity $\rho$ for {\Sb}.
At $T > 80$~K, $\rho$ decreases with increasing temperature, showing semiconducting behavior, which is consistent with the diamagnetic susceptibility.
In addition, the $\rho(T)$ curve has a plateau below 80~K, which may originate from extrinsic semiconductivity, with the carriers being supplied by impurities acting as dopants. This type of $\rho-T$ dependence has been observed in narrow-gap semiconductors such as SrSi$_2$~\cite{Imai07c} and FeSi~\cite{Pasc97}.
The conductivity $1/\rho$ could not be fitted above 100~K using either the exponential function $\exp(-{\eg}/{\kb}T)$ or the variable range hopping model function $\exp(-(T_{\rm 0}/T)^{1/(1+d)})$~\cite{Mott68}, where {\kb} is the Boltzmann constant, $T_{\rm 0}$ is a characteristic temperature, and $d$ is the dimension ($d = 1, 2$ and 3). Fitting with the exponential function above 270~K, as shown in the inset of Fig.~{\ref{physProp}}(b), gives an energy gap of ${\eg} = 0.08$~eV, which indicates that {\eg} is greater than 0.08~eV.
\section{Discussion}
The La site occupancy $x$ is approximately the same for La$_x$Rh$_4X_{12}$ synthesized at high pressures: {\RhP}~\cite{Shir05}, {\As}~\cite{Arii09}, and {\Sb}. These are classified into two groups according to their electronic properties; {\RhP} is a superconductor, while {\As} and {\Sb} are semiconductors with ${\eg} = 0.03$~eV~\cite{Arii10} and ${\eg} > 0.08$~eV, respectively.
The corresponding unfilled skutterudites are similarly classified; an unfilled skutterudite RhP$_3$, which corresponds to {\RhP}, has been experimentally determined to be a metal~\cite{Odil78,Saet09}. RhAs$_3$ and RhSb$_3$, which correspond to {\As} and {\Sb}, respectively, are semiconductors with ${\eg} > 0.85$~eV~\cite{Cail95} and ${\eg} = 0.8$~eV~\cite{Nola96,Cail96}, respectively.
Thus, {\As} and {\Sb} not only have similar electrical properties, but also exhibit similar changes in the electrical properties as a result of La filling. This suggests that La insertion into unfilled metallic skutterudites will be a future strategy for producing new superconductors.
The three compounds LaRu$_4X_{12}$ ($X =$ P, As, Sb) are all superconductors~\cite{Meis81,Shir97,Uchi99}, which implies that the superconductivity is insensitive to the difference of the pnictogen species with high La-site-occupancy.
Thus, if higher occupancy is realized by higher pressures, superconductivity may be induced in {\Sbx} and {\Asx}, as in {\RhP}. On the assumption that $x$ increases in proportion to the pressure, full occupancy with x = 1 in {\Sbx} would be realized at approximately 15~GPa, which is beyond the achievable pressure of the apparatus used in the present study.
\section{Summary}
A {\Sb} filled skutterudite with a high La concentration was successfully synthesized by utilizing a high-pressure technique. This adds a new example to the filled skutterudites that include cobalt group elements. In addition, the presence of the LaSb$_3$ binary phase at high pressure was confirmed. Electrical resistivity and magnetic susceptibility measurements revealed that {\Sb} is not a superconductor, but a semiconductor with an energy gap greater than 0.08~eV. The electrical properties are more similar to those of {\As} than those of {\RhP}.
\section{Acknowledgments}
The authors thank M. Nishio for experimental support and H. Kitaguchi for allowing use of the SQUID magnetometer. This work was supported by the Funding Program for World-Leading Innovative R\&D on Science and Technology (FIRST), Japan.
\bibliographystyle{model1-num-names}
|
\section{Introduction}
The random restriction method and the polynomial method are powerful tools in computational complexity \cite{Jun12}; the former is applied to make a function become easier subfunctions (such as the classical switching lemma \cite{Has86}) and the polynomials are good computation or approximation models of Boolean functions \cite{Am06,BL14,Raz87}.
A subfunction of a given Boolean function is obtained by fixing some variables $0/1$ assignments (a.k.a. a restriction).
In is work we propose an idea of dealing with general Boolean functions by finding a well-studied while non-trivial substructure (for instance, symmetric or monotone subfunctions). To do this, we need the subfunctions maintain some property after fixing some variables.
In the literature (such as \cite{Has86,Jun12,Tal14}), arguments via random or adaptive restrictions on a Boolean function are targeted to prove a very low probability of sustaining hard subfunctions after some restrictions.
That is, a Boolean function is very likely to degenerate after restrictions.
For example, if any variable is assigned as $1$ then the $\mathsf{OR}_n$ function becomes a constant. It is not clear that for arbitrary Boolean function depending on all variables, does there exist a subfunction that depends on the rest variables (the so-called no-junta subfunction)? This is important as most of complexity measures are conditioned on the number of influential variables. We give a positive answer in this work.
As an application of the existence of no-junta subfunction, we consider the degree lower bound of general Boolean functions. The degree of a polynomial exactly representing a boolean function relates to many classical complexity measures \cite{Jun12}, including
the decision tree complexity, the circuit complexity \cite{Raz87} and the quantum complexity \cite{Am06}, etc. It is easy to acquire the representing degree for a specific function since the polynomial can be constructed by the interpolation based on the truth table. However, this cannot give a nontrivial degree lower bound for all functions. Nisan and Szegedy \cite{NS94} first proved a lower bound for real value representing polynomials. When the polynomial coefficients are in finite fields, a breakthrough was made by Gopalan et al. \cite{GLS10}.
In this work we give lower bounds with simpler forms and different proofs.
The rest of this paper is organized as follows. In Section 2, we introduce some notations and definitions. In Section 3 we prove any $n$-variate function has a subfunction depending on the rest of $n-1$ variables. Section 4 shows the degree bounds. Section 5 concludes the paper.
\section{Preliminaries}
Define $\mathcal{F}_n=\left\{ f:\{0,1\}^{n} \to \{0,1\} \right\}$, which is the collection of all
$n$-variable Boolean functions.
Let $[n]=\{1,\cdots,n\}$, $\bfv{x}=(x_1,\cdots,x_n)\in \{0,1\}^n$. Note that $\mathcal{F}_n$ contains juntas. A $k$-junta is a function $f(\bfv{x})$ represented as a formula of $n$ variables while in fact there exists a proper subset $K\subset [n]$ of size $k<n$ that decides the value of $f$, and all variables in $[n]\setminus K$ have no influence on the output of $f$. Denote $\mathcal{F}_{n}^*$ as the subset of $\mathcal{F}_n$ that excludes all juntas.
Hence, if $f\in \mathcal{F}_{n}^*$, then $f$ depends on all $n$ variables.
Consider an index subset $I\subseteq[n]$ and a partial assignment $\rho_{I}$ on the variables indexed by $I$. For a given $\rho_I$, we call a variable in $I$ the assigned variable and a variable in $I^c$ the free variable.
We define the subfunction $f(\bfv{x})|_{\rho_I}$ to be the function derived by restricting the variables according to $\rho_I$ and keeping the variables in $I^c$ free. Take the address function \cite{Jun12}
as an example: $\mathsf{Address}(x_1,x_2,x_3,x_4,x_5,x_6)$ outputs the value of $x_{z}$ where $z={x_1}\cdot 2^0+{x_2}\cdot 2^1+3$. Let $\rho_{\{1,6\}}$ be a restriction such that $x_1=1, x_6=0$ then $\mathsf{Address}(x_1,x_2,x_3,x_4,x_5,x_6)|_{\rho_{ \{1,6\} }}=x_4-x_2x_4$.
Note this subfunction is a 2-junta not depending on the free variables $x_3$ and $x_5$.
As mentioned, for an arbitrary $n$-variate function the existence of its no-junta subfunction was not clear.
Let $\mathbb{Z}_m$ be the ring of $\{0,\cdots,m-1\}$, where the integer $m\geq 2$.
Let $\mathbb{Z}_m[\bfv{x}]$ be the set of all polynomials over $\mathbb{Z}_m$, and similarly for $\mathbb{R}[\bfv{x}]$ and $\mathbb{Z}[\bfv{x}]$.
For $f \in \mathcal{F}_n$, let $\left\{ f(D):D\subseteq [n] \right\}$ be its truth table.
Then $f$ can be represented by $F(\bfv{x})=\sum_{D\subset [n]}{f(D)\Pi_{i\in D}x_i\Pi_{i\not\in D}(1-x_i)}\in \mathbb{Z}[\bfv{x}]$ (also in $\mathbb{R}[\bfv{x}]$). This implies the existence of the exact representation and the uniqueness is easy to prove \cite{GLS10,Jun12}.
Furthermore, since all coefficients in the expansion of $F(\bfv{x})$ are integers and $x_i$'s are in $\{0,1\}$, $F(\bfv{x})$ is a multilinear polynomial. Also observe that $F(\bfv{x})\pmod{m}\in\mathbb{Z}_m[\bfv{x}]$. Denote $F(\bfv{x})\pmod{m}$ by $P(\bfv{x})$. Obviously we have $\deg(F(\bfv{x}))\geq \deg(P(\bfv{x}))$. That is, for a given $f$, $d_{\mathbb{Z}}(f)\geq d_{m}(f)$.
For $f\in \mathcal{F}_{n}^*$,
the Nisan-Szegedy bound \cite{NS94} states that $d_{\mathbb{R}}(f)\geq \log{n}-O(\log{\log{n}})$.
Observe that $\mathsf{PARITY}(\bfv{x})$ over $\mathbb{Z}_2$ can be computed by $\sum_{i=1}^nx_i\pmod{2}$, i.e. $d_{2}(\mathsf{PARITY})=1$, which shows the difference between $d_{\mathbb{R}}(f)$ and $d_{m}(f)$.
Also note that Gopalan et al. \cite{GLS10} proved $ {\lceil \lg p\rceil}\cdot p^{2 d_{p}(f)} \cdot d_{p}(f)\cdot d_{q}(f) \geq n$.
Suppose the total degree of $P(\bfv{x})$ is $d$, and write $P(x_1,\cdots,x_n)=\sum_{D\subseteq [n];|D|\leq d}{c_D\Pi_{j\in D}x_j}\pmod{m}$, where $c_D\in \mathbb{Z}_m$ for all $D\subseteq [n]$.
Sometimes we treat $D\subseteq [n]$ as a variable, such as $f(D)$ or $P(D)$; by this we mean $f(D)=f(x_1,\cdots,x_n)$ where $x_i=1\Leftrightarrow i\in D$.
Denote $|\bfv{x}|=\sum_{i=1}^{n}x_i$ while $|D|$ is the cardinality of $D\subseteq [n]$. If $f$ is symmetric then the output value of $f$ is decided by $|\bfv{x}|$.
The symmetry makes all monomials of the same degree have the same coefficient, that is, for all $D$ and $D'$ with $|D|=|D'|=k$, we have $c_D=c_{D'}=c_k$. It is easy to prove
\begin{fact} \label{sympoly}
If $P$ is a symmetric polynomial of degree $d$, then for any $A\subseteq [n]$,
$$
P(A)=\sum_{k=0}^{d}{c_k}\binom{|A|}{k}
$$
\end{fact}
The following property is helpful to expose the relation between the binomial coefficients and the degree of a polynomial.
\begin{lemma}(See \cite{Tsai96}.)\label{cycle}
Let $m=\Pi_{i=1}^{r}{p_i^{e_i}}$ be a positive integer, where $p_i$'s are distinct primes and $e_i$'s are positive integers.
Let $L_k=\Pi_{i=1}^{r}{p_i^{e_i+\lfloor \log_{p_i}{k} \rfloor}}$. Then for any given nonnegative integers $s$, $j$ and $k$, we have
$$
\binom{sL_k+j}{k}\equiv \binom{j}{k}\pmod{m}.
$$
\end{lemma}
For a given symmetric polynomial of degree $d$, we will consider $k=0,1,\cdots,d$.
For saving space, we abbreviate $a\equiv b \pmod{m}$ as $a\equiv_{m} b$.
Observe that if $k\leq d$ then $L_k|L_d$. This means $\forall k\leq d$, $\binom{sL_d+j}{k}\equiv_{m} \binom{j}{k}$.
Besides, note that $\binom{j}{k}=0$ if $j<k$.
\section{Existence of No-Junta Subfunction}
Let
$f\in \mathcal{F}_n$
be a Boolean function on $n$ variables $x_1, \cdots, x_n$. We say $f$ depends
on the $i$-th variable $x_i$ if there is some input $a$ such that $f(a)\neq f(a^{(i)})$, where $a^{(i)}$
is obtained from $a$ by flipping the value of its $i$-th coordinate. The function $f$ is called \textit{nondegenerate} if $f$ depends on all its variables(, i.e. $f\in \mathcal{F}_n^{*}$).
The variable $x_i$ is said to be \textit{useful} for $f$ if $f$ depends on $x_i$; otherwise $x_i$ is \textit{useless} for $f$. We will denote the set of (indices of) useless variables for f by $U(f) := \{i : f \mbox{ does not depend on } x_i \}$.
Note that the notion of a useless variable presupposes (often implicitly) a universe of variables
on which f is defined.
For $i\in [n]$ and $b\in\{0, 1\}$,
the restriction $ f|_{x_i =b} : \{0, 1\}^{[n]\setminus\{i\}}\to\{0, 1\}$
is defined as the subfunction on variables $x_j$, $j\neq i$, obtained by fixing $x_i = b$. For notational convenience, we
denote $f|_{x_i =b}$ by $f_{ib}$.
Our main theorem is
\begin{theorem} \label{no_junta}
If $f\in \mathcal{F}_n^{*}$ then there exists an $i\in [n]$ such that at least one of $f_{i0}$ or $f_{i1}$ is nondegenerate, i.e., it depends on all the variables $[n]\setminus \{i\}$.
\end{theorem}
We start with an obvious observation:
\begin{prop} \label{juntabasic}
(i) If $x_i$ is useless for $f$, it is useless for any restriction of $f$. In particular,
$i\in U(f)$, then for all $j\in [n]\setminus\{i\}$, $i\in U(f_{j0})\cap U(f_{j1})$.\\
(ii) On the other hand, if $i$ is useless for both $f_{j0}$ and $f_{j1}$ for some $i$ and $j$, then $i$ is useless
for $f$ as well. In notation, $\exists j$ such that $i\in U(f_{j0})\cap U(f_{j1})\Rightarrow i\in U(f)$.
\end{prop}
\begin{proof}
For $i\neq j\in [n]$, the conclusions are clear by observing $f=(1-x_j)\cdot f_{j0}+x_j\cdot f_{j1}$.
\end{proof}
Given a boolean function $f\in \mathcal{F}_n$, we construct a digraph $G_f=(V,E)$ with $V=[n]$ and a directed edge ${i}\overset{b}{\longrightarrow} j$ labeled by $b$ if and only if $j\in U(f_{ib})$, i.e.
$E=\{i\overset{b}{\longrightarrow} j: i\in [n], j\in U(f_{ib}),b=0\mbox{ or }1\}$. The digraph $G_f$ has the following properties:
\begin{prop} \label{trans_cycle}
(\textbf{Transitivity}:)
Suppose in $G_f$ we have a path $i\overset{b_i}{\longrightarrow} j \overset{b_j}{\longrightarrow}k$ for distinct $i,j,k\in [n]$. Then the edge $i\overset{b_i}{\longrightarrow} k$ is also in $G_f$, i.e., $k\in U(f_{ib_i})$. More generally, if a path of distinct variables $i\overset{b_i}{\longrightarrow}j_1\overset{b_{j_1}}{\longrightarrow}\cdots \overset{b_{j_{t-1}}}{\longrightarrow} j_t$ exists in $G_f$, then $\{{j_1},\cdots,{j_t}\}\subseteq U(f|_{x_i=b_0})$.\\
(\textbf{Cycles}:)
If there are two cycles through $i$ containing edges $i\overset{0}{\longrightarrow} j$ and $i\overset{1}{\longrightarrow} k$, all three variables $x_i, x_j, x_k$ are useless for $f$.
\end{prop}
Observe that if $G_f$ has a cycle, say ${5}\overset{0}{\longrightarrow} {2}\overset{1}{\longrightarrow} {4}\overset{0}{\longrightarrow}5$, then we can apply Transitivity to reduce it to a collection of cycles of two vertices, e.g., ${5}\overset{0}{\longrightarrow}{4}\overset{0}{\longrightarrow}5$, etc.
\begin{proof} (Transitivity:)
Note that $i,j,k$ are distinct, otherwise it is in fact a cycle.
In the path, by definition of $j \overset{b_j}{\longrightarrow}k$ we have $f|_{x_j=b_j,x_k=0}=f|_{x_j=b_j,x_k=1}$.
Observe that
$$\aligned
& f|_{x_{j}=b_j,x_{k}=0}= f|_{x_{j}=b_j,x_{k}=1} \\
& \Rightarrow
f|_{x_i=b_i,x_{j}=b_j,x_{k}=0}= f|_{x_i=b_i,x_{j}=b_j,x_{k}=1}
\endaligned
$$
Besides, $i \overset{b_i}{\longrightarrow}j$ implies $f|_{x_i=b_i,x_j=b_j}=f|_{x_i=b_i,x_j=1-b_j}$.
Hence,
$$\aligned
& f|_{x_i=b_i,x_j=b_j,x_k=0}= f|_{x_i=b_i,x_j=1-b_j,x_k=0} \\
& =
f|_{x_i=b_i,x_j=b_j,x_k=1}= f|_{x_i=b_i,x_j=1-b_j,x_k=1}
\endaligned
$$
That is, $k\in U(f_{i b_i})$.
For the general case of $i\overset{b_i}{\longrightarrow}j_1\overset{b_{j_1}}{\longrightarrow}\cdots \overset{b_{j_{t-1}}}{\longrightarrow} j_t$, it is easy to prove by induction on $t$ with almost the same argument as above.
(Cycles:)
First observe that for a given cycle $i\overset{b_i}{\longrightarrow}j\overset{b_j}{\longrightarrow}i $, by definition we have $f|_{x_i=b_i,x_j=b_j}=f|_{x_i=b_i,x_j=1-b_j}$ because of $i\overset{b_i}{\longrightarrow}j$, and $f|_{x_i=b_i,x_j=b_j}=f|_{x_i=1-b_i,x_j=b_j}$ because of $j\overset{b_j}{\longrightarrow}i$. Combining these two equations, it is clear to have $f|_{x_i=b_i,x_j=b_j}=f|_{x_i=b_i,x_j=1-b_j}=f|_{x_i=1-b_i,x_j=b_j}$. Note that there is no requirement on $f|_{x_i=1-b_i,x_j=1-b_j}$.
Now, for the given pair of cycles $i\overset{0}{\longrightarrow}j\overset{b_j}{\longrightarrow}i$ and $i\overset{1}{\longrightarrow}k\overset{b_k}{\longrightarrow}i$ in $G_f$, we can list the corresponding equations:
$$
\left\{
\aligned
& f|_{x_i=0,x_j=b_j,x_k=b_k}=f|_{x_i=0,x_j=1-b_j,x_k=b_k}=f|_{x_i=1,x_j=b_j,x_k=b_k} \\
& f|_{x_i=0,x_j=b_j,x_k=1-b_k}=f|_{x_i=0,x_j=1-b_j,x_k=1-b_k}=f|_{x_i=1,x_j=b_j,x_k=1-b_k} \\
& f|_{x_i=1,x_j=b_j,x_k=b_k}=f|_{x_i=1,x_j=b_j,x_k=1-b_k}=f|_{x_i=0,x_j=b_j,x_k=b_k} \\
& f|_{x_i=1,x_j=1-b_j,x_k=b_k}=f|_{x_i=1,x_j=1-b_j,x_k=1-b_k}=f|_{x_i=0,x_j=1-b_j,x_k=b_k} \\
\endaligned
\right.
$$
Note that the above equations are related by $f|_{x_i=0,x_j=b_j,x_k=b_k}$ in the first and third equations, $f|_{x_i=0,x_j=1-b_j,x_k=b_k}$ in the first and fourth, and $f|_{x_i=1,x_j=b_j,x_k=1-b_k}$ in the second and third.
Besides, they contain all eight configurations.
This shows that $f$ is independent of $x_i, x_j$ and $x_k$.
I.e., $f$ is not sensitive on variables $x_i,x_j$ and $x_k$.
\end{proof}
\begin{definition}
For $f\in \mathcal{F}_n$ and $S\subseteq [n]$, $U1(f,S):=\bigcup_{i\in S}(U(f_{i0})\cup U(f_{i1}))$. In words, a variable $j$ is in $U1(f,S)$ if it is useless for a restriction of $f$ that fixes one variable to some value.
\end{definition}
\begin{prop} \label{closure}
(\textbf{Closure Property}:) For any $i\in[n]$ and $b\in \{0,1\}$, we have
$$
U1(f,U(f_{ib}))\subseteq U(f_{ib})\cup \{i\}.
$$
Furthermore, if for every $k\in U(f_{ib})$ and every $b'\in \{0,1\}$, $i\not\in U(f_{kb'})$, then
$$
U1(f,U(f_{ib}))\subseteq U(f_{ib}).
$$
\end{prop}
\begin{proof}
For any $j\in U(f_{ib_i})$ and $j'\in U(f_{jb_j})\setminus \{i\}$ (where $b_i,b_j\in\{0,1\}$) we can append ${j}\overset{b_j}{\longrightarrow} j'$ to $i\overset{b_i}{\longrightarrow} j$ and obtain a path.
If for some $j\in U(f_{ib_i})$, $i\in U(f_{j0})\cup U(f_{j1})$, then we have $U(f_{j0})\cup U(f_{j1})\subseteq U(f_{ib_i})\cup \{i\}$. Hence,
$$
U1(f;U(f_{ib_i}))
=\bigcup_{j\in U(f_{ib_i}),b=0,1}{U(f_{jb})}
\subseteq U(f_{ib_0})\cup \{i\}.
$$
On the other hand, if $\forall j\in U(f_{ib_i}), i\not\in U(f|_{j0})\cup U(f|_{j1})$, then $U(f_{j0})\cup U(f_{j1})\subseteq U(f_{ib_i})$. This concludes that
$U1(f,U(f_{ib}))\subseteq U(f_{ib})$.
\end{proof}
\begin{proof} (\textbf{of Theorem \ref{no_junta}})
We prove the theorem by contraction. So, assume $f$ is nondegenerate but every restriction $f_{ib}$, $i\in [n]$ is degenerate.
Then we have $U(f|_{x_i=0}) \neq \emptyset$ and $U(f|_{x_i=1})\neq \emptyset$ for all $i\in [n]$.
Thus, we can assume that $f$ satisfies the following property:\\
$$
(*)\mbox{ } \forall i\in [n], U(f|_{x_i=0}) \neq \emptyset \mbox{ and } U(f|_{x_i=1})\neq \emptyset.
$$
We prove the claim below:\\
\textbf{Main Claim:} {If $f$ has property $(*)$ and $U(f;S)\subseteq S$ for some $S\subseteq [n]$ with $|S|\geq 2$, then $\exists j\in S$ such that $f|_{j0}= f|_{j1}$, i.e., $j$ is useless for $f$.}
Since $U1(f,[n])\subseteq [n]$ and $n\geq 2$, the claim can be applied with $S=[n]$ and it implies $f$ is degenerate. This is a contraction.
\end{proof}
\begin{proof} (\textbf{of Main Claim}) We prove this by induction on $|S|$.
Base case ($|S|=2$):
Note that $|S|=2$ with $U(f;S)\subseteq S$ implies $U(f|_{i0})=\{j\}$ and $U(f|_{i1})=\{j\}$. Since $j\in U(f_{i0})\cap U(f_{i1})$, Proposition \ref{juntabasic}(ii) shows $j\in S$ is useless for $f$.
Inductive step:
Assume $|S|=s\geq 3$ and the claim holds for all $S$ with $|S|\leq s-1$.
Consider an arbitrary vertex $j^{*}\in S$. By $(*)$ there is at least one $0$-edge $j^{*}\overset{0}{\longrightarrow} u$ and at least one $1$-edge $j^{*}\overset{1}{\longrightarrow} v$. As shown in Figure \ref{case1}, if $u=v$ then $u\in U(f_{j^{*}0})\cap U(f_{j^{*}1})$ and by Proposition \ref{juntabasic}(ii), $u$ is useless and we are done. So, we may assume $u\neq v$. If there is a directed path from $u$ back to $j^{*}$ and a directed path from $v$ back to $j^{*}$ (as shown in Figure \ref{case2}), then $j^{*},u,v$ satisfy the Cycles property of Proposition \ref{trans_cycle} and all of them are useless for $f$ and we are again done.
Hence, we can assume that there is no cycle \textit{via}, w.l.o.g., any $1$-edge leaving $j^{*}$ back to $j^{*}$. This case is shown in Figure \ref{case3}, where we use the dashed edge to emphasize $G_f$ does not have such edge. Let $S':=U(f_{j^{*}1})$. By the foregoing assumption, no $k\in S'$ can have an edge going back to $j^{*}$ and hence $j^{*}\not\in S'$. By the Closure property (Proposition \ref{closure}, the "furthermore" part), we have $U1(f,S')\subseteq S'$. We also have that $|S'|\leq s-1$. Moreover, $|S'|\geq 2$ since there is at lest one $1$-edge $j^{*}\overset{1}{\longrightarrow} v$ and by $(*)$, there are two outgoing edges out of $v$ neither of which can go back to $j^{*}$ (the heads of those two edges could coincide $--$ however, that'd immediately imply the claim). Hence we can apply the induction hypothesis to $S'$ and conclude there is a vertex $w\in S'$ that is useless for $f$. Note that $S'\subset S$ by the Transitivity property since $j^{*}\in S$ and $U1(f,S)\subseteq S$. Hence $w\in S$ and is useless for $f$.
\begin{figure}
\centering
\subfigure[$u=v$]{
\begin{tikzpicture}[shorten >=1pt,node distance=2.8cm,on grid,auto]
\node[state] (j) {$j^*$};
\node[state] (u) [right=of j]{$u$};
\path[->]
(j) edge [bend left] node {$0$} (u)
edge [bend right] node {$1$} (u);
\end{tikzpicture}
\label{case1}}
\quad
\subfigure[$u\neq v$, cycles]{
\begin{tikzpicture}[shorten >=1pt,node distance=2.8cm,on grid,auto]
\node[state] (j) {$j^*$};
\node[state] (u) [above right =of j]{$u$};
\node[state] (v) [below right =of j]{$v$};
\path[->]
(j) edge [bend left] node {$0$} (u)
edge [bend right] node [swap] {$1$} (v)
(u) edge [bend left] node {$b_u$} (j)
(v) edge [bend right] node [swap] {$b_v$} (j) ;
\end{tikzpicture}
\label{case2}}
\quad
\subfigure[$u\neq v$, no back edge]{
\begin{tikzpicture}[shorten >=1pt,node distance=2.8cm,on grid,auto]
\node[state] (j) {$j^*$};
\node[state] (u) [right=of j]{$u$};
\node[state] (v) [below=of j]{$v$};
\node[state] (k) [right=of v]{$k$};
\path[->]
(j) edge [bend left] node {$0$} (u)
edge [bend right] node {$1$} (v)
(u) edge [bend left] node {$b_u$} (j)
(v) edge [bend right] node {$0,1$} (k);
\path[->, dashed]
(v) edge [bend right] node [swap]{$0,1$} (j);
\end{tikzpicture}
\label{case3}}
\caption{ Graphs for Inductive step}
\end{figure}
\end{proof}
Many complexity measures are hard to be obtained for general functions but easier for that with some property (such as monotone or symmetric). This suggests we can get some complexity bounds of $f$ through its subfunction $f|_{\rho_I}$.
However, as the example of the $\mathsf{Address}$ function illustrates, even $f\in \mathcal{F}_{n}^*$, usually
$f|_{\rho_I}\not\in \mathcal{F}_{n-|I|}^*$.
The implication of Theorem \ref{no_junta} promises that for given $0<k<n$ there is one restriction ${\rho_I}$ with $n-|I|=k$ such that $f|_{\rho_I}$ is no-junta.
An crucial observation is that any no-junta and symmmetric $f|_{\rho_I}\in \mathcal{F}_{n-|I|}^*$; this motivates us to derive a degree lower bound for non-symmetric Boolean functions from symmetric subfunctions.
\section{Degree Bounds}
\begin{theorem}
Let $f\in \mathcal{F}_{n}^*$ be symmetric, $p_i$'s be distinct primes, and $r$ and $e_i$'s be positive integers.
Let $m=\Pi_{i=1}^{r}p_i^{e_i}$. Then
$$
m\cdot d_{p_1^{e_1}}(f)\cdots d_{p_r^{e_r}}(f) > n.
$$
\end{theorem}
The idea of the proof is to find a particular subset $A$ such that the output values of representing polynomials on $A$ are distinct over two $\mathbb{Z}_{p_i^{e_i}}$'s, which is a contradiction; hence this $A$ must not exist.
\begin{proof}
We only have to consider $f\in \mathcal{F}_{n}^*$ satisfying $f(\emptyset)=0$ since $\forall i$, $d_{p_i^{e_i}}(f)=d_{p_i^{e_i}}(\neg f)$.
Besides, $f$ is no-junta and symmetric means there exists a weight $\tau\in [n]$ s.t. $\forall D\subseteq [n]$ with $|D|=\tau \leq d_{min}$, $f(D)=1$, where $d_{min}=\min_{i\in [r]}(d_{p_i^{e_i}}(f))$. W.L.O.G. let $d_{min}=d_1$.
Let $P_{i}(\bfv{x})\in \mathbb{Z}_{p_i^{e_i}}[\bfv{x}]$ s.t. $P_{i}(\bfv{x})\pmod{p_i^{e_i}} = f(\bfv{x})$ for all $\bfv{x}\in \{0,1\}^n$. Denote $d_i=\deg(P_i(\bfv{x}))$.
Since $f$ is symmetric, each $P_i(A)$ can be written as $\sum_{k=0}^{d_i}{c_{P_i,k}\binom{|A|}{k}}$ by Fact \ref{sympoly}.
Define $L_{i}=p_i^{e_i+\lfloor \log_{p_i}{d_i} \rfloor}$.
Then by Lemma \ref{cycle} for any nonnegative integers $s$, $j$ and $k$, we have
$$
\forall k\leq d_{min}, \binom{sL_{i}+j}{k} \equiv_{p_i^{e_i}} \binom{j}{k}.
$$
Consider $A\subseteq [n]$ such that
$|A|\equiv \tau \pmod{L_{1}}$ and for $i\geq 2$, $|A|\equiv 0 \pmod{L_{i}}$.
Let $L:=\Pi_{i=1}^r {L_i}$ and $L_i^{'}=L/L_i$.
Then by the CRT, the unique solution (in the sense of modular convergence) is
$$
\aligned
|A|=
& \sum_{i=2}^r 0\cdot L_i^{'}\cdot [{L_i^{'}}^{-1}\pmod{L_i}]\\
& +\tau\cdot L_1^{'}\cdot [{L_1^{'}}^{-1}\pmod{L_1}] \pmod{L}
\leq L
\endaligned
$$
For this $|A|$ and by Fact \ref{sympoly},
$$\aligned
P_1(A)&=\sum_{k=0}^{d_1}{c_{P_1,k}\binom{|A|}{k}}
\equiv_{p_1^{e_1}} \sum_{k=0}^{d_1}{c_{P_1,k}\binom{\tau}{k}}
=P_1(D_{\tau}
\equiv_{p_1^{e_1}}1\\
P_2(A)&=\sum_{k=0}^{d_2}{c_{P_2,k}\binom{|A|}{k}}
\equiv_{p_2^{e_2}} \sum_{k=0}^{d_2}{c_{P_2,k}\binom{0}{k}}
=P_2(\emptyset
\equiv_{p_2^{e_2}}0\\
\cdots \\
P_r(A)&=\sum_{k=0}^{d_r}{c_{P_r,k}\binom{|A|}{k}}
\equiv_{p_r^{e_r}} \sum_{k=0}^{d_r}{c_{P_r,k}\binom{0}{k}}
=P_r(\emptyset
\equiv_{p_r^{e_r}}0
\endaligned
$$
where $D_{\tau}$ is any subset of $[n]$ with size $\tau$.
This is a contradiction unless $|A|>n$. Hence we have
$n<|A|\leq L=\Pi_{i=1}^r {L_i}\leq m\cdot\Pi_{i=1}^r{d_i}$.
\end{proof}
Consider $r=2$ and $e_1=e_2=1$ in above Theorem, we have $d_{p_1}(f)d_{p_2}(f)=\Omega(n)$.
\begin{corollary}
Let $f\in \mathcal{F}_{n}^*$ be symmetric, then $f$ can has degree $o(\sqrt{n})$ in at most one finite field.
\end{corollary}
Note that for a fixed $I^c$ of size $|I^c|=k$, there are exactly $2^{n-k}$ partial assignments that leave the variables in $I^c$ free. Furthermore, for fixed $k$ there are ${\binom {n}{k}}2^{n-k}$ $k$-subcubes. We denote by $\mathcal{R}_{k,n}$ the set of restrictions that leave $k$ of $n$ variables free, i.e. $|\mathcal{R}_{k,n}|={\binom {n}{k}}2^{n-k}$.
For a fixed $\rho\in \mathcal{R}_{k,n}$, we choose $f\in \mathcal{F}_n$ uniformly at random.
Let $\mathbb{I}_{\rho}(f)$ be the indicator variable of the event that $f|_{\rho}$ is a non-junta symmetric subfunction on the corresponding $k$-subcube.
Therefore, if $\mathbb{I}_{\rho}(f)=1$ then $f|_{\rho}\in \mathcal{F}_{k}^*$. Observe that $\Pr_{f\in \mathcal{F}_n}\left[\mathbb{I}_{\rho}(f)=1 \right]=(2^{k+1}-2)\big/{2^{2^k} }$.
A natural question is how small can be $k$ to make sure the existence of a no-junta symmetric subfunction. We are going to apply the second moment method which in fact is a corollary of $\Pr[X=0]\le Var[X]/{E[X]^2}$.
\begin{lemma}(See \cite{AS08}.)\label{probmd}
Let $X=\sum_{i=1}^{M}X_i$, where $X_i$ is the indicator random variable for the event $A_i$. Denote $i\backsim j$ for the events $A_i$, $A_j$ that are not independent and define $\Delta=\sum_{i\backsim j}\Pr[A_i\wedge A_j]$. Thus $Var[X]\leq E[X]+\Delta$. If $E[X]\to \infty$ and $\Delta=o({E[X]}^2)$ then $X>0$ almost always.
\end{lemma}
Here we consider the above $X_i$ as $\mathbb{I}_{\rho_{i}}(f)$ and $A_i$ as the event that $f|_{\rho_i}$ is no-junta and symmetric. To acquire an upper bound of $\Delta$, we need the following fact.
\begin{fact}\label{dep}
For a fixed $k$-subcube $S$ in an $n$-cube, there are at most $\binom{n}{k}\cdot 2^k$ $k$-subcubes intersecting with $S$.
\end{fact}
\begin{proof}
For each vertex of the $n$-cube, it can only belong to at most $\binom{n}{k}$ different $k$-subcubes. A fixed $k$-subcubes contains $2^k$ vertices, so there are at most $\binom{n}{k}\cdot 2^k$ $k$-subcubes intersecting with the given $S$.
\end{proof}
\begin{lemma} \label{aalgn}
For large $n$, if $k\leq \lg{n}-1$ then
almost always each $f\in \mathcal{F}_n$ has at least one $k$-variable symmetric subfunction.
\end{lemma}
\begin{proof}
Define $\mathbb{I}(f)=\sum_{\rho\in \mathcal{R}_{k,n}}{\mathbb{I}_{\rho}(f)}$.
The goal is to show $E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\to \infty$ and $\Delta=o({E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]}^2)$. Then by Lemma \ref{probmd}, we have $\mathbb{I}(f)>0$ almost always. Since $\mathbb{I}(f)$ is the summation of indicators, this means almost always $\mathbb{I}(f)\geq 1$.
First observe that
$$
E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]
= {\binom {n}{k}}2^{n-k}\cdot\frac{ (2^{k+1}-2)}{2^{2^k}} $$
$$ > \frac{{\binom {n}{k}}2^{n-k}\cdot 2^{k}}{2^{2^k}}
> \frac{n^k 2^n}{k^k 2^{2^k}},
$$
where the last term is by $\binom {n}{k}>n^k/k^k$. It is easy to see that if $k\leq \lg{n}$ then $2^{2^{k}}\leq 2^n$ and $n^k\gg k^k$. Therefore, as $n\to \infty$ and $k\leq \lg{n}$, $E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\to \infty$.
Furthermore,
$$
\frac{1}{\left(E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\right)^2}
<\left( \frac{2^{2^k}}{ {\binom {n}{k}}\cdot 2^{n} } \right)^2 .
$$
On the other hand, by Fact \ref{dep},
$$
\Delta=\sum_{i\backsim j}\Pr[A_i\wedge A_j]
< {\binom {n}{k}}2^{n-k} \cdot \binom{n}{k}2^k \cdot \frac{2^{k+1}}{2^{2^k}}.
$$
Combining the above two inequalities, we get
$$
\frac{\Delta}{\left(E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\right)^2}
<\frac{2^{2^k+k+1}}{2^n}
$$
Note that $k= \lg{n}-1\Rightarrow 2^k= n/2$. This makes ${\Delta}/{\left(E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\right)^2}\to 0$, i.e. $\Delta=o({\left(E_{f\in \mathcal{F}_n}\left[\mathbb{I}(f) \right]\right)^2})$. Hence by Lemma \ref{probmd}, if $n$ is large enough then $\mathbb{I}(f)\gg 1$ almost always.
\end{proof}
\begin{theorem}\label{thm_nonsymexact}
Let $f\in \mathcal{F}_{n}$ be a non-symmetric function.
Let $m=\Pi_{i=1}^{r}p_i^{e_i}$, where $p_i$'s are distinct primes, and $r$ and $e_i$'s are positive integers.
Then almost always
$$
m\cdot d_{p_1^{e_1}}(f)\cdots d_{p_r^{e_r}}(f) > \lg{n}-1.
$$
\end{theorem}
\begin{proof}
For a given non-symmetric $f$, let $k=\lg{n}-1$, then the above lemma implies that almost always we can find a restriction $\rho\in \mathcal{R}_{k,n}$ for this $f$ such that $f|_{\rho}$ is no-junta and symmetric. Obviously $\forall i$ $d_{p_i^{e_i}}(f)\geq d_{p_i^{e_i}}(f|_{\rho})$.
Furthermore, $f|_{\rho}\in \mathcal{F}_{k}^*$. Hence, for $f|_{\rho}$ we have
$m\Pi_{i=1}^{r}{d_{p_i^{e_i}}(f)} \geq m\Pi_{i=1}^{r}{d_{p_i^{e_i}}(f|_{\rho})} > \lg{n}-1$.
\end{proof}
\begin{corollary}
With the same setting of the above theorem, and let $d_{max}=\max_{i\in [r]}(d_{p_i^{e_i}}(f))$ and $M=\max_{i\in [r]}(p_i^{e_i})$, then almost always
$$
d_{max}>\frac{\left(\lg{n}-1\right)^{1/r}}{M}.
$$
\end{corollary}
For the $\mathsf{OR}_n$ function, although in different representing models, this corollary is analogue to a result of \cite{BL14}, which states the degree lower bound of the weak representation by nonclassical polynomials is $\Omega{((\lg{n})^{1/r})}$.
It is interesting to know why different proofs meet the same barrier of $\lg{n}$; does this relate to the entropy?
By taking $m=pq$ with distinct primes $p$ and $q$, we immediately have $d_p(f)d_q(f)=\Omega(\lg{n})$ almost always.
\begin{corollary}
It is almost always true that $f\in \mathcal{F}_n$ has exact representing polynomial of degree $o(\sqrt{\lg{n}})$ over at most one finite field.
\end{corollary}
Note the lower bound in \cite{GLS10} means $f$ has degree $o(\lg{n})$ in at most one finite field. Although our result is weaker than \cite{GLS10}, our proof and the bound are both much simpler.
\section{Conclusions}
We prove any $n$-variate Boolean has no-junta subfunctions. It shows a possible way to deal with general Boolean functions via its subfunctions of under some restrictions. Besides,
for symmetric boolean functions, we proved $pqd_p(f)d_q(f)>n$, where $p$ and $q$ are distinct primes. This means any no-junta symmetric $f$ can have degree $o(\sqrt{n})$ in at most one finite field. In the nonsymmetric case, we prove that for a random function $f:\{0,1\}^n\to \{0,1\}$ almost always $pqd_p(f)d_q(f)>\lg(n)-1$, which means almost always $f$ can have degree $o(\sqrt{\lg{n}})$ in at most one finite field.
|
\section{Introduction}
Dark Matter (DM) is the dominant mass component in the universe. The hierarchical cold DM model with a
cosmological constant ($\Lambda$CDM) is successful in explaining many observations on large-scale structures.
The Navarro-Frenk-White (NFW) density profile (Navarro et al. 1997), which has been derived from DM only simulations,
provides an accurate description of the dark matter density profile over most radii, can be described
by just two parameters: the concentration and the mass enclosed in a region with a mean density above some
given threshold.
Simulations have shown that these two parameters are correlated, with the average concentration of a
halo being a weakly decreasing function of mass (e.g. NFW; Bullock et al. 2001; Hennawi et al. 2007; Duffy et al. 2008; Prada et al. 2012).
With the latest MultiDark simulations, Klypin et al. (2014) provide a more accurate analytical fit for dark matter halo density
profiles and concentrations.
However, the mass-concentration (\textit{m-c} hereafter) relation is different in the exact form as measured in
simulations and estimated from observations (e.g. Bullock et al. 2001; Comerford \& Natarajan 2007; Hennawi et
al. 2007; Duffy et al. 2008; Maccio et al. 2008; Mandelbaum et al. 2008; Oguri et al. 2009, 2012; Okabe et al.
2010; Sereno \& Covone 2013; Fo\"{e}x et al. 2014). For example, considering a sample of $28$ galaxy clusters whose radial profiles
are well constrained from combining weak and strong lensing data, Oguri et al. (2012) found a steeper function
of mass, with the slope $\sim -0.59\pm0.12$. Merten et al. (2014) and Umetsu et al. (2014) studied the
\textit{m-c}
relation with the Cluster Lensing and Supernova Survey with Hubble (CLASH) data, which is in excellent
agreement with simulations when the CLASH X-ray selection function and projection effects are taken into
account. Du \& Fan (2014) studied the bias from the miscentering problem, selection effects and shape noise from intrinsic
ellipticity of background galaxies on the measurement of \textit{m-c} relation with individual clusters. They found that
the existence of noise can compensate the miscentering problem and reduce the systematic bias although the scatters
of mass and concentration get considerably larger.
Mandelbaum et al. (2008) performed a statistical analysis to estimate the \textit{m-c} relation over three
orders of magnitude in mass using the stacked weak lensing signals of galaxies, groups and clusters in
Sloan Digital Sky Survey (SDSS).
Their analysis used the information outside the halo central region in order to avoid regions where
baryons dominate, and to
minimize errors from mis-estimation of clusters' centers. They found an \textit{m-c} relation with a slope consistent
with simulations, but an amplitude that is about $2\sigma$ below theoretical expectations. With the
SDSS-III/Constant Mass (CMASS) galaxies in the Canada-France-Hawaii Telescope Lensing Survey (CFHTLenS),
Miyatake et al. (2013) also showed a $20\%$ smaller
concentration than the theoretical predictions. With the CFHTLenS catalog and photometrically selected clusters, Covone
et al. (2014) found an over-concentrated but flat relation.
Galaxy properties, such as
luminosity or stellar mass, are tightly coupled to the DM halo mass. The link between the stellar and DM
halos is critical for understanding the formation and evolution of galaxies. The observed luminosity of
the central galaxy is always used to fix the stellar mass component. There are several methods to measure the
stellar-to-halo mass relation (\textit{SHMR}) including stellar kinematics (Conroy et al. 2007; More et al. 2011),
abundance matching (Moster et al. 2010; Behroozi et al. 2010), the Tully-Fisher relation (Pizagno et al. 2007;
Miller et al. 2014) and weak lensing (Mandelbaum et al. 2006; Leauthaud et al. 2012; Velander et al. 2014;
Hudson et al. 2015; Han et al. 2015). In particular, Leauthaud et al. (2012) found that the slope of the
\textit{SHMR} rises sharply at $M_*>\sim 4.5\times 10^{10}M_{\odot}$.
In this paper, we use a sample of redMaPPer clusters (Rykoff et al. 2014), the LOWZ and CMASS galaxies of SDSS-III BOSS DR10
and the shear catalog from the CFHT Stripe 82 dataset to measure the tangential shear, and quantify the
\textit{m-c} relation and the \textit{SHMR}.
\section{OBSERVATIONAL DATA} \label{sec: data}
\subsection{The Source catalog}
The source galaxies used in our measurement are taken from the CFHT Stripe 82
Survey (CS82: Erben et al. 2015), which is an $i$-band imaging survey covering a large fraction of SDSS Stripe 82
region with a median seeing $0.59"$. The survey contains a total of $173$ tiles with $165$ from CS82 and $8$ f
rom CFHT-LS Wide. Each CS82
tile was obtained from $4$ consecutive dithered observations each with an exposure time
of $410$ seconds. The completeness magnitude is $i_{AB}\sim24.0$.
The shapes of faint galaxies are measured with the \textit{lensfit} method (Miller et al 2007, 2013).
In this work we use the same pipeline as the CFHTLenS collaboration. Therefore the shear calibration
factors described in Heymans et al. (2012) can be directly applied to our data.
We use all galaxies
with weight $w>0$, FITCLASS$=$0 and MASK$\leq$1, in which $w$ represents an inverse variance weight assigned
to each source galaxy by \textit{lensfit}, FITSCLASS is a star/galaxy classification provided by
\textit{lensfit}.
The parameter MASK describes the mask information at an object's position. Objects with MASK$\leq$1 can
safely be used for most weak lensing analyses (Erben et al. 2013). After masking out
bright stars and other image artifacts, the effective sky coverage is $\sim$130$\rm deg^2$.
As CS82 contains only $i$-band, we derive the photometric redshifts (photo-z) using \textit{EAZY}
(Brammer et al. 2008) for the source galaxies from the overlapping multi-color co-add data from
SDSS (Annis et al. 2014). The total number of source galaxies that pass the aforementioned cuts
and have an available photo-z is $4,449,618$ ($\sim 9.5 gals/arcmin^2$).
\subsection{Lens selection}
We use the redMaPPer clusters and LOWZ/CMASS galaxies in two redshift bins, $0.2<z<0.4$ and $0.4<z<0.6$, as
lenses:
(1) The RedMaPPer Cluster Catalog:
The redMaPPer cluster catalog by Rykoff et al. (2014) is based on
the optimized red-sequence richness estimator $\lambda$. The richness $\lambda$ is the number of red sequence
galaxies brighter than $0.2L_{*}$ at the redshift of the cluster within a scaled aperture. The redMaPPer
cluster catalog in CS82, which is obtained from the SDSS Stripe 82 coadd using the CS82 object detections,
covers the redshift range $0.1<z<0.6$. The objects with richness $\lambda>15$ correspond to
the halo mass $\sim8\times10^{13}~M_{\odot}$ (Erben et al. 2015). In order to reduce the effects of
miscentering problem, we choose
the clusters with well-defined centers $\rm P_{\rm cen}>0.9$, which is the center probability. We then divide
the clusters into two redshift bins with redshift $0.2<z<0.4$ and $0.4<z<0.6$. For both low and high redshift
clusters, we divide
them into two richness bins ($20<\lambda<30, 30<\lambda<200$) with ($57, 31$) and ($96, 75$) systems, respectively.
(2) LOWZ \& CMASS Galaxy Catalog:
We also use the LOWZ and CMASS galaxy catalog from SDSS-III BOSS DR10 as lens galaxies\footnote{\tt https://www.sdss3.org/dr10/spectro/galaxy.php}. The two samples are both the main
SDSS-III BOSS BAO tracers (Dawson et al. 2013). LOWZ samples consist of red galaxies at $z<0.4$ from the SDSS DR8 (Aihara et al. 2011)
imaging data; CMASS samples are selected with an approximately constant stellar mass threshold (Eisenstein et al. 2011).
Stellar masses of the two samples are from the Portsmouth SED-fitting (Maraston et al. 2013) based on the BOSS
spectroscopic redshift.
We choose the LOWZ and CMASS galaxies within redshift $0.2<z<0.4$ and $0.4<z<0.6$.
We also divide them into $3$ and $4$ stellar mass bins, which will be used on the \textit{SHMR} measurement,
with ($2350, 2492, 2325$) and ($5966, 6168, 7481, 7611$) systems for the low-z and high-z range, respectively (see Figure~1).
\section{Lensing Signal}
We measure the \textit{m-c} relation with stacked lensing signal. Stacking many
halos reduces the fluctuations due to noise caused by uncorrelated structures along the line-of-sight,
shape measurement noise, substructures, and the shape variations of individual halos. Thus, the measurement
can determine the average mass profile. Furthermore, it allows for the lensing measurement of lower mass
halos, where individual detection is impossible due to their smaller shears relative to clusters.
Individual massive cluster observations and those based on stacked analysis of many halos are thus
complementary, drastically increasing the available baseline in mass.
\begin{figure*}
\includegraphics[width=0.49\textwidth]{gglensing_z0204}
\includegraphics[width=0.49\textwidth]{gglensing_z0406}
\caption{Best-fit model for the low (left) and high (right) redshift samples (black solid lines). Black
circles show the measured excess surface mass density $\Delta \Sigma$ of both LOWZ/CMASS galaxies and
clusters with different bins (clusters: Bottom-middle and bottom-right panels; LOWZ/CMASS: others).
The model components are the central baryonic matter (red), the dark matter halo profile (green), the
miscentering halo component (blue) and neighboring halos (magenta).}
\label{fig:gglensing}
\end{figure*}
Except the dark matter halos, the models for the lensing signals include the following components:
(1) Gavazzi et al. (2007) found that a central baryonic component is required to fit both the strong and weak
lensing profiles of early-type galaxies in the SLACS survey. This component can be simply modeled as a point
mass (e.g., Leauthaud et al. 2012).
(2) In order to measure the lensing signals, locating the centers of dark matter halos is very important because
miscentering can add significant systematic uncertainties. In general, halo centers are typically assumed to be traced by a
central galaxy. However, it has been argued that these assumptions may not be entirely
true and that central galaxies could be ``sloshing'' around in the dark matter halo potential well (e.g. van den Bosch et al. 2005; Gao \& White
2006; Johnston et al. 2007; Hikage et al. 2013). As in Johnston et al. (2007), Leauthaud et al. (2010),
George et al. (2012) and Li et al. (2014), centroid errors will lead to a smoothing of the lensing signal on
small scales and to an underestimate of halo masses. The distribution of miscentering can be described by a 2D Gaussian
distribution:
\begin{equation}
P(R_{\rm off})=\frac{R_{\rm off}}{\sigma_{\rm off}}\exp{\left(-\frac{1}{2}(R_{\rm off}/\sigma_{\rm off})^2 \right)},
\end{equation}
where $R_{\rm off}$ is the distance between the halo center and central galaxy position,
$\sigma_{\rm off}$ is the effective scale length. This model is from mock catalogs by
Johnston et al. (2007), which is consistent with the measurement from
2D lensing shear maps of $25$ massive clusters by Oguri et al. (2010).
\begin{table*}
\centering
\caption{Density Profile Models in Figure~1.}
\begin{tabular}{ccccccccc}
\hline
\hline
& z & $M_{\rm vir}$ & $c_{\rm vir}$ & $M_{200}$ & $c_{200}$ & $M_*$ & $\sigma_{\rm off}$ & $p_{\rm cen}$ \\
& & $10^{14} M_{\odot}/h$ & & $10^{14} M_{\odot}/h$ & & $10^{12} M_{\odot}$ & $\rm Mpc$ & \\
\hline
$(10^{11.0}, 10^{11.2})M_{\odot}$ & $0.2<z<0.4$ & $0.12\pm0.04$ & $4.66\pm1.29$ & $0.10\pm0.03$ & $3.80\pm1.09$ & $0.13\pm0.10$ & $0.45\pm0.26$ & $0.76\pm0.18$\\
$(10^{11.2}, 10^{11.35})M_{\odot}$ & $0.2<z<0.4$ & $0.15\pm0.04$ & $4.19\pm0.85$ & $0.13\pm0.03$ & $3.41\pm0.71$ & $0.19\pm0.17$ & $0.30\pm0.17$ & $0.82\pm0.13$\\
$(10^{11.35}, 10^{12.0})M_{\odot}$ & $0.2<z<0.4$ & $0.37\pm0.07$ & $3.60\pm0.66$ & $0.31\pm0.06$ & $2.91\pm0.55$ & $0.30\pm0.27$ & $0.19\pm0.15$ & $0.89\pm0.10$\\
$20<\lambda<30$ & $0.2<z<0.4$ & $1.01\pm0.09$ & $3.15\pm1.02$ & $0.83\pm0.073$ & $2.54\pm0.85$ & $0.32\pm0.25$ & $0.55\pm0.28$ & $0.95\pm0.12$\\
$30<\lambda<200$ & $0.2<z<0.4$ & $1.37\pm0.36$ & $3.66\pm1.20$ & $1.14\pm0.30$ & $2.96\pm1.01$ & $0.37\pm0.21$ & $0.63\pm0.25$ & $0.96\pm0.11$\\
\hline
$(10^{10.0}, 10^{11.1})M_{\odot}$ & $0.4<z<0.6$ & $0.074\pm0.03$ & $5.40\pm1.51$ & $0.067\pm0.03$ & $4.62\pm1.31$ & $0.095\pm0.056$ & $0.32\pm0.17$ & $0.86\pm0.15$\\
$(10^{11.1}, 10^{11.2})M_{\odot}$ & $0.4<z<0.6$ & $0.083\pm0.03$ & $5.00\pm0.97$ & $0.074\pm0.03$ & $4.26\pm0.84$ & $0.14\pm0.11$ & $0.23\pm0.12$ & $0.87\pm0.13$\\
$(10^{11.2}, 10^{11.35})M_{\odot}$ & $0.4<z<0.6$ & $0.12\pm0.05$ & $4.69\pm0.91$ & $0.11\pm0.04$ & $3.99\pm0.79$ &$0.19\pm0.15$ & $0.40\pm0.17$ & $0.86\pm0.12$\\
$(10^{11.35}, 10^{12.0})M_{\odot}$ & $0.4<z<0.6$ & $0.24\pm0.06$ & $3.88\pm0.82$ & $0.22\pm0.04$ & $3.28\pm0.70$ & $0.30\pm0.18$ & $0.24\pm0.14$ & $0.88\pm0.21$\\
$20<\lambda<30$ & $0.4<z<0.6$ & $0.68\pm0.11$ & $3.62\pm1.11$ & $0.60\pm0.084$ & $3.06\pm0.96$ & $0.33\pm0.21$ & $0.57\pm0.33$ & $0.91\pm0.21$\\
$30<\lambda<200$ & $0.4<z<0.6$ & $1.39\pm0.30$ & $3.35\pm0.52$ & $1.20\pm0.26$ & $2.83\pm0.45$ & $0.47\pm0.28$ & $0.68\pm0.21$ & $0.90\pm0.33$\\
\hline
\hline
\end{tabular}
\leftline{Notes: $M_*$ is the stellar mass from the Portsmouth SED-fitting (Maraston et al. 2013) based
on the SDSS-III BOSS DR10 data.}
\label{tab:tab1}
\end{table*}
(3) At the large scales, the neighboring halos contribute to the lensing profile (Johnston et al. 2007;
Cacciato et al. 2009, Li et al. 2009, Oguri \& Hamana 2011). The tangential shear profile due to the neighboring halos is
\begin{equation}
\gamma_{2h}(\theta; M,z)=\int \frac{ldl}{2\pi}J_2(l\theta)\frac{\bar{\rho}_m(z)b_h(M)}{(1+z)^3\Sigma_{\rm crit}D_A^2(z)}P_m(k_l;z),
\end{equation}
where $J_2$ is the second order Bessel function, $\bar{\rho}_m(z)$ is the mass density at $z$,
$D_A(z)$ is the angular diameter distance, $P_m(k)$ is the
linear matter power spectrum, $\Sigma_{\rm crit}$ is the critical mass density, and
$b_h(M)$ is the halo bias that we take from Tinker et al. (2010).
Combining all the contributions, the lensing signal $\Delta\Sigma$ can be written as:
\begin{eqnarray}
\Delta\Sigma(R)=\frac{M_p}{\pi R^2}+p_{\rm cen}\Delta\Sigma_{\rm NFW}(R)+\\
(1-p_{\rm cen})\Delta\Sigma^{\rm off}_{\rm NFW}(R|P_{\rm off})+\Delta\Sigma_{2h}(R), \nonumber
\end{eqnarray}
where sequentially, the terms come from the contribution of central baryonic components, the dark matter halos, the miscentering halos, and the neighboring halos.
The model is described by $5$ parameters: the halo mass $M_{\rm vir}$, the concentration $c_{\rm vir}$,
the central baryonic mass $M_p$, the effective scale length $\sigma_{\rm off}$ and the probability $p_{\rm cen}$
that the centers we use here are accurately centered on the dark matter halos.
To obtain $\Delta\Sigma$ we stack lens-source pairs in $8$ logarithmic radial ($R$) bins from $0.1$ to
$10~\rm Mpc~h^{-1}$. For a sample of selected lenses, $\Delta\Sigma(R)$ is estimated using
\begin{equation}
\Delta\Sigma(R)=\frac{\sum_{ls}w_{ls}\gamma_t^{ls}\Sigma_{\rm crit}}{\sum_{ls}w_{ls}}\,,
\end{equation}
where $\gamma_t^{ls}$ is the tangential shear, $w_{ls}=w_n\Sigma_{\rm crit}^{-2}$, and $w_n$ is a
weight factor introduced to account for intrinsic scatter in ellipticity and shape measurement error.
With the multiplicative errors $m$ taken into account statistically (Miller et al. 2013), we apply it to
the average shear measurement in both the low-z and high-z samples using the correction as Velander et
al. (2014) and Hudson et al. (2015):
\begin{equation}
1+K(z_l)=\frac{\sum_{ls}w_{ls}(1+m)}{\sum_{ls}w_{ls}}.
\end{equation}
The lensing signal can be calibrated as follows:
\begin{equation}
\Delta\Sigma^{\rm cal}(R)=\frac{\Delta\Sigma(R)}{1+K(z_l)}.
\end{equation}
The effect of this correction term is to increase the average lensing signals by $\sim 5.5\%$ and $6.2\%$
at $0.2<z<0.4$ and $0.4<z<0.6$, respectively.
To minimize the contamination due to photo-z uncertainties of the source galaxies, we remove the
lens-source pairs with $z_s-z_l<0.1$.
\section{Results}
We show the observed galaxy-galaxy lensing signals for the low and high redshift samples in Figure~1.
The black circles with error bars represent the measured excess surface mass density. We fit them to the
model described in section~3. The black solid curves shows the best-fit results with all the
components. The different model components are shown in dashed curves: the central baryonic matter (red),
the dark matter halo profile (green), the miscentering halo component (blue) and neighboring halos (magenta).
At small scales ($0.1-1 \rm Mpc$), the dark matter halo is the dominant component. At large scales ($>3 \rm Mpc$
for galaxies and $>5 \rm Mpc$ for clusters), the neighboring halos contribute more. The best-fit results are
given in Table~1. We convert the measured $M_{\rm vir}$ and $c_{\rm vir}$ to $M_{200}$ and $c_{200}$ with
the formula by Johnston et al. (2007).
The stellar masses $M_*$ of LOWZ/CMASS galaxies are from the Portsmouth SED-fitting (Maraston et al. 2013) based
on the SDSS-III BOSS DR10 data. $M_*$ of the redMaPPer clusters are
estimated by matching the positions of central galaxies with the galaxy catalog of SDSS DR10.
The fitted probability $p_{\rm cen}$ are all close to $1$, which indicates that: (1) the centers of redMaPPer
clusters are well defined; (2) LOWZ/CMASS galaxies are almost all central galaxies, $\sim 10\%$ of them may be
the most luminous satellite galaxies as the clustering analysis by Parejko et al. (2013) and White et al. (2011).
\subsection{\textit{m-c} relation}
The \textit{m-c} relation from the simulation predictions is given as:
\begin{equation}
c_{200}(z,M)=A\left(\frac{M_{200}}{M_0}\right)^{B}(1+z)^{C},
\end{equation}
with $M_0=2.0 \times 10^{12}M_{\odot}h^{-1}$. As we cannot get a redshift dependence relation with only
two redshift bins, the redshift dependence is calculated with the mass and concentration
with Multidark simulations of Klypin et al. (2014). Because of the upturn of the concentrations at large
masses regime of high redshift, the redshift dependence $\rm C\sim-0.67$ is fitted with the DM halos of
Multidark in the redshift region $0.0<z<0.6$, which is suitable to the data in our analysis.
The best-fit \textit{m-c} relations are shown in Table~2.
Because of the bigger errors on low-z sample, the \textit{m-c} relations of both samples are consistent.
Giocoli et al. (2012) found
that the halo triaxiality and the substructures within the host halo virial radius can bias the
\textit{m-c} relation. They propose a method for correcting the \textit{m-c} relation for the projection
effects and also for adiabatic contraction:
\begin{equation}
c_{2D}(M)=c_{3D}(M)\times 1.630M^{-0.018},
\end{equation}
In Figure~2, we show the results on the 3D corrected \textit{m-c} relation after rescaling the results to $z=0$.
The 3D correction is used on all the observational dataset and simulation predictions.
\begin{table}
\centering
\caption{Best-fit Mass-Concentration relations.}
\begin{tabular}{cccc}
\hline
\hline
z & A & B \\
\hline
$0.2<z<0.4$ & $5.25\pm1.67$ & $-0.13\pm0.12$ \\
$0.4<z<0.6$ & $6.77\pm1.13$ & $-0.15\pm0.06$ \\
\hline
\hline
\end{tabular}
\label{tab:tab2}
\end{table}
\begin{figure}
\includegraphics[width=0.5\textwidth]{mc_cs82_3d}
\caption{3D corrected \textit{m-c} relation with clusters and galaxies in CS82 after rescaling all the results to $z=0$.
Red and blue circles denote the stacked lensing signals from low and high redshift samples,
respectively. The red curve is the best-fit \textit{m-c} relation of the low redshift sample.
The blue curve and cyan-shaded area is best-fit \textit{m-c} relation of the high redshift sample and
its $1\sigma$ uncertainty. Black
curves are the simulation predictions by Duffy et al. (2008) (dashed) and Klypin et al. (2014) (solid).
The black symbols denote the measurement on mass and concentration by Mandelbaum et al. (2008) with
SDSS; Covone et al. (2014) with CFHTLenS; Umetsu et al. (2014) and Merten et al. (2014) with CLASH
cluster sample.}
\label{fig:mc}
\end{figure}
With the 3D corrections by Giocoli et al. (2012), we can directly compare our measurements with the
three-dimensional (3D) N-body simulation results.
Comparing with the results by Duffy et al. (2008) and Klypin et al. (2014), the amplitudes $\rm A$ and
slopes $\rm B$ of both
redshift samples are consistent with the simulation predictions within the errors.
We then compare our results with previous measurements. The 3D corrections by Giocoli et al. (2012) are
used on these measurements.
By stacking weak lensing signals of galaxies, groups and clusters from SDSS dataset, Mandelbaum et al. (2008)
fit the data to the model assuming a spherical NFW profile excluding small scales to reduce the effects of
baryons and miscentering halos. As shown in Figure~2, their measurements are comparable to ours, but more
scattered.
With photo-z selected clusters (Wen et al. 2009) in the CFHTLenS data, Covone et al. (2014)
measured the mass and concentration by stacking shear profiles. The measured concentrations in
four richness bins are higher than the simulation predictions and generate a flat and
over-concentrated mass-concentration relation.
A possible reason might be the choice of the cluster catalog. Wen's cluster finder purely based on the
photo-z of galaxies. The richness of some clusters may be overestimated due to the inaccuracy
of photo-z. A detail comparison will need to be carried on in future.
The recent CLASH measurements for individual clusters (Merten et al. 2014) and stacking (Umetsu et al. 2014)
focus on more massive strong lensing clusters $\sim 10^{15}M_{\odot}h^{-1}$, which give large concentrations.
Accounting for the CLASH selection function based on X-ray morphology and projections effects inherent in lensing
observations (Meneghetti et al. 2014), a better agreement can be found between observations and theoretical
predictions.
\begin{table*}
\centering
\caption{Best-fit results of \textit{SHMR}.}
\begin{tabular}{cccccc}
\hline
\hline
z & $\log_{10}(M_1)$ & $\log_{10}(M_{*,0})$ & $\beta$ & $\delta$ & $\gamma$ \\
\hline
$0.2<z<0.4$ & $12.51 \pm 0.047$ & $10.97\pm0.032$ & $0.47\pm0.019$ & $0.43\pm0.11$ & $1.45\pm0.27$\\
$0.4<z<0.6$ & $12.69 \pm 0.056$ & $11.13\pm0.035$ & $0.48\pm0.022$ & $0.44\pm0.15$ & $1.90\pm0.31$\\
\hline
\hline
\end{tabular}
\label{tab:tab3}
\end{table*}
\begin{figure}
\includegraphics[width=0.5\textwidth]{ms_cs82}
\caption{Halo-to-stellar mass ratio as a function of stellar mass.
Red and blue dots denote the measurements from low and high redshift samples,
respectively. The red solid curve and cyan-shaded area is the best-fit \textit{SHMR}
relation of our low redshift sample combined with H15 and its $1\sigma$ uncertainty.
The blue solid curve is the best-fit \textit{SHMR} relation of our high redshift sample combined with H15.
The red and blue dashed curves are the fitted relation by the COSMOS measurements
We compare our measurements with other
galaxy-galaxy lensing measurements, including SDSS (Mandelbaum et al. 2006), CFHTLenS (H15),
COSMOS (Leauthaud et al. 2012), COSMOS X-ray clusters (Leauthaud et al. 2010), and GAMA (Han et al. 2015).}
\label{fig:ms}
\end{figure}
\subsection{\textit{SHMR} measurement}
With the measurement of stellar mass, we can also estimate the \textit{SHMR}.
Following Leauthaud et al. (2012), $f_{SHMR}(M_h)$ is mathematically defined via its inverse
function:
\begin{eqnarray}
\log_{10}(f^{-1}_{SHMR}(M_h))=\log_{10}(M_1)+\beta \log_{10}\left(\frac{M_*}{M_{*,0}}\right)+ \\
\frac{\left(\frac{M_*}{M_{*,0}}\right)^{\delta}}{1+\left(\frac{M_*}{M_{*,0}}\right)^{-\gamma}}-\frac{1}{2}, \nonumber
\end{eqnarray}
where $M_1$ is a characteristic DM halo mass, $M_{*,0}$ is a characteristic stellar mass,
$\beta$ is the low-mass \textit{SHMR} slope, $\delta$ and $\gamma$ is the high-mass \textit{SHMR} slope.
In Figure~3, we show the results of this paper in addition to other galaxy-galaxy lensing studies, including the
measurement in SDSS (Mandelbaum et al. 2006), CFHTLenS (Hudson et al. 2015 (H15)),
the Cosmic Evolution Survey (COSMOS: Leauthaud et al. 2010, 2012) and the Galaxy And Mass Assembly (GAMA:
Han et al. 2015). Our measurement of $M_{200}/M_*$ is ideally combined with the results of H15, as they
provide very good estimates of the \textit{SHMR} at the same redshift bins as ours and covering smaller masses.
Our best fit results for this combined dataset are given in Table~3.
We compare our results with the best-fit in Leauthaud et al. (2012). At low stellar mass
($M_*<4\times10^{10}M_{\odot}$), the constraints are only from H15. The behaviors of \textit{SHMR}
are similar. The major halo mergers are
minor galaxy mergers in this regime. The growth of stellar mass through the effects of mergering will be limited
compared to the growth of DM halo mass. At $M_*\sim4-6\times10^{10}M_{\odot}$,
$M_{200}/M_*$ of the results reaches a minimum, which reflects the most efficient star formation rate at such a
mass. The minimum values are consistent with the COSMOS measurements within the errors.
At low stellar mass regime, the DM halos are dominated by single central galaxies, the supernovae feedback
could be the main reason to prevent the star formation. At high stellar mass, we find a relatively flatter slope.
The scatter should be due to the appearance of hot gas in the massive DM halos, making more complex merger
history.
\section{Conclusions}
Combing two different redshift samples of redMaPPer clusters and LOWZ/CMASS galaxies with the shear catalog of CS82,
we fit the lensing signals around these clusters and massive galaxies including the effects of baryons,
miscentering halos, and neighboring halos, and study the \textit{m-c} relation and also the \textit{SHMR}.
Our main results are as follows:
1. We find that the concentration decreases weakly with mass, $c=A(M/M_0)^{B}$ with the amplitude A and
slope $\rm B$ in good agreement with the previous measurement by Mandelbaum et al. (2008). Including
the 3D correction suggested by Giocoli et al. (2012), the two parameters are also consistent with
simulation predictions within the errors.
Klypin et al. (2014) also shows that the NFW profile can systematically overpredict the halo concentration
by $10-20\%$ for the massive halo density profile. The Einasto profile can provide a more accurate measurement.
However, the current precision of data cannot distinguish between the two profiles. The lensing signals in the
inner region of dark matter halos have large error bars. Mandelbaum et al. (2008) found a very similar
\textit{m-c} relation for the NFW and Einasto profiles with SDSS data. Using the recent CLASH,
Umetsu et al. (2014) found similar mass and concentration of the two profiles. In order to compare the two
profiles, we also fit the lensing signals of the redMaPPer clusters with the Einasto profile, and find similar
\textit{m-c} relation as with the NFW profile. We conclude that the current data can not distinguish the two
profiles.
2. Combining previous galaxy-galaxy lensing measurements, we measure a relatively flatter slope of the
\textit{SHMR} at high stellar mass. With COSMOS data, Leauthaud et al. (2012) found a steeper slope of
the \textit{SHMR} at $M_*>4.5\times10^{10}M_{\odot}$.
The flatter slope of the \textit{SHMR} at high stellar mass implies the massive DM halos contain more complex
structures and merger history. The scatter at the high stellar mass regime could be due to the appearance of hot
gas in the massive DM halos, making a more complex merger history.
In our analysis, we use the stellar mass of central galaxy for the redMaPPer clusters. There could
be more than one central galaxy in a cluster, however, making our measurement a lower limit on the total stellar
mass in a halo. Such an effect could raise
the \textit{SHMR} at high stellar mass regime and increase the scatter.
With future wide surveys (DES/HSC/LSST/EUCLID), we expect to measure more accurate the \textit{m-c} relation and
the \textit{SHMR}, and test for possible measured bias.
\section*{Acknowledgments}
This work is based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the
Canada-France-Hawaii Telescope (CFHT), which is operated by the National Research Council (NRC) of Canada,
the Institut National des Science de l'Univers of the Centre National de la Recherche Scientifique (CNRS) of
France, and the University of Hawaii. The Brazilian partnership on CFHT is managed by the Laborat\'{o}rio Nacional
de Astrof\'isica (LNA). This work made use of the CHE cluster, managed and funded by ICRA/CBPF/MCTI, with financial
support from FINEP and FAPERJ. We thank the support of the Laborat\'{o}rio Interinstitucional de e-Astronomia (LIneA).
We thank the CFHTLenS team for their pipeline development and verification upon which much of this surveys pipeline was
built.
The authors thank Francisco Prada and Charling Tao for useful discussions.
This research was supported by a Marie Curie International Incoming
Fellowship within the $7^{th}$ European Community Framework Programme. HYS acknowledges the support from Swiss National Science
Foundation (SNSF) and NSFC of China under grants 11103011. JPK acknowledges support from the ERC advanced grand LIDA.
LR acknowledges the NSFC grants (No.11303033) and the support from Youth Innovation Promotion Association of CAS.
JC acknowledges financial support from MINECO (Spain) under project number AYA2012-31101.
TE is supported by the Deutsche Forschungsgemeinschaft through the Transregional
Collaborative Research Centre TR 33 - 'The Dark Universe'.
BM acknowledges financial support from the CAPES Foundation grant 12174-13-0.
MM is partially supported by CNPq (grant 486586/2013-8) and FAPERJ (grant E-26/110.516/2-2012).
|
\section{Introduction}
Analysis of online activity of the internet users has proved its worth in various disciplines, most notably in psychology \cite{McCarthy2010,Page2011,Sueki2011,Yang2011,Hagihara2012}, ecology \cite{McCallum2013,McCallum2014,Ficetola2013,Verissimo2014}, epidemiology \cite{polgreen2008,ginsberg2009,Carneiro2009,Seifter2010,Dugas2012}, medicine \cite{Linkov2014,Telem2014} linguistics \cite{Mocanu2013}, politology \cite{Metaxas2012}, sociology \cite{Rybski2009,Grabowicz2014} and in a wide range of economics, marketing and finance \cite{Preis2010,mondria2010,Goel2010,Vosen2011,Drake2012,Bank2011,Bordino2012,Vlastakis2012,Dzielinski2012,Preis2012a,Preis2013,Kristoufek2013,Kristoufek2013a,Moat2013,Curme2014}. In the economic and financial applications, the focus has been primarily put on the search queries on various search engines such as Google, Yahoo! and Baidu. Bank \textit{et al.} \cite{Bank2011} find connection between Google searches and liquidity at the German stock market. Bordino \textit{et al.} \cite{Bordino2012} study traded volume of the NASDAQ-100 index component stocks and they report that it is correlated with the related searches of the Yahoo! engine. Vlastakis \& Markellos \cite{Vlastakis2012} find positive correlation between internet search queries for NASDAQ and NYSE stocks and their traded volume and volatility. Dzielinski \cite{Dzielinski2012} introduces an uncertainty measure based on the financial online search queries. Preis \textit{et al.} \cite{Preis2013} show that Google searches for financial terms can be used for profitable trading strategies. Kristoufek \cite{Kristoufek2013} utilizes popularity of the Dow Jones stocks measured by Google search queries for portfolio diversification. Kristoufek \cite{Kristoufek2013a} further studies dynamics between Google searches, Wikipedia page views and dynamics of the Bitcoin crypto-currency uncovering a strong relationship between these. Moat \textit{et al.} \cite{Moat2013} report that even Wikipedia page views can be utilized for the trading strategy construction. And Curme \textit{et al.} \cite{Curme2014} cluster the online searches into groups and show that mainly politics and business oriented searches are connected to the stock market movements.
The most frequently reported relationship between the online searches, traded volume and volatility directs further to the dynamic characteristics of the online searches time series. As traded volume and volatility have been repeatedly studied for their power-law correlation structures \cite{Cont2001,Poon2003,Poon2005}, the same research line is at hand for the online searches as well. Potential long-term memory of the online activity has further implications for modeling and correct inspection of dynamics between the searches and other series. Here, we examine the correlation structure of the Google searches related to the Dow Jones Industrial Average (DJIA) index components. Daily Google searches data are utilized for the components of DJIA and as such, we present the first such study of the correlation structure of the online searches. To do so, we apply the rescaled range and rescaled variance tests to uncover the power-law correlations structure and we further proceed with the detrended fluctuation analysis of the search queries series. As it turns out that the DJIA-related Google queries are in fact power-law correlated, we reinvestigate a popular topic of cross-correlations between the searches, traded volume and volatility of the examined stocks. As we find the online searches to be power-law correlated and on the edge of (non)stationarity, we utilize the newly proposed correlation coefficients based on the detrended cross-correlation and detrending moving-average cross-correlation analyses.
The paper is organized as follows. In Section 2, we describe the used methodology, specifically the rescaled range and rescaled variance tests together with the moving block bootstrap significance criterion, and the detrended fluctuation analysis as well as the correlation coefficients. Section 3 introduces the dataset and presents the results. Section 4 concludes. We show that the Google searches related to the DJIA component stocks show scaling characteristic for the power-law correlated processes. This is supported by all utilized methods. General interest in the publicly traded companies thus shares similar properties to the variance and traded volume series -- there are profound periods of high interest followed by long-lived periods of low interest. However, the search series always revert back to a long-term trend so that no explosive behavior is observed. After taking the long-term memory aspect of the online query series into consideration, the correlations between the searches, traded volume and volatility become quite unstable and no universal relationship is found. The initial long-term memory analysis thus proves to be crucial for a correct treatment of cross-correlations between the online searches and various possibly connected series.
\section{Methodology}
\subsection{Long-term memory and its tests}
Long-term memory (or alternatively long-range dependence and long-range correlations) is defined through a power-law decay of the auto-correlation function $\rho(k)$ which scales as $\rho(k) \propto k^{2H-2}$ for lag $k \rightarrow +\infty$ \cite{Hurst1951,Beran1994,Samorodnitsky2006}. The series are then referred to as the power-law (auto-)correlated processes as well. The characteristic parameter of such processes is Hurst exponent $H$, or alternatively parameter $\alpha$, which takes values between 0 and 1 for stationary processes. The breaking value of $H=0.5$ characterizes a process with no long-term memory. Processes with $H>0.5$ are usually referred to as persistent processes whereas the ones with $H<0.5$ as anti-persistent processes. The former ones are reminiscent of locally trending processes which, however, keep their stationarity (for $H<1$) and return to their mean value quickly enough. The latter ones are very erratic in behavior as they switch their direction more frequently than uncorrelated processes. Integrating the stationary long-range dependent processes once creates an additional category of processes which have interesting properties. For $1 \le H < 1.5$, we have non-stationary yet still mean reverting processes. The frontier of $H=1.5$ marks a unit root process and $H>1.5$ characterizes processes which are non-stationary and not mean reverting, i.e. explosive processes. The long-term memory property of time series has far-reaching consequences for the time series modeling and forecasting mainly due to its implication of a non-summable auto-correlation function \cite{Samorodnitsky2006}. Therefore, it is essential to distinguish between long-range dependence with its power-law correlations and short-range dependence with its exponential correlation structure. For this purpose, we utilize the modified rescaled range test and the rescaled variance test.
The modified rescaled range test \cite{Lo1991} is an adjusted version of the original rescaled range analysis \cite{Hurst1951}. Both methods are based on scaling of the rescaled ranges with an increasing time series length. For the time series $\{x_t\}$ with $t=1,2,\ldots,T$, the testing statistic $V_T$ is defined as
\begin{equation}
\label{MRS}
V_{T}=\frac{R}{S\sqrt{T}} \nonumber
\end{equation}
where $R$ is a range of the profile of the analyzed series,
\begin{equation}
R=\max_{t=1,\ldots,T}\left(\sum_{i=1}^{t}{(x_i-\bar{x}})\right)-\min_{t=1,\ldots,T}\left(\sum_{i=1}^{t}{(x_i-\bar{x}})\right), \nonumber
\end{equation}
with $\bar{x}$ being the time series average, $S$ is a heteroskedasticity and autocorrelation consistent (HAC) estimator of the standard deviation of the original series, defined as
\begin{equation}
\label{S}
S^2=\widehat{\gamma}(0)+2\sum_{k=1}^q{\left(1-\frac{k}{q+1}\right)\widehat{\gamma}(k)},
\end{equation}
with $\widehat{\gamma}(k)$ being an estimated auto-correlation with lag $k$ using the Barlett-kernel weights. Note that $\widehat{\gamma}(0)$ is an estimated variance. The crucial difference between the original and the modified version of the test stems in Eq. \ref{S} which is constructed to control for a possible short-term memory bias. Selection of the parameter $q$ then becomes crucial as an overshot parameter $q$ can suppress even long-term memory whereas an undershot $q$ parameter can direct to a misleadingly found long-term memory which in fact is only a strong short-term memory. We stick to an automatic selection criterion of the parameter as proposed by Lo \cite{Lo1991}
\begin{equation}
\label{eq2}
q^{\ast}=\left\lfloor\left(\frac{3T}{2}\right)^{\frac{1}{3}}\left(\frac{2|\widehat{\rho}(1)|}{1-\widehat{\rho}(1)^2}\right)^{\frac{2}{3}}\right\rfloor
\end{equation}
where $\widehat{\rho}(1)$ is the sample first order autocorrelation and $\lfloor \rfloor$ is the lower integer operator.
The rescaled variance test \cite{Giraitis2003} is based on a very similar idea as the previous one but, as the name suggests, it is based on the profile variance rather than the profile range so that it is less sensitive to extreme values. The testing statistic $M_T$ is then defined as
\begin{equation}
M_T=\frac{\text{var}(X)}{TS^2} \nonumber
\end{equation}
where $\text{var}(X)$ is the variance of the profile of the original series. To control for the short-term memory bias, the HAC standard deviation from Eq. \ref{S} and the optimal $q^{\ast}$ from Eq. \ref{eq2} are used here as well.
Even though both $V_T$ and $M_T$ have well defined asymptotic critical values \cite{Lo1991,Giraitis2003}, we opt for an alternative approach utilizing the moving block bootstrap methodology \cite{Efron1993,Srinivas2000} due to a finite sample, a very heterogenous dynamics of the analyzed series as well as their distributional properties. In the procedure, surrogate series are formed by shuffling the blocks of a fixed size from the original series. This way, the short-term correlations and distributional properties are kept but the long-term correlations are shuffled away creating a distribution of the testing statistic under a more realistic null hypothesis. In our application, we fix the block size to 25 observations and we bootstrap 1000 surrogate series to obtain statistical significance.
\subsection{Detrended fluctuation analysis}
Detrended fluctuation analysis (DFA) \cite{Peng1993,Peng1994,Kantelhardt2002} is the most popular and the most frequently applied time domain estimator of Hurst exponent. This is mainly due to the fact that DFA works under various settings such as non-stationarity and trends \cite{Kantelhardt2002}, periodic cycles and seasonalities \cite{Hu2001}, and heavy tails \cite{Barunik2010}.
The procedure is based on the following steps. We work with the profile $X(t)$ of the series $\{x_t\}$ with $t=1,\ldots,T$ defined as
\begin{equation}
X(t)=\sum_{i=1}^{t}{(x_i-\bar{x})}. \nonumber
\end{equation}
The profile is divided into $T_s\equiv \lfloor T/s \rfloor$ non-overlapping windows of length $s$ which is referred to as a scale. The time series length $T$ may be non-divisible by $s$ which creates an issue with the end of the series which would not be used in the procedure. For this purpose, the series is in addition divided into boxes from the end of the series so that we obtain $2T_s$ boxes of size $s$. In each of these boxes, we calculate a mean squared deviation from the linear time trend inside the box. This means that for the $j$th box of size $s$, we obtain
\begin{equation}
F^2(j,s)=\frac{1}{s}\sum_{i=1}^{s}{(X(s[j-1]+i)-\widehat{X_j(i)})^2} \nonumber
\end{equation}
where $\widehat{X_j(i)}$ is a linear fit of a time trend at position $i$ in window $j$. In a similar manner, we obtain the fluctuation for the boxes formed from the end of the series as
\begin{equation}
F^2(j,s)=\frac{1}{s}\sum_{i=1}^{s}{(X(T-s[j-T_s]+i)-\widehat{X_j(i)})^2}. \nonumber
\end{equation}
We then construct a fluctuation for specific scale $s$ as
\begin{equation}
F(s)=\left(\frac{1}{2T_s}\sum_{j=1}^{2T_s}{[F^2(j,s)]} \right)^{\frac{1}{2}} \nonumber
\end{equation}
and finally, we obtain Hurst exponent via the scaling law
\begin{equation}
\label{HH}
F(s) \propto s^H.
\end{equation}
In the application, we estimate the exponent for scales between $s_{min}=10$ and $s_{max}=500 \approx T/5$. Moreover for better illustrational purposes, we base the estimation and the results on scales $s$ which are powers of 10 to a single decimal point. Note, however, that the results do not change qualitatively for other specifications of scales and box splitting procedures and such approach is thus kept primarily for a straightforward presentation of the results.
\subsection{DCCA and DMCA coefficients}
The detrended cross-correlation coefficient $\rho_{DCCA}(s)$ for scale $s$ as proposed by Zebende \cite{Zebende2011} is a combination of the detrended fluctuation analysis (DFA) \citep{Peng1993,Peng1994,Kantelhardt2002} and the detrended cross-correlation analysis (DCCA) \citep{Podobnik2008,Zhou2008,Jiang2011}. The coefficient is defined as
\begin{equation}
\rho_{DCCA}(s)=\frac{F^2_{DCCA}(s)}{F_{DFA,x}(s)F_{DFA,y}(s)}, \nonumber
\label{rho}
\end{equation}
where $F^2_{DCCA}(s)$ is a detrended covariance between profiles of series $\{x_t\}$ and $\{y_t\}$ based on a window of size $s$, and $F^2_{DFA,x}$ and $F^2_{DFA,y}$ are detrended variances of profiles of the separate series, respectively, for a window size $s$\footnote{DCCA is a bivariate generalization of DFA presented in the previous section.}. For time series of length $T$, the series is divided into non-overlapping boxes of length $s$. In each box, fluctuation functions are computed for linearly detrended series which are in turn averaged over all boxes of the same length. In the case when $T$ is not divisible by $s$, the series is divided from the beginning as well as from the end and the averages are based on these sub-periods as in the case of DFA. More details about the methods and some alternative specifications can be found in Refs. \cite{Kantelhardt2002,Podobnik2008,Kristoufek2014,He2011,He2011b,Cao2014,Wang2014}.
The detrending moving-average cross-correlation coefficient $\rho_{DMCA}(\lambda)$ for scale $\lambda$ has been introduced by Kristoufek \cite{Kristoufek2014a} as an alternative to the DCCA coefficient. The method builds on a connection between the detrending moving average (DMA) procedure \citep{Vandewalle1998,Alessio2002} and detrending moving-average cross-correlation analysis (DMCA) \citep{Arianos2009,He2011a}. The coefficient is defined as
\begin{equation}
\rho_{DMCA}(\lambda)=\frac{F_{DMCA}^2(\lambda)}{F_{x,DMA}(\lambda)F_{y,DMA}(\lambda)}, \nonumber
\end{equation}
where $F^2_{DMCA}(\lambda)$, $F^2_{DMA,x}(\lambda)$ and $F^2_{DMA,y}(\lambda)$ are a detrended covariance between profiles of the examined series and detrended variances of the separate series, respectively, with a moving average parameter $\lambda$. Fluctuation functions are based on series detrended by a centered moving average of length $\lambda$. Various specifications can be utilized for the detrending but the centered averaging has been shown to outperform the contenders \cite{Carbone2003}. Contrary to the DCCA coefficient, the DMCA coefficient is not based on a box-splitting and it is thus computationally more efficient. More details can be found in Refs. \cite{Kristoufek2014a,Alessio2002,Arianos2009}.
In a series of papers, Kristoufek \cite{Kristoufek2014,Kristoufek2014a} shows that the statistical properties of both methods depend strongly on long-term memory properties of the separate series. Moreover, reliability of the methods is not constant for different levels of correlation between the studied processes either. To control for such effects, we apply the Theiler's Amplitude Adjusted Fourier Transform (TAAF) \citep{Theiler1992}. This method reconstructs the series with the same spectral as well as distributional properties as the original one. This way, we obtain two series with an unchanged auto-correlation and distributional structure which are, however, pairwise uncorrelated. Statistical significance of estimated correlations based on DCCA and DMCA can be then obtained and tested.
Specifically for each studied pair of processes, we obtain TAAF transformed series which are not cross-correlated but retain the auto-correlation and distribution properties of the original series. The DCCA and DMCA coefficients are then estimated for such series. As the series are not cross-correlated, the expected value of the coefficients is zero. However, variance of the estimates can be possibly high. Therefore, we estimate the coefficients on 1000 surrogate series to obtain a finite sample distribution under the null hypothesis of no cross-correlations between series which controls for both long-term auto-correlations and distributional properties.
\section{Data and results}
Google provides search query time series for specified terms from the year of 2004 onwards. However, the series are not reported as a pure number of searches for a given term but these are renormalized according to the Google algorithm which can be in essence seen as rescaling the searches into the 0--100 interval so that the number represents the proportion of the specified searched term among all searched terms in time being kept between 0 and 100. Moreover, the obtained numbers are based on sampling from all searched terms so that these represent an estimated rescaled proportion. Even though such rescaling procedure can somewhat dilute the information content of the series, the empirical results summarized in the introductory section show otherwise.
The Google data can be downloaded freely from the Google Trends website (trends.google.com) at a weekly frequency. To obtain the data at a higher frequency, specifically the daily one, one needs to download the series in three-months sections and the series further need to be rescaled and chained together. We apply such procedure for the component stocks of the Dow Jones Industrial Average (DJIA) index between years 2004 and 2013 (apart from Exxon Mobil, J. P. Morgan and Procter \& Gamble for which the series are several months shorter which will be evident later in the text) and thus obtaining 2516 observations for most series. The most severe issue with the Google queries data is its relative arbitrariness in defining the searched terms. Further, the sampling and thresholding procedure applied by Google for its search series quite frequently ends up with reporting incomplete series. If the specified term is not searched for frequently enough, the series is practically useless. We thus analyze only the component stocks which provide reliable search query series. Out of 30 DJIA stocks, we end up with 18 stocks for which the Google series are reliable without discontinuities. The analyzed stocks are summarized in Tab \ref{query}. We have tried various combinations and specifications of the searched terms and we report the ones which provided the most complete series.
The Google searches for the analyzed stocks are illustrated in Figs. \ref{fig_series1} and \ref{fig_series2}. These uncover that the searching frequencies for the component stocks are very heterogenous. The trends are sharply decreasing (IBM, Merck, Microsoft), slowly decreasing (3M, Boeing, Du Pont, GE, Intel), or reversely increasing rapidly (McDonald's) or slowly (Caterpillar, Coca Cola, Exxon Mobil, Home Depot), or remains quite stable in time (Johnson \& Johnson, J. P. Morgan, United Technologies, Walt Disney). Most of the series show strong seasonal patterns (hence the choice of the DCCA and DMCA techniques which are constructed for such series) mainly connected to the end of the year but also some stronger patterns as for Home Depot. The examined dataset thus provides a complex selection of various dynamic behaviors.
Before we get to the estimated values of Hurst exponent and thus to the type of memory in question, we first test whether the analyzed series are in fact power-law correlated. In Tab. \ref{H}, we present the testing statistics as well as the corresponding $p$-values for the rescaled range and rescaled variance tests as described in the previous section. Apart from two cases (Coca Cola and IBM), the power-law correlations are reported for all series (the null hypothesis is rejected by at least one of the tests at at least 10\% level). It needs to be stressed that levels of the optimal $q$ parameter climb high for all and very high for some cases, sometimes taking into consideration as much as 372 lags of the covariance function (here specifically for IBM). This only strengthens the claim that the analyzed Google series are long-term correlated. This is due to the fact that taking into account already tens of lags practically means considering long-term memory, even more so for hundreds of lags.
Tab. \ref{H} also reports the estimated Hurst exponents which are further supported by Figs. \ref{fig_scaling1} and \ref{fig_scaling2}. In the figures, we report an evident power-law scaling of the fluctuation functions according to Eq. \ref{HH}. For all series, the scaling is very stable and the estimated Hurst exponents are thus reliable. Tab. \ref{H} shows that Hurst exponents vary between 0.8 and 1.1. The Google searches are thus strongly persistent for all the analyzed series. Even though the memory is very strong for these series, Hurst exponents still remain below 1.5 which implies that the series stay mean reverting. In the DFA context, this means that even though the online queries series tend to wander away from the long-term trend, they always return to it and they never explode. The fact that the series remain on the edge of stationarity and non-stationarity (around $H=1$) only highlights the need for a careful treatment of such series in multivariate settings which are standardly applied in the empirical literature.
To further illustrate the usefulness of the presented results, we reinvestigate the relationship between Google searches, traded volume and volatility. The traded volume for each component stock of the DJIA index is directly available at finance.yahoo.com as well as are the open, close, high and low prices. We utilize the provided information and construct volatility series using the Garman-Klass variance estimator \cite{Garman1980} defined as
\begin{equation}
\widehat{\sigma}^2_{GK,t}=\frac{(\log(H_t/L_t))^2}{2}-(2\log2-1)(\log(C_t/O_t))^2
\end{equation}
where $H_t$ and $L_t$ are daily highs and lows, respectively, and $C_t$ and $O_t$ are daily closing and opening prices, respectively. The estimator possesses very good statistical properties and serves as an excellent choice without a need of using high-frequency data \cite{Chou2010}. We study a logarithmic transformation of both $\widehat{\sigma}^2_{GK}$ and the traded volume series which is a standard procedure in the applied literature. The transformation of the original variance series allows us to comment on both variance and volatility as the logarithmic variance becomes just twice the logarithmic volatility.
We examine the correlations between Google searches, traded volume and volatility at various scales using the DCCA and DMCA coefficients. For the DCCA coefficient, we study the correlation between the searches and traded volume, and between the searches and volatility for scales between 10 and 250 with a step of 10. For the DMCA coefficient, we use moving window lengths between 11 and 251 with a step of 10 as well. This way, we obtain comparable results using these methodologies.
Figs. \ref{fig1} and \ref{fig2} depict the results for variance and traded volume, respectively, for both methods. Only significant correlations with the $p$-value below 0.10 are reported, the insignificant ones are set to zero. We find several interesting results. First, the DMCA method reports more stable results with more significant coefficients. This is well in hand with the numerical results presented by Kristoufek \cite{Kristoufek2014,Kristoufek2014a}. Second, the correlations for traded volume are in general higher than the ones for volatility. Third, a majority of significant correlations occur at the lower scales. There thus seems to be rather short-term or medium-term relationship between the online searches and the examined financial indicators. In the long-term, only few correlations are identified as significant. And fourth, the level of correlations varies considerably across stock titles. There thus seems to be no universality in the relationship between the searches, and volatility and volume. Tables \ref{rho1} and \ref{rho2} further illustrate the heterogeneity of the results. There, we present the average DCCA and DMCA coefficients across scales together with their significance level. The above mentioned results are supported. First, the significance, level and sign of the correlations vary widely. Second, the DMCA procedure delivers more significant results. And third, the correlations are higher for volume than for volatility. Nevertheless, many of the significant correlations are still below a level of 0.05 and practically all the correlation coefficients fall between -0.2 and 0.2. The correlations are thus very weak even if found statistically significant.
There are still some interesting results mainly connected to various signs of the correlations. For example Microsoft shows some unorthodox behavior for volatility. A positive relationship is usually reported, whereas the search queries for Microsoft are negatively correlated with volatility. Conversely, traded volume shows a positive correlation. It thus seems that general interest in Microsoft is mainly tied with positive news which stabilize the stock price rather than with negative news that would make the price more volatile. Similar dynamics is found for Johnson \& Johnson. The only stock which gives insignificant results for both financial quantities is Merck. Other stocks show either positive and thus expected correlations or only weak negatives ones.
\section{Discussion and conclusions}
We have analyzed the power-law correlations in the online search queries for the DJIA stock components. By reconstructing the daily Google search series, we have been able to obtain enough observations for a valid analysis of long-range dependence. Using the combination of the rescaled range and rescaled variance tests and the detrended fluctuation analysis, we have shown that the online searches are indeed power-law correlated. Importantly, the level of long-term memory is very high with Hurst exponents around unity (between 0.8 and 1.1) for all the analyzed stocks. Such results suggest that the finance-related online searches have similar dynamic properties to stock variance and traded volume which are themselves power-law correlated. The information flow coming into the stock markets evidently enters the general public interest and keeps it for quite long periods and its dissipation is thus not immediate. The fact that the online searches and implied attention are usually attributed to retail and small investors, such information and attention dissipation fits into the picture of a small investor using the information for decision-making in a longer term. Such persistent dynamics of the series might also arise from indecisiveness of the small investors which would think twice before investing into a specific stock. Online searches then cluster and keep their level for longer time intervals. The results remain fascinatingly universal across the analyzed stocks. Even though the global dynamics of the series is very heterogenous with different speeds of trends or various volatility levels, they all remain strongly persistent with a smooth scaling of fluctuations.
In addition, we have studied the relationship between Google searches, traded volume and volatility using the recently proposed DCCA and DMCA coefficients. The results primarily suggest that there is no universal relationship between the online search queries and the analyzed financial measures. Even though we confirm positive correlation for a majority of pairs, there are several pairs with insignificant or even negative correlations. Further, the correlations vary quite strongly across scales. The online searches have thus retained their potential for financial modeling and various applications but our findings suggest that one needs to carefully study each stock or asset separately as the usefulness of the queries can fluctuate considerably. The reported results do not necessarily contradict some previous findings which find statistically significant connections \cite{Bank2011,Bordino2012,Vlastakis2012,Preis2013} or time varying correlations \cite{Curme2014}. However, we stress that there seems to be no universal and global relationship between the online searches and relevant financial variables (traded volume and volatility).
Our results open an interesting area of further research of the topic. First, the power-law properties of the correlation structure might be observed also in different types of search queries, not necessarily only for stocks or financial markets in general. This would show how information or information seeking dissipates in time and how such behavior connects to other specific phenomena of the relevant time series. Second, the results indicate that the online searches are strongly persistent and on the edge of (non-)stationarity. Such characteristic implies that the simple correlation studies reported in the literature should take this property into consideration as an inappropriate analysis of persistent data using tools for short-range dependent series can produce spurious and in turn misleading results. And third, knowing the basic dynamic properties of the series helps to construct the forecasting models which are of high interest for practitioners, specifically in risk management.
\section*{Acknowledgements}
Support from the Czech Science Foundation under project No. 14-11402P and
the Research Councils UK via Grant EP/K039830/1 is gratefully acknowledged. I would also like to thank Tomas Vakrman for help with a dataset collection. Google data are registered trademarks of Google Inc., used with permission.
\section*{References}
|
\section*{Introduction}
Optical trapping of atoms is essential to realize atomic Bose-Einstein condensates \cite{bagnato1991}. Similarly, the 2D spatial confinement of excitons or exciton-polaritons by inhomogeneous strains \cite{balili2007}, natural defects and potential fluctuations \cite{lai2004,sanvitto2009,degiorgi2014,nguyen2014}, nanofabrication \cite{kaitouni2006,lai2007,bajoni2008,wertz2010,wertz2012,maragkou2010,zajac2012}, and optical potentials \cite{manni2011,tosi2012,cristofolini2013,dreismann2014} in a semiconductor heterostructure/microcavity is considered conducive for the formation and control of the condensates of these quasiparticles. These dynamic condensates can form a meta-stable state in a finite-momentum state \cite{lai2007,nelsen2013} and multiple spatial modes \cite{krizhanovskii2009}. The optical visualization of macroscopic interacting quantum states in solid-state systems has been demonstrated in condensates of excitons (bound electron-hole pairs) and kindred quasiparticles. Such light--matter hybrid condensates are typically formed at cryogenic temperatures in a photoexcited density much below the Mott transition where constituent quasiparticles can be considered boson-like. Optically defined potentials enabled by the effective repulsive interactions of polaritons result in further real-time manipulation of these light-matter fluids in a steady state \cite{tosi2012,cristofolini2013,dreismann2014}. On the other hand, relaxation oscillations with $\sim$10--50 ps period caused by the interplay between reservoir feeding and Bose stimulation have been reported in localized condensates in natural potential fluctuations \cite{degiorgi2014}.
In this study, we report the optical visualization of dynamic quantized states in a highly photoexcited microcavity at \emph{room temperature}. Photoexcitation creates mainly free pairs of electrons and holes in a III-V-based semiconductor quantum well (QW) as a result of thermal ionization at high temperature \cite{chemla1985,colocci1990a}. In the highly photoexcited microcavity studied here, the nonlinearly photomodulated refractive index results in a sizable effective cavity resonance shift, which enabls optically induced confinement. In an optical confinement initiated by spatially modulated nonresonant ps pulse excitation, sequential multiple-pulse lasing commences at several quantized energy levels. The transverse optical modes in a spatially modulated refractive index can also be understood as an optically induced potential for a fictitious quasiparticle (see Supplemental Material), e.g., correlated \emph{e-h} pairs (\emph{cehp}) resulting from the effective coupling to the cavity light field. Multiple-pulse lasing from discrete states in a harmonic confinement (potential) manifests as a result of the time- and energy-dependent competition between the gain and reservoir carrier relaxation.
The Fabry--P\'erot microcavity sample consists of a $\lambda$ GaAs cavity layer containing three sets of three InGaAs/GaAs QWs embedded within GaAs/AlAs distributed Bragg reflectors (DBRs) (see also Supplemental Material). The sample is nonresonantly excited by a 2-ps pulse pump laser at $E_p$ = 1.58 eV at room temperature. The pump energy is about 250 meV above the QW band gap ($E_g' \approx$ 1.33 eV) and 170 meV above the cavity resonance ($E_c \approx $ 1.40--1.41 eV). The high-density \emph{e-h} plasmas of a density $\approx1-5 \times 10^{12} \, \text{cm}^{-2}$ per QW per pulse are formed momentarily after \emph{non-resonant} pulse excitation as a result of rapid ($<$10 ps) energy dissipation through optical phonons. The radiative recombination rate of these \emph{e-h} carriers in the reservoir is suppressed because the cavity resonance $E_c$ is detuned to $\sim$ 70--80 meV above the QW bandgap $E_g'$, i.e., the $e1hh1$ transition between the first quantized electron and (heavy-)hole states in a QW. Below the lasing threshold, the high-density \emph{e-h} plasmas in the microcavity are subject to non-radiative loss with a long decay time ($>$500 ps) (Supplemental Fig.~S1). Therefore, the chemical potential of the \emph{e-h} plasmas ($\mu_0$) appears to be stationary within $\sim$100 ps after pulse excitation. The bare cavity resonance $E_c$ of the sample studied here is close to $E_g''$, the $e2hh2$ transition between the second quantized electron and hole levels of InGaAs/GaAs QWs.
When $\mu_0$ advances toward $E_c$, the average refractive index near $E_c$ can be considerably modified, and this results in a sizable blueshift of the effective cavity resonance ($E'_c$) and consequent emission energy (Supplemental Fig.~S2a). In general, $E'_c$ increases with the photoexcited density and can be seen as an effective potential $V$ for \emph{cehp}s (Supplemental Material section S3). Therefore, a quasi-stationary confining potential $V(r)$ for \emph{cehps} can be established by a ring-shaped spatial distribution of photoexcited carrier density. In our experiments, we use a \emph{double-hump-shaped} beam to fixate the orientation of the optically defined potential (Fig. \ref{fig:rk_spectra}a--b).
Real-space (r-space) imaging spectra provide direct visualization of the optical potential. Fig.~\ref{fig:rk_spectra}c--d shows the r-space imaging spectra of a narrow cross-sectional stripe across the trap. A nearly parabolic potential well of $\sim$10 meV across 3 $\mu$m is revealed. Such a quasi-1D harmonic potential, $V(x) = 1/2 \ \alpha x^2$, is spontaneously formed under 2 ps pulse excitation. The resultant radiation appears in-between the two humps, a few micrometers away from the pump spot. The r-space intensity profiles agree with the probability distributions of the quantized states of a harmonic oscillator (harmonic states). The deviations of the actual potential from a perfectly harmonic trap result in slightly asymmetric luminescence intensity distributions. These standing wave patterns form in the self-induced harmonic optical potential when a macroscopic coherent state emerges. At a critical density, the occupation number of \emph{cehp}s ($n_i$) in these quantized harmonic states approaches unity when the conversion from the reservoir ($N_{eh}$) overcomes the decay of the harmonic states. Consequently, when a stimulated process ($\propto N_{eh} n_i$) prevails, $n_i$ and the resultant radiation increase nonlinearly by a few orders of magnitude with the increasing $N_{eh}$.
\begin{figure}[htb]
\centering
\includegraphics[width= 0.5 \textwidth]{fig2_ek_size.pdf}
\caption{\label{fig:ek_size}\textbf{Quantized states in optically controlled confining potentials.} K-space imaging spectra under below-threshold (a--b) and above-threshold (d--e) for two double-hump-shaped pump beams with peak-to-peak distances of 5 $\mu$m and 3 $\mu$m, respectively. For comparison, the k-space imaging spectra under a flat-top pump beam are shown in (c) and (f). The corresponding r-space images and spectra are shown in Supplemental Fig. S1.
}
\end{figure}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.6\textwidth]{fig3_dynamics.pdf}
\caption{\label{fig:dynamics}\textbf{Dynamics.} (a) Time-dependent luminescence in k-space at $P = 1.0 $, $1.1$, $1.2$ and $1.6 \, P_{th}$. The $E_3$, $E_2$ and $E_1$ states appear sequentially with the increasing pump flux. The rise times decrease with the increasing pump flux for all states. (b) Time-dependent spectra in r-space. The false color represents normalized intensities.
}
\end{figure*}
An even more regular pattern appears in k-space imaging spectra (Fig.~\ref{fig:rk_spectra}e--f). The $E$ vs. $k_\parallel$ dispersion measured below the threshold allows the direct measurement of the effective mass $m^{*} = 3\times 10^{5} \, m_e$, where $m_e$ is the electron rest mass. Moreover, the strength of the optically defined harmonic potential $\alpha$ can be tuned through a variation in the spatial distance between two humps of the pump beam (Fig.~\ref{fig:ek_size}). The energy quantization ($\hbar \omega$) varies with $\sqrt{\alpha/m^*}$, while the emission patterns evolve into the probability distributions for a particle with an effective mass $m^*$ in $V(x)$ (Fig.~\ref{fig:ek_size} and Supplemental Fig.~S3).
\begin{figure}[htb]
\centering
\includegraphics[width=0.35\textwidth]{fig4_spectra_analysis.pdf}
\caption{\label{fig:density}\textbf{Density dependence.} (a) Temporally and spectrally integrated emission flux vs. pump flux. All three modes display non-linear increases in intensity by more than two orders of magnitude, and saturate at $1.1$, $1.2$ and $1.3 \, P_{th}$, respectively. (b) Peak energy (solid shapes) and linewidths 2$\Delta E$ (error bars) vs. pump flux. These states spectrally blue shift by 1 to 4 meV. The spectral linewidths ($\Delta E$) and pulsewidths ($\Delta t$) are reciprocal with a product of $\Delta E \times \Delta t \approx 4 \hbar$ ($\hbar$) for $E_3$ and $E_1$ ($E_2$), which is closed to the uncertainty (Fourier-transform) limit. (c) Rise time vs. pump flux for the three states $E_1$ (blue), $E_2$ (red) and $E_3$ (black). The error bar represents 2$\Delta t$.}
\end{figure}
Temporally, these harmonic states emerge sequentially and display distinct density-dependent dynamics. In Fig. \ref{fig:dynamics}a, we study the time evolutions of three harmonic states in k-space. The corresponding time-integrated imaging spectra in k-space are shown in Supplemental Fig.~S4. The energy relaxation of these three states is revealed in the time-resolved spectra in Fig.~\ref{fig:dynamics}b. At a critical photoexcited density, the high-energy $E_3$ state arises $\sim$ 25 ps after pulse excitation and lasts for $\sim$ 20 ps. The corresponding pump flux is defined as the threshold ($P_{th}$). With the increasing pump flux, the $E_2$ state emerges $\sim$ 25 ps after $E_3$ at 1.1 $P_{th}$, whereas the \emph{ground} $E_1$ state appear 50 ps after $E_2$ at 1.2 $P_{th}$. In the optically induced harmonic confinement studied here, the effective cavity resonance $E'_c$ decreases with time as a result of the decay of reservoir carriers. However, the conversion from the reservoir to specific confined $E_i$ state can be efficient only when $E'_c$ is resonant with $E_i$. Such a temporally decrease in $E'_c$ results in a time- and energy-dependent conversion efficiency for these confined harmonic states (Fig. S1) and consequent multiple-pulse lasing above the critical density threshold.
We further analyze the emission flux and energy of these harmonic states by using time-integrated spectra measured with increasing photoexcited density (Fig. \ref{fig:density} and Supplemental Fig.~S2). Far below the threshold ($P < 0.4 \ P_{th}$), emission is dominated by luminescence from GaAs spacer layers. When the pump flux is increased, the emissions from the \emph{cehp} states become increasingly dominant, and the $E_3$ state eventually lases at the threshold. The emission fluxes of all three states increase nonlinearly by more than two orders of magnitude across a threshold and then reach a plateau at a saturation density (Fig.~\ref{fig:density}a). On the other hand, the emission energy of these three states increases to a constant with the increasing pump flux (Fig.~\ref{fig:density}b). The energy spacing $\hbar \omega$ only increases slightly with density. The spectral linewidths increase slightly for the $E_1$ state but by about a factor of 10 for the $E_3$ state. Next, we study the density-dependent dynamics. Fig.~\ref{fig:density}c shows the rise times and pulse durations for $E_3$, $E_2$, and $E_1$. The product of the variances of the spectral linewidth ($\Delta E$) and the pulse duration ($\Delta t$) is found to be close to that of a transform-limited pulse: $\gtrsim 4 \hbar$ and $\approx 1 \hbar$ for the $E_3/E_1$ and $E_2$ states, respectively. These harmonic states are macroscopically coherent states with finite phase and intensity fluctuations induced by interactions.
We use a rate-equation model to describe the dynamic formation of quantized states in an optically defined harmonic confinment (Supplemental Material Sec. S4Supplemental Material). This phenomenological model reproduces qualitatively the dynamics and integrated emission flux of the harmonic states when photoexcited density is varied (Supplemental Figs.~S5, S6, and S7). Non-equilibrium polariton condensates have been modeled by a modified Gross--Pitaevskii (GP) (or complex Ginzberg--Landau [cGL]) equation that accounts for the finite lifetime of polaritons \cite{szymanska2006,wouters2007}. However, cGL-type equations are inapplicable for the multiple dynamic states examined in this study. Additionally, the formation of a BEC-like condensate that underpins the GP- or cGL-type equation is not necessarily justified in our room-temperature experiments.
Transverse light-field patterns and confined optical modes have been identified in nonlinear optical systems \cite{mandel2004}, vertical-cavity surface-emitting lasers (VCSELs) \cite{chang-hasnain1991,zhang2004}, and microscale photonic structures \cite{reithmaier1997}. In principle, the multiple transverse mode lasing in a high-density \emph{e-h}-plasma described in the present study can be modeled by a self-consistent numerical analysis with Maxwell-Bloch equations developed for conventional semiconductor lasers \cite{koch1995,sarzala2012}, provided that the strong optical nonlinearities induced by Coulomb many-body effects, such as screening, bandgap renormalization, and phase-space filling, are all considered. For example, one can consider the formation of index-guided multiple transverse modes as a result of an optically-induced refractive index reduction ($\Delta n_c(x)$). The cavity resonance shift ($\delta E$) can be estimated from $\delta E_c/E_c = - \Delta n_c/n_c$, where $n_c$ is the effective refractive index averaged over the longitudinal cavity photon mode which spans over $\sim\mu$m in the growth direction. Therefore, a cavity resonance shift $\delta E \sim$10 meV (Figs. \ref{fig:rk_spectra} and \ref{fig:ek_size}) corresponds to $|\Delta n_c/n_c| \sim$ 1\%. Such a significant refractive index is probable with resonance-enhance optical nonlinearity from the \emph{e-h} correlation \cite{schmitt-rink1986a,campi1998,tanguy1999,kamide2012} or carrier-induced change in refractive index at a high carrier density (of $\sim10^{19}$ cm$^{-3}$ or more) \cite{lee1986,bennett1990,huang1998}. To uncover the microscopic formation mechanisms of such a sizable spatial modulation in refractive index or equivalent effective harmonic confinement under pulse excitation, further characterizing the photoexcited density distribution with the use of other ultrafast spectroscopic techniques, such as a pump-probe spectroscopy, is necessary.
We identify sequential multiple $\sim$10 ps pulse lasing in an optically induced harmonic confinement in a semiconductor microcavity at room temperature. Laser radiation emerges at the quantized states of an optically induced harmonic confinement. The lasing frequency, rise time, pulse width, and radiation angle can be controlled through a variation in the photoexcited density or optical pump spot dimensions in real time. The sample has a composition structure similar to those used for studies of polariton condensates/lasers and the widely used VCSELs. In the highly photoexcited microcavity studied here, harmonic confinement can be optically induced and controlled as a result of photomodulated refractive index changes even in the weak-coupling plasma limit. Our demonstration of macroscopic harmonic states in a room-temperature optically induced harmonic confinement improves our understanding of nonlinear laser dynamics and should stimulate studies on emergent ordered states near the Fermi edge of a high-density \emph{e-h} plasma in a semiconductor cavity \cite{nozieres1985,keeling2005,yamaguchi2013}.
\begin{acknowledgments}
We thank Cheng Chin, Mark Dykman, Brage Golding, Peter B. Littlewood, John A. McGuire, David W. Snoke and Carlo Piermarocchi for the discussions. This work was supported by NSF grant DMR-0955944 and J. Cowen endowment at Michigan State University.
\end{acknowledgments}
|
\section{Introduction}
Bekenstein and Hawking showed that the entropy of a black hole is proportional
to the area of the
horizon\cite{IN-Bekenstein:1972tm,IN-Bekenstein:1973ur,IN-Hawking:1976de}.
Although all the evidences suggest that the Bekenstein-Hawking entropy is
truly the thermodynamic entropy, the statistical origin of the black-hole
entropy has not yet been fully understood. It seems that an unavoidable
candidate for the statistical origin is the entropy of the thermal atmosphere
of the black hole, which can also be thought of as the entanglement entropy
across the horizon\cite{IN-Bombelli:1986rw}.
To be generic, we will consider a spherically symmetric background metric of
the form
\begin{equation}
ds^{2}=f\left( r\right) dt^{2}-\frac{dr^{2}}{f\left( r\right)
-r^{2}\left( d\theta^{2}+\sin^{2}\theta d\phi^{2}\right) ,
\label{eq:SchwarzschildMetric
\end{equation}
where $f\left( r\right) $ has a simple zero at $r=r_{h}$ with $f^{\prime
}\left( r_{h}\right) $ being finite and nonzero. The vanishing of $f\left(
r\right) $ at point $r=r_{h}$ indicates the presence of an event horizon.
Thus, the atmosphere entropy of the black hole with the metric $\left(
\ref{eq:SchwarzschildMetric}\right) $ can be expressed in the for
\begin{equation}
S=\sum_{l=0}^{\infty}\left( 2l+1\right) \int dn\left( \omega,l\right)
s\left( \frac{\omega}{T}\right) , \label{eq:EntropySum
\end{equation}
where $\omega$ is the Killing energy associated with $t$, $l$ is the angular
momentum, $T$ is the Hawking temperature of the black hole, $n\left(
\omega,l\right) $ is the number of one-particle states not exceeding $\omega$
with fixed value of angular momentum $l$, and $s\left( \frac{\omega
{T}\right) $ is the thermal entropy per mode. However, the entropy diverges
when we attempt to sum eqn. $\left( \ref{eq:EntropySum}\right) $ over all
the modes. There are two kinds of divergences. The first one is due to
infinite volume of the system, which has to do with the contribution from the
vacuum surrounding the system at large distances and is of little relevance
here. The second one arises from the infinite volume of the deep throat region
near the horizon. In order to regulate the two divergences, t' Hooft
\cite{IN-'tHooft:1984re}\ proposed the brick wall model for a scalar field
$\phi$, where two brick wall cutoffs are introduced at some small distance
$r_{\varepsilon}$ from the horizon and at a large distance $L\gg r_{h}$,
\begin{equation}
\phi=0\text{ \ at }r=r_{h}+r_{\varepsilon}\text{ and }r=L\text{.}
\label{eq:DirichletBoundary
\end{equation}
The minimally coupled scalar field satisfies the Klein-Gordon equatio
\begin{equation}
\left( \frac{\nabla^{2}}{\hbar^{2}}-m^{2}\right) \phi=0.
\label{eq:KleinGordon
\end{equation}
Since WKB approximation is reliable as long as the black hole's mass $M\gg
1$(in Planck units), t' Hooft took the ansatz for $\phi$
\begin{equation}
\phi=\exp\left[ -\frac{i}{\hbar}\omega t+\frac{i}{\hbar}\int p_{r}dr\right]
Y_{lm}\left( \theta,\phi\right) ,
\end{equation}
and solved for $p_{r}$ to the WKB leading order. Define the radial wave number
$k\left( r,l,\omega\right) $ b
\begin{equation}
k\left( r,l,\omega\right) =\left\vert p_{r}\right\vert , \label{eq:K
\end{equation}
as long as $p_{r}^{2}\geq0$, and $k\left( r,l,\omega\right) =0$ otherwise.
With the two Dirichlet boundaries $\left( \ref{eq:DirichletBoundary}\right)
$, $n\left( \omega,l\right) $ can be expressed a
\begin{equation}
n\left( \omega,l\right) =\frac{1}{\pi\hbar}\int_{r_{h}+r_{\varepsilon}
^{L}k\left( r,l,\omega\right) dr. \label{eq:N
\end{equation}
It would appear that the entanglement entropy is sensitive to the
ultraviolet(UV) behavior of quantum fields, where quantum gravity effects
become important. Even though there is still no complete and consistent
quantum theory of gravity, one could still rely on effective models to study
the UV behavior of the entanglement entropy. For example, this issue was
studied in \cite{IN-Chang:2003sa,IN-Jacobson:2007jx} in the context of the
possible modifications of the standard dispersion relation. On the other hand,
various theories of quantum gravity, such as string theory, loop quantum
gravity and quantum geometry, predict the existence of a minimal
length\cite{IN-Townsend:1977xw,IN-Amati:1988tn,IN-Konishi:1989wk}. The
generalized uncertainty principle(GUP)\cite{IN-Kempf:1994su} is a simple way
to realize this minimal length. In
\cite{IN-Li:2002xb,IN-Kim:2006rx,IN-Kim:2007bx,IN-Kim:2007nh}, the authors
considered the generalized uncertainty relatio
\begin{equation}
\Delta x\Delta p\geq\hbar+\frac{\lambda}{\hbar}\left( \Delta p\right) ^{2},
\label{eq:LiForm
\end{equation}
which gives a minimal length, $2\sqrt{\lambda}$. As a consequence, the number
of quantum states should be changed to
\begin{equation}
\frac{d^{3}xd^{3}p}{\left( 2\pi\hbar\right) ^{3}\left( 1+\lambda
p^{2}\right) ^{3}},
\end{equation}
where $p^{2}=p_{i}p^{i}$. Therefore, they had for massless particle
\begin{equation}
n\left( \omega,l\right) =\frac{1}{\pi\hbar}\int_{r_{h}+r_{\varepsilon}
^{L}\frac{k\left( r,l,\omega\right) }{\left( 1+\lambda\omega^{2}/f\right)
^{3}}dr, \label{eq:NLamda
\end{equation}
where $k\left( r,l,\omega\right) $ is given by eqn. $\left( \ref{eq:K
\right) $. The all order generalized uncertainty relation was also considered
in \cite{IN-Kim:2007if}. It is found there that the artificial cutoff
parameter in the brick wall model located just outside the horizon can be
avoided if GUP is considered. Alternatively, the modified Klein-Gordon
equation for $\phi$ was considered in the framework of Horava-Lifshitz gravity
and GUP in \cite{IN-Eune:2010kx}. The WKB leading term of $p_{r}$ for the
modified Klein-Gordon equation was obtained and $n\left( \omega,l\right) $
was given by eqn. $\left( \ref{eq:N}\right) $. The entanglement entropy was
then calculated and the result is consistent with previous studies.
Furthermore, GUP should modify the Hawking temperature $T$ in eqn. $\left(
\ref{eq:EntropySum}\right) $ as well as $n\left( \omega,l\right) $. Indeed,
the GUP deformed Hamilton-Jacobi equation in curved spacetime have been
introduced and the corrected Hawking temperatures have been derived in
\cite{IN-Chen:2013pra,IN-Chen:2013tha,IN-Chen:2013ssa,IN-Chen:2014xsa,IN-Mu:2015qta
. The GUP corrections to the Hawking temperature were found to depend not only
on the black hole's mass but the mass, the energy, and the angular momentum of
emitted particles as well. In \cite{IN-Mu:2015qta}, we derived the deformed
Klein-Gordon and Dirac equation incorporating the GUP form proposed in
\cite{DHM-Hossenfelder:2003jz,DHM-Hassan:2002qk}. The GUP modified Hawking
temperatures for scalars and fermions were then obtained and the black hole's
evaporation was also discussed in \cite{IN-Mu:2015qta}. In this paper, we will
investigate the entropy of the thermal atmosphere of the black hole in the
context of \cite{IN-Mu:2015qta}. Taking GUP corrections to both $n\left(
\omega,l\right) $ and Hawking temperature $T$ into consideration, we will
calculate the entanglement entropy of a massless scalar field via eqn.
$\left( \ref{eq:EntropySum}\right) $ in the brick wall model.
The organization of this paper is as follows. In section \ref{Sec:DHM}, we
review some results of \cite{IN-Mu:2015qta}, which are necessary for
calculating the entanglement entropy. In section \ref{Sec:EEBH}, the
entanglement entropy of a massless scalar field near the horizon is calculated
in the brick wall model. Section \ref{Sec:Con} is devoted to our discussion
and conclusion. In this paper, we take Geometrized units $c=G=1$, where the
Planck constant $\hbar$ is square of the Planck Mass $m_{p}$. For simplicity,
we assume that the emitted particles are massless and neutral.
\section{Deformed Hamilton-Jacobi Method}
\label{Sec:DHM}
In \cite{IN-Mu:2015qta} and this paper, we consider an effective model of the
GUP in one dimensional quantum mechanics given
by\cite{DHM-Hossenfelder:2003jz,DHM-Hassan:2002qk}
\begin{equation}
L_{f}k(p)=\tanh\left( \frac{p}{M_{f}}\right) , \label{tanh1
\end{equation
\begin{equation}
L_{f}\omega(E)=\tanh\left( \frac{E}{M_{f}}\right) , \label{tanh2
\end{equation}
where the generators of the translations in space and time are the wave vector
$k$ and the frequency $\omega$, $L_{f}$ is the minimal length, and $L_{f
M_{f}=\hbar$. From eqn. $\left( \ref{tanh1}\right) $, it is noted that
although one can increase $p$ arbitrarily, $k$ has an upper bound which is
$\frac{1}{L_{f}}$. The upper bound on $k$ implies that particles could not
possess arbitrarily small Compton wavelength $\lambda=2\pi/k$ and that there
exists a minimal length $\sim L_{f}$. The quantization in position
representation $\hat{x}=x$ leads to
\begin{equation}
k=-i\partial_{x},\text{ }\omega=i\partial_{t}.
\end{equation}
In the $\left( 3+1\right) $ dimensional flat spacetime, the relations
between $\left( p_{i},E\right) $ and $\left( k_{i},\omega\right) $ can
simply be generalized t
\begin{align}
L_{f}k_{i}(p) & =\tanh\left( \frac{p_{i}}{M_{f}}\right)
,\label{eq:wavevector}\\
L_{f}\omega(E) & =\tanh\left( \frac{E}{M_{f}}\right) ,
\label{eq:frequency
\end{align}
where one has for $\vec{k}$ in the spherical coordinate
\begin{equation}
\vec{k}=-i\left( \hat{r}\frac{\partial}{\partial r}+\frac{\hat{\theta}
{r}\frac{\partial}{\partial\theta}+\frac{\hat{\phi}}{r\sin\theta
\frac{\partial}{\partial\phi}\right) .
\end{equation}
Expanding eqn. $\left( \ref{eq:wavevector}\right) $ and eqn. $\left(
\ref{eq:frequency}\right) $ for small arguments to the third order gives the
energy and momentum operator in position representatio
\begin{gather}
E=i\hbar\partial_{t}\left( 1-\frac{\hbar^{2}}{M_{f}^{2}}\partial_{t
^{2}\right) ,\\
\vec{p}=\frac{\hbar}{i}\left[ \hat{r}\left( \partial_{r}-\frac{\hbar
^{2}\partial_{r}^{3}}{M_{f}^{2}}\right) +\hat{\theta}\left( \frac
{\partial_{\theta}}{r}-\frac{\hbar^{2}\partial_{\theta}^{3}}{r^{3}M_{f}^{2
}\right) +\hat{\phi}\left( \frac{\partial_{\phi}}{r\sin\theta}-\frac
{\hbar^{2}\partial_{\phi}^{3}}{r^{3}\sin^{3}\theta M_{f}^{2}}\right) \right]
,
\end{gather}
where we also omit the factor $\frac{1}{3}$. Substituting the above energy and
momentum operators into the energy-momentum relation, the deformed
Klein-Gordon equation satisfied by the massless scalar field is
\begin{equation}
E^{2}\phi=p^{2}\phi, \label{eq:KGFD
\end{equation}
where $p^{2}=\vec{p}\cdot\vec{p}.$ Making the ansatz for $\phi$
\begin{equation}
\phi=\exp\left( \frac{iI}{\hbar}\right) ,
\end{equation}
and substituting it into eqn. $\left( \ref{eq:KGFD}\right) $, one expands
eqn. $\left( \ref{eq:KGFD}\right) $ in powers of $\hbar$ and then finds that
the lowest order gives the deformed scalar Hamilton-Jacobi equation in the
flat spacetime
\begin{gather}
\left( \partial_{t}I\right) ^{2}\left( 1+\frac{2\left( \partial
_{t}I\right) ^{2}}{M_{f}^{2}}\right) -\left( \partial_{r}I\right)
^{2}\left( 1+\frac{2\left( \partial_{r}I\right) ^{2}}{M_{f}^{2}}\right)
-\frac{\left( \partial_{\theta}I\right) ^{2}}{r^{2}}\left( 1+\frac{2\left(
\partial_{\theta}I\right) ^{2}}{r^{2}M_{f}^{2}}\right) \nonumber\\
-\frac{\left( \partial_{\phi}I\right) ^{2}}{r^{2}\sin^{2}\theta}\left(
1+\frac{2\left( \partial_{\phi}I\right) ^{2}}{r^{2}\sin^{2}\theta M_{f}^{2
}\right) =0, \label{eq:HJFD
\end{gather}
which is truncated at $\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $. The
Hamilton-Jacobi equation in the metric $\left( \ref{eq:SchwarzschildMetric
\right) $ can be obtained from that in flat spacetime by making replacements
$\partial_{r}I\rightarrow\sqrt{f\left( r\right) }\partial_{r}I$ and
$\partial_{t}I\rightarrow\frac{\partial_{t}I}{\sqrt{f\left( r\right) }}$.
Therefore, the deformed Hamilton-Jacobi equation in flat spacetime, eqn.
$\left( \ref{eq:HJFD}\right) $, leads to the deformed Hamilton-Jacobi
equation in the metric $\left( \ref{eq:SchwarzschildMetric}\right) $, which
is to $\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $
\begin{gather}
\frac{\left( \partial_{t}I\right) ^{2}}{f\left( r\right) }\left(
1+\frac{2\left( \partial_{t}I\right) ^{2}}{f\left( r\right) M_{f}^{2
}\right) -f\left( r\right) \left( \partial_{r}I\right) ^{2}\left(
1+\frac{2f\left( r\right) \left( \partial_{r}I\right) ^{2}}{M_{f}^{2
}\right) -\frac{\left( \partial_{\theta}I\right) ^{2}}{r^{2}}\left(
1+\frac{2\left( \partial_{\theta}I\right) ^{2}}{r^{2}M_{f}^{2}}\right)
\nonumber\\
-\frac{\left( \partial_{\phi}I\right) ^{2}}{r^{2}\sin^{2}\theta}\left(
1+\frac{2\left( \partial_{\phi}I\right) ^{2}}{r^{2}\sin^{2}\theta M_{f}^{2
}\right) =0. \label{eq:HJCG
\end{gather}
Taking into account the Killing vectors of the background spacetime, we can
employ the following ansatz for the action
\begin{equation}
I=-\omega t+W\left( r,\theta\right) +p_{\phi}\phi, \label{eq:Ianzatz
\end{equation}
where $\omega$ and $p_{\phi}$ are constants and they are the energy and the
$z$-component of angular momentum of emitted particles, respectively.
Inserting eqn. $(\ref{eq:Ianzatz})$ into eqn. $(\ref{eq:HJCG}),$ we find that
the deformed Hamilton-Jacobi equation becomes
\begin{equation}
p_{r}^{2}\left( 1+\frac{2f\left( r\right) p_{r}^{2}}{M_{f}^{2}}\right)
=\frac{1}{f^{2}\left( r\right) }\left[ \omega^{2}\left( 1+\frac
{2\omega^{2}}{f\left( r\right) M_{f}^{2}}\right) -f\left( r\right)
\frac{L^{2}}{r^{2}}\right] , \label{eq:pr
\end{equation}
where we neglect terms higher than $\mathcal{O}\left( \frac{1}{M_{f}^{2
}\right) $. In the above equation, we define $p_{r}=\partial_{r}W$,
$p_{\theta}=\partial_{\theta}W$, $L^{2}=p_{\theta}^{2}+\frac{p_{\phi}^{2
}{\sin^{2}\theta}$, and $L^{2}=\left( l+1\right) l\hbar^{2}$ which is the
magnitude of the angular momentum of the particle. Solving eqn. $(\ref{eq:pr
)$ for $p_{r}$ to $\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $give
\begin{equation}
p_{r}^{\pm}=\pm\left( \frac{\Omega^{\frac{1}{2}}}{f\left( r\right)
-\frac{\Omega^{\frac{3}{2}}}{f^{2}\left( r\right) M_{f}^{2}}\right) ,
\label{eq:PrSolution
\end{equation}
where +/$-$ represent the outgoing/ingoing solutions and we define
\begin{equation}
\Omega=\omega^{2}\left( 1+\frac{2\omega^{2}}{f\left( r\right) M_{f}^{2
}\right) -f\left( r\right) \frac{\left( l+1\right) l\hbar^{2}}{r^{2}}.
\end{equation}
Using the residue theory for semi circles, we obtain for the imaginary part of
$W_{\pm}$ to $\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $
\begin{equation}
\operatorname{Im}W_{\pm}=\pm\frac{\pi\omega}{f^{\prime}\left( r_{h}\right)
}\left[ 1+\frac{2\left( l+1\right) l\hbar^{2}}{M_{f}^{2}r_{h}^{2}}\right]
.
\end{equation}
Taking both the spatial contribution and the temporal contribution into
account\cite{DHM-Akhmedova:2008au,DHM-Akhmedova:2008dz,DHM-Akhmedov:2008ru,DHM-Chowdhury:2006sk,DHM-Akhmedov:2006un
, one finds that the tunnelling rate of the particle crossing the horizon is
\begin{align}
\Gamma & \propto exp\left[ -\frac{1}{\hbar}\left( \operatorname{Im}\left(
\omega\Delta t\right) +\operatorname{Im}\oint p_{r}dr\right) \right]
\nonumber\\
& =exp\left\{ -\frac{4\pi\omega}{f^{\prime}\left( r_{h}\right) }\left[
1+\frac{2\left( l+1\right) l\hbar^{2}}{M_{f}^{2}r_{h}^{2}}\right] \right\}
. \label{eq:Gamma
\end{align}
This is the expression of Boltzmann factor with an effective temperature to
$\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $
\begin{equation}
T=\frac{T_{0}}{1+\Delta}, \label{HKT1
\end{equation}
where $T_{0}=\frac{\hbar f^{\prime}\left( r_{h}\right) }{4\pi}$ is the
original Hawking temperature and we defin
\begin{equation}
\Delta=\frac{\left( l+1\right) l\hbar^{2}}{M_{f}^{2}r_{h}^{2}}.
\label{eq:Delta
\end{equation}
.
\section{Entanglement Entropy of Black Hole}
\label{Sec:EEBH}
For scalar particles emitted in a wave mode labelled by energy $\omega$ and
$l,$ we find from eqn. $\left( \ref{eq:Gamma}\right) $ that
\cite{EEBH-Hartle:1976tp
\begin{align*}
& \left( \text{Probability for a black hole to emit a particle in this
mode}\right) \\
& =\exp\left( -\frac{\omega}{T}\right) \times(\text{Probability for a black
hole to absorb a particle in the same mode}),
\end{align*}
where $T$ is given by eqn. $\left( \ref{HKT1}\right) $. Neglecting
back-reaction, detailed balance condition requires that the ratio of the
probability of having $N$ particles in a particular mode with $\omega$ and $l$
to the probability of having $N-1$ particles in the same mode is $\exp\left(
-\frac{\omega}{T}\right) .$ One then follows the standard textbook procedure
to get the average number $n_{\omega,l}$ in the mod
\begin{equation}
n_{\omega,l}=n\left( \frac{\omega}{T}\right) ,
\end{equation}
where we defin
\begin{equation}
n\left( x\right) =\frac{1}{\exp x-1}.
\end{equation}
The von Neumann entropy for the mode i
\begin{equation}
s_{\omega,l}=\left( n_{\omega,L}+1\right) \ln\left( 1+n_{\omega,L}\right)
-n_{\omega,L}\ln n_{\omega,L}, \label{eq:EntropyForMode
\end{equation}
where $s_{\omega,l}$ is $s\left( \frac{\omega}{T}\right) $ in eqn. $\left(
\ref{eq:EntropySum}\right) $. The $s\left( x\right) $ is given b
\begin{equation}
s\left( x\right) =\frac{\exp x}{\exp x-1}\ln\left( \frac{\exp x}{\exp
x-1}\right) +\frac{\ln\left( \exp x-1\right) }{\exp x-1}.
\end{equation}
Define the radial wave number $k\left( r,l,\omega\right) $ b
\begin{equation}
k\left( r,l,\omega\right) =\left\vert p_{r}^{\pm}\right\vert ,
\end{equation}
as long as $p_{r}^{2}\geq0$, and $k\left( r,l,\omega\right) =0$ otherwise.
The $p_{r}^{\pm}$ is given by eqn. $\left( \ref{eq:PrSolution}\right) $.
Taking two Dirichlet conditions at $r=r_{h}+r_{\varepsilon}$ and $r=L$ in the
brick wall model into account, we find that the number of one-particle states
not exceeding $\omega$ with fixed value of angular momentum $l$ is given b
\begin{equation}
n\left( \omega,l\right) =\frac{1}{\pi\hbar}\int_{r_{h}+r_{\varepsilon}
^{L}k\left( r,l,\omega\right) dr.
\end{equation}
Thus, we get for the total entropy of radiatio
\begin{align}
S & =\sum\limits_{\omega,l,m}s_{\omega,l}=\int\left( 2l+1\right) dl\int
d\omega\frac{dn\left( \omega,l\right) }{d\omega}s_{\omega,l}\nonumber\\
& =-\frac{1}{\pi\hbar^{3}}\int d\left[ \left( l+1\right) l\hbar
^{2}\right] \int d\omega\frac{\partial s_{\omega,l}}{\partial\omega
\int_{r_{h}+r_{\varepsilon}}^{L}drk\left( r,l,\omega\right) .
\end{align}
Defining $u=\frac{\omega}{T_{0}}$ and expanding $s_{\omega,l}$ to
$\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right)
\begin{equation}
s_{\omega,l}\approx s\left( u\right) +s^{\prime}\left( u\right) u\Delta,
\end{equation}
we find that the entropy to $\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $
is
\[
S\approx-\frac{M_{f}^{2}r_{h}^{2}}{\pi\hbar^{3}}\int du\int_{r_{h
+r_{\varepsilon}}^{L}dr\int d\Delta\left[ s^{\prime}\left( u\right)
+\left( s^{\prime}\left( u\right) u\right) ^{\prime}\Delta\right] \left(
\frac{\Omega^{\frac{1}{2}}}{f\left( r\right) }-\frac{\Omega^{\frac{3}{2}
}{f^{2}\left( r\right) M_{f}^{2}}\right) ,
\]
where we use eqn. $\left( \ref{eq:Delta}\right) $ for $\Delta$ and eqn.
$\left( \ref{eq:PrSolution}\right) $ for $k\left( r,l,\omega\right) $.
Performing $\Delta$ integral which runs over the region where $\Omega>0$ give
\begin{equation}
S\approx\frac{2r_{h}^{2}T_{0}^{3}}{\pi\hbar^{3}}\int s\left( u\right)
du\int_{r_{h}+r_{\varepsilon}}^{L}dr\frac{r^{2}}{r_{h}^{2}}\frac{u^{2}
{f^{2}\left( r\right) }\left[ 1+\frac{4T_{0}^{2}u^{2}}{f\left( r\right)
M_{f}^{2}}-\frac{6T_{0}^{2}u^{2}}{5f\left( r\right) M_{f}^{2}}\frac{r^{2
}{r_{h}^{2}}\right] ,
\end{equation}
where the second term in the bracket comes from the GUP corrections to
$n\left( \omega,l\right) $ and the third term from the GUP corrections to
Hawking temperature $T$. Focusing on the divergent parts near horizon, we fin
\begin{align}
\int_{r_{h}+r_{\varepsilon}}^{L}dr\frac{r^{2}}{r_{h}^{2}}\frac{1}{f^{2}\left(
r\right) } & \sim\frac{1}{4\kappa^{2}r_{\varepsilon}}-\left( \frac
{1}{\kappa r_{h}}-\eta\right) \frac{\ln\kappa r_{\varepsilon}}{2\kappa
},\nonumber\\
\int_{r_{h}+r_{\varepsilon}}^{L}dr\frac{r^{2}}{r_{h}^{2}}\frac{1}{f^{3}\left(
r\right) } & \sim\frac{1}{4\kappa^{3}r_{\varepsilon}^{2}}+\frac{\frac
{2}{\kappa r_{h}}-3\eta}{8\kappa^{2}r_{\varepsilon}}-\left( \frac{1
{\kappa^{2}r_{h}^{2}}+3(2\eta^{2}-\theta)-\frac{6\eta}{\kappa r_{h}}\right)
\frac{\ln\kappa r_{\varepsilon}}{8\kappa},\nonumber\\
\int_{r_{h}+r_{\varepsilon}}^{L}dr\frac{r^{4}}{r_{h}^{4}}\frac{1}{f^{3}\left(
r\right) } & \sim\frac{1}{4\kappa^{3}r_{\varepsilon}^{2}}+\frac{\frac
{4}{\kappa r_{h}}-3\eta}{8\kappa^{2}r_{\varepsilon}}-\left( \frac{6
{\kappa^{2}r_{h}^{2}}+3(2\eta^{2}-\theta)-\frac{12\eta}{\kappa r_{h}}\right)
\frac{\ln\kappa r_{\varepsilon}}{8\kappa}, \label{eq:FIntegrals
\end{align}
where we expand $f\left( r\right) $ and $\frac{r^{2}}{r_{h}^{2}}$ at
$r=r_{h}$
\begin{align}
f\left( r\right) & \sim2\kappa\left( r-r_{h}\right) \left[ 1+\eta
\kappa\left( r-r_{h}\right) +\theta\kappa^{2}\left( r-r_{h}\right)
^{2}\right] ,\nonumber\\
\frac{r^{2}}{r_{h}^{2}} & \sim1+\frac{2\kappa\left( r-r_{h}\right)
}{\kappa r_{h}}+\frac{\kappa^{2}\left( r-r_{h}\right) ^{2}}{\kappa^{2
r_{h}^{2}}. \label{eq:Expansion
\end{align}
In eqn. $\left( \ref{eq:FIntegrals}\right) ,\,$\ we neglect finite terms as
$\kappa r_{\varepsilon}\rightarrow0$ and terms involving $L$. Note that we
define $\kappa=\frac{f^{\prime}\left( r_{h}\right) }{2}$ which is the
surface gravity for the black hole and hence $T_{0}=\frac{\hbar\kappa}{2\pi}$.
Thus, the divergent part of entropy near the horizon i
\begin{align}
S & \sim\frac{2r_{h}^{2}T_{0}^{3}}{\pi\hbar^{3}\kappa}\int u^{2}s\left(
u\right) du\left\{ \frac{1}{4\kappa r_{\varepsilon}}-\left( \frac{1}{\kappa
r_{h}}-\eta\right) \frac{\ln\kappa r_{\varepsilon}}{2}+\right. \nonumber\\
& \left. \frac{T_{0}^{2}u^{2}}{M_{f}^{2}}\left[ \frac{14}{20}\frac
{1}{\kappa^{2}r_{\varepsilon}^{2}}+\frac{\frac{8}{\kappa r_{h}}-21\eta
}{20\kappa r_{\varepsilon}}+\frac{2}{5}\left( \frac{28}{\kappa^{2}r_{h}^{2
}+39(2\eta^{2}-\theta)-\frac{96\eta}{\kappa r_{h}}\right) \frac{\ln\kappa
r_{\varepsilon}}{8}\right] \right\} . \label{eq:EntropyDivergentR
\end{align}
Defining the proper distance distance $\varepsilon$ between the brick wall and
the horizo
\[
\varepsilon=\int_{r_{h}}^{r_{h}+r_{\varepsilon}}\frac{dr}{\sqrt{f}}\sim
\frac{\sqrt{r_{\varepsilon}}}{\sqrt{2\kappa}}\left( 2-\frac{\eta\kappa
r_{\varepsilon}}{3}\right) ,
\]
one could express eqn. $\left( \ref{eq:EntropyDivergentR}\right) $ in terms
of $\varepsilon
\begin{align}
S & \sim\frac{r_{h}^{2}}{8\pi^{4}}\int u^{2}s\left( u\right) du\left\{
\frac{1}{\varepsilon^{2}}-2\kappa^{2}\left( \frac{1}{\kappa r_{h}
-\eta\right) \ln\kappa\varepsilon\right. \nonumber\\
& \left. +\frac{L_{f}^{2}u^{2}\kappa^{2}}{2\pi^{2}\varepsilon^{2}}\left[
\frac{14}{5\kappa^{2}\varepsilon^{2}}+\frac{2\left( 5\eta-\frac{39}{\kappa
r_{h}}\right) }{45}+\frac{2\kappa^{2}\varepsilon^{2}}{5}\left( \frac
{28}{\kappa^{2}r_{h}^{2}}+39(2\eta^{2}-\theta)-\frac{96\eta}{\kappa r_{h
}\right) \frac{\ln\kappa\varepsilon}{4}\right] \right\} ,
\end{align}
where we use $L_{f}=\frac{\hbar}{M_{f}}$. A natural choice for $\varepsilon$
is the minimal length $L_{f}$. If we take $\varepsilon=L_{f}$, we have for the
entanglement entropy near the horizon
\begin{align}
S & \sim\frac{r_{h}^{2}}{8\pi^{4}L_{f}^{2}}\int_{0}^{\infty}\left(
1+\frac{14u^{2}}{10\pi^{2}}\right) u^{2}s\left( u\right) du\nonumber\\
& \left. -\frac{r_{h}\kappa}{4\pi^{4}}\left( 1-\eta\kappa r_{h}\right)
\int u^{2}s\left( u\right) du\ln\kappa L_{f}+\text{Finite terms as
}\varepsilon L_{f}\rightarrow0\right. \nonumber\\
& \sim\frac{17A}{1800\pi L_{f}^{2}}-\frac{r_{h}\kappa\left( 1-\eta\kappa
r_{h}\right) }{45}\ln\kappa L_{f}+\text{Finite terms as }\varepsilon
L_{f}\rightarrow0, \label{eq:EntropyD
\end{align}
where $A=4\pi r_{h}^{2}$ is the horizon area.
\section{ Discussion and Conclusion}
\label{Sec:Con}
In \cite{IN-'tHooft:1984re,CON-Solodukhin:2011gn}, the entanglement entropy of
a massless scalar field near the horizon of the metric $\left(
\ref{eq:SchwarzschildMetric}\right) $ was calculated in the brick wall model
without the GUP. It has been shown the entropy in this case i
\begin{equation}
S\sim\frac{A}{360\pi\varepsilon^{2}}-\frac{\kappa r_{h}}{45}\left(
1-\kappa\eta r_{h}\right) \ln\kappa\varepsilon, \label{eq:EntropyNo
\end{equation}
where $\kappa$ and $\eta$ are defined in eqn. $\left( \ref{eq:Expansion
\right) $ and $\varepsilon$ is the proper distance between the wall and the
horizon. In our paper, we take $\varepsilon=L_{f}$ which is the minimal
length. The numerical factor of the leading term of the entropy in eqn.
$\left( \ref{eq:EntropyD}\right) $ has been changed from $\frac{1}{360\pi}$
to $\frac{17}{1800\pi}$ due to the minimal length effects. The leading terms
of eqn. $\left( \ref{eq:EntropyD}\right) $ and eqn. $\left(
\ref{eq:EntropyNo}\right) $ are both proportional to the horizon area. The
area law in the brick wall model was first obtained in
\cite{IN-'tHooft:1984re} and later was also studied in
\cite{CON-Mann:1990fk,CON-Susskind:1994sm,CON-Barbon:1994ej,CON-Mukohyama:1998rf
. In \cite{IN-Li:2002xb} where the GUP form $\left( \ref{eq:LiForm}\right) $
was used, the leading term of the entanglement entropy for the spherical
metric $\left( \ref{eq:SchwarzschildMetric}\right) $ was calculated up to
$\mathcal{O}\left( \lambda\right) $. The result was
\begin{equation}
S\sim\frac{3}{L_{f}^{2}\pi}\frac{A}{4},
\end{equation}
where $L_{f}=2\sqrt{\lambda}$ is the minimal length. In \cite{IN-Kim:2007if}
where the all order GUP form
\begin{equation}
\Delta x\Delta p\geq\frac{\hbar}{2}\exp\left[ \frac{\lambda}{\hbar^{2
}\left( \Delta p\right) ^{2}\right] ,
\end{equation}
was considered, the leading term of the all order entanglement entropy for the
spherically symmetric metric $\left( \ref{eq:SchwarzschildMetric}\right) $
was given b
\begin{equation}
S\sim\frac{e^{3}\zeta\left( 3\right) }{4\pi L_{f}^{2}}\frac{A}{4},
\end{equation}
where $L_{f}=\sqrt{\frac{e\lambda}{2}}$ and $\zeta\left( 3\right)
{\displaystyle\sum\limits_{n=1}^{\infty}}
\frac{1}{n^{3}}\approx1.202$. In both cases, the area law in the brick wall
model is preserved after the GUP corrections to $n\left( \omega,l\right) $
are considered. Our result shows that the area law is still valid after the
GUP corrections to both $n\left( \omega,l\right) $ and the Hawking
temperature $T$ are included. If we choose the invariant cutoff $\varepsilon
=\frac{m_{p}}{\sqrt{90\pi}}$ in t' Hooft's original calculation, then the
dominant entropy term becomes the Bekenstein-Hawking entropy, $S\sim
S_{BH}\equiv\frac{A}{4m_{p}^{2}}$. Similarly, if we have $L_{f}=\sqrt
{\frac{17}{450\pi}}m_{p}\approx0.1m_{p}$ in our calculation, $S_{BH}$ also
appears for the leading entropy term. On the other hand, it is normally
assumed that the minimal $L_{f}$ is of the order of $m_{p}$. If $L_{f}\sim
m_{p}$, one finds the entanglement entropy for one scalar field $S\sim
10^{-2}S_{BH}$. Since the entanglement entropy depends on the number and
nature of quantum fields while $S_{BH}$ is universal, the \textquotedblleft
species problem\textquotedblright\ arises when the the entanglement entropy
makes important contributions to $S_{BH}$. Taking $L_{f}\sim m_{p}$ would
provide a resolution to the species problem.
The subleading logarithmic part of the entanglement entropy is also calculated
in our paper. It turns out that the subleading logarithmic part is universally
given by $-\frac{\kappa r_{h}}{45}\left( 1-\kappa\eta r_{h}\right) \ln
\kappa\varepsilon$ in scenarios both with and without the GUP. For the
Schwarzschild metric with $f\left( r\right) =1-\frac{2M}{r}$, the subleading
logarithmic part becomes $-\frac{\ln\kappa\varepsilon}{45}$, which was first
obtained in \cite{CON-Solodukhin:1994yz}. On the other hand, we can estimate
the entropy of the black hole using eqn. $\left( \ref{HKT1}\right) $. \ In
fact, the angular momentum of the particle $L\sim pr_{h}\sim\omega r_{h}$ near
the horizon of the the black hole. Thus, one can rewrite $T
\begin{equation}
T\sim\frac{T_{0}}{1+\frac{2\omega^{2}}{M_{f}^{2}}}, \label{eq:Temp
\end{equation}
where $T_{0}=\frac{\hbar}{8\pi M}$ for the Schwarzschild black hole. As
reported in \cite{CON-AmelinoCamelia:2005ik}, the authors obtained the
relation $\omega\gtrsim\frac{\hbar}{\delta x}$ between the energy of a
particle and its position uncertainty in the framework of GUP. Near the
horizon of the the Schwarzschild black hole, the position uncertainty of a
particle will be of the order of the Schwarzschild radius of the black hole
\cite{CON-Bekenstein:1973ur} $\delta x\sim r_{h}$. Thus, one finds for $T
\begin{equation}
T\sim\frac{T_{0}}{1+\frac{m_{p}^{4}}{2M^{2}M_{f}^{2}}}, \label{eq: Temp
\end{equation}
where we use $\hbar=m_{p}^{2}$. Using the first law of the black hole
thermodynamics, we find the corrected black hole entropy i
\begin{equation}
S=\int\frac{dM}{T}\sim\frac{A}{4m_{p}^{2}}+\frac{4\pi L_{f}^{2}}{m_{p}^{2}
\ln\left( \frac{A}{16\pi}\right) , \label{eq:entropy
\end{equation}
where $A=4\pi r_{h}^{2}=16\pi M^{2}$ is the area of the horizon. The
logarithmic term in eqn. $\left( \ref{eq:entropy}\right) $ is the well known
correction from quantum gravity to the classical Bekenstein-Hawking entropy,
which have appeared in different studies of GUP modified thermodynamics of
black holes\cite{CON-Bina:2010ir,CON-Chen:2002tu,CON-Xiang:2009yq}. Comparing
eqn. $\left( \ref{eq:EntropyD}\right) $\ with eqn. $\left( \ref{eq:entropy
\right) $, we note that there are two discrepancies in the subleading
logarithmic term, one of which is the sign and the other dependence on the
minimal length $L_{f}$. These discrepancies would imply that the entanglement
entropy could not solely account for the entropy of the black hole.
It has been shown in
\cite{IN-Li:2002xb,IN-Kim:2006rx,IN-Kim:2007bx,IN-Kim:2007nh,IN-Kim:2007if}
that the artificial cutoff parameter in the brick wall model located just
outside the horizon can be avoided if GUP is considered. However, this is not
the case in our paper. Actually, if one attempts to let $r_{\varepsilon
=0\,$\ in eqn. $\left( \ref{eq:EntropyDivergentR}\right) $, the entropy
diverges and a brick wall is still needed. How can we reconcile the
contradiction? One might note that we calculate the entanglement entropy to
$\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $. For the typical energy
$\omega\sim T_{0}=\frac{\hbar\kappa}{2\pi}\,$, one finds that the
$\mathcal{O}\left( \frac{1}{M_{f}^{2}}\right) $ GUP corrections to the
entropy $\sim\frac{\omega^{2}}{f\left( r\right) M_{f}^{2}}\sim\frac
{\hbar^{2}\kappa^{2}}{4\pi^{2}f\left( r\right) M_{f}^{2}}$. At the wall at
$r_{\varepsilon}=r_{h}+2\kappa L_{f}^{2},$ we have $\frac{\hbar^{2}\kappa^{2
}{f\left( r_{\varepsilon}\right) M_{f}^{2}}\sim\frac{\hbar^{2}\kappa^{2
}{16\pi^{2}\kappa^{2}L_{f}^{2}M_{f}^{2}}\sim\frac{1}{16\pi^{2}}$. Thus, our
perturbative method is valid outside the wall at $r_{\varepsilon
=r_{h}+2\kappa L_{f}^{2}$. However, the perturbation would break down deep
within the wall and one needs to consider higher order contributions. The
divergence of the entropy eqn. $\left( \ref{eq:EntropyDivergentR}\right) $
as $r_{\varepsilon}\rightarrow0$ is more like due to the breaking down of our
perturbative method. In
\cite{IN-Li:2002xb,IN-Kim:2006rx,IN-Kim:2007bx,IN-Kim:2007nh}, it is crucial
for $\lambda\omega^{2}/f$ to be in the denominator of the integrand in eqn.
$\left( \ref{eq:NLamda}\right) $ to get rid of the wall. If one replaces
$\left( 1+\lambda\omega^{2}/f\right) ^{-3}$ with $1-3\lambda\omega^{2}/f$ in
eqn. $\left( \ref{eq:NLamda}\right) ,$ the integral diverges as
$r\rightarrow r_{h}.$ The divergence arises simply because the Taylor
expansion of $\left( 1+\lambda\omega^{2}/f\right) ^{-3}$ breaks down when
$\lambda\omega^{2}/f>1$.
In this paper, we calculate the entanglement entropy of a massless scalar
field near the horizon of a 4D spherically symmetric black hole in the brick
wall model incorporating the minimal length effects. We show that the leading
term of the entropy is proportional to the horizon area. If the minimal length
$L_{f}\sim m_{p}$, the entanglement entropy makes a small contribution to the
Bekenstein-Hawking entropy, which might resolve the \textquotedblleft species
problem\textquotedblright. The subleading logarithmic term is also calculated.
The result is the same as the one in the usual brick wall model without the
minimal length and independent of the minimal length $L_{f}$.\
\noindent\textbf{Acknowledgements. }
We would like to acknowledge useful discussions with Y. He, Z. Sun and H. W.
Wu. This work is supported in part by NSFC (Grant No. 11005016, 11175039 and
11375121) and SYSTF (Grant No. 2012JQ0039).
|
\section{\label{introduction} Introduction}
Since the discovery of a light Higgs boson at the LHC,
interest in beyond-the-Standard-Model scenarios
has focused on models in which the Higgs is naturally light
compared to the typical scale of new physics.
One approach postulates the existence of a new strongly interacting sector,
which we will refer to as hypercolor in this paper. The Higgs doublet of the Standard Model (SM)
emerges among the Nambu--Goldstone bosons (NGBs)
originating from dynamical symmetry breaking of the flavor symmetry group $G$
of the hypercolor theory.
The electroweak gauge bosons as well as the SM fermions then couple to these
NGBs, breaking the symmetry group $G$ explicitly to a smaller group,
thereby generating an effective potential for the NGBs.
Under suitable conditions, this radiatively induced effective potential
leads to electroweak symmetry breaking, with
the Higgs field acquiring an expectation value as in the SM.
This framework still allows for many different possibilities.
For reviews that span the evolution of this field, as well as for
generic features of these models, we refer to Refs.~\cite{MP,GGPR,BBRV,RC,BCS}.
We will be interested in composite-Higgs models in which
the sector external to the hypercolor gauge theory,
which includes the SM gauge bosons and fermions, is as simple as possible.
For instance, we do not wish to introduce any weakly coupled gauge
bosons besides the electroweak gauge bosons,
as in Little Higgs models \cite {MP}.
The electroweak gauge bosons have to stay massless at the
dynamical symmetry breaking scale of the hypercolor theory,
and therefore they have to couple to generators in the unbroken
flavor subgroup $H\subset G$. As a result,
the effective potential generated for the
hypercolor NGBs by the electroweak gauge bosons will not lead to
electroweak symmetry breaking, a phenomenon often referred to as
vacuum alignment \cite{vac}.
Electroweak symmetry breaking must therefore originate
in the effective potential generated by the top quark,
being the SM fermion with the strongest coupling to the Higgs,
and, hence, to the hypercolor theory. We will postulate that the top quark
couples linearly to hyperbaryons (the baryons of the hypercolor theory),
as first proposed in Ref.~\cite{KaplanB}.
This idea is attractive from the point of view of $CP$ violation
and flavor-changing neutral currents (FCNCs) \cite{RC}.
Here, we will limit ourselves to a discussion of the top quark sector,
where the main concerns are to generate the experimentally measured value
of the top quark's mass naturally,%
\footnote{Without causing problems for $Z\to b\overline{b}$ decays
\cite{ferretti,ACDP}.
}
together with a Higgs potential that triggers electroweak symmetry breaking.
It is generally acknowledged that the mass of the top quark sets it
apart from the other SM fermions as it is the only SM fermion with a mass
of the order of the electroweak symmetry breaking scale,
$v\sim 250$~GeV. This suggests that the top quark may play an essential
role in generating electroweak symmetry breaking, whereas the origin
of the other SM fermion masses, and the strength of other symmetry breakings
such as $CP$ violation and FCNCs, might be very different.
The concrete hypercolor theory we will study in this article was proposed in
Ref.~\cite{ferretti}. It was preceded by a general study
that highlighted what makes that theory particularly attractive \cite{FK}.%
\footnote{See also Ref.~\cite{Vecchi}.}
The hypercolor theory is a vector-like $SU(4)$ gauge theory
with fermions in two different {\it irreps}\ (irreducible representations).
One of these {\it irreps}, the six-dimensional two-index antisymmetric {\it irrep},
is real. With 5 Majorana (or Weyl) fermions in this {\it irrep},
dynamical symmetry breaking in that sector of the theory
gives rise to an $SU(5)/SO(5)$ non-linear sigma model as its
low-energy effective theory. As we will see in detail below,
the Higgs field lives in this non-linear sigma model.
Generally speaking, composite-Higgs models often rely on an
$SU(N_w)/SO(N_w)$ non-linear sigma model, which can arise from chiral symmetry
breaking in a theory containing $N_w$ Weyl (or Majorana) fermions in
a real {\it irrep}. An alternative coset structure is $SU(N_w)/Sp(N_w)$,
for which the $N_w$ Weyl fermions should be in a pseudoreal {\it irrep}\ \cite{vac}.
For recent lattice work involving the pseudoreal fundamental {\it irrep}\
of $SU(2)$ gauge theory we refer to Refs.~\cite{LPS,DCP}. For a review
on beyond-the-SM lattice work, see Ref.~\cite{Kuti}.
The most familiar example of a real {\it irrep}\ is the adjoint representation,
which occurs for example in supersymmetric theories \cite{susy}.
However, 5 Majorana fermions in the adjoint {\it irrep}\ would most likely
push the theory from being confining to being conformal, even before
the introduction of any fermions in another {\it irrep}.%
\footnote{See Ref.~\cite{Kuti} and references therein.}
Avoiding the adjoint {\it irrep},
the smallest instance of a real {\it irrep}\ is the sextet of $SU(4)$.
Our goals are as follows. First, a gauge theory
such as this $SU(4)$ hypercolor theory is amenable
to investigations using the methods of lattice gauge theory.
The effective theory below the hypercolor scale,
relevant for SM phenomenology, can be parametrized
in terms of low-energy couplings (LECs).
These LECs can be expressed in terms of correlation functions
in the hypercolor theory, which, in turn, allows for their computation
on the lattice. While this is well understood
for the electroweak gauge sector, a similar careful derivation
of the LECs controlling the top sector has to our knowledge not been given
to date. We derive the necessary correspondence
using spurion techniques.
Second,
once the connection between the effective theory and the hypercolor theory
has been established, we find that
it is possible to obtain semi-quantitative estimates of
the size of these LECs, using large-$N$ methods and factorization.
In particular, we show that the contribution of the top quark
to the Higgs effective potential indeed drives electroweak symmetry breaking
in a particular large-$N$ limit.
We expect that the techniques developed in this article can be easily extended
to similar hypercolor models. In this sense, our choice of the model
of Ref.~\cite{ferretti} should be considered as a useful example.
This article is organized as follows. Sec.~\ref{ferretti-model}
introduces and reviews the hypercolor theory \cite{ferretti},
including its field content, symmetries, and the effective non-linear fields
that will be needed for the low-energy effective theory.
In Sec.~\ref{weak-effpot} we briefly discuss the contribution of the
electroweak gauge bosons to the Higgs effective potential.
The main part of this article is Sec.~\ref{top},
where we discuss the top quark sector in detail.
We introduce the top quark spurions and the hyperbaryons in Sec.~\ref{baryons}.
We discuss the top Yukawa coupling in Sec.~\ref{top-mass},
and the top quark contribution to the Higgs effective potential
in Sec.~\ref{top-effpot}. In Sec.~\ref{large-N} we define a large-$N$ limit
of the model, and show that for large enough $N$ the top-induced
Higgs potential will lead to electroweak symmetry breaking.
For simplicity, we assume a minimal explicit breaking
of the flavor group of the hypercolor theory by the couplings
to the SM. In Sec.~\ref{pheno} we briefly comment on the more general
situation that arises if we relax this assumption.
In Sec.~\ref{lattice} we discuss the similarities between the
hypercolor theory and QCD, thus arguing that techniques developed to
study QCD on the lattice should be sufficient for hypercolor theories as well.
Section~\ref{conclusion} contains our conclusions.
A short appendix collects some of our conventions.
\section{\label{ferretti-model} Ferretti's model}
In Ref.~\cite{FK}, several requirements were put forward for a class of
composite Higgs models based on a hypercolor gauge theory as a UV completion.
We begin by listing these requirements.
The gauge group is assumed to be simple, and the dynamical symmetry
breaking pattern, $G\to H$, to be such that
\begin{eqnarray}
\label{dynsym}
H &\supset& SU(3)_{\rm color}\times SU(2)_L\times SU(2)_R\times U(1)_X
\\
&\supset& SU(3)_{\rm color}\times SU(2)_L\times U(1)_Y\ ,
\nonumber
\end{eqnarray}
with the SM gauge group in the last line. The group
$SU(2)_R$ is the familiar custodial symmetry of the SM, and
the hypercharge is $Y=T^3_R+X$. The SM Higgs doublet,
with quantum numbers $({\bf 1},{\bf 2},{\bf 2})_0$ under
$SU(3)_{\rm color}\times SU(2)_L\times SU(2)_R\times U(1)_X$, should be contained
in the NGB multiplet associated with the symmetry breaking $G\to H$.
In order to accommodate a partially composite top quark \cite{KaplanB},
\mbox{\it i.e.}, for the top quark to acquire its mass through linear couplings
to hyperbaryons, there must exist hyperbaryons with quantum numbers
that match those of the SM quarks.
This includes a set of right-handed, spin-$1/2$ hyperbaryons
with quantum numbers $({\bf 3},{\bf 2})_{1/6}$ of the SM gauge group
$SU(3)_{\rm color}\times SU(2)_L\times U(1)_Y$, which serve
as partners of the SM quark doublet $q_L$;
and left-handed, spin-$1/2$ hyperbaryons
with the quantum numbers $({\bf 3},{\bf 1})_{2/3}$, to serve as partners
of the SM quark singlet $t_R$. Finally, the hypercolor theory
should be asymptotically free, and both the hypercolor gauge group and the
SM gauge group should be free of anomalies.
The hypercolor model with the smallest gauge group
that satisfies all these requirements is an $SU(4)$ gauge theory \cite{FK}.
The hyperfermion content consists of
five Majorana fermions $\chi_i$, $i=1,\dots,5$, transforming in the
six-dimensional two-index antisymmetric {\it irrep}\ of hypercolor,
which is a real representation; and three Dirac fermions $\psi_a$, $a=1,2,3$,
in the fundamental representation.
The Majorana field $\chi$ can be written in terms of a Weyl fermion $\Upsilon$ as
\begin{subequations}
\label{maj}
\begin{eqnarray}
\chi_{ABi} &=&
\left(
\begin{array}{c}
\vspace{1ex}
\Upsilon_{ABi} \\
{1\over 2}\e_{ABCD}\,\e\,(\bar{\U}^{CD}_i)^T
\end{array}
\right) \ ,
\label{maja}\\
\overline{\c}^{AB}_i &=& {1\over 2} \e^{ABCD}\chi^T_{CDi} \, C \ = \ \rule{0ex}{3.5ex}
\left(
\begin{array}{cc}
\! -{1\over 2}\e^{ABCD}(\Upsilon_{CDi})^T \e & \bar{\U}^{AB}_i
\end{array}
\right) \ .
\label{majb}
\end{eqnarray}
\end{subequations}
We use capital letters for the $SU(4)$ hypercolor indices,
with lower indices for the fundamental {\it irrep}, and upper indices
for the anti-fundamental {\it irrep}. Several lower or upper indices
will always be fully antisymmetrized.
A Dirac fermion $\psi$ in the fundamental {\it irrep}\ can be written in terms
of two right-handed Weyl fermions, $\Psi$ in the fundamental {\it irrep}\
and $\tilde\J$ in the anti-fundamental, as
\begin{equation}
\psi_{Aa} = \left( \begin{array}{c}
\Psi_{Aa} \\ \e \bar{\tilde\J}^T_{Aa}
\end{array} \right) \ , \qquad
\overline{\j}_a^A = \left( -(\tilde\J_a^A)^T\e \ \ \bar{\J}_a^A \right) \ .
\label{DW}
\end{equation}
We suppress spinor indices. $C$ is the charge-conjugation matrix,
$\e=i\s_2$ is the two-dimensional $\e$-tensor acting on the Weyl
spinor index,
and the superscript $T$ denotes the transpose in spinor space.
With the lattice in mind, we work in euclidean space, choosing our Dirac
matrices to be hermitian and using the chiral representation, see App.~\ref{conv}.
The hypercolor theory possesses a flavor symmetry group
\begin{equation}
\label{flavor}
G = SU(5)\times SU(3)\times SU(3)'\times U(1)_X\times U(1)'\ ,
\end{equation}
with quantum numbers $({\bf 5},{\bf 1},{\bf 1})_{(0,-1)}$ for $\Upsilon$;
$({\bf 1},{\bf \bar{3}},{\bf 1})_{(1/3,5/3)}$ for $\Psi$;
and $({\bf 1},{\bf 1},{\bf 3})_{(-1/3,5/3)}$ for $\tilde\J$.%
\footnote{Compare Table~1 of Ref.~\cite{ferretti}.}
We assume that dynamical symmetry breaking takes place,
generating a condensate $\langle\overline{\c}_i\chi_j\rangle\propto\delta_{ij}$
that breaks $SU(5)\to SO(5)$.
Consistent with the general considerations of Ref.~\cite{vac},
the Majorana bilinear $\overline{\c}_i\chi_j$ is antisymmetric
on its spinor indices and symmetric on its hypercolor indices,
and so it is symmetric on its flavor indices.
In addition, there is a condensate $\langle\overline{\j}_a\psi_b\rangle\propto\delta_{ab}$
that breaks $SU(3)\times SU(3)'$ to its diagonal subgroup, which we
identify with $SU(3)_{\rm color}$. Both condensates also break $U(1)'$.
The unbroken group is
\begin{equation}
\label{breaking}
H = SO(5)\times SU(3)_{\rm color}\times U(1)_X\ .
\end{equation}
For heuristic arguments supporting this pattern of symmetry breaking,
see Refs.~\cite{vac,ferretti}. Of course, whether this is the actual
symmetry breaking pattern is something that can be investigated
on the lattice. Indeed the symmetry breaking pattern
of the Dirac fermions, with $SU(3)\times SU(3)'$ breaking to
the diagonal $SU(3)$ subgroup, is consistent with all known lattice results.
A first study of the real-{\it irrep}\ symmetry breaking pattern,
in a similar theory except with four, instead of five, Majorana fermions,
has recently appeared in Ref.~\cite{taco}.
The effective theory at energy scales much below the hypercolor scale
$\Lambda_{\rm HC}$ thus contains NGBs parametrizing the $U(1)'$ group manifold,
and the cosets $SU(3)\times SU(3)'/SU(3)_{\rm color}$ and $SU(5)/SO(5)$,
amounting to 1, 8 and 14 NGBs for each of these factors, respectively.
These NGBs are massless when all couplings of the hypercolor theory
to the SM are turned off. A non-trivial effective potential
is induced both by the SM gauge bosons, as we briefly review
in Sec.~\ref{weak-effpot}, and by the coupling to the third-generation quarks.
The latter, which is the main subject of this paper, will be studied
in Sec.~\ref{top}.
The Higgs doublet is a subset of the NGB multiplet parametrizing the coset
$SU(5)/SO(5)$. In more detail,
the 14 NGBs corresponding to the generators in this coset
are described by a non-linear field $\Sigma\in SU(5)/SO(5)$
obtained by considering fluctuations
around the vacuum $\langle\Sigma\rangle=\Sigma_0= {\bf 1}$,
\begin{equation}
\label{Sigma}
\Sigma = u\,\Sigma_0\,u^T = \mbox{exp}(i\Pi/f)\,\Sigma_0\,\mbox{exp}(i\Pi/f)^T
= \mbox{exp}(2i\Pi/f)\ ,
\end{equation}
with%
\footnote{Note that in Ref.~\cite{ferretti}, the notation $\Sigma$ is used for the
field $u$ of Eq.~(\ref{Sigma}).
}
\begin{equation}
\label{Sigmasymm}
\Sigma = \Sigma^T\quad\Rightarrow\quad\Pi=\Pi^T\ .
\end{equation}
Under $g\in SU(5)$, $\Sigma$ transforms as $\Sigma\to g\Sigma g^T$.
At the level of the algebra, $SU(2)_L\times SU(2)_R$ in Eq.~(\ref{dynsym}) is
equivalent to the $SO(4) \subset SO(5)$ associated with the first
four rows and columns. The explicit form of the generators is given
in the appendix. With this choice,
the field $\Pi$ can be written as
\begin{equation}
\label{Pi}
\Pi = \Theta+\Theta^\dagger+\Phi_0+\Phi_++\Phi_+^\dagger+\eta\ ,
\end{equation}
with $\Theta$ containing the Higgs doublet $H=(H_+,H_0)^T$,
\begin{equation}
\label{H}
\Theta=\left(\begin{array}{ccccc}
0 & 0 & 0 & 0 & -iH_+/\sqrt{2} \\
0 & 0 & 0 & 0 & H_+/\sqrt{2} \\
0 & 0 & 0 & 0 & iH_0/\sqrt{2} \\
0 & 0 & 0 & 0 & H_0/\sqrt{2} \\
-iH_+/\sqrt{2} & H_+/\sqrt{2} & iH_0/\sqrt{2}
& H_0/\sqrt{2} & 0
\end{array}\right)\ .
\end{equation}
For the explicit parametrization of the rest of $\Pi$,
we refer to Ref.~\cite{ferretti}, as we will not need it here.
The Higgs doublet comprises four of the NGBs,
and the $SU(2)_L$ triplets $\phi_0$, $\phi_+$ and $\phi_-=(\phi_+)^\dagger$
comprise 9 more NGBs.%
\footnote{In the notation of Ref.~\cite{ferretti},
$\phi_0=(\phi_0^-,\phi_0^0,\phi_0^+)$ with $\phi_0^0$ real and
$\phi_0^+ = (\phi_0^-)^*$, while $\phi_+=(\phi_+^-,\phi_+^0,\phi_+^+)$,
with all components complex.
}
Finally, $\Pi$ has a component $\eta$ proportional to
the generator $\mbox{diag}(1,1,1,1,-4)$, which is neutral with
respect to the entire SM model gauge group, and which completes the multiplet
of 14 NGBs.
While they play only a small role, we will need also
the non-linear fields associated with the other broken symmetries.
We account for the $SU(3)\times SU(3)'/SU(3)_{\rm color}$ coset
by a non-linear field $\Omega\in SU(3)$, transforming as $\Omega\to g\Omega h^\dagger$
for $g\in SU(3)$ and $h\in SU(3)'$.
A similar non-linear field arises in the familiar chiral lagrangian of
3-flavor QCD, but the reader should keep in mind the different
physical roles of the various $SU(3)$ groups in the case at hand.
Throughout most of this paper, we will assume that the only source
of \textit{explicit} breaking of $SU(3)\times SU(3)'$ to $SU(3)_{\rm color}$
arises from the coupling of the $SU(3)_{\rm color}$ currents to the SM gluons.
A wider range of possibilities than what is the main focus of this article
is allowed if we relax this assumption.
This is briefly discussed in Sec.~\ref{pheno}.
The 8 NGBs associated with the non-linear $\Omega$ field transform in
the adjoint {\it irrep}\ of $SU(3)_{\rm color}$.
They are singlets under both $SU(2)_L$ and $U(1)_Y$.
Finally, to account for the spontaneous breaking of $U(1)'$
we introduce a non-linear field $\Phi\in U(1)$
with unit charge under $U(1)'$.
The associated NGB, $\eta'$, is neutral under the SM gauge interactions.
Using a $\sim$ sign to indicate
identical transformation properties under the entire flavor group $G$,
we thus have
\begin{subequations}
\label{transG}
\begin{eqnarray}
\label{transGa}
\Phi^{-2}\, \Sigma_{ij} &\sim& \overline{\c}_i P_R \chi_j
\ \sim \ \e^{ABCD} (\Upsilon_{CDi})^T \,\e\, \Upsilon_{ABj} \ ,
\\
\label{transGb}
\Phi^{-10/3}\, \Omega_{ab} &\sim& \overline{\j}_a P_L \psi_b
\ \sim \ \bar{\J}_a^A \,\e\, (\bar{\tilde\J}_{Ab})^T \ .
\end{eqnarray}
\end{subequations}
\section{\label{weak-effpot} Higgs effective potential from electro-weak gauge bosons}
In this section, we briefly review the contribution
from the SM gauge bosons to the effective potential for the NGBs,
starting with the effective potential for
the $SU(5)/SO(5)$ non-linear field $\Sigma$ generated by the electroweak
gauge bosons. This part of the effective potential takes the form
\begin{equation}
\label{VeffEW}
V_{\rm eff}^{\rm EW}(\Sigma) = C_{LR}\sum_Q\,{\rm tr}\left(Q\Sigma Q^*\Sigma^*\right)\ ,
\end{equation}
if we work to leading (\mbox{\it i.e.}, quadratic) order in the SM gauge couplings.
The sum over $Q$ runs over the $SU(2)_L$ generators
$gT^a_L$ with $T^a_L$ given in Eq.~(\ref{embed}),
and the hypercharge generator $g'Y=g'\left(T^3_R+X\right)$,
with $X=0$ for the $\Pi$ field. Here
\begin{equation}
\label{CLR}
C_{LR} = \frac{3}{(4\p)^2}\int_0^\infty dq^2\,q^2\,\Pi_{LR}(q^2)\ ,
\end{equation}
and
\begin{equation}
\label{PiLR}
\left(q^2\delta_{\mu\nu}-q_\mu q_\nu\right)\Pi_{LR}(q^2)
= \int d^4x\,e^{iqx}\,
\,{\rm tr} \svev{\gamma_\mu P_R[\chi(x)\overline{\c}(0)]\gamma_\nu P_L[\chi(0)\overline{\c}(x)]} \ ,
\end{equation}
where $[\chi(x)\overline{\c}(y)]$ is the Majorana fermion propagator for the field $\chi$
of Eq.~(\ref{maj}), and $\langle\dots\rangle$ indicates the expectation value with
respect to the hypercolor gauge fields. This type of effective potential
goes back to the well-known formula for the mass difference
between the charged and neutral pions in QCD.
For further explanations and a derivation of this result in the present
context we refer to the review article Ref.~\cite{RC}
and to the appendix of Ref.~\cite{vacalignW}.%
\footnote{The factor of 3 in Eq.~(\ref{CLR}) comes from tracing over the transversal
projector. This factor is erroneously missing in the published versions
of Eqs.~(A11a) and~(A11b) of Ref.~\cite{vacalignW}, and Eq.~(A5) of
Ref.~\cite{vacalignS}. Also, in the published versions,
the term $gW_\mu Q_a J_{\mu A}$ in Eq.~(A2) of Ref.~\cite{vacalignW}
should be multiplied by $i$, and similarly for the term
$gW_\mu (Q_a^L J_{\mu a}^L + Q_a^R J_{\mu a}^R)$ in Eq.~(A2) of Ref.~\cite{vacalignS}.
All these factors have been corrected in the current archive versions.
}
The proof of Ref.~\cite{EW} that $C_{LR}>0$ applies also in this case.
Using the explicit form~(\ref{embed}),
the minimum of $V_{\rm eff}^{\rm EW}(\Sigma)$ is equal to
$-C_{LR}(3g^2+g'^2)$, which is attained at $\Sigma={\bf 1}$. This
part of the effective potential does not rotate the vacuum of the hypercolor
theory, exhibiting the phenomenon of vacuum alignment \cite{vac}.
Expanding $\Sigma$ to quadratic order in $\Pi$ inside $V_{\rm eff}^{\rm EW}$,
we find the mass terms
\begin{equation}
\label{massNGBsEW}
V_{\rm eff}^{{\rm EW}(2)} = \frac{C_{LR}}{f^2}
\left\{(3g^2+g'^2)\left(2H^\dagger H+\frac{16}{3}\,\phi_+^\dagger
\phi_+\right)+8g^2\phi_0^\dagger\phi_0\right\}\ .
\end{equation}
The explicit symmetry breaking by
the electroweak gauge bosons produces a positive mass-squared for all components
of the NGB multiplet $\Pi$ except $\eta$.
To summarize, when we couple the hypercolor theory to the SM gauge bosons,
an effective potential for the non-linear field $\Sigma$ is generated.
At lowest order, it is proportional to the squares of the electroweak couplings
$g_{\rm EW}=g$ for $SU(2)_L$, or $g_{\rm EW}=g'$ for $U(1)_Y$,
where it is understood that all SM couplings are evaluated
at the hypercolor scale $\Lambda_{\rm HC}$.
Because of vacuum alignment, the expectation value $\Sigma_0$ will remain
equal to one, but the Higgs doublet and the three $SU(2)_L$ triplets
will acquire a mass proportional to $g_{\rm EW}f$, while the
singlet $\eta$ will remain massless. To avoid confusion,
the contribution to the effective potential from the top quark has not yet
been included,
and will be discussed in the next section.
Similarly, when we turn on the QCD interactions,
an effective potential for the $\Omega$ non-linear field is generated,
\begin{equation}
\label{VeffQCD}
V_{\rm eff}^{\rm QCD}(\Omega) = -C_{LR}^{\rm QCD}\sum_Q\,{\rm tr}\left(Q\Omega Q\Omega^\dagger\right)\ ,
\end{equation}
where now $Q$ runs over the 8
generators $g_{\rm s}\lambda_a$ of $SU(3)_{\rm color}$,
and $g_{\rm s}$ is the QCD coupling (at the hypercolor scale).
In the underlying hypercolor theory, $C_{LR}^{\rm QCD}$ has
a representation analogous to Eq.~(\ref{CLR}),
except that the Majorana-fermion propagator in Eq.~(\ref{PiLR})
is replaced by the Dirac-fermion propagator $[\psi(x)\overline{\j}(y)]$.
Once again there is vacuum alignment, nailing down the vacuum at
$\svev{\Omega_{ab}}=\delta_{ab}$, and giving the octet of NGBs a mass
of order $g_{\rm s}f$. Thus,
as long as $f$ is much larger than the electroweak scale,
both the $SU(2)_L$ triplet NGBs and the color-octet NGBs are much heavier
than the electroweak gauge bosons or the top quark.
As already noted,
the NGB $\eta'$ of the spontaneously broken $U(1)'$ is inert
under all the SM gauge interactions. Moreover, the coupling of the
hypercolor sector to the SM considered in the next section does not break
$U(1)'$ explicitly. Therefore, no effective potential will be generated for
the associated non-linear field $\Phi$, and $\eta'$ will remain exactly
massless.
\section{\label{top} The top quark sector}
We now proceed to the main part of this article,
which is the study of the dynamics arising from the coupling
of the top quark to the hypercolor theory.
There are two aspects of interest: the contribution of the top quark to the
effective potential for the non-linear field $\Sigma$ containing
the Higgs field, analogous to the
contribution from the weak gauge bosons in Eq.~(\ref{VeffEW}); and the mass
of the top quark itself.
We will begin with the coupling of the top quark to the hypercolor theory at
the ``microscopic'' level, which involves only the elementary fields:
the hypercolor gauge fields, and the hyperfermions of Eqs.~(\ref{maj}) and~(\ref{DW}).
We introduce fermionic spurions which transform in complete
representations of the flavor group $G$ of Eq.~(\ref{flavor}), and which contain the
$SU(2)_L$ doublet $q_L$ of the left-handed top and bottom quarks
$t_L$ and $b_L$, as well as the right-handed top quark $t_R$.
Following the mechanism proposed in Ref.~\cite{KaplanB},
the spurions will be coupled linearly to suitable hyperbaryon fields,
which are three-fermion operators in the hypercolor theory.
We demand that the spurion--hyperbaryon interactions
are invariant under $G$, because the microscopic theory does not
know about the dynamical breaking $G\to H$.
The spurion--hyperbaryon interaction terms are
four-fermion operators, and are assumed to arise from some extended
hypercolor (EHC) sector with a dynamical scale $\Lambda_{\rm EHC}\gg
\Lambda_{\rm HC}$, the origin of which we will not specify.
We will return to this point in the conclusion section.
The hyperbaryon operators and the four-fermion couplings
are constructed in Sec.~\ref{baryons}.
We then turn to the effective low-energy theory.
We demand that also the effective theory is invariant under $G$,
but it can now depend on the effective fields:
the $SU(5)/SO(5)$ coset field $\Sigma$, which plays
a central role since it contains the Higgs field,
as well as the $SU(3)\times SU(3)'/SU(3)_{\rm color}$ field $\Omega$
and the $U(1)$-valued field $\Phi$.
Note that we do not allow the effective theory to contain
any effective fields for hyperbaryons. This strategy generalizes
the standard construction of the chiral lagrangian for QCD.%
\footnote{See for instance Ref.~\cite{MGchpt}.}
We proceed in two steps. First, in Sec.~\ref{top-mass}, we consider the
coupling of the SM quarks $q_L$ and $t_R$ to the effective non-linear fields,
integrating out all other states in the hypercolor theory.
This will lead to an expression for the top Yukawa coupling in terms of
a hyperbaryon two-point function in the hypercolor theory.
Next, in Sec.~\ref{top-effpot}, we consider the contribution
to the effective potential obtained by integrating also
over the third-generation quarks to leading order in the top Yukawa coupling.
This involves restricting the spurions to their SM values,
in which all components except those corresponding to
$q_L=(t_L,b_L)$ and to $t_R$ are set equal to zero, and
integrating over $q_L$ and $t_R$.
Like the coupling to the SM gauge bosons (Sec.~\ref{weak-effpot}),
this breaks explicitly the flavor group $G$.
However, in the approximation in which we work, only $SU(5)$ is broken explicitly,
whereas all other factors in Eq.~(\ref{flavor}) are not.
As a result, no effective potential is generated for $\Omega$ or $\Phi$.
\begin{table}[t]
\vspace*{3ex}
\begin{center}
\begin{tabular}{ c | clccc | c } \hline
& $SU(5)$ & $SU(3)\times SU(3)'$ & $SU(3)_c$ & $U(1)_X$ & $U(1)'$ &
\\ \hline\hline
$\Upsilon (\Psi\J)$ & \mbox{\bf 5} & $(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})\times(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})\to(\mbox{\bf 3},\mbox{\bf 1})$
& \mbox{\bf 3} & 2/3 & 7/3 & $B_R$ \\
$\Upsilon (\bar{\tilde\J}\btJ)$ & \mbox{\bf 5} & $(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})\times(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})\to(\mbox{\bf 1},\mbox{\bf 3})$
& \mbox{\bf 3} & 2/3 & -13/3 & $B'_R$ \\
$\bar{\U} (\bar{\J}\bJ)$ & \mbox{${\bf \overline{5}}$} & $(\mbox{\bf 3},\mbox{\bf 1})\times(\mbox{\bf 3},\mbox{\bf 1})\to(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})$
& \mbox{${\bf \overline{3}}$} & -2/3 & -7/3 & $\overline{B}_R$ \\
$\bar{\U} (\tilde\J\tJ)$ & \mbox{${\bf \overline{5}}$} & $(\mbox{\bf 1},\mbox{\bf 3})\times(\mbox{\bf 1},\mbox{\bf 3})\to(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})$
& \mbox{${\bf \overline{3}}$} & -2/3 & 13/3 & $\overline{B}'_R$ \\
\hline\hline
$\bar{\U} (\Psi\J)$ & \mbox{${\bf \overline{5}}$} & $(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})\times(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})\to(\mbox{\bf 3},\mbox{\bf 1})$
& \mbox{\bf 3} & 2/3 & 13/3 & $B_L$ \\
$\bar{\U} (\bar{\tilde\J}\btJ)$ & \mbox{${\bf \overline{5}}$} & $(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})\times(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})\to(\mbox{\bf 1},\mbox{\bf 3})$
& \mbox{\bf 3} & 2/3 & -7/3 & $B'_L$ \\
$\Upsilon (\bar{\J}\bJ)$ & \mbox{\bf 5} & $(\mbox{\bf 3},\mbox{\bf 1})\times(\mbox{\bf 3},\mbox{\bf 1})\to(\mbox{${\bf \overline{3}}$},\mbox{\bf 1})$
& \mbox{${\bf \overline{3}}$} & -2/3 & -13/3 & $\overline{B}_L$ \\
$\Upsilon (\tilde\J\tJ)$ & \mbox{\bf 5} & $(\mbox{\bf 1},\mbox{\bf 3})\times(\mbox{\bf 1},\mbox{\bf 3})\to(\mbox{\bf 1},\mbox{${\bf \overline{3}}$})$
& \mbox{${\bf \overline{3}}$} & -2/3 & 7/3 & $\overline{B}'_L$ \\
\hline\hline
\end{tabular}
\end{center}
\vspace*{-3ex}
\begin{quotation}
\floatcaption{tabHC}{%
Local hyperbaryon operators. The leftmost column gives the Weyl-fermion
content, and the rightmost column the notation used for the operator.
The remaining columns list the quantum numbers.
}
\end{quotation}
\vspace*{-4.5ex}
\end{table}
The explicit breaking of $SU(5)$ generates an effective potential for $\Sigma$.
This new contribution is parametrized by one new LEC,
$C_{\rm top}$, analogous to $C_{LR}$ in Eq.~(\ref{VeffEW}). We will show
that $C_{\rm top}$ can be expressed as an integral over a hyperbaryon
four-point function convoluted with two free, massless fermion propagators.
Up to this point, our analysis is from first principles.
In Sec.~\ref{large-N} we turn to physical but non-rigorous considerations.
We show that
a large-$N$ limit exists (where $N=4$ for the hypercolor group $SU(4)$ we
consider here) in which the hyperbaryon four-point function factorizes,
leading to a simple result for $C_{\rm top}$, and ultimately to a
non-trivial expectation value for the Higgs field.
Finally, in Sec.~\ref{pheno} we comment on the phenomenological consequences
of our analysis. This includes a brief discussion of the more general
situation where the explicit breaking of $SU(3)\times SU(3)'$
to its diagonal subgroup $SU(3)_{\rm color}$ is allowed to come from
other sources than the QCD gluons.
\subsection{\label{baryons} Top quark spurions and hyperbaryons}
We begin with introducing the top-quark spurions,
a left-handed spurion $T_L$, which we choose in the {\bf 5} {\it irrep}\ of $SU(5)$,
and a right-handed spurion $T_R$, which we choose
in the ${\bf \bar{5}}$ {\it irrep}. Both {\it irreps}\ reduce to the
{\bf 5} of $SO(5)$. This choice ensures that
terms like $\overline{T}_L T_R$ in the effective potential are disallowed by
$SU(5)$, but allowed by $SO(5)$.%
\footnote{We may switch {\bf 5} with ${\bf \bar{5}}$, but the key point
is that the two spurions are chosen to be in different $SU(5)$ {\it irreps}.
}
Both $T_L$ and $T_R$ will have $U(1)_X$ charge $2/3$,
as this will yield the correct hypercharges for $q_L$ and $t_R$.
The SM values for these spurions are
\begin{equation}
\label{fspurions}
T_L = \hat{T}_L \equiv \frac{1}{\sqrt{2}}\left(\begin{array}{c}
ib_L\\beta_L\\iota t_L\\-t_L\\0
\end{array}\right)\ ,\qquad
T_R = \hat{T}_R \equiv \left(\begin{array}{c}
0\\0\\0\\0\\iota t_R
\end{array}\right)\ .
\end{equation}
Using Eq.~(\ref{embed}), it is straightforward to verify that the pair $(t_L,b_L)$
transforms as an $SU(2)_L$ doublet, and has hypercharge $Y=1/6$.
The $SU(2)_L$ singlet $t_R$ has hypercharge $Y=2/3$.
The quantum numbers of the spurions under the remaining flavor symmetries
will be discussed shortly.
The spurions couple to the hypercolor theory through
the $G$-invariant lagrangian
\begin{equation}
{\cal L}_{\rm EHC} = \lambda_1\overline{T}_L B_R +\lambda_1^*\overline{B}_R T_L
+ \lambda_2 \overline{T}_R B_L + \lambda_2^*\overline{B}_L T_R\ ,
\label{lagSU5}
\end{equation}
where $B_{L,R}$ are hyperbaryon fields with appropriate
quantum numbers. Setting the spurions $T_{L,R}$
equal to their SM values~(\ref{fspurions}) then tells us how the SM quarks
$t_L$, $b_L$ and $t_R$ couple to the hypercolor theory.
Since we will use three-hyperfermion local interpolating fields
for the hyperbaryons, the four-fermion interactions in ${\cal L}_{\rm EHC}$ have
engineering dimension six. ${\cal L}_{\rm EHC}$ originates from some other
sector with scale $\Lambda_{\rm EHC}\gg\Lambda_{\rm HC}$, with effective couplings
$\lambda_{1,2}\sim O(\Lambda_{\rm EHC}^{-2})$ just below that scale.
Limiting ourselves to local hyperbaryon fields,
all operators that can be used in the construction of a $G$-invariant
${\cal L}_{\rm EHC}$ are listed in Table~\ref{tabHC}.
The schematic structure in terms of Weyl fields
is indicated in the first column of the table, followed by the quantum numbers
under the flavor group $G$. The column labeled as $SU(3)_c$
gives the $SU(3)_{\rm color}$ {\it irrep}. The spinor index of the hyperbaryon
field is always carried by the Majorana fermion $\chi$.
Using Eqs.~(\ref{maj}) and~(\ref{DW}), explicit expressions for the
unprimed operators in Table~\ref{tabHC} are
\begin{subequations}
\label{bar4}
\begin{eqnarray}
B_{Ria} &=& -{1\over 2} \e^{ABCD} \e_{abc}\, P_R\, \chi_{ABi}\,
\left( \psi_{Cb}^T \,C P_R\, \psi_{Dc} \right)
\label{bar4a}\\
&=& {1\over 2} \e^{ABCD} \e_{abc}\, \Upsilon_{ABi}
\left(\Psi^T_{Cb}\,\e\,\Psi_{Dc}\right) \ ,
\nonumber\\
\overline{B}_{Ria} &=& {1\over 2} \e_{ABCD} \e_{abc}\, \overline{\c}^{AB}_i P_L\,
\left( \overline{\j}^C_b C P_L \Big(\overline{\j}^D_c\Big)^T \right)
\label{bar4b}\\
&=& {1\over 2} \e_{ABCD} \e_{abc}\, \bar{\U}^{AB}_i
\left(\bar{\J}^C_b\,\e\,\Big(\bar{\J}^D_c\Big)^T\right) \ ,
\nonumber\\
B_{Lia} &=& -{1\over 2} \e^{ABCD} \e_{abc}\, P_L \chi_{ABi}\,
\left( \psi_{Cb}^T \,C P_R\, \psi_{Dc} \right)
\label{bar4c}\\
&=& \e_{abc}\,\e\,\Big(\bar{\U}^{AB}_i\Big)^T
\left(\Psi^T_{Ab}\,\e\,\Psi_{Bc}\right) \ ,
\nonumber\\
\overline{B}_{Lia} &=& {1\over 2} \e_{ABCD} \e_{abc}\, \overline{\c}^{AB}_i P_R\,
\left( \overline{\j}^C_b C P_L \Big(\overline{\j}^D_c\Big)^T \right)
\label{bar4d}\\
&=& \e_{abc}\,\Upsilon^T_{ABi}\,\e
\left(\bar{\J}^A_b\,\e\,\Big(\bar{\J}^B_c\Big)^T\right) \ .
\end{eqnarray}
\end{subequations}
The primed operators in Table~\ref{tabHC} are obtained from Eq.~(\ref{bar4})
by interchanging $P_R\leftrightarrow P_L$ inside the $\psi\j$
and $\overline{\j}\bj$ bilinears.
When the spurions $T_{L,R}$ are restricted to their SM values,
the phases of $\lambda_{1,2}$ in Eq.~(\ref{lagSU5})
can be removed by (non-anomalous) $SU(2)_L$ and
$SU(2)_R$ transformations on the spurion fields,
implying that, from now on, we may take $\lambda_{1,2}$ to be real and positive.
This allows us to require that the lagrangian~(\ref{lagSU5}) be $CP$ invariant.
The $CP$ transformation acts as
\begin{equation}
\psi \to \gamma_2 \overline{\j}^T \ , \qquad \overline{\j} \to \psi^T \gamma_2 \ ,
\label{CP}
\end{equation}
for both Dirac and Majorana fermions (see App.~\ref{conv} for our Dirac matrices
conventions). The sign choices we have made in Eq.~(\ref{bar4})
imply that a $CP$ transformation applied
to the elementary fields $\chi$, $\psi$ and $\overline{\j}$ induces a $CP$ transformation
of the same form
on the hyperbaryon fields as well, thereby ensuring the $CP$ invariance
of ${\cal L}_{\rm EHC}$.
The unprimed fields in Table~\ref{tabHC} transform non-trivially
under $SU(3)$ and are singlets under $SU(3)'$, whereas for the primed fields
the opposite is true. Choosing either the primed or the unprimed version
for each hyperbaryon field gives rise to a total of four different
possibilities for ${\cal L}_{\rm EHC}$.
The quantum numbers of the spurions $T_L$ and $T_R$
are chosen accordingly, so as to ensure the $G$ invariance of ${\cal L}_{\rm EHC}$.
The SM quark fields $q_L$ and $t_R$ are endowed with same
$SU(3)\times SU(3)'\times U(1)_X\times U(1)'$ quantum numbers
as their parent spurion. This construction is consistent with the SM,
since the resulting quantum numbers under $SU(3)_{\rm color}$
will always be the same. In addition, it follows that
the entire theory, including the hypercolor sector, the SM lagrangian,
and their coupling via ${\cal L}_{\rm EHC}$, is invariant under both $SU(3)$
and $SU(3)'$, provided that the QCD interactions can be neglected.
This is indeed the case in this section, because we calculate the Higgs
effective potential to second order in all SM couplings,
and since the result will be quadratic in the top Yukawa coupling,
any corrections that involve an additional dependence
on the QCD coupling $g_{\rm s}$ are neglected.
Note that we cannot generalize Eq.~(\ref{lagSU5}) to include,
simultaneously, terms that couple a given spurion to both of
the unprimed and primed hyperbaryons, as this will not allow for any
consistent assignment of $SU(3)\times SU(3)'$ quantum numbers.
For example, $\overline{T}_L$ can couple to either $B_R$ or $B'_R$, but not to both.
It is because of this fact that ${\cal L}_{\rm EHC}$ depends on only two
coupling constants $\lambda_{1,2}$. This will lead to considerable simplification
in our analysis. In Sec.~\ref{pheno} we briefly comment on the
more general case, where ${\cal L}_{\rm EHC}$ is restricted only by
$SU(3)_{\rm color}$.
\subsection{\label{top-mass} The top quark Yukawa coupling}
Our next task is to construct the electroweak effective field theory.
As a first step, we integrate only over the gauge fields and fermions
of the hypercolor theory, and obtain an effective theory that depends on the
spurions $T_L$ and $T_R$, and on the non-linear fields,
including in particular the $SU(5)/SO(5)$ field $\Sigma$.
We assume that the electroweak scale $m_W\sim m_t$ is much smaller
than the hypercolor scale $f\sim M\sim\Lambda_{\rm HC}$, where $M$ is
of order the mass of the hyperbaryons which are assumed
to couple to the top quark in Eq.~(\ref{lagSU5}). This provides us with
a power counting, and, in particular, the effective theory can be
organized according to a derivative expansion. We will be concerned
with the lowest non-trivial order in this expansion.
Demanding full $G$ invariance, the leading order spurion potential is
\begin{equation}
\label{topmasseff}
V_{\rm top} = \mu_L\Phi^2\,\overline{T}_R\Sigma^*T_L + \mu_R\Phi^{-2}\,\overline{T}_L\Sigma T_R\ .
\end{equation}
Terms like $\overline{T}_L T_L$ vanish because of chiral projectors,
while terms like $\overline{T}_L\overline{T}_R^T$ are not allowed by $U(1)_X$ symmetry.
Bilinear terms independent of $\Sigma$ are possible,
but thanks to $SU(5)$ invariance, they
have an $LL$ or $RR$ structure, and need an insertion of $\gamma_\mu$.
Therefore, they contain at least one derivative, and their role is to
renormalize the kinetic terms for the top and bottom quarks,
which are present when these SM fields are made dynamical.
It can be checked that the correction is of order $y$,
where $y$ is the top quark Yukawa coupling introduced in Eq.~(\ref{y}) below.
$V_{\rm top}$ depends on two effective fields, $\Sigma$ and $\Phi$.
The role of $\Phi$ is to reinstate $U(1)'$ invariance ({\it cf.} Eq.~(\ref{transG})).
When we choose both hyperbaryons in Eq.~(\ref{lagSU5}) to be unprimed ones,
the hyperbaryons and the spurions
transform non-trivially only under $SU(3)$, and are singlets of $SU(3)'$.
Therefore, $V_{\rm top}$ is invariant under $SU(3)\times SU(3)'$ as it stands,
without having to introduce any dependence on the effective field $\Omega$.
If we set $\Sigma=\Phi=1$, and substitute the SM values
$\hat{T}_L$ and $\hat{T}_R$ defined in Eq.~(\ref{fspurions}) for $T_L$ and $T_R$,
we find that $V_{\rm top}$ vanishes.
$\Sigma$ will need to develop a non-trivial expectation value for the SM
top quark to acquire a non-zero mass.
This will be discussed in Sec.~\ref{top-effpot} below.
In order to find the LECs $\mu_{L,R}$ in Eq.~(\ref{topmasseff}),
we consider the second derivatives
\begin{equation}
\label{LRder}
\frac{\partial^2}{\partial T_L(y)\partial\overline{T}_R(x)}\log{Z}\ ,\qquad
\frac{\partial^2}{\partial T_R(y)\partial\overline{T}_L(x)}\log{Z}\ ,
\end{equation}
where $Z$ is the partition function of either the effective or
the microscopic theory. Requiring the effective theory to match
the microscopic theory (and noting that the fermionic spurions are Grassmann)
yields the relations
\begin{eqnarray}
\label{mus}
-\mu_L P_L \langle\Phi^2\Sigma^*\rangle\delta(x-y) + \cdots
&=&\lambda_1\lambda_2 P_L\langle B_L(x)\overline{B}_R(y)\rangle P_L\ ,
\\
-\mu_R P_R \langle\Phi^{-2}\Sigma\rangle\delta(x-y) + \cdots
&=&\lambda_1\lambda_2 P_R\langle B_R(x)\overline{B}_L(y)\rangle P_R\ .
\nonumber
\end{eqnarray}
The ellipses on the left-hand side indicate that
the leading-order low-energy theory given by $V_{\rm top}$
reproduces the correlation functions on the right-hand side only to leading
order in a derivative expansion.
By assumption, symmetry breaking in the hypercolor theory yields
$\svev{\overline{\c}_i\chi_j}\propto\svev{\Sigma_{ij}}=\delta_{ij}$,
up to a symmetry transformation. (When the SM fields become dynamical
we may in general have $\svev{\Sigma_{ij}}\ne\delta_{ij}$, but these
corrections are of higher order in the SM gauge and Yukawa couplings.)
Setting $\svev{\Sigma_{ij}}=\delta_{ij}$ and $\svev{\Phi}=1$
in Eq.~(\ref{mus}) provides us with expressions for the parameters $\mu_{L,R}$.
Assembling the chiral baryon fields together as
\begin{equation}
\label{deftB}
B = B_R+B_L\ , \qquad \overline{B} = \overline{B}_R + \overline{B}_L \ ,
\end{equation}
where $B$ is a Dirac field with quantum numbers $({\bf 5},{\bf 3})$
under the unbroken $SO(5)\times SU(3)_{\rm color}$, and writing
\begin{eqnarray}
\label{FT}
\delta(x-y) &=& \int\frac{d^4p}{(2\p)^4}\,e^{ip(x-y)}\ ,
\\
\svev{B(x)\overline{B}(y)}
&=& \int\frac{d^4p}{(2\p)^4}\,e^{ip(x-y)}S_{B}(p)\ ,
\nonumber
\end{eqnarray}
we find,
to leading order in the momentum expansion of the effective theory,
\begin{eqnarray}
\label{lecs}
\mu_L P_L &=& -\lambda_1\lambda_2 P_LS_{B}(0)P_L\ ,
\\
\mu_R P_R &=& -\lambda_1\lambda_2 P_RS_{B}(0)P_R\ .
\nonumber
\end{eqnarray}
Apart from the chiral projectors, these two expressions must be equal, because
any hyperbaryon fields occurring in Eq.~(\ref{lagSU5}) can only have
a Dirac mass; Majorana masses (such as $B_R^T \e B_R$ or $B_L^T \e B_L$)
are forbidden by $U(1)_X$ symmetry.
Therefore,
\begin{equation}
\label{mu}
\mu=\mu_L=\mu_R=-\lambda_1\lambda_2 S_{B}(0)\ ,
\end{equation}
where we have used that at zero momentum $S_{B}(0)$ is proportional to
the unit matrix in spinor space.
We may now introduce the top quark Yukawa coupling $y$ by writing
\begin{equation}
\mu=yf/2 \ ,
\label{y}
\end{equation}
with $f$ the decay constant of the hypercolor theory.
If the Higgs field will now develop a non-zero expectation value,
\begin{equation}
\label{Hh}
\langle H_0\rangle = \langle H_0^\dagger\rangle = h/\sqrt{2} \ ,
\end{equation}
this will induce a top quark mass
\begin{equation}
\label{topmass}
m_t = \frac{1}{2\sqrt{2}}\,yf\sin{(2h/f)}\approx \frac{1}{\sqrt{2}}\,yh\ ,
\end{equation}
where we have used Eqs.~(\ref{Sigma}),~(\ref{H}),~(\ref{fspurions})
and~(\ref{topmasseff}), and the approximate equality holds for $h/f\ll 1$.
We may introduce an effective hyperbaryon field $\tilde{B}$ which is
canonically normalized by writing
\begin{equation}
\label{B}
B = f^3\sqrt{Z_B} \tilde{B}\ .
\end{equation}
We define $M$ as the zero-momentum mass of the canonically normalized
$\overline{B}$ field, corresponding to a term $M\overline{\tB} \tilde{B}$.
In other words, $S_{B}(0) = f^6 Z_B/M$. This gives
\begin{equation}
\label{yukawa}
y = -2\lambda_1\lambda_2\,Z_B\,f^5/M\ .
\end{equation}
We comment that the field $B$ does not necessarily correspond to any
baryon mass eigenstate of the hypercolor theory. Still,
generically we might expect it to couple to the lightest
hyperbaryon with quantum numbers that match those of the SM quarks,
in which case $M$ will be a quantity of the order of this smallest
hyperbaryon mass.
We conclude this subsection with a technical comment.
If in Eq.~(\ref{lagSU5}) we choose one unprimed and one primed hyperbaryon field,
this implies that one of the spurions transforms non-trivially
under $SU(3)$ while the other under $SU(3)'$.
In this case, $V_{\rm top}$ will depend on $\Omega$.
For definiteness, replacing $B_L$ in Eq.~(\ref{lagSU5}) by $B'_L$
implies that now $T_R$ transforms non-trivially under $SU(3)'$,
and Eq.~(\ref{topmasseff}) gets replaced by
\begin{equation}
\label{veffprimea}
V_{\rm top} = \mu_L\Phi^{-14/3}\,\overline{T}_R \Sigma^* \Omega^\dagger T_L
+ \mu_R\Phi^{14/3}\,\overline{T}_L\Sigma \Omega T_R\ .
\end{equation}
This hardly changes our analysis, because,
in order to obtain expressions for the parameters $\mu_{L,R}$
in Eq.~(\ref{topmasseff}) we are setting all non-linear fields
equal to the identity anyway.
In the next subsection, we will work out
the effective potential after integrating over the third-generation quarks.
In this calculation, any dependence on both $\Omega$ and $\Phi$ will drop out
regardless of our choice of hyperbaryon fields in Eq.~(\ref{lagSU5}), as it must be
because both $SU(3)\times SU(3)'$ and $U(1)'$ are not explicitly broken
in the top sector to the order we work, and therefore no effective potential
can be generated for those non-linear fields.
In particular, when $V_{\rm top}$ depends on $\Omega$ as in Eq.~(\ref{veffprimea}),
then expression~(\ref{vefftermsc}) below,
which is the only term that will contribute to the effective potential,
gets multiplied by $\,{\rm tr}(\Omega\O^\dagger)=\,{\rm tr}\,{\bf 1}$, showing that indeed
the $\Omega$ dependence cancels out.
\subsection{\label{top-effpot} Higgs effective potential induced by the top quark}
We now integrate over the SM top quark in order to obtain the associated
contribution $V_{\rm eff}^{\rm top}(\Sigma)$ to the effective potential.
Adding this to Eq.~(\ref{VeffEW}) gives the complete effective potential
for $\Sigma$ to second order in the SM gauge and Yukawa couplings.
We will disregard all the other SM fermions, including the bottom quark,
on the grounds that their Yukawa couplings are much smaller,
and so their contribution to the
effective potential will be much smaller as well.
We begin by splitting the spurions $T_{L,R}$ as follows%
\footnote{Whether or not $t_L$ in Eq.~(\ref{fsrewrite}) coincides with
the component with the same name of $\hat{T}_L$ in Eq.~(\ref{fspurions}) depends
on the value we choose for $v_L$, as we will see below.
}
\begin{equation}
\label{fsrewrite}
T_L(x) = t_L(x) v_L\ ,\qquad T_R(x) = t_R(x) v_R\ .
\end{equation}
The new global spurions $v_{L,R}$ carry the $SU(5)$ quantum numbers, which
contains the SM symmetry $SU(2)_L$. We also assign $U(1)_X$ to these spurions,
because the hypercharge $Y$ is the sum of the charge $X$ and the third component
of $SU(2)_R$, with the latter being a subgroup of $SU(5)$ as well.
The (Grassmann) fields $t_{L,R}$ carry the spin index.
They also inherit the $SU(3)$ and $SU(3)'$ quantum numbers from $T_{L,R}$.
We promote $t_{L,R}$ to dynamical fields by adding tree-level kinetic terms
$\overline{t}_L {\sl\partial} t_L + \overline{t}_R {\sl\partial} t_R$.
The effective potential $V_{\rm eff}^{\rm top}$ at order $y$
is obtained by substituting Eq.~(\ref{fsrewrite}) into $V_{\rm top}$
of Eq.~(\ref{topmasseff}), and integrating over the top quark, leading to
a contribution with the form of $\Phi^{-2} \overline{v}_L \Sigma v_R + \mbox{h.c.}$.
This contribution vanishes, however,
because the only non-zero tree-level top propagators
are $\svev{t_L \overline{t}_L}$ and $\svev{t_R \overline{t}_R}$.
The leading contribution to $V_{\rm eff}^{\rm top}$ is of order $y^2$.
It involves four global spurions.
Momentarily suppressing any dependence on the $\Omega$ and $\Phi$ fields,
the possible terms that depend on $\Sigma$ are
\begin{subequations}
\label{veffterms}
\begin{eqnarray}
\label{vefftermsa}
(\overline{v}_L\Sigma v_R)^2+\mbox{h.c.}\ ,
\\
\label{vefftermsb}
\quad (\overline{v}_R\Sigma^* v_L)^2+\mbox{h.c.}\ ,
\\
\label{vefftermsc}
(\overline{v}_L\Sigma v_R)(\overline{v}_R\Sigma^* v_L)\ .
\end{eqnarray}
\end{subequations}
The tree-level top propagators allow for the generation of the last term only.
If the effective potential~(\ref{vefftermsc}) arises as the product
of the two interactions in Eq.~(\ref{topmasseff}),
the $\Phi$ dependence evidently cancels out. Moreover, as we have explained in
the previous section, regardless of the choice of hyperbaryon fields
we make in Eq.~(\ref{lagSU5}), $V_{\rm eff}^{\rm top}$ will be
independent of $\Omega$ and $\Phi$, because, to the order we are working,
$SU(3)$, $SU(3)'$ and $U(1)'$ are not broken explicitly.
Promoting the fields $t_L$ and $t_R$ in Eq.~(\ref{fspurions})
to be dynamical amounts to setting
\begin{equation}
\label{vLvR}
v_L = \hat{v}_L \equiv \frac{1}{\sqrt{2}}\left(\begin{array}{c}
0\\0\\iota \\-1\\0
\end{array}\right)\ ,\qquad
v_R = \hat{v}_R \equiv \left(\begin{array}{c}
0\\0\\0\\0\\iota
\end{array}\right)\ ,
\end{equation}
with $\hat{\overline{v}}_{L,R} = \hat{v}_{L,R}^\dagger$.
The resulting contribution to the effective potential is
\begin{equation}
\label{vtopeff}
y^2 C_{\rm top}\,(\hat{\overline{v}}_L\Sigma \hat{v}_R)(\hat{\overline{v}}_R\Sigma^* \overline{v}_L)
= \frac{y^2}{2}\,C_{\rm top}
\left(\Sigma_{35}-i\Sigma_{45}\right) \left(\Sigma_{35}^*+i\Sigma_{45}^*\right)\ .
\end{equation}
$C_{\rm top}$ is a new LEC. We have factored out the square of
the Yukawa coupling $y$ to make explicit the order at which we work.
However, we are not done yet. In order to arrive at this result
we have used Eq.~(\ref{vLvR}) for the global spurions, which projects onto
a particular component of the $SU(2)_L$ doublet $q_L$, the one denoted $t_L$
in Eq.~(\ref{fspurions}). In order to add the contribution of the other component,
denoted $b_L$, we replace $\hat{v}_L$ of Eq.~(\ref{vLvR}) by $(i,1,0,0,0)^T/\sqrt{2}$
(the right-handed singlet spurion $\hat{v}_R$ is unchanged).
Adding the two contributions together we arrive at the $SU(2)_L$ invariant
effective potential
\begin{equation}
\label{vtopeffb}
V_{\rm eff}^{\rm top} = \frac{y^2}{2}\,C_{\rm top}
\left( |\Sigma_{35}-i\Sigma_{45}|^2 + |\Sigma_{15}+i\Sigma_{25}|^2 \right)\ .
\end{equation}
This is the leading contribution of dynamical third-generation quarks
to the effective potential.
Expanding the non-linear $\Sigma$ field to quadratic order in the NGB fields gives
\begin{equation}
\label{massNGBstop}
V_{\rm eff}^{{\rm top}(2)} = 4y^2 \frac{C_{\rm top}}{f^2}\,H^\dagger H\ .
\end{equation}
When the Higgs field $H=(H_+,H_0)^T$ acquires an expectation value,
conventionally it is assigned to the lower component $H_0$, as in Eq.~(\ref{Hh}).
This selects $t_L$ (rather than any other linear combination
of the doublet fields $t_L$ and $b_L$) as the left-handed field that,
together with the right-handed field $t_R$, forms the physical top quark.
Let us pause to consider these results.
The full effective potential $V_{\rm eff}(\Sigma)$
is the sum of Eqs.~(\ref{VeffEW}) and~(\ref{vtopeffb}).
If $C_{\rm top}$ is positive, the global minimum
is attained for $\Sigma=\Sigma_0={\bf 1}$ (with $\Sigma_{i5}=0$ for $i=1,\ldots,4$).
For electroweak symmetry breaking to take place, $C_{\rm top}$
must therefore be negative. To second order in the NBG fields,
the effective potential is the sum of Eqs.~(\ref{massNGBsEW}) and~(\ref{massNGBstop}).
As already observed in Ref.~\cite{ferretti}, the curvature at the origin
can become negative only in the direction of the $H$ field. This happens when
\begin{equation}
\label{condition}
\frac{C_{\rm top}}{C_{LR}}
< -\frac{3g^2+g'^2}{2y^2}
= -\frac{2m_W^2+m_Z^2}{m_t^2}
\approx -0.7 \ ,
\end{equation}
triggering a non-zero expectation value for the Higgs field.
If we use Eq.~(\ref{Hh}) and, moreover, assume that all other NGB fields
in Eq.~(\ref{Pi}) remain zero, the total effective potential is
\begin{equation}
\label{Veffh}
V_{\rm eff}(h) = -C_{LR}\left(3g^2+g'^2\right)\cos^2{(h/f)}
+ \frac{y^2}{2}\,C_{\rm top}\,\sin^2{(2h/f)}\ .
\end{equation}
While the global minimum of $V_{\rm eff}(\Sigma)$
must occur at non-zero $h$ if Eq.~(\ref{condition}) is satisfied,
due to the complexity of $V_{\rm eff}(\Sigma)$ we have not been able to prove that,
given arbitrary values of the SM couplings or the LECs, the global minimum
will never involve non-zero expectation values for any other NGBs.
We next turn to the calculation of the low-energy constant $C_{\rm top}$.
As in the previous subsection, this is done by matching the
effective potential~(\ref{vtopeffb}) to the underlying theory
with top-hyperbaryon couplings as given in Eq.~(\ref{lagSU5}).
The difference is that now we also integrate over the SM top quark field.
After splitting the $T_{L,R}$ spurions as in Eq.~(\ref{fsrewrite}),
the matching will involve taking four derivatives, one with respect
to each of the global spurions $v_{L,R}$ and $\overline{v}_{L,R}$.%
\footnote{We will discuss later on what values to choose for the
global spurions in Eq.~(\ref{fsrewrite}), or equivalently, with respect
to which component of each global spurion one would
choose to differentiate.
}
Once again we will set $\Sigma={\bf 1}$.
This implies that we must take into account terms independent of $\Sigma$
that have a similar dependence on the global spurions as Eq.~(\ref{vefftermsc}).
There are two such terms,
\begin{equation}
\label{Sindep}
y^2 C_1(\overline{v}_L v_L)(\overline{v}_R v_R) + y^2 C_2(\overline{v}_L\overline{v}_R^T)(v_R^T v_L)\ ,
\end{equation}
where we introduced new LECs $C_1$ and $C_2$, and, for convenience,
separated out a factor of $y^2$, as we did in Eq.~(\ref{vtopeff}).
Taking the four derivatives in the effective theory we find, after
setting $\Sigma={\bf 1}$,
\begin{equation}
\label{d4Zeff}
\frac{\partial^4}{\partial\overline{v}_{Li}\partial v_{Rj}\partial\overline{v}_{Rk}
\partial v_{L\ell}} \log Z_{\rm eff}
=
-y^2 V \left(C_{\rm top}\,\delta_{ij}\delta_{k\ell}+C_1\,\delta_{i\ell}\delta_{jk}
+C_2\,\delta_{ik}\delta_{j\ell}\right)\ ,
\end{equation}
where $V$ is the volume. In the microscopic theory
we find
\begin{eqnarray}
\label{d4ZUV}
&& \hspace{-10ex}
\frac{\partial^4}{\partial\overline{v}_{Li}\partial v_{Rj}\partial\overline{v}_{Rk}
\partial v_{L\ell}} \log{Z}
\\
&=&(\lambda_1\lambda_2)^2
\int d^4x_1\,d^4x_2\,d^4x_3\,d^4x_4
\nonumber\\
&& \times \svev{(\overline{B}_{R\ell} t_L)(x_4)(\overline{t}_L B_{Ri})(x_1)
(\overline{B}_{Lj} t_R)(x_2)(\overline{t}_R B_{Lk})(x_3)}
\nonumber\\
&=&-(\lambda_1\lambda_2)^2\int d^4x_1\,d^4x_2\,d^4x_3\,d^4x_4
\int \frac{d^4p}{(2\p)^4}\frac{d^4q}{(2\p)^4}\,
\frac{p_\mu}{p^2}\,\frac{q_\nu}{q^2}\, e^{ip(x_4-x_1)+iq(x_2-x_3)}
\nonumber\\
&& \times \svev{\left(\overline{B}_{R\ell}(x_4)\gamma_\mu P_RB_{Ri}(x_1)\right)
\left(\overline{B}_{Lj}(x_2)\gamma_\nu P_LB_{Lk}(x_3)\right)}\ .
\nonumber
\end{eqnarray}
In the last equality we have integrated over the top quark,
substituting free massless fermion propagators for its two-point functions.
The remaining expectation value on the last line
is to be computed in the pure hypercolor theory. We may now project onto the
$C_{\rm top}$ term in Eq.~(\ref{d4Zeff}) by choosing $i=j\ne k=\ell$, obtaining
\begin{eqnarray}
\label{Ctopmain}
C_{\rm top} &=&
\frac{(\lambda_1\lambda_2)^2}{y^2}\frac{1}{V}\int d^4x_1\,d^4x_2\,d^4x_3\,d^4x_4
\int \frac{d^4p}{(2\p)^4}\frac{d^4q}{(2\p)^4}\,
\frac{p_\mu}{p^2}\,\frac{q_\nu}{q^2}\, e^{ip(x_4-x_1)+iq(x_2-x_3)}
\nonumber\\
&& \times \svev{\left(\overline{B}_{Rk}(x_4)\gamma_\mu P_RB_{Ri}(x_1)\right)
\left(\overline{B}_{Li}(x_2)\gamma_\nu P_LB_{Lk}(x_3)\right)}_{i\ne k}\ .
\end{eqnarray}
In terms of the Fourier transform
\begin{eqnarray}
\label{FTB}
&& \hspace{-15ex}
\svev{\left(\overline{B}_{R\ell}(k_4)\gamma_\mu P_RB_{Ri}(k_1)\right)
\left(\overline{B}_{Lj}(k_2)\gamma_\nu P_LB_{Lk}(k_3)\right)}
\\
&=& \int d^4x_1\,d^4x_2\,d^4x_3\,d^4x_4\,e^{-ik_1x_1+ik_2x_2-ik_3x_3+ik_4x_4}
\nonumber\\
&& \times \svev{\left(\overline{B}_{R\ell}(x_4)\gamma_\mu P_RB_{Ri}(x_1)\right)
\left(\overline{B}_{Lj}(x_2)\gamma_\nu P_LB_{Lk}(x_3)\right)} \ ,
\nonumber
\end{eqnarray}
we may write this in momentum space as%
\footnote{In finite volume, the momentum integral $\int d^4p/(2\p)^4$
is to be understood as a momentum average, $V^{-1} \sum_{p_\mu}$.
Alternatively, in infinite volume, $V$ is to be interpreted
as $(2\p)^4 \delta(0)$ in momentum space.
}
\begin{equation}
\label{Ctopfinal}
C_{\rm top}
= \frac{(\lambda_1\lambda_2)^2}{y^2}\frac{1}{V}\!
\int\frac{d^4p}{(2\p)^4}\frac{d^4q}{(2\p)^4}\,
\frac{p_\mu}{p^2}\,\frac{q_\nu}{q^2}
\svev{\left(\overline{B}_{Rk}(p)\gamma_\mu P_RB_{Ri}(p)\right)
\!\left(\overline{B}_{Li}(q)\gamma_\nu P_LB_{Lk}(q)\right)}_{i\ne k}\ .
\end{equation}
This is our main result.
The top sector effective potential, Eq.~(\ref{vtopeffb}),
depends on the experimentally known top Yukawa coupling, and on $C_{\rm top}$.
Using Eq.~(\ref{yukawa}), we may reexpress the ratio $\lambda_1\lambda_2/y$
in Eq.~(\ref{Ctopfinal}) in terms of quantities that are calculable
in the pure hypercolor theory. This determines $V_{\rm eff}^{\rm top}$
completely. The dependence on the extended hypercolor sector,
coming from the couplings $\lambda_{1,2}$, has dropped out.
Diagrammatically, each term on the right-hand side of Eq.~(\ref{d4Zeff})
originates from diagrams of the microscopic theory with a distinct topology.
This is shown in Fig.~\ref{topB}, where we have kept only
the propagators of the top quark (solid lines) and of the Majorana fermions
$\chi$ (dashed-dot lines). All other fields, including the Dirac fermions
in the fundamental {\it irrep}\ of hypercolor, have been suppressed.
With these conventions, $C_{\rm top}$ arises from the class
of diagrams represented by Fig.~\ref{topB}(a), while $C_1$ and $C_2$
arise from Fig.~\ref{topB}(b) and \ref{topB}(c) respectively.
\begin{figure}[t]
\begin{center}
\includegraphics*[height=4.5cm]{topBa.eps}
\hspace{8ex}
\includegraphics*[height=4.5cm]{topBb.eps}
\hspace{8ex}
\includegraphics*[height=4.5cm]{topBc.eps}
\end{center}
\begin{quotation}
\floatcaption{topB}%
{The three possible Majorana-fermion contractions contributing to
Eq.~(\ref{d4ZUV}), corresponding to (a) $C_{\rm top}$, (b) $C_1$, and (c) $C_2$.
Only the top quark (solid lines) and Majorana fermions (dashed lines)
are shown. The vertices $x_1,\ldots,x_4$ of Eq.~(\ref{d4ZUV})
correspond to a clock-wise motion starting at the lower-left corner.}
\end{quotation}
\vspace*{-4ex}
\end{figure}
\begin{boldmath}
\subsection{\label{large-N} Large-$N$ estimate of $y^2C_{\rm top}$}
\end{boldmath}
Determining $C_{\rm top}$ using Eq.~(\ref{Ctopfinal}) requires knowledge
of the ratio $\lambda_1\lambda_2/y$, and a strong-coupling calculation
that can be done using lattice gauge theory.
Such a lattice calculation, however, would be a major undertaking
(see Sec.~\ref{lattice}).
In this subsection, we resort to analytic techniques hoping to shed some light
on the most interesting question, which is whether $C_{\rm top}$ could
indeed be negative, and large enough in size to cause electroweak
symmetry breaking.
We will first consider what can be said if
we assume that the hyperbaryon four-point function in Eq.~(\ref{Ctopfinal})
factorizes into the product of two hyperbaryon two-point functions.
We will show that $C_{\rm top}$ is negative in this case.
We will then argue that a large-$N$ limit exists in which the factorized
contribution dominates, and thus the Higgs field acquires a non-zero
expectation value.
Assuming factorization, and using%
\footnote{On the right-hand side, $\delta_{ij}$ follows from $SO(5)$ invariance.
}
\begin{eqnarray}
\label{hbprops}
P_R\langle B_{Ri}(p)\overline{B}_{Lj}(q)\rangle P_R
&=& (2\p)^4 \delta(p-q) \delta_{ij} P_RS_{B}(p)P_R\ ,
\\
P_L\langle B_{Li}(q)\overline{B}_{Rj}(p)\rangle P_L
&=& (2\p)^4 \delta(p-q) \delta_{ij} P_LS_{B}(p)P_L\ ,
\nonumber
\end{eqnarray}
where $S_{B}(p)$ was defined in Eqs.~(\ref{deftB}) and~(\ref{FT}),
Eq.~(\ref{Ctopfinal}) leads to
\begin{equation}
\label{Ctopfact}
C_{\rm top}^{\rm fact}
= -\frac{(\lambda_1\lambda_2)^2}{y^2}
\int \frac{d^4p}{(2\p)^4}\,\frac{p_\mu p_\nu}{(p^2)^2}\,
\,{\rm tr}\left(\gamma_\mu P_RS_{B}(p)\gamma_\nu P_LS_{B}(p)\right)\ .
\end{equation}
Using a dispersive representation for the hyperbaryon propagator,
\begin{equation}
\label{baryondisp}
S_{B}(p)=\int_0^\infty \frac{ds}{2\p}\,\r(s)\,\frac{-i\sl{p}+\sqrt{s}}{p^2+s}\ ,
\end{equation}
with $\r(s)\ge 0$ for all $s\ge 0$, this becomes
\begin{equation}
\label{Ctopfactdisp}
C_{\rm top}^{\rm fact}
= -\frac{1}{8\p^2}\,\frac{(\lambda_1\lambda_2)^2}{y^2}
\int_0^\infty \frac{ds}{2\p}\int_0^\infty
\frac{dt}{2\p}\,\r(s)\r(t)\,\frac{\sqrt{ts}}{t-s}\,\log{\frac{t}{s}}\ .
\end{equation}
This result is negative, because the integrand is manifestly positive.%
\footnote{Only the factorizable contribution appears to have been
considered in Refs.~\cite{RC,ferretti}.
}
As a simple example, if we take $\r(s)=2\p f^6Z_B\delta(s-M^2)$,
Eq.~(\ref{Ctopfactdisp}) reduces to
\begin{equation}
\label{Ctopfactexpl}
C_{\rm top}^{\rm fact}
= -\frac{1}{8\p^2}\frac{(\lambda_1\lambda_2)^2}{y^2}\,f^{12}Z_B^2
= -\frac{1}{32\p^2}\,f^2 M^2\ ,
\end{equation}
where we used Eq.~(\ref{yukawa}).
We next consider a large-$N$ limit in which factorization
can be shown to hold. Of course,
in the model of Ref.~\cite{ferretti}, the number of (hyper)colors is $N=4$.
What makes this generalization non-trivial is that, unlike the $SU(4)$ case
where the two-index antisymmetric {\it irrep}\ is real, for any $N>4$ this
{\it irrep}\ is complex. This means that the Majorana condition~(\ref{majb})
cannot be imposed without violating gauge invariance.
In order to cope with this,
in addition to Weyl fields $\Upsilon_{ABi}$ in the antisymmetric
representation of $SU(N)$, we introduce Weyl fields $\tilde\U_i^{AB}$
belonging to the {\it irrep}\ made out of the antisymmetrized product
of two anti-fundamentals.%
\footnote{This is the same as the antisymmetric product of $N-2$ fundamentals,
since $\tilde\U^{AB} \sim \e^{A_1 A_2\dots A_{N-2} AB} \Upsilon'_{A_1 A_2\dots A_{N-2}}$.
}
Instead of Majorana fermions, we now construct Dirac fermions
out of these Weyl fields according to
\begin{eqnarray}
\label{Diracchi}
\tilde\o_{ABi} &=& \left(\begin{array}{c}
\Upsilon_{ABi}\\ \e\bar{\tilde\U}_{ABi}^T
\end{array}\right)\ ,
\\
\bar{\tilde\o}^{AB}_i &=& \left(\begin{array}{cc}
-(\tilde\U^{AB}_i)^T\e& \bar{\U}^{AB}_i
\end{array}\right)\ ,
\nonumber
\end{eqnarray}
as well as their charge conjugates
\begin{eqnarray}
\label{defomega}
\omega^{AB}_i&=&C(\bar{\tilde\o}^{AB}_i)^T_i=\left(\begin{array}{c}
\tilde\U^{AB}_i\\ \e(\bar{\U}^{AB}_i)^T
\end{array}\right)\ ,\\
\overline{\o}_{ABi}&=&\tilde\o^T_{ABi} C=\left(\begin{array}{cc}
-(\Upsilon_{ABi})^T\e& \bar{\tilde\U}_{ABi}
\end{array}\right)\ .\nonumber
\end{eqnarray}
The index $i=1,\ldots,N_f$, now counts the number of Dirac fermions.
Going back to the $SU(4)$ theory, we have been forced to consider
an even number, $2N_f$, of Majorana fermions. There is no
large-$N$ generalization that would involve the desired odd number
of five Majorana fermions for the $SU(4)$ theory. Even more,
for any $N>4$ the symmetry breaking pattern becomes that of complex-{\it irrep}\
Dirac fermions, namely, $SU(N_f)\times SU(N_f)\to SU(N_f)$ \cite{vac}.
We have shown in Fig.~\ref{topB} the contractions of the
antisymmetric-{\it irrep}\ fermions that contribute to $C_{\rm top}$.
According to Eq.~(\ref{Ctopfinal}), we should choose fixed values $i\ne k$.
The minimal number of Dirac fermions we need in order to distinguish
Fig.~\ref{topB}(a) from Figs.~\ref{topB}(b) and~\ref{topB}(c)
is two, and so we will take
$N_f=2$ Dirac fermions in the antisymmetric {\it irrep}. For $N=4$,
this is equivalent to a theory with four Majorana fermions (instead of five).
This is the best we can do in terms of a large-$N$ generalization.
According to Eq.~(\ref{Ctopfinal}), for the left side of Fig.~\ref{topB}(a)
we need the contraction $P_R\svev{\omega_i\overline{\o}_i}P_R$,
whereas for the right side we need $P_L\svev{\omega_k\overline{\o}_k}P_L$.
Once these contractions have been fixed we can drop the indices $i$ and $k$,
and forget about the flavor index. This suggests that the
number of flavors is not crucial if our goal is to obtain a large-$N$ estimate
of the class of diagrams depicted in Fig.~\ref{topB}(a),
and thus that the necessary transition from Majorana fermions
to Dirac fermions for $N>4$ is inconsequential.
We are now ready to give the generalization of the hyperbaryon operators.
In the $N>4$ theory with $N_f=2$ Dirac flavors they are defined by
\begin{eqnarray}
\label{cornerslN}
B_{Ria} &=& \e_{abc}\,P_R\omega_i^{AB}\,(\Psi_{Ab}^T\,\e\,\Psi_{Bc})\ ,
\\
\overline{B}_{Ria} &=& \e_{abc}\,\overline{\o}_{iAB}P_L\,\Big(\bar{\J}^{Ab}\,\e\,(\bar{\J}^{Bc})^T\Big)\ ,
\nonumber\\
B_{Lia} &=& \e_{abc}\,P_L\omega_i^{AB}\,(\Psi_{Ab}^T\,\e\,\Psi_{Bc})\ ,
\nonumber\\
\overline{B}_{Lia} &=& \e_{abc}\,\overline{\o}_{iAB}P_R\,\Big(\bar{\J}^{Ab}\,\e\,(\bar{\J}^{Bc})^T\Big)\ .
\nonumber
\end{eqnarray}
We have used index conventions similar to the previous sections.
For $N=4$ we may impose the Majorana condition $\omega=\tilde\o$,
and then these definitions reproduce Eq.~(\ref{bar4}).
\begin{figure}[t]
\begin{center}
\includegraphics*[height=5cm]{baryonN.eps}
\end{center}
\begin{quotation}
\floatcaption{baryonN}%
{A hyperbaryon in large $N$: diagrams without hypergluons (a,b)
and with them (c,d). Diagrams (a) and (c) are leading, while
(b) and (d) are subleading.}
\end{quotation}
\vspace*{-4ex}
\end{figure}
Before we work out the more complicated case of Fig.~\ref{topB}(a),
let us consider the behavior of a single hyperbaryon in large $N$.
The hyperbaryons are bound by interchanging hypergluons between their
elementary constituents. The situation here is different from the conventional
large-$N$ limit of baryons made only of fundamental-{\it irrep}\ fermions,
where the number of constituents grows linearly with $N$ \cite{WLN}.
For our hyperbaryons, the number of elementary constituents (as well as
the number of hypercolor indices of each field) is fixed.
This resembles the behavior of mesons within the usual large-$N$ treatment.
In Fig.~\ref{baryonN} we show a few examples. The bottom end of each diagram
represents a hyperbaryon, say $B_R$ (first line of Eq.~(\ref{cornerslN})),
and the top end the corresponding anti-baryon,
say $\overline{B}_L$ (last line of Eq.~(\ref{cornerslN})).
Starting with Fig.~\ref{baryonN}(a)
the two vertical lines in the middle represent the propagation
of the double-indexed fermion of the antisymmetric {\it irrep},
and the lines on the sides
the propagation of the two fundamental-{\it irrep}\ fermions, one on each side.
The lines are oriented: the arrows point from a superscript index to a
subscript index. Fig.~\ref{baryonN}(b) shows
an alternative index contraction, still without hypergluon fields.
Note that Fig.~\ref{baryonN}(a) dominates over Fig.~\ref{baryonN}(b)
in large $N$, since the former is of order $N^2$ and the latter of
order $N$. In Fig.~\ref{baryonN}(c) we have added hypergluon interactions.
Introducing the `t~Hooft coupling $\lambda=g^2 N$ it follows that
that planar diagram is again of order $N^2$. The factors of $g=\sqrt{\lambda/N}$
from each hypergluon vertex are compensated by
a matching increase in the number of index loops.
While the diagram appears disconnected
to the eye, this is really not the case, because the two central vertical
lines correspond to the two-point function of a single two-index fermion,
$\svev{\omega^{AB}\overline{\o}_{CD}}$.
Fig.~\ref{baryonN}(d) shows a different arrangement of hypergluon
interactions. The hypergluon that is exchanged at the center of the diagram
represents a self-energy correction for $\svev{\omega^{AB}\overline{\o}_{CD}}$,
which, to be consistent with the directionality of the index lines,
gives rise to a non-planar diagram. This diagram is subleading
in the large-$N$ counting.
The upshot is that,
in large $N$, the dominant diagrams that bind the hyperbaryon
are planar diagrams of order $N^2$, such as for example those in
Fig.~\ref{baryonN}(a) and Fig.~\ref{baryonN}(c).
The hyperbaryon two-point function in Eq.~(\ref{FT}) will exhibit this large-$N$
behavior, much like the two-point function of NGBs made out of
antisymmetric-{\it irrep}\ fermions \cite{taco}, which, in turn,
leads to $f\sim N$ for large $N$.%
\footnote{Note the difference with QCD, where the decay constant $f_\p$
of the fundamental-{\it irrep}\ NGBs scales like $\sqrt{N}$.
}
Using Eqs.~(\ref{B}) and~(\ref{yukawa}),
it follows that $M$ is independent of $N$, while $Z_B\sim 1/N^4$,
and the top Yukawa coupling behaves like $y\sim N$.
In contrast to QCD, where $m_{\rm nucleon}/f_\p$ grows like $\sqrt{N}$,
we find that $M/f$ decreases like $1/N$.
If $M$ is indeed related to the mass of the lightest hyperbaryon in the theory,
this suggests that the lightest hyperbaryon could be relatively light
compared to $\Lambda_{\rm HC}$.
With the contractions of the double-index hyperfermions fixed to be those in
Fig.~\ref{topB}(a), let us now study the large-$N$ behavior of the various
possible contractions of the single-index, fundamental-{\it irrep}\ fields.
Since $U(1)_X$ is not broken spontaneously, the Wick contractions
have to comply with this symmetry.
Let us start, for example, with the two $\Psi$ fields of the $B_R$
hyperbaryon at the bottom left of Fig.~\ref{topB}(a).
They can be contracted with the corresponding
fields in the top left or the bottom right, but not with those in the top right.
There are three possibilities: (1) both $\Psi$ fields are contracted with those
at the top left, (2) both are contracted with those at the bottom right,
or (3) one is contracted with a field at the top left and the other
with a field at the bottom right.
\begin{figure}[t]
\begin{center}
\includegraphics*[height=5cm]{ctopN.eps}
\end{center}
\begin{quotation}
\floatcaption{ctopN}%
{Different large-$N$ contributions to Fig.~\ref{topB}(a).
The curved lines at the top and the bottom represent the two top-quark
propagators. The diagrams scale like (a) $N^4$, (b) $N^2$, and (c) $N^3$}
\end{quotation}
\vspace*{-4ex}
\end{figure}
Let us consider these three cases in turn.
First we contract both $\Psi$ fields at the lower left corner
with the $\bar{\J}$ fields of the $\overline{B}_L$ hyperbaryon
at the upper left corner, \mbox{\it i.e.}, case (1) above.
Remembering that also the two-index field at the lower left corner of
Fig.~\ref{topB}(a) is contracted with the two-index field at the
upper left corner, this gives a contribution proportional to
\begin{equation}
\label{contr1}
\left(\delta^A_C\delta^B_D-\delta^A_D\delta^B_C\right)
\left(\delta_A^C\delta_B^D-\delta_A^D\delta_B^C\right)=2N(N-1)\ ,
\end{equation}
where we label the $SU(N)$ indices as $\omega^{AB}\Psi_A\Psi_B$ at the lower left corner
and as $\overline{\o}_{CD}\bar{\J}^C\bar{\J}^D$ at the upper left corner. We get a
similar factor from the right side of the diagram, so that the total diagram
is of order $(N(N-1))^2\sim N^4$. In Fig.~\ref{ctopN}(a)
we show as an example the order-$N^4$ contribution coming
from using twice the diagram of Fig.~\ref{baryonN}(a).
For case (2), the $\Psi$ fields at the lower left corner are both contracted
with the corresponding fields at the lower right corner.
Using $SU(N)$ indices $AB$ at the lower left,
$CD$ at the upper left, $EF$ at the lower right, and $GH$ at the upper
right corners, this contraction leads to a contribution of order
\begin{equation}
\label{contr2}
\delta^{[A}_C\delta^{B]}_D\delta^{[G}_E\delta^{H]}_F\delta^{[E}_A\delta^{F]}_B\delta^{[C}_G\delta^{D]}_H\sim N^2\ .
\end{equation}
The notation $[\dots]$ denotes antisymmetrization in the pair of indices inside
the brackets. An example is shown in Fig.~\ref{ctopN}(b).
Finally we consider the mixed case (3), where one of the $\Psi$ fields is
contracted with the lower right corner, and the other with the upper left one.
It is straightforward to see that this gives a contribution
\begin{equation}
\label{contr3}
\delta^{[A}_C\delta^{B]}_D\delta^{[G}_E\delta^{H]}_F\delta_A^{[C}\delta_G^{D]}\delta_B^{[E}\delta_H^{F]} \sim N^3\ .
\end{equation}
An example is shown in Fig.~\ref{ctopN}(c).
We should also consider diagrams ``dressed''
with hypergluons. The interesting case is that of Fig.~\ref{ctopN}(a),
where hypergluons exchanged between the two hyperbaryons on the
left and on the right make these hyperbaryons interact. Such interactions
would spoil the factorization of the four-point function in Eq.~(\ref{Ctopmain}).
However, these interactions are suppressed in large $N$ for the same
reason that meson-meson interactions are suppressed in large-$N$ QCD.
Any hypergluon connecting the left and right sides of Fig.~\ref{ctopN}(a)
will reduce the number of hypercolor loops by one,
in addition to adding a factor of $g^2/N$.
The key point here is that the number of hyperfermion
constituents of the hyperbaryons is fixed to three, in contrast to the
case of baryons in QCD, where the number of constituent quarks grows
linearly with $N$.
We conclude that the factorizable part of Fig~\ref{topB}(a),
which grows like $N^4$, is the dominant large-$N$ contribution of the
top-quark sector. $C_{\rm top}$ itself grows like $N^2$,
since $V_{\rm eff}^{\rm top}$ is proportional to $y^2 C_{\rm top}$
and $y$ scales like $N$ (this is consistent with the large-$N$ behavior
of the right-hand side of Eq.~(\ref{Ctopfactexpl})).
According to the discussion in the beginning of this subsection,
this produces a negative
value for $C_{\rm top}$, which, in turn, generates a negative curvature
for the Higgs field in Eq.~(\ref{Veffh}).
We recall that the other contribution to the effective potential~(\ref{Veffh})
for the Higgs field is coming from a single electroweak gauge-boson exchange.
This contribution, which is parametrized by $C_{LR}$, always produces
a positive-curvature term in the effective potential.
It is easily seen that $C_{LR}$ is subleading in the large-$N$ counting.
The electroweak gauge group $SU(2)_L$ is a subgroup of $SU(5)$,
whereas hypercharge is a subgroup of $SU(5)\times U(1)_X$.
Therefore all the electroweak gauge bosons interact with the
antisymmetric-{\it irrep}\ hyperfermions, from which the $SU(5)/SO(5)$ coset fields
are formed. $C_{LR}$ involves a single closed fermion loop,
and so for the antisymmetric-{\it irrep}\ fields it is of order $N^2$.
(For the fundamental {\it irrep}, $C_{LR}$ would be of order $N$.
Observe that the electroweak gauge bosons do not carry hypercolor,
and therefore their exchange has no effect on the large-$N$ counting.)
This is subleading to the factorizable contribution of the top sector,
which is of order $N^4$.
It follows that, in large $N$, Eq.~(\ref{Veffh}) is dominated
by the contribution of the top-quark sector, developing a negative curvature
at the origin that triggers electroweak symmetry breaking.
\bigskip
\subsection{\label{pheno} Phenomenological consequences}
Our analysis in this section was based on the lagrangian~(\ref{lagSU5}).
While ${\cal L}_{\rm EHC}$ depends on two coupling constants $\lambda_1$ and $\lambda_2$,
only their product enters the determination of the top Yukawa coupling
in Sec.~\ref{top-mass}, and of the induced effective potential for the Higgs
studied in Sec.~\ref{top-effpot}. An additional free parameter of the model
is the scale of the hypercolor theory itself, $\Lambda_{HC}$.
Using the results of Sec.~\ref{top-mass} and Sec.~\ref{top-effpot},
we may in principle perform a lattice calculation that will
fix the values of $\lambda_1\lambda_2$ and of $\Lambda_{HC}$
in terms of the experimentally known values of the top Yukawa coupling $y$,
and of the Higgs field's expectation value,
provided that $C_{\rm top}$ turns out to be large enough compared
to $C_{LR}$ to trigger condensation (see Eq.~(\ref{condition})).%
\footnote{We are assuming that none of the $SU(2)_L$ triplet NGBs acquires
an expectation value.
}
The only remaining uncertainty then arises from the four different choices
for the hyperbaryon fields (that each can be a primed or an unprimed one,
{\it cf.\ }\ Table~\ref{tabHC})
in Eq.~(\ref{lagSU5}). This gives rise to a discrete four-fold ambiguity
in the predicted values of $\lambda_1\lambda_2$ and of $\Lambda_{HC}$.
For each of these four possibilities, one can proceed to
compare other predictions of the hypercolor theory
with experimental constraints, which can in principle rule out, or rule in,
that particular version of the hypercolor theory.
A less constrained hypercolor model can be obtained by relaxing the assumption
that $SU(3)\times SU(3)'$ is broken explicitly to $SU(3)_{\rm color}$ only by
the QCD interactions. There are at least two alternative ways
to introduce such an explicit breaking.
First, one can introduce a Dirac mass term $m\overline{\j}\psi$ for the fundamental-{\it irrep}\
hyperfermions, with $m\,\raisebox{-1.0ex}{$\stackrel{\textstyle >}{\sim}$}\,\Lambda_{HC}$. We may think of $m$ as arising
from the expectation value of a three by three global spurion $M_{ab}=m\delta_{ab}$,
where $M_{ab}$ has the same transformation properties as the
$SU(3)\times SU(3)'/SU(3)_{\rm color}$ coset field $\Omega_{ab}$.
Using the global spurion $M_{ab}$ allows for the coupling of each SM spurion
field $T_{L,R}$ to both of the unprimed an primed hyperbaryon fields
from Table~\ref{tabHC}. In fact,
if the hyperbaryon field is assumed to have well defined
transformation properties under $SU(3)_{\rm color}$ only,
some additional operators besides those shown in Table~\ref{tabHC}
may occur. For their structure, see Table~2 of Ref.~\cite{ferretti}.
The bottom line is that ${\cal L}_{\rm EHC}$ can now depend on four (or more)
coupling constants, instead of just two.
An alternative mechanism involves the introduction of two more spurions
$T'_{L,R}$, and assuming that the unprimed hyperbaryon fields and spurions
transform under $SU(3)$, while the primed ones transform under $SU(3)'$.
Twice as many terms are then allowed by $SU(3)\times SU(3)'$ invariance
of ${\cal L}_{\rm EHC}$, with each term involving either unprimed or primed fields.
The explicit breaking to $SU(3)_{\rm color}$
then occurs by assigning to all spurions, both primed and unprimed,
the same SM values as in Eq.~(\ref{fspurions}).
In any one of these more general schemes, the matching
of the top Yukawa coupling and of $C_{\rm top}$ to the underlying theory
can be done using the same techniques as before.
However, the predictive power will be reduced, since this analysis
would still provide just two constraints on the larger set of parameters,
which includes $\Lambda_{HC}$ and all the coupling constants that may now occur
in ${\cal L}_{\rm EHC}$.
Finally, note that if we expand both $\Sigma$ and $\Phi$ in Eq.~(\ref{topmasseff})
to first order in the NGB fields, and use the SM values $\hat{T}_{L,R}$,
we obtain the dimension-five operator
$(y/f') \eta'\overline{t}_R H_i \e_{ij} q_{Lj}$ and its hermitian conjugate,
where $f'$ is the decay constant of $\eta'$. This couples $\eta'$
to the SM fields. If we set the Higgs field to its vacuum expectation value,
the above operator reduces to $(m_t/f') \eta' \overline{t}_R t_L$.
The phenomenological implications of these interactions
have to be looked into. If they turn out to be incompatible with experiment,
this would necessitate the introduction of an explicit breaking of $U(1)'$
that makes the $\eta'$ sufficiently heavy. One possible source for this
explicit breaking is the Dirac mass term $m\overline{\j}\psi$ discussed above.
\section{\label{lattice} Lattice aspects}
In this short section, we explain why a lattice computation of $C_{\rm top}$
would be a ``QCD-like'' computation. The question to address
is why the lattice formulation of hypercolor theories such as
considered here resembles the lattice formulation of QCD, in view of the
well-known complications with fermion doubling and
chirality on the lattice. The fermion doubling problem has its roots
in the observation that a single Weyl fermion cannot live on the lattice
\cite{KS,NN}. Since the hypercolor theory contains an odd number of Weyl
fermions in the two-index antisymmetric {\it irrep},
this might seem to imply that the model we consider
here cannot be easily discretized.
The first observation is that the integration over the SM quark fields
$q_L$ and $t_R$ is done analytically, as we did in this article.
The results for the top Yukawa coupling $y$ (Eqs.~(\ref{mu}) and~(\ref{y}))
and for $C_{\rm top}$ (Eq.~(\ref{Ctopfinal})) are obtained in terms of
pure hypercolor correlation functions. Hence only the hypercolor theory
needs to be considered on the lattice, and the lattice action
does not contain the four-fermion lagrangian ${\cal L}_{\rm EHC}$ of Eq.~(\ref{lagSU5}).
Let us start from the sector that resembles QCD most closely, namely, the
Dirac fermions $\psi_a$ in the fundamental {\it irrep}\ of $SU(4)$ hypercolor.
These can be treated in exactly the same way as the quark fields of
$N_f$-flavor QCD.
In the Wilson formulation of the theory, the gauge invariant Wilson mass term
removes the fermion doublers at the price of breaking the symmetry group
$SU(N_f)_L\times SU(N_f)_R$ of the continuum theory explicitly to
its diagonal $SU(N_f)$ subgroup.
Tuning the bare mass appropriately, one then recovers the massless
theory with the full chiral symmetry group in the continuum limit \cite{KS}.
In the hypercolor theory we have $N_f=3$ Dirac fermions.
The chiral flavor group $SU(3)_L\times SU(3)_R$ has been renamed
$SU(3)\times SU(3)'$, and the unbroken diagonal subgroup was identified
with $SU(3)_{\rm color}$.
We next turn to the novel feature of the hypercolor theory, which
is the 5 real-{\it irrep}\ Weyl fermions $\Upsilon_{ABi}$.
The key point is that each of these Weyl fermions can be assembled
together with its antifermion into a Majorana fermion field $\chi_{ABi}$.
A Majorana mass term of the form $\overline{\c}^{AB}_i\chi_{ABi}$ is allowed by the $SU(4)$
gauge symmetry, just like a Dirac mass term is allowed in the familiar case
of fundamental-{\it irrep}\ fermions.
Therefore, a Wilson formulation of the Majorana fermions
is possible, with a Wilson mass term that once again removes
fermion doublers without breaking gauge invariance.
The hypercolor theory is chiral with respect to the $SU(5)$ flavor symmetry
of the Majorana or Weyl fields, and consequently,
the Majorana--Wilson mass term breaks $SU(5)$ explicitly down to $SO(5)$.
Again, the full $SU(5)$ chiral symmetry
should be recovered in the continuum limit after appropriate tuning
of the bare mass.%
\footnote{The lattice formulation of an adjoint Majorana fermion
was studied extensively in the context of supersymmetric theories,
see, \mbox{\it e.g.}, Ref.~\cite{susy}.
}
The situation with respect to (complex-{\it irrep})
Dirac fermions and to (real-{\it irrep}) Majorana fermions is thus
completely parallel. In both cases,
what we anticipate as the spontaneous symmetry breaking pattern
of the continuum theory turns into an explicit breaking in the Wilson
formulation. The explicit breaking disappears, and the full chiral
symmetry group can be recovered, in the continuum limit by tuning
the bare mass terms. In particular, the flavor group $SO(5)$
of the Majorana-Wilson fermion action in the hypercolor theory
enlarges to $SU(5)$ in the continuum limit, much like the
diagonal $SU(N_f)$ symmetry of the Dirac-Wilson action of QCD enlarges
to the full $SU(N_f)_L\times SU(N_f)_R$ symmetry in the continuum limit.
In short, it is precisely the unbroken flavor symmetry group $H$
of Eq.~(\ref{breaking}) that is preserved in a lattice formulation
with Wilson fermions. It is actually possible
to gauge the SM group $SU(3)_{\rm color}\times SU(2)_L\times U(1)_Y$
in this lattice formulation, because this group is contained in $H$.
The difficulties caused on the lattice with chiral gauge symmetries%
\footnote{For a review, see \cite{MGlat2000}.}
would only appear were one to couple also the SM fermions to this model.
The lattice formulation of the hypercolor theory is not without technical
challenges. First, two fermion {\it irreps}\ need to be introduced simultaneously.
While clearly a coding task, this is
something that to our knowledge has not been done to date. Also,
the computation of a four-point function as in Eq.~(\ref{Ctopmain})
would be very demanding. Because there are two independent momentum variables,
the cost is expected to grow like the square of the 4-volume of the lattice.
The computation of $S_{B}(0)$ in Eq.~(\ref{mu}),
and of the factorizable contribution to $C_{\rm top}$, would already be
interesting. Here only the hyperbaryon two-point function is required,
and the cost grows linearly with the 4-volume.
\section{\label{conclusion} Conclusion}
In this article, we discussed a recently proposed composite Higgs model
\cite{ferretti}, concentrating on the top quark sector.
This ``hypercolor'' model is an
$SU(4)$ gauge theory, with a quintuplet of two-index antisymmetric
Majorana fermions and a color triplet of $SU(4)$-fundamental Dirac fermions.
We considered the
top Yukawa coupling and the top contribution to the effective potential
for the Higgs, showing how to match the relevant low-energy constants
$\mu=yf/2$ and $C_{\rm top}$ to correlation functions of the hypercolor theory.
This matching is the starting point for any non-perturbative evaluation
of the low-energy constants. The needed computations can in principle
be done on the lattice, using methods that are much the same as those
employed in lattice QCD. The main differences are that now the gauge
group is $SU(4)$ instead of $SU(3)$, and that there are fermions in
more than one {\it irrep}\ of the gauge group.
For the Higgs effective potential, a four-point hyperbaryon correlation
function needs to be considered; this appears to have been overlooked in
at least some of the literature on models of this type. A lattice calculation
of this four-point function would be very demanding, and how to do it
efficiently is a question that goes beyond the scope of this paper.
As mentioned in Sec.~\ref{lattice}, only a two-point function is needed
for the low-energy constant $\mu$,
as well as for the factorizable contribution to $C_{\rm top}$.
This computation would be comparable in scope with a lattice computation
of $C_{LR}$, the LEC controlling the
contribution to the Higgs effective potential from
the SM gauge bosons ({\it cf.\ }\ Sec.~\ref{weak-effpot}).
We found that a large-$N$ limit exists in which the factorizable
contribution to the hyperbaryon four-point function dominates $C_{\rm top}$.
We also showed that the factorizable contribution generates a negative
curvature at the origin. This is a necessary condition for electroweak
symmetry breaking, because the effective potential induced by
the SM gauge bosons does not break the SM symmetries,
a manifestation of vacuum alignment.
For very large $N$, the top sector dominates the whole
Higgs effective potential. This maximizes the symmetry breaking,
with the minimum of $V_{\rm eff}(h)$ presumably occurring for $\sin(2h/f)=1$
in Eq.~(\ref{Veffh}). Phenomenologically, this is not allowed \cite{RC,ferretti}.
The hope is that for $N=4$, electroweak symmetry breaking
with a phenomenologically acceptable value of $h/f$ takes place.
Only the lattice can address this question quantitatively.
In order to couple the SM fermions to the hypercolor theory,
an extended hypercolor sector is necessary. While the detailed structure
of the EHC theory is very important for phenomenology,
the lattice setup is largely blind to these details. At energy scales
much below $\Lambda_{\rm EHC}$,
the coupling of SM and hypercolor fermions is summarized by
the four-fermion lagrangian ${\cal L}_{\rm EHC}$ of Eq.~(\ref{lagSU5}).
Much like the familiar treatment of hadronic matrix elements
of the electroweak interactions,
${\cal L}_{\rm EHC}$ is not taken as part of the lattice action.
Instead, one evaluates the hypercolor-theory correlation functions
that arise from working to leading order in the four-fermion lagrangian
${\cal L}_{\rm EHC}$.
In the most constrained case,
which is the one we have worked out in detail in this paper,
${\cal L}_{\rm EHC}$ depends on only two couplings $\lambda_{1,2}$.
A lattice computation can then in principle determine
their product $\lambda_1\lambda_2$ together with the hypercolor scale $\lambda_{\rm HC}$
in terms of the experimental values of the top Yukawa coupling
and the Higgs expectation value, up to a four-fold ambiguity.
In a less constrained setup, a similar lattice computation would supply two
constraints among the parameters of the hypercolor theory.
The parameters $\lambda_{1,2}$ have mass dimension two.
According to naive dimensional analysis, their values would be
of order $(\Lambda_{\rm HC}/\Lambda_{\rm EHC})^2$, making the top Yukawa coupling
of order $y \sim (\Lambda_{\rm HC}/\Lambda_{\rm EHC})^4$. For comparison,
we recall that in classic (walking) technicolor,
the top Yukawa coupling is naively of order $(\Lambda_{\rm TC}/\Lambda_{\rm ETC})^2$,
where $\Lambda_{\rm TC}$ and $\Lambda_{\rm ETC}$ are the scales of the technicolor
and of the extended technicolor theories. Naively, the case
for a partially-composite top seems even worse than for technicolor.
On the other hand,
it might be that the experimental constraints allow $\Lambda_{\rm EHC}$ to be
much closer to $\Lambda_{\rm HC}$ than $\Lambda_{\rm ETC}$ to $\Lambda_{\rm TC}$.
We also found that the hyperbaryons that couple to the top quark
in ${\cal L}_{\rm EHC}$ might be relatively light, as suggested by large-$N$ counting,
and this helps boost the value of the top Yukawa $y$ as well. The
key reason for the possible lightness of these hyperbaryons is that
they are composed of hyperfermions in two different {\it irreps}\ of the
gauge group. As far as large-$N$ counting goes, these hyperbaryons behave
more like mesons than like baryons in the usual large-$N$ limit of QCD.
Yet another feature that may be needed for a phenomenologically viable
partial-compositeness model is large anomalous dimensions
for the hyperbaryon fields \cite{RC}. It should be possible to
study on the lattice whether or not this is the case.
\vspace{3ex}
\noindent {\bf Acknowledgments}
\vspace{3ex}
We thank Gabriele Ferretti for explanations about phenomenological
aspects of the model, Will Jay for a discussion of hyperbaryon operators,
and Tom DeGrand for his comments on the manuscript.
We also thank Ben Svetitsky, Yuzhi Liu and Ethan Neil
for discussions. MG and YS thank the Department of Physics of
the University of Colorado, Boulder, and
YS thanks the Department of Physics and Astronomy of San Francisco
State University for hospitality.
MG is supported in part by the US Department of Energy, and
YS is supported by the Israel Science Foundation
under grant no.~449/13.
|
\section{Introduction}
\label{sect-intro}
In~\cite{BM13a}, we presented an approach to computational complexity in
which algorithmic problems are viewed as Boolean function families that
consist of one $n$-ary Boolean function for each natural number $n$ and
the complexity of such problems is assessed in terms of the length of
finite single-pass instruction sequences acting on Boolean registers
that compute the members of these families.
The instruction sequences concerned contain only instructions to set
and get the content of Boolean registers, forward jump instructions,
and a termination instruction.
Moreover, each Boolean register used serves as either input register,
output register or auxiliary register.
Auxiliary Boolean registers are not needed to compute Boolean functions.
The question whether shorter instruction sequences are possible with the
use of auxiliary Boolean registers was not answered in~\cite{BM13a}.
In~\cite{BM14e}, we showed that, in the case of the parity functions,
shorter instruction sequences are possible with the use of an auxiliary
Boolean register provided the instruction set is extended with
instructions to complement the content of auxiliary Boolean registers.
In the current paper, we consider all instructions for Boolean
registers that are possible in the setting in which the work presented
in~\cite{BM13a,BM14e} has been carried out and investigate instruction
sequence size bounded functional completeness of instruction sets for
Boolean registers.
Intuitively, a given instruction set for Boolean registers is
$n$-size-bounded functionally complete if the effects of each possible
instruction for Boolean registers can be obtained by an instruction
sequence whose length is at most $n$ and that contains only instructions
from the given instruction set for Boolean registers, forward jump
instructions, and a termination instruction.
A given instruction set for Boolean registers is functionally
complete if it is $n$-size-bounded functionally complete for some $n$.
We have identified one of the $256$ smallest instruction sets for
Boolean registers that is $1$-size-bounded functionally complete
(Corollary~\ref{corollary-bounded-complete-1}), and we have found that
there is a large subset of this $1$-size-bounded functionally complete
instruction set with the following properties:
(i)~each of its proper subsets that does not include the instructions to
complement the content of Boolean registers, but includes the
instructions to set and get the content of Boolean registers, is
$4$-size-bounded functionally complete and not $3$-size-bounded
functionally complete and
(ii)~each of its proper subsets that includes the instructions to
complement the content of Boolean registers is
$3$-size-bounded functionally complete and not $2$-size-bounded
functionally complete (Corollary~\ref{corollary-bounded-complete-2}).
The use of a $1$-size-bounded functionally complete instruction
set, such as the one referred to in the previous paragraph, gives rise
to the smallest instruction sequence sizes.
However, the use of many instruction sets that are not $1$-size-bounded
functionally complete, e.g.\ the ones referred to under~(i) above, gives
rise to instruction sequence sizes that are at most $4$ times larger.
The work presented in~\cite{BM14e} triggered the work presented in this
paper because of the choice to use an extension of the instruction set
used in~\cite{BM13a}.
Since the results from the latter paper, with the exception of one
auxiliary result, are concerned with upper bounds of instruction
sequence size complexity, these results go through if instruction
sequences may also contain instructions to complement the content of
auxiliary Boolean registers.
However, for the work presented in~\cite{BM14e}, the instruction set did
matter in the sense that we were not able to prove the main result of
the paper, which is concerned with a lower bound of instruction sequence
size complexity, using the instruction set used in~\cite{BM13a}.
We consider the work presented in the current paper to be useful to
adequately assess that result as it is.
We expect that it will turn out to be also useful to adequately assess
results of future work that is concerned with lower bounds of
instruction sequence size complexity.
Like the work presented in~\cite{BM13a,BM14e}, the work presented in
this paper is carried out in the setting of \PGA\ (ProGram Algebra).
\PGA\ is an algebraic theory of single-pass instruction sequences that
was taken as the basis of the approach to the semantics of programming
languages introduced in~\cite{BL02a}.
As a continuation of the work presented in~\cite{BL02a},
(i)~the notion of an instruction sequence was subjected to systematic
and precise analysis and
(ii)~issues relating to diverse subjects in computer science and
computer engineering were rigorously investigated in the setting of
\PGA.
The subjects concerned include programming language expressiveness,
computability, computational complexity, algorithm efficiency,
algorithmic equivalence of programs, program verification, program
compactness, micro-architecture, and probabilistic programming.
For a comprehensive survey of a large part of this work,
see~\cite{BM12b}.
An overview of all the work done to date in the setting of \PGA\ and
open questions originating from this work can be found on~\cite{SiteIS}.
This paper is organized as follows.
First, we present the preliminaries to the work presented in this paper
(Sections~\ref{sect-PGA-and-BTA} and~\ref{sect-TSI}) and introduce the
possible instructions for Boolean registers
(Section~\ref{sect-instr-br}).
Next, we define an equivalence relation on these instructions that
identifies instructions that have the same effects
(Section~\ref{sect-eqv-instr-br}) and study instruction sequence
size bounded functional completeness of instruction sets for Boolean
registers (Section~\ref{sect-bounded-complete}).
Finally, we make some concluding remarks (Section~\ref{sect-concl}).
Some familiarity with the basic notions related to algebraic theories
and their models is assumed in this paper.
The relevant notions are explained in handbook chapters and books on the
foundations of algebraic specification,
e.g.~\cite{EM85a,ST99a,ST12a,Wir90a}.
The following should be mentioned in advance.
The set $\Bool$ is a set with two elements whose intended
interpretations are the truth values \emph{false} and \emph{true}.
As is common practice, we represent the elements of $\Bool$ by the bits
$0$ and $1$.
In line with generally accepted conventions, we use terminology based on
identification of the elements of $\Bool$ with their representation
where appropriate.
For example, where a better link up with commonly used terminology is
expected, the elements of $\Bool$ are loosely called bits and the
elements of $\Bool^n$ are loosely called bit strings of length $n$.
The preliminaries to the work presented in this paper
(Sections~\ref{sect-PGA-and-BTA} and~\ref{sect-TSI}) are almost the
same as the preliminaries to the work presented in~\cite{BM14a} and
earlier papers.
For this reason, there is some text overlap with those papers.
Apart from the preliminaries, the material in this paper is new.
\section{Program Algebra and Basic Thread Algebra}
\label{sect-PGA-and-BTA}
In this section, we give a survey of \PGA\ (ProGram Algebra) and \BTA\
(Basic Thread Algebra) and make precise in the setting of \BTA\ which
behaviours are produced by the instruction sequences considered in \PGA\
under execution.
The greater part of this section originates from~\cite{BM13a}.
A comprehensive introduction to \PGA\ and \BTA, including examples, can
among other things be found in~\cite{BM12b}.
In \PGA, it is assumed that there is a fixed but arbitrary set $\BInstr$
of \emph{basic instructions}.
The intuition is that the execution of a basic instruction may modify a
state and produces a reply at its completion.
The possible replies are $\False$ and $\True$.
The actual reply is generally state-dependent.
The set $\BInstr$ is the basis for the set of instructions that may
occur in the instruction sequences considered in \PGA.
The elements of the latter set are called \emph{primitive instructions}.
There are five kinds of primitive instructions:
\begin{itemize}
\item
for each $a \in \BInstr$, a \emph{plain basic instruction} $a$;
\item
for each $a \in \BInstr$, a \emph{positive test instruction} $\ptst{a}$;
\item
for each $a \in \BInstr$, a \emph{negative test instruction} $\ntst{a}$;
\item
for each $l \in \Nat$, a \emph{forward jump instruction} $\fjmp{l}$;
\item
a \emph{termination instruction} $\halt$.
\end{itemize}
We write $\PInstr$ for the set of all primitive instructions.
On execution of an instruction sequence, these primitive instructions
have the following effects:
\begin{itemize}
\item
the effect of a positive test instruction $\ptst{a}$ is that basic
instruction $a$ is executed and execution proceeds with the next
primitive instruction if $\True$ is produced and otherwise the next
primitive instruction is skipped and execution proceeds with the
primitive instruction following the skipped one --- if there is no
primitive instruction to proceed with,
inaction occurs;
\item
the effect of a negative test instruction $\ntst{a}$ is the same as
the effect of $\ptst{a}$, but with the role of the value produced
reversed;
\item
the effect of a plain basic instruction $a$ is the same as the effect
of $\ptst{a}$, but execution always proceeds as if $\True$ is produced;
\item
the effect of a forward jump instruction $\fjmp{l}$ is that execution
proceeds with the $l$th next primitive instruction --- if $l$ equals $0$
or there is no primitive instruction to proceed with, inaction occurs;
\item
the effect of the termination instruction $\halt$ is that execution
terminates.
\end{itemize}
\PGA\ has one sort: the sort $\InSeq$ of \emph{instruction sequences}.
We make this sort explicit to anticipate the need for many-sortedness
later on.
To build terms of sort $\InSeq$, \PGA\ has the following constants and
operators:
\begin{itemize}
\item
for each $u \in \PInstr$,
the \emph{instruction} constant $\const{u}{\InSeq}$\,;
\item
the binary \emph{concatenation} operator
$\funct{\ph \conc \ph}{\InSeq \x \InSeq}{\InSeq}$\,;
\item
the unary \emph{repetition} operator
$\funct{\ph\rep}{\InSeq}{\InSeq}$\,.
\end{itemize}
Terms of sort $\InSeq$ are built as usual in the one-sorted case.
We assume that there are infinitely many variables of sort $\InSeq$,
including $X,Y,Z$.
We use infix notation for concatenation and postfix notation for
repetition.
A closed \PGA\ term is considered to denote a non-empty, finite or
eventually periodic infinite sequence of primitive instructions
\footnote
{An eventually periodic infinite sequence is an infinite sequence with
only finitely many distinct suffixes.}
The instruction sequence denoted by a closed term of the form
$t \conc t'$ is the instruction sequence denoted by $t$
concatenated with the instruction sequence denoted by $t'$.
The instruction sequence denoted by a closed term of the form $t\rep$
is the instruction sequence denoted by $t$ concatenated infinitely
many times with itself.
Closed \PGA\ terms are considered equal if they represent the same
instruction sequence.
The axioms for instruction sequence equivalence are given in
Table~\ref{axioms-PGA}
\begin{table}[!t]
\caption{Axioms of \PGA}
\label{axioms-PGA}
\begin{eqntbl}
\begin{axcol}
(X \conc Y) \conc Z = X \conc (Y \conc Z) & \axiom{PGA1} \\
(X^n)\rep = X\rep & \axiom{PGA2} \\
X\rep \conc Y = X\rep & \axiom{PGA3} \\
(X \conc Y)\rep = X \conc (Y \conc X)\rep & \axiom{PGA4}
\end{axcol}
\end{eqntbl}
\end{table}
In this table, $n$ stands for an arbitrary natural number from
$\Natpos$
\footnote
{We write $\Natpos$ for the set $\set{n \in \Nat \where n \geq 1}$ of
positive natural numbers.}
For each $n \in \Natpos$, the term $t^n$, where $t$ is a \PGA\ term, is
defined by induction on $n$ as follows: $t^1 = t$, and
$t^{n+1} = t \conc t^n$.
A typical model of \PGA\ is the model in which:
\begin{itemize}
\item
the domain is the set of all finite and eventually periodic infinite
sequences over the set $\PInstr$ of primitive instructions;
\item
the operation associated with ${} \conc {}$ is concatenation;
\item
the operation associated with ${}\rep$ is the operation ${}\srep$
defined as follows:
\begin{itemize}
\item
if $U$ is finite, then $U\srep$ is the unique infinite sequence $U'$
such that $U$ concatenated $n$ times with itself is a proper prefix of
$U'$ for each $n \in \Nat$;
\item
if $U$ is infinite, then $U\srep$ is $U$.
\end{itemize}
\end{itemize}
It is immediately clear that this model has no proper subalgebra.
Moreover, we know from~\cite[Section~3.2.2]{BL02a} that the axioms of
\PGA\ are complete with respect to satisfaction of equations between
closed terms in this model.
Hence, this model is an initial model of \PGA\ (see e.g.~\cite{ST99a}).
We confine ourselves to this model of \PGA\ for the interpretation of
\PGA\ terms.
In the sequel, we use the term \emph{PGA instruction sequence} for the
elements of the domain of this model.
Below, we will use \BTA\ to make precise which behaviours are produced
by \PGA\ instruction sequences under execution.
In \BTA, it is assumed that a fixed but arbitrary set $\BAct$ of
\emph{basic actions} has been given.
The objects considered in \BTA\ are called threads.
A thread represents a behaviour which consists of performing basic
actions in a sequential fashion.
Upon each basic action performed, a reply from an execution environment
determines how the thread proceeds.
The possible replies are the values $\False$ and $\True$.
\BTA\ has one sort: the sort $\Thr$ of \emph{threads}.
We make this sort explicit to anticipate the need for many-sortedness
later on.
To build terms
of sort $\Thr$, \BTA\ has the following constants and operators:
\begin{itemize}
\item
the \emph{inaction} constant $\const{\DeadEnd}{\Thr}$;
\item
the \emph{termination} constant $\const{\Stop}{\Thr}$;
\item
for every $a \in \BAct$, the binary \emph{postconditional composition}
operator $\funct{\pcc{\ph}{a}{\ph}}{\Thr \x \Thr}{\Thr}$.
\end{itemize}
Terms of sort $\Thr$ are built as usual in the one-sorted case.
We assume that there are infinitely many variables of sort $\Thr$,
including $x,y$.
We use infix notation for postconditional composition.
We introduce \emph{basic action prefixing} as an abbreviation:
$a \bapf t$, where $t$ is a \BTA\ term, abbreviates
$\pcc{t}{a}{t}$.
We identify expressions of the form $a \bapf t$ with the \BTA\
term they stand for.
The thread denoted by a closed term of the form $\pcc{t}{a}{t'}$
will first perform $a$, and then proceed as the thread denoted by
$t$ if the reply from the execution environment is $\True$ and proceed
as the thread denoted by $t'$ if the reply from the execution
environment is $\False$.
The thread denoted by $\Stop$ will do no more than terminate and the
thread denoted by $\DeadEnd$ will become inactive.
Closed \BTA\ terms are considered equal if they are syntactically the
same.
Therefore, \BTA\ has no axioms.
Each closed \BTA\ term denotes a finite thread, i.e.\ a thread with a
finite upper bound to the number of basic actions that it can perform.
Infinite threads, i.e.\ threads without a finite upper bound to the
number of basic actions that it can perform, can be defined by means of
a set of recursion equations (see e.g.~\cite{BM09k}).
We are only interested in models of \BTA\ in which sets of recursion
equations have unique solutions, such as the projective limit model
of \BTA\ presented in~\cite{BM12b}.
We confine ourselves to this model of \BTA, which has an initial model
of \BTA\ as a submodel, for the interpretation of \BTA\ terms.
In the sequel, we use the term \emph{BTA thread} or simply \emph{thread}
for the elements of the domain of this model.
Regular threads, i.e.\ finite or infinite threads that can only be in a
finite number of states, can be defined by means of a finite set of
recursion equations.
Provided that the set $\BInstr$ of basic instructions is identified with
the set $\BAct$ of basic actions, the behaviours produced by \PGA\
instruction sequences under execution are exactly the behaviours
represented by regular threads and the behaviours produced by finite
\PGA\ instruction sequences are exactly the behaviours represented by
finite threads.
Henceforth, we will identify $\BInstr$ with $\BAct$.
Intuitively, this means that we will not distinguish the basic action
that takes place when a basic instruction is executed from that
basic instruction.
We combine \PGA\ with \BTA, identifying $\BInstr$ with $\BAct$, and
extend the combination with the \emph{thread extraction} operator
$\funct{\extr{\ph}}{\InSeq}{\Thr}$, the axioms given in
Table~\ref{axioms-thread-extr}
\begin{table}[!tb]
\caption{Axioms for the thread extraction operator}
\label{axioms-thread-extr}
\begin{eqntbl}
\renewcommand{\arraystretch}{1.3}
\begin{eqncol}
\extr{a} = a \bapf \DeadEnd \\
\extr{a \conc X} = a \bapf \extr{X} \\
\extr{\ptst{a}} = a \bapf \DeadEnd \\
\extr{\ptst{a} \conc X} =
\pcc{\extr{X}}{a}{\extr{\fjmp{2} \conc X}} \\
\extr{\ntst{a}} = a \bapf \DeadEnd \\
\extr{\ntst{a} \conc X} =
\pcc{\extr{\fjmp{2} \conc X}}{a}{\extr{X}}
\end{eqncol}
\qquad
\begin{eqncol}
\extr{\fjmp{l}} = \DeadEnd \\
\extr{\fjmp{0} \conc X} = \DeadEnd \\
\extr{\fjmp{1} \conc X} = \extr{X} \\
\extr{\fjmp{l+2} \conc u} = \DeadEnd \\
\extr{\fjmp{l+2} \conc u \conc X} = \extr{\fjmp{l+1} \conc X} \\
\extr{\halt} = \Stop \\
\extr{\halt \conc X} = \Stop
\end{eqncol}
\end{eqntbl}
\end{table}
and the rule that $\extr{X} = \DeadEnd$ if $X$ has an infinite chain of
forward jumps beginning at its first primitive instruction
\footnote
{This rule, which can be formalized using an auxiliary structural
congruence predicate (see e.g.~\cite{BM12b}), is unnecessary when
considering only finite \PGA\ instruction sequences.
}
In Table~\ref{axioms-thread-extr}, $a$ stands for an arbitrary basic
instruction from $\BInstr$, $u$ stands for an arbitrary primitive
instruction from $\PInstr$, and $l$ stands for an arbitrary natural
number from $\Nat$.
For each closed \PGA\ term $t$, $\extr{t}$ denotes the behaviour
produced by the instruction sequence denoted by $t$ under execution.
\section{Interaction of Threads with Services}
\label{sect-TSI}
Services are objects that represent the behaviours exhibited by
components of execution environments of instruction sequences at a high
level of abstraction.
A service is able to process certain methods.
For the purpose of the extension of \BTA\ that will be presented in this
section, it is sufficient to know the following about methods:
(i)~the processing of a method by a service may involve a change of the
service and
(ii)~at completion of the processing of a method by a service, the
service produces a reply value.
The possible reply values are $\False$ and $\True$.
Execution environments are considered to provide a family of
uniquely-named services.
A thread may interact with the named services from the service family
provided by an execution environment.
That is, a thread may perform a basic action for the purpose of
requesting a named service to process a method and to return a reply
value at completion of the processing of the method.
In this section, we give a survey of the extension of \BTA\ with
services, service families, a composition operator for service families,
and operators that are concerned with this kind of interaction.
This section originates from~\cite{BM09k}.
A comprehensive introduction to the presented extension of \BTA,
including examples, can among other things be found in~\cite{BM12b}.
First, we introduce an algebraic theory of service families called
\SFA\ (Service Family Algebra).
In \SFA, it is assumed that a fixed but arbitrary set $\Meth$ of
\emph{methods} has been given.
Moreover, the following is assumed with respect to services:
\begin{itemize}
\item
a signature $\Sig{\Services}$ has been given that includes the following
sorts:
\begin{itemize}
\item
the sort $\Serv$ of
\emph{services};
\item
the sort $\Repl$ of \emph{replies};
\end{itemize}
and the following constants and operators:
\begin{itemize}
\item
the
\emph{empty service} constant $\const{\emptyserv}{\Serv}$;
\item
the \emph{reply} constants $\const{\False,\True,\Div}{\Repl}$;
\item
for each $m \in \Meth$, the
\emph{derived service} operator $\funct{\derive{m}}{\Serv}{\Serv}$;
\item
for each $m \in \Meth$, the
\emph{service reply} operator $\funct{\sreply{m}}{\Serv}{\Repl}$;
\end{itemize}
\item
a $\Sig{\Services}$-algebra $\ServAlg$ that has no proper subalgebra has
been given in which the following holds:
\begin{itemize}
\item
$\False \neq \True$, $\True \neq \Div$, $\Div \neq \False$;
\item
for each $m \in \Meth$,
$\derive{m}(z) = \emptyserv \Liff \sreply{m}(z) = \Div$.
\end{itemize}
\end{itemize}
The intuition concerning $\derive{m}$ and $\sreply{m}$ is that on a
request to service $s$ to process method $m$:
\begin{itemize}
\item
if $\sreply{m}(s) \neq \Div$, $s$ processes $m$, produces the reply
$\sreply{m}(s)$, and then proceeds as $\derive{m}(s)$;
\item
if $\sreply{m}(s) = \Div$, $s$ is not able to process method $m$ and
proceeds as $\emptyserv$.
\end{itemize}
The empty service $\emptyserv$ itself is unable to process any method.
It is also assumed that a fixed but arbitrary set $\Foci$ of
\emph{foci} has been given.
Foci play the role of names of services in a service family.
\SFA\ has the sorts, constants and operators from $\Sig{\Services}$ and
in addition the sort $\ServFam$ of \emph{service families} and the
following constant and operators:
\begin{itemize}
\item
the
\emph{empty service family} constant $\const{\emptysf}{\ServFam}$;
\item
for each $f \in \Foci$, the unary
\emph{singleton service family} operator
$\funct{\mathop{f{.}} \ph}{\Serv}{\ServFam}$;
\item
the binary
\emph{service family composition} operator
$\funct{\ph \sfcomp \ph}{\ServFam \x \ServFam}{\ServFam}$;
\item
for each $F \subseteq \Foci$, the unary
\emph{encapsulation} operator $\funct{\encap{F}}{\ServFam}{\ServFam}$.
\end{itemize}
We assume that there are infinitely many variables of sort $\Serv$,
including $z$, and infinitely many variables of sort $\ServFam$,
including $u,v,w$.
Terms are built as usual in the many-sorted case
(see e.g.~\cite{ST12a}).
We use prefix notation for the singleton service family operators and
infix notation for the service family composition operator.
We write $\Sfcomp{i = 1}{n} t_i$, where $t_1,\ldots,t_n$ are
terms of sort $\ServFam$, for the term
$t_1 \sfcomp \ldots \sfcomp t_n$.
The service family denoted by $\emptysf$ is the empty service family.
The service family denoted by a closed term of the form $f.t$ consists
of one named service only, the service concerned is the service denoted
by $t$, and it is named $f$.
The service family denoted by a closed term of the form
$t \sfcomp t'$ consists of all named services that belong to either the
service family denoted by $t$ or the service family denoted by $t'$.
In the case where a named service from the service family denoted by
$t$ and a named service from the service family denoted by $t'$ have
the same name, they collapse to an empty service with the name
concerned.
The service family denoted by a closed term of the form $\encap{F}(t)$
consists of all named services with a name not in $F$ that belong to
the service family denoted by $t$.
The axioms of \SFA\ are given in
Table~\ref{axioms-SFA}
\begin{table}[!t]
\caption{Axioms of \SFA}
\label{axioms-SFA}
{
\begin{eqntbl}
\begin{axcol}
u \sfcomp \emptysf = u & \axiom{SFC1} \\
u \sfcomp v = v \sfcomp u & \axiom{SFC2} \\
(u \sfcomp v) \sfcomp w = u \sfcomp (v \sfcomp w) & \axiom{SFC3} \\
f.z \sfcomp f.z' = f.\emptyserv & \axiom{SFC4}
\end{axcol}
\qquad
\begin{saxcol}
\encap{F}(\emptysf) = \emptysf & & \axiom{SFE1} \\
\encap{F}(f.z) = \emptysf & \mif f \in F & \axiom{SFE2} \\
\encap{F}(f.z) = f.z & \mif f \notin F & \axiom{SFE3} \\
\multicolumn{2}{@{}l@{\quad}}
{\encap{F}(u \sfcomp v) =
\encap{F}(u) \sfcomp \encap{F}(v)} & \axiom{SFE4}
\end{saxcol}
\end{eqntbl}
}
\end{table}
In this table, $f$ stands for an arbitrary focus from $\Foci$ and
$F$ stands for an arbitrary subset of $\Foci$.
These axioms simply formalize the informal explanation given
above.
For the set $\BAct$ of basic actions, we now take
$\set{f.m \where f \in \Foci, m \in \Meth}$.
Performing a basic action $f.m$ is taken as making a request to the
service named $f$ to process method $m$.
We combine \BTA\ with \SFA\ and extend the combination with the
following operators:
\pagebreak[2]
\begin{itemize}
\item
the binary \emph{abstracting use} operator
$\funct{\ph \sfause \ph}{\Thr \x \ServFam}{\Thr}$;
\item
the binary \emph{apply} operator
$\funct{\ph \sfapply \ph}{\Thr \x \ServFam}{\ServFam}$;
\end{itemize}
and the axioms given in Tables~\ref{axioms-abstracting-use}
and~\ref{axioms-apply}
\begin{table}[!t]
\caption{Axioms for the abstracting use operator}
\label{axioms-abstracting-use}
\begin{eqntbl}
\begin{saxcol}
\Stop \sfause u = \Stop & & \axiom{AU1} \\
\DeadEnd \sfause u = \DeadEnd & & \axiom{AU2} \\
(\pcc{x}{f.m}{y}) \sfause \encap{\set{f}}(u) =
\pcc{(x \sfause \encap{\set{f}}(u))}
{f.m}{(y \sfause \encap{\set{f}}(u))} & & \axiom{AU3} \\
(\pcc{x}{f.m}{y}) \sfause (f.t \sfcomp \encap{\set{f}}(u)) =
x \sfause (f.\derive{m}t \sfcomp \encap{\set{f}}(u))
& \mif \sreply{m}(t) = \True & \axiom{AU4} \\
(\pcc{x}{f.m}{y}) \sfause (f.t \sfcomp \encap{\set{f}}(u)) =
y \sfause (f.\derive{m}t \sfcomp \encap{\set{f}}(u))
& \mif \sreply{m}(t) = \False & \axiom{AU5} \\
(\pcc{x}{f.m}{y}) \sfause (f.t \sfcomp \encap{\set{f}}(u)) = \DeadEnd
& \mif \sreply{m}(t) = \Div & \axiom{AU6}
\end{saxcol}
\end{eqntbl}
\end{table}
\begin{table}[!t]
\caption{Axioms for the apply operator}
\label{axioms-apply}
\begin{eqntbl}
\begin{saxcol}
\Stop \sfapply u = u & & \axiom{A1} \\
\DeadEnd \sfapply u = \emptysf & & \axiom{A2} \\
(\pcc{x}{f.m}{y}) \sfapply \encap{\set{f}}(u) = \emptysf
& & \axiom{A3} \\
(\pcc{x}{f.m}{y}) \sfapply (f.t \sfcomp \encap{\set{f}}(u)) =
x \sfapply (f.\derive{m}t \sfcomp \encap{\set{f}}(u))
& \mif \sreply{m}(t) = \True & \axiom{A4} \\
(\pcc{x}{f.m}{y}) \sfapply (f.t \sfcomp \encap{\set{f}}(u)) =
y \sfapply (f.\derive{m}t \sfcomp \encap{\set{f}}(u))
& \mif \sreply{m}(t) = \False & \axiom{A5} \\
(\pcc{x}{f.m}{y}) \sfapply (f.t \sfcomp \encap{\set{f}}(u)) = \emptysf
& \mif \sreply{m}(t) = \Div & \axiom{A6}
\end{saxcol}
\end{eqntbl}
\end{table}
In these tables, $f$ stands for an arbitrary focus from $\Foci$, $m$
stands for an arbitrary method from $\Meth$, and $t$ stands for an
arbitrary term of sort $\Serv$.
The axioms formalize the informal explanation given below and in
addition stipulate what is the result of abstracting use and apply if
inappropriate foci or methods are involved.
We use infix notation for the abstracting use and apply operators.
The thread denoted by a closed term of the form $t \sfause t'$ and the
service family denoted by a closed term of the form $t \sfapply t'$ are
the thread and service family, respectively, that result from processing
the method of each basic action performed by the thread denoted by $t$
by the service in the service family denoted by $t'$ with the focus
of the basic action as its name if such a service exists.
When the method of a basic action performed by a thread is processed by
a service, the service changes in accordance with the method concerned
and the thread reduces to one of the two threads that it can possibly
proceed with dependent on the reply value produced by the service.
The projective limit model of the extension of the combination of \BTA\
and \SFA\ with the abstracting use operator, the apply operator, and the
axioms for these operators is a reduct of the projective limit model
presented in~\cite[Section~3.1.9]{BM12b}.
The reduct of this model to the constants and operators of \BTA\ is
the projective limit model of \BTA.
\section{Instructions for Boolean Registers}
\label{sect-instr-br}
The primitive instructions that concern us in the remainder of this
paper are primitive instructions for Boolean registers.
We introduce in this section the possible primitive instructions for
Boolean registers.
It is assumed that, for each $\funct{p,q}{\Bool}{\Bool}$,
$\mbr{p}{q} \in \Meth$.
These methods can be explained as follows:
\begin{quote}
when $\mbr{p}{q}$ is processed by a Boolean register service whose
register content is $b$, the reply is $p(b)$ and the register content
becomes $q(b)$.
\end{quote}
We write $\Methbr$ for the set
$\set{\mbr{p}{q} \where \funct{p,q}{\Bool}{\Bool}}$.
Every method that a Boolean register service could possibly process is a
method from $\Methbr$.
For $\Sig{\Services}$, we take the signature that consists of the sorts,
constants and operators that are mentioned in the assumptions with
respect to services made in Section~\ref{sect-TSI} and a constant
$\BR^M_b$ for each $M \subseteq \Methbr$ and $b \in \Bool$.
Informally, $\BR^M_b$ denotes the Boolean register service with register
content $b$ that is able to process precisely all methods from $M$.
For $\ServAlg$, we take the $\Sig{\Services}$-algebra that has no
proper subalgebra and that satisfies the conditions that are mentioned
in the assumptions with respect to services made in
Section~\ref{sect-TSI} and the following conditions for each
$M \subseteq \Methbr$ and $b \in \Bool$:
\begin{ldispl}
\begin{gceqns}
\derive{\mbr{p}{q}}(\BR^M_b) = \BR^M_{q(b)} & \mif \mbr{p}{q} \in M\;,
\eqnsep
\sreply{\mbr{p}{q}}(\BR^M_b) = p(b) & \mif \mbr{p}{q} \in M\;,
\end{gceqns}
\qquad
\begin{gceqns}
\derive{m}(\BR^M_b) = \emptyserv & \mif m \notin M\;,
\eqnsep
\sreply{m}(\BR^M_b) = \Div & \mif m \notin M\;.
\end{gceqns}
\end{ldispl
$\Bool \to \Bool$, the set of all unary Boolean functions, consists of
the following four functions:
\begin{itemize}
\item
the function $\FFunc$, satisfying
$\FFunc(\False) = \False$ and $\FFunc(\True) = \False$;
\item
the function $\TFunc$, satisfying
$\TFunc(\False) = \True$ and $\TFunc(\True) = \True$;
\item
the function $\IFunc$, satisfying
$\IFunc(\False) = \False$ and $\IFunc(\True) = \True$;
\item
the function $\CFunc$, satisfying
$\CFunc(\False) = \True$ and $\CFunc(\True) = \False$.
\end{itemize}
In~\cite{BM13a}, we actually used the methods $\mbr{\FFunc}{\FFunc}$,
$\mbr{\TFunc}{\TFunc}$, and $\mbr{\IFunc}{\IFunc}$, but denoted them by
$\setbr{0}$, $\setbr{1}$ and $\getbr$, respectively.
In~\cite{BM14e}, we actually used, in addition to these methods, the
method $\mbr{\CFunc}{\CFunc}$, but denoted it by $\negbr$.
We define, for each $M \subseteq \Methbr$, the following sets:
\begin{ldispl}
\BIbr(M) = \set{f.m \where f \in \Foci \Land m \in M}\;,
\eqnsep
\PIbr(M) =
\BIbr(M) \union \set{\ptst{a} \where a \in \BIbr(M)} \union
\set{\ntst{a} \where a \in \BIbr(M)}\;.
\end{ldispl
$\BIbr(\Methbr)$ consists of 16 basic actions per focus and
$\PIbr(\Methbr)$ consists of 48 primitive instructions per focus.
For Boolean registers that serve as input register, we used
in~\cite{BM13a,BM14e} only primitive instructions from
$\PIbr(\set{\mbr{\IFunc}{\IFunc}})$.
For Boolean registers that serve as output register, we used
in~\cite{BM13a,BM14e} only primitive instructions from
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc}})$.
For Boolean registers that serve as auxiliary register, we used
in~\cite{BM13a} only primitive instructions from
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc}})$
and in~\cite{BM14e} only primitive instructions from
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc}})$.
However, in the case of auxiliary registers,
other possible instruction sets are eligible.
In Section~\ref{sect-bounded-complete}, we study instruction sequence
size-bounded functional completeness of instruction sets for Boolean
registers.
We expect that the results of that study will turn out to be useful to
adequately assess results of work that is concerned with lower bounds of
instruction sequence size complexity in cases where auxiliary Boolean
registers may be used.
We write $\ISbr(M)$, where $M \subseteq \Methbr$, for the set of all
finite \PGA\ instruction sequences in the case where $\BIbr(M)$ is taken
for the set $\BInstr$ of basic instructions.
\section{Equivalence of Instructions for Boolean Registers}
\label{sect-eqv-instr-br}
There exists a model of the extension of the combination of \PGA, \BTA,
and \SFA\ with the thread extraction operator, the abstracting use
operator, the apply operator, and the axioms for these operators such
that the initial model of \PGA\ is its reduct to the signature of \PGA\
and the projective limit model of the extension of the combination of
\BTA\ and \SFA\ with the abstracting use operator, the apply operator,
and the axioms for these operators is its reduct to the signature of
this theory.
This follows from the disjointness of the signatures concerned by the
amalgamation result about expansions presented as Theorem 6.1.1
in~\cite{Hod93a} (adapted to the many-sorted case).
Henceforth, we work in the model just mentioned, and denote the
interpretations of constants and operators in it by the constants and
operators themselves.
However, we could work in any model for which the axioms are complete
with respect to satisfaction of equations between closed terms.
On execution of an instruction sequence, different primitive
instructions from $\PIbr(\Methbr)$ do not always have different effects.
We define an equivalence on $\PIbr(\Methbr)$ that identifies primitive
instructions if they have the same effects.
Let $u,v \in \PIbr(\Methbr)$.
Then $u$ and $v$ are \emph{effectually equivalent}, written $u \eeqv v$,
if there exists an $f \in \Foci$ such that, for each $b \in \Bool$ and
$n \in \set{1,2}$:
\begin{center}
\lststretch
\begin{tabular}[t]{@{}l@{}}
$\extr{u \conc \halt^n} \sfause f.\BR^\Methbr_b =
\extr{v \conc \halt^n} \sfause f.\BR^\Methbr_b\;$,
\\
$\extr{u \conc \halt^n} \sfapply f.\BR^\Methbr_b \hsp{.2} {} =
\extr{v \conc \halt^n} \sfapply f.\BR^\Methbr_b\;$.
\end{tabular}
\end{center}
Let $u,v \in \PIbr(\Methbr)$ be such that $u \neq v$, and
let $f \in \Foci$ be such that, for some $m \in \Methbr$,
$v \equiv f.m$ or $v \equiv \ptst{f.m}$ or $v \equiv \ntst{f.m}$.
Then $u \eeqv v$ only if
$\extr{u \conc \halt^n} \sfause f.\BR^\Methbr_b =
\extr{v \conc \halt^n} \sfause f.\BR^\Methbr_b$ and
$\extr{u \conc \halt^n} \sfapply f.\BR^\Methbr_b =
\extr{v \conc \halt^n} \sfapply f.\BR^\Methbr_b$
for each $b \in \Bool$ and $n \in \set{1,2}$.
From this and the definition of $\eeqv$, it follows immediately that
$\eeqv$ is transitive.
Moreover, it follows immediately from the definition of $\eeqv$ that
$\eeqv$ is reflexive and symmetric.
Hence, $\eeqv$ is an equivalence relation indeed.
Replacement of primitive instructions in an instruction sequence by
effectually equivalent ones does not change the functionality of the
instruction sequence on execution.
Let $X,Y \in \ISbr(\Methbr)$.
Then $X$ and $Y$ are \emph{functionally equivalent},
written $X \feqv Y$, if, for some $n \in \Natpos$,
there exist $f_1,\ldots,f_n \in \Foci$ such that,
for each $b_1,\ldots,b_n \in \Bool$:
\begin{center}
\lststretch
\begin{tabular}[t]{@{}l@{}}
$\extr{X} \sfause \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i} = \Stop$ or
$\extr{X} \sfause \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i} = \DeadEnd$,
\\
$\extr{X} \sfause \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i} =
\extr{Y} \sfause \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i}$,
\\
$\extr{X} \sfapply \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i} \hsp{.2} {} =
\extr{Y} \sfapply \Sfcomp{i = 1}{n} f_i.\BR^\Methbr_{b_i}$.
\end{tabular}
\end{center}
The proof that $\feqv$ is an equivalence relation goes along similar
lines as the proof that $\eeqv$ is an equivalence relation.
Here, $X \feqv Y$ only if the equations from the definition hold in the
case where we take the foci of primitive instructions from
$\PIbr(\Methbr)$ that occur in $Y$ for $f_1,\ldots,f_n$.
\begin{proposition}
\label{proposition-eff-equiv}
Let $u,v \in \PIbr(\Methbr)$, and
let $X,Y \in \ISbr(\Methbr)$ be such that $Y$ is $X$ with every
occurrence of $u$ replaced by $v$.
Then $u \eeqv v$ implies $X \feqv Y$.
\end{proposition}
\begin{proof}
It is easily proved by induction on the length of $X$ that $u \eeqv v$
implies, for each $l,n \in \Nat$,
$\fjmp{l} \conc X \conc \halt^n \feqv \fjmp{l} \conc Y \conc \halt^n$
\footnote
{We use the convention that $t'\conc t^0$ stands for $t'$.}
From this, it follows immediately that $u \eeqv v$ implies $X \feqv Y$.
\qed
\end{proof}
Axioms for effectual equivalence are given in
Table~\ref{axioms-eff-equiv}.
\begin{table}[!t]
\caption{Axioms for effectual equivalence}
\label{axioms-eff-equiv}
\begin{eqntbl}
\begin{eqncol}
\ptst{f.\mbr{\FFunc}{p}} \eeqv \ntst{f.\mbr{\TFunc}{p}} \\
\ptst{f.\mbr{\TFunc}{p}} \eeqv \ntst{f.\mbr{\FFunc}{p}} \\
\ptst{f.\mbr{\IFunc}{p}} \eeqv \ntst{f.\mbr{\CFunc}{p}} \\
\ptst{f.\mbr{\CFunc}{p}} \eeqv \ntst{f.\mbr{\IFunc}{p}}
\end{eqncol}
\qquad
\begin{eqncol}
\ptst{f.\mbr{\TFunc}{p}} \eeqv f.\mbr{q}{p}
\end{eqncol}
\qquad
\begin{eqncol}
u \eeqv u \\
u \eeqv v \Limpl v \eeqv u \\
u \eeqv v \Land v \eeqv w \Limpl u \eeqv w
\end{eqncol}
\end{eqntbl}
\end{table}
In this table, $f$ stands for an arbitrary focus from $\Foci$, $p$ and
$q$ stand for arbitrary functions from $\Bool \to \Bool$, and $u$, $v$,
and $w$ stand for arbitrary primitive instructions from
$\PIbr(\Methbr)$.
Moreover, we use $\eeqv$ in this table as a predicate symbol (and not as
the symbol that denotes the effectual equivalence relation on
$\PIbr(\Methbr)$ defined above).
\begin{theorem}
\label{theorem-eff-equiv}
The axioms in Table~\ref{axioms-eff-equiv} are sound and complete for
the effectual equivalence relation on $\PIbr(\Methbr)$ defined above.
\end{theorem}
\begin{proof}
The soundness of the axioms follows immediately from the definition of
effectual equivalence, using the conditions on $\ServAlg$ laid down in
Section~\ref{sect-instr-br}.
The following conclusions can be drawn from the definition of effectual
equivalence:
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(a)} &
$\ptst{f.\mbr{p}{q}} \eeqv f.\mbr{p'}{q'} \Limpl
p = \TFunc \Land q = q'$\,;
\\
\textup{(b)} &
$\ntst{f.\mbr{p}{q}} \eeqv f.\mbr{p'}{q'} \Limpl
p = \FFunc \Land q = q'$\,;
\\
\textup{(c)} &
$\ptst{f.\mbr{p}{q}} \eeqv \ntst{f.\mbr{p'}{q'}} \Limpl
p = C(p') \Land q = q'$\,,
\\ &
where $C(\FFunc) = \TFunc$, $C(\TFunc) = \FFunc$, $C(\IFunc) = \CFunc$,
$C(\CFunc) = \IFunc$.
\end{tabular}
\end{flushleft}
The completeness of the axioms follows easily by case distinction
between the different forms that a formula $u \eeqv v$ can take, making
use of~(a), (b), and~(c).
\qed
\end{proof}
The equivalence classes of $\PIbr(\Methbr)$ with respect to $\eeqv$ are
the following for each $f \in \Foci$:
\begin{ldispl}
\set{\ul{\ptst{f.\mbr{\FFunc}{\FFunc}}},
\ntst{f.\mbr{\TFunc}{\FFunc}}}\;, \\
\set{\ptst{f.\mbr{\FFunc}{\TFunc}},
\ul{\ntst{f.\mbr{\TFunc}{\TFunc}}}}\;, \\
\set{\ptst{f.\mbr{\FFunc}{\IFunc}},
\ul{\ntst{f.\mbr{\TFunc}{\IFunc}}}}\;, \\
\set{\ptst{f.\mbr{\FFunc}{\CFunc}},
\ul{\ntst{f.\mbr{\TFunc}{\CFunc}}}}\;,
\\
\set{\ptst{f.\mbr{\TFunc}{\FFunc}},\ntst{f.\mbr{\FFunc}{\FFunc}},
\ul{f.\mbr{\FFunc}{\FFunc}},f.\mbr{\TFunc}{\FFunc},
f.\mbr{\IFunc}{\FFunc},f.\mbr{\CFunc}{\FFunc}}\;, \\
\set{\ptst{f.\mbr{\TFunc}{\TFunc}},\ntst{f.\mbr{\FFunc}{\TFunc}},
f.\mbr{\FFunc}{\TFunc},\ul{f.\mbr{\TFunc}{\TFunc}},
f.\mbr{\IFunc}{\TFunc},f.\mbr{\CFunc}{\TFunc}}\;, \\
\set{\ptst{f.\mbr{\TFunc}{\IFunc}},\ntst{f.\mbr{\FFunc}{\IFunc}},
f.\mbr{\FFunc}{\IFunc},f.\mbr{\TFunc}{\IFunc},
\ul{f.\mbr{\IFunc}{\IFunc}},f.\mbr{\CFunc}{\IFunc}}\;, \\
\set{\ptst{f.\mbr{\TFunc}{\CFunc}},\ntst{f.\mbr{\FFunc}{\CFunc}},
f.\mbr{\FFunc}{\CFunc},f.\mbr{\TFunc}{\CFunc},
f.\mbr{\IFunc}{\CFunc},\ul{f.\mbr{\CFunc}{\CFunc}}}\;,
\\
\set{\ul{\ptst{f.\mbr{\IFunc}{\FFunc}}},
\ntst{f.\mbr{\CFunc}{\FFunc}}}\;, \\
\set{\ul{\ptst{f.\mbr{\IFunc}{\TFunc}}},
\ntst{f.\mbr{\CFunc}{\TFunc}}}\;, \\
\set{\ul{\ptst{f.\mbr{\IFunc}{\IFunc}}},
\ntst{f.\mbr{\CFunc}{\IFunc}}}\;, \\
\set{\ptst{f.\mbr{\IFunc}{\CFunc}},
\ul{\ntst{f.\mbr{\CFunc}{\CFunc}}}}\;,
\\
\set{\ptst{f.\mbr{\CFunc}{\FFunc}},
\ul{\ntst{f.\mbr{\IFunc}{\FFunc}}}}\;, \\
\set{\ptst{f.\mbr{\CFunc}{\TFunc}},
\ul{\ntst{f.\mbr{\IFunc}{\TFunc}}}}\;, \\
\set{\ptst{f.\mbr{\CFunc}{\IFunc}},
\ul{\ntst{f.\mbr{\IFunc}{\IFunc}}}}\;, \\
\set{\ul{\ptst{f.\mbr{\CFunc}{\CFunc}}},
\ntst{f.\mbr{\IFunc}{\CFunc}}}\;.
\end{ldispl
We have underlined one representative of each equivalence class in order
to refer to them easily in the proof of the following theorem.
\begin{theorem}
\label{theorem-min-meth-set}
\mbox{}
\begin{itemize}
\item[\textup{(1)}]
The set
$\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc},
\mbr{\TFunc}{\IFunc},\mbr{\TFunc}{\CFunc}}$
is a minimal $M \subseteq \Methbr$ such that $\PIbr(M)$ contains at
least one representative from each of the equivalence classes of
$\PIbr(\Methbr)$ with respect to $\eeqv$.
\item[\textup{(2)}]
Each minimal $M \subseteq \Methbr$ such that $\PIbr(M)$ contains at
least one representative from each of the equivalence classes of
$\PIbr(\Methbr)$ with respect to $\eeqv$ consists of $8$ methods.
\end{itemize}
\end{theorem}
\begin{proof}
By uniformity, it is sufficient to look at the equivalence classes of
$\PIbr(\Methbr)$ for an arbitrary focus from $\Foci$.
\begin{itemize}
\item[(1)]
Let
$M'= \set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc},
\mbr{\TFunc}{\IFunc},\mbr{\TFunc}{\CFunc}}$.
Then the representatives of the different equivalence classes of
$\PIbr(\Methbr)$ that are underlined above belong to the set
$\PIbr(M')$.
Moreover, each method from $M'$ occurs in a primitive instruction from
$\PIbr(\Methbr)$ that belongs to an equivalence class of
$\PIbr(\Methbr)$ that contains only one other primitive instruction, but
the method that occurs in this other primitive instruction is not from
$M'$.
Hence $M'$ is minimal.
\item[(2)]
First, we consider the first and last four equivalence classes above.
Each of them consists of two primitive instructions.
Each method that occurs in the primitive instructions from one of them
does not occur in the primitive instructions from another of them.
Consequently, exactly eight methods are needed for representatives from
these equivalence classes.
Next, we consider the remaining eight equivalence classes.
For each of them, the methods that occur in the primitive instructions
from it include the methods that occur in the primitive instructions
from one of the equivalence classes that we considered first.
Consequently, no additional methods are needed for representatives from
the remaining equivalence classes.
Hence, exactly eight methods are needed for representatives from all
equivalence classes. \qed
\end{itemize}
\end{proof}
Theorem~\ref{theorem-min-meth-set} tells us that each primitive
instruction from $\PIbr(\Methbr)$ has the same effects as one with a
method from
$\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc},
\mbr{\TFunc}{\IFunc},\mbr{\TFunc}{\CFunc}}$
and that there does not exist a smaller set with this
property.
The methods that we used in~\cite{BM13a,BM14e} are included in this set.
We have the following corollary of the proof of part~(2) of
Theorem~\ref{theorem-min-meth-set}.
\begin{corollary}
\label{corollary-min-meth-set}
There exist $256$ minimal $M \subseteq \Methbr$ such that $\PIbr(M)$
contains at least one representative from each of the equivalence
classes of $\PIbr(\Methbr)$ with respect to $\eeqv$.
\end{corollary}
\section{Bounded Functional Completeness of Instruction Sets}
\label{sect-bounded-complete}
Not all methods from the minimal set mentioned in
Theorem~\ref{theorem-min-meth-set} are needed to obtain the effects of
each primitive instruction from $\PIbr(\Methbr)$ in the case where
instruction sequences instead of instructions are used to obtain the
effects.
In this section, we look at the case where instruction sequences are
used.
We begin by defining the notion of $k$-size-bounded functional
completeness.
Let $M \subseteq \Methbr$ and $k \in \Natpos$.
Then the instruction set $\PIbr(M)$ is
\emph{$k$-size-bounded functionally complete} if
there exists a function $\funct{\psi}{\PIbr(\Methbr)}{\ISbr(M)}$ such
that, for each $u \in \PIbr(\Methbr)$, $\len(\psi(u)) \leq k$ and
there exists an $f \in \Foci$ such that,
for each $b \in \Bool$ and $n \in \Nat$:
\begin{center}
\lststretch
\begin{tabular}[t]{@{}l@{}}
$\extr{u \conc \halt^n} \sfause f.\BR^\Methbr_b =
\extr{\psi(u) \conc \halt^n} \sfause f.\BR^\Methbr_b$,
\\
$\extr{u \conc \halt^n} \sfapply f.\BR^\Methbr_b \hsp{.2} {} =
\extr{\psi(u) \conc \halt^n} \sfapply f.\BR^\Methbr_b$.
\end{tabular}
\end{center}
$\PIbr(M)$ is called
\emph{strictly $k$-size-bounded functionally complete} if $\PIbr(M)$ is
$k$-size-bounded functionally complete and there does not exist a
$k' < k$ such that $\PIbr(M)$ is $k'$-size-bounded functionally
complete.
The following proposition illustrates the relevance of the notion of
$k$-size-bounded functionally completeness.
\begin{proposition}
\label{proposition-bounded-complete}
Let $M \subseteq \Methbr$ and $k \in \Natpos$.
Let $\funct{\psi}{\PIbr(\Methbr)}{\ISbr(M)}$ be as in the definition of
$k$-size-bounded functional completeness given above.
Let $\funct{\psi'}{\ISbr(\Methbr)}{\ISbr(M)}$ be such that
$\psi'(u_1 \conc \ldots \conc u_n) = u_1' \conc \ldots \conc u_n'$,
where
\begin{center}
\lststretch
\begin{tabular}[t]{@{}l@{\,\,}l@{}}
$u_i' \equiv \halt$ & if $u_i \equiv \halt$;
\\
$u_i' \equiv \fjmp{l'}$ \\ \,\,\,\, with
$l' =
l + \sum_{j \in \set{i,\ldots,i+l-1}
\mathrm{\,s.t.\,} u_j \in \PIbr(\Methbr)}
(\len(\psi(u_j)) - 1)$ & if $u_i \equiv \fjmp{l}$;
\\
$u_i' \equiv \psi(u_i)$ & otherwise.
\end{tabular}
\end{center}
Assume that $\psi$ restricted to $\ISbr(M)$ is the identity function on
$\ISbr(M)$.
Then, for each $X \in \ISbr(\Methbr)$, $\psi'(X) \feqv X$ and
$\len(\psi'(X)) \leq \len(X) + (k - 1) \mul p$, where $p$ is the number
of occurrences of primitive instructions from
$\PIbr(\Methbr) \diff \PIbr(M)$ in $X$.
\end{proposition}
\begin{proof}
It is easily proved by induction on the length of $X$ that, for each
$l,n \in \Nat$,
$\fjmp{l} \conc \psi'(X) \conc \halt^n \feqv
\fjmp{l} \conc X \conc \halt^n$.
From this, it follows immediately that $\psi'(X) \feqv X$.
Suppose that $X = u_1 \conc \ldots \conc u_n$.
Let $p$ be the number of occurrences of primitive instructions from
$\PIbr(\Methbr) \diff \PIbr(M)$ in $X$.
Then
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{}}
$\sum_{i \in \set{1,\ldots,n}
\mathrm{\,s.t.\,} u_i \notin \PIbr(\Methbr) \diff \PIbr(M)}
\len(u_i) \phantom{\psi()} =
\len(X) - p\;$;
\\
$\sum_{i \in \set{1,\ldots,n}
\mathrm{\,s.t.\,} u_i \in \PIbr(\Methbr) \diff \PIbr(M)}
\len(\psi(u_i)) \leq
k \mul p\;$.
\end{tabular}
\end{flushleft}
Hence,
$\len(\psi'(X)) \leq \len(X) - p + k \mul p = \len(X) + (k - 1) \mul p$.
\qed
\end{proof}
We have the following corollary of part~(1) of
Theorem~\ref{theorem-min-meth-set} and the definition of
$k$-size-bounded functional completeness.
\begin{corollary}
\label{corollary-bounded-complete-1}
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc},
\mbr{\TFunc}{\IFunc},\mbr{\TFunc}{\CFunc}})$
is $1$-size-bounded \sloppy functionally complete.
\end{corollary}
The following theorem concerns the $k$-size-bounded functional
completeness of a few subsets of this $1$-size-bounded functionally
complete instruction set, including the ones that we used
in~\cite{BM13a,BM14e}.
\begin{theorem}
\label{theorem-bounded-complete}
\mbox{}
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{}l@{\hsp{.3}}l@{\hsp{.15}}c@{\hsp{.2}}r@{}}
\textup{(1)} &
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}})$
& is & strictly $2$-size-bounded funct.\ compl.
\\
\textup{(2)} &
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc}})$
& is & strictly $3$-size-bounded funct.\ compl.
\\
\textup{(3)} &
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc}})$
& is & strictly $4$-size-bounded funct.\ compl.
\\
\textup{(4)} &
$\PIbr(\set{\mbr{\CFunc}{\CFunc}})$
& is & strictly $3$-size-bounded funct.\ compl.
\\
\textup{(5)} &
$\PIbr(\set{\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}})$
& is & strictly $4$-size-bounded funct.\ compl.
\end{tabular}
\end{flushleft}
\end{theorem}
\begin{proof}
\sloppy
We assume that, for each $M \subseteq \Methbr$ and
$k \in \Natpos$, the restriction to $\PIbr(M)$ of a function $\psi$ that
witnesses $k$-size-bounded functional completeness of $\PIbr(M)$ is the
identify function on $\PIbr(M)$ and the restriction to
$\PIbr(\Methbr) \diff \PIbr(M)$ has the same instruction sequence from
$\ISbr(M)$ as value for primitive instruction from the same equivalence
class of $\PIbr(\Methbr)$.
It is clear that this assumption can be made without loss of generality.
Below, for each individual part of the theorem, first a function $\psi$
that witnesses the stated size-bounded functional completeness is
uniquely characterized by giving the instruction sequences for the
primitive instructions for Boolean registers that are not covered by the
assumption and then the strictness of the stated size-bounded functional
completeness is established by checking for one of the primitive
instructions concerned for which the given instruction sequence was of
the greatest length that the given instruction sequence cannot be
replaced by a shorter one.
We say that a $u \in \PIbr(\Methbr)$ cannot be replaced by a jump
instruction if there exists a $b \in \Bool$ and $n \in \Natpos$ such
that, for each $v \in \set{\fjmp{l} \where l \in \Nat}$,
$\extr{u \conc \halt^n} \sfause f.\BR^\Methbr_b \neq
\extr{v \conc \halt^n} \sfause f.\BR^\Methbr_b$.
\begin{itemize}
\item[(1)]
Let
$M = \set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}}$.
Then instruction sequences from $\ISbr(M)$ are needed for
$\ntst{f.\mbr{\TFunc}{\IFunc}}$ and $\ntst{f.\mbr{\TFunc}{\CFunc}}$.
Take $\psi$ such that
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(a)} &
$\psi(\ntst{f.\mbr{\TFunc}{\IFunc}}) = \fjmp{2}$,
\\
\textup{(b)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
f.\mbr{\CFunc}{\CFunc} \conc \fjmp{2}$.
\end{tabular}
\end{flushleft}
Then $\psi$ witnesses the $2$-size-bounded functional completeness of
$\PIbr(M)$.
Because $\ntst{f.\mbr{\TFunc}{\CFunc}}$ cannot be replaced by a jump
instruction and there exists no $u \in \PIbr(M)$ such that
$u \eeqv \ntst{f.\mbr{\TFunc}{\CFunc}}$, $\PIbr(M)$ is not
$1$-size-bounded functionally complete.
Hence, $\PIbr(M)$ is strictly $2$-size-bounded functionally complete.
\item[(2)]
Let
$M = \set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc}}$.
Then instruction sequences from $\ISbr(M)$ are needed for
$\ntst{f.\mbr{\TFunc}{\IFunc}}$, $\ntst{f.\mbr{\TFunc}{\CFunc}}$,
$\ptst{f.\mbr{\IFunc}{\FFunc}}$, $\ntst{f.\mbr{\IFunc}{\FFunc}}$,
$\ptst{f.\mbr{\IFunc}{\TFunc}}$, and $\ntst{f.\mbr{\IFunc}{\TFunc}}$.
Take $\psi$ such that (a), (b), and
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(c1)} &
$\psi(\ptst{f.\mbr{\IFunc}{\FFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
\ptst{f.\mbr{\FFunc}{\FFunc}}$ or \\
\textup{(c2)} &
$\psi(\ptst{f.\mbr{\IFunc}{\FFunc}}) =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
\ptst{f.\mbr{\FFunc}{\FFunc}}$ or \\
\textup{(c3)} &
$\psi(\ptst{f.\mbr{\IFunc}{\FFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\CFunc}{\CFunc}} \conc
\fjmp{2}$ or \\
\textup{(c4)} &
$\psi(\ptst{f.\mbr{\IFunc}{\FFunc}}) =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc \fjmp{2} \conc
\ptst{f.\mbr{\CFunc}{\CFunc}}$,
\\
\textup{(d1)} &
$\psi(\ntst{f.\mbr{\IFunc}{\FFunc}}) =
\ntst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
\ptst{f.\mbr{\FFunc}{\FFunc}}$ or \\
\textup{(d2)} &
$\psi(\ntst{f.\mbr{\IFunc}{\FFunc}}) =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
\ptst{f.\mbr{\FFunc}{\FFunc}}$ or \\
\textup{(d3)} &
$\psi(\ntst{f.\mbr{\IFunc}{\FFunc}}) =
\ntst{f.\mbr{\IFunc}{\IFunc}} \conc \fjmp{2} \conc
\ptst{f.\mbr{\CFunc}{\CFunc}}$ or \\
\textup{(d4)} &
$\psi(\ntst{f.\mbr{\IFunc}{\FFunc}}) =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc \ptst{f.\mbr{\CFunc}{\CFunc}} \conc
\fjmp{2}$,
\\
\textup{(e1)} &
$\psi(\ptst{f.\mbr{\IFunc}{\TFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ntst{f.\mbr{\TFunc}{\TFunc}} \conc
\ntst{f.\mbr{\TFunc}{\TFunc}}$ or \\
\textup{(e2)} &
$\psi(\ptst{f.\mbr{\IFunc}{\TFunc}}) =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc \ntst{f.\mbr{\TFunc}{\TFunc}} \conc
\ntst{f.\mbr{\TFunc}{\TFunc}}$ or \\
\textup{(e3)} &
$\psi(\ptst{f.\mbr{\IFunc}{\TFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\CFunc}{\CFunc}}$ or \\
\textup{(e4)} &
$\psi(\ptst{f.\mbr{\IFunc}{\TFunc}}) =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc \ntst{f.\mbr{\CFunc}{\CFunc}} \conc
\fjmp{2}$,
\end{tabular}
\end{flushleft}
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(f1)} &
$\psi(\ntst{f.\mbr{\IFunc}{\TFunc}}) =
\ntst{f.\mbr{\IFunc}{\IFunc}} \conc \ntst{f.\mbr{\TFunc}{\TFunc}} \conc
\ntst{f.\mbr{\TFunc}{\TFunc}}$ or \\
\textup{(f2)} &
$\psi(\ntst{f.\mbr{\IFunc}{\TFunc}}) =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc \ntst{f.\mbr{\TFunc}{\TFunc}} \conc
\ntst{f.\mbr{\TFunc}{\TFunc}}$ or \\
\textup{(f3)} &
$\psi(\ntst{f.\mbr{\IFunc}{\TFunc}}) =
\ntst{f.\mbr{\IFunc}{\IFunc}} \conc \ntst{f.\mbr{\CFunc}{\CFunc}} \conc
\fjmp{2}$ or \\
\textup{(f4)} &
$\psi(\ntst{f.\mbr{\IFunc}{\TFunc}}) =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\CFunc}{\CFunc}}$.\footnotemark
\end{tabular}
\footnotetext
{For several instruction sequences that start with a test instruction,
there is a counterpart with the same methods in the same numbers that
starts with the opposite test instruction.
We refrain from mentioning these counterparts as alternatives.}
\end{flushleft}
Then $\psi$ witnesses the $3$-size-bounded functional completeness of
$\PIbr(M)$.
To obtain the effects of $\ptst{f.\mbr{\IFunc}{\FFunc}}$, an instruction
sequence from $\ISbr(M)$ is needed that contains a test instruction from
$\PIbr(M)$ with $\mbr{\IFunc}{\IFunc}$ or $\mbr{\CFunc}{\CFunc}$ as
method and a primitive instruction from $\PIbr(M)$ with
$\mbr{\FFunc}{\FFunc}$ or $\mbr{\CFunc}{\CFunc}$ as method.
Because there does not exist such an instruction sequence of length $2$
with the right effects, $\PIbr(M)$ is not $2$-size-bounded functionally
complete.
Hence, $\PIbr(M)$ is strictly $3$-size-bounded functionally complete.
\item[(3)]
Let
$M = \set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc}}$.
Then instruction sequences from $\ISbr(M)$ are needed for
$\ntst{f.\mbr{\TFunc}{\IFunc}}$, $\ptst{f.\mbr{\IFunc}{\FFunc}}$,
$\ntst{f.\mbr{\IFunc}{\FFunc}}$, $\ptst{f.\mbr{\IFunc}{\TFunc}}$,
$\ntst{f.\mbr{\IFunc}{\TFunc}}$, $f.\mbr{\CFunc}{\CFunc}$,
$\ptst{f.\mbr{\CFunc}{\CFunc}}$, $\ntst{f.\mbr{\CFunc}{\CFunc}}$,
and $\ntst{f.\mbr{\TFunc}{\CFunc}}$.
Take $\psi$ such that (a), (c1) or (c3), (d1) or (d3), (e1) or (e3),
(f1) or (f3), and
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(g)} &
$\psi(f.\mbr{\CFunc}{\CFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
f.\mbr{\TFunc}{\TFunc}$,
\\
\textup{(h)} &
$\psi(\ptst{f.\mbr{\CFunc}{\CFunc}}) =
\ntst{f.\mbr{\IFunc}{\IFunc}} \conc \ntst{f.\mbr{\TFunc}{\TFunc}} \conc
\ptst{f.\mbr{\FFunc}{\FFunc}}$,
\\
\textup{(i)} &
$\psi(\ntst{f.\mbr{\CFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
\ntst{f.\mbr{\TFunc}{\TFunc}}$,
\\
\textup{(j)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\IFunc}} \conc \ptst{f.\mbr{\FFunc}{\FFunc}} \conc
f.\mbr{\TFunc}{\TFunc} \conc \fjmp{2}$.
\end{tabular}
\end{flushleft}
Then $\psi$ witnesses the $4$-size-bounded functional completeness of
$\PIbr(M)$.
To obtain the effects of $\ntst{f.\mbr{\TFunc}{\CFunc}}$, an instruction
sequence from $\ISbr(M)$ is needed that contains a test instruction from
$\PIbr(M)$ with $\mbr{\IFunc}{\IFunc}$ as method, a primitive
instruction from $\PIbr(M)$ with $\mbr{\FFunc}{\FFunc}$ as method, and a
primitive instruction from $\PIbr(M)$ with $\mbr{\TFunc}{\TFunc}$ as
method.
Because there does not exist such an instruction sequence of length $3$
with the right effects, $\PIbr(M)$ is not $3$-size-bounded functionally
complete.
Hence, $\PIbr(M)$ is strictly $4$-size-bounded functionally complete.
\item[(4)]
Let
$M = \set{\mbr{\CFunc}{\CFunc}}$.
Then instruction sequences from $\ISbr(M)$ are needed for
$\ntst{f.\mbr{\TFunc}{\IFunc}}$, $\ntst{f.\mbr{\TFunc}{\CFunc}}$,
$\ptst{f.\mbr{\IFunc}{\FFunc}}$, $\ntst{f.\mbr{\IFunc}{\FFunc}}$,
$\ptst{f.\mbr{\IFunc}{\TFunc}}$, $\ntst{f.\mbr{\IFunc}{\TFunc}}$,
$f.\mbr{\FFunc}{\FFunc}$, $\ptst{f.\mbr{\FFunc}{\FFunc}}$,
$f.\mbr{\TFunc}{\TFunc}$, $\ntst{f.\mbr{\TFunc}{\TFunc}}$,
$f.\mbr{\IFunc}{\IFunc}$, $\ptst{f.\mbr{\IFunc}{\IFunc}}$, and
$\ntst{f.\mbr{\IFunc}{\IFunc}}$.
Take $\psi$ such that (a), (b), (c4), (d4), (e4), (f4), and
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(k)} &
$\psi(f.\mbr{\FFunc}{\FFunc}) \phantom{+} =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc f.\mbr{\CFunc}{\CFunc}$,
\\
\textup{(l)} &
$\psi(\ptst{f.\mbr{\FFunc}{\FFunc}}) =
\ptst{f.\mbr{\CFunc}{\CFunc}} \conc f.\mbr{\CFunc}{\CFunc} \conc
\fjmp{2}$,
\\
\textup{(m)} &
$\psi(f.\mbr{\TFunc}{\TFunc}) \phantom{+} =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc f.\mbr{\CFunc}{\CFunc}$,
\\
\textup{(n)} &
$\psi(\ntst{f.\mbr{\TFunc}{\TFunc}}) =
\ntst{f.\mbr{\CFunc}{\CFunc}} \conc f.\mbr{\CFunc}{\CFunc} \conc
\fjmp{2}$,
\end{tabular}
\end{flushleft}
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(o)} &
$\psi(f.\mbr{\IFunc}{\IFunc}) \phantom{+} =
f.\mbr{\CFunc}{\CFunc} \conc f.\mbr{\CFunc}{\CFunc}$,
\\
\textup{(p)} &
$\psi(\ptst{f.\mbr{\IFunc}{\IFunc}}) =
f.\mbr{\CFunc}{\CFunc} \conc \ptst{f.\mbr{\CFunc}{\CFunc}}$,
\\
\textup{(q)} &
$\psi(\ntst{f.\mbr{\IFunc}{\IFunc}}) =
f.\mbr{\CFunc}{\CFunc} \conc \ntst{f.\mbr{\CFunc}{\CFunc}}$.
\end{tabular}
\end{flushleft}
Then $\psi$ witnesses the $3$-size-bounded functional completeness of
$\PIbr(M)$.
To obtain the effects of $\ptst{f.\mbr{\FFunc}{\FFunc}}$, an instruction
sequence from $\ISbr(M)$ is needed.
Because there does not exist such an instruction sequence of length $2$
with the right effects, $\PIbr(M)$ is not $2$-size-bounded functionally
complete.
Hence, $\PIbr(M)$ is strictly $3$-size-bounded functionally complete.
\item[(5)]
Let
$M = \set{\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}}$.
Then instruction sequences from $\ISbr(M)$ are needed for
$\ntst{f.\mbr{\TFunc}{\IFunc}}$, $f.\mbr{\FFunc}{\FFunc}$,
$\ptst{f.\mbr{\FFunc}{\FFunc}}$, $f.\mbr{\TFunc}{\TFunc}$,
$\ntst{f.\mbr{\TFunc}{\TFunc}}$, $f.\mbr{\IFunc}{\IFunc}$,
$\ptst{f.\mbr{\IFunc}{\IFunc}}$, $\ntst{f.\mbr{\IFunc}{\IFunc}}$,
$f.\mbr{\CFunc}{\CFunc}$, $\ptst{f.\mbr{\CFunc}{\CFunc}}$,
$\ntst{f.\mbr{\CFunc}{\CFunc}}$, and $\ntst{f.\mbr{\TFunc}{\CFunc}}$.
Take $\psi$ such that (a) and
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(r)} &
$\psi(f.\mbr{\FFunc}{\FFunc}) \phantom{+} =
f.\mbr{\IFunc}{\FFunc}$,
\\
\textup{(s)} &
$\psi(\ptst{f.\mbr{\FFunc}{\FFunc}}) =
f.\mbr{\IFunc}{\FFunc} \conc \fjmp{2}$,
\\
\textup{(t)} &
$\psi(f.\mbr{\TFunc}{\TFunc}) \phantom{+} =
f.\mbr{\IFunc}{\TFunc}$,
\\
\textup{(u)} &
$\psi(\ntst{f.\mbr{\TFunc}{\TFunc}}) =
f.\mbr{\IFunc}{\TFunc} \conc \fjmp{2}$,
\\
\textup{(v1)} &
$\psi(f.\mbr{\IFunc}{\IFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\ntst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(v2)} &
$\psi(f.\mbr{\IFunc}{\IFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\TFunc}} \conc
\ptst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(v3)} &
$\psi(f.\mbr{\IFunc}{\IFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\fjmp{1}$ or
\\
\textup{(v4)} &
$\psi(f.\mbr{\IFunc}{\IFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2} \conc
\ptst{f.\mbr{\IFunc}{\FFunc}}$,
\\
\textup{(w1)} &
$\psi(\ptst{f.\mbr{\IFunc}{\IFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\ptst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(w2)} &
$\psi(\ptst{f.\mbr{\IFunc}{\IFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\TFunc}} \conc
\ntst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(w3)} &
$\psi(\ptst{f.\mbr{\IFunc}{\IFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\fjmp{2}$ or
\\
\textup{(w4)} &
$\psi(\ptst{f.\mbr{\IFunc}{\IFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\IFunc}{\FFunc}}$,
\\
\textup{(x1)} &
$\psi(\ntst{f.\mbr{\IFunc}{\IFunc}}) =
\ntst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\FFunc}} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(x2)} &
$\psi(\ntst{f.\mbr{\IFunc}{\IFunc}}) =
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(x3)} &
$\psi(\ntst{f.\mbr{\IFunc}{\IFunc}}) =
\ntst{f.\mbr{\IFunc}{\FFunc}} \conc \fjmp{2} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(x4)} &
$\psi(\ntst{f.\mbr{\IFunc}{\IFunc}}) =
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\fjmp{2}$,
\\
\textup{(y1)} &
$\psi(f.\mbr{\CFunc}{\CFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\FFunc}} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(y2)} &
$\psi(f.\mbr{\CFunc}{\CFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(y3)} &
$\psi(f.\mbr{\CFunc}{\CFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(y4)} &
$\psi(f.\mbr{\CFunc}{\CFunc}) \phantom{+} =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\fjmp{1}$,
\\
\textup{(z1)} &
$\psi(\ptst{f.\mbr{\CFunc}{\CFunc}}) =
\ntst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\ptst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(z2)} &
$\psi(\ptst{f.\mbr{\CFunc}{\CFunc}}) =
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\TFunc}} \conc
\ntst{f.\mbr{\IFunc}{\FFunc}}$ or
\\
\textup{(z3)} &
$\psi(\ptst{f.\mbr{\CFunc}{\CFunc}}) =
\ntst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\TFunc}} \conc
\fjmp{2}$ or
\\
\textup{(z4)} &
$\psi(\ptst{f.\mbr{\CFunc}{\CFunc}}) =
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\IFunc}{\FFunc}}$,
\end{tabular}
\end{flushleft}
\begin{flushleft}
\lststretch
\begin{tabular}[t]{@{\hsp{5}}l@{\hsp{.7}}l@{}}
\textup{(aa1)} &
$\psi(\ntst{f.\mbr{\CFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\FFunc}} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(aa2)} &
$\psi(\ntst{f.\mbr{\CFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(aa3)} &
$\psi(\ntst{f.\mbr{\CFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \fjmp{2} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}}$ or
\\
\textup{(aa4)} &
$\psi(\ntst{f.\mbr{\CFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\fjmp{2}$,
\\
\textup{(ab1)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \ptst{f.\mbr{\IFunc}{\FFunc}} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2}$ or
\\
\textup{(ab2)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2}$ or
\\
\textup{(ab3)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\FFunc}} \conc \fjmp{2} \conc
\ntst{f.\mbr{\IFunc}{\TFunc}} \conc \fjmp{2}$ or
\\
\textup{(ab4)} &
$\psi(\ntst{f.\mbr{\TFunc}{\CFunc}}) =
\ptst{f.\mbr{\IFunc}{\TFunc}} \conc \ntst{f.\mbr{\IFunc}{\FFunc}} \conc
\fjmp{1} \conc \fjmp{2}$.
\end{tabular}
\end{flushleft}
Then $\psi$ witnesses the $4$-size-bounded functional completeness of
$\PIbr(M)$.
To obtain the effects of $\ntst{f.\mbr{\TFunc}{\CFunc}}$, an instruction
sequence from $\ISbr(M)$ is needed.
Because there does not exist such an instruction sequence of length $3$
with the right effects, $\PIbr(M)$ is not $3$-size-bounded functionally
complete.
Hence, $\PIbr(M)$ is strictly $4$-size-bounded functionally complete.
\qed
\end{itemize}
\end{proof}
Theorem~\ref{theorem-bounded-complete} tells us among other things
that the instruction sets
$\PIbr(\set{\mbr{\CFunc}{\CFunc}})$
and
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc}})$
are both strictly $3$-size-bounded functionally complete.
However, the latter instruction set often gives rise to shorter
instruction sequences than the former instruction set because the
effects of a primitive instruction from the set
$\PIbr(\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc}})$
do not have to be obtained by means of two or three primitive
instructions from the set
$\PIbr(\set{\mbr{\CFunc}{\CFunc}})$.
We have the following corollary of the proof of
Theorem~\ref{theorem-bounded-complete}.
\begin{corollary}
\label{corollary-bounded-complete-2}
Let $M \subset \set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},
\mbr{\IFunc}{\IFunc},\mbr{\CFunc}{\CFunc},
\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}}$.
Then:
\begin{itemize}
\item[\textup{(1)}]
$\PIbr(M)$ is strictly $4$-size-bounded functionally complete
if $\mbr{\CFunc}{\CFunc} \notin M$ and
either
$\set{\mbr{\FFunc}{\FFunc},\mbr{\TFunc}{\TFunc},\mbr{\IFunc}{\IFunc}}
\subseteq M$
or
$\set{\mbr{\IFunc}{\FFunc},\mbr{\IFunc}{\TFunc}} \subseteq M$;
\item[\textup{(2)}]
$\PIbr(M)$ is strictly $3$-size-bounded functionally complete
if $\mbr{\CFunc}{\CFunc} \in M$.
\end{itemize}
\end{corollary}
\section{Concluding Remarks}
\label{sect-concl}
We have investigated instruction sequence size bounded functional
completeness of instruction sets for Boolean registers.
Our main results are Corollaries~\ref{corollary-bounded-complete-1}
and~\ref{corollary-bounded-complete-2}.
The latter corollary covers 44 instruction sets.
The covered instruction sets include the instruction sets that we used
earlier in~\cite{BM13a,BM14e} and many other relatively obvious
instruction sets.
The covered instruction sets belong to the 255 instruction sets that are
non-empty subsets of one of the 256 instruction sets with the
property that each possible instruction has the same effects as one
from the set (see Corollary~\ref{corollary-min-meth-set}).
It is still an open question what is the smallest $k$ such that each of
these 255 instruction sets is $k$-size-bounded functionally complete if
it is $k'$-size-bounded functionally complete for some $k'$.
In our work on instruction sequence size complexity presented
in~\cite{BM13a}, we have established several connections between
instruction sequence based complexity theory and classical complexity
theory.
For example, we have introduced instruction sequence based counterparts
of the complexity classes P/poly and NP/poly and we have formulated an
instruction sequence based counterpart of the well-known
complexity-theoretic conjecture that \mbox{NP $\not\subseteq$ P/poly}
\footnote
{The non-uniform complexity classes P/poly and NP/poly, as well as the
conjecture that NP $\not\subseteq$ P/poly, are treated in many
textbooks on classical complexity theory
(see e.g.~\cite{AB09a,HS11a,Weg05a}).}
However, for many a question that arises naturally with the approach to
complexity based on instruction sequence size, it is far from obvious
whether a comparable question can be raised in classical complexity
theory based on Turing machines or Boolean circuits.
In particular, this is far from obvious for questions concerning
instruction sets for Boolean registers.
\subsection*{Acknowledgement}
We thank three anonymous referees for their helpful suggestions.
\bibliographystyle{splncs03}
|
\section{Introduction}
Modelling the dynamical coupling of electrons and nuclei beyond the Born-Oppenheimer (BO), or adiabatic, regime is currently among the most challenging problems in the fields of Theoretical Chemistry and Condensed Matter Physics. Within the BO framework, molecular systems are visualized as a set of nuclei moving on a single potential energy surface that represents the effect of the electrons in a given eigenstate. Many interesting phenomena, however, such as vision~\cite{cerulloN2010, ishidaJPCB2012}, charge separation in organic photovoltaic materials~\cite{rozziNC2013, jailaubekovNM2013} or Joule heating in molecular junctions~\cite{todorovJPCMN2007, horsfield_JPhysCM2004}, occur in non-adiabatic conditions. In these situations, solving exactly the time-dependent Schr\"odinger equation (TDSE) for the coupled system of electrons and nuclear is not feasible, as the computational cost scales exponentially with the number of degrees of freedom. However, since the full quantum treatment requires to represent the problem in terms of adiabatic states and transitions among them in regions of strong non-adiabatic coupling, wave packet propagation techniques have been developed, retaining the quantum character of the nuclear dynamics~\cite{cederbaumCPL1990, cederbaumJCP1992, burghardtJCP1999, worthTCA2003, thossCM2004, mayJCP2006, wastermannJCP2013, thossJCP2003, thossJPCC2010, mengJCP2012, brownPCCP, martinezCPL1996, martinezJPC1996, martinezACR2006}. And these techniques are presently the state of the art in quantum dynamics computational methods, proving the benchmark for approximate methods. In fact, for large systems, the dimensionality of the problem does not allow to employ a quantum mechanical description, thus the only feasible approach is to combine a classical description for the nuclei (or ions) with a quantum treatment of a few other essential degrees of freedom, e.g. electrons or protons . In this context, the question of how to model the coupling between the quantum and classical subsystems still remains open, despite the fact that several schemes~\cite{ehrenfest, pechukasPR1969_2, TSH_1990, Thoss_PRL1997, Miller_JCP1997_2, truhlarFD2004, ciccottiJCP2000, kapralJCP2013, bonellaJCP2005, cokerJCP2012, marxPRL2002, martinezACR2006, tavernelliPCCP2011, wyattJCP2001, burghardtJCP2005, prezhdoPRL2001, tannorJCP2012, burghardtJCP2011,thossJCP2013, ananthJCP2013} have been proposed in the literature trying to settle this issue.
In recent work, we have addressed this problem in the context of the exact factorization of the electron-nuclear wave function~\cite{AMG,AMG2}. In such a treatment of quantum dynamics, the solution of the TDSE is written as a single product of a nuclear wave function and an electronic factor, that parametrically depends on the nuclear configuration. Several advantages of this reformulation have been pointed out. First of all, it has been shown~\cite{steps,long_steps} that the nuclear wave function evolves according to a modified TDSE where a time-dependent vector potential and a time-dependent scalar potential represent the effect of the electrons on the nuclei, beyond the adiabatic regime. When a classical treatment of the nuclear degrees of freedom is introduced, the coupling to the electrons is exactly represented by the force determined from the gradient of the time-dependent potentials~\cite{long_steps,long_steps_mt}. On the other hand, the electronic factor evolves according to a (less standard) evolution equation, coupled to the nuclear TDSE, where the effect of the nuclei is represented by an \textsl{electron-nuclear coupling operator} explicitly depending on the nuclear wave function.
The analysis of the time-dependent potentials and of the classical nuclear force has been the subject of previous work. We have been able to analyze a simple model system to pinpoint some relevant features~\cite{steps,long_steps,long_steps_mt} of the potentials that should be accounted for when developing approximations. Based on these observations, starting from the exact formulation, we have proposed~\cite{mqc, long_mqc} a novel mixed quantum-classical algorithm to solve the coupled electronic and nuclear evolution equations in a fully approximated way. The classical limit is considered as the lowest-order, in a $\hbar$-expansion, of the nuclear wave function in the complex-phase representation~\cite{VanVleck_PNAS1928}.
In the present paper, the focus is directed towards the analysis of the \textsl{electron-nuclear coupling operator} that, in the electronic equation, mediates the coupling to the nuclei and that depends explicitly on (the gradient of) the nuclear wave function. It is fundamental to be able to correctly approximate such term, as it is responsible to induce electronic non-adiabatic transitions~\cite{mqc,long_mqc} and decoherence~\cite{MQC_min}. When a classical description of the nuclei is adopted, the concept of wave function is somehow lost and problems arise when approximating this operator. If a distribution of trajectories~\cite{long_steps_mt} is used to mimic the evolution of the nuclear wave function, its modulus and phase cannot be smooth functions of space. The numerical error thus introduced affects the calculations, but it can be cured if refined approximations are considered. The goal of this paper is to describe a procedure to avoid the above issue and to test its efficiency against exact calculations for a simple model system. Therefore, we propose here (i) to employ a representation of the nuclear density in terms of evolving frozen gaussians (FGs)~\cite{Heller_JCP1981_2}, rather than trajectories, following the scheme presented in Ref.~\cite{long_steps_mt} and (ii) to estimate the phase of the nuclear wave function adopting such a FGs picture, based on a simplified form of the semiclassical Herman-Kluk~\cite{Kluk_CP1984, Davis_JCP1986, Kay_CP2006, Miller_MP2002} propagation scheme within the initial value representation (IVR) theory~\cite{Miller_JCP1970, Wang_ARPC2004, Kay_ARPC2005, Miller_JPCA2001}. The model system for non-adiabatic charge transfer of Ref.~\cite{MM} allows for an exact numerical solution of the full quantum mechanical problem, thus providing a benchmark to any approximation that will be considered. Starting from these results, we will compute the exact time-dependent scalar potential, also referred to as time-dependent potential energy surface (TDPES), in a gauge where the vector potential can be set to zero. The effect of the electrons is then fully accounted for by this TDPES, that is adopted to evolve the FGs. Since the electronic part of the problem is solved exactly, the only source of error will be in this semiclassical approximation. It is worth stressing that this procedure does not result in the development of a new algorithm, but is a test of the performance of the FGs approximation of the nuclear motion.
The paper is organized as follows. In Section~\ref{sec: factorization} we briefly recall the factorization formalism and we focus on the analysis of the electron-nuclear coupling (ENC) operator in the electronic evolution equation. Section~\ref{sec: semiclassics} is devoted to a discussion on the apronximations employed in the calculations. Numerical results are presented in Section~\ref{sec: results}, by comparing situations with different non-adiabatic coupling strengths and testing the approximations employed to evaluate the ENC term. Conclusions are stated in Section~\ref{sec: conclusion}.
\section{The exact factorization framework}\label{sec: factorization}
In the absence of an external field, the non-relativistic Hamiltonian $ \hat H = \hat T_n+\hat H_{BO}$ describes a system of interacting nuclei and electrons. Here, $\hat T_n$ denotes the nuclear kinetic energy and $\hat H_{BO}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})=\hat
T_e({\underline{\underline{\bf r}}})+\hat V_{e,n}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})$ is the BO Hamiltonian, containing the electronic kinetic energy $\hat T_e({\underline{\underline{\bf r}}})$ and all interactions $\hat V_{e,n}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}})$. As recently proven~\cite{AMG, AMG2}, the full wave function, $\Psi({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}},t)$, solution of the TDSE
\begin{equation}\label{eqn: tdse}
\hat H\Psi({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}},t)=i\hbar\partial_t\Psi({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}},t),
\end{equation}
can be written as the product
\begin{equation}\label{eqn: factorization}
\Psi({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}},t)=\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)\chi({\underline{\underline{\bf R}}},t),
\end{equation}
of the nuclear wave function, $\chi({\underline{\underline{\bf R}}},t)$, and the electronic wave function, $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$, which parametrically depends of the nuclear
configuration~\cite{hunter, Gross_PTRSA2014}. Throughout the paper the symbols ${\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}}$ indicate the coordinates of the $N_e$ electrons and
$N_n$ nuclei, respectively. Eq.~(\ref{eqn: factorization}) is unique under the partial normalization condition (PNC)
\begin{equation}\label{eqn: pnc}
\int d{\underline{\underline{\bf r}}}\left|\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)\right|^2=1\,\,\,\forall \,{\underline{\underline{\bf R}}},t
\end{equation}
up to within a gauge-like phase transformation. The evolution equations for $\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)$ and $\chi({\underline{\underline{\bf R}}},t)$,
\begin{align}
\left(\hat H_{el}-\epsilon({\underline{\underline{\bf R}}},t)\right)\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t) = i\hbar \partial_t\Phi_{\underline{\underline{\bf R}}}({\underline{\underline{\bf r}}},t)\label{eqn: electronic eqn} \\
\hat H_n\chi({\underline{\underline{\bf R}}},t) = i\hbar \partial_t\chi({\underline{\underline{\bf R}}},t),\label{eqn: nuclear eqn}
\end{align}
are derived by applying Frenkel's action principle~\cite{frenkel,mclachlan} with respect to the two wave functions and are exactly equivalent~\cite{AMG, AMG2} to the TDSE~(\ref{eqn: tdse}). Eqs.~(\ref{eqn: electronic eqn}) and~(\ref{eqn: nuclear eqn}) are obtained by imposing the PNC~\cite{alonsoJCP2013, AMG2013} by means of Lagrange multipliers.
The electronic equation~(\ref{eqn: electronic eqn}) contains the electronic Hamiltonian
\begin{equation}
\hat H_{el}=\hat H_{BO}({\underline{\underline{\bf r}}},{\underline{\underline{\bf R}}}) + \hat U_{en}^{coup}[\Phi_{\underline{\underline{\bf R}}},\chi],
\end{equation}
which is the sum of the BO Hamiltonian and the ENC operator $\hat U_{en}^{coup}[\Phi_{\underline{\underline{\bf R}}},\chi]$,
\begin{align}
\hat U_{en}^{coup}&[\Phi_{\underline{\underline{\bf R}}},\chi]=\sum_{\nu=1}^{N_n}\frac{1}{M_\nu}\left[
\frac{\left[-i\hbar\nabla_\nu-{\bf A}_\nu({\underline{\underline{\bf R}}},t)\right]^2}{2} \right.\label{eqn: enco} \\
& \left.+\left(\frac{-i\hbar\nabla_\nu\chi}{\chi}+{\bf A}_\nu({\underline{\underline{\bf R}}},t)\right)
\left(-i\hbar\nabla_\nu-{\bf A}_{\nu}({\underline{\underline{\bf R}}},t)\right)\right].\nonumber
\end{align}
In Eq.~(\ref{eqn: electronic eqn}), $\epsilon({\underline{\underline{\bf R}}},t)$ is the TDPES, defined as
\begin{equation}\label{eqn: tdpes I}
\epsilon({\underline{\underline{\bf R}}},t)=\left\langle \Phi_{\underline{\underline{\bf R}}}(t)\right|\hat H_{el}-i\hbar\partial_t\left|\Phi_{\underline{\underline{\bf R}}}(t)\right\rangle_{\underline{\underline{\bf r}}}.
\end{equation}
$\hat U_{en}^{coup}$ and $\epsilon({\underline{\underline{\bf R}}},t)$, along with the vector potential ${\bf A}({\underline{\underline{\bf R}}},t)$,
\begin{equation}\label{eqn: vector potential}
{\bf A}({\underline{\underline{\bf R}}},t) = \left\langle \Phi_{\underline{\underline{\bf R}}}(t)\right| \left. -i\hbar\nabla_\nu\Phi_{\underline{\underline{\bf R}}}(t)\right\rangle_{\underline{\underline{\bf r}}},
\end{equation}
mediate the coupling between electrons and nuclei in a formally exact way. Here, the symbol $\langle\,\cdot\,|\,\cdot\,\rangle_{\underline{\underline{\bf r}}}$ stands for an integration over electronic coordinates.
The nuclear evolution is generated by the Hamiltonian
\begin{equation}\label{eqn: nuclear H}
\hat H_n({\underline{\underline{\bf R}}},t) = \sum_{\nu=1}^{N_n} \frac{\left[-i\hbar\nabla_\nu+{\bf A}_\nu({\underline{\underline{\bf R}}},t)\right]^2}{2M_\nu} + \epsilon({\underline{\underline{\bf R}}},t),
\end{equation}
according to the TDSE~(\ref{eqn: nuclear eqn}).
The TDPES and the vector potential are uniquely determined up to within gauge-like transformations~\cite{AMG,AMG2}. The uniqueness can be straightforwardly proven by following the steps of the current density version~\cite{Ghosh-Dhara} of the Runge-Gross theorem~\cite{RGT}. In this paper, as a choice of gauge, we introduce the additional constraint ${\bf A}_{\nu}\left({\underline{\underline{\bf R}}},t\right)=0$ (see Ref.~\cite{long_steps_mt} for a detailed discussion on how this condition can be imposed)~\cite{CI_MAG}.
As discussed in the introduction, we study the properties of the ENC term $-i\hbar\nabla_\nu\chi/\chi$ that in the expression of the operator $\hat U_{en}^{coup}[\Phi_{\underline{\underline{\bf R}}},\chi]$ explicitly depends on the nuclear wave function. This analysis is based on the interest in developing a procedure to approximate it when a classical or semiclassical treatment of the nuclear motion is adopted. For instance, in the classical limit, we have derived~\cite{mqc,long_mqc} its expression in terms of the nuclear momentum. This has been done by writing the nuclear wave function as $\chi({\underline{\underline{\bf R}}},t)=\exp{\left[\frac{i}{\hbar}\mathcal S({\underline{\underline{\bf R}}},t)\right]}$, with $\mathcal S({\underline{\underline{\bf R}}},t)$ a complex function~\cite{VanVleck_PNAS1928}. If we now suppose~\cite{mqc, long_mqc} that this function can be expanded as an asymptotic series in powers of $\hbar$, namely $\mathcal S({\underline{\underline{\bf R}}},t)=\sum_{\alpha}\hbar^{\alpha}S_{\alpha}({\underline{\underline{\bf R}}},t)$, the ENC term becomes
\begin{equation}\label{eqn: zeroth order of nabla chi / chi}
\frac{-i\hbar\nabla_\nu\chi({\underline{\underline{\bf R}}},t)}{\chi({\underline{\underline{\bf R}}},t)} = \nabla_\nu S_0({\underline{\underline{\bf R}}},t),
\end{equation}
at the lowest-order in $\hbar$. On the right-hand-side (RHS), the function $\nabla_\nu S_0({\underline{\underline{\bf R}}},t)$ is the classical nuclear momentum evaluated along the trajectory, since~\cite{mqc,long_mqc} $S_0$ satisfies a Hamilton-Jacobi equation with Hamiltonian
\begin{equation}\label{eqn: classical hamiltonian}
H_n = \sum_{\nu=1}^{N_n}\frac{\left[\nabla_\nu S_0({\underline{\underline{\bf R}}},t)+{\bf A}_\nu({\underline{\underline{\bf R}}},t)\right]^2}{2M_\nu}+ \epsilon({\underline{\underline{\bf R}}},t).
\end{equation}
The vector potential appears in the above expression of the classical Hamiltonian because this result has general validity, not only in the gauge adopted in the following calculations.
Alternatively, if the nuclear wave function is written in terms of its modulus and phase, $\chi=|\chi|e^{iS/\hbar}$, the (exact) expression of the ENC term becomes
\begin{align}\label{eqn: polar nabla chi over chi}
\frac{-i\hbar\nabla_\nu\chi({\underline{\underline{\bf R}}},t)}{\chi({\underline{\underline{\bf R}}},t)} = \nabla_\nu S({\underline{\underline{\bf R}}},t) +i\frac{-\hbar\nabla_\nu\left|\chi({\underline{\underline{\bf R}}},t)\right|}
{\left|\chi({\underline{\underline{\bf R}}},t)\right|}.
\end{align}
It is clear at this point that a good estimate of the ENC term is only possible when both the modulus and the phase of the nuclear wave function are correctly described. A classical treatment, as in Eq.~(\ref{eqn: zeroth order of nabla chi / chi}), only provides an approximation to the real part of the ENC term, while the information about the imaginary part is lost~\footnote{To be precise, the nuclear density is approximated, within the classical treatment, as $\delta$-function, centered at all times at the classical trajectory. But this contribution is totally omitted in the expression of the ENC term.}.
In the following, we will introduce a FG-based approach to determine an approximation to Eq.~(\ref{eqn: polar nabla chi over chi}). FGs~\cite{Heller_JCP1981_2} evolving on the exact TDPES are used to reconstruct the nuclear density, thus allowing to calculate the second term on the RHS, as $|\chi|$ is a smooth function of the nuclear coordinates. The phase information is instead encoded in the classical action accumulated over time and associated to each FG, as we will show in Section~\ref{sec: semiclassics}.
Henceforth, we will drop the bold-double underlined notation for electronic and nuclear positions as we will deal with one-dimensional (1D) quantities.
\section{Semiclassical approximation}\label{sec: semiclassics}
\subsection{Nuclear density}\label{sec: nuclear density}
According to the procedure presented in Ref.~\cite{long_steps_mt}, a set of independent classical trajectories evolving on the exact TDPES are able to reproduce the nuclear density in almost perfect agreement with quantum results. As pointed out in the Introduction, however, constructing a histogram from the distribution of the trajectories does not allow to compute the second term on the RHS of Eq.~(\ref{eqn: polar nabla chi over chi}), that involves the gradient of $|\chi|$, without a large numerical error. The reason is that the ``classical'' density is not a smooth function of space. The solution proposed here is to improve the previous approximation of nuclear dynamics, by propagating the mean positions and momenta of a set of FGs on the exact TDPES, rather than classical trajectories.
Given a set of $N_{traj}$ initial positions and momenta, $R_0,P_0$, sampled as described in Section~\ref{sec: results}, complex gaussians, also referred to as \textsl{coherent states}, are constructed as
\begin{align}\label{eqn: coherent states}
g\left(R;R_{l,0}P_{l,0},\gamma\right) =
\left(\frac{\gamma}{\pi}\right)^{\frac{1}{4}}e^{-\frac{\gamma}{2}\left(R-R_{l,0}\right)^2+\frac{i}{\hbar}P_{l,0}(R-R_{l,0})},
\end{align}
with width $\gamma$ to be determined below. Each FG is also associated a ``weight'',
\begin{align}\label{eqn: weight}
w_l = \int dRg^*\left(R;R_{l,0}P_{l,0},\gamma\right)\chi_0(R),
\end{align}
corresponding to the projection of the initial nuclear wave function $\chi_0(R)$ on the FGs. The nuclear density at each time is then obtained as
\begin{align}\label{eqn: density as sum of gaussians}
\left|\chi(R,t)\right|^2 \simeq \sum_{l=1}^{N_{traj}} |w_l|^2\left|g\left(R;R_{l}(t)P_{l}(t),\gamma\right)\right|^2,
\end{align}
where $R_l(t),P_l(t)$ are the time-evolved positions and momenta of $R_{l,0},P_{l,0}$. In comparison to the purely classical approximation, Eq.~(\ref{eqn: density as sum of gaussians}) allows not only to reproduce the nuclear density in very good agreement with quantum results, as will be shown in Section~\ref{sec: results}, but also to calculate (analytically) the gradient of the nuclear density (or of the modulus, as it appears in Eq.~(\ref{eqn: polar nabla chi over chi})).
It is important to notice that Eq.~(\ref{eqn: density as sum of gaussians}) is an approximation to the nuclear density when the nuclear wave function is represented as a superposition of coherent states. In fact, coherent states form an overcomplete basis and, in writing Eq.~(\ref{eqn: density as sum of gaussians}), we neglect the overlaps of coherent states. The reason for this further approximation is related to the choice of the initial set of positions and momenta, the mean positions and mean momenta of the FGs. On one hand, we want to maintain the same choice of initial conditions done for the classical propagation (see Section~\ref{sec: results}), in order to be able to directly compare classical and FG results. On the other hand, we want to obtain an initial nuclear density as close as possible to the exact density. We have thus computed the root mean square deviation (RMSD) for the nuclear density when either (1) Eq.~(\ref{eqn: density as sum of gaussians}) is employed or (2) the full expression is considered, i.e. with overlap terms. The best agreement achieved in case (1) is for $\gamma=$ 7.0~a$_0^{-2}$ (used in Section~\ref{sec: results}), with RMSD = 0.020, 0.017, 0.006 for $N_{traj}=$ 2000, 5000, 10000, respectively, whereas in case (2) is for $\gamma=$ 3.0~a$_0^{-2}$ with RMSD = 0.037, 0.024 0.0013. It is evident that a better agreement at the initial time is obtained for case (1), namely when the approximation in Eq.~(\ref{eqn: density as sum of gaussians}) is used, for all the values of $N_{traj}$. Once the parameters defining the coherent states are selected, namely $R_l, P_l,\gamma$, the weights $w_l$ associated to each FG are automatically determined by Eq.~(\ref{eqn: weight}) and kept constant throughout the propagation.
\subsection{Nuclear phase}
The phase of the nuclear wave function will be determined according to
\begin{align}\label{eqn: semiclassical phase}
S(R,t)\simeq\arctan\frac{\Im \left[\chi_{SC}(R,t)\right]}{\Re \left[\chi_{SC}(R,t)\right]},
\end{align}
where $\chi_{SC}(R,t)$ is a semiclassical approximation to the exact $\chi(R,t)$. Following the Herman-Kluk procedure~\cite{Kluk_CP1984, Davis_JCP1986, Kay_CP2006, Miller_MP2002} to approximate the quantum propagator, the expression of the time-evolved wave function at time $t$ is
\begin{align}\label{eqn: semiclassical propagator}
\chi_{SC}(R,t)= \int \frac{dR_0dP_0}{2\pi\hbar}C_t&(R_0,P_0) e^{\frac{i}{\hbar}S(R_0,P_0;t)}\nonumber \\
&\left\langle R|R_tP_t,\gamma\right\rangle\left\langle R_0P_0,\gamma|\chi_0\right\rangle.
\end{align}
Here, $\left\langle R_0P_0,\gamma|\chi_0\right\rangle$ denotes the projection of the initial nuclear wave function on the coherent states, similarly to Eq.~(\ref{eqn: weight}), while $\left\langle R|R_tP_t,\gamma\right\rangle$ is an alternative expression for the coherent states (see Eq.~(\ref{eqn: coherent states})). $R_t,P_t$ are the (classically) evolved positions and momenta corresponding to the initial conditions $R_0,P_0$. $S(R_0,P_0;t)$ is the classical action accumulated up to time $t$ along the trajectory whose initial conditions are $R_0,P_0$. The Herman-Kluk pre-factor is indicated here with the symbol $C_t(R_0,P_0)$~\cite{Kluk_CP1984, Davis_JCP1986, Kay_CP2006, Miller_MP2002}, but it will be set equal to unity throughout the calculations. Therefore, the symbol $\chi_{SC}(R,t)$ will be replaced by $\chi_{FG}(R,t)$, since we will use a FGs approximation rather than a rigorous semiclassical approximation.
In the procedure employed here, the semiclassical nuclear wave function is estimated as a sum over trajectories of time-evolved coherent states, namely
\begin{align}
\chi_{FG}(R,t) = \sum_{l=1}^{N_{traj}} w_l\frac{e^{\frac{i}{\hbar}S_l}}{2\pi\hbar}g(R;R_l(t),P_l(t),\gamma),
\end{align}
where the $l$-th classical action is calculated as
\begin{align}
S_l = \int_0^t d\tau \left(\frac{P_l^2(\tau)}{2M}-\epsilon(R_l(\tau),\tau)\right).
\end{align}
Here, the (exact) TDPES is evaluated, at time $\tau$, at the classical position $R_l(\tau)$ and its expression is obtained by solving the TDSE for the full wave function $\Psi$ and by directly calculating Eq.~(\ref{eqn: tdpes I}) once the factorization~(\ref{eqn: factorization}) is applied.
\section{Numerical results}\label{sec: results}
The expression of the TDPES according to Eq.~(\ref{eqn: tdpes I}) is determined by calculating the electronic wave function $\Phi_R(r,t)$ from the
full wave function $\Psi(r,R,t)$, which is known at all times by solving the TDSE~(\ref{eqn: tdse}) for the model Hamiltonian
\begin{align}
\hat{H}(r,R)= -\frac{1}{2}\frac{\partial^2}{\partial r^2}-\frac{1}{2M}\frac{\partial^2}{\partial R^2} +
\frac{1}{\left|\frac{L}{2}-R\right|}+\frac{1}{\left|\frac{L}{2} + R\right|}\label{eqn: metiu-hamiltonian}\\
-\frac{\mathrm{erf}\left(\frac{\left|R-r\right|}{R_f}\right)}{\left|R - r\right|}
-\frac{\mathrm{erf}\left(\frac{\left|r-\frac{L}{2}\right|}{R_r}\right)}{\left|r-\frac{L}{2}\right|}
-\frac{\mathrm{erf}\left(\frac{\left|r+\frac{L}{2}\right|}{R_l}\right)}{\left|r+\frac{L}{2}\right|}.\nonumber
\end{align}
This system has been introduce by Shin and Metiu~\cite{MM} as a prototype for non-adiabatic charge transfer. The system is 1D and consists of three
ions and a single electron, as depicted in Fig.~\ref{fig: metiu model}.
\begin{figure}[h!]
\centering
\includegraphics*[width=.35\textwidth]{Figure1.pdf}
\caption{Schematic representation of the model system described by the Hamiltonian~(\ref{eqn: metiu-hamiltonian}).}
\label{fig: metiu model}
\end{figure}
Two ions are fixed at a distance of $L=19.0$~$a_0$, the third ion and the electron are free to move in 1D along the line joining the two fixed ions. Here, the symbols $r$ and $R$ are the coordinates of the electron and the movable ion measured from the center of the two fixed ions. The ionic mass is chosen as $M=1836$, the proton mass, whereas the other parameters are tuned in order to make the system essentially a two-electronic-state model. We present here the results obtained by choosing two different sets of parameters, producing strong and weak non-adiabatic couplings, between the first, $\epsilon^{(1)}_{BO}$, and the second BOPES, $\epsilon^{(2)}_{BO}$, around the avoided crossing at $R_{ac}=-1.90$~a$_0$. The values of the parameters in the Hamiltonian~(\ref{eqn: metiu-hamiltonian}) are: $R_f=5.0$~a$_0$, $R_l=3.1$~a$_0$ and $R_r=4.0$~a$_0$, for the strong coupling case; $R_f=3.8$~a$_0$, $R_l=2.9$~a$_0$ and $R_r=5.5$~a$_0$, for the weak coupling case. The BO surfaces are shown in Fig.~\ref{fig: BO} (upper panels).
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=.4\textwidth]{Figure2.pdf}
\caption{Upper panels: Lowest four BO surfaces, as functions of the nuclear coordinate, for strong (left) and weak (right) non-adiabatic coupling strengths. The first (red line) and second (green line) surfaces correspond to the adiabatic states that are populated during the dynamics, whereas the third and fourth surfaces (dashed black lines) are shown for reference. The squared modulus (reduced by ten times and rigidly shifted in order to superimpose it on the energy curves) of the initial nuclear wave packet is also shown (thin black line). Lower panels: Populations of the two lowest adiabatic states ($\rho_1$ and $\rho_2$) as functions of time. The arrows represent the time-steps shown in the following figures.}
\label{fig: BO}
\end{center}
\end{figure}
We study the time evolution of this system by choosing the initial wave function as the product of a real-valued normalized Gaussian wave packet, centered at $R_c=-4.0$~$a_0$ with variance $\sigma=1/\sqrt{2.85}$~$a_0$ (thin black line in Fig.~\ref{fig: BO}, upper panels), and the second BO electronic state. To calculate the TDPES, we first solve the TDSE~(\ref{eqn: tdse}) for the complete system, with Hamiltonian~(\ref{eqn: metiu-hamiltonian}), and obtain the full wave function, $\Psi(r,R,t)$. This is done by numerical integration of the TDSE using the split-operator-technique~\cite{spo}, with time-step of $2.4\times10^{-3}$~fs (or $0.1$~a.u.).
The mean positions and momenta of the FGs evolve along classical trajectories, generated according to Hamilton's equations
\begin{align}
\dot R_l(t) = P_l(t)/M;\,\,\dot P_l(t) = -\partial_R \epsilon\left(R,t\right)\big|_{R_l(t)},
\label{eqn: rp dot}
\end{align}
integrated by using the velocity-Verlet algorithm with a time-step of $12.0\times10^{-3}$~fs (or $0.5$~a.u.). The initial conditions are sampled from the Wigner phase-space distribution corresponding to the initial nuclear wave function. The initial coherent states used in Eq.~(\ref{eqn: weight}) are thus constructed to determine the (complex-valued) weights $w_l$. Numerical results have been obtained for sets of $N_{traj}=2000, 5000, 10000$ trajectories, with the value~\footnote{The value of $\gamma$ is chosen to have the best overlap between the initial semiclassical density and the exact density, in order to impose initial conditions as close as possible to the exact ones.} $\gamma=7.0$~a$_0^{-2}$ for the width of the FGs. The coherent states used to represent the nuclear wave function form an overcomplete basis, therefore the sum in Eq.~(\ref{eqn: density as sum of gaussians}) has to be truncated. In order to choose the adequate number of basis functions to include in the sum, energy conservation has been tested and confirmed for all values of $N_{traj}$. The results will be presented only for the case $N_{traj}=5000$. We have computed the RMSD between the exact nuclear density and its approximation in Eq.~(\ref{eqn: density as sum of gaussians}) for a set of values of $\gamma$ and we have chosen the value of this parameter for which the RMSD is minimum (see also the discussion at the end of Section~\ref{sec: nuclear density}), but clearly other values can be selected.
We will confirm below that the semiclassical FG scheme, adapted to the factorization approach, is an accurate and efficient way to approximate the nuclear wave function. Before presenting the FG results, let us first show different snapshots taken along the dynamics, showing the TDPES. Its feature have been extensively discussed in previous work~\cite{steps,long_steps,long_steps_mt}, but for the sake of completeness, we report here a few configurations. Moreover, we would like to underline that here results for different non-adiabatic coupling strengths will be presented, in order to test the
efficiency of the method also in the weak coupling regime.
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=.4\textwidth]{Figure3.pdf}
\caption{TDPES at different time-steps, as indicated by the arrows in Fig.~\ref{fig: BO}, for strong (left) and weak (right) non-adiabatic coupling strengths. The two lowest BO surfaces (black) are plotted for reference. The GI part of the TDPES (red) presents the well-studied~\cite{steps,long_steps} dynamical steps that bridge piecewise adiabatic shapes potential energy surface, whereas the GI part~\cite{long_steps_mt} (cyan) is piecewise constant. The nuclear density (blue) is shown for reference.}
\label{fig: tdpes}
\end{center}
\end{figure}
Fig.~\ref{fig: tdpes} shows the gauge-invariant (GI) and gauge-dependent (GD) components of the TDPES, namely the two terms that can be identified in Eq.~(\ref{eqn: tdpes I}) as $\epsilon_{GI}\left({\underline{\underline{\bf R}}},t\right)=\langle \Phi_{\underline{\underline{\bf R}}}(t)|\hat H_{el}|\Phi_{\underline{\underline{\bf R}}}(t)\rangle_{\underline{\underline{\bf r}}}$ and $\epsilon_{GD}\left({\underline{\underline{\bf R}}},t\right)=\langle \Phi_{\underline{\underline{\bf R}}}(t)|-i\hbar\partial_t|\Phi_{\underline{\underline{\bf R}}}(t)\rangle_{\underline{\underline{\bf r}}}$. The snapshots shown in Fig.~\ref{fig: tdpes} are taken along the evolution at the times indicated by the arrows in Fig.~\ref{fig: BO} (lower panels) and the two lowest adiabatic surfaces are shown for reference. As previously discussed~\cite{steps,long_steps,long_steps_mt}, before the splitting of the nuclear wave packet, the GI part of the TDPES is diabatic, whereas the GD part is constant, and after the splitting the GI component develops steps that bridge between different adiabatic surfaces, whereas the GD part is piecewise constant. The TDPES is only calculated in the regions where the nuclear density is (numerically) not zero. The lack of reliable information beyond the regions shown in the figures is not an issue when classical trajectories or FGs are employed to mimic the nuclear density, as the regions where the exact density is exponentially small are not, or are poorly, sampled.
It is evident from Fig.~\ref{fig: tdpes} that we will discuss results for short dynamics, limited to the first half of the oscillation period of the nuclear wave packet in the potential well. Interesting dynamics may arise at later times, when for instance the nuclear wave packet crosses a second time the non-adiabatic coupling region. However, here we focus on the initial non-adiabatic event and test how the FG approximation capture this process. Due to the fact that the analysis reported below is the first attempt to incorporate semiclassically nuclear quantum effects in the exact factorization formalism, we study a simple situation that nonetheless captures the main features of a non-adiabatic event. Also, situations where, for instance, reproducing tunneling dynamics might represent a problem for classically evolving FGs~\cite{:/content/aip/journal/jcp/129/2/10.1063/1.2949818,ankerhold_book,AnkerholdPRA2003,KayPRA2013} will be the subject of further study and not addressed here.
A first set of results is shown in Fig.~\ref{fig: nwf2}, where we compare the nuclear density, $|\chi(R,t)|^2$, from three different calculations: quantum (cyan), employing the full electron-nuclear wave function; classical (green), where the histogram is constructed from the distribution of classical trajectories evolving on the TDPES according to Eqs.~(\ref{eqn: rp dot}) (as in Ref.~\cite{long_steps_mt}); FG (red), with the nuclear density given in Eq.~(\ref{eqn: density as sum of gaussians}).
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=.4\textwidth]{Figure4.pdf}
\caption{Nuclear density for strong (left) and weak (right) non-adiabatic coupling strengths. Exact results (cyan) are compared with the
semiclassical density (red), expressed as sum of FGs, and with the histogram (green) constructed from the distribution of classical positions
along the trajectories.}
\label{fig: nwf2}
\end{center}
\end{figure}
As expected from previous calculations~\cite{long_steps_mt}, the use of classical trajectories seems to be enough accurate to reproduce the nuclear density. It is however important to stress again, that these results are not obtained by solving a fully approximate form of the coupled electronic and nuclear equations~(\ref{eqn: electronic eqn}) and~(\ref{eqn: nuclear eqn}). They only represent a benchmark for any quantum-classical algorithm, since the effect of the electrons, via the TDPES, is treated exactly. The semiclassical density, constructed as the weighted sum of FGs given in Eq.~(\ref{eqn: density as sum of gaussians}), is also accurate. In comparison to the classical histogram, the gain here is the smoothness of the density, not achievable with purely classical trajectories. This feature is extremely important for the calculation of the ENC term containing the gradient of the nuclear density, via the term $-i\hbar\nabla_\nu|\chi|/|\chi|$ in Eq.~(\ref{eqn: polar nabla chi over chi}).
The second important characteristic of the semiclassical approach is that each FG contributes a phase factor to the full nuclear wave function, thus allowing to determine also the first term on the RHS of Eq.~(\ref{eqn: polar nabla chi over chi}). We show this term, i.e. $\nabla_\nu S(R,t)$, in Fig.~\ref{fig: Re ENC}, comparing once again exact results with the corresponding FG and classical approximations. The semiclassical value of $\nabla_\nu S(R,t)$ is determined as the gradient of the phase in Eq.~(\ref{eqn: semiclassical phase}), whereas the function $\nabla_\nu S(R,t)$ is classically interpreted as the nuclear momentum evaluated along each trajectory. While the agreement between quantum and FG results is remarkable, the phase-space points corresponding to the classical trajectories do not allow to reconstruct a smooth function of $R$. Even if the number of classical trajectories is increased, the phase-space points are too ``noisy'' to allow for reconstructing a smooth function. A smoothing algorithm should then be employed, but the numerical efficiency of the whole procedure might become questionable.
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=270,width=.4\textwidth]{Figure5.pdf}
\caption{Real part of the ENC term for strong (left) and weak (right) non-adiabatic coupling strengths. The nuclear density (black) is shown for reference. Exact results (cyan) are compared with semiclassical calculations (red) and with the classical phase-space points (green).}
\label{fig: Re ENC}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[angle=270,width=.4\textwidth]{Figure6.pdf}
\caption{Imaginary part of the ENC term for strong (left) and weak (right) non-adiabatic coupling strengths. The color code is the same as in Fig.~\ref{fig: Re ENC}.}
\label{fig: Im ENC}
\end{center}
\end{figure}
Fig.~\ref{fig: Im ENC} shows the imaginary part of the ENC term from Eq.~(\ref{eqn: polar nabla chi over chi}) at different time-steps during the dynamics, as in previous figures. The exact results (cyan) are compared only with the approximation (red) based on the semiclassical propagation of FGs. Even if we employ a simplified form of the Herman-Kluk propagator, where the pre-factor is set to 1, semiclassical results are in satisfactory good agreement with exact results.
One could consider improving the approach developed here by, for instance, explicitly computing the Herman-Kluk pre-factor, at the expenses of increasing the computational cost.
\section{Conclusion}\label{sec: conclusion}
We have reported our first semiclassical procedure adapted to the formalism of the exact factorization of the electron-nuclear wave function~\cite{AMG,AMG2}. The approach has been used to estimate the ENC term that explicitly depends on the nuclear wave function (gradient of its modulus and phase) in the electronic equation. In previous work~\cite{mqc,long_mqc} on the development of a mixed quantum-classical algorithm in the context of the exact factorization, such term has been treated fully classically and identified as the nuclear momentum. However, we observed that, despite the fact that this approximation is widely used in the literature~\cite{TSH_1990, truhlarFD2004, ciccottiJCP2000}, correction terms naturally arise in the factorization framework. The extent and role of the corrections may depend on the process that we intend to study, therefore we have shown in the present paper how to evaluate such corrections based on the semiclassical propagation of FGs, for a simple model case of non-adiabatic charge transfer. The main gain in using FGs, rather than a purely classical approach, lies in the possibility of obtaining smooth functions whose gradients can be easily determined without introducing large numerical errors.
The semiclassical results shown in the paper have been obtained by evolving FGs on the exact TDPES, that is known for the simple model system studied here as the outcome of exact calculations based on the numerical solution of the full TDSE. This procedure is a test, since the only approximation is the semiclassical treatment of the nuclear dynamics, whereas the electrons are treated exactly, via the information encoded in the TDPES. Moreover, this study provides a benchmark for future development, aiming at improving the initial, and lowest-order, mixed quantum-classical algorithm derived from the factorization~\cite{mqc,long_mqc}. The procedure described here can be easily implemented in an algorithm, as we will present elsewhere~\cite{MQC_min}, resulting in a novel mixed quantum-semiclassical scheme for solving coupled electron-nuclear dynamics.
\section*{Acknowledgements}
The authors would like to thank Neepa T. Maitra for her help in improving the presentation of the results. Partial support from the Deutsche Forschungsgemeinschaft (SFB 762) and from the European Commission (FP7-NMP-CRONOS) is gratefully acknowledged.
|
\section{Introduction}
Given a graph $G$, one of the most basic tasks one would like to perform
on $G$ is to find a spanning tree of it or perhaps some other sparse spanning subgraph $G'$.
This task can be easily accomplished using numerous well-known algorithms such as DFS (depth-first search), BFS (breadth-first search)
and more. What all of these algorithms have in common is that in order to
decide whether a given edge $e$ belongs to the spanning subgraph $G'$, one has to construct the entire spanning tree. Suppose however that one is not interested
in constructing the entire spanning subgraph $G'$, but rather to be able to ``quickly''
tell if a given edge $e$ belongs to $G'$ or not.
By ``quickly'' we mean using a {\em constant\/} number of operations.
Such algorithms are of importance in distributed settings, where processors reside on the
vertices of the graph and the goal is to select as few communication links (edges) as possible
while maintaining connectivity. Another relevant setting is one in which the graph resides in a centralized database, but different,
uncoordinated, servers have access to it, and are interested in only parts of a common sparse spanning subgraph.
In both cases we would like the decision regarding any given edge to be made after inspecting
only a very small portion of the whole graph, but all decisions must be consistent with the same spanning
subgraph. Such algorithms belong to a growing family of local algorithms
for solving classical problems in graph theory. We elaborate on relevant related works
in Subsection~\ref{subsec:related}.
Let us make a simple observation regarding the task of locally constructing a spanning subgraph.
Note that if one insists on locally constructing a spanning {\em tree} $G'$, then it is easy to see that the
task cannot be performed in general without inspecting
almost all of $G$; that is, this task cannot be achieved using a {\em constant} number of queries to $G$.
To see this, observe that if $G$ consists of a single path, then the algorithm must answer positively on all edges, while if $G$ consists of a single cycle then the
algorithm must answer negatively on one edge. However, the
two cases cannot be distinguished without inspecting a
linear number of edges.
So suppose we allow the
algorithm some slackness, and rather than requiring
that $G'$ be a tree, only require that it be relatively sparse,
i.e., contain at most $(1+\eps)n$ edges. Summarizing, the question is then, given $\epsilon > 0$,
for which graphs $G$ can we locally construct a spanning subgraph $G'$ consisting
of $(1+\epsilon)n$ edges, such that given an edge $e \in E(G)$ one can
determine if $e \in G'$ using a constant (that may depend on $\epsilon$ but not on $n$) number
of queries to $G$?
Our main result in this paper, stated informally as Theorem~\ref{theo:informal} below,
shows that the answer to the above question is given by a certain variant of graph expansion, which
we now turn to define. For a graph $G$ and a subset $S \sub V(G)$, we write $\partial_G(S)$
for the set of edges of $G$ with precisely one endpoint in $S$.
We write $\phi_G$ for the (edge) \emph{expansion} of $G$, that is, $\phi_G = \min_S \s{\partial_G(S)}/\s{S}$ where
the minimum is taken over all $S \subseteq V(G)$ of size $1 \leq |S| \leq |V(G)|/2$.
Note that a graph may have small expansion yet contain (large) subgraphs with large expansion.
For example, a vertex-disjoint union of cliques has expansion $0$, yet it contains complete graphs that have the largest expansion possible (for graphs of their order). Let us thus say that
a graph is \emph{$f$-non-expanding} if every $t$-vertex subgraph $H$ satisfies $\phi_H \le f(t)$ (we assume $t > 2 $).
Our main result in this paper can be informally stated as follows.
\BT[{\bf Informal Statement}]\label{theo:informal}
We have the following dichotomy:
\begin{itemize}
\item If $G$ is $f$-non-expanding for $f(t) \ll 1/\log t$ then one can
locally construct a sparse spanning subgraph of $G$. The algorithm is deterministic.
\item There is a family of $3$-regular graphs $G_n$ that are (roughly) $1/\log t$-non-expanding
so that every (possibly randomized) local algorithm for constructing a sparse spanning subgraph
of $G_n$ must accept every edge of $G_n$.
\end{itemize}
\ET
We refer the reader to Definition~\ref{dfn:SSG-alg} for the precise definition of what it means to locally construct
a sparse spanning subgraph, and to Theorems~\ref{cor:f-non-expand-ub} and~\ref{theo:main} for the precise statements of
the two assertions in Theorem~\ref{theo:informal}.
We note that there are numerous families of graphs that satisfy the condition in the first
item of Theorem~\ref{theo:informal}. For example, it follows from the planar separator theorem of Lipton and Tarjan~\cite{LT79} and its extension by Alon, Seymour and Thomas~\cite{AST90} that planar graphs (and more generally, $H$-minor-free graphs) of bounded degree satisfy the condition of the first item.
Also, observe that since the graphs $G_n$ in the second item of Theorem~\ref{theo:informal} have $3n/2$ edges,
there is no algorithm that can locally construct a spanning subgraph of $G_n$ with $(1+\epsilon)n$ edges
for $\epsilon < 1/2$.
We make two comments regarding the results which appeared in the preliminary conference version of this paper~\cite{LRR14}.
First, it was shown in~\cite{LRR14} that
there are graphs such that any algorithm has to inspect
$\Omega(\sqrt{n})$ edges in order to decide whether a given edge belongs to a spanning subgraph $G'$
containing $(1+\eps)n$ edges, for a constant $\eps$.
However, those graphs resulted from random graphs, which have expansion $\Theta(1)$,
and so could not be used in order to show that the non-expansion requirement
given in the first item of Theorem~\ref{theo:main} cannot be relaxed.
Second, it was shown in~\cite{LRR14} that for certain families of graphs, one can solve
the sparse spanning subgraph problem in time $O(\sqrt{n})$. It is an interesting open problem
to decide whether this can be extended to hold for all bounded-degree graphs.
In fact, it would even be interesting to show that for any bounded-degree graph $G$,
one can find a sparse spanning subgraph using $o(n)$ queries\footnote{Note that if we are allowed to make $\Theta(n)$ queries, then
we can just use the standard BFS or DFS algorithms, which find the entire spanning tree of $G$.}.
\subsection{Related work}\label{subsec:related}
As is evident from the above description of the problem,
the model we study here is similar to both classical models, such as distributed and parallel computation,
and to more recent ones. In what follows, we describe these models and some related results,
so as to provide a broad context for our work.
\subsubsection{Local algorithms for other graph problems}\label{subsec:local}
The model of {\em local computation algorithms} as considered in
this work, was defined by Rubinfeld et al.~\cite{RTVX} (see
also Alon et al.~\cite{ARVX12}). Such algorithms for maximal
independent set, hypergraph coloring, $k$-CNF and maximum
matching are given in~\cite{RTVX,ARVX12,MRVX12,MV13}.
This model generalizes other models that have been studied
in various contexts, including locally decodable codes
(e.g.,~\cite{STV99}), local decompression~\cite{DLRR13},
and local filters/reconstructors
\cite{ACC+08,SS10,B08,KPS08,JR11,CS06a}. Local computation
algorithms that give approximate solutions for various
optimization problems on graphs, including vertex cover,
maximal matching, and other packing and covering problems,
can also be derived from sublinear time algorithms for
parameter estimation~\cite{PR,MR,NO,HKNO09,YYI}.
The model of local computation is related to several other models, including property testing and online algorithms.
To give a notable example, Mansour et al.~\cite{MRVX12} proposed a general scheme for converting a large family of online algorithms into local computation algorithms, consequently, improving the complexity of hypergraph 2-coloring and $k$-CNF in the local computation model.
In the related field of local reconstructors, Campagna et al.~\cite{CGR13} study the property of connectivity.
Namely, under the promise that the input graph is almost
connected, their reconstructor provides oracle access to
the adjacency matrix of a connected graph which is close to
the input graph. We emphasize that our model is
different from theirs, in that they allow the addition of
new edges to the graph, whereas our algorithms must provide
spanning graphs whose edges are present in the original
input graph.
\subsubsection{Distributed and parallel algorithms}\label{subsec:dist}
The name {\em local algorithms} is also used in the
distributed context~\cite{MNS95,NS95,Lin92}. As observed by
Parnas and Ron~\cite{PR}, local distributed algorithms can
be used to obtain local computation algorithms as defined
in this work, by simply emulating the distributed algorithm
on a sufficiently large subgraph of the graph $G$. However,
while the main complexity measure in the distributed
setting is the number of rounds (where it is usually
assumed that each message is of length $O(\log n)$), our
main complexity measure is the number of queries performed
on the graph $G$. By this standard reduction, the bound on
the number of queries (and hence running time) depends on
the size of the queried subgraph and may grow exponentially
with the number of rounds. Therefore, this reduction gives
meaningful results only when the number of rounds is
significantly smaller than the diameter of the graph.
While the problem of computing a spanning graph has not been studied in the distributed model, the problem of computing a minimum-weight spanning tree is a central one in this model.
Kutten and Peleg~\cite{KP98}
provided an algorithm that works in $O(\sqrt{n} \log^* n +
D)$ rounds, where $D$ denotes the diameter of the graph.
Their result is nearly optimal in terms of the complexity
in $n$, as shown by Peleg and Rubinovich~\cite{PR00} who
provided a lower bound of $\Omega(\sqrt{n}/\log n)$ rounds
(when the length of the messages must be bounded).
Another problem studied in the distributed setting that is
related to the one studied in this paper, is finding a
sparse spanner. The requirement for spanners is much
stronger since the distortion of the distance should be as
small as possible. Thus, to achieve this property, it is
usually the case that the number of edges of the spanner is
super-linear in $n$. Pettie~\cite{Pet10} was the first to
provide a distributed algorithm for finding a low
distortion spanner with $O(n)$ edges without requiring
messages of unbounded length or $O(D)$ rounds. The number
of rounds of his algorithm is $\log^{1+o(1)}n$. Hence, the
standard reduction of~\cite{PR} yields a local algorithm
with a trivial linear bound on the query complexity.
The problems of computing a spanning tree and a minimum
weight spanning tree were studied extensively in the
parallel computing model as well (see, e.g.,~\cite{BC05}, and the
references therein). However, these parallel algorithms
have time complexity which is at least logarithmic in $n$
and therefore do not yield an efficient algorithm in the
local computation model. See~\cite{RTVX,ARVX12} for further
discussion on the relationship between the ability to
construct local computation algorithms and the parallel
complexity of a problem.
\subsubsection{Local cluster algorithms}\label{subsec:clust}
Local algorithms for graph theoretic problems have also
been given for PageRank computations on the web
graph~\cite{JW03,Ber06,SBC+06,ACL06,ABC+08}. Local graph
partitioning algorithms have been presented
in~\cite{ST04,ACL06,AP09,ZLM13,OZ13}, which find subsets of
vertices whose internal connections are significantly
richer than their external connections in time that depends
on the size of the cluster that they output.
For instance, Andersen and Peres~\cite{AP09} provide an algorithm which, given a starting vertex $v$, finds a cluster of $v$ of small conductance,
whose complexity depends on the volume of the cluster it outputs but has only polylogarithmic dependence in the size of the graph.
However, even when the
size of the cluster is guaranteed to be small, it is not
obvious how to use these algorithms in the local
computation setting where the cluster decompositions must
be consistent among queries to all vertices.
\subsubsection{Other related sublinear-time approximation algorithms for graphs}\label{subsec:sublin}
The problem of estimating the weight of a minimum-weight
spanning tree in sublinear time was considered by Chazelle,
Rubinfeld and Trevisan~\cite{CRT}. They describe an
algorithm whose running time depends on the approximation
parameter, the average degree and the range of the weights,
but does not directly depend on the number of vertices.
\subsection{Organization}
The rest of the paper is organized as follows. In Section~\ref{sec:prel}
we formally define the local sparse spanning subgraph problem which we consider in this paper.
In Section~\ref{sec:upper} we prove the first item of Theorem~\ref{theo:informal}, which is formally stated as Theorem~\ref{cor:f-non-expand-ub}. The proof of this part has two main steps. In the first one, we show that if $G$ is $f$-non-expanding with $f \ll 1/\log t$ then one can remove from $G$ only
a relatively small number of
edges and thus partition it into connected components of size $O(1)$ each. We then show that if a graph can be so partitioned, then one can solve on it the local spanning subgraph problem by executing a ``localized'' version of Kruskal's~\cite{Kruskal56} famous algorithm for finding minimum-weight spanning tress\footnote{Recall that if $G$ is a graph with weights assigned to its edges, then Kruskal's algorithm finds a spanning tree of minimal {\em total} weight}.
The proof of the second paper of Theorem~\ref{theo:informal}, which is the more challenging part of this paper, is
given in Section~\ref{sec:lower} and formally stated as Theorem~\ref{theo:main}. It establishes that the $1/\log t$-non-expansion
requirement from the first item of Theorem~\ref{theo:informal} is essentially tight. What we show is that there
are graphs which are (about) $1/\log t$-non-expanding, and have the property that any local algorithm for constructing
a spanning subgraph using a constant number of queries must accept {\em every} edge of the graph. To prove this result we describe a construction
of certain extremal graphs that might be of independent interest. These are $3$-regular graphs, that on one
hand have unbounded \emph{girth}\footnote{As usual, the girth of a graph is the minimum length of a cycle in it.}, but on the other hand are (about) $1/\log t$-non-expanding.
We make no serious attempt to optimize the constants obtained in the various statements. In fact, the $f$-non-expansion requirements in our upper and lower bound results (Theorems~\ref{cor:f-non-expand-ub} and~\ref{theo:main}), which are about $(1/\log t)(1/\log\log t)^2$ and $(1/\log t)(\log\log t)^2$ respectively, can each be improved by replacing the $(\log\log t)^2$ term by $(\log\log t)^{1+o(1)}$.
We opted for proving our results with the slightly weaker bounds in order to simplify the presentation.
We henceforth write $\log(\cdot)$ for $\log_2(\cdot)$.
\section{Preliminaries}\label{sec:prel}
Let us now give the precise definition of the algorithmic problem we are addressing in this paper.
As in most cases where one tries to design a local/distributed/sublinear algorithm, we will assume
that the input graph is given via an {\em oracle access to its incidence-list representation}, meaning the following:
First, we assume that the input graph $G=(V,E)$
is given via incidence-lists representation, that is, for each vertex $v \in V(G)$, there is an
ordered list of its neighbors in $G$. Second, the algorithm is supplied
with integers $n$ and $d$, that represent the number of vertices, and an upper bound on the degrees of vertices of $G$.
Finally, given a pair $(v,i)$ with $1 \leq v \leq n$ and $1 \leq i \leq d$, the oracle either returns the
$i^{th}$ neighbour of $v$ (in the incidence list representation) or an indication that $v$ has less than $i$ neighbours.
We will assume that each vertex $v$ has an id, $id(v)$, where there is a full order over the ids. We will think of the ids of vertices
in the graphs simply as the integers $\{1,\ldots,n\}$.
We now turn to formally define the algorithmic problem we consider in this paper.
\renewcommand{\labelenumi}{\theenumi}
\renewcommand{\theenumi}{\roman{enumi}.}
\BD
\label{dfn:SSG-alg}
An algorithm $\calA$ is an {\em $(\e,q)$-local sparse spanning graph algorithm} if, given $n,d \ge 1$ and oracle access to the incidence-lists representation of a connected graph $G=(V,E)$ on $n$ vertices and degree at most $d$,
it provides
query access to a subgraph $G'=(V, E')$ of $G$ such that:
\BE
\item $G'$ is connected.
\item $\s{E'} < (1+\eps)\cdot n$ with probability at least
$2/3$ (over the
internal randomness of $\mathcal{A}$).
\item $E'$ is determined by $G$ and the internal randomness
of $\mathcal{A}$.
\item $\calA$ makes at most $q$ queries to $G$.
\EE
By ``providing query access to $G'$'' we mean that
on input $(u, v)\in E$,
$\calA$ returns whether $(u, v) \in E'$ and for any
sequence of queries, $\calA$ answers consistently with
the same $G'$.
An algorithm $\calA$ is an {\em $(\e,q)$-local sparse spanning graph
algorithm for a family of graphs $\mathcal{C}$}
if the above conditions hold, provided that the input graph
$G$ belongs to $\mathcal{C}$.
\ED
\renewcommand{\labelenumi}{\theenumi}
\renewcommand{\theenumi}{\arabic{enumi}.}
We note that the choice of the required success probability being $2/3$ is of course arbitrary
and can be replaced by any probability smaller than $1$. Having said this, let us stress that the positive results
we obtain here (i.e., the algorithmic results) in Theorem~\ref{cor:f-non-expand-ub} are deterministic
(i.e., hold with probability $1$), whereas our lower bound in Theorem~\ref{theo:main} holds for \emph{any} positive success probability.
We also note that even though Definition~\ref{dfn:SSG-alg} considers only the number of queries performed by the algorithm, our algorithm in Theorem~\ref{cor:f-non-expand-ub} runs in time polynomial in the number of queries, and in particular, independent of $n$.
We are interested in
local algorithms that have query complexity which is
independent of $n$, namely, that perform a constant number
of queries to the graph (for each edge they are queried on)
and whose running time (per queried edge) is small as well.
In the next section, we show that the family of graphs that are
$f$-non-expanding with $f \ll 1/\log t$ have a local sparse spanning graph
algorithm. In the following section, we will show that one cannot prove
such a result when $f$ is only slightly larger.
\section{Upper bound}\label{sec:upper}
In this section we prove the following theorem, which formalizes the first assertion of Theorem~\ref{theo:informal}.
\BT\label{cor:f-non-expand-ub}
For every $C$ there is a function $q:\R_+\times\N\to\N$ so that for every $\epsilon >0$ there is an $(\e,q(\e,d))$-local
sparse spanning graph algorithm for the family of $f$-non-expanding graphs with
\begin{equation}\label{eq:f}
f(x) = \frac{C}{\log x\cdot (\log\log x)^2} \;,
\end{equation}
where $d$ is the input degree-bound. Furthermore, the algorithm is deterministic.
\ET
\subsection{Decomposition of non-expanding graphs}
The first step in the proof of Theorem~\ref{cor:f-non-expand-ub} is a decomposition result stated in Lemma~\ref{lemma:CSS} below.
It shows that if $G$ is $f$-non-expanding, with $f$ as in Equation~(\ref{eq:f}),
then $G$ can be decomposed into connected components of bounded size by removing only $\e n$ edges.
This extends a result of~\cite{CSS09} that applies for somewhat larger $f$.
As mentioned earlier, there are many families of graphs which are $f$-non expanding with $f$ as in Equation~(\ref{eq:f}).
For example, planar graphs of bounded degree are $f$-non-expanding with $f=O(1/\sqrt{x})$ by the famous planar separator theorem of Lipton and Tarjan~\cite{LT79}. More generally, a result of Alon, Seymour and Thomas~\cite{AST90} implies that
for any fixed $H$, the family of $H$-minor-free graphs of bounded degree is $f$-non-expanding with $f=O(1/\sqrt{x})$.
Hence, Lemma~\ref{lemma:CSS} applies to these families of graphs in particular.
We note that the reason why the bound in Lemma~\ref{lemma:CSS} is doubly exponential in $\epsilon$ is that we insist on assuming that $f$ is very close to the threshold of $1/\log x$ (which by Theorem~\ref{theo:main} is essentially tight).
For example, the details of the proof of Lemma~\ref{lemma:CSS} show that if $f=x^{-c}$ for some $0 < c < 0$ (as is the case
with planar graphs, say), then the bound can be improved to polynomial in $1/\epsilon$.
We note that in such cases we can also set $k=\poly(1/\epsilon)$ in step $1$ of our algorithm (Algorithm~\ref{alg:local-kruskal} below), thus obtaining a much
more efficient algorithm.
\BL\label{lemma:CSS}
If $G$ is an $n$-vertex $f$-non-expanding graph with
$f(x) = C/\log x(\log\log x)^2$, then one can remove $\e n$ edges from $G$
so that each connected component of the remaining graph is of size at most
$2^{2^{2(C/\e) + 3}}$.
\EL
\BPF
First, we claim that any $f$-non-expanding $n$-vertex graph $G=(V,E)$ has
a subset $S \subset V(G)$ of size $n/3 \le \s{S} \le (2/3)n$
and expansion $\phi_{G}(S) \eqdef \s{\partial_G(S)}/\s{S} \le f(n/3)$.
For the proof we iteratively construct subsets $S_1,\ldots,S_k \sub V(G)$ as follows.
To obtain $S_i$, we consider the induced subgraph $G_i = G[V \sm \bigcup_{j=1}^{i-1} S_j]$
and let $S_i \sub V(G_i)$ satisfy $\s{S_i} \le n_i/2$ and $\phi_{G_i}(S_i) \le f(n_i)$, where $n_i = \s{V(G_i)}$.
We stop once $S \eqdef \bigcup_{i=1}^k S_i$ is of size $\s{S} \ge n/3$.
Note that
$$\s{S} \le \sum_{i=1}^{k-1} \s{S_i} + n_k/2 = (n+\sum_{i=1}^{k-1} \s{S_i})/2 \le 2n/3 \;.$$
It remains to bound $\phi_{G}(S)$.
Observe that every edge in the edge boundary $\partial_G(S)$ is a member of some edge boundary $\partial_{G_i}(S_i)$. Hence,
$$\frac{\s{\partial_G(S)}}{\s{S}} \le
\frac{\sum_{i=1}^k \s{\partial_{G_i}(S_i)}}{\s{S}} =
\sum_{i=1}^k \frac{\s{S_i}}{\s{S}} \phi_{G_i}(S_i)
\le \max_{1\le i \le k} \phi_{G_i}(S_i) \le \max_{1\le i \le k} f(n_i) = f(n_k) \le f(n/3) \;,$$
where in the last inequality we used the fact that $n_k \ge n-\s{S} \ge n/3$.
This proves our claim.
Fix an integer $k \geq 50$ and let $G=(V,E)$ be any $f$-non-expanding graph on $n \geq k/3$ vertices.
Consider the following process; take any subset $S \subset V$ of size $n/3 \le \s{S} \le n/2$ and expansion $\phi_{G}(S) \le 2f(n/3)$
(whose existence follows from the claim above), remove all its outgoing edges and proceed recursively on the two induced subgraphs $G[S]$ and $G[V \sm S]$, which are $f$-non-expanding as well.
The recursion stops
whenever we reach a graph on at most $k$ vertices.
It is clear that at the end of this process, the edges removed from $G$ leave a graph whose connected components have at most $k$ vertices each. Let $r_k(G)$ be the number of edges
removed by the above process. We will shortly prove that if $G$ has $n$ vertices, then $r_k(G) \leq Cn/\ln\ln(k/3)$. Hence, setting $k=2^{2^{2(C/\e) + 3}} \ge \max\{50, 3 \cdot e^{e^{C/\e}}\}$
enables us to remove no more than $\epsilon n$ edges and break $G$ into connected components of size at most $k$, thus
proving the lemma.
In order to facilitate an inductive proof, it will be more convenient to prove the following slightly stronger claim:
\begin{equation}\label{eq:ind-claim}
r_k(G) \leq \beta(n) \eqdef \frac{Cn}{\ln\ln(k/3)} - \frac{Cn}{\ln\ln n}\;.
\end{equation}
Set $h(x) = x/\ln\ln x$ and $f^*(x) = f(x)/C=(\log x)^{-1}(\log \log x)^{-2}$.
First, we
establish some properties of $h$.
It is easy to verify that the derivative of $h$ is
$h'(x) = (\ln\ln x)^{-1} - (\ln x)^{-1}(\ln\ln x)^{-2}$,
and moreover, $h''(x) \le 0$ for $x \ge 20$.
It follows that for every $n \ge 50$ and $n/3 \le s \le n/2$ we have
\begin{equation}\label{eq:h-bound}
h(n) - h(n-s) \le s \cdot h'(n-s) \le h(s)-s \cdot 2f^*(n/3)
\end{equation}
where in the first inequality we used the concavity of $h$ on the interval $(20,\infty)$,
and in the second inequality we used the fact that $n \geq 50$ and $n/3 \leq s \leq n/2$ and that in this range
$$
(\log(n/3)(\log\log(n/3))^2/2 \geq \ln(2n/3)(\ln\ln(2n/3))^2\geq \ln(n-s)(\ln\ln(n-s))^2\;.
$$
We prove Equation~(\ref{eq:ind-claim}) by induction on $n$. For the base case(s) where $(k/3 \le)$ $n \le k$ we have that $\beta(n) \ge \beta(k/3) = 0 = r_k(G)$, as needed.
For the induction step we have
\begin{align*}
r_k(G) &\le \max_{\substack{S \subset V:\\ n/3 \le \s{S} \le n/2}} \s{S} \cdot 2f(n/3) + r_k(G[S]) + r_k(G[V \sm S]) \\
&\le \max_{n/3 \le s \le n/2} s \cdot 2f(n/3) + \beta(s) + \beta(n-s)\\
&= C\Big(n/\ln\ln (k/3) + \max_{n/3 \le s \le n/2} s \cdot 2f^*(n/3) - h(s) - h(n-s) \Big)\\
&\le C\Big(n/\ln\ln (k/3) - h(n) \Big) = \beta(n)
\end{align*}
where the first inequality follows from the definition of the process described in the second paragraph of the proof,
the second inequality follows from the induction hypothesis since $k/3 \le s,n-s \le n-1$, and in the third inequality we used~(\ref{eq:h-bound}) since $n \ge k \ge 50$.
This completes the proof of Equation~(\ref{eq:ind-claim}).
\EPF
\subsection{The algorithm}
The algorithm we design in order to prove Theorem~\ref{cor:f-non-expand-ub} is based on Kruskal's minimum-weight spanning tree algorithm~\cite{Kruskal56}. The idea is to assign weights to the edges of the graph in a way that will determine some fixed spanning tree $T$. The algorithm will always accept the edges of $T$ but will also accept a few other edges.
We will pick the weights of the edges in a way that will make it possible to determine the edges of a sparse spanning subgraph in a ``local'' fashion.
Recall that Kruskal's algorithm for finding a minimum-weight spanning tree in a
weighted connected graph works as follows. First it sorts the edges of the graph $e_1,\dots e_m$ from minimum to maximum weight (breaking ties arbitrarily).
It then goes over the edges in this order, and adds $e_i$ to the spanning tree if and only if it does not close a cycle with the previously selected edges.
Put differently:
\BF\label{fact:Kruskal}
Edge $e$ is picked by Kruskal's algorithm if and only if for any cycle $C$ of $G$ containing $e$, the edge $e$ does not
have the largest weight among the edges of $C$.
\EF
It is well known (and easy to verify)
that if the weights of the edges are distinct, then there is a single minimum-weight spanning tree in the graph. For an unweighted graph $G$, consider the order
defined over its edges by the order of the ids of the vertices. Namely, we define
a ranking $r$ of the edges as follows:
$r(u,v) < r(u',v')$ if and only if $\min\{id(u),id(v)\} < \min\{id(u'),id(v')\}$
or $\min\{id(u),id(v)\} = \min\{id(u'),id(v')\}$
and $\max\{id(u),id(v)\} < \max\{id(u'),id(v')\}$.
If we run Kruskal's algorithm using the rank $r$ as the weight function (where
there is a single ordering of the edges), then
we obtain a (well-defined) spanning tree of $G$.
While the local algorithm described next (Algorithm~\ref{alg:local-kruskal})
is based on the aforementioned
global algorithm, it does not exactly emulate it, but rather emulates a certain {\em relaxed\/}
version of it which can be executed {\em locally}. In particular, it will answer YES for every edge selected by the global algorithm
(ensuring connectivity), but may answer YES also on edges not selected by the global algorithm.
We will thus need to show that it does not answer YES on too many edges that are not selected
by the global algorithm.
In the description and analysis of the algorithm we will use
the following standard notation; for a vertex
$v \in V$ and an integer $k$, we denote by $C_k(v,G)$
the subgraph of $G$ induced by the set of vertices at distance at most $k$ from $v$.
\medskip
\begin{algorithm}
\caption{{\bf (Kruskal-based Algorithm)}}
(The algorithm works for some fixed $\epsilon>0$.)
\textbf{Input:} $n,d \ge 1$, query access to a graph $G$ on $n$ vertices and degree at most $d$.
\textbf{Queried edge:} $(x,y) \in E(G)$.
\BE
\item Set $k=2^{2^{2(C/\e) + 3}}$.
\item Perform a BFS to depth $k$ from $x$, thus obtaining the subgraph $C_k(x,G)$.
\item If $(x,y)$ is the edge with largest rank on some cycle in $C_k(x,G)$ then answer NO;\\ Otherwise, answer YES.
\EE
\label{alg:local-kruskal}
\end{algorithm}
\BPFOF{Theorem~\ref{cor:f-non-expand-ub}}
We will show that if $G=(V,E)$ is $C/\log x(\log \log x)^2$-non-expanding then Algorithm~\ref{alg:local-kruskal} is an $(\epsilon,q(\epsilon,d))$-local sparse spanning subgraph algorithm, where $q(\epsilon,d)=d^{k+1}$ with $k$ being the constant from step $1$ of the algorithm.
By the description of Algorithm~\ref{alg:local-kruskal} it directly follows that
the algorithm is deterministic and that its answers are consistent with a connected subgraph $G'$.
Indeed, if $T$ is the tree returned by Kruskal's algorithm, then Fact~\ref{fact:Kruskal} and step 3 of Algorithm~\ref{alg:local-kruskal} guarantee that each edge of $T$ will be accepted by Algorithm~\ref{alg:local-kruskal}. Observe that the number of queries to $G$
performed by Algorithm~\ref{alg:local-kruskal} is at most $d^{k+1}$.
We now complete the proof by showing that the algorithm returns YES on fewer than $(1+\epsilon)n$ edges.
Let $R$ be a set of at most $\epsilon n$ edges whose removal disconnects $G$ into connected components of size at most $k$. The existence of such a set is guaranteed by Lemma~\ref{lemma:CSS}.
Let $G_R$ be the graph obtained by removing $R$ from $G$; that is, $G_R=(V,E \setminus R)$.
We note (crucially) that while the analysis of the algorithm uses properties of $G_R$, the algorithm does not actually compute $R$. We will now show that $G'$ does not contain a cycle of $G_R$. Since $|R| \leq \epsilon n$, this means
that $G'$ has fewer than $(1+\epsilon)n$ edges.
Let $\sigma$ be a cycle in $G_R$.
Suppose $(w,v)$ is the edge of $\sigma$ with the largest rank. Since the connected components of $G_R$ are of size at most $k$, we infer that $\sigma$ has at most $k$ vertices, implying that $C_k(w,G)$ contains $\sigma$. It follows that on query $(w,v)$ the algorithm will return NO. Thus, $G'$ does not contain $\sigma$.
\EPFOF
\section{Lower bound}\label{sec:lower}
\renewcommand{\labelenumi}{\theenumi}
\renewcommand{\theenumi}{\roman{enumi}.}
The next theorem shows that there are graphs for which any local sparse spanning graph
algorithm must perform a number of queries that grows with $n$,
yet these graphs are $f$-non-expanding with $f(x)$ only slightly larger than $1/\log x$.
This is essentially the best one can hope for in light of Theorem~\ref{cor:f-non-expand-ub}.
\BT\label{theo:main}
For infinitely many $n$, there is an
$f$-non-expanding $n$-vertex graph $G$ with
$$f(x) = \frac{1}{\log x}\cdot (70\log\log x)^2$$
such that every $(\frac12,q)$-local sparse spanning graph algorithm
for the graphs isomorphic to $G$ satisfies $q \ge \log\log(n)/8000$.
\ET
\subsection{A regular non-expanding graph}
The main result in this subsection (stated in
Lemma~\ref{lemma:cube-cycle}) is a construction of
regular non-expanding graphs that we will use in
Subsection~\ref{sec:LB-proof} to prove
Theorem~\ref{theo:main}. A main ingredient is a result
from~\cite{MS15} showing that, roughly speaking, there are
graphs
that simultaneously have large
girth and small hereditary expansion (in fact, small edge
separators). While the degree of these graphs may grow with $n$, their maximum
degree is at most a constant times their average degree.
We will use this in order to construct a regular graph
with similar properties. We note that the regularity
condition is crucial for proving Theorem~\ref{theo:main}.
The following theorem was proved in~\cite{MS15}.
\BT[\cite{MS15}]\label{theo:cube}
For any $n,k$ with $2 \le k \le \frac{1}{648}\log\log n$ there is an $n$-vertex graph
$G=G_{n,k}$ satisfying:
\begin{enumerate}
\item $G$ has average degree at least $k$ and maximum degree at most $6k$.
\item $G$ has girth at least $\log n/(6k)^2$.
\item For every $t$-vertex subgraph $H$ of $G$ that is not a forest, there exists a subset $S \sub V(H)$ of size $(1/3)t \le \s{S} \le (2/3)t$ such that
$$\s{\partial_H(S)} \le \frac{t}{\log t}\cdot (\log\log t)^2 \;.$$
\end{enumerate}
\ET
We note that each of the parameters in
Theorem~\ref{theo:main} is quantitatively essentially
optimal (see~\cite{MS15} for further discussion).
The main result in this section is the following.
\BL\label{lemma:cube-cycle} For any $n_0$ there is a
connected graph $G_\circ$ on $n \ge n_0$ vertices
satisfying:
\begin{enumerate}
\item $G_\circ$ is $3$-regular.
\item $G_\circ$ has girth at least $\log\log(n)/2000$.
\item $G_\circ$ is $f$-non-expanding with $f(x) = (1/\log x)\cdot (4\log\log x)^2$.
\end{enumerate}
\EL
For the proof we will need the weighted version of the
well-known vertex separator theorem for trees. For
completeness, we give a short proof below.
\BCM\label{clm:tree} Let $T=(V,E)$ be a tree, and let
$w:V\to\R^+$ be a nonnegative weight function over the vertices of $T$.
There is a
vertex $v\in V$ whose removal disconnects $T$ into
connected components of weight at most $w(V)/2$
each.\footnote{For a subset $X \sub V$ we write
$w(X)=\sum_{v\in X} w(v)$.}
\ECM
\BPF
Start a walk in $T$ from an arbitrary vertex, in each step moving from a vertex $u$ to a neighbor $u'$ if the weight of the tree rooted at $u'$, when the edge $(u,u')$ is removed, is strictly greater than $w(V)/2$.
Since $T$ has no cycles and since the walk never reverts the last step taken, the walk eventually stops at some vertex $v$. This means that when $v$ is removed from $T$, the weight of the tree rooted at each of the neighbors of $v$ is at most $w(V)/2$. Since these trees are the connected components resulting
from the removal of $v$, we are done.
\EPF
\BPFOF{Lemma~\ref{lemma:cube-cycle}}
Set $k=\log\log m/648$ and let $G_{m,k}$ be the graph from Theorem~\ref{theo:cube}, where we take $m$ to be large enough such that
$k \ge \min\{n_0, 2\}$.
We note that in the rest of the proof we will use the inequality
\begin{equation}\label{eq:relations}
(6k)^4 \le \log m
\end{equation}
which holds since $m$ is sufficiently large.
As is well known, by iteratively removing vertices of $G_{m,k}$ of degree at most $k/2$, one obtains a (non-empty) graph of minimum degree at least $k/2$.
Let $G$ be a connected component of the largest average degree in the obtained graph.
Note that the average degree of $G$ is at least $k$, the maximum degree is still at most $6k$, and the girth is still at least $\log m/(6k)^2$.
Finally, $G$ still satisfies item $(iii)$ of Theorem~\ref{theo:cube}, being a subgraph of $G_{m,k}$.
Let $G_\circ$ be obtained by taking the replacement product of $G$ with a cycle. That is, $G_\circ$ is obtained from $G$ by replacing each vertex of degree $x$ by a cycle on $x$ new vertices -- which we henceforth refer to as a ``cloud'' -- and further adding edges as follows: if $u,v$ are adjacent in $G$, with $u$ being the $\ith$ neighbor of $v$ and $v$ being the $\jth$ neighbor of $u$ (under a fixed arbitrary enumeration of the neighbors of each vertex), then the $\ith$ vertex in the cloud corresponding to $v$ is connected by an edge to the $\jth$ vertex in the cloud corresponding to $u$.
So for example, it is easy to see that there is a one-to-one correspondence between the edges of $G$ and those edges of $G_\circ$ that connect vertices from different clouds.
Note that our graph $G_\circ$ is connected, as needed. Letting $n$ denote the number of its vertices,
note that $n$ equals the sum of the degrees of all vertices of $G$, so $n \ge k\s{V(G)} \ge n_0$, as needed.
Furthermore, $G_\circ$ is $3$-regular, since each vertex has two neighbors in its cloud and one neighbor in precisely one other cloud, as required by item~$(i)$ of the statement.
Let us now prove that the girth of $G_\circ$ is equal to the minimum between the girth of $G$ and the minimum degree of $G$. First, note that any cycle $C$ in $G_\circ$, other than a cloud, naturally determines a closed trail in $G$ (i.e., where vertices may be visited more than once, but not edges).
Indeed, for each edge of $C$ that connects two different clouds, the trail simply moves along the corresponding edge in $G$.. Note that the length of $C$ is at least the length (i.e., number of edges) of the trail. Since the length of the shortest closed trail in $G$ is its girth, we conclude that the length of any cycle in $G_\circ$ is at least the girth of $G$, unless that cycle is a cloud. Furthermore, since the smallest number of vertices in a cloud of $G_\circ$ equals the minimum degree of $G$, our claim follows.
That is, the girth of $G_\circ$ is at least
$$\min\{\log m/(6k)^2,\,k/2\} =\log\log(m)/1296 \ge \log\log(n)/2000 \;,$$
where we used the setting of $k$, Equation~(\ref{eq:relations}) and the fact that $n \le 6k\s{V(G)} \le 6km \le m^2$.
This proves item~$(ii)$ of the statement.
It remains to show that $G_\circ$ satisfies item~$(iii)$ of the statement as well.
Let $H_\circ$ be a $t$-vertex subgraph of $G_\circ$. Our goal is to bound $\phi_{H_\circ}$ from above.
Let $H$ be the induced subgraph of $G$ obtained by retaining only those vertices whose corresponding cloud has at least one vertex in $H_\circ$. Put $h=\s{V(H)}$, and notice $t \ge h$.
We next consider two cases, depending on whether $H$ is a forest or not.
First, suppose that $H$ is not a forest.
Hence, by item~$(iii)$ of Theorem~\ref{theo:cube} (a property which is also satisfied by $G$, as mentioned above) there is a partition $V(H)=S \cup S'$ with $\s{S},\s{S'} \ge h/3$ satisfying $\s{\partial_H(S)},\s{\partial_H(S')} \le (h/\log h) \cdot (\log\log h)^2$.
Let $S_{\circ}$ be the subset of $V(H_\circ)$ corresponding to $S$ (i.e., obtained by replacing each vertex in $S$ with the vertices of its cloud in $H_\circ$).
Assume without loss of generality that $\s{S_{\circ}} \le t/2$ (otherwise take $S'_{\circ}$, which is defined from $S'$ in a similar fashion).
Observe that $\s{\partial_{H_\circ}(S_{\circ})} \le \s{\partial_{H}(S)}$, since any edge in $\partial_{H_\circ}(S_{\circ})$ must go between two different clouds,
and there is a unique edge in $\partial_{H}(S)$ connecting the two vertices corresponding to these clouds.
Therefore,
$$\phi_{H_\circ} \le \frac{\s{\partial_{H_\circ}(S_{\circ})}}{\s{S_{\circ}}} \le
\frac{(h/\log h) \cdot (\log\log h)^2}{h/3} = \frac{3(\log\log h)^2}{\log h} \le \frac{3(\log\log t)^2}{\log (t/6k)} \le \frac{6(\log\log t)^2}{\log t} \;,$$
where
in the second inequality we used the fact that $\s{S_{\circ}} \ge \s{S} \ge h/3$, in the third inequality we used the fact that $h \le t \le 6k\cdot h$, and in the last inequality we used the fact that $t/6k \ge \sqrt{t}$ (i.e., $\sqrt{t} \ge 6k$); the latter follows from the fact that since $H$ is not a forest, $h$ is at least the girth of $G$, so $t \ge h \ge \log m/(6k)^2 \ge (6k)^2$ by Equation~(\ref{eq:relations}).
This proves item~$(iii)$ of the statement under the assumption that $H$ is not a forest.
Suppose next that $H$ is a forest. Notice we may assume that $H$ is a connected graph since otherwise $H_\circ$ is also not connected, meaning that $\phi_{H_\circ} = 0$ so we are done.
We apply Claim~\ref{clm:tree} on the tree $H$, where we set the weight of each vertex in $H$ to be the number of vertices in the corresponding cloud in $H_\circ$. Let $v$ be the vertex guaranteed by Claim~\ref{clm:tree}, and let $v_1,\ldots,v_d$ be the vertices of the cloud/cycle corresponding to $v$, in their order on the cycle. For each $1 \le i \le d$, let $S_i \sub V(H_\circ)$ be the set of vertices in $H_\circ$ corresponding to the $\ith$ connected components of $H-v$ (i.e., so that $v_i$ is the unique vertex in the cloud of $v$ that is connected to $S_i$).
Put $S'_i = S_i\cup\{v_i\}$. Then $\sum_{i=1}^d \s{S'_i} = t$, and our choice of $v$ guarantees that $\s{S'_i} \le t/2 + 1$.
We claim that there is an index $1 \le j \le d$ such that $(1/4)t \le \sum_{i=1}^j \s{S'_i} \le (3/4)t$.
Indeed, if $1\le j\le d$ is the smallest index such that $\sum_{i=1}^j \s{S'_i} \ge (1/4)t$ then
$$\sum_{i=1}^j \s{S'_i} = \sum_{i=1}^{j-1} \s{S'_i} + \lvert S'_j \rvert \le (t/4-1) + (t/2+1) = (3/4)t \;.$$
Now, let $S_\circ \sub V(H_\circ)$ be the smallest between $\bigcup_{i=1}^j S'_i$ and its complement, so that $t/4 \le \s{S_\circ} \le t/2$.
Observe that since $\{1,2,\ldots,j\}$ is an interval, $\s{\partial_{H_\circ}(S_\circ)} \le 2$.
We conclude that
$$\phi_{H_\circ} \le 2/(t/4) = 8/t \le (1/\log t)\cdot (4\log\log t)^2 \;,$$
where in the last inequality we used the fact that $(x/\log x)\cdot (\log\log x)^2 \ge 1/2$, which can be verified to hold for any real $x\ge 3$ (and thus for any integer $t>2$).
This completes the proof.
\EPFOF
\subsection{Lower bound proof}\label{sec:LB-proof}
For our proof of Theorem~\ref{theo:main} we will need the graph witnessing the lower bound to contain a bridge.
The following lemma shows that one can modify a given graph so as to contain a bridge while preserving high girth and small hereditary expansion.
\BL\label{lemma:bridge} Suppose there is a $3$-regular
connected $n$-vertex graph $G$ with girth $g$ that is
$f$-non-expanding, where $f:[1/2,\infty)\to\R$ is monotone
decreasing. Then there is a $3$-regular connected
$(2n+2)$-vertex graph that contains a bridge, and moreover,
has girth at least $g$ and is $h$-non-expanding with
$h(x)=3f(x/2-1)$.
\EL
\BPF Let $G_1,G_2$ be two vertex-disjoint copies of $G$.
Let $e_i$ be an arbitrary edge of $G_i$, $i\in\{1,2\}$, and
let $G_i'$ be obtained by subdividing $e_i$. That is,
$G_i'$ is obtained from $G_i$ by adding a new vertex $w_i$,
removing the edge $e_i=(u_i,v_i)$ and adding the edges
$(u_i,w_i),(w_i,v_i)$. It is clear that subdividing an edge
does not decrease the girth. Now, construct the graph $F$
from the union of $G_1'$ and $G_2'$ by adding the bridge
$(w_1,w_2)$. It is clear that $F$ is $3$-regular, connected
and has girth at least $g$. It therefore remains to show
that $F$ is $h$-non-expanding. Let $H$ be a $t$-vertex
subgraph of $F$ with $t>2$. We need to show that $\phi_H \le h(t)$.
Without loss of generality, $H$
has at least $t/2$ vertices in $G_1'$. Let $H'$ be the
subgraph of $H$ induced by those vertices, where we remove
the subdividing vertex $w_1$ if $w_1 \in V(H)$. Note that
$H'$ is a subgraph of $G_1$. Let $t' \ge t/2-1$ denote the
number of vertices of $H'$. Since $H'$ is
$f$-non-expanding, there is a subset $S \sub V(H')$ with
$\s{S} \le t'/2$ and $\s{\partial_{H'}(S)}/\s{S} \le f(t')$. Note
that $\s{\partial_{H}(S)} \le \s{\partial_{H'}(S)}+2$,
since the only edges in $H$ connecting a vertex in $H'$ and
a vertex not in $H'$ are $(u_1,w_1)$ and $(w_1,v_1)$. We
conclude that
$$\phi_H \le 3f(t') \le 3f(t/2 - 1) = h(t) \;,$$
where in the second inequality we used the monotonicity of $f$ for $t \ge 1/2$.
\EPF
\medskip
For a local sparse spanning graph algorithm $\calA$, we denote by $\calA(G,u,v) \in \{0,1\}$ the output of $\calA$ when the input graph is $G=(V,E)$ and the input edge is $(u,v) \in E$.
The \emph{query-answer transcript} of $\calA$ on $G$, where $\calA$ makes $q$ queries and $G$ is $d$-regular, is the sequence of triples $((x_j,i_j,y_j))_{j=1}^q$ where $(x_j,i_j) \in V\times[d]$ is the $\jth$ query and $y_j\in V$ is the corresponding answer.
Finally, for a permutation $\sigma$ on $V$,
we denote by $\sigma(G)$ the graph isomorphic to $G$ on the same vertex set, for which $(u,v) \in E(\sigma(G))$
if and only if $(\sigma(u),\sigma(v)) \in E(G)$.
We stress that in what follows, the graph $\sigma(G)$ will not necessarily have the same neighborhood ordering as that of $G$.
That is, if $y$ is the $\ith$ neighbor of $x$ in $G$ and $\sigma(v)=x,\sigma(u)=y$ then $u$ is {\em not} necessarily the $\ith$ neighbor of $v$ in $\sigma(G)$.
\BL
\label{lemma:SSG_girth}
Let $G$ be a $3$-regular
connected graph of girth $g$ that contains a bridge. Any
$(\frac12,q)$-local sparse spanning graph algorithm for the
graphs isomorphic to $G$ satisfies $q \ge g/2$.
\EL
\BPF
Let $\calA$ be an $(\frac12,q)$-local sparse
spanning graph algorithm for the graphs isomorphic to $G$,
and assume, contrary to the claim in the lemma, that $q < g/2$.
We shall say that $\calA$ {\em accepts\/}
an edge $(u,v)$ in $G$ if it gives a positive answer when queried on $(u,v)$ (that is, $(u,v)$ belongs to the sparse spanning graph $G'$).
We will show that with probability $1$
over its random coins, $\calA$ accepts every edge of $G$. This
will complete the proof as it means that the number
of edges of $G$ that $\calA$ accepts is $(1+\frac12)n$,
where $n$ is the number of vertices of $G$,
contradicting condition~$(ii)$ in
Definition~\ref{dfn:SSG-alg}.
Let $(u,v) \in E(G)$ and assume for contradiction that there is a sequence $r$ of random coins for $\calA$ such that the
corresponding deterministic algorithm $\calA_r$ satisfies $\calA_r(G,u,v)=0$.
Suppose, without loss of generality, that the vertex
set of $G$ is $[n]$ and that
$(1,2)$ is a bridge in $G$.
We will construct a permutation $\sigma$ on $[n]$ with $\sigma(u)=1,\, \sigma(v)=2$
so that the graph $\sigma(G)$ (with an appropriate way of ordering the neighbors of each vertex)
has the property that the query-answer transcript of $\calA_r(G,u,v)$ is identical to that of $\calA_r(\sigma(G),u,v)$.
Note that $\calA_r(\sigma(G),u,v)$ is well defined since the input edge $(u,v)$ is indeed an edge of $\sigma(G)$, and since $\sigma(G)$ is a valid input graph to $\calA_r$ being isomorphic to $G$.
Since $\calA_r$ is deterministic, whether or not $\calA_r$ accepts $(u,v)$ depends solely on the
query-answer transcript.
Therefore, the existence of $\sigma$ as above would imply that $\calA_r(\sigma(G),u,v)=0$. However, this would contradict condition~$(i)$ in Definition~\ref{dfn:SSG-alg} since $(u,v)$ is a bridge in $\sigma(G)$.
Let $Q=(x_j,i_j,y_j)_{j=1}^q$ be the query-answer transcript of $\calA_r(G,u,v)$.
We first claim that if a permutation $\sigma$ and an ordering of the neighbors of each vertex
of $\sigma(G)$, are such that $\sigma(u)=1,\sigma(v)=2$ and
for every $1 \leq j \leq q$ the $i_j$-th neighbor of vertex $x_j$ in $\sigma(G)$ is vertex $y_j$ then the query-answer transcript of $\calA_r(G,u,v)$ is
identical to the query-answer transcript of $\calA_r(\sigma(G),u,v)$.
To see this, let the query-answer transcript of $\calA_r(\sigma(G),u,v)$ be denoted by $(x_j',i_j',y_j')_{j=1}^{q'}$.
We prove, by induction on $j$, that the two query-answer transcripts are the same when restricted to the first $1\le j\le q$ queries, that is, $(x'_j,i'_j)=(x_j,i_j)$ and $y'_j=y_j$ for every $1\le j\le q$.
Note that this will also imply that $q=q'$ (i.e., that the number of queries is identical).
For $j=1$ we have $(x'_1,i'_1)=(x_1,i_1)$ since $\calA_r$ is deterministic and in both cases the input is $(u,v)$.
Our assumption on $\sigma$ thus guarantees that we also have $y'_1=y_1$.
Suppose our claim holds for the first $j-1$ queries.
Again, since $\calA_r$ is deterministic, the $\jth$ query is determined only by the query-answer transcript of the first $j-1$ queries (and the input edge).
Hence, the induction hypothesis implies that $(x'_j,i'_j)=(x_j,i_j)$ and our assumption on $\sigma$
again implies that we also have $y'_j=y_j$. This completes the inductive proof.
It follows that in order to
complete the proof it suffices to find a permutation $\sigma$ and an ordering of
the neighbors of each vertex, as above.
Let again $Q=(x_j,i_j,y_j)_{j=1}^q$ be the query-answer transcript of $\calA_r(G,u,v)$,
and let $F$ be the (labeled) graph spanned by the edge set\footnote{$E(F)$ might contain the edge $(x,y)$ twice if $y$ is the $\ith$ neighbor of $x$, $x$ is the $\jth$ neighbor of $y$ and the algorithm queried both $(x,i)$ and $(y,j)$. In this case we will keep just one copy of $(x,y)$ thus making sure
that $E(F)$ is indeed a set, and not a multi-set.}
$$E(F)= \{(x_j,y_j)\}_{j=1}^q \cup \{(u,v)\} \;.$$
Since
\begin{equation}\label{eq:EF}
\s{E(F)} \le q+1 \le g/2 \;,
\end{equation}
we have that $F$ is a forest.
Let $T_1,\dots,T_k$ be the (labeled) trees in $F$. For the sake of defining
$\sigma$ it will be convenient to consider a single tree $T$.
The edge-set of $T$ consists of $E(F)$ and $k-1$ additional
edges. The additional edges do not necessarily belong to $G$, and are selected as follows.
For each labeled tree $T_i$, let $t_i$ denote an
arbitrary vertex of degree smaller than $3$. For every
$i\in [k-1]$, add the edge $(t_i, t_{i+1})$.
Observe that Equation~(\ref{eq:EF}) implies that $\s{E(T)} <
g$. Consider a rooted version of $T$ where $u$ is the root,
and construct $\sigma$ as follows. Set $\sigma(u)=1$,
$\sigma(v)=2$, and define the neighborhood relation between
$u,v$ in $\sigma(G)$ as it is in $G$. That is, if $u$ is
the $\ith$ neighbor of $v$ and $v$ is the $\jth$ neighbor
of $u$ in $G$ then the same holds in $\sigma(G)$. Suppose
we have already defined $\sigma(x)$ for all $x$ at distance
at most $d-1$ from $u$ (in $T$) as well as for some
vertices at distance $d$, and let $y$ be a vertex at
distance $d$ for which $\sigma(y)$ has not been defined
yet. Let $x$ be the parent of $y$ in $T$ (whose distance
from $v$ is thus $d-1$) and let us set $\sigma(y)$ to be a
neighbor of $\sigma(x)$ in $G$ which is not the image of
any vertex under the $\sigma$ we have defined thus far.
Such a vertex exists since $G$ is $3$-regular and the
degree in $T$ is at most $3$. If the edge $(x, y)$ is in
$F$ then we define the neighborhood relation between
$\sigma(x)$ and $\sigma(y)$ as $x$ and $y$ in $G$. Once we
define $\sigma$ for all vertices of $T$ we arbitrarily
extend $\sigma$ to a permutation, and extend the
neighborhood relation between the vertices in a consistent
manner.
\EPF
We are now ready to prove Theorem~\ref{theo:main}.
\BPFOF{Theorem~\ref{theo:main}}
Let
$$h(x)=\frac{1}{\log(3x)}\cdot 32(\log\log(8x))^2 \;.$$
It is not hard to check that $h:[1/2,\infty)\to\R$ is monotone decreasing.
Note that the graph in Lemma~\ref{lemma:cube-cycle} is $h$-non-expanding, since for $x\ge 3$,
$$h(x) \ge \frac{1}{\log(x^2)}\cdot 32(\log\log x)^2 = \frac{1}{\log x}\cdot (4\log\log x)^2 \;.$$
Apply Lemma~\ref{lemma:bridge} on the graph(s) in Lemma~\ref{lemma:cube-cycle}.
We get a $3$-regular connected $n$-vertex graph, for infinitely many $n$, that contains a bridge, has girth at least
$$\frac{\log\log((n-1)/2)}{2000} \ge \frac{\log\log(n/4)}{2000} \ge \frac{\log\log(n)}{4000}$$
and is $f$-non-expanding with
$$f(x)= 3h(x/2-1) \le \frac{1}{\log (x/2)} \cdot 96(\log\log(4x))^2
\le \frac{3}{\log x} \cdot 96(4\log\log x)^2
\le \frac{1}{\log x} \cdot (70\log\log x)^2$$
where we assumed $x\ge 3$.
The proof now follows immediately from Lemma~\ref{lemma:SSG_girth}.
\EPFOF
\bibliographystyle{plain}
|
\section{Introduction}
Although network functionalities play an important role in enterprise network to make it robust, fast, and secure, they often burden a company with great hardware financial pressure and management complexity. Network middleboxes, such as firewalls and intrusion detection systems (IDS), are the customized appliances to implement these sophisticated functionalities, and they are also often hard to deploy and upgrade. A recent survey \cite{sherry2012making} reveals that the investments in network infrastructures deployment of these middleboxes as well as in the personnel cost of managing and maintaining them are substantial.
The emergence of SDN helps us separate the logic control from the basic traffic processing, so we can relieve some of the above ``pain points'' to a certain extent by taking the advantage of SDN\cite{casado2007ethane,gember2012toward,qazi2013simple,DPIaS14}.
But though the management of networks can be simplified by SDN, it still has many shortcomings. For instance, because these functionalities are still implemented within the enterprise network, the hardware cost and the everyday maintenance cost are still there.
Thus, to further reduce the cost and complexity faced by local networks, some SDN researchers have attempted to outsource these network functionalities or middleboxes out to cloud providers, in the same way that the computation and storage services have been successfully outsourced \cite{sherry2012making, gibb2012outsourcing}. After migrating middleboxes to cloud, the local enterprises will receive better services with less expenditure and management complexity since the cloud provider can deploy more advanced hardware and hire professional experts to offer better network functionality services by taking advantage of economies of scale. Moreover, the local enterprise can also have a clearer and more elastic network infrastructure.
However, the outsourcing of local SDN middleboxes may bring in serious privacy issues\cite{sherry2012making}. Because at present, an enterprise has to provide the detailed underlying configurations of these middleboxes, which may embody very sensitive and important enterprise secrets, when outsourcing middleboxes. For example, the configurations of firewalls or IDS may reveal what the enterprise network topology looks like and how different sectors within an enterprise are regulated. These policies or configurations are set and known by a restricted few administrators even before outsourcing. Were these confidential pieces of information hidden in these polices to be exposed to adversaries, they may be utilized to impose a great threat to the enterprise\cite{SRDS_Liu2012}.
In brief, the new challenges in this scenario now we face are: 1) we need to protect the private information hidden in the configurations of these outsourced middleboxes; 2) we need to allow the cloud provider to perform the network functionalities without mistakes ( or with acceptable errors). Note that in line with existing outsourcing architecture in SDN\cite{sherry2012making,karaoglu2013offloading,gibb2012outsourcing}, we assume that the cloud provider for an enterprise is its Internet Service Providers (ISP), so we will not consider the protection of the packet flows of the enterprise here.
In this paper, we try to address these challenges by designing a privacy preserving scheme for network middlebox outsourcing in SDN, exploiting a \textit{cryptographic multilinear map}, a powerful cryptographic tool that was recently invented \cite{zvika2013obfus,coron2013practical,garg2013candidate}. To keep things concrete, in this paper we will take firewalls, a typical and indispensable network middlebox, as a case study.
The main idea of our scheme is to use cryptographic multilinear map to obfuscate the rules in a firewall efficiently such that the original configurations are protected while their functionality is preserved. The theoretical foundation of our idea is cryptographic \textit{program obfuscation} \cite{barak2012possibility,programobf}.
In our work, we first present framework SOFA (\underline{S}ecure framework for \underline{O}utsourcing \underline{F}irew\underline{A}ll). This framework consists of two phases. First, the controller in the enterprise SDN constructs a cryptographic obfuscator for the firewall based on its rules. This obfuscator encodes the firewall rules such that the underlying confidential information cannot be derived. Then this enterprise transmits this obfuscated firewall to a provider.
In the second phase, the cloud provider performs the firewall functionality by filtering incoming and outgoing network traffics with this obfuscated firewall without knowing the detailed configurations.
Based on this overall framework, a basic scheme and two enhanced schemes are proposed to address the detailed technical issues.
To summarize, we have made three main contributions in our work:
\begin{itemize}
\item To the best of our knowledge, we are the first to deal with the privacy problem about network functionality outsourcing (NFO) in SDN. The obfuscation approach we propose has a solid theoretical foundation. And framework SOFA is compatible with existing SDN outsourcing architecture, such as APLOMB\cite{sherry2012making}.
\item To outsource firewalls as a concrete example, we propose a basic scheme which is efficient in the first phase, and two enhanced schemes which can further improve the efficiency in the second phase.
We also analyse the complexity and security of our algorithms to show that they are efficient, secure, and reliable.
\item We implement a prototype of our schemes, and evaluate its performance on real-world firewall rules, the results show that our proposed schemes are practical.
\end{itemize}
Our paper are organized as follows. First, we review some related work in section \ref{sec2}. In section \ref{sec3}, we give formal definition to the privacy issues in network functionality outsourcing and introduce some basic cryptographic backgrounds. In section \ref{sec4}, we present framework SOFA and a basic scheme. Some approaches which can boost its efficiency are discussed in \ref{Improve}. The security of these algorithms is analysed in section \ref{sec5}. Evaluation of SOFA with firewall rules is presented in section \ref{sec6}. Finally, in section \ref{sec7} we conclude our work and discuss some potential future research directions.
\section{Related Works}
\label{sec2}
In this section, we review some recent works on outsourcing frameworks in SDN, some privacy-preserving outsourcing techniques in cloud computing or traditional enterprise networks.
We can see that no existing work can effectively address the novel challenges we have proposed.
\subsection{Outsourcing network functionality in SDN}
There has already been much research done about the outsourcing of network middleboxes services exploiting the advantage of SDN\cite{gember2013stratos,sherry2012making,karaoglu2013offloading,gibb2012outsourcing,kotronis2012outsourcing}.
Sherry et al.\cite{sherry2012making} design and implement a prototype of APLOMB architecture, by which the middlebox processing within enterprises can be outsourced to clouds. This architecture is mainly intended to deal with the problem of how to redirect the traffic to/from the enterprise SDN to the cloud without greatly damaging performance. It is compatible with many specific middleboxes, such as firewalls.
Gibb et al.\cite{gibb2012outsourcing} also envision outsourcing these network functionalities to external \textit{Feature Providers}. The main goal of their work is to provide anyone with chances to be a network service provider without any limitation on location.
\textit{Cloud-Assisted Routing}\cite{karaoglu2013offloading} is another scheme aimed at offloading the local computation-intensive and memory-intensive operations in the routing functionality to the cloud.
There is only a little current research concerning the privacy and security problem of outsourcing network functionality in SDN. Fayazbakhsh et al.\cite{fayazbakhsh2013verifiable} mainly discuss the roadmap to construct a verifiable NFO architecture that can verify the correctness on intended functionality, performance assurance and can audit the actual workload in the cloud.
However, none of the these works about outsourcing in SDN can address the privacy challenges we have stated.
\subsection{Privacy Preserving outsourcing in Cloud Computing}
In literature, many schemes are designed to tackle the privacy problem of outsourcing data storage and computation rather than the network middlebox\cite{wang2010privacy,wei2014security}. However, because the network functionality is often performed on real-time packet traffics in the cloud, its requirement is different. For example, the well-known public auditing framework \cite{wang2010privacy} introduces a third party auditor to help cloud users check the integrity of their outsourced data. However, in this framework the user files are sent to the cloud directly without encryption, so it will not be applicable to network middlebox outsourcing. So these existing approaches for cloud computing or storage are incapable of addressing this new problem.
Khakpour and Liu\cite{SRDS_Liu2012} present Landon, a framework to address the problem of how to protect the firewall policies when it is outsourced to a cloud provider in traditional enterprise networks. The main idea of this framework is to convert the raw access control lists to an equivalent Firewall Decision Diagram (FDD) and then to anonymize the FDD with Bloom Filters. This framework works well after some complex optimization methods are used. The limitation of this framework is that it is based on a specific presentation (FDD) of firewall and the Bloom Filter is not very secure and has some drawbacks\cite{rottenstreich2012bloom}.
\subsection{Related Cryptographical Research}
In cryptography, there is also some theoretical work dealing with similar problems, such as methods based on Yao's \emph{garbled circuits} for secure multi-party computation\cite{yao1986generate}. In particular, recently Brakerski and Rothblum\cite{zvika2013obfus} propose an elegant construction for obfuscating conjunctions based on an asymmetric multilinear map. This is the fundamental cryptographic tool we will use in this paper.
But their work focuses on the theoretical construction and security proof of obfuscating a single conjunction. So their result, though significant in theory, will be impractical in practice in directly obfuscating numerous firewall rules. We will discuss this problem at length in section \ref{sec4}.
\section{Problem formulation and Preliminaries}
\label{sec3}
In this section, we formulate the privacy problems of NFO in SDN. To this end, we first review the overall architecture for outsourcing network functionality, and give its threat model. Specifically, we give a definition for privacy preserving outsourcing firewalls.
In addition, we also recall some important cryptographic backgrounds about multilinear map.
\subsection{System Model}
The overall architecture for outsourcing network functionality to cloud has been proposed, such as APLOMB\cite{sherry2012making} and Jingling\cite{gibb2012outsourcing}. In such architecture, enterprise administrators specify the policies of middleboxes with its local SDN controller in a simple way. Then by the communication mechanism between the enterprise clients and the cloud provider, these policies are transmitted to the cloud controller and service requirements are also negotiated. The controller in cloud SDN then translate these policies into low-level network configurations for middleboxes, and then manage the cloud network resources to carry out the tasks. The traffic redirection of the enterprise network flows is also managed by this architecture.
We will focus on the privacy protection of middlebox configurations in the above related procedures.
\subsection{Threat Model}
We assume that the cloud middlebox provider is \textit{semi-honest} since in the outsourcing system model, the cloud providers are commonly the ISPs for enterprises. Because the provider performs network tasks on packet traffics of the enterprise, it will be able to observe the input packets and the output actions of them. So the \textit{semi-honest} provider may be curious to learn information from these input-output histories. But the provider should not actively generate tremendous packets to attack the outsourced middlebox $\mathcal{MB'}$ for such an attack will affect the service quality it provides. So the two main threats we consider are the analysis of the outsourced $\mathcal{MB'}$ itself, and the analysis of its running input-output history.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.3}
\centering
\caption{ Firewall Rule Examples }\label{table1}
\begin{tabular}{@{} c |c |c |c |c |c |c @{}}
\hline
\multirow{2}{0.5cm}{rule \#} & \multicolumn{2}{c|}{Source} & \multicolumn{2}{c|}{Destination} & \multirow{2}{0.8cm}{Protocol} & \multirow{2}{0.8cm}{Action} \\
\cline{2-3} \cline{4-5}
& IP & Port & IP & Port & & \\
\hline
1 & 192.168.45.* & * & * & * & * & deny \\
\hline
2 & 10.*.*.* & * &192.168.4.* &80 &TCP &permit \\
\hline
3 & 10.56.*.* &* &192.168.*.* &[22,88] &TCP &permit \\
\hline
4 & 114.212.190.* &8000 &* &8090 &UDP &deny \\
\hline
\end{tabular}
\end{table}
\subsection{Problem Definition}
Different middleboxes have their own characteristic in detailed configurations, thus in the following we focus on the firewall as an example.
The configuration of a typical IP firewall \textit{f} is an ordered sequence of rules, and they are usually presented in the format of access control lists (ACLs). Some examples are shown in Table \ref{table1}. Technically, the rules can be classified as two types: standard ACLs and extended ACLs. The standard rules only allow you to deny or permit packets by specifying the source IP addresses, like the first one in Table \ref{table1}. Extended rules also allow you to specify the protocol type, the destination address and the port ranges, like the other examples in table \ref{table1}. Once the configuration are settled, the firewall can perform packet filtering based on these rules to protect an enterprise network.
The rules of the firewall are what we need to protect since they may reveal, for example, much information of the enterprise's inner network topology and its security vulnerabilities. Specifically, we think that the IP address is more important than TCP port since the IP address determines the host end while the TCP ports are for specific processes in a host. Formally, we give the following definition:
\begin{definition}(\textbf{Privacy Preserving Firewall Problem})
\label{PPFP}
The Privacy preserving of a firewall \textit{f} is the construction of a semantically equivalent secure firewall \textit{f\'} such that (1) for every packet \textit{p}, the probability of the decisions on it by \textit{f\'} and \textit{f} are different is negligible, (2) it should be computationally efficient to perform the filtering tasks with \textit{f\'}, and (3) there should be no computationally efficient ways to get the whole original plain rules of \textit{f} based on \textit{f\'} itself.
\end{definition}
It should be easy for us to give privacy preserving definitions for other network middleboxes similarly.
\subsection{Preliminaries}
Here, we first review the definition of \textit{ cryptographic multilinear map}, which is the cornerstone of our algorithms.
\begin{definition}(\textbf{$\kappa$-multilinear Map}\cite{garg2013candidate,boneh2003applications})
For $\kappa+1$ cyclic groups $G_1,\dots, G_\kappa, G_T$ of the same order $p$, a \textit{$\kappa$-multilinear map} $e : G_1\times \dots \times G_\kappa \rightarrow G_T$ has the following properties:
(1) For elements $g_1 \in G_1,\dots, g_\kappa \in G_\kappa$, index $i \in \left[ \kappa \right]$ and an integer $\alpha \in \mathbb{Z}_p $, we have:
\[e(g_1,\dots,\alpha \cdot g_i,\dots,g_\kappa)= \alpha\cdot e(g_1,\dots,g_\kappa)\]
(2) The map $e$ is non-degenerate, which means that if $g_i \in G_i\; (i=1,\dots,\kappa) $ is a generator of $G_i $, then $e(g_1,\dots,g_\kappa)$ is a generator of the target group $G_T$.
In the above definition, if $G_i \; (i=1,\dots,\kappa)$ are all identical groups, it is called a symmetric multilinear map.
\end{definition}
A cryptographic multilinear map has been a long sought after and powerful tool. Not until recently was it approximately constructed in the form of Graded Encoding System \cite{garg2013candidate,coron2013practical}. In particular, we will design our algorithms based on the construction over integers, which is a more practical one devised by Coron, Lepoint, and Tibouchi\cite{coron2013practical}. Now we briefly recall the definition of Graded Encoding System (GES) and its associated efficient procedures for manipulating encodings on which the description of our algorithms is based directly.
Note that we only consider the symmetric case here.
\begin{definition}(\textbf{$\kappa-$Graded Encoding System}\cite{garg2013candidate})
\label{Def2} A $\kappa$-Graded Encoding System consists of a ring $R$ and a system of sets $\mathcal{S} = \lbrace{S^{(\alpha)}_v \in {\lbrace 0,1\rbrace}^* : v \in \mathbb{N}, \alpha \in R \rbrace}$, with the following properties:
1. For every $v \in \mathbb{N}$, the sets $\lbrace S^{(\alpha)}_v : \alpha \in R \rbrace$ are disjoint.
2. There is an associate binary operation `+' and a self-inverse unary operation `$-$' (on ${\lbrace 0,1 \rbrace}^*$) such that for every $\alpha_1, \alpha_2 \in R$, every index $i \leq \kappa$, and every $u_1 \in S^{(\alpha_1)}_i$ and $u_2 \in S^{(\alpha_2)}_i$, it holds that
\[ u_1 + u_2 \in S^{(\alpha_1 + \alpha_2)}_i \; , \; -u_1 \in S^{(-\alpha1)}_i\]
where $\alpha_1 + \alpha_2$ and $-\alpha_1$ are addition and negation in $R$.
3. There is an associate binary operation `$\times$' (on ${\lbrace 0,1 \rbrace}^*$) such that for every $\alpha_1, \alpha_2 \in R$, every $i_1, i_2$ with $i_1 + i_2 \leq \kappa$, and every $u_1 \in S^{(\alpha_1)}_{i_1}$ and $u_2 \in S^{(\alpha_2)}_{i_2}$, it holds that $u_1 \times u_2 \in S^{(\alpha_1 \cdot \alpha_2)}_{i_1 + i_2}$. Here $\alpha_1 \cdot \alpha_2$ is multiplication in $R$, and $i_1 + i_2$ is integer addition.
\end{definition}
\begin{definition}(\textbf{Efficient Procedures for $\kappa$-Graded Encoding System}\cite{garg2013candidate,coron2013practical})
\label{Def3} For graded encoding system, we have the following efficient procedures:
\textbf{Instance Generation}:$(params,P_{zt})\leftarrow\mathnormal{InstGen}(1^\lambda, 1^\kappa)$, where $params$ is a description of a $\kappa$-Graded Encoding System with security parameter $\lambda$, and $P_{zt}$ is a zero-test parameter for level $\kappa$.
\textbf{Ring Sampler}: $c \leftarrow \mathnormal{samp(params)}$. This procedure outputs a ``level-zero encoding" $c \in S^{(\alpha)}_0$ for a nearly uniform element $\alpha \in_R R$.
\textbf{Encoding}: $c_k \leftarrow \mathnormal{enc(params,k,c)}$. This procedure outputs the ``level-i encoding" $c_k \in S^{(\alpha)}_i$ for a ``level-zero encoding" $c \in S^{(\alpha)}_0$.
\textbf{Re-Randomization}: $c' \leftarrow \mathnormal{reRand(params,i,c)}$. Procedure $reRand$ can re-randomize encodings relative to the same level i.
\textbf{Addition and negation}: $u' \leftarrow \mathnormal{add(params,i,u_1,u_2)}$ and $u' \leftarrow \mathnormal{neg(params,i,u_1)}$. The two procedures are corresponding to operation `+' and `$-$' in the above definition.
\textbf{Multiplication}: $u' \leftarrow \mathnormal{mul(params, i_1, u_1, i_2, u_2)}$. This procedure is for operation `$\times$' in the above definition.
\textbf{Zero Testing}: $\mathnormal{isZero(params,P_{zt},u_{\kappa}) \overset{?}{=} 0/1}$. This procedure will check whether $u_{\kappa} \in S^{(0)}_{\kappa}$.
\textbf{Extraction}: $sk \leftarrow \mathnormal{ext(params, P_{zt},u_{\kappa})}$. This procedure extracts a ``canonical'' and ``random" representation of ring elements from their level-$\kappa$ encoding.
\end{definition}
\section{SOFA: Framework and A Basic Scheme}
\label{sec4}
As a concrete example for secure network middlebox outsourcing, we present SOFA in detail in this section. First, we will introduce our framework which is comprised of two phases, the obfuscating phase and the online execution phase. Then we give a naive scheme by directly applying the theoretical conjunction obfuscation technique \cite{zvika2013obfus}. The discussion of its lack of feasibility will lead to our basic scheme, which is much more efficient in obfuscating the original firewall. And in the next section, we will further present another two approaches to boost our scheme's efficiency in execution phase.
\subsection{Fundamental Framework}
A \textit{cryptographic obfuscator }of a program can make a program perform the same functionality while the detailed underlying instructions are hidden \cite{barak2012possibility}.
Combining the idea of \textit{program obfuscation} and the Definition \ref{PPFP}, we observe that the privacy preserving firewall problem can be solved by constructing a cryptographic obfuscator $\mathcal{O}$ for firewall \textit{f} such that \textit{f\'}=$\mathcal{O}(f)$. The secrets are the underlying firewall rules, and they can be hidden in \textit{f\'}. Generally, this relationship between privacy preserving outsourcing requirement and cryptographic obfuscator construction can be extended to other middleboxes.
Our framework SOFA consists of two phases. The first phase is the construction of an obfuscator for original firewall \textit{f}. And this work is done by the control plane of SDN in the local enterprise. The second phase is the execution of the obfuscated firewall \textit{f\'}=$\mathcal{O}(f)$ in the cloud. And this task should be done by the control plane of cloud firewall provider.
\textbf{Obfuscation Phase}: In the first phase, the controller of the enterprise SDN can set up the basic cryptographic parameters of the multilinear map. And then for each rule \textit{r} in the firewall, such as these in Table \ref{table1}, the control plane generates a corresponding atomic rule ``ciphertext" \textit{r\'}. Subsequently, the control plane can aggregate these atomic ciphertexts to be an integral secure firewall \textit{f\'}. And at last it sends \textit{f\'} and some necessary public parameters to the cloud provider. This obfuscation phase for firewall can be done offline and once for all. And if there are some updates for the firewall \textit{f}, we can just reconstruct the related atomic rules, reassemble them and then replace the old ones in \textit{f\'} with them. But this process still should be done as efficiently as possible since one main purpose of outsourcing is to reduce costs for enterprises.
\textbf{Execution Phase}: After receiving \textit{f\'}, the cloud firewall provider performs the packet filtering task on the inbound and outbound traffics of the protected enterprise. For each packet, the high-performance devices in the cloud, under the supervision of its controller, securely checks its packet header according to \textit{f\'}. If the packet is found satisfying a rule \textit{r\'}, the cloud firewall devices will take the corresponding actions associated with this rule, such as permitting it, or denying it.
The firewall obfuscation is implemented with cryptographic multilinear map. So in the following we model the firewall rule and traffic packet header in the bitwise level. For each rule, we present it in the format of $R = (\vec{v},W,A)$, where $\vec{v} \in {\lbrace 0,1\rbrace}^n$ specifies the expected bit, 0 or 1, on non-wildcard bits, the set $W \subseteq \left[ n \right] $ specifies the ``wildcard'' bits, and $A$ indicates the associated actions. Note that actually set $W$ is often the subnet mask in firewall rules. For ease of presentation, we denote $\vec{v} \left[ i \right]=0$ if $i \in W$. For instance, the rule \#1 in Table \ref{table1}, ``192.168.45.*'', can be presented as $\vec{v}=``1100000010101000..."$,
$W={\lbrace 25,26,\dots,32\rbrace}$ and $A=\lbrace deny \rbrace$. In addition, $A$ is the output decision on a packet, it leaks little information about the detailed enterprise network, so we do not consider protecting it in this paper. Please see section \ref{sec5} for a detailed analysis of our schemes' security. Similarly, for a tarffic packet's TCP/IP header, we also present its related bits as a n-bit vector $\vec{p} \in {\lbrace 0,1\rbrace}^n$.
\subsection{A Naive Scheme}
A naive and intuitive idea is to directly apply the theoretical construction of conjunction obfuscator\cite{zvika2013obfus} in obfuscating the firewall rules.
Here we give a concise description of this scheme. For each rule \textit{r} in firewall \textit{f}, we present it as $\textit{r} \triangleq (\vec{v},W,A)$. Then for the $i$-th bit of it, we generate two pairs of level-1 encodings in the GES (We term them ``encoding pairs unit"):
\begin{align}
(u_{i,0},v_{i,0})&\in({S^{\rho_{i,0}}_1, S^{\rho_{i,0} \times \alpha_{i,0}}_1}), \label{Eq1}\\
(u_{i,1},v_{i,1})&\in({S^{\rho_{i,1}}_1, S^{\rho_{i,1} \times \alpha_{i,1}}_1})\label{Eq2}
\end{align}
In Equation (\ref{Eq1}) and (\ref{Eq2}), $\rho_{i,0},\alpha_{i,0},\rho_{i,1},\alpha_{i,1}$ are all randomly and uniformly chosen if $i \notin W$, but for $i \in W$, they should also satisfy $\alpha_{i,0} = \alpha_{i,1}$. In addition to two pairs for each bit, an encoding pair is also generated for a whole rule:
\begin{equation}\label{Eq3}
(u_{n+1},v_{n+1}) \in(S^{\rho_{n+1}}_1,S^{\rho_{n+1}\times\prod_{i \in \left[ n \right] }\alpha_{i,\vec{v}\left[ i \right]}}_1)
\end{equation}
In the execution phase, for each obfuscated rule, we pick and multiply encodings from each pair according to $\vec{p}$ and get the following two (n+1)-level encodings:
\begin{equation}\label{Eq4}
LHS=u_{n+1}\times \prod_{i \in \left[ n \right] } v_{i,\vec{p}\left[ i \right]},
RHS=v_{n+1}\times \prod_{i \in \left[ n \right] } u_{i,\vec{p}\left[ i \right]}
\end{equation}
At last procedure $\mathnormal{isZero}$ is performed to determine whether $LHS$ (Left-hand side) and $RHS$ (Right-hand side) are equal to judge whether packet $\vec{p}$ satisfies this rule and which actions to take.
Actually, the scheme we briefly presented above is the simplified symmetric counterpart for the original asymmetric construction in \cite{zvika2013obfus}, and the asymmetry is vital in its theoretical security proof to defend the DDH attack which may leak the wildcard locations. It is much harder and inefficient to construct such an asymmetric multilinear map. More importantly, as we all know, the wildcard part in a IP address is for indicating the subnet range, and these bits are always continuous in the later part. Nothing additional useful information about specific subnets or hosts, except for the the size of an unknown subnet, can be discovered without $\vec{v}$. So leaking this subnet mask rather than the specific prefixes ( $\vec{v}$ ) is acceptable in most firewall scenarios. Thus in this paper, we do not attempt to perfectly defend such DDH attack and herein we use symmetric multilinear map in all our schemes and leave such protection as future work. But we point out that following schemes are also fit for asymmetric one.
\textbf{Discussion}. Now we argue that this naive scheme will have serious issues in practical application.
(1) Its time and space consumption in the obfuscating phase will be too large. If we have $l$ rules in a firewall, then the obfuscation will take $\Theta(ln+l)$ times encoding and re-randomization, and need store $\Theta(ln+l)$ encoding ciphertexts. In addition, when $n$ is large, the singe procedures will take more time and the storage will also be larger. So such privacy preserving scheme will be impractical due to the computation and storage cost in obfuscating phase is hardly acceptable.
(2) When $n$ is large, it also will be much harder to construct the GES due to the noise limit. Thus to make the $2(n+1)$ multiplication operations and $\mathnormal{isZero}$ viable, the underlying GES parameters have to be set large enough. Consequently the real-time execution in cloud will be too inefficient to have any practical usability if we adopt the naive scheme directly.
Therefore we have to make up for these limitations to make it more practical.
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{algorithm}[!t]
\caption{Basic Scheme For Obfuscating Firewall}
\label{Basic_phase1
\begin{algorithmic
\REQUIRE A original plain firewall \textit{f} with rule sets $\lbrace r\rbrace $.
\ENSURE A obfuscated firewall \textit{f\'} with the same functionality.
\STATE
$(params,P_{zt})\leftarrow\mathnormal{InstGen}(1^\lambda, 1^{n+1})$.
\FOR {$i \in \{1,\ldots,M+N\}$}
\STATE
$C_{\rho_{i,0}} \leftarrow \mathnormal{samp(\cdot)} ,\quad C_{\rho_{i,1}}\leftarrow \mathnormal{samp(\cdot)}$.
\IF{$i \leq M$}
\STATE $C_{\alpha_{i,0}} = C_{\alpha_{i,1}} \leftarrow \mathnormal{samp(\cdot)}$.
\ELSE
\STATE $C_{\alpha_{i,0}}\leftarrow \mathnormal{samp(\cdot)},\quad C_{\alpha_{i,1}} \leftarrow \mathnormal{samp(\cdot)}$.
\ENDIF
$C_\beta\leftarrow enc(\cdot,1,C_{\rho_{i,0}})$,
$u_{i,0}\leftarrow reRand(\cdot,1,C_\beta)$.
$C_\gamma \leftarrow enc(\cdot,1,C_{\rho_{i,0}} \times C_{\alpha_{i,0}})$,
$v_{i,0} \leftarrow reRand(\cdot,1,C_\gamma)$.
$C_\delta \leftarrow enc(\cdot,1,C_{\rho_{i,1}})$,
$u_{i,1} \leftarrow reRand(\cdot,1,C_\delta)$.
$C_\epsilon \leftarrow enc(\cdot,1,C_{\rho_{i,1}} \times C_{\alpha_{i,1}})$,
$v_{i,1} \leftarrow reRand(\cdot,1,C_\epsilon)$.
$C[i] \triangleq \lbrace(u_{i,0},v_{i,0}),(u_{i,1},v_{i,1})\rbrace$.
\ENDFOR
\STATE $C$=$\sigma(C)$.
\STATE $E=\lbrace \sigma(1),\ldots,\sigma(M)\rbrace$,
\STATE $UE=\lbrace \sigma(M+1),\ldots,\sigma(M+N)\rbrace$.
\FOR { $\textit{r} \triangleq (\vec{v},W,A) \in$ \textit{f} }
\FOR {$i \in \{1,\ldots,n\}$}
\IF{$i \in W$}
\STATE $P_r[i]\overset{R}{\nleftarrow} E$.
\ELSE
\STATE $P_r[i]\overset{R}{\nleftarrow} UE$.
\ENDIF
\ENDFOR
\STATE $\rho_{i,n+1}\leftarrow \mathnormal{samp(\cdot)}$,
\STATE $C_\theta \leftarrow enc(\cdot,1,C_{\rho_{i,n+1}})$,
$u_{n+1} \leftarrow reRand(\cdot,1,C_\theta)$.
$C_\mu \leftarrow enc(\cdot,1,C_{\rho_{i,n+1}} \times\prod_{i \in \left[ n \right] }C[P_r[i]]. C_{\alpha_{i,\vec{v}\left[ i \right]}})$
$v_{n+1} \leftarrow reRand(\cdot,1,C_\mu)$.
\STATE \textit{r\'} $\triangleq$
$\big(P_r,u_{n+1},v_{n+1}, A\big)$.
\ENDFOR
\RETURN \textit{f\'}$\triangleq \big( {\lbrace C[i] \rbrace}_{i \in [M+N]}, \lbrace r' \rbrace, P_{zt}, params'\big)$.
\end{algorithmic}
\end{algorithm}
\subsection{Our Basic Scheme}
\label{OC}
Here, we first give a basic scheme by compressing the encodings to address the issue (1) discussed above. The computation and space complexity will be reduced from $\Theta(ln+l)$ to $\Theta(n+l)$ encodings. This scheme is not as secure as the naive one, but the amount of leaked information is limited and acceptable ( please see section \ref{sec5} for details).
In the naive scheme, the encoding pairs $(u_{i,0},v_{i,0})$ and $(u_{i,1},v_{i,1})$ for every rule are dependent on $W$ while the pair $(u_{n+1},v_{n+1})$ is dependent on $\vec{v}$. Now we give our basic scheme, in which the encoding pairs for each bit every rule are extracted and shared.
The obfuscating steps are summarized as Algorithm \ref{Basic_phase1}\footnote{For conciseness, we denote the parameter $params$ as `` $\cdot$ " in every primitive procedure in all our following algorithms. In addition, The $C$ we outsource do not contains $C_{\alpha_{i,\vec{v}\left[ i \right]}}$, but in the obfuscating phase, we use $C[P_r[i]]. C_{\alpha_{i,\vec{v}\left[ i \right]}}$ to denote the access for presentation convenience.} utilizing primitive procedures in GES. We first get $M+N$ encoding pairs units, of which the first $M$ units have equal random ratios in two encoding pairs while in the last $N$ units such ratios are different. Then we permute ($\sigma$) these encoding pairs units to shuffle them, and then collect two indexes sets:
\begin{align}
E&=\lbrace \sigma(1), \sigma(2),\ldots,\sigma(M)\rbrace,\\
UE&=\lbrace \sigma(M+1), \sigma(M+2),\ldots,\sigma(M+N)\rbrace.
\end{align}
Set $E$ collects the indexes of encoding pairs units with $\alpha_{i,0}= \alpha_{i,1}$ while set $UE$ collects those with $\alpha_{i,0}\neq \alpha_{i,1}$.
To obfuscate a rule $\textit{r} \triangleq (\vec{v},W,A)$, we attach an encoding pairs unit indexes array $p_r$ to \textit{r\'} such that if $i \in W$, then $p_r[i]$ is randomly sampled without replacement ($\overset{R}{\nleftarrow}$) from set $E$, otherwise it is randomly sampled without replacement from set $UE$. Thus $p_r$ specifies the $n$ encoding pairs units for rule $r$. Then $(u_{n+1},v_{n+1})$ for this rule is calculated according to Equation (\ref{Eq3}) by choosing encoding pairs in $C[P_r[i]]$ with $\vec{v}$. At last, public parameters $params'$ of GES and zero-test parameter $P_{zt}$ are also included in the obfuscated firewall \textit{f\'}. We emphasize that we should sample randomly without replacement rather than directly sampling with replacement. This trick is used to eliminate the possibility of false positive which may happen when two non-wildcard bits with different values chose the same encoding pair unit.
\begin{algorithm}[!t]
\caption{Execution with Outsourced Firewall \textit{f\'} }
\label{Basic_phase2}
\begin{algorithmic
\REQUIRE A packet $\vec{p}={\lbrace 0,1\rbrace}^n$, and obfuscated firewall \textit{f\'}.
\ENSURE Decisions on $\vec{p}$.
\FOR { \textit{r\'} $\triangleq$
$\big(P_r,u_{n+1},v_{n+1}, A\big) \in $ \textit{f\'} }
\STATE
$LHS=u_{n+1}\times \prod_{i \in \left[ n \right] } C[P_r[i]].v_{i,\vec{p}\left[ i \right]} $,
$RHS=v_{n+1}\times \prod_{i \in \left[ n \right] } C[P_r[i]].u_{i,\vec{p}\left[ i \right]}$.
\IF {$\mathnormal{isZero}(params',P_{zt},LHS - RHS)$}
\RETURN $A$ and exit.
\ELSE
\STATE \textit{Continue} checking $\vec{p}$ with the next rule.
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
In the second phase, the execution of outsourced firewall \textit{f\'} is straightforward. The basic steps are summarized in Algorithm \ref{Basic_phase2}. For an input packet $\vec{p}$, we check whether it satisfies an underlying rule in \textit{f\'} sequentially. Note that for each obfuscated rule \textit{r\'}, the multiplications are also performed on encodings chose from encoding pairs units $C$ specified by $\vec{p}$ and $P_r$.
Our scheme's correctness can be verified easily. In Algorithm \ref{Basic_phase1}, only two encodings are generated to a specific rule. The overheads of $P_r$ generation is negligible compared to encoding and re-randomization, thus it is easy to figure out the complexity of Algorithm \ref{Basic_phase1} is $\Theta(n+l)$. Thus the basic scheme is efficient in the obfuscation phase. The efficiency in execution phase is not bad, but needs further improvement.
\section{Online Efficiency Improvement}
\label{Improve}
In this section, we try to improve the efficiency and usability of our scheme in the online execution phase. As we have discussed previously, the online efficiency is vital for middlebox outsourcing scheme. The two enhanced schemes we propose here can reduce its complexity from $\Theta(n)$ to $\Theta(k)$ or $\Theta(n/k)$ on faster encodings' multiplication operations. The cost is $\Theta(1)$ increasement in encoding and re-randomization in the first phase.
The main guideline for our improvement is to reduce overheads of single multiplication and the times of multiplication. To this end, in both of the following approaches, we partition an original bit-wise rule into several parts. In the blocking approach, we then construct an encoding pair for an element in each part rather than for each rule bit, and thus we can check a packet in block level rather bit level. In the divide-and-conquer approach, we then test if a packet satisfies the rule in each part separately and sequentially with less encoding levels.
\subsection{Blocking Scheme}
We find that if we view a firewall rule in a higher level rather than the primitive bit level, there will be fewer occurrences of multiplication needed in online phase. If the block is small enough, the single multiplication will also take less time. To this end, we take another approach of firewall obfuscation by constructing in block level rather than bitwise level.
The basic steps in the first phase are described as algorithm \ref{practical_phase1}.
\begin{algorithm}[!t]
\caption{ Firewall Obfuscation (Blocking Approach)}
\label{practical_phase1}
\begin{algorithmic
\REQUIRE A original plain firewall \textit{f} with rule sets $\lbrace r\rbrace $.
\ENSURE A obfuscated firewall \textit{f\'} with the same functionality.
\STATE Generate parameters:
$(params,P_{zt})\leftarrow\mathnormal{InstGen}(1^\lambda, 1^{k+1})$.
\FOR { $\textit{r}\triangleq ({\lbrace F_i\rbrace}_{i \in [k]}, A) \in$ \textit{f}}
\FOR {$i \in \{1,\ldots,k\}$}
\STATE
$C_{\eta_i} \leftarrow \mathnormal{samp(\cdot)}$.
\FOR { $j \in I_i$}
\STATE
$C_{\beta_j} \leftarrow \mathnormal{samp(\cdot)}$
\IF{$j \in S_i$}
\STATE $C_{\alpha_j} = C_{\eta_i}$.
\ELSE
\STATE
$C_{\alpha_j} \leftarrow \mathnormal{samp(\cdot)}$.
\ENDIF
\STATE
$C_\gamma \leftarrow enc(\cdot,1,C_{\beta_j})$,
$C_\delta \leftarrow enc(\cdot,1,C_{\beta_j}\times C_{\alpha_j})$,
$u_{i,j} \leftarrow reRand(\cdot,1,C_\gamma)$,
$ v_{i,j} \leftarrow reRand(\cdot,1,C_\delta)$.
\ENDFOR
\ENDFOR
\STATE
$C_{\rho_{k+1}}\leftarrow \mathnormal{samp(\cdot)}$.
\STATE
$C_\epsilon \leftarrow enc(\cdot,1,C_{\rho_{k+1}})$,
$u_{k+1} \leftarrow reRand(\cdot,1,C_\epsilon)$,
$C_\theta \leftarrow enc(\cdot,1,C_{\rho_{k+1}}\times\prod_{i \in \left[ k \right] }C_{\eta_i})$,
$v_{k+1} \leftarrow reRand(\cdot,1,C_\theta)$.
\STATE \textit{r\'} $\triangleq
\big({{\lbrace ( u_{i,j},v_{i,j} )\rbrace}}_{i \in \left[ k \right], j \in I_i}, u_{k+1},v_{k+1}, A \big)$.
\ENDFOR
\RETURN \textit{f\'} $\triangleq \big(\lbrace r' \rbrace, P_{zt}, params' \big)$.
\end{algorithmic}
\end{algorithm}
Note that for brevity and highlighting the key idea, this algorithm is described based on the naive scheme. The encoding pairs compression idea in basic scheme can be easily integrated in algorithm \ref{practical_phase1}. We omit this integration here.
Specifically, in the first phase, for a firewall rule, we split it into $k$ fields, each with a small integer range. Take the second rule in table \ref{table1} as an example, we can split its IP source address header into four fields, each with the same domain $I_{1,2,3,4}=[0,255]$ (Just as its dotted decimal notation). In each field, the rule specifies a target filter set, such as $F_{1}={192},F_{2}={168},F_3={45}$ and $F_4=[0,255]$ for the above example. Another advantage of this approach is that it enables us to process some rare extended ACLs with port range. For example, for the destination port range in the rule \#3 in table \ref{table1}, we can directly set the corresponding target filter set as $[22,88]$. So now a rule can be presented as $\textit{r}\triangleq ({\lbrace F_i\rbrace}_{i \in [k]}, A)$.
After this partition, for each field $i$ within the domain $I_i$, we ``encrypt" each element in it with a level-1 encoding pair. The key point is that, for all elements that are in the filter set $F_i$, the underlying ratios of the two encodings in their encoding pairs are all equal to a constant number $\eta_i$, which is specified for each field in this obfuscation. Finally, just similar to basic scheme, these ratios in each field are aggregated in a product of them, which is also hidden as a ratio in another level-1 encoding pair $(u_{k+1},v_{k+1})$:
\begin{equation}
(u_{k+1},v_{k+1}) \in (S^{\rho_{k+1}}_1,S^{\rho_{k+1} \times \prod_{i \in \left[ k \right] }\eta_i}_1)
\end{equation}
Then, in the second phase, as a consequence of the simplification in the first phase, the burden of the cloud firewall provider will be greatly relieved. For a packet $\vec{p}$, cloud firewall provider also splits its packet header into $k$ blocks, which means that we can present it in the format of an integer tuple $(p_1,\dots,p_k)$. Then an encodings pair in each field is chosen according to the integer tuple value followed by the computation of two level-($k+1$) encodings,$LHS$ and $RHS$, by multiplying them. At last the equality of $LHS$ and $RHS$ is tested and thus a decision upon this packet can be made. The modification of algorithm 2 to coordinate with algorithm 3 is easy, so we leave out the details here.
\subsection{Divide-and-Conquer Scheme}
To reduce overheads of a single multiplication, we can decrease the maximum levels in GES. So another strategy we can take is divide-and-conquer. We divide the whole bit-string rule $(\vec{v},W, A)$ into $k$ parts: $(\vec{v},W,A)=(\lbrace\vec{v}_i,W_i\rbrace_{1\leq i\leq k},A)$. And in each part, we construct a obfuscated $r'_i$ in the same way as algorithm 1. Thus the corresponding obfuscated firewall will be $f'\triangleq \big( \big \lbrace {\lbrace r'_i \rbrace}_{1\leq i\leq k} \big\rbrace, {\lbrace P_{zt}^i\rbrace}_{1\leq i\leq k}, params'\big)$. Similarly, in the second phase, we also partition the packet in the same way, and get $\vec{p}=(\vec{p}_{i})_{1\leq i\leq k}$. Then we can check whether $\vec{p}$ matches the underlying $r$ by checking $\vec{p}_{i}$ with $r'_i$ in each part sequentially with algorithm 2.
This approach is simple, but effective, as will be demonstrated in our experiments. It is not hard to modify the basic scheme in this approach, so we omit the details here.
\section{privacy analysis}
\label{sec5}
In this section, we will take a close look at the security of the privacy preserving schemes we proposed.
\subsubsection{Analysis of the basic scheme}
In \cite{zvika2013obfus}, the authors have proved the naive scheme is secure when the inputting function is drawn from a distribution with a high superlogarithmic entropy given the GXDH assumption and the GCAN assumption \cite{zvika2013obfus} are true or in the \textit{generic model} where the adversary is only allowed to manipulate encodings via oracle access.
When parameters $M$ and $N$ are large enough, our basic scheme achieves the same security as the naive scheme and the security proof is roughly the same as proofs provided in \cite{zvika2013obfus}. Due to both completeness considerations and the page limit, below we present the proof sketch for the high entropy input distribution case only and neglect the proof in the generic model.
\begin{theorem}\label{thm:alg12sec}
When $M$ and $N$ are large enough so that all $P_r[i]$ are different, Algorithms 1 and 2 are secure when $(\vec{v},W)$ is drawn from a distribution where the entropy of $\vec{v}$ given $W$ is superlogrithmic given the GXDH assumption and the GCAN assumption are true.
\end{theorem}
\begin{proof}
Same as \cite{zvika2013obfus}, by claiming Algorithm 1 and 2 are secure we mean that, for each rule $r$, anything that the cloud firewall provider or the adversary learns from the execution of the two algorithms can be learned from a blackbox firewall by constructing a simulator $\mathcal{S}$ that simulates the adversary's view from algorithms 1 and 2 given only the access of a blackbox firewall.
In particular, the adversary's view is the obfuscated firewall output by Algorithm 1:$\big( {\lbrace C[i] \rbrace}_{i \in [M+N]}, \lbrace r' \rbrace, P_{zt}, params'\big)$.
$\mathcal{S}$ simulates $ {\lbrace C[i] \rbrace}_{i \in [M+N]}$ with $4(M+N)$ uniformly random level-1 encodings: $(\{u'_{i,0},v'_{i,0}, u'_{i,1},v'_{i,1}\}_{i\in[M+N]})$. $S$ first runs \textit{InstGen} and get a random $P'_{zt}$ and random public parameters $params''$. For each $r'$, $S$ can simulate $P_r$ with $n$ different random samplings in $[M+N]$ , simulate $(u_{n+1},v_{n+1})$ with two uniformly random level-1 encodings: $(u'_{n+1},v'_{n+1})$, simulates $A$ with $A$ itself, and simulates $P_{zt}$ and $params'$ with $P'_{zt}$ and $params''$. Due to the GCAN assumption and the entropy of $\vec{v}$ given $W$ is superlogrithmic, we know $u_{n+1},v_{n+1}$ is indistinguishable from uniform thus is independent with any pair in $\{C[i]\}$. In addition, due to the GXDH assumption, we know each $C[i]$ is indistinguishable from $(u'_{i,0},v'_{i,0}, u'_{i,1},v'_{i,1})$, and $(u_{n+1},v_{n+1})$ is indistinguishable from $(u'_{n+1},v'_{n+1})$. Thus we know $\mathcal{S}$ simulates the algorithm 1's output for ($(\vec{v},W)$) input.
For algorithm 2's output, the $S$ can simply use the access of the blackbox firewall to get it.
\end{proof}
When $M$ and $N$ are not large enough, Algorithm 1 reveal partial information about the rules. The leakage could happen
when for different rules $r_1$ and $r_2$, there exist $i,j\in \{1,\ldots,n\}$ such that $P_{r_1}[i]=P_{r_2}[j]$.
In this case, the adversary would notice similar ``wildcard or not'' information between two bit locations in two rules.
We can figure out that the probability for such leakage is
$1-\big( \frac{\binom{M}{|W_1|} \cdot \binom{M-|W_1|}{|W_2|}} {\binom{M}{|W_1}\cdot \binom{M}{|W_2|}} \big) \cdot
\big( \frac{\binom{N}{m_1} \cdot \binom{N-m_1}{m_2}} {\binom{N}{m_1}\cdot \binom{N}{m_2}} \big)
= 1- \big( \frac{(M-|W_1|)!\cdot (M-|W_2|)!}{M!\cdot (M-|W_1|-|W_2|)!}\big) \cdot
\big( \frac{(N-m_1)!\cdot (N-m_2)!}{N!\cdot (N-m_1-m_2)!}\big)
$, in which the $W_1$, $W_2$ is the wildcard bits set in $r_1$ and $W_2$ each, and $m1=n-|W_1|, m2=n-|W_2|$. So if such leak is unacceptable, we may increase the value of $M$ and $N$ in practice.
\subsubsection{Analysis of the Blocking Scheme}
The main difference between the basic and blocking scheme is that Algorithm 3 ``encrypt"s each integer in each field rather than ``0" or ``1" on each bit.
For similar reasons as Algorithm 1, we also give a proof sketch for Algorithm 3's security in the case of high entropy inputs and leave out the detailed proofs. Note that Algorithm 3 itself achieves the same security as the naive scheme's security. So when integrated with basic scheme, the blocking scheme is as secure as Algorithm 1.
\begin{theorem}Algorithm 3 is secure when $(\vec{v},W)$ is drawn from a distribution where the entropy of $\vec{v}$ given $W$ is superlogrithmic given the GXDH assumption and the GCAN assumption are true.
\end{theorem}
\begin{proof}
The proof goes similarly as the Proof of Theorem \ref{thm:alg12sec}. The only difference here is $\mathcal{S}$ simulates algorithm 3's output $\big({{\lbrace ( u_{i,j},v_{i,j} )\rbrace}}_{i \in \left[ k \right], j \in I_i}, u_{k+1},v_{k+1}\big)$ with $2\Sigma^k_1{|I_i|}+2$ uniformly random level-0 encodings: $(\{u'_{i,j},v'_{i,j}\}_{i\in[n],j\in I_i}, u'_{k+1},v'_{k+1})$. Due to the GXDH assumption, we know for each $i\in[n]$, $\{<u'_{i,j},v'_{i,j}>\}_{j\in I_i}$ is still indistinguishable with $\{<u_{i,j},v_{i,j}>\}_{j\in I_i}$. Due to the GCAN assumption and the entropy assumption, we know each $\{<u_{i,j},v_{i,j}>\}_{j\in I_i}$ and $(u_{k+1},v_{k+1})$ are independent with each other, and $(u_{k+1},v_{k+1})$ is indistinguishable with $(u'_{k+1},v'_{k+1})$. Therefore, $\mathcal{S}$ simulates algorithm 1's output for ($(\vec{v},W)$) input.
\end{proof}
\subsubsection{Analysis of the Divide-and-Conquer Scheme}
The divide-and-conquer scheme break the matching of one rule into several $k$ parts. It is easy to see it has the same amount of security except it reveals the matching results of all parts in addition.
\section{Experiments}
\label{sec6}
To evaluate the feasibility and performance of SOFA, we implement our algorithms in C++, and carry out experiments on real-world firewall rules. The cryptographic multilinear map we employ is the one constructed over integers\cite{coron2013practical, mmapcode}.
\subsection{Experiment Setup}
Our experiments are conducted on a PC running Linux with 2.90 GHz \textit{Inter(R) Core(TM) i7} CPU and 8 GB RAM. And we use 50 standard real-world firewall rules to evaluate the scalability of our algorithms. Note that our blocking scheme can also deal with the extended rules with minor modification. To leave out unnecessary details, these rules are preprocessed such that only the fields in the IP/TCP header are included in $r$. In reality, many firewall rules belong to this type.
The generation of an instance of a practical multilinear map involves the configuration of many parameters\cite{coron2013practical}. And the configuration will affect its performance and usability, which is related to the control of the ``noise". So it should be tuned with specific application. Thus we do not present the tedious detailed parameters of our experiments here. Instead our settings ensure the ``noise" in our experiments is under control such that no correctness will be sacrificed for efficiency.
\begin{figure}[!t]
\centering
\includegraphics[height=1.7in,width=2.5in]{prim}
\caption{ Computation overheads of primitive procedures in GES}
\label{primitive}
\end{figure}
\subsection{Primitive Cryptographic Overheads}
The procedures of GES that we have defined are the basic steps in our algorithms, so we first measure the efficiency of these primitive cryptographic operations. We set the multilinearity level $\kappa$=5, 9 and 33, which are the parameters for obfuscating a standard rule each for the basic scheme, the divide-and-conquer scheme and the one with blocking approach. The results are shown in figure \ref{primitive}.
We can observe from the results that the \textit{ring sampler} takes only less than 1 millisecond. The \textit{encoding and re-randomization} procedure are always performed together, so we measure them together, and we can observe that the multilinearity level affects its running time greatly.
The \textit{ multiplication} procedure is the major operation in the second phase, and it can be seen that its efficiency is also strongly affected by the multilinearity level, which determines the specific system parameters. We can also see that the \textit{zero testing} procedure is relatively lightweight in computation overheads.
\subsection{Performance in Obfuscation Phase}
The first phase, the obfuscation of original firewall rules, is performed locally by the enterprise SDN controller. We first measure the impact of the bit numbers $n$ (note that the corresponding multilinearity level in GES is $n+1$) on the performance of naive scheme (Fig.\ref{fig_first_case}). The time overheads of procedure \textit{Instance Generation} and the obfuscating time of a single standard rule are measured separately. From Fig.\ref{fig_first_case}, we can see that both of them grow linearly with respect to the bit size. Specifically, when $n=32$ ( which we need when naive scheme is adopted to protect the standard rules), the first takes about 0.8 seconds, while the obfuscating for a rule takes about 4.1 seconds. This shows the inefficiency of naive scheme.
\begin{figure}[!t]
\centering
\subfloat[Naive Scheme with Various Bits]{\includegraphics[width=1.7in]{Bit}%
\label{fig_first_case}}
\hfil
\subfloat[ Naive and Basic scheme]{\includegraphics[width=1.7in]{naive-basic}%
\label{fig_second_case1}}
\hfil
\subfloat[ Divide-and-Conquer Scheme]{\includegraphics[width=1.7in]{divide-and-conquer}%
\label{fig_second_case2}}
\hfil
\subfloat[ Blocking Scheme]{\includegraphics[width=1.7in]{blocking}%
\label{fig_second_case3}}
\caption{ Time Performance in the First Phase}
\label{obfuscation}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[height=2.0in,width=2.5in]{p2}
\caption{Time Performance in the Second Phase}
\label{execution}
\end{figure}
And we also measure the scalability of the schemes we proposed (Fig.\ref{fig_second_case1}, Fig.\ref{fig_second_case2}, Fig.\ref{fig_second_case3}) and the effect of basic scheme on other approaches. The results show that by the encodings compression of our basic scheme, the obfuscation of original rules can be greatly accelerated. This is in line with our theoretical analysis. These results also show that our framework is elastic with different number of rules.
\subsection{Performance in Execution Phase}
The performance in the second phase, the real-time execution of the obfuscated firewalls in the cloud, is more crucial. We measure our solution's performance in this phase with different number of rules, the result is shown in Fig.\ref{execution}.
From the result we can see that the blocking scheme is much more efficient. Actually, we find that the running time to check a packet with them is separately 300, 100, 6 milliseconds. Considering that our experiments are performed on a ordinary PC and no modern firewall process acceleration techniques (such as FDD) are adopted, our schemes, especially blocking scheme, are practical. And in practice, when adopted in cloud processing, many parallelism and acceleration means can be taken to make our scheme faster.
\section{Conclusion}
\label{sec7}
In this paper, we propose solutions to make it secure for an enterprise to outsource firewalls. The security of our schemes have solid theoretical foundation.
And the experiment results show that our schemes are also practical to use.
There are more other network functionalities, such as IDS, IPS, that can be outsourced, and their privacy issues must be considered too when outsourcing. We think the framework we propose also works for them because many of them are also based on packet header filtering. But since their configurations are different from firewalls, so the specific solutions to them will also vary. We leave this as our future work.
\bibliographystyle{IEEEtran}
|
\section*{Introduction}\noindent
This paper is about the bar recursion operator \cite{Spc}, in the context of classical realizability \cite{JLK2,JLK3}. It is a sequel to
the three papers~\cite{BBC,BO,Str}. We use the definitions and notations of the theory of classical realizability
as expounded in~\cite{JLK1,JLK2,JLK3}.\\
In \cite{BBC}, S.~Berardi, M. Bezem and T. Coquand have shown that a form of the bar recursion operator can be used,
in a proof-program correspondence, to interpret {\em the axiom of dependent choice} in proofs of $\Pi^0_2$-formulas
of arithmetic. Their work was adapted to the theory of domains by U. Berger and P. Oliva in \cite{BO}.
In \cite{Str}, T. Streicher has shown, by using the bar recursion operator of \cite{BO}, that the models of ZF\/,
associated with realizability algebras~\cite{JLK1,JLK3} obtained from usual models of $\lbd$-calculus
(Scott domains, coherent spaces,~\ldots), satisfy the axiom of dependent choice.\\
We give here a proof of this result, but for a realizability algebra which is built following the presentation
of~\cite{BBC}, which we call the BBC-algebra.\\
In section~\ref{BBCalg}, we define and study this algebra~; we define also the bar recusion operator,
which is a closed $\lbd$-term.\\
In sections~\ref{chden} and~\ref{chdep}, which are very similar, we show that this operator
realizes the axiom of countable choice (CC), then the axiom of dependent choix (DC). The proof is
a little simpler for CC.\\
In section~\ref{WOR}, we deduce from this result, using results of \cite{JLK4} that, in the model of ZF associated
with this realizability algebra,
\emph{every real (more generally, every sequence of ordinals) is constructible}.\\
The formulas ``$\mathbb{R}$\emph{ is well ordered}''\ and ``\emph{Continuum hypothesis}'' are therefore
realized, in these models, by closed $\lbd_c$-terms (i.e. $\lbd$-terms containing the control instruction
$\ccc$ of Felleisen-Griffin).\\
We show also that \emph{every true formula of analysis is realized by a closed $\lbd_c$-term}.\\
In this way, we show how to obtain a program (closed $\lbd_c$-term) from any proof of a $\Pi^0_2$ arithmetical
formula in the theory \ ZF + ``Dependent choice'' + ``Every real is constructible'' (and therefore
``Well ordering of $\RR$'' and ``Continuum hypothesis'').
\section{The BBC realizability algebra}\label{BBCalg}\noindent
The definition and general properties of realizability algebras are given at the beginning of~\cite{JLK1}. In particular,
closed $\lbd$-terms are interpreted as terms in these algebras.
\smallskip\noindent
The BBC realizability algebra $\gB=(\LLbd,\PPi,\bbot)$ is defined as follows~:
\smallskip\noindent
$\bullet$~~The set of \emph{processes} $\LLbd\star\PPi$ is $\LLbd\fois\PPi$.
\smallskip\noindent
$\bullet$~~The set of \emph{terms} $\LLbd$ is the smallest set which contains the following
\emph{constants of term}~:
\smallskip
$\BBB,\,\CCC,\,\III,\,\KKK,\,\WWW$ \emph{(Curry's combinators)}, $\ccc$ \emph{(Felleisen-Griffin instruction)},
$\AAA$ \emph{(abort instruction)}, $\pp,\qq_0,\ldots,\qq_N$ \emph{(variables)} where $N$ is a fixed integer~;
\smallskip\noindent
and is such that~:
if $\xi,\eta\in\LLbd$ then $(\xi)\eta\in\LLbd$ \ \emph{(application)}~;
with each sequence $\xi_i(i\in\NN)$ of closed elements of $\LLbd$ (i.e. which contain no variable
$\pp,\qq_0,\ldots,\qq_N$) \ is associated, in a one-to-one (and well founded) way, a constant of term
denoted by $\bigwedge_i\xi_i$.
\smallskip\noindent
Therefore, each term $\xi\in\LLbd$ is a finite sequence of constants of term and parentheses.\\
$\LLbd$ is defined by an induction of length $\aleph_1$ and is of cardinality $\aleph_1$.
\smallskip\noindent
{\bfseries Notations.}\\
The application $(\ldots((\xi_1)\xi_2)\ldots)\xi_n$ will be often written $(\xi_1)\xi_2\ldots\xi_n$
or even $\xi_1\xi_2\ldots\xi_n$.\\
The finite sequence $\qq_0,\ldots,\qq_N$ will be often written $\vec{\qq}$.
\smallskip\noindent
$\bullet$~~The set of \emph{stacks} $\PPi$ is defined as follows~: a stack $\pi$ is a finite sequence
$t_0\ps\ldots\ps t_{n-1}\ps\pi_0$ with $t_0,\ldots,t_{n-1}\in\LLbd$~; it is terminated by the symbol $\pi_0$ which
represents the \emph{empty stack}.
\smallskip\noindent
For each stack $\pi$, the \emph{continuation} $\kk_\pi$ is a term which is defined by recurrence~:\\
$\kk_{\pi_0}=\AAA$~; \ $\kk_{t\ps\pi}=\ell_t\,\kk_\pi$,
with $\ell_t=((\CCC)(\BBB)\CCC\BBB)t$ or $\lbd k\lbd x(k)(x)t$.\\
Thus, if the stack $\pi$ is $t_0\ps\ldots\ps t_{n-1}\ps\pi_0$, we have~:\\
\centerline{$\kk_\pi=(\ell_{t_0})\ldots(\ell_{t_{n-1}})\AAA$ or $\lbd x(\AAA)(x)t_0\ldots t_{n-1}$.}
\smallskip\noindent
The \emph{integer} $\ul{n}$ is defined as follows~:\\
\centerline{$\ul{0}=(\KKK)\III$ or $\lbd x\lbd y\,y$~; $\ul{n+1}=(\sig)\ul{n}$
with $\sig=(\BBB\WWW)(\CCC)(\BBB)\BBB\BBB$ or $\lbd n\lbd f\lbd x(f)(n)fx$.}
\smallskip\noindent
The relation of \emph{execution} $\succ$ is the least preorder on $\LLbd\star\PPi$ defined by the following rules
(with $\xi,\eta,\zeta\in\LLbd,\pi\in\PPi$ and $n\in\NN$)~:
\smallskip\noindent
1.~~$(\xi)\eta\star\pi\succ\xi\star\eta\ps\pi$~; \ \emph{(push)}\\
2.~~$\BBB\star\xi\ps\eta\ps\zeta\ps\pi\succ\xi\star(\eta)\zeta\ps\pi$~; \ \emph{(apply)}\\
3.~~$\CCC\star\xi\ps\eta\ps\zeta\ps\pi\succ\xi\star\zeta\ps\eta\ps\pi$~; \emph{(switch)}\\
4.~~$\III\star\xi\ps\pi\succ\xi\star\pi$~; \ \emph{(no operation)}\\
5.~~$\KKK\star\xi\ps\eta\ps\pi\succ\xi\star\pi$~; \ \emph{(delete)}\\
6.~~$\WWW\star\xi\ps\eta\ps\pi\succ\xi\star\eta\ps\eta\ps\pi$~; \ \emph{(copy)}\\
7.~~$\ccc\star\xi\ps\pi\succ\xi\star\kk_\pi\ps\pi$~; \ \emph{(save the stack)}\\
8.~~$\AAA\star\xi\ps\pi\succ\xi\star\pi_0$~; \emph{(abort)} or \emph{(delete the stack)}\\
9.~~$\bigwedge_i\xi_i\star\ul{n}\ps\pi\succ\xi_n\star\pi$~; \ \emph{(oracle)}
\smallskip\noindent
When $\xi,\eta\in\LLbd$, we set $\xi\succ\eta$ iff $(\pt\pi\in\PPi)(\xi\star\pi\succ\eta\star\pi)$.
\smallskip\noindent
$\bullet$~~\emph{Proof-like terms}.\\
Let $\PL_0$ be the countable set of terms built with the constants $\BBB,\,\CCC,\,\III,\,\KKK,\,\WWW,\,\ccc$ \
and the application. It is the smallest possible set of \emph{proof-like terms}.\\
We shall also consider the set $\PL$ of \emph{closed terms} (i.e. with no occurrence of $\pp,\vec{\qq}$)
which is of cardinality $\aleph_1$.
\smallskip\noindent
$\bullet$~~\emph{Execution of processes~; definition of $\bbot$}.\\
For every process $\xi\star\pi$, at most one among the rules 1 to 9 applies. By iterating these rules,
we obtain the \emph{reduction} or the \emph{execution} of the process $\xi\star\pi$. This execution stops
if and only if the stack is insufficient (rules 2 to 8) or does not begin with an integer (rule 9) or else
if the process has the form $\pp\star\varpi$ or $\qq_i\star\varpi$.
\smallskip\noindent
Finally, we set $\bbot=\{\xi\star\pi\in\LLbd\star\PPi\;;\;(\ex\varpi\in\PPi)(\xi\star\pi\succ\pp\star\varpi)\}$.
\begin{lemma}
$\gB$ is a coherent realizability algebra.
\end{lemma}\noindent
\begin{proof}
\emph{$\gB$ is a realizability algebra}~:\\
It remains to check that $\kk_\pi\star\xi\ps\varpi\succ\xi\star\pi$, which is done by recurrence on $\pi$~:\\
if $\pi=\pi_0$, it is rule~8~;\\
if $\pi=t\ps\rho$ we have $\kk_\pi\star\xi\ps\varpi=\kk_{t\ps\rho}\star\xi\ps\varpi=\ell_t\kk_\rho\star\xi\ps\varpi
\succ(\kk_\rho)(\xi)t\star\varpi\succ\kk_\rho\star\xi t\ps\varpi\succ\xi t\star\rho$
(recurrence hypothesis) $\succ\xi\star t\ps\rho$.
\smallskip\noindent
\emph{$\gB$ is coherent}~:\\
If $\theta\in\PL$ then $\theta\star\pi_0\notin\bbot$~; indeed, $\,\pp$ does not appear
during the execution of $\theta\star\pi_0$.
\end{proof}
\subsection*{Models and functionals}
A coherent realizability algebra is useful in order to give \emph{truth values} to formulas of ZF\/. In fact, we use
a theory called \ZFe\ \cite{JLK2} which is a conservative extension of ZF\/.
This theory has an additional strong membership relation symbol $\veps$ \emph{which is not extensional}.
\smallskip\noindent
For each closed formula $F$ of \ZFe, we define \emph{two truth values}, denoted $\|F\|$ and $|F|$, with $\|F\|\subset\PPi$ and
$|F|\subset\LLbd$, with the relation $\xi\in|F|\Dbfl(\pt\pi\in\|F\|)(\xi\star\pi\in\bbot)$.\\
The relation $\xi\in|F|$ is also written $\xi\forces F$ and reads \emph{``the term $\xi$ realizes the formula $F$''.}
\smallskip\noindent
All the necessary definitions are given in~\cite{JLK1,JLK2,JLK3}.
\smallskip\noindent
The following lemma~\ref{eta_eq} is a useful property of the BBC realizability algebra $\gB$.
\begin{lemma}\label{eta_eq}
For all formulas $A,B$ of \ \ZFe, and all terms $\,\xi\in\LLbd$, we have~:\\
$\xi\forces A\to B$ \ iff \ $(\pt \eta\in\LLbd)(\eta\forces A\Fl\xi\eta\forces B)$.
\end{lemma}\noindent
Indeed, by the general definition of $\forces$, we have~:\\
\centerline{$(\xi\forces A\to B)\Dbfl(\pt \eta\forces A)(\pt\pi\in\|B\|)(\xi\star\eta\ps\pi\in\bbot)$.}
\noindent
Now, by the above definition of $\bbot$, it is clear that $(\xi\star\eta\ps\pi\in\bbot)\Dbfl(\xi\eta\star\pi\in\bbot)$
from which the result follows.
\qed
\smallskip\noindent
Classical realizability is an extension of forcing. As in forcing, we start with an ordinary model
${\cal M}$ of ZFC (or even ZF + V = L) which we call the \emph{ground model},
and we build a \emph{realizability model} ${\cal N}$ which satisfies \ZFe\ in the following sense~:\\
${\cal M}$ and ${\cal N}$ have the same domain, but neither the same language, nor the same truth values.
The language of ${\cal N}$ has the additional binary symbol $\veps$ of \emph{strong membership}. The truth
values of ${\cal N}$ are not $0,1$ as for ${\cal M}$, but are taken in ${\cal P}(\PPi)$
endowed with a suitable structure of Bolean algebra \cite{JLK1,JLK3}. We say that \emph{${\cal N}$ satisfies
a formula $F$} iff there is a proof-like term $\theta$ which realizes $F$ or equivalently\/, if the truth
value $\|F\|$ of $F$ is the unit of the Boolean algebra ${\cal P}(\PPi)$.
\smallskip\noindent
A \emph{functional} on the ground model ${\cal M}$ is a formula $F(\vec{x},y)$ of ZF with parameters
in ${\cal M}$, such that ${\cal M}\models\pt\vec{x}\,\ex!y\,F(\vec{x},y)$. Denoting such a functional
by $f$, we write $y=f(\vec{x})$ for $F(\vec{x},y)$.
\smallskip\noindent
Since ${\cal M}$ and ${\cal N}$ have the same domain, all the functionals defined on ${\cal M}$ are also
defined on ${\cal N}$ and \emph{they satisfy the same equations} and even the same \emph{Horn formulas}
i.e. formulas of the form $\pt\vec{x}(f_1(\vec{x})=g_1(\vec{x}),\ldots,f_n(\vec{x})=g_n(\vec{x})\to
f(\vec{x})=g(\vec{x}))$.
\smallskip\noindent
A particularly useful binary functional on ${\cal M}$ (and thus also on ${\cal N}$) is the \emph{application},
denoted by app, which is defined as follows~: \ $\mbox{app}(f,x)=\{y\;;\;(x,y)\in f\}$.\\
We shall often write $f[x]$ for $\mbox{app}(f,x)$. This allows to consider each set in ${\cal M}$
(and in ${\cal N}$) as a unary functional.
\smallskip\noindent
{\small{\bfseries{}Remark.}
We can define a set $f$ in ${\cal M}$ by giving $f[x]$ for every $x$, provided that there exists
a set $X$ such that $f[x]=\vide$ for all $x\notin X$~: take $f=\bigcup_{x\in X}\{x\}\fois f[x]$.\\
In the ground model ${\cal M}$, every function is defined in this way but in general,
\emph{this is false in ${\cal N}$}.}
\subsection*{Quantifiers restricted to $\NN$}\noindent
In \cite{JLK3}, we defined the quantifier $\pt x\indi$, by setting~:\\
$\|\pt x\indi F[x]\|=\bigcup_{n\in\NN}\|\{\ul{n}\}\to F[n]\}\|=\{\ul{n}\ps\pi\;;\;n\in\NN,\pi\in\|F[n]\|\}$,
so that we have~:\\
$\xi\forces\pt x\indi F[x]\;\Dbfl\;\xi\ul{n}\forces F[n]$ for all $n\in\NN$~;\\
and it is shown that it is a correct definition of the restricted quantifier to $\NN$.\\
Indeed the equivalence \ $\pt x\indi F[x]\dbfl\pt x(\mbox{int}[x]\to F[x])$ is realized by a closed
$\lbd$-term independent of $F$, called a \emph{storage operator}.\\
The formula $\mbox{int}[x]$ is any formula of ZF which says that $x$ is an integer.
\begin{theorem}\label{omega-modele}
If we take $\PL$ for the set of proof-like terms, and if the ground model ${\cal M}$ is transitive and
countable, then there exists a countable realizability model ${\cal N}$ which has only standard integers,
i.e. which is an $\omega$-model.
\end{theorem}\noindent
Let ${\cal T}$ be the theory formed with closed formulas, with parameters in ${\cal M}$, which are
realized by a proof-like term. ${\cal T}$ is $\omega$-complete~: indeed, if $\theta_n\in\PL$ and
$\theta_n\force F[n]$ for $n\in\NN$, let us set $c=\bigwedge_i\theta_i$. Then $c\ul{n}\force F[n]$
for all $n\in\NN$ and therefore $c\force\pt n\indi F[n]$, i.e. $\pt n\indi F[n]\in{\cal T}$.
It follows that ${\cal T}$ has a countable $\omega$-model.
\qed
\begin{proposition}\ \\
Let $f:\NN\to2$ and $\theta\in\PL$, $\theta\forces\ex n\indi(f(n)=1)$.
Then $\;\theta\star\pp\ps\pi_0\succ\pp\star\ul{n}\ps\varpi$ with $f(n)=1$.
\end{proposition}\noindent
There exists $\tau\in\LLbd$ such that $\tau\ul{n}\succ\pp$ if $f(n)=1$ and $\tau\ul{n}\succ\qq_0$ if $f(n)=0$~:
set $\tau=\lbd x(\bigwedge_i\xi_i)x\,\pp\qq_0$ with $\xi_n=\KKK$ if $f(n)=1$ and $\xi_n=\KKK\III$ if $f(n)=0$.\\
Then we have $\tau\forces\pt n\indi(f(n)\ne1)$ and therefore $\theta\tau\forces\bot$. We necessarily have~:\\
$\theta\star\tau\ps\pi_0\succ\tau\star\ul{n}\ps\pi$ for some $n$~; furthermore, we have $\tau\ul{n}\succ\pp$,
otherwise we should have $\tau\ul{n}\succ\qq_0$, and thus $\;\theta\star\tau\ps\pi_0\notin\bbot$.
Therefore $f(n)=1$.
\qed
\smallskip\noindent
{\small{\bfseries{}Remark.}
This shows that, from any proof-like term which realizes a given $\Sigma^0_1$ arithmetical formula, we obtain
a program which computes an integer satisfying this formula. Such a realizer is given by any proof of this formula by means of axioms which have themselves such realizers.\\
The theory of classical realizability gives realizers for the axioms of ZF\/. We show below that the bar recursion operator
realizes the axiom of dependent choice. Finally, in section~\ref{WOR}, we get (rather complicated) proof-like realizers
for the axioms ``$\RR$ is well ordered'' and ``Continuum hypothesis''.}
\subsection*{Execution of processes}
{\bfseries Notation.} If $\pi=t_0\ps\ldots\ps t_{n-1}\ps\pi_0$, we shall write
$\pi\ps t$ for $t_0\ps\ldots\ps t_{n-1}\ps t\ps\pi_0$.\\
Thus, we obtain $\kk_{\pi\ps t}$ by replacing, in $\kk_\pi$, the last occurrence of $\AAA$ by $\ell_t\AAA$.
\begin{lemma}\label{pi-ps-t}
If $\,\xi\star\pi\in\bbot$, then $\,\xi'\star\pi'\in\bbot$ and $\,\xi'\star\pi'\ps t\in\bbot$, where $\,\xi'\star\pi'$ is obtained by replacing, in $\,\xi\star\pi$, some occurrences of $\,\AAA$ by $(\ell_u)\AAA=\kk_{u\ps\pi_0}$ and some occurrences of the variabless
$\,\qq_0,\ldots,\qq_N$ by $\,t_0,\ldots,t_N$~; $t_0,\ldots,t_N,t,u$ are arbitrary terms.
\end{lemma}\noindent
{\small{\bfseries{}Remark.}
In particular, it follows that $\,\xi\star\pi_0\in\bbot\Fl\xi\forces\bot$.}
\smallskip\noindent
Proof by recurrence on the length of the execution of $\xi\star\pi\in\bbot$ by means of rules 1 to~9.
We consider the last used rule. There are two non trivial cases~:
\smallskip\noindent
$\bullet$~~Rule~7 (execution of $\ccc$)~; we must show
$\ccc\star\xi'\ps\pi'\ps t\in\bbot$.\\
We apply the recurrence hypothesis to $\,\xi\star\kk_\pi\ps\pi$, in which we replace~:\\
- $\pi_0$ by $t\ps\pi_0$ (thus $\pi$ becomes $\pi\ps t$)~;\\
- the last occurrence of $\AAA$ in $\kk_\pi=(\ell_{t_0})\ldots(\ell_{t_{n-1}})\AAA$ by $(\ell_t)\AAA$
(thus $\kk_\pi$ becomes $\kk_{\pi\ps t}$).\\
Then, we make the substitutions in $\xi,\pi$, which gives $\xi'\star\kk_{\pi'\ps t}\ps\pi'\ps t$.
\smallskip\noindent
$\bullet$~~Rule~8 (execution of $\AAA$)~; we must show $(\ell_u)\AAA\star\xi'\ps\pi'\ps t\in\bbot$.\\
We apply the recurrence hypothesis to $\,\xi\star\pi_0$, which gives $\xi'\star u\ps\pi_0\in\bbot$, thus
$\xi' u\star\pi_0\in\bbot$ and therefore $\AAA\star\xi'u\ps\pi'\ps t\in\bbot$ (rule~8)~; finally, we obtain
$(\ell_u)\AAA\star\xi'\ps\pi'\ps t\in\bbot$.
\qed
\smallskip\noindent
In each process $\,\xi\star\pi\in\bbot$, we define an occurrence of $\pp$, which is called \emph{efficient},
by recurrence on the length of its reduction. If $\xi=\pp$, it is this very occurrence. Otherwise, we
consider the last rule used in the reduction, and the definition is clear~; for example, if it is rule~7,
and if the efficient occurrence in $\xi\star\kk_\pi\ps\pi$ is in $\kk_\pi$ or in $\pi$, then we take
the corresponding occurrence in $\ccc\star\xi\ps\pi$.
\begin{lemma}\label{efficace}
If $\;\xi\star\pi\in\bbot$, then~:
\begin{itemize}
\item $\xi'\star\pi'\in\bbot$, where $\xi'\star\pi'$ is obtained by substituting arbitrary terms for
the non efficient occurrences of $\,\pp$.
\item $\xi'\star\pi'\notin\bbot$ and indeed $\xi'\star\pi'\succ\qq_0\star\varpi$, where $\xi'\star\pi'$
is obtained by substituting $\;\qq_0$ for the efficient occurrence of $\,\pp$, and arbitrary terms for
the non efficient occurrences of $\,\pp$.
\end{itemize}
\end{lemma}\noindent
The proof is immediate, by recurrence on the length of the reduction of
$\xi\star\pi$ by means of rules 1 to 9~: consider the last used rule.
\qed
\begin{corollary}\label{gl2nt}\ \\
If $\xi\forces\top,\bot\to\bot$ and $\xi\forces\bot,\top\to\bot$, then $\xi\forces\top,\top\to\bot$,
and thus~:\\
$\lbd x(x)\III\,\III\,\forces\neg\pt x^{\gl2}(x\ne0,x\ne1\to\bot)$ and\\
$\WWW\,\forces\pt x^{\gl2}(\pt y^{\gl2}(y\ne0,y\ne x\to y\not\le x),x\ne0\to\bot)$.
\end{corollary}\noindent
{\small{\bfseries{}Remark.} These two formulas express respectively that the Boolean algebra $\gl2$ is
non trivial and that it is atomless.}
\smallskip\noindent
We apply lemma~\ref{efficace} to $\xi\star\pp\ps\pp\ps\pi_0$. We have $\xi\star\qq_0\ps\pp\ps\pi_0\in\bbot$
and $\xi\star\pp\ps\qq_0\ps\pi_0\in\bbot$, which shows that the efficient occurrence of $\pp$ is in $\xi$.
Therefore $\xi\star t\ps u\ps\pi_0\in\bbot$ for every $t,u\in\LLbd$, again by lemma~\ref{efficace}.\\
The last two assertions follow from the fact that~:\\
$\|\pt x^{\gl2}(x\ne0,x\ne1\to\bot)\|=\|\top,\bot\to\bot\|\cup\|\bot,\top\to\bot\|$ and therefore~:\\
$|\pt x^{\gl2}(x\ne0,x\ne1\to\bot)|=|\top,\top\to\bot|$.
\qed
\begin{theorem}\label{den_cont}
For every sequence $\xi_i\in\LLbd$ $(i\in\NN)$, there exists $\;\phi\in\LLbd$ such that~:\\
$\bullet$~~$\phi\ul{i}\succ\xi_i$ for every $i\in\NN$~;\\
$\bullet$~~for every $\;U\in\LLbd$ such that $\;U\phi\forces\bot$, there exists $k\in\NN$ such that
$\;U\psi\forces\bot$ for every $\;\psi\in\LLbd$ such that \ $\psi\ul{i}\succ\xi_i$ for every $i<k$.
\end{theorem}\noindent
{\small{\bfseries{}Remark.} Theorem~\ref{den_cont} will be used in order to show properties of the
bar recursion operator. In fact, the following weaker formulation is sufficient~:
\noindent
For every sequence $\xi_i\in\LLbd$ $(i\in\NN)$ and every $U\in\LLbd$ such that~:\\
\centerline{$(\pt k\in\NN)(\ex\psi\in\LLbd)\{U\psi\nforces\bot,(\pt i<k)(\psi\ul{i}\succ\xi_i)\}$}
\noindent
there exists $\phi\in\LLbd$ such that $U\phi\nforces\bot$ and $(\pt i\in\NN)(\phi\ul{i}\succ\xi_i)$.\\
In the particular case of forcing, this is exactly the \emph{decreasing chain condition}~:
every decreasing sequence of (non false) conditions has a lower bound (which is non false).}
\smallskip\noindent
We set $\eta_i=\lbd\pp\lbd\vec{\qq}\,\xi_i$~; thus, we have $\eta_i\in\PL$ and $\eta_i\pp\vec{\qq}\succ\xi_i$.\\
Let $\eta=\bigwedge_i\eta_i$ and $\phi=\lbd x(\eta)x\,\pp\vec{\qq}$.
Thus, we have $\eta\in\PL$ and $\phi\ul{i}\succ\xi_i$.\\
We may assume that $\eta$ does not appear in $U$.\\
We have $U\phi\forces\bot\Dbfl U\star\phi\ps\pi_0\in\bbot$ (lemma~\ref{pi-ps-t}).
During the execution of the process $U\star\phi\ps\pi_0$, the constant $\eta$ arrives in head position a finite
number of times, always through~$\phi$ (since it is deleted each time it arrives in head position),
therefore as follows~:\\
\centerline{$\eta\star\ul{i}\ps\pp\ps\vec{\qq}\ps\pi\succ\xi_i\star\pi$.}
\noindent
Let $k$ be an integer, greater than all the arguments of $\eta$ during this execution and let $\;\psi\in\LLbd$
be such that $\,\psi\ul{i}\succ\xi_i$ for all $i<k$. Let us set $\tau=\lbd x\lbd\pp\lbd\vec{\qq}\;\psi x$~;
thus, we have $\tau\ul{i}\,\pp\vec{\qq}\succ\,\psi\ul{i}\succ\xi_i$ for $i<k$.
In the process $U\star\phi\ps\pi_0$, let us replace the constant $\eta$ by the term $\tau$~; we obtain
$U\star\psi\ps\pi_0$. The execution is the same, and therefore $U\star\psi\ps\pi_0\in\bbot$ and $U\psi\forces\bot$.
\qed
\subsection*{The bar recursion operator}\noindent
We define below two \emph{proof-like terms} $\chi$ and $\Psi$ (which are, in fact, closed $\lbd$-terms).\\
In these definitions, the variables $i,k$ represent (intuitively) integers and the variable~$f$
represents a function of domain $\NN$, with arbitrary values in $\LLbd$.
\smallskip\noindent
$\bullet$~~We want a $\lbd$-term $\chi$ such that~:
\smallskip
\centerline{$\chi\ul{k}fz\ul{i}\succ f\ul{i}$ \ if \ $i<k$~; $\chi\ul{k}fz\ul{i}\succ z$ \ if \ $i\ge k$.}
\smallskip\noindent
Therefore, we set~:\\
\centerline{$\chi=\lbd k\lbd f\lbd z\lbd i((i\ppt k)(f)i)z$}
\noindent
where the boolean $(i\ppt k)$ is defined by~:
\noindent\centerline{
$(i\ppt k)=((kA)\lbd d\,\mathbf{0})(iA)\lbd d\,\mathbf{1}$}
\noindent
with $\mathbf{0}=\lbd x\lbd y\,y$ or $\KKK\,\III$, $\mathbf{1}=\lbd x\lbd y\,x$ or $\KKK$ and
$A=\lbd x\lbd y\,yx$ or $\CCC\,\III$.
\smallskip\noindent
The term $\chi\ul{k}f$ is a representation, in $\lbd$-calculus, of the finite sequence
$(f\ul{0},f\ul{1},\ldots,f\ul{k-1})$.
\smallskip\noindent
$\bullet$~~We want a $\lbd$-term $\Psi$ such that~:
\smallskip
\centerline{$\Psi gu\ul{k}f\succ(u)(\chi\ul{k}f)(g)\lbd z(\Psi gu\ul{k}^+)(\chi)\ul{k}fz$}
\smallskip\noindent
where $k^+=((\BBB\WWW)(\CCC)(\BBB)\BBB\BBB)k$ or $\lbd f\lbd x(f)(k)fx$ is the successor of the integer $k$.\\
Thus, we set~:
\smallskip\noindent
\centerline{$\Psi=\lbd g\lbd u(\Y)\lbd h\lbd k\lbd f(u)(\chi kf)(g)\lbd z(hk^+)(\chi)kfz$.}
\smallskip\noindent
where $\Y$ is the Turing fix point operator~:\\
\centerline{$\Y=XX$ \ with $X=\lbd x\lbd f(f)(x)xf=(\WWW)(\BBB)(\BBB\WWW)(\CCC)\BBB$.}
\smallskip\noindent
The term $\Psi$ will be called the \emph{bar recursion operator}.
\section{Realizing countable choice}\label{chden}\noindent
The \emph{axiom of countable choice} is the following formula~:
\smallskip\noindent
(CC)\hspace{7em} $\pt n\ex x\, F[n,x]\to\ex f\pt n\indi F[n,f[n]]$
\smallskip\noindent
where $F[n,x]$ is an arbitrary formula
of \ZFe(see~\cite{JLK2}), with parameters and two free variables.
The notation $f[n]$ stands for $\mbox{app}(f,n)$ (the functional app has been defined above).
\smallskip\noindent
{\small{\bfseries{}Remark.}
This is a strong form of countable choice which shows that, in the realizability model~${\cal N}$, every
countable sequence has the form $n\mapsto f[n]$ for some $f$. This will be used in section~\ref{WOR}.}
\begin{theorem}\label{CC}
$\lbd g\lbd u(\Psi)gu\,\ul{0}\,\ul{0}\,\forces\mbox{\rm CC}$.
\end{theorem}\noindent
The axiom of countable choice is therefore realized in the model of ZF associated with the BBC realizability algebra
(in fact, it is sufficient that the realizability algebra satisfies the property formulated in the remark following
theorem~\ref{den_cont}).
\smallskip\noindent
We write the axiom of countable choice as follows~:
\smallskip\noindent
(CC)\hspace{7em} $\pt n\neg\pt x\neg F[n,x],
\pt f\neg\pt n\indi F[n,f[n]]\to\bot$
\smallskip\noindent
Let $G,U\in\LLbd$ be such that \ $G\forces\pt n\neg\pt x\neg F[n,x]$ \ and \
$U\forces\pt f\neg\pt n\indi F[n,f[n]]$.\\
We set $H=\Psi GU$ and we have to show that $H\ul{0}\,\ul{0}\forces\bot$. In fact, we shall show that \
$H\ul{0}\xi\forces\bot$ for every $\xi\in\LLbd$.
\begin{lemma}\label{rec-Psi2}
Let $k\in\NN$ and \ $\phi\in\LLbd$ be such that \ $(\pt i<k)\ex a_i(\phi \ul{i}\forces F[i,a_i])$.\\
If $H\ul{k}\,\phi\nforces\bot$, then there exist a set $a_k$ and a term $\;\zeta_{k,\;\phi}\in\LLbd$ such that~:\\
\centerline{$\zeta_{k,\;\phi}\forces F[k,a_k]$ \ and \ $(H\ul{k}^+)(\chi)\ul{k}\,\phi\,\zeta_{k,\;\phi}\nforces\bot$.}
\end{lemma}\noindent
Define $\eta_{k,\;\phi}=\lbd z(H\ul{k}^+)(\chi)\ul{k}\,\phi z$, so that
$H\ul{k}\,\phi\succ(U)(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}$.\\
If \ $\eta_{k,\;\phi}\forces\pt x\neg F[k,x]$ \ then, by hypothesis on $G$, we have $G\eta_{k,\;\phi}\forces\bot$.
Let us check that~:\\
\centerline{$(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}\forces\pt n\indi F[n,f_k[n]]$}
\noindent
where $f_k$ is defined by~: \ $f_k[i]=a_i$ if $i<k$ (i.e. $i\in k$)~; $f_k[i]=\vide$ if $i\notin k$.\\
Indeed, if we set \ $\phi'=(\chi\ul{k}\phi)(G)\eta_{k,\;\phi}$, we have~:\\
$\phi'\ul{i}\succ\phi\ul{i}\forces F[i,a_i]$ for $i<k$ and $\phi'\ul{i}\succ(G)\eta_{k,\;\phi}\forces\bot$
for $i\ge k$, and therefore $\phi'\ul{i}\forces F[i,\vide]$.\\
By hypothesis on $U$, it follows that $(U)(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}\forces\bot$, in other words $H\ul{k}\,\phi\forces\bot$.\\
Thus, we have shown that, if $H\ul{k}\,\phi\nforces\bot$, then $\eta_{k,\;\phi}\nforces\pt x\neg F[k,x]$,
which gives immediately the desired result.
\qed
\smallskip\noindent
Let $\phi_0\in\LLbd$ be such that $H\ul{0}\phi_0\nforces\bot$. By means of lemma~\ref{rec-Psi2},
we define $\phi_{k+1}\in\LLbd$ and $a_k$ recursively on $k$, by setting \
$\phi_{k+1}=\chi\ul{k}\,\phi_k\,\zeta_{k,\;\phi_k}$.\\
By definition of $\chi$, we have $\phi_{k+1}\ul{i}\succ\zeta_{k,\phi_k}$ for $i\ge k$.\\
Then, we show easily, by recurrence on $k$~:
\smallskip\centerline{
$\phi_{k+1}\ul{i}\succ\phi_{i+1}\ul{i}\succ\zeta_{i,\phi_i}\forces F[i,a_i]$ \ for $i\le k$~; \ $H\ul{k}\phi_k\nforces\bot$.}
\smallskip\noindent
Therefore, we can define~:\\
a function $f$ of domain $\NN$ such that $f[i]=a_i$ for every $i\in\NN$~;\\
and, by theorem~\ref{den_cont}, a term $\phi\in\LLbd$ such that $\phi\ul{k}\succ\zeta_{k,\;\phi_k}$
for all $k\in\NN$.\\
Therefore, we have $\phi\ul{i}\forces F[i,f[i]]$ for every $i\in\NN$, that is to say \
$\phi\forces\pt n\indi F[n,f[n]]$.\\
By hypothesis on $U$, it follows that \ $U\phi\forces\bot$.
Therefore, by theorem~\ref{den_cont}, applied to the sequence $\xi_i=\zeta_{i,\;\phi_i}$, there exists an
integer~$k$ such that $U\psi\forces\bot$, for every
term $\psi\in\LLbd$ such that $\psi\ul{i}\succ\zeta_{i,\;\phi_i}$ for $i<k$.
\smallskip\noindent
Thus, in particular, we have $(U)(\chi\ul{k}\,\phi_k)\xi\forces\bot$ for every $\xi\in\LLbd$.
\smallskip\noindent
Now, by definition of $H$, we have \ $H\ul{k}\,\phi_k\succ(U)(\chi\ul{k}\,\phi_k)\xi$ with
$\xi=(G)\lbd z(H\ul{k}^+)(\chi)\ul{k}\,\phi_kz$, and therefore \ $H\ul{k}\,\phi_k\forces\bot$,
that is a contradiction.
\smallskip\noindent
Thus, we have shown that $H\ul{0}\phi_0\forces\bot$ for every $\phi_0\in\LLbd$.
\qed
\section{Realizing dependent choice}\noindent\label{chdep}
The \emph{axiom of dependent choice} is the following formula~:\\
(DC)\hspace{8em} $\pt x\ex y\,F[x,y]\to\ex f\pt n\indi F[f[n],f[n+1]]$\\
where $F[x,y]$ is an arbitrary formula of \ZFe, with parameters and two free variables.\\
The notation $f[n]$ stands for $\mbox{app}(f,n)$ as defined above.
\begin{theorem}\label{DC}
$\lbd g\lbd u(\Psi)gu\,\ul{0}\,\ul{0}\forces\mbox{\rm DC}$.
\end{theorem}\noindent
The axiom of dependent choice is therefore realized in the model of ZF associated with the BBC realizability algebra
(or, more generally, with any realizability algebra satisfying the property formulated in the remark after
theorem~\ref{den_cont}).
\smallskip\noindent
The proof of theorem~\ref{DC} is almost the same as theorem~\ref{CC}.\\
We write the \emph{axiom of dependent choice} as follows~:\\
(DC)\hspace{8em} $\pt x\,\neg\pt y\,\neg F[x,y],\,\pt f\neg\pt n\indi F[f[n],f[n+1]]\to\bot$.
\smallskip\noindent
Let $G,U\in\LLbd$ be such that \ $G\forces\pt x\,\neg\pt y\,\neg F[x,y]$ and
$U\forces\pt f\neg\pt n\indi F[f[n],f[n+1]]$.\\
We set $H=\Psi GU$ and we have to show that $H\ul{0}\,\ul{0}\forces\bot$. In fact, we shall show that \
$H\ul{0}\xi\forces\bot$ for every $\xi\in\LLbd$.
\begin{lemma}\label{rec-Psi3}\ \\
Let $a_0,\ldots,a_k$ be a finite sequence in ${\cal M}$ and $\phi\in\LLbd$ be such that \
$(\pt i<k)(\phi \ul{i}\forces F[a_i,a_{i+1}])$.\\
If $H\ul{k}\,\phi\nforces\bot$, then there exist $\zeta\in\LLbd$ and $a_{k+1}$ in ${\cal M}$ such that~:\\
\centerline{$\zeta\forces F[a_k,a_{k+1}]$ and $(H\ul{k}^+)(\chi)\ul{k}\,\phi\,\zeta\nforces\bot$.}
\end{lemma}\noindent
Define $\eta_{k,\;\phi}=\lbd z(H\ul{k}^+)(\chi)\ul{k}\,\phi z$, so that
$H\ul{k}\,\phi\succ(U)(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}$.\\
If \ $\eta_{k,\;\phi}\forces\pt y\,\neg F[a_k,y]$ \ then, by hypothesis on $G$, we have
$G\eta_{k,\;\phi}\forces\bot$. We check that~:\\
\centerline{$(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}\forces\pt n\indi F[f_k[n],f_k[n+1]]$}
\noindent
where $f_k$ is defined by $f_k[i]=a_i$ for $i\le k$ (i.e. $i\in k+1$)~; $f_k[i]=\vide$ for $i\notin k+1$.\\
Indeed, if we set \ $\phi'=(\chi k\phi)(G)\eta_{k,\;\phi}$, we have~:\\
$\phi'\ul{i}\succ\phi\ul{i}\forces F[a_i,a_{i+1}]$ for $i<k$ and
$\phi'\ul{i}\succ(G)\eta_{k,\;\phi}\forces\bot$ for $i\ge k$.\\
Therefore, we have $\phi'\ul{i}\forces F[f_k[i],f_k[i+1]]$ for every $i\in\NN$.
\smallskip\noindent
By hypothesis on $U$, it follows that $(U)(\chi\ul{k}\,\phi)(G)\eta_{k,\;\phi}\forces\bot$, that is
$H\ul{k}\,\phi\forces\bot$.
\smallskip\noindent
Thus, we have shown that, if $H\ul{k}\,\phi\nforces\bot$, then $\eta_{k,\;\phi}\nforces\pt y\,\neg F[a_k,y]$,
which gives immediately the desired result.
\qed
\smallskip\noindent
Let $\phi_0\in\LLbd$ be such that $H\ul{0}\phi_0\nforces\bot$ and let $a_0=\vide$. Using lemma~\ref{rec-Psi3},
we define $\phi_{k+1}\in\LLbd$ and $a_{k+1}$ in ${\cal M}$ recursively on $k$, by setting \
$\phi_{k+1}=\chi\ul{k}\,\phi_k\,\zeta_{k,\,\phi_k}$, where $\zeta_{k,\,\phi_k}$ is given by lemma~\ref{rec-Psi3},
where we set $\phi=\phi_k$. By definition of $\chi$, we have $\phi_{k+1}\ul{i}\succ\zeta_{k,\phi_k}$ for $i\ge k$.\\
Then, we show easily, by recurrence on $k$~:
\smallskip\centerline{
$\phi_{k+1}\ul{i}\succ\phi_{i+1}\ul{i}\succ\zeta_{i,\phi_i}\forces F[a_i,a_{i+1}]$ \ for $i\le k$~; \
$H\ul{k}\phi_k\nforces\bot$.}
\smallskip\noindent
Therefore, we can define~:\\
a function $f$ of domain $\NN$ such that $f[i]=a_i$ for every $i\in\NN$~;\\
and, by means of theorem~\ref{den_cont}, a term $\phi\in\LLbd$ such that
$\phi\ul{k}\succ\zeta_{k,\,\phi_k}$ for every $k\in\NN$.\\
Thus, we have $\phi\ul{i}\forces F[f[i],f[i+1]]$ for every $i\in\NN$, that is to say \
$\phi\forces\pt n\indi F[f[n],f[n+1]]$.\\
By hypothesis on $U$, it follows that \ $U\phi\forces\bot$.
Therefore, by theorem~\ref{den_cont}, applied to the sequence $\xi_i=\zeta_{i,\,\phi_i}$,
there exists an integer~$k$ such that $U\psi\forces\bot$, for every
term $\psi\in\LLbd$ such that $\psi\ul{i}\succ\zeta_{i,\,\phi_i}$ for $i<k$.
\smallskip\noindent
Thus, in particular, we have $(U)(\chi\ul{k}\,\phi_k)\xi\forces\bot$ for every $\xi\in\LLbd$.
\smallskip\noindent
But, by definition of $H$, we have \ $H\ul{k}\,\phi_k\succ(U)(\chi\ul{k}\,\phi_k)\xi$ with
$\xi=(G)\lbd z(H\ul{k}^+)(\chi)\ul{k}\,\phi_kz$, and therefore \ $H\ul{k}\,\phi_k\forces\bot$,
that is a contradiction.
\smallskip\noindent
Thus, we have shown that $H\ul{0}\phi_0\forces\bot$ for every $\phi_0\in\LLbd$.
\qed
\section{A well ordering on $\mathbb{R}$}\label{WOR}\noindent
In this section, we use the notations and the results of \cite{JLK3} and \cite{JLK4}.\\
If $F$ is a closed formula of \ZFe, the notation $\forces F$ \ means that there exists a proof-like
term \ $\theta\in\PL_0$ (i.e. a closed $\lbd_c$-term) such that $\theta\forces F$.
\smallskip\noindent
In section~\ref{chden}, we have realized the axiom of countable choice (CC).
We replace $F[n,x]$ with $\mbox{int}(n)\to F[n,x]$ and we add a parameter $\phi$~; we obtain~:
\smallskip
\centerline{$\forces\pt \phi\left(\pt n\indi\ex x\,F[n,x,\phi]\to
\ex f\pt n\indi F[n,f[n],\phi]\right)$}
\smallskip\noindent
for every formula $F[n,x,\phi]$ of \ZFe.
\smallskip\noindent
In particular, taking $\phi\eps2^\NN$ and $F[n,x,\phi]\equiv(x=\phi(n))\land(x=0\lor x=1)$\\
(i.e. $(n,x)\eps\phi\land(x=0\lor x=1)$), we find~:
\smallskip
\centerline{$\forces(\pt \phi\eps2^\NN)\ex f\pt n\indi\left((f[n]=\phi(n))\land(f[n]=0\,\lor f[n]=1)\right)$.}
\smallskip\noindent
For any set $f$ in the ground model ${\cal M}$, let $g=\{x\;;\;f[x]=1\}$.\\
We have trivially \ $\III\,\forces\lbr n\in g\rbr=\lbr f[n]=1\rbr$
\footnote{The notations $\gl2$ and $\lbr F\rbr$ where $F$ is a closed formula of ZF\/,
with parameters in the realizability model ${\cal N}$, are defined in \cite{JLK3,JLK4}.
$\gl2$ is called the \emph{characteristic Boolean algebra of ${\cal N}$}. We have $\lbr F\rbr\eps\gl2$.}\\
It follows that~: $\forces\pt f\ex g\pt n\left((f[n]=0\,\lor f[n]=1)\to f[n]=\lbr n\in g\rbr\right)$.
\smallskip\noindent
We have shown that~: $\forces(\pt \phi\eps2^\NN)\ex g\pt n\indi(\phi(n)=\lbr n\in g\rbr)$.
\smallskip\noindent
Now, in \cite{JLK4}, we have built an ultrafilter ${\cal D}:\gl2\to2$ on the Boolean algebra $\gl2$,
with the following property~: the model ${\cal N}$, equipped with the binary relations
${\cal D}(\lbr x\in y\rbr)$, ${\cal D}(\lbr x=y\rbr)$, is a model of ZF\/, denoted ${\cal M}_{\cal D}$,
which is an elementary extension of the ground model ${\cal M}$. Moreover, ${\cal M}_{\cal D}$ is isomorphic
to a transitive submodel of ${\cal N}$ (considered as a model of ZF), which contains every ordinal of~${\cal N}$.\\
${\cal M}_{\cal D}$ satisfies the axiom of choice, because we suppose that ${\cal M}\models \mbox{ZFC}$.\\
If we suppose that ${\cal M}\models\mbox{ V = L}$, then ${\cal M}_{\cal D}$ is isomorphic to the class $L^{\cal N}$
of constructible sets of~${\cal N}$.
\smallskip\noindent
For every $\phi:\NN\to2$, we have obviously ${\cal D}(\phi(n))=\phi(n)$. It follows that~:
\smallskip
\centerline{$\forces(\pt \phi\eps2^\NN)\ex g\pt n\indi\left(\phi(n)={\cal D}\lbr n\in g\rbr\right)$.}
\smallskip\noindent
This shows that the subset of $\NN$ defined by $\phi$ is in the model ${\cal M}_{\cal D}$~:
indeed, it is the element $g$ of this model.\\
We have just shown that \emph{${\cal N}$ and ${\cal M}_{\cal D}$ have the same reals}.\\
Therefore, $\mathbb{R}$ is well ordered in~${\cal N}$, and we have~: $\forces(\mathbb{R}$ \emph{is well ordered}).
\smallskip\noindent
Moreover, if the ground model ${\cal M}$ satisfies V = L, we have~: $\forces$ (\emph{every real is
constructible}). Therefore, the continuum hypothesis is realized.
\smallskip\noindent
Since the models ${\cal N}$ and ${\cal M}_{\cal D}$ have the same reals, every formula of analysis
(closed formula with quantifiers restricted to $\NN$ or $\mathbb{R}$) has the same truth value in
${\cal M}_{\cal D},\,{\cal M}$ or ${\cal N}$.\\
It follows that~:
\smallskip\noindent
\emph{For every formula $F$ of analysis, we have \ ${\cal M}\models F$ \ if and only if \ $\forces F$.}\\
In particular, we have $\forces F$ or $\forces\neg F$.
|
\section{Statement of results}
Here we are interested in minors of ribbon graphs. For ribbon graph minors it is necessary to contract loops and doing so can create additional vertices and components \cite{MR1906909,MR2507944}. Thus there is a fundamental difference between graph and ribbon graph minors.
While it is known that every minor-closed family of graphs can be characterised by a finite set of excluded minors \cite{MR2099147}, the corresponding result for ribbon graphs and their minors is currently only a conjecture \cite{Moffatt:2013ty}.
In this note we provide some support for this conjecture by giving an explicit list of excluded minors for an interesting class of ribbon graphs: those that admit partial duals of low genus. Our main result is the following.
\begin{theorem}\label{t1}
Let $X_1$--$X_3$ be the ribbon graphs of Figure~\ref{f1}.
Then a ribbon graph has a partial dual of Euler genus at most one if and only if it has no ribbon graph minor equivalent to $X_1$, $X_2$, or $X_3$.
\end{theorem}
Although we assume a familiarity with ribbon graphs, ribbon graph minors and partial duals (see, for example, \cite{MR3086663} for a leisurely overview of these topics, or Sections~2 and~3.2 of \cite{Moffatt:2013ty} for a brief review), we briefly recall that the partial dual $G^A$ of a ribbon graph $G$ is the ribbon graph obtained by, roughly speaking, forming the geometric dual of $G$ but only with respect to a given set $A$ of edges of $G$. Partial duality was defined by Chmutov in \cite{MR2507944}. It appears to be a funtamental construction, arising as a natural operation in knot theory, topological graph theory, graph polynomials, matroid theory, and quantum field theory. The \emph{Euler genus} of a connected ribbon graph $G$ is its genus if it is non-orientable, and twice its genus otherwise. For non-connected ribbon graphs it is defined as the sum of the Euler genera of its components.
A connected ribbon graph of Euler genus 0 is a \emph{plane ribbon graph}, and of Euler genus 1 is an \emph{\ensuremath{\mathbb{R}\mathrm{P}^2} ribbon graph}.
Our approach to Theorem~\ref{t1} is as follows. We use a compatibility between partial duals and minors (Equation~\ref{e1}) to reduce the proof of the theorem to showing that a bouquet (i.e., a 1-vertex ribbon graph)
whose partial duals are all of Euler genus at least two contains $X_1$ or $X_2$ as a minor. We then apply a `rough structure theorem' (Theorem~\ref{t2}) that tells us that such a ribbon graph $G$ does not have a particular type of decomposition into a set of plane graphs and one \ensuremath{\mathbb{R}\mathrm{P}^2} graph. $X_1$ is an obstruction to finding an appropriate \ensuremath{\mathbb{R}\mathrm{P}^2} graph ribbon subgraph in $G$ for the decomposition, $X_2$ is an obstruction to finding an appropriate set of plane ribbon subgraphs for the decomposition, and, somewhat surprisingly, $X_1$ is also an obstruction to plane and \ensuremath{\mathbb{R}\mathrm{P}^2} subgraphs fitting together in a way required by the rough structure theorem to allow for a low-genus partial dual.
\begin{figure}[t]
\begin{subfigure}[b]{.3\linewidth}
\centering
\includegraphics[scale=0.4]{f1}
\caption{$X_1$.}
\label{f1a}
\end{subfigure}
\begin{subfigure}[b]{.3\linewidth}
\centering
\includegraphics[scale=0.4]{f2}
\caption{$X_2$.}
\label{f1b}
\end{subfigure}
\begin{subfigure}[b]{.3\linewidth}
\centering
\includegraphics[scale=0.4]{f3}
\caption{$X_3$.}
\label{f1c}
\end{subfigure}
\caption{The excluded minors.}
\label{f1}
\end{figure}
\section{Proof of the main theorem}
The following constructions and results will be needed for the proof of Theorem~\ref{t1}.
A vertex $v$ of a ribbon graph $G$ is a {\em separating vertex} if there are non-trivial ribbon subgraphs $G_1$ and $G_2$ of $G$ such that $G=G_1\cup G_2$ and $G_1\cap G_2=\{v\}$. Let $G$ be a ribbon graph and $A\subseteq E(G)$. We use $G|_A$ to denote the restriction of $G$ to $A$ (i.e., the ribbon subgraph with edge set $A$ and vertices incident to edges in $A$), and $A^c$ to denote the complement of $A$ in $E(G)$.
Following \cite{MR2994404}, we say that $A$ {\em defines a biseparation} of $G$ if every vertex of $G$ that is in both $G|_A$ and $G|_{A^c}$ is a separating vertex of $G$.
If, in addition, every component of $G|_A$ and of $G|_{A^c}$ is plane then we say $A$ defines a \emph{plane-biseparation}; and if exactly one component of $G|_A$ or of $G|_{A^c}$ is \ensuremath{\mathbb{R}\mathrm{P}^2} and all of the other components are plane then $A$ defines an \emph{\ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation}.
We will use the following rough structure theorem for ribbon graphs with low genus partial duals from \cite{MR2994404} (the plane case first appeared in \cite{MR2928906}).
\begin{theorem}\label{t2}
Let $G$ be a connected ribbon graph and $A\subseteq E(G)$. Then $G^A$ is a plane ribbon graph (respectively, \ensuremath{\mathbb{R}\mathrm{P}^2} ribbon graph) if and only if $A$ defines a plane-biseparation (respectively, \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation) of $G$.
\end{theorem}
We will also need the following result from \cite{Moffatt:2013ty} that says the partial duals of minors are the minors of a partial dual.
For a ribbon graph $G$ and $A\subseteq E(G)$, we have
\begin{equation}\label{e1}
\{ J^A \mid J \text{ is a minor of } G \} = \{ H \mid H \text{ is a minor of } G^A \} .
\end{equation}
Here we use the convention that if $H$ is a minor of $G$, and $A\subseteq E(G)$, then by $H^A$ we mean $H^{A\cap E(H)}$.
A \emph{bouquet} is a ribbon graph that has exactly one vertex. Two edges $e$ and $f$ in a bouquet are \emph{interlaced} if their ends are met in the cyclic order $e\, f\, e\; f$ when travelling round the boundary of the unique vertex of $G$. The \emph{intersection graph} $\mathcal{I}(G)$ of a bouquet $G$ is the vertex weighted simple graph whose vertices set is $E(G)$ and which two vertices $e$ and $f$ of $\mathcal{I}(G)$ are adjacent if and only if the edges $e$ and $f$ are interlaced in $G$. A vertex $e$ of $\mathcal{I}(G)$ has weight ``+'' if $e$ is an orientable loop in $G$, and has weight $``-$'' if it is non-orientable.
\begin{proof}[Proof of Theorem~\ref{t1}]
For one implication observe that $X_1$--$X_3$ do not have partial duals of Euler genus less than 1. It was shown in \cite{Moffatt:2013ty} that for each $k\in \mathbb{N}_0$ the set
of all ribbon graphs that have a partial dual of Euler genus at most $k$ is minor-closed, and so $X_1$--$X_3$ cannot be minors of ribbon graphs that have partial duals of Euler genus less than 1.
For the converse we prove that if a ribbon graph does not admit a plane- or \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation, then it has an $X_1$--$X_3$-minor. The result then follows from Theorem~\ref{t2}. If $G$ is an orientable ribbon graph that does not admit a plane-biseparation then, from \cite{MR2928906}, it contains $X_2$ or $X_3$ as a minor. Now suppose that $G$ is a non-orientable ribbon graph that does not admit an \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation. To complete the proof we need to show that $G$ has an \mbox{$X_1$--$X_3$-minor.} By Equation~\eqref{e1} it is enough to show that when $G$ is a bouquet it has $X_1$ or $X_2$ as a minor. (To see why, suppose $G$ has no \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation, and let $T$ be the edge set of a spanning tree of $G$. Then $G^T$ is a bouquet. If $G^T$ has $X_1$ or $X_2$ as a minor, by Equation~\ref{e1}, $G=(G^T)^T$ has $X_1^T$ or $X_2^T$ as a minor. Finally, $X_1^T=X_1$, and $X_2^T = X_2$ or $X_2^T=X_3$.)
Assume that $G$ is a non-orientable bouquet that does not admit an \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation.
We can write $G=G_O \sqcup G_N$ where $G_O$ is the ribbon subgraph of $G$ consisting of all of the orientable loops, and $G_N$ the non-orientable loops. If $G_N$ has two edges that are not interlaced then these two edges induce an $X_1$-minor and we are done. Otherwise $G_N$ consists of $q\geq 1$ non-orientable loops that all interlace each other (i.e., the edges are met in the cyclic order $1\,2\cdots q\,1\,2\cdots q$). Assume this is the case.
Next consider $G_O$. Suppose the intersection graph $\mathcal{I}(G_O)$ is not bipartite. It then contains an odd cycle of length at least 3. This odd cycle corresponds to a ribbon subgraph of $G$ that is an orientable bouquet with $p\geq 3$ edges, for $p$ odd, that meet the vertex in the cyclic order
$ 2\,1\,3\,2\,4\,3\cdots p\,(p-1)\,1\,p$. Contracting the edges corresponding to $4, 5, \ldots ,p$ results in a copy of $X_2$ (see the example in \cite{Moffatt:2013ty}).
Now suppose $\mathcal{I}(G_O)$ is bipartite. Consider the graph $\mathcal{I}'(G)$ which is obtained from $\mathcal{I}(G)$ by identifying all of the negatively weighted vertices into a single negatively weighted vertex $v$, and deleting any loops created in this process (so all of the edges in $G_N$ are represented by a single vertex $v$).
$\mathcal{I}'(G)$ cannot be bipartite since otherwise $G$ would admit an \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation (if $\mathcal{I}'(G)$ is 2-coloured then the set $A$ of all edges of $G$ of a single colour defines an \ensuremath{\mathbb{R}\mathrm{P}^2}-biseparation).
We therefore have that $\mathcal{I}'(G)$ is not bipartite and therefore contains an odd cycle. Let $C$ be a minimal odd cycle of $\mathcal{I}'(G)$. Since $\mathcal{I}(G_O)=\mathcal{I}'(G) \setminus v$ is bipartite, this odd cycle must contain $v$. Furthermore, by minimality, the subgraph of $\mathcal{I}'(G)$ induced by $C$ must also be $C$ (since adding any other edge between the vertices of $C$ will create a smaller cycle). Suppose the vertices of $C$ are $1,2,\cdots ,2m,v$ in that order. Consider the ribbon subgraph $H$ of $G$ corresponding to $C$. Since $C$ is minimal, $H$ consists of $2m$ orientable loops, which we name $1, \ldots, 2m$, whose ends appear in the order
$1\,2\,1\,3\,2\, 4\, 3 \cdots (2m-1) \,(2m-2) \,(2m) \,(2m-1) \,(2m)$, together with some number of non-orientable loops such that a non-orientable loop interlaces with 1, a (not necessarily distinct) orientable loop interlaces with $2m$, and all of the non-orientable loops interlace with each other (since the non-orientable part is $G_N$). See Figure~\ref{f2}.
With respect to this order, we name the arcs on the boundary of the unique vertex of $H$ that are bounded by the orientable edges on $H$ as follows:
\begin{multline*}
\varepsilon \,1 \, \alpha \, 2\, \gamma_1 \,1 \, \delta_1\, 3\, \gamma_2 \,2\, \delta_2 \,4\, \gamma_3 3 \cdots (2m-1)\, \gamma_{(2m-2)} \,
\\
(2m-2) \, \delta_{(2m-2)}\,
(2m) \,\gamma_{(2m-1)}\, (2m-1) \, \beta \,(2m) \, \varepsilon,
\end{multline*}
as in Figure~\ref{f2}.
The ends of the non-orientable edges of $H$ lie on these arcs, and we will use the names of these arcs to discuss the possible ways that the non-orientable edges in $H$ can be interlaced with its orientable edges, and for each case to describe how the find the required ribbon graph minor.
\begin{figure}[t]
\centering
\labellist \small\hair 2pt
\pinlabel {$1$} at 12 162
\pinlabel {$2$} at 14 80
\pinlabel {$3$} at 75 15
\pinlabel {$2m$} [l] at 273 162
\pinlabel {$\alpha$} at 87 166
\pinlabel {$\beta$} at 201 166
\pinlabel {$\gamma_1$} at 85 133
\pinlabel {$\gamma_2$} at 102 98
\pinlabel {$\gamma_3$} at 133 82
\pinlabel {$\gamma_{(2m-1)}$} [r] at 215 136
\pinlabel {$\varepsilon$} at 144 203
\pinlabel {$\delta_1$} at 49 96
\pinlabel {$\delta_2$} at 93 52
\endlabellist
\includegraphics[scale=0.6]{f5}
\caption{$H$ without its non-orientable loops, and the labelled arcs on its vertex.}
\label{f2}
\end{figure}
First suppose that $m=1$. If there is a single non-orientable loop $e$ interlacing both $1$ and $2$ then the ribbon subgraph with edges $1,2,e$ is given by the order of the ends $ 1 \,e \,2 \,1 \,e \,2$ or $1 \,2 \,e \,1 \,2 \,e$ around the vertex. In either case contracting $e$ results in an $X_1$-minor. So suppose that no non-orientable loop interlaces both 1 and 2. Then there is a non-orientable loop $e$ interlacing $1$ but not 2, and a non-orientable loop $f$ interlacing $2$ but not 1. Then the ends of $e$ are therefore on $\alpha$ and $\varepsilon$, or $\gamma_1$ and $\beta$ (otherwise it interlaces 2). Similarly, the ends of $f$ are therefore on $\beta$ and $\varepsilon$, or $\gamma_1$ and $\alpha$. Using the fact that all non-orientable loops interlace, $H$ must therefore contain one of the four ribbon subgraphs with edges $1,2,e,f$ and their positions given by one of $1\,e\,2\,1\,f\,2\,e\,f$, or $1\,f\,e\,2\,f\,1\,2\,e$, or $1\,2\,e\,1\,f\,e\,2\,f$, or $1\,f\,2\,e\,f\,1\,e\,2$. In all of these cases contracting the edges 1 and 2 results in a copy of $X_1$, as required. Thus we can obtain the required minor when $m=1$.
Now suppose that $m>1$. First consider the case where there is a single non-orientable loop $e$ interlacing both $1$ and $2m$. Then $e$ cannot have ends on either of the arcs $\gamma_1$ or $\gamma_{(2m-1)}$, as this could mean that $C$ is not a minimal odd cycle in $\mathcal{I}'(G)$ (given by the edges $1,2,e$ or $(2m-1),(2m),e$). Thus $e$ has ends on $\alpha$ and $\beta$. Deleting all of the non-orientable loops of $H$ except for $e$, then contracting the edges $e, 2,3,\ldots , (2m-1)$ gives an $X_1$-minor.
Now consider the case when no non-orientable loop interlaces both $1$ and $2m$.
Then there is a non-orientable loop $e$ interlacing $1$ but not $2m$, and a non-orientable loop $f$ interlacing $2m$ but not 1,
$e$ has an end on $\alpha$ or $\gamma_1$, and $f$ has an end on $\beta$ or $\gamma_{(2m-1)}$.
First suppose that $e$ has an end on $\gamma_1$. Then its other end must be on $\delta_1$ or $\gamma_2$. (This is since either $e$ would not be interlaced with 1, or the edges $1,2,e$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.) Consider $f$. Since all non-orientable loops interlace and $f$ does not interlace with $1$, we have that if $e$ has an end on $\delta_1$ then $f$ has an end on $\delta_1$, and if $e$ has an end on $\gamma_2$ then $f$ has an end on $\delta_1$ or $\gamma_2$. In all three cases the edges $1,2,e,f$ induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$. Thus $e$ cannot have an end on $\gamma_1$, and must have an end on $\alpha$.
A similar argument shows that $f$ cannot have an end on $\gamma_{(2m-1)}$, and must have an end on $\beta$. We now examine where the other ends of $e$ and $f$ can lie.
If $e$ has an end on some $\gamma_i$ for $i\in \{2, \ldots ,(2m-1)\}$ (we have shown that it cannot have an end on $\gamma_1$) then the edges $e, i, i+1$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.
Similarly, if $f$ has an end on some $\gamma_i$ for $i\in \{1, \ldots ,(2m-2)\}$ (we have shown that it cannot have an end on $\gamma_{(2m-1)}$) then the edges $f, i, i+1$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.
Thus $e$ has one end on $\alpha$, and one on $\varepsilon$ or on some $\delta_i$; and $f$ has one end on $\beta$ and one on $\varepsilon$ or on some $\delta_j$.
If one of $e$ or $f$ has an end on $\delta_i$, then, by interlacement, the other has an end on some $\delta_j$.
If $e$ has an end on $\delta_{2k-1}$, for some $k$, then the edges $e,1, \ldots , 2k$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.
Similarly, if $f$ has an end on $\delta_{2k}$, for some $k$, then the edges $2k+1, \ldots , 2m$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.
It follows that if $e$ has an end on $\delta_{2k}$ then $f$ has an end on $\delta_{2j-1}$ for some $j\leq k$, but then the edges $e,f, 2j, \ldots , (2k+1)$ will induce an odd cycle smaller that $C$ in $\mathcal{I}'(G)$.
Thus we have that $e$ has ends on $\alpha$ and $\varepsilon$ and $f$ has ends on $\beta$ and $\varepsilon$, with $e$ and $f$ interlaced. Deleting all of the non-orientable loops of $H$ except for $e$ and $f$, and contracting all of the orientable loops results in a copy of $X_1$. Thus we have shown in all cases that $H$, and therefore $G$ has an $X_1$-minor. This completes the proof of the theorem.
\end{proof}
We conclude with an application of Theorem~\ref{t1} to knot theory. Dasbach, Futer, Kalfagianni, Lin and Stoltzfus, in \cite{MR2389605}, described how every alternating link can be represented by a ribbon graph. This construction readily extends to link diagrams on other surfaces. Using, from \cite{SM}, that a ribbon graph represents a checkerboard colourable diagram of a link in real projective space, $\mathbb{R}\mathrm{P}^3$, if and only if it has a plane or \ensuremath{\mathbb{R}\mathrm{P}^2} partial dual immediately gives the following corollary.
\begin{corollary}
A ribbon graph $G$ represents a checkerboard colourable diagram of a link in real projective space if and only if it has no minor equivalent to $X_1$, $X_2$, or $X_3$.
\end{corollary}
Finding the corresponding result for non-checkerboard colourable diagrams is an interesting open problem.
\bibliographystyle{plain}
|
\section{Fractional minimum positive semidefinite rank} \label{sec-frac-PSD-mr}
In this section, we introduce $(d;r)$ faithful orthogonal subspace representations, $r$-fold minimum positive semidefinite rank, and fractional minimum positive semidefinite rank, extending the definitions of faithful orthogonal representations and minimum positive semidefinite rank. We then introduce faithful $d/r$-projective representations and connect everything to projective rank. A connection to positive semidefinite matrices is explored, and properties of our new quantities are proven.
Unless otherwise specified, all matrices and vectors in this section are assumed to be complex-valued (the literature on minimum positive semidefinite rank is mixed, with both real and complex cases studied).
\subsection{Faithful orthogonal subspace representations and fractional minimum positive semidefinite rank}
Given a graph $G$ and $d, r \in \mathbb{N}$ with $r \leq d$, a \emph{$(d;r)$ faithful orthogonal subspace representation}, or $(d;r)$-FOSR, for $G$ is a set of subspaces $\{S_u\}_{u \in V(G)}$ where
\begin{itemize}
\item for each $u \in V(G)$, $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^d$; and
\item $S_u \perp S_v$ if and only if $uv \notin E(G)$.
\end{itemize}
A faithful orthogonal representation (as defined in Section \ref{sec-intro-defs}) generates a $(d;1)$ faithful orthogonal subspace representation, and vice versa. Further, a $(d;r)$-FOSR for a graph $G$ is a $(d;r)$-OSR for its complement $\overline{G}$, but the reverse statement is not true in general.
Now that we have defined an $r$-fold analogue of a faithful orthogonal representation, it is natural to consider a corresponding version of $\mr^+(G)$. The \emph{$r$-fold minimum positive semidefinite rank} of $G$ is
\[ \mrr{r}^+(G) = \min \{ d : G \text{ has a $(d;r)$ faithful orthogonal subspace representation} \}. \]
In particular, we have $\mrr{1}^+(G) = \mr^+(G)$, using definition \eqref{def-mrplus} of $\mr^+$; we caution the reader that this coincides with the definitions of faithful orthogonal representation and minimum positive semidefinite rank in the literature (e.g. \cite{BHH11, HLA2ch46}) if and only if $G$ has no isolated vertices.
We note that $\mrr{r}^+(G)$ is subadditive. The proof is analogous to the proof of Lemma \ref{xirsubadditive} and is omitted, as are the proofs for other results in this section that parallel those for the non-faithful case (i.e., the $\xi$-family of parameters).
\begin{Lemma} \label{mrrsubadditive}
$\mrr{r}^+$ is a subadditive function of $r$, i.e., for every graph $G$ and all $r, s \in \mathbb{N}$,
\[ \mrr{r+s}^+(G) \leq \mrr{r}^+(G) + \mrr{s}^+(G). \]
\end{Lemma}
As in the non-faithful case, an immediate corollary relates $\mrr{r}^+$ to $\mr^+$.
\begin{Corollary} \label{mrrleqrtimesmr}
For every graph $G$ and all $r \in \mathbb{N}$,
\[ \frac{\mrr{r}^+(G)}{r} \leq \mr^+(G). \]
\end{Corollary}
For any graph $G$, we define the \emph{fractional minimum positive semidefinite rank} of $G$ as
\[ \mr^+_f(G) = \inf_r \left\{ \frac{\mrr{r}^+(G)}{r} \right\}. \]
Notice that if $G$ has a $(d;r)$ faithful orthogonal subspace representation, then $\mrr{r}^+(G) \leq d$, so $\mr_f^+(G) \leq \frac{d}{r}$.
\smallskip
We can upper bound fractional minimum positive semidefinite rank by the non-fractional version by using Corollary \ref{mrrleqrtimesmr}. Again, recall that this coincides with the literature if and only if the graph $G$ has no isolated vertices.
\begin{Corollary} \label{mrfleqmr} For every graph $G$,
\[ \mr_f^+(G) \leq \mr^+(G). \]
\end{Corollary}
Since $\mrr{r}^+(G)$ is subadditive, we have the following corollary, which follows from Lemma \ref{fgtA41} (\cite{fgt}, Lemma A.4.1).
\begin{Corollary} \label{mrf-equals-lim} For every graph $G$,
\[ \mr_f^+(G) = \lim_{r \to \infty} \frac{\mrr{r}^+(G)}{r} \, , \]
and this limit exists.
\end{Corollary}
We conclude this section with an example that gives further insight into these new parameters.
\begin{Example} \label{ex-mrr-of-P4}
{\rm
Let $r \in \mathbb{N}$ and consider the graph $G = P_4$ with $V(P_4) = \{1,2,3,4\}$ and $E(P_4) = \{12, 23, 34\}$. With $\bv{e}_i$ as the $i^{th}$ standard basis vector in $\mathbb{C}^{2r+1}$, we can verify that the following is a valid $(2r+1;r)$-FOSR for $P_4$: $S_1 = \range([\bv{e}_1, \, \bv{e}_2, \, \ldots, \, \bv{e}_{r}])$, $S_2 = \range([ \bv{e}_2, \, \bv{e}_3, \, \ldots, \, \bv{e}_{r+1}])$, $S_3 = \range([\bv{e}_{r+1}, \, \bv{e}_{r+2}, \, \ldots, \, \bv{e}_{2r}])$, $S_4 = \range([\bv{e}_{r+2}, \, \bv{e}_{r+3}, \, \ldots, \, \bv{e}_{2r+1}])$. Therefore, $\mrr{r}^+(P_4) \leq 2r+1$. Suppose that $\{Q_u\}_{u \in V(P_4)}$ is a $(2r; r)$-FOSR for $P_4$; we show that such a representation cannot exist. Since $13, 14 \notin E(P_4)$, $Q_1 \perp Q_3$ and $Q_1 \perp Q_4$. The underlying space is $\mathbb{C}^{2r}$ and each subspace $Q_i$ is $r$-dimensional, so we must therefore have $Q_3 = Q_4 = Q_1^\perp$. Now, $23 \in E(P_4)$, so $Q_2 \not\perp Q_3$, but $24 \notin E(P_4)$, so it also follows that $Q_2 \perp Q_4$. Since $Q_3 = Q_4$, this is a contradiction; thus there is no $(2r;r)$-FOSR for $P_4$, and so $\mrr{r}^+(P_4) = 2r+1$. Using the limit characterization of $\mr_f^+$, it follows that $\mr_f^+(P_4) = \lim_{r \to \infty} \frac{2r+1}{r} = 2$.
}
\end{Example}
This example demonstrates that the infimum in the definition of the fractional minimum positive semidefinite rank cannot be replaced with a minimum, even when $\mr_f^+$ is a rational number. Additionally, since $\mr^+(P_4) = 3$, the graph $G = P_4$ satisfies $\mr_f^+(G) < \mr^+(G)$.
\subsection{Faithful $\mathbit{d/r}$-projective representations}
Let $G$ be a graph and $d, r \in \mathbb{N}$ with $r \leq d$. A \emph{faithful $d/r$-projective representation}, or faithful $d/r$-representation for short, is an assignment of matrices $\{P_u\}_{u \in V(G)}$ to the vertices of $G$ such that
\begin{itemize}
\item for each $u \in V(G)$, $P_u \in \mathbb{C}^{d \times d}$, $\rank P_u = r$, $P_u^* = P_u$, and $P_u^2 = P_u$; and
\item $P_u P_v = 0$ if and only if $uv \notin E(G)$.
\end{itemize}
A faithful $d/r$-representation for $G$ is a $d/r$-representation for $\overline{G}$, but the reverse is not necessarily true.
It is convenient to note that a $(d;r)$ faithful orthogonal subspace representation for $G$ is equivalent to a faithful $d/r$-representation. The proof is analogous to that of Proposition \ref{drOSRiffdr}; as before, we will omit such parallel proofs.
\begin{Proposition} \label{drFOSRifffaithfuldr}
A graph $G$ has a $(d;r)$ faithful orthogonal subspace representation if and only if $G$ has a faithful $d/r$-representation.
\end{Proposition}
An immediate corollary gives an alternate definition for $\mr_f^+(G)$.
\begin{Corollary} For every graph $G$,
\[ \mr^+_f(G) = \inf_{d,r} \left\{ \frac{d}{r} : G \text{ has a faithful $d/r$-representation} \right\}. \]
\end{Corollary}
\begin{Corollary} \label{xifleqmrf}
For any graph $G$ with complement $\overline{G}$,
\[ \xi_f(\overline{G}) \leq \mr_f^+(G) \leq \mr^+(G). \]
\end{Corollary}
\begin{proof}
This follows from the fact that any faithful $d/r$-representation for $G$ is also a $d/r$-\break representation for $\overline{G}$, as well as from Corollary \ref{mrfleqmr}. \qquad
\end{proof}
\subsection{Relation to positive semidefinite matrices}
In this section, we connect $(d;r)$ faithful orthogonal subspace representations to positive semidefinite matrices, thus generalizing the known results for the $r = 1$ case (when the graph in question has no isolated vertices) and connecting $\mrr{r}^+(G)$ to the rank of a positive semidefinite matrix.
We begin with some definitions. Let $G$ be a graph on $n$ vertices and suppose that $V(G) = \brackets{1}{n}$. For some $r \in \mathbb{N}$, let $A \in \mathbb{C}^{nr \times nr}$ be partitioned into an $n \times n$ block matrix $[A_{ij}]$, where $A_{ij}$ is the $r \times r$ submatrix in (block) row $i$ and (block) column $j$ of $A$. We say that the matrix $A$ \emph{$r$-fits} $G$ if $A_{ii} = I_r$ for each $i \in V(G)$ and $A_{ij} = 0$ if and only if $ij \notin E(G)$, and define the set
\[ \Hr{r}^+(G) = \left\{ A \in \mathbb{C}^{nr \times nr} : A \mgeq 0 \text{ and $A$ $r$-fits $G$} \right\}. \]
\begin{Example} {\rm
We provide a simple example for the $r = 2$ case. Let $G = P_3$, the path on 3 vertices, with $V(G) = \{1,2,3\}$ and $E(G) = \{ 12, 23 \}$. Choosing $X = [ \bv{e}_1 \, \bv{e}_2 \, | \, \bv{e}_1 \, \bv{e}_4 \, | \, \bv{e}_3 \, \bv{e}_4]$, where $\bv{e}_j$ is the $j^{th}$ standard basis vector in $\mathbb{C}^{4}$, we can verify that
\[ A = X^* X = \left[ \begin{array}{cc|cc|cc}
1 & 0 & 1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\ \hline
1 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 1 \\ \hline
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 & 0 & 1
\end{array} \right] \in \Hr{2}^+(P_3). \]
}
\end{Example}
This constructive example gives an intuitive feel for one direction of the proof of the main result of this section.
\begin{Theorem} For every graph $G$ on $n$ vertices and any $r \in \mathbb{N}$,
\[ \mrr{r}^+(G) = \min \left\{ \rank A : A \in \Hr{r}^+(G) \right\}. \]
\end{Theorem}
\begin{proof}
Let $d = \mrr{r}^+(G)$ and let $\ell = \min \left\{ \rank A : A \in \Hr{r}^+(G) \right\}$.
First, assume that $\{S_i \}$ is a $(d;r)$ faithful orthogonal subspace representation for $G$ and for each $i \in V(G)$ let $X_i \in \mathbb{C}^{d \times r}$ be a basis matrix for $S_i$. Define $X = [ X_1 \, | \, X_2 \, | \cdots | \, X_n ] \in \mathbb{C}^{d \times nr}$ and let $B = X^* X \in \mathbb{C}^{nr \times nr}$. We see immediately that $B \mgeq 0$ and $\rank B = \rank X \leq d$. Partitioning $B$ into an $n \times n$ block matrix with blocks $[B_{ij}]$ of size $r \times r$, we have $B_{ij} = X_i^* X_j$. Since $S_i \perp S_j$ if and only if $X_i^* X_j = 0$, we have $B_{ij} = 0$ if and only if $S_i \perp S_j$, which occurs if and only if $ij \notin E(G)$. Additionally, since $X_i$ has orthonormal columns, we have $B_{ii} = I_r$ for each $i$. Therefore, $B \in \Hr{r}^+(G)$, so $\min \left\{ \rank A : A \in \Hr{r}^+(G) \right\} \leq \rank B \leq d = \mrr{r}^+(G)$.
For the reverse inequality, suppose that $B \in \Hr{r}^+(G)$ and $\rank B = \ell$. Then there exists a matrix $X \in \mathbb{C}^{\ell \times nr}$ such that $B = X^* X$. Partition $B$ into $r \times r$ blocks $[B_{ij}]$ and partition $X$ into $\ell \times r$ blocks as $X = [ X_1 \, | \, X_2 \, | \cdots | \, X_n ]$. For each vertex $i \in V(G)$, let $S_i = \range(X_i) \subseteq \mathbb{C}^\ell$. Since $X_i^* X_i = I_r$, we have $\rank X_i = r$, so each $S_i$ is an $r$-dimensional subspace of $\mathbb{C}^\ell$. Additionally, $X_i^*X_j = B_{ij} = 0$ if and only if $ij \notin E(G)$, so $S_i \perp S_j$ if and only if $ij \notin E(G)$. Therefore, $\{S_i\}$ is an $(\ell;r)$ faithful orthogonal subspace representation for $G$, so $\mrr{r}^+(G) \leq \ell = \min \left\{ \rank A : A \in \Hr{r}^+(G) \right\}$ and thus equality holds. \qquad
\end{proof}
This matrix-based representation is a powerful theoretical tool that allows us to simplify the proofs of some properties of $r$-fold minimum positive semidefinite rank, as well as to more clearly draw parallels to the existing and well-established $r=1$ case (although again, the connection to the literature requires that the graph in question has no isolated vertices).
The condition that $A_{ii} = I_r$ if $A$ $r$-fits a graph $G$ is a strong one, so we conclude this section with a weaker condition that will be used to further simplify proofs without sacrificing utility. We say that $A$ \emph{weakly $r$-fits} $G$ if $A_{ii}$ is a diagonal matrix with strictly positive diagonal entries for each $i \in V(G)$ and $A_{ij} = 0$ if and only if $ij \notin E(G)$. Clearly, any matrix that $r$-fits $G$ also weakly $r$-fits $G$.
\begin{Remark} {\rm
Suppose that $A$ weakly $r$-fits a graph $G$ and let $D = D_1 \oplus \cdots \oplus D_n$, where each $D_i$ is the inverse of the positive square root of $A_{ii}$, i.e., $D_i = A_{ii}^{-\frac{1}{2}}$. Then the matrix $B = DAD$ $r$-fits $G$, since $D$ is a diagonal matrix with strictly positive diagonal entries, so multiplication by $D$ does not change the zero pattern of $A$. Further, $\rank B = \rank A$, since $D$ has full rank.
}
\end{Remark}
This remark yields an immediate corollary to the previous theorem.
\begin{Corollary} For every graph $G$ on $n$ vertices and any $r \in \mathbb{N}$,
\[ \mrr{r}^+(G) = \min \left\{ \rank A : A \in \mathbb{C}^{nr \times nr}, \, A \mgeq 0 \text{ and $A$ weakly $r$-fits $G$} \right\}. \]
\end{Corollary}
\subsection{Properties of mr$\mathbit{_{[r]}^+(G)}$ and mr$\mathbit{_f^+(G)}$}
In this section, we prove numerous results regarding properties of $r$-fold and fractional minimum positive semidefinite rank, many of which extend known properties of $\mr^+$ to the new parameters.
\begin{Observation} \label{mrr-geq-alpha}
For every graph $G$ and all $r \in \mathbb{N}$, $\mrr{r}^+(G) \geq r \cdot \alpha(G)$.
\end{Observation}
\begin{Proposition} \label{mrr-subgraph}
Let $r \in \mathbb{N}$ and let $H$ be an induced subgraph of $G$. Then $\mrr{r}^+(H) \leq \mrr{r}^+(G)$.
\end{Proposition}
\begin{proof}
For any $u,v \in V(H)$, $uv \in E(H)$ if and only if $uv \in E(G)$, since $H$ is induced. Therefore any $(d;r)$ faithful orthogonal subspace representation for $G$ provides a $(d;r)$ faithful orthogonal subspace representation for $H$, and the result follows immediately. \qquad
\end{proof}
\begin{Proposition} \label{mrr-dcup}
If $G = \bigdcup_{i=1}^t G_i$ for some graphs $\{G_i\}_{i=1}^t$, then $\mrr{r}^+(G) = \sum_{i=1}^t \mrr{r}^+(G_i)$ for each $r \in \mathbb{N}$.
\end{Proposition}
\begin{proof}
Suppose that $V(G) = \brackets{1}{n}$ and that $|V(G_i)| = n_i$ for $i = 1, 2, \ldots, t$. Further assume that $V(G_i) = \brackets{1 + \sum_{j=1}^{i-1} n_j}{\sum_{j=1}^i n_j}$, so that if $A \in \Hr{r}^+(G)$, then $A = A_1 \oplus A_2 \oplus \cdots \oplus A_t$, where $A_i \in \Hr{r}^+(G_i)$ for each $i$. Note that $\rank A = \sum_{i=1}^t \rank A_i$. We therefore have
\begin{align*}
\mrr{r}^+(G)
&= \min \left\{ \rank A : A \in \Hr{r}^+(G) \right\} \\
&= \min \left\{ \sum_{i=1}^t \rank A_i : A_i \in \Hr{r}^+(G_i) \text{ for each $i$} \right\} \\
&= \sum_{i=1}^t \min \left\{ \rank A_i : A_i \in \Hr{r}^+(G_i) \right\} \\
&= \sum_{i=1}^t \mrr{r}^+(G_i). \qedhere
\end{align*}
\end{proof}
\begin{Theorem}
If $G = \bigcup_{i=1}^t G_i$ for some graphs $\{G_i\}_{i=1}^t$, then $\mrr{r}^+(G) \leq \sum_{i=1}^t \mrr{r}^+(G_i)$ for each $r \in \mathbb{N}$.
\end{Theorem}
\begin{proof}
We prove the result for the case where $t = 2$ and note that recursive application of this case will prove the more general one.
Let $V(G) = \brackets{1}{n}$ where $n > 0$ and assume that $V(G_1) \setminus V(G_2) = \brackets{1}{n_1}$, $V(G_1) \cap V(G_2) = \brackets{n_1+1}{n_1+c}$, and $V(G_2) \setminus V(G_1) = \brackets{n_1+c + 1}{n_1+c+n_2}$, where $n_1, n_2, c \geq 0$ (it is not assumed that each of these is strictly nonzero). Note that $n = n_1 + c + n_2$, and this ordering asserts that the first $n_1$ vertices (enumerating in the natural order) lie exclusively in $G_1$, the next $c$ are common to both graphs, and the last $n_2$ lie exclusively in $G_2$.
For $i = 1,2$, let $\mrr{r}^+(G_i) = d_i$ and let $A_i \in \Hr{r}^+(G_i)$ be chosen so that $\rank A_i = d_i$. Notice that $A_1 \in \mathbb{C}^{(n_1+c)r \times (n_1+c)r}$ has its rows and columns indexed by $V(G_1) = \brackets{1}{n_1+c}$ and $A_2 \in \mathbb{C}^{(n_2+c)r \times (n_2+c)r}$ has its rows and columns indexed by $V(G_2) = \brackets{n_1+1}{n}$.
Let
\[ \hat A_1 = \left[ \begin{array}{cc} A_1 & 0 \\ 0 & 0 \end{array} \right] \in \mathbb{C}^{nr \times nr}, \qquad \hat A_2 = \left[ \begin{array}{cc} 0 & 0 \\ 0 & A_2 \end{array} \right] \in \mathbb{C}^{nr \times nr} \]
and define $A = \hat A_1 + \beta \hat A_2 \in \mathbb{C}^{nr \times nr}$, where $\beta > 0$ is chosen so that if $A$, $\hat A_1$, and $\hat A_2$ are partitioned into $n \times n$ block matrices with block size $r \times r$, then $A_{ij} = 0$ if and only if $(\hat A_1)_{ij} = 0$ and $(\hat A_2)_{ij} = 0$ (i.e., no cancellation of an entire block occurs).
Since $A$ is a positive linear combination of positive semidefinite matrices, $A \mgeq 0$, and by our choice of $\beta$ we see that $A$ weakly $r$-fits $G$. Therefore,
\[ \mrr{r}^+(G) \leq \rank A \leq \rank \hat A_1 + \rank \hat A_2 = d_1 + d_2 = \mrr{r}^+(G_1) + \mrr{r}^+(G_2). \qedhere \]
\end{proof}
All of the results we have proven for $r$-fold minimum positive semidefinite rank can be extended to results for fractional minimum positive semidefinite rank. The proof is analogous to that of Theorem \ref{xifproperties} and is omitted.
\begin{Theorem} For every graph $G$:
\begin{enumerate}[i)]
\item $\mr_f^+(G) \geq \alpha(G)$.
\item If $H$ is an induced subgraph of $G$, then $\mr_f^+(H) \leq \mr_f^+(G)$.
\item If $G = \bigdcup_{i=1}^t G_i$ for some graphs $\{ G_i \}_{i=1}^t$, then $\mr_f^+(G) = \sum_{i=1}^t \mr_f^+(G_i)$.
\item If $G = \bigcup_{i=1}^t G_i$ for some graphs $\{ G_i \}_{i=1}^t$, then $\mr_f^+(G) \leq \sum_{i=1}^t \mr_f^+(G_i)$.
\end{enumerate}
\end{Theorem}
\smallskip
Let $G$ be a connected graph of order at least two. A standard technique for computing the minimum positive semidefinite rank of $G$ is cut-vertex reduction \cite{BHH11, HLA2ch46, Hol09}: Suppose that $v \in V(G)$ is a cut-vertex and $(G - v)$ has connected components $\{H_i\}_{i=1}^t$. For each $i$, let $G_i$ be the subgraph of $G$ induced by the union of the vertices of $H_i$ with $v$, that is, $G_i = G[V(H_i) \cup \{v\}]$. Then $\mr^+(G) = \sum_{i=1}^t \mr^+(G_i)$. Unfortunately, this technique does not carry over to the $r$-fold case when $r > 1$, as the following example shows.
\begin{Example} {\rm
Consider the graph $G = P_4$, the path on 4 vertices, with $V(G) = \{x,y,v,z\}$ in path order; recall from Example \ref{ex-mrr-of-P4} that $\mrr{r}^+(G) = 2r+1$ for any $r \in \mathbb{N}$. Taking $v$ as a cut-vertex, we have $G_1 = P_3$ with $V(G_1) = \{x,y,v\}$ and $G_2 = P_2$ with $V(G_2) = \{v, z\}$. Fix $r > 1$. Since $\alpha(G_1) = 2$, any valid $(d;r)$-FOSR for $G_1$ must have $d \geq 2r$. Further, it is easy to see that $\mr^+(G_1) = 2$, so $4 \leq \mrr{r}^+(G_1) \leq 2 \cdot \mr^+(G_1) = 2r$. Hence equality holds and $\mrr{r}^+(G_1) = 2r$. Next, since $\mr^+(G_2) = 1$ and $d \geq r$ for any valid $(d;r)$-FOSR, we have $r \leq \mrr{r}^+(G_2) \leq r \cdot \mr^+(G_2) = r$, so $\mrr{r}^+(G_2) = r$. Hence if $r > 1$, then $\mrr{r}^+(G) = 2r+1 < 2r + r = \mrr{r}^+(G_1) + \mrr{r}^+(G_2)$, so cut-vertex reduction does not apply.
}
\end{Example}
\subsection{Fractional minimum positive semidefinite rank and projective rank} \label{subsec-equality}
Recall that any $(d;r)$-FOSR for $G$ is a $(d;r)$-OSR for $\overline{G}$, but the reverse statement does not apply in general. It thus follows that $\xir{r}(\overline{G}) \leq \mrr{r}^+(G)$ for any graph $G$ and $r \in \mathbb{N}$, and the next example demonstrates that this inequality can be strict.
\begin{Example} { \rm
Consider the graph $G = P_4$ with $V(P_4) = \{1,2,3,4\}$ and $E(P_4) = \{12, 23, 34 \}$ and fix $r \in \mathbb{N}$. Since $\omega(P_4) = 2$, we have $\xir{r}(P_4) \geq 2r$. With $\bv{e}_i$ as the $i^{th}$ standard basis vector for $\mathbb{C}^{2r}$, it is easy to verify that the following is a $(2r;r)$-OSR for $P_4$: $S_1 = S_3 = \range([\bv{e}_1, \, \bv{e}_2, \, \ldots, \, \bv{e}_{r}])$, $S_2 = S_4 = \range([ \bv{e}_{r+1}, \, \bv{e}_{r+2}, \, \ldots, \, \bv{e}_{2r}])$. Therefore, $\xir{r}(P_4) = 2r$. Since $\overline{P_4} = P_4$ and $\mrr{r}^+(P_4) = 2r+1$ (Example \ref{ex-mrr-of-P4}), we have $2r = \xir{r}(\overline{P_4}) < \mrr{r}^+(P_4) = 2r+1$.
}
\end{Example}
Recall from Corollary \ref{xifleqmrf} that $\xi_f(\overline{G}) \leq \mr_f^+(G)$ for any graph $G$. While strict inequality may hold in the $r$-fold case for an arbitrary graph $G$, we now demonstrate that equality always holds in the ``fractional case" for any graph $G$. For this result, we require the following lemma.
\begin{Lemma} \label{repswithineps}
Let $G$ be a graph with complement $\overline{G}$. Let $\{P_u\}_{u \in V(G)}$ be a $d/r$-representation for $\overline{G}$ and let $\{R_u\}_{u \in V(G)}$ be a faithful $b/1$-representation for $G$. Then for any $k \in \mathbb{N}$, $G$ has a faithful $(kd+b)/(kr+1)$-representation $\{Q_u\}_{u \in V(G)}$. Further, given any $\eps > 0$, $k$ can be chosen such that $\left| \frac{d}{r} - \frac{kd+b}{kr+1} \right| < \eps$, i.e., the value of the faithful representation $\{Q_u\}$ for $G$ is within $\eps$ of the value of the (non-faithful) representation $\{P_u\}$ for $\overline{G}$.
\end{Lemma}
\begin{proof}
Since $\{P_u\}$ is a $d/r$-representation for $\overline{G}$, we have $P_u \in \mathbb{C}^{d \times d}$ with $\rank P_u = r$ for each $u \in V(\overline{G}) = V(G)$, and $P_u P_v = 0$ if $uv \in E(\overline{G})$, so $P_u P_v = 0$ if $uv \notin E(G)$.
Let $\eps > 0$ be arbitrary and choose $k > \left( \frac{|d-rb|}{r^2 \eps} - \frac{1}{r} \right)$.
\smallskip
For each vertex $u \in V(G)$, let $Q_u \in \mathbb{C}^{(kd+b) \times (kd+b)}$ be the block diagonal matrix constructed from $k$ copies of $P_u$ and one copy of $R_u$, i.e.,
\[ Q_u = \left(\bigoplus_{i=1}^k P_u \right) \oplus R_u. \]
We see immediately that $\rank Q_u = kr+1$, and since $P_u$ and $R_u$ are projectors, so is $Q_u$. Since $P_u P_v = 0$ if $uv \notin E(G)$ and $R_u R_v = 0$ if \emph{and only if} $uv \notin E(G)$, we conclude that $Q_u Q_v = 0$ if and only if $uv \notin E(G)$. Therefore, $\{Q_u\}_{u \in V(G)}$ is a faithful $(kd+b)/(kr+1)$-representation for $G$, which verifies the first claim.
By choice of $k$, we have $kr+1 > \frac{|d-rb|}{r \eps}$. Consider
\begin{align*}
\left| \frac{d}{r} - \frac{kd+b}{kr+1} \right|
&= \left| \frac{d(kr+1) - r(kd+b)}{r(kr+1)} \right| \\
&= \frac{|d-rb|}{r} \cdot \frac{1}{kr+1} \\
&< \frac{|d-rb|}{r} \cdot \frac{r\eps}{|d-rb|} \\
&= \eps,
\end{align*}
which verifies the second claim. \qquad
\end{proof}
It was previously noted that any faithful $d/r$-representation for $G$ is also $d/r$-representation for $\overline{G}$. Lemma \ref{repswithineps} is a partial converse in the sense that, given any $d/r$-representation for $\overline{G}$, we can construct a faithful $d_1/r_1$-representation for $G$ such that the two representations have essentially the same value. This yields the next result.
\begin{Theorem}
For every graph $G$ with complement $\overline{G}$,
\[ \xi_f(\overline{G}) = \mr_f^+(G). \]
\end{Theorem}
\begin{proof}
Let
\[ R = \left\{ \frac{d}{r} : \overline{G} \text{ has a $d/r$-representation} \right\}, \]
\[ F = \left\{ \frac{d}{r} : G \text{ has a faithful $d/r$-representation} \right\}. \]
For any $\frac{d}{r} \in R$ and $\eps > 0$, Lemma \ref{repswithineps} asserts that there exists some $\frac{d_1}{r_1} \in F$ such that $\left| \frac{d}{r} - \frac{d_1}{r_1} \right| < \eps$. It follows that $\inf R = \inf F$, i.e., $\xi_f(\overline{G}) = \mr_f^+(G)$. \qquad
\end{proof}
\section*{Acknowledgements}
Some of this work was done while Leslie Hogben was a general member of the Institute for Mathematics and its Applications (IMA) and during a week-long visit of Kevin Palmowski to IMA; they thank IMA both for financial support (from NSF funds) and for providing a wonderful collaborative research environment.
David E. Roberson is supported in part by the Singapore National Research Foundation under NRF RF Award No. NRF-NRFF2013-13.
Simone Severini is supported by the Royal Society and EPSRC.
\section{Introduction}
This paper deals with fractional versions of graph parameters defined by orthogonal representations, including minimum positive semidefinite rank. In Section \ref{sec-OSR-and-PR}, we extend the existing idea of an orthogonal representation for a graph via a ``higher-dimensional" construction. With this, we introduce a new parameter, $r$-fold orthogonal rank, that is to orthogonal rank as $b$-fold chromatic number is to chromatic number (see Section \ref{sec-intro-frac} for the definition of $b$-fold chromatic number and other terms related to fractional chromatic number). This allows us to formally characterize projective rank as ``fractional orthogonal rank," a concept that was previously understood (e.g., in \cite{Robersonthesis, MR12}) but not rigorously presented (formal definitions of projective rank and other parameters are given in Section \ref{sec-intro-defs}).
In Section \ref{sec-frac-PSD-mr}, we apply this ``fractionalization" process to the minimum positive semidefinite rank problem (viewed via faithful orthogonal representations) and develop two new graph parameters, namely, $r$-fold and fractional minimum positive semidefinite rank. We also provide an alternate definition of $r$-fold minimum positive semidefinite rank that is based on the minimum rank of a matrix that ``$r$-fits" a graph, allowing us to view the ``higher-dimensional" problem through either of the two viewpoints traditionally associated with the classical minimum positive semidefinite rank problem.
Our final result, found in Section \ref{subsec-equality}, shows that the fractional minimum positive semidefinite rank of a graph is equal to the projective rank of the complement of the graph. This result serves to connect the two seemingly different problems; moving forward, this will allow the extensive existing literature on minimum positive semidefinite rank to be used to inform new developments in the more recently introduced area of projective rank.
In the remainder of this introduction we discuss applications of the fractional parameters discussed (Section \ref{sec-intro-apps}), give a brief introduction to the fractional approach to chromatic number to motivate our definitions (Section \ref{sec-intro-frac}), and provide necessary notation and terminology (Section \ref{sec-intro-defs}).
\subsection{Applications} \label{sec-intro-apps}
Linear algebraic structures and associated graph theoretic frameworks have recently become more important tools to
study the fundamental differences that characterize theories of nature, like
classical mechanics, quantum mechanics, and general probabilistic theories.
Matrices, graphs, and their related combinatorial optimization techniques turn out to
provide a surprisingly general language with which to approach questions connected with
foundational ideas, such as the analysis of contextual inequalities and non-local
games \cite{cabello, cameron}, and with concrete aspects, such as quantifying
various capacities of entanglement-assisted channels \cite{duan, leung}, and
the overhead needed to classically simulate quantum computation \cite{emerson}.
A point of strength of such frameworks is their ability to reformulate
mathematical questions in a coarser manner that is nonetheless effective, in
some cases, to single out specific facts. Tsirelson's problem \cite{tsirelson}
provides a remarkable example: deciding whether the mathematical models of
non-relativistic quantum mechanics, where observers have linear operators acting on a
finite dimensional tensor product space, and algebraic quantum field theory,
where observers have commuting linear operators on a single (possibly infinite dimensional) space, produce the same set of correlations. We know that if Tsirelson's
problem has a positive answer then the notorious Connes' Embedding conjecture
\cite{capraro, ozawa}, originally concerned with an approximation property for
finite von Neumann algebras, is true.
Tsirelson's problem can be seen from a combinatorial matrix point of view by working with
graphs and their associated algebraic structures \cite{paulsen1}. Roughly
speaking, instead of constructing sets of correlation matrices, we can try
looking for various patterns of zeroes in the sets, as in the spirit of
combinatorial matrix theory. The projective rank, denoted $\xi_{f}$,
is a
recently introduced graph parameter with the potential for settling the above
discussion. Indeed, it has been shown that if there exists a graph whose
projective rank is irrational, then Tsirelson's problem has a negative answer
\cite{paulsen2}.
Projective representations and projective rank were originally defined in \cite{MR12} as a tool for studying quantum colorings and quantum homomorphisms of graphs. Quantum colorings and the quantum chromatic number give quantitative measures of the advantage that quantum entanglement provides in performing distributed tasks and in distinguishing scenarios related to classical and quantum physics, respectively. In fact, the existence of a quantum $n$-coloring for a given graph is equivalent to the existence of a projective representation of value $n$ for the Cartesian product of the graph with a complete graph on $n$ vertices.
It was also shown in \cite{MR12} that projective rank is monotone with respect to quantum homomorphisms, i.e., if there exists a quantum homomorphism from a graph $G$ to a graph $H$, then $\xi_f(G) \leq \xi_f(H)$. This shows that projective rank is a lower bound for quantum chromatic number, and more generally provides a method for forbidding the existence of quantum homomorphisms. Indeed, this approach was used to determine the quantum odd girth of the Kneser graphs in \cite{Robersonthesis}. Projective rank has also been studied from a purely graph theoretic point of view, and in \cite{Stahlke14} it was shown that this parameter is multiplicative with respect to the lexicographic and disjunctive graph products. Using this fact the authors were able to find a separation between quantum chromatic number and a recently defined semidefinite relaxation of this parameter, answering a question posed in \cite{paulsen1}.
This paper takes a linear algebraic approach to these questions, building connections between recent graph theoretical approaches to quantum questions and existing literature on orthogonal representations and minimum positive semidefinite rank.
\subsection{A fractional approach} \label{sec-intro-frac}
To demonstrate the fractional approach that we use with orthogonal representations and minimum positive semidefinite rank, consider the following derivation of the fractional chromatic number as found in \cite{fgt}. The \emph{chromatic number} $\chi(G)$ of a graph $G$ is the least number $c$ such that $G$ can be colored with $c$ colors; that is, we can assign to each vertex of $G$ one of $c$ colors in such a way that adjacent vertices receive different colors. A coloring with $c$ colors can be generalized to a \emph{$b$-fold coloring with $c$ colors}, or a $c$:$b$-coloring: from a palette of $c$ colors, assign $b$ colors to each vertex of $G$ such that adjacent vertices receive disjoint sets of colors. For a fixed $b$, the \emph{$b$-fold chromatic number} of $G$, $\chi_b(G)$, is the smallest $c$ such that G has a $c$:$b$-coloring. With this, we can define the \emph{fractional chromatic number} of $G$ as
\[ \chi_f(G) = \inf_b \frac{\chi_b(G)}{b} \, . \]
While it is not obvious, it can be shown that $\chi_f(G)$ is always a rational number, as there is an alternative linear programming formulation for the parameter for which strong duality holds. For further information on fractional coloring, including a time-scheduling interpretation of the problem, see the discussions in the Preface and Chapter 3 of \cite{fgt}.
The process of assigning objects to the vertices of a graph, subject to certain constraints, is a key element common to the problems we examine in this work, and the procedure of generalizing from assigning one object to assigning $b$-many objects (or, in our case, $b$-dimensional or rank-$b$ objects) is an underlying theme. At each stage of the process, we are interested in graph parameters that give information about the ``most efficient" set of objects we can use, with the end goal of developing fractional versions of existing parameters (in the spirit of \cite{fgt}) and connecting the more recent work on projective rank with existing ideas from the realm of minimum positive semidefinite rank.
Rather than the colors used for coloring problems, the objects that we assign to the vertices of a graph are vectors and matrices, which adds a distinctly linear algebraic flavor to both the problems and the constraints: the idea of ``different colors" translates to orthogonality conditions on our objects. As such, our results often see linear algebra and graph theory working hand-in-hand, with structure found in one discipline influencing results that are based in the other.
\subsection{Background, definitions, and notation} \label{sec-intro-defs}
The natural numbers, $\mathbb{N}$, start at 1. We use the notation $\brackets{a}{b}$ to denote the set of integers $\{a, a+1, \ldots, b-1, b\}$. Throughout, $d$ and $r$ are used to represent natural numbers. Vectors are denoted by boldface font, typically $\bv{x}$, and matrices are capital letters, typically $A$, $B$, $P$, or $X$, depending on context. The symbol $0$ denotes either the scalar zero or a zero matrix, and an identity matrix is denoted by $I$; any of these may be subscripted to clarify their sizes. We follow the usual convention of denoting the $j^{th}$ standard basis vector in $\mathbb{C}^d$ (for some $d$) as $\bv{e}_j$. Rows and columns of matrices may be indexed either by natural numbers or by vertices of a graph, depending on context. The elements of a matrix $A$ are denoted $a_{ij}$; if $A$ is a block matrix, then its blocks are denoted $A_{ij}$. Graphs are usually denoted by $G$ or $H$, vertices by $u, v$ or $i, j$, and edges by $uv$ or $ij$.
If $A \in \mathbb{C}^{p \times p}$ and $B \in \mathbb{C}^{q \times q}$, then the \emph{direct sum} of $A$ and $B$, denoted $A \oplus B$, is the block diagonal matrix
\[ \left[ \begin{array}{cc} A & 0 \\ 0 & B \end{array} \right] \in \mathbb{C}^{(p+q) \times (p+q)}. \]
We denote the conjugate transpose of $A$ by $A^*$. A \emph{Hermitian} matrix satisfies $A = A^*$. A Hermitian matrix $A \in \mathbb{C}^{n \times n}$ is \emph{positive semidefinite}, denoted $A \mgeq 0$, if $\bv{x}^* A \bv{x} \geq 0$ for all $\bv{x} \in \mathbb{C}^n$, or equivalently, if all of its eigenvalues are nonnegative.
Typically, $G = (V, E)$ will denote a simple undirected graph on $n$ vertices, where $V = V(G)$ is the set of vertices of $G$ and $E = E(G)$ is the set of edges of $G$. An \emph{isolated vertex} is a vertex that is not adjacent to any other vertex of $G$. A \emph{subgraph} of a graph $G$ is a graph $H$ such that $V(H) \subseteq V(G)$ and $E(H) \subseteq E(G)$. An \emph{induced subgraph} of a graph $G$, denoted $G[W]$ for some set $W \subseteq V(G)$, is a subgraph with vertex set $W$ such that if $u,v \in W$ and $uv \in E(G)$, then $uv \in E(G[W])$. The \emph{union} of graphs $G$ and $H$, denoted $G \cup H$, is the graph with vertex set $V(G \cup H) = V(G) \cup V(H)$ and edge set $E(G \cup H) = E(G) \cup E(H)$. If $V(G) \cap V(H) = \emptyset$, then this union is \emph{disjoint} and denoted $G \dcup H$. The \emph{complement} of $G$, denoted $\overline{G}$, is the graph with $V(\overline{G}) = V(G)$ and $E(\overline{G}) = \{ uv : u \neq v, uv \notin E(G) \}$. An \emph{independent set} in $G$ is a set $W \subseteq V(G)$ such that if $u, v \in W$, then $uv \notin E(G)$. The \emph{independence number} of $G$, denoted $\alpha(G)$, is the largest possible cardinality of an independent set in $G$. A \emph{clique} in $G$ is an induced subgraph $H$ that is a complete graph, i.e., $uv \in E(H)$ for every $u,v \in V(H)$. The \emph{clique number} of $G$, denoted $\omega(G)$, is the largest possible order of a clique in $G$. A \emph{clique-sum} of graphs $G$ and $H$ on $K_t$, i.e., the graph $G \cup H$ where $G \cap H = K_t$, is denoted by $G \cs{t} H$; this is also called a $t$-clique-sum of $G$ and $H$. A \emph{chordal graph} is a graph that does not have any induced cycles of length greater than 3; any chordal graph can be constructed as clique-sum(s) of complete graphs. A \emph{perfect graph} is a graph $G$ for which every induced subgraph $H$ of $G$ satisfies $\omega(H) = \chi(H)$. A \emph{cut-vertex} of a connected graph $G$ is a vertex whose deletion disconnects $G$. A graph with a cut-vertex can be viewed as a $1$-clique-sum.
We work in the vector space $\mathbb{C}^d$ for some $d \in \mathbb{N}$. We use $S$ to denote a subspace of a vector space. A \emph{basis matrix} for an $r$-dimensional subspace $S$ of $\mathbb{C}^d$ is a matrix $X \in \mathbb{C}^{d \times r}$ that has orthonormal columns and satisfies $S = \range(X)$. We say that two subspaces $S_1$ and $S_2$ of $\mathbb{C}^d$ are \emph{orthogonal}, denoted $S_1 \perp S_2$, if $\bv{u}_1^* \bv{u}_2 = 0$ for all $\bv{u}_1 \in S_1$ and all $\bv{u}_2 \in S_2$; an equivalent condition is that $X_1^* X_2 = 0$, where $X_1$ and $X_2$ are basis matrices for $S_1$ and $S_2$, respectively.
Given some graph $G$ and $d \in \mathbb{N}$, an \emph{orthogonal representation} in $\mathbb{C}^d$ for $G$ is a set of unit vectors $\{\bv{x}_u\}_{u \in V(G)} \subset \mathbb{C}^d$ such that $\bv{x}_u^* \bv{x}_v = 0$ if $uv \in E(G)$. It is clear that such a representation always exists for $d = |V(G)|$. Provided that $G$ has at least one edge, it is clear that such a representation cannot be made for $d = 1$. We define the \emph{orthogonal rank} of $G$ to be
\[ \xi(G) = \min\left\{ d : G \text{ has an orthogonal representation in $\mathbb{C}^d$} \right\}. \]
Let $d, r \in \mathbb{N}$ with $r \leq d$. A \emph{$d/r$-projective representation}, or $d/r$-representation, is an assignment of matrices $\{P_u\}_{u \in V(G)}$ to the vertices of $G$ such that
\begin{itemize}
\item for each $u \in V(G)$, $P_u \in \mathbb{C}^{d \times d}$, $\rank P_u = r$, $P_u^* = P_u$, and $P_u^2 = P_u$; and
\item if $uv \in E(G)$, then $P_u P_v = 0$.
\end{itemize}
In words, a $d/r$-representation is an assignment of rank-$r$ $(d \times d)$ orthogonal projection matrices (projectors) to the vertices of $G$ such that adjacent vertices receive projectors that are orthogonal. The \emph{projective rank} of $G$ is defined as
\[ \xi_f(G) = \inf_{d,r} \left\{ \frac{d}{r} : G \text{ has a $d/r$-representation} \right\}. \]
Projective rank was first introduced in 2012 by Roberson and Man\v cinska, where it is noted that $\xi_f(G) \leq \xi(G)$; see \cite{Robersonthesis} and \cite{MR12} for additional information, properties, and applications.
Complementary to the idea of an orthogonal representation is that of a \emph{faithful} orthogonal representation (here we follow the complementary usage in the minimum rank literature). In order for the definitions given next to coincide with those in the minimum rank literature, we must assume that the graph $G$ has no isolated vertices. A \emph{faithful orthogonal representation} in $\mathbb{C}^d$ for a graph $G$ is a set of unit vectors $\{ \bv{x}_u \}_{u \in V(G)} \subset \mathbb{C}^d$ such that $\bv{x}_u^*\bv{x}_v = 0$ if and only if $uv \notin E(G)$. We define the \emph{minimum positive semidefinite rank} of $G$ as
\begin{equation} \label{def-mrplus}
\mr^+(G) = \min \left\{ d : G \text{ has a faithful orthogonal representation in $\mathbb{C}^d$} \right\}.
\end{equation}
We say that a matrix $A \in \mathbb{C}^{n \times n}$ \emph{fits} the order-$n$ graph $G$ if $a_{ii} = 1$ for all $i \in \brackets{1}{n}$, and for all $i \neq j$, we have $a_{ij} = 0$ if and only if $ij \notin E(G)$. Let $\mathcal{H}^+ (G) = \left\{ A \in \mathbb{C}^{n \times n} : A \mgeq 0 \text{ and $A$ fits $G$} \right\}$. A faithful orthogonal representation in $\mathbb{C}^d$ for $G$ corresponds to a matrix $A \in \mathcal{H}^+(G)$ with $\rank A \leq d$, and a matrix $A \in \mathcal{H}^+(G)$ with rank $d$ can be factored as $A = B^* B$ for some $B \in \mathbb{C}^{d \times n}$. Thus an alternate characterization (see, e.g., \cite{HLA2ch46}) of $\mr^+(G)$ is
\[ \mr^+(G) = \min\{ \rank A : A \in \mathcal{H}^+(G) \}, \]
(and in fact, this is the customary definition of this parameter).
The definitions and explanation given here coincide with those in the literature provided that the graph $G$ has no isolated vertices. The most common definition of $\mathcal{H}^+(G)$ in the literature does not contain the assumption that $a_{ii}=1$. If vertex $i$ is adjacent to at least one other vertex, then properties of positive semidefinite matrices require $a_{ii} > 0$, and so $A$ can be scaled by a positive diagonal congruence to a matrix of the same rank and nonzero pattern that has all diagonal entries equal to one. However, consider the case where $G$ consists of $n$ isolated vertices (no edges): then as defined in \cite{BHH11, HLA2ch46}, etc., $\mr^+(G) = 0$, whereas with our definition $\mr^+(G) = n$. The two definitions of minimum positive semidefinite rank coincide precisely when $G$ has no isolated vertices. Our definition facilitates connections to the use of orthogonal rank in the study of quantum issues, and the assumption of no isolated vertices is needed only when connecting to the minimum rank literature, so we omit it except when discussing connections to such work (where we state either this assumption or one that implies it, such as the graph being connected and of order at least two). We also note that for any graph the values of the parameters studied can be computed from their values on the connected components of the graph (see Section \ref{sec-frac-PSD-mr}), which facilitates handling cases with isolated vertices separately.
\section{Orthogonal subspace representations and projective rank} \label{sec-OSR-and-PR}
In this section, we introduce and discuss $(d;r)$ orthogonal subspace representations for a graph $G$, which are extensions of orthogonal representations in the spirit of fractional graph theory \cite{fgt}. The $r$-fold orthogonal rank of a graph, $\xir{r}(G)$, is defined and some properties of this quantity are examined. We then relate these representations to $d/r$-projective representations and tie projective rank into the new theory, formalizing the existing understanding that projective rank and ``fractional orthogonal rank" are one and the same.
Unless otherwise specified, all matrices and vectors in this section are assumed to be complex-valued.
\subsection{Orthogonal subspace representations and $\mathbit{r}$-fold orthogonal rank}
Let $G$ be a graph and let $d, r \in \mathbb{N}$ with $d \geq r$. A \emph{$(d;r)$ orthogonal subspace representation}, or $(d;r)$-OSR, for $G$ is a set of subspaces $\{S_u\}_{u \in V(G)}$ such that
\begin{itemize}
\item for each $u \in V(G)$, $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^d$; and
\item if $uv \in E(G)$, then $S_u \perp S_v$.
\end{itemize}
The \emph{$r$-fold orthogonal rank} of a graph $G$ is defined by
\[ \xir{r}(G) = \min \left\{ d : G \text{ has a $(d;r)$ orthogonal subspace representation} \right\}. \]
An orthogonal representation in $\mathbb{C}^d$ naturally generates a $(d;1)$ orthogonal subspace representation, and vice versa, so $\xi(G) = \xir{1}(G)$.
We now explore some properties of $\xir{r}(G)$.
\begin{Lemma} \label{xirsubadditive}
$\xir{r}$ is a subadditive function of $r$, i.e., for every graph $G$ and all $r, s \in \mathbb{N}$,
\[ \xir{r+s}(G) \leq \xir{r}(G) + \xir{s}(G). \]
\end{Lemma}
\begin{proof}
Let $d_r = \xir{r}(G)$ and $d_s = \xir{s}(G)$. Then $G$ has a $(d_r ;r)$ orthogonal subspace representation containing $r$-dimensional subspaces of $\mathbb{C}^{d_r}$, say $\{S^r_u\}_{u \in V(G)}$, and a $(d_s; s)$ orthogonal subspace representation containing $s$-dimensional subspaces of $\mathbb{C}^{d_s}$, say $\{S^s_u\}_{u \in V(G)}$. We show by construction that there exists an orthogonal subspace representation for $G$ containing $(r+s)$-dimensional subspaces of $\mathbb{C}^{d_r + d_s}$.
For each $u \in V(G)$, let $X_u^r \in \mathbb{C}^{d_r \times r}$ and $X_u^s \in \mathbb{C}^{d_s \times s}$ be basis matrices for $S_u^r$ and $S_u^s$, respectively. Define
\[ X_u = \left[ \begin{array}{cc} X_u^r & 0_{d_r \times s} \\ 0_{d_s \times r} & X_u^s \end{array} \right] \in \mathbb{C}^{(d_r+d_s) \times (r+s)} \]
and let $S_u = \range(X_u)$. We immediately see that $S_u$ is a subspace of $\mathbb{C}^{d_r + d_s}$, $X_u$ is a basis matrix for $S_u$, and $\dim(S_u) = \rank X_u = \rank X_u^r + \rank X_u^s = r+s$.
Suppose $u, v \in V(G)$ and let $X_u^r$, $X_v^r$, $X_u^s$, $X_v^s$, $X_u$, and $X_v$ be as above; then
\[ X_u^* X_v = \left[ \begin{array}{cc}
(X_u^r)^*(X_v^r) & 0 \\
0 & (X_u^s)^*(X_v^s)
\end{array} \right]. \]
Suppose $uv \in E(G)$. Since $\{S_u^r\}$ is an orthogonal subspace representation, we have $(X_u^r)^*(X_v^r) = 0$; similarly, $(X_u^s)^*(X_v^s) = 0$, so $X_u^* X_v = 0$. Since $X_u$ and $X_v$ are basis matrices for $S_u$ and $S_v$, respectively, we conclude that if $uv \in E(G)$, then $S_u \perp S_v$.
Thus $\{S_u\}_{u \in V(G)}$ is a $(d_r + d_s; r+s)$ orthogonal subspace representation for $G$, so $\xir{r+s}(G) \leq d_r + d_s = \xir{r}(G) + \xir{s}(G)$. \qquad
\end{proof}
\begin{Corollary}
For every graph $G$ and all $r \in \mathbb{N}$, $\frac{\xir{r}(G)}{r} \leq \xi(G)$.
\end{Corollary}
\begin{proof}
Since $\xir{1}(G) = \xi(G)$, we have
\[ \xir{r}(G) \leq \xir{r-1}(G) + \xi(G) \leq \ldots \leq r \cdot \xi(G) . \qedhere\]
\end{proof}
\begin{Observation} \label{xir-geq-omega}
For every graph $G$ and all $r \in \mathbb{N}$, $\xir{r}(G) \geq r \cdot \omega(G)$.
\end{Observation}
\begin{Proposition} \label{xir-subgraph}
Let $r \in \mathbb{N}$ and let $H$ be a subgraph of $G$. Then $\xir{r}(H) \leq \xir{r}(G)$.
\end{Proposition}
\begin{proof}
Since every edge of $H$ is an edge of $G$, any $(d;r)$ orthogonal subspace representation for $G$ provides a $(d;r)$ orthogonal subspace representation for $H$, and the result is immediate. \qquad
\end{proof}
\begin{Proposition} \label{xirdisjointmax}
Suppose $r \in \mathbb{N}$ and $G = \bigdcup_{i=1}^t G_i$ for some graphs $\{G_i\}_{i=1}^t$. Then $\xir{r}(G) = \max_i \left\{ \xir{r}(G_i) \right\}$.
\end{Proposition}
\begin{proof}
Since each $G_i$ is an induced subgraph of $G$, we have $\xir{r}(G_i) \leq \xir{r}(G)$ for each $i$, so $\max_i \left\{ \xir{r}(G_i) \right\} \leq \xir{r}(G)$.
\smallskip
For each $i \in \brackets{1}{t}$, let $d_i = \xir{r}(G_i)$ and let $d = \max_i \{d_i\}$. Let $\{S_u^i\}_{u \in V(G_i)}$ be a $(d_i; r)$ orthogonal subspace representation for $G_i$ and for each vertex $u \in V(G_i)$ let $X_u^i \in \mathbb{C}^{d_i \times r}$ be a basis matrix for $S_u^i$. For each $u \in V(G)$, we have $u \in V(G_i)$ for some $i$; define
\[ S_u = \range \left[ \begin{array}{c} X_u^i \\ 0_{(d-d_i) \times r} \end{array} \right]. \]
Each $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^d$, and if $uv \in E(G)$, then $uv \in E(G_k)$ for some $k$, so $S_u^k \perp S_v^k$, which implies that $S_u \perp S_v$ (by construction). Therefore, $\{ S_u \}_{u \in V(G)}$ is a $(d;r)$-OSR for $G$, so $\xir{r}(G) \leq d = \max_i \{ \xir{r}(G_i) \}$ and equality follows. \qquad
\end{proof}
This result does not hold for arbitrary graph unions, as the following example for the $r = 1$ case shows.
\begin{Example} {\rm
Let $G = C_5$ with $V(G) = \{1,2,3,4,5\}$ and $E(G) = \{12, 23, 34, 45, 51\}$. Define $G_1 = P_4$ with $V(G_1) = \{1,2,3,4\}$ and $E(G_1) = \{12, 23, 34 \}$ and define $G_2 = P_3$ with $V(G_2) = \{4,5,1\}$ and $E(G_2) = \{45, 51\}$. We see that $G = G_1 \cup G_2$, but since $\xi(P_3) = \xi(P_4) = 2$ and $\xi(C_5) = 3$, it is not true that $\xi(G) = \max\{ \xi(G_1), \, \xi(G_2)\}$.
}
\end{Example}
While the maximum property observed in Proposition \ref{xirdisjointmax} may not carry over to the case when $G$ is a nondisjoint union of graphs, we are still able to obtain a weaker result, which follows.
\begin{Proposition}
Suppose $r \in \mathbb{N}$ and $G = \bigcup_{i=1}^t G_i$, where $G_i$ is an induced subgraph of $G$ for each $i$. Then $\xir{r}(G) \leq \sum_{i=1}^t \xir{r}(G_i)$.
\end{Proposition}
\begin{proof}
We prove the result for the case where $t = 2$ and note that recursive application of this case will prove the more general one.
For each $i \in \{1,2\}$, let $d_i = \xir{r}(G_i)$ and $\{S_u^i\}_{u \in V(G_i)}$ be a $(d_i;r)$-OSR for $G_i$, and for each $u \in V(G_i)$, let $X_u^i \in \mathbb{C}^{d_i \times r}$ be a basis matrix for $S_u^i$.
We partition $V(G) = V(G_1) \cup V(G_2)$ into three disjoint sets and consider vertices in each set. If $u \in V(G_1) \setminus V(G_2)$, let
\[ X_u = \left[ \begin{array}{c} X_u^1 \\ 0_{d_2 \times r} \end{array} \right]; \]
if $u \in V(G_2) \setminus V(G_1)$, let
\[ X_u = \left[ \begin{array}{c} 0_{d_1 \times r} \\ X_u^2 \end{array} \right]; \]
and if $u \in V(G_1) \cap V(G_2)$, let
\[ X_u = \left[ \begin{array}{c} X_u^1 \\ X_u^2 \end{array} \right]. \]
For each $u \in V(G)$, let $S_u = \range(X_u)$. Each $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^{d_1+d_2}$.
We consider multiple cases to show that if $uv \in E(G)$, then $X_u^* X_v = 0$, so $S_u \perp S_v$. Throughout, we assume that $uv \in E(G)$.
First, suppose that $u \in V(G_1) \setminus V(G_2)$; then either $v \in V(G_1) \setminus V(G_2)$ or $v \in V(G_1) \cap V(G_2)$. In either case, $uv \in E(G_1)$ (since $G_1$ is an induced subgraph), and block multiplication yields $X_u^* X_v = (X_u^1)^* X_v^1$. Since $S_u^1 \perp S_v^1$, this quantity equals the zero matrix, so $S_u \perp S_v$. The case where $u \in V(G_2) \setminus V(G_1)$ is similar.
If $u, v \in V(G_1) \cap V(G_2)$, then $uv \in E(G_1) \cap E(G_2)$ since $G_1$ and $G_2$ are induced subgraphs. Then $X_u^* X_v = (X_u^1)^* X_v^1 + (X_u^2)^* X_v^2$. Since $S_u^1 \perp S_v^1$ and $S_u^2 \perp S_v^2$, this quantity is again the zero matrix, so $S_u \perp S_v$.
Therefore, $\{S_u\}_{u \in V(G)}$ is a $(d_1+d_2;r)$-OSR for $G$, so $\xir{r}(G) \leq d_1 + d_2 = \xir{r}(G_1) + \xir{r}(G_2)$. \qquad
\end{proof}
\begin{Lemma} \label{stdcliquerep}
Suppose that the complete graph $K_t$ is a subgraph of $G$ with $V(K_t) = \brackets{1}{t}$ and $G$ has a $(d;r)$ orthogonal subspace representation. Then $d \geq rt$ and $G$ has a $(d;r)$ orthogonal subspace representation in which the vertex $i \in V(K_t)$ is represented by
\[ \vspan \left\{ \bv{e}_{(i-1)r+1}, \ldots , \bv{e}_{(i-1)r+r-1},\bv{e}_{ir} \right\}. \]
\end{Lemma}
\begin{proof}
By Observation \ref{xir-geq-omega}, $d \geq r \cdot \omega(G) \geq rt$.
If $M \in \mathbb{C}^{d \times \ell}$ for some $\ell \leq d$ and the columns of $M$ are orthonormal, then by a change of orthonormal basis there exists a unitary matrix $U \in \mathbb{C}^{d \times d}$ such that $UM = [ \bv{e}_{1}, \ldots, \bv{e}_{\ell}]$.
Let $\{S_u\}_{u \in V(G)}$ be a $(d;r)$ orthogonal subspace representation for $G$ and for each $u \in V(G)$ let $X_u$ be a basis matrix for $S_u$. Define $M = [X_1, \ldots, X_t]$ and choose $U$ so that $UM=[ \bv{e}_1 , \ldots, \bv{e}_{tr}]$. Define $S_u' = \range(UX_u)$. Then $\left\{ S_u' \right\}_{u\in V(G)}$ is a $(d;r)$ orthogonal subspace representation for $G$ with the desired property. \qquad
\end{proof}
\begin{Theorem}
If $G = G_1 \cs{t} G_2$ and $r \in \mathbb{N}$, then $\xir{r}(G) = \max \left\{ \xir{r}(G_1), \xir{r}(G_2) \right\}$.
\end{Theorem}
\begin{proof}
Without loss of generality, let $d_1 = \xir{r}(G_1) \geq d_2 = \xir{r}(G_2)$ and $V(K_t) = \brackets{1}{t}$. Then by Lemma \ref{stdcliquerep}, for $i = 1,2$, each $G_i$ has a $(d_1; r)$ orthogonal subspace representation, $\{S_u^i\}_{u\in V(G)}$, in which vertex $v \leq t$ is represented by $S_v^i = \vspan \left\{ \bv{e}_{(v-1)r+1}, \ldots, \bv{e}_{(v-1)r+r-1}, \bv{e}_{vr} \right\}$. Thus for $v \in \brackets{1}{t}$, $S_v^1 = S_v^2$; denote this common subspace by $S_v$.
For vertices $u\in V(G_i) \setminus \brackets{1}{t}$, define $S_u = S_u^i$ (observe that $u > t$ is in only one of $V(G_1)$ or $V(G_2)$). Then $\{S_u\}_{u\in V(G)}$ is a $(d_1;r)$ orthogonal subspace representation for $G$. \qquad
\end{proof}
\begin{Proposition} \label{prop-xir-equals-romega}
If $G$ is a graph with $\omega(G) = \chi(G)$, then $\xir{r}(G) = r \cdot \omega(G)$ for every $r \in \mathbb{N}$.
\end{Proposition}
\begin{proof}
It is well-known that $\xi(G) \leq \chi(G)$ (see, e.g., \cite{Robersonthesis}). Therefore,
\[ r \cdot \omega(G) \leq \xir{r}(G) \leq r \cdot \xi(G) \leq r \cdot \chi(G) = r \cdot \omega(G) \]
and thus equality holds throughout.
\end{proof}
We note that perfect graphs and chordal graphs are among those that satisfy $\omega(G) = \chi(G)$, and so Proposition \ref{prop-xir-equals-romega} applies to these classes.
\begin{Remark} {\rm
Since $\xir{1}(G) = \xi(G)$ for every graph $G$, the previous properties of $r$-fold orthogonal rank also apply to orthogonal rank, where appropriate.
}
\end{Remark}
\subsection{Projective rank as fractional orthogonal rank}
It is easy to see that $(d;r)$ orthogonal subspace representations are closely related to $d/r$-\break representations; in fact, they are in one-to-one correspondence.
\begin{Proposition} \label{drOSRiffdr}
A graph $G$ has a $(d;r)$ orthogonal subspace representation if and only if $G$ has a $d/r$-representation.
\end{Proposition}
\begin{proof}
Suppose that $G$ has a $(d;r)$ orthogonal subspace representation $\{S_u\}_{u \in V(G)}$, so each $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^d$. For each $u \in V(G)$, define $P_u =X_u X_u^*$, where $X_u \in \mathbb{C}^{d \times r}$ is a basis matrix for $S_u$. It is then easy to verify that $P_u \in \mathbb{C}^{d \times d}$, $\rank P_u = \rank X_u = r$, $P_u^* = P_u$, and $P_u^2 = P_u$.
Let $uv \in E(G)$, so $S_u \perp S_v$. We see that
\[ S_u \perp S_v \iff X_u^* X_v = 0 \iff X_u X_u^* X_v X_v^* = 0 \iff P_u P_v = 0. \]
Thus if $uv \in E(G)$, then $P_u P_v = 0$. We conclude that $\{P_u\}_{u \in V(G)}$ is a $d/r$-representation for $G$.
For the converse, suppose that $\{P_u\}_{u \in V(G)}$ is a $d/r$-representation for $G$. For each $u \in V(G)$, let $P_u = X_u I_r X_u^*$ be a reduced singular value decomposition of the projector $P_u$ (where $X_u \in \mathbb{C}^{d \times r}$) and define $S_u = \range(P_u) = \range(X_u)$. Clearly $S_u$ is an $r$-dimensional subspace of $\mathbb{C}^d$. If $uv \in E(G)$, then $P_u P_v = 0$, so by the above chain of equivalences $S_u \perp S_v$. Therefore, $\{S_u\}_{u \in V(G)}$ is a $(d;r)$ orthogonal subspace representation for $G$. \qquad
\end{proof}
With this in mind, we obtain the following ``fractional" definition of projective rank.
\begin{Theorem} \label{xifandxir} For every graph $G$,
\[ \xi_f(G) = \inf_r \left\{ \frac{\xir{r}(G)}{r} \right\}. \]
\end{Theorem}
\begin{proof}
\begin{align*}
\inf_r \left\{ \frac{\xir{r}(G)}{r} \right\}
&= \inf_r \left\{ \frac{ \min\{ d : G \text{ has a $(d;r)$-OSR} \}}{r} \right\} \\
&= \inf_r \left\{ \min_d \left\{ \frac{d}{r} : G \text{ has a $(d;r)$-OSR} \right\} \right\} \\
&= \inf_{d,r} \left\{ \frac{d}{r} : G \text{ has a $(d;r)$-OSR} \right\} \\
&= \inf_{d,r} \left\{ \frac{d}{r} : G \text{ has a $d/r$-representation} \right\} \\
&= \xi_f(G). \qedhere
\end{align*}
\end{proof}
Given that this expression of $\xi_f(G)$ is similar to that of $\chi_f(G)$ given in \cite{fgt}, it is not unreasonable to hope that this could shed some light on the question of the rationality of $\xi_f(G)$ for all graphs.\footnote{Recall that $\chi_f(G)$ is rational for any graph $G$.} Unfortunately, finding a $b$-fold coloring with $c$ colors for $G$ is ultimately a far different problem from finding a $(d;r)$ orthogonal subspace representation for $G$. In the $b$-fold coloring problem, we have a restriction on the number of available colors, which adds a certain finiteness to the problem: each vertex is assigned a subset of the available $c < \infty$ colors. In contrast, restricting the subspaces to lie in $\mathbb{C}^d$ in the orthogonal subspace representation problem does not impose this same type of finiteness: each vertex is assigned a finite dimensional subspace of $\mathbb{C}^d$, and $d < \infty$, but there are infinitely many subspaces that can be assigned to each vertex.
We provide one additional equivalent definition of projective rank, for which we need the following utility result from \cite{fgt}, also commonly known as Fekete's Lemma.
\begin{Lemma}[\cite{fgt}, Lemma A.4.1] \label{fgtA41}
Suppose $g : \mathbb{N} \to \mathbb{R}$ is subadditive and $g(n) \geq 0$ for all $n$. Then the limit
\[ \lim_{n \to \infty} \frac{g(n)}{n} \]
exists and is equal to the infimum of $g(n) / n$ $(n \in \mathbb{N})$.
\end{Lemma}
Since $\xir{r}$ is subadditive, this yields the following corollary to the previous theorem.
\begin{Corollary} \label{xirinfequalslim} For every graph $G$,
\[ \xi_f(G) = \inf_r \left\{ \frac{\xir{r}(G)}{r} \right\} = \lim_{r \to \infty} \frac{\xir{r}(G)}{r} \, , \]
and this limit exists.
\end{Corollary}
With this result, we see that many of the properties of $\xir{r}(G)$ also apply to $\xi_f(G)$.
\begin{Theorem} \label{xifproperties}
For every graph $G$:
\begin{enumerate}[i)]
\item {\rm \cite{Robersonthesis, MR12} } $\xi_f(G) \geq \omega(G)$.
\item If $H$ is a subgraph of $G$, then $\xi_f(H) \leq \xi_f(G)$.
\item If $G = \bigdcup_{i=1}^t G_i$ for some graphs $\{ G_i \}_{i=1}^t$, then $\xi_f(G) = \max_i \left\{ \xi_f(G_i) \right\}$.
\item If $G = \bigcup_{i=1}^t G_i$ for some induced subgraphs $\{ G_i \}_{i=1}^t$, then $\xi_f(G) \leq \sum_{i=1}^t \xi_f(G_i)$.
\item If $G = G_1 \cs{t} G_2$, then $\xi_f(G) = \max \left\{ \xi_f(G_1), \xi_f(G_2) \right\}$.
\item If $G$ satisfies $\omega(G) = \xi(G)$, then $\xi_f(G) = \omega(G)$.
\end{enumerate}
\end{Theorem}
\begin{proof}
Consider the second claim. By Proposition \ref{xir-subgraph}, for any $r \in \mathbb{N}$, $\xir{r}(H) \leq \xir{r}(G)$, so $\frac{\xir{r}(H)}{r} \leq \frac{\xir{r}(G)}{r}$. Taking the limit as $r$ approaches $\infty$ and applying Corollary \ref{xirinfequalslim}, we have $\xi_f(H) \leq \xi_f(G)$.
The remaining claims follow by applying similar arguments to the corresponding $r$-fold results.
\end{proof}
|
\section{Introduction}
In nature and in our daily life, many soft materials are
formed due to the dynamical arrest of the constituent particles
\cite{book,witten,tartaglia}. Usually they are labelled as gels if
the particle density is low and as glasses if the density is large.
However, the difference between these two states is at present not very
well understood and therefore it is not always easy to distinguish them.
Despite this difficulty quite a few features in this glass-gel cross-cover
regime have been studied extensively.
In dense glass forming liquids the slowing down of dynamics is related to
the mutual steric-hindrance in the motion of the constituent particles.
The dynamical properties of these glass formers are characterized by
a stretched-exponential shape of relaxation functions \cite{book},
or similarly by the anomalous, exponential tails in the van Hove
distributions of particle displacements \cite{Pinaki}. Many of
these dynamical features are described well by mode-coupling theory
\cite{book}. At even lower temperatures, the relaxation dynamics can
be understood by means of the random first order transition theory
\cite{rfot}. In these glass formers, the structure is usually close-packed
for hard sphere or van der Waals type interactions and is accompanied
by a super-Arrhenius increase of viscosity. Or, if there are covalent
bondings, they form network-like structures and exhibit an Arrhenius
increase in viscosity.
On the other hand, chemical gels or rubbers are soft solids having random
network structures \cite{witten,larson,zaccarellirev}. The cross linking
of the permanent bonds between the constituent monomers happens during the
synthesis process inducing a vulcanization transition once the density of
the links exceeds the percolation threshold~\cite{zippeliusrev}. Different
static and dynamical properties in the vicinity of this transition
have been studied, both using simulations and theoretical models (e.g.,
see \cite{coniglio1, zippelius1}). Physical gels are on the other hand
low-density network structures with bonds that can be broken/realigned by
thermal fluctuations within finite timescales~\cite{zaccarellirev}. One
possible nonequilibrium path to physical gelation is via a thermal
quench across the liquid-gas spinodal leading to dynamically
arrested states \cite{zaccarelli2, paddy0, paddy1,paddy2}, that
show complex aging phenomena \cite{testard,suarez}. In general,
these paths lead to spatially heterogeneous structures. However,
in recent times, considerable effort has been made to devise
ways by which spatially homogeneous physical gels can be formed
\cite{pablo,porte,pablo2,pre2010,shibu,delgado1,coniglio1,coniglio2,miller,patch1,patch2}.
For such gel-forming systems, a wide variety of relaxation functions have
been reported: logarithmic \cite{zaccarelli3, zaccarelli4, puertas,log},
stretched \cite{suarez} or compressed exponentials \cite{duri, shibu}.
While theoretical models \cite{fuchs, dawson, kroy, patch2} have been
proposed to account for such dynamical properties, they are at this time
certainly not yet comprehensive.
Of particular interest is the interplay between these different arrest
mechanisms, viz. gel and glass, since their competition can be used
to engineer materials with novel functionalities \cite{pre2010, zhao, hoffmann, royalltanaka}.
Similar studies have been
carried out in systems with competing lengthscales \cite{moreno,sperl}
or interactions \cite{reviewfuchs,moreno-polymer,zaccarelli3}. Yet,
only few models do allow to study the low density gel phase and high density
glass at the same time. Accessing this regime is, however, necessary for
investigating the structure and dynamics at those intermediate densities
where the gel transforms into a glass and vice versa. Here, we study
a simple model with direct experimental relevance \cite{porte}, and
which permits us to traverse the density regime of interest and hence
to study the interplay of different processes which lead to either
gelation or glassiness. Furthermore, our model also allows us to tune
the lifetime of bonds, a feature that is usually not present in other
models (for example, see the recent work \cite{Nagi}). On one hand,
this facilitates a wider exploration of the relaxation dynamics of such
model physical gels, but also allows to disentangle the origin of the
apparently anomalous relaxation observed in these systems. In fact,
mode-coupling theory predicts, e.g., logarithmic relaxation whenever
two different arrest lines meet, one gel-like and another glass-like,
and relates it to an underlying higher-order singularity in the theory
\cite{dawson}. The versatility of our model, and in particular the
possibility of tuning at will the bonds lifetime, allows us to explore
the different mechanisms at play and hence to elucidate the interplay
of various lengths and time-scales \cite{pre2010}.
The paper is structured as follows. In Section II we explain the
details of our model transient-network fluid, together with the
numerical schemes used to simulate its dynamics. The phase diagram
of the model system and its structural properties are discussed in
Sections III and IV, respectively. In Section V, we analyze in
details the fluid's dynamics, quantified by the mean squared
displacement and the incoherent scattering function. For that, we
follow different routes across the phase diagram, which allows us
to clearly understand the interplay between the gel and the glass
regimes. Finally, the dynamical heterogeneities which characterize
the slow dynamics in the gel and glass phases are studied in Section
VI, followed by a summary of our results and a broader perspective
in Section VII.
\section{Model and details of simulation}
Our model system is a coarse-grained representation~\cite{pablo} of a
transient gel which has been studied in experiments~\cite{porte}. In this
system, an equilibrium low-density gel is obtained by adding telechelic
polymers to an oil-in-water microemulsion. Since the polymer end-groups
are hydrophobic, the polymers effectively act as (attractive) bridges
between the oil droplets they connect. The strength, lengthscale and
typical lifetime of these bridging polymers can be controlled at will.
Denoting by $C_{ij}$ the number of polymers connecting droplets $i$
and $j$, we have established in Refs.~\cite{pablo,pablo2}, that the
following interaction is a reasonable coarse-grained representation of
this ternary system:
\begin{equation}
V = \epsilon_1 \sum_{j > i}
\left( \frac{\sigma_{ij}}{r_{ij}} \right)^{14}
+ \epsilon_2 \sum_{j > i}
C_{ij} V_{\rm FENE}(r_{ij})
+\epsilon_{0} \sum_{i} C_{ii}.
\label{model}
\end{equation}
\noindent
The first term is a soft repulsion acting between bare oil droplets, where
$\sigma_{ij} = (\sigma_i + \sigma_j)/2$, $\sigma_i$ is the diameter of
droplet $i$, and $r_{ij}$ is the distance between the droplet centers.
The second term describes the entropic attraction induced by the
telechelic polymers, which has the standard ``FENE'' (finitely-extensible
nonlinear elastic) form known from polymer physics~\cite{witten},
$V_{\rm FENE}(r_{ij}) = \ln ( 1 - (r_{ij}-\sigma_{ij})^2/\ell^2)$,
and accounts for the maximal extension $\ell$ of the polymers. The
last term introduces the energy penalty $\epsilon_0$ for polymers
that have both end-groups in the same droplet. The most drastic
approximation of the model (\ref{model}) is the description of the
polymers as effective bonds between the droplets, which is justified
whenever the typical lifetime of the bonds is much larger than the
timescale for polymer dynamics in the solvent~\cite{porte}. Thus,
for exploring the different properties of such a model, the relevant
variables are the droplet volume fraction $\phi$ and the number of
polymer heads per droplet $R$ \cite{pablo}. In order to describe the
dynamics of the system, we use a combination of molecular dynamics to
propagate the droplets with the interaction (\ref{model}), and local
Monte Carlo moves with Metropolis acceptance rates $\tau_{\rm link}^{-1}
{\rm min}[1,\exp(\Delta V/k_BT)]$ to update the polymer connectivity
matrix $C_{ij}$, where $\Delta V$ is the difference in potential energy
of the system for the two bond configurations~\cite{pablo,pablo2}. Thus
$\tau_{\rm link}$ is the timescale governing the renewal of the polymer
network topology. In order to prevent crystallization at high volume
fractions (which would be the case for the monodisperse model studied
earlier \cite{pablo}), we use a polydisperse emulsion with a flat
distribution of particle sizes in the range $\sigma_i \in [0.75, 1.25]$
(having a mean diameter $\sigma=1$). The units of length, energy and time
are respectively $\sigma$, $\epsilon_1$ and $\sigma\sqrt{m/\epsilon_1}$
where $m$ is the mass of the particles. The space of control parameters
is quite large. Therefore we set $\ell=3.5 \, \sigma$ as measured in
experiments~\cite{porte}, $T=1$, and $\epsilon_0=1$ and $\epsilon_2=50$,
and vary the remaining parameters $\{\phi, R, \tau_{\rm link}\}$.
These choices for the parameter values
leads to a phase diagram which is similar
to the one obtained in experiments.
Our numerical simulations are done for a three dimensional system
of $N=1000$ particles. The equations of motions of these particles
are integrated using a velocity Verlet scheme with a time step of
$\delta{t}=0.005$. Here, most of the results are reported for
$\tau_{\rm link}=10^2$, although we also explore other values of
$\tau_{\rm link}$: $1,10,10^3,10^4$ to illustrate some of the dynamical
features of the system. At each volume fraction $\phi$, we first
equilibrate the system of particles without any links ($R=0$). Once
equilibrated, bonds are introduced corresponding to the required
value of $R$ and then the system is again equilibrated to obtain
the proper distribution of bonds per particle. Since the structure
of the network is independent of the choice of $\tau_{\rm link}$
\cite{pablo2}, we use a small value of $\tau_{\rm link}=1$ to expedite the
equilibration process. Subsequently, data is generated by continuing the
simulations with different $\tau_{\rm link}$ values when required
and the averages are typically calculated over 100 different time origins.
We also do simulations for the
case when the bonds between two particles are completely frozen. In order
to do a proper sampling of the network configurations for this situation,
we use 6 initial configurations (positions,
connectivities) from the simulations with a finite $\tau_{\rm link}$
as initial inputs for subsequent evolution of the particle positions
using molecular dynamics with the connections now permanently fixed.
\section{Phase diagram}
\begin{figure}[]
\includegraphics[width=80mm,clip=true]{phasediagram_new.eps}
\caption{(Color online)
Phase diagram of the system obtained by varying $R$ and
$\phi$. The diamonds correspond to the coexistence region between gas
and liquid, the circles to the sol phase and the squares to the gel phase.
The horizontal dashed line corresponds to onset of percolation, the dot-dashed line
indicates the phase-coexistence boundary and the thick line
marks the predicted glass line. Snapshots of typical configurations in
each phase have been published elsewhere, see Fig. 3 in \cite{pablo2}.}
\label{phasediagram}
\end{figure}
We begin by summarizing our earlier findings for the phase diagram
(shown in Fig.~\ref{phasediagram}) for this model. If the number of
polymers is small (i.e $R<2$), the system is in a simple liquid phase
(the sol) at small values of $\phi$. In this regime, the distribution
of connectivities per particle is just an exponential \cite{pablo2}.
With increasing $\phi$, the dynamics in the sol regime becomes slow and
one eventually enters a glassy phase at large $\phi$, characterized by
a very strong increase of the timescales for structural relaxation. If
the number of bonds is large and $\phi$ is small, phase separation
is observed due to strong attractions between the droplets. Here, the
distribution of connectivities becomes bimodal, with one of the two peaks
corresponding to a well-connected liquid and the other to free particles.
For intermediate values of $R$ the system is in a gel phase. In this
region of the phase diagram the particles are connected together and
form a percolating cluster and the spatial density is homogeneous. Here,
the connectivity distribution is peaked around the average value and has
an exponential tail. Note that this gel is an equilibrium phase since the
polymer network is constantly rewired on the timescale $\tau_{\rm link}$.
However, if we go to large enough volume fractions, we observe again a
glassy system for all connectivities, with the corresponding divergence
of relaxation timescales.
At low $\phi$, in the gel phase, the main slow relaxation process is
related to the connections between the droplets by means of the polymers
and the timescale associated with its reconfiguration. At low $R$, as the
system becomes glassy at large $\phi$, the origin of the slow dynamics
is the steric hindrance caused by the caging of each particle by its
neighbors. In the region where both $R$ and $\phi$ act as a source for
slow dynamics, we have shown that the generic relaxation process has three
steps but with proper tuning of the two relaxation timescales, one can
also obtain logarithmic decays of the relaxation function~\cite{pre2010}.
In the following we will discuss in details the interplay between these
two relaxation processes and show the consequences on the nature of
different dynamical quantities as we move around in the phase diagram.
\section{Structure}
\begin{figure}[h]
\includegraphics[width=80mm,clip=true]{sq.eps}
\caption{(Color online) Main panel: Structure factor $S(q)$ computed for the particles at a
volume fraction of $\phi=0.70$ for different connectivities $R=0,4,8,$ and 12.
Inset: $S(q)$ at $\phi=0.50$ for $R=0,12,30,$ and 50. The dashed line corresponds to $q^{-4}$.
}
\label{sq70}
\end{figure}
Before we discus the different dynamical properties of the system, we
briefly look at its structure. In Fig.~\ref{sq70} we plot the static
structure factor, $S(q)$, for a system that is dense, $\phi=0.70$,
varying the connectivity $R$. The general shape of $S(q)$ is very similar
to the one of a simple liquid and hence we can conclude that the system
is homogeneous. Also, we see no significant dependence of $S(q)$ on $R$,
thus showing that at high densities the structure is mainly governed by
steric hindrance. For intermediate densities, however, we do note the
weak dependence on $R$ in the regime of small wave-vectors, illustrated
in the inset of Fig.~\ref{sq70} where we show $S(q)$ for $\phi=0.50$.
This is at a volume fraction at which for large $R$ the system approaches
the co-existence region, and hence one starts to see the emergence of
a power-law behavior at small values of $q$ with increasing $R$; the
data for $R=50$ can be approximated by $q^{-4}$ (which is expected in
proximity to phase co-existence \cite{furukawa}).
\section{Relaxation dynamics}
\subsection{Dependence on volume fraction $\phi$}
\begin{figure}[t]
\includegraphics[width=80mm,clip=true]{msdpv50.eps}
\includegraphics[width=80mm,clip=true]{pv61msd.eps}
\includegraphics[width=80mm,clip=true]{pv70msd.eps}
\caption{(Color online) Variation of mean squared displacements
$\Delta^2(t)$ with changing $R$ for (a) $\phi= 0.50$, (b) $\phi=0.61$, and (c)
$\phi=0.70$, using a bond lifetime of $\tau_{\rm link}=10^2$. }
\label{msdt100}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=80mm,clip=true]{fig2_pv50fsq6t100_a.eps}
\includegraphics[width=80mm,clip=true]{fig3a_pv61fsq6t100_a.eps}
\includegraphics[width=80mm,clip=true]{pv70fsq6t100.eps}
\caption{(Color online) Variation of the self intermediate scattering
function with changing $R$. (a) $\phi= 0.50$, (b) $\phi=0.61$, and (c)
$\phi=0.70$, using $\tau_{\rm link}=10^2$. Note the change of time span
in different panels, from a maximum time of $10^4$ in (a) to $10^6$ in
(c). The inset in panel (c) shows the collapse of correlation functions
for $\phi=0.70$ after rescaling by the relaxation time, similar to
time-temperature superposition principle.
}
\label{fsq6t100}
\end{figure}
\begin{figure}[h]
\includegraphics[width=80mm,clip=true]{diffusion.eps}
\caption{(Color online) Variation of diffusion constant $D$
with R, for different $\phi$ (shown in Fig.~\ref{msdt100}). The dotted
line, corresponding to an exponential function, is drawn as a guide to the eye.}
\label{difft100}
\end{figure}
We now characterize the dynamical properties of the system by focusing on
two quantities: (i) the mean squared displacement, defined as $\Delta^2(t)
= \langle\frac{1}{N} \sum_i \left\langle | {\bf r}_i(t) - {\bf r}_i(0)
|^2 \right\rangle$ and (ii) the self-intermediate scattering function,
defined as $F_s(q,t) = \langle\frac{1}{N}\sum_j \exp \left( i {\bf q}
\cdot [ {\bf r}_j(t) - {\bf r}_j(0) ] \right)\rangle$. Here ${\bf
r}_i(t)$ is the position of particle $i$ at time $t$, $q$ is the
wave-vector, and $\langle .\rangle$ corresponds to the ensemble average.
For increasing volume fraction $\phi=0.50, 0.61, 0.70$, we discuss
simultaneously the data for $\Delta^2(t)$, shown in Fig.~\ref{msdt100},
and $F_s(q,t)$, shown in Fig.~\ref{fsq6t100}, computed at a wave-vector
value $q=6$. Thus, the measured $F_s(q,t)$ probes the relaxation
dynamics on length scales that are slightly larger than the average
particle diameter (the peak in the structure factor $S(q)$ occurs at
$q\approx{7.3}$).
We start in the pure gel phase ($\phi=0.50$). The mean squared
displacement, Fig.~\ref{msdt100}a, shows that the increasing number of
bonds restricts the motion of the particles in that for $R>0$ we see the
emergence of an intermediate regime which develops into a well-defined
plateau at $R=10$. The height of this plateau depends significantly on
$R$, showing that this cage motion is directly related to the transient
bonds between the particles. The self intermediate scattering function,
Fig.~\ref{fsq6t100}a, shows for small $R$ a very rapid decay. This
changes in that for $R$ around 4-6 a plateau develops at intermediate
and long times, the height of which depends strongly on $R$. The presence
of this increasing plateau height, which is reminiscent to the so called
type-A transition of mode-coupling theory~\cite{dawson}, indicates
that the relaxation mechanism is changing: For small $R$ the motion of
the particles is only weakly slowed down by the presence of the bonds,
which typically break on the time scale of $\tau_{\rm link}$. However,
for larger $R$ breaking a few bonds is not enough to allow the particles
to move since the remaining bond still allow to maintain the particle
inside its cage. Hence this makes that at large $R$ the relaxation
dynamics does not depend very strongly on $R$ anymore. This effect is
seen in Fig.~\ref{difft100} where we show the diffusion constant of the
particles, $D$, (as obtained from the mean squared displacement at long
times) as a function of $R$. For small $R$, $D$ shows a rather strong
$R-$dependence, whereas for $R>5$ this dependence becomes weaker. Below
we will discuss the $R-$dependence of the relaxation time in more detail.
Next, we look at the data for an intermediate density, viz. $\phi=0.61$,
and in Fig.~\ref{msdt100}b we show the corresponding $\Delta^2(t)$. Like
for $\phi=0.50$, the longtime diffusion decreases with increasing $R$,
and the $R-$dependence of the diffusion constant shows again a break at
around $R\approx 5$ (see Fig.~\ref{difft100}). For this value of $\phi$
we observe, however, for $R>2$, at intermediate times a shoulder in
$\Delta^2(t)$. This shoulder, clearly visible for $R=4$, is related to
the presence of the bonds that lead to a caging of the particles on the
length scale related to $\ell$, the maximum extension of a bond. Thus
the hint of the short-time plateau (when $\Delta^2(t)\approx{0.1}$)
and again one at later time (when $\Delta^2(t) \approx 1$) reflects
the presence of the two different mechanisms for constraining
particle motion, viz. local steric hindrance and the network bonds.
Since each type of caging leads to a plateau in the intermediate
scattering function~\cite{pre2010}, the existence of the two competing
mechanisms makes that at intermediate times $F_s(q,t)$, shown in
Fig.~\ref{fsq6t100}b, has a very slow, almost logarithmic, decay,
if $R$ is around 6. More details on this particular case are give in
the context of Fig.~\ref{decon}. If one compares the data for $R=10$
at the two volume fractions $\phi=0.50$ and $\phi=0.61$, we recognize
that the height of the plateaus in $F_s(q,t)$ increases with $\phi$
from which one can conclude that the proximity of the particles leads
to increased tightening of the cage.
If the density is increased further to $\phi=0.70$, the relaxation
dynamics becomes strongly dominated by the steric hindrance
mechanism. Already for $R=0$ one sees a weak plateau in $\Delta^2(t)$
and its length grows rapidly with increasing $R$ without changing much
its height (see Fig.\ref{msdt100}c). This is the typical behavior of
simple glass-forming liquids~\cite{kob_95}. At the same time the self
intermediate scattering function shows the growth of a shoulder with
{\it finite} height and this height depends again only weakly on $R$,
Fig.~\ref{fsq6t100}c. In contrast to the case at lower densities,
here the shape of the correlator is basically independent of $R$. This
is demonstrated in the inset of Fig.~\ref{fsq6t100}c where we plot
$F_s(q,t)$ as a function of $t/\tau$, with the relaxation time $\tau$
defined by $F_s(q,\tau)=0.03$. The fact that this presentation of the
curves leads to a nice master curve shows that we have for this system a
time-$R$ superposition, in analogy to the time-temperature superposition
found in simple glass-forming systems~\cite{book}.
Despite the qualitative changes seen in $\Delta^2(t)$ and $F_s(q,t)$
if $\phi$ is increased, the $R-$dependence of the diffusion constant
for $\phi=0.70$ is very similar to the one seen at lower densities,
see Fig.~\ref{difft100}. Also for this high value of $\phi$ we see that
this dependence is relatively strong at small $R$ and becomes weaker if
$R>6$. Hence the fact that at low $R$ just few bonds have to be broken
in order to allow a particle to move whereas at high $R$ this is not a
sufficient condition, is reflected in the $R-$dependence of $D$ at {\it
all} $\phi$.
\subsection{Varying the bond life-time $\tau_{\rm link}$}
\begin{figure}[]
\includegraphics[width=80mm,clip=true]{fig5b_r63v61t1000fsqall.eps}
\includegraphics[width=80mm,clip=true]{fig5b_r63v61t100fsqall.eps}
\includegraphics[width=80mm,clip=true]{fig5a_r63v61t10fsqall.eps}
\caption{(Color online) $F_s(q,t)$ for $R=6.3, \phi=0.61$, a) $\tau_{\rm link}=10^3$ b) $\tau_{\rm link}=10^2$
and c) $\tau_{\rm link}=10$.}
\label{r63pv61fsqall}
\end{figure}
We will now explore further the interplay between the two processes
leading to slow relaxation, i.e. the nearest-neighbor caging and
the constrained motions due to the polymer bonds. Since the relative
importance of these two processes depends on the lifetime of the polymer
bonds, we will in the following vary this lifetime and consider values
of $\tau_{\rm link}=10, 10^2$, and $10^3$, and at fixed state-point of
$R=6.3,\phi=0.61$, where logarithmic decay in the time-correlation
function is observed. We study how the shape of $F_s(q,t)$ changes with
varying $\tau_{\rm link}$ and will relate this to the interplay between the
two processes. This is done for different values of wave-vector in order
to see the how relaxation timescales vary over different lengthscales.
We begin by looking at the case of the large $\tau_{\rm link}=10^3$,
i.e. when the polymer bonds hinder the motion on a time scale longer
than the steric hindrance effect, see Fig.~\ref{r63pv61fsqall}a. The
correlation function reflects three different relaxation processes
which can be seen for all value of $q$. Initially, the particles rattle
inside the cage, resulting in partial relaxation of the correlation
function on a timescale $\tau_\beta \approx 1$. Later on, the particles escapes
from the cage of neighboring particles, (which for this value of $\phi$
not very pronounced), but the relaxation process is then held up by the
polymer bonds.
Eventually, the polymer network rewires on a timescale which is proportional to
$\tau_{\rm link}=10^3$ and the particles start the final relaxation process.
The height of the plateau in $F_s(q,t)$ increases with decreasing $q$,
which is the typical behavior for a glassy system \cite{kob_95,nauroth}.
However, for the range of wave-vectors explored, the final timescale for
decay of $F_s(q,t)$ depends only weakly on $q$. Thus, over these length
scales, the relaxation process is determined by the reconfiguration
of the network. However, at larger length scales, one can expect that
hydrodynamic effects will eventually dominate and this will then determine
the relaxation timescales.
For intermediate values of the bond lifetime, $\tau_{\rm link}=10^2$
in Fig.~\ref{r63pv61fsqall}b, the final decay has moved to shorter
times and makes that now the interplay between the two processes
results in a logarithmic decay of $F_s(q,t)$ as discussed elsewhere
\cite{pre2010}. This logarithmic dependence is seen for a range of $q$
values (see Fig.~\ref{r63pv61fsqall}b), with the time-window over which
it exists decreasing with decreasing $q$. This implies that this form
for the correlation function occurs only for specific combination of
the relaxation timescales of the two processes of steric hindrance and
eventual network relaxation. At sufficiently small $q$, hydrodynamics
makes that the relaxation becomes so slow that the second relaxation step
is no longer visible out and hence the logarithmic $t-$dependence is no
longer observed. Finally we mention that the logarithmic shape in the
relaxation function as discussed here is not related to any underlying
higher order mode-coupling transition, in contrast to the case of
certain colloidal systems for which similar relaxation functions have
been observed \cite{zaccarelli3}.
If we set $\tau_{\rm link}$ to a small value, this three-step relaxation
can no longer be observed, as is seen for the case of $\tau_{\rm link}=10$
in the bottom panel of Fig.~\ref{r63pv61fsqall}: For all $q$'s the
curve show a (seemingly) simple two step relaxation, since the third step
(related to the bonds) starts already when the second step (related to
steric hindrance) is not yet completed and hence the two processes become
completely mixed in time. Below we will briefly come back to this effect.
\subsection{When the bonds are permanent}
A useful way to check the influence of the polymer bonds on the relaxation
dynamics is by comparing the relaxation functions for the case when the
bonds are permanent, i.e.~$\tau_{\rm link}=\infty$, to those when the
lifetime is finite. In Fig.~\ref{fsq6t100tinf}, we do this comparison for
different connectivities ($R=4,6,$ and 8) and different values of $\phi$.
In Fig.~\ref{fsq6t100tinf}a, we show $F_s(q=6,t)$ for the case of
$R=4$ with the two different bond lifetimes $\tau_{\rm link}=10^2$
and $\tau_{\rm link}=\infty$. For $\phi=0.50,$ 0.60, and 0.65,
the timescales for overcoming the steric hindrance are the same for
both lifetimes. However, for all $\phi$, we find that the correlation
function for $\tau_{\rm link}=\infty$ shows a plateau at long times (not
visible in this plot), which is due to the fact that the frozen bonds
in the percolating gel-network prevent the complete relaxation of the
system. In contrast the curves for the finite $\tau_{\rm link}$
vanish at long times. For small and intermediate $\phi$ the two
sets of curves are very similar, indicating that the presence of a few
bonds does not change the dynamics significantly. Only for $\phi=0.7$ one
sees a substantial difference in that the correlator for the permanent
links decays slower than the one with $\tau_{\rm link}=10^2$. It is
reasonable that these differences are noticeable for times somewhat longer
than $10^2$, i.e. the time scale of $\tau_{\rm link}$.
\begin{figure}[ht]
\includegraphics[width=80mm,clip=true]{fig6a_r4fsq6t100tinf.eps}
\includegraphics[width=80mm,clip=true]{fig6b_r6fsq6t100tinf.eps}
\includegraphics[width=80mm,clip=true]{fig6c_r8polyfsq6t100tinf.eps}
\caption{(Color online)
$F_s(q,t)$ for different $\phi$ and $R=4,6,$ and 8 when the
bond lifetime is $\tau_{\rm link}=100$ (lines
with symbols) and when $\tau_{\rm link}=\infty$ (lines). The wave-vector is $q=6.0.$}
\label{fsq6t100tinf}
\end{figure}
If we increase the connectivity to $R=6$ (Fig.~\ref{fsq6t100tinf}b),
we see that the behavior is qualitatively similar to $R=4$ in that for
all values of $\phi$ the two curves track each other up to times around
$10^2$, i.e. the time of the finite $\tau_{\rm link}$. For larger
times the correlators for $\tau_{\rm link}=10^2$ decay to zero whereas the
ones for $\tau_{\rm link}=\infty$ show at long times a marked plateau.
The height of this plateau depends now more strongly on $\phi$ than it was
the case for $R=4$, showing that if $R$ is increased the life time of the
bonds becomes more influential. This is reasonable since it is related
to the general observation that in glass-forming system small changes influence
the relaxation dynamics increasingly more the slower the dynamics is.
We also note that for $\phi=0.70$ the correlator for the permanent bonds
becomes very stretched. This sluggish relaxation might be related
to the fact that for this value of $R$ there are, in addition to the
percolating cluster, clusters of different sizes (see Ref.~\cite{pablo2}
for typical distributions), thus giving rise to relaxation dynamics that
spans many orders of magnitude in time and hence to a very stretched
average correlation function. The stretching of the correlator for
the frozen bonds could, however, also be due to the fact that these
different clusters hinder each other resulting in the overall slowdown
of the dynamics \cite{puertas2}.
Next, we increase the number of bonds even more, viz. $R=8$, as shown
in Fig.~\ref{fsq6t100tinf}c. We see that for the case of permanent
bonds the height of the asymptotic plateau has increased strongly
in comparison to the case of $R=6$. As a result the correlator for
$\phi=0.70$ shows only a negligible decay of the correlation function
if the bonds are permanent. The motion is so constrained by these bonds
that the height of the asymptotic plateau, caused by the permanent
bonds, becomes comparable to the one related to the steric hindrance.
Also for $\phi=0.5$ the height of the second plateau has increased so
much that the relaxation from caging is now completely masked. However,
for the intermediate values of $\phi$, one does notice a difference
between the two plateau heights and the correlations functions decay in
a very stretched fashion from one to the other. In fact, the decay is
so slow that the time-dependence is seen to be logarithmic (nearly for
five decades in the case of $\phi=0.65$). Note that this logarithmic
decay is due to the heterogeneous relaxation of the floppy clusters of
frozen bonds, which we will discuss later in further detail. Thus this
mechanism is different from the one leading to the logarithmic relaxation
seen in Fig.~\ref{fsq6t100} which was due to a unique combination of
the two {\it finite} relaxation timescales.
\begin{figure}[]
\includegraphics[width=70mm,clip=true]{pv60r2afqtinf.eps}
\includegraphics[width=70mm,clip=true]{pv60r4afqtinf.eps}
\includegraphics[width=8cm,clip=true]{fqpn.eps}
\caption{(Color online) (a),(b): Variation of $F_s(q,t)$ with wave-vector
$q$ for $R=2$ and 4 at $\phi=0.60$ with $\tau_{\rm link}=\infty$. (c):
Height of plateau, $f_q^{\rm PN}$, of $F_s(q,t)$ at long times as a function
of $q$ for $R=2$, 4, and 6. The dashed lines are fits of the form
$\exp{(-q/\xi)}$ with the corresponding $\xi=0.55$, 0.99, and 2.27.
(d),(e): $f_q^{\rm PN}$ vs. $q$ for different values of $\phi$ for $R=4$ and 6. }
\label{ap1}
\end{figure}
We now study the floppiness of this network of particles connected by the
permanent bonds by probing the wave-vector dependence of the relaxation
functions $F_s(q,t)$. In Figs.~\ref{ap1}a and b, we show for $\phi=0.60$
the variation of $F_s(q,t)$ for $R=2$ and 4, i.e. in the region of
the phase diagram where gelation sets in. The height of the plateau at
long times, also called non-ergodicity parameter, is a measure for the
stiffness of the network on the length scale $q$ considered. Comparing
the two panels we recognize that, for a given $q$, the height of the
plateau increases with increasing $R$. Denoting this height by $f_q^{\rm
PN}$, we show in Fig.~\ref{ap1}c that $f_q^{\rm PN}$ shows basically
an exponential decrease in $q$ with a slope that decreases rapidly
with increasing $R$. That $f_q^{\rm PN}$ decreases with increasing $q$
is of course reasonable since on small length scales the particles have
more leeway to flop around than on large length scales. Note, however,
that this exponential dependence is in contrast to the one found for
the height of the plateau due to the steric hindrance, the latter being
basically a gaussian function~\cite{nauroth}. Since in the representation
of Fig.~\ref{ap1}c such a gaussian dependence is given by a parabola,
we see that such a curve will intersect the one for $f_q^{\rm PN}$ at
a certain value of $q_x$. For $q< q_x$ the plateau due to the steric
hindrance is above $f_q^{\rm PN}$, thus making that one observes two
plateaus in the correlator. However, for $q>q_x$ the plateau at long
times is higher than the steric one, thus making that the latter one
will be completely masked by the former and thus the correlator will
show only one plateau.
We have also studied how the $q-$dependence of $f_q^{\rm PN}$ changes with
the volume fraction and in Figs.~\ref{ap1}d and e, we show $f_q^{\rm PN}$
for $R=4$ and 6, respectively. For both cases we see that the rigidity
of the network at large scales, i.e.~small $q$, is not affected by the
volume fraction which is not surprising. For $R=4$ we see that this is
also true at small length scales whereas for $R=6$ we note a significant
$\phi-$dependence if $q$ is large. This difference is likely related
to the fact that the system with $R=6$ is much more sluggish than the
one for $R=4$, see Fig.\ref{fsq6t100tinf}, and hence small changes
(here in $\phi$) will have a stronger impact on the dynamics.
Finally we disentangle the dynamics of the particles that belong to the
percolating cluster from the one that are not attached to it. In the
following discussion these are referred as clustered and non-clustered
particles, respectively. The objective is to clarify the respective
contributions to the different dynamical quantities that we have
discussed above. We do this comparison for an increasing number of
connections $R$ at a fixed (large) volume fraction of $\phi=0.65$.
In Fig.~\ref{figclustfree}, we show the data for $\Delta^2(t)$ and $F_s(q,t)$
for these two families of particles.
\begin{figure}[h]
\includegraphics[width=80mm,clip=true]{msd_diffR_pv65_freecluster.eps}
\includegraphics[width=80mm,clip=true]{pv65fsq6tinf_free.eps}
\caption{(Color online) Relaxation dynamics of the clustered (dashed lines
with open symbols) and non-cluster particles (full lines with filled
symbols) for the case of frozen bonds (i.e $\tau_{\rm link}=\infty$),
at $R=4$, 6, and 8 for volume fraction $\phi=0.65$. (a): Mean squared
displacement $\Delta^2(t)$. (b) Self intermediate scattering function
$F_s(q,t)$ for $q=6$.}
\label{figclustfree}
\end{figure}
The mean squared displacement of the clustered particles shows after the
ballistic regime at short times a shoulder that is related to the cage
of the steric hindrance (Fig.~\ref{figclustfree}a). This localization
is, however, only temporary and is followed by a further increase of
$\Delta^2(t)$. Only at longer times $\Delta^2(t)$ saturates at a height
that depends on $R$. We note that the approach to this asymptotic
height becomes increasingly slow with increasing $R$ and in fact for
$R=8$ the time dependence is close to logarithmic and does not end
within the time window of our simulation. This behavior
is also seen in the self intermediate scattering function $F_s(q,t)$
(Fig.~\ref{figclustfree}b). At short times the correlator decays quickly
onto a plateau (not very pronounced) before the relaxation of the
steric hindrance starts. For $R=4$ and 6, this process ends in that the
correlator reaches the final plateau (which is given by $f_q^{\rm PN}$
discussed above). However, for $R=8$ the final decay is so slow, again
compatible with a logarithmic time dependence, that we do not see the
asymptotic behavior.
Finally we look at the motion of the non-clustered particles and compare
it with the one for $R=0$. From Fig.~\ref{figclustfree}a we recognize that
at this volume fraction also these particles are slowed down by the cage
effect in that one sees for all values of $R$ a shoulder in $\Delta^2(t)$
at time $t\approx 1$. For $R=0$ the $\Delta^2(t)$ shows then immediately
the diffusive behavior, i.e. it is proportional to $t$. However, if $R$
is increased, the $t-$dependence of $\Delta^2(t)$ for the un-clustered
particles follows first the one of the clustered particles. Only once
the latter starts to reach the plateau discussed above, do the former cross
over to the diffusive behavior. Hence we can conclude that
before this crossover the relaxation dynamics of the two population
of particles are strongly coupled. This result is reasonable because in
order to move, the un-clustered particles have to explore the holes within
the percolating cluster formed by the particles which are permanently linked.
Also the self intermediate scattering function of the non-clustered
particles tracks the one of the clustered particles at short times
(Fig.~\ref{figclustfree}b). However, once the latter starts to show at
long times a plateau that has a significant height, the two correlators
differ strongly since the one for the non-clustered particles decays
to zero at long times. From the graph we also see that for $R=8$
the correlator is extremely stretched and shows almost a logarithmic
$t-$dependence. This very slow decay indicates that the mobile clusters
can move around the percolating cluster only with great difficulty, a
behavior that is similar to the relaxation dynamics of particles moving
in random porous media~\cite{kim,kurzidim}. It is also interesting that
for the highest $R$ the mean squared displacement shows for the last two decades
in time a nice diffusive behavior, whereas $F_s(q,t)$ is far from having
decayed to zero. This apparent contradiction is related to the fact that
$\Delta^2(t)$ is dominated by the particles that move relatively fast
(i.e. they are in the small clusters) whereas $F_s(q,t)$ is dominated by
the slowly moving particles (i.e. the large clusters). This cluster-size
dependent dynamics leads to a so-called dynamical heterogeneity and
in Section~\ref{sec_vi} we will discuss this phenomenon in more detail.
Although we show in~Fig.~\ref{figclustfree} the comparison between the dynamics
of clustered and non-clustered particles for the case that $\tau_{\rm
link}$ is infinitely large, it is evident that for a very large but {\it
finite} value of $\tau_{\rm link}$ the relaxation dynamics will be very
similar. Hence if, e.g., $\tau_{\rm link}$ is on the order of $10^5$,
basically none of the shown curves will change significantly and thus
the conclusions drawn from Fig.~\ref{figclustfree} will be apply also
for such value of $\tau_{\rm link}$.
\subsection{Relaxation timescales}
We now investigate how the two different relaxation timescales, one due
to breaking of local cages and the other due to the reconfiguration
of the network-bonds, vary with the volume fraction $\phi$ and the
connectivity $R$.
To start we consider the case of structural relaxation related to the
steric hindrance. In order to avoid that this relaxation process is
influenced by the one of the network, we consider the case in which
the latter is completely suppressed, which can be achieved by choosing
$\tau_{\rm link}=\infty$. In the following we will study the relaxation
times associated with the intermediate scattering function for wave-vector
$q=6$. As discussed above, this correlator shows at long times an
asymptotic plateau the height of which, $f_q^{\rm PN}$, depends on $R$ and
$\phi$. To take this into account we define the relaxation time $\tau_{\rm
SH}$, to be the time at which $F_s(q,t)-f_q^{\rm PN}=0.03$. The evolution
of $\tau_{\rm SH}$ with $\phi$ is shown in Fig.~\ref{figrelaxtime}a,
for different value of $R$.
We see that in the absence of any bonds, i.e. for $R=0$, we have the usual
slowing down of dynamics with increasing $\phi$. The $\phi-$dependence
of the relaxation time can be fitted by a Vogel-Fulcher-Tammann-law of
the form $\tau_{\rm SH}\sim{\exp[A/(\phi_c-\phi)^\beta]}$, with $\beta
\approx 1$, from which we can estimate the volume fraction $\phi_c$ at
which the relaxation times would diverge. If we increase the number of
bonds among the particles, we see that $\tau_{\rm SH}$ increases. That
this increase is not just a constant ($R-$dependent) factor but depends
also on $\phi$ is demonstrated in the inset of the figure where we
have normalized the relaxation times to its value at $\phi=0.5$. Using
the Vogel-Fulcher-Tammann-law we can thus extract from these data the
$R-$dependence of $\phi_c$, which can be considered as a proxy for the
$R-$dependence of the glass transition temperature. This $\phi_c(R)$ line
is included in Fig.~\ref{phasediagram} as well and we see that it has a
weak negative slope with $\phi_c(R=0)=0.847$ and $\phi_c(R=6)=0.808$. Thus
we can conclude that the glass transition as induced by the steric
hindrance mechanism does not depend strongly on the value of $R$.
However, since the prefactor of the Vogel-Fulcher-law does depend strongly
on $R$, we can conclude that a line of iso-relaxation time will bend
significantly more.
\begin{figure}[]
\includegraphics[width=70mm,clip=true]{fig7a_tauSHvsph.eps}
\includegraphics[width=72mm,clip=true]{lntaupnRtaulink1a.eps}
\caption{(Color online)
(a) Main panel: Plot of $\tau_{\rm SH}$ vs. $\phi$ for different values
of $R$. Inset: Same data scaled by $\tau_{\rm SH}(q,\phi=0.5)$. (b)
Main panel : $\tau_{\rm PN}$ vs. $R$, for different values of $\phi$.
Inset: Same data as in the main panel but normalized by the relaxation
time for $R=0$, leading to a master curve. Note the double logarithmic
scale in this plot.}
\label{figrelaxtime}
\end{figure}
Next, we investigate how the timescale for the full relaxation of the
network of particles depends on the average number of bonds, $R$.
For this we define a relaxation time using the $t-$dependence of
$F_s(q=6,t)$ with a very short lifetime for the bonds ($\tau_{\rm
link}=1$). $\tau_{\rm PN}$ is then defined via $F_s(q=6,\tau_{\rm
PN})=0.03$ and the $R-$dependence of this relaxation time is shown
in Fig.~\ref{figrelaxtime}b. The figure shows that for small and
intermediate values of $R$ the $R$-dependence is independent of the
volume fraction in that the curves for the different values of $\phi$
seem to be just shifted vertically. That this is indeed the case is
demonstrated in the inset where we show that a plot of the same data,
but now normalized to the relaxation time for $R=0$, gives a master
curve. The main panel shows that at small concentration of bonds the
$R-$dependence is close to an exponential, a result that is likely
related to the fact that increasing $R$ leads to a tighter cage for the
steric hindrance and hence a slower dynamics. However, for intermediate
values of $R$ the curve $\tau_{\rm PN}$ start to bend over towards a
weaker $R-$dependence. In order to investigate this effect better we
have carried out simulations for $\phi=0.50$ at very high values of $R$:
16, 30, and 50. We find that in this regime the relaxation time follows
closely an exponential (see main Fig.~\ref{figrelaxtime}b), but with an
exponential scale that is smaller than the one seen at small $R$.
Note that this dependence implies that there is no singularity in the
relaxation dynamics at any finite value of $R$, at least for this volume
fraction. Instead the dynamics shows a behavior that is similar to an
Arrhenius law in that the barrier for the relaxation depends only on the
number of bonds between the particles. This observation is in agreement
with earlier simulations for equilibrium gels \cite{demichele}.
\begin{figure}
\includegraphics[width=80mm]{Fs-Dx2-q_6-phi_0_61-R_6_3-several-tau2.eps}
\caption{ (Color online) \label{decon} $F_s(q,t)$ vs. $t$ for $R=6.3,
\phi=0.61$, and different $\tau_{\rm link}$. The inset shows $\Delta^2(t)$
for the same parameters. The dotted lines in the main figure mark
the plateaus in $F_s(q,t)$: $f_q^{\rm SH}$ and $f_q^{\rm PN}$, which
respectively correspond to the localization lengths $l_{\rm SH}$ and
$l_{\rm PN}$ indicated in the inset.}
\end{figure}
We conclude this discussion by using the different relaxation timescales
to develop a criterion that tells whether or not the correlator $F_s(q,t)$
will show the logarithmic $t-$dependence discussed in the context of
Fig.~\ref{fsq6t100}b. For this we define $f_q^{\alpha}$ as the height
of the plateau associated to the $\alpha$-process, with $\alpha\in
\{{\rm SH, PN}\}$, and recall that $\tau_{\beta}$, $\tau_{\rm SH}$
and $\tau_{\rm PN}\propto \tau_{\rm link}$ are the timescales for
the three different relaxation processes described above, namely the
rattling inside the cage, the escape from the local cage, and the network
renewal process. We can now define a simple criterion for the anomalous
logarithmic relaxation to be observed by requiring that the slope of
the two segments defined by pairs $[(\ln \tau_{\beta},f_q^{\rm SH}),(\ln
\tau_{\rm SH},f_q^{\rm PN})]$ and $[(\ln \tau_{\rm SH},f_q^{\rm PN}),(\ln
\tau_{\rm PN},0)]$ is the same, see Fig. \ref{decon}. It is known that
for glass-forming systems the plateau related to steric hindrance is a
Gaussian functions of the wave-vector $q$, $f_q^{\rm SH}\sim \exp(-q^2
\ell_{\rm SH}^2)$~\cite{nauroth}. We showed earlier, see Fig.~\ref{ap1}c,
that for the PN-process, $f_q^{\rm PN}$ has an exponential shape at large
$q$ (the regime corresponding to anomalous logarithmic behavior). Putting
these elements together we thus obtain
\begin{equation}
(q\ell_{\rm PN} - q^2\ell_{\rm SH}^2) = \ln \left[1 +
\frac{\ln (\tau_{\rm SH}\tau_{\beta}^{-1})}{\ln (\tau_{\rm PN}\tau_{\rm SH}^{-1})} \right] \, .
\label{eq1}
\end{equation}
\noindent
This equation relates a purely structural observable which depends
on the competition between two localization lengthscales with
dynamical information encoded in the different relaxation times,
predicting a precise connection between structure and dynamics
whenever anomalous logarithmic relaxation is to be expected.
\section{Heterogeneities in dynamical properties}
\label{sec_vi}
Typical glass formers show a heterogeneous dynamics of the particles when
the system is increasingly supercooled. It reflects the broad distribution
in the timescales for local structural relaxations \cite{chi4ref}. On the
other hand, we have reported earlier \cite{pablo} that also the dynamics
in the gel phase can be heterogeneous: The particles in the percolating
cluster and those that are unattached have different nobilities till
the timescales at which the bonds in the network are reconfigured. In
the following we will thus explore how these two different sources of
heterogeneity interact as we increase the density of particles as well
as the number of connectivities in the system.
We use two different measures of dynamical heterogeneity. The
first one is the non-Gaussian parameter $\alpha_2(t)$, defined as
$\alpha_2(t)=3\langle r^{4}(t) \rangle/5\langle r^{2}(t) \rangle ^{2}-1$,
where $\langle r^{2}(t) \rangle$ and $\langle r^{4}(t) \rangle$ are the
second and fourth moments of the distribution function of single particle
displacements $G_s(r,t)=N^{-1}\sum_i\langle \delta(r-|{\bf r}_i(t) -
{\bf r}_i(0)|)\rangle$, i.e. of the self part of the van Hove function.
A non-zero value of $\alpha_2(t)$ quantifies the extent of deviation
from a Gaussian shape for $G_s(r,t)$ \cite{alpha2}. Note that $G_s(r,t)$
is Gaussian at short times, i.e. in the ballistic regime, and again at
long times when the particles are diffusive.
\begin{figure}[ht]
\includegraphics[width=80mm,clip=true]{fig10a_alpha2_phi0.50.eps}
\includegraphics[width=80mm,clip=true]{fig10b_alpha2_phi0.70.eps}
\caption{(Color online)
Time dependence of $\alpha_2$ for different values of $R$ (as marked
in the graphs) for $\tau_{\rm link}=10^2$. (a): $\phi=0.50$ in
double-logarithmic representation and (b): $\phi=0.70$.}
\label{alpha2phi}
\end{figure}
The second measure is the dynamic susceptibility $\chi_4(q,t)$,
computed via the fluctuations of time-correlation function:
$\chi_4(q,t)=N[\langle{F^2_s(q,t)}\rangle-\langle{F_s(q,t)}\rangle^2]$.
It is designed to capture the spatiotemporal correlations of particle
nobilities and provides, from its peak value, an estimate of the dynamic
correlations \cite{glotzer, chi4ref}. Such functions have also been
analyzed for gels with permanent bonds \cite{coniglio3} or low density
gels \cite{colombo2}.
\subsection{Non-Gaussian parameter $\alpha_2(t)$}
We begin by looking at how the non-Gaussian parameter varies with changing
connectivities $R$, either when we go from the liquid to the gel phase, at
$\phi=0.50$, or when we are in a strongly glassy region, $\phi=0.70$. For
the case $\tau_{\rm link}=10^2$ the data is shown in Fig.~\ref{alpha2phi}.
For $\phi=0.50$, Fig.~\ref{alpha2phi}a, $\alpha_2(t)$ shows at $t\approx
1$ a peak if $R=0$ and a shoulder for $R>0$. This feature is thus related
to the dynamical heterogeneity due to the steric hindrance mechanism,
as in usual glass-forming systems~\cite{kob_95}. For $R>0$ we find
in addition a very prominent peak the location of which is basically
independent of $R$, which shows that for this packing fraction the time
at which the system is maximally heterogeneous does not depend on $R$.
This observation is in tune with our earlier discussion that the timescale
for structural relaxation at small values of $\phi$ does not change with
$R$ if the number of bonds is not very large (see Fig.~\ref{fsq6t100}a).
The maximum occurs around $t\approx{10^3}$, a time which corresponds to
a timescale for which a significant number of particles have broken the
bonds with their neighbors and exit the constraints of the network. The
location of the peak is thus somewhat larger than $\tau_{\rm link}$. As
has been documented in Ref.~\cite{pablo}, at small packing fractions
one has on this timescale two families of particles, one for the
mobile particles and the other related to those that are immobile
and as a consequence the shape of $G_s(r,t)$ deviates strongly from a
Gaussian. Figure~\ref{alpha2phi}a indicates thus that the same behavior
persists to the larger volume fraction of $0.50$. One interesting feature
is that the peak height is non-monotonous in $R$ in that it increases
till $R=6$ and then decreases again for larger $R$ values. The reason
for the growth is that an increasing $R$ allows for more diverse values
of the connectivity for the particles, and hence to a stronger variety
in the dynamical behavior. On the other hand, if $R$ is very large most
of the particles are strongly connected all the times and hence show a
much smaller variation in their relaxation dynamics, i.e. they behave
like in a mean-field like regime.
Next we study the behavior for $\phi=0.70$, Fig.~\ref{alpha2phi}b.
When there are no bonds, $R=0$, we see that $\alpha_2(t)$ is peaked
at around $t=8$. This peak, which is related to the steric hindrance,
corresponds thus to the one seen for $\phi=0.50$ at $t\approx 1$ and
which, due to the higher density has shifted to larger times. For $R=2$
(when the percolating cluster of the bonds develops), the location of
this peak slightly shifts to larger times and one sees the appearance
of a second peak at $t \approx 2000$ (corresponding to the timescales
for renewal of the network connectivities). However, the dominant
heterogeneity is still due to the local steric hindrances. As $R$ is
increased, the location of the first peak continues to shift to longer
timescales, in track with the increasing structural relaxation timescales
(see Fig.~\ref{fsq6t100}c), and also its height increases, in qualitative
agreement with the behavior found in simple glass-formers if the coupling
is increased~\cite{kob_95}. For $R=6$, the two peaks have merged, since
the timescales for the two sources of heterogeneity are nearly the same,
and thus $\alpha_2(t)$ has a single peak. If $R$ is increased even more,
the position of the peak moves to larger times, but its height starts to
decrease. The reason for this decrease is likely the same as the one we
indicated when we discussed the data for $\phi=0.50$, i.e. that for large
$R$ the system starts to become mean-field like and hence heterogeneities
are suppressed.
We also note that the height of the peak is significantly smaller
than the one for $\phi=0.50$. Thus, at large density, steric hindrance
dominates and even the faster particles have less space to move around
resulting in significantly less non-Gaussian shapes for $G_s(r,t)$. As
a consequence, the dynamics is more homogeneous in this regime compared
to the gel at lower $\phi$.
\begin{figure}[]
\includegraphics[width=80mm,clip=true]{gsrtr2pv70.eps}
\caption{(Color online) Self part of the van Hove function for $R=2$
and $\phi=0.70$, parameters for which $\alpha_2(t)$ shows two peaks (see
Fig.~\ref{alpha2phi}b).}
\label{vanhover4pv70}
\end{figure}
To elucidate the origin of the two peaks in the non-Gaussian
parameter at $R=2, \phi=0.70$, we determine how the distribution of
particle displacements $G_s(r,t)$ evolves with time. This is shown in
Fig.~\ref{vanhover4pv70}. At short times, particle motion is restricted
by the local cage of neighbors and thus $G_s(r,t)$ is a Gaussian at small
$r$. At larger distances the distribution has an exponential tail which
is due to a few particles that have escaped the steric hindrance cage,
in agreement with the usual heterogeneous glassy dynamics in supercooled
systems~\cite{Pinaki}. The presence of these two processes gives rise
to the maximum in $\alpha_2$ at short times. At around $t=400$, most of
the particles have escaped from this steric hindrance cage and thus the
dynamics of the system becomes more homogeneous with $G_s(r,t)$ assuming
a Gaussian form and hence $\alpha_2$ decreases again. But with time, the
motion of the particles is again restricted, this time by the bonds which
have not yet relaxed. Thus again $G_s(r,t)$ develops a Gaussian shape
(with a larger width, determined by the length of the connecting bonds)
and an exponential tail, implying an increase in $\alpha_2$. Later,
the network eventually relaxes, the particles are diffusive and thus
$\alpha_2$ decreases again.
\begin{figure}[]
\includegraphics[width=80mm,clip=true]{fig11_alpha2_taulink_new.eps}
\caption{(Color online) Time dependence of $\alpha_2$ for different
values of $\tau_{\rm link}$ at $R=6, \phi=0.65$. Inset: For the case
of fixed bonds ($\tau_{\rm link}=\infty$), $\alpha_2$ for the full system
(maroon) along with those for the clustered (green open symbols) and
non-clustered particles (orange filled symbols). }
\label{alpha2taulink}
\end{figure}
The presence of these two different contributions to the shape of
$\alpha_2(t)$ can be further clarified by varying the lifetime of the
network, i.e. $\tau_{\rm link}$, and in Fig.~\ref{alpha2taulink} we
show this for the case $\phi=0.65, R=6$. Since we scan a large span
of timescales and $\alpha_2$ changes strongly, the data is shown in
logarithmic scales. For $\tau_{\rm link}=1$, we see only a small peak at
$t\approx 10$, which corresponds to the one due to the steric hindrance.
For $\tau_{\rm link}=10^2$, we find that this peak has increased a bit
in height and its location has also shifted to $t \approx 30$. This
change is a consequence of the increased effective coupling between the
particles due to the increased lifetime of the bonds. In addition, we
see the development of a new (weak) peak at $t\approx{10^3}$, which is
caused by the network renewal process. Now, if we increase $\tau_{\rm
link}$ further, the location of the first peak remains at the same place,
since the local crowding effects are unaffected by the network dynamics.
In contrast to this, the location of the second peak shifts to longer
timescales and its height increases strongly. Since the position of
this peak scales with $\tau_{\rm link}$, we recognize that this peak
is related to the renewal process. Finally, if we take the case of
permanent bonds between the particles, the curve is nearly identical to
that for $\tau_{\rm link}=10^4$, except that we see no signature of it
eventually decreasing with time. The reason for this runaway effects is
the fact that with increasing lifetime of the network, the unattached
particles diffuse away and travel long distances before the network is
again renewed. This results in very long extended tails in $G_s(r,t)$
which shows up as very large values for the non Gaussian parameter.
For the case that $\tau_{\rm link}$ diverges the dynamics never becomes
Gaussian and $\alpha_2(t)$ diverges at long times.
It is interesting that if one computes, for $\tau_{\rm link}=\infty$,
$\alpha_2(t)$ separately for the clustered and non-clustered particles,
only the peak at short times is observed, i.e. the one which is due to the
crowding effects. This is shown in the Inset of Fig.~\ref{alpha2taulink}
for the case of the system of particles where there is a permanently
frozen percolating cluster and few unattached particles. In this case,
the $\alpha_2(t)$ for the clustered particles follows the curve for
the full system, shows the bump at short times and becomes at long
times a constant (see the corresponding data for $\Delta^2(t)$ in
Fig.~\ref{figclustfree}). The result that at long times $\alpha_2(t)$
does not go to zero, as it would be the case for vibrations in a typical
amorphous solid with $R=0$, is related to the fact that the spanning
cluster is a disordered network of particles that have different
local environments. Thus the large variety of slow floppy motions,
related to the slow relaxation in $F_s(q,t)$ over long times (see
Fig.~\ref{figclustfree}), gives a non-Gaussian shape for $G_s(r,t)$
even at very long times. This non-Gaussianity is also the reason for
the exponential shape of $f_q^{\rm PN}(q)$, as observed earlier, see
Fig.~\ref{ap1}. For the non-clustered particles, $\alpha_2(t)$ does show
a slight decrease beyond the short-time maximum which shows that the
relaxation dynamics starts to become a bit more homogeneous. However,
it cannot be expected that the non-Gaussian parameter will go to zero
even at very long times, since the mobile clusters do have a spread in
size and hence a different diffusion constant.
\subsection{Four point susceptibility $\chi_4(q,t)$}
\begin{figure}
\includegraphics[width=80mm,clip=true]{fig12a_chi4R0.eps}
\includegraphics[width=80mm,clip=true]{fig12b_chi4phi0.7.eps}
\includegraphics[width=80mm,clip=true]{fig13_chi4_phi65_taulink.eps}
\caption{(Color online)
Time dependence of $\chi_4$ for $\tau_{\rm link}=100$. (a) $R=0$ and different values of $\phi$.
(b) $\phi=0.70$ and different values of $R$ (as marked on plot)
(c) $R=4, \phi=0.65$ for different values of $\tau_{\rm link}$ (as marked on plot).}
\label{chi4}
\end{figure}
We conclude by measuring to what extent the observed heterogeneous
dynamics is related to a correlated motion of the particles. This can
be quantified by $\chi_4(q,t)$, defined above, measured for $q=6.0$,
i.e. we look for relaxations over distances which are slightly larger
than the average particle diameter.
To have a reference we consider first the case $R=0$, i.e. when
there are no polymers and the system is just a standard glass-forming
system. The time dependence of $\chi_4(q,t)$ is shown in Fig.~\ref{chi4}a
for different packing fractions. In agreement with earlier studies on
similar systems, Ref.~\cite{glotzer}, we find that $\chi_4(q,t)$ shows
a maximum at a time $t$ that increases with $\phi$ and which tracks the
increasing relaxation time $\tau_{\rm SH}$. The fact that the height of
the peak increases with increasing $\phi$ indicates that the relaxation
dynamics become more cooperative, also this in agreement with previous
studies for other glass-forming systems~\cite{chi4ref}.
Next, we contrast this behavior with the case of increasing gelation,
i.e. we fix a volume fraction and increase the number of polymer
bonds. For this, we choose $\phi=0.70$, i.e. the same packing fraction
for which we have studied the correlation functions and the non-Gaussian
parameter, since for this $\phi$ the effects due to the steric hindrance
as well as the network constraints are clearly observed. The corresponding
$\chi_4(q,t)$ is shown in Fig.~\ref{chi4}b for $\tau_{\rm link}=10^2$. For
$0<R\leq 6$, the location of the peak in $\chi_4(q,t)$ shifts to larger
times and the height of the maximum increases. Thus this trend follows the
behavior observed in the data for $F_s(q,t)$ as well as $\alpha_2(t)$,
indicating that steric hindrance dominates the relaxation process and
that the dynamics becomes more and more collective. However, for $R>6$,
the peak height decreases again and its location becomes independent of
$R$. We can thus conclude that in this regime the relaxation dynamics
becomes less cooperative, a result that makes sense since the rewiring
of the network is not really a collective process. This result is also in
qualitative agreement with the decrease of the maximum in $\alpha_2(t)$,
see Fig.~\ref{alpha2phi}b, which showed that the dynamics becomes more
homogeneous.
We conclude by investigating how the lifetime of the network-bonds
influences the function $\chi_4(q,t)$. For this we have calculated this
observable for the case $\phi=0.65$ and $R=6$, i.e. at intermediate
density and in the gel phase for which one sees a significant
$\tau_{\rm link}$-dependence in the relaxation behavior (see
Fig.~\ref{fsq6t100tinf}b). We see, Fig.~\ref{chi4}c, that at small and
intermediate $\tau_{\rm link}$ the location of the peak in $\chi_4(q,t)$
moves to larger time and that its height increases somewhat, i.e. a
behavior that is directly related to the steric hindrance mechanism in
which the effective cage becomes stiffer due to the increased lifetime
of the bonds. However, once $\tau_{\rm link}$ exceeds $10^4$, we see that
$\chi_4(q,t)$ no longer depends on this lifetime, i.e. the rewiring of
the network no longer affects the cooperativity of the dynamics and
the latter is solely dependent on the steric hindrance.
\section{Summary and Conclusions}
In this paper, we have studied a coarse-grained model for a transient
network fluid, a system that can be realized in experiments by a
oil-in-water emulsion. By tuning the volume fraction $\phi$ of the
constituent particles, the number of bonds $R$ between the mesoparticle,
and the lifetime $\tau_{\rm link}$ of these connections, we have scanned
across the phase diagram to investigate the relaxation dynamics of this
system, in particular the interplay between the gel-transition and the
glass transition induced by steric hindrance effects.
By analyzing the mean-squared displacement and self intermediate
scattering function we find that the nature of the slowing down depends
on the packing fraction: At low $\phi$ an increase of $R$ leads to a
two step relaxation with a plateau that increases continuously with $R$
and an $\alpha-$relaxation time that is independent of $R$ and that
is related to the lifetime of the bonds. In contrast to this we find at
large $\phi$ a two step relaxation with a plateau height that is basically
independent of $R$ whereas the relaxation time does depend on the density
of bonds. Thus these two different behaviors show that the system can
show a glassy dynamics that is related on one hand to a gel transition
and on the other hand to a glass transition associated with the steric
hindrance mechanism. For intermediate densities and a certain range of $R$
and $\tau_{\rm link}$ the interplay between these two mechanisms leads to
a decay of the time correlation function that is logarithmic over several
decades in time, whereas for other combinations it gives rise to a three
step relaxation. We note here that a similar multistep relaxation pattern
has been recently reported for mixtures of \emph{multiarm} telechelic
polymers and oil-in-water microemulsions \cite{malo}, suggesting the
existence of a competition of different arrest length- and time-scales
in these systems. We have also studied how the height of the plateau at
long times, i.e. the Debye-Waller factor, depends on the wave-vector
and found that it decays in an exponential manner in $q$, i.e. very different
from the Gaussian decay observed in more standard glass-forming systems.
Furthermore we have determined the relaxation times of the system and
find that these can approximately be factorized into a function that
depends strongly on $R$ and a Vogel-Fulcher type dependence on $\phi$.
The $R-$dependent factor shows for $R\leq 10$ a strong exponential
dependence that is related to the escape of the particles from the
local cage, whereas for larger values of $R$ the dependence is weaker
and linked to an Arrhenius-like process for bond-breaking.
By studying the non-Gaussian parameter we have probed to what extent
the relaxation dynamics of the system is heterogeneous. At low packing
fraction this dynamics becomes extremely heterogeneous, if $R$ is not too
large, since some of the particles are very strongly connected to their
neighbors (and hence are immobile) whereas others can move almost freely.
However, if $R$ becomes larger than 6, the relaxation dynamics becomes
again quite homogeneous since at any time all particles are well connected
to their neighbors. At large packing fractions the non-Gaussian parameter
shows a double peak structure and, by monitoring the van Hove function,
we can show that this feature is directly related to the two relaxation
processes, i.e. the steric hindrance and the connectivity of the network.
Whether or not the relaxation dynamics is cooperative can be characterized
by the four-point correlation function $\chi_4(q,t)$. We find that for
large packing fractions the height of the peak in $\chi_4(q,t)$ increases
rapidly with $R$, showing that the strengthened coupling leads to an
enhanced cooperative motion. However, if $R$ is increased beyond 6, this
cooperativity decreases again, since the relaxation dynamics is strongly
dominated by the rewiring of the network, i.e. a non-cooperative process.
Summarizing we see that the interplay between the two different mechanisms
giving rise to glassy behavior can lead a rather complex and unusual
relaxation dynamics. In the present study we have focused on a system in
which the particles are connected by polymers having a fixed extension
length $\ell=3.5\sigma$. For much smaller extension lengths, one can
expect to recover the re-entrant scenario observed in colloidal gels
\cite{zaccarellirev}. For longer extension lengths, the particles will
instead have more space to explore and it will likely result in a even
more floppy network. In the future, it would certainly be worthwhile to
explore also systems that have polymers with different value of $\ell$,
since this implies multiple localization lengths and thus provide novel
relaxation scenarios that can subsequently also studied in experiments.
And finally we recall that our model is motivated by an experimental
system which has highly nontrivial rheological properties, e.g. this
material flows like a liquid but eventually breaks as a brittle solid
\cite{fracture,ligoure}, being capable of self-healing through thermal
fluctuations. Thus, in the future we plan to study the rheological
properties of our model in order to understand the microscopic mechanisms
that lead to the experimental observations.
\acknowledgments
Financial support from ANR's TSANET is acknowledged. PIH acknowledges
financial support from Spanish MICINN project FIS2009-08451, FIS2013-43201-P, University
of Granada, Junta de Andaluc\'{\i}a projects P06-FQM1505, P09-FQM4682
and GENIL PYR-2014-13 project. WK is member of the Institut Universitaire
de France.
|
\section{Introduction}
Since the dawn time the chaos is an indispensable part of human life. One of the most famous models with chaotic behavior is the logistic map.
Its simple analytic form meant that it was used in many scientific disciplines such as biology, cryptography, communication chemical physics and stock market ~\cite{pareeka, ferrettia, pellicer, andrecut}.
One of the mathematical models of chaos was discovered by Mitchell
Feigenbaum (1978)~\cite{Fei1, Fei2, Thom1, Steph1}. Feigenbaum considered ordinary difference
equations used for example in biology to describe the development
of population in its dependence on time. He discovered that
population oscillates in time between stable powers (fixed
points), the number of which doubles according to changes in the
power of external parameter. The generalization of the Feigenbaum
model contains all first-order difference equations
$f(x_n)=x_{n+1}$. The condition for the existence of chaos is the
single maximum of a function $f(x_n)$. Feigenbaum also proved that
the transition to chaos is described by two universal constans
$\alpha$ and $\delta$, later named the Feigenbaum numbers.\\
We quote here a few definitions which are basic in the
development of this exposition.
\bf Definition 1.\rm ~[cf. \cite{Smital}, p. 62] \\
Let $I=[a,b]\subset\Re$, $a<b$, $f(I)\subset I$. Discrete
iterations of a function $f:I\rightarrow I$ are called functions
$f^n:I\rightarrow I$ defined inductively: $f^0=id$,
$f^{n+1}=f\circ {f^n}$,
$n=1,2,3,...$.
\bf Definition 2.\rm ~[\cite{Smital}, p. 62] \\
The sequence $\{x_n\}$, where $x_n=f^n(x_0)$ is called the
sequence of iterates of a function $f$ generated by point
$x_0\in I$.
\bf Definition 3.\rm ~[cf. \cite{Smital}, p. 65]\\
A point $s\in I$, with $f(s)=s$ is called a first-order fixed
point of $f$. A fixed point $s$ of a function $f$ is locally
stable if there exists a neighbourhood $U_s$ such that all
sequences of iterates are convergent to $s$ for $x_0\in U_s$.
\\
$\mathbf{Theorem 1.}$ [see \cite{Smital}, p. 68] \\
If function $f$ is differentiable in fixed point $s$ and $|f^{'}(s)|<1$, then $s$ is attractive.
\bf Definition 4.\rm ~[cf. \cite{Smital}, p. 72]\\
Let $f:I\rightarrow I$. We say that $s$ is a periodic point of the
function $f$, or that $s$ generates a cycle of order $k\geq 2$, if
$f^k(s)=s$ and $\forall_{i<k} f^{(i)}(s)\not =s$.
\\
$\mathbf{Theorem 2.}$ \cite{Li1}\\
If function $f$ has a cycle of order different to $2^k$, then $f$ is chaotic.
\\
$\mathbf{Theorem 3.}$ [cf. \cite{Smital}, p. 79]\\
For arbitrary continuous function $f:I\rightarrow I$ existence cycles of order $m$
involve the existence cycles of order $n$ in the following
order:\\
$3>>5>>7>>2\cdot 3>>2\cdot 5>>2\cdot 7...>>2^k\cdot 3>>2^k\cdot
7>>...>>2^k>>...>>8>>4>>2>>1$.
\begin{center}
{\it \large {The conventional logistic map}}
\end{center}
The Logistic map ~\cite{verhulst, Ott1, Schu1},
\begin{equation}
f(x_n)=x_{n+1}=rx_n(1-x_n),\; x\in[0,1],\; (p,q)>0,\; n=0,1,2,...
\end{equation}
is one of the most simple forms of a chaotic process. Basically,
this map, like any one-dimensional map, is a rule for getting a
number from a number. The parameter $r$ is fixed, but if one
studies the map for different values of $r$ (up to 4, else the
unit interval is no longer invariant) it is found that $r$ is the
catalyst for chaos. After many iterations $x$ reaches some values
independent of its starting value 3 regimes:
$r<1:\; x=0$ for large $n$,\quad $1<r<3:\; x=constant$ for large
$n$,\quad $3<r<r_{\infty}:\;$ cyclic behavior, where $r_{\infty}\approx3.56$,\quad
$r_{\infty}<r\leq 4:\;$ mostly chaotic,\quad \\
An interesting thing happens if a value of $r$ greater than 3 is
chosen. The map becomes unstable and we get a pitchfork
bifurcation with two stable orbits of period two corresponding to
the two stable fixed points of the second iteration of $f$. With
$r$ slightly bigger than 3.54, the population will oscillate
between 8 values, then 16, 32, etc. The lengths of the parameter
intervals which yield the same number of oscillations decrease
rapidly; the ratio between the lengths of two successive such
bifurcation intervals approaches the $Feigenbaum~~constant$
$\delta =4.669...$. The period doubling bifurcations come faster
and faster ( 8, 16, 32, ...), then suddenly break off. Beyond a
certain point, known as the accumulation point $r_{\infty}$,
periodicity gives way to chaos. In the middle of the complexity, a
window suddenly appears with a regular period like 3 or 7 as a
result of mode locking. The 3-period bifurcation occurs at
$r=1+2\sqrt{2}$, and period doubling then begin again
with cycles of 6,12,... and 7,14,28,... and then once again break
off to chaos.
\section{Generalized logistic map}
\quad One-dimensional representations are the simplest
arrangements capable to producing a chaotic movement.
Let us consider $p,q>0$. We postulate non-linear function:
\begin{equation}
f_r(x)=rx^p(1-x^q),\; x\in[0,1]
\label{glm}
\end{equation}
set by difference equation:
\begin{equation}
x_{n+1}=rx^p_n(1-x^q_n),\;x\in[0,1],\; n=0,1,2,....
\label{genlogmap}
\end{equation}
For the first time the function (\ref{glm}) has been presented in \cite{rak}.
The maximum of the function (\ref{glm}) is reached for
$x=\left(p \over p+q\right)^{1/q}$, because
$f^{\;'}_{r}(x)=r(px^{p-1}-(p+q)x^{p+q-1})$
and $f^{\;'}_{r}(x)=0\; \Leftrightarrow\; x=\left(p \over
p+q\right)^{1/q}$, $\Re\ni(p,q)>0$.\\
It results that $f_{r}\left(\left(p \over p+q\right)^{1\over q
}\right)$=$rqp^{p\over q}\over (p+q)^{p+q\over q}$ and for $r\in
\left(0,\; {({p+q\over p})^{p \over q}(p+q)\over q}\right)$
where\\
$\left(0,\; {({p+q\over p})^{p \over q}(p+q)\over
q}\right)=r_{max}$ function $f_r$ is $f([0,1])\subset [0,1]$.\\
Parameter $r_{max}$ depending on $p$ and $q$ accepts the following
limit values (Fig.\ref{fig1}):
$$\lim_{q\to \infty} r_{max}=1, \; \; \lim_{q\to
0^+}r_{max}=\infty$$
$$\lim_{p\to \infty}r_{max}=\infty, \; \;
\lim_{p\to 0^+}r_{max}=1$$
\begin{figure}[!ht]
\begin{center}
\epsfxsize 12cm
\hspace*{0.3in}
\epsffile{fig1.eps}
\end{center}
\caption{Values of $r_{max}$ for generalized logistic map depending on $p$ and $q$ (left) and depending on $k$ (right).}
\label{fig1}
\end{figure}
For the sake of parameters $p$ and $q$ the function (3) generates
a lot of new functions which can be analyzed using mathematical
methods, just like for a logistic map,
thus obtaining very interesting results.\\
a)\quad Let us assume that $k={q\over p}$ and $k>0$. Then $r$ is depend
on the value $k$ and $r\in\left[0,\frac{(k+1)^{\frac{k+1}{k}}}{k}\right].$\\
b) \quad If $k=1$, then $p=q$ and $r_{max}$ is independent
on $p$, $q$ and $r_{max}=4$ always - for logistic map, we have $p=q=1$ and $r_{max}=4$.\\
c) \quad If $\quad k \rightarrow {\infty}$, then $r_{max} \rightarrow 1$, because
$\lim_{k\to \infty } \,\frac{(k+1)^{\frac{k+1}{k}}}{k}=1.$\\
d) \quad If $k\rightarrow {0^{+}}$, then
$r_{max}\rightarrow {\infty}$, because
$\lim_{k\to \infty } \,\frac{(k+1)^{\frac{k+1}{k}}}{k}=+\infty.$\\
e) \quad Now let us consider a case, when $p=1$ and $q>0$. Then we have
\begin{equation}
f_{r}(x)=rx(1-x^q),~~x\in[0,1],~~ n=0,1,2,....
\end{equation}
and $r\in\left[0,\frac{(q+1)^{\frac{q+1}{q}}}{q}\right].$
Taking into account the above relationship we can observe change maximum of the function (5) depending on $q$ (Fig.\ref{fig2}).
\begin{figure}[!ht]
\epsfxsize 10cm
\hspace*{1in}
\epsffile{fig2.eps}
\caption{$f_r(x)=rx(1-x^q)$ for $r=r_{max}$ depending on $q$.}
\label{fig2}
\end{figure}
\newpage
\bf A)\quad The property of $\bf x_{n+1}=rx^p_n(1-x^q_n)$ for $\bf p=1$, $\bf q=2$ in the periodic region.\rm\\
In the following part of this paper we will show that the
Feigenbaum model of transition to chaos is correct for
representation (4) for given positive values of parameters $p$
and $q$. Therefore, we will study the dynamics of this map in the
case of $p=1$ and $q=2$. Then the representation (5) assumes the
following form:
\begin{equation}
f_{r}(x)=rx(1-x^2),\;\; x\in[0,1],\;\; n=0,1,2,...
\label{eq12}
\end{equation}
A function $f_r$ in the point $x={\sqrt3\over 3}$ has a local maximum
and the value of function in this point is equal to
$f_{r}({\sqrt3\over 3})={2r\sqrt3\over9}$. From this relationship, we have
$f([0,1])\subset [0,1]$ for $r\in(0,{3\sqrt3\over 2}]$. The
trajectories behavior generated by $f_r$ depends on $r$ and,
therefore, we have to distinguish several intervals of this
value:\\
1)\quad For $r\in(0,1]$ all the sequences of iterates
$x_1,x_2,x_3,...,x_n$ depend on the initial value $x_0$ tend to
$s=0$. This point is the fixed point of the map for $r\in(0,1]$
and according to \it Definition 3 \rm it is also the stable point
of this map.\\
2)\quad Let $r=r_{1}>1$. If $x=rx(1-x^2)$ then, $x_{1}=0$,
$x_{2}=\sqrt{r-1\over r}$, so for $r>1$ function (6) has two fixed
points. Using \it Theorem 1 \rm we get: $f^{\; '}_{r}(x)=r-3rx^2$,
$|f^{\; '}_{r}(0)|=r$ and $|f^{\; '}_{r}(\sqrt{r-1\over
r})|=|3-2r|$.\\
Therefore, we have that $x_1=0$ is an unstable fixed point,
whereas $x_{2}=\sqrt{r-1\over r}$ is an attractive fixed point for
$r\in (1,2)$.\\
Change $r$ from 1 to 2 has caused the migration of the attractive
point from 0 to ${\sqrt2\over 2}$.\\
\it Conclusion 1. \rm \\
For $1<r<2$ there exists a stable fixed
point attracting all points. When $r=r_{1}=2$, then $|f^{\;
'}_{r}(\sqrt{r-1\over
r})|=|3-2r|=1$ and this point stops being an attractive point.\\
3)\quad For $r=r_{2}\geq 2$ we will look into the stability of
fixed points $f_r$ and $f_r^{(2)}$ of the map (5) as a function of
parameter $r$, because the Feigenbaum model is generated by
bifurcations connected with iterations of a function.
Let us consider the second iteration of $f_r$, namely
$f^{(2)}_{r}(x)=r^2x - r^2x^3 - r^4x^3 + 3r^4x^5 - 3r^4x^7 + r^4x^9$, with fixed points: $x_{1}=0$, $x_{2}=\sqrt{r-1\over r}$,
$x_{3}=\sqrt {{1\over 2} - {\sqrt {r^2-4}\over 2r}}$, $x_{4}=\sqrt
{{1\over 2} + {\sqrt {r^2-4}\over 2r}}$.
Because $f^{\; '(2)}_{r}(x)=r^2 - 3r^2x^2 - 3r^4x^2 + 15r^4x^4 -21r^4x^6 + 9r^4x^8$
we have $f^{\; '(2)}_{r}(x_1)=r^2$, $f^{\;'(2)}_{r}(x_2)=(3-2r)^2$, $f^{\; '(2)}_{r}(x_3)=f^{\;'(2)}_{r}(x_4)=9-2r^2$ whereas for $2<r_2<\sqrt 5$ fixed points
$x_3$ and $x_4$ are attractive points according to condition in
\it Theorem 1 \rm.\\
\it Conclusion 2.\rm \\
The fixed point $x=\sqrt{r-1\over r}$ of $f_r^{(1)}$ is also the
fixed point of $f_r^{(2)}$ as well as of all higher iterations.\\
\it Conclusion 3.\rm \\
If the fixed point of $f_r^{(1)}$ becomes unstable, then it is
also the unstable fixed point of $f_r^{(2)}$ and of all next
iterations.\\
From the inequality $|f^{\; '}(s)|>1$ we have
$|f^{\; '(2)}(s)|=|f^{\; '}[f(s)]f^{\; '}(s)|=|f^{\; '}(s)|>1$.\\
If $\sqrt 5<r_3<r_4$, then the fixed points of $f^{(2)}_r$ become
repulsive at the same time. Following this instability, the fourth
iteration shows two new pitchfork bifurcations giving cycle $2^2$
order for four attractive fixed points, which is called period
duplication.
To generalize the above examples we get:\\
\bf a)\rm \quad For $r_{n-1}<r<r_n$ there exists a stable cycle
$2^{n-1}$, whose elements
$\overline{x}_0,\overline{x}_1,\overline{x}_2,\overline{x}_3,...,\overline{x}_{2^{n-1}-1}$
are defined by the following
connections:\\
$$f_{r}(\overline{x} _i)=\overline{x}_{i+1},\quad \quad
f_{r}^{2^{n-1}}(\overline{x} _i)=\overline{x}_{i},\quad \quad
|\prod_{i}f_{r}^{\; '}(\overline{x}_i)|<1$$\\
\bf b)\rm \quad For $r_n$ all points of the cycle of $2^{n-1}$
become unstable simultaneously and the bifurcation creates a new
stable cycle of $2^n$ order for
$f_{r}^{2^n}=f_{r}^{2^{n-1}}(f_{r}^{2^{n-1}})$ existing for
$r_{n-1}<r<r_n$.\\
\bf c)\rm \quad For a map (6) the value of parameter $r_{n+1}$ we can
describe the following connection: $r_{n+1}\approx \sqrt{3+r_n},$
where $r_{0}=1,\; \; r_1=2,r_2=\sqrt{5},...,r_{\infty}$.\\
\bf d)\rm \quad There exists a point of accumulation of an infinite
number of the bifurcation of period duplication for the finite value
$r$ which we appointed as $r_{\infty}$: $r_{\infty}=\lim_{n\to
\infty}r_{n}={1+\sqrt{13}\over
2}\approx{2.303}$.\\
Taking into account the above, $r_{n+1}\approx r_{n}$,\quad $\sqrt{3+r}=r$,
$r^{2}-r-3=0$,\quad thus $r={1+\sqrt{13}\over 2}$.\\
\bf e)\rm \quad For the representation (6) there exists a constant,
whose value corresponds to the Feigenbaum constant:
$$\delta= \lim_{n\to \infty}{r_{n}-r_{n-1}\over{r_{n+1}-r_{n}}}\approx{4.61}.$$\\
\it Proof.\rm \quad Let us assume that $r_{n-1}=a$. Then $r_{n}=\sqrt{3+r_{n-1}}=\sqrt{3+a}$,
$r_{n+1}=\sqrt{3+r_{n}}=\sqrt{3+\sqrt{3+a}}$ and
$\delta=\lim_{n\to\infty}{r_{n}-r_{n-1}\over{r_{n+1}-r_{n}}}=\sqrt{14+2\sqrt{13}}\approx{4.61}$.\\
A table of $2^n$ cycles type and values of $r_n$ is given below. It is clearly visible that doubling bifurcations come faster and
faster (8, 16, 32, ...) and suddenly break off beyond a certain point known as the accumulation point and periodicity gives way to chaos.
\begin{table}[!ht]
\epsfxsize 12.5cm \hspace{0cm} \epsffile{tab1.eps} \caption{The
algebraic orders of the values of $r_n$ for $n=1,2,...$ are given
by $1,2,4,8,...$.} \label{tab1}
\end{table}
\\
\bf B)\quad The property of $\bf x_{n+1}=rx^p_n(1-x^q_n)$
for $\bf p=1$, $\bf q=2$ in chaotic region.\rm
\\
\bf a)\rm \quad For $r_{\infty}$ the bifurcation sequence ends
with a set of the infinitely numerous points, which is called the
set of Feigenbaum attraction. \\
\bf b)\rm \quad If $r_{\infty}<r\leq {3\sqrt{3}\over{2}}$ then we
observe the irregular dynamics of map (6) with narrow ranges in
which the set of attraction has a periodical character.\\
\bf c)\rm \quad Numeric calculations for the representation (6) show
that the largest range is for $r=r_{c}\approx 2.451$. Then
there exists 3-cycle (Fig.\ref{fig3}), which according to th.2 and th.3,
implicates the chaos as well as the existence of cycles of
different orders.\\
\bf d)\rm \quad For the map (6) at $r=r_{max}={3\sqrt 3\over 2}$,\\
\begin{figure}[!ht]
\epsfxsize 14cm
\hspace{2.4cm}
\epsffile{fig3.eps}
\caption{Iterates of map (6) strting from $x_{0}=0.99$. Left: In the
stable 3-cycle region for $r=r_c$;~~Right: In the chaotic region for
$r=r_{c}-\Delta r,\; \Delta r=0.001$ .}
\label{fig3}
\end{figure}
\begin{equation}
x_{n+1}={3\sqrt 3\over 2}x_{n}(1-x_{n}^2)\equiv
f_{r_{max}}(x_{n}),
\end{equation}
can be solved by the simple change of
variables:
\begin{equation}
x_{n}=\cos{(y_{n})}\equiv g(y_{n}).
\end{equation}
Then (7) can be converted into\\
$$\cos{(y_{n+1})}={3\sqrt 3\over 2}\cos{(y_n)}[1-\cos{^2(y_n)}]={3\sqrt 3\over
2}\cos{(y_n)}\sin{^2(y_n)},$$\\
which has one solution $y_{n+1}=\arccos{\left[{3\sqrt 3\over
2}\cos{(y_n)}\sin{^2(y_n)}\right]}$ and \\
$$x_{n}=\arccos (\frac{1}{{\sqrt{3}}\,{\left( -{y_0} +{\sqrt{-1 + {{y_0}}^2}} \right) }^{\frac{1}{3}}}+
\frac{{\left( -{y_0} + {\sqrt{-1 + {{y_0}}^2}} \right) }^
{\frac{1}{3}}}{{\sqrt{3}}}).$$\\
The invariant density function $\rho_{{3\sqrt 3\over 2}}(x)$ of $f_{{3\sqrt
3\over 2}}(x)$ can be calculated using the following definition:\\
\begin{equation}
\lim_{N\to \infty}{1\over N}\sum_{n = 0}^{N-1}
\delta \,\left( x - {x_n} \right)=\lim_{N\to \infty}{1\over N}\sum_{n = 0}^{N-1}
\delta \,\left( x - {g(y_n)} \right)
\end{equation} Using $\rho(y)=1$, (9) becomes: $\rho_{{3\sqrt 3\over 2}}(x)=\int_0^{1}dy\rho(y)\delta[x-h(y)]$\\
i.e.
\begin{equation}
\rho_{{3\sqrt 3\over 2}}(x)=\frac{\left(\sqrt{x^2-1}-x\right)^{2/3}-1}{3 \sqrt{x^2-1}
\sqrt[3]{\sqrt{x^2-1}-x}
\sqrt{-\left(\sqrt{x^2-1}-x\right)^{2/3}-\frac{1}{\left(\sqrt{x^2-1}-x
\right)^{2/3}}+1}} .
\end{equation}
The above function is shown in Fig.\ref{fig4}.
\\
\begin{figure}[!ht]
\epsfxsize 14cm
\hspace{2.5cm}
\epsffile{fig4.eps}
\caption{The invariant density function of map (6) for $r=r_{max}={3\sqrt 3\over 2}$. Left: An analytical form represented by the equation (10). Right: Numerical simulation - the histogram for 50000 iterations.}
\label{fig4}
\end{figure}
The invariant density function of the logistic map for $r=r_{max}=4$ is given by $\frac{1}{\pi \sqrt{(1-x) x}}$.
\\
\\
\bf C)\quad Ordered chaos - construction of an attractor for\\
$\bf x_{n+1}=rx_{n}(1-x_{n}^2)$ and $\bf r=r_{max}={3\sqrt
3\over2}$.\rm \\
\\
At the end of this considerations we will present a method of the
{\it ordering of chaos}. A lot of numeric experiments have led
to the situation where it is possible to rank chaos in such a way
that the successive iterations generated by chaotic dynamic system
would reach an attractor of a regular shape. This method also has
it, that from the sequence of several thousand iterations (for
representation $x_{n+1}=rx_{n}(1-x_{n}^2)$ where $r=r_{max}={3\sqrt
3\over 2}$ defined recursively) we draw up the following points
$(x_{n+1},x_{n}x_{n+1})$ on a plane - we construct an
attractor. In this way an attractor comes into being, which is visited irregularly by points of successive iterations (Fig.\ref{fig5}).\\
\begin{figure}[!ht]
\epsfxsize 13.5cm
\hspace{-0.3cm}
\epsffile{fig5.eps}
\caption{The set of attraction for representation (6) [60000 iterations]: a) for $r=r_{max}$(red), $r=2.56$(blue), $r=2.53$(green), $r=2.45$(orange);~~b) for $r=r_{max}$(red), $r=2.42$(black), $r=2.4$(blue).}
\label{fig5}
\end{figure}
\newpage
\bf D)\quad Bifurcation diagram and Lyapunov exponent.\rm \\
\\
Nowadays the Lyapunov exponent is applicable in various scientific disciplines such as biology, engineering, bioengineering and informatics ~\cite{lap11, lap1, lap2, lap3, lap4, lap5}.
A quantitative measure of the sensitive dependence to initial conditions is the Lyapunov exponent, which is a measure of the exponential separation of nearby orbits. For discrete time system $x_{n+1}=f(x_n)$, for an orbit starting with $x_0$ the Lyapunov exponent $\lambda(x_0)$ can be calculated by
\begin{equation}
\lambda(x_0)=\lim_{n\to \infty} \frac{1}{n}\sum_{i = 0}^{n-1}\ln|f'(x_i)|,
\label{lyapunov}
\end{equation}
A positive Lyapunov exponent can be considered as an indicator of chaos, whereas negative exponents are associated with regular and periodic behavior of the system.
Figure \ref{fig6} shows the bifurcation diagram and Lyapunov exponent for the equation (\ref{eq12}). It is clear that even in chaotic areas ($r_{\infty}<r\leq {3\sqrt{3}\over{2}}$) there are many periodic intervals ($\lambda<0$).
\begin{figure}[!ht]
\epsfxsize 9cm
\hspace{2cm}
\epsffile{fig6.eps}
\caption{The bifurcation diagram and the Lyapunov exponent for the map ~(\ref{eq12}) and $2.2<r\leq {3\sqrt{3}\over{2}}$.}
\label{fig6}
\end{figure}
\section{Summary}
Currently, modeling real-world systems is very popular. A well-known model that takes into account the deterministic chaos is the logistic map.
However, this one-dimensional map can be controlled only by one parameter. In the present contribution, we postulate a generalization of the classical logistic map which is generally controlled by three parameters. This could allow to adapt to the conditions prevailing in the modeled system.
We focus on the $r_{max}$ parameter (depending on $p$ and $q$) which is responsible for the dynamics of the system - among others shown that if $p=q$ then $r_{max}=4$.
As an example, we carry out analytical and quantitative analysis in the case where $p=1$ and $q=2$ both in the periodic and chaotic regime where the Lyapunov exponent has positive values. It turns out that the dynamics of this equation is faster than in the case of the logistic map - the value of the parameter $r_{max}\approx2.3$ is less than the value of $r_{max}=4$ for the logistic map.
In the periodic regime we identified Feigenbaum constant $\delta$ which is typical for all dissipative systems (also for the logistic map).
For chaotic area we have found an analytical form of the invariant density function. For both regimes we have proposed specific form of data representation which allowed to obtain a non-trivial structure of an attractor set.\\
\bf Footnotes \rm \\
This work was partially supported by the Centre for Innovation and Transfer of Natural Sciences and Engineering Knowledge (University of Rzeszow).
|
\section{Introduction}
The general relativity (GR) is a very successful theory of gravity,
but it is perhaps not an ultimate theory. One of the reasons is that
the fourth derivative terms in the action of gravity become necessary
as the UV completion of the theory at the semiclassical level \cite{UtDW}
(see also \cite{birdav,book} for the introduction and \cite{PoImpo} for
a recent pedagogical review). The same
fourth-derivative terms make the theory of quantum gravity (QG)
renormalizable \cite{Stelle-77}. On the other side, fourth
derivatives lead to the massive ghosts in the physical spectrum
of the theory, leading to the violation of of unitarity.
The consistency of the fourth derivative quantum gravity (QG)
can be, in principle, achieved by dealing with the dressed propagator
instead of the classical one \cite{Tomboulis-77,salstr,antomb}. The
main expectation is that the massive ghost poles become unstable
and decay in the far future, such that the asymptotic {\it out}-state
becomes free of ghosts. Unfortunately, the final conclusion concerning
this approach requires a complete non-perturbative knowledge of the
dressed propagator \cite{johnston}, which is unavailable.
Some years ago a completely different approach was proposed by
Tomboulis \cite{Tomboulis-97}. The action of this new theory of
QG has an infinite amount of derivatives. It was discovered
a few years earlier by Tseytlin \cite{Tseytlin-95} that for some
specially tuned form of the non-local action such a theory is
free of ghosts at the tree level while the exponential form
factors remove Newtonian singularity, similar to the much simpler
fourth derivative gravity \cite{Stelle-77}. The approach of
\cite{Tseytlin-95} (see also \cite{Siegel-03}) was to use this
action in the framework of string theory, as an alternative of
the Zweibach ghost-killing transformation of the background
fields \cite{zwei,dere,Tse}. In string theory the
ghost-free non-local action is a kind of a ``final product'',
which is not supposed to gain further quantum
corrections\footnote{However, this does not make it free of
ambiguity related to the third and higher powers of curvature,
similar to the one discussed in \cite{marot}.}.
On the contrary, if one takes the same model as a basis of
quantum gravity \cite{Tomboulis-97}, the following three
important questions should be answered:
\begin{itemize}
\item
First, how to quantize the non-local theory?
\item Second, what is the power counting in a theory with infinite
amount of derivatives?
\item
The third and most difficult question is what happens with the
ghost-free structure of the theory after the quantum corrections
are taken into account?
\end{itemize}
Concerning the first point, the quantization of non-local theories
has been discussed in the literature \cite{Efimov} and is
relatively well-understood. The second issue has been explored in
\cite{Tomboulis-97} and in the more recent publications
\cite{Mode-fin,LMT,BiMa}. The main conclusion is that the power
counting in the non-local theory of \cite{Tomboulis-97} is the same
as in the local higher derivative superrenormalizable QG suggested
earlier in \cite{highderi}. Moreover, in both cases there is a chance
to make such a QG theory finite. This can be certainly achieved in
the local case \cite{highderi} and very likely in the non-local
one\footnote{In the odd space-time dimensions this can be easily
proved \cite{Mode-fin}.}.
In the present work we will mainly address the third question.
There are strong arguments that at least the most simple example
of the non-local theory suggested in \cite{Tomboulis-97} does not
remain free of ghost-like states at the quantum level. The last
means that the quantum corrections lead to an infinite amount of
the ghost-like states in the dressed propagator. The relation
between ghosts in the classical (naked) and dressed propagators
is almost opposite to what was expected in the fourth-derivative
renormalized theory of QG \cite{Stelle-77,Tomboulis-77,salstr,antomb}.
The paper is organized as follows. In Sect. 2 we present a brief
review of the non-local gravity which is ghost-free at the tree-level.
In Sect. 3 we explain the power counting in the non-local model, here
our consideration mainly follows previous publications
\cite{Tomboulis-97}, \cite{Mode-fin} and \cite{BiMa}, but we try to
make it more transparent, especially by comparing to the local
superrenormalizable QG case \cite{highderi}. Some relevant details
concerning Lagrangian quantization of the non-local theory are
settled in the Appendix. In Sect. 4 it is shown how the ghost-free
structure is violated by quantum corrections to the propagator.
In Sect. 5 we discuss the modified Newtonian limit in the non-local
theory and the possible role played by the ``hidden'' ghosts.
Finally, in the last section we draw our conclusions.
\section{Non-local ghost-free models}
The simplest way to count degrees of freedom in QG is based on
the analysis of the tree-level propagator on the flat background.
In most of the theories this procedure gives the same result as
canonical quantization \cite{GitTyu,book}. In order to explore
the flat-space propagator, the relevant part of the classical
action is at most bilinear in the curvature tensor,
\beq
\label{act}
S = \int d^4 x \sqrt{-g} \,
\Big\{ - \frac{1}{\kappa^2} \,R
+ R \, F_1(\Box) \, R
+ R_{\mu\nu} \, F_2 (\Box)\, R^{\mu\nu}
+ R_{\mu\nu\alpha\beta} \, F_3 (\Box) \, R^{\mu\nu\alpha\beta}\Big\}\,.
\eeq
Here $\kappa^2=16\pi G$ and $F_{1,2,3}$ are functions of d'Alembertian
operator. The cosmological constant term is set to zero, following
the standard treatment \cite{Stelle-77}.
In order to simplify the action, let us note that the difference
between the term $R_{\mu\nu\alpha\beta} F_3 (\Box) R^{\mu\nu\alpha\beta}$ and
the combination $4R_{\mu\nu} F_3 (\Box) R^{\mu\nu} - R F_3 (\Box) R$
is proportional to the term of the third power in curvature,
${\cal O}(R^3_{\dots})$ (see, e.g., \cite{highderi,NewSing}).
Therefore one can cast the relevant part of
the action (\ref{act}) in the form
\beq
\label{act Phi}
S = \int d^4 x \sqrt{-g} \,
\Big\{ - \frac{1}{\kappa^2} \,R
+ \frac12\,C_{\mu\nu\alpha\beta} \, \Phi(\Box) \, C_{\mu\nu\alpha\beta}
+ \frac12\,R \, \Psi (\Box)\, R\Big\}\,,
\eeq
where $C_{\mu\nu\alpha\beta}$ is the Weyl tensor.
The function $\Psi$ is responsible for the spin-0 part of the
propagator and the function $\Phi$ for the spin-2 part. For the sake
of simplicity, we can mainly concentrate on the spin-2 sector. The
consideration for the $\Psi$-part would be very similar. After the
Fourier transformation, the relevant equation for defining the poles
of the propagator is \cite{Tomboulis-97}
\beq
p^2\,\big[1 \,+\, \kappa^2 p^2\Phi(-p^2)\big] &=& 0\,.
\label{spin-2}
\eeq
One can see that there is always a massless pole corresponding
to gravitons. For a constant $\Phi$ there is also a massive pole
corresponding to a spin-2 ghost, which may be also a tachyon.
For a non-constant polynomial function $\Phi$ there are always
ghost-like poles, real or complex \cite{highderi}. However, one
can choose the function $\Phi$ in such a way that there will not
be any other spin-2 pole, except the graviton $p^2=0$. The simplest
example of this sort is \cite{Tseytlin-95}
\beq
1 \,+\, \kappa^2 p^2\Phi(-p^2) &=& e^{\alpha p^2}\,,
\label{exp}
\eeq
where $\,\alpha\,$ is some constant of the dimension $mass^{-2}$.
One can find other entire functions which have the same
features \cite{Tomboulis-97,Mode-fin}, but for the sake of
simplicity we consider only (\ref{exp}).
Let us remember that the exponential function has two remarkable
properties. The equation $\exp z = 0$ has no real solutions and
only one very peculiar solution
\beq
z = - \infty + i\times 0
\label{inf}
\eeq
on the extended complex plane. At the same time, already the
equation $\exp z = A \neq 0$ has infinitely many complex solutions,
the same is true for
\beq
e^z = A z^2 \log z\,,
\label{exp-brok}
\eeq
which is the typical case for the exponential theory with
logarithmic quantum corrections.
These well-known features of exponential function mean, in our
case, that the absence of massive ghosts in the spin-2 part of
the propagator of the theory (\ref{exp}) is the result of an
absolutely precise tuning of the function $\Phi(-p^2)$. If this
tuning is violated by the loop corrections, then the ghosts-like
states will emerge in an infinite number. For instance, any
polynomial addition to the exponential function produce infinitely
many complex solutions.
One important note is in order. The expression ``ghosts-like
states'' mean that these states are not exactly the ``classical''
massive ghosts, that means states with positive square of mass
and negative kinetic energy. In the present case there are mostly
complex poles, that means a complex ``square of mass'' and complex
``kinetic energy''. This situation makes the particle interpretation
of these states rather complicated. We postpone the discussion of
this issue until another publication and will call these states
simply ghosts in what follows.
If the theory with more ghosts should be qualified worst, then
the exponential gravity (\ref{exp}) with violated absolute tuning
is worst than the polynomial version of superrenormalizable QG
\cite{highderi} (see also the next section), because the last
has only finite amount of ghosts. So, the main question
concerning the theory of exponential gravity (\ref{exp}) is
whether one can preserve an absolute tuning of (\ref{exp}) at
the quantum level. In the next sections we consider this issue
starting from the strongest effect related to the UV divergences
and related logarithmic running. For comparison, we also present
considerations for the mentioned polynomial model of QG.
\section{Power-counting in local and non-local QG}
Before discussing the dressed propagator
and possible violation of the absolute tuning in (\ref{exp}), let
us shortly review the renormalization properties of the theory
(\ref{act Phi}) and some its natural extensions. A brief survey of
the Lagrangian quantization of the theories such as (\ref{act Phi})
or (\ref{superre}) with some details related to non-local
versions of the theory can be found in the Appendix.
\subsection{Polynomial higher-derivative gravity}
The action of the general superrenormalizable polynomial model can
be written as
\beq
S
&=& S_{EH} \,+\,\int d^4x\sqrt{-g}\,\Big\{
d_1R_{\mu\nu\alpha\beta}^2 + d_2R_{\mu\nu}^2 + d_3R^2 + \, ...
\label{superre}
\\
&+&
c_1 R_{\mu\nu\alpha\beta} \Box^k R^{\mu\nu\alpha\beta}
+ c_2 R_{\mu\nu}\Box^k R^{\mu\nu}
+ c_3 R\Box^k R
\, + \dots + \, b_{1,2,..}R_{...}^{k+2}
\Big\}\,.
\nonumber
\eeq
where the omitted terms and $b_{1,2,..}R_{...}^{k+2}$ denote the set
of all covariant local terms with the derivatives up to
the order $2k+4$. The action includes not only quadratic in curvature
terms, but also generic ${\cal O}(R^3_{...})$ terms, and so on.
$\,d_{1,2,3},\,c_{1,2,3},\,...b_{1,2,\,...}\,$
are arbitrary coefficients.
In order to explore the superficial degree of divergence of the
theory one needs two relations, namely
\beq
D + d \,=\,\sum\limits_{l_{int}}(4-r_l)
\,-\,4n \,+\,4\,+\,\sum\limits_{\nu}K_\nu
\label{dD}
\eeq
for the power counting, and the topological relation
\beq
l_{int} = p + n - 1\,.
\label{topo}
\eeq
In these formulas $l_{int}$ is the number of internal lines with
the inverse power of momenta $r_l$ in the propagator, $n$ is the
number of vertices with $K_\nu$ derivatives and $p$ is the number
of loops.
On the {\it l.h.s.}, $d$ is the number of derivatives acting on the
external lines of a given diagram and $D$ is its superficial degree
of divergence.
In the theory (\ref{superre}) the
most divergent diagrams correspond to the vertices with maximal
number of derivatives, $K_\nu=2k+4$. One can always formulate
the theory (see \cite{Stelle-77,frts-82,highderi} and Appendix
of the present work) in such a way that $r_l \equiv 2k+4$ for all
fields. Then
it is an easy exercise to combine (\ref{dD}) and (\ref{topo}),
and the result is \cite{highderi}
\beq
d &=& 4\,+\,k(1-p)
\label{sup-pow}
\eeq
for the logarithmically divergent diagrams with $D=0$. The last
relation shows that the versions of QG with $k \geq 3$ have
only one-loop divergences. This means, the higher order
contributions may be also divergent, but they become finite
after we renormalize the one-loop sub-diagrams. Furthermore,
the possible counterterms may have only four, two and zero
mass dimensions. In other words, only the terms in the first
line of (\ref{superre}) needs to be renormalized. All terms with
derivatives higher than
four are not running. At the same time, the coefficients of these
higher derivative terms define the running of the cosmological
and Newton constants and of the coefficients $d_1$, $d_2$ and $d_3$.
The last two observations which will be used in the rest of the
paper and (as we shall see in what follows) can be applied also
to the exponential gravity, are as follows:
\begin{itemize}
\item
The running of the parameters $\,G$,
$\,\rho_\Lambda\,$ and $\,d_{1,2,3}\,$
is gauge-fixing independent, because the classical equations
of motion have more derivatives than the counterterms. In order
to understand this statement, let us remember that the
gauge-fixing dependence disappears on-shell (see, e.g.,
\cite{FRG-Lavrov} for further references on the subject).
The practical application of this feature to QG was discussed
in \cite{a}.
\item
The $\beta$-functions for the Newton constant and the ones of
$\,d_{1,2,3}\,$ are given by bi-linear and linear combinations
of the coefficients of the ${\cal O}(R^3_{...})$ and
${\cal O}(R^4_{...})$ terms in the action (\ref{superre}).
Therefore, one can provide to these $\beta$-functions any
desirable values by changing the corresponding
coefficients\footnote{
Of course, this does not mean that the explicit
derivation of these $\beta$-functions would not be interesting.
Since the potential result is a possibility to obtain exact
$\beta$-functions in some model of quantum gravity, this calculation
would worth the requested hard work anyway. From the physical side,
different choices of these coefficients may correspond to different
physical properties of the theory, so such a calculation would be
quite relevant.}.
The remarkable exception is the $\beta$-function for the cosmological
constant derived in \cite{highderi}. This unique $\beta$-function is
completely defined by the coefficients $c_{1,2,3}$ in (\ref{superre}).
\end{itemize}
\subsection{Exponential gravity}
In the exponential gravity theory the power counting
formula (\ref{dD}) has no much sense, because it leads to an
indefinite output of the $\infty-\infty$ type. At the same
time the topological relation (\ref{topo}) is working well
and shows that the theory is superrenormalizable
\cite{Tomboulis-97,Mode-fin,BiMa}. Let us consider
this point.
Each propagator gives contribution of
infinite negative powers of momenta, let us call it
${\cal I}$. With the vertices the situation is more complex,
because there are vertices with different powers of momenta.
Without loss of generality one can consider only the
diagrams with maximal divergence, when each vertex gives
contribution $-{\cal I}$. It is important that the two
symbols ${\cal I}$ and $-{\cal I}$ correspond to the same
power of infinity, for otherwise the relation should become
more complicated\footnote{This means, in particular, that
the value of $\alpha$ in the Eq. (\ref{exp}) must be identical
for both functions $\Phi(\Box)$ and $\Psi(\Box)$ in
(\ref{act Phi}), for otherwise the theory would be badly
non-renormalizable.}.
Then it is clear that
the diagram with more internal lines than vertices will be
automatically convergent and the diagram with more vertices
than internal lines will be strongly divergent. The relation
(\ref{topo}) tells us that the difference is
$\,l_{int} - n = p - 1$. This means that only the one-loop
diagrams with $p=1$ can be divergent. At the same time, the
presence of the exponential form factor does not change the
degree of divergence of the one-loop diagrams.
The power counting in the exponential gravity is performed by
the topological relation (\ref{topo}), without the formula
(\ref{dD}). Nevertheless, the result is exactly the same as
in the polynomial theory (\ref{superre}) for $k \geq 3$.
Namely, the divergences show up only at the
one-loop level, and the counterterms have zero, two and
four derivatives of the metric only. In other words, the
possible counterterms have the form
\beq
\Delta S
&=& \int d^4x\sqrt{-g}\,\big\{
a_1R_{\mu\nu\alpha\beta}^2 + a_2R_{\mu\nu}^2 + a_3R^2
+ a_4 \Box R + a_5R + a_6\big\}\,.
\label{DeltaS}
\eeq
The divergent coefficients $\,a_{1,2,...,6}\,$ are at most
${\cal O}(1/(n-4))$ in dimensional regularization.
Similar to the polynomial case, there is a chance to specially tune
the $\,{\cal O}(R^2_...)$, $\,{\cal O}(R^3_...)\,$ and
$\,{\cal O}(R^4_...)$-terms in the action (\ref{superre}) such that
the divergent coefficients $\,a_{1,2,...,6}\,$ in (\ref{DeltaS})
can be adjusted to have desirable values. In the case of
exponential QG one should try to provide these divergent
coefficients to become zero, because the possible running would
violate an absolute tuning requested by the ghost-free structure
of the exponential gravity. In case of the logarithmic divergences
of the form (\ref{DeltaS}), the equation for the poles of the
propagator has the form (\ref{exp-brok}) with $z=p^2$ and
$A=4a_1+a_2$ for the spin-2 sector of the propagator. As we already
know, if such corrections take place, then the dressed propagator
has infinitely many ghost-like states. So, if we intend to keep the
ghost-free structure at the quantum level, the first thing to do
is to require that the theory should be finite. We shall discuss
this subject further in the next section.
\section{Quantum corrections and dressed propagator}
As we already know, the power counting in the theory with
exponential form factors (\ref{exp}) is exactly the same as in the
polynomial theory (\ref{superre}) with $k \geq 3$. One can see this
similarity in the following way. Imagine we replace the exponential
function in (\ref{exp}) by the partial sum of its Taylor expansion,
\beq
P_N(\alpha p^2) &=& \sum\limits_{l=0}^{N}
\frac{\big(\alpha p^2\big)^l}{l!}\,.
\label{series}
\eeq
For a sufficiently large $N$ the theory will be superrenormalizable,
exactly as in the exponential case. At the same time, there will be
$N$ roots of the polynomial $P_N(\alpha p^2)$ at the complex plane.
The number of these roots is growing with larger $N$. When
$N\to\infty$ the power counting remains the same. At the same time,
the number of poles becomes infinite, so the theory gains an infinite
amount of ghost-like poles, most of them complex. However, it happens
that all these poles converge to the very special infinite point
(\ref{inf}). Then the theory is ghost-free at the tree-level, but
there is a danger that the absolute tuning may be broken by the
one-loop corrections, which require very special attention.
The one-loop effective action in the theory
(\ref{act Phi}) is given by the expression \cite{frts-82}
(see Appendix for details)
\beq
\Gamma_{div}^{(1)} &=&
\frac{i}{2}\,\Tr \Log {\hat H}
\,-\,i\Tr \Log {\hat H}_{ghost}
\,+\,\frac{i}{2}\,\Tr \Log {\hat Y}\,.
\label{one}
\eeq
The last two terms are contributions of ghost and weight operators.
Both of them have standard form plus some part related to the term
$\Tr \Box$.
This term is discussed in the Appendix, where we argue that its
contribution has the form (\ref{DeltaS}) with fixed coefficients.
Therefore, the main question is whether the finiteness can be
provided by changing the action $S$ in such a way that the operator
\beq
{\hat H} &=&
\frac{1}{2\,\sqrt{-g}}\,\frac{\delta^2 S}{\delta g_{\alpha\beta}\delta g_{\rho\sigma}}
\label{hess}
\eeq
provides a cancellation of the first term in the formula (\ref{one})
with the divergences coming from the last two terms of the same
expression.
It is obvious that one can not achieve this goal by using the
original action (\ref{act Phi}), because both functions
$\Psi$ and $\Phi$ are proportional to the same expression
\beq
\Psi = c_1\,e^{-\alpha\Box}\,,\qquad
\Phi = c_2\,e^{-\alpha\Box}\,.
\label{PsiPhi}
\eeq
It is easy to see that by changing the coefficients $c_1$ and
$c_2$ one modify only the cosmological constant - type counterterm,
and not the fourth-derivative ones, which are relevant for Eq.
(\ref{exp-brok}). Therefore, in order to provide finiteness one
has to generalize the action (\ref{act Phi}). As it was discussed
for the polynomial QG, this can be done by
adding $\,{\cal O}(R^3_{...})$- and $\,{\cal O}(R^4_{...})$-type
terms. The explicit calculation in this theory would be quite
difficult and also there is no real need to make it. Let us instead
present a general evaluation of the possible effect of the
${\cal O}(R^3_{...})$-type terms. The general form of the terms
with a minimal possible non-local insertion is
\beq
\frac{1}{M^2}\,\int \sqrt{-g}\,
R_{\dots}
R_{\dots}
\,e^{-\alpha\Box}\,
R_{\dots}\,,
\label{R3}
\eeq
where $M$ is a new massive parameter.
Then the operator (\ref{hess}) will have a general non-minimal
structure (after an appropriate gauge-fixing)
\beq
{\hat H} &\propto&
{\hat \Box} \,+\, M^{-2}\,\hat{\cal D}^{\mu\nu} \na_\mu\na_\nu
\,+\, {\hat \Pi}\,,
\label{hess-nm}
\eeq
where all the operators act in the space of quantum metrics, and
$\hat{\cal D}^{\mu\nu}$ and ${\hat \Pi}$ are proportional to the
curvature tensor. The contribution of $\Tr\Log {\hat H}$ for the
operators of the form (\ref{hess-nm}) is known, in particular it
was elaborated recently in \cite{CPTL-Tib} by means of the
generalized Schwinger-DeWitt technique \cite{bavi-85}. One meets
the one-loop divergences which are given by an infinite series in
curvatures $\hat{\cal D}^{\mu\nu}$ and ${\hat \Pi}$, and the
(super)renormalizability of the theory is completely broken, so
we have to look for some generalization.
Another
possibility is to modify the expression (\ref{R3}) by introducing
further non-localities. The possible solution is to consider
the non-local terms of the general form
\beq
\int \sqrt{-g}\,
R_{\dots}
\frac{1}{\Box}
R_{\dots}
\,e^{-\alpha\Box}\,
R_{\dots}\,.
\label{R3-N}
\eeq
The one-loop divergences in the theories of similar type were already
considered in the literature \cite{Ishinose}.
Let us note that the expression (\ref{R3-N}) still leaves us a lot of
freedom in the choice of the action, because of the numerous possible
tensor structures and corresponding coefficients. Since the number
of the possible tensor structures in the operator ${\hat H}$ is
restricted, there is a good chance to meet such a combination of
terms in (\ref{R3-N}) which would lead to the operator
\beq
{\hat H} &\propto&
{\hat \Box}^2 \,+\, \hat{V}^{\mu\nu} \na_\mu\na_\nu
\,+\, {\hat U}\,,
\label{hess-N}
\eeq
plus some contribution of the operator ${\hat \Box}^{-1}$ which can
be factorized out in a standard way (see, e.g., Chapter 9 of
\cite{book}). In the expression (\ref{hess-N}) one still has the
freedom to choose the operator $\hat{V}^{\mu\nu}$, which is
proportional to the curvature tensor. As a result, it is possible
to manipulate the divergent part of effective action (\ref{DeltaS})
and to provide the desirable pre-fixed values for the coefficients
$a_{1,2\dots 6}$. In particular, there is a chance to obtain a
finite QG in this way.
The situation may be even more simple if we include the
${\cal O}(R^4_{...})$-type terms with an additional $\,\Box^{-2}$
insertion. In this case the relevant operator will have the form
\beq
{\hat H} &\propto&
{\hat \Box}^3 \,+\, \hat{V}^{\mu\nu\alpha\beta} \na_\mu\na_\nu \na_\alpha\na_\beta
\,+\, \hat{U}^{\mu\nu} \na_\mu\na_\nu
\,+\, {\hat W}\,,
\label{hess-N4}
\eeq
similar to the one we dealt with in \cite{highderi}. The operator
$\hat{U}^{\mu\nu}$ will be linearly proportional to the
coefficients of the ${\cal O}(R^4_{...})$-type terms. On the
other hand, linear dependence will also take place between
$\hat{U}^{\mu\nu}$ and the fourth-derivative terms in (\ref{DeltaS}).
Therefore, there are pretty good chances to provide finiteness in the
exponential QG theory by means of a special choice of the coefficients
of the ${\cal O}(R^4_{...})$-type terms with an appropriate non-local
insertion.
Indeed, the possibility to have a finite theory in the non-local
case is not so certain as in the polynomial QG (\ref{superre}).
In case of the non-finite theory the ghost-free structure will
be certainly violated. So, let us be generous to the exponential
QG and simply assume that the non-local theory can be made finite
in the way we described above. As we shall see right now, this
is still not sufficient to prevent the theory from the ghost-like
states. The consistent theory of QG should include quantization
of matter fields, not only the metric. The matters fields of the
spin-0, spin-$1/2$ and spin-$1$ contribute to the divergences in
the form of Eq. (\ref{DeltaS}) \cite{birdav,book}. For the
illustration purpose, let us reproduce the complete form factors
of the one-loop quantum corrections to the $\Phi$-function in Eq.
(\ref{act Phi}), derived in \cite{apco,fervi} for massive scalar
and fermion fields,
\beq
{\bar \Gamma}^{(1)}_{scal}
&=&
\frac{1}{32\pi^2}\,\int d^4x \sqrt{-g}\,
C_{\rho\sigma\alpha\beta} \Big[\frac{1}{60\,(4-n)}
+ \frac{1}{120}\,\ln \Big(\frac{4\pi \mu^2}{m^2}\Big)
+ \frac12\,k^s_W(a)
\Big] C^{\rho\sigma\alpha\beta}
\label{final}
\eeq
and
\beq
{\bar \Gamma}^{(1)}_{ferm}
&=&
\frac{1}{32\pi^2}\,\int d^4x \sqrt{-g}\,
C_{\rho\sigma\alpha\beta} \,\Big[\,\frac{1}{10\,(4-n)}
+ \frac{1}{20}\ln \Big(\frac{4\pi \mu^2}{m_f^2}\Big)+\frac12\,k^f_W(a)
\Big] C^{\rho\sigma\alpha\beta}\,,
\label{final-f}
\eeq
where
\beq
k_W(a)
&=&
\frac{240A + 20a^2 + 3a^4}{450\,a^4}\,,
\label{ff-scal}
\\
k^f_W(a)
&=&
\frac{300Aa^2-480\,A
\,-\, 40 a^2 \,+\, 19\,a^4}{225\,a^4}
\label{ff-ferm}
\eeq
and we used notations
\beq
A &=&
1-\frac{1}{a}\ln \frac{1+a/2}{1-a/2}
\quad {\rm and} \quad
a^2 \,=\,
\frac{4\Box^2}{\Box^2-4m^2}\,.
\label{A}
\eeq
The contributions to the $\Psi$-function are qualitatively similar
\cite{fervi}, but we do not reproduce them here for the sake of brevity.
The first observation is that the divergences in the Weyl-squared
sector have the same sign independent on whether we take scalar,
fermion, massless or massive vector. This well-known feature of
the divergences \cite{birdav} means that no cancelation of the
overall contribution to the $\Phi$-function due to supersymmetry
is possible. Therefore, if there is no cancelation with the
divergences coming from the QG sector, the expression (\ref{exp})
gains $\,p^4\log p^2\,$ contribution due to the matter fields loops
and this is certainly sufficient to have infinitely many ghost-like
excitations at the quantum level.
The second important point is that, even if the cancelation of
the divergences really takes place, it is not sufficient to
preserve the ghost-free structure of the theory even at the
one-loop level. The reason is that
both expressions (\ref{ff-scal}) and (\ref{ff-ferm}) have
an infinite sets of sub-logarithmic contributions, and those
can not be canceled by the QG part. The situation may be, in
principle, different in a strictly massless theory of matter,
when the cancelation in the leading-log part may be sufficient.
Furthermore, even if one can adjust the QG contribution to
make the theory completely free of divergences in the
presence of matter fields, this would be true only at the
one-loop level. Let us remember that the exponential QG has
only one-loop divergences, but for the matter loops this is
not so. Starting from the second loop, matter fields produce
the form factors with higher powers of $\log \big(\Box/\mu^2\big)$
in the UV. Then the compensation seems to be completely
impossible. Let us note that the $\beta$-functions in the
matter sector are not affected by QG in all superrenormalizable
models, as it was explained in Sect. 5 of \cite{highderi}.
One can naturally expect that the same breaking will take place in
the case of pure QG without matter. Indeed, any quantum theory,
including (\ref{act Phi}) may produce sub-logarithmic
contributions in the dressed propagator, and then an absolute
fine-tuning leading to the ghost-free structure will be violated.
Unfortunately, the explicit results concerning sub-logarithmic
contributions in QG are not available, but there can not be much
doubts about their existence. And the last is sufficient for
violating an absolute tuning leading to the ghost-free structure
of exponential gravity (\ref{exp}). So we have to conclude that
the exponential QG has an infinite set of massive ghost-like
states at the quantum level.
\section{Note concerning Newtonian singularity}
There is a very interesting and simple relation between the
renormalizability of the QG theory and the absence of Newtonian
singularity at the classical level. This relation was first
noted by Stelle in \cite{Stelle-77}, the main formula for the
modified Newtonian potential is
\beq
\varphi (r) = - GM \left[\frac{1}{r}
- \frac{4}{3} \frac{e^{-m_{(2)} r}}{r}
+ \frac{1}{3} \frac{e^{-m_{(0)} r}}{r} \right]\,.
\label{eq9}
\eeq
Here $m_{(2)}$ and $m_{(0)}$ are masses of the spin-2 ghost
and spin-0 massive particle which are present in the spectrum
of the fourth-derivative gravity. It is easy to see that near
the origin $r=0$ the contribution of these two massive degrees
of freedom exactly cancels the one of the graviton, such that
the limit of the modified Newtonian potential $\varphi(r)$ at
$r \to 0$ is free of singularity. In our recent work
\cite{NewSing} it was shown that the same cancelation takes
place in the more general theory of gravity with the action
(\ref{superre}), if additional degrees of freedom in this
theory correspond to the non-degenerate real massive
poles\footnote{The spin-0 contribution was elaborated much
earlier in \cite{HJS}}. For a while, there is no proof that
the same effect takes place in the case of complex massive
poles.
It is remarkable that the non-singular modified Newtonian limit
takes place also in the exponential gravity theory (\ref{exp}).
This was originally found by Tseytlin in \cite{Tseytlin-95}.
Recently, exactly the same non-singular solution has been
rediscovered in the papers \cite{Mode-fin,BGKM,BiMa}. An unfortunate
detail apparently related to the Ref. \cite{BGKM} is the evaluation
of our preprint \cite{NewSing} which was done by Biswas et al in
\cite{BiMa}. On page 3 of their manuscript authors say ``{\it It
was not until recently though, that concrete criteria for any
covariant gravitational theory (including infinite-derivative
theories) to be free from ghosts and tachyons around the Minkowski
vacuum was obtained by Biswas,
Gerwick, Koivisto and Mazumdar (BGKM) [23,24], see also [25]
for a recent re-derivation of the same results using auxiliary
field methods.}'' \footnote{Here {\it [23,24]} correspond to
\cite{BGKM} and \cite{BGKMsemG} and {\it [25]} to the citation
\cite{NewSing} of the present work. Let me stress that our
paper \cite{NewSing} is devoted to the Newtonian singularity
in the polynomial theory (\ref{superre}) and hence we did not
repeat the result of \cite{Tseytlin-95} and consequent works
such as \cite{Siegel-03} and \cite{BGKM}, which calculated the
modified Newtonian limit in the exponential theory case.}.
I believe that the evaluation of our work which was done in
\cite{BiMa} is not correct for at least two reasons. First, in
our Ref. \cite{NewSing} there is no use of auxiliary fields.
Second, we did not explore or discuss ``concrete criteria
for any covariant gravitational theory (including
infinite-derivative theories) to be free from ghosts and
tachyons around the Minkowski vacuum'' and moreover we even
did not deal with the ``infinite-derivative theories'' in
\cite{NewSing}. As it was already mentioned above, our work
\cite{NewSing} is about Newtonian singularity in the polynomial
theory (\ref{superre}), so it is not easy to understand what
the observation of \cite{BiMa} actually means.
Coming back to the relation between (super)renormalizability of
QG and the absence of Newtonian singularity, perhaps the most
intriguing aspect is that the theory without real poles can be
free of singularities. One can note that the polynomial theory
with a form factor given by the partial sums of the exponential
function (\ref{series}) has no real poles in the propagator.
From the other side, the limit at $N\to \infty$ is free of
Newtonian singularity, according to \cite{Tseytlin-95} and to
the consequent publications on the subject \cite{Siegel-03,BGKM}.
Let us remember that the polynomials with growing $N$ have
growing amount of massive poles. One most natural physical
interpretation is as follows. These poles are organized in such
a way that they lead to the cancelation of singularity in a way
similar to (\ref{eq9}) and to the more complicated relations
discussed in \cite{NewSing}. Then the cancelation of singularity
in the exponential case of \cite{Tseytlin-95} is nothing
else but the same effect coming from an infinite amount of
the ``hidden'' ghosts. This conjecture is something
interesting to verify, in our opinion it would give better
understanding of the relation between local and non-local
models of QG.
\section{Conclusions and discussions}
The main conflict of QG is between renormalizability and unitarity.
In order to have renormalizable or superrenormalizable QG, one has
to include higher derivatives into the starting action. Higher
derivatives lead to ghosts and/or tachyons and excluding these
unphysical states from the spectrum produce violation of unitarity.
Since higher derivatives emerge already at the semiclassical level,
apparently there is no way to avoid them, so the question is
how to deal with the massive ghosts.
Some qualitatively new approach to the problem of QG was suggested
in \cite{Tomboulis-97} and recently elaborated further in
\cite{Mode-fin,LMT} and \cite{BiMa}. This new approach assumes
that the starting theory is chosen in such a way that the higher
derivative theory is free of ghosts from the very beginning.
Such a choice implies that the theory must be non-local, in
a way explored earlier by Efimov et al \cite{Efimov}.
We have shown that in the case of exponential QG the usual power
counting evaluation must be modified and the main role is played
by the topological relation between the number of vertices and
internal lines. After all, the renormalization properties of the
non-local theory of \cite{Tomboulis-97} are very similar to the
ones of the superrenormalizable QG, introduced earlier in
\cite{highderi}. In particular, this means that the $\beta$-functions
of the matter fields are not affected by QG and that the
$\beta$-functions in the gravitational sector can vanish,
rending the theory finite.
While the classical theory of exponential gravity is ghost-free,
the quantum corrections may easily lead to the dressed propagator
which has infinitely many complex ghost-like poles. We have shown
that this scenario is unavoidable if the theory is not finite.
The finiteness in such a theory is possible by tuning the
${\cal O}(R^3_...)$ and ${\cal O}(R^4_...)$-type non-local terms
in the action. This may guarantee the absence of the strongest
logarithmic corrections and the result can be extended even to
the one-loop theory with quantum matter included. However, at
higher loops this tuning breaks down. On the other hand, even
the one-loop contributions of massive matter fields have
well-established sub-logarithmic contributions, which can
not be cancelled in the exponential QG model. The final conclusion
is that an infinite amount of unphysical complex poles emerge in
the theory at the quantum level.
In our opinion, an improved understanding of the role of ghosts is
one of the most relevant issues for QG. In particular, the main
lesson which one should learn from the comparison of polynomial
and exponential models of QG is the importance of the models with
complex poles, which were not covered by the consideration of
\cite{highderi}. It would be
interesting to treat this case in details in both quantum field
theory framework and in the more phenomenological way proposed
recently in \cite{HD-Stab}.
\section*{Acknowledgements}
Author is grateful to CNPq, FAPEMIG and ICTP for partial support
of his work and to the D\'epartement de Physique Th\'eorique and
Center for Astroparticle Physics at the Universit\'e de Gen\`eve
for partial support and kind hospitality during his sabbatical
period. I am also grateful to Theoretisch-Physikalisches-Institut
of the Friedrich-Schiller-Universit${\ddot {\rm a}}$t in Jena for the
warm hospitality during the time when the first version of this
paper was completed. The observations of anonymous referee were
very useful in improving presentation and I would like to
express gratitude to his/her contribution.
\section*{Appendix. Brief review of Lagrangian quantization}
Let us consider the Faddeev-Popov procedure for the theory
of QG based on the action (\ref{act Phi}) with possible
additional terms of the ${\cal O}(R^3_...)$- and
${\cal O}(R^4_...)$-type. Higher order terms have no much
importance, because they do not affect the divergences, in
case of the ``right'' distribution of non-local exponential
factors. Our treatment will cover both polynomial and
exponential choices of the form factors $\Phi=\Phi(\Box)$ and
$\Psi=\Psi(\Box)$ in Eq. (\ref{act Phi}). As we have noted
in the main text, in the exponential case the theory can be
superrenormalizable only if the functions $\Phi$ and $\Psi$
have the same exponential factor, let's call it $\exp(-\alpha\Box)$.
In the polynomial QG of Ref. \cite{highderi} the requirement
is less rigid, namely $\Phi$ and $\Psi$ should be polynomials
of the same order. For the sake of simplicity, let us assume
that $\Phi=\Psi$. Our consideration will be partially
repeating the one of \cite{Stelle-77,book} and \cite{highderi}
and we include it mainly to provide consistent presentation and
to discuss special features of the non-local case.
We assume that the quantum metric $h_{\mu\nu}$ is defined as
$\,h_{\mu\nu}=g_{\mu\nu}-\eta_{\mu\nu}$.
Let us introduce the gauge fixing condition in the form
\beq
S_{\rm gf}
&=& \int d^4x \,\, \chi_\mu \,Y^{\mu\nu} \, \chi_\nu\,,
\label{11}
\eeq
with the following form of the gauge-fixing and weight
functions:
\beq
\chi_\mu &=& \pa_\lambda h_\mu^\lambda - \beta \,\pa_\mu h
\nonumber
\\
Y^{\mu\nu}
&=&
- \frac{1}{\tau} \,\Omega(\Box)\,\left( g^{\mu\nu} \,\Box +
\gamma \,\na^\mu\, \na^\nu \right) \,.
\label{12}
\eeq
Here $h=h_{\mu\nu} \eta_{\mu\nu}$ and $\,\beta, \,\tau,\,\gamma\,$ are
gauge-fixing parameters. Furthermore, $\Omega(\Box)$ is a function,
which can be chosen by the convenience criteria.
Our first purpose is to have a non-degenerate bilinear form of
the action, therefore it is useful to choose
$\Omega(\Box)=\Phi(\Box)=\Psi(\Box)$. As it was discussed in the main
text, the divergences do not depend on the choice of the gauge
fixing. For this reason we will not discuss
the most general form of the weight function, which may be
dependent on the curvature tensor. Also, the gauge-fixing
parameters can be chosen to make the consideration simpler.
The most general bilinear form of the action on the flat
background is
\beq
S^{(2)} &=&
\frac12\,\int d^4x\,\,h^{\kappa\omega}
\Big\{
k_1\delta_{\kappa\omega , \rho\sigma}\,\Box^2
+ k_2g_{\kappa\omega} g_{\rho\sigma}\,\Box^2
+ k_3 \big(g_{\kappa\omega} \pa_\rho\pa_\sigma
+ g_{\rho\sigma} \pa_\kappa\pa_\omega\big)
\nonumber
\\
&+& k_4 \big(g_{\kappa\rho} \pa_\omega\na_\sigma
+ g_{\kappa\sigma} \pa_\omega\na_\rho
+ g_{\omega\rho} \pa_\rho\pa_\sigma
+ g_{\omega\sigma} \pa_\kappa\pa_\rho\big)
+ k_5 \pa_\omega\pa_\kappa\pa_\rho\pa_\sigma
\Big\}h^{\rho\sigma}\,,
\label{biline}
\eeq
Here $\,k_i,\,\,i=1,\dots,5\,$ are some functions of $\Box$, which
depend on the choice of the theory. In case of the polynomial theory
(\ref{superre}) they are polynomials, while in the case of
the exponential theory they are all proportional to
$\,\Box\,\exp\big(-\alpha\Box \big)$. The explicit form of these
functions can be found in \cite{BGKM} and \cite{Mode-fin}, but we
do not need them here.
Since the gauge-fixing parameters $\,\beta$, $\,\tau$
and $\,\gamma\,$ do not affect divergences, one can chose them
in such a way that the bilinear form of the overall action
$\,S^{(2)} + S_{\rm gf}\,$ becomes minimal. This means that the
tensor structures proportional to $\,k_3$, $\,k_4\,$ and $\,k_5\,$
cancel. Then, since the remaining $k_1$ and $k_2$ will be
proportional to $\Phi(\Box)$, the propagator of the
quantum metric has $r_l = 4+2N$ for the $\Phi(\Box)$ being
polynomial of order $N$, and $r_l={\cal I}$ for the exponential.
In order to apply the power counting relations (\ref{dD}) and
(\ref{topo}), one has to provide that the Faddeev-Popov ghosts
have the same power of momenta $r_l$ in the bilinear form of
their action. This can be achieved by using the method suggested
by Fradkin and Tseytlin \cite{frts-82}. The presence of extra
derivatives in the form factors of the initial action does not
affect the scheme \cite{highderi}, in both polynomial and
exponential models. Let us introduce the modified form of the
ghost action,
\beq
S_{gh} &=&
\int d^4x\sqrt{-g}\,
{\bar C}_\alpha\,Y^\alpha_{\,\,\beta}\,\,M^\beta_{\,\,\gamma}\,C^\gamma\,,
\quad
\mbox{where}
\quad
M^\beta_{\,\,\gamma} \,=\, \frac{\delta \chi^\beta}{\delta h_{\rho\sigma}}
\,R_{\rho\sigma.\gamma}\,,
\label{FP}
\eeq
where $R_{\rho\sigma.\gamma}$ is the generator of gauge transformations
of $\,h_{\mu\nu}\,$ in the background field method. An extra
insertion of the weight function in the definition of the ghost
action in (\ref{FP}) provides that for the quantum metric and
for the Faddeev-Popov ghosts there will be the same $r_l$ in the
formula (\ref{dD}).
The effective action $\Gamma$ is defined as \cite{frts-82}
\beq
e^{i\Gamma[g_{\mu\nu}]}
&=&
\big( \Det Y_\alpha^\beta \big)^{-1/2} \int dh_{\mu\nu} d{\bar C}_\alpha
dC^\beta\, e^{i S + iS_{gf} + iS_{gh}}\,.
\label{Z}
\eeq
The remaining problem is how to evaluate the
functional determinant $\,\Det Y_\alpha^\beta\,$ in the last expression.
For the polynomial QG theory (\ref{superre}), this operator is of
the standard sort,
\beq
Y_\alpha^\beta &=&
\Box^{2k}\,\big(\Box^2\,\delta^\alpha_\beta
+ \lambda \na^\alpha\na_\beta\big)
\,+\,{\cal O}(R_{\dots})\times \na^{2k}_{\dots}\,+\dots\,
\,,
\label{op}
\eeq
considered in \cite{highderi}. This type of operator can be
elaborated by the generalized Schwinger-DeWitt technique of
Barvinsky and Vilkovisky \cite{bavi-85}. The divergent
contribution of this expression is ${\cal O}(R^2_{\dots})$,
confirming the power counting - based analysis.
In the exponential case the contribution of the weight operator
in (\ref{op}) is more complicated. It is easy to see
that $\,\Det Y_\alpha^\beta\,$ is factorized into the product of
determinants of the two operators. One of these operators is
trivial and for the second one has to evaluate
\beq
\log \Det \Omega(\Box) &=&
\log \Det \exp \big(-\alpha\Box\big)
\,=\,
\Tr \log \exp \big(-\alpha\Box\big)
\,=\,
\Tr \big(-\alpha\Box\big)\,.
\label{expo}
\eeq
The last expression has quadratic divergences, but it does not mean
there are no logarithmic ones too, as usual. The evaluation of
it can be performed by local momentum representation or by the
Schwinger-DeWitt technique. Unfortunately, the operator
(\ref{expo}) seems to be unappropriate for the
technique of \cite{bavi-85}. However, there are strong reasons
to suppose that the result will be qualitatively the same as for
the polynomial case. In order to see this, one has to regard the
$\,\exp \big(-\alpha\Box\big)\,$ as a limit of the expression
(\ref{series}). In order to complete
the story, one has to note that a qualitatively different result
for the divergent part of (\ref{expo}) would mean one more
addition to Eq. (\ref{exp-brok}). This would further enforce the
main conclusion of the present paper, concerning the presence
of the infinite set of ghosts in the dressed propagator of the
exponential theory.
|
\section{INTRODUCTION}
In the study of the excitation function of superheavy ions their energy losses have to be known along with their masses, which can be taken from one of the mass tables \cite{NPA96Myers, AD95Moller, NPA03Audi}. No energy loss data are available to heavy ions with $\it{Z}$$>$103. The electronic stopping-power and range tables for ions with 1$\leq$$\it{Z}$$\leq$103 at 38 energies ranging from 0.0125 to 12 MeV/u in 24 media have been reported by Northcliffe and Schilling \cite{NDT70Northcliffe}. Hubert $\it{et}$ $\it{al.}$ \cite{AD90Hubert} reported electronic stopping-power and range tables for 2.5-500 MeV/u ions with 2$\leq$$\it{Z}$$\leq$103 in 36 solid materials. However, nuclear stopping powers were not listed in both tables. Electronic and nuclear stopping powers of ions with $\it{Z}$$\leq$92 in various media can be obtained from SRIM \cite{SRIM}, but stopping powers of charged particles with $\it{Z}$$>$92 are not available. There are scarce experimental data and expected values of stopping powers for heavy ions with $\it{Z}$$>$103, but the stopping power of $^{289}$Fl ($\it{Z}$=114) in Mylar was reported \cite{NIM10Wittwer}.
In the previous work \cite{IJMPE10Chung} electronic and nuclear stopping powers of a superheavy ion with $\it{Z}$=120 and $\it{A}$=300 at 26 energies ranging from 0.0177 to 0.141 MeV/u in UF$_4$, He, Mylar and butane have been estimated. Since hot fusion could be used in producing elements with $\it{Z}$$\geq$120, the energy range has been expanded to 0.20 MeV/u to estimate stopping powers of a heavy ion with $\it{Z}$=120 and $\it{A}$=300 in Al and UO$_2$ \cite{CJP14Chung}. In this work stopping powers of superheavy ions with $\it{Z}$$>$120 in Al and those with $\it{Z}$$\geq$120 in U are estimated at 35 energies ranging from 0.01 to 0.20 MeV/u. The estimation has been evaluated for light ions.
\section{Energy loss of superheavy ions in media }
Energy loss of superheavy ions in media arises from their electronic and nuclear stopping powers. The electronic stopping powers were deduced by fitting data from SRIM with a function with six parameters \cite{IJMPE10Chung} or by fitting them with a following function \cite{CJP14Chung}:
\begin{equation}
(-dE/dx)_e/Z^2 = a_1 +a_2 $exp$[-a_3/Z^{2/3}]
\label{eq:one}
\end{equation}
The nuclear stopping powers of the superheavy ion were deduced with a fifth-order polynomial fit with respect to $\it{Z}$ \cite{IJMPE10Chung,CJP14Chung}. Electronic stopping power of $^{289}$Fl in Mylar at 24.7$\pm$4.0 MeV deduced by Eq. \eqref{eq:one} was 30.2$\pm$2.7 MeV/(mg/cm$^2$) with its corresponding nuclear stopping powers of 8.3$\pm$1.2 MeV/(mg/cm$^2$) \cite{CJP14Chung}. The resulting total stopping power of 38.5$\pm$3.0 MeV/(mg/cm$^2$) agreed quite well with the experimental one of 34.9$\pm$15.4 MeV/(mg/cm$^2$), implying that this estimation is applicable to superheavy ions with $\it{Z}$=120 or even higher.
\subsection{Stopping powers of superheavy ions with $\it{Z}$$>$120 and $\it{A}$=300 in Al}
Electronic and nuclear stopping powers of 77 ions ranging from C to U in Al at 35 energies have been obtained from SRIM \cite{SRIM}. Their atomic numbers are 6, 10, 14, 18, 20, 21, 22,..., 91, and 92. Their 35 energies range from 0.01 to 0.20 MeV/u. The electronic and nuclear stopping powers at each energy have been used to extrapolate the corresponding stopping powers of heavy ions with $\it{Z}$=120 and $\it{A}$=300 with Eq. \eqref{eq:one} and the fifth-order polynomial fit, respectively \cite{CJP14Chung} and to further extrapolate those of heavy ions with $\it{Z}$=122 and 124 and $\it{A}$=300. The fits and the resulting variations at 0.015, 0.10 and 0.20 MeV/u are shown in Figs. \ref{fig1} and \ref{fig2}, respectively. They show that the fits work well with most of the ions except five ions with variations larger than 35 \% at 0.015 MeV/u and three ions with variations larger than 15 \% at 0.10 MeV/u. The ions at 0.015 MeV/u are Ge with 35 \%, Cs with 35 \%, Ce with 36 \%, Pr with 37 \% and Sm with 46 \% variation, while those at 0.10 MeV/u are Ni with 48 \%, Pm with 19 \% and Au with 25 \% variation \cite{CJP14Chung}.
\begin{figure}
\includegraphics[width=10cm]{fig1.png}
\caption{Electronic stopping powers of the ions in Al divided by $\it{Z^2}$ at 0.015, 0.10 and 0.20 MeV/u. Open symbols are obtained from SRIM \cite{SRIM} and the lines refer to their corresponding fits with Eq. \eqref{eq:one}.}
\label{fig1}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig2.png}
\caption{Variations of the fits in Fig. \ref{fig1}.}
\label{fig2}
\end{figure}
Electronic stopping powers of $^{16}$O and $^{19}$F ions in Al deduced by Eq. \eqref{eq:one} have been compared with the corresponding experimental data \cite{NIM00Xiting}. The results are shown in Table \ref{table:o16f19} and Figs. \ref{fig3} and \ref{fig4}. The estimated stopping powers of the $^{16}$O ion agree with the corresponding experimental values within less than 5 \% variation above 2 MeV, 6-14 \% between 0.81 and 1.87 MeV (0.05 and 0.12 MeV/u) and 17-19 \% below 0.8 MeV, while the SRIM data agree with them with 5-7 \% variation between 1.11 and 3.14 MeV (0.07 and 0.20 MeV/u) and less than 5 \% below 1 MeV. The estimated stopping powers of the $^{19}$F ion agree with the corresponding experimental values within 3-9 \% variation above 1.9 MeV, 13-25 \% between 0.58 and 1.75 MeV (0.03 and 0.09 MeV/u), and 31-47 \% below 0.5 MeV, while the SRIM data agree with them with 6-10 \% variation above 1 MeV, 15-33 \% between 0.58 and 0.88 MeV (0.03 and 0.05 MeV/u) and 33-52 \% below 0.5 MeV. The estimation works even better for the $^{16}$O and $^{19}$F ions above 0.12 MeV/u than the SRIM data.
\begin{figure}
\includegraphics[width=10cm]{fig3.png}
\caption{Electronic stopping powers of $^{16}$O in Al.}
\label{fig3}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig4.png}
\caption{Electronic stopping powers of $^{16}$F in Al.}
\label{fig4}
\end{figure}
\begin{table}[ht]
\caption{Electronic stopping powers of $^{16}$O and $^{19}$F ions in Al}
\centering
\begin{tabular}{l c c c c l c c c}
\hline\hline
$^{16}$O &&&&& $^{19}$F \\ [0.5ex]
\hline
E ~~~~~~~~~~& $\it{S_e}$ & $\it{S_e ^a}$ & $\it{S_e ^b}$ & ~~~~~~~& E~~~~~~~~~~ & $\it{S_e}$ & $\it{S_e ^a}$ & $\it{S_e ^b}$ \\
MeV & & MeV/(mg/cm$^2$) & && MeV & & MeV/(mg/cm$^2$) & \\ [0.5ex]
\hline
0.59 &3.52 &2.96 &3.09 &&0.41 &2.59 &1.76 &2.68 \\
0.70 &3.89 &3.34 &3.43 &&0.49 &2.90 &2.22 &2.96 \\
0.81 &4.21 &3.71 &3.74 &&0.58 &3.25 &2.60 &3.24 \\
0.92 &4.48 &3.97 &4.04 &&0.67 &3.57 &2.96 &3.50 \\
1.11 &4.84 &4.25 &4.49 &&0.78 &3.94 &3.31 &3.80 \\
1.29 &5.11 &4.59 &4.85 &&0.88 &4.25 &3.64 &4.05 \\
1.48 &5.33 &4.86 &5.18 &&1.04 &4.69 &4.06 &4.43 \\
1.67 &5.51 &5.17 &5.46 &&1.21 &5.07 &4.42 &4.81 \\
1.87 &5.67 &5.36 &5.71 &&1.39 &5.41 &4.72 &5.17 \\
2.09 &5.82 &5.58 &5.93 &&1.57 &5.70 &5.00 &5.49 \\
2.29 &5.94 &5.74 &6.10 &&1.75 &5.94 &5.27 &5.79 \\
2.50 &6.06 &5.95 &6.26 &&1.93 &6.14 &5.61 &6.06 \\
2.71 &6.16 &6.02 &6.38 &&2.11 &6.32 &5.82 &6.30 \\
2.92 &6.26 &6.20 &6.48 &&2.30 &6.48 &6.06 &6.53 \\
3.14 &6.36 &6.23 &6.57 &&2.50 &6.76 &6.17 &6.75 \\
& & && &2.71 &6.78 &6.31 &6.94 \\
& & &&&2.90 &6.90 &6.47 &7.10 \\
& & &&&3.23 &7.09 &6.71 &7.33 \\
& & &&&3.49 &7.22 &6.87 &7.48 \\
& & &&&3.76 &7.34 &7.15 &7.61 \\ [1ex]
\hline
\end{tabular}
\\{Values of $\it{S_e ^a}$ were taken from Ref. \cite{NIM00Xiting} and those of $\it{S_e ^b}$ obtained from SRIM \cite{SRIM}.}
\label{table:o16f19}
\end{table}
Nuclear stopping powers of the superheavy ions with $\it{Z}$=122 and 124 and $\it{A}$=300 have been extrapolated by a fifth-order polynomial fit \cite{IJMPE10Chung,CJP14Chung}. The fits and the resulting variations at 0.015, 0.10 and 0.20 MeV/u are shown in Figs. \ref{fig5} and \ref{fig6}, respectively. The fits work well with all the ions within 0.1 \% variation except four light ions, C with 1.2-2.1 \% , Ne with 0.4-0.8 \%, Si with 0.2-0.5 \% and Ar with 0.2 \% variation. The results are listed in Table \ref{table:SheAl} and the total stopping powers are shown in Fig. \ref{fig7}. The electronic and nuclear stopping powers for the $^{300}$122 ion at 0.20 MeV/u are 36.1 and 4.2 MeV/(mg/cm$^2$), respectively, while those at 0.015 MeV/u are 14.4 and 15.6 MeV/(mg/cm$^2$). The corresponding stopping powers of the $^{300}$124 ion at 0.20 MeV/u are 36.5 and 4.3 MeV/(mg/cm$^2$), while those at 0.015 MeV/u are 14.7 and 15.8 MeV/(mg/cm$^2$). The electronic stopping powers are smaller than their corresponding nuclear ones at $\it{E}$$\geq$0.017 MeV/u. Fig. \ref{fig7} shows that total stopping powers of the $^{300}$122 and $^{300}$124 ions tend to decrease slightly until 0.043 MeV/u, that is a little higher than 0.04 MeV/u in the $^{300}$120 ion \cite{CJP14Chung}, and then increase smoothly as the energy increases.
\begin{figure}
\includegraphics[width=10cm]{fig5.png}
\caption{Fifth-order polynomial fits for nuclear stopping powers of the ions in Al at 0.015, 0.10 and 0.20 MeV/u. Open symbols are obtained from SRIM \cite{SRIM} and the lines refer to their corresponding fits.}
\label{fig5}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig6.png}
\caption{Variations of the fits in Fig. \ref{fig5}.}
\label{fig6}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig7.png}
\caption{Total stopping powers of the superheavy ions with $\it{Z}$=122 and 124 and $\it{A}$=300 in Al at 35 energies ranging from 0.01 to 0.20 MeV/u.}
\label{fig7}
\end{figure}
\begin{table}[ht]
\caption{Stopping powers of superheavy ions with $\it{Z}$=120, 122, and 124 and $\it{A}$=300 in Al}
\centering
\begin{tabular}{l l l l l l l }
\hline\hline
&~~~$\it{Z}$=120\footnote{The data are taken from Ref. \cite{CJP14Chung}. } & &~~~$\it{Z}$=122&&~~~$\it{Z}$=124 \\ [-0.5ex]
\hline
E ~~~&~~~ $\it{S_e}$ & $\it{S_n}$ ~~~~~&~~~ $\it{S_e}$ & $\it{S_n}$~~~~~ & ~~~$\it{S_e}$ & $\it{S_n}$ \\ [-1.5ex]
\multicolumn{7}{l}{MeV/u~~~~~~~~~~~~~~~~~~~~~MeV/(mg/cm$^2$)} \\ [-0.5ex]
\hline
0.010 &~~~12.745 &17.002 &~~~13.045 &17.237 &~~~13.348 &17.457 \\ [-1ex]
0.011 &~~~12.987 &16.714 &~~~13.290 &16.955 &~~~13.597 &17.183 \\ [-1ex]
0.012 &~~~13.221 &16.436 &~~~13.527 &16.683 &~~~13.837 &16.918 \\ [-1ex]
0.013 &~~~13.448 &16.182 &~~~13.758 &16.437 &~~~14.070 &16.681 \\ [-1ex]
0.013 &~~~13.669 &15.872 &~~~13.981 &16.123 &~~~14.296 &16.362 \\ [-1ex]
0.015 &~~~14.093 &15.326 &~~~14.411 &15.575 &~~~14.733 &15.813 \\ [-1ex]
0.017 &~~~14.498 &14.858 &~~~14.822 &15.112 &~~~15.149 &15.357 \\ [-1ex]
0.018 &~~~14.886 &14.433 &~~~15.214 &14.692 &~~~15.546 &14.943 \\ [-1ex]
0.020 &~~~15.253 &14.019 &~~~15.586 &14.279 &~~~15.922 &14.533 \\ [-1ex]
0.022 &~~~15.584 &13.605 &~~~15.920 &13.861 &~~~16.259 &14.111 \\ [-1ex]
0.023 &~~~15.942 &13.225 &~~~16.283 &13.479 &~~~16.627 &13.727 \\ [-1ex]
0.027 &~~~16.564 &12.547 &~~~16.910 &12.797 &~~~17.260 &13.042 \\ [-1ex]
0.030 &~~~17.130 &11.940 &~~~17.481 &12.184 &~~~17.835 &12.424 \\ [-1ex]
0.033 &~~~17.647 &11.411 &~~~18.000 &11.651 &~~~18.356 &11.889 \\ [-1ex]
0.037 &~~~18.119 &10.914 &~~~18.474 &11.147 &~~~18.832 &11.377 \\ [-1ex]
0.040 &~~~18.552 &10.471 &~~~18.907 &10.698 &~~~19.265 &10.922 \\ [-1ex]
0.043 &~~~18.956 &10.070 &~~~19.310 &10.291 &~~~19.668 &10.510 \\ [-1ex]
0.047 &~~~19.331 &9.719 &~~~19.685 &9.938 &~~~20.041 &10.155 \\ [-1ex]
0.050 &~~~19.682 &9.365 &~~~20.033 &9.575 &~~~20.387 &9.784 \\ [-1ex]
0.053 &~~~20.015 &9.054 &~~~20.364 &9.260 &~~~20.716 &9.464 \\ [-1ex]
0.057 &~~~20.334 &8.771 &~~~20.681 &8.973 &~~~21.030 &9.174 \\ [-1ex]
0.060 &~~~20.640 &8.505 &~~~20.984 &8.702 &~~~21.331 &8.899 \\ [-1ex]
0.067 &~~~21.232 &8.026 &~~~21.571 &8.215 &~~~21.913 &8.404 \\ [-1ex]
0.075 &~~~21.954 &7.503 &~~~22.287 &7.681 &~~~22.622 &7.859 \\ [-1ex]
0.083 &~~~22.677 &7.072 &~~~23.006 &7.245 &~~~23.336 &7.418 \\ [-1ex]
0.092 &~~~23.420 &6.680 &~~~23.746 &6.844 &~~~24.074 &7.009 \\ [-1ex]
0.100 &~~~24.195 &6.337 &~~~24.520 &6.495 &~~~24.847 &6.652 \\ [-1ex]
0.108 &~~~25.002 &6.038 &~~~25.328 &6.190 &~~~25.656 &6.342 \\ [-1ex]
0.117 &~~~25.845 &5.762 &~~~26.175 &5.908 &~~~26.506 &6.054 \\ [-1ex]
0.125 &~~~26.723 &5.511 &~~~27.058 &5.651 &~~~27.395 &5.790 \\ [-1ex]
0.133 &~~~27.633 &5.294 &~~~27.975 &5.429 &~~~28.319 &5.565 \\ [-1ex]
0.150 &~~~29.537 &4.907 &~~~29.897 &5.034 &~~~30.260 &5.162 \\ [-1ex]
0.167 &~~~31.518 &4.580 &~~~31.903 &4.700 &~~~32.290 &4.821 \\ [-1ex]
0.183 &~~~33.561 &4.295 &~~~33.974 &4.409 &~~~34.390 &4.523 \\ [-1ex]
0.200 &~~~35.630 &4.045 &~~~36.075 &4.151 &~~~36.522 &4.259 \\
\hline
\end{tabular}
\label{table:SheAl}
\end{table}
\subsection{Stopping powers of superheavy ions with $\it{Z}$$\geq$120 and $\it{A}$=300 in U}
Electronic and nuclear stopping powers of superheavy ions with $\it{Z}$=120, 122 and 124 and $\it{A}$=300 in U have been estimated similarly as stated in the previous section. The fits for the electronic stopping powers and the resulting variations at 0.015, 0.10 and 0.20 MeV/u are shown in Figs. \ref{fig8} and \ref{fig9}, respectively. They show that the fits work well with most of the ions except six ions with variations larger than 30 \% at 0.015 MeV/u and three ions with variations larger than 14 \% at 0.10 MeV/u. The ions at 0.015 MeV/u are Ge with 35 \%, Cs with 35 \%, Ba with 31 \%, La with 36 \%, Pr with 37 \% and Sm with 46 \% variation, while those at 0.10 MeV/u are Ni with 48 \%, Pm with 19 \% and Au with 25 \% variation. Electronic stopping powers of the Lr ($\it{Z}$=103) ion estimated at 0.015, 0.10 and 0.20 MeV/u are 2.0, 5.9 and 10.9 MeV/(mg/cm$^2$), respectively, while their corresponding values in Northcliffe and Schilling$'$s tables are 1.3, 4.6 and 8.2 MeV/(mg/cm$^2$). Their discrepancies are within a 40 \% variation above 0.04 MeV/u and increase to more than 60 \% variation at the lower energy region.
\begin{figure}
\includegraphics[width=10cm]{fig8.png}
\caption{Electronic stopping powers of the ions in U divided by $\it{Z^2}$ at 0.015, 0.10 and 0.20 MeV/u. Open symbols are obtained from SRIM \cite{SRIM} and the lines refer to their corresponding fits with Eq. \eqref{eq:one}.}
\label{fig8}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig9.png}
\caption{Variations of the fits in Fig. \ref{fig8}.}
\label{fig9}
\end{figure}
The fits for the nuclear stopping powers and the resulting variations at 0.015, 0.10 and 0.20 MeV/u are shown in Figs. \ref{fig10} and \ref{fig11}, respectively. They work well with all the ions within 0.1 \% variation except four light ions, C with 1.7-2.4 \% , Ne with 0.5-1.0 \%, Si with 0.3-0.5 \% and Ar with 0.1-0.2 \% variation. The results are listed in Table \ref{table:SheU} and the total stopping powers with those of $^{238}$U are shown in Fig. \ref{fig12}. The electronic and nuclear stopping powers for the $^{300}$120 ion at 0.20 MeV/u are 12.1 and 2.2 MeV/(mg/cm$^2$), respectively, while those at 0.01 MeV/u are 2.0 and 7.4 MeV/(mg/cm$^2$). The corresponding stopping powers of the $^{300}$122 ion at 0.20 MeV/u are 12.3 and 2.3 MeV/(mg/cm$^2$), while those at 0.01 MeV/u are 2.1 and 7.5 MeV/(mg/cm$^2$). The corresponding stopping powers of the $^{300}$124 ion at 0.20 MeV/u are 12.4 and 2.3 MeV/(mg/cm$^2$), while those at 0.01 MeV/u are 2.1 and 7.5 MeV/(mg/cm$^2$). The total stopping powers decrease slightly until about 0.040 MeV/u for $\it{Z}$=120, 0.043 MeV/u for $\it{Z}$=122 and 0.046 MeV/u for $\it{Z}$=124 ion that are higher than 0.02 MeV/u for the $^{238}$U ($\it{Z}$=92) ion and increase smoothly as the energy increases as shown in Al, except a little oscillation below 0.0125 MeV/u.
\begin{figure}
\includegraphics[width=10cm]{fig10.png}
\caption{Fifth-order polynomial fits for nuclear stopping powers of the ions in Al at 0.015, 0.10 and 0.20 MeV/u. Open symbols are obtained from SRIM \cite{SRIM} and the lines refer to their corresponding fits.}
\label{fig10}
\end{figure}
\begin{figure}
\includegraphics[width=10cm]{fig11.png}
\caption{Variations of the fits in Fig. \ref{fig10}.}
\label{fig11}
\end{figure}
\begin{table}[ht]
\caption{Stopping powers of superheavy ions with $\it{Z}$=120, 122, and 124 and $\it{A}$=300 in U}
\centering
\begin{tabular}{l l l l l l l }
\hline\hline
&~~~$\it{Z}$=120& &~~~$\it{Z}$=122&&~~~$\it{Z}$=124 \\ [-0.5ex]
\hline
E ~~~&~~~ $\it{S_e}$ & $\it{S_n}$ ~~~~~&~~~ $\it{S_e}$ & $\it{S_n}$~~~~~ & ~~~$\it{S_e}$ & $\it{S_n}$ \\ [-1.5ex]
\multicolumn{7}{l}{MeV/u~~~~~~~~~~~~~~~~~~~~~MeV/(mg/cm$^2$)} \\ [-0.5ex]
\hline
0.010 &~~~2.041 &7.363 &~~~2.089 &7.512 &~~~2.138 &7.512 \\ [-1ex]
0.011 &~~~2.119 &7.284 &~~~2.168 &7.364 &~~~2.218 &7.437 \\ [-1ex]
0.012 &~~~2.193 &7.231 &~~~2.244 &7.318 &~~~2.295 &7.398 \\ [-1ex]
0.013 &~~~2.265 &7.158 &~~~2.317 &7.248 &~~~2.369 &7.331 \\ [-1ex]
0.013 &~~~2.333 &7.076 &~~~2.386 &7.168 &~~~2.440 &7.252 \\ [-1ex]
0.015 &~~~2.463 &6.926 &~~~2.519 &7.022 &~~~2.575 &7.111 \\ [-1ex]
0.017 &~~~2.586 &6.776 &~~~2.643 &6.874 &~~~2.702 &6.966 \\ [-1ex]
0.018 &~~~2.700 &6.632 &~~~2.760 &6.733 &~~~2.820 &6.829 \\ [-1ex]
0.020 &~~~2.810 &6.498 &~~~2.871 &6.602 &~~~2.933 &6.701 \\ [-1ex]
0.022 &~~~2.914 &6.359 &~~~2.977 &6.463 &~~~3.041 &6.563 \\ [-1ex]
0.023 &~~~3.014 &6.234 &~~~3.079 &6.340 &~~~3.144 &6.442 \\ [-1ex]
0.027 &~~~3.205 &5.991 &~~~3.272 &6.098 &~~~3.339 &6.202 \\ [-1ex]
0.030 &~~~3.384 &5.754 &~~~3.453 &5.859 &~~~3.523 &5.962 \\ [-1ex]
0.033 &~~~3.556 &5.543 &~~~3.627 &5.647 &~~~3.699 &5.748 \\ [-1ex]
0.037 &~~~3.721 &5.352 &~~~3.794 &5.455 &~~~3.867 &5.557 \\ [-1ex]
0.040 &~~~3.882 &5.168 &~~~3.956 &5.269 &~~~4.031 &5.368 \\ [-1ex]
0.043 &~~~4.032 &5.019 &~~~4.107 &5.122 &~~~4.183 &5.223 \\ [-1ex]
0.047 &~~~4.194 &4.858 &~~~4.270 &4.958 &~~~4.348 &5.057 \\ [-1ex]
0.050 &~~~4.346 &4.705 &~~~4.424 &4.802 &~~~4.502 &4.897 \\ [-1ex]
0.053 &~~~4.498 &4.575 &~~~4.576 &4.670 &~~~4.655 &4.765 \\ [-1ex]
0.057 &~~~4.648 &4.459 &~~~4.728 &4.555 &~~~4.808 &4.649 \\ [-1ex]
0.060 &~~~4.799 &4.343 &~~~4.879 &4.437 &~~~4.959 &4.530 \\ [-1ex]
0.067 &~~~5.100 &4.121 &~~~5.181 &4.211 &~~~5.263 &4.299 \\ [-1ex]
0.075 &~~~5.480 &3.896 &~~~5.563 &3.985 &~~~5.647 &4.072 \\ [-1ex]
0.083 &~~~5.868 &3.678 &~~~5.953 &3.761 &~~~6.039 &3.843 \\ [-1ex]
0.092 &~~~6.267 &3.502 &~~~6.354 &3.582 &~~~6.442 &3.662 \\ [-1ex]
0.100 &~~~6.677 &3.356 &~~~6.766 &3.436 &~~~6.857 &3.517 \\ [-1ex]
0.108 &~~~7.097 &3.202 &~~~7.190 &3.279 &~~~7.283 &3.356 \\ [-1ex]
0.117 &~~~7.529 &3.064 &~~~7.625 &3.137 &~~~7.722 &3.210 \\ [-1ex]
0.125 &~~~7.971 &2.945 &~~~8.071 &3.016 &~~~8.171 &3.087 \\ [-1ex]
0.133 &~~~8.421 &2.847 &~~~8.525 &2.918 &~~~8.630 &2.989 \\ [-1ex]
0.150 &~~~9.339 &2.648 &~~~9.453 &2.714 &~~~9.568 &2.780 \\ [-1ex]
0.167 &~~~10.273 &2.487 &~~~10.399 &2.550 &~~~10.525 &2.613 \\ [-1ex]
0.183 &~~~11.21 0 &2.343 &~~~11.348 &2.403 &~~~11.487 &2.462 \\ [-1ex]
0.200 &~~~12.140 &2.222 &~~~12.292 &2.279 &~~~12.444 &2.337 \\
\hline
\end{tabular}
\label{table:SheU}
\end{table}
\begin{figure}
\includegraphics[width=10cm]{fig12.png}
\caption{Total stopping powers of the superheavy ions with $\it{Z}$=120, 122 and 124 and $\it{A}$=300 in U at the same energies as in Fig. \ref{fig7}.}
\label{fig12}
\end{figure}
\section{Conclusions}
Electronic and nuclear stopping powers of superheavy ions with $\it{Z}$=122 and 124 and $\it{A}$=300 in Al and $\it{Z}$=120, 122 and 124 and $\it{A}$=300 in U at 35 energies ranging from 0.01 to 0.20 MeV/u have been deduced using the data obtained from SRIM. The total stopping powers of the $^{300}$122 and $^{300}$124 ions in Al tend to decrease slightly until 0.043 MeV/u, that is a little higher than 0.04 MeV/u in the $^{300}$120 ion, and then increase smoothly as the energy increases. The total stopping powers in U decrease slightly until about 0.040 MeV/u for $\it{Z}$=120, 0.043 MeV/u for $\it{Z}$=122 and 0.046 MeV/u for $\it{Z}$=124 ion that are higher than 0.02 MeV/u for the $^{238}$U ion and increase smoothly as the energy increases. Electronic stopping powers of the Lr ion in U estimated at 0.02, 0.10 and 0.20 MeV/u are compared with their corresponding values in Northcliffe and Schilling$'$s tables. Their discrepancies are within a 40\% variation above 0.04 MeV/u and tend to increase at the lower energy region.
The estimated electronic stopping powers of $^{16}$O and $^{19}$F ions agreed well with the experimental values above 0.10 MeV/u, while there were larger deviations near 0.01 MeV/u. This implies that this estimation would be applicable to superheavy ions as well as light ions above 0.10 MeV/u.
|
\section{Introduction}
\label{sec:Introduction}
The antisymmetric spin-orbit coupling, which is induced by inversion symmetry breaking of the crystal structure, has attracted interest because it leads to fascinating phenomena, such as the magnetoelectric effect~\cite{curie1894symetrie,dzyaloshinskii1960magneto,kimura2003magnetic,wang2003epitaxial,KhomskiiPhysics.2.20} and noncentrosymmetric superconductivity~\cite{Bauer_PhysRevLett.92.027003,Bauer_Sigrist201201,Frigeri_PhysRevLett.92.097001,Samokhin_PhysRevB.69.094514,Fujimoto_PhysRevB.72.024515,Fujimoto_doi:10.1143/JPSJ.75.083704}.
In such noncentrosymmetric systems, the Hamiltonian for the antisymmetric spin-orbit coupling is written by
\begin{align}
\label{eq:Ham_ASOC_def}
\mathcal{H}_{{\rm ASOC}} (\bm{k})
=\bm{g}(\bm{k}) \cdot \bm{\sigma} \sim (\bm{k} \times \nabla V_{{\rm pot}}) \cdot \bm{\sigma},
\end{align}
where $\bm{g}(\bm{k})$ is the polar vector with respect to the wave vector $\bm{k}$,
$\bm{\sigma}$ is the spin operator, and $\nabla V_{{\rm pot}}$ is a potential gradient due to the inversion symmetry breaking.
This antisymmetric spin-orbit coupling often provides a clue for understanding of peculiar electronic and transport properties, such as the spin splitting in the band structure~\cite{rashba1960properties,bychkov1984oscillatory,Dresselhaus_Dresselhaus_Jorio}.
In such spin-orbital coupled phenomena, not only uniform but also local (site) inversion symmetry breaking has recently drawn considerable attention~\cite{sakhnenko2012magnetoelectric,ter2014cation,Yanase:JPSJ.83.014703}.
The local inversion symmetry breaking means that the inversion symmetry is broken at every lattice site in a different manner from site to site.
For instance, a two-dimensional honeycomb lattice breaks the local inversion symmetry at every lattice site in an alternating way on the two sublattices, while it preserves the inversion symmetry at each center of the nearest-neighbor bonds and the hexagons.
In these systems, the parity mixing between different orbitals occurs locally in a sublattice-dependent form, which we call the local parity mixing.
In the presence of the local parity mixing, $\bm{g}(\bm{k})$ in Eq.~(\ref{eq:Ham_ASOC_def}), also depends on the sublattice.
This suggests that new spin-orbital coupled phenomena are expected by using the sublattice degree of freedom in crystals.
For example, once a sublattice-dependent magnetic or electronic order is induced in the system with local parity mixing, multipoles extending over different sublattice sites are simultaneously activated in odd-parity sectors, such as a magnetic quadrupole and an electric octupole~\cite{Spaldin:0953-8984-20-43-434203,kopaev2009toroidal}.
Such unusual odd-parity multipoles have been recently studied for unconventional superconductivity~\cite{Yoshida_doi:10.7566/JPSJ.82.074714,Yoshida_PhysRevB.86.134514,Maruyama_doi:10.1143/JPSJ.81.034702} and electronic orders~\cite{Yanase:JPSJ.83.014703,Hayami_PhysRevB.90.024432,Hayami_PhysRevB.90.081115,hayami2014quantum,hayami2014toroidal,hitomi2014electric}.
A typical case was theoretically discussed on a zig-zag chain, where the local inversion symmetry is broken and the sublattice-dependent antisymmetric spin-orbit coupling is present~\cite{Yanase:JPSJ.83.014703}.
In this case, an antiferromagnetic order can be regarded as an odd-parity multipole order, accompanying both magnetic toroidal and quadrupole components~\cite{Spaldin:0953-8984-20-43-434203,kopaev2009toroidal}.
Accordingly, the antiferromagnetic order acquires an unusual magnetoelectric coupling to an electric current~\cite{Yanase:JPSJ.83.014703,Hayami_PhysRevB.90.024432}.
Therefore, such odd-parity multipoles are intriguing for exploring new spin-orbital coupled phenomena.
However, the microscopic theory is not fully investigated.
In particular, it is still unclear when and how such multipole orders are stabilized.
It is highly desired to develop a general framework of the microscopic theory for the systems with local parity mixing.
Experimentally, the local inversion symmetry breaking is widely found in multi-orbital systems on particular lattice structures.
For instance, there are several $f$-electron materials possessing the lattice structures without local inversion symmetry, such as
ferromagnetic superconductors UGe$_2$~\cite{oikawa1996crystal,saxena2000superconductivity,Huxley_PhysRevB.63.144519}, URhGe~\cite{aoki2001coexistence,levy2005magnetic,Hardy_PhysRevLett.94.247006}, and UCoGe~\cite{huy2007superconductivity,Huy_PhysRevLett.100.077002,aoki2009extremely}, the zig-zag chain compounds $LnM$$_2$Al$_{10}$ ($Ln$$=$Ce, Nd, Gd, Dy, Ho, and Er; $M$$=$Fe, Ru, and Os)~\cite{thiede1998ternary,reehuis2003magnetic,KhalyavinPhysRevB.82.100405,TanidaPhysRevB.84.115128,tanida2010possible,kondo2010magnetization,muro2011magnetic,mignot2011neutron}, the distorted honeycomb compounds $\alpha$- and $\beta$-YbAlB$_4$~\cite{macaluso2007crystal,nakatsuji2008superconductivity,matsumoto2011quantum}, and the diamond-structure compounds $R$$T$$_2$$X$$_{20}$ ($R$$=$Pr, La, Yb, and U, $T$$=$Fe, Co, Ti, V, Nb, Ru, Rh, and Ir, and $X=$Al and Zn)~\cite{moze1998crystal,Bauer_PhysRevB.74.155118,torikachvili2007six,onimaru2010superconductivity,sakai2011kondo,higashinaka2011single,onimaru2012nonmagnetic,onimaru2012simultaneous,matsubayashi2012pressure,sakai2012superconductivity,Tsujimoto_PhysRevLett.113.267001,ikeura2014anomalous}.
Although these materials can be candidates for unusual odd-parity multipole ordering through the local parity mixing, their properties have not been studied from such a viewpoint.
In order to stimulate experiments and theories, it is important to clarify the microscopic mechanism for odd-parity multipole ordering.
In this paper, we investigate a microscopic model by taking into account the
hybridization between conduction and localized orbitals with different parity; e.g., between a $d$ component in conduction electrons and localized $f$ orbitals.
Starting from the periodic Anderson model with the antisymmetric hybridization defined on a quasi-one-dimensional lattice structure composed of zig-zag chains, we derive an effective low-energy model with antisymmetric exchange couplings between conduction electrons and localized spins.
This is an extension of the Kondo lattice model to the case with local parity mixing.
We show that the effective antisymmetric exchange couplings stabilize a
N{\'e}el-type antiferromagnetic order with the spin polarization perpendicular to the zig-zag plane.
This is an odd-parity multipole order, which gives rise to a peculiar band deformation and magnetoelectric effects~\cite{Yanase:JPSJ.83.014703,Hayami_PhysRevB.90.024432}.
Using several numerical calculations, we clarify that the multipole order is widely stabilized at and near half filling in the extended Kondo lattice model.
The organization of this paper is as follows.
In Sect.~\ref{sec:Model}, we describe the derivation of the Kondo lattice model including antisymmetric exchange couplings from an extended periodic Anderson model.
In Sect.~\ref{sec:results}, we clarify how the
antisymmetric exchange couplings favor a multipole order associated with an antiferromagnetic order.
We also compute the electronic band structure and magnetoelectric effect under this multipole ordering.
Moreover, we examine the stability of the multipole order systematically by the variational calculation for the ground state, simulated annealing, and Monte Carlo simulation at finite temperatures.
The last section is devoted to a summary of this paper.
The canonical transformation leading to the extended Kondo model is given in Appendix.
\section{Model}
\label{sec:Model}
\subsection{Periodic Anderson model with antisymmetric hybridization}
\label{subsec:PAM-ASOC}
In this paper, we consider the effect of the antisymmetric spin-orbit coupling in the presence of local parity mixing.
For this purpose, we introduce an extended periodic Anderson model with the antisymmetric hybridization between different parity orbitals~\cite{Yanase_doi:10.1143/JPSJ.77.124711}, whose Hamiltonian is given by
\begin{align}
\label{eq:Ham_PAMSO}
\mathcal{H} &= \mathcal{H}_0 +\mathcal{H}_1,
\end{align}
where
\begin{align}
\label{eq:Ham_PAMSO1}
\mathcal{H}_0
&= -\sum_{i,j,\sigma}
(t_{ij} c^{\dagger}_{i \sigma} c^{}_{j \sigma} + {\rm H.c.})
+ U \sum_i n_{i \uparrow}^f n_{i \downarrow}^f + E_0 \sum_{i, \sigma} n_{i \sigma}^f, \\
\label{eq:Ham_PAMSO2}
\mathcal{H}_1
&= \sum_{l,\bm{k} ,\sigma} V_l({\bm{k}})
( c^{\dagger}_{l\bm{k} \sigma}f^{}_{l\bm{k} \sigma}+{\mathrm{H.c.}} )
+ \sum_{l,\bm{k},\sigma}
\bm{g}^{cf}_l(\bm{k}) \cdot \bm{s}^{cf}_l (\bm{k}).
\end{align}
Here, $c^{\dagger}_{i \sigma}$($c^{}_{i \sigma}$) and $f^{\dagger}_{i \sigma}$($f^{}_{i \sigma}$)
are the creation (annihilation) operators of conduction and localized electrons with spin $\sigma$ at site $i=(p,l )$ ($p$ and $l$ denote the indices for the unit cell and sublattice, respectively);
$c^{\dagger}_{l\bm{k} \sigma}$($c^{}_{l\bm{k} \sigma}$) and $f^{\dagger}_{l\bm{k} \sigma} (f^{}_{l\bm{k} \sigma})$
are their Fourier transformations, respectively.
The first term in Eq.~(\ref{eq:Ham_PAMSO1}) represents the kinetic energy of conduction electrons; $t_{ij}$ is the hopping matrix element from site $j$ to $i$.
The second and third terms are the on-site Coulomb interaction for localized electrons and the atomic energy of localized electrons, respectively ($n^f_{i\sigma} = f^{\dagger}_{i \sigma} f^{}_{i \sigma}$).
Meanwhile, the first term in Eq.~(\ref{eq:Ham_PAMSO2}) represents the (symmetric) hybridization between conduction and localized electrons.
We note that $\mathcal{H}_0$ and the first term in $\mathcal{H}_1$ comprise the conventional periodic Anderson model~\cite{Anderson_PhysRev.124.41}.
On the other hand, the second term in Eq.~(\ref{eq:Ham_PAMSO2}) is introduced to describe an antisymmetric hybridization between $c$ and $f$ electrons;
$\bm{g}_l^{cf} (\bm{k})$ is the polar vector describing the antisymmetric part, and
\begin{align}
\bm{s}_l^{cf}(\bm{k})
= \sum_{\sigma, \sigma'} (c^{\dagger}_{l\bm{k} \sigma} \bm{\sigma}_{\sigma \sigma'} f^{}_{l\bm{k}\sigma'}
+ {\rm H.c.}),
\end{align}
where $\bm{\sigma}=(\sigma^x, \sigma^y, \sigma^z)$ is the vector of Pauli matrices, as in Eq.~(\ref{eq:Ham_ASOC_def}).
This term originates from the cooperation between the odd-parity crystalline electric field, atomic spin-orbit coupling, and off-site hybridization between orbitals with different parity~\cite{Yanase_kotai,Hayami_PhysRevB.90.024432}.
Although similar hybridizations may appear between the same $c$-$c$ or $f$-$f$ orbitals, we omit them and consider only the hybridization between different orbitals $c$ and $f$ in the present model.
This is justified in some realistic situations, e.g., when the atomic spin-orbit coupling for conduction electrons is weak and the intersite overlap integrals between localized electrons are small.
The model in Eq.~(\ref{eq:Ham_PAMSO}) is applicable to the generic systems without spatial inversion symmetry.
In the present study, we focus on a specific situation where the antisymmetric hybridization originates from local parity mixing.
As mentioned in the introduction, the local parity mixing exists in the lattice structures in which the inversion symmetry is broken at each site in a sublattice-dependent form;
typical examples are a zig-zag chain, honeycomb lattice, and diamond lattice.
On these lattices, the sublattice-dependent antisymmetric hybridization is present because each site is affected by the sublattice-dependent odd-parity crystalline electric field~\cite{Hayami_PhysRevB.90.024432}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=1.0 \hsize]{fig1.eps}
\caption{
\label{Fig:zigzag_chain}
Schematic pictures of (a) a single zig-zag chain and (b) a three-dimensional lattice composed of the zig-zag chains.
The zig-zag chain is in the $xy$ plane:
the $x$ and $y$ axes are taken in the parallel and perpendicular directions to the chain, respectively.
The dashed rectangle represents the unit cell including two sublattice sites, A and B.
$a$, $b$, and $c$ are the lattice constants;
$t_1$, $t_2$, $t_3$, and $t_4$ represent the hopping integrals.
}
\end{center}
\end{figure}
In the following sections, we consider one of the simplest realizations of the local parity mixing, a three-dimensional system composed of weakly-coupled one-dimensional zig-zag chains, as shown in Fig.~\ref{Fig:zigzag_chain}.
We set the $x$ and $y$ axes in the plane on which the chains lie, with taking the $x$ axis in the chain direction;
we set the lattice constants $a=b=c=1$.
For this setting, the inversion symmetry is broken at each lattice site, and the odd-parity crystalline electric field is present along the $y$ direction with an alternating sign for two sublattices.
Then, the local parity mixing results in the sublattice-dependent antisymmetric vector $\bm{g}^{cf}_{l}(\bm{k})$.
For the current system with the quasi-one-dimensional structure, $\bm{g}_l^{cf}(\bm{k})$ is approximately given by
\begin{align}
\label{eq:gsimple}
\bm{g}_l^{cf}(\bm{k}) =
(-1)^l g \hat{\bm{z}} \sin k_x,
\end{align}
where $l=0(1)$ for the A(B) sublattice, $g$ is the coupling constant, and $\hat{\bm{z}}$ is the unit vector
in the $z$ direction.
The form of $\bm{g}_l^{cf}(\bm{k}) \parallel \hat{z}$ is understood from Eq. (1) by considering the quasi one-dimensional kinetic motion $\bm{k} \parallel \hat{x}$ and $\bm{\nabla}V_{\rm pot} \parallel \hat{y}$.
Note that $\bm{g}_l^{cf}(\bm{k})$ acts only between $c$ and $f$ electrons on the same sublattice.
\subsection{
Extended Kondo lattice model}
\label{subsec:generalized KLM}
With these preliminaries in the previous subsection, we derive an effective low-energy model by treating ${\mathcal H}_1$ as the perturbation to ${\mathcal H}_0$.
This is done by a standard procedure on the basis of the canonical transformation, that is the Schrieffer-Wolff transformation~\cite{Schrieffer_PhysRev.149.491}.
Note that the procedure for the model without the antisymmetric hybridization term (i.e., $g=0$) leads to the standard Kondo lattice model~\cite{Schrieffer_PhysRev.149.491,Lacroix_PhysRevB.20.1969,Hewson199704}.
In other words, we here extend the Kondo lattice model by including the effect of the antisymmetric hybridization.
In the derivation, we assume that the number of $f$ electrons is one at each site, as in the standard procedure.
We also assume
that the hybridization $V_l(\bm{k})$ has only the on-site component and sublattice independent: we drop the $\bm{k}$ and $l$ dependences.
By the straightforward calculations given in Appendix in detail, we obtain the effective Hamiltonian as
\begin{align}
\label{eq:Ham_zigzag}
&\mathcal{H}_{{\rm ex-KLM}} =
\sum_{i,j,\sigma} (t_{ij} c_{i\sigma}^{\dagger}c^{}_{j \sigma} + {\rm H.c.})
+ \frac{J}{2} \sum_{i,\sigma,\sigma'} c_{i \sigma}^{\dagger} \bm{\sigma}_{\sigma \sigma'} c^{}_{i \sigma'} \cdot \bm{S}_i\nonumber \\
&\qquad+
\sum_{l,p, \bm{k}, \bm{k}'}
\left\{ \frac{D_l}{2}
{\rm s}_{k_x}
(S_{pl}^{+} s^-_{l\bm{k}'\bm{k}}
-S_{pl}^{-} s^+_{l\bm{k}'\bm{k}}
+S_{pl}^{z} n_{l\bm{k}'\bm{k}}) \delta_{\bm{R}_p} + {\rm H.c} \right\}
\nonumber \\
&\qquad+ \frac{G}{2}\sum_{l,p, \bm{k}, \bm{k}'}
{\rm s}_{k_x} {\rm s}_{k'_x} (2S_{pl}^{z} s^z_{l\bm{k}'\bm{k}}-S_{pl}^{+} s^-_{l\bm{k}'\bm{k}}
-S_{pl}^{-} s^+_{l\bm{k}'\bm{k}}) \delta_{\bm{R}_p},
\end{align}
where
${\rm s}_k = \sin k$,
$\delta_{\bm{R}_p} = e^{-{\rm i}(\bm{k}'-\bm{k})\cdot \bm{R}_{p}}$,
$s^+_{l\bm{k}'\bm{k}} = c_{l\bm{k}'\uparrow}^{\dagger} c^{}_{l\bm{k} \downarrow}$,
$s^-_{l\bm{k}'\bm{k}} = c_{l\bm{k}'\downarrow}^{\dagger} c^{}_{l\bm{k} \uparrow}$,
$s^z_{l\bm{k}'\bm{k}} = (c_{l\bm{k}'\uparrow}^{\dagger} c^{}_{l\bm{k} \uparrow}-c_{l\bm{k}'\downarrow}^{\dagger} c^{}_{l\bm{k} \downarrow})/2$, and
$n_{l\bm{k}'\bm{k}} = c_{l\bm{k}'\uparrow}^{\dagger} c^{}_{l\bm{k} \uparrow}+c_{l\bm{k}'\downarrow}^{\dagger} c^{}_{l\bm{k} \downarrow}$;
$\bm{S}_{pl}=(S^x_{pl},S^y_{pl},S^z_{pl})$ represents the localized spins at the unit cell $p$ and sublattice $l$ and $S^{\pm}_{pl}=S^x_{pl} \pm {\rm i} S^y_{pl}$.
In Eq.~(\ref{eq:Ham_zigzag}), the first and second terms are the kinetic energy of conduction electrons and the on-site exchange coupling between localized spins and conduction electrons, respectively.
These two terms represent the standard Kondo lattice model.
Here, $J$ is in the second order of $V$, as shown in Eqs.~(\ref{eq:J_expression}) and (\ref{eq:J_simple}).
In the first term, we take into account four different hopping matrix elements: intrachain hoppings between nearest and next-nearest neighbor sites, $t_1$ and
$t_2$, respectively [see Fig.~\ref{Fig:zigzag_chain}(a)], and interchain hoppings between the same sublattices for the neighboring sites in the $y$ and $z$ directions, $t_3$ and $t_4$, respectively
[see Fig.~\ref{Fig:zigzag_chain}(b)].
The third and fourth terms in Eq.~(\ref{eq:Ham_zigzag}) are the antisymmetric exchange couplings, derived from the antisymmetric hybridization in Eq.~(\ref{eq:Ham_PAMSO2}).
The third term results from the combination of $V$ and $\bm{g}_l^{cf}(\bm{k})$.
The coefficient $D_l$ is given by
\begin{align}
D_l=
(-1)^l \sqrt{JG},
\end{align}
where $G$ is proportional to $g^2$ [see Eqs.~(\ref{eq:G_expression}) and (\ref{eq:G_simple})]:
$D_l$ is proportional to $gV$.
The $s_{k_{x}}=\sin k_x$ dependence in Eq.~(\ref{eq:Ham_zigzag}) comes from Eq.~(\ref{eq:gsimple}), which is the source of the band deformation and magnetoelectric effect, similar to the case of toroidal ordering~\cite{Yanase:JPSJ.83.014703,Hayami_PhysRevB.90.024432};
this is demonstrated in Sect.~\ref{sec:Electronic Structure}.
The fourth term in Eq.~(\ref{eq:Ham_zigzag}) comes from the second order of $\bm{g}_l^{cf}(\bm{k})$.
It plays no essential role in the peculiar electronic and transport properties, since it is symmetric with respect to $(\bm{k},\bm{k})\to (-\bm{k},-\bm{k}')$.
Hereafter, assuming the situation $V>g$, we mainly consider the case with $J>|D_l|>G$; we take $J$ and $G$ to be positive.
For simplicity, we treat $\bm{S}_i$ as a classical spin with $|\bm{S}_i|=1/2$ in the following analysis.
\section{Results}
\label{sec:results}
\subsection{Two-site problem}
\label{subsec:2site}
Let us first examine the effect of the third and fourth terms in Eq.~(\ref{eq:Ham_zigzag}) by considering a simple two-site problem.
We discuss what type of spin configuration is stabilized by these antisymmetric exchange couplings arising from the antisymmetric hybridization.
We consider two sites connected by $t_2$ in a single chain, $1$ and $2$ (belonging to the same sublattice, say A-sublattice), as shown in Fig.~\ref{Fig:zigzag_spin_ponti}(a).
For the two sites, the exchange couplings in Eq.~(\ref{eq:Ham_zigzag}) are explicitly written in the matrix form as
\begin{align}
\label{eq:2site_mat}
\tilde{\mathcal{H}}_{
2 \, {\rm site}} =
\left(
\begin{array}{cc}
\tilde{\mathcal{H}}^{11}_{2 \, {\rm site}}
&
\tilde{\mathcal{H}}^{12}_{2 \, {\rm site}} \\
\tilde{\mathcal{H}}^{21}_{2 \, {\rm site}}
&
\tilde{\mathcal{H}}^{22}_{2 \, {\rm site}}
\end{array}
\right),
\end{align}
where
\begin{align}
\tilde{\mathcal{H}}^{11}_{2 \, {\rm site}} &= \left( \frac{J}{2} S^{z}_{1}+\frac{G}{2} {\rm s}_{k_x}^2 S^z_{2} \right) \sigma^z
+ \left( \frac{J}{2}S_{1}^x -\frac{G}{2} {\rm s}_{k_x}^2 S^x_{2} \right) \sigma^x, \\
\tilde{\mathcal{H}}^{12}_{2 \, {\rm site}} & =\frac{\sqrt{JG}}{2} {\rm s}_{k_x} (S^z_{1}+S^z_{2}) \sigma^0
+ {\rm i} \frac{\sqrt{JG}}{2} {\rm s}_{k_x} (S_{1}^x-S_{2}^x) \sigma^y.
\label{eq:2site_mat_offdiag}
\end{align}
$\tilde{\mathcal{H}}^{21}_{2 \, {\rm site}}=\tilde{\mathcal{H}}^{12\,\dagger}_{2 \, {\rm site}}$ and $\tilde{\mathcal{H}}^{22}_{2 \, {\rm site}}$ is given by $\tilde{\mathcal{H}}^{11}_{2 \, {\rm site}}$ with $1 \leftrightarrow 2$.
We take the basis $(c_{1 \uparrow}, c_{1 \downarrow}, c_{2 \uparrow}, c_{2 \downarrow})$ in Eq.~(\ref{eq:2site_mat}), and $\sigma^0$ is the $2 \times 2$ unit matrix.
Here, we omit the $y$ component of localized spins, $S^y$, without loss of generality, as it gives an equivalent contribution to that from $S^x$ due to the spin rotational symmetry in the $xy$ plane.
Equation~(\ref{eq:2site_mat}) indicates that
there is an effective hopping of conduction electrons, induced by the antisymmetric exchange coupling proportional to $\sqrt{JG}$.
The effective hopping depends on the directions of localized spins, as shown in Eq.~(\ref{eq:2site_mat_offdiag}).
This differentiates the energies for different spin configurations.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=1.0 \hsize]{fig2.eps}
\caption{
\label{Fig:zigzag_spin_ponti}
Schematic pictures of magnetic orders on the zig-zag chain: (a) the antiferromagnetic order with up-down spin moments along the $z$ direction ($z$-UD) and (b) the up-up-down-down antiferromagnetic order with the moments along the $x$ direction ($x$-UUDD).
In (a), the dashed oval shows the two neighboring sites in the same sublattice, $1$ and $2$.
}
\end{center}
\end{figure}
We first assume two different types of ferromagnetic configurations for the two A-sublattice sites: one is the state with spin polarization in the $x$ direction, and the other in the $z$ direction.
For these two cases, the matrix elements of the Hamiltonian in Eq.~(\ref{eq:2site_mat}) are given by
\begin{align}
\tilde{\mathcal{H}}^{x{\rm -F}}_{
2 \, {\rm site}} &=\left(
\begin{array}{cc}\displaystyle
\left(\frac{J}{4}-\frac{G}{4}
{\rm s}_{k_x}^2 \right) \sigma^x
& 0 \\
0
&\displaystyle \left( \frac{J}{4}-\frac{G}{4}
{\rm s}_{k_x}^2 \right) \sigma^x \\
\end{array}
\right),
\\
\tilde{\mathcal{H}}^{z{\rm -F}}_{2 \, {\rm site}} &=\left(
\begin{array}{cc}\displaystyle
\left( \frac{J}{4} + \frac{G}{4} {\rm s}_{k_x}^2 \right) \sigma^z
&\displaystyle -\frac{\sqrt{JG}}{2}
{\rm s}_{k_x} \sigma^0 \\\displaystyle
-\frac{\sqrt{JG}}{2} {\rm s}_{k_x} \sigma^0
&\displaystyle \left( \frac{J}{4}+\frac{G}{4}
{\rm s}_{k_x}^2 \right) \sigma^z \\
\end{array}
\right).
\end{align}
The eigenvalues are
obtained as
\begin{align}
\label{eq:energy-Fx-A-sublattice}
\varepsilon^{x{\rm -F}}_{
2 \, {\rm site}} (k_x) &=
\pm \left( \frac{J}{4}- \frac{G}{4}{\rm s}_{k_x}^2 \right), \\
\label{eq:energy-Fz-A-sublattice}
\varepsilon^{z{\rm -F}}_{
2 \, {\rm site}} (k_x) &=\pm \left( \frac{J}{4} + \frac{G}{4} {\rm s}_{k_x}^2\right) \pm \frac{\sqrt{JG}}{2} {\rm s}_{k_x}.
\end{align}
The results in Eqs.~(\ref{eq:energy-Fx-A-sublattice}) and~(\ref{eq:energy-Fz-A-sublattice}) clearly show that the antisymmetric exchange couplings prefer to the ferromagnetic configuration with magnetic moments in the $z$ direction for the two A-sublattice sites.
Next, let us consider a N\'eel-type antiferromagnetic spin configuration for the A-sublattice sites.
In this case, the Hamiltonian is given by
\begin{align}
\tilde{\mathcal{H}}^{x{\rm -AF}}_{
2 \, {\rm site}} &=\left(
\begin{array}{cc}\displaystyle
\left( \frac{J}{4}+\frac{G}{4} {\rm s}^2_{k_x} \right) \sigma^x
&\displaystyle {\rm i} \frac{\sqrt{JG}}{2} {\rm s}_{k_x} \sigma^y \\\displaystyle
- {\rm i} \frac{\sqrt{JG}}{2} {\rm s}_{k_x} \sigma^y
&\displaystyle -\left( \frac{J}{4}+\frac{G}{4} {\rm s}^2_{k_x} \right) \sigma^x
\end{array}
\right), \\
\tilde{\mathcal{H}}^{z{\rm -AF}}_{
2 \, {\rm site}} &=\left(
\begin{array}{cc}\displaystyle
\left( \frac{J}{4}-\frac{G}{4} {\rm s}^2_{k_x} \right) \sigma^z
& 0 \\
0
&\displaystyle -\left( \frac{J}{4}-\frac{G}{4} {\rm s}^2_{k_x} \right) \sigma^z
\end{array}
\right),
\end{align}
for the cases with the antiferromagnetic moments along the $x$ and $z$ directions, respectively.
The eigenvalues are obtained as
\begin{align}
\label{eq:energy-AFx-A-sublattice}
\varepsilon^{x{\rm -AF}}_{
2 \, {\rm site}} (k_x) &=
\pm \left( \frac{J}{4} +\frac{G}{4} {\rm s}^2_{k_x} \right) \pm \frac{\sqrt{JG}}{2} {\rm s}_{k_x}, \\
\label{eq:energy-AFz-A-sublattice}
\varepsilon^{z{\rm -AF}}_{
2 \, {\rm site}} (k_x) &=
\pm \left( \frac{J}{4}-\frac{G}{4} {\rm s}^2_{k_x} \right).
\end{align}
Thus, the stability of the $x$ and $z$ orders is opposite to the ferromagnetic case:
the antisymmetric exchange couplings favor the antiferromagnetic configuration with magnetic moments in the $x$ direction.
On the other hand, the spin configuration between the A and B sublattices will depend on an effective exchange coupling induced by the intersublattice hopping $t_1$.
For instance, when $t_1$ is much larger than $t_2$, the antiferromagnetic configuration between two sublattices is induced at and near half filling in the strongly-correlated region.
This is partly understood by considering the second-order perturbation with respect to $t_1$ at half filling, which leads to an effective antiferromagnetic interaction between localized spins.
Meanwhile, the ferromagnetic configuration is favored at low and high fillings due to the double-exchange mechanism, which is an effective ferromagnetic interaction between localized spins induced by the kinetic motion of itinerant electrons~\cite{Zener_PhysRev.82.403,anderson1955considerations}.
The above considerations suggest that the antisymmetric exchange couplings in Eq.~(\ref{eq:Ham_zigzag}) stabilize a specific spin configuration with a preference on the direction of magnetic moments.
For the ferromagnetic configuration in the same sublattice, the $z$-component order is preferred, which results in the up-down type antiferromagnetic state ($z$-UD) at and near half filling [see Fig.~\ref{Fig:zigzag_spin_ponti}(a)], and the ferromagnetic state ($z$-F) at low and high fillings.
On the other hand, for the antiferromagnetic configuration for the same sublattice, the $x$-component order is favored, leading to the up-up-down-down type antiferromagnetic state ($x$-UUDD) irrespective of the sign in the effective exchange coupling between different sublattices[see Fig.~\ref{Fig:zigzag_spin_ponti}(b)].
In the following sections, we focus on the $z$-UD order, since it accompanies an odd-parity multipole order composed of magnetic toroidal and quadrupole components~\cite{Yanase:JPSJ.83.014703}.
In this case, the ferroic toroidal order $\bm{T}$ is induced in the $x$ direction:
$\bm{T}\propto (\bm{\nabla}_{l}V_{{\rm pot}})\times\langle\bm{S}_{l}\rangle$.
In the following calculations, for simplicity, we restrict ourselves to the ordered states with the ordering vector, $\bm{q}=(q_{x},0,0)$.
\subsection{Band deformation and magnetoelectric effect}
\label{sec:Electronic Structure}
In this subsection, we examine the nature of the $z$-UD ordered state in Fig.~\ref{Fig:zigzag_spin_ponti}(a).
As shown in the previous study~\cite{Yanase:JPSJ.83.014703}, the $z$-UD order is expected to cause a band deformation with a shift of the band bottom.
Also, the system may exhibit the magnetoelectric effect in which a staggered magnetic moment is induced by an electric current.
We note that the magnetoelectric effect takes place even in the paramagnetic state, due to the presence of the antisymmetric exchange couplings.
We here confirm that these effects occur also in our extended Kondo lattice model in Eq.~(\ref{eq:Ham_zigzag}).
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.8 \hsize]{fig3.eps}
\caption{
\label{Fig:band}
Energy dispersion along the $k_x$ direction of the Hamiltonian in Eq.~(\ref{eq:Ham_zigzag}) under the $z$-UD order, as shown in the inset.
Each band is doubly degenerate.
The result is calculated at $t_1=1$, $t_2=0.1$, $t_3=0.2$, $t_4=0.2$, $J=4$, and $G=0.5$.
The dotted red lines indicate the band bottoms.
The dashed black curves represent the energy dispersion at $G=0$.
}
\end{center}
\end{figure}
Figure~\ref{Fig:band} shows the band structure in our model in Eq.~(\ref{eq:Ham_zigzag})
along the $k_x$ direction from $(-\pi,0,0)$ to $(\pi,0,0)$.
The result is calculated by assuming the presence of the $z$-UD type antiferromagnetic order with full polarization in localized electrons: $\bm{S}_{pl} = (-1)^l(\hat{\bm{z}}/2)$.
We take $t_1=1$, $t_2=0.1$, $t_3=0.2$, $t_4=0.2$, $J=4$, and $G=0.5$.
We also show the result at $G=0$ for comparison.
As shown in Fig.~\ref{Fig:band},
the $z$-UD order in the presence of the antisymmetric exchange couplings modifies the band structure in a peculiar way, with shifting the band bottom in the $k_x$ direction.
This is similar to the previous results~\cite{Yanase:JPSJ.83.014703,Hayami_PhysRevB.90.024432}:
although both spatial-inversion $\mathcal{P}$ and time-reversal $\mathcal{T}$ symmetries are broken in this ordered state,
the combined $\mathcal{PT}$ symmetry is retained, which ensures $E_\sigma(\bm{k}) = E_{-\sigma}(\bm{k})$ (two-fold degeneracy in each band), while there is no guarantee to satisfy $E_\sigma(\bm{k}) = E_{\sigma}(-\bm{k})$.
In our extended Kondo lattice model in Eq.~(\ref{eq:Ham_zigzag}),
this band deformation is understood as follows.
The Hamiltonian for the $z$-UD ordered state is represented by
\begin{align}
\label{eq:Ham_chain}
\tilde{\mathcal{H}}_{{\rm chain}} =\left(
\begin{array}{cc}
u+v
& \varepsilon_{{\rm 1}} \\
\varepsilon_{{\rm 1}}
& u-v
\\
\end{array}
\right),
\end{align}
where
$u = \varepsilon_{{\rm 2}} + G\sigma {\rm s}^2_{k_x}/4$ and
$v = (J\sigma/2 + \sqrt{JG} {\rm s}_{k_x} )/2$ ($\sigma = \pm 1$ for up and down spins);
$\varepsilon_1= - 2 t_1 \cos (k_x/2)$ and $\varepsilon_2 = -2t_2 \cos k_x -2 t_3 \cos k_y -2 t_4 \cos k_z$ represent the energy dispersions for the different and same sublattices, respectively.
We take the basis as $(c_{{\rm A}\bm{k}\sigma}, c_{{\rm B}\bm{k}\sigma})$, where $c_{{\rm A}\bm{k} \sigma}$($c_{{\rm B}\bm{k} \sigma}$) is the annihilation operator at A(B) sublattice with wave vector $\bm{k}$ and spin $\sigma$.
By diagonalizing the Hamiltonian in Eq.~(\ref{eq:Ham_chain}), the eigenvalues are obtained as
\begin{align}
\label{eq:energy_updown}
E (\bm{k}) &= \varepsilon_2 + \frac{G}{4} {\rm s}^{2}_{k_x}
\pm \sqrt{\varepsilon_1^2 + \frac14 \left(\frac{J}{2} + \sqrt{JG} {\rm s}_{k_x}\right)^2}.
\end{align}
This indicates that the band deformation with the band bottom shift is due to $\sqrt{JG} {\rm s}_{k_x}$, which comes from the $ D_{l}{\rm s}_{k_x}$ term in Eq.~(\ref{eq:Ham_zigzag}).
Equation~(\ref{eq:energy_updown}) also indicates that the peculiar band deformation does not occur for $G=0$ (the standard Kondo lattice model), as shown in Fig.~\ref{Fig:band}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.85 \hsize]{fig4.eps}
\caption{
\label{Fig:ME}
Correlation between the $z$-UD moment and electric current in the $x$ direction, $K_{zx}^{\rm (s)}$, in Eq.~(\ref{Eq:Magnetoelectric effect_sec3}).
The data are obtained for several values of $t_1$ and $t_2$ indicated in the figure, at $t_3=0.2$, $t_4=0.2$, $J=4$, $G=0.5$, $T=0.01$, and $\eta=0.01$.
The inset shows a schematic picture for the magnetoelectric response expected from the correlation $K_{zx}^{\rm (s)}$.
}
\end{center}
\end{figure}
Next, let us discuss the magnetoelectric effect by computing the linear response of the staggered magnetization in the form of $z$-UD order by an electric current in the $x$ direction.
The response is calculated by the Kubo formula in the form:
\begin{align}
K_{zx}^{\rm (s)} =\frac{g \mu_{{\rm B}}}{2}\frac{1}{{\rm i} V_{{\rm s}}} \sum_{m,n,\bm{k}} \frac{f(\varepsilon_{n \bm{k}})-f(\varepsilon_{m \bm{k}})}{\varepsilon_{n \bm{k}}-\varepsilon_{m \bm{k}}}
\frac{
\sigma_{z,\bm{k}}^{nm} J_{x,\bm{k}}^{mn} }{
\varepsilon_{n \bm{k}}-\varepsilon_{m \bm{k}}+{\rm i} \eta},
\label{Eq:Magnetoelectric effect_sec3}
\end{align}
where $V_{{\rm s}}$ is the system volume, $f(\varepsilon)$ is the Fermi distribution function, $\sigma_{z,\bm{k}}^{nm}=\langle n \bm{k} | (-1)^l \sigma^{z} | m\bm{k} \rangle$, and $J_{x,\bm{k}}^{mn}=\langle m \bm{k} |J_{x} | n\bm{k} \rangle$ is the matrix element of the current operator
$J_x = (e/\hbar)\partial \mathcal{H}_{{\rm ex-KLM}}
/\partial k_x$;
$\varepsilon_{n \bm{k}}$ and $| n \bm{k} \rangle$ are the $n$-th eigenvalue and eigenstate of $\mathcal{H}_{{\rm ex-KLM}}$ by assuming the $z$-UD order with full polarization in the localized spins.
We set $g \mu_{{\rm B}} e/2h=1$.
Figure~\ref{Fig:ME} shows the result of $K_{zx}^{{\rm (s)}}$ as a function of the electron density $n_{{\rm e}}=(1/N)\sum_{i \sigma} \langle c^{\dagger}_{i \sigma} c^{}_{i \sigma} \rangle$, where $N$ is the total number of sites.
The results are calculated at $t_3=0.2$, $t_4=0.2$, $J=4$, $G=0.5$, while changing $t_1$ and $t_2$, which may correspond to a deformation of the zig-zag chain (see below).
We set $T=0.01$ by taking the damping factor $\eta=0.01$.
As shown in the inset of Fig.~\ref{Fig:ME}, the
staggered moments in the $z$ direction are induced by the electric current in the $x$ direction, consistent with the previous study~\cite{Yanase:JPSJ.83.014703}.
This is due to the presence of the toroidal component $T^{x}$ in the multipole order~\cite{Hayami_PhysRevB.90.024432}.
As shown in Fig.~\ref{Fig:ME}, the magnetoelectric response tends to be larger for larger $t_1/t_2$.
This tendency is clearly seen in the low density region where the Fermi surface has a simple shape.
This suggests that the shallower zig-zag structure, where
the intersublattice hopping becomes dominant, gives rise to a larger magnetoelectric effect.
It also implies that the magnetoelectric response can be controlled by a uniaxial pressure along or perpendicular to the zig-zag chain.
The results are consistent with those in the single-band Hubbard model including the effect of the antisymmetric spin-orbit coupling~\cite{Yanase:JPSJ.83.014703}.
Moreover, the response also depends on $J$ and $G$ in a peculiar manner (not shown here).
The $G$ dependence is rather simple, at least, in the low density region.
This is because the increase of $G$ results in the increment of the antisymmetric spin-orbit coupling as in the third term of Eq.~(\ref{eq:Ham_zigzag}), which is the origin of the staggered magnetoelectric response.
Meanwhile, the $J$ dependence is complicated because the increase of $J$ enhances not only the antisymmetric spin-orbit coupling but also the exchange coupling as in the second term of Eq.~(\ref{eq:Ham_zigzag}).
\subsection{Phase diagram}
\label{subsec:phase diagram}
In the previous subsection, we have studied the electronic and transport properties in the assumed $z$-UD ordered state.
Now, we examine when and how such an ordered state is realized in the model in Eq.~(\ref{eq:Ham_zigzag}).
For that purpose, we perform three numerical calculations which are complementary to each other: variational calculation for the ground state in Sect.~\ref{sec:Variational Calculation}, simulated annealing in Sect.~\ref{sec:Simulated Annealing}, and Monte Carlo simulation at finite temperatures in Sect.~\ref{sec:Monte Carlo Simulation}.
\subsubsection{Variational calculation}
\label{sec:Variational Calculation}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=1.0 \hsize]{fig5.eps}
\caption{
\label{Fig:zigzag_variational}
Schematic pictures of ordering patterns of localized spins used in the variational calculations.
U and D represent up- and down-spin polarization, respectively, and F represents the ferromagnetic state.
The prefices $x$ and $z$ indicate the direction of the magnetic moments.
}
\end{center}
\end{figure}
First, we examine the ground state of the model in Eq.~(\ref{eq:Ham_zigzag}) by
a variational calculation.
Namely, for each parameter set of the Hamiltonian, we compare the zero-temperature grand potential per site, $\Omega = E - \mu n_{{\rm e}}$ ($E = \langle \mathcal{H}_{{\rm ex-KLM}} \rangle /N$ is the internal energy per site and $\mu$ is the chemical potential) for different magnetically-ordered states and determine the most stable one which gives the lowest $\Omega$.
In the present calculations, we assume a collection of typical magnetic orders in the localized spins, up to the eight-site unit cell in the single chain as candidates for the ground state;
the states considered in this study are shown in Fig.~\ref{Fig:zigzag_variational}.
We consider only uniform $\bm{q}=\bm{0}$ orders for all these patterns; namely, we assume that each magnetically ordered pattern is composed of a uniform arrangement of the magnetic unit cell in all the directions.
The data are computed by approximating the integral over the folded Brillouin zone using the sum over grid points of $64 \times 64 \times 64$.
The phase separated regions are determined by the method in Ref.~\citen{Hayami_Conference}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.9 \hsize]{fig6.eps}
\caption{
\label{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu}
Ground-state phase diagram of the model in Eq.~(\ref{eq:Ham_zigzag}) obtained by the variational calculation at (a) $G=0$ and (b) $G=0.5$.
Other parameters are taken to be $t_1=1$, $t_2=0.1$, $t_3=0.2$, and $t_4=0.2$.
PS indicates a phase separated region.
Ordering patterns are shown in Fig.~\ref{Fig:zigzag_variational}.
}
\end{center}
\end{figure}
Figure~\ref{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu} shows the phase diagram obtained by the variational calculation, as a function of $n_{{\rm e}}$ and $J$.
Figure~\ref{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu}(a) corresponds to the result for the standard Kondo lattice model
without the antisymmetric exchange couplings, i.e., at $G=0$.
In this case, there is no spin anisotropy because the model in Eq.~(\ref{eq:Ham_zigzag}) retains the spin rotational symmetry.
In the low and high density regions, the ferromagnetic metallic phase appears and becomes wider as $J$ increases.
The ferromagnetic phase is stabilized by the double-exchange mechanism~\cite{anderson1955considerations}.
On the other hand, a staggered antiferromagnetic order along the chain is stabilized at and near half-filling ($n_{{\rm e}}=1$) due to the effective antiferromagnetic interaction mentioned in Sect.~\ref{subsec:2site}.
Note that an incommensurate order might take over the antiferromagnetic state in the weak-coupling region, which cannot be described in the present variational scheme within the limited sizes of magnetic unit cells; we will reexamine this point in the following subsections.
In the intermediate $n_{{\rm e}}$ region, there are several phases characterized by other ordering wave vectors: UUDD and 4U4D antiferromagnetic orders (see Fig.~\ref{Fig:zigzag_variational}).
Note that, in all these phases, the common direction of the magnetic moments can be taken arbitrarily owing to spin rotational symmetry of the system.
Next, we discuss the effect of the antisymmetric exchange couplings by turning on $G$.
When $G\neq 0$, the antisymmetric exchange couplings introduce the spin anisotropy, and hence, the system has a preference in the direction of the magnetic moments in each phase.
Figure~\ref{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu}(b) shows the result at $G=0.5$.
At and near half filling, the quantized axis in the up-down antiferromagnetic phase prefers to be fixed in the $z$ direction.
This is the $z$-UD phase with multipole ordering discussed in Sect.~\ref{sec:Electronic Structure}, which gives rise to a band deformation with a band bottom shift and exhibits the magnetoelectric effect.
The result in Fig.~\ref{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu} indicates that this UD phase tends to be stabilized by $G$, as was discussed in Sect.~\ref{subsec:2site}.
Meanwhile, the UUDD phases near quarter and three quarter fillings are also present at $G=0.5$, as shown in Fig.~\ref{Fig:t1=1.0_t2=0.1_G=0_t3=0.2_souzu}(b).
In this case, however, the magnetic moments are aligned within the $xy$ plane by the spin anisotropy;
we denote it by UUDD with the prefix $x$, while the moment can point to any direction within the $xy$ plane.
On the other hand, in the ferromagnetic phases in the low and high filling regions, the moments are polarized in the $z$ direction.
For all the phases we obtained,
the direction of the magnetic moments as well as the stable parameter region is consistent with the arguments for the two-site problem discussed in Sect.~\ref{subsec:2site}.
The tendency does not change while changing $t_1/t_2$ (not shown).
\subsubsection{Simulated annealing}
\label{sec:Simulated Annealing}
In order to confirm the stability of the $z$-UD state by a more unbiased method than the variational calculation in Sect.~\ref{sec:Variational Calculation}, we adopt simulated annealing.
The simulated annealing is an optimization method for finding the global minimum of a function that possesses many local minima~\cite{kirkpatrick1983optimization}.
By using the method, we obtain the accurate magnetic order in the ground state within the unit cell we set in the calculation; we do not need to assume a specific magnetic order,
in contrast to the variational calculation in Sect.~\ref{sec:Variational Calculation}.
We optimize the classical localized spins by the simulated annealing as follows:
temperature $T$ is decreased gradually, and for each $T$, the spin configuration is updated by the Monte Carlo sampling with the single-spin flip algorithm.
In the present calculations, we performed the simulated annealing while decreasing $T$ in a geometrical way, $T_{n+1} = \alpha T_n$, where $T_n$ is the temperature in the $n$-th step.
We started from the initial temperature $T_0=1.0$ by taking the coefficient of geometrical cooling $\alpha = 0.97$ and the total steps of cooling 270: the final temperature $T_{270}$ reaches down to $\sim 3.0\times 10^{-4}$.
We considered the systems with $N=16 \times 1 \times 1$ and $24 \times 1 \times 1$ while introducing a supercell consisting of $N_k = 8^3$ copies of the $N$-site lattice to reduce the finite-size effect.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.8 \hsize]{fig7.eps}
\caption{
\label{Fig:annealing_Kondo_zigzag}
$q_x$ dependence of the spin structure factor at $q_y=q_z=0$ divided by the system size $N$ obtained by the simulated annealing.
See Ref.\citen{comment_wave_number}.
The data are calculated at $t_1=1$, $t_2=0.1$, $t_3=0.2$, $t_4=0.2$, and $J=4$.
The results for $G=0$ and $G=0.5$ with the system size $N=16$ and 24 are shown in each figure: (a) $\mu=-0.4$ and (b) $\mu=-0.8$.
Electron fillings for each $\mu$ are also shown in the figure.
}
\end{center}
\end{figure}
Figure~\ref{Fig:annealing_Kondo_zigzag} shows the spin structure factor obtained by the simulated annealing.
The spin structure factor is defined by
\begin{align}
\label{eq:Sq_sec3}
S_{\bm{q}} = \frac{1}{N} \sum_{i,j} (\bm{S}_i \cdot \bm{S}_j) e^{{\rm i} \bm{q}\cdot (\bm{r}_i - \bm{r}_j)},
\end{align}
where $\bm{q}$ is a wave vector and $\bm{r}_i$ is the position vector at site $i$.
In Fig.~\ref{Fig:annealing_Kondo_zigzag}, $S_{\bm{q}}/N$ is plotted as a function of $q_x$ at $\bm{q} = (q_x,0,0)$.
Note that, for the perfect UD order, $S_{\bm{Q}}/N = 0.25$ at $\bm{Q}=(\pi, 0, 0)$, and otherwise zero~\cite{comment_wave_number}.
At $\mu=0.0$ (almost half filling, $n_{\rm e} \simeq 1.0$), the spin structure factor shows a sharp peak at $\bm{Q}=(\pi, 0, 0)$ for both $G=0$ and $G=0.5$ (not shown here).
The peak values do not depend on the system size, which suggests the UD order is stable even when allowing other magnetic orders in larger unit cells than in the variational calculations.
Similar to the results in Sect.~\ref{sec:Variational Calculation}, the moments for the UD order are aligned in the $z$ direction when introducing $G$.
As shown in Fig.~\ref{Fig:annealing_Kondo_zigzag}(a), the peaks remain at $\bm{Q}=(\pi, 0, 0)$ against a slight decrease of $n_{{\rm e}}$ by a few percent.
However, the peak for $G=0$ slightly decreases ($\sim 1.5\%$) when increasing the system size from $N=16$ to $24$, while the peak value for $G=0.5$ remains almost unchanged (the change is less than 0.3\%).
These results imply that the UD state survives against a slight doping, and the antisymmetric exchange couplings stabilize the $z$-UD state.
While further decreasing $n_{{\rm e}}$, as shown in Fig.~\ref{Fig:annealing_Kondo_zigzag}(b), the peak position for $G=0$ shifts to a smaller $q_x$.
This suggests the commensurate UD state is no longer stable and taken over by an incommensurate order with a longer period, as anticipated in the variational arguments in the previous subsection.
Meanwhile, for $G=0.5$, the $z$-UD state remains stable, as shown in Fig.~\ref{Fig:annealing_Kondo_zigzag}(b).
Thus, the results of the simulated annealing confirm the variational results in the previous subsection: the $z$-UD ordered state appears as a stable phase at and near half filling in the presence of the antisymmetric exchange couplings.
\subsubsection{Monte Carlo simulation}
\label{sec:Monte Carlo Simulation}
In this subsection, we investigate the stability of the $z$-UD phase at finite temperatures by Monte Carlo simulation.
In the Monte Carlo calculations, we adopt the standard method for the spin-charge coupled systems with classical localized moments~\cite{Yunoki_PhysRevLett.80.845}.
We typically performed
10,000-90,000 Monte Carlo steps after 10,000 steps for thermalization.
The statistical errors were estimated by dividing the data into five to ten bins and calculating the standard deviation among the bins.
The calculations were performed on the $N=4 L \times L \times L$-site lattice with $L=2$ and $4$ under the periodic boundary conditions in all the directions.
For $L=2$, we introduced a supercell consisting of $N_k = 2^3$ copies of the $N$-site lattice so that the finite-size effect is reduced.
In order to stabilize the $q_y=q_z=0$ order as in the previous sections,
we introduce the additional ferromagnetic exchange interaction between localized spins $\mathcal{H}_{{\rm F}} = -J_{{\rm F}} \sum_{\langle i, j \rangle_{yz}} \bm{S}_i \cdot \bm{S}_j$;
we take $J_{{\rm F}} = 0.1$ and the sum of $\langle i,j \rangle_{yz}$ over the nearest-neighbor sites for the $y$ and $z$ directions.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.9 \hsize]{fig8.eps}
\caption{
\label{Fig:zika_Monte}
Monte Carlo results for (a) the order parameter for the $z$-UD order, $m_{\bm{Q}}$, and (b) the electron density $n_{{\rm e}}$ at $\mu=-0.1$ and $\mu = -0.5$.
The calculations were done at $t_1=1$, $t_2=0.1$, $t_3=0.2$, $t_4=0.2$ $J=4$, and $G=0.5$ for the system sizes $N=8 \times 2 \times 2$ and $16 \times 4 \times 4$.
}
\end{center}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.9 \hsize]{fig9.eps}
\caption{
\label{Fig:souzu_Monte}
Finite-temperature phase diagram of the model in Eq.~(\ref{eq:Ham_zigzag}) as a function of the electron density $n_{{\rm e}}$ obtained by the Monte Carlo simulation.
The symbols indicate the critical temperatures, which are estimated by the inflection points of the order parameter $m_{\bm{Q}}$ plotted in Fig.~\ref{Fig:zika_Monte}.
The horizontal error bars indicate the statistical errors of $n_{\rm e}$ at the chemical potential for the critical temperature.
The results are calculated at $t_1=1$, $t_2=0.1$, $t_3=0.2$, $t_4=0.2$, $J=4$, and $G=0.5$.
}
\end{center}
\end{figure}
Figure~\ref{Fig:zika_Monte}(a) shows the temperature dependence of the order parameter for the UD order, $m_{\bm{Q}}$.
Here, $m_{\bm{Q}}$ is defined by $m_{\bm{Q}}=[S_{\bm{Q}}/N]^{1/2}$, where $S_{\bm{Q}}$ is the spin structure factor in Eq.~(\ref{eq:Sq_sec3}) at $\bm{Q} =(\pi,0,0)$.
The data are calculated for the two different system sizes at $J=4$ and $G=0.5$ for two different values of the chemical potential, $\mu=-0.1$ and $\mu=-0.5$.
The results show that $m_{\bm{Q}}$ develops rapidly below a particular temperature.
We confirmed that the magnetic moments are along the $z$ direction by analyzing the spin component of $S_{\bm{Q}}$.
These indicate a phase transition from the high-temperature
paramagnetic state to the low-temperature $z$-UD ordered state.
The rough estimates of the critical temperature $T_c$ can be obtained from the inflection point of $m_{\bm{Q}}(T)$:
$T_c \simeq 0.09$ at $\mu=-0.1$ ($n_{{\rm e}}\sim 0.96$) and $T_c \simeq 0.075$ at $\mu=-0.5$ ($n_{{\rm e}}\sim 0.85$) [see also Fig.~\ref{Fig:zika_Monte}(b)].
With further decreasing temperature, the order parameter $m_{\bm{Q}}$ approaches its saturated value $0.5$ in the ground state.
This confirms that the multipole ordered state found in the variational calculations remains stable against thermal fluctuations as well as the carrier doping.
By calculating $m_{\bm{Q}}(T)$ while varying $\mu$ in a similar way, we obtain the finite-temperature phase diagram in Fig.~\ref{Fig:souzu_Monte}.
The $z$-UD phase is stable around half filling, with a dome-like shape of $T_c$ with its maximum around $n_{\rm e}=1$.
Although it is difficult to determine the phase boundary at $n_{{\rm e}} \sim 0.7$ within the limited system sizes in the current calculations,
the result indicates that the multipole ordered state is robustly stable around half filling, consistent with the results by the variational calculation and simulated annealing.
\section{Summary}
\label{sec:Summary}
In summary, we have investigated the odd-parity multipole ordering that is spontaneously induced by the antisymmetric spin-orbit coupling in the systems with local parity mixing.
Starting from a general form of the site-dependent antisymmetric hybridization between different parity orbitals, we derived an effective low-energy model with site-dependent antisymmetric exchange couplings.
This is an extended Kondo lattice model, which is a fundamental model for considering the effect of antisymmetric hybridization in $d$- and $f$-electron systems.
We have analyzed the model on a quasi-one-dimensional zig-zag lattice, as the minimal lattice structure describing local parity mixing.
We found that the antisymmetric exchange couplings induce the effective hopping of conduction electrons depending on the configurations of localized spins.
One of the stable configurations is a N\'eel type antiferromagnetic order with the moments perpendicular to the zig-zag plane.
The N\'eel order accompanies an odd-parity multipole order composed of magnetic toroidal and quadrupole components.
This unusual multipole order exhibits a band deformation with a band bottom shift and magnetoelectric response, due to the activated toroidal moment.
We have investigated the stability of the multipole ordered state by the complementary numerical calculations, i.e., the variational calculation for the ground state, the simulated annealing, and the Monte Carlo simulation at finite temperatures.
We found that the multipole ordered state is indeed stabilized by the antisymmetric exchange couplings in a wide parameter range at and near half filing.
Our results will stimulate further studies of odd-parity multipole ordering in systems with local parity mixing.
Multipole orders similar to that in the present study will be widely observed in the materials to meet the following conditions: (i) local inversion symmetry breaking due to the lattice structure (zig-zag, honeycomb, diamond, \dots), (ii) strong spin-orbit coupling, (iii) hybridization between different parity orbitals, e.g., $s$-$f$, $p$-$d$, and $d$-$f$.
There are many candidate materials to meet such conditions in $f$-electron systems, as mentioned in the introduction.
We anticipate a similar situation also in 4$d$- and 5$d$-electron systems where localized levels are expected under some crystal field splitting.
It is desired to systematically study such systems from the viewpoint of odd-parity multipole ordering for further understanding of the exotic magnetism, accompanying peculiar electronic and transport properties.
\begin{acknowledgements}
SH is supported by Grant-in-Aid for JSPS Fellows.
This work was supported by Grants-in-Aid for Scientific Research (No.~24340076), the Strategic Programs for Innovative Research (SPIRE), MEXT, and the Computational Materials Science Initiative (CMSI), Japan.
\end{acknowledgements}
|
\section{Introduction}
In several earlier papers \cite{R7, R25, R21} I showed how to give
a precise meaning to statements in the literature of high-energy
physics and string theory of the kind ``matrix algebras
converge to the sphere''. (See the references to the quantum physics
literature given in
\cite{R7, R17, Stn2, ChZ, DiG, AHI, AcD}.) I did this by introducing
and developing a concept of ``compact quantum metric spaces'', and a corresponding quantum Gromov-Hausdorff-type distance
between them. The compact quantum spaces are unital C*-algebras,
and the metric data is given by putting on the
algebras seminorms that play the role of the usual Lipschitz
seminorms on the algebras of continuous functions on ordinary
compact metric spaces. The natural setting for ``matrix algebras
converge to the sphere'' is that of coadjoint orbits of compact
semi-simple Lie groups.
But physicists need much more than just the algebras. They need vector
bundles, gauge fields, Dirac operators, etc. So I now seek to give
precise meaning to statements in the physics literature of the kind
``here are the vector bundles over the matrix algebras that
correspond to the monopole bundles on the sphere''. (See \cite{R17}
for many references.) In \cite{R17} I studied convergence of ordinary
vector bundles on ordinary compact metric spaces for ordinary
Gromov-Hausdorff distance. From that study it became clear that one
needed Lipschitz-type seminorms on all the matrix algebras over the
underlying algebras, with these seminorms coherent in the
sense that they form a ``matrix seminorm'' (defined below).
The purpose of this paper is to define and develop the properties
of such matrix seminorms for the setting of coadjoint orbits, and
especially to study how these matrix seminorms mesh with
quantum Gromov-Hausdorff distance.
Very recently {Latr\'emolie\`re} \ introduced an improved version of quantum
Gromov-Hausdorff distance \cite{Ltr2} that he calls ``propinquity''. We
show that propinquity works very well for our setting of coadjoint orbits,
and so propinquity is the form of quantum Gromov-Hausdorff
distance that we use in this paper. {Latr\'emolie\`re} \ defines his propinquity
in terms of an improved version of the ``bridges'' that I had used
in my earlier papers. For our matrix seminorms we need
corresponding ``matricial bridges'', and we show how to
construct natural ones for the setting of coadjoint orbits.
It is crucial to obtain good upper bounds for the lengths of
the bridges that we construct. In the matricial setting the
calculations become somewhat complicated. In order to ease
the calculations we introduce a notion of ``bridges with
conditional expectations'', and develop their general
theory, including the matricial case, and including
bounds for their lengths in the matricial case.
The main theorem of this paper, Theorem \ref{thmmat1q}, states
in a quantitative way that for the case of coadjoint orbits
the lengths of the matricial bridges goes to 0 as the size
of the matrix algebras goes to infinity.
We also discuss a closely related class of examples
coming from \cite{R25}, for which we construct bridges
between different matrix algebras associated to a given
coadjoint orbit. This provides further motivation for our
definitions and theory of bridges with conditional
expectation.
\tableofcontents
\section{The first basic class of examples}
\label{basic1}
In this section we describe the first of the two basic classes of examples
underlying this paper.
It consists of the main class of examples studied in the
papers \cite{R7, R21}.
We begin by describing the common setting for the two
basic classes of examples.
Let $G$ be a compact group (perhaps even finite, at first). Let $U$ be an
irreducible unitary representation of $G$ on a
(finite-dimensional) Hilbert space ${\mathcal
H}$. Let ${\mathcal B} = {\mathcal L}({\mathcal H})$ denote the $C^*$-algebra of all linear
operators on ${{\mathcal H}}$ (a ``full matrix algebra'', with its operator
norm). There is a natural action, $\a$, of $G$ on ${\mathcal B}$ by conjugation by
$U$, that is, $\a_x(T) = U_xTU_x^*$ for $x \in G$ and $T \in {\mathcal B}$. Because
$U$ is irreducible, the action $\a$ is ``ergodic'', in the sense that the
only $\a$-invariant elements of ${\mathcal B}$ are the scalar multiples of the
identity operator.
Let $P$ be a rank-one projection in ${\mathcal B}({{\mathcal H}})$ (traditionally
specified by giving a non-zero vector in its range). For any $T \in {\mathcal B}$ we
define its Berezin covariant symbol \cite{R7}, $\sigma_T$, with
respect to $P$, by
\[
\sigma_T(x) = {\mathrm{tr}}(T\a_x(P)),
\]
where ${\mathrm{tr}}$ denotes the usual (un-normalized) trace on ${\mathcal B}$. (When the
$\a_x(P)$'s are viewed as giving states on ${\mathcal B}$ via ${\mathrm{tr}}$, they
form a family of ``coherent states'' \cite{R7} if a few additional
conditions are satisfied.) Let $H$ denote the
stability subgroup of $P$ for $\a$. Then it is evident that $\sigma_T$ can be
viewed as a (continuous) function on $G/H$. We let $\l$ denote the action
of $G$ on $G/H$, and so on ${\mathcal A} = C(G/H)$, by left-translation. If we
note that
${\mathrm{tr}}$ is $\a$-invariant, then it is easily seen that $\sigma$ is a unital,
positive, norm-nonincreasing, $\a$-$\l$-equivariant map from ${\mathcal B}$ into ${\mathcal A}$.
Fix a continuous length function, $\ell$, on $G$ (so
$G$ must be metrizable). Thus $\ell$ is non-negative, $\ell(x) = 0$ iff
$x = e_G$ (the identity element of $G$), $\ell(x^{-1}) = \ell(x)$, and
$\ell(xy) \le \ell(x) + \ell(y)$. We also require
that $\ell(xyx^{-1}) = \ell(y)$ for all $x$, $y \in G$. Then in
terms of $\a$ and $\ell$ we can define a
seminorm, $L^{\mathcal B}$, on ${\mathcal B}$ by the formula
\begin{equation}
\label{lipn}
L^{\mathcal B}(T) = \sup\{ \|\a_x(T) - T \|/\ell(x): x \in G \quad \mathrm{and}
\quad x \neq e_G\} .
\end{equation}
Then $({\mathcal B},L_{\mathcal B})$
is an example of a compact C*-metric-space,
as defined in definition 4.1 of \cite{R21},
and in particular $L_{\mathcal B}$ satisfied the conditions
given there for being a ``Lip-norm''.
Of course, from $\l$ and $\ell$ we also obtain a seminorm,
$L^{\mathcal A}$, on ${\mathcal A}$ by the evident analog of
formula \ref{lipn}, except that we must permit $L^{\mathcal A}$
to take the value $\infty$. It is shown in proposition 2.2
of \cite{R4} that the set of functions for which
$L^{\mathcal A}$ is finite (the Lipschitz functions) is a dense
$*$-subalgebra of ${\mathcal A}$. Also, $L^{\mathcal A}$
is the restriction to ${\mathcal A}$ of the seminorm on
$C(G)$ that we get from $\ell$ when we view $C(G/H)$ as a subalgebra of
$C(G)$, as we will often do when convenient.
From $L^{\mathcal A}$ we can use equation \ref{metr} below
to recover the usual quotient metric \cite{Wvr2}
on $G/H$ coming from the metric on $G$ determined by $\ell$.
One can check easily that $L^{\mathcal A}$ in turn comes from
this quotient metric. Thus $({\mathcal A},L^{\mathcal A})$ is the compact C*-metric-space
associated to this ordinary compact metric space.
Then for any bridge
from ${\mathcal A}$ to ${\mathcal B}$ we can use $L^{\mathcal A}$ and $L^{\mathcal B}$ to define the
length of the bridge in the way given by {Latr\'emolie\`re}, which we will describe
soon below.
For any two unital C*-algebras ${\mathcal A}$ and ${\mathcal B}$ a
bridge from ${\mathcal A}$ to ${\mathcal B}$ in the sense of {Latr\'emolie\`re} \ \cite{Ltr2} is a quadruple
$\{{\mathcal D}, \pi_{\mathcal A}, \pi_{\mathcal B}, \o \}$ for which ${\mathcal D}$ is a unital C*-algebra,
$\pi_{\mathcal A}$ and $\pi_{\mathcal B}$ are unital injective homomorphisms of
${\mathcal A}$ and ${\mathcal B}$ into ${\mathcal D}$, and $\o$ is a self-adjoint element
of ${\mathcal D}$ such that 1 is an element of the spectrum
of $\o$ and $\|\o \| = 1$. Actually, {Latr\'emolie\`re} \ only requires a looser
but more complicated condition on $\o$, but the
above condition will be appropriate for our examples. Following
{Latr\'emolie\`re} \ we will call $\o$ the ``pivot'' for the bridge.
We will often omit mentioning the injections $\pi_{\mathcal A}$ and $\pi_{\mathcal B}$
when it is clear what they are from the context, and accordingly we
will often write as though ${\mathcal A}$ and ${\mathcal B}$ are unital subalgebras of ${\mathcal D}$.
For our first class of examples, in which ${\mathcal A}$ and ${\mathcal B}$ are
as described in the paragraphs above,
we take ${\mathcal D}$ to be the C*-algebra
\[
{\mathcal D} = {\mathcal A} \otimes {\mathcal B} = C(G/H, {\mathcal B}) .
\]
We take $\pi_{\mathcal A}$ to be the injection of ${\mathcal A}$ into ${\mathcal D}$
defined by
\[
\pi_{\mathcal A}(a) = a \otimes 1_{\mathcal B}
\]
for all $a \in {\mathcal A}$, where $1_{\mathcal B}$ is the identity element of ${\mathcal B}$.
The injection $\pi_{\mathcal B}$ is defined similarly.
From the many calculations done in \cite{R7, R21} it is
not surprising that we define the pivot $\o$ to be the function in
$C(G/H, {\mathcal B})$ defined by
\[
\o(x) = \a_x(P)
\]
for all $x \in G/H$. We notice that $\o$ is actually a projection
in ${\mathcal D}$, and so it satisfies the requirements for being a pivot.
We will denote the bridge $\{{\mathcal D}, \o\}$ by $\Pi$.
For any bridge between two unital C*-algebras ${\mathcal A}$ and ${\mathcal B}$
and any choice of
seminorms $L^{\mathcal A}$ and $L^{\mathcal B}$ on ${\mathcal A}$ and ${\mathcal B}$,
{Latr\'emolie\`re} \ \cite{Ltr2} defines the ``length'' of the
bridge in terms of these seminorms. For this
he initially puts relatively weak requirements on the seminorms,
but for the purposes of the matricial bridges that we will
define later, we need somewhat different weak requirements.
To begin with, {Latr\'emolie\`re} \ only requires his seminorms, say
$L^{\mathcal A}$ on a unital C*-algebra ${\mathcal A}$, to be
defined on the subspace of self-adjoint elements of the
algebra, but we need ${\mathcal A}$ to be defined on
all of ${\mathcal A}$. To somewhat compensate for this we require
that $L^{\mathcal A}$ be a $*$-seminorm. As with {Latr\'emolie\`re},
our $L^{\mathcal A}$ is permitted to take value $+\infty$.
{Latr\'emolie\`re} \ also requires the subspace on which $L^{\mathcal A}$
takes finite values to be dense in the algebra. We do not really
need this here, but for us there would be no harm in assuming it,
and all interesting examples probably will satisfy this. Finally,
{Latr\'emolie\`re} \ requires that the null space of $L^{\mathcal A}$ (i.e
where it takes value 0) be exactly ${\mathbb C} 1_{\mathcal A}$.
We must loosen this to simply requiring that $L^{\mathcal A}(1_{\mathcal A}) = 0$,
but permitting $L^{\mathcal A}$ to also take value 0 on elements not
in ${\mathbb C} 1_{\mathcal A}$. We think of such seminorms as ``semi-Lipschitz
seminorms''. To summarize all of this we make:
\begin{defn}
\label{slip}
By a \emph{slip-norm} on a unital C*-algebra ${\mathcal A}$ we mean
a $*$-seminorm, $L$, on ${\mathcal A}$ that is permitted to take the
value $+\infty$, and is such that $L(1_{\mathcal A}) = 0$.
\end{defn}
Because of these weak requirements on $L^{\mathcal A}$, various
quantities in this paper may be $+\infty$, but most interesting
examples will satisfy stronger requirements that will
result in various quantities being finite.
{Latr\'emolie\`re} \ defines the length of a bridge by first defining its ``reach''
and its ``height''. We apply his definitions to slip-norms.
\begin{defn}
\label{reach}
Let ${\mathcal A}$ and ${\mathcal B}$ be unital C*-algebras and let
$\Pi = \{{\mathcal D}, \o\}$ be a bridge from ${\mathcal A}$ to ${\mathcal B}$ .
Let $L^{\mathcal A}$ and $L^{\mathcal B}$ be slip-norms on ${\mathcal A}$ and ${\mathcal B}$. Set
\[
{\mathcal L}_{\mathcal A}^1 = \{a \in {\mathcal A}: a = a^* \quad \mathrm{and} \quad L^{\mathcal A}(a) \leq 1\} ,
\]
and similarly for ${\mathcal L}_{\mathcal B}^1$. (We can view these as subsets of ${\mathcal D}$.)
Then the \emph{reach} of $\Pi$ is given by:
\[
\mathrm{reach}(\Pi) = \mathrm{Haus}_{\mathcal D}\{{\mathcal L}_{\mathcal A}^1 \o \ , \ \o {\mathcal L}_{\mathcal B}^1\} ,
\]
where $\mathrm{Haus}_{\mathcal D}$ denotes the Hausdorff distance with respect
to the norm of ${\mathcal D}$, and where the product defining ${\mathcal L}_{\mathcal A}^1 \o$ and
$ \o {\mathcal L}_{\mathcal B}^1$ is that of ${\mathcal D}$.
\end{defn}
{Latr\'emolie\`re} \ shows
just before definition 3.14 of \cite{Ltr2} that,
under conditions that include the case in which
$({\mathcal A}, L^{\mathcal A})$ and $({\mathcal B}, L^{\mathcal B})$ are
C*-metric spaces, the reach of $\Pi$ is finite.
To define the height of $\Pi$ we need to consider the state space, $S({\mathcal A})$,
of ${\mathcal A}$, and similarly for ${\mathcal B}$ and ${\mathcal D}$. Even more, we set
\[
S_1(\o) = \{\phi \in S({\mathcal D}): \phi(\o) = 1\} ,
\]
the ``level-1 set of $\o$''. The elements of $S_1(\o)$ are ``definite''
on $\o$ in the sense \cite{KR1} that
for any $d \in {\mathcal D}$ we have
\[
\phi(d\o) = \phi(d) = \phi(\o d) .
\]
Let $\rho_{\mathcal A}$ denote the metric on $S({\mathcal A})$ determined by $L^{\mathcal A}$
by the formula
\begin{equation}
\label{metr}
\rho_{\mathcal A}(\mu, \nu) = \sup\{|\mu(a) - \nu(a)|: L^{\mathcal A}(a) \leq 1\}.
\end{equation}
(Without further conditions on $L_{\mathcal A}$ we must
permit $\rho_{\mathcal A}$ to take the value $+\infty$. Also, it is
not hard to see that the supremum can be taken equally well
just over ${\mathcal L}^1_{\mathcal A}$.)
Define $\rho_{\mathcal B}$ on $S({\mathcal B})$ similarly.
\begin{notation}
We denote by $S_1^{\mathcal A}(\o)$ the restriction of the
elements of $S_1(\o)$ to ${\mathcal A}$. We define $S_1^{\mathcal B}(\o)$
similarly.
\end{notation}
\begin{defn}
\label{height}
Let ${\mathcal A}$ and ${\mathcal B}$ be unital C*-algebras and let
$\Pi = \{{\mathcal D}, \o\}$ be a bridge from ${\mathcal A}$ to ${\mathcal B}$ .
Let $L^{\mathcal A}$ and $L^{\mathcal B}$ be slip-norms on ${\mathcal A}$ and ${\mathcal B}$.
The \emph{height} of the bridge $\Pi$ is given by
\[
\mathrm{height}(\Pi) =
\max\{\mathrm{Haus}_{\rho_{\mathcal A}}(S_1^{\mathcal A}(\o), S({\mathcal A})) , \
\mathrm{Haus}_{\rho_{\mathcal B}}(S_1^{\mathcal B}(\o) , S({\mathcal B}))\} ,
\]
where the Hausdorff distances are with respect to the indicated
metrics (and value $+\infty$ is allowed).
The length of $\Pi$ is then defined by
\[
\mathrm{length}(\Pi) = \max\{\mathrm{reach}(\Pi), \mathrm{height}(\Pi)\} .
\]
\end{defn}
In Section \ref{app1} we will show how to obtain a useful upper bound
on the length of $\Pi$ for our first class of examples.
\section{The second basic class of examples}
\label{basic2}
Our second basic class of examples has the same starting point
as the first class, consisting of $G$, ${\mathcal H}$, $U$ and $P$ as before,
with ${\mathcal B} = {\mathcal L}({\mathcal H})$. But now we will also have a second
irreducible representation. The more concrete class of examples
motivating this situation, but for which we will not need the details,
is that in \cite{R7, R25, R21} in
which $G$ is a compact semi-simple Lie group, $\l$ is
a positive integral weight, and our two representations of $G$
are the representations with highest weights $m\l$ and $n\l$
for positive integers $m$ and $n$, $m \neq n$.
Furthermore, the projections
$P$ are required to be those along highest weight vectors. The
key feature of this situation that we do need to remember here
is that the stability subgroups $H$ for the two projections
coincide.
Accordingly, for our slightly more general situation, we will denote
our two representations by $({\mathcal H}^m, U^m)$ and $({\mathcal H}^n, U^n)$,
where now $m$ and $n$ are just labels. Our two C*-algebras will
be ${\mathcal B}^m = {\mathcal L}({\mathcal H}^m)$ and ${\mathcal B}^n = {\mathcal L}({\mathcal H}^n)$. We will
denote the action of $G$ on these two algebras just by $\a$,
since the context should always make clear which algebra
is being acted on. The
corresponding projections will be $P^m$ and $P^n$. The crucial
assumption that we make is that the stability subgroups of
these two projections coincide. We will denote this common
stability subgroup by $H$ as before.
We construct a bridge from ${\mathcal B}^m$ to ${\mathcal B}^n$ as follows.
We let ${\mathcal A} = C(G/H)$ as in our first class of examples,
and we define ${\mathcal D}$ by
\[
{\mathcal D} = {\mathcal B}^m \otimes {\mathcal A} \otimes {\mathcal B}^n \ = \ C(G/H, {\mathcal B}^m \otimes {\mathcal B}^n) .
\]
We view ${\mathcal B}^m$ as a subalgebra of ${\mathcal D}$ by sending
$b \in {\mathcal B}^m$ to $b \otimes 1_{\mathcal A} \otimes 1_{{\mathcal B}^n}$, and similarly for ${\mathcal B}^n$.
From the many calculations done in \cite{R25} it is
not surprising that we define the pivot, $\o$, to be the function in
$C(G/H, {\mathcal B}^m \otimes {\mathcal B}^n)$ defined
by
\[
\o(x) = \a_x(P^m) \otimes \a_x(P^n) .
\]
We let $L^m$ be the Lip-norm defined on ${\mathcal B}^m$ determined
by the action $\a$ and the length function $\ell$ as in
Section \ref{basic1}, and similarly for $L^n$ on ${\mathcal B}^n$. In
terms of these Lip-norms the length of any bridge from
${\mathcal B}^m$ to ${\mathcal B}^n$ is defined. Thus the length of the bridge
described above is defined. In Section \ref{app2} we will see how
to obtain useful upper bounds on the length of this bridge.
\section{Bridges with conditional expectations}
\label{brce}
We will now seek a somewhat general framework for obtaining
useful estimates for the lengths of bridges such as
those of our two basic classes of examples.
To discover this framework we will explore some properties
of our two basic classes of examples.
We will summarize what we find at the end of this section.
On $G/H$ there is a unique probability measure that is invariant
under left translation by elements of $G$. We denote the
corresponding linear functional on ${\mathcal A} = C(G/H)$ by $\t_{\mathcal A}$,
and sometimes refer to it as the canonical tracial state on ${\mathcal A}$.
On ${\mathcal B} = {\mathcal L}({\mathcal H})$ there is a unique tracial state, which we
denote by $\t_{\mathcal B}$. These combine to form a tracial state,
$\t_{\mathcal D} = \t_{\mathcal A} \otimes \t_{\mathcal B}$ on ${\mathcal D} = {\mathcal A} \otimes {\mathcal B}$. Similarly,
we have the unique tracial
states $\t_m$ and $\t_n$ on ${\mathcal B}^m$ and ${\mathcal B}^n$, which
combine with $\t_{\mathcal A}$ to give a tracial state on
${\mathcal D} = {\mathcal B}^m \otimes {\mathcal A} \otimes {\mathcal B}^n$.
For ${\mathcal D} = {\mathcal A} \otimes {\mathcal B}$, the tracial state $\t_{\mathcal B}$
determines a conditional expectation, $E^{\mathcal A}$,
from ${\mathcal D}$ onto its subalgebra ${\mathcal A}$, defined on
elementary tensors by
\[
E^{\mathcal A}(a \otimes b) = a \t_{\mathcal B}(b)
\]
for any $a \in {\mathcal A}$ and $b \in {\mathcal B}$. (This is an example of
a ``slice map'' as discussed in \cite{Blk2}, where
conditional expectations
are also discussed.) This conditional expectation
has the property that for any $d \in {\mathcal D}$ we have
\[
\t_{\mathcal A}(E^{\mathcal A}(d)) = \t_{\mathcal D}(d) ,
\]
and it is the unique conditional expectation with this property. (See
corollary II.6.10.8 of \cite{Blk2}.)
In the same way the tracial state $\t_{\mathcal A}$
determines a canonical conditional expectation, $E^{\mathcal B}$ from
${\mathcal D}$ onto its subalgebra ${\mathcal B}$.
For the case in
which ${\mathcal D} = {\mathcal B}^m \otimes {\mathcal A} \otimes {\mathcal B}^n$,
the tracial state $\t_{\mathcal A} \otimes \t_n$ on ${\mathcal A} \otimes {\mathcal B}^n$
determines a canonical conditional
expectation, $E^m$, from ${\mathcal D}$ onto ${\mathcal B}^m$ in the same
way as above,
and the tracial state $\t_m \otimes \t_{\mathcal A}$ determines
a canonical conditional expectation, $E^n$, from ${\mathcal D}$ onto
${\mathcal B}^n$ .
These conditional expectations relate well to the pivots of the
bridges. For the case in which ${\mathcal D} = {\mathcal A} \otimes {\mathcal B}$ we find
that for any $F \in {\mathcal D} = C(G/H, {\mathcal B})$ we have
\[
E^{\mathcal A}(F\o)(x) = \t_{\mathcal B}(F(x)\a_x(P)) .
\]
In particular, for any $T \in {\mathcal B}$ we have
\[
E^{\mathcal A}(T\o)(x) = \t_{\mathcal B}(T\a_x(P))
\]
for all $x \in G/H$. Aside from the fact that we are here using the
normalized trace instead of the standard trace on the matrix
algebra ${\mathcal B}$, the right-hand side is exactly the definition of
the Berezin covariant symbol of $T$ that plays such an important
role in \cite{R7, R21} (beginning in section 1 of \cite{R7}),
and that is denoted there by
$\sigma_T$. This indicates that for general ${\mathcal A}$ and ${\mathcal B}$ a
map $b \mapsto E^{\mathcal A}(b\o)$ might be of importance to us.
For our specific first basic class of examples we note
the following favorable properties:
\begin{enumerate}
\item Self-adjointness \quad i.e. $E^{\mathcal A}(F^*\o) = (E^{\mathcal A}(\o F))^*$
for all $F \in {\mathcal D}$.
\item $E^{\mathcal A}(F\o) = E^{\mathcal A}(\o F)$ \quad for all $F \in {\mathcal D}$.
\item Positivity, \quad i.e. if $F \geq 0$ then $E^{\mathcal A}(F\o) \geq 0$.
\item $E^{\mathcal A}(1_{\mathcal D}\o) = r^{-1}1_{\mathcal A}$ \quad where ${\mathcal B}$ is an $r \times r$
matrix algebra .
\end{enumerate}
However, if we consider $E^{\mathcal B}$ instead $E^{\mathcal A}$,
then for any $F \in {\mathcal D}$ we have
\[
E^{\mathcal B}(F\o) = \int_{G/H} F(x)\a_x(P) \ dx ,
\]
and we see that in general properties 1-3 above fail, although property
4 still holds, with the same constant $r$. But if we restrict $F$ to be
any $f \in {\mathcal A}$, we see that properties 1-3 again hold. Even more,
the expression
\[
\int_{G/H} f(x)\a_x(P) \ dx
\]
is, except for normalization of the trace, the formula involved in the Berezin
contravariant symbol that in \cite{R7} is denoted by $\breve \sigma$.
For our second class of examples, in which
${\mathcal D} = {\mathcal B}^m \otimes {\mathcal A} \otimes {\mathcal B}^n$, we find that for
$F \in {\mathcal D} = C(G/H, {\mathcal B}^m \otimes {\mathcal B}^n)$ we have
\[
E^m(F\o) = \int_{G/H} (\iota_{\mathcal A}
\otimes \t_n)(F(x)(\a_x(P^m) \otimes \a_x(P^n))) \ dx .
\]
Again we see that properties 1-3 above are not in general satisfied.
But if we restrict $F$ to be any $T \in {\mathcal B}^n$ then the above
formula becomes
\[
\int_{G/H} \a_x(P^m) \t_n(T\a_x(P^n)) \ dx ,
\]
which up to normalization of the trace
is exactly the second displayed formula in section 3 of \cite{R25}.
It is not difficult to
see that properties 1-3 above are again satisfied under this restriction.
We remark that it is easily seen that the maps
$T \mapsto E^m(T\o)$ from ${\mathcal B}^n$ to ${\mathcal B}^m$ and
$S \mapsto E^n(S\o)$ from ${\mathcal B}^m$ to ${\mathcal B}^n$ are
each other's adjoints when they
are viewed as being between the Hilbert spaces ${\mathcal L}^2({\mathcal B}^m, \t_m)$
and ${\mathcal L}^2({\mathcal B}^n, \t_n)$. A similar statement hold for our first basic
class of examples.
With these observations in mind, we begin to formulate a somewhat
general framework. As before, we assume that we
have two unital C*-algebras
${\mathcal A}$ and ${\mathcal B}$, and a bridge $\Pi = ({\mathcal D}, \o)$ from ${\mathcal A}$ to
${\mathcal B}$. We now require that we are given conditional expectations
$E^{\mathcal A}$ and $E^{\mathcal B}$ from ${\mathcal D}$ onto its subalgebras ${\mathcal A}$ and
${\mathcal B}$. (We do not require that they be associated to any tracial
states.) We require that they relate well to $\o$. To begin with,
we will just require that $\o \geq 0$ so that $\o^{1/2}$ exists.
Then the map
\[
D \mapsto E^{\mathcal A}(\o^{1/2}D\o^{1/2})
\]
from ${\mathcal D}$ to ${\mathcal A}$ is positive.
Once we have slip-norms $L^{\mathcal A}$ and $L^{\mathcal B}$ on ${\mathcal A}$ and ${\mathcal B}$,
we need to require that the conditional expectations are
compatible with these slip-norms. To begin with, we
require that if $L^{\mathcal B}(b) = 0$ for some $b \in {\mathcal B}$
then $L^{\mathcal A}(E^{\mathcal A}(\o^{1/2}b\o^{1/2})) = 0$.
But one of the conditions on a Lip-norm is that it takes
value 0 exactly on the scalar multiples of the identity element,
and the case of Lip-norms is important to us. For Lip-norms
we see that the above requirement implies that
$E^{\mathcal A}(\o) \in {\mathbb C} 1_{\mathcal A}$, and so $E^{\mathcal A}(\o) = r_\o1_{\mathcal A}$
for some positive real number $r_\o$. We require the same
of $E^{\mathcal B}$ with the same real number, so that we require that
\[
E^{\mathcal A}(\o) = r_\o1_{\mathcal A} = E^{\mathcal B}(\o) .
\]
We then define a map, $\Phi^{\mathcal A}$, from ${\mathcal D}$ to ${\mathcal A}$ by
\[
\Phi^{\mathcal A}(d) = r_\o^{-1} E^{\mathcal A}(\o^{1/2}d\o^{1/2}) .
\]
In a similar way we define $\Phi^{\mathcal B}$ from ${\mathcal D}$ to ${\mathcal B}$.
We see that $\Phi^{\mathcal A}$ and $\Phi^{\mathcal B}$ are unital positive maps, and
so are of norm 1 (as seen by composing them with states).
Then the main compatibility requirement that we need
is that for all $b \in {\mathcal B}$ we have
\[
L^{\mathcal A}(\Phi^{\mathcal A}(b)) \leq L^{\mathcal B}(b) ,
\]
and similarly for ${\mathcal A}$ and ${\mathcal B}$ reversed. Notice that this
implies that if $b \in {\mathcal L}^1_{\mathcal B}$ then $\Phi^{\mathcal A}(b) \in {\mathcal L}^1_{\mathcal A}$.
We now show how to obtain an upper bound for the reach of
the bridge $\Pi$ when the above requirements are satisfied.
Let $b \in {\mathcal L}^1_{\mathcal B}$ be given. As an approximation to $\o b$ by
an element of the form $a\o$ for some $a \in {\mathcal L}^1_{\mathcal A}$ we take
$a = \Phi^{\mathcal A}(b)$. It is indeed in ${\mathcal L}^1_{\mathcal A}$ by the
requirements made just above. This prompts us to set
\begin{equation}
\label{gamb1}
\gamma^{\mathcal B} =
\sup\{\|\Phi^{\mathcal A}(b)\o - \o b\|_{\mathcal D} : b \in {\mathcal L}^1_{\mathcal B} \} ,
\end{equation}
and we see that $\o b$ is then in the $\gamma^{\mathcal B}$-neighborhood
of ${\mathcal L}^1_{\mathcal A} \o$. Note that without further assumptions
on $L^{\mathcal B}$ we could have $\gamma^{\mathcal B} = +\infty$.
Interchanging the roles of ${\mathcal A}$ and ${\mathcal B}$, we define $\gamma^{\mathcal A}$
similarly. We then see that
\[
\mathrm{reach}(\Pi) \leq \max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}\} .
\]
We will explain in Sections \ref{app1} and \ref{app2}
why this upper bound is
useful in the context of \cite{R7, R25, R21}.
We now consider the height of $\Pi$. For this we need to consider $S_1(\o)$
as defined in Section \ref{basic1}. Let $\mu \in S({\mathcal A})$. Because
$\Phi^{\mathcal A}$ is positive and unital, its
composition with $\mu$ is in $S({\mathcal D})$. When we
evaluate this composition at $\o$ to see if it is in $S_1(\o)$, we
obtain $\mu(r_\o^{-1} E^{\mathcal A}(\o^2))$, and we need this to equal 1.
Because $\mu(r_\o^{-1} E^{\mathcal A}(\o)) = 1$, it follows
that we need $\mu(r_\o^{-1} E^{\mathcal A}(\o-\o^2)) = 0$. If this is to hold
for all $\mu \in S({\mathcal A})$, we must have $E^{\mathcal A}(\o-\o^2) = 0$.
If $E^{\mathcal A}$ is a faithful conditional expectation, as is true for
our basic examples, then because $\o \geq \o^2$
it follows that $\o^2 = \o$ so that
$\o$ is a projection, as is also true for our basic examples.
These arguments are reversible, and so it is easy to see
that if $\o$ is a projection, then for every $\mu \in S({\mathcal A})$
we obtain an element, $\phi_\mu$, of $S_1(\o)$, defined by
\[
\phi_\mu(d) = \mu(r_\o^{-1} E^{\mathcal A}(\o d\o)) = \mu(\Phi^{\mathcal A}(d)) .
\]
This provides us with a substantial collection of elements
of $S_1(\o)$.
Consequently, since to estimate the height of $\Pi$ we need to
estimate the distance from each $\mu \in S({\mathcal A})$ to
$S_1^{\mathcal A}(\o)$, we can hope that $\phi_\mu$ restricted
to ${\mathcal A}$ is relatively close to $\mu$. Accordingly, for any
$a \in {\mathcal A}$ we
compute
\[
|\mu(a) - \phi_\mu(a)| = |\mu(a) - \mu(\Phi^{\mathcal A}(a))|
\leq \|a - \Phi^{\mathcal A}(a)\| .
\]
Set
\[
\d^{\mathcal A} = \sup \{\|a - \Phi^{\mathcal A}(a)\| : a \in {\mathcal L}^1_{\mathcal A}\}.
\]
Then we see that
\[
\rho_{L^{\mathcal A}}(\mu, \ \phi_\mu |_{\mathcal A}) \leq \d^{\mathcal A} .
\]
We define $\d^{\mathcal B}$ in the same way, and obtain the
corresponding estimate for the distances from elements
of $S({\mathcal B})$ to the restriction of $S_1(\o)$ to ${\mathcal B}$.
In this way we see that
\[
\mathrm{height}(\Pi) \leq \max\{\d^{\mathcal A}, \d^{\mathcal B}\} .
\]
(Notice that $\d^{\mathcal A}$ involves what $\Phi^{\mathcal A}$ does on ${\mathcal A}$,
whereas $\gamma^{\mathcal A}$ involves what $\Phi^{\mathcal B}$ does on ${\mathcal A}$.)
While this bound is natural within this context, it turns out not to be
so useful for our two basic classes of example. In
Proposition \ref{alth} below we will give a different bound that
does turn out to be useful for our basic examples. But perhaps
other examples will arise for which the above bound is useful.
We now summarize the main points discussed in this section.
\begin{defn}
\label{defexp}
Let ${\mathcal A}$ and ${\mathcal B}$ be unital C*-algebras and
let $\Pi=({\mathcal D}, \o)$ be a bridge from
${\mathcal A}$ to ${\mathcal B}$. We say that $\Pi$ is a \emph{bridge with conditional
expectations} if conditional expectations $E^{\mathcal A}$ and $E^{\mathcal B}$ from
${\mathcal D}$ onto ${\mathcal A}$ and ${\mathcal B}$ are specified, satisfying the following
properties:
\begin{enumerate}
\item The conditional expectations are faithful.
\item The pivot $\o$ is a projection.
\item There is a constant, $r_\o$, such that
\[
E^{\mathcal A}(\o) = r_\o1_{\mathcal D} = E^{\mathcal B}(\o) .
\]
\end{enumerate}
For such a bridge with conditional expectations we define
$\Phi^{\mathcal A}$ on ${\mathcal D}$ by
\[
\Phi^{\mathcal A}(d) = r_\o^{-1} E^{\mathcal A}(\o d\o) .
\]
We define $\Phi^{\mathcal B}$ similarly, with the
roles of ${\mathcal A}$ and ${\mathcal B}$ reversed.
We will often write $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ for a bridge
with conditional expectations.
\end{defn}
I should mention here that at present I do not see how the class of
examples considered by {Latr\'emolie\`re} \ that involves
non-commutative tori \cite{Ltr3}
fits into the setting of bridges with conditional expectations, though
I have not studied this matter carefully. It would certainly be
interesting to understand this better. I also do not see how the
general case of ordinary compact metric spaces, as discussed
in theorem 6.6 of \cite{Ltr2}, fits into the setting of
bridges with conditional expectations
\begin{defn}
\label{admis}
With notation as above, let $L^{\mathcal A}$ and
$L^{\mathcal B}$ be slip-norms on ${\mathcal A}$ and ${\mathcal B}$.
We say that a bridge with conditional
expectations $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$
is \emph{admissible for $L^{\mathcal A}$ and
$L^{\mathcal B}$} if
\[
L^{\mathcal A}(\Phi^{\mathcal A}(b)) \leq L^{\mathcal B}(b)
\]
for all $b \in {\mathcal B}$, and
\[
L^{\mathcal B}(\Phi^{\mathcal B}(a)) \leq L^{\mathcal A}(a)
\]
for all $a \in {\mathcal A}$ .
\end{defn}
We define the reach, height and length of a bridge with
conditional expectations $({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ to
be those of the bridge $({\mathcal D}, \o)$.
From the earlier discussion we obtain:
\begin{thm}
\label{thmadmis}
Let $L^{\mathcal A}$ and
$L^{\mathcal B}$ be slip-norms on unital C*-algebras ${\mathcal A}$ and ${\mathcal B}$,
and let $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ be a bridge with conditional expectations from ${\mathcal A}$ to ${\mathcal B}$
that is admissible for $L^{\mathcal A}$ and
$L^{\mathcal B}$. Then
\[
\mathrm{reach}(\Pi) \leq \max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}\},
\]
where
\[
\gamma^{\mathcal A} = \sup\{\|a\o - \o\Phi^{\mathcal B}(a)\|_{\mathcal D} : a \in {\mathcal L}^1_{\mathcal A} \},
\]
and similarly for $\gamma^{\mathcal B}$, while
\[
\mathrm{height}(\Pi) \leq \max\{\d^{\mathcal A}, \d^{\mathcal B}\} ,
\]
where
\[
\d^{\mathcal A} = \sup \{\|a - \Phi^{\mathcal A}(a)\|: a \in {\mathcal L}^1_{\mathcal A}\}
\]
and similarly for $\d^{\mathcal B}$. Consequently
\[
\mathrm{length}(\Pi) \leq \max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}, \d^{\mathcal A}, \d^{\mathcal B}\} .
\]
(Consequently the propinquity between $({\mathcal A}, L^{\mathcal A})$ and $({\mathcal B}, L^{\mathcal B})$,
as defined in \cite{Ltr2}, is no greater than the right-hand side above.)
\end{thm}
We could axiomitize the above situation in terms of just $\Phi^{\mathcal A}$ and
$\Phi^{\mathcal B}$, without requiring that they come from conditional expectations,
but at present I do not know of examples for which this would be
useful. It would not suffice to require that $\Phi^{\mathcal A}$ and
$\Phi^{\mathcal B}$ just be positive (and unital) because for the matricial
case discussed in the next
section they would need to be completely positive.
The following result is very pertinent to our first
class of basic examples.
\begin{prop}
\label{abel} With notation as above, suppose that our bridge
$\Pi$ has the quite special property that $\o$ commutes
with every element of ${\mathcal A}$, or at least that
$E^{\mathcal A}(\o a\o) = E^{\mathcal A}(a\o)$ for all $a \in {\mathcal A}$. Then
$\Phi^{\mathcal A}(a) = a$ for all $a \in {\mathcal A}$. Consequently
$\d^{\mathcal A} = 0$, and the
restriction of $S_1(\o)$ to ${\mathcal A}$ is all of $S({\mathcal A})$.
\end{prop}
\begin{proof}
This depends on the conditional expectation property
of $E^{\mathcal A}$. For $a \in {\mathcal A}$ we have
\[
\Phi^{\mathcal A}(a) = r_\o^{-1}E^{\mathcal A}(a\o)
= ar_\o^{-1}E^{\mathcal A}(\o) = a .
\]
\end{proof}
The following steps might not initially seem useful, but
in Sections \ref{app1} and \ref{app2} we will
see in connection with our basic examples that they are quite
useful. Our
notation is as above. Let $\nu \in S({\mathcal B})$. Then
as seen above, $\nu \circ \Phi^{\mathcal B} \in S_1(\o)$, and so its
restriction to ${\mathcal A}$ is in $S({\mathcal A})$. But then
$\nu \circ \Phi^{\mathcal B} \circ \Phi^{\mathcal A} \in S_1(\o)$. Let us
denote it by $\psi_\nu$. Then the restriction of $\psi_\nu$
to ${\mathcal B}$ can be used as an
approximation to $\nu$ by an element of $S_1(\o)$. Now
for any $b \in {\mathcal B}$ we have
\[
|\nu(b) - \psi_\nu(b)| =
|\nu(b) - (\nu \circ \Phi^{\mathcal B} \circ \Phi^{\mathcal A})(b)|
\leq \|b - \Phi^{\mathcal B}(\Phi^{\mathcal A}(b))\| .
\]
\begin{notation}
\label{bert}
In terms of the above notation we set
\[
\hat\d^{\mathcal B} = \sup \{\|b - \Phi^{\mathcal B}(\Phi^{\mathcal A}(b))\| : b \in {\mathcal L}^1_{\mathcal B}\}.
\]
\end{notation}
We note that $L^{\mathcal B}( \Phi^{\mathcal B}(\Phi^{\mathcal A}(b))) \leq L^{\mathcal B}(b)$ because
of the admissibility requirements of Definition \ref{admis}. It follows that
\[
\rho_{L^{\mathcal B}}(\nu, \ \psi_\mu) \leq \hat \d^{\mathcal B} .
\]
We define $\hat \d^{\mathcal A}$ in the same way, and obtain the
corresponding estimate for the distances from elements
of $S({\mathcal A})$ to the restriction of $S_1(\o)$ to ${\mathcal A}$.
In this way we obtain:
\begin{prop}
\label{alth}
For notation as above,
\[
\mathrm{height}(\Pi) \leq
\max\{\min\{ \d^{\mathcal A}, \hat \d^{\mathcal A}\}, \min\{\d^{\mathcal B}, \hat \d^{\mathcal B}\}\} .
\]
\end{prop}
We will see in Section \ref{app1} that for our first class of
basic examples, $ \Phi^{\mathcal B} \circ \Phi^{\mathcal A}$ is exactly
a term that plays an important role
in \cite{R7, R21}. It is essentially
an ``anti-Berezin-transform''.
\section{The corresponding matricial bridges}
\label{matricial}
Fix a positive integer $q$. We let $M_q$ denote the C*-algebra
of $q \times q$ matrices with complex entries. For any C*-algebra
${\mathcal A}$ we let $M_q({\mathcal A})$ denote the C*-algebra of $q \times q$ matrices with entries in ${\mathcal A}$. We often identify it in the evident way with the C*-algebra
$M_q \otimes {\mathcal A}$.
Let ${\mathcal A}$ and ${\mathcal B}$ be unital C*-algebras, and let $\Pi = ({\mathcal D}, \o)$ be
a bridge from ${\mathcal A}$ to ${\mathcal B}$. Then $M_q({\mathcal A})$ can be viewed as
a subalgebra of $M_q({\mathcal D})$, as can $M_q({\mathcal B})$. Let
$\o_q = 1_q \otimes \o$, where $1_q$ is the identity element
of $M_q$, so $\o_q$ can be viewed as the diagonal
matrix in $M_q({\mathcal D})$ with $\o$ in each diagonal entry.
Then it is easily seen that $\Pi^q = (M_q({\mathcal D}), \o_q)$ is a bridge
from $M_q({\mathcal A})$ to $M_q({\mathcal B})$.
In order to measure the length of $\Pi^q$ we need slip-norms
$L^{\mathcal A}_q$ and $L^{\mathcal B}_q$ on $M_q({\mathcal A})$ and $M_q({\mathcal B})$.
It is reasonable to want these slip-norms to be coherent
in some sense as $q$ varies. The discussion that we will
give just after Theorem \ref{thmht} suggests that the
coherence requirement be that the sequences
$\{L^{\mathcal A}_q\}$ and $\{L^{\mathcal B}_q\}$ form ``matrix slipnorms''.
To explain what
this means, for any positive integers $m$ and $n$ we let
$M_{mn}$ denote the linear space of $m \times n$
matrices with complex entries, equipped with the norm obtained
by viewing such matrices as operators from the Hilbert
space ${\mathbb C}^n$ to the Hilbert space ${\mathbb C}^m$. We then note
that for any $A \in M_n({\mathcal A})$, any $\a \in M_{mn}$,
and any $\b \in M_{nm}$ the
usual matrix product $\a A \b$ is in $M_m({\mathcal A})$. The
following definition, for the case of Lip-norms,
is given in definition 5.1 of \cite{Wuw2} (and see
also \cite{Wuw1, Wuw3, R21, ER}).
\begin{defn}
\label{mtx}
A sequence $\{L^{\mathcal A}_n\}$
is a \emph{matrix slip-norm} for ${\mathcal A}$ if
$L^{\mathcal A}_n$ is a $*$-seminorm (with value $+\infty$ permitted)
on $M_n({\mathcal A})$
for each integer $n \geq 1$, and this family of
seminorms has the following properties:
\begin{enumerate}
\item For any $A \in M_n({\mathcal A})$, any $\a \in M_{mn}$, and
any $\b \in M_{nm}$, we have
\[
L_m^{\mathcal A}(\a A \b) \leq \|\a\|L_n^{\mathcal A}(A)\|\b\|.
\]
\item For any $A \in M_m({\mathcal A})$ and any $C \in M_n({\mathcal A})$
we have
\[
L_{m+n}^{\mathcal A}\left(
\begin{bmatrix}
A & 0 \\
0 & C
\end{bmatrix}
\right)
= \max(L_m^{\mathcal A}(A), L_n^{\mathcal A}(C)) .
\]
\item $L_1^{\mathcal A}$ is a slip-norm.
\end{enumerate}
\end{defn}
We remark that the properties above imply that for $n \geq 2$
the null-space of $L_n^{\mathcal A}$ contains all of $M_n$,
not just the scalar multiples of the identity. This is why
our definition of slip-norms does not require that the
null-space is exactly the scalar multiples of the identity.
Now let $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ be a bridge with
conditional expectations. For any integer $q \geq 1$ set
$E^{\mathcal A}_q = \iota_q \otimes E^{\mathcal A}$, where $\iota_q$
is the identity map from $M_q$ onto itself. Define
$E^{\mathcal B}_q$ similarly. Then it is easily seen that $E^{\mathcal A}_q$
and $E^{\mathcal B}_q$ are faithful conditional expectations from
$M_q({\mathcal D})$ onto its subalgebras $M_q({\mathcal A})$ and $M_q({\mathcal B})$
respectively. Furthermore, $E^{\mathcal A}_q(\o^q)$ is the diagonal
matrix each diagonal entry of which is $E^{\mathcal A}(\o) = r_\o 1_{\mathcal D}$, and
from this we see that $E^{\mathcal A}_q(\o_q) = r_\o 1_{M_q({\mathcal A})}$.
Thus $r_{\o_q} = r_\o$.
It is also clear that $\o_q$ is a projection. Putting this
all together, we obtain:
\begin{prop}
\label{promatx}
Let $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ be a bridge with
conditional expectations from ${\mathcal A}$ to ${\mathcal B}$. Then
\[
\Pi^q = (M_q({\mathcal D}), \o_q, E^{\mathcal A}_q, E^{\mathcal B}_q)
\]
is a bridge with conditional expectations from
$M_q({\mathcal A})$ to $M_q({\mathcal B})$. It has the same constant
$r_\o$ as does $\Pi$.
\end{prop}
We can then set $\Phi^{\mathcal A}_q = \iota_q \otimes \Phi^{\mathcal A}$,
and similarly for $\Phi^{\mathcal B}_q$. Because $\Pi^q$
has the same constant
$r_\o$ as does $\Pi$, we see that for any $D \in M_q({\mathcal D})$
we have
\[
\Phi^{\mathcal A}_q(D) = r_\o^{-1} E^{\mathcal A}_q(\o_q D\o_q) .
\]
Suppose now that ${\mathcal A}$ and ${\mathcal B}$ have matrix slip-norms
$\{L^{\mathcal A}_n\}$ and $\{L^{\mathcal B}_n\}$.
We remark that a matrix slip-norm $\{L^{\mathcal A}_n\}$
is in general not
at all determined by $L^{\mathcal A}_1$. Thus a bridge that is
admissible for $L^{\mathcal A}_1$ as in Definition \ref{admis}
need not relate well to the seminorms $L^{\mathcal A}_n$ for higher $n$.
\begin{defn}
\label{matadm}
With notation as above, let $\{L_n^{\mathcal A}\}$ and
$\{L_n^{\mathcal B}\}$ be matrix slip-norms
on ${\mathcal A}$ and ${\mathcal B}$. We say that
a bridge with conditional expectations $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$
is \emph{admissible for $\{L_n^{\mathcal A}\}$} and
$\{L_n^{\mathcal B}\}$ if for all integers $n \geq 1$ the bridge $\Pi^n$
is admissible for $L_n^{\mathcal A}$ and
$L_n^{\mathcal B}$; that is, for all integers $n \geq 1$
we have
\[
L_n^{\mathcal A}(\Phi_n^{\mathcal A}(B)) \leq L_n^{\mathcal B}(B)
\]
for all $B \in M_n({\mathcal B})$,
where
\[
\Phi_n^{\mathcal A}(B) = r_\o^{-1} E_n^{\mathcal A}(\o_nB\o_n) ,
\]
and similarly with the roles of ${\mathcal A}$ and ${\mathcal B}$ reversed.
\end{defn}
We assume now that $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$
is admissible for $\{L_n^{\mathcal A}\}$ and
$\{L_n^{\mathcal B}\}$.
Since for a fixed integer $q$ the bridge $\Pi^q$ is
admissible for $L_q^{\mathcal A}$ and
$L_q^{\mathcal B}$, the length of $\Pi^q$ is
defined. We now show how to obtain an upper bound for
the length of $\Pi^q$ in terms of the data used in
the previous section to get an upper bound on the
length of $\Pi$.
We consider first the reach of $\Pi^q$. Set, much as earlier,
\[
{\mathcal L}_{\mathcal A}^{1q} = \{A \in M_q({\mathcal A}):
A = A^* \quad \mathrm{and} \quad L^{\mathcal A}_q(A) \leq 1\} ,
\]
and similarly for ${\mathcal L}_{\mathcal B}^{1q}$.
Then the reach of $\Pi^q$ is defined to be
\[
\mathrm{Haus}_{M_q({\mathcal D})}\{{\mathcal L}_{\mathcal A}^{1q} \o_q \ , \ \o_q {\mathcal L}_{\mathcal B}^{1q}\} .
\]
Suppose that $B \in {\mathcal L}_{\mathcal B}^{1q}$. Then $\Phi^{\mathcal A}_q(B) \in {\mathcal L}_{\mathcal A}^{1q}$
by the admissibility requirement. So we want to bound
\[
\|\Phi^{\mathcal A}_q(B)\o_q - \o_q B\|_{M_q({\mathcal D})} .
\]
I don't see any better way to bound this in terms of the data used
in Theorem \ref{thmadmis} for $\Pi$ than by using an entry-wise estimate
as done in the third paragraph before
lemma 14.2 of \cite{R21}). We use the fact that
for a $q \times q$ matrix
$C = [c_{jk}]$ with entries in a C*-algebra
we have $\|C\| \leq q \max_{jk}\{\|c_{jk}\|\}$ (as is
seen by expressing $C$ as the sum of the $q$ matrices whose
only non-zero entries are the entries $c_{jk}$ for which $j-k$
is a given constant mod $q$). In this way,
for $B \in M_q({\mathcal B})$ with $B = [b_{jk}]$ we find that the last
displayed term above is
\[
\leq q \max_{jk}\{\|\Phi^{\mathcal A}(b_{jk})\o - \o b_{jk}\|\} .
\]
The small difficulty is that the $b_{jk}$'s need not be
self-adjoint. But for any $b \in {\mathcal B}$, if we denote its
real and imaginary parts by $b_r$ and $b_i$, then
because $L^{\mathcal B}$ is a $*$-seminorm it follows that
$L^{\mathcal B}(b_r) \leq L^{\mathcal B}(b)$ and similarly for $b_i$.
Consequently
\begin{eqnarray*}
\|\Phi^{\mathcal A}(b)\o - \o b\|
\leq \|\Phi^{\mathcal A}(b_r)\o - \o b_r\| + \|\Phi^{\mathcal A}(b_i)\o - \o b_i\| \\
\leq \gamma^{\mathcal B} L^{\mathcal B}(b_r) + \gamma^{\mathcal B} L^{\mathcal B}(b_i) \leq 2\gamma^{\mathcal B} L^{\mathcal B}(b).
\end{eqnarray*}
Thus the term displayed just before is
\[
\leq 2q\gamma^{\mathcal B} L^{\mathcal B}(b_{jk}) .
\]
But
$\{L^{\mathcal B}_n\}$ is a matrix slip-norm, and by the
first property of such seminorms given in Definition \ref{mtx}, we have
\[
\max \{L^{\mathcal B}(b_{jk})\} \leq L^{\mathcal B}_q(B) .
\]
Thus for $B \in {\mathcal L}_{\mathcal B}^{1q}$ we see that
\[
\|\Phi^{\mathcal A}_q(B)\o_q - \o_q B\|_{M_q({\mathcal D})} \leq 2q\gamma^{\mathcal B},
\]
so that $\o_q B$ is in the $2q \gamma^{\mathcal B}$-neighborhood of
$ {\mathcal L}_{\mathcal A}^{1q} \o_q$. In the same way
$A\o_q $ is in the $2q \gamma^{\mathcal A}$-neighborhood of
$\o_q {\mathcal L}_{\mathcal B}^{1q}$ for every $A \in {\mathcal L}_{\mathcal A}^{1q}$.
We find in this way that
\[
\mathrm{reach}(\Pi^q) \ \leq \ 2q \max \{ \gamma^{\mathcal A}, \gamma^{\mathcal B}\} .
\]
We now consider the height of $\Pi^q$. We argue
much as in the discussion of height before Definition \ref{defexp}.
For any $\mu \in S(M_q({\mathcal A}))$
its composition with $\Phi^{\mathcal A}_q$ is an element, $\phi_\mu$, of $S_1(\o_q)$,
specifically defined by
\[
\phi_\mu(D) = \mu(\Phi^{\mathcal A}_q(D)) = \mu(r_\o^{-1} E^{\mathcal A}_q(\o_q D\o_q)) .
\]
We take $\phi_\mu |_{M_q({\mathcal A})}$ as an approximation to $\mu$,
and estimate
the distance between these elements of $S(M_q({\mathcal A}))$.
For $A \in {\mathcal L}_{\mathcal A}^{1q}$ we calculate
\[
|\mu(A) - \phi_\mu(A)| = |\mu(A) - \mu(\Phi^{\mathcal A}_q(A))|
\leq \|A - \Phi^{\mathcal A}_q(A)\| .
\]
Again I don't see any better way to bound this in terms of the data used
in Theorem \ref{thmadmis} for $\Pi$ than by using an entry-wise estimates.
For $A \in M_q({\mathcal A})$ with $A = [a_{jk}]$ we find (by using arguments
as above to deal with the fact that the $a_{jk}$'s need
not be self-adjoint) that the last
displayed term above is
\[
\leq q \max_{jk}\{\|a_{jk} - \Phi^{\mathcal A}(a_{jk})\|\}
\leq 2q \d^{\mathcal A} \max \{L^{\mathcal A}(a_{jk})\} .
\]
But again
$\{L^{\mathcal A}_n\}$ is a matrix slip-norm, and so by the
first property of such seminorms given in Definition \ref{mtx} we have
\[
\max \{L^{\mathcal A}(a_{jk})\} \leq L^{\mathcal A}_q(A) .
\]
Since our assumption is that $A \in {\mathcal L}_{\mathcal A}^{1q}$,
we see in this way that
\[
\rho_{L^{\mathcal A}_q}(\mu, \ \phi_\mu |_{\mathcal A}) \leq 2q\d^{\mathcal A} .
\]
Thus $S(M_q({\mathcal A}))$ is in the $2q\d^{\mathcal A}$-neighborhood of the restriction
to $M_q({\mathcal A})$ of $S_1(\o_q)$.
We find in the same way that $S(M_q({\mathcal B}))$ is in the $2q\d^{\mathcal B}$-neighborhood of the restriction
to $M_q({\mathcal B})$ of $S_1(\o_q)$. Consequently,
\[
\mathrm{height}(\Pi^q) \ \leq \ 2q \max \{ \d^{\mathcal A}, \d^{\mathcal B}\} .
\]
We can instead use $\hat \d^{\mathcal A}$ in the way done in Proposition
\ref{alth}. Using reasoning much like that used above, we find
that for any $B \in M_q({\mathcal B})$ we have:
\begin{eqnarray*}
\|B-\Phi^{\mathcal B}_q(\Phi^{\mathcal A}_q(B))\| \leq q \max_{jk}\{\|b_{jk} - \Phi^{\mathcal B}(\Phi^{\mathcal A}(b_{jk}))\| \\
\leq 2q \hat\d^{\mathcal B} \max\{L^{\mathcal B}(b_{jk})\} \leq 2q\hat\d^{\mathcal B} L^{\mathcal B}_q(B)
\end{eqnarray*}
Consequently we see that
\[
\mathrm{height}(\Pi^q) \leq 2q\max\{\hat\d^{\mathcal A}, \hat\d^{\mathcal B}\} .
\]
We summarize what we have found by:
\begin{thm}
\label{thmmatad}
Let $\{L^{\mathcal A}_n\}$ and
$\{L^{\mathcal B}_n\}$ be matrix slip-norms
on unital C*-algebras ${\mathcal A}$ and ${\mathcal B}$,
and let $\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ be a bridge with conditional expectations
from ${\mathcal A}$ to ${\mathcal B}$ that is admissible for $\{L^{\mathcal A}_n\}$ and
$\{L^{\mathcal B}_n\}$. For any fixed positive integer $q$ let $\Pi^q$ be the
corresponding bridge with conditional expectations
from $M_q({\mathcal A})$ to $M_q({\mathcal B})$. Then
\[
\mathrm{reach}(\Pi^q) \leq 2q\max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}\},
\]
where as before
\[
\gamma^{\mathcal A} =
\sup\{\|a\o - \o\Phi^{\mathcal B}(a)\|_{\mathcal D} : a \in {\mathcal L}^1_{\mathcal A} \},
\]
and similarly for $\gamma^{\mathcal B}$; while
\[
\mathrm{height}(\Pi^q) \leq
2q\max\{ \min\{\d^{\mathcal A}, \hat\d^{\mathcal A}\}, \min\{\d^{\mathcal B}\, \hat\d^{\mathcal B}\}\} ,
\]
where as before
\[
\d^{\mathcal A} = \sup \{\|\Phi^{\mathcal A}(a) - a\|: a \in {\mathcal L}^1_{\mathcal A}\}
\]
and
\[
\hat\d^{\mathcal A} = \sup \{\|a - \Phi^{\mathcal A}(\Phi^{\mathcal B}(a))\| : a \in {\mathcal L}^1_{\mathcal A}\},
\]
and similarly for $\d^{\mathcal B}$ and $\hat \d^{\mathcal B}$.
Consequently
\[
\mathrm{length}(\Pi^q) \leq
2q\max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}, \min\{\d^{\mathcal A}, \hat\d^{\mathcal A}\}, \min\{\d^{\mathcal B}\, \hat\d^{\mathcal B}\}\} .
\]
\end{thm}
\section{The application to the first class of basic examples}
\label{app1}
We now apply the above general considerations to our first
class of basic examples, described in Section \ref{basic1}.
Thus we have $G$, ${\mathcal H}$, ${\mathcal B}$, $P$, $H$, ${\mathcal A}$, ${\mathcal D}$
and $\o$ as defined there, as well as $L^{\mathcal A}$ and $L^{\mathcal B}$
given by equation \eqref{lipn}.
We proceed to obtain an upper bound for the length of
the bridge $\Pi = ({\mathcal D}, \o)$, where ${\mathcal D} = C(G/H, {\mathcal B})$
and $\o(x) = \a_x(P)$. We begin by considering
its reach.
As seen in Section \ref{brce} ,
for any $F \in {\mathcal D} = C(G/H, {\mathcal B})$ we have
\[
E^{\mathcal A}(\o F\o)(x) = \t_{\mathcal B}(F(x)\a_x(P)) .
\]
From this it is easily seen that $r_\o^{-1}$
is the dimension of ${\mathcal H}$, and so $r_\o^{-1}\t_{\mathcal B}$
is the usual unnormalized trace on ${\mathcal B}$, which
we now denote by ${\mathrm{tr}}_{\mathcal B}$.
In particular, for any $T \in {\mathcal B}$ we have
\begin{equation}
\Phi^{\mathcal A}(T)(x) = r_\o^{-1}E^{\mathcal A}(\o T\o)(x) = \mathrm{tr}_{\mathcal B}(T\a_x(P))
\label{phia}
\end{equation}
for all $x \in G/H$. But this is exactly the covariant Berezin symbol of $T$ (for this general context)
as defined early in section 1 of \cite{R7}
and denoted there by $\sigma_T$.
It is natural to put on ${\mathcal D}$ the action $\l\otimes \a$ of $G$. One
then easily checks that $\Phi^{\mathcal A}$ is equivariant for
$\l\otimes \a$ and $\l$. From this
it is easy to verify that
\[
L^{\mathcal A}(\Phi^{\mathcal A}(T)) = L^{\mathcal A}(\sigma_T) \leq L^{\mathcal B}(T)
\]
for all $T \in {\mathcal B}$, which is exactly the content of
proposition 1.1 of \cite{R7}. Thus that part of
admissibility is satisfied.
Now
\[
(\Phi^{\mathcal A}(T)\o - \o T)(x) = \a_x(P)(\sigma_T(x)1_{\mathcal B} - T).
\]
Consequently
\begin{equation*}
\|\Phi^{\mathcal A}(T)\o - \o T\|_{\mathcal D} = \sup\{\|\a_x(P)(\sigma_T(x)1_{\mathcal B} - T)\|_{\mathcal B}: x \in G/H\} .
\end{equation*}
As discussed in the text before proposition 8.2 of \cite{R21}, by
equivariance this is
\[
= \sup\{\|P(\sigma_{\a_x(T)}1_{\mathcal B} - \a_x(T))\|_{\mathcal B} : x \in G\} .
\]
Then because $\a_x$ is isometric on ${\mathcal B}$ for $L^{\mathcal B}$
(as well as for the norm),
we find for our present example
that $\gamma^{\mathcal B}$, as defined in equation \eqref{gamb1}, is given by
\begin{IEEEeqnarray}{rCl}
\label{gamb}
\gamma^{\mathcal B} & = & \sup\{\|\Phi^{\mathcal A}(T)\o - \o T\|_{\mathcal D} : T \in {\mathcal L}^1_{\mathcal B} \} \\
& = & \sup\{\|P(\mathrm{tr}(PT) 1_{\mathcal B} - T)\|_{\mathcal B} : T \in {\mathcal L}^1_{\mathcal B} \} .
\nonumber
\end{IEEEeqnarray}
This last term is exactly
the definition of $\gamma^{\mathcal B}$ given
in proposition 8.2 of \cite{R21}.
We next consider $\gamma^{\mathcal A}$. For any $f \in {\mathcal A}$ we have
\begin{equation}
\Phi^{\mathcal B}(f) = r_\o^{-1}E^{\mathcal B}(f)
= d_{\mathcal H}\int_{G/H} f(x)\a_x(P) \ dx .
\label{phib}
\end{equation}
where $d_{\mathcal H} = \mathrm{dim}({\mathcal H})$.
But this is exactly the formula used for the Berezin contravariant
symbol, as indicated in Section \ref{brce}. Early in section 2 of \cite{R7} this $\Phi^{\mathcal B}$ is denoted by
$\breve \sigma _f$, that is,
\begin{equation}
\label{brsig}
\Phi^{\mathcal B}(f) = \breve \sigma_f
\end{equation}
for the present class of examples. One
easily checks that $\Phi^{\mathcal B}$ is equivariant for
$\l\otimes \a$ and $\a$. From this
it is easy to verify that
\[
L^{\mathcal B}(\Phi^{\mathcal B}(f)) = L^{\mathcal B}(\breve \sigma_f) \leq L^{\mathcal A}(f)
\]
for all $f \in {\mathcal A}$, as shown in section 2 of \cite{R7}.
Thus we obtain:
\begin{prop}
\label{adm1}
The bridge with conditional expectations \\
$\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ is admissible for
$L^{\mathcal A}$ and $L^{\mathcal B}$.
\end{prop}
Much as in the statement of proposition 8.1 of \cite{R21} set
\begin{equation}
\label{gama}
\breve\gamma^{\mathcal A} = d_{\mathcal H} \int \rho_{G/H}(e, y)\|P\a_y(P)\| dy .
\end{equation}
In the proof of proposition 8.1 of \cite{R21}
(given just before the statement of the proposition,
and where $\breve\gamma^{\mathcal A}$ is denoted
just by $\gamma^{\mathcal A}$) it is shown, with different notation,
that
\begin{equation}
\|f\o - \o \breve \sigma_f\| \leq \breve\gamma^{\mathcal A} L^{\mathcal A}(f) .
\label{gamin}
\end{equation}
Thus if $f \in {\mathcal L}^1_{\mathcal A}$ then
\[
\|f\o - \o \Phi^{\mathcal B}(f)\| \leq \breve\gamma^{\mathcal A} .
\]
It follows that $\gamma^{\mathcal A} \leq \breve\gamma^{\mathcal A} $. We
have thus obtained:
\begin{prop}
\label{prore1}
For the present class of examples, with notation as above,
we have
\[
\mathrm{reach}(\Pi) \leq
\max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}\} \leq \max\{\breve\gamma^{\mathcal A}, \gamma^{\mathcal B}\}
\]
where $\breve\gamma^{\mathcal A}$ is defined in equation \ref{gama}
and $\gamma^{\mathcal B}$ is defined in equation \ref{gamb}
(and \ref{gamb1}) above.
\end{prop}
Even more, for the case mentioned
at the beginning of Section \ref{basic2} in which $G$ is a
compact semisimple Lie group and
$\l$ is
a positive integral weight, for each positive integer $m$
let $({\mathcal H}^m, U^m)$ be the irreducible representation of $G$
with highest weight $m\l$. Then let ${\mathcal B}^m = {\mathcal L}({\mathcal H}^m)$
with action $\a$ of $G$, and let $P^m$ be the projection on the
highest weight vector in ${\mathcal H}^m$. All the $P^m$'s will have
the same $\a$-stability group, $H$. As before, we let
${\mathcal A} = C(G/H)$. Then for each $m$ we can construct
as in Section \ref{brce} the bridge with conditional
expectations, $\Pi_m = ({\mathcal D}_m, \o^m, E^{\mathcal A}_m, E^{\mathcal B}_m)$.
From a fixed length function $\ell$ on $G$ we will obtain
Lip-norms $\{L^{{\mathcal B}^m}\}$ which together
with $\{L^{{\mathcal A}}\}$ give meaning to the lengths of the bridges
$\Pi_m$. In turn the constants $\gamma^{\mathcal A}_m$, $\breve \gamma^{\mathcal A}_m$,
$\gamma^{{\mathcal B}^m}$, $\d^{\mathcal A}_m$, $\d^{{\mathcal B}^m}$ will be defined.
Now it follows from the discussion of $\breve\gamma^{\mathcal A}$
above that $\gamma^{\mathcal A}_m \leq \breve\gamma^{\mathcal A}_m$ for each $m$.
But section 10 of \cite{R21} gives a proof
that the sequence
$\breve \gamma^{\mathcal A}_m$ converges to 0 as $m$ goes to $\infty$.
It follows that $\gamma^{\mathcal A}_m$
converges to 0 as $m$ goes to $\infty$.
Then section 12 of \cite{R21} gives a proof that the sequence
$\gamma^{{\mathcal B}^m}$ converges to 0 as $m$ goes to $\infty$.
Putting together these results for $\gamma^{\mathcal A}_m$
and $\gamma^{{\mathcal B}^m}$, we obtain:
\begin{prop}
\label{progam}
The reach of the bridge $\Pi_m$ goes to 0 as $m$ goes to $\infty$.
\end{prop}
We now consider the height of $\Pi$. For $\d^{\mathcal A}$ something
quite special happens. It is easily seen that ${\mathcal A} = C(G/H)$ is
the center of ${\mathcal D} = {\mathcal A} \otimes {\mathcal B}$, and so all
elements of ${\mathcal A}$ commute with $\o$. Thus we can
apply Proposition \ref{abel} to conclude that
$\d^{\mathcal A} = 0$.
In order to deal with ${\mathcal B}$ we use $\hat \d^{\mathcal B}$ of
Notation \ref{bert} and the discussion surrounding it.
For any $T \in {\mathcal B}$ we have
\begin{align*}
\Phi^{\mathcal B}(\Phi^{\mathcal A}(T)) &= r_\o^{-1}E^{\mathcal B}(\o (r_\o^{-1}E^{\mathcal A}(\o T \o)) \o) \\
&= d_{\mathcal H} \int \a_x(P) (d_{\mathcal H} \t_{\mathcal B}(\a_x(P)T\a_x(P))\a_x(P) dx \\
&= d_{\mathcal H} \int \a_x(P) (tr_{\mathcal B}(\a_x(P)T) dx = \breve \sigma(\sigma_T) .
\end{align*}
where for the last term we use notation
from \cite{R7,R21}. The term $\breve \sigma(\sigma_T)$ plays an important
role there. See theorem 6.1 of \cite{R7}
and theorem 11.5 of \cite{R21}.
The $\hat \d^{\mathcal B}$ of our Notation \ref{bert} is for the present
class of examples exactly the $\d^{\mathcal B}$
of notation 8.4 of \cite{R21}. For use in the next section we
here denote it by $\tilde \d^{\mathcal B}$, that is:
\begin{notation}
\label{tdel}
For $G$, ${\mathcal A} = C(G/H)$, ${\mathcal B} = {\mathcal L}({\mathcal H})$,
$\sigma$, $\breve \sigma$, etc as above,
we set
\[
\tilde \d^{\mathcal B} = \sup \{\|T - \breve \sigma(\sigma_T)\| : T \in {\mathcal L}^1_{\mathcal B}\}.
\]
\end{notation}
When we combine this with Propositions \ref{alth} and \ref{prore1} we
obtain:
\begin{thm}
\label{thmprod}
For the present class of examples,
with notation as above, we have
\[
\mathrm{height}(\Pi) \leq \tilde \d^{\mathcal B} .
\]
Consequently
\[
\mathrm{length}(\Pi) \leq \max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}, \min\{\d^{\mathcal B}, \tilde \d^{\mathcal B}\}
\leq \max\{\breve\gamma^{\mathcal A}, \gamma^{\mathcal B}, \min\{\d^{\mathcal B}, \tilde \d^{\mathcal B}\} \} .
\]
\end{thm}
We will indicate in Section \ref{trek} {Latr\'emolie\`re}'s definition of his
propinquity between compact quantum metric spaces, but
it is always no larger than the length of any bridge between
the two spaces. He denotes his propinquity simply by $\L$,
but we will denote it here by ``Prpq''. Consequently,
from the above theorem we obtain:
\begin{cor}
\label{corpr}
With notation as above,
\[
\mathrm{Prpq}(({\mathcal A}, L^{\mathcal A}), \ ({\mathcal B}, L^{\mathcal B})) \leq
\max\{\breve\gamma^{\mathcal A}, \gamma^{\mathcal B}, \min\{\d^{\mathcal B}, \tilde \d^{\mathcal B}\} \} .
\]
\end{cor}
For
the case of highest weight representations discussed just
above, theorem 11.5 of \cite{R21} gives a proof
that the sequence
$\tilde \d^{\mathcal B}_m$ (in our notation)
converges to 0 as $m$ goes to $\infty$.
It follows from the above proposition that:
\begin{prop}
\label{prodel}
The height of the bridge $\Pi_m$ goes to 0 as $m$ goes to $\infty$.
\end{prop}
Combining this with Proposition \ref{progam}, we obtain:
\begin{thm}
\label{thmht}
The length of the bridge $\Pi_m$ goes to 0 as $m$ goes to $\infty$.
Consequently $\mathrm{Prpq}(({\mathcal A}, L^{\mathcal A}), \ ({\mathcal B}^m, L^{{\mathcal B}^m}))$
goes to 0 as $m$ goes to $\infty$.
\end{thm}
We now treat the matricial case, beginning with the general
situation in which $G$ is some compact group. We must
first specify our matrix slip-norms. This is essentially done
in example 3.2 of \cite{Wuw2} and section 14 of \cite{R21}.
As discussed in Section \ref{basic1}, we have the
actions $\l$ and $\a$ on ${\mathcal A} = C(G/H)$ and ${\mathcal B} = {\mathcal B}({\mathcal H})$
respectively. For any $n$ let $\l^n$ and $\a^n$ be the
corresponding actions $\iota_n\otimes \l$ and $\iota_n \otimes \a$
on $M_n\otimes {\mathcal A} = M_n({\mathcal A})$ and $M_n\otimes {\mathcal B} = M_n({\mathcal B})$.
We then use the length function $\ell$ and formula \ref{lipn} to
define seminorms $L^{\mathcal A}_n$ and $L^{\mathcal B}_n$ on $M_n({\mathcal A})$
and $M_n({\mathcal B})$. It is easily verified that $\{L^{\mathcal A}_n\}$ and
$\{L^{\mathcal B}_n\}$ are matrix slip-norms. Notice that here $L_1^{\mathcal A} = L^{\mathcal A}$
and $L_1^{\mathcal B} = L^{\mathcal B}$ are actually lipnorms, and so, by property
1 of Definition \ref{mtx}, for each $n$ the null-spaces of $L^{\mathcal A}_n$ and
$L^{\mathcal B}_n$ are exactly $M_n$.
Now fix $q$ and take $n=q$. From our bridge with conditional
expectations
$\Pi = ({\mathcal D}, \o, E^{\mathcal A}, E^{\mathcal B})$ we define the bridge
with conditional expectations
$\Pi^q = (M_q({\mathcal D}), \o_q, E_q^{\mathcal A}, E_q^{\mathcal B})$
between $M_q({\mathcal A})$ and $M_q({\mathcal B})$ in the
way done in Proposition \ref{promatx}. We then define
$\Phi^{\mathcal A}_q = \iota_q \otimes \Phi^{\mathcal A}$,
and similarly for $\Phi^{\mathcal B}_q$ as done right after
Proposition \ref{promatx}.
Because $\l$ and $\a$ and $\Phi^{\mathcal A}$ and $\Phi^{\mathcal B}$
act entry-wise on $M_q({\mathcal D})$, and because $\Phi^{\mathcal A}$ and $\Phi^{\mathcal B}$
are equivariant for $\l\otimes\a$ and $\l$ and for $\l\otimes\a$ and $\a$
respectively, it is easily seen that
$\Pi^q$ is admissible for $\{L^{\mathcal A}_q\}$ and
$\{L^{\mathcal B}_q\}$.
We are thus in position to apply Theorem \ref{thmmatad}.
From it and Theorem \ref{thmprod} we conclude that:
\begin{thm}
\label{thmmat1}
With notation as above, we have
\[
\mathrm{reach}(\Pi^q) \leq 2q\max\{\gamma^{\mathcal A}, \gamma^{\mathcal B}\}
\leq 2q\max\{\breve\gamma^{\mathcal A}, \gamma^{\mathcal B}\},
\]
where $\breve\gamma^{\mathcal A}$ is defined by formula \ref{gama}.
Furthermore
\[
\mathrm{height}(\Pi^q) \leq 2q \min\{\d^{\mathcal B}, \tilde \d^{\mathcal B}\},
\]
where $\tilde \d^{\mathcal B}$ is defined in Notation \ref{tdel}.
Thus
\[
\mathrm{length}(\Pi^q) \leq 2q
\max\{\breve\gamma^{\mathcal A}, \gamma^{\mathcal B}, \min\{\d^{\mathcal B}, \tilde \d^{\mathcal B}\} \} .
\]
\end{thm}
We remark that we could improve slightly on the above theorem by
using a calculation given in section 14 of \cite{R21}
in the middle of the discussion there of Wu's results.
Let $F \in M_q({\mathcal A})$ be given, with $F = \{f_{jk}\}$,
and set
\[
t_{jk} = \breve \sigma_{f_{jk}} = d_{\mathcal H} \int f_{jk}(y)\a_y(P) dy ,
\]
and let $T = \{t_{jk}\}$. Then
\begin{eqnarray*}
(F\o_q - \o_q T)(x) = \{\a_x(P)(f_{jk}(x) - d_{\mathcal H} \int f_{jk}(y)\a_y(P) dy)\} \\
=\{d_{\mathcal H} \int (f_{jk}(x) - f_{jk}(y))\a_x(P)\a_y(P) dy\} .
\end{eqnarray*}
To obtain a bound on $\gamma^{\mathcal A}_q$ we need to take the supremum
of the norm of this expression over all $x$ and over all $F$
with $L^{\mathcal A}_q(F) \leq 1$. By translation by $x$, in the way done
shortly before proposition 8.1 of \cite{R21}, it suffices to
consider
\begin{align*}
\sup\{ \| \{d_{\mathcal H} & \int (f_{jk}(e) - f_{jk}(y))P\a_y(P) dy\} \| \} \\
&\leq d_{\mathcal H} \int \| F(e) - F(y)\| \|P\a_y(P)\| dy \\
&\leq L^{\mathcal A}_q(F) d_{\mathcal H} \int \rho_{G/H}(e, y) \| P \a_y(P)\| dy
= L^{\mathcal A}_q(F) \breve\gamma^{\mathcal A} .
\end{align*}
In this way we see that $\gamma^{\mathcal A}_q \leq \breve \gamma^{\mathcal A}$, with
no factor of $2q$ needed.
We can apply Theorem \ref{thmmat1}
to the situation considered before
Proposition \ref{progam} in which $G$ is a
compact semisimple Lie group, $\l$ is
a positive integral weight, and
$({\mathcal H}^m, U^m)$ is the irreducible representation of $G$
with highest weight $m\l$ for each positive $m$,
with ${\mathcal B}^m = {\mathcal L}({\mathcal H}^m)$. We can then form the bridge with conditional
expectations, $\Pi_m = ({\mathcal D}_m, \o^m, E^{\mathcal A}_m, E^{\mathcal B}_m)$ that
is discussed there. For any positive integer $q$ we then have
the matricial version involving $M_q({\mathcal A})$, $M_q({\mathcal B}_m)$,
and the corresponding bridge $\Pi^q_m$. On applying
Theorem \ref{thmmat1} together with the results mentioned
above about the convergence of the quantities
$\breve\gamma^{\mathcal A}_m, \gamma^{\mathcal B}_m$, and $\tilde \d^{\mathcal B}_m$
to 0, we obtain one of the two main
theorems of this paper:
\begin{thm}
\label{thmmat1q}
With notation as above, we have
\[
\mathrm{length}(\Pi^q_m) \leq 2q
\max\{\breve\gamma^{\mathcal A}_m, \gamma^{\mathcal B}_m, \min\{\d^{\mathcal B}_m, \tilde \d^{\mathcal B}_m\} \}
\]
where $\breve\gamma^{\mathcal A}_m$ is defined as in formula \ref{gama}, and
where $\tilde \d^{\mathcal B}_m$ is defined as in Notation \ref{tdel}.
Consequently $\mathrm{length}(\Pi^q_m)$ converges
to 0 as $m$ goes to $\infty$, for each fixed $q$.
\end{thm}
We remark that because of the factor $q$ in the right-hand
side of the above bound for $\mathrm{length}(\Pi^q_m)$,
we do not obtain convergence to 0 that is uniform in $q$.
I do not have a counter-example to the convergence being
uniform in $q$, but it seems to me very possible that the convergence
will not be uniform.
\section{The application to the second class of basic examples}
\label{app2}
We now apply our general considerations to our second
basic class of examples, described in Section \ref{basic2}.
We use the notation of that section.
We also use much of the notation of Section \ref{app1}, but now
we have two representations, $({\mathcal H}^m, U^m)$ and $({\mathcal H}^n, U^n)$
(where for the moment $m$ and $n$ are just labels). We have
corresponding C*-algebras ${\mathcal B}^m$ and ${\mathcal B}^n$,
and projections $P^m$ and $P^n$.
We let $L^m$ be the Lip-norm defined on ${\mathcal B}^m$ determined
by the action $\a$ and the length function $\ell$ as in
equation \ref{lipn}, and similarly for $L^n$ on ${\mathcal B}^n$. In
terms of these Lip-norms the length of any bridge from
${\mathcal B}^m$ to ${\mathcal B}^n$ is defined.
As in Section \ref{basic2} we consider the bridge $\Pi = ({\mathcal D}, \o)$ for which
\[
{\mathcal D} = {\mathcal B}^m \otimes {\mathcal A} \otimes {\mathcal B}^n \ = \ C(G/H, \ {\mathcal B}^m \otimes {\mathcal B}^n) ,
\]
and the pivot, $\o$, in $C(G/H, \ {\mathcal B}^m \otimes {\mathcal B}^n)$, is defined by
\[
\o(x) = \a_x(P^m) \otimes \a_x(P^n) .
\]
We view ${\mathcal B}^m$ as a subalgebra of ${\mathcal D}$ by sending
$T \in {\mathcal B}^m$ to $T \otimes 1_{\mathcal A} \otimes 1_{{\mathcal B}^n}$, and similarly for
${\mathcal B}^n$.
As seen in Section \ref{brce},
the tracial state $\t_{\mathcal A} \otimes \t_n$ on ${\mathcal A} \otimes {\mathcal B}^n$
determines a canonical conditional
expectation, $E^m$, from ${\mathcal D}$ onto ${\mathcal B}^m$,
and the tracial state $\t_m \otimes \t_{\mathcal A}$ determines
a canonical conditional expectation, $E^n$, from ${\mathcal D}$ onto
${\mathcal B}^n$ .
We find that for any $F \in {\mathcal D}$ we have
\[
E^m(F) = \int_{G/H} (\iota_m
\otimes \t_n)(F(x)) \ dx ,
\]
and similarly for $E^n$, where here $\iota_m$ is the
identity map from ${\mathcal B}^m$ to itself.
From this it is easily seen that
\[
r_\o^{-1} = d_md_n
\]
where $d_m$ is the dimension of ${\mathcal H}^m$ and similarly for $d_n$.
Thus $\Pi = \{{\mathcal D}, \o, E^m, E^n\}$ is a bridge with conditional
expectations.
Then
\[
\Phi^m(F)(x)
= r_\o^{-1}E^m(\o F\o)(x) = d_m d_n E^m(\o F\o)(x) .
\]
But, if we set $\a_x(P^m \otimes P^n) = \a_x(P^m) \otimes \a_x(P^n)$, we have
\[
E^m(\o F\o)(x)
= \int (\iota_m \otimes \t_n)(\a_x(P^m \otimes P^n)F(x)\a_x(P^m \otimes P^n)) dx .
\]
In particular, for any $T \in {\mathcal B}^n$ we have
\[
E^m(\o T\o)(x) = \int \a_x(P^m)\t_n(T\a_x(P^n)) dx ,
\]
and so since $d_n \t_{{\mathcal B}^n}$
is the usual unnormalized trace ${\mathrm{tr}}_n$ on ${\mathcal B}^n$, we have
\[
\Phi^m(T) = d_m \int \a_x(P^m) {\mathrm{tr}}_n(T\a_x(P^n)) dx .
\]
This is essentially the formula obtained in Section \ref{brce}, and
is exactly the second displayed formula in section 3 of \cite{R25}.
Even more, with notation as in
Section \ref{app1}, especially the $\Phi^{\mathcal A}$ of equation \eqref{phia},
except for our different ${\mathcal B}^m$
and ${\mathcal B}^n$ etc, we see that we can write
\begin{equation}
\label{dphi}
\Phi^m(T) = \Phi^{{\mathcal B}^m}(\Phi^{\mathcal A}(T)) = \breve \sigma^m(\sigma^n_T) .
\end{equation}
We have a similar equation for $\Phi^n(T)$, and we see
that we depend on the context to make clear on which
of the two algebras ${\mathcal B}^m$ and
${\mathcal B}^n$ we consider $\Phi^{\mathcal A}$ to be defined.
As in the proof of Proposition \ref{adm1}, we can use
the fact that $E^m$ and $E^n$ are equivariant, where the
action of $G$ on ${\mathcal D}$ is given by $\a \otimes \l \otimes \a$,
to obtain:
\begin{prop}
\label{adm2}
The bridge with
conditional expectations $\Pi$ is admissible for $L^m$ and $L^n$.
\end{prop}
The formula \eqref{dphi} suggests the following steps for obtaining a
bound on the reach of $\Pi$ in terms of the data of
the previous section. Let $S \in {\mathcal B}^m$,
$f \in C(G/H)$, and $T \in {\mathcal B}^n$. Then, for the
norm of ${\mathcal B}^m \otimes {\mathcal B}^m$ and for any $x \in G/H$, we have
\begin{align}
\|(S \o - \o & T)(x)\|
=\|(S\a_x(P^m))\otimes \a_x(P^n) -
\a_x(P^m)\otimes (\a_x(P^n)T)\| \nonumber \\
& \leq \|(S\a_x(P^m))\otimes \a_x(P^n) -
f(x)(\a_x(P^m)\otimes (\a_x(P^n))\| \nonumber \\
& \quad + \|f(x)(\a_x(P^m)\otimes (\a_x(P^n)) -
\a_x(P^m)\otimes (\a_x(P^n)T) \| \nonumber \\
& = \|S\a_x(P^m) - f(x)\a_x(P^m)\| +
\|f(x)\a_x(P^n) - \a_x(P^n)T \| .
\label{bigineq}
\end{align}
Notice that the last two norms are in ${\mathcal B}^m$ and
${\mathcal B}^n$ respectively.
We will also use the $\Phi^{\mathcal B}$ of equation
\eqref{phib}, but now, to distinguish it from the $\Phi^m$
above, we indicate that it is defined on ${\mathcal A}$ (and maps to ${\mathcal B}^m$) by
writing $\Phi_{\mathcal A}^{{\mathcal B}^m}$.
For fixed $T \in {\mathcal B}^n$ let us set
$f(x) = \Phi^{\mathcal A}(T) = {\mathrm{tr}}_n(\a_x(P^n)T)$, and then
let us set $S = \Phi_{\mathcal A}^{{\mathcal B}^m}(f) = d_m\int f(x)\a_x(P^m)$.
Thus $S = \breve \sigma^m_f$ by equation \ref{brsig}.
When we substitute these into the inequality \eqref{bigineq},
we obtain
\begin{align*}
&\|(\Phi_{\mathcal A}^{{\mathcal B}^m}(f)\o - \o T)(x)\| \\ &\leq
\|(\Phi_{\mathcal A}^{{\mathcal B}^m}(f)(x) - f(x))\a_x(P^m)\| +
\|\a_x(P^n)(f(x) -T) \|
\end{align*}
In view of the definition of $f$, we recognize
that the supremum over $x \in G/H$ of
the second term on the right of the inequality sign is
the kind of term involved in the supremum in the
right-hand side of equality \eqref{gamb}.
Consequently that second term above is no greater than
$\gamma^{{\mathcal B}^n}L^n(T)$.
To indicate that this comes from
equality \eqref{gamb} we write $\gamma^{{\mathcal B}^n}_{\mathcal A}$ instead
of just $\gamma^{{\mathcal B}^n}$.
Because $\Phi_{\mathcal A}^{{\mathcal B}^m}(f) = \breve \sigma^m_f$,
we also recognize that the supremum over $x \in G/H$ of
the first term above on the right of the inequality sign is
exactly
(after taking adjoints to get $P^m$ on the correct side)
the left hand side of inequality \eqref{gamin},
where the $\o$ there is that of Section \ref{app1}.
Consequently that term is no greater than
$\breve \gamma^{\mathcal A}_m L^{\mathcal A}(f)$, where the subscript
$m$ on $\breve \gamma^{\mathcal A}_m$ indicates that
$P^m$ should be used in equation \eqref{gama}.
But from the admissibility in Proposition \ref{adm1}
involving $\Phi_{\mathcal A}^{{\mathcal B}^m} = \Phi^{{\mathcal B}^m}$ and $\Phi^{\mathcal A}$ we have
\[
L^m(S) = L^m(\Phi_{\mathcal A}^{{\mathcal B}^m}(f)) \leq L^{\mathcal A}(f) \leq L^n(T) .
\]
Notice that it follows that if $T \in {\mathcal L}^1_{{\mathcal B}^n}$ then
$S \in {\mathcal L}^1_{{\mathcal B}^m}$.
Anyway, on taking the supremum over $x \in G/H$, we obtain
\[
\|\Phi_{\mathcal A}^{{\mathcal B}^m}(f)\o - \o T\|
\leq (\breve \gamma^{\mathcal A}_m \ + \ \gamma^{{\mathcal B}^n}_{\mathcal A}) L^n(T) .
\]
We see in this way
that the distance from $\o T$ to ${\mathcal L}^1_{{\mathcal B}^m} \o$
is no bigger than $\breve \gamma^{\mathcal A}_m + \gamma^{{\mathcal B}^n}_{\mathcal A}$.
The role of ${\mathcal A}$ in Theorem \ref{thmadmis} is here being
played by ${\mathcal B}^m$. So to reduce confusion we will here write
$\gamma^m_n$ for the $\gamma^{\mathcal A}$ of Theorem \ref{thmadmis}, showing
also the dependence on $n$. Thus by definition
\[
\gamma^m_n = \sup\{\|T\o - \o \Phi^n(T)\|: T \in {\mathcal L}^1_{B^m}\}.
\]
We define $\gamma^n_m$ similarly. Then in terms of this notation,
what we have found above is that
\[
\gamma^n_m \leq \breve \gamma^{\mathcal A}_m + \gamma^{{\mathcal B}^n}_{\mathcal A}
\]
We now indicate the dependence of $\Pi$ on $m$ and
$n$ by writing $\Pi_{m,n}$.
The situation just above
is essentially symmetric in $m$ and $n$,
and so, on combining this with the first inequality
of Theorem \ref{thmadmis}, we obtain:
\begin{prop}
\label{prorch2}
With notation as above, we have
\[
\mathrm{reach}(\Pi_{m,n}) \leq \max\{\gamma^n_m, \gamma^m_n\} \leq
\max\{\breve \gamma^{\mathcal A}_m +
\gamma^{{\mathcal B}^n}_{\mathcal A}, \breve \gamma^{\mathcal A}_n + \gamma^{{\mathcal B}^m}_{\mathcal A}\},
\]
\end{prop}
As mentioned in the previous section,
section 10 of \cite{R21} gives a proof
that the sequence
$\breve \gamma^{\mathcal A}_m$ converges to 0 as $m$ goes to $\infty$,
while section 12 of \cite{R21} gives a proof that the sequence
$\gamma^{{\mathcal B}^m}_{\mathcal A}$ converges to 0 as $m$ goes to $\infty$.
We thus see that we obtain:
\begin{prop}
\label{progamn}
The reach of the bridge $\Pi_{m,n}$ goes to 0 as $m$
and $n$ go to $\infty$ simultaneously.
\end{prop}
We now obtain an upper bound for the height of $\Pi_{m,n}$.
For this we will again use Proposition \ref{alth}. We
calculate as follows, using equation \eqref{dphi}.
For $T \in {\mathcal B}^n$ we have
\begin{align*}
\Phi^n(\Phi^m(T)) = \Phi^n(\breve \sigma^m(\sigma^n_T)) =
\breve \sigma^n(\sigma^m(\breve \sigma^m(\sigma^n_T)))
\end{align*}
Thus
\begin{align}
\label{2term}
\|T &- \Phi^n(\Phi^m(T))\| \\
&\leq \|T - \breve \sigma^n(\sigma^n_T)\|
+ \|\breve \sigma^n(\sigma^n_T) -
\breve \sigma^n ((\sigma^m\circ \breve \sigma^m)(\sigma^n_T)\| \nonumber \\
&\leq \tilde \d^{{\mathcal B}^n}_{\mathcal A} L^{{\mathcal B}^n}(T)
+\|\sigma^n_T - \sigma^m(\tilde \sigma^m(\sigma^n_T))\| , \nonumber
\end{align}
where the first term of the last line comes from
Notation \ref{tdel} and we write $\tilde \d^{{\mathcal B}^n}_{\mathcal A}$
for the $ \tilde \d^{{\mathcal B}^n}$ there.
But $\sigma^n_T$ is just an element of ${\mathcal A}$, and
in inequality 11.2 of \cite{R21} it is shown that
for any $f \in {\mathcal A}$ we have
\[
\|f - \sigma^m(\breve \sigma^m(f)\| \leq \tilde\d^{\mathcal A}_m L^{\mathcal A}(f),
\]
where $\tilde\d^{\mathcal A}_m$ is defined in equation 11.1 of \cite{R21}
by
\begin{equation}
\label{dela}
\tilde\d^{\mathcal A}_m =
\int_{G/H} \rho_{G/H}(e, x)d_m {\mathrm{tr}}(P^m \a_x(P^m))\ dx .
\end{equation}
(In equation 11.1 of \cite{R21} $\tilde\d^{\mathcal A}_m$ is denoted
just by $\d^{\mathcal A}_m$. Also, $ \sigma^m \circ\breve \sigma^m$ is,
within our setting, the usual Berezin transform.)
Thus we see that the second
term of the last line of inequality \eqref{2term} is no bigger
than $\tilde \d^{\mathcal A}_m L^{\mathcal A}(\sigma^n_T)$.
But $L^{\mathcal A}(\sigma^n_T) \leq L^{{\mathcal B}^n}(T)$.
From all of this we see that if $T \in {\mathcal L}^1_{{\mathcal B}^n}$ then
\[
\|T - \Phi^n(\Phi^m(T))\| \leq \tilde \d^{{\mathcal B}^n}_{\mathcal A} + \tilde\d^{\mathcal A}_m .
\]
Again, the role of ${\mathcal B}$ in Notation \ref{bert} is being played
here by ${\mathcal B}^n$, and so to reduce confusion we will here
write $\hat \d^n_m$ for the $\hat \d^{\mathcal B}$ of Notation \ref{bert}.
We then see that for for our
present class of examples, that depend on $m$ and $n$,
we have
\[
\hat \d^n_m \leq \tilde \d^{{\mathcal B}^n}_{\mathcal A} + \tilde\d^{\mathcal A}_m .
\]
The situation is essentially symmetric in $m$ and $n$, and so,
combining this with Propositions \ref{alth} and \ref{prorch2},
we obtain:
\begin{thm}
\label{thmht2}
With notation as above, we have
\[
\mathrm{height}(\Pi_{m,n}) \leq \max\{\hat \d^n_m, \hat \d^m_n\}
\leq \max\{\tilde \d^{{\mathcal B}^n}_{\mathcal A} + \tilde\d^{\mathcal A}_m, \
\tilde \d^{{\mathcal B}^m}_{\mathcal A} + \tilde\d^{\mathcal A}_n \} .
\]
Consequently
\[
\mathrm{length}(\Pi_{m,n}) \leq \max\{\breve \gamma^{\mathcal A}_m +
\gamma^{{\mathcal B}^n}_{\mathcal A}, \ \breve \gamma^{\mathcal A}_n + \gamma^{{\mathcal B}^m}_{\mathcal A}, \
\tilde \d^{{\mathcal B}^n}_{\mathcal A} + \tilde\d^{\mathcal A}_m, \
\tilde \d^{{\mathcal B}^m}_{\mathcal A} + \tilde\d^{\mathcal A}_n \} .
\]
\end{thm}
As mentioned in the previous section,
theorem 11.5 of \cite{R21} gives a proof
that the sequence
$\tilde \d^{{\mathcal B}^m}_{\mathcal A}$ (in our notation)
converges to 0 as $m$ goes to $\infty$, while
theorem 3.4 of \cite{R7} shows that the sequence
$\d^{\mathcal A}_m$ (where it was denoted by $\gamma_m$)
converges to 0 as $m$ goes to $\infty$.
Thus when we combine this with Proposition \ref{progamn}
we obtain:
\begin{thm}
\label{thmmn}
The height of the bridge $\Pi_{m,n}$ goes to 0 as $m$
and $n$ go to $\infty$ simultaneously. Consequently
the length of the bridge $\Pi_{m,n}$ goes to 0 as $m$
and $n$ go to $\infty$ simultaneously, and thus
$\mathrm{Prpq}(({\mathcal B}^m, L^m), ({\mathcal B}^n, L^n))$
goes to 0 as $m$
and $n$ go to $\infty$ simultaneously.
\end{thm}
We now consider the matricial case. For any natural
number $q$ we apply the constructions
of Section \ref{matricial} to obtain the bridge with
conditional expectations
\[
\Pi^q_{m,n} = (M_q({\mathcal D}), \o_q, E^m_q, E^n_q)
\]
from $M_q({\mathcal B}^m)$ to $M_q({\mathcal B}^n)$.
From this we then obtain the corresponding maps
$\Phi^m_q$ and $\Phi_q^n$.
We have the actions $\a^q$ of $G$ on $M_q({\mathcal B}^m)$ and
$M_q({\mathcal B}^n)$, much as discussed after Theorem \ref{thmht}.
From these actions and the length function $\ell$ we obtain
the slip-norms $L^m_q$ and $L^n_q$. As $q$ varies, these
result in
matrix slip-norms. One shows that $\Pi^q_{m,n}$ is
admissible for $L^m_q$ and $L^n_q$ by arguing in much
the same way as done after formula \eqref{dphi}.
We are thus in a position to apply Theorem \ref{thmmatad},
as well as the convergence to 0 indicated above
for the various constants,
to obtain the second main
theorem of this paper:
\begin{thm}
\label{thmmat2q}
With notation as above, we have
\[
\mathrm{length}(\Pi^q_{m,n}) \leq 2q \max\{\breve \gamma^{\mathcal A}_m +
\gamma^{{\mathcal B}^n}_{\mathcal A}, \breve \gamma^{\mathcal A}_n + \gamma^{{\mathcal B}^m}_{\mathcal A},
\tilde \d^{{\mathcal B}^n}_{\mathcal A} + \tilde\d^{\mathcal A}_m, \
\tilde \d^{{\mathcal B}^m}_{\mathcal A} + \tilde\d^{\mathcal A}_n \}
\]
where $\breve\gamma^{\mathcal A}_m$ is defined as in formula \ref{gama} while
$\gamma^{{\mathcal B}^m}_{\mathcal A}$ is the $\gamma^{{\mathcal B}^m}$ of equation \eqref{gamb},
and where $\tilde \d^{{\mathcal B}^m}_{\mathcal A} = \tilde \d^{{\mathcal B}^m}$
is defined in Notation \ref{tdel}
while $\tilde\d^{\mathcal A}_m$ is defined by equation \eqref{dela},
and similarly for $n$.
Consequently $\mathrm{length}(\Pi^q_{m,n})$ converges
to 0 as $m$ and $n$ go to $\infty$ simultaneously,
for each fixed $q$.
\end{thm}
\section{Treks}
\label{trek}
{Latr\'emolie\`re} \ defines his propinquity in terms of ``treks''. We will not give here the
precise definition (for which see definition 3.20 of \cite{Ltr2}), but the notion
is quite intuitive. A trek is a finite ``path'' of bridges, so that the ``range''
of the first bridge should be the ``domain'' of the second, etc. The length
of a trek is the sum of the lengths of the bridges in it. The propinquity between
two quantum compact metric spaces is the infimum of the lengths of
all the treks between them. {Latr\'emolie\`re} \ shows in \cite{Ltr2} that propinquity
is a metric on the collection of isometric isomorphism classes of quantum
compact metric spaces. Notably, he proves the striking fact that
if the propinquity between two quantum compact metric spaces is
0 then they are isometrically isomorphic.
There is an evident trek associated with our second class of examples.
In this section we will briefly examine this trek. Let the notation be as
in the early parts of the previous section. Thus we have ${\mathcal A} = C(G/H)$,
and the operator algebras ${\mathcal B}^m$ and ${\mathcal B}^n$. In Section \ref{app1}
we have the bridge $\Pi_m = ({\mathcal A} \otimes {\mathcal B}^m, \o_m)$ from ${\mathcal A}$ to
${\mathcal B}^m$, and the corresponding bridge $\Pi_n$ from ${\mathcal A}$ to
${\mathcal B}^n$. But by reversing the roles of ${\mathcal A}$ and ${\mathcal B}^m$ we obtain
a bridge from ${\mathcal B}^m$ to ${\mathcal A}$. We do this by still
viewing ${\mathcal A}$ and ${\mathcal B}^m$
as subalgebras of ${\mathcal D}_m = C(G/H, {\mathcal B}^m)$, but we now let ${\mathcal A}$
act on the right of ${\mathcal D}_m$ and we let ${\mathcal B}^m$ act on the left. We will
denote this bridge by ${\mathcal D}_m^{-1}$, which is consistent with the
notation of {Latr\'emolie\`re} \ at the beginning of the proof of proposition 4.7
of \cite{Ltr2}. Of course ${\mathcal D}_m^{-1}$ has the ``same'' conditional
expectations $E^{\mathcal A}$ and $E^{{\mathcal B}^m}$ as those of $\Pi_m$.
We will write $E^{\mathcal A}$ as $E^{\mathcal A}_m$ to distinguish it from the
$E^{\mathcal A}$ from ${\mathcal D}_n$, which we will denote by $E^{\mathcal A}_n$.
Then ${\mathcal D}_m^{-1}$ is a bridge with conditional expectations,
which is easily seen to be admissible for $L^{{\mathcal B}^m}$ and
$L^{\mathcal A}$. It is then easily seen that
\[
\mathrm{length}({\mathcal D}_m^{-1}) \ = \ \mathrm{length}({\mathcal D}_m) .
\]
The pair $\Gamma_{m,n} = ({\mathcal D}_m^{-1}, \ {\mathcal D}_n)$ then forms a trek
from ${\mathcal B}^m$ to ${\mathcal B}^n$, and
\[
\mathrm{length}(\Gamma_{m,n}) \ = \ \mathrm{length}({\mathcal D}_m^{-1}) \ + \
\mathrm{length}({\mathcal D}_n) .
\]
From Theorem \ref{thmprod} it follows that
\begin{align*}
\mathrm{length}(\Gamma_{m,n})
&\leq \
\max\{\breve\gamma^{\mathcal A}_m, \gamma^{{\mathcal B}^m}, \min\{\d^{{\mathcal B}^m}, \tilde \d^{{\mathcal B}^m}\} \} \\
&+ \ \max\{\breve\gamma^{\mathcal A}_n, \gamma^{{\mathcal B}^n}, \min\{\d^{{\mathcal B}^n}, \tilde \d^{{\mathcal B}^n}\} \} .
\end{align*}
Note that $\tilde\d_m^{\mathcal A}$ and $\tilde\d_n^{\mathcal A}$ do not appear in the
above expression, in contrast to their appearance in the estimate
in Theorem \ref{thmht2} for $\mathrm{length}(\Pi_{m,n})$. This
opens the possibility that in some cases
$\mathrm{length}(\Gamma_{m,n})$ gives a
smaller bound for Prpq(${\mathcal B}^m, {\mathcal B}^n$) than does
$\mathrm{length}(\Pi_{m,n})$, and, even more, that this might
give examples for which the lengths of certain multi-bridge
treks are strictly smaller that the lengths of any single-bridge
treks. But I have not tried to determine if this happens for
the examples in this paper.
We can view the situation slightly differently as follows. Although
{Latr\'emolie\`re} \ does not mention it, it is natural to define the reach of a
trek as the sum of the reaches of the bridges it contains,
and similarly for the height of a trek. One could then
give a new definition of the length of a trek as simply the
max of its reach and height. This definition is no bigger
that the original definition, and might be smaller. I have not
examined how this might affect the arguments in \cite{Ltr2},
but I imagine that the effect would not be very significant. Anyway,
for the above examples we see from Proposition \ref{prore1} that
we would have
\[
\mathrm{reach}(\Gamma_{m,n}) \ \leq \
\max\{\breve \gamma^{\mathcal A}_m, \gamma_{\mathcal A}^{{\mathcal B}^m}\} +
\max\{\breve \gamma^{\mathcal A}_n, \gamma_{\mathcal A}^{{\mathcal B}^n}\},
\]
so that the bound for $\mathrm{reach}(\Pi_{m,n})$ given in
Proposition \ref{prorch2} is no bigger than that above for
$\mathrm{reach}(\Gamma_{m,n})$. But from Theorem \ref{thmprod}
we see that
\[
\mathrm{height}(\Gamma_{m,n}) \ \leq \ \tilde \d^{{\mathcal B}^m}_{\mathcal A}
+ \tilde \d^{{\mathcal B}^n}_{\mathcal A}
\]
(where $\tilde \d^{{\mathcal B}^m}_{\mathcal A} = \tilde \d^{{\mathcal B}^m}$),
and this can clearly be less than the right-most bound
for $\mathrm{height}(\Pi_{m,n})$
given in Theorem \ref{thmht2}.
\def\leavevmode\hbox to 0pt{\hskip.2ex \accent"16\hss}d{\leavevmode\hbox to 0pt{\hskip.2ex \accent"16\hss}d}
\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{Introduction and preliminaries}
In the papers (\cite{EK1}, \cite{EK2}) Kazhdan and Etingof constructed the equivalence between the categories of quantum groups and Lie bialgebras over $\mathbb{C}[[\hslash]]$. The precise statement is in the following theorem.
\begin{thm}\label{KazEti}
Let $\mathrm{Qgroup}$ be the category of quantum groups, i.e. topologically free cocommutative $\mathrm{mod}\ \hslash$ Hopf algebras over $\mathbb{C}[[\hslash]]$ such that $H/\hslash H$ is a universal enveloping algebra of some Lie algebra $\mathfrak{g}$ over $\mathbb{C}$. Let $\mathrm{LieBialg}$ be the category of topologically free Lie bialgebras over $\mathbb{C}[[\hslash]]$ with $\delta = 0\ \mathrm{mod}\ \hslash$. Then there exists a functor $Lie: \mathrm{Qgroup} \to \mathrm{LieBialg}$ that is an equivalence of categories.
\end{thm}
This theorem can be seen as a quantization of the classical Lie theorem on the equivalence between the categories of simply connected Lie groups and Lie algebras. Lie theorem reduces the problem of classification for (say, semisimple) Lie groups to the same problem for Lie algebras that is much more simple. The same approach works in the quantum group case.
Our goal is to classify quantum groups whose quasiclassical limit is a given semisimple Lie algebra $\mathfrak{g}$. In order to do this we will classify Lie bialgebra structures on $\mathfrak{g}(\mathbb{C}[[\hslash]])$. From Theorem \ref{KazEti} we see that this classification will indeed be a classification of quantum groups.
Any Lie bialgebra structure on $\mathfrak{g}(\mathbb{C}[[\hslash]])$ can be lifted to a Lie bialgebra structure on $\mathfrak{g}(\mathbb{C}((\hslash)))$. Also any Lie bialgebra structure on $\mathfrak{g}(\mathbb{C}((\hslash)))$ defines a Lie bialgebra structure on $\mathfrak{g}(\mathbb{C}[[\hslash]])$ after multiplication by an appropriate power of $\hslash$. So we are left with the problem of classification of Lie bialgebra structures on $\mathfrak{g}(\mathbb{C}((\hslash)))$.
Summarizing the discussion above we have obtained the following: the classification of quantum groups is reduced to the classification of Lie bialgebras over $\mathbb{C}((\hslash))$.
Recall the Belavin-Drinfeld classification:
\begin{thm}[\cite{BD}, \cite{ES}]
Let $\mathfrak{g}$ be a simple Lie algebra over an algebraically closed field of characteristic zero. Then any Lie bialgebra structure on $\mathfrak{g}$ is a coboundary one. Let $r$ be a corresponding $r$-matrix. If $r$ is not skew-symmetric then for some root decomposition we have
$$r = r_0 + \sum _ {\alpha > 0} e_{-\alpha}\otimes e_{\alpha} + \sum_{\alpha \in \mathrm{Span}( \Gamma_1)^{+}}\sum_{k \in \mathbb{N}}e_{-\alpha} \wedge e_{\tau^k(\alpha)} .$$
Here $\Gamma_1, \Gamma_2$ are subsets of the set of simple roots, $\tau: \Gamma_1 \to \Gamma_2$ is an isometric bijection, and for every $\alpha \in \Gamma_1$ there exists $k \in \mathbb{N}$ such that $\tau^k(\alpha) \in \Gamma_1 \setminus \Gamma_2$. The triple $(\Gamma_1, \Gamma_2,\tau)$ is called \emph{admissible}.
The tensor $r_0 \in \mathfrak{h} \otimes \mathfrak{h}$ must satisfy the following two conditions:
(1) $r_0+r_0^{21}=\sum t_k \otimes t_k$, where $t_k$ is an orthonormal basis of $\mathfrak{h}$,
(2) for any $\alpha \in \Gamma_1$ we have $(\tau(\alpha)\otimes \mathrm{id} + \mathrm{id} \otimes \alpha)r_0 = 0 $.
\end{thm}
Our goal is to obtain the version of this theorem over non-closed field.
\subsection{Setting}
In many cases we would prove the results for $B,C$ and $D$ series simultaneously. For this we will introduce some general notation, and then specify it for each case. Note that further in the text we would also make distinction between $D_n$ algebras for odd and even $n$.
By $\mathfrak{g}(F)$ we denote the split simple Lie algebra of rank $n$ over a field $F$ of characteristic zero. We will work with algebras from the $B,C,D$ series. By $G$ we denote a corresponding connected algebraic group, by $H$ a Cartan subgroup of $G$. The simple roots of $\mathfrak{g}$ are denoted $\alpha_1, ..., \alpha_n$. Let $e^{\alpha}$ be the character of $H$ that corresponds to the root $\alpha$. In all cases we will specify the isomorphism of $H$ with the standard torus $(F^*)^n$. Abusing notation, by $(d_1, ..., d_n)$ we denote the element of $H$ that corresponds to an element $(d_1, ..., d_n) \in (F^*)^n $.
We will now specify the choice of matrix representation and root decomposition for the series $B,C,D$. By $M$ we will denote the size of matrices in each matrix representation.\\
{\bf $B_n$ series}
Let $G$ be the subgroup of $SL(2n+1)$ consisting of the matrices satisfying $XBX^T=B$ (we have $M=2n+1$). Here $B$ is the matrix with $1$'s on the anti-diagonal and zeros elsewhere. Then $\mathfrak{g}$ is the algebra of matrices satisfying $XB+BX^T=0$. The Cartan subgroup $H$ (Cartan subalgebra, respectively) can be chosen to be the subset of diagonal matrices of $SL(2n+1)$ (respectively $\mathfrak{sl}(2n+1)$).
Simple root vectors are given by $e_{\alpha_i}=e_{i, i+1}-e_{2n+2-i, 2n+1-i}$. The Cartan subgroup is the subgroup of diagonal matrices of the form $\diag(d_1, ..., d_n, 1, d_n^{-1}, ..., d_1^{-1})=:(d_1, ..., d_n)$. We have $e^{\alpha_i}(d_1, ..., d_n)=d_id_{i+1}^{-1}$ for $i<n$ and $e^{\alpha_n}(d_1, ..., d_n)=d_n$.\\
{\bf $C_n$ series}
Let $G$ be the subgroup of $SL(2n)$ consisting of the matrices satisfying $XBX^T=B$ (we have $M=2n$). Here $B$ is the matrix with $1$'s on the upper half of the anti-diagonal, minus ones on the lower and zeros elsewhere. E.g. for $n=2$
$$B=\begin{pmatrix}
0& 0& 0& 1 \\
0& 0& 1& 0 \\
0& -1& 0& 0\\
-1& 0& 0& 0
\end{pmatrix}.$$
Then $\mathfrak{g}$ is the algebra of matrices satisfying $XB+BX^T=0$. The Cartan subgroup $H$ (Cartan subalgebra, respectively) can be chosen to be the subset of diagonal matrices of $SL(2n)$ (respectively $\mathfrak{sl}(2n)$).
Simple root vectors are given by $e_{\alpha_i}=e_{i, i+1}-e_{2n+2-i, 2n+1-i}$ for $i <n$ and $e_{\alpha_n}=e_{n, n+1}-e_{n+1, n}$. The Cartan subgroup is the subgroup of diagonal matrices of the form $\diag(d_1, ..., d_n, d_n^{-1}, ..., d_1^{-1})=:(d_1, ..., d_n)$. We have $e^{\alpha_i}(d_1, ..., d_n)=d_id_{i+1}^{-1}$ for $i<n$ and $e^{\alpha_n}=d_n^2$.\\
{\bf $D_n$ series}
Let $G$ be the subgroup of $SL(2n)$ consisting of the matrices satisfying $XBX^T=B$ (we have $M=2n$). Here $B$ is the matrix with $1$'s on the anti-diagonal and zeros elsewhere. Then $\mathfrak{g}$ is the algebra of matrices satisfying $XB+BX^T=0$. The Cartan subgroup $H$ (Cartan subalgebra, respectively) can be chosen to be the subset of diagonal matrices of $SL(2n)$ (respectively $\mathfrak{sl}(2n)$).
Simple root vectors are given by $e_{\alpha_i}=e_{i, i+1}-e_{2n+2-i, 2n+1-i}$ for $i <n$ and $e_{\alpha_n}=e_{n-1, n+1}-e_{n+2, n}$. The Cartan subgroup is the subgroup of diagonal matrices of the form $\diag(d_1, ..., d_n, d_n^{-1}, ..., d_1^{-1})=:(d_1, ..., d_n)$. We have $e^{\alpha_i}(d_1, ..., d_n)=d_id_{i+1}^{-1}$ for $i<n$ and $e^{\alpha_n}=d_{n-1}d_n$.
\subsection{Properties of $r$-matrices}
Let $Q: \mathfrak{g} \to \mathfrak{g}^*$ be the natural isomorphism given by Killing form $K$, i.e. $Q(x)(y)=K(x, y)$. Then $\Phi := \mathrm{id} \otimes Q : \mathfrak{g} \otimes \mathfrak{g} \to \mathfrak{g} \otimes \mathfrak{g}^*$ provides an isomorphism between $\mathfrak{g} \otimes \mathfrak{g}$ and $\mathrm{End}(\mathfrak{g})$.
\begin{thm}[\cite{SP}] \label{rstuff}
Let $r_{BD}$ be a non-skewsymmetric r-matrix from the Belavin-Drinfeld list. Denote by $S$ and $N$ the semisimple and nilpotent parts of $\Phi(r_{BD})$. Then $S=\sum_{\alpha>0}e_{-\alpha} \otimes e_{\alpha}+ H$, where $H$ acts by zero on each $\mathfrak{g}_{\alpha}$. The normalizers of the eigenspaces of $S$ with eigenvalues $0$ and $1$ are the Borel subalgebras $\mathfrak{b}_+$ and $\mathfrak{b}_{-}$ respectively. The normalizer of any other eigenspace is the Cartan subalgebra $\mathfrak{h}$.
\end{thm}
\begin{cor}\label{Omega+}
If $r_1, r_2$ are two non-skewsymmetric $r$-matrices, $r_1+r_1^{21}=\alpha \Omega$ and $r_2=r_1 + \beta\Omega$ then $\beta=-\alpha$.
\end{cor}
\begin{proof}
From Theorem \ref{rstuff} we know that normalizers of $0$ and $\alpha$ eigenspaces of $\Phi(r_1)$ are ``big''. Therefore, they must be mapped to the $0$ and $\alpha+2\beta$ eigenspaces of $\Phi(r_2)$ (we have $\Phi(\Omega)=\mathrm{id}$ and $r_2+r_2^{21}=(\alpha+2\beta)\Omega$). On the other hand $\Phi(r_2)=\Phi(r_1)+\beta \mathrm{id}$. The result follows.
\end{proof}
We will frequently work with the following objects.
\begin{defn}
Let $(\Gamma_1, \Gamma_2, \tau)$ be an admissible triple. Then sets of the form
$\alpha, \tau(\alpha), ..., \tau^{k}(\alpha)$, where $\alpha \in \Gamma_1 \setminus \Gamma_2$ and $\tau^{k}(\alpha) \in \Gamma_2\setminus \Gamma_1$ will be called strings of $\tau$.
\end{defn}
\begin{defn}
The \emph{centralizer} $C(r)$ of an $r$-matrix $r$ is the set of all $X \in G(\overline{F})$ such that $\mathrm{Ad}_Xr=r$.
\end{defn}
The centralizer of a matrix from the Belavin-Drinfeld list can be explicitly described.
\begin{thm} [\cite{SP}] \label{CenterinH}
For any simple Lie algebra $\mathfrak{g}$ and for any Belavin-Drinfeld matrix $r_{BD}$ we have $C(r_{BD}) \subset H$. If $(\Gamma_1, \Gamma_2, \tau)$ is an admissible triple corresponding to $r_{BD}$, then $X \in C(r_{BD})$ iff for any root $\alpha \in \Gamma_1 \setminus \Gamma_2$ and for any $k \in \mathbb{N}$ we have $e^{\alpha}(X)=e^{\tau^k(\alpha)}(X)$, i.e. $e^{\alpha}(X)$ is constant on the strings of $\tau$.
\end{thm}
\begin{thm}
Let $r$ be a non-skewsymmetric $r$-matrix that defines a Lie bialgebra structure on $\mathfrak{g}(F)$, $r+r^{21}=a\Omega$. Then $a^2 \in F$.
\end{thm}
\begin{proof}
Recall that a Lie bialgebra structure $\delta$ corresponding to $r$ is given by the formula $\delta(a)=[r, a \otimes 1+ 1 \otimes a]$. Fix an arbitrary $\sigma \in \Gal(\overline{F}/F)$. Consider the $r$-matrix $\sigma(r)$. As $r$ defines a Lie bialgebra structure on $\mathfrak{g}(F)$ we have $\delta(a)=\sigma(\delta(a))$ for all $a \in \mathfrak{g}(F)$. Therefore $\sigma(r)=r+\alpha \Omega$. From Corollary \ref{Omega+} we get $\alpha=0$ or $\alpha=2a$. On the other hand $\sigma(a)\Omega=\sigma(r+r^{21})=\sigma(r)+\sigma(r^{21})=r+r^{21}+2\alpha\Omega=(a+2\alpha) \Omega$. Therefore $\sigma(a)=\pm a$. Thus $\sigma(a^2)=a^2$ for all $\sigma$.
\end{proof}
\subsection{Matrix lemmas}
\begin{lem}\label{blocks}
If for some $X \in GL(n+m, \overline{F})$ and for any $\sigma \in \Gal(\overline{F}/ F)$ we have $X^{-1}\sigma(X)$ is a $n \times n, m \times m$ block matrix, then there exists $Q \in GL(n+m, F)$ and a block matrix $K$ such that $X=QK$.
\end{lem}
\begin{proof}
Let $X= \left( \begin{array}{cc}
A & B \\
C & D \\
\end{array} \right)$. Rearranging rows if necessary, we can assume that $A$ and $D$ are non-degenerate (because of Laplace formula). Then $\sigma(X)=XK$, where $K=\left( \begin{array}{cc}
K_1 & 0 \\
0 & K_2 \\
\end{array} \right) \in GL(n+m, \overline{F})$. Then we have $\sigma(A)=A K_1$ and $\sigma(C)=CK_1$. Since $CA^{-1}$ is $\Gal(\overline{F}/ F)$-stable we have $C=F_1 A$, $F_1 \in GL(m \times n, F)$. Similarly $B=F_2 D$, $F_2 \in GL(n \times m, F)$. Finally $\left( \begin{array}{cc}
A & B \\
C & D \\
\end{array} \right) =
\left( \begin{array}{cc}
I & F_2 \\
F_1 & I \\
\end{array} \right)
\left( \begin{array}{cc}
A& 0 \\
0 & D \\
\end{array} \right)$.
\end{proof}
\begin{cor}\label{diagblocks}
If for some $X \in GL(n, F)$ and for any $\sigma \in \Gal(\overline{F}/ F)$ we have $X^{-1}\sigma(X)$ is diagonal, then $X=QD$ where $D$ is diagonal and $Q \in GL(n, F)$.\hfill $\Box$
\end{cor}
\section{Overview of the paper}
Our main goal is to classify Lie bialgebra structures on $\mathfrak{g}(F)$. We will mostly be interested in the case $F=\mathbb{C}((\hslash))$; however, many general results hold for an arbitrary field of characteristic zero. Because any Lie bialgebra structure on $\mathfrak{g}(F)$ extends to a Lie bialgebra structure on $\mathfrak{g}(\overline{F})$, we have
\begin{thm}
Any Lie bialgebra structure on $\mathfrak{g}(F)$ is a coboundary one given by an $r$-matrix. If $r$ is not skewsymmetric then it has the form $r=\alpha \mathrm{Ad}_X r_{BD}$ where $\alpha \in \overline{F}^*$, $X \in G(\overline{F})$, $r_{BD}$ is an $r$-matrix from the Belavin-Drinfeld list.
\end{thm}
However, not all $r$-matrices of the form $r=\alpha \mathrm{Ad}_X r_{BD}$ define a Lie bialgebra structure on $\mathfrak{g}(F)$.
In what follows we will be interested only in non-skewsymmetric $r$-matrices.
We will classify all (non-skewsymmetic) $r$-matrices that do define a Lie bialgebra structure on $\mathfrak{g}(F)$ up to the following equivalence
\begin{defn}
Two $r$-matrices $r_1$ and $r_2$ are called equivalent if for some $a \in F^*$ and $X \in G(F)$ we have $r_1=a\mathrm{Ad}_X r_2$.
\end{defn}
By the Belavin-Drinfeld classification every $r$-matrix is equivalent to $r_{BD}$ over $\overline{F}$, and $r_{BD}$'s for different triples $(\Gamma_1, \Gamma_2, \tau)$ are not equivalent. However, over $F$ each $\overline{F}$-equivalence class can split into several equivalence classes.
It turns out that if $a \mathrm{Ad}_X r_{BD}$ defines a Lie bialgebra structure on $\mathfrak{g}(F)$ then $a^2 \in F$. Therefore
we have a class of $r$-matrices for each $a^2 \in F^* / (F^*)^2 $. Elements from different classes are nor equivalent, but each class can further split into several equivalence classes. For $F=\mathbb{C}((\hslash))$ we have only two elements in $F^* / (F^*)^2 $: the class of $1$ and the class of $\sqrt{\hslash}$. For every element $u \in F^* / (F^*)^2 $ we will introduce the Belavin-Drinfeld cohomologies that parametrize equivalence classes of $r$-matrices of the form $a \mathrm{Ad}_X r_{BD}$, where $a^2=u$ in $F^* / (F^*)^2$. The cohomologies are called twisted if $u \neq 1$ and non-twisted otherwise.
\subsection{Non-twisted cohomologies}
The following theorem holds:
\begin{thm}[\cite{SP}]
$Ad_X r_{BD}$ defines a Lie bialgebra structure on $\mathfrak{g}(F)$ if and only if for any $\sigma \in \Gal(\overline{F}/F)$ we have $X^{-1}\sigma(X) \in C(r)$.
\end{thm}
With that theorem in mind we can define Belavin-Drinfeld cohomologies without referring to Lie bialgebra structures.
\begin{defn}
$X \in G(\overline{F})$ is called a non-twisted Belavin-Drinfeld cocycle for $r_{BD}$ if for any $\sigma \in \Gal(\overline{F}/F)$ we have $X^{-1}\sigma(X) \in C(r_{BD})$. The set of non-twisted cocycles will be denoted $Z(r_{BD})=Z(G, r_{BD})$.
\end{defn}
In other words $X \in Z(r_{BD})$ if $\mathrm{Ad}_X r_{BD}$ defines a Lie bialgebra structure on $\mathfrak{g}(F)$.
\begin{defn}
Two cocycles $X_1, X_2 \in Z(r_{BD})$ are called equivalent if there exist $Q \in G(F)$ and $C \in C(r_{BD})$ such that $X_1=QX_2C$.
\end{defn}
It is easy to see that $X_1$ and $X_2$ are equivalent if and only if $\mathrm{Ad}_{X_1} r_{BD}$ and $\mathrm{Ad}_{X_2} r_{BD}$ are equivalent as $r$-matrices.
\begin{defn}
The set of equivalence classes of non-twisted cocycles is denoted $H_{BD}^1(r_{BD})=H_{BD}^1(G, r_{BD})$ and is called non-twisted Belavin-Drinfeld cohomologies.
\end{defn}
\begin{rem}
The equivalent way to treat Belavin-Drinfeld cohomology is to understand it as a Galois cohomology of the Galois module $C(r)$ (over the absolute Galois group of $F$). This justifies the use of the word ``cohomology''. Also this approach enables us to put a natural group structure on $H_{BD}^1(r)$. Indeed, for any $X \in Z(r)$ the map $\Gal(\overline{F}/F) \to C(r)$ given by $\sigma \mapsto X^{-1}\sigma(X)$ is a 1-cocycle. If such a map is equivalent to identity then there exists $C \in C(r)$ such that $(XC^{-1})^{-1}\sigma(XC^{-1})=1$ for all $\sigma$. Then one can find $Q \in G(F)$ such that $XC^{-1}=Q$. This means that our equivalence relation is the same as the equivalence of Galois cocycles. We also need to prove that any Galois cocycle is of the given form. We do not know how to prove this in general, however for the algebras of $B,C,D$ types we sketch a proof.
First one observes that by Hilbert's theorem 90 it is always possible to find $Y \in GL(n, \overline{F})$ such that $Y^{-1}\sigma(Y)=X^{-1}\sigma(X) \in C(r)$ for all $\sigma \in \Gal(\overline{F}/F)$. With the suitable matrix representation of the group $G$ we can assume that $C(r)$ is a subset of the set of diagonal matrices. Then by Lemma \ref{blocks} we can assume that $Y$ is diagonal. By looking at the equations that define $H$ inside the set of diagonal matrices we can decompose $Y$ as $Y=QY'$ with $Y' \in C(r)$ and $Q \in \diag(n, F)$. Then $Y'$ is in $C(r)$ and gives rise to the same cocycle as $Y$.
\end{rem}
Computation of $H_{BD}^1(G, r_{BD})$ is possible because of the easy description of $C(r_{BD})$ (see Theorem \ref{CenterinH}).
\subsection{Twisted cohomologies}
Twisted case is somewhat similar to the non-twisted one, however computations become more complicated.
Fix a non-zero $a\in \overline{F}$ such that $a^2 \in F$. We will mostly work with $F=\mathbb{C}((\hslash))$ and $a=\sqrt{\hslash}$. We have the following theorem.
\begin{thm}
$a \mathrm{Ad}_X r_{BD}$ defines a Lie bialgebra structure on $\mathfrak{g}(F)$ if and only if $X$ is a non-twisted cocycle for the field $F[a]$ and $\mathrm{Ad}_{X^{-1}\sigma_0(X)}r_{BD}=r_{BD}^{21}$. Here $\sigma_0$ is the nontrivial element of $\Gal(F[a]/F)$.
\end{thm}
To deal with the condition $\mathrm{Ad}_{X^{-1}\sigma_0(X)}r_{BD}=r_{BD}^{21}$ we will classify all triples $(\Gamma_1, \Gamma_2, \tau)$ such that $r_{BD}^{21}$ and $r_{BD}$ are conjugate. In each case we will find suitable $S \in G(F)$ such that $r_{BD}^{21}=\mathrm{Ad}_S r_{BD}$. Then we can define Belavin-Drinfeld cocycles and cohomologies similar to the non-twisted case. In all cases $S^2=\pm 1$.
\begin{defn}
$X \in G(\overline{F})$ is called the Belavin-Drinfeld twisted cocycle if for any $\sigma \in \Gal(\overline{F}/F[a])$ we have $X^{-1}\sigma(X) \in C(r_{BD})$ and $SX^{-1}\sigma_0(X) \in C(r_{BD})$. The set of Belavin-Drinfeld twisted cocycles is denoted $\overline{Z}(r_{BD})=\overline{Z}(G, r_{BD})$.
\end{defn}
\begin{defn}
Two twisted cocycles $X_1, X_2$ are called equivalent if there exist $Q \in G(F)$ and $C \in C(r_{BD})$ such that $X_1=QX_2C$. The set of equivalence classes of twisted cocycles is called twisted Belavin-Drinfeld cohomologies and is denoted by $\overline{H}_{BD}^1(r_{BD})=\overline{H}_{BD}^1(G, r_{BD})$.
\end{defn}
\begin{prop}
Two $r$-matrices $a\mathrm{Ad}_X r_{BD}$ and $a\mathrm{Ad}_Y r_{BD}$ are conjugate by an element of $G(F)$ if and only if $X$ and $Y$ are equivalent as twisted cocycles. Thus the set of equivalence classes of twisted Lie bialgebra structures is $\overline{H}^{1}_{BD}(G,r)$.
\end{prop}
To find $\overline{H}_{BD}^1(r_{BD}, G)$ we proceed as follows.
First we find a matrix $J$ such that $J \in GL(F[a])$ and $\overline{J}=SJ$ (where $\overline{T}:=\sigma_0(T)$). Then we will show that all Belavin-Drinfeld cocycles are of the form $RJD$ with $R \in GL(F)$ and $D$ diagonal. We will then show that for a diagonal matrix $D$ that satisfies some additional assumptions we can find $R$ such that $RJD$ is a cocycle. We will also see that the equivalence of $R_1JD_1$ and $R_2JD_2$ can be understood purely in terms of $D_1$ and $D_2$. Therefore the problem reduces to classification of classes of diagonal matrices under some equivalence. This will be done explicitly.
\subsection{Results}
Here we summarize the results of the paper.
\begin{table}[H]
\begin{tabular}{|c|c|c|c|}
\hline
Algebra & Triple type & \begin{tabular}[c]{@{}c@{}}$H_{BD}^1$ for \\an arbitrary field\end{tabular} & $H_{BD}^1$ for $\mathbb{C}((\hslash))$ \\ \hline
$B_n$ & & trivial & \\ \hline
$C_n$ & & trivial & \\ \hline
\multirow{2}{*}{$D_n$} & \begin{tabular}[c]{@{}c@{}}there exists a \\ string of $\tau$ that\\ contains $\alpha_{n-1}$\\ and $\alpha_n$\end{tabular} & $F^*/(F^*)^2$ & 2 elements \\ \cline{2-4}
& \begin{tabular}[c]{@{}c@{}}$\alpha_{n-1}$ and $\alpha_n$\\ do not belong to \\ the same string of $\tau$\end{tabular} & trivial & \multicolumn{1}{l|}{} \\ \hline
\end{tabular}
\end{table}
\begin{table}[H]
\begin{tabular}{|c|c|c|c|}
\hline
Algebra & \multicolumn{2}{c|}{Triple type} & $\overline{H}_{BD}^1$ for $\mathbb{C}((\hslash))$ \\ \hline
\multirow{2}{*}{$B_n$} & \multicolumn{2}{c|}{Drinfeld-Jimbo} & one element \\ \cline{2-4}
& \multicolumn{2}{c|}{not DJ} & empty \\ \hline
\multirow{2}{*}{$C_n$} & \multicolumn{2}{c|}{Drinfeld-Jimbo} & one element \\ \cline{2-4}
& \multicolumn{2}{c|}{not DJ} & empty \\ \hline
\multirow{5}{*}{$D_n$} & \multirow{2}{*}{even $n$} & Drinfeld-Jimbo & one element \\ \cline{3-4}
& & not DJ & empty \\ \cline{2-4}
& \multirow{3}{*}{odd $n$} & \begin{tabular}[c]{@{}c@{}}$\Gamma_1=\{\alpha_{n-1}\}$\\ $\tau(\alpha_{n-1})=\alpha_n$;\\ $\Gamma_1=\{\alpha_n\}$\\ $\tau(\alpha_n)=\alpha_{n-1}$;\\ $\Gamma_1=\{\alpha_{n-1}, \alpha_k\}$, $k \neq n$\\ $\tau(\alpha_{n-1})=\alpha_k, \tau(\alpha_k)=\alpha_n$;\\ $\Gamma_1=\{\alpha_{n}, \alpha_k\}$, $k \neq n-1$\\ $\tau(\alpha_{n})=\alpha_k, \tau(\alpha_k)=\alpha_{n-1}$\end{tabular} & two elements \\ \cline{3-4}
& & Drinfeld-Jimbo & one element \\ \cline{3-4}
& & other $r_{BD}$ &empty \\ \hline
\end{tabular}
\end{table}
\section{Non-twisted cohomologies}
\begin{lem}\label{DJtriv}
The Belavin-Drinfeld cohomology associated to the Drinfeld-Jimbo $r$-matrix is trivial.
\end{lem}
\begin{proof}
Let $X \in Z(r_{DJ})$. By Theorem \ref{CenterinH} and Lemma \ref{diagblocks} we have $X=QD$ where $Q \in GL(M, F)$ and $D \in \diag(M, \overline{F})$. Hence, for any $\sigma \in \Gal(\overline{F}/F)$ we have $D^{-1}\sigma(D) \in C(r_{DJ}) \subset H$. This implies $d_id_{M+1-i}=k_i \in F$ (for $B_n$ series we also have $d_{[(M+1)/2]} \in F$). Let $K=\diag(k_1, ..., k_{[M/2]}, 1, 1, ...,1)$ (for $B_n$ let $K=\diag(k_1, ..., k_{[M/2]}, d_{[(M+1)/2]}, 1, ...,1)$). Then $D_1=K^{-1}D \in G$. We have $X=(QK) \cdot D_1$. Thus $X \sim I$.
\end{proof}
\begin{rem}
It is easy to see that $C(r_{DJ})=H$. Therefore for an arbitrary $r$-matrix $r$ we have $Z(r) \subset Z(r_{DJ})$.
\end{rem}
Let $r$ be an $r$-matrix from the Belavin-Drinfeld list associated to a triple $(\Gamma_1, \Gamma_2, \tau)$.
\begin{thm}\label{non-tw}
For any $r$ matrix in the $B_n, C_n$ series $H_{BD}^1(r)$ contains one element. For the $D_n$ series if there exists $k \in \mathbb{Z}$ such that $\tau^{k}(n-1)=n$, then $H_{BD}^1(r)$ is parametrized by the elements of $F^*/(F^*)^2$. Otherwise $H_{BD}^1(r)$ is trivial.
\end{thm}
\begin{proof}
First note that in the $B_n$ and $C_n$ series there is no $k$ such that $\tau^{k}(n-1)=n$, because $\alpha_{n-1}$ and $\alpha_n$ have different length.
Consider a homomorphism $T:H(\overline{F})\to \diag(n, \overline{F})\subset GL(n, \overline{F})$ defined by $T(d_1, ..., d_n)=(s_1, ..., s_{n})$, $s_i=e^{\alpha_i}(d_1,..., d_n)$. Then $T$ is surjective and its kernel is the subset of $(d_1, ...,d_n) \in H$ such that $d_1=...=d_n$. For $s_1, ..., s_{n} \in F$ we can find an element of $H(F)$ in $T^{-1}(s_1, ..., s_{n})$ if and only if the following holds: for the $D_n$ case $s_{n-1}s_n$ is a perfect square, for the $B_n$ case $s_n$ is a perfect square, and always for the $C_n$ series.
For a positive number $i$ such that $\alpha_i \in \Gamma_1$ let us consider the string of $\tau$ that contains $\alpha_i$. By $\eta(i)$ we will denote the smallest number $k$ such that $k>i$ and $\alpha_k$ lies in the same string as $\alpha_i$. If $\alpha_i \not\in \Gamma_1$, then we define $\eta(i)=0$. For $D \in H$ let $s_i=e^{\alpha_i}(D)$.
For any $X \in Z(r_{BD})$ we have a decomposition $X=QD$. Let $D=(d_1, ..., d_n)$ by Lemma \ref{DJtriv}.
{\bf Case 1}. $\eta(n-1)=0$. If $D=(d_1, ..., d_n)$ is a diagonal matrix then the tuple $(s_1, ..., s_{n-1})$ defines $D$ up to a scalar. $D \in C(r_{BD})$ iff $s_i$'s are equal on the strings of $\tau$. We have $D^{-1}\sigma(D) \in C(r_{BD})$. This means that for all $\alpha_i, \alpha_j$ such that $\tau^k(\alpha_i)=\alpha_j$ we have $s_i=k_is_j, k_i \in F$. We can find a matrix $D'=(d_1', ..., d_n') \in H(F)$ for which $s_i'=k_is_j'$ for all $i, j$ such that $\tau^k(\alpha_i)=\alpha_j$ for some $k$. Then the decomposition $X=(QD') \cdot I \cdot (D'^{-1}D)$ gives an equivalence between $X$ and $I$.
{\bf Case 2.} $\eta(n-1)=n$ (possible only in the $D_n$ algebra). Again we have $s_i=k_is_{\eta(i)}$ for $k_i \in F$. The equations $s_i=k_is_{\eta(i)}$ for $i=1, ..., n-1$ and $s_{n-1}=k_{n-1}^2s_n$ have a solution $d_1', ..., d_n' \in F$. Let $I'=(1,1, ..., 1, \sqrt{k_{n-1}})$. Then $X=(QD')I'(D'^{-1}D)$. Therefore, every cocycle is equivalent to one of the form $(1, ..., 1, \sqrt{k})$ for some $k \in F$. Let $I'=(1, .., 1, \sqrt{k})$ and $I''=(1, 1, .., 1, \sqrt{l})$. If $I'' \sim I'$, then $I'=QI''C$, where $Q \in H(F)$ and $C \in C(r_{BD})$. But then $k/l$ must be a square in $F$. Therefore, $H_{BD}^1(r)$ is parametrized by the elements of $F^*/(F^*)^2$.
\end{proof}
\section{Twisted cohomologies}
In this section we work over the field $\mathbb{K}=\mathbb{C}((\hslash))$.
From now on we will make distinction not only between $B,C,D$ series, but also between the algebras from $D_n$ series for odd and even $n$.
By $S$ we denote the following matrix:\\
the matrix with $1$'s on the antidiagonal for the $D_{n}$, $n$ even, and $B_n$ algebras;\\
the matrix with $1$'s on the antidiagonal but with central $2 \times 2$ block replaced with a unit matrix for $D_{n}$ with odd $n$;\\
the matrix with $-1$'s on the upper half of the antidiagonal and $1$'s on the lower half for the $C_n$ algebra.
\begin{ex}
The matrices $S$ for the algebras $D_2, D_3, B_2$ and $C_2$ respectively are\\
$\begin{pmatrix}
& & & 1\\
& &1 & \\
&1 & & \\
1 & & &
\end{pmatrix},
\begin{pmatrix}
& & & & & 1\\
& & & &1 & \\
& &1 &0 & & \\
& &0 &1 & & \\
& 1& & & & \\
1 & & & & &
\end{pmatrix},
\begin{pmatrix}
& & & &1 \\
& & &1 & \\
& &1 & & \\
&1 & & & \\
1 & & & &
\end{pmatrix},
\begin{pmatrix}
& & & -1\\
& &-1 & \\
& 1 & & \\
1 & & &
\end{pmatrix}.
$
\end{ex}
Note that we always have $S \in G$ and $S^2= \pm \mathrm{id}$ with $S^2=-\mathrm{id}$ only for $C_n$ algebra.
Recall that if $\overline{H}^1_{BD}(G,r)$ is nonempty then $r^{21}$ and $r$ are conjugate.
\begin{lem}\label{r-conj}
If $r$ corresponds to a triple $(\Gamma_1, \Gamma_2, \tau)$ and $r$ and $r^{21}$ are conjugate, then $r^{21}$ and $\mathrm{Ad}_S r$ are conjugate by an element of $C(r_{DJ})$.
\end{lem}
\begin{proof}
It is easy to see that $\mathrm{Ad}_Se_{\alpha}=e_{-\alpha}$ for all cases except $D_n$ with odd $n$. For $D_n$ and odd $n$ we have $\mathrm{Ad}_{S}e_{\alpha_i}=e_{-\alpha_i}$ for $i=1,...,n-1$ and $\mathrm{Ad}_{S}e_{\alpha_n}=e_{-\alpha_{n-1}}, \mathrm{Ad}_{S}e_{\alpha_{n-1}}=e_{-\alpha_n}$. Note that $\mathrm{Ad}_S$ acts as $-1$ on $\mathfrak{h} \wedge \mathfrak{h}$. Thus $\mathrm{Ad}_{S}r_{DJ}=r_{DJ}^{21}$. This means that $\mathrm{Ad}_Sr$ and $r^{21}$ have the same semisimple parts. As they are conjugate we can conclude that they are conjugate by an element of $C(r_{DJ})$.
\end{proof}
\begin{lem}\label{r-list}
If $r$ corresponds to a triple $(\Gamma_1, \Gamma_2, \tau)$ and $r$ and $r^{21}$ are conjugate, then the triple is one of the following:\\
for $B_n, C_n$ or $D_n$ with even $n$: $\Gamma_1=\Gamma_2=\emptyset$;\\
for odd $n$:
$\Gamma_1=\Gamma_2=\emptyset$ or one of the following cases:\\
1. $\Gamma_1=\{\alpha_{n-1}\}$, $\tau(\alpha_{n-1})=\alpha_n$;\\
2. $\Gamma_1=\{\alpha_{n}\}$, $\tau(\alpha_{n})=\alpha_{n-1}$;\\
3. $\Gamma_1=\{\alpha_{n-1}, \alpha_k\}$, $\tau(\alpha_{n-1})=\alpha_k, \tau(\alpha_k)=\alpha_{n}$;\\
4. $\Gamma_1=\{\alpha_{n}, \alpha_k\}$, $\tau(\alpha_{n})=\alpha_k, \tau(\alpha_k)=\alpha_{n-1}$.\\
In these cases $r^{21}=\mathrm{Ad}_S r$.
\end{lem}
\begin{proof}
Form Lemma \ref{r-conj} we know that $\mathrm{Ad}_S r$ and $r^{21}$ are conjugate by an element of $C(r_{DJ})$.
Note that $e_\alpha \wedge e_\beta$ is a common eigenvector for $C(r_{DJ})$. If $e_{-\alpha} \wedge e_{\beta}$, $\beta\neq-\alpha$ is a summand of $r$ then $\mathrm{Ad}_Sr$ contains $e_{\alpha} \wedge e_{-\beta}$ and $r^{21}$ contains $e_{-\alpha} \wedge e_{\beta}$. Therefore, in all cases except $D_n$ for odd $n$ if $\alpha=\tau(\beta)$, then $\beta=\tau^{k}(\alpha)$ for some $k$, i.e. $\Gamma_1=\Gamma_2=\emptyset$. For $D_n$ with odd $n$ the same works if $\alpha,\beta \notin \{\alpha_{n-1}, \alpha_n \}$. The only Belavin-Drinfeld triples that do not have such elements are listed in the statement. It is easy to see that for each of them $r^{21}=\mathrm{Ad}_{S}r$.
\end{proof}
\begin{rem}
Assume that $X\in G(\overline{\mathbb{K}})$ is a Belavin--Drinfeld twisted cocycle associated to $r$. The identity $(\mathrm{Ad}_{X^{-1}\sigma_{2}(X)}\otimes \mathrm{Ad}_{X^{-1}\sigma_{2}(X)})(r)=r^{21}$ is equivalent to
$(\mathrm{Ad}_{S^{-1}X^{-1}\sigma_{2}(X)}\otimes \mathrm{Ad}_{S^{-1}X^{-1}\sigma_{2}(X)})(r)=r$, which implies that $S^{-1}X^{-1}\sigma_{2}(X)\in C(r)$. Thus $X^{-1}\sigma_{2}(X)=SD$ with $D\in C(r)$.
On the other hand, $(\mathrm{Ad}_{X^{-1}\sigma(X)}\otimes \mathrm{Ad}_{X^{-1}\sigma(X)})(r)=r$ holds for any $\sigma\in \Gal(\overline{\mathbb{K}}/\mathbb{K}[\sqrt{\hslash}])$. This in turn is equivalent to $X^{-1}\sigma(X)=D_{\sigma}\in C(r)$.
\end{rem}
We will now introduce a specific matrix $J \in GL(M, \mathbb{K}[\sqrt{\hslash}])$ for the $B,C,D$ series. In every case we will have $\overline{J}:=\sigma_{0}(J)=JS$. Again, notice the distinction between $D_n$ for odd and even $n$.
For $D_n$, $n$ even denote by $J=J(D_n)\in GL(2n,\mathbb{K}[\sqrt{\hslash}])$ the matrix
with the following entries:
$a_{kk}=1$ for $k \leq n$, $a_{kk}=-\sqrt{\hslash}$ for $k \geq n+1$, $a_{k,2n+1-k}=1$
for $k \leq n$, $a_{k,2n+1-k}=\sqrt{\hslash}$ for $k \geq n+1$.
For $D_n$, $n$ odd let $J$ be a matrix $J(D_{n-1})$ with the unit matrix inserted as the central $2 \times 2$ block.
For $B_n$ let $J$ be a matrix $J(D_n)$ with the unit matrix inserted as the central
$1 \times 1$ block.
For $C_n$ take $J$ to be $J(D_n)$.
E.g.\ for $D_2$ and $C_2$ we have $$J=\left(\begin{array}{cccc}
1 & 0 &0 & 1\\
0 & 1 & 1 & 0\\
0 &\sqrt{\hslash} &-\sqrt{\hslash} &0\\
\sqrt{\hslash} & 0 & 0 & -\sqrt{\hslash} \end{array}\right),$$ for $D_3$ we have
$$J= \left(\begin{array}{cccccc}
1 & & &... & & 1\\
& 1& &... &1 &\\
... & & 1 & 0 & ... &\\
... & & 0 & 1 & ... &\\
& \sqrt{\hslash} & &... & - \sqrt{\hslash} &\\
\sqrt{\hslash} & & &... & & -\sqrt{\hslash} \end{array}\right),$$ and for $B_2$ we have
$$J=\left(\begin{array}{ccccc}
1 & 0 & 0 &0 & 1\\
0 & 1 &0 & 1 & 0\\
0 &0 &1 &0 & 0\\
0 &\sqrt{\hslash} & 0 &-\sqrt{\hslash} &0\\
\sqrt{\hslash} & 0 &0 & 0 & -\sqrt{\hslash} \end{array}\right).$$
\begin{lem}\label{o_RJD}
Let $X\in \overline{Z}(G,r)$. Then there exist $R\in GL(M,\mathbb{K})$
and $D\in \mathrm{diag}(M,\overline{\mathbb{K}})$ such that $X=RJD$. Moreover,
$D=\mathrm{diag}(d_{1},..., d_{M})$, where $d_{i}d_{M-i}\in \mathbb{K}$
for all $i$ (for $B_n$ series we also have $d_{n+1}=1$, for $C_n$ series we have $d_id_{M-i} \in \sqrt{\hslash} \mathbb{K}$), and for any $\sigma \in \Gal(\overline{\mathbb{K}}/\mathbb{K}[\sqrt{\hslash}])$ we have $D^{-1}\sigma(D) \in C(r)$.
\end{lem}
\begin{proof}
By Theorem \ref{CenterinH} we have $C(r)\subset H$.
Note that in the equation $\overline{X}=XSD$ the matrix $SD$ is a block matrix with $2 \times 2$ blocks. Applying Lemma \ref{blocks} we see that $X=RX_1$ with $X_1$ having the same block form. If we would provide an appropriate decomposition for each $2 \times 2$ block we would obtain the desired decomposition.
In each block the equation can be equivalent to the non-twisted cohomology equation (if in $S$ we have a unit block) or as one of the equations considered below. For $\sigma \in \Gal(\overline{\mathbb{K}}/\mathbb{K}[\sqrt{\hslash}])$ we have $RJDC=\sigma(RJD)=RJ\sigma(D)$, where $C \in C(r)$.
Let $X=\left(\begin{array}{cc} a & b \\
c & d \end{array}\right)\in GL(2,\mathbb{K}[\sqrt{\hslash}])$
satisfy $\overline{X}=XWD$ with $D=\mathrm{diag}(d_{1},d_{2})\in SO(2,\mathbb{K}[\sqrt{\hslash}])$, where $W=\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix}$. The identity is equivalent to the following system: $\overline{a}=bd_{1}$, $\overline{b}=ad_{2}$, $\overline{c}=dd_{1}$,
$\overline{d}=cd_{2}$. Note that $cd\neq 0$ (otherwise both $c$ and $d$ should be zero). Let $a/c=a'+b'\sqrt{\hslash}$. Then $b/d=a'-b'\sqrt{\hslash}$. One can immediately check that $X=RJ'D'$, where
$R=\left(\begin{array}{cc} a' & b' \\
1 &0 \end{array}\right)\in GL(2,\mathbb{K})$,
$J'= \left(\begin{array}{cc} 1 & 1 \\
\sqrt{\hslash} &-\sqrt{\hslash} \end{array}\right)$, $D'=\mathrm{diag}(c,d)\in \mathrm{diag}(2,\mathbb{K}[\sqrt{\hslash}])$. Note that $\overline{cd}=cd d_1d_2=cd$, thus $cd \in \mathbb{K}$.
Let $X=\left(\begin{array}{cc} a & b \\
c & d \end{array}\right)\in GL(2,\mathbb{K}[\sqrt{\hslash}])$
satisfy $\overline{X}=XWD$ with $D=\mathrm{diag}(d_{1},d_{2})\in SO(2,\mathbb{K}[\sqrt{\hslash}])$, where $W=\begin{pmatrix}
0 & -1\\
1 & 0
\end{pmatrix}$. The identity is equivalent to the following system: $\overline{a}=bd_{1}$, $\overline{b}=-ad_{2}$, $\overline{c}=dd_{1}$,
$\overline{d}=-cd_{2}$. Note that $cd\neq 0$ (otherwise both $c$ and $d$ should be zero). Let $a/c=a'+b'\sqrt{\hslash}$. Then $b/d=a'-b'\sqrt{\hslash}$. One can immediately check that $X=RJ'D'$, where
$R=\left(\begin{array}{cc} a' & b' \\
1 &0 \end{array}\right)\in GL(2,\mathbb{K})$,
$J'= \left(\begin{array}{cc} 1 & 1 \\
\sqrt{\hslash} &-\sqrt{\hslash} \end{array}\right)$, $D'=\mathrm{diag}(c,d)\in \mathrm{diag}(2,\mathbb{K}[\sqrt{\hslash}])$. Note that $\overline{cd}=-cd d_1d_2=-cd$, thus $cd \in \sqrt{\hslash} \mathbb{K}$.
\end{proof}
\begin{lem}
Let $D \in \mathrm{diag}(M, \overline{\mathbb{K}})$ satisfy $ d_i d_{M+1-i} \in \mathbb{K}$ for all $i$ (and $d_{n+1}=1$ for $B_n$ series; for $C_n$ we have $ d_i d_{M+1-i} \in \sqrt{\hslash} \mathbb{K}$ instead). Then there exists $R \in GL(M, \mathbb{K})$ such that $RJD \in G$.
\end{lem}
\begin{proof}
Let $X=RJD \in G$. Then $X$ must satisfy $X^{T}SX=S$. Denote $Y:=R^{-1}$. Then
$YSY^{T}=JDS(JD)^{T}$. Therefore, we can find the desired $R$ if and only if the quadratic forms with matrices $S$ and $JDS(JD)^T$ are equivalent over $\mathbb{K}$. Note that symplectic forms are always equivalent over a field. Henceforth we consider the cases of $B_n$ and $D_n$ series. We have $$JDS(JD)^T=\mathrm{diag}(2d_{1}d_{2n},2d_{2}d_{2n-1},...,2d_{n}d_{n+1},-2\hslash d_{n}d_{n+1},...,-2\hslash d_{1}d_{2n})$$ for $D_n$ and even $n$,
$$JDS(JD)^T= \mathrm{diag} (2d_{1}d_{2n},2d_{2}d_{2n-1},..., d_nd_{n+1}, d_{n}d_{n+1}, ...,-2\hslash d_{1}d_{2n})$$ for $D_n$ and odd $n$, and $$JDS(JD)^T=\mathrm{diag}(2d_{1}d_{2n},2d_{2}d_{2n-1},...,2d_{n}d_{n+1}, 1, -2\hslash d_{n}d_{n+1},...,-2\hslash d_{1}d_{2n})$$ for the $B_n$ algebra. Thus it suffices to show that for $d \in \mathbb{K}$ the quadratic forms with matrices $\left(\begin{array} {cccc}
0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
1 & 0 & 0 & 0 \\ \end{array}\right)$,
$\left(\begin{array} {cccc}
d &0 & 0 & 0\\
0 & d & 0 & 0\\
0 &0 & -\hslash d & 0\\
0 &0 & 0 & -\hslash d \end{array}\right)$
and $\left(\begin{array} {cccc}
d &0 & 0 & 0 \\
0 & d & 0 & 0 \\
0 & 0 & d & 0 \\
0 & 0 & 0 & d \\
\end{array}\right)$ are equivalent. This is straightforward.
\end{proof}
For $D \in \diag(M, \overline{\mathbb{K}})$ define $T(D)=SD^{-1}S\overline{D}$ (for the $C_n$ series we should take $T(D)=-SD^{-1}S'\overline{D}$ where $S'$ is the matrix with $1$'s on the antidiagonal, i.e. we ``forget'' the signs).
We will denote by $\mathbf{Z}$ the set of all diagonal matrices $D=\diag(d_1, ..., d_{M}) \in \diag(M, \overline{\mathbb{K}})$ satisfying $d_id_{M+1-i} \in \mathbb{K}$ (also $d_{n+1}=1$ for the $B_n$ algebra, $d_id_{2n+1-i} \in \sqrt{\hslash} \mathbb{K}$ for the $C_n$ algebra) and $D^{-1}\sigma(D) \in C(r)$ for all $\sigma \in \Gal(\overline{\mathbb{K}}/\mathbb{K}[\sqrt{\hslash}])$.
\begin{lem}
Let $X=RJD \in G$, $D \in \mathbf{Z}$. Then $X$ is a twisted cocycle if and only if $T(D) \in C(r)$.
\end{lem}
\begin{proof}
We have $\overline{X}=RJS\overline{D}=XST(D)$.
\end{proof}
\begin{thm}
We have
$$\overline{H}_{BD}^1(G,r)=\frac{T^{-1}(C(r)) \bigcap \mathbf{Z}}{C(r)\Ker T}$$
\end{thm}
\begin{proof}
For an arbitrary $D \in T^{-1}(C(r)) \bigcap \mathbf{Z}$ we can find $R \in GL(M, \mathbb{K})$ such that $X=RJD$ is a cocycle.
Assume that $X=R_1JD_1 \sim Y=R_2JD_2$. Then $Y=QXC$, $Q=R_2JD_2C^{-1}D_1J^{-1}R_1^{-1}$. Since $Q=\overline{Q}$, we have $D_2C^{-1}D_1^{-1}=S\overline{D_2C^{-1}D_1^{-1}}S$ (for $C_n$ series one should take $S'$ instead of $S$ in the last formula). Thus $D_2D_1^{-1} \in C(r)\Ker T$. If $D_2D_1^{-1} \in C(r)\Ker T$, then $D_2D_1^{-1}=D_0C$. If we denote $Q=R_2JD_0J^{-1}R_1^{-1}$, then we have $Y=QXC$.
\end{proof}
\begin{thm}
Let $r$ be from the list of Lemma \ref{r-list}. Then $|\overline{H}_{BD}^1 (G, r)|=1$ for $r=r_{DJ}$ and $|\overline{H}_{BD}^1 (G, r)|=2$ for $r \neq r_{DJ}$.
\end{thm}
\begin{proof}
We will first consider the case of the Drinfeld-Jimbo $r$-matrix. Let $D=\diag(d_1, ..., d_{M})$. Then for $D_n$ with even $n$ we have $$T(D)=\diag(d_{2n}^{-1}\overline{d_1}, d_{2n-1}^{-1}\overline{d_2}, ..., d_1^{-1}\overline{d_{2n}}),$$ for $D_n$ with odd $n$ we have $$T(D)=(d_{2n}^{-1}\overline{d_1}, ..., d_{n}^{-1}\overline{d_n}, d_{n+1}^{-1}\overline{d_{n+1}}, ..., d_{1}^{-1}\overline{d_{2n}}),$$ for $B_n$ we have $$T(D)=\diag(d_{2n}^{-1}\overline{d_1}, d_{M-1}^{-1}\overline{d_2}, ..., d_{n+1}^{-1}d_n, 1, d_n^{-1}d_{n+1}, ..., d_1^{-1}\overline{d_{2n}}),$$ anf for the $C_n$ algebra $$ T(D)=(d_{2n}^{-1}\overline{d_1}, ..., d_{n+1}^{-1}\overline{d_n}, -d_{n-1}^{-1}\overline{d_{n+1}}, ..., -d_{1}^{-1}\overline{d_1}) .$$ If $D \in \mathbf{Z}$, then $T(D) \in G$, thus $T(D) \in C(r)$. It is easy to see that an arbitrary matrix from $\mathbf{Z}$ can be decomposed as a product of a matrix from $G$ and a matrix from $\Ker T$. Therefore $|H_{BD}^1(G, r_{DJ})|=1$. For the $D_n$ series decomposing $D \in \mathbf{Z}$ as $D=WD_{1}$ with $W \in \Ker T$ and $D_1 \in G$ one gets the following: if $D\in \mathbb{K}$ (resp.\ $\mathbb{K}[\sqrt{\hslash}]$), then $W, D_1 \in \mathbb{K}$ (resp.\ $\mathbb{K}[\sqrt{\hslash}]$).
Now let $r \neq r_{DJ}$ (this means that we deal with the case $D_n$ for odd $n$). Let $D \in \mathbf{Z}$. Then $D=KF$, where $K \in \diag(2n, \mathbb{K}) \bigcap \Ker T$, $F \in G$. Then $F$ is a non-twisted cocycle for $r$ over $\mathbb{K}[\sqrt{\hslash}]$. Hence, because of factorization by $C(r)$ we can assume that either $F \in \diag(2n,\mathbb{K}[\sqrt{\hslash}])$ or $F=PF_1$, where $F_1 \in \diag(2n,K[\sqrt{\hslash}])$, $P=\diag(1,1, ..., \sqrt[4]{\hslash}, 1/\sqrt[4]{\hslash}, 1, ..., 1)$ (see Theorem \ref{non-tw}). But it is easy to see that $P \in \Ker T$. Therefore in what follows we assume that $D \in GL(2n, K[\sqrt{\hslash}])$, $d_id_{2n+1-i} \in \mathbb{K}$. Finally $$ \overline{H}_{BD}^1(SO(2n),r)=\frac{T^{-1}(C(r)) \bigcap \mathbf{Z}}{C(r)\Ker T}=\frac{A}{C(r)\Ker T},$$
where $A=T^{-1}(C(r))\bigcap O(2n, \mathbb{K}[\sqrt{\hslash}])$.
Recall that we define $(d_1, ..., d_n)=\diag(d_1, ..., d_n, 1/d_n, ..., 1/d_1)$. We will write the defining relations for $A$ and $C(r)$ in each case. Defining relations for $\Ker T$ are $d_1\overline{d_1}=1, ...,d_{n-1}\overline{d_n}=1, \overline{d_n}=d_n$.
\noindent{\bf Case 1.}\\
$A:$ $(d_{n}^{-1}\overline{d_n})^2=1$.\\
$C(r):$ $d_n^2=1$.\\
If $d_n \in \mathbb{K}$, then $(d_1, ..., d_n)=(d_1, ..., d_{n-1}, 1)(1, 1, ..., d_n)\in C(r)\Ker T$. \\
If $d_n \in \sqrt{\hslash} \mathbb{K}$, then $(d_1, ..., d_n)=(1, ..., \sqrt{\hslash})(1, ..., 1, d_n/\sqrt{\hslash})(d_1, ..., d_{n-1},1)$. We have $(1, ..., 1)\not\sim (1, ..., \sqrt{\hslash})$ because if $(a_1, ..., a_n) \in \Ker T \cup C(r)$ then $a_n \in \mathbb{K}$.
Hence $H_{BD}^1$ contains two elements.
Case 2 is similar to Case 1.
\noindent{\bf Case 3.1} (Case 3 with $k=n-2$).\\
$A: (d_n^{-1}\overline{d_n})^2=1;\ (d_{n-1}\overline{d_{n-1}})^2d_n^{-1}\overline{d_n}=d_{n-2}\overline{d_{n-2}}$.\\
$C(r): d_n^2=1, d_{n-1}^2d_n=d_{n-2}$.\\
If $d_n \in \mathbb{K}$, then
$$(d_1, ..., d_n)=(d_1, ..., d_{n-3}, d_{n-1}^2, d_{n-1},1)(1, ..., 1, d_{n-2}/d_{n-1}^2, 1, d_n).$$
If $d_n \in \sqrt{\hslash}\mathbb{K}$, then
$$(d_1, ..., d_n)=(1, ..., \sqrt{\hslash})(d_1, ..., d_{n-1}^2, d_{n-1},1)(1, ..., d_{n-2}/d_{n-1}^2, 1, d_n/\sqrt{\hslash}).$$
\noindent{\bf Case 3.2} (Case 3 with $k\neq n-2$).\\
To simplify the notation we assume $k=n-3$ (general case is completely similar)\\
$C(r): d_n^2=1; d_{n-3}=d_{n-2}d_{n-1}d_n$.\\
$A: (d_n^{-1}\overline{d_n})^2=1; d_{n-3}\overline{d_{n-3}}=d_{n-2}d_{n-1}d_n^{-1} \overline{d_{n-2}d_{n-1}d_n}$. \\
If $d_n \in \mathbb{K}$, then
$$(d_1, ..., d_n)=(d_1, ..., d_{n-2}d_{n-1}, d_{n-2}, d_{n-1}, 1) (1, ..., d_{n-3}/d_{n-2}d_{n-1}, 1, 1, d_n). $$
If $d_n \in \sqrt{\hslash}\mathbb{K}$, then
$$(d_1, ..., d_n)=(1,..., \sqrt{\hslash})(d_1, ..., d_{n-2}d_{n-1}, d_{n-2}, d_{n-1}, 1) (1, ..., d_{n-3}/d_{n-2}d_{n-1}, 1, 1, d_n/\sqrt{\hslash}). $$
Case 4 is similar to Case 3.
\end{proof}
\section{Concluding Remarks}
Consider the isomorphosm $\mathfrak{so}(3,3) \cong \mathfrak{sl}(4)$. Under this isomorphism $F$-points are mapped to $F$-points. We know that $|\overline{H}_{BD}^1(r_{BD}, \mathfrak{sl}(4), GL(4))|=1$ (see \cite{SP}). Our results imply that $\overline{H}_{BD}^1(r_{BD}, \mathfrak{so}(3,3))$ contains $2$ elements. This means that we have two twisted Lie bialgebra structure on $\mathfrak{so}(3,3)$ that are non-conjugate but isomorphic. This motivates the following conjecture:
\begin{conj}
If $r_{BD}$ is an $r$-matrix from the list of Lemma \ref{r-list} then all twisted Lie bialgebra structures on $\mathfrak{g}$ corresponding to $r_{BD}$ are isomorphic.
\end{conj}
Consider the problem of quantizing Belavin-Drinfeld $r$-matrices. In \cite{ESS} Etingof, Schedler and Schiffman explicitely quantized the whole Belavin-Drinfeld list. Let us see how this is related to our considerations.
Note that any $r$-matrix of twisted type satisfies $r+r^{21}=\sqrt{\hslash} \Omega$, where $\Omega$ is a Casimir element. This means that the quasi-classical limit of the twisted type $r$-matrix is skew. Therefore, Belavin-Drinfeld non-twistes cohomologies parametrize the set of quantizations of a given Belavin-Drinfeld matrix (up to some equivalence that is weaker then isomorphism). Thus, in most cases considered (except for special BD-triples for $D_n$) the Etingof-Schedler-Schiffman quantization is unique.
\medskip
\textbf{Acknowledgment.} The authors are grateful to J.-H. Lu for valuable suggestions.
|
\section{Introduction}
Helical Dirac fermions protected by topology and found at the surface
of topological insulators (TI)\cite{Hasan,Qi1,Moore,Hsieh1} exhibit
spin locked perpendicular to their momentum.\cite{Chen1,Hsieh1,Hsieh2,Jozwiak,Xu}
Doping with magnetic impurities can break time reversal symmetry and
create massive Dirac fermions as has been experimentally demonstrated\cite{Chen2}
in $Bi_{2}Se_{3}$. A gap can also be introduced in thin films when
the distance between top and bottom surfaces is of the order of the
extend in space of the surface states. \cite{Lu,Linden} Gapped Dirac fermions
arise in many other systems, for example in two dimensional membranes
such as the dichalcoginide $MoS_{2}$ \cite{Mak,Splend,Lee,Feng,Li1}and
silicene\cite{Aufray,Padova,Stille,Ezawa1,Ezawa2} with buckled honeycomb
lattice. In both these cases pseudospin plays the analogous role to
the real spin of topological insulators.
The electron-phonon interaction renormalizes quasiparticle dynamics
and leads to important observable changes in electronic properties
\cite{Carbotte,Li2,Li3,Li4,Stauber,Pound2,Ion,Leblanc1,Leblanc2,Pound3}
which illustrate the effects of many body renormalizations not captured
in single particle theories. For the Dirac electrons in graphene as
an example, features observed in the density of states \cite{Pound3,Miller,GLi,Nicol}
and in the dispersion curves measured in angular resolved photo emission
spectroscopy \cite{Zhou} have been interpreted as phonon structure.
\cite{Pound}The optical properties of graphene are also renormalized
in a non trivial way.\cite{Carbotte,Li4,Stauber,zli,Carbotte1} In
a simple bare band picture there is no optical absorption in the photon
energy region between the Drude intraband contribution centered about
$\omega=0$ and the interband onset at twice the value of the chemical
potential $\omega=2\mu$. In reality the real part of the dynamic
longitudinal conductivity is observed to be finite and almost one
third its universal background\cite{Gusynin1,Gusynin2} value in this
photon region.\cite{zli} This absorption is due to many body renormalizations
and is at least partially assigned \cite{Carbotte,Stauber} to the
electron-phonon interaction which provides boson assisted processes
referred to as Holstein processes. A phonon is created along with
an electron-hole pair.
Understanding the transport properties of the Dirac electrons on the surface of
a topological insulator is important for possible device applications.
At finite temperature the electron-phonon interaction is expected to be an
important scattering chanel \cite{Hatch}, X. Zhu et al. \cite{Zhu1} have studied the surface phonons
on the (001) surface of $Bi_{2}Se_3$ and in particular have found a giant Kohn anomaly
associated with a branch having a maximum of 18 THz. From measurements of the phonon self energy,
the same group \cite{Zhu2} determined the size of
the electron-phonon interaction report a coupling constant of 0.43 for a particular branch which is
much larger than reported in angular resolved photo-emission spectroscopy. One such study
by Z. H. Pan et al. \cite{Pan} gives a mass enhancement $\lambda$ of 0.08
while another by R. L. Hatch et al. \cite{Hatch} found $\lambda\sim$ 0.25.
A more recent report \cite{Chen} gives $\sim$ 0.17 with characteristic phonon energy $\sim$18 meV.
The results of X. Zhu et al. are further supported by an infrared study by
A. D. LaForge et al. \cite{LaForge} which found a strong electron-phonon
coupling to a 7.6 meV optical phonon while S. Giraud et al. \cite{Giraud1}
provide arguments for coupling to acoustic phonon with $\lambda\sim$ 0.42 in their films and even
larger in other geometries. \cite{Giraud2}
Very recently J. A. Sobota et al. \cite{Sobota}, using time-resolved photoemission
spectroscopy, find evidence for coupling to a 2.05 ThZ surface phonon mode with the Dirac
electrons in $Bi_{2}Se_3$. Possible complications in the interpretation of optical
pumping on time resolved data were discussed by S. Ulstrup \cite{Ulstrup}
and need to be kept in mind. Finally, the recent transport measurements
of M.V. Costache et al. \cite{Costache} were interpreted
with strong coupling to a single optical phonon mode of energy $\sim$ 8 meV.
In this paper we will emphasize the effect of the electron-phonon
interaction on the spin texture of gapped Dirac helical fermions.
We will present for comparison, results for the familiar spectral
density ($A_{I}(k,\omega)$) associated with quasiparticle renormalization
which we compared with the much less familiar equivalent function
which enters gap renormalizations ($A_{z}(k,\omega)$). This is the
essential quantity for the calculation of the z-axis (perpendicular
to the surface states) properties. While only $A_{I}(k,\omega)$ is
needed in calculations of the density of states $N(\omega)$ both
$A_{z}$ and $A_{I}$ enter the dynamic conductivity $\sigma_{xx}(\omega)$.
In section II we present the necessary formalism for both the self
energy associated with quasiparticle renormalizations $\Sigma_{I}(\omega))$
and with the renormalization of the gap $\Sigma_{z}(\omega))$ which
determine the needed spectral functions $A_{I}(k,\omega)$ and $A_{z}(k,\omega)$.
Numerical results are presented in section III which deals with spectral
densities, density of states and optical conductivity. In section
IV we consider both in plane and out of plane spin texture. A summary
and conclusions are given in section V.
\section{Formalism}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=2.6in,height=3.2in,angle=-90]{Fig1.eps}
\par\end{centering}
\caption{(Color online) Illustrative representation of Dirac cones (left frame)
and including a gap $\Delta$ (right frame). }
\label{fig1}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig2.eps}
\par\end{centering}
\caption{(Color online) The real and imaginary part of the electron-phonon
self energies $\Sigma^{I}(\omega)$ and $\Sigma^{z}(\omega)$ in meV
as a function of $\omega$ in meV. $\Sigma^{I}(\omega)$ (real part,
solid black), and imaginary part (dashed red) gives the quasiparticle
renormalization and $\Sigma^{z}(\omega)$ (real part, dotted blue),
and imaginary part (dash-dotted purple) gives the gap renormalization.
Our choice of g=10 corresponds to a mass enhancement $\lambda\approxeq0.3$.}
\label{fig2}
\end{figure}
We begin with a model hamiltonian for the electronic states at the
surface of a topological insulator which has the form
\begin{equation}
H_{0}=\hbar v_{F}[k_{x}\sigma_{y}-k_{y}\sigma_{x}]+\Delta\sigma_{z}+E_{0}(\mathbf{k})\label{Ham}
\end{equation}
where $\sigma$'s are Pauli spin matrices for real electron spin,
$v_{F}$ is the Fermi velocity of the Dirac electrons, $\mathbf{k}$
is momentum, $\Delta$ is a gap which can be introduced for example
by doping with magnetic impurities and $E_{0}(\mathbf{k})$ is a quadratic
term $E_{0}(\mathbf{k})=\frac{\hbar^{2}k^{2}}{2m}$ with $m$ the
electron mass. The introduction of $E_{0}$ gives particle hole asymmetry.
Here we will assume that $m$ is very large, as a first approximation,
and will ignore this term.
The non interacting Green's function takes on the form
\begin{equation}
G_{0}(\mathbf{k},s,i\omega_{n})=\frac{1}{i\omega_{n}+\mu-s\sqrt{\hbar^{2}v_{F}^{2}k^{2}+\Delta^{2}}}
\end{equation}
with $i\omega_{n}$ the Matsubara frequencies and $\mu$ is the non
interacting chemical potential. The eigen energies are $\varepsilon_{k,s}=s\sqrt{\hbar^{2}v_{F}^{2}k^{2}+\Delta^{2}}$
with $s=\pm$ for conduction and valence band respectively. These
dispersion curves are shown schematically in Fig. 1 with (right frame)
and without (left frame) a gap.
The matrix ($2\times2$) Green's function $\hat{G}_{0}(\mathbf{k},i\omega_{n})$
can be written in terms of the scalar $G_{0}(\mathbf{k},s,i\omega_{n})$
as
\begin{equation}
\hat{G}_{0}(\mathbf{k},i\omega_{n})=\frac{1}{2}\sum_{s=\pm}(1+s\mathbf{F}_\mathbf{k}\cdot\mathbf{\sigma})G_{0}(\mathbf{k},s,i\omega_{n})
\end{equation}
with the vector $\mathbf{F}_{k}$ defined as
\begin{equation}
\mathbf{F}_\mathbf{k}=\frac{(-\hbar v_{F}k_{y},\hbar v_{F}k_{x},\Delta)}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+\Delta^{2}}}
\end{equation}
We want to include an electron-phonon interaction. The simplest is
the Holstein model for coupling to a phonon mode of energy $\omega_{E}$
with matrix element between electronic and phonon assumed to be a
constant ($g$). It is written as
\begin{equation}
H_{e-ph}=-g\omega_{E}\sum_{\mathbf{k},\mathbf{k}^{\prime},s}c_{\mathbf{k},s}^{\dag}c_{\mathbf{k}^{\prime},s}(b_{\mathbf{k}^{\prime}-\mathbf{k}}^{\dag}+b_{\mathbf{k}-\mathbf{k}^{\prime}})\label{phonon}
\end{equation}
where $c_{\mathbf{k},s}^{\dag}$ creates an electron of momentum $\mathbf{k}$
and spin $s$ and $b_{\mathbf{k}^{\prime}-\mathbf{k}}^{\dag}$ creates
a phonon of momentum $\mathbf{k}^{\prime}-\mathbf{k}$ reduced to
the first surface state Brillouin zone and energy $\omega_{E}$ with
$g$ coupling constant. The perturbing Hamiltonian (5) provides a
self energy to electron motion which has two components one proportional
to the unit matrix $\widehat{I}$ and the other to $\widehat{\sigma}_{z}$.
Writing $\widehat{\Sigma}(i\omega_{n})=\Sigma^{I}(i\omega_{n})\widehat{I}+\Sigma^{Z}(i\omega_{n})\widehat{\sigma}_{z}$
we get in lowest order perturbation theory
\begin{eqnarray}
& & \Sigma^{I}(i\omega_{n})=\frac{g^{2}\omega_{E}^{2}}{2}\sum_{\mathbf{q,}s}\notag\nonumber \\
& & \left[\frac{f_{F}(\varepsilon_{q,s})+f_{B}(\omega_{E})}{i\omega_{n}+\mu+\omega_{E}-\varepsilon_{q,s}}+\frac{f_{B}(\omega_{E})+1-f_{F}(\varepsilon_{q,s})}{i\omega_{n}+\mu-\omega_{E}-\varepsilon_{q,s}}\right]\label{sigmaI}
\end{eqnarray}
and
\begin{eqnarray}
& & \Sigma^{Z}(i\omega_{n})=\frac{g^{2}\omega_{E}^{2}}{2}\sum_{\mathbf{q,}s}\frac{s\Delta}{\sqrt{\hbar^{2}v_{F}^{2}q^{2}+\Delta^{2}}}\times\notag\nonumber \\
& & \left[\frac{f_{F}(\varepsilon_{q,s})+f_{B}(\omega_{E})}{i\omega_{n}+\mu+\omega_{E}-\varepsilon_{q,s}}+\frac{f_{B}(\omega_{E})+1-f_{F}(\varepsilon_{q,s})}{i\omega_{n}+\mu-\omega_{E}-\varepsilon_{q,s}}\right]\label{sigmaZ}
\end{eqnarray}
where $f_{F}$ and $f_{B}$ are fermion and boson distribution functions
$1/[e^{(\omega-\mu)/k_{B}T}\pm1]$ respectively. Here $\mu$ is the
chemical potential which applies only to the electron distribution. Note that $\Sigma^{Z}(i\omega_{n})$ in Eq. (7)
is directly proportional to the gap $\Delta$ and will vanish for
$\Delta=0$. In terms of the self energies $\Sigma^{I}(i\omega_{n})$
and $\Sigma^{Z}(i\omega_{n})$ the interacting matrix Green's function
takes on the form
\begin{equation}
\hat{G}(\mathbf{k},i\omega_{n})=\frac{1}{2}\sum_{s=\pm}(1+s\mathbf{H}_{\mathbf{k}}\cdot\mathbf{\sigma})G(k,s,i\omega_{n})\label{Greenf}
\end{equation}
with
\begin{equation}
\mathbf{H}_{\mathbf{k}}=\frac{(-\hbar v_{F}k_{y},\hbar v_{F}k_{x},\Delta+\Sigma^{Z}(i\omega_{n}))}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+(\Delta+\Sigma^{Z}(i\omega_{n}))^{2}}}\label{Hk}
\end{equation}
and
\begin{eqnarray}
&&G(k,s,i\omega_{n})=\nonumber\\
&&\frac{1}{i\omega_{n}+\mu-\Sigma^{I}(i\omega_{n})-s\sqrt{(\Delta+\Sigma^{Z}(i\omega_{n}))^{2}+\hbar^{2}v_{F}^{2}k^2}}\label{Green}
\end{eqnarray}
In Ref. (19) a factor $\Sigma^{Z\ast}(i\omega_{n})$ was mistakenly
introduced instead of $\Sigma^{Z}(i\omega_{n})$ in (9) and (10).
This leads to small numerical differences but has no qualitative significance.
The spectral function associated with $\widehat{I}$ and $\widehat{\sigma}_{z}$
matrix are
\begin{eqnarray}
A_{I}(k,s,\omega) & = & -\frac{1}{\pi}ImG(k,s,i\omega_{n}\rightarrow\omega+i\delta)\nonumber \\
& \equiv & A(k,s,\omega)
\end{eqnarray}
which is the quantity measured in angular resolved photo emission
spectroscopy ARPES.\cite{Zhou} Further
\begin{eqnarray}
& & A_{z}(k,s,\omega)\notag\nonumber \\
& = & -\frac{1}{\pi}ImTr[\sigma_{z}\hat{G}(k,s,i\omega_{n}\rightarrow\omega+i\delta)]\notag\nonumber \\
& = & -\frac{1}{\pi}Im\{\frac{s[\Delta+\Sigma^{Z}(\omega+i\delta)]G(k,s,\omega+i\delta)}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+(\Delta+\Sigma^{Z}(\omega+i\delta))^{2}}}\}.
\end{eqnarray}
A considerable and instructive mathematical simplification of these
complicated expressions for (11) and (12) result when the imaginary
part of the z-component of the self energy $\Sigma^{Z}(i\omega_{n}\rightarrow\omega+i\delta)$
is ignored. We get
\begin{equation}
A(k,s,\omega)=\frac{1}{\pi}\frac{Im\Sigma^{I}(\omega)}{[\widetilde{\omega}-s\sqrt{M}]^{2}+[Im\Sigma^{I}(\omega)]^{2}}
\end{equation}
\begin{eqnarray}
& & A_{z}(k,s,\omega)\notag\nonumber \\
& = & \frac{s[\Delta+Re\Sigma^{Z}(\omega)]A(k,s,\omega)}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+(\Delta+Re\Sigma^{Z}(\omega))^{2}}}
\end{eqnarray}
with
\begin{equation}
\widetilde{\omega}=\omega+\mu-Re\Sigma^{I}(\omega)
\end{equation}
and
\begin{equation}
M=\hbar^{2}v_{F}^{2}k^{2}+(\Delta+Re\Sigma^{Z}(\omega))^{2}
\end{equation}
where $M$ is a function of $k$ and $\omega$. If we further take
the imaginary part of $\Sigma^{I}(\omega)$ to be zero (13) reduces
to a Dirac delta function but the real part of both $\Sigma^{I}$
and $\Sigma^{Z}$ remain which renormalize the single particle energies
and the gap. The probability of occupation of the state $\mathbf{k}$
denoted by $n(k)$ at temperature $T$ is given by\cite{Carbotte}
\begin{equation}
n(k)=\int_{-\infty}^{+\infty}d\omega A(k,s,\omega)f(\omega)
\end{equation}
and this is to be compared with the corresponding expression for the
z-component of spin $S_{z}(k)$
\begin{equation}
S_{z}(k)=\int_{-\infty}^{+\infty}d\omega A_{z}(k,s,\omega)f(\omega).
\end{equation}
We can also calculate the average value of the square root of the
sum of the squares of $x$ and $y$ component of spin which remains
locked perpendicular to momentum but its magnitude is changed by the
electron-phonon coupling
\begin{equation}
\sqrt{S_{x}^{2}(k)+S_{y}^{2}(k)}=\int_{-\infty}^{+\infty}d\omega A_{x-y}(k,s,\omega)f(\omega)
\end{equation}
with
\begin{equation}
A_{x-y}(k,s,\omega)=-\frac{1}{\pi}Im\{\frac{[\hbar v_{F}k]G(k,s,\omega)}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+(\Delta+\Sigma^{Z}(\omega))^{2}}}\}
\end{equation}
The density of electronic states $N(\omega)$ follows from $A(k,s,\omega)$
on integration over $k$
\begin{equation}
N(\omega)=\sum_{\mathbf{k,}s}A(k,s,\omega).\label{DOS}
\end{equation}
This quantity enters scanning tunneling microscopy experiments (STM).\cite{Miller,GLi,Nicol,Zhou,Pound}
The real part of the dynamic longitudinal conductivity which gives
the absorption spectrum for light follows as \cite{Li4}
\begin{eqnarray}
& & Re\sigma_{xx}(\omega)=\frac{e^{2}v_{F}^{2}\pi^{2}}{\omega}\sum_{\mathbf{k}}\int_{-\infty}^{\infty}\frac{d\omega}{2\pi}[f(\omega)-f(\omega+\Omega)]\notag\nonumber \\
& & \times\lbrack A_{I}(\mathbf{k},\omega)A_{I}(\mathbf{k,}\omega+\Omega)-A_{z}(\mathbf{k},\omega)A_{z}(\mathbf{k,}\omega+\Omega)]
\end{eqnarray}
with $A_{I}(\mathbf{k},\omega)=\sum_{s}A_{I}(k,s,\omega)$ and $A_{z}(\mathbf{k},\omega)=\sum_{s}A_{z}(k,s,\omega)$.
Note that both $A_{z}$ and $A_{I}$ enter this quantity even though
it is an in-plane property.
\section{Numerical results for self energy and derived quantities}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig3a.eps} \includegraphics[width=3.2in,height=3.2in]{Fig3b.eps}
\par\end{centering}
\caption{(Color online) The spectral density $A(k,\omega)$ as a function of
energy $\omega$ in meV for three values of momentum $k$ namely $k=0$
(solid black), $k=0.01$ (dotted blue) and $k=0.02$ (dashed red)
in units of the inverse lattice parameter $a$. The top frame is for
$A_{z}(k,\omega)$ (associated with the gap renormalization) and the
bottom frame is for $A_{I}(k,\omega)$ associated with quasiparticle
renormalization. Vertical dotted black lines are at $\omega=\pm\omega_{E}$
with phonon Einstein energy set at 7.5 meV. The gap $\Delta=20$ meV and the bare chemical potential
is $\mu=25$ meV.}
\label{fig3}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.3in,height=3.3in,angle=-90]{Fig4.eps}
\par\end{centering}
\caption{(Color online) A color plot of the gap spectral density $A_{z}(k,\omega)$
as a function of energy $\omega$ in (eV) vertical axis and momentum
$k$ in ($1/a$) horizontal axis, with $a$ the lattice parameter.
The gap $\Delta=10$ meV and the Einstein phonon energy $\omega_{E}=7.5$
meV. }
\label{fig4}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3in,height=3.2in]{Fig5.eps}
\par\end{centering}
\caption{(Color online) The density of states $N(\omega)$ in inverse eV versus
energy $\omega$ in eV. The dotted red line includes the electron-phonon
interaction with $g=10$ in Eq. (6)-(7). The solid black line which
is for comparison is for bare bands ($g=0$). The top frame has a
gap $\Delta=10$ meV while the bottom frame has a gap of 20 meV. }
\label{fig5}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig6.eps}
\par\end{centering}
\caption{(Color online) The real part of the longitudinal dynamic optical conductivity
$Re\sigma_{xx}(\omega)$ in units of $e^{2}/\hbar$ as a function
of photon energy $\omega$ in meV. The dotted red line includes the
electron-phonon interaction with $g=10$ in Eq. (6)-(7). The solid
black line, which is for comparison, is for bare bands ($g=0$). The
top frame has no gap $\Delta=0$ and the bottom frame includes a gap
$\Delta=20$ meV. In both cases the bare chemical potential $\mu=35$
meV. The arrow identifies the phonon structure.}
\label{fig6}
\end{figure}
In lowest order perturbation theory the chemical potential $\mu$
in Eq. (6) and (7) is to be interpreted as having its non interacting,
independent particle value and will be quoted in all the results to
be presented in this paper. With electron-phonon coupling the chemical
potential will in general be shifted from its free band value to $\mu_{interacting}=\mu+Re\Sigma^{I}(i\omega_{n}\rightarrow\omega+i\delta)$.
In Fig. 2 we show our results for $\Sigma(i\omega_{n}\rightarrow\omega+i\delta)$
in meV as a function of energy $\omega$ in meV where we have analytically
continued from Matsubara to real frequencies. The solid black curve
is $Re\Sigma^{I}$ with $Im\Sigma^{I}$ given by the dashed
red curve while the dotted blue is for $Re\Sigma^{Z}(\omega)$
with its imaginary part $Im\Sigma^{Z}(\omega)$ given by the purple
dash-dotted curve. First the real part of
the quasiparticle self energy is not zero at $\omega\simeq0$ but
rather is of the order of $\simeq-5$meV which reduces the chemical
potential of the interacting picture by about 1/4 of its bare band
value here taken to be 25 meV with the gap $\Delta$ equal to 20 meV.
With $g$ set to a value of 10 in Eq. (6) and (7), the mass enhancement parameter $\lambda\sim$ 0.3 as can be seen
in Fig. 2. The parameter $\lambda$ is the slope of the real part of $\Sigma^I(\omega)$ at $\omega=0$, i.e. $\lambda=(d/d\omega)Re\Sigma^I(\omega)$
at $\omega=0$. We have also taken the Einstein phonon energy to be 7.5 meV which is representative of what is seen in many
but not all experiments in $Bi_{2}Se_3$ and has been identified in the work of X. Zhu et al. \cite{Zhu1} as a surface phonon.
As discussed in the introduction coupling to specific optical phonons have been found in other experiments \cite{Zhu2,Pan,Hatch,Chen,LaForge}.
Of course there can also be coupling to acoustic phonons as studied by Giraud et al..\cite{Giraud1,Giraud2}
These authors use a continuum model for the phonons and a deformation potential model to describe their coupling
to the helical surface Dirac fermions with due attention to the modifications brought about by helicity of the charge carriers.
In our simplified model these complications enter only in determining the size of $g$ in Eq. (6) and (7) which we set
through consideration of experiment (i.e. $\lambda\sim$ 0.3). In principle these could also be coupling to electronic modes
but this is not considered in this work which deals with electron-phonon coupling alone. Our choice of gap $\Delta=10$ mev is representative
of the work of Chen et al. \cite{Chen2} who found in ARPES
experiments a gap of $\simeq-7$ mev in $Bi_{2}Se_3$ magnetically dopped with
$Mn$. We note prominent phonon structure at $\omega=\pm\omega_{E}=\pm7.5$
meV by choice, with additional structures at $-\mu+\Delta-\omega_{E}$
and at $-\mu-\Delta-\omega_{E}$. For $\omega>0$ the imaginary
part of the quasiparticle self energy is zero in the interval ($0,\omega_{E}$)
after which its absolute magnitude is finite and increases linearly
with increasing value of $\omega$. This dependence reflects the linear
dependence on energy of the underlying bare density of states. Also
it needs to be negative because $-Im\Sigma^{I}(i\omega_{n}\rightarrow\omega+i\delta)$
is a scattering rate, and hence positive. By contrast the imaginary part of the gap self energy is flat because the extra factor of
$\Delta/\sqrt{\hbar^2v_F^2q^2+\Delta^2}$ appearing in (7) but not in (6) compensates for the density of states variarion which comes from the sum
over $q$. Note also that both self energies have a finite imaginary part in the interval $\omega\in(0,-\omega_E)$ and $\omega<-\mu-\Delta-\omega_E$.
While the real part of the quasiparticle self energy can be both positive or negative, the real
part of the gap self energy is everywhere negative and so decreases
the bare gap $\Delta$ at all $\omega$ considered here. Its magnitude
is everywhere of order 1/4 of the input value of $\Delta$ with some
$\omega$ dependence and is encoded with boson structure. These self
energies have a profound effect on the corresponding spectral densities
of Eq. (11) and (12) as we see in Fig. 3 where the top frame applies
to $A_{z}(k,\omega)=A_{z}(k,s>0,\omega)$ (conduction band) and the
bottom frame is for $A_{I}(k,\omega)=A_{I}(k,s>0,\omega)$. For the
range of values of $k$ shown namely, $k=0.0$, $0.01$ and $0.02$
in units of the inverse of the lattice parameter ($a$), these two
functions do not differ much from each other. This is traced to the
fact that $A_{z}(k,\omega)$ has the extra factor $\frac{\Delta}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+\Delta^{2}}}$
as compared with $A_{I}(k,\omega)$ where we have, for simplicity, neglected renormalizations as these do not change the argument in an important
way. As long as $\hbar v_{F}k$ is small compared with $\Delta$ the
relevant additional factor is near one. For large $k$ however, larger than
those shown in Fig. 3, $A_{z}(k,\omega)$ will become much smaller
in magnitude than $A_{I}(k,\omega)$ by a factor of $\Delta/(\hbar v_{F}k)$.
Returning to the top frame of Fig. 3 we first note in the solid black
curve for $k=0$ (which is the momentum of the bottom of the bare
conduction band) a large, only slightly broaden and shifted in energy,
quasiparticle peak centered approximately at $\simeq6$ meV below
the Fermi surface at $\omega\simeq0$. For the bare bands it would
be at $5$ meV instead. But with correlations there are also additional
features. There is a large increase in $A_{z}(k,\omega)$ starting
at $\omega=-\omega_{E}=-7.5$meV followed by a large boson structure
extending down to $\simeq-12.5$meV where it drops to zero. Such a
feature is also seen in the dashed red curve but it is much smaller.
The lower cut off is identified to correspond to the energy $-\mu-\omega_{E}+\Delta$.
If we had shown the valence band contribution to the spectral density
it would have a further boson structure at $-\mu-\omega_{E}-\Delta$
as in the self energy of Fig. 2. It is clear from this description
that the effect of the electron-phonon interaction on the bare band
is much more complicated than a simple constant shift in gap value
and a slight shift in quasiparticle energies with small broadening.
Note also the clear phonon sidebands at energies above $\omega=\omega_{E}$
in all three curves. The black vertical dotted lines identify energies
$\omega=\pm\omega_{E}$ for easy reference. A different representation
of these changes is given in Fig. 4 where we show a false color plot
of $A_{z}(k,\omega)$ as a function for energy $\omega$ along the
vertical axis and momentum $k$ along the horizontal. As seen in Eq. (12) and (14), $A_{z}$ carries the sign of $s$ and is
positive (red) in the renormalized conduction band and negative (blue) in the renormalized valence band. The phonon at
$\omega=\pm\omega_{E}$ is clearly identified as is the bottom of
the renormalized conduction band at $-\mu-\omega_{E}+\Delta$ and the valence band
phonon structure at $-\mu-\omega_{E}-\Delta$. Particularly striking
is the large modification of the bare band dispersion curves in the
region just above the renormalized conduction band minimum.
The large renormalization effects seen in Fig. 3 and Fig. 4 can have
a profound effect on certain quantities while at the same time have
much more modest manifestations in others as we will now describe.
In Fig. 5 we show results for the density of states (DOS) $N(\omega)$
given in Eq. (21). The top frame is for a gap $\Delta=10$ meV and
the bottom for $\Delta=20$ meV. The red dotted curve is the renormalized
quasiparticle DOS while the black solid curve is the bare band case
shown for comparison. In both frames the bare chemical potential is
set at $25$ meV, which is $5$ meV above the gap in the conduction
band (for the case $\Delta=20$ meV). The energy variable $\omega$ is set so that
$\omega=0$ corresponds to the Fermi surface so that the bottom of
the conduction band is at $-5$ meV in bottom frame and $-15$ meV
in top frame with the top of the valence band at $-45$ meV and $-35$
meV respectively. In both cases the remaining gap between valence
and conduction band has been very much reduced over its bare band
value. The four phonon structures at $\omega=\pm\omega_{E},-\mu-\omega_{E}-\Delta$
and $-\mu-\omega_{E}+\Delta$ are clearly seen with the bottom of
the conduction band given by the energy $-\mu-\omega_{E}+\Delta$.
By comparison, the corresponding boson structures in the real part
of the dynamic longitudinal optical conductivity $Re\sigma_{xx}(\omega)$
which is more closely related to a convolution of two DOS factors and
is given in Eq. (22), are much more modest as seen in Fig. 6. The top
frame is for gapless Dirac fermions ($\Delta=0$) and is included
for comparison while the bottom frame is for the gapped case with
$\Delta=20$ meV. The solid black curves are for the bare band case
with a small residual scattering rate included to broaden out the
Drude peak due to intraband transitions. This peak is large only at small
$\omega$ and is centered at $\omega=0$. The bare chemical potential
is $35$ meV and we see the onset of a second absorption
band coming from the interband transitions which start at $\omega=2\mu$.
These transitions continue up to large energies and provide the so
called universal background. In our units for $Re\sigma_{xx}(\omega)$
which is $e^{2}/\hbar$, this background\cite{Gusynin1,Gusynin2}
has a height of $1/16$. As is clear in the top frame for $\Delta=0$
this height is almost unaffected by the electron-phonon interaction
(red dotted curve). Staying with the top frame we see however the
appearance of the Holstein processes above $\omega=\omega_{E}$ which
provides significant phonon assisted absorption in the photon region
above $\omega_{E}$ and below the main interband absorption edge at
$2\mu$. The other feature to be noted is that, for the correlated
case, the onset of the interband transitions has moved to lower energies
and is now at twice the value of the interacting chemical potential.
The lower frame which is for the gap fermion case has another important
element. As is well known,\cite{Stille,Gusynin1,Gusynin2} when $\Delta\neq0$
there is a peak in the interband transitions just above the threshold
energy which persists up to a few $\Delta$ above the threshold before
the value of the background is reestablished at its universal value.
In the clean limit we have the analytic result
\begin{eqnarray}
Re\sigma_{xx}(\omega) & = & \frac{e^{2}}{4\hbar}[\delta(\omega)\frac{\mu^{2}-\Delta^{2}}{|\mu|}\theta(\mu^{2}-\Delta^{2})+\notag\nonumber \\
& & \frac{\omega^{2}+4\Delta^{2}}{4\omega^{2}}\theta(\omega-2\max(|\mu|,\Delta))]
\end{eqnarray}
where $\theta$ is a heaviside function. We see that, just above the
interband onset the conductivity is larger than its universal value
of $\frac{e^{2}}{16\hbar}$ as is seen most clearly in the solid black curve of
the lower frame of Fig. 6 which is the bare band result. The horizontal
straight line segment indicate $1/16$. We note that, with electron-phonon
(dotted red curve), the magnitude of the absorption in the region of
the edge still remains above the universal background value but now
there is also a small phonon structure highlighted by the vertical
arrow. No such structure is seen in the top frame for $\Delta=0$.
It is the variation with energy of the background (in the presence
of a finite gap) which allows for the phonon structure to be revealed.
\section{Spin texture}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig7.eps}
\par\end{centering}
\caption{(Color online)The z-component of spin $S_{z}(k)$ as a function of
momentum $k$ normalized to $k_{F}$. The solid black line is for
$g=0$ (no electron-phonon interaction) and is for comparison with
the dotted blue curve which includes the self energy with coupling
$g=10$ (Eq. (6)-(7)). }
\label{fig7}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig8.eps}
\par\end{centering}
\caption{(Color online) The magnitude of the in plane spin $S_{x-y}(k)$ as
a function of momentum $k$ normalized to its value at the Fermi surface
$k_{F}$. The solid black curve is for comparison and involves the
bare bands ($g=0$) with gap $\Delta=20$ meV. The dotted blue has
the same gap but includes the electron-phonon renormalizations with
coupling $g=10$ in Eq. (6)-(7). The dashed red is for a smaller value
of gap $\Delta=10$ meV. }
\label{fig8}
\end{figure}
\begin{figure}[tp]
\begin{centering}
\includegraphics[width=3.2in,height=3.2in]{Fig9.eps}
\par\end{centering}
\caption{(Color online) Comparison of the z-component of spin $S_{z}(k)$ of
Eq. (18) as a function of momentum $k$ normalized to its value at
the Fermi surface $k_{F}$ (short dashed red curve) with the probability
of occupation of state $n(k)$ of Eq. (17) (long dashed black curve).
The heavy solid black line gives the case of bare bands for comparison. }
\label{fig9}
\end{figure}
The z-axis spin texture is determined from the spectral density
$A_{z}(k,s,\omega)$ of Eq. (12). Results are presented in Fig. 7
for $S_{z}(k)$ as a function of momentum $k$ normalized to its value
at the Fermi surface ($k_{F}$). $S_{z}(k)$ involves an integral
over frequency $\omega$ of the overlap of $A_{z}(k,s,\omega)$ and
the Fermi function $f(\omega)$ as given in Eq. (18). We only show
results for the conduction band $s=+$. For the valence band $s=-1$
and $S_{z}(k)$ will change sign. The gap $\Delta=20$ meV and the
chemical potential $\mu=25$ meV as before in section III. The solid
black curve is the bare band result for temperature $T=0.1K$. We
see an abrupt drop to zero at $k/k_{F}=1$ because of the thermal
factor which here is basically a Heaviside function with cut off at
the Fermi surface. By contrast the dotted blue curve includes the
electron-phonon renormalization discussed at length in the previous
section. Now the magnitude of $S_{z}(k)$ at $k=0$ is reduced below
its bare band value. More importantly the jump at $k/k_{F}$ is smaller than it is in the solid curve and further, there are finite
tails beyond this momentum which are entirely due to many body correlation
effects that go beyond a bare band description. They can be taken
as representative of other correlation effect such as those due to
electron-electron interactions rather than electron-phonon. Compared
to the modifications found in the spectral densities, the density
of states and the conductivity, these effects are certainly not as
spectacular. Mathematically this results from the fact that $S_{z}(k)$
involves an integration over energy $\omega$ while it is this energy
dependence which reflects most directly the details of the correlation
effects.
The in-plane spin texture is also changed, although the perpendicular
spin momentum locking remains. In Fig. 8 we show results for the momentum
dependence (in units of $k/k_{F}$) of $S_{x-y}(k)$ defined in Eq.
(19). It gives the magnitude of the in-plane spin at momentum $k$.
Without interactions it would be equal to $\frac{\hbar v_{F}k}{\sqrt{\hbar^{2}v_{F}^{2}k^{2}+\Delta^{2}}}\theta(k_{F}-k)$
and is shown as the solid black curve for a gap of $20$ meV and temperature
$T=0.1$K. It starts at zero for $k=0$ and rises monotonically with
a sharp cut off at $k/k_{F}=1$ at zero temperature. When the electron-phonon
interaction is included through Eq. (20) we get the dotted blue
curve which follows closely the solid black curve with the important
difference that it has tails beyond $k=k_{F}$ characteristic of correlation
effects. The dashed red curve is for the same value of $g$
but now the gap has been reduced to $10$ meV. In this case the curve
rises more sharply out of $k=0$, flattens before showing a sharp
drop at $k=k_{F}$ and the correlation tails beyond this drop off
are now considerably larger. The correlation tails seen in $S_{z}(k)$
as well as in $S_{x-y}(k)$ are very similar to the much more familiar
case\cite{Carbotte} of the momentum distribution in a fermion system
with electron-phonon interaction. The probability of occupation of
the state $k$ at zero temperature denoted by $n(k)$ is given in
Eq. (17) and in the bare band picture is a step function $\theta(k_{F}-k)$.
As shown in Fig. 9 long dashed black curve for $T=0.1$K, $g=10$
and a gap $\Delta=20$ meV, $n(k)$ is considerably reduced from value
one throughout the occupied states. It still has a finite discontinuous
jump at $k=k_{F}$ with correlation tail for $k>k_{F}$. The short
red dashed curve is $S_{z}(k)$ previously calculated and presented
here for comparison with $n(k)$. This shows the qualitatively similar
effect of the electron-phonon interaction on these two quantities.
\section{Conclusions}
We calculated the effect of electron-phonon coupling on a system of
helical gapped Dirac fermions. A simple Holstein model with coupling
to a single Einstein phonon of energy $\omega_{E}$ was used. For
massive Dirac fermions two self energies need to be introduced. There
is the familiar quasiparticle renormalizarion which changes the bare
band energies to dressed dispersion curves and provides damping. In
addition there is a second self energy directly associated with modifications
of the gap. It is energy ($\omega$) dependent and complex with the
real part directly modulating the magnitude of the bare band gap.
Both these self energies enter the quasipaticle spectral density $A(k,s,\omega)$
with $k$ momentum $\omega$ energy and $s=\pm$ giving conduction
and valence band respectively. In direct analogy, a second spectral
density $A_{z}(k,s,\omega)$ associated with the gap channel is also
introduced. The first function $A(k,s,\omega)$ can be measured directly
in angular resolved photo-emission spectroscopy (ARPES). Its average
over momentum $k$ determines the electronic density of states $N(\omega)$
measured in scanning tunneling spectroscopy (STM). The second $A_{z}(k,s,\omega)$
enters along with $A(k,s,\omega)$ the expressions for the dynamic
longitudinal optical conductivity. It also determines the out of plane
spin texture. Both $A$ and $A_{z}$ are functions of $\omega$ and
are encoded with sharp phonon structures at $\omega=\pm\omega_{E},-\mu-\omega_{E}-\Delta$
and $-\mu-\omega_{E}+\Delta$. These phonon structures manifest directly
in the density of states and in a somewhat different way in the optical
conductivity.
The out of plane and in plane spin textures are modified by the electron-phonon
interaction. But these quantities involve an integration over energy
of an overlap of spectral density and thermal factor. Consequently
the manifestation of phonon-electron coupling in these quantities
is more subtle and not as direct. Nevertheless important corrections
to a bare band picture arise. Just as the probability of occupation
of a state of momentum $k$ is reduced from one for $k$ less than
$k_{F}$, and extended (in momentum) tails appear for $k>k_{F}$,
the magnitude of $S_{z}(k)$ follows the same trend. A similar picture
applies for the magnitude of the in plane spin component. While the
spin remains locked in the direction perpendicular to momentum, correlation
tails appear beyond $k=k_{F}$. While the calculations are for the explicit
case of the electron-phonon interaction, they serve to illustrate
how many body interactions in general modify the bare band picture.
\begin{acknowledgments} This work was supported by the Natural Sciences
and Engineering Research Council of Canada (NSERC) and the Canadian
Institute for Advanced Research (CIFAR). This material is also based upon
work supported by the National Science Foundation under the NSF EPSCoR Cooperative
Agreement No. EPS-1003897 with additional support from the Louisiana Board of Regents.
JPC initiated the work and wrote much of the text, ZL did
all the calculations.
\end{acknowledgments}
\section*{References}
|
\section{Introduction}\label{intro}
Detailed modeling and simulation of large, complex networks are increasingly being used in several fields (from biology, epidemiology and economics, to traffic management and supply-chains).
As a result, there have been several empirical studies of the interaction between the structure of such networks and their dynamics (a ``network'' analog of
structure-property relations in materials science).
Detecting, understanding, and eventually exploiting this structure-dynamics relation in networks is thus a subject of intense interdisciplinary research.
Clearly, the construction of networks with prescribed structure is an integral part of this effort, and the purpose of this paper is
to present and illustrate a computer-assisted, mixed integer linear programming (MILP) approach to this important task.
There are many important examples where the topology of real world networks has been analyzed in great detail in order to
understand their formation and dynamics - the Internet holds probably pride of place among these.
A brief history of the development of Internet ``topology generators'' (the construction of networks that hopefully
model the Internet) can be found in Tangmunarunkit~\textit{et al.}~\cite{Tang02network}.
The earliest internet topology generators were based on the \emph{Erd\"{o}s-R\'{e}nyi} model.
Realistic simulations of dynamical models of the internet, however, required topology generators that more accurately represented its structure. Hence, better models that take into account the hierarchy in the internet were later proposed~\cite{Calv97modeling}, and these were widely accepted until the breakthrough results by Faloutsos~\textit{et al.}~\cite{Falo99power-law} revealed that the topology of the internet exhibits power-law dependence in terms of the degrees of the nodes.
Networks with such power-law degree distributions are called \emph{scale-free} networks, and are in stark contrast with networks featuring the Poisson distribution of degrees created by topology generators based on the Erd\"{o}s-R\'{e}nyi model.
This suggests that local properties of real networks are not accurately reproduced by simple graph models.
While a lot of attention has focused on designing realistic models of the Internet, the utility of graph models that produce power-law degree distributions is not restricted to the Internet domain alone.
Several systems have been found to have power law degree distributions with different degree exponents.
Newman~\cite{Newm03structure} and Albert and Barab\'{a}si~\cite{Albe02statistical} review graphs from numerous systems that adhere to a power-law degree distribution.
These discoveries create the need for better alternative approaches to generating models of real world networks.
In response to this need, Barab\'{a}si and Albert~\cite{Bara99emergence} themselves proposed an evolutionary process to create scale-free networks.
In their model, the network grows by the addition of nodes which preferentially attach to the existing nodes that are already well connected. This model is termed in the literature as the ``preferential attachment'' model.
A different approach is used by Aiello~\textit{et al.}~\cite{Aiel00random} to create random graphs with power-law degree distribution.
They first generate sequences of degrees from a power law distribution and then create networks that satisfy this degree distribution.
These two approaches exemplify two distinct (yet at some level related) approaches by which realistic network properties (i.e., the scale-free structure in this case)
can be ``built in'' model networks.
The first approach, which one might name \emph{complex network growth models} is to use {\em dynamical construction procedures} (evolutionary or growth processes) that give rise to networks with required properties.
Examples of such models that have been extensively used in complex systems research are the Watts-Strogatz model for small-world networks~\cite{Watt98collective} and the aforementioned preferential attachment model that creates scale-free networks~\cite{Bara99emergence}.
While such complex network growth models may provide insights into the microscopic processes that govern the formation of real world networks,
they will in general create graphs that do not fully capture several properties of interest.
For example, Bu~\textit{et al.}~\cite{Bu02distinguishing} show that Internet topology generators produce graphs which have different values of characteristic path lengths and clustering coefficients compared to the actual internet graph.
The second approach to network generation is built around the requirement to create graphs with specified properties.
Such methods might be named \emph{network generation algorithms}.
The main feature of this second approach is the fact that {\em combinations of several properties} (beyond, for example, just the degree distribution)
can be explicitly taken into account.
Newman~\cite{Newm03structure} provides a list of different graph properties that can be ``built in'' network generation algorithms,
and discusses their relevance in real systems.
For instance, the average path length of social networks is found to be small for even very large networks, which characterizes what is known as the \emph{small-world property}. This has direct implications on the diffusion of information or the spread of epidemics on such networks.
Shirley and Rushton~\cite{Shir05impacts} study the impact of the structure of four different types of network topologies, namely Erd\"{o}s-R\'{e}nyi, regular lattices, small-world, and scale-free, on epidemic dynamics.
They show that in scale-free networks with very small clustering and short path lengths, the spread of the disease is faster, while in regular lattices with high clustering coefficients and long path lengths, the spread is the slowest among all other types of networks.
Badham and Stocker~\cite{Badh10impact} report that increases in assortativity and
clustering coefficient are associated with smaller epidemics.
The effect of correlations between neighboring node degrees in a network has also been found to be a crucial property in determining system behavior in a variety of physical systems.
Miao {\em et al.}~\cite{Miao08effects} study the effect of degree-degree correlations on the pinning controllability of networks and find that dissortative mixing enhances controllability. The results of Pechenick {\em et al.}~\cite{Pech12influence} show that chaotic
gene regulatory networks exhibit increasingly robust phenotypes with increasing assortativity.
Pastor-Satorras {\em et al.}~\cite{Past01dynamical} report that disassortative networks enhance transport through them.
They find that ``the pressure congestion for disassortative (assortative) networks is lower (bigger) than the one for uncorrelated networks.''
Watts and Strogatz~\cite{Watt98collective} report that oscillators connected by a small-world structure synchronize more readily than those connected by other network topologies.
Hong~\textit{et al.}~\cite{Hong04factors} find other important properties that affect the synchronizability of oscillators connected by a Watts-Strogatz network~\cite{Watt98collective}.
They find that ``better synchronizability for the WS small-world network is induced as the heterogeneity
of the degree distribution is increased, as the characteristic path length is decreased, as the maximum betweenness centrality is reduced, or as the maximum degree is raised.''
Thus, we find that properties like the maximum degree and the maximum betweenness centrality are also relevant for some dynamical network systems.
Ravasz and Barab\'{a}si~\cite{Rava03hierarchical} find that the clustering coefficients of the nodes are proportional to the inverse of the nodal degrees for four real networks (actor network, semantic web, world wide web and the internet).
They propose this to be related to the hierarchical organization of the network.
They suggest that ``the fundamental discrepancy between models and empirical measurements is rooted in a previously disregarded, yet generic feature of many real networks: their hierarchical topology.''
Thus, it may be important to maintain the ``degree-dependent clustering'' in models of network topology.
All these results clearly point to the necessity of being able to generate graphs with specified properties, not only to replicate realistic structures of real networks but also to create artificial networks that will result in desired functions.
One of the main difficulties in accurately studying the effect of network properties on the function of networked systems is disassociating the effect of one property from the effect of others.
Wrong conclusions will be derived when the network topology generator induces artificial correlations between different sets of properties.
It is clear that graph-generating algorithms must be able to create networks with controllably-tunable combinations of graph properties.
A majority of existing network generation algorithms focus on constructing networks with specified degree statistics.
The Havel-Hakimi algorithm~\cite{Haki62realizability,Have55remark} is a deterministic algorithm that guarantees a network for any given graphical sequence of degrees. However, the algorithm always creates the same network for a given input degree sequence, which may be a limitation in certain applications.
The ``pairing'' or ``configuration'' model~\cite{Boll80probabilistic,Worm80some} is a stochastic algorithm that alleviates this problem, but it suffers from non-uniform sampling.
Strategies to generate uniformly distributed random graphs with specified degree sequences are discussed in a number of publications~\cite{Blit06sequential,Mil03uniform,Vige05efficient}.
%
The Chung-Lu algorithm~\cite{Chun02connected,Mill11efficient} is another probabilistic algorithm that creates a random graph from an ensemble of graphs with a given {\em expected} degree sequence.
Besides the degree sequence, clustering and assortativity are two other properties that have received considerable attention in the context of network topology generation algorithms.
Algorithms that tune clustering coefficients in scale-free networks are given in~\cite{Fu02evolving,Holm02growing}, while an algorithm for small-world networks can be found in~\cite{Wang08evolving}.
Several algorithms have also been proposed to construct graphs with specified clustering coefficients in conjunction with given degree distributions~\cite{Bans09exploring,Deij09random,Kim04performance,Volz04random}.
Procedures to create networks with specified assortativity have also been introduced~\cite{Guo06growing, Newm02assortative, Newm03mixing}.
An algorithm that controls both the degree distribution and assortativity can be found in Xulvi-Brunet and Sokolov~\cite{Xulv04reshuffling}. The problem of constructing graphs with given maximum degree and diameter that has the largest number of vertices has also been studied~\cite{Aleg86some,Berm84large,Fell95large}.
In addition, Nakano~\textit{et al.}~\cite{nakano} discuss an algorithm to construct weighted graphs such that the total weight of the edges incident to a node is at least the given weight of the node.
Finally, other property combinations that have been addressed in the context of network generation algorithms include the
(a) degree distribution and degree-dependent clustering~\cite{OPTLpaper,BTER_2014,Serr05tuning,BTER_2012},
(b) joint degree-degree distribution~\cite{Doro02how,Webe07generation},
(c) joint degree-degree distribution and modularity~\cite{Zhou10generation},
(d) degree-degree correlations and degree-dependent clustering~\cite{Pusc08generating}, and
(e) eigenvector centrality~\cite{Nico12controlling}.
The majority of the above algorithms cannot guarantee the generation of networks that satisfy exactly the properties specifications, or that lie within bounded deviations from the desired values.
In our earlier work~\cite{OPTLpaper}, we introduced a method to specify degree distribution in conjunction with a specified number of triangles among nodes of given degree (i.e., degree-dependent clustering levels).
The method provided theoretical guarantees that a network satisfying the imposed specification would be constructed, if such a network existed.
It also constituted the first time in the open literature that a network generation approach based on mathematical optimization was proposed. In particular, the problem was posed as a Mixed-Integer Linear Problem (MILP), and as a result, the method offered a number of advantages over traditional approaches.
For example, in addition to identifying a network that possesses a collection of desired properties, an MILP-based framework is typically suitable to
(a) certify if no such network exists,
(b) identify multiple, or even all, non-isomorphic such networks,
(c) identify the network that is most similar --under a suitable metric-- to a reference network,
(d) identify the network that minimizes/maximizes some application-specific objective,
(e) determine the smallest size $N_{0}$ for a network to admit all properties, and
(f) identify networks that approximately satisfy the properties within given tolerances.
Posing the problem as an MILP allows for employing standard, well-developed solution algorithms (e.g., branch-and-bound, branch-and-cut) to deal with the complex decision-making that is involved.
It is thus a promising approach towards developing the capability to systematically design networks that possess desired properties with explicit guarantees of achieving the specifications exactly or within bounded deviations.
In this paper, we demonstrate the scope and generality of the MILP-based approach, and in particular the associated modeling techniques required to handle a wide variety of properties.
The remainder of the paper is structured as follows: after formalizing the Network Generation Problem, Section~\ref{modeling} introduces the modeling techniques that can be used to formulate the mathematical expression and specification of any network property of interest in an optimization context.
Section~\ref{popular} then develops model ``snippets'' that correspond to a number of popular network properties.
Section~\ref{enhancements} addresses modeling and implementation issues that can enhance the performance of optimization algorithms in the context of the Network Generation Problem.
Section~\ref{studies} discusses how to put together complete optimization formulations for the design of networks that have tunable levels of connectivity, spread, assortativity and robustness.
Although by no means exhaustive for the capabilities of the proposed optimization-based framework, these illustrative applications were selected so as to showcase its versatility and flexibility with regards to network properties that can be addressed.
A number of computational case studies are also provided for illustration.
Finally, Section~\ref{conclusions} presents our concluding remarks.
Throughout the paper, we use boldface to indicate vector quantities.
\section{Modeling the Network Generation Problem}\label{modeling}
The Network Generation Problem (NGP) calls for the identification of the connections (edges) that need to exist between its nodes in order for the network to attain specific values, or values within specific ranges, for a set of properties of interest.
By ``properties'' we refer to various characteristic quantities of networks including global metrics (e.g., the global clustering coefficient), local metrics (e.g., a node's local clustering coefficient), nodal distributions (e.g., the degree sequence), network distances (e.g., shortest paths among nodes), and densities of distinctive patterns (e.g., total number of triangles).
We remark that, as different properties result in different admissible (feasible) network configurations, the NGP seeks configurations that are simultaneously feasible with respect to all specified properties.
In order to address the NGP, we will formulate optimization models according to two major aims.
The first is to concisely mathematically express each of the desired properties so that they can be conveniently evaluated given any network configuration that is explored during the solution process.
The second aim is to appropriately restrict the feasible solution space into only those network configurations that satisfy all property specifications.
Both aims are to be met with an appropriate selection of optimization variables and equality or inequality constraints, something that is discussed in detail in Sections~\ref{express} and~\ref{specify}.
Throughout the analysis, we preserve an overall MILP structure for our models and strive to involve expressions that have as tight LP-relaxations as possible.
We also focus on creating modular formulations, which we achieve by describing each property with an independent set of constraints that can be included as a building block to an overall formulation only when the
property in question is part of the problem specifications.
The modularity of our formulations is illustrated later in Section~\ref{studies}, where complete formulations are put together by utilizing the constraints presented as stand-alone examples in Sections~\ref{popular} and~\ref{enhancements}.
Let $V = \left\{1,2,\cdots,N\right\}$ be the set of the network's nodes, where $N$ is an {\em a priori} known constant - here the maximum allowable number of nodes.
Depending on the application, a set of properties may explicitly or implicitly impose that all $N$ nodes are non-isolated (i.e., have degree of at least one), or even that the network has a single component (i.e., is connected).
Let also $E=\left\{(i,j) \in V \times V : i < j\right\}$ be the set of all (undirected) edges that can possibly form between the nodes in $V$.
As the cases $N=0$ and $N=1$ are trivial for the NGP, we always assume for convenience that $N \geq 2$.
For ease of exposition, this paper focuses on undirected, simple networks.
Extensions to more general classes of networks can be handled by appropriately augmenting the
set $E$.
For example, loops can be considered by allowing $i =j$, while directed arcs can be introduced by
allowing $j > i$.
Furthermore, we consider nodes to be unlabeled; that is, the nodes are permutable with
each other and the specification does not provide any particular reference to node labels.
We remark that the unlabeled case is in fact the more general and ``difficult'' variant of
the NGP, as achieving a more abstract specification must in general be more involved.
\subsection{Expressing Properties}\label{express}
The basic set of optimization variables utilized in our framework is a set of binary variables $x_{ij}$ to denote the existence of each edge $(i,j) \in E$:
\begin{equation}
x_{ij} = \left\{ \begin{array}{r@{\quad:\quad}l}
1 & \text{if edge $(i,j)$ exists in the solution}\\
0 & \text{otherwise.} \\
\end{array} \right.\label{xijdef}
\end{equation}
The $x$-variables suffice to fully define the configuration of the network.
Once their values are determined, every property of the network can be uniquely evaluated, and whether the property specifications are satisfied or not can be unambiguously determined.
However, only a handful of properties can be expressed in closed form as a linear function of these variables.
A notable example is the degree of a node, which can be expressed as the sum of all edges incident to that node.
Other properties will typically necessitate the introduction of additional auxiliary variables to facilitate their mathematical description.
For example, properties such as the global clustering coefficient require the encoding of whether particular features (in this case, triangles) have been formed or not, while expressing properties such as the network diameter would require the encoding of the shortest-path lengths among each pair of nodes.
To that end, it is often convenient to introduce a new set of variables $y_f$ to capture the formation of each feature $f \in F$, where $F$ is the set of features of interest, as well as a new set of variables $w_{ij}$ to express the length of the shortest path between each pair of nodes $(i,j) \in E$.
Restricting these variables to their intended values is discussed in Sections~\ref{features} and~\ref{shortestpaths}.
In general, most properties of interest can be expressed directly as a linear
combination of the $x$-, $y$- and $w$-variables:
\begin{equation}
p = \sum\limits_{(i,j) \in E} \alpha_{ij} x_{ij} + \sum\limits_{f \in F} \beta_f y_f + \sum\limits_{(i,j) \in E} \gamma_{ij} w_{ij} + \delta \label{pidef},
\end{equation}
where $\mathbf{\alpha}$, $\mathbf{\beta}$, $\mathbf{\gamma}$ and $\delta$ are parameters that depend on the type of property, and $p$ is an auxiliary variable to capture the property's value.
The use of an auxiliary $p$-variable helps the model from a human-readability and modularity perspective.
Furthermore, it facilitates the expression (in a linear fashion) of more complex, higher-level properties.
For example, once the local clustering coefficients of each node are represented by an appropriately defined set of $p$-variables, the average clustering coefficient of the entire network can be expressed based on their sum.
We note that, irrespectively of whether it represents a discrete or a continuous quantity, every $p$-variable may be declared to the MILP solver as a continuous variable within appropriate lower and upper bounds, $p^L$ and $p^U$.
These bounds can be readily derived based on interval arithmetic considerations.
We further remark that most commercial MILP solvers will automatically detect the special manner in which these additional variables participate in the formulation and remove them during a preprocessing phase.
Therefore, it is not expected that the computational performance of the solution process will be burdened by these types of modeling simplifications.
Examples of expressing a number of popular properties are provided in Section~\ref{popular}.
\subsubsection{Encoding the Existence of Features (Motifs)}\label{features}
The existence or absence in the designed network of characteristic patterns (e.g., triplets, triangles, cliques, stars) may be of interest as a result of some higher-level property (e.g., global clustering coefficient, clique number).
In such cases, the explicit encoding of the existence of these features is required to enable the expression of the desired property.
We observe that the formation of any feature (motif) corresponds to the concurrent existence and/or absence of certain edges.
For example, a triangle between nodes $1$, $2$ and $3$ exists if and only if all three edges $(1,2)$, $(1,3)$ and $(2,3)$ exist or, equivalently, if and only if $x_{12}x_{13}x_{23} = 1$.
Similarly, an open triplet along trajectory $2$--$1$--$3$ exists if and only if edges $(1,2)$ and $(1,3)$ exist, while edge $(2,3)$ does not, or if and only if $x_{12}x_{13}\left(1-x_{13}\right) = 1$.
For every feature $f \in F$, where $F$ is the set of all features of interest, let $E^{+}_f \subseteq E$ and $E^{-}_f \subseteq E$ be the corresponding collections of edges that need exist and be absent, respectively, and let a variable $y_f$ be defined as
\begin{equation}
y_{f} = \left\{ \begin{array}{r@{\quad:\quad}l}
1 & \text{if feature $f$ exists in the solution}\\
0 & \text{otherwise.} \\
\end{array} \right.\label{yfdef}
\end{equation}
In order to maintain linear structure in the model, the variable $y_f$ is not expressed directly as the product $\prod\limits_{(i,j) \in E^{+}_f} x_{ij} \prod\limits_{(i,j) \in E^{-}_f} \left(1-x_{ij}\right)$; rather it is restricted to an appropriate binary value through the following set of constraints:
\begin{eqnarray}
y_{f} &\leq& x_{ij} \;\;\;\;\;\;\;\;\;\;\;\;\; \forall (i,j) \in E^{+}_f \label{yf1a} \\
y_{f} &\leq& 1 - x_{ij} \;\;\;\;\;\; \forall (i,j) \in E^{-}_f \label{yf1b} \\
y_{f} &\geq& 1 - \sum\limits_{(i,j) \in E^{+}_f} \left(1-x_{ij}\right) - \sum\limits_{(i,j) \in E^{-}_f} x_{ij} \label{yf2} \\
y_{f} &\geq& 0 \label{yf3}.
\end{eqnarray}
We remark that, at any feasible solution (where all $x$-variables have attained binary values), the above constraints restrict the value of $y_f$ to either $0$ or $1$ without any explicit claim about its integrality.
This means that variables $y_f$ need not be defined as binary --rather as mere continuous-- in the context of the MILP solution algorithm.
An alternative to the $O (|E^{+}_f|+|E^{-}_f|)$ constraints~(\ref{yf1a}) and~(\ref{yf1b}) are the two aggregate constraints
\begin{eqnarray}
y_{f} &\leq& \frac{1}{|E^{+}_f|} \sum\limits_{(i,j) \in E^{+}_f} x_{ij} \label{yf1a'}\\
y_{f} &\leq& 1 - \frac{1}{|E^{-}_f|} \sum\limits_{(i,j) \in E^{-}_f} x_{ij}. \label{yf1b'}
\end{eqnarray}
Use of constraints~(\ref{yf1a})--(\ref{yf1b}) is preferred, as they result into tighter LP-relaxations.
Furthermore, their total-unimodularity endowes the overall LP-relaxations with an increased tendency to admit integral solutions.
The alternative constraints~(\ref{yf1a'})--(\ref{yf1b'}) are attractive in cases where the size of the formulation becomes an issue and has to be limited within an acceptable level.
We note that when using the aggregate constraints, binarity of variable $y_f$ is not guaranteed and must be explicitly declared to the MILP solver.
Examples of defining a number of common features are provided for illustration in Section~\ref{popular}.
\subsubsection{Encoding Shortest Paths}\label{shortestpaths}
A number of properties (e.g., betweeness-centrality, network diameter) depend on shortest-path information.
To that end, the explicit encoding of the shortest path between any node pair of interest $i \in V$ and $j \in V \setminus{\left\{i\right\}}$ is required.
Since a shortest path is itself an optimal quantity (out of all the paths that connect nodes $i$ and $j$, it is the one with the minimum length), it can be obtained via the solution of a mathematical optimization formulation.
Given a network configuration $\mathbf{x}$, the following formulation applies:
\begin{eqnarray}
&\min\limits_{\mathbf{f^{ij}}} & \sum\limits_{k \in V} \sum\limits_{l \in V} f^{ij}_{kl} \label{sp-p0} \\
& \text{s.t.} & f^{ij}_{kl} + f^{ij}_{lk} \leq x_{kl} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (k,l) \in E \label{sp-p1} \\
& & \sum\limits_{l \in V} f^{ij}_{kl} - \sum\limits_{l \in V} f^{ij}_{lk} = \xi^{ij}_k \;\;\;\;\;\; \forall k \in V \label{sp-p2} \\
& & f^{ij}_{kk} = 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall k \in V \label{sp-p3} \\
& & f^{ij}_{kl} \geq 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\, \forall (k,l) \in V \times V \label{sp-p4},
\end{eqnarray}
where parameters $\xi^{ij}_i = +1$, $\xi^{ij}_j = -1$ and $\xi^{ij}_k = 0$, if $k \neq i,j$.
In the above formulation, each variable $f^{ij}_{kl}$ represents whether arc $(k \rightarrow l)$ participates or not in the shortest path between nodes $i$ and $j$.
Constraint~(\ref{sp-p1}) restricts participation to only those arcs that are eligible; that is, to only those arcs that correspond to edges that exist in the network.
The formulation ensures the integrality of the objective function at optimality.
To that end, it suffices that the $f$-variables be declared to the optimization solver as continuous.
We remark however that, due to the possibility of degeneracy (i.e., existence of more than one distinct shortest paths), some $f$-variables may admit fractional (i.e., neither $0$ nor $1$) values at optimality.
In cases when we are not only interested in the length of the shortest path but also interested in encoding the actual shortest path, the $f$-variables should be declared as binary.\footnote{For better performance of the branch-and-bound-based search process, it is advisable to deprioritize the $f$-variables during branching considerations. By doing so, the $f$-variables will effectively be treated as continuous throughout the search process, and only at the very end (and only if necessary) will the solver take branching decisions on $f$ to resolve the degeneracy.}
An alternative formulation that can be used to obtain the shortest path between nodes $i$ and $j$ is the dual of the above (primal) formulation, which can be cast as follows:
\begin{eqnarray}
&\max\limits_{\mathbf{t^{ij}},\mathbf{u^{ij}},\mathbf{v^{ij}}} & \sum\limits_{(k,l) \in E} u^{ij}_{kl} + t^{ij}_i - t^{ij}_j \label{sp-d0} \\
& \text{s.t.} & \hspace{-0.23cm} \left.\begin{array}{rcl}
v^{ij}_{kl} + t^{ij}_{k} - t^{ij}_{l} &\leq& 1\\
v^{ij}_{kl} + t^{ij}_{l} - t^{ij}_{k} &\leq& 1\end{array}\right\} \;\;\;\; \forall (k,l) \in E \label{sp-d1} \\
& & u^{ij}_{kl} \leq v^{ij}_{kl} + U \left( 1 - x_{kl} \right) \;\;\;\;\;\; \forall (k,l) \in E \label{sp-d2} \\
& & t^{ij}_k \in {\mathbb R} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall k \in V \label{sp-d3} \\
& & u^{ij}_{kl}, v^{ij}_{kl} \leq 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (k,l) \in E \label{sp-d4},
\end{eqnarray}
where $\mathbf{t^{ij}}$, $\mathbf{u^{ij}}$ and $\mathbf{v^{ij}}$ are continuous variables, and $U=N-2$ is an appropriate ``big-M'' coefficient.
Although the aforementioned two formulations (the primal and the dual) are available to us for computing the shortest path between two given nodes of interest, none of the two can be readily utilized in the context of the NGP.
The complication arises due to the fact that the shortest path is only encoded when the formulations
are solved to optimality.
This cannot be guaranteed in general, because the objective function of the overall NGP formulation may have to be reserved for some other application-specific functional.
Furthermore, when shortest paths between additional pairs of nodes need to be encoded at the same time, as is often the case, then it would not be possible to concurrently optimize multiple interdependent --and generally competing-- objectives.
To address this shortfall, we use the strong duality theorem of linear programming, which ensures that when no gap exists between a feasible solution of the primal problem and a feasible solution of the dual problem (i.e., when these two solutions have equal objective values), then both solutions are the optimal solutions of their respective problems.
Given this fact, an explicit encoding of shortest-path information that does not utilize the overall formulation's objective function can be achieved by appending all constraints of both the primal and dual shortest path formulations (i.e., constraints~(\ref{sp-p1})--(\ref{sp-p4}) and~(\ref{sp-d1})--(\ref{sp-d4})), and in addition, appending the two equality constraints~(\ref{sp-o1}) and~(\ref{sp-o2}) to enforce that the objectives of the primal and dual formulations simultaneously attain a common value, $w_{ij}$.
\begin{eqnarray}
w_{ij} &=& \sum\limits_{k \in V} \sum\limits_{l \in V} f^{ij}_{kl} \label{sp-o1} \\
w_{ij} &=& \sum\limits_{(k,l) \in E} u^{ij}_{kl} + t^{ij}_i - t^{ij}_j \label{sp-o2},
\end{eqnarray}
where $w_{ij} \in \left[ 1, N-1 \right]$ are new variables that can be declared to the MILP solver as continuous variables (the constraints guarantee that they will attain discrete values).
We remark that an implication of encoding the shortest path between a pair of nodes $i$ and $j$ is the fact that the search is restricted to networks in which $i$ and $j$ participate in the same network component.\footnote{If shortest paths are encoded for all pairs $(i,j) \in E$, then the search is restricted to connected networks.}
This is due to the implicit requirement of the optimization process that any variable of the formulation attains a finite value.
However, in applications where the specified properties involve shortest paths (or some other property that depends on shortest paths), it is typically the case that the existence of the shortest paths is indeed desirable.
\subsection{Specifying Properties}\label{specify}
In the context of the NGP, the set of properties of interest should in general attain values within specified ranges.
Note that this generic specification admits also the cases where there is only a single (either lower or upper) required bound, as well as the case where a property must equal some specified value exactly.
In the following, we describe how range restrictions can be imposed for linear, fractional, and sequence properties, as well as restrictions on collective statistics of those.
Examples of imposing restrictions for a number of popular properties will be presented in Section~\ref{slacks}.
We remark that the larger the number of property specifications, the smaller the feasible solution space is.
To that end, property combinations should be chosen judiciously so as not to inadvertedly correspond to infeasible problem instances (i.e., instances where no feasible network configuration exists).
\subsubsection{Linear Properties}\label{specify_lp}
We define a linear property $lp$ to be any property that can be directly expressed as a linear combination of any of the formulation's declared variables (e.g., the total number of triplets or triangles in the network can be expressed as sums of appropriately defined $y$-variables).
This definition obviously admits any property that can be expressed as a $p$-variable according to constraint~(\ref{pidef}).
Specifying that $lp$ attains some value in the range $\left[\pi^{L},\pi^{U}\right]$ can be trivially achieved by appending in the overall formulation the explicit bounds
\begin{equation}
\pi^L \leq lp \leq \pi^U \label{lp_bounds}.
\end{equation}
It is assumed that the linear equality constraint that constitutes the definition of property $lp$ has also been included in the formulation.
\subsubsection{Fractional Properties}\label{specify_fp}
We define a fractional property $fp$ to be any property that can be expressed as the ratio of two linear properties, $fp = \frac{lp^A}{lp^B}$, where the linear property $lp^B$ cannot attain the value of zero.
An example of a fractional property would be the global clustering coefficient, which is defined as thrice the ratio between the total triangles and triplets in the network.
Specifying that a fractional property attains some value in the range $\left[\pi^{L},\pi^{U}\right]$ can be achieved by appending in the overall formulation the constraints\footnote{Constraints~(\ref{fp_bounds}) apply for strictly positive $lp^B$.
If instead $lp^B$ is strictly negative, the inequality signs should be reversed.}
\begin{equation}
\pi^L lp^B \leq lp^A \leq \pi^U lp^B \label{fp_bounds}.
\end{equation}
It is assumed that the linear equality constraints that constitute the definitions of properties $lp^A$ and $lp^B$ have also been included in the formulation.
\subsubsection{Sequence Properties}\label{specify_seq}
Most specified properties are scalar quantities (e.g., the clustering coefficient of a node, the girth of the network).
However, a property may also be a vector; a case in point would be the scalar values that a collection of nodes attains for some other, scalar property.
For example, the degree sequence of all neighbors of a given node $i$ is a vector property.
If we want to impose it on the network to be constructed, we need to have explicitly encoded which node $j$ (that is a neighbor of node $i$) attains which value from the sequence.
More formally, let a node-specific property of interest and let variables $p_i$ be this property's values for each node $i$, given a network configuration $\mathbf{x}$.
It is assumed that the $p$-variables attain appropriate values at any feasible solution (e.g., as a result of constraints of type~(\ref{pidef}) having been included in the formulation).
In addition, let a subset of $M$ nodes ($M \leq N$), and let participation of a node $i \in V$ in this subset be dictated by an appropriately defined binary variable $z_i$; that is, $z_i = 1$, if node $i$ participates in the subset, and $z_i = 0$, if it does not, such that $\sum\limits_{i \in V} z_i = M$ (the activation of the $z$-variables is formally discussed in Section~\ref{subsets}).
Finally, let a sequence of ranges $\mathbf{\Pi} = \left\{ \left[\pi_1^L,\pi_1^U\right], \left[\pi_2^L,\pi_2^U\right], \cdots, \left[\pi_M^L,\pi_M^U\right] \right\}$ that is to be specified on the node subset for the property of interest.
In other words, it is desired that the solution exhibits a one-to-one correspondence between the $p_i$ values for the $M$ nodes and the ranges in the sequence $\mathbf{\Pi}$.
We note that the admissible correspondence may not be unique, if any two ranges in $\mathbf{\Pi}$ overlap.
We introduce a new set of binary variables $q_{im}$ to denote the assignment of nodes to elements of the sequence:
\begin{equation}
q_{im} = \left\{ \begin{array}{r@{\quad:\quad}l}
1 & \text{if node $i$ is assigned to element $m$}\\
0 & \text{otherwise,} \\
\end{array} \right.\label{qimdef}
\end{equation}
where $z_i = 1$ and $\pi_m^L \leq p_i \leq \pi_m^U$ are necessary conditions for assignment.
The definition of the $q$-variables and the specification of the sequence is enforced via the following set of constraints:
\begin{eqnarray}
&\left.\begin{array}{rcl}
p_i &\geq& p_i^L - \left(p_i^L - \pi_m^L \right) q_{im} \\
p_i &\leq& p_i^U - \left(p_i^U - \pi_m^U \right) q_{im}\end{array}\right\}& \forall (i,m) \in V \times \left\{1,2,\cdots,M\right\} \label{qim1} \\
& \sum\limits_{i \in V} q_{im} = 1& \forall m \in \left\{1,2,\cdots,M\right\} \label{qim2} \\
& \sum\limits_{m = 1}^M q_{im} \leq z_{i}& \forall i \in V \label{qim3}\\
& \sum\limits_{m = 1}^M q_{im} \geq z_{i}& \forall i \in V \label{qim4},
\end{eqnarray}
where $p_i^L$ and $p_i^U$ are appropriate lower and upper limits for the value of variable $p_i$.
We remark that in cases where the sequence refers to nodes belonging to the intersection of multiple subsets, the $q$-variables need to be appropriately restricted.
This can be achieved by (i) appending separate constraints~(\ref{qim3}) for each applicable subset and (ii) replacing the right-hand side in constraints~(\ref{qim4}) with an aggregate term $1 - \sum\limits_e \left(1 - z_i^e\right)$ ($e$ is used here to index over the collection of subsets of interest and $z_i^e$ are the corresponding $z$-variables for subset $e$).
\subsubsection{Statistical Properties}\label{specify_st}
Instead of specific values or sequences, one may want to specify collective statistics,
such as the arithmetic mean or the median of a given property.
Typically these statistical metrics are to be evaluated across the complete set of nodes (e.g., the case of the network's average clustering coefficient).
However, in certain contexts they may refer only to a subset of nodes, which could itself be a variable quantity that depends on the network configuration $\mathbf{x}$ (e.g., the case of the average degree among the neighbors of a given node).
More formally, let a node-specific property of interest and let variables $p_i$ attain appropriate property values for each node $i \in V$.
In addition, let variables $z_i$ define participation in a node subset of interest.
To assist us in the expression of statistical metrics, we introduce a new set of continuous variables $p'_i$ to be equal to $p_i z_i + \tilde{p} \left(1 - z_i \right)$; that is, to be equal to the value of property $p_i$, if $z_i = 1$, or equal to some constant value $\tilde{p}$, if $z_i = 0$.
This definition of the $p'$-variables is enforced with the following set of constraints:
\begin{eqnarray}
&\left( p_i^L - \tilde{p} \right) z_i \leq p'_i - \tilde{p} \leq \left( p_i^U - \tilde{p} \right) z_i& \;\;\;\; \forall i \in V \label{pip1}\\
&\left( p_i^L - \tilde{p} \right) \left(1 - z_i\right) \leq p_i - p'_i \leq \left( p_i^U - \tilde{p} \right) \left(1 - z_i\right) & \;\;\;\; \forall i \in V \label{pip2},
\end{eqnarray}
where $p_i^L$ and $p_i^U$ are appropriate lower and upper limits for the value of variable $p_i$. We remark that, as a result of these constraints, $p'_i \in \left[ \min\left\{\tilde{p}, p_i^L\right\}, \max\left\{\tilde{p}, p_i^U\right\}\right]$, for all $i \in V$.
Furthermore, we note that in cases where the sequence refers to nodes belonging to the intersection of multiple subsets, the $p'$-variables need to be appropriately restricted.
This can be achieved by (i) appending separate constraints~(\ref{pip1}) for each applicable subset and (ii) replacing the term $\left(1 - z_i\right)$ in constraints~(\ref{pip2}) with an aggregate term $\sum\limits_e \left(1 - z_i^e\right)$ ($e$ is used here to index over the collection of subsets of interest and $z_i^e$ are the corresponding $z$-variables for subset $e$).
Given the constraint set~(\ref{pip1}) and~(\ref{pip2}), the sum of the property values across the subset of interest can be linearly expressed by setting $\tilde{p} = 0$ and summing $p'$-variables across the constant node set $V$, $\sum\limits_{i \in V} p'_i$.
The mean across the subset of interest can then be expressed as the ratio $\sum\limits_{i \in V} p'_i / \sum\limits_{i \in V} z_i$.
Therefore, the mean constitutes a fractional property and specifying it can be achieved according to the discussion in Section~\ref{specify_fp}.
Higher-order moments could be specified along the same lines, assuming integral powers of the properties ($p_i^2$, $p_i^3$, etc.) can also be expressed in a linear fashion.
We remark that the specification of higher-order moments simplifies if the moments are taken around a constant value, or if the corresponding lower-order moments are also specified to fixed values.
We will conclude this discussion by addressing the case of specifying the
median\footnote{Given an ordered set of $n$ values $\alpha_i$, $i \in \left\{1,2,\cdots,n\right\}$,
we define the median to be the value $\alpha_{\left(n+1\right)/2}$, if $n$ is odd,
or any value in the range $\left[\alpha_{n/2},\alpha_{\left(n+2\right)/2}\right]$, if $n$ is even.}
$\hat{p}$ of a set of property values $p_i$, $i \in V \setminus \left\{i : z_i = 0\right\}$.
We will show that the median can be
linearly expressed in the formulation and, thus, can be directly specified as a linear property.
We first note that the values $\hat{p}^L = \min\limits_{i \in V} \left\{p_i^L\right\}$ and $\hat{p}^U = \max\limits_{i \in V} \left\{p_i^U\right\}$ are appropriate lower and upper limits for the value of the median, irrespectively of the actual participation of nodes in the subset of interest.
We then define $p'$-variables based on $\tilde{p} = \hat{p}^U$ and introduce two new sets of binary variables, $\mathbf{r^{+}}$ and $\mathbf{r^{-}}$, to indicate respectively whether the median is larger or smaller than each of the $p'$-variables.
Similarly to~(\ref{zi}), the following constraints enforce the definitions of the $r$-variables:
\begin{eqnarray}
&\left( p_i^L - \hat{p}^U \right) r_i^{+} \leq p'_i - \hat{p} \leq \left( p_i^U - \hat{p}^L \right) \left(1 - r_i^{+}\right)& \label{rip}\\
&\left( p_i^U - \hat{p}^L \right) r_i^{-} \geq p'_i - \hat{p} \geq \left( p_i^L - \hat{p}^U \right) \left(1 - r_i^{-}\right)& \label{rim}.
\end{eqnarray}
Given this definition, the median $\hat{p}$ will attain the proper value as a result of including in the formulation the inequalities
\begin{eqnarray}
&\sum\limits_{i \in V} z_i \leq 2 \sum\limits_{i \in V} r_i^{+} \leq 1 + \sum\limits_{i \in V} z_i& \label{medianp}\\
&2N - \sum\limits_{i \in V} z_i \leq 2 \sum\limits_{i \in V} r_i^{-} \leq 1 + 2N - \sum\limits_{i \in V} z_i& \label{medianm}\\
&\sum\limits_{i \in V} z_i \geq 1 ,& \label{median1}
\end{eqnarray}
where the last constraint requires explicitly that the subset of interest is not empty. We note that the LP-relaxation can be tightened by adding constraints $r_i^{-}~\geq~1-z_i$ and $r_i^{+}~\leq~z_i$ for all $i \in V$.
\subsubsection{Participation of Nodes in Subsets}\label{subsets}
Every property refers to an appropriately defined subset of the network's nodes.
Most often properties are node-specific; that is, given any network configuration $\mathbf{x}$, each node $i \in V$ attains a separate value for this property (e.g., the properties of degree and clustering coefficient).
In this case, the applicable subset consists of a single node.
Properties may also refer to the network as a whole (e.g., the degree sequence), in which case the subset consists of the complete set $V$.
There are additional cases, however, where a property may refer to some other subset of the nodes (e.g., the case of specifying the maximum shortest-paths among the nodes with a given clustering coefficient).
We note that, in the general case, the participation of nodes in a specified subset of interest is itself variable; that is, it depends on the network configuration $\mathbf{x}$ and it is not a-priori known which nodes belong to the subset.
In such cases, whether a node belongs or not to the subset needs to be appropriately encoded in the formulation.
Given a node-specific property, we define a subset of interest as the collection of those nodes $i \in V$ for which the corresponding values of this property, $p_i$, evaluate above some specified threshold $\pi$.
To indicate whether each node $i \in V$ belongs or not to the subset of interest, we introduce a new set of binary variables
\begin{equation}
z_{i} = \left\{ \begin{array}{r@{\quad:\quad}l}
1 & \text{if node $i$ participates in the subset}\\
0 & \text{otherwise,} \\
\end{array} \right.\label{zidef}
\end{equation}
where $p_i \geq \pi$ is the necessary and sufficient condition for participation.
The definition of the $z$-variables is enforced via the constraints\footnote{We note that constraints~(\ref{zi}) do not completely enforce the definition~(\ref{zidef}), as variable $z_i$ remains unrestricted in the case when variable $p_i$ exactly attains the value $\pi$. However, in cases where $p_i$ attains values from a discrete set, one can always identify some small $\epsilon$ to subtract from $\pi$ in~(\ref{zi}) so that the equality $p_i = \pi$ induces the restriction $z_i = 1$. In cases where $p_i$ attains values from a continuum, there is no practical concern, as there is zero likelihood that $p_i$ attains any specific value exactly.}
\begin{equation}
\left( p_i^L - \pi \right) \left(1 - z_i\right) \leq p_i - \pi \leq \left( p_i^U - \pi \right) z_i \label{zi},
\end{equation}
where $p_i^L$ and $p_i^U$ are appropriate lower and upper limits for the value of variable $p_i$.
It is assumed that the $p$-
variables attain appropriate values as a result of constraints~(\ref{pidef}) having been included in the formulation.
We note that the subset definition presented above can readily admit the case where $\pi$ refers to an upper limit; that is, the case where participation in the subset is realized when $p_i \leq \pi$.
In such a case, one need use $-p_i$ as the property and $-\pi$ as the threshold value, which is equivalent to performing the substitution $z_i \leftarrow \left(1 - z_i \right)$ in constraints~(\ref{zi}).
More complex subsets can then be expressed as intersections of these two simpler cases.
For example, a subset that is defined based on a property evaluating within some specified range, $\pi^L \leq p_i \leq \pi^U$, can be described as the intersection of the two simple subsets $p_i \geq \pi^L$ and $p_i \leq \pi^U$.
\section{Popular Properties}\label{popular}
In this section, we use the modeling elements presented in Section~\ref{modeling} in order to describe a number of popular network properties and associated features, which can then be used in various specifications for the NGP.
Although not an exhaustive list of properties, the compilation of examples illustrates the generality of the proposed MILP-based approach.
We begin by illustrating how a number of common features can be defined.
\begin{example}
{\bf (2-Path).} Let variable $ytp_{ijk}$ indicate whether a 2-path (i.e., open or closed triplet) exists along trajectory $j$--$i$--$k$, where $i \in V$ and $(j,k) \in \left\{\left(V \setminus{\left\{i\right\}}\right)^2 : j < k\right\}$. This variable can be defined with the constraint set
\begin{equation}
\begin{array}{c}
ytp_{ijk} \leq x_{j'j''} \;;\;ytp_{ijk} \leq x_{k'k''} \\
ytp_{ijk} \geq x_{j'j''} + x_{k'k''} - 1\\
ytp_{ijk} \geq 0,
\end{array}\label{ytp}
\end{equation}
where $j' = \min\left\{i,j\right\}$, $j'' = \max\left\{i,j\right\}$, $k' = \min\left\{i,k\right\}$ and $k'' = \max\left\{i,k\right\}$.
Generalizations to longer paths can be readily addressed. \qed
\end{example}
\begin{example}
{\bf (3-Cycle).} Let variable $ytr_{ijk}$ indicate whether a 3-cycle (i.e., triangle) exists between three nodes $(i,j,k) \in T$, where $T = \left\{(i,j,k) \in V^3 : i < j < k\right\}$. This variable can be defined with the constraint set
\begin{equation}
\begin{array}{c}
ytr_{ijk} \leq x_{ij} \;;\;ytr_{ijk} \leq x_{ik} \;;\;ytr_{ijk} \leq x_{jk} \\
ytr_{ijk} \geq x_{ij} + x_{ik} + x_{jk} - 2\\
ytr_{ijk} \geq 0.
\end{array}\label{ytr}
\end{equation}
Generalizations to longer cycles (i.e., squares) can be readily addressed. \qed
\end{example}
\begin{example}
{\bf (4-Clique).} Let variable $yclq_{ijkl}$ indicate whether a clique exists between four nodes $(i,j,k,l) \in \left\{V^4 : i < j < k < l\right\}$. This variable can be defined with the constraint set
\begin{equation}
\begin{array}{c}
yclq_{ijkl} \leq x_{ij} \;;\;yclq_{ijkl} \leq x_{ik} \;;\;yclq_{ijkl} \leq x_{il} \\
yclq_{ijkl} \leq x_{jk} \;;\;yclq_{ijkl} \leq x_{jl} \;;\;yclq_{ijkl} \leq x_{kl} \\
yclq_{ijkl} \geq x_{ij} + x_{ik} + x_{il} + x_{jk} + x_{jl} + x_{kl} - 5\\
yclq_{ijkl} \geq 0,
\end{array}\label{yclq}
\end{equation}
or equivalently, with the constraint set
\begin{equation}
\begin{array}{c}
yclq_{ijkl} \leq \frac{1}{6} \left(x_{ij} + x_{ik} + x_{il} + x_{jk} + x_{jl} + x_{kl}\right)\\
yclq_{ijkl} \geq x_{ij} + x_{ik} + x_{il} + x_{jk} + x_{jl} + x_{kl} - 5\\
yclq_{ijkl} \in \left\{0,1\right\}.
\end{array}\label{yclq'}
\end{equation}
Generalizations to larger cliques can be readily addressed. \qed
\end{example}
\begin{example}
{\bf (4-Star).} Let variable $ystr_{ijkl}$ indicate whether node $i \in V$ corresponds to the center of a star involving nodes $(j,k,l) \in \left\{\left(V \setminus{\left\{i\right\}}\right)^3 : j < k < l\right\}$. This variable can be defined with the constraint set
\begin{equation}
\begin{array}{c}
ystr_{ijkl} \leq x_{j'j''} \;;\;ystr_{ijkl} \leq x_{k'k''} \;;\;ystr_{ijkl} \leq x_{l'l''} \\
ystr_{ijkl} \leq 1 - x_{jk} \;;\;ystr_{ijkl} \leq 1 - x_{jl} \;;\;ystr_{ijkl} \leq 1 - x_{kl} \\
ystr_{ijkl} \geq x_{j'j''} + x_{k'k''} + x_{l'l''} - x_{jk} - x_{jl} - x_{kl} - 2\\
ystr_{ijkl} \geq 0,
\end{array}\label{ystr}
\end{equation}
or equivalently, with the constraint set
\begin{equation}
\begin{array}{c}
ystr_{ijkl} \leq \frac{1}{3} \left(x_{j'j''} + x_{k'k''} + x_{l'l''} \right) \\
ystr_{ijkl} \leq 1 - \frac{1}{3} \left(x_{jk} + x_{jl} + x_{kl} \right) \\
ystr_{ijkl} \geq x_{j'j''} + x_{k'k''} + x_{l'l''} - x_{jk} - x_{jl} - x_{kl} - 2\\
ystr_{ijkl} \in \left\{0,1\right\},
\end{array}\label{ystr'}
\end{equation}
where $j' = \min\left\{i,j\right\}$, $j'' = \max\left\{i,j\right\}$, $k' = \min\left\{i,k\right\}$, $k'' = \max\left\{i,k\right\}$, $l' = \min\left\{i,l\right\}$ and $l'' = \max\left\{i,l\right\}$.
Generalizations involving more nodes can be readily addressed. \qed
\end{example}
Given such definitions of features, a number of properties can be expressed in a linear or fractional fashion. Some examples follow.
\begin{example}
{\bf (Degree).} The degree of each node $i \in V$ can be captured by the variable
\begin{equation}
pd_i = \sum\limits_{(j,k) \in E: \atop \left\{(j = i) \vee (k = i)\right\}} x_{jk}. \label{pd}
\end{equation} \qed
\end{example}
\begin{example}
{\bf (Number of Triplets).} The number of open and closed triplets (i.e., 2-paths) centered at each node $i \in V$ can be captured by the variable
\begin{equation}
pntp_i = \sum\limits_{(j,k) \in E} ytp_{ijk}. \label{pntp} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Number of Triangles).} The number of triangles (i.e., 3-cycles) in which each node $i \in V$ participates can be captured by the variable
\begin{equation}
pntr_i = \sum\limits_{(j,k,l) \in T: \atop \left\{(j = i) \vee (k = i) \vee (l = i)\right\}} ytr_{jkl}. \label{pntr} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Clustering Coefficient).} The (local) clustering coefficient of a node $i \in V$ can be captured by the variable
\begin{equation}
pcc_i = \frac{pntr_i}{pntp_i} \label{pcc}
\end{equation}
and, thus, constitutes a fractional property. However, when node $i$ has a specified (fixed) degree $d_i$, then the local clustering coefficient can also be captured by the linear expression
\begin{equation}
pcc_i = \frac{1}{\left(d_i \atop 2 \right)} pntr_i, \label{pcc_fix}
\end{equation}
where $\left(d_i \atop 2 \right)$ is the constant number of 2-paths centered at node $i$. \qed
\end{example}
\begin{example}
{\bf (Average Clustering Coefficient).} The average clustering coefficient can be captured by the variable
\begin{equation}
pacc = \frac{1}{N} \sum\limits_{i \in V} pcc_i. \label{pacc} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Global Clustering Coefficient).} The global clustering coefficient of a network can be captured by the variable
\begin{equation}
pgcc = \frac{\sum\limits_{i \in V} pntr_i}{\sum\limits_{i \in V} pntp_i} \label{pgcc}
\end{equation}
and, thus, constitutes a fractional property.
However, when the network has a specified (fixed) degree sequence $\mathbf{d} = \left\{d_1, d_2, \cdots, d_N\right\}$, then the global clustering coefficient can also be captured by the linear expression
\begin{equation}
pgcc = \frac{3}{\sum\limits_{i \in V} \left(d_i \atop 2 \right)} \sum\limits_{i \in V} pntr_i, \label{pgcc_fix}
\end{equation}
where $\sum\limits_{i \in V} \left(d_i \atop 2 \right)$ is the constant number of total 2-paths in the network. \qed
\end{example}
\begin{example}
{\bf (Closeness Centrality).} The inverse of the closeness centrality of each node $i \in V$ can be captured by the variable
\begin{equation}
piclc_i = \frac{1}{N-1} \sum\limits_{(j,k) \in E: \atop \left\{(j = i) \vee (k = i)\right\}} w_{jk}. \label{piclc}
\end{equation}
Specifying bounds $\left[\pi^L,\pi^U\right]$ for the closeness centrality of a node can be achieved through specifying bounds $\left[\frac{1}{\pi^U},\frac{1}{\pi^L}\right]$ for its inverse. \qed
\end{example}
\begin{example}
{\bf (Average Path Length).} The average path length of a network can be captured by the variable
\begin{equation}
papl = \frac{1}{|E|} \sum\limits_{(i,j) \in E} w_{ij}, \label{papl}
\end{equation}
where $|E| = N \left(N-1\right)/2$ is the cardinality of the set of edges $E$. \qed
\end{example}
\begin{example}
{\bf (Characteristic Path Length).} The characteristic path length of a network (i.e., defined here as the median of all its shortest paths) can be captured by a variable $pcpl$, which assumes its appropriate value via the constraints\footnote{Constraints~(\ref{pcpl}) are the equivalent of constraints~(\ref{medianp}) and~(\ref{medianm}) for the case of a fixed subset of interest with cardinality $|E|$ (in particular, the set of edges $E$).}
\begin{equation}
\begin{array}{c}
\hspace{-0.23cm} \left.\begin{array}{c}
\begin{array}{rcr}
-U_1 \; rcpl_{ij}^{+} &\leq w_{ij} - pcpl \leq& U_2 \left(1 - rcpl_{ij}^{+}\right)\\
U_2 \; rcpl_{ij}^{-} &\geq w_{ij} - pcpl \geq& -U_1 \left(1 - rcpl_{ij}^{-}\right)
\end{array}\\
rcpl_{ij}^{+},rcpl_{ij}^{-} \in \left\{0,1\right\}
\end{array}\right\} \;\;\; \forall (i,j) \in E \\
2\sum\limits_{(i,j) \in E} rcpl_{ij}^{+} = 2\sum\limits_{(i,j) \in E} rcpl_{ij}^{-} = |E| + \mathbbm{1}_{\left\{|E| \text{ is odd}\right\} },
\end{array} \label{pcpl}
\end{equation}
where $U_1 = \left \lceil{ \frac{N}{2} }\right \rceil - 1$ and $U_2 = N - 2$ are appropriate ``big-M'' coefficients, $|E| = N \left(N-1\right)/2$ is the cardinality of the set of edges $E$, and $\mathbbm{1}_{\left\{\cdot\right\}}$ is the indicator function. \qed
\end{example}
\begin{example}
{\bf (Sum of Degrees of Neighbors).} The sum of degrees of all neighbors of each node $i \in V$ can be captured by variables $psdn_i$, which assume their appropriate values via the constraints
\begin{equation}
\begin{array}{cl}
\left.\begin{array}{rcl}
d^L \; x_{kl} \leq &pd'_{kl} &\leq d^U \; x_{kl}\\
d^L \; \left( 1- x_{kl}\right) \leq &pd_l - pd'_{kl} &\leq d^U \; \left(1 - x_{kl}\right)\\
d^L \; x_{kl} \leq &pd'_{lk} &\leq d^U \; x_{kl}\\
d^L \; \left( 1- x_{kl}\right) \leq &pd_k - pd'_{lk} &\leq d^U \; \left(1 - x_{kl}\right)
\end{array}\right\} & \forall (k,l) \in E \\
\hspace{-0.70cm} psdn_i = \sum\limits_{j \in V} pd'_{ij} & \forall i \in V,
\end{array} \label{psdn}
\end{equation}
where $pd'_{kl}$, $pd'_{lk}$ are $p'$-variables defined for each $(k,l) \in E$ with $\tilde{p}=0$ and with $x_{kl}$ serving as the corresponding $z$-variables, and $d^L$, $d^U$ are applicable lower and upper bounds for the degrees of each node $i \in V$ (such that $0 \leq d^L \leq d^U \leq N-1$).
However, when the network has a specified (fixed) degree sequence $\mathbf{d} = \left\{d_1, d_2, \cdots, d_N\right\}$, one can instead use the simpler expression
\begin{equation}
psdn_i = \sum\limits_{j \in V : \atop \left\{(i,j) \in E\right\}} d_j x_{ij} + \sum\limits_{j \in V : \atop \left\{(j,i) \in E\right\}} d_j x_{ji}.\label{psdn_fix} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Number of Nodes with Given Degree).} Given applicable lower and upper bounds $d^L$ and $d^U$ (such that $0 \leq d^L \leq d^U \leq N-1$) for the degree of each node in the network, the number of nodes that attain each possible degree $q \in Q$, where $Q = \{d^L,d^L+1,\cdots,d^U-1,d^U\}$, can be captured by variables $pnnd_q$.
These variables assume their appropriate values via the constraints
\begin{equation}
\left.\begin{array}{c}
\left.\begin{array}{c}
\left( d^L \; - q + \epsilon \right) \left(1 - z_{qi}\right) \leq pd_i - q + \epsilon \leq \left( d^U - q + \epsilon\right) z_{qi}\\
z_{qi} \in \left\{0,1\right\}
\end{array} \right\} \; \forall i \in V \\
\hspace{-1.75cm} pnnd_q = \sum\limits_{i \in V} z_{qi} - \sum\limits_{i \in V} z_{\left(q+1\right)i}\\
\end{array} \right\} \; \forall q \in Q, \label{pnnd}
\end{equation}
where $z_{qi}$, $q \in Q$, are $z$-variables that are activated when $pd_i \geq q$ and deactivated when $pd_i < q$,
$z_{\left(d^U+1\right)i} = 0$ for all $i \in V$ are parameters defined for notational convenience,
and $\epsilon \in \left(0,1\right)$ is a small parameter to induce activation of variable $z_{qi}$ when $pd_i$ equals $q$ exactly. \qed
\end{example}
\section{Algorithmic Enhancements}\label{enhancements}
As will become evident with the case studies we will present in Section~\ref{studies}, the modeling techniques discussed so far provide considerable amount of flexibility in putting together NGP formulations for a wide variety of property specifications of interest.
However, the NGP is in general a highly-complex combinatorial problem and solving it via standard optimization algorithms may be challenging for all but the smallest instances.
In this section, we discuss a number of algorithmic enhancements that we have found to improve the performance of solution algorithms in the NGP setting.
More specifically, in Section~\ref{slacks} we discuss the utilization of an appropriate objective function that attempts to minimize the cumulative deviation of the solution from the allowable ranges dictated by the specifications.
In Section~\ref{symmbreak}, we discuss the issue of breaking the symmetry embedded inherently in any NGP formulation by appropriately restricting the search space to solutions that are non-isomorphic to each other.
\subsection{Deviation of Properties from Specified Values}\label{slacks}
The discussion in Section~\ref{modeling} addressed efficient ways to express and specify (restrict) properties of interest in an overall MILP formulation.
In all these cases, setting up the modeling of the problem did not require
the formulation's objective function.
This allows us to formulate application-specific objective functions.
For example, out of all those network configurations $\mathbf{x}$ that meet a certain specification, one may seek to identify the network that maximizes or minimizes some particular property, in which case the objective must be the appropriate expression quantifying that property.
Nevertheless, there are many applications for which an explicit metric to assess the fitness of a network configuration does not exist, and the problem at hand is merely a feasibility problem; that is, to identify one or more configurations $\mathbf{x}$ that meet the specification, without consideration to how ``good'' or ``bad'' each of the obtained solutions are.
In such cases, a suitable objective may be devised to enhance the computational performance of the solution process.
In particular, variables can be introduced to encode the deviation of each of the properties from its nearest acceptable (as per the specification) value.
The objective would then require that the sum of all these ``slack'' variables be minimized.
Such an objective quantifies the total deviation from the feasible space of the problem and allows the solver to efficiently guide the search and locate feasible solutions.
As discussed, the property specifications come in the general form $g_m^L \leq g_m(\mathbf{x}) \leq g_m^U$, where $m$ is used here to index over each specified property, $g_m$ is the corresponding functional of the formulation's variables and $g_m^L$, $g_m^U$ are appropriate constants (see, e.g., the constraints~(\ref{lp_bounds}), (\ref{fp_bounds}), (\ref{qim1}), and~(\ref{symbreaklite1})).
Continuous slack variables $\mathbf{s}$ are introduced in the formulation in a way that relaxes each of these specification constraints:
\begin{eqnarray}
&g_m^L \leq g_m(\mathbf{x}) + s_m^{-} - s_m^{+}\leq g_m^U& \label{s1}\\
&s_m^{-}, s_m^{+} \geq 0,& \label{s2}
\end{eqnarray}
where $s_m^{-}$ and $s_m^{+}$ are pairs of variables to encode respectively the negative and positive deviation of quantities $g_m$ from their allowable ranges $\left[g_m^L,g_m^U\right]$.
The sum of all $s$-variables, $\sum\limits_m \left(s_m^{-} + s_m^{+}\right)$, should constitute the objective function to be minimized.
We remark that for a feasible solution all slack variables need to be zero at optimality.\footnote{Note that corresponding variables $s^{-}$ and $s^{+}$ cannot be both non-zero at an optimal solution.}
An optimal solution with a strictly positive objective (as defined above) corresponds to an infeasible NGP, and would signify that the property specification is not attainable.
Examples of specifying popular properties via use of $s$-variables are presented below.
Whenever the corresponding specifications are required, these constraints are to be appended during the modular build-up of a complete formulation.
In Section~\ref{studies}, we present a number of illustrative formulations, and we elaborate further on how these constraints contribute in the context of a complete optimization model.
\begin{example}
{\bf (Specification of Degree).} Let $pd_i$ be the variable capturing the degree of a node $i \in V$ as per constraint~(\ref{pd}).
A lower bound $d^L$ and an upper bound $d^U$ can be specified on the degrees of this node via the constraints
\begin{equation}
d^L \leq pd_i + sd_i^{-} - sd_i^{+} \leq d^U \label{spec-deg},
\end{equation}
where $sd_i^{-}, sd_i^{+} \geq 0$ are a pair of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Degree Sequence).} Let $pd_i$ be the variables capturing the degree of each node $i \in V$ as per constraint~(\ref{pd}).
A degree sequence $\mathbf{d} = \left\{d_1, d_2, \cdots, d_N\right\}$ can be specified via the constraints
\begin{equation}
d_i \leq pd_i + sd_i^{-} - sd_i^{+} \leq d_i \;\; \forall i \in V \label{spec-degseq},
\end{equation}
where $sd_i^{-}, sd_i^{+} \geq 0$ are two sets of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Average Clustering Coefficient).} Let $pacc$ be the variable capturing the average clustering coefficient as per constraint~(\ref{pacc}).
A lower bound $acc^L$ and an upper bound $acc^U$ can be specified on the average clustering coefficient via the constraints
\begin{equation}
acc^L \leq pacc + sacc^{-} - sacc^{+} \leq acc^U \label{spec-acc},
\end{equation}
where $sacc^{-}$, $sacc^{+} \geq 0$ are a pair of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Global Clustering Coefficient).} Let the global clustering coefficient of a network as defined in constraint~(\ref{pgcc}).
Since this quantity constitutes a fractional property, a lower bound $gcc^L$ and an upper bound $gcc^U$ can be specified via the constraints
\begin{equation}
gcc^L \sum\limits_{i \in V} pntp_i \leq \sum\limits_{i \in V} pntr_i + sgcc^{-} - sgcc^{+} \leq gcc^U \sum\limits_{i \in V} pntp_i \label{spec-gcc},
\end{equation}
where $sgcc^{-}$, $sgcc^{+} \geq 0$ are a pair of slack variables to be minimized.
However, when the network has a specified (fixed) degree sequence, then the global clustering coefficient can be specified as a linear property.
Let $pgcc$ be the variable capturing the global clustering coefficient as per constraint~(\ref{pgcc_fix}). We can directly utilize this variable in the constraints
\begin{equation}
gcc^L \leq pgcc + sgcc^{-} - sgcc^{+} \leq gcc^U \label{spec-gcc_fix},
\end{equation}
where $sgcc^{-}$, $sgcc^{+} \geq 0$ are a pair of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Average Path Length).} Let $papl$ be the variable capturing the characteristic path length as per constraint~(\ref{papl}).
A lower bound $apl^L$ and an upper bound $apl^U$ can be specified on the characteristic path length via the constraints
\begin{equation}
apl^L \leq papl + sapl^{-} - sapl^{+} \leq apl^U \label{spec-apl},
\end{equation}
where $sapl^{-}$, $sapl^{+} \geq 0$ are a pair of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Characteristic Path Length).} Let $pcpl$ be the variable capturing the characteristic path length as per constraint~(\ref{pcpl}).
A lower bound $cpl^L$ and an upper bound $cpl^U$ can be specified on the characteristic path length via the constraints
\begin{equation}
cpl^L \leq pcpl + scpl^{-} - scpl^{+} \leq cpl^U \label{spec-cpl},
\end{equation}
where $scpl^{-}$, $scpl^{+} \geq 0$ are a pair of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Diameter).} A lower bound $D^L$ and an upper bound $D^U$ can be specified on the network's diameter via the constraints
\begin{eqnarray}
&1+\left(D^L-1\right)\psi_{ij} \leq w_{ij} + sD^{-} - sD^{+} \leq D^U \;\;\; \forall (i,j) \in E& \label{spec-diam1} \\
&\sum\limits_{(i,j) \in E} \psi_{ij} \geq 1& \label{spec-diam2}
\end{eqnarray}
where $\psi_{ij}$ are auxiliary binary variables defined for each $(i,j) \in E$, and $sD^{-}$, $sD^{+} \geq 0$ are a pair of slack variables to be minimized.
Specifying the diameter to a given value $D$ can be achieved by setting $D^L = D^U = D$.\qed
\end{example}
In the above example, constraints~(\ref{spec-diam1}) impose the strict upper bound of $D^U$ for all shortest paths $w_{ij}$ as well as a lower bound of either $D^L$ or $1$ (a trivial value), depending on whether variable $\psi_{ij}$ assumes the value of $1$ or $0$, respectively. At the same time, constraint~(\ref{spec-diam2}) ensures that ``at least one'' of these lower bounds will be $D^L$.
\begin{example}
{\bf (Specification of Closeness Centrality Sequence).} Let $piclc_i$ be the variables capturing the inverse of the closeness centrality of each node $i \in V$ as per constraint~(\ref{piclc}).
A sequence $\mathbf{clc} = \left\{\left[clc_1^L,clc_1^U\right], \left[clc_2^L,clc_2^U\right], \cdots, \left[clc_N^L,clc_N^U\right] \right\}$ for the closeness centralities of the nodes in $V$ can be specified via the constraint set\footnote{Constraints~(\ref{spec-clc}) correspond to constraints~(\ref{qim1})--(\ref{qim4}) for the case of the applicable subset being the complete node set $V$.}
\begin{equation}
\hspace{-0.23cm}\left.\begin{array}{c}
\hspace{-0.23cm}\left.\begin{array}{l}
piclc_i + siclc_i^{-} - siclc_i^{+} \geq piclc^L - \left(piclc^L - \frac{1}{clc_m^U} \right) q_{im}\\
piclc_i + siclc_i^{-} - siclc_i^{+} \leq piclc^U - \left(piclc^U - \frac{1}{clc_m^L} \right) q_{im}
\end{array}\right\} \forall (i,m) \in V \times V \\
\hspace{-0.23cm}\left.\begin{array}{l}
\;\sum\limits_{i \in V} q_{im} = 1 \;\;\; \forall m \in V \\
\sum\limits_{m \in V} q_{im} = 1 \;\;\; \forall i \in V,
\end{array}\right.
\end{array}\right.\label{spec-clc}
\end{equation}
where $piclc^L = \min\limits_{j\in V} \left\{\frac{1}{clc_j^U}\right\}$ and $piclc^U = \max\limits_{j\in V} \left\{\frac{1}{clc_j^L}\right\}$ are respectively lower and upper limits for variables $piclc_i$,
$q_{im}$ are $q$-variables to indicate the assignment of the $m^\text{th}$ interval from the sequence $\mathbf{clc}$ to the $i^\text{th}$ node, indicating that the closeness centrality of node $i$, $1/piclc_i$, is to attain a value within the range $\left[clc_m^L,clc_m^U\right]$,
and $siclc_i^{-}, siclc_i^{+} \geq 0$ are two sets of slack variables to be minimized.\qed
\end{example}
\begin{example}
{\bf (Specification of Average Degree of Neighbors).} Let $psdn_i$ be the variables capturing the sum of degrees of the neighbors of each node $i \in V$ as per constraints~(\ref{psdn}), $pnnd_q$ be the variable capturing the number of nodes that attain degree $q$ as per constraints~(\ref{pnnd}), and
$z_{qi}$ be the $z$-variables activated if $pd_i \geq q$ as per constraints~(\ref{pnnd}).
A lower bound $adn_q^L$ and an upper bound $adn_q^U$ can be specified on the average degree of the neighbors of all nodes that attain degree $q$ via the constraints
\begin{eqnarray}
\left.\begin{array}{rcl}
0 \leq &psdn'_{qi}& \leq \left(d^U \right)^2 \left( z_{iq} - z_{i\left(q+1\right)}\right) \\
0 \leq &psdn_i - psdn'_{qi}& \leq \left(d^U \right)^2 \left( 1 - z_{iq} + z_{i\left(q+1\right)}\right)
\end{array}\right\} \; \forall i \in V \label{spec-adn1} \\
adn_q^L \; q \; pnnd_q \leq \sum\limits_{i \in V} psdn'_{qi} + sadn_q^{-} - sadn_q^{+} \leq adn_q^U \; q \; pnnd_q \label{spec-adn2},
\end{eqnarray}
where $psdn'_{qi}$ are sets of $p'$-variables corresponding to variables $psdn_i$ (defined for $\tilde{p} = 0$ and activated if node $i$ attains degree $q$),
$d^U$ is the maximum possible degree attainable by any node,
and $sadn_q^{-}, sadn_q^{+} \geq 0$ are two sets of slack variables to be minimized.\qed
\end{example}
In the above example, the variables $psdn'_i$ assume the value $psdn'_i = psdn_i$, if $pd_i = q$ (i.e., if both $z_{iq} = 1$ and $z_{i\left(q+1\right)} = 0$), while otherwise they assume the value $psdn'_i = 0$.
This facilitates the convenient expression of the sum of degrees of all neighbors of nodes that attain a given degree $q$ as the sum of variables $psdn'_i$ over all nodes $i \in V$, irrespective of whether a node $i$ attains degree $q$ or not.
The sum is utilized in constraint~(\ref{spec-adn2}) that imposes the required specification of average degree of the neighbors as a fractional property involving the ratio of this sum over the total number of neighbors of the nodes that attain degree $q$ (which is equal to the product of $q$ times $pnnd_q$).
\subsection{Removing Isomorphism}\label{symmbreak}
A prominent characteristic of the NGP is the potential isomorphism of its solutions; that is, the admittance of multiple solutions that correspond to the same configuration up to permutation (relabeling) of the node set $V$ and are, thus, practically indistinguishable to each other.
For example, consider a network with $N = 3$ nodes and two configurations $\mathbf{x} = \left(x_{12},x_{13},x_{23}\right)$ such that $\mathbf{x}_A = \left(1,1,0\right)$ and $\mathbf{x}_B = \left(1,0,1\right)$.
It is easy to see that both configurations correspond to the same network, a simple open triplet, the difference being the (arbitrary) label of its central node (node $1$ in $\mathbf{x}_A$ and node $2$ in $\mathbf{x}_B$).
It is not uncommon for networks with as few as ten nodes to have hundreds of isomorphic counterparts.
Isomorphism constitutes a major challenge in addressing the problem via an MILP approach, as the associated branch-and-bound-based search process would be subject to significant burden in exploring unnecessarily enlarged feasible spaces.
Furthermore, the attempt to identify and record multiple distinct solutions will be challenged by the continued recurrence of solutions that are isomorphic to the ones already encountered.
It is therefore desirable to remove all or part of the isomorphism by introducing constraints that will restrict the search to only those solutions that are ``nominal,'' according to some appropriate criterion, while of course ensuring that the full space of distinct, non-isomorphic configurations remains attainable (feasible).
As the goal is essentially to break the symmetry among the nodes in set $V$, the criterion can be based on a node-specific property.
We can then consider that a solution is nominal only if the nodes $i \in V$ attain values for this property, $p_i$, in a sorted (w.l.o.g., non-increasing) fashion:
\begin{equation}
p_i - p_{i+1} \geq 0 \;\;\;\; \forall i \in V \setminus \left\{N\right\}. \label{symbreak1}
\end{equation}
The degree property is most often the obvious choice of property upon which to reduce isomorphism, since it is easily expressed as a linear sum of the $x$-variables (the formulation's basic variables).
However, constraints~(\ref{symbreak1}) may not suffice to fully eliminate isomorphism due to the lack of a specification that is adequately restrictive (e.g., the specification of a degree sequence with two or more of its elements being equal).
To further break ties among nodes that attain equal values of the ``primary'' property, $p_i$, some ``secondary'' property, $pp_i$, must be considered in tandem.
More specifically, in addition to~(\ref{symbreak1}), we utilize the constraints
\begin{equation}
U \left( p_i - p_{i+1} \right) + pp_i - pp_{i+1} \geq 0 \;\;\;\; \forall i \in V \setminus \left\{N\right\}, \label{symbreak2}
\end{equation}
where $U$ is an appropriate ``big-M'' coefficient that will allow the variables $pp_i$ and $pp_{i+1}$ to vary freely for even a small, yet ``sufficiently'' strictly positive, difference $p_i - p_{i+1}$.
For the case of the the primary symmetry-breaking property $p$ being the degree $p_d$, which is an integral property, the value $U = pp_{i+1}^U - pp_i^L$ would suffice to achieve the desired behavior.
Tertiary and additional, lower-level symmetry-breaking criteria could be imposed in a similar fashion so as to reinforce the attempt to alleviate isomorphism.
However, in most applications there would be little or no residual symmetry to justify the associated increase in model size or the potential scaling challenges due to the introduction of increasingly large coefficients $U^2$, $U^3$, etc.
We remark that breaking symmetry is greatly simplified in the special case where a complete sequence is specified for the primary property (e.g., a full degree sequence). Let $\mathbf{\Pi} = \left\{\pi_1,\pi_2,\cdots,\pi_N\right\}$ be the desired sequence, which w.l.o.g. we consider to be sorted in non-increasing order.
Primary symmetry-breaking can be achieved by directly specifying that $\pi_i$ be assigned to node $i$, for all nodes $i \in V$.
The secondary symmetry-breaking criterion could then be considered only for those successive node pairs $i$ and $i+1$ that we know a-priori would attain the same degree.
By doing so, not only can symmetry-breaking be imposed without the need for ``big-M'' coefficients, but also the sequence is explicitly assigned without the need for $q$-variables.
Constraints~(\ref{symbreaklite1}) and~(\ref{symbreaklite2}) apply.
The first constraint is a specification of a linear property to an exact value (see constraint~(\ref{lp_bounds}) with $\pi^L = \pi^U \rightarrow \pi_i$) and the second constraint selectively applies the second criterion wherever necessary.
\begin{eqnarray}
&\pi_i \leq p_i \leq \pi_i& \;\;\;\;\; \forall i \in V \label{symbreaklite1} \\
&pp_i \geq pp_{i+1}& \;\;\;\;\; \forall i \in V \setminus \left\{N\right\} : \left\{\pi_i = \pi_{i+1}\right\} \label{symbreaklite2}
\end{eqnarray}
\begin{example}
{\bf (Symmetry-breaking based on Degrees and Clustering Coefficients).} Let $pd_i$ be the variables capturing the degree of each node $i \in V$ as per constraint~(\ref{pd}), and let $pcc_i$ be the variables capturing the (local) clustering coefficient of each node $i \in V$ as per constraints~(\ref{pcc}) or~(\ref{pcc_fix}).
Breaking symmetry among nodes based primarily on their degrees and secondarily on their (local) clustering coefficients can be achieved via the constraint set
\begin{equation}
\left.\begin{array}{r}
pd_i - pd_{i+1} \geq 0 \\
pd_i - pd_{i+1} + pcc_i - pcc_{i+1} \geq 0
\end{array} \right\} \forall i \in V \setminus \left\{N\right\},\label{symbreak_deg_cc}
\end{equation}
where the implied coefficient $U=1$ reflects the fact that two clustering coefficients can at most differ by one.
However, when a (non-increasing) degree sequence has been specified as via constraints~(\ref{spec-degseq}), then it suffices to use the simpler constraint set
\begin{equation}
pcc_i \geq pcc_{i+1} \;\;\;\;\; \forall i \in V \setminus \left\{N\right\} : \left\{d_i = d_{i+1}\right\}. \label{symbreak_deg_cc_fix} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Symmetry-breaking based on Degrees and Shortest-Path Distances).} Let $pd_i$ be the variables capturing the degree of each node $i \in V$ as per constraint~(\ref{pd}), and let $w_{ij}$ be the variables capturing the shortest-path distance between two nodes $(i,j) \in E$ as per the discussion of Section~\ref{shortestpaths}.
Breaking symmetry among nodes based primarily on their degrees and secondarily on their shortest-path distances from the last node, $i = N$, can be achieved via the constraint set
\begin{equation}
\begin{array}{rl}
pd_i - pd_{i+1} \geq 0 & \;\; \forall i \in V \setminus \left\{N\right\} \\
U \left( pd_i - pd_{i+1} \right) + w_{iN} - w_{\left(i+1\right)N} \geq 0 & \;\; \forall i \in V \setminus \left\{N-1,N\right\},
\end{array} \label{symbreak_deg_wN}
\end{equation}
where $U=N-2$ is the appropriate ``big-M'' coefficient representing the maximum possible difference between two shortest paths in a graph of $N$ nodes.
However, when a (non-increasing) degree sequence has been specified as via constraints~(\ref{spec-degseq}), then it suffices to use the simpler constraint set
\begin{equation}
w_{iN} \geq w_{\left(i+1\right)N} \;\;\;\;\;\; \forall i \in V \setminus \left\{N-1,N\right\} : \left\{d_i = d_{i+1}\right\}. \label{symbreak_deg_wN_fix} \qed
\end{equation}
\end{example}
We remark that, in the above example, any node could have been selected as the reference node for such symmetry breaking purposes. Yet, in all cases, it would not have been possible to differentiate among the neighbors of the reference node via this set of constraints, since all neighbors feature the same shortest distance of $1$ from the reference node.
Our choice of the last node, $i = N$, as the reference node is thus justified by the fact that it connects to the smallest amount of neighbors (given that the primary breaking of symmetry is done via degrees), which in turn translates to a formulation with as minimal residual isomorphism as possible.
\begin{example}
{\bf (Symmetry-breaking based on Degrees and Sums of Degrees of Neighbors).} Let $pd_i$ be the variables capturing the degree of each node $i \in V$ as per constraint~(\ref{pd}), and let $psdn_i$ be the variables capturing the sum of degrees of the neighbors of each node $i \in V$ as per constraints~(\ref{psdn}).
Breaking symmetry among nodes based primarily on their degrees and secondarily on the sums of degrees of their neighbors can be achieved via the constraint set
\begin{equation}
\left.\begin{array}{r}
pd_i - pd_{i+1} \geq 0 \\
U \left( pd_i - pd_{i+1} \right) + psdn_i - psdn_{i+1} \geq 0
\end{array} \right\} \forall i \in V \setminus \left\{N\right\}, \label{symbreak_deg_sdn}\\
\end{equation}
where $U=\left(d^U \right)^2$ is the appropriate ``big-M'' coefficient representing the maximum possible difference between the sums of degrees of neighbors of two given nodes.\footnote{Consider the first node being of maximum degree and connecting to $d^U$ nodes that all attain maximum degree themselves (e.g., as part of a clique), and consider the second node being an isolated node (sum of degrees of neighbors being $0$ in this case).}
However, when a (non-increasing) degree sequence has been specified as via constraints~(\ref{spec-degseq}), then it suffices to use the simpler constraint set
\begin{equation}
psdn_i \geq psdn_{i+1} \;\;\;\;\; \forall i \in V \setminus \left\{N\right\} : \left\{d_i = d_{i+1}\right\}. \label{symbreak_deg_sdn_fix} \qed
\end{equation}
\end{example}
\begin{example}
{\bf (Symmetry-breaking based on Degrees and Closeness Centralities).} Let $pd_i$ be the variables capturing the degree of each node $i \in V$ as per constraint~(\ref{pd}), and let $piclc_i$ be the variables capturing the (inverse of) closeness centrality of each node $i \in V$ as per constraints~(\ref{piclc}).
Breaking symmetry among nodes based primarily on their degrees and secondarily on their closeness centralities can be achieved via the constraint set
\begin{equation}
\left.\begin{array}{r}
pd_i - pd_{i+1} \geq 0 \\
U \left( pd_i - pd_{i+1} \right) + piclc_{i+1} - piclc_{i} \geq 0
\end{array} \right\} \forall i \in V \setminus \left\{N\right\}, \label{symbreak_deg_clc}\\
\end{equation}
where $U = \frac{N}{2}-1$ is the appropriate ``big-M'' coefficient representing the maximum possible difference between the inverse closeness centralities of two nodes in any (connected) network,\footnote{The value $\frac{N}{2}-1$ corresponds to the difference between the inverse closeness centrality of a terminal node in a path graph ($cl.c. = \frac{2}{N}$) and that of a node in a clique ($cl.c. = 1$).} and $d^L$, $d^U$ are respectively the minimum and maximum possible degrees attainable by any node.
However, when a (non-increasing) degree sequence has been specified as via constraints~(\ref{spec-degseq}), then it suffices to use the simpler constraint set
\begin{equation}
piclc_i \leq piclc_{i+1} \;\;\;\;\; \forall i \in V \setminus \left\{N\right\} : \left\{d_i = d_{i+1}\right\}. \label{symbreak_deg_clc_fix} \qed
\end{equation}
\end{example}
\section{Illustrative Applications}\label{studies}
In this section we discuss the development of complete formulations for a number of informative applications.
We demonstrate the modularity of the approach by utilizing constraint sets introduced previously in Sections~\ref{popular} and~\ref{enhancements}.
All case studies were solved on a Win7 machine with an Intel i5-3230M (2.60GHz) processor and 3GB RAM. Gurobi~5.6~\cite{GUROBI} was used as the MILP optimization solver, and all runs were explicitly limited to utilize a single CPU core.
Aside from customized branching prioritization for binary variables according to the earlier discussion, default solver options were used throughout this study.
The node labels in all figures represent the degrees of the depicted nodes.
\subsection{Tuning Network Connectivity}
Two primary characteristics of a network in terms of the connectivity among its nodes are the degree distribution and the level of clustering.
As elaborated in the introduction, there exist many well-known algorithms for constructing networks with a specified degree sequence.
However, none of these algorithms can guarantee that the constructed networks adhere also to pre-specified clustering statistics, which can in general vary significantly among networks of the given degree sequence.
In fact, networks generated by approaches that randomly connect vertices tend to have a low number of triangles and, thus, low clustering coefficients.
In contrast, many networks that develop in the real world feature larger clustering coefficients.
Take social networks, for instance, where two randomly chosen friends (neighbors) of a person (node) are also more likely to know each other than two randomly chosen people in the network.
Furthermore, networks whose nodes have an underlying spatial meaning (e.g., power-grid networks) tend to develop more edges between nodes that are closer to each other, leading to the presence of more triangles than what would be expected by mere chance.
To that end, it is sometimes necessary to be able to control the statistics of both degrees and clustering of the nodes in the constructed network.
In this application, we consider specifying the degree sequence in conjunction with the network's average and global clustering coefficients.
Let $\mathbf{d} = \left\{d_1, d_2, \cdots, d_N\right\}$ be a specification for the degree sequence\footnote{For the clustering coefficients to be well-defined, it should hold that $d_N \geq 2$.} (w.l.o.g., we assume $d_i \geq d_{i+1} \; \forall i \in V \setminus \left\{N\right\}$), and let $\left[acc^L,acc^U\right]$ and $\left[gcc^L,gcc^U\right]$ be respectively the acceptable bounds for the average and global clustering coefficient.
After picking the appropriate sets of constraints from examples presented previously in the paper, we formulate the following optimization model:
\begin{eqnarray}
&\min\limits_{\mathbf{x}} & \sum\limits_{i \in V} \left(sd_i^{-} + sd_i^{+}\right) + sacc^{-} + sacc^{+} + sgcc^{-} + sgcc^{+} \label{formclu-obj} \\
& \text{s.t.} & \text{Constraints~(\ref{ytr})} \;\;\;\;\;\;\; \forall (i,j,k) \in T \nonumber \\
& & \text{Constraints~(\ref{pd})} \;\;\;\;\;\;\; \forall i \in V \nonumber \\
& & \text{Constraints~(\ref{pntr})} \;\;\;\;\;\;\; \forall i \in V \nonumber \\
& & \text{Constraints~(\ref{pcc_fix})} \;\;\;\;\;\;\; \forall i \in V \nonumber \\
& & \text{Constraint~\,\,(\ref{pacc})} \nonumber \\
& & \text{Constraint~\,\,(\ref{pgcc_fix})} \nonumber \\
& & \text{Constraint~\,\,(\ref{spec-degseq})} \nonumber \\
& & \text{Constraint~\,\,(\ref{spec-acc})} \nonumber \\
& & \text{Constraint~\,\,(\ref{spec-gcc_fix})} \nonumber \\
& & \text{Constraints~(\ref{symbreak_deg_cc_fix})} \nonumber \\
&& x_{ij} \in \left\{0,1\right\} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (i,j) \in E \label{formclu-xbinary}\\
&& sd_i^{-},sd_i^{+} \geq 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall i \in V \label{formclu-slackpos1}\\
&& sacc^{-},sacc^{+},sgcc^{-},sgcc^{+} \geq 0 \label{formclu-slackpos2}.
\end{eqnarray}
Constraints~(\ref{ytr}) through~(\ref{pgcc_fix}) restrict all applicable $p$-variables (variables $pd_i$, $pacc$ and $pgcc$) to their appropriate values (the degree of each node and the two clustering coefficients, respectively), while constraints~(\ref{spec-degseq}), (\ref{spec-acc}) and~(\ref{spec-gcc_fix}) impose the specifications.
Obviously, if no specification is desired for either of the two clustering coefficients, the corresponding constraint~(\ref{spec-acc}) or~(\ref{spec-gcc_fix}) should be omitted (or simply ``disabled'' by setting $\left[0,1\right]$ as the allowable bounds).
Constraints~(\ref{spec-degseq}) are also used in conjunction with the ordered degree sequence specification to reduce isomorphism.
Symmetry is primarily broken based on node's degrees, while residual symmetry is broken based on local clustering coefficients.
Finally, constraints~(\ref{formclu-xbinary}) declare the binarity of the $x$-variables, while constraints~(\ref{formclu-slackpos1}) and~(\ref{formclu-slackpos2}) declare the non-negativity of the $s$-variables.
The objective~(\ref{formclu-obj}) minimizes the sum of every slack variable associated with a specification.
We remark that although a number of $y$-, $p$- and $s$-variables participate in the formulation, we only list $\mathbf{x}$ under the $\min$ operator to highlight the fact that the $x$-variables suffice to fully define the solution (i.e., the values of all other variables can be uniquely defined once values for $\mathbf{x}$ are determined).
Case Study~\ref{case1} demonstrates the use of this formulation.
We provide a specific (randomly chosen) degree sequence, and we desire to construct networks with various values for the average and the global clustering coefficient.
To that end, we provide three ranges (``low,'' ``medium'' and ``high'') from which we would require both coefficients to attain values.
\begin{casestudy}\label{case1}
Let $N=10$ nodes, and let $\mathbf{d} =\left\{5, 4, 4, 3, 3, 3, 2, 2, 2, 2\right\}$ be a specification for the degree sequence.
Let also three separate specifications with $\left[0.000,0.250\right]$, $\left[0.250,0.500\right]$
and $\left[0.500,0.750\right]$ as the allowable bounds for the average and global
clustering coefficients (bounds common to both metrics). %
Figure~\ref{fig:case1} depicts representative solutions.
Each of the three networks was constructed in less than 0.1 $sec$ CPU. \qed
\end{casestudy}
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 2.0cm 0 2.5cm,clip=true]{./1A.pdf}
}
\subfloat[]{
\includegraphics[totalheight=0.30\textheight,trim=0 5.0cm 0 3.5cm,clip=true]{./1B.pdf}
}
\subfloat[]{
\includegraphics[totalheight=0.26\textheight,trim=0 1.5cm 0 1.5cm,clip=true]{./1C.pdf}
}
\caption{Representative optimal solutions to the three specifications from Case Study~\ref{case1}. From left to right, the networks possess the same degree sequence but are respectively tuned to low ($acc = 0.077, gcc = 0.086$), medium ($acc = 0.353, gcc = 0.257$) and high ($acc = 0.713, gcc = 0.514$) level of clustering.}
\label{fig:case1}
\end{figure}
Instead of trying to locate network realizations that adhere to a particular specification, a problem cast as a slack-minimization problem, Case Study~\ref{case2} discusses the situation where an application-specific objective function is available.
In particular, we seek to identify the network with maximum clustering coefficient (either average or global) that adheres to the specified degree sequence.
The applicable formulations can be derived from the previous one per the following steps: (i) remove all $s$-variables (i.e., set them to a fixed value of zero), (ii) disable the specifications for the two coefficients (i.e., set $acc^L=gcc^L=0$ and $acc^U=gcc^U=1$ in constraints~(\ref{spec-acc}) and~(\ref{spec-gcc_fix}), or completely remove these constraints from the formulation), and (iii) replace the objective with either $\left\{\max\limits_{\mathbf{x}} pacc\right\}$ or $\left\{\max\limits_{\mathbf{x}} pgcc\right\}$, depending on which of the two coefficients we seek to maximize.
\begin{casestudy}\label{case2}
Let $N=10$ nodes, and let $\mathbf{d} =\left\{5, 4, 4, 3, 3, 3, 2, 2, 2, 2\right\}$ be a specification for the degree sequence.
We seek the networks that maximize the average clustering coefficient as well as the networks that maximize the global clustering coefficient.
Figure~\ref{fig:case2} depicts a common solution that happens to maximize both metrics at the same time.
The network was obtained in about 1 $sec$ CPU. \qed
\end{casestudy}
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 4.5cm 0 4.5cm,clip=true]{./2A.pdf}
}
\caption{A common optimal solution to the problems of maximizing the average and global clustering coefficients given a specified degree sequence. The network possesses maximum values for the average and global clustering coefficients, $acc = 0.833$ and $gcc = 0.686$, respectively.}
\label{fig:case2}
\end{figure}
We observe that Case Study~\ref{case2} resulted in a network that features two disconnected components. The optimizer was able to discern that, in this way, ``more clustered'' communities can be formed while at the same time adhering to the overall specified degree sequence.
\subsection{Tuning Network Spread}
Beyond local connectivity in the network, which can be controlled through properties such as the degrees and clustering coefficients, another informative network quality to consider during
construction is the network's {\em long-range spread}.
The spread is especially important in networks where the edges represent conduits of information flow, such as the case of communication networks.
It relates to the relative size of shortest paths across the network and can thus be characterized by a number of relevant properties, most notably the characteristic path length (cpl) and the network diameter (i.e., the median and maximum shortest-paths, respectively).
It is important to note that the network spread may sometimes correlate with node connectivity, but tuning connectivity does not suffice for also tuning the spread.
Randomly constructed networks typically have a small number of triangles (low clustering) and low spread.
On the other hand, small-world networks also exhibit low spread, but are known to feature a large number of triangles.
In general, networks with different possible combinations of local connectivity and spread exist, and it is important to be able to independently control both qualities during network construction.
To that end, in this application we study the construction of networks that adhere to a specific degree sequence and at the same time possess pre-specified values for the two network spread metrics, cpl and diameter.
Let $\mathbf{d} = \left\{d_1, d_2, \cdots, d_N\right\}$ be a specification for the degree sequence (w.l.o.g., we assume $d_i \geq d_{i+1} \; \forall i \in V \setminus \left\{N\right\}$), and let $\left[cpl^L,cpl^U\right]$ and $\left[D^L,D^U\right]$ be respectively the acceptable bounds for the network's cpl and diameter.
The following formulation applies:
\begin{eqnarray}
&\min\limits_{\mathbf{x}} & \sum\limits_{i \in V} \left(sd_i^{-} + sd_i^{+}\right) + scpl^{-} + scpl^{+} + sD^{-} + sD^{+} \label{formspread-obj} \\
& \text{s.t.} & \text{Constraints~(\ref{sp-p1})--(\ref{sp-p4}),~(\ref{sp-d1})--(\ref{sp-d4}),~(\ref{sp-o1}),~(\ref{sp-o2})} \;\;\; \forall (i,j) \in E \nonumber \\
& & \text{Constraints~(\ref{pd})} \;\;\;\;\;\;\;\;\;\;\;\;\;\, \forall i \in V \nonumber \\
& & \text{Constraints~(\ref{pcpl})} \nonumber \\
& & \text{Constraints~(\ref{spec-degseq})} \nonumber \\
& & \text{Constraints~(\ref{spec-cpl})} \nonumber \\
& & \text{Constraints~(\ref{spec-diam1}),~(\ref{spec-diam2})} \nonumber \\
& & \text{Constraints~(\ref{symbreak_deg_wN_fix})} \nonumber \\
&& x_{ij}, \psi_{ij} \in \left\{0,1\right\} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (i,j) \in E \label{formspread-xbinary} \\
&& sd_i^{-},sd_i^{+} \geq 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall i \in V \label{formspread-slackpos1} \\
&& scpl^{-},scpl^{+},sD^{-},sD^{+} \geq 0 \label{formspread-slackpos2}.
\end{eqnarray}
Constraints~(\ref{sp-p1}) through~(\ref{sp-o2}) correspond to the encoding in $w$-variables of shortest path lengths via the primal/dual formulation analysis presented in Section~\ref{shortestpaths}.
Constraints~(\ref{pd}) and~(\ref{pcpl}) define the applicable $p$-variables for the node degrees and the cpl, while constraints~(\ref{spec-degseq}) and~(\ref{spec-cpl}) impose the corresponding specifications. The network diameter is specified via constraints~(\ref{spec-diam1}) and~(\ref{spec-diam2}).
Obviously, if no specification is desired for either the cpl or the diameter, the corresponding constraints should be omitted (or simply ``disabled'' by setting sufficiently ``loose'' bounds).
Isomorphism is primarily eliminated based on node degrees, while residual symmetry in the problem is broken via constraints~(\ref{symbreak_deg_wN_fix}), which order nodes of the same degree according to their shortest path to the last node (node $i = N$).
Finally, constraints~(\ref{formspread-xbinary}) declare the binarity of the $x$- and $\psi$-variables, while constraints~(\ref{formspread-slackpos1}) and~(\ref{formspread-slackpos2}) declare the non-negativity of the $s$-variables.
The objective~(\ref{formspread-obj}) minimizes the sum of every slack variable associated with a specification.
Case Study~\ref{case3} demonstrates the use of this formulation. We provide a degree sequence as in the previous case studies, and we desire to construct networks that attain various combinations of values for the cpl and the diameter.
\begin{casestudy}\label{case3}
Let $N=10$ nodes, and let $\mathbf{d} =\left\{5, 5, 4, 4, 3, 3, 2, 2, 1, 1\right\}$ be a specification for the degree sequence.
In addition, let a specification of $cpl^L = cpl^U = 2$ for the characteristic path length, and let three separate specifications for the diameter requiring that it attains the values of $D=3$, $D=4$ and $D=5$.
Figure~\ref{fig:case3} depicts representative solutions for the three cases, which were respectively obtained in approximately 20, 23 and 198 $sec$ CPU. \qed
\end{casestudy}
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 4.5cm 0 4.5cm,clip=true]{./3A.pdf}
}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 0.0cm 0 0.0cm,clip=true]{./3B.pdf}
}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 3.5cm 0 3.5cm,clip=true]{./3C.pdf}
}
\caption{Representative optimal solutions to the three specifications from Case Study~\ref{case3}. From left to right, the networks possess diameter values of 3, 4 and 5. All three networks possess a characteristic path length value of 2.}
\label{fig:case3}
\end{figure}
\subsection{Tuning Network Assortativity}
In the previous applications we specified node degrees, clustering information and statistics on the lengths of shortest paths.
In this section we will consider mixing patterns between nodes that form constituent links; that is, correlations between node properties of neighbors.
We specifically focus on capturing the correlations between degrees of neighbors in a network.
A network is said to have assortative mixing if high (low) degree nodes preferentially form links with high (low) degree nodes.
In contrast, disassortative networks are those in which high degree nodes preferentially link with low degree ones, and vice versa.
Empirical studies of real world networks have revealed that social networks tend to be assortative, while technological and biological networks are usually disassortative.
In either case, having the ability to accurately capture degree-degree correlations of real world networks is crucial to understanding the dynamic processes that govern the formation of these networks.
Degree-degree correlations can be quantified in different ways.
We pose the problem such that we specify bounds for the average degree among all neighbors of nodes of a given degree.
For convenience, we further specify an overall minimum and maximum degree for the whole network (which may just be the trivial values of 0 and $N-1$, respectively).
Note that we do not specify the exact degree sequence as in the previous applications.
More formally, let $Q = \{d^L,d^L+1,\cdots,d^U-1,d^U\}$ be the set of possible degrees, where $0 \leq d^L \leq d^U \leq N-1$, and let also $Q^+ = Q \setminus \left\{0\right\}$ defined for notational convenience.
We focus on each possible degree $q \in Q^+$ and require that the average degree of the
neighbors of all nodes that attain degree $q$, denoted as $adn_q$, be
restricted in the range $\left[adn_q^L,adn_q^U\right]$.
Note that we do not provide a specification for $q=0$, as isolated nodes do not
possess any neighbors.
We formulate the following optimization model:
\begin{eqnarray}
&\min\limits_{\mathbf{x}} & \sum\limits_{i \in V} \left(sd_i^{-} + sd_i^{+}\right) + \sum\limits_{q \in Q^+} \left(sadn_q^{-} + sadn_q^{+}\right) \label{formassort-obj} \\
& & \text{Constraints~(\ref{pd})} \;\;\;\;\;\;\;\;\;\;\;\;\; \forall i \in V \nonumber \\
& & \text{Constraints~(\ref{psdn})} \nonumber \\
& & \text{Constraints~(\ref{pnnd})} \nonumber \\
& & \text{Constraints~(\ref{spec-deg})} \;\;\;\;\;\;\;\;\;\;\;\;\; \forall i \in V \nonumber \\
& & \text{Constraints~(\ref{spec-adn1}),~(\ref{spec-adn2})} \;\;\;\; \forall q \in Q^+ \nonumber \\
& & \text{Constraints~(\ref{symbreak_deg_sdn})} \nonumber \\
&& pd_1 \geq 1 \label{formassort-nonull}\\
&& x_{ij} \in \left\{0,1\right\} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (i,j) \in E \label{formassort-xbinary} \\
&& sd_i^{-},sd_i^{+} \geq 0 \;\;\;\;\;\;\;\;\;\;\;\;\;\;\, \forall i \in V \label{formassort-slackpos1} \\
&& sadn_q^{-},sadn_q^{+} \geq 0 \;\;\;\;\;\;\,\forall q \in Q^+ \label{formassort-slackpos2}.
\end{eqnarray}
Constraints~(\ref{pd}),~(\ref{psdn}) and~(\ref{pnnd}) define variables $pd_i$, $psdn_i$ and $pnnd_q$ for the degrees, the sums of neighbors' degrees and the number of nodes that attain a given degree, respectively.
Constraints~(\ref{spec-deg}) impose the degree specification, as it applies to each node independently, and they are only relevant when bounds other than $d^L = 0$ and $d^U = N-1$ are chosen.
Constraints~(\ref{spec-adn1}) and~(\ref{spec-adn2}) impose the specification for the average degree of neighbors of the nodes of each degree $q \in Q^+$.
Constraints~(\ref{symbreak_deg_sdn}) break symmetry, primarily based on degrees, with residual symmetry among nodes that attain the same degree being broken based on the sum of degrees of their neighbors.
Constraint~(\ref{formassort-nonull}) is added to avoid the trivial solution of a null graph, which may otherwise be feasible under certain specifications.
Finally, constraints~(\ref{formassort-xbinary}) through~(\ref{formassort-slackpos2}) declare binarity and positivity of variables.
As before, the objective~(\ref{formassort-obj}) minimizes the sum of the slack variables used in the formulation.
For the offset parameter in constraints~(\ref{pnnd}) we chose the value $\epsilon = 0.01$, which is small enough to not adversely affect the tightness of the LP-relaxations but at the same time large-enough to sufficiently exceed the optimization solver's (default) feasibility tolerance setting of $10^{-6}$.
In Case Study~\ref{case4}, we design an assortative network by specifying bounds that are increasing as a function of $q$.
Conversely, in Case Study~\ref{case5}, we construct a disassortative network by specifying bounds that decrease with $q$.
\begin{casestudy}\label{case4}
Let $N=10$ nodes and let $d^L = 0$ and $d^U = N-1=9$ (i.e., no restriction on node degrees).
The set of possible degrees is, thus, $Q = \left\{0,1,\cdots,9\right\}$.
Let also $adn^L_q = 1/2 + \left(2/3\right)q$ and $adn^U_q = 1 + \left(2/3\right)q$, for all $q \in Q^+$, be lower and upper allowable values for the average degree among the neighbors of all nodes that attain degree $q$.
Figure~\ref{fig:case4and5} (top-left) depicts a representative solution to this specification.
We observe that only eight nodes connect in the network, leaving two nodes isolated.
Let us now choose $d^L = 1$, enforcing that no isolated nodes exist.
A representative solution is also shown in Figure~\ref{fig:case4and5} (top-right).
It took less than 0.3 $sec$ CPU to obtain either of these two solutions. \qed
\end{casestudy}
We observe that the solution we obtained for the case of no isolated nodes is in fact a network that is {\em regular of degree 3}; that is, a network where all nodes attain the same degree 3. This is indeed a valid solution since $\left\{3\right\} \in \left[adn^L_3, adn^U_3\right]=\left[2.5,3.5\right]$.
The specification for the remaining degree classes, $q \in Q^+ \setminus \left\{3\right\}$, becomes irrelevant in this solution, as these classes were not sampled by any node. We remark that if more degree diversity is desired, regular graphs can be trivially excluded from consideration by appending the constraint $pd_{1}-pd_{N} \geq n$, where $n \geq 1$ is the minimum span of degrees acceptable for a solution.
\begin{casestudy}\label{case5}
Let $N=10$ nodes and let $d^L = 0$ and $d^U = N-1=9$ (i.e., no restriction on node degrees).
The set of possible degrees is, thus, $Q = \left\{0,1,\cdots,9\right\}$.
Let also $adn^L_q = 4 - \left(2/3\right)q$ and $adn^U_q = 5 - \left(2/3\right)q$, for all $q \in Q^+$, be lower and upper allowable values for the average degree among the neighbors of all nodes that attain degree $q$.
Figure~\ref{fig:case4and5} (bottom-left) depicts a representative solution to this specification. We observe that only nine nodes connect in the network, leaving one node isolated.
Let us now choose $d^L = 1$, enforcing that no isolated nodes exist.
A representative solution is also shown in Figure~\ref{fig:case4and5} (bottom-right).
It took less than 0.5 $sec$ CPU to obtain either of these two solutions. \qed
\end{casestudy}
We observe that the solutions of Case Study~\ref{case5} do not feature any nodes with degree $7$ or above. In fact, this was enforced implicitly by the specification, despite the looser setting $d^U = 9$.
Since only connected nodes (i.e., nodes of degree at least $1$) can serve as neighbors of other nodes, the setting $adn^U_{q} < 1$ for all $q \geq 7$ implied that nodes of degree $7$ or above may not have any neighbors, i.e., that no such nodes may exist in the solution. Additionally, no nodes of degree $6$ exist in these solutions, as the setting $adn^U_{6} = 1$ implies that such nodes may only connect to nodes of degree $1$, forming $6$-star components, which apparently is very restrictive given the rest of the specification (in particular, it is impossible to achieve $adn_{1} \leq adn^U_{1}=4.33$ in a $10$-node network that features a $6$-star component).
We finally remark that, similarly to Case Study~\ref{case2}, in Case Studies~\ref{case4} and~\ref{case5} we were able to obtain solutions that featured disconnected components. This was possible due to the fact that no explicit restriction was imposed on network properties pertaining to connectivity and/or length of paths between nodes as, for example, was the setting in Case Study~\ref{case3}.
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 5.0cm 0 4.0cm,clip=true]{./4A.pdf}
}
\hspace{1cm}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 5.0cm 0 5.0cm,clip=true]{./4B.pdf}
}\hfil
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 5.0cm 0 4.0cm,clip=true]{./5A.pdf}
}
\hspace{1cm}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 5.0cm 0 5.0cm,clip=true]{./5B.pdf}
}
\caption{Representative optimal solutions to Case Studies~\ref{case4} (top) and~\ref{case5} (bottom). The networks at the top are assortative, while the ones at the bottom are disassortative. The networks to the left feature solutions with isolated nodes.}
\label{fig:case4and5}
\end{figure}
\subsection{Tuning Network Robustness}
Robustness (a.k.a., resilience) of complex networks to attacks or failures is an important subject of study in many fields involving technological networks, most notably in the context of defense applications.
Robustness is usually studied by measuring the effect of removing one or more nodes on the network's functional properties.
In this context, the importance of the individual nodes to the integrity of the network is generally quantified by
using some measure of {\em node centrality}.
The node degrees capture one aspect of node centrality.
However, degree centrality only addresses the importance of the node in its immediate neighborhood.
The {\em closeness centrality} of a node (i.e., the inverse of the average shortest-path length from this node to every other node in the network) can be used as a better metric to measure the node's importance to the network as a whole.
A value of closeness centrality close to 1 indicates that the node is centrally-located, and thus integral to the structural robustness of the network, while a value close to 0 indicates that the node is peripheral to the network, and thus less important from a robustness point of view.
Therefore, by specifying different distributions of closeness centrality for the nodes of a network, we can control the overall network robustness.
In this application, we construct networks with a specified sequence of closeness centralities.
More formally, let $d^L$ and $d^U$, where $0 \leq d^L \leq d^U \leq N-1$, be lower and upper bounds on the node degrees, and let $\mathbf{clc} = \left\{\left[clc_1^L,clc_1^U\right], \left[clc_2^L,clc_2^U\right], \cdots, \left[clc_N^L,clc_N^U\right] \right\}$ be a sequence of ranges for the values of closeness centrality that are to be assumed by the network's nodes.
In other words, we require that there is a one-to-one correspondence between the set of closeness centralities of each node with the $N$ ranges in the sequence $clc$.
The following formulation applies:
\begin{eqnarray}
&\min\limits_{\mathbf{x}} & \sum\limits_{i \in V} \left(sd_i^{-} + sd_i^{+} + siclc_i^{-} + siclc_i^{+}\right) \label{formrob-obj} \\
& \text{s.t.} & \text{Constraints~(\ref{sp-p1})--(\ref{sp-p4}),~(\ref{sp-d1})--(\ref{sp-d4}),~(\ref{sp-o1}),~(\ref{sp-o2})} \;\;\; \forall (i,j) \in E \nonumber \\
&&\hspace{-0.23cm}\left.\begin{array}{l}
\text{Constraints~(\ref{pd})} \\
\text{Constraints~(\ref{piclc})}\\
\text{Constraints~(\ref{spec-deg})}
\end{array} \;\; \right\} \;\;\;\;\;\;\;\;\;\;\, \forall i \in V \nonumber \\
&& \text{Constraints~(\ref{spec-clc})} \nonumber\\
&& \text{Constraints~(\ref{symbreak_deg_clc})} \nonumber \\
&& x_{ij} \in \left\{0,1\right\} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; \forall (i,j) \in E \label{formrob-xbinary}\\
&& q_{im} \in \left\{0,1\right\} \;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\, \forall (i,m) \in V \times V \label{formrob-qbinary}\\
&& sd_i^{-},sd_i^{+},siclc_i^{-},siclc_i^{+} \geq 0 \;\;\;\; \forall i \in V \label{formrob-slackpos1}.
\end{eqnarray}
Constraints~(\ref{sp-p1}) through~(\ref{sp-o2}) correspond to the encoding in $w$-variables of shortest-path lengths via the primal/dual formulation analysis presented in Section~\ref{shortestpaths}.
Constraints~(\ref{pd}) and~(\ref{piclc}) define the applicable $p$-variables for the node degrees and the (inverse of) closeness centrality, respectively.
Constraints~(\ref{spec-deg}) impose the specification of overall degree bounds, as they apply to each node independently, while constraints~(\ref{spec-clc}) impose the specification for the closeness centrality sequence.
Constraints~(\ref{symbreak_deg_clc}) eliminate isomorphism. This is primarily based on node degrees, with nodes of the same degree being arranged in order of decreasing closeness centrality (equivalently, in order of increasing inverse of closeness centrality).
Finally, constraints~(\ref{formrob-xbinary}) and (\ref{formrob-qbinary}) declare the binarity of the $x$- and $q$-variables, while constraints~(\ref{formrob-slackpos1}) declare the non-negativity of the $s$-variables.
The objective~(\ref{formrob-obj}) minimizes the sum of every sum of all slack variables participating in the formulation.
Case Studies~\ref{case6} and~\ref{case7} demonstrate the use of this formulation.
We aim to design networks that attain two separate sequences for closeness centrality.
The first case study creates a network where the nodes are progressively important for robustness.
The second case study results in a network that possesses a vulnerable link.
\begin{casestudy}\label{case6}
Let $N=10$ nodes and let $d^L = 0$ and $d^U = N-1=9$ (i.e., no restriction on node degrees).
In addition, let the following specification for the closeness centrality sequence, which requires the network nodes to be of increasing importance for robustness:
\begin{equation*}
\left[clc^L_m,clc^U_m \right] = \left[0.35 + m/30, 0.42 + m/30 \right], \;\; \forall m \in V.
\end{equation*}
Figure~\ref{fig:case6and7a} (left) depicts a representative solution for this specification, which was obtained in about 50 $min$ CPU.\footnote{Enabling multi-threaded optimization and allowing the MILP optimizer to utilize all 4 cores of our machine, this number was reduced to less than 4 $min$ CPU.}\qed
\end{casestudy}
\begin{casestudy}\label{case7}
Let $N=10$ nodes and let $d^L = 0$ and $d^U = N-1=9$ (i.e., no restriction on node degrees).
In addition, let the following specification for the closeness centrality sequence, which requires two nodes to be distinctively more important for robustness than the rest of the network:
\begin{equation*}
\left[clc^L_m,clc^U_m \right] = \left\{
\begin{array}{l}
\left[0.10, 0.50 \right], \;\; \forall m \in \left\{1,2,\cdots,8\right\} \\
\left[0.60, 1.00 \right], \;\; \forall m \in \left\{9,10\right\}.
\end{array}\right.
\end{equation*}
Figure~\ref{fig:case6and7a} (right) depicts a representative solution for this specification, which was obtained in 28 $sec$ CPU.\qed
\end{casestudy}
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.32\textheight,trim=0 7.0cm 0 7.0cm,clip=true]{./6A.pdf}
}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 0.0cm 0 0.0cm,clip=true]{./7A.pdf}
}
\caption{Representative optimal solutions to Case Studies~\ref{case6} (left) and~\ref{case7} (right). {\em Left:} The node of degree 6 at the top is the most important for robustness, exhibiting a closeness centrality value of $clc=0.74$, while the node of degree 1 at the right is the least important, with a value of $clc=0.39$. {\em Right:} The two nodes of degree 5 attain a closeness centrality value of $clc = 0.69$ and are distinctively more critical for robustness than the other nodes, all of which attain a much smaller value of $clc = 0.43$.}
\label{fig:case6and7a}
\end{figure}
In the above case studies, the degree sequence was not specified. We will conclude this application by addressing the case where a specific degree sequence is desired in conjunction with the specification of a sequence for the closeness centralities.
Solutions to this problem can be obtained by simply replacing $d^L$ and $d^U$ in constraints~(\ref{spec-deg}) with the desired degree values $d_i$, for all $i \in V$ (equivalently, by replacing constraints~(\ref{spec-deg}) with constraints~(\ref{spec-degseq})). We further replace the symmetry-breaking constraints~(\ref{symbreak_deg_clc}) with their simpler version~(\ref{symbreak_deg_clc_fix}).
Figure~\ref{fig:case6and7b} depicts two networks that are also solutions to the specifications of Case Studies~\ref{case6} and~\ref{case7}, but at the same possess the degree sequence specified in Case Study~\ref{case3} (i.e., $\mathbf{d} =\left\{5, 5, 4, 4, 3, 3, 2, 2, 1, 1\right\}$).
These solutions were respectively obtained in about 122 $sec$ and 16 $sec$ CPU. We note that the former is significantly faster than the corresponding run of Case Study~\ref{case6}, which did not entail a specification for the degree sequence.
Specifying the degrees--and utilizing this fact explicitly in order to break some of the formulation's symmetry--appears in this case to have considerably reduced the feasible search space and to have made the task of identifying a solution easier.
\begin{figure}[!ht]
\captionsetup[subfigure]{labelformat=empty}
\centering
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 4.0cm 0 4.0cm,clip=true]{./6B.pdf}
}
\hspace{2.5cm}
\subfloat[]{
\includegraphics[totalheight=0.27\textheight,trim=0 4.0cm 0 4.0cm,clip=true]{./7B.pdf}
}
\caption{Optimal solutions to Case Studies~\ref{case6} (left) and~\ref{case7} (right), which additionally possess the degree sequence specified in Case Study~\ref{case3}.}
\label{fig:case6and7b}
\end{figure}
\section{Conclusions}\label{conclusions}
This paper demonstrated the potential and versatility of using Mixed-Integer Linear Optimization to solve the Network Generation Problem, which is the problem of the ``on-demand'' construction of networks that possess specified collective properties.
It was shown how to mathematically express a number of network properties in a format consistent with a MILP optimization formulation,
and how to enforce their specification by restricting them into appropriate values in the context of the optimization formulation.
The particular network properties covered in this paper were chosen due to their popularity among network researchers and were meant to serve as a paradigm.
The modeling techniques that were presented can be utilized in extending the framework to handle a broad class of additional properties that may be relevant to some particular application of interest.
We further compiled complete optimization formulations to specify various levels of connectivity, spread, assortativity and robustness on the networks to be generated, and we solved a number of relevant case studies involving small (10-node) networks to illustrate the overall framework.
Solutions to these case studies were obtained within a few seconds or minutes --depending on the instance-- through the use of uncustomized, ``off-the-shelf,'' optimization software.
There are multiple factors that affect the computational performance of an MILP solution process when dealing with instances of the Network Generation Problem.
These include the size of the network to be constructed, the number and type of variables required to encode the problem (which depends on the application), as well as the restrictiveness of the desired specification.
Future research will be directed towards establishing the limitations of available ``off-the-shelf'' optimization software and, more importantly, towards developing customized solution methodologies in order to expand the current performance envelope.
\section*{Acknowledgements}\label{funding}
This work was supported by the United States Department of Energy (DE-SC0002097), the Defense Threat Reduction Agency (HDTRA1-07-1-0005) and the National Science Foundation (CBET-0827907). CEG further acknowledges support from the Berkman Faculty Development Fund.
\linespread{1}
\small
\bibliographystyle{plain}
|
\section{Introduction}
Uniform random intersection graphs have received much attention and been used in various applications \cite{r1,Perfectmatchings,NikoletseasHM,r4,ryb3,zz,2013arXiv1301.0466R,yagan,ZhaoYaganGligor,ZhaoCDC,ANALCO}. A uniform random intersection graph with $n$ nodes is defined as follows: each node picks a set of $K_n$ different items \emph{uniformly at random} from the same pool of $P_n$ distinct items, and an \emph{undirected} edge is put between
any two nodes which share at least one item. We will denote a uniform random intersection graph by $G(n, K_n, P_n)$.
Uniform random intersection graphs belong to a wider class of graphs called random intersection graphs in which each node selects some items in a random manner and any two nodes have an undirected edge upon sharing a certain number of items \cite{zz,2013arXiv1301.0466R,yagan,ZhaoYaganGligor,ZhaoCDC,ANALCO}.
Uniform random intersection graphs are also referred to as random key graphs due to their applications to the Eschenauer--Gligor key predistribution scheme \cite{virgil}, which is a recognized approach to ensure secure
communications in wireless sensor networks. In the Eschenauer--Gligor scheme for a wireless sensor network with $n$ sensors,
before deployment,
each sensor is assigned a set of $K_n$ distinct cryptographic keys selected uniformly at random from the same key pool containing $P_n$ different keys. After deployment, two sensors establish secure communication if
and only if they have at least one common key. Clearly the induced topology is a uniform random intersection graph. In addition to secure sensor networks, uniform random intersection graphs have been used for
recommender systems \cite{r4}, social networks \cite{ZhaoYaganGligor}, and circuit design \cite{2013arXiv1301.0466R}.
In this paper, we study monotone increasing properties and their phase transitions in uniform random intersection graphs. The studied properties include $k$-connectivity, perfect matching containment, and Hamilton cycle containment. First, $k$-connectivity means that each pair of nodes has at least $k$
internally node-disjoint path(s) between them \cite{ZhaoYaganGligor}. Second, a perfect matching in a graph with an even number of nodes means a
matching covering all nodes, where a matching in a graph is a set of edges without common nodes \cite{Perfectmatchings}.
We use the generalized notion of perfect matching: for a graph with an odd number of nodes, a perfect matching
means a
matching covering all nodes except one \cite{Perfectmatchings}. Finally, a Hamiltonian cycle in a graph is a closed loop that visits each node once \cite{EfthymiouaHM}.
The above studied properties of uniform random intersection graphs have various applications. In the use of uniform random intersection graphs for secure wireless sensor networks \cite{virgil}, $k$-connectivity enables multi-path routing and load balancing,
and is useful for consensus \cite{ZhaoYaganGligor};
perfect matchings have been used for the analysis of wireless information flow \cite{Amaudruz2009} and network coverage \cite{Wang20091427},
and the optimal allocation of rate and power \cite{4407772}; and Hamilton cycles have been used for cyclic routing which along with distributed optimization achieves efficient in-network data processing \cite{1413472}.
We make the following contributions in this paper: i) we provide an exact analysis on the probabilities of
a uniform random intersection graph $G(n, K_n, P_n)$ having a perfect matching and having a Hamilton cycle respectively, and ii) we compute the phase transition widths of $G(n, K_n, P_n)$ for $k$-connectivity, perfect matching containment, and Hamilton cycle containment, respectively.
Note that
when we say a graph has a perfect matching (respectively, Hamilton cycle), we mean the graph has \emph{at least one} perfect matching (respectively, Hamilton cycle).
The rest of the paper is organized as follows. Section \ref{sec:main:res} presents the main results as
theorems. Then, we introduce several auxiliary lemmas in
Section \ref{sec:factlem}, before establishing the theorems in
Section \ref{sec:thmprf:kcon}. Section
\ref{related} reviews related work, and Section \ref{sec:Conclusion}
concludes the paper. The Appendix
details the proofs of the lemmas.
\section{The Main Results} \label{sec:main:res}
We present the main results in Theorems \ref{thm:PM}--\ref{thm:PHW}. We use the standard asymptotic notation $\Omega(\cdot), \omega(\cdot), O(\cdot),
o(\cdot),\Theta(\cdot)$. All asymptotics and limits are taken with $n \to \infty$. Also, $\mathbb{P}[\cdot]$
denotes an event \vspace{-1pt} probability. An event happens \emph{with high probability} if its probability converges to $1$ as $n\to\infty$.
\subsection{An exact analysis on perfect matching containment and Hamilton cycle containment}
\begin{thm} \label{thm:PM}
For a uniform random intersection graph
$G(n,K_n,P_n)$, if there is a sequence $\alpha_n$ with $\lim_{n \to \infty}{\alpha_n} \in [-\infty, \infty]$
such that
\begin{align}
\frac{{K_n}^2}{P_n} & = \frac{\ln n +
{\alpha_n}}{n}, \label{thm:PM:pe}
\end{align}
then under $P_n = \omega\big(n (\ln n)^5\big)$, it holds that
\begin{align}
& \lim_{n \to \infty} \mathbb{P}[\hspace{2pt}G(n, K_n, P_n)\text{ has a perfect matching.}\hspace{2pt}] \nonumber \\ & \hspace{-1pt}=\hspace{-1.5pt} e^{- e^{-\lim\limits_{n \to \infty}{\alpha_n}}} \hspace{-1.5pt}=\hspace{-1.5pt} \begin{cases} 0,&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\alpha_n} \hspace{-1.5pt}=\hspace{-1.5pt}- \infty, \\
1,&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\alpha_n} \hspace{-1.5pt}=\hspace{-1.5pt} \infty, \\
e^{-e^{- \alpha^{*}}},&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\alpha_n} \hspace{-1.5pt}=\hspace{-1.5pt} \alpha^{*}\hspace{-1.5pt}\in\hspace{-1.5pt} (-\infty, \hspace{-1.5pt}\infty). \end{cases}\nonumber
\end{align}
\end{thm}
\begin{thm} \label{thm:HC}
For a uniform random intersection graph
$G(n,K_n,P_n)$, if there is a sequence $\beta_n$ with $\lim_{n \to \infty}{\beta_n} \in [-\infty, \infty]$
such that
\begin{align}
\frac{{K_n}^2}{P_n} & = \frac{\ln n + \ln \ln n +
{\beta_n}}{n}, \label{thm:HC:pe}
\end{align}
then under $P_n = \omega\big(n (\ln n)^5\big)$, it holds that
\begin{align}
& \lim_{n \to \infty} \mathbb{P}[\hspace{2pt}G(n, K_n, P_n)\text{ has a Hamilton cycle.}\hspace{2pt}] \nonumber \\ & \hspace{-1pt}=\hspace{-1.5pt} e^{- e^{-\lim\limits_{n \to \infty}{\beta_n}}} \hspace{-1.5pt}=\hspace{-1.5pt} \begin{cases} 0,&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\beta_n} \hspace{-1.5pt}=\hspace{-1.5pt}- \infty, \\
1,&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\beta_n} \hspace{-1.5pt}=\hspace{-1.5pt} \infty, \\
e^{-e^{- \beta^{*}}},&\text{\hspace{-4pt}if }\lim_{n \to \infty}{\beta_n} \hspace{-1.5pt}=\hspace{-1.5pt} \beta^{*}\hspace{-1.5pt}\in\hspace{-1.5pt} (-\infty, \hspace{-1.5pt}\infty). \end{cases}\nonumber
\end{align}
\end{thm}
Theorems \ref{thm:PM} and \ref{thm:HC} show that uniform random intersection graphs exhibit phase transitions for perfect matching containment and Hamilton cycle containment.
By \cite[Lemma 8]{ZhaoYaganGligor}, the term $ \frac{{K_n}^2}{P_n}$ in (\ref{thm:PM:pe}) and (\ref{thm:HC:pe})
is an asymptotic value of the edge probability $1-\binom{P_n-K_n}{K_n}\big/\binom{P_n}{K_n}$ (i.e., the probability for the existence of an edge between two nodes). If we \vspace{1.5pt} replace $ \frac{{K_n}^2}{P_n}$ in (\ref{thm:PM:pe}) and (\ref{thm:HC:pe}) with the edge probability, Theorems \ref{thm:PM} and \ref{thm:HC} still follow; see the explanations given in the proofs of Theorems \ref{thm:PM} and \ref{thm:HC}.
Note that a difference between a uniform random intersection graph and an Erd\H{o}s--R\'enyi graph \cite{erdosPF} which is constructed by assigning an edge between each pair of nodes independently with the same probability, is that not all edges in the former graph are independent, while in the latter graph all edges are mutually independent. Despite of this difference, the results above along with Lemmas \ref{lem:ER:PM} and \ref{lem:ER:HC} in Section \ref{sec:factlem} show that in both graphs,
common critical thresholds of the edge probability are $\frac{\ln n }{n}$ for perfect matching containment, and $\frac{\ln n + \ln \ln n }{n}$ for Hamilton cycle containment.
\subsection{Phase transition widths for $k$-connectivity, perfect matching containment and Hamilton cycle containment}
For each $n$, given $P_n$, the probability that $G(n, K_n, P_n)$ has a monotone increasing graph property $\mathcal{I}$ increases as $K_n$ increases \cite{zz}, so we define for a positive constant $\epsilon < \frac{1}{2}$ that\vspace{-2pt}
\begin{align}
K_n^{-}(\mathcal{I}, \epsilon) :=\vspace{-2pt} \min \bigg\{K_n \bigg| \begin{array}{l}G(n, K_n, P_n)\text{ has property }\mathcal{I}\\ \text{ with probability at least }\epsilon .\end{array}\bigg\} \label{KnI-minus}
\end{align}
and\vspace{-2pt}
\begin{align}
K_n^{+}(\mathcal{I}, \epsilon) :=\vspace{-2pt} \min \bigg\{K_n \bigg| \begin{array}{l}G(n, K_n, P_n)\text{ has property }\mathcal{I}\\ \text{ with probability at least }1-\epsilon .\end{array}\bigg\}.\label{KnI-plus}
\end{align}
The phase transition width $d_n(\mathcal{I}, \epsilon)$ of graph $G(n,K_n,P_n)$ for a monotone increasing graph property $\mathcal{I}$ and a positive constant $\epsilon < \frac{1}{2}$ is defined by\vspace{-2pt}
\begin{align}
d_n(\mathcal{I}, \epsilon)\vspace{-2pt} = K_n^{+}(\mathcal{I}, \epsilon) - K_n^{-}(\mathcal{I}, \epsilon). \label{phase-transition-width}
\end{align}
\begin{thm} \label{thm:PHW}
For a uniform random intersection graph
$G(n,K_n,P_n)$,
the following results hold for any positive constant $\epsilon < \frac{1}{2}$ and any positive constant integer $k$:\\
(i) If $P_n = \Omega(n)$ and $P_n = o(n\ln n)$, $d_n(\text{$k$-connectivity}, \epsilon)$ equals either $0$ or $1$ for each $n$ sufficiently large.\\
(ii) If $P_n = \Theta(n\ln n)$, then $d_n(\text{$k$-connectivity}, \epsilon)=\Theta(1)$.\\
(iii) If $P_n = \omega(n\ln n)$, then $d_n(\text{$k$-connectivity}, \epsilon)=\omega(1)$.\\
(iv) If $P_n \hspace{-.5pt}=\hspace{-.5pt} \omega\big(n (\ln n)^5\big)$, $d_n(\text{perfect matching containment}, \epsilon)$ and $d_n(\text{Hamilton cycle containment}, \epsilon)$ can both be written as $\omega(1)$. \vspace{-2pt}
\end{thm}
The result (i) in Theorem \ref{thm:PHW} above shows that under $P_n = \Omega(n)$ and $P_n = o(n\ln n)$, a uniform random intersection graph exhibits a sharp phase transition for $k$-connectivity for each $n$ sufficiently large: if $G(n,K_n,P_n)$ has a probability of at least $\epsilon$ to be $k$-connected, then $G(n,K_n+1,P_n)$ has a probability of at least $1-\epsilon$ to be $k$-connected, since $d_n(\text{$k$-connectivity}, \epsilon)$ is at most $1$. Note that $d_n(\text{$k$-connectivity}, \epsilon)$ could be $0$; e.g., for some $K_n^*$, if the probabilities of $G(n,K_n^*,P_n)$ and $G(n,K_n^*+1,P_n)$ being $k$-connected are less than $\epsilon$, and at least $1-\epsilon$, respectively, then $d_n(\text{$k$-connectivity}, \epsilon) = 0$ follows since $K_n^{-}(\text{$k$-connectivity}, \epsilon)$ and $K_n^{+}(\text{$k$-connectivity}, \epsilon)$ given by (\ref{KnI-minus}) and (\ref{KnI-plus}) both equal $K_n^*+1$ in this case.
For a uniform random intersection graph
$G(n,K_n,P_n)$, the results on $k$-connectivity in Lemma \ref{thm:ZhaoCDCk} in the next section use $P_n = \Omega(n)$, while the results on perfect matching containment and Hamilton cycle containment in Theorems \ref{thm:PM} and \ref{thm:HC} rely on $P_n = \omega\big(n (\ln n)^5\big)$. Since Theorem \ref{thm:PHW} above is established from Lemma \ref{thm:ZhaoCDCk}, and Theorems \ref{thm:PM} and \ref{thm:HC}, then Theorem \ref{thm:PHW} has results for $k$-connectivity under $P_n = \Omega(n)$, and for perfect matching containment and Hamilton cycle containment under a narrower range of $P_n = \omega\big(n (\ln n)^5\big)$; i.e., for $P_n = \Omega(n)$ and $P_n = O\big(n (\ln n)^5\big)$, we have phase transition results
for $k$-connectivity, but not for perfect matching containment and Hamilton cycle containment. A future direction is to extend our results to a wider range of $P_n$.
\section{Auxiliary Lemmas} \label{sec:factlem}
We provide a few lemmas used to establish the
theorems. The proofs of Lemmas
\ref{thm:ZhaoCDCk} and \ref{cp_rig_er} are given in the Appendix, while Lemmas \ref{lem-only-prove-lnlnn}, \ref{lem:ER:PM}, \ref{lem:ER:HC}, and \ref{lem:subseq} are results directly from prior work, and the proof of Lemma \ref{lem:subsequenc} is omitted since it is straightforward.
Lemma
\ref{thm:ZhaoCDCk} presents an exact analysis of $k$-connectivity in a uniform random intersection graph
$G(n,K_n,P_n)$. Since this lemma easily follows from our work \cite[Lemma 1]{ANALCO} and \cite[Theorem 1]{ZhaoCDC}, it is not emphasized in our contributions, but we still explain its proof in the Appendix for clarity.
%
\iffalse
\begin{thm} \label{thm:Rybarczykk1}
\cite{ryb3}
For a uniform random intersection graph
$G(n,K_n,P_n)$, with a sequence $\alpha_n$ for all $n $ defined through
\begin{align}
\frac{{K_n}^2}{P_n} & = \frac{\ln n +
{\alpha_n}}{n}, \label{thm:pe:k1}
\end{align}
then if $K_n \geq 2$ for all $n$ sufficiently large, it holds for a positive constant integer $k$ that
\begin{align}
\lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G(n,K_n,P_n)\textrm{
is connected}.\hspace{2pt}\right] & = e^{- e^{-\lim_{n \to \infty}{\alpha_n}}},
\end{align}
\end{thm}
\fi
\iffalse
\begin{align}
\hspace{-2pt} \lim_{n \to \infty}\hspace{-2pt}\mathbb{P} \hspace{-2.5pt}\left[\hspace{-4pt}\begin{array}{l}G(n, K_n, P_n)\\ \text{is connected}.\end{array}\hspace{-4.5pt}\right] &\hspace{-2pt} = \hspace{-2pt}e^{- \frac{e^{-\alpha ^*}}{(k-1)!}},
\text{ if $\lim_{n \to \infty}\hspace{-2pt}{\alpha_n}
\hspace{-2pt}=\hspace{-2pt}\alpha ^* \hspace{-2pt}\in\hspace{-2pt} [-\infty, \hspace{-1.5pt} \infty]$}. \nonumber
\end{align}
\begin{align}
\lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G(n,K_n,P_n)\textrm{
is $k$-connected}.\hspace{2pt}\right] & = e^{- \frac{e^{-\lim_{n \to \infty}{\alpha_n}}}{(k-1)!}},
\end{align}
\begin{align}
\lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G(n,K_n,P_n)\textrm{
is $k$-connected}.\hspace{2pt}\right] & = e^{- \frac{e^{-\lim_{n \to \infty}{\alpha_n}}}{(k-1)!}},
\end{align}
\fi
\begin{lem} \label{thm:ZhaoCDCk}
For a uniform random intersection graph
$G(n,K_n,P_n)$, if there is a sequence $\gamma_n$ with $\lim_{n \to \infty}{\gamma_n} \in [-\infty, \infty]$
such that
\begin{align}
\frac{{K_n}^2}{P_n} & = \frac{\ln n + {(k-1)} \ln \ln n +
{\gamma_n}}{n}, \label{thm:pe:kgencdc}
\end{align}
then under $P_n = \Omega(n)$, it holds for a positive constant integer $k$ that\begin{align}
& \lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G(n,K_n,P_n)\textrm{
is $k$-connected}.\hspace{2pt}\right] \nonumber \\
& = \lim_{n \to \infty}\mathbb{P} \left[\hspace{2pt}G(n,K_n,P_n)\textrm{
has a minimum degree at least $k$}.\hspace{2pt}\right] \nonumber \\ & = e^{- \frac{e^{-\lim_{n \to \infty}{\gamma_n}}}{(k-1)!}}. \label{thm:pe:kgencdcres}
\end{align}
\end{lem}
\begin{lem}[\hspace{0pt}{Our work \cite[Lemma 1]{ANALCO}}\hspace{0pt}] \label{lem-only-prove-lnlnn}
For a uniform random intersection graph
$G(n,K_n,P_n)$ under $P_n = \Omega(n)$ and (\ref{thm:pe:kgencdc}), the following results hold:
(i) If $\lim_{n \to \infty}\gamma_n = -\infty$, there exists graph $G(n,\widetilde{K_n},\widetilde{P_n})$ under $\widetilde{P_n} = \Omega(n)$ and
$\frac{{\widetilde{K_n}}^2}{\widetilde{P_n}} = \frac{\ln n + {(k-1)} \ln \ln n + {\widetilde{\gamma_n}}}{n}$ with $\lim_{n \to \infty}\widetilde{\gamma_n} = -\infty$ and $\widetilde{\gamma_n} = -O(\ln \ln n)$,
such that there exists a graph coupling\footnote{As used
by Rybarczyk \cite{zz,2013arXiv1301.0466R}, a coupling of two random graphs $G_1$ and
$G_2$ means a probability space on which random graphs $G_1'$ and
$G_2'$ are defined such that $G_1'$ and $G_2'$ have the same
distributions as $G_1$ and $G_2$, respectively. If $G_1'$ is a spanning subgraph
(resp., spanning supergraph) of $G_2'$, we say that under the coupling, $G_1$ is a spanning subgraph
(resp., spanning supergraph) of $G_2$, which yields that for any monotone increasing property $\mathcal {I}$, the probability of $G_1$ having $\mathcal {I}$ is at most (reap., at least) the probability of $G_2$ having $\mathcal {I}$.} under which
$G(n,K_n,P_n)$ is a spanning subgraph of $G(n,\widetilde{K_n},\widetilde{P_n})$.
(ii) If $\lim_{n \to \infty}\gamma_n = \infty$, there exists graph $G(n,\widehat{K_n},\widehat{P_n})$ under $\widehat{P_n} = \Omega(n)$ and
$ \frac{{\widehat{K_n}}^{2}}{{\widehat{P_n}}} = \frac{\ln n + {(k-1)} \ln \ln n + {\widehat{\gamma_n}}}{n}$
with $\lim_{n \to \infty}\widehat{\gamma_n} = \infty$ and $\widehat{\gamma_n} = O(\ln \ln n)$,
such that there exists a graph coupling under which
$G(n,K_n,P_n)$ is a spanning supergraph of $G(n,\widehat{K_n},\widehat{P_n})$.
\end{lem}
\begin{lem} \label{cp_rig_er}
If $K_n = \omega\big((\ln n)^3\big)$, $ \frac{K_n}{P_n} = o\left( \frac{1}{n\ln n} \right)$\vspace{1pt} and $ \frac{{K_n}^2}{P_n} =
o\left( \frac{1}{\ln n} \right)$, then there exists $s_n =
\frac{{K_n}^2}{P_n} \cdot \left[1-
o\left(\frac{1}{ \ln n}\right)\right]$ such that for any monotone increasing graph property $\mathcal{I}$,
\begin{align}
\mathbb{P}[\hspace{2pt}G(n,K_n,P_n)
\textrm{ has }\mathcal{I}.
\hspace{2pt}]
& \geq
\mathbb{P}[\hspace{2pt}G_{ER}(n,s_n)
\textrm{ has }\mathcal{I}.\hspace{2pt} ] - o(1). \label{cp_res_rig_er}
\end{align}
\end{lem}
Lemmas \ref{lem:ER:PM} and \ref{lem:ER:HC} below present results on Erd\H{o}s--R\'enyi graphs \cite{erdosPF}, where
an $n$-node Erd\H{o}s--R\'enyi graph denoted by $G_{ER}(n,s_n)$ is constructed by assigning an edge between each pair of nodes independently with the same probability $s_n$.
\begin{lem}[\hspace{-.1pt}{\cite[Theorem 1]{erdosPF}}\hspace{0pt}] \label{lem:ER:PM}
For an Erd\H{o}s--R\'enyi graph $G_{ER}(n,s_n)$, if there is a sequence $\alpha_n$ with $\lim_{n \to \infty}{\alpha_n} \in [-\infty, \infty]$
such that $s_n = \frac{\ln n +
{\alpha_n}}{n}$,
then it holds that
\begin{align}
\lim_{n \to \infty} \mathbb{P}[ G_{ER}(n,s_n)\text{ has a perfect matching.} ] & = e^{- e^{-\lim\limits_{n \to \infty}{\alpha_n}}}. \nonumber
\end{align}
\end{lem}
\begin{lem} [\hspace{-.1pt}{\cite[Theorem 1]{erdosHC}}\hspace{0pt}] \label{lem:ER:HC}
For an Erd\H{o}s--R\'enyi graph $G_{ER}(n,s_n)$, if there is a sequence $\beta_n$ with $\lim_{n \to \infty}{\beta_n} \in [-\infty, \infty]$
such that $s_n = \frac{\ln n + \ln \ln n +
{\beta_n}}{n}$,
then it holds that
\begin{align}
\lim_{n \to \infty} \mathbb{P}[ G_{ER}(n,s_n)\text{ has a Hamilton cycle.} ] & = e^{- e^{-\lim\limits_{n \to \infty}{\beta_n}}}. \nonumber
\end{align}
\end{lem}
\begin{lem} [\hspace{-.1pt}{\cite[Lemma 1]{Xuan}}\hspace{0pt}] \label{lem:subseq}
Every sequence $a_n|_{n=1,2,\ldots}$ has a subsequence $a_{M_j} |_{j=1,2,\ldots}$ with $\lim_{j \to \infty} a_{M_j} \in [-\infty, \infty]$, where $M_1<M_2<\ldots$ and $M_1, M_2, \ldots \in \mathbb{N}$ ($\mathbb{N}$ is the set of
all positive integers).
\end{lem}
\begin{lem} \label{lem:subsequenc}
In Theorems \ref{thm:PM} and \ref{thm:HC}, and Lemma \ref{thm:ZhaoCDCk}, if the conditions hold for just a subsequence of $\mathbb{N}$ instead of all $n \in \mathbb{N}$ , then the corresponding asymptotic results also hold for the subsequence.
\end{lem}
\section{Establishing Theorems \ref{thm:PM}--\ref{thm:PHW}}
\label{sec:thmprf:kcon}
\subsection{Proof of Theorem \ref{thm:PM}}
Let $\texttt{PerfMatch}$ be short for perfect matching.
Theorem \ref{thm:PM} follows once we prove the two inequalities below:
\begin{align}
\hspace{-1pt} \mathbb{P}[\hspace{.5pt}G(n, K_n, P_n)\text{ has a \texttt{PerfMatch}.}\hspace{.5pt}] & \hspace{-2pt} \leq \hspace{-2pt}e^{- e^{-\lim\limits_{n \to \infty}{\alpha_n}}} \hspace{-2pt} \cdot \hspace{-2pt} [1\hspace{-2pt}+\hspace{-2pt}o(1)]\label{PM-leq}
\end{align}
and
\begin{align}
\hspace{-1pt} \mathbb{P}[\hspace{.5pt}G(n, K_n, P_n)\text{ has a \texttt{PerfMatch}.}\hspace{.5pt}] & \hspace{-2pt} \geq \hspace{-2pt}e^{- e^{-\lim\limits_{n \to \infty}{\alpha_n}}} \hspace{-2pt}\cdot\hspace{-2pt} [1\hspace{-2pt}-\hspace{-2pt}o(1)]. \label{PM-geq}
\end{align}
(\ref{PM-leq}) clearly holds from Lemma \ref{thm:ZhaoCDCk} with $k=1$ and the fact that a necessary condition for a graph to contain a perfect matching is that the minimum degree is at least $1$ (i.e., there is no isolated node).
Now we establish (\ref{PM-geq}). From Lemma \ref{lem-only-prove-lnlnn} and the fact that perfect matching containment is a monotone increasing graph property, we can introduce an auxiliary condition $|\alpha_n |= O(\ln \ln n)$. We will use Lemmas \ref{cp_rig_er} and \ref{lem:ER:PM} to prove (\ref{PM-geq}). We first show that the conditions of Lemma \ref{cp_rig_er} all hold given the conditions of Theorem \ref{thm:PM} and the auxiliary condition $|\alpha_n |= O(\ln \ln n)$. From $|\alpha_n |= O(\ln \ln n)$ and (\ref{thm:PM:pe}), it follows that
\begin{align}
\frac{{K_n}^2}{P_n} & = \frac{\ln n}{n} \cdot [1 \pm o(1)]. \label{KnPnnlnn}
\end{align}
(\ref{KnPnnlnn}) implies $ \frac{{K_n}^2}{P_n} =
o\left( \frac{1}{\ln n} \right)$. Furthermore, (\ref{KnPnnlnn}) and $P_n = \omega\big(n (\ln n)^5\big)$
together yield $K_n = \sqrt{ \frac{{K_n}^2}{P_n} \cdot P_n }= \omega\big((\ln n)^3\big)$ and $ \frac{K_n}{P_n} = \sqrt{\frac{{K_n}^2}{P_n}} / \sqrt{P_n} = o\big( \frac{1}{n(\ln n)^2} \big) = o\left( \frac{1}{n\ln n} \right)$. Then all conditions in Lemma \ref{cp_rig_er} hold. From (\ref{thm:PM:pe}) and (\ref{KnPnnlnn}), the sequence $s_n$ in Lemma \ref{cp_rig_er} satisfies
\begin{align}
s_n & = \frac{\ln n +
{\alpha_n}}{n} - \frac{\ln n}{n} \cdot [1 \pm o(1)] \cdot o\left(\frac{1}{ \ln n}\right) \nonumber \\ & = \frac{\ln n +
{\alpha_n}-o(1)}{n} ,\label{KnPnnlnnres2}
\end{align}
which is used in Lemma \ref{lem:ER:PM} to induce
\begin{align}
\lim_{n \to \infty} \mathbb{P}[ G_{ER}(n,s_n)\text{ has a \texttt{PerfMatch}.} ] & = e^{- e^{-\lim\limits_{n \to \infty}{\alpha_n}}}. \label{KnPnnlnn3}
\end{align}
Then (\ref{PM-geq}) clearly follows from (\ref{KnPnnlnn3}) and Lemma \ref{cp_rig_er}.
We have established Theorem \ref{thm:PM} by showing (\ref{PM-leq}) and (\ref{PM-geq}).
We now explain that Theorem \ref{thm:PM} still follows if $ \frac{{K_n}^2}{P_n}$ in (\ref{thm:PM:pe}) is replaced by the edge probability denoted by $q_n$; i.e., $q_n = \frac{\ln n +
{\alpha_n}}{n}$. First, from \cite[Proof of Lemma 14]{ANALCO}, we can also introduce an auxiliary condition $|\alpha_n |= O(\ln \ln n)$, which along with $q_n = \frac{\ln n +
{\alpha_n}}{n}$ yields $q_n = \frac{\ln n}{n} \cdot [1 \pm o(1)]$. Then
from \cite[Lemma 8]{ZhaoYaganGligor}, it is straightforward to derive $ \frac{{K_n}^2}{P_n} = \frac{\ln n +
{\alpha_n} \pm o(1)}{n} $. Thus Theorem \ref{thm:PM} clearly still follows.
\subsection{Proof of Theorem \ref{thm:HC}}
The proof of Theorem \ref{thm:HC} is similar to that of Theorem \ref{thm:PM}.
Let $\texttt{HamiCycle}$ be short for Hamilton cycle.
Theorem \ref{thm:HC} follows once we prove the two inequalities below:
\begin{align}
\hspace{-1pt} \mathbb{P}[\hspace{.5pt}G(n, K_n, P_n)\text{ has a \texttt{HamiCycle}.}\hspace{.5pt}] & \hspace{-2pt} \leq \hspace{-2pt}e^{- e^{-\lim\limits_{n \to \infty}{\beta_n}}} \hspace{-2pt} \cdot \hspace{-2pt} [1\hspace{-2pt}+\hspace{-2pt}o(1)] \label{HC-leq}
\end{align}
and
\begin{align}
\hspace{-1pt} \mathbb{P}[\hspace{.5pt}G(n, K_n, P_n)\text{ has a \texttt{HamiCycle}.}\hspace{.5pt}] & \hspace{-2pt} \geq \hspace{-2pt}e^{- e^{-\lim\limits_{n \to \infty}{\beta_n}}} \hspace{-2pt}\cdot\hspace{-2pt} [1\hspace{-2pt}-\hspace{-2pt}o(1)]. \label{HC-geq}
\end{align}
(\ref{HC-leq}) clearly holds from Lemma \ref{cp_rig_er} with $k=2$ and the fact that a necessary condition for a graph to contain a Hamilton cycle is that the minimum degree is at least $2$.
Now we establish (\ref{HC-geq}). From Lemma \ref{lem-only-prove-lnlnn} and the fact that Hamilton cycle containment is a monotone increasing graph property, we can introduce an auxiliary condition $|\beta_n |= O(\ln \ln n)$. We will use Lemmas \ref{cp_rig_er} and \ref{lem:ER:HC} to prove (\ref{HC-geq}). We first show that the conditions of Lemma \ref{cp_rig_er} all hold given the conditions of Theorem \ref{thm:HC} and the auxiliary condition $|\beta_n |= O(\ln \ln n)$. From $|\beta_n |= O(\ln \ln n)$ and (\ref{thm:HC:pe}), (\ref{KnPnnlnn}) still follows here.
As explained before, we have $ \frac{{K_n}^2}{P_n} =
o\left( \frac{1}{\ln n} \right)$ and $ \frac{K_n}{P_n} = o\left( \frac{1}{n\ln n} \right)$ from (\ref{KnPnnlnn}) and $P_n = \omega\big(n (\ln n)^5\big)$.
Then all conditions in Lemma \ref{cp_rig_er} hold. From (\ref{thm:HC:pe}) and (\ref{KnPnnlnn}), the sequence $s_n$ in Lemma \ref{cp_rig_er} satisfies
\begin{align}
s_n & = \frac{\ln n + \ln \ln n +
{\beta_n}}{n} - \frac{\ln n}{n} \cdot [1 \pm o(1)] \cdot o\left(\frac{1}{ \ln n}\right) \nonumber \\ & = \frac{\ln n + \ln \ln n +
{\beta_n}-o(1)}{n} ,\label{KnPnnlnnres2}
\end{align}
which is used in Lemma \ref{lem:ER:HC} to induce
\begin{align}
\lim_{n \to \infty} \mathbb{P}[\hspace{1pt}G_{ER}(n,s_n)\text{ has a \texttt{HamiCycle}.}\hspace{1pt}] & = e^{- e^{-\lim\limits_{n \to \infty}{\beta_n}}}. \label{KnPnnlnn3}
\end{align}
Then (\ref{HC-geq}) clearly follows from (\ref{KnPnnlnn3}) and Lemma \ref{cp_rig_er}.
We have established Theorem \ref{thm:HC} by showing (\ref{HC-leq}) and (\ref{HC-geq}).
We now explain that Theorem \ref{thm:HC} still follows if $ \frac{{K_n}^2}{P_n}$ in (\ref{thm:HC:pe}) is replaced by the edge probability denoted by $q_n$; i.e., $q_n = \frac{\ln n + \ln \ln n +
{\beta_n}}{n}$. First, from \cite[Proof of Lemma 14]{ANALCO}, we can also introduce an auxiliary condition $|\beta_n |= O(\ln \ln n)$, which along with $q_n = \frac{\ln n + \ln \ln n +
{\beta_n}}{n}$ yields $q_n = \frac{\ln n}{n} \cdot [1 \pm o(1)]$. Then
from \cite[Lemma 8]{ZhaoYaganGligor}, it is straightforward to derive $ \frac{{K_n}^2}{P_n} = \frac{\ln n + \ln \ln n +
{\beta_n} \pm o(1)}{n} $. Thus Theorem \ref{thm:HC} still follows.
\subsection{Proof of Theorem \ref{thm:PHW}}
\iffalse
\begin{lem} \label{lem:HC:subsequenc}
In Theorems \ref{thm:HC}, if the conditions hold for a subsequence of $\mathbb{N}$, then the asymptotic results hold for $n \to \infty$ along the same subsequence.
\end{lem}
\fi
We define $F_n(\mathcal{I})$ as
\begin{align}
\hspace{-1.5pt}F_n(\mathcal{I}) \hspace{-1.5pt} = \hspace{-1.5pt} \begin{cases} \ln n,~\hspace{2.5pt}\text{if $\mathcal{I}$ is perfect matching containment}, \\
\ln n + \ln \ln n,~\hspace{2.5pt}\text{if $\mathcal{I}$ is Hamilton cycle containment}, \\
\ln n + (k-1) \ln \ln n,~\hspace{2.5pt}\text{if $\mathcal{I}$ is $k$-connectivity}. \end{cases} \label{eq-FnI}
\end{align}
Recalling $K_n^{-}(\mathcal{I}, \epsilon)$ specified in (\ref{KnI-minus}), we further define $\alpha_n^{-}(\mathcal{I}, \epsilon)$ such that
\begin{align}
\frac{{[K_n^{-}(\mathcal{I}, \epsilon)]}^2}{P_n} & = \frac{F_n(\mathcal{I}) +
{\alpha_n^{-}(\mathcal{I}, \epsilon)}}{n}.\label{KnI-minus-alpha}
\end{align}
We now use
\begin{align}
\mathbb{P}[\hspace{1pt}G(n, K_{n}^{-}(\mathcal{I}, \epsilon), P_{n})\text{ has property $\mathcal{I}$.}\hspace{1pt}] & \geq \epsilon \label{geqepsilon}
\end{align}
to establish for any positive constant $\delta_1 < \epsilon$ that
\begin{align}
\frac{{[K_n^{-}(\mathcal{I}, \epsilon)]}^2}{P_n} \hspace{-2pt} \geq \hspace{-2pt} \frac{F_n(\mathcal{I}) \hspace{-2pt}-\hspace{-2pt}\ln (\hspace{-1pt}-\hspace{-1pt}\ln \delta_1)}{n}\text{ for all $n$ sufficiently large,} \label{PMalpha-lowerboundsb}
\end{align}
where (\ref{geqepsilon}) holds due to (\ref{KnI-minus}).
With $\alpha_n^{-}(\mathcal{I}, \epsilon)$ defined in (\ref{KnI-minus-alpha}), (\ref{PMalpha-lowerboundsb}) is equivalent to
\begin{align}
\alpha_n^{-}(\mathcal{I}, \epsilon) \geq -\ln (-\ln \delta_1)\text{ for all $n$ sufficiently large.} \label{PMalpha-lowerbound}
\end{align}
By contradiction, if (\ref{PMalpha-lowerbound}) is not true,
there exists a subsequence $N_i |_{i=1,2,\ldots}$ of $\mathbb{N}$ (the set of
all positive integers) such that
$\alpha_{N_i}^{-}(\mathcal{I}, \epsilon) < -\ln (-\ln \delta_1)$ for $i=1,2,\ldots$. By Lemma \ref{lem:subseq}, there exists a subsequence $M_j |_{j=1,2,\ldots}$ of $N_i |_{i=1,2,\ldots}$ such that $\lim_{j \to \infty} \alpha_{M_j}^{-}(\mathcal{I}, \epsilon) \in [-\infty, \infty]$. From $\alpha_{N_i} < -\ln (-\ln \delta_1)$ for $i=1,2,\ldots$, we have $\lim_{j \to \infty} \alpha_{M_j}^{-}(\mathcal{I}, \epsilon)\in [-\infty, -\ln (-\ln \delta_1)]$. Then from Lemma \ref{lem:subsequenc}, it follows that
\begin{align}
& \lim_{j \to \infty} \mathbb{P}[\hspace{1pt}G(M_j, K_{M_j}^{-}(\mathcal{I}, \epsilon), P_{M_j})\text{ has property $\mathcal{I}$.}\hspace{1pt}] \nonumber \\ & \quad = e^{- e^{-\lim\limits_{j \to \infty}{\alpha_{M_j}^{-}(\mathcal{I}, \epsilon)}}} \leq e^{-e^{\ln (-\ln \delta_1)}} = \delta_1,\nonumber
\end{align}
which contradicts (\ref{geqepsilon}).
Therefore, (\ref{PMalpha-lowerboundsb}) and (\ref{PMalpha-lowerbound}) are established.
Similar to the above analysis of using (\ref{geqepsilon}) to prove (\ref{PMalpha-lowerboundsb}), we can use
\begin{align}
\mathbb{P}[\hspace{1pt}G(n, K_{n}^{-}(\mathcal{I} \hspace{-2pt}- \hspace{-2pt}1, \epsilon), P_{n})\text{ has property $\mathcal{I}$.}\hspace{1pt}] & < \epsilon \label{geqepsilon-a}
\end{align}
to prove
for any positive constant $\delta_2 > \epsilon$ that
\begin{align}
\hspace{-2pt}\frac{{[K_n^{-}(\mathcal{I},\hspace{-1.5pt} \epsilon) \hspace{-2pt}-\hspace{-2pt} 1]}^2}{P_n} \hspace{-2.5pt}\leq \hspace{-2.5pt} \frac{F_n(\mathcal{I}) \hspace{-2.5pt}-\hspace{-2.5pt}\ln (\hspace{-1pt}-\hspace{-1.5pt}\ln \delta_2)}{n}\text{ \hspace{-1pt}for \hspace{-1.5pt}all \hspace{-1.5pt}$n$ \hspace{-1.5pt}sufficiently \hspace{-1.5pt}large,} \label{PMalpha-lowerboundsb-a}
\end{align}
use
\begin{align}
\mathbb{P}[\hspace{1pt}G(n, K_{n}^{+}(\mathcal{I}, \epsilon), P_{n})\text{ has property $\mathcal{I}$.}\hspace{1pt}] & \geq 1-\epsilon \label{geqepsilon-b}
\end{align}
to prove
for any positive constant $\delta_3 < 1 - \epsilon$ that
\begin{align}
\frac{{[K_n^{+}(\mathcal{I}, \epsilon)]}^2}{P_n} \hspace{-2pt} \geq \hspace{-2pt} \frac{F_n(\mathcal{I})\hspace{-2pt} -\hspace{-2pt}\ln (\hspace{-1pt}-\hspace{-1pt}\ln \delta_3)}{n}\text{ for all $n$ sufficiently large,} \label{PMalpha-lowerboundsb-b}
\end{align}
and use
\begin{align}
\mathbb{P}[\hspace{1pt}G(n, K_{n}^{+}(\mathcal{I} - 1, \epsilon), P_{n})\text{ has property $\mathcal{I}$.}\hspace{1pt}] & < 1-\epsilon \label{geqepsilon-c}
\end{align}
to prove
for any positive constant $\delta_4 > 1 - \epsilon$ that
\begin{align}
\hspace{-2pt}\frac{{[K_n^{+}(\mathcal{I},\hspace{-1.5pt} \epsilon) \hspace{-2pt}-\hspace{-2pt} 1]}^2}{P_n} \hspace{-2.5pt}\leq \hspace{-2.5pt} \frac{F_n(\mathcal{I}) \hspace{-2.5pt}-\hspace{-2.5pt}\ln (\hspace{-1pt}-\hspace{-1.5pt}\ln \delta_4)}{n}\text{ \hspace{-1pt}for \hspace{-1.5pt}all \hspace{-1.5pt}$n$ \hspace{-1.5pt}sufficiently \hspace{-1.5pt}large,} \label{PMalpha-lowerboundsb-c}
\end{align}
Considering that the proofs of (\ref{PMalpha-lowerboundsb-a}) (\ref{PMalpha-lowerboundsb-b}) and (\ref{PMalpha-lowerboundsb-c}) are very similar to that of (\ref{PMalpha-lowerboundsb}), we omit the details here due to space limitation. Note that (\ref{geqepsilon-a}) holds from (\ref{KnI-minus}), while (\ref{geqepsilon-b}) and (\ref{geqepsilon-c}) follow from (\ref{KnI-plus}).
\iffalse
For a constant $c$, we define $A_n(c)$ such that
\begin{align}
\frac{{[K_n^{-}(\mathcal{I}, \epsilon)+c]}^2}{P_n}& = \frac{F_n(\mathcal{I}) +
{A_n(c)}}{n}.\label{KnPn:An}
\end{align}
Then
from (\ref{KnI-minus-alpha}) and (\ref{KnPn:An}), we have
\begin{align}
\hspace{-1pt}A_n(c) &\hspace{-2pt} = \hspace{-1pt} \left\{ \sqrt{ \frac{P_n [F_n(\mathcal{I}) +
{\alpha_n^{-}(\mathcal{I}, \epsilon)}]}{n}}+c\right\}^2 \cdot n/P_n \hspace{-1pt}-\hspace{-1pt}F_n(\mathcal{I}) \nonumber \\ & \hspace{-1pt} = \hspace{-1pt}
\alpha_n^{-}(\mathcal{I}, \epsilon) \hspace{-1pt}+ \hspace{-1pt}2c\sqrt{n[F_n(\mathcal{I}) \hspace{-1pt}+\hspace{-2pt} \alpha_n^{-}(\mathcal{I}, \epsilon)]/P_n} \hspace{-1pt}+ \hspace{-1pt}c^2n/P_n .\label{KnI-minus-alpha-plus-c}
\end{align}
\fi
From (\ref{PMalpha-lowerboundsb}) and (\ref{PMalpha-lowerboundsb-a}), it follows that
\begin{align}
\hspace{-1.5pt} \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_1)]}{n}} &\hspace{-1.5pt} \leq \hspace{-1.5pt}K_n^{-}(\mathcal{I}, \epsilon) \nonumber \\
& \hspace{-1.5pt}\leq \hspace{-1.5pt}\sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_2)]}{n}}\hspace{-1.5pt}+\hspace{-1.5pt}1. \label{KnI-minus-alpha-plus-x}
\end{align}
From (\ref{PMalpha-lowerboundsb-b}) and (\ref{PMalpha-lowerboundsb-c}), it follows that
\begin{align}
\hspace{-1.5pt}\sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_3)]}{n}} & \hspace{-1.5pt} \leq \hspace{-1.5pt}K_n^{+}(\mathcal{I}, \epsilon) \nonumber \\
& \hspace{-1.5pt} \leq\hspace{-1.5pt} \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_4)]}{n}}\hspace{-1.5pt}+\hspace{-1.5pt}1. \label{KnI-minus-alpha-plus-y}
\end{align}
With the phase transition width $ d_n(\mathcal{I}, \epsilon)$ defined in (\ref{phase-transition-width}), we obtain from (\ref{KnI-minus-alpha-plus-x}) and (\ref{KnI-minus-alpha-plus-y}) that
\begin{align}
& d_n(\mathcal{I}, \epsilon) +1 \nonumber \\
& \geq \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_3)]}{n}} - \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_2)]}{n}}\label{dnlower}
\end{align}
and
\begin{align}
&d_n(\mathcal{I}, \epsilon) - 1 \nonumber \\
& \leq \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_4)]}{n}} - \sqrt{\frac{P_n[F_n(\mathcal{I}) -\ln (-\ln \delta_1)]}{n}} \label{dnupper}
\end{align}
To use (\ref{dnlower}) and (\ref{dnupper}), we compute
for constants $c_1$ and $c_2$ that
\begin{align}
& \sqrt{\frac{P_n(F_n(\mathcal{I})+c_1)}{n}} - \sqrt{\frac{P_n(F_n(\mathcal{I})+c_2)}{n}} \vspace{-2pt} \nonumber \\
& \quad = \left[\frac{P_n(F_n(\mathcal{I})+c_1)}{n} - \frac{P_n(F_n(\mathcal{I})+c_2)}{n}\right] \vspace{-2pt} \nonumber \\
& \quad\quad\times \left[ \sqrt{\frac{P_n(F_n(\mathcal{I})+c_1)}{n}} + \sqrt{\frac{P_n(F_n(\mathcal{I})+c_2)}{n}}\hspace{2pt}\right]^{-1} \vspace{-2pt} \nonumber \\
&\quad \sim \frac{P_n(c_1-c_2)}{n} \times \left(2\sqrt{\frac{P_n F_n(\mathcal{I})}{n}}\hspace{2pt}\right)^{-1} \vspace{-2pt} \nonumber \\
&\quad \sim \frac{1}{2} (c_1-c_2) \sqrt{\frac{P_n}{n \ln n}} . \label{dnlowerupper}
\end{align}
where the last step uses
$F_n(\mathcal{I}) \sim \ln n$ from (\ref{eq-FnI}) (an asymptotic equivalence $x_n \sim y_n$ means \vspace{.5pt} $\lim_{n \to \infty} \frac{x_n}{y_n} = 1$).
With R.H.S. meaning the right hand side, we get from (\ref{dnlowerupper}) that \\a) if $P_n = o(n\ln n)$, then $\text{R.H.S. of (\ref{dnlower})}=o(1)$ and $\text{R.H.S. of (\ref{dnupper})}=o(1)$, \\b) if $P_n = \Theta(n\ln n)$, then $\text{R.H.S. of (\ref{dnlower})}=\Theta(1)$ and $\text{R.H.S. of (\ref{dnupper})}=\Theta(1)$, and \\c) if $P_n = \omega(n\ln n)$, then $\text{R.H.S. of (\ref{dnlower})}=\omega(1)$ and $\text{R.H.S. of (\ref{dnupper})}=\omega(1)$.
In view that the results in Lemma \ref{thm:ZhaoCDCk} on $k$-connectivity has the condition $P_n = \Omega(n)$, and that the results in
Theorems \ref{thm:PM} and \ref{thm:HC} on perfect matching containment and Hamilton cycle containment respectively both rely on the condition $P_n = \omega\big(n (\ln n)^5\big)$, we obtain the desired results (i)--(iv) from (\ref{dnlower}) (\ref{dnupper}), the above a), b) and c), and $d_n(\mathcal{I}, \epsilon)\geq 0$.
\section{Related Work} \label{related}
For connectivity in $G_u(n,P_n,K_n)$, Blackburn and Gerke \cite{r1}, and Ya\u{g}an and Makowski
\cite{yagan} obtain different granularities of zero--one laws, and
Rybarczyk
\cite{ryb3} establishes the asymptotically exact probability result. For $k$-connectivity, Rybarczyk \cite{zz}
implicitly shows a zero--one law, and we \cite{ZhaoCDC} derive the asymptotically exact probability, as used in Lemma \ref{thm:ZhaoCDCk} of this current paper. For perfect matching containment, Bloznelis and {\L}uczak \cite{Perfectmatchings} give the asymptotically exact probability result,
but their result after a rewriting applies to a different set of conditions on $P_n$ compared with our Theorem \ref{thm:PM} which is for $P_n = \omega\big(n (\ln n)^5\big)$: they consider $P_n = \Omega\big(n (\ln n)^{-1}\big)$ and $P_n = o\big(n (\ln n)^{-\frac{3}{5}}\big)$ instead.
In terms of Hamilton cycle containment, Nikoletseas \emph{et al. } \cite{NikoletseasHM} proves that
$G(n, K_n, P_n)$ under $K_n \geq 2$ has a
Hamilton cycle with high probability
if it holds for some constant $\delta>0$ that $n \geq (1+\delta) \binom{P_n}{K_n} \ln \binom{P_n}{K_n} $, which implies that $P_n$ is much smaller than $n$ ($P_n = O(\sqrt{n}\hspace{1.5pt})$ given $K_n \geq 2$,
$P_n = O(\sqrt[3]{n}\hspace{1.5pt})$ if $K_n \geq 3$, $P_n = O(\sqrt[4]{n}\hspace{1.5pt})$ if $K_n \geq 4$, etc.). Different from the result of Nikoletseas \emph{et al. } \cite{NikoletseasHM}, our Theorem \ref{thm:HC} is for $P_n = \omega\big(n (\ln n)^5\big)$. Furthermore, Theorem \ref{thm:HC} presents the asymptotically exact probability, whereas Nikoletseas \emph{et al. } \cite{NikoletseasHM} only derive conditions for $G_u(n,P_n,K_n)$ to have a Hamilton cycle with high probability. They do not provide conditions for $G_u(n,P_n,K_n)$ to have no Hamilton cycle with high probability, or to have a Hamilton cycle with an asymptotic probability in $(0,1)$.
A graph model related to uniform random intersection graphs is the so-called \emph{binomial random intersection graph model}: each item in a pool is assigned to each node independently with the same probability, and two nodes establish an undirected edge upon sharing at least one item. Note that here the number of items on each node follows a binomial distribution. This graph model has also been studied in the literature as noted below.
For connectivity, Rybarczyk presents a zero--one law
\cite{zz} and later obtains the stronger result of the asymptotically exact probability \cite{2013arXiv1301.0466R}. For $k$-connectivity,
she establishes zero--one laws \cite{zz,2013arXiv1301.0466R}, and we compute the asymptotically exact probability \cite{ZhaoCDC}. For perfect matching containment, Rybarczyk provides a zero--one law
\cite{zz} and derives the asymptotically exact probability \cite{2013arXiv1301.0466R}.
For Hamilton cycle containment, Efthymioua and Spirakis \cite{EfthymiouaHM}, and Rybarczyk \cite{zz,2013arXiv1301.0466R} show zero--one laws.
\iffalse
A graph model related to uniform random intersection graphs is the so-called \emph{binomial random intersection graph model}. In a binomial random intersection graph with $n$ nodes, each item in a pool of $P_n$ different items is assigned to each node independently with a probability $t_n$, and two nodes establish an undirected edge upon sharing at least one item (we denote the induced graph by $G_{\text{binomial}}(n, t_n, P_n)$ ). Note that in $G_{\text{binomial}}(n, t_n, P_n)$,
the number of items on each node follows a binomial distribution with $P_n$ as the number of trials and $t_n$ as the
success probability in each trial, which is different from the case in a uniform random intersection graph $G(n, K_n, P_n)$ where each node has $K_n$ items. The graph $G_{\text{binomial}}(n, t_n, P_n)$ has also been studied in the literature in terms of properties including ($k$-)connectivity, perfect matching containment, and Hamilton cycle containment.
For connectivity, Rybarczyk presents a zero--one law
\cite{zz} and later obtains the stronger result of the asymptotically exact probability \cite{2013arXiv1301.0466R}. For $k$-connectivity,
she shows different granularities of zero--one laws \cite{zz,2013arXiv1301.0466R}, and we compute the asymptotically exact probability \cite{ZhaoCDC}. For perfect matching containment, Rybarczyk provides a zero--one law
\cite{zz} and later derives the asymptotically exact probability \cite{2013arXiv1301.0466R}.
For Hamilton cycle containment in $G_{\text{binomial}}(n, t_n, P_n)$, Efthymioua and Spirakis \cite{EfthymiouaHM}, and Rybarczyk \cite{zz,2013arXiv1301.0466R} establish different granularities of zero--one laws.
binomial random intersection graphs $G_{\text{binomial}}(n, t_n, P_n)$
Efthymioua and Spirakis \cite{EfthymiouaHM} investigate Hamilton cycle containment in a binomial random intersection graph $G_{\text{binomial}}(n, t_n, P_n)$. Using their result \cite[Theorem 3]{EfthymiouaHM}, and \cite[Lemma 4]{Rybarczyk} which relates $G_{\text{binomial}}(n, t_n, P_n)$ and a uniform random intersection graph $G(n, K_n, P_n)$, we can derive the corresponding results on Hamilton cycle containment in $G(n, K_n, P_n)$; see the full version of this paper \cite{ISITfullver} for the details. However, due to the limitation of the results of Efthymioua and Spirakis \cite{EfthymiouaHM} on $G_{\text{binomial}}(n, t_n, P_n)$, the obtained results on $G(n, K_n, P_n)$ are weaker than Theorem \ref{thm:HC} in this paper, and are given as follows: $G(n, K_n, P_n)$ under $P_n = \omega\big(n (\ln n)^5\big)$
has a Hamilton cycle with high probability if $ \frac{{K_n}^2}{P_n} \geq \frac{c \ln n}{n}$ with a constant $c>1$, and
has no Hamilton cycle with high probability if $ \frac{{K_n}^2}{P_n} \leq \frac{c \ln n}{n}$ with a positive constant $c<1$. The above results are implied by Theorem \ref{thm:HC} since $\beta_n$ defined in (\ref{thm:HC:pe}) converges to $\infty$ if $ \frac{{K_n}^2}{P_n} \geq \frac{c \ln n}{n}$ with a constant $c>1$, and converges to $-\infty$ if $ \frac{{K_n}^2}{P_n} \leq \frac{c \ln n}{n}$; moreover, two points that Theorem \ref{thm:HC} uniquely presents but the above results fail to provide are i) a critical threshold for Hamilton cycle containment is $ \frac{{K_n}^2}{P_n} = \frac{\ln n + {(k-1)} \ln \ln n }{n}$, and ii) conditions under which the asymptotic probability of $G(n, K_n, P_n)$ having a Hamilton cycle is neither $0$ nor $1$, but some value in $(0,1)$.
\fi
\section{Conclusion}
\label{sec:Conclusion}
In a uniform random intersection graph,
for perfect matching containment and Hamilton cycle containment, we derive the asymptotically exact probabilities and the phase transition widths. In addition, for $k$-connectivity, we use the asymptotically exact probability result in our prior work \cite{ZhaoCDC} to compute the phase transition width.
\renewcommand\baselinestretch{.95}
\small
|
\section{Introduction}
\label{sec:1}
Since the discovery of quantized conductance in atomic-size contacts, there has been considerable interest in studying the electronic transport in nanoscale systems. So far, a number of calculation methods to investigate the electron transport properties through a nanostructure embedded by semi-infinite electrodes have been proposed such as the nonequilibrium Green's function (NEGF) method \cite{negf}, recursion transfer matrix method \cite{rtm}, Lippmann-Schwinger method \cite{ls} and overbridging boundary-matching (OBM) method \cite{book,obm}. Among them, the NEGF approach is most commonly used in connection with tight-binding models and first-principles methods for solving transport problems. In the NEGF formalism, localized basis sets consisting of either atomic orbitals or Gaussians are extensively utilized because it is straightforward to partition a whole system into a transition region and electrodes. However, the incompleteness of the basis sets sometimes degrades the accuracy \cite{ghost} and the diffuse functions in the basis sets prevent us from achieving high-performance computing on massively parallel computers.
On the other hand, the calculation method using the real-space finite-difference (RSFD) approach~\cite{book,cheliko,tsdg} does not suffer from these problems. However, the linear-equation solver used to obtain the perturbed Green's functions of a transition region becomes computationally intensive as the numbers of grid points and sampling energy points increase. More efficient linear-equation solvers are required to perform large-scale first-principles transport calculations by the NEGF approach using the RSFD scheme.
In this study, we propose an efficient scheme to compute perturbed Green's functions that combines the procedure developed in the OBM method and the shifted conjugate orthogonal conjugate gradient (COCG) algorithm \cite{sCOCG,ss}. In the NEGF method, the matrix to be inverted includes the self-energy terms of electrodes, which are energy-dependent and destroy the shift invariance of the Krylov subspace with respect to the energy of electrons. The total computational cost of obtaining perturbed Green's functions is proportional to the number of energy points. On the other hand, unperturbed Green's functions are first calculated by the OBM method and then the effect of the electrodes is introduced in the wavefunction-matching procedure.
Since the matrix to be inverted does not include nonlinear energy dependent terms, the unperturbed Green's functions are obtained by solving a set of shifted linear equations with respect to the energy. The shifted linear equations can be solved by the shifted COCG method, in which the calculation is actually carried out only for a reference energy point and the solutions for the other shifted energy points are obtained with a moderate computational cost owing to the shift invariance of the Krylov subspace.
Employing the mathematically strict relationship between perturbed and unperturbed Green's functions developed in the OBM method \cite{rsfd-negf}, one can greatly decrease the computational cost of the linear-equation solver.
The rest of this paper is organized as follows. In Sec.~\ref{sec:2}, the NEGF method using the RSFD approach is briefly introduced. In Sec.~\ref{sec:3}, we state the problem of obtaining perturbed Green's functions and introduce the computational scheme combining the procedure developed in the OBM method and the shifted COCG algorithm together with a numerical example. In Sec.~\ref{sec:4}, to demonstrate the efficiency and applicability of the newly developed method, we present a transport calculation of a C$_{60}$@(10,10) carbon nanotube (CNT) peapod connected to CNT electrodes. Finally, we summarize our work in Sec.~\ref{sec:5}.
\section{Computational scheme}
\label{sec:2}
\subsection{Hamiltonian of whole system}
\begin{figure}[h]
\includegraphics{junction.png}
\caption{Schematic of transport calculation model. The transition region is suspended between the left and right semi-infinite electrodes.}
\label{fig:junction}
\end{figure}
We consider ballistic transport through a nanostructure sandwiched between two semi-infinite crystalline electrodes as shown in Fig. \ref{fig:junction}. The periodic boundary condition is applied in the $x$ and $y$ directions, while the semi-infinite boundary condition is imposed in the $z$ direction. The system under investigation can be divided into three regions: the left electrode (L), the transition region (T) and the right electrode (R). By introducing localized basis sets or real-space grids, we can decompose a whole Hamiltonian $\hat{H}$ into the following matrix form:
\begin{equation}
\label{eqn: H matrix}
\hat{H}=
\left
[\begin{array}{c|c|c}
\hat{H}_L&\hat{B}_{LT}&0 \\ \hline
\hat{B}_{LT}^{\dagger}&\hat{H}_{T} &\hat{B}_{TR} \\ \hline
0 & \hat{B}_{TR}^{\dagger}&\hat{H}_{R}
\end{array}
\right],
\end{equation}
where $\hat{H}_L$, $\hat{H}_T$ and $\hat{H}_R$ are the Hamiltonian matrices for the left electrode, transition region and right electrode, respectively. $\hat{B}_{LT}$ ($\hat{B}_{TR}$) is the coupling term between the transition region and the left (right) electrode. Because of electronic screening effects, the coupling between the left and right electrodes is zero when the transition region is sufficiently large.
We next derive the Kohn-Sham (KS) Hamiltonian \cite{ks} for the RSFD scheme. Since the direction of current flow is assumed to be along the $z$ axis, the KS Hamiltonian is composed of the block-matrix elements with a dimension of $N_{xy}(=N_x\times{N_y})$, where $N_{x}$ and $N_{y}$ are the numbers of grid points in the $x$ and $y$ directions, respectively. Wavefunction values on the $x$-$y$ plane at the point $z = z_k$ are expressed by the column vector ${\Psi(z_k)}$ with $N_{xy}$ components given by \{$\psi(x_i,y_j,z_k):i=1,\ldots,N_x,j=1,\ldots,N_y$\}. To simplify the explanation, we here restrict ourselves to the case of the central finite-difference approximation, i.e., $N=1$ in Eq.~(1) of Ref.~\onlinecite{cheliko}, and local pseudopotentials. The extensions to higher-order finite-difference approximations and the inclusion of nonlocal parts of pseudopotentials are given in Ref.~\onlinecite{book}. The KS equation for a whole system is expressed as a relation between three adjacent terms along the $z$ direction
\begin{equation}
-B_z^{\dagger}\Psi(z_{k-1})+[E-A(z_k)]\Psi(z_k)-B_z\Psi(z_{k+1})=0 \;\;(k = -\infty, \ldots, -1, 0, 1, \ldots, \infty),
\label{eqn:sabun1}
\end{equation}
where $E$ is the energy of an electron, $A(z_k)$ is an $N_{xy}$-dimensional block tridiagonal matrix including the potentials on the $x$-$y$ plane at $z=z_{k}$ and the coefficients of the finite-difference approximation, and $B_z$ is a constant matrix that is proportional to the $N_{xy}$-dimensional unit matrix $I$.
The discretized KS equation of the whole system is written as
\begin{equation}
\label{eqn:matrix eq.}
[E-\hat{H}]
\left[
\begin{array}{c}
\vdots \\ \Psi(z_0) \\ \Psi(z_1) \\ \vdots \\ \Psi(z_m) \\ \Psi(z_{m+1}) \\ \vdots
\end{array}
\right]
=0,
\end{equation}
where $\hat{H}$ is the block tridiagonal matrix
\begin{equation}
\label{eqn:matrix eq.}
\hat{H}=
\left
[\begin{array}{ccc|ccccc|ccc}
\ddots & \ddots& 0&&&&&&&& \\
\ddots & A_{z_{-2}} &B_{z} &&&&&&& \\
0 & B_{z}^{\dagger} & A_{z_{-1}} & B_{z_{1}} &&&&&&& \\ \hline
&& B_{z}^{\dagger} & A_{z_0} &B_{z} &&&&&& \\
&&& B_{z}^{\dagger} &A_{z_1} &B_{z} &&&&& \\
&&&& \ddots & \ddots & \ddots &&&& \\
&&&&& B_{z}^{\dagger} & A_{z_{m}}&B_{z} &&& \\
&&&&&& B_{z}^{\dagger} & A_{z_{m+1}} & B_{z} && \\ \hline
&&&&&&& B_{z}^{\dagger} & A_{z_{m+2}} & B_{z} & 0 \\
&&&&&&&& B_{z}^{\dagger} & A_{z_{m+3}} & \ddots \\
&&&&&&&& 0 & \ddots & \ddots \\
\end{array}
\right]
\end{equation}
with $A_{z_{k}}=A(z_{k})$. The partitioning in Eq. (\ref{eqn:matrix eq.}) corresponds to those in Fig. \ref{fig:junction} and Eq. (\ref{eqn: H matrix}). Since the KS Hamiltonian matrix for the RSFD scheme is very sparse and block tridiagonal, we can employ efficient iterative algorithms such as the steepest descent (SD) or conjugate gradient (CG) method to solve the eigenvalue and matrix inversion problems.
\subsection{NEGF method using RSFD approach}
This subsection gives a brief explanation of the NEGF method for later convenience. The NEGF method gives the electron transport properties without calculating the scattering wavefunction explicitly. The Green's function of a whole system can be defined as the resolvent of the Hamiltonian matrix of Eq. (\ref{eqn:matrix eq.}),
\begin{equation}
\label{eqn:GF1}
\hat{G}(Z)=
\left[
\begin{array}{c}
Z-\hat{H}
\end{array}
\right]^{-1},
\end{equation}
where $Z$($=E+i\eta$) is a complex energy with $\eta$ being a positive infinitesimal. The important quantity determining electron transport properties is the perturbed Green's function of the transition region written as
\begin{equation}
\label{eqn:GF2}
\hat{G}_T(Z)=
\left[
\begin{array}{c}
Z-\hat{H}_T-\hat{\Sigma}_L(Z)-\hat{\Sigma}_R(Z)
\end{array}
\right]^{-1},
\end{equation}
where
\begin{equation}
\label{eqn:self-energy}
\hat{\Sigma}_L(Z)=\hat{B}_{LT}^{\dagger}\hat{\mathcal{G}}_L(Z)\hat{B}_{LT} \:\: \mbox{and}\:\: \hat{\Sigma}_R(Z)=\hat{B}_{TR}\hat{\mathcal{G}}_R(Z)\hat{B}_{TR}^{\dagger}
\end{equation}
are the self-energy terms of the left and right electrodes, respectively. Here
\begin{equation}
\label{eqn:surface green's function}
\hat{\mathcal{G}}_L(Z)=[Z-\hat{H}_{L}]^{-1} \:\: \mbox{and}\:\: \hat{\mathcal{G}}_R(Z)=[Z-\hat{H}_{R}]^{-1}
\end{equation}
are the surface Green's functions of the semi-infinite left and right electrodes, respectively. In the RSFD-NEGF scheme, since $\hat{B}_{LT}$ ($\hat{B}_{TR}$) has only one nonzero $N_{xy}$-dimensional block-matrix element $B_z$ [see Eq.~(\ref{eqn:matrix eq.})], the self-energy terms are found to take the very simple forms of
\begin{eqnarray}
{\textstyle \hat{\Sigma}_L}(Z) & = &
\left[
\begin{array}{cccc}
{\textstyle \Sigma_L}(z_0;Z) & 0 & \cdots & 0 \\
0 & 0 & \cdots & 0 \\
\vdots & & \ddots & \\
0 & 0 & \cdots & 0
\end{array}
\right]
\end{eqnarray}
and
\begin{eqnarray}
{\textstyle \hat{\Sigma}_R}(Z) & = &
\left[
\begin{array}{cccc}
0 & \cdots & 0 & 0 \\
& \ddots & & \vdots \\
0 & \cdots & 0 & 0 \\
0 & \cdots & 0 & {\textstyle \sum _R}(z_{m+1};Z)
\end{array}
\right],
\label{eqn:sec2-01}
\end{eqnarray}
where
\begin{eqnarray}
{\textstyle \Sigma_L}(z_0;Z) & = & B_z^{\dagger} \mathcal{G}_L(z_{-1},z_{-1};Z) B_z
\end{eqnarray}
and
\begin{eqnarray}
{\textstyle \Sigma_R}(z_{m+1};Z) & = & B_z \mathcal{G}_R(z_{m+2},z_{m+2};Z) B_z^{\dagger}
\label{eqn:sec2-02}
\end{eqnarray}
with $\mathcal{G}_{\{L,R\}}(z_k,z_l;Z)$ being the $N_{xy}$-dimensional $(k,l)$ block-matrix element of $\hat{\mathcal{G}}_{\{L,R\}}(Z)$. The perturbed Green's function of the transition region $\hat{G}_T(Z)$ can be obtained by inverting the finite-size matrix given as
\begin{equation}
\label{eqn:NEGF Hamiltonian}
\left[
\begin{array}{ccccc}
D({z_0};Z)-{\textstyle \Sigma_L}(z_0;Z) & -B_z &&&0 \\
-B_z^{\dagger} &D({z_1};Z)&-B_z && \\
&\ddots &\ddots &\ddots& \\
&& -B_{z}^{\dagger}&D({z_m};Z)&-B_{z} \\
0&& &-B_{z}^{\dagger}&D({z_{m+1}};Z)-{\textstyle \Sigma_R}(z_{m+1};Z)\\
\end{array}
\right],
\end{equation}
where $D({z_l};Z)=Z-A(z_l)$.
The expression for the conductance at a real energy $E$ is given by the Fisher-Lee formula \cite{fisherlee}:
\begin{equation}
\label{eqn:TF}
T(E)=
\mathrm{Tr}\left[
\begin{array}{c}
\Gamma_L(z_0;E)G_T^a(z_0,z_{m+1};E)\Gamma_R(_{m+1};E)G_T^r(z_{m+1},z_0;E)
\end{array}
\right],
\end{equation}
where $G_T^r(E)$ is the retarded Green's function of the transition region defined as
\begin{equation}
G_T^r(z_k,z_l;E)=\lim_{\eta \to 0^+}G_T(z_k,z_l;E+i\eta),
\end{equation}
$G_T^a(E)(=[G_T^r(E)]^{\dagger})$ is the advanced Green's function, and $\Gamma_{L(R)}$ is the coupling matrix calculated from the imaginary part of the self-energy term,
\begin{equation}
\Gamma_{X}(z_k;E)=i[\Sigma_X(z_k;E)-\Sigma_X^{\dagger}(z_k;E)] \;\;\;\;\ (X\in{L,R}).
\end{equation}
\section{Efficient procedure to obtain perturbed Green's function}
\label{sec:3}
\subsection{Difficulties in obtaining perturbed Green's function by iterative method}
The simplest way to obtain $\hat{G}_T(Z)$ is the direct inversion of the matrix given by Eq. (\ref{eqn:NEGF Hamiltonian}). However, direct matrix inversion becomes impractical when the matrix size increases. In addition, iterative solvers for linear equations, such as the SD and CG methods, are not efficient in this case because the matrix to be inverted given by Eq. (\ref{eqn:NEGF Hamiltonian}) is neither Hermitian nor very sparse owing to the self-energy terms. In addition, the a computation of $\hat{G}_T(Z)$, which is carried out independently at each energy, is a computationally demanding task. To circumvent these difficulties, we propose a computational procedure employing the unperturbed Green's function used in the OBM method \cite{obm,book} and the shifted COCG algorithm \cite{sCOCG,ss} that uses the collinear residual theorems \cite{frommer} in the following subsection.
\subsection{Unperturbed Green's function and shifted COCG method}
The unperturbed Green's function of the transition is written as
\begin{equation}
\label{eqn:GF4}
\hat{\mathcal{G}}_T(Z)=[Z-\hat{H}_{T}]^{-1}.
\end{equation}
The relationship between the unperturbed and perturbed Green's functions is expressed as the following equation, which was derived in Ref. \onlinecite{rsfd-negf}.
\begin{equation}
\left[
\begin{array}{c}
G_T(z_0,z_l;Z) \\
G_T(z_1,z_l;Z) \\
\\
\vdots \\
G_T(z_l,z_l;Z) \\
\vdots \\
\\
G_T(z_m,z_l;Z) \\
G_T(z_{m+1},z_l;Z) \\
\end{array}
\right]
=\hat{\mathcal{G}}_T(Z) \left[
\begin{array}{c}
\sum_L(z_0;Z)G_T(z_0,z_l;Z) \\
0 \\
\vdots \\
0 \\
I \\
0 \\
\vdots \\
0 \\
\sum_R(z_{m+1};Z)G_T(z_{m+1},z_l;Z) \\
\end{array}
\right]
\begin{array}{l}
\\
\\
\\
\\
.\leftarrow \mbox{the} \hspace{2mm} l\mbox{th} \\
\\
\\
\\
\\
\end{array}
\label{eqn:sec3-01a}
\end{equation}
For $l=0$ and $m+1$,
\begin{eqnarray}
G_T(z_0,z_0;Z) &=& \tilde{\mathcal{G}}_T(z_0,z_0;Z)\left[ I-{\textstyle \sum_L}(z_0;Z)\tilde{\mathcal{G}}_T(z_0,z_0;Z) \right]^{-1},\label{eqn:sec3-02}
\\
G_T(z_{m+1},z_0;Z) &=& \Bigl[I - \mathcal{G}_T(z_{m+1},z_{m+1};Z){\textstyle \sum_R}(z_{m+1};Z)\Bigr]^{-1} \nonumber \\
&& \times\ \mathcal{G}_T(z_{m+1},z_{0};Z) \left[ I-{\textstyle \sum_L}(z_0;Z)\tilde{\mathcal{G}}_T(z_0,z_0;Z) \right]^{-1}, \\
G_T(z_0,z_{m+1};Z) &=& \Bigl[ I - \mathcal{G}_T(z_{0},z_{0};Z){\textstyle \sum_L}(z_{0};Z)\Bigr]^{-1}\mathcal{G}_T(z_{0},z_{m+1};Z) \nonumber \\
& & \times \left[ I-{\textstyle \sum_R}(z_{m+1};Z)\tilde{\mathcal{G}}_T(z_{m+1},z_{m+1};Z) \right]^{-1},
\end{eqnarray}
and
\begin{eqnarray}
G_T(z_{m+1},z_{m+1};Z) &=& \tilde{\mathcal{G}}_T(z_{m+1},z_{m+1};Z)\left[ I-{\textstyle \sum_R}(z_{m+1};Z)\tilde{\mathcal{G}}_T(z_{m+1},z_{m+1};Z) \right]^{-1},
\label{eqn:sec3-03}
\end{eqnarray}
where $\tilde{\mathcal{G}}_T(z_i,z_j;Z)$ is the modified $\mathcal{G}_T(z_i,z_j;Z)$ under the influence of the self-energy terms, ${\textstyle \sum_R}$ and ${\textstyle \sum_L}$, which is expressed as
\begin{eqnarray}
\tilde{\mathcal{G}}_T(z_0,z_0;Z) &=& \mathcal{G}_T(z_0,z_0;Z) \nonumber \\
&& \hspace{-20mm} + \mathcal{G}_T(z_0,z_{m+1};Z) {\textstyle \sum_R}(z_{m+1};Z) \nonumber \\
&& \hspace{-20mm} \times\ \Bigl[ I - \mathcal{G}_T(z_{m+1},z_{m+1};Z){\textstyle \sum_R}(z_{m+1};Z)\Bigr]^{-1}\mathcal{G}_T(z_{m+1},z_{0};Z)
\end{eqnarray}
and
\begin{eqnarray}
\tilde{\mathcal{G}}_T(z_{m+1},z_{m+1};Z) &=& \mathcal{G}_T(z_{m+1},z_{m+1};Z) \nonumber \\
&& \hspace{-20mm} + \mathcal{G}_T(z_{m+1},z_{0};Z) {\textstyle \sum_L}(z_0;Z) \nonumber \\
&& \hspace{-20mm} \times\ \Bigl[ I - \mathcal{G}_T(z_{0},z_{0};Z){\textstyle \sum_L}(z_{0};Z)\Bigr]^{-1} \mathcal{G}_T(z_{0},z_{m+1};Z).
\label{eqn:sec3-06}
\end{eqnarray}
The advantage of computing the unperturbed Green's function is that the matrix inverted in Eq. (\ref{eqn:GF4}), $Z-\hat{H}_T$, is very sparse and complex symmetric. Moreover, $Z_k-\hat{H}_T$ are shifted matrices for a given set of energy points $Z_k$ $(k=1,2,...)$, which is in contrast to $Z_k-\hat{H}_T-\hat{\Sigma}_L(Z_k)-\hat{\Sigma}_R(Z_k)$ in Eq.~(\ref{eqn:GF2}). Therefore, shifted methods such as the shifted COCG method \cite{sCOCG} can be applied to obtain unperturbed Green's functions.
To solve the complex-symmetric linear equations $(Z_k-\hat{H}_T)x^{(k)}=e_j $, where $e_{j}$ denotes the $j$th unit vector, by the conventional COCG method \cite{COCG}, we build up the Krylov subspaces independently at each complex energy $Z_k$, which requires time-consuming matrix-vector operations for each energy. On the other hand, the shifted COCG method makes use of the fact that Krylov subspaces are shift-invariant; to solve the linear equations for a given set of sampling energy points, the spanned Krylov subspace is common to all the energy points. Once the Krylov subspace has been built up for the reference energy point, we can obtain the Green's functions for all the sampling energy points without matrix-vector operations during the iteration steps.
It is noteworthy that we do not need to store the whole matrices of the unperturbed Green's functions $\hat{\mathcal{G}}_T(Z)$ in memory. To calculate the transport properties by Eq. (\ref{eqn:TF}), the block-matrix elements of the perturbed Green's functions, $G_T(z_{m+1},z_0;Z)$ and $G_T(z_0,z_{m+1};Z)$, are required. In addition, the density of states is relevant to the diagonal elements of the perturbed Green's functions, $G_T(z_{l},z_{l};Z)$. The evaluation of $G_T(z_{l},z_{l};Z)$ requires the additional computation of the off-diagonal block elements $G_T(z_0,z_{l};Z)$ and $G_T(z_{m+1},z_l;Z)$, which are analytically determined in an analogous manner to the derivation of Eqs. (\ref{eqn:sec3-02})-(\ref{eqn:sec3-03}) from Eq. (\ref{eqn:sec3-01a}). It is sufficient to store a couple of block-matrix elements of the unperturbed Green's functions so as to perform matrix-matrix operations between the $N_{xy}$-dimensional block-matrix elements of the Green's functions and the self-energy terms. Thus, we can obtain the perturbed Green's functions with a moderate computation time and memory load.
There is no guarantee that the residual norms of all the sampling energy points satisfy the convergence criterion when that of the reference energy point is sufficiently small. Because the iterations are performed until all the residual norms satisfy the convergence criterion, the total number of iterations increases, resulting in the slight increase of CPU time in the shifted COCG method with respect to the number of sampling energy points. In addition, when the norm of the residual vector of the seed system becomes small, the numerical precision of the residual vectors for the sampling energy points degrades. The seed switching technique \cite{ss}, which switches the reference energy point to the point where the residual norm is largest among the sampling energy points after the solution of the reference energy point converges, enables us to build up the residual vectors for the sampling energy points without loss of numerical precision and significant increase of the computational cost. The detail of the seed switching technique is introduced in Appendix.
\subsection{Numerical test}
To demonstrate the efficiency of the present scheme, we apply the shifted COCG method to the calculation of the unperturbed Green's functions of the transition region for a Na atomic wire. The valence electron-ion interaction is described using norm-conserving pseudopotentials \cite{norm} generated by the scheme proposed by Troullier and Martins \cite{tm}. Exchange-correlation effects are treated within the local density approximation \cite{lda} of density functional theory \cite{dft,ks}. The central finite-difference approximation is adopted for the second-order derivation arising from the kinetic energy operator in the KS equation. The dimensions of the supercell are $20.0\times20.0\times7.0$ bohr and a single Na atom is contained in the supercell. The real-space grid spacing is chosen to be $\sim0.45$ bohr, which corresponds to a total Hamiltonian dimension of 12800. All calculations are carried out on a workstation with an Intel Xeon E5-2667v2 CPU.
\begin{figure}[h]
\centering
\includegraphics{calctime.png}
\caption{CPU time required to obtain unperturbed Green's functions of Na atomic wire. The red squares are the results obtained by the conventional COCG method and the black triangles are those obtained by the shifted COCG method. The energy points are shifted from the Fermi energy by 0.1 eV.}
\label{fig:calctime}
\end{figure}
Figure \ref{fig:calctime} shows the CPU time versus the number of sampling energy points for the shifted COCG and conventional COCG methods. The initial reference energy point is chosen to be the Fermi energy. For the conventional COCG method, the CPU time is proportional to the number of energy points because the Green's functions are calculated independently at each energy point. On the other hand, in the shifted COCG method, the matrix-vector operations are carried out only at the reference energy point and the cost of the scalar-vector operations to update the approximate solutions for all the shifted energy points is negligibly small. The increase in the computational time is attributed to the time-consuming matrix-vector operations in the additional iteration steps to reach global convergence.
\section{Application}
\label{sec:4}
\begin{figure}[h]
\includegraphics{peapod.png}
\caption{Atomic structure of (10,10)CNT with single C$_{60}$ in transition region. The dashed lines are the boundaries of the supercell used to determine the atomic structure and KS effective potential.}
\label{fig:peapod}
\end{figure}
We next compute the conductance spectrum of a C$_{60}$@(10,10)CNT peapod, in which a (10,10)CNT encapsulates a C$_{60}$ molecule, as an application of the proposed method to a large system. Figure \ref{fig:peapod} shows the computational model. This system has been observed by transmission electron microscopy \cite{peapod3} and its electronic structure has been intensively studied by theoretical calculations \cite{peapod,peapod2,ohno}. However, the transport properties of peapods have mostly been investigated using the tight-binding approximation \cite{ohno} because of the large system size required for calculations.
The size of the transition region is taken to be $42.2\times42.2\times32.6$ bohr and the transition region contains 340 carbon atoms. The initial carbon-carbon distance is set to 2.68 bohr \cite{peapod2} and structural relaxation is carried out. To determine the KS effective potential for the transport calculation, the conventional supercell indicated by the rectangle in Fig. \ref{fig:peapod} is used, where the periodic boundary condition in imposed in all directions, and the grid spacing is $\sim0.45$ bohr, which gives a matrix of $774144\times774144$ for the Hamiltonian of the transition region. The number of sampling energy points in the transport calculation is 101. Calculations are implemented using 64 Intel Xeon E5-2697v2 CPUs. The total elapsed time required to calculate the unperturbed Green's functions for 101 energy points is 15294 second, in which the average number of iterations required to satisfy the convergence criterion is 6713, while the elapsed time and the average number of iterations are 6029 second and 3122, respectively, when the unperturbed Green's function is only calculated for the Fermi energy.
The electronic band structures of the (10,10)CNT and C$_{60}$@(10,10)CNT peapod are shown in Figs.~\ref{fig:bands}(a) and \ref{fig:bands}(b), respectively. In the band structure of the C$_{60}$@(10,10)CNT peapod, there are three flat bands above the Fermi level, which originate from the $t_{1u}$ orbitals of the isolated C$_{60}$ molecule. Owing to the hybridization between the $t_{1u}$ orbitals of the C$_{60}$ molecule and $\pi$ orbitals of the (10,10)CNT, the $\pi$ bands of the (10,10)CNT are divided by the bands originating from the $t_{1u}$ orbitals.
\begin{figure}[h]
\begin{center}
\includegraphics{cband.png}
\\
\includegraphics{pband.png}
\end{center}
\caption{Electronic band structures of (a) (10,10)CNT and (b) C$_{60}$@(10,10)CNT. The Fermi level is marked by the dotted line.}
\label{fig:bands}
\end{figure}
The conductance spectrum of the C$_{60}$@(10,10)CNT is plotted in Fig. \ref{fig:spectra}. Owing to the use of the shifted COCG method, we can sample a large number of energy points so as to identify the spiky dips in the spectrum. It is found that three dips appear near the energy range of the flat bands. Figure \ref{fig:swf} shows a real-space picture of the charge densities of scattering wavefunctions, which is useful for understanding the spatial behaviors of the transport phenomenon. The charge density distribution with the energy at $T(E)\approx2$ spreads around the (10,10)CNT, while that with the energy at the dip of the conductance spectrum accumulates in the vicinity of the C$_{60}$ molecule. These results imply the incident electrons are scattered by the $t_{1u}$ orbitals of the encapsulated C$_{60}$ molecule. Since peapods have recently become experimentally accessible structures, this information should be helpful in understanding and predicting the resonant scattering in experimental transport measurements \cite{sts}.
\begin{figure}[h]
\includegraphics{spectrum.png}
\caption{Conductance spectrum of C$_{60}$@(10,10)CNT.}
\label{fig:spectra}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics{swf1.png}
\\
\includegraphics{swf2.png}
\end{center}
\caption{Charge density distributions of scattering wavefunctions of C$_{60}$@(10,10)CNT. (a) and (b) correspond to the energies indicated by the arrows in Fig. \ref{fig:spectra}. The spheres represent the positions of carbon atoms. Each contour represents twice or half the charge density of the adjacent contour lines. The lowest-density contour represents a density of $5.0\times10^{-4}$ e/$\AA^3$}
\label{fig:swf}
\end{figure}
\section{Conclusion}
\label{sec:5}
An efficient scheme to compute perturbed Green's functions, which combines the procedure developed in the OBM method based on the RSFD approach and the shifted COCG algorithm, has been proposed. Since the number of grid points in the RSFD approach is larger than the number of bases in the tight-binding approach, the computational cost of obtaining the perturbed Green's functions of the transition region has been the bottleneck. In this procedure, unperturbed Green's functions are calculated and perturbed Green's functions are obtained by employing the mathematically strict relationship between the perturbed and unperturbed Green's functions. The main advantage of the present scheme for computing unperturbed Green's functions is that the spanned Krylov subspace is common to all the sampling energy points. Therefore, once the Krylov subspace for the reference energy point has been built up, the Green's functions for all the sampling energy points can be obtained without time-consuming matrix-vector operations. To demonstrate the potential power of the present method, the transport properties of a C$_{60}$@(10,10)CNT peapod suspended between (10,10)CNTs are investigated. By sampling a large number of energy points owing to the use of the present method, the scattering attributed to the coupling between the states of the CNT and C$_{60}$ can be observed in the calculated conductance spectrum.
The present technique enables us to efficiently obtain perturbed Green's functions in the framework of the RSFD approach. The RSFD scheme of first-principles calculations is free of problems concerning the completeness of the basis set that appear in the methods using localized basis sets of either atomic orbitals or Gaussians. Moreover, this scheme has the advantage of being scalable to massively parallel computers without compromising on precision. We believe that this development will be important for executing large-scale transport calculations using massively parallel computers.
\section*{Acknowledgment}
The authors would like to thank Professor Shao-Liang Zhang of Nagoya University, Professor Tomohiro Sogabe of Aichi Prefectural University, and Professor Susumu Yamamoto of Tokyo University of Technology for contributing to our fruitful discussions. This research was partially supported by the Computational Materials Science Initiative (CMSI) and a Grant-in-Aid for Scientific Research on Innovative Areas (Grant No. 22104007) from the Ministry of Education, Culture, Sports, Science and Technology, Japan. The numerical calculations were carried out using the computer facilities of the Institute for Solid State Physics at the University of Tokyo, the Center for Computational Sciences at University of Tsukuba, and K computer under a trial use at Advanced Institute for Computational Science at RIKEN.
|
\section{Introduction}
\baselineskip=0.6cm
A perfect map is a kind of continuous function between
topological spaces. Perfect maps are weaker than homeomorphisms, but strong enough to preserve some
topological properties such as local compactness that are not always preserved by continuous maps.
Let $X, Y$ be two topological spaces such that $X$ is Hausdorff. A continuous map $f:X\to Y$ is said to be a perfect map provided that $f$ is closed, surjective and each fiber $f^{-1}(y)$ is compact in $X$. In \cite{KNS} the authors are introduced a generalization of compactness, i.e., $\lambda$-compactness. Motivated by this concept we are led to generalize the concept of perfect mapping. In the current section we recall some preliminary definitions and related results such as $\lambda$-compact spaces and $P_\lambda$-spaces which are presented in \cite{KNS}. In Section 2, we will introduce $\lambda$-perfect maps and generalize some classical results which are related to the perfect maps and finally we prove that if the composition of two continuous mappings is $\lambda$-perfect, then its components are $\mu$-compact and $\beta$-compact, where $\mu, \beta\leq\lambda$.
\begin{definition}\label{def 0-1}
A topological space $X$ (not necessarily Hausdorff) is said to be $\lambda$-compact whenever each open cover of $X$ can be reduced to an open cover of $X$ whose cardinality is less than $\lambda$, where $\lambda$ is the least infinite cardinal number with this property. $\lambda$ is called the compactness degree of $X$ and we write $d_c(X)=\lambda$.
\end{definition}
We note that compact spaces, Lindel\"{o}f non-compact spaces are $\aleph_\circ$-compact, $\aleph_1$-compact spaces, respectively and in general every topological space $X$ is $\lambda$-compact for some infinite cardinal number $\lambda$, see the concept of Lindel\"{o}f number in \cite[p 193]{Engelking}.
The following useful Lemma is somehow a generalization of the fact that continuous function $f:X\to Y$ takes compact sets in $X$ to compact sets in $Y$.
\begin{lemma}\label{lem 0-2}
Let $X, Y$ be two topological spaces. If $f:X\to y$ is a continuous function and $A\subseteq X$ is $\lambda$-compact, then $d_c(f(A))\leq\lambda$.
\end{lemma}
\begin{proposition}\label{pro 0-3}
Let $F$ be a closed subset of topological space $X$, then $d_c(F)\leq d_c(X)$.
\end{proposition}
Next we are going to introduce a generalization of the concept $P$-space (Pseudo-discrete space). $P$-spaces are very important in the contexts of rings of continuous functions which were introduced by Gillman and Henriksen, see \cite{GH}.
\begin{definition}\label{def 0-4}
Let $X$ be a topological space. The intersection of any family with cardinality less than $\lambda$ of open sets of $X$ is called a $G_\lambda$-set. Obviously, $G_\delta$-sets are precisely $G_{\aleph_1}$-sets.
\end{definition}
\begin{definition}\label{def 0-5}
The topological space $X$ is said to be a $P_\lambda$-space whenever each $G_\lambda$-set in $X$ is open.
It is manifest that every arbitrary space is a $P_{\aleph_0}$-space and $P_{\aleph_1}$-spaces are precisely $P$-spaces. See Example 1.8 and Example 1.9 in \cite{KNS}
\end{definition}
\section{$\lambda$-perfect maps}
\begin{lemma}\label{lemma-1}
If A is a $\beta$-compact subspace of a space $X$ and $y$ a point of a
$P_{\lambda} $-space $Y$ such that $\beta\leq\lambda$ then for every open set
$W\subseteq X\times Y$ containing $A\times\{y\}$ there exist open sets
$U\subseteq X$ and $V\subseteq Y$ such that $A\times\{y\}\subseteq U\times
V\subseteq W.$
\end{lemma}
\begin{proof}
For every $x\in A$ the point $(x,y)$ has a neighborhood of the form $U_x\times
V_x$ contained in $W$. Clearly $A\times\{y\}\subseteq\bigcup_{x\in
A}U_x\times V_x$. since $A$ is $\beta$-compact; there exists a family
$\{x_i\}_{i\in I}$ with $|I|<\beta\leq\lambda$ such that
$A\times\{y\}\subseteq\bigcup_{i\in I}U_{x_i}\times V_{x_i}$. Clearly the sets
$U=\bigcup_{i\in I}U_{x_i}$ and $V=\bigcap_{i\in I}V_{x_i}$ have the required
properties.
\end{proof}
Let us recall Kuratowski theorem.
\begin{theorem}\label{theorem-0-1}
For a topological space $X$ the following conditions are equivalent:\\
(i) The space $X$ is compact.\\
(ii) For every topological space $Y$ the projection $p:X\times Y\to Y$ is
closed. \\
(iii) For every $T_4$-space $Y$ the projection $p:X\times Y\to
Y$ is closed.
\end{theorem}
Next we generalize the above theorem.
\begin{theorem}\label{theorem-1}
For an infinite cardinal
number $\lambda$ and a topological space $X$ the following conditions are equivalent.\\
(i) The space $X$ is $\beta$-compact for $\beta\leq\lambda$. \\
(ii) For every $P_{\lambda} $-space $Y$ the projection $p:X\times Y\to Y$ is
closed. \\
(iii) For every $T_4$, $P_{\lambda} $-space $Y$ the projection $p:X\times Y\to
Y$ is closed.
\end{theorem}
\begin{proof}
(i)$\Rightarrow$(ii) Let $X$ be a $\beta$-compact space and $F$ be a closed
subspace of $X\times Y$. Take a point $y\notin p(F)$; as
$X\times\{y\}\subseteq(X\times Y)\setminus F$, it follows from the lemma \ref{lemma-1}
that $y$ has a neighborhood $V$ such that $(X\times V)\bigcap F=\emptyset$.
We then have $p(F)\bigcap V=\emptyset$, which shows that $p(F)$ is a closed
subset of $Y$. \\
(ii)$\Rightarrow$(iii) It is obvious\\
(iii)$\Rightarrow$ (i) Suppose that there exists a family
$\{F_s\}_{s\in S}$ of closed subset of $X$ with the $\lambda$- intersection
property such that $\bigcap_{s\in S}F_s=\emptyset$. Take a point $y_0\notin X$
and on the set $Y=X\bigcup\{y_0\}$ consider the topology $T$ consisting of all
subsets of $X$ and of all sets of the form$$\{y_0\}\cup(\bigcap_{i\in
I}F_{s_i})\cup K~~ \mbox{where}~~ s_i\in S~~ ,~~ |I|<\beta~~
\mbox{and}~~K\subseteq X. $$
since $\bigcap_{s\in S}F_s=\emptyset$ for every $x\in X$ there exists $s_x\in S$
such that $x\notin F_{s_x}$. Notice that the set $$\{y_0\}\cup
F_{s_x}\cup(X\setminus \{x\})=Y\setminus\{x\}$$ belong to $T$
and since $X\in T$, $\{y_0\}$ is closed, therefor $Y$ is $T_1$-space.\\ If $A$
and $B$ are two disjoint closed subsets of $Y$ then at least one of them, say
$A$, does not contain $y_0$ and so is open; therefor $A$ and $Y\setminus A$ are
two disjoint open subset of $Y$ that contains $A$ and $B$ respectively. Hence
$Y$ is normal.
Take $F=\overline{\{(x,x):x\in X\}}\subseteq X\times Y$. By hypothesis $p(F)$
is closed in $Y$. Since $\{y_0\}$ is not open, every open subset of $Y$ that
contains $y_0$ meets $X$. Hence $\overline X=Y$.\\
If $x\in X$ then $(x,x)\in F$ therefor $x=p(x,x)\in p(F)$. Hence $X\subseteq
p(F)$ $$ \Rightarrow y_0\in Y=\overline X\subseteq\overline{p(F)}=p(F).$$
Therefor there exists $x_0\in X$ such that $(x_0,y_0)\in F.$ For every
neighborhood $U\subseteq X$ of $x_0$ and every $s\in S$ the set
$U\times(\{y_0\}\cup F_s) $ is open in $X\times Y$; hence
$$[U\times(\{y_0\}\cup F_s)]\cap\{(x,x):x\in X\}\neq\emptyset,$$ i.e., $U\cap
F_s\neq\emptyset$, which shows that $x_0\in F_s$ for every $s\in S$. This
implies that $\bigcap_{s\in S}F_s\neq\emptyset$, and we have a contradiction.
\end{proof}
\begin{definition}
Let $X$ and $Y$ be two topological spaces and $f:X\to Y$ be a continuous
function.
We say that $f$ is $\lambda$-perfect whenever $f$ be closed and $\lambda$ be the least infinite cardinal number such that for every $y\in Y$ the compactness degree of $f^{-1}(y)$ is less than or equal to $\lambda$ .
\end{definition}
\begin{theorem}
If $Y$ is a $P_{\lambda} $-space and $X$ is a $\beta$-compact space such that
$\beta\leq\lambda$, then the projection $p:X\times Y\to Y$ is $\gamma$-perfect
for $\gamma\leq\beta$.
\end{theorem}
\begin{proof}
Theorem \ref{theorem-1} implies that $p$ is closed. For every $y\in Y$ the set
$p^{-1}(y)$
is of the form $A\times\{y\}$ for some $A\subseteq X$ and it is closed in $X\times Y$,
by the continuity of $p$, So it is closed in $X\times\{y\}$ and since this set is $\beta$-
compact, we infer that $p^{-1}(y)$ is $\alpha$-compact for $\alpha\leq\beta$ by \cite[Proposition 2.6]{NS}.
Consequently $p$ is $\gamma$-perfect for $\gamma\leq\beta$.
\end{proof}
The following theorem is an extension of Theorem 3.7.2 in \cite{Engelking}
\begin{theorem}
Let $\lambda$ and $\beta$ be two regular cardinal numbers and $f:X\to Y$ be a $\lambda$-perfect mapping, then for every $\beta$-compact
subspace $Z\subseteq Y$ the inverse image $f^{-1}(Z)$ is $\mu$-compact for
$\mu\leq\max\{\lambda , \beta\}$.
\end{theorem}
\begin{proof}
Let $\{U_s\}_{s\in S}$ be a family of open subsets of $X$ whose union contains
$f^{-1}(Z)$, $K$ be the family of all subsets of $S$ with cardinality less than
$\lambda$ and $ U_T=\bigcup_{s\in T}U_s$ for $T\in K$. For each $z\in Z$, $d_{c}(f^{-1}(z))\leq\lambda$ and thus is contained in the set $U_T$ for
some $T\in K$; it follows that $z\in Y\setminus f(X\setminus U_T)$, and thus:
$$Z\subseteq\bigcup_{T\in K}(Y\setminus f(X\setminus U_{T})).$$ Since $f$ is
closed; the sets $Y\setminus f(X\setminus U_{T})$ being open. Hence there
exists
$I\subseteq K$ with $|I|< \beta$ such that $Z\subseteq\bigcup_{T\in I}
(Y\setminus f(X\setminus U_{T}))$. Hence
$$f^{-1}(Z)\subseteq\bigcup_{T\in I}f^{-1}(Y\setminus f(X\setminus
U_T))=\bigcup_{T\in I}(X\setminus f^{-1}f(X\setminus
U_T))\subseteq$$$$\bigcup_{T\in I}(X\setminus(X\setminus
U_T))=\bigcup_{T\in I}U_T=\bigcup_{s\in S_0}U_s,$$ where
$S_0=\bigcup_{T\in I}T.$\\
We notice that $S_0$ is the union of a family with cardinality less than $\beta$
of member of $K$ which has cardinality less than $\lambda$. Since $\lambda$
and $\beta$ are regular cardinals; $|S_0|<\max\{\lambda , \beta\}$, i.e.,
$f^{-1}(Z)$ is $\mu$-compact for $\mu\leq\max\{\lambda , \beta\}$.
\end{proof}
The following corollary is now immediate.
\begin{corollary}
If $f:X\to Y$ is an onto $\lambda$-perfect mapping, then for every $\lambda$-
compact subspace $Z\subseteq Y$ the inverse image $f^{-1}(Z)$ is $\lambda$-
compact too.
\end{corollary}
\begin{proof}
By previous theorem $f^{-1}(Z)$ is $\mu$-compact for $\mu\leq\lambda$; and
by ( ) $Z=f(f^{-1}(Z))$ is $\gamma$-compact for $\gamma\leq\mu$. But by
hypothesis $\gamma=\lambda$, so, $\mu=\lambda$.
\end{proof}
\begin{corollary}
If $g:X\to Z$ be a $\lambda$-perfect and $f:Z\to Y$ be a $\beta$-perfect
mapping then the composition $fg:X\to Y$ is $\gamma$-perfect for
$\gamma\leq\max\{\lambda , \beta\}$.
\end{corollary}
It is well-known that in every regular space we can separate compact sets from closed sets by the open sets. Next we have an extension of this fact.
\begin{theorem}
If $A$ is a $\beta$-compact subspace of a regular $P_{\lambda}$-space $X$, for $\beta\leq\lambda$, then for every closed set $B$ disjoint from $A$ there exist open sets $U , V\subseteq X$ such that $A\subseteq U$ , $B\subseteq V$ and $U\cap V=\emptyset$.\\ If, moreover, $d_c(B)=\alpha\leq\lambda$ then it suffices to assume that $X$ is a Hausdorff $P_{\lambda}$-space .
\end{theorem}
\begin{proof}
Since $X$ is regular, for every $x\in A$ there exist open sets$~U_x ,~ V_x\subseteq X$ such that $x\in U_x~~,~~B\subseteq V_x$ and \\$ U_x\cap V_x=\emptyset$. Clearly $A\subseteq\bigcup_{x\in A}U_x$, therefor there exists $J\subseteq A$ whit $|J|\leq \beta$ such that $A\subseteq\bigcup_{x\in J}U_x$. Obviously the sets $U=\bigcup_{x\in J}U_x$ and $V=\bigcap_{x\in J}V_x$ have the required properties. Now let we assume that $d_c(B)=\alpha\leq\lambda$ and $X$ is only a Hausdorff $P_{\lambda}$-space. For every $x\in B$ we consider the sets $A$ and $\{x\}$ and clearly similar to the first part we will obtain the disjoint open sets $U_x$ and $V_x$ such that $x\in V_x$ and $A\subseteq U_x$. Clearly $B\subseteq\bigcup_{x\in B}V_x$
\end{proof}
\begin{lemma}
If $X$ is a Hausdorff $P_{\lambda} $-space then a $\beta$-perfect mapping
$f:X\to Y$, whenever $\beta\leq\lambda$, can not be continuously extended over
any Hausdorff space $Z$ that contains $X$ as a dense proper subspace.
\end{lemma}
\begin{proof}
Without loss of generality we can suppose that $Z=X\cup\{x\}$. Now if $F:Z\to
Y$ is a continuous extension of $f$ then $d_c(f^{-1}(F(x)))\leq\beta$ and clearly $f^{-1}(F(x))$ dose not contain $x$. Therefor by the second part of
theorem 0.9. there exist open sets $U,V\subseteq Z$ such that $x\in U,f^{-1}
(F(x))\subseteq V$ , and $U\cap V=\emptyset$. Clearly the sets $(Z\setminus
V)\cap X=X\setminus V$ , $f(X\setminus V)$ and $F^{-1}(f(X\setminus V))$
are closed. Thus we have $$cl_z(X\setminus V)\subseteq F^{-1}(f(X\setminus
V))=f^{-1}(f(X\setminus V))\subseteq X.$$ Since $x\notin cl_zV$, we have
$cl_zV\subseteq X$ therefor $cl_zX=cl_z(X\setminus V)\cup cl_zV\subseteq X$.
Hence $X$ does not dense in $Z$.
\end{proof}
\begin{proposition}
If the composition $gf$ of continuous mappings $f:X\to Y$ and $g:Y\to Z$ is
$\lambda$-perfect then the mappings $g_f=g|f(X)$ and $f$ are $\beta$-perfect
and $\mu$-perfect respectively for $\mu , \beta\leq\lambda$.
\end{proposition}
\begin{proof}
Since for every point $z\in Z$ we have $d_{c}((gf)^{-1}(z))\leq\lambda$; it
follows from proposition () that the compactness degree of the set $g_f^{-1}(z)=f(X)\cap g^{-1}
(z)=f[(gf)^{-1}(z)]$ is less than or equal to $\lambda$. The fact that
$g_f$ is a closed mapping follows from [proposition 2.1.3 Eng] and thus the
mapping $g_f$ is $\mu$-perfect for $\mu\leq\lambda$.\\
We note that for every point $y\in Y$, $f^{-1}(y)$ is closed subset of $(gf)^{-1}(g(y))$ then $d_c(f^{-1}(y))\leq d_c(gf)^{-1}(g(y))\leq\lambda$. thus to conclude the proof it suffices to show that $f$ is closed. For this end we consider the arbitrary closed closed set $F\subseteq X$ and the mapping $h=(gf)|F: F\to Z.$ For every $z\in Z, h^{-1}(z)=(gf)^{-1}(z)\cap F$ is closed subset of $(gf)^{-1}(z)$ then $d_c(h^{-1}(z))\leq d_c((gf)^{-1}(z))\leq\lambda.$ this means that $h$ is $\alpha$-perfect for $\alpha\leq\lambda.$ so that, by the first part of our proof, the restriction $g|f(F)$ is $\theta$-perfect for $\theta\leq\alpha$. Clearly $g|f(F)$ can be continuously extended over $\overline{f(F)}.$ (note that the restriction of $g$ to $f(F)$ is the continuous extension of $g|f(F)$ over $\overline {f(F)}.$) It follows from the previous lemma that $f(F)=\overline {f(F)}$ and thus $f$ is a closed mapping.
\end{proof}
|
\section{Introduction}
Electromagnetically induced transparency (EIT) is a phenomenon whereby a medium
can be switched between states of transparency and opacity through the controllable
application of a weak probe field~\cite{Harris1997},
and applications of EIT include slow light~\cite{Hau1999, Harris99,Kash1999}, optical switching~\cite{Harris1998}, and optical quantum memory~\cite{Lvovsky2009}.
Double EIT extends the notion of EIT from controllable creation of transparency at a given frequency
to creation of two transparency windows at two different frequencies~\cite{Rebic2004}.
Double-double EIT (DDEIT) extends EIT even further to the case that each of a signal and a probe field
experience two transparency windows with the transparency window for the signal controlled by a coupling field and a probe field and the transparency window for the probe field controlled by the coupling field and the signal field~\cite{Hessa2013}.
DDEIT introduces the possibility of controlling propagation of and interaction between two bichromatic fields.
One advantage of controlling light is slowing it down.
Slow light is especially important for optical communication
and for quantum information processing.
For optical communication, slow light enhances light-matter interaction times and thereby
leads to an increase in nonlinear interactions~\cite{Harris99, Kash1999,Schmidt1996},.
In quantum computing, slow light enables storage of the quantum state of light for a sufficiently long time to enable quantum memory~\cite{Lvovsky2009}.
Our main objective is to propose a mechanism to slow the probe field in the second transparency window
of Doppler-broadened DDEIT.
For this slowing to be achieved,
we need to balance two competing requirements.
One requirement for slowing the probe pulse in
Doppler-broadened EIT is to reduce the driving field intensity
because the group velocity is proportional the driving-field intensity.
On the other hand,
the driving-field intensity must be sufficiently large to circumvent
inhomogeneous broadening~\cite{Taichenachev2000, Yuri2002, Javan2002, Ye2002, Figueroa2006}.
We derive an analytical expression to enable us to find a parameter regime where these competing requirements can simultaneously be satisfied.
Our analytical technique is based on approximating the Maxwell-Boltzmann velocity
distribution for atoms by Lorentzian distributions over the narrow but relevant domain of small atomic velocities~\cite{Javan2002}.
We find that the nonlinear interaction between the probe and signal field maintain the width of the second window constant for high Doppler width. This result permits us to lower the intensity of the signal field further without losing the EIT transparency window and get lower probe field group velocity at the second window compared to the first window.
We apply our scheme to the case of ${}^{87}$Rb
under realistic experimental conditions
and show that group-velocity reduction of the probe field is feasible.
We present our work in the following order.
In Sec.~\ref{sec:stationaryatom},
we introduce the optical density matrix element describing the optical properties of
the transition $\ket{1}\rightarrow\ket{4}$ and the resultant atomic susceptibility,
and we use these results to calculate the widths
of the Lorentzian-shaped transparency windows and the
corresponding group velocities for the probe field.
Doppler broadening due to temperature is incorporated into the expression for susceptibility in Sec.~\ref{sec:Doppler}. We solve this susceptibility numerically and analytically.
Our analytical solution is based on ignoring quadratic dependence of the probe-field Rabi
frequency and employing a Lorentzian approximation for a narrow
band around the Gaussian Maxwell-Boltzmann distribution. This
approximate expression enables an intuition about how to control group velocities' reduction at the second window.
\begin{figure}
\includegraphics [width=.7\columnwidth]{FTripodfigure}
\caption{Four-level tripod electronic structure with high-energy state~$|4\rangle$
and lower-energy levels~$\ket{3}$,~$\ket{2}$, and~$\ket{1}$ in order of
decreasing energy. Transitions are driven by probe~(p), coupling~(c) and
signal~(s) fields with frequencies $\omega_\text{x}$ and detunings $\delta_\text{x}$
with x$\in\{\text{p,c,s}\}$.
Dephasing rates are $\gamma_{\phi i}$ for $i\in\{2,3\}$.}
\label{fig:tripod}
\end{figure}
In Sec.~\ref{sec:group},
we present the procedure to reduce the group velocity in the second window. Finally, we summarize in Sec.~\ref{sec:summary}.
\section{Stationary-Atom Optical Susceptibility}
\label{sec:stationaryatom}
Consider the closed $\pitchfork$ atomic model scheme depicted in Fig.~\ref{fig:tripod}
with electronic level~$|4\rangle$ coupled to the lower levels~$\ket{1}$,~$\ket{2}$, and~$\ket{3}$ by three coherent fields,
namely the probe, coupling, and signal Rabi frequencies,
$\Omega_\text{p}$, $\Omega_\text{c}$,~$\Omega_\text{s}$, respectively. The $\ket{1}\leftrightarrow\ket{2}$, $\ket{1}\leftrightarrow\ket{3}$, and $\ket{2}\leftrightarrow\ket{3}$ transitions are dipole forbidden.
The three fields are detuned from the electronic transition frequency~$\omega_{\imath\jmath}$ between states $|\imath\rangle$ and $|\jmath\rangle$ by
\begin{align}
\label{eq:deltap}
\delta_\text{p}:=&\omega_{41}-\omega_\text{p}, \\
\label{eq:deltac}
\delta_\text{c}:=&\omega_{42}-\omega_\text{c},\\
\label{eq:deltas}
\delta_\text{s}:=&\omega_{43}-\omega_\text{p},
\end{align}
respectively.
The analytical steady-state density-matrix element ($\rho_{\imath\jmath}$) solution for a stationary atom,
to first order in the probe-field Rabi frequency,
can be approximated from the exact expression~\cite{Hessa2013} as
\begin{widetext}
\begin{equation}
\begin{split}
\rho^{(1)}_{14}\approx\text{i}\Omega_\text{p}\frac{\left(\rho_{11}-\rho_{44}\right)(\Gamma_{43}+2\text{i}\delta_\text{s}+\frac{\vert\Omega_\text{c}\vert^2}{\Gamma_{32}+2\text{i}\delta_\text{sc}})+\left(\rho_{44}-\rho_{33}\right)\frac{\vert\Omega_\text{s}\vert^2}{\gamma_3-2\text{i}\delta_\text{ps}}}{(\Gamma_{43}+2\text{i}\delta_\text{s}+\frac{\vert\Omega_\text{c}\vert^2}{\Gamma_{32}+2\text{i}\delta_\text{sc}})(\gamma_4-2\text{i}\delta_\text{p}+\frac{\vert\Omega_\text{c}\vert^2}{\gamma_2-2\text{i}\delta_\text{pc}}+\frac{\vert\Omega_\text{s}\vert^2}{\gamma_{3}-2\text{i}\delta_\text{ps}})},
\end{split}
\label{eq:rho143}
\end{equation}
\end{widetext}
where
\begin{equation}
\label{eq:deltaxy}
\delta_\text{xy}:=\delta_\text{x}-\delta_\text{y}
\end{equation}
is the two-photon detuning.
Details concerning the derivation of Eq.~(\ref{eq:rho143}) appear in Appendix~\ref{sec:stationary}.
We verified this expression numerically for weak signal and weaker probe Rabi frequencies, i.e.,
for the condition
\begin{equation}
\left|\Omega_\text{c}\right|^2\gg\left|\Omega_\text{s}\right|^2\gg|\Omega_\text{p}|^2.
\end{equation}
The decay rates in Eq.~(\ref{eq:rho143}) are
\begin{equation}
\gamma_\jmath:=\sum_{\imath<\jmath}(\gamma_{\jmath\imath}+\gamma_{\phi\jmath}),
\end{equation}
and the coherence decay rate is
\begin{equation}
\Gamma_\text{kl}=\gamma_\text{k}+\gamma_\text{l}.
\end{equation}
The dephasing rate between the forbidden transitions is not zero;
therefore, $\gamma_2=\gamma_{\phi 2}$ and $\gamma_3=\gamma_{\phi 3}$.
In our system, we impose the equal-population condition
\begin{equation}
\label{eq:equalpop}
\rho_{11}\approx\rho_{33}\approx0.5.
\end{equation}
Condition~(\ref{eq:equalpop}) makes the equations approximately solvable analytically
as the equations of motion for population~(\ref{eq:Droh14}) are effectively decoupled from the equations of motion for coherence~(\ref{eq:Droh44}).
Condition~(\ref{eq:equalpop}) is achieved by incoherent excitation from
ground state~$\ket{1}$ to the excited state~$\ket{4}$ with constant pumping rate. Thus, the diagonal matrix elements are held constant by conditions~(\ref{eq:equalpop}). See Appendix~\ref{subsec:population} for details
concerning the atomic population in our scheme.
The optical linear susceptibility for an atomic gas in three dimensions with~$\mathcal{N}$ the atomic density
and~$\bm{d}_{14}$ the dipole moment is
\begin{equation}
\label{eq:chi1p}
\chi^{(1)}_\text{p}
=\eta\frac{\rho^{(1)}_{14}}{\Omega_\text{p}},
\eta=\frac{\mathcal{N}\left|\bm{d}_{14}\right|^2}{\epsilon_0\hslash}.
\end{equation}
We can substitute Eq.~(\ref{eq:rho143})
into the numerator for~$\chi^{(1)}_\text{p}$ in Eq.~(\ref{eq:chi1p}),
which is complicated so we express~$\chi^{(1)}_\text{p}$ as
\begin{equation}
\chi^{(1)}_\text{p}=\frac{\text{i}\eta}{2(B_1+2\text{i}B_2)}\left(1-\frac{C_1+2\text{i}C_2}{A_1-2\text{i}A_2}\right)
\label{eq:suscep}
\end{equation}
with the terms~$A_{1,2}$, $B_{1,2}$, and~$C_{1,2}$ explained in the following.
To simplify Eq.~(\ref{eq:rho143}),
we fix the value $\rho_{44}=0$. This is always true because the atoms are trapped to the dark state leaving level~$|4\rangle$ unpopulated. The population of the other three levels depends on the Rabi frequency of the driving fields.
See Appendix~\ref{subsec:population} for more details of the dark-state analysis and state populations.
The variables in Eq.~(\ref{eq:suscep}) are
\begin{align}
A_1:=&\Gamma_{43}+\frac{\left|\Omega_\text{c}\right|^2\Gamma_{32}}{\Gamma_{32}^2+4\delta^2_\text{sc}},\;
A_2:=\frac{\left|\Omega_\text{c}\right|^2\delta_\text{sc}}{\Gamma_{32}^2+4\delta^2_\text{sc}}-\delta_\text{s},
\nonumber\\
B_1:=&\gamma_4+\frac{\left|\Omega_\text{c}\right|^2\gamma_2}{\gamma_2^2
+4\delta^2_\text{pc}}+\frac{\left|\Omega_\text{s}\right|^2\gamma_3}{\gamma^2_3+4\delta^2_\text{ps}},
\nonumber\\
B_2:=&\frac{\left|\Omega_\text{c}\right|^2\delta_\text{pc}}{\gamma_2^2
+4\delta^2_\text{pc}}+\frac{\left|\Omega_\text{s}\right|^2\delta_\text{ps}}
{\gamma^2_3+4\delta^2_\text{ps}}-\delta_\text{p},
\\
C_1:=&\frac{\left|\Omega_\text{s}\right|^2\gamma_3}{\gamma^2_3+4\delta^2_\text{ps}},
C_2:=\frac{\left|\Omega_\text{s}\right|^2\delta_\text{ps}}{\gamma^2_3+4\delta^2_\text{ps}},\nonumber
\end{align}
We now have expressions for the steady-state solution~(\ref{eq:rho143})
and the corresponding susceptibilities for the probe field~(\ref{eq:suscep}).
\begin{figure}
\centering
\subfloat[\label{ImChip}]{%
\includegraphics[width=0.4\textwidth]{FImrho.eps}
}
\hfill
\subfloat[\label{ReChip}]{%
\includegraphics[width=0.4\textwidth]{FRerho.eps}
}
\caption{%
(a)~Absorption and~(b)~dispersion as a function of the probe detuning $\delta_\text{p}$, with
numerical (dotted line line), analytical (solid line),
and approximate linear equation (dashed line line)
for $\gamma_4=18$ MHz, $\gamma_3=10$ kHz,
$\gamma_2=40$ kHz, $\Omega_\text{c}=\gamma_4$,
$\Omega_\text{s}=0.3\gamma_4$, $\Omega_\text{p}=0.05\gamma_4$, $\delta_\text{s}=9$MHz, and $\delta_\text{c}=0$.%
}
\label{fig:Re0k}
\end{figure}
Expression~(\ref{eq:suscep}) is used to calculate and plot the susceptibility,
whose imaginary part is shown in Fig.~\ref{fig:Re0k}(a), and whose real part is shown in Fig.~\ref{fig:Re0k}(b).
This absorption plot clearly displays the first probe window centered at~$\delta_\text{p}=\delta_\text{c}$
and the second EIT window centered at~$\delta_\text{p}=\delta_\text{s}\neq\delta_\text{c}$.
\subsection{Linewidth and Group Velocity}
\label{subsec:linewidthgroup}
The linewidth of each transparency window $\imath\in\{1,2\}$
is given by the half-width at half-maximum (HWHM)~$\daleth_\imath$.
The half-maximum values~$\varkappa_\imath$
are determined first by finding the maximum $h_{\text{max}_\imath}$
and the minimum $h_{\text{min}_\imath}$ values of window $\imath$
as shown in Appendix~\ref{subsec:stationaryatom},
and then calculating
\begin{equation}
\varkappa_\imath
=\frac{h_{\text{max}_\imath}
+h_{\text{min}_\imath}}{2}.
\end{equation}
By solving
\begin{equation}
\text{Im}[\chi^{(1)}_\text{p}] =\varkappa_\imath
\end{equation}
for $\delta_\text{pc}$ and $\delta_\text{ps}$ separately,
$\daleth_1$ and $\daleth_2$ are determined respectively,
with $\delta_\text{pc}=\daleth_1$ and $\delta_\text{ps}=\daleth_2$.
In Appendix~\ref{subsec:stationaryatom} we also discuss the requirements
\begin{equation}
\label{eq:hom1}
\left|\Omega_\text{c}\right|^2\gg\gamma_2\gamma_4
\end{equation}
and
\begin{equation}
\label{eq:hom2}
\left|\Omega_\text{s}\right|^2\gg\gamma_3\gamma_4
\end{equation}
to overcome homogeneous broadening~\cite{Fleischhauer2005} at the first and second windows respectively, and to observe the presence of the EIT windows.
If conditions~(\ref{eq:hom1}) and~(\ref{eq:hom2})
are satisfied,
the HWHMs of the first and second window are
\begin{equation}
\daleth_1=\frac{\left|\Omega_\text{c}\right|^2}{\gamma_4+\sqrt{4\left|\Omega_\text{c}\right|^2+\gamma^2_4}},\;
\daleth_2=\frac{\left|\Omega_\text{s}\right|^2}{2\sqrt{\gamma^2_4+\left|\Omega_\text{s}\right|^2}},
\label{eq:fw2}
\end{equation}
respectively.
Probe dispersion is shown in Fig.~\ref{fig:Re0k}(b). For detuning~$\delta_\text{p}$
chosen at the center of each window,
dispersion is zero or close to zero.
For each of windows~1 and~2,
group velocity is~\cite{Rebic2004}
\begin{equation}
\label{eq:vg}
v_\text{g}
\approx\frac{2c}{n_\text{g}},\;
n_\text{g}=(\omega_0-\delta_\text{p})
\left.\frac{\partial \text{Re}[\chi^{(1)}_\text{p}]}{\partial\delta_\text{p}}\right\vert_{\delta_\text{cen}}
\end{equation}
for~$n_\text{g}$ the group index,
~$\delta_\text{cen}$ the detuning at the center of each window (1 and~2),
and $\omega_0$ the transition frequency between levels~$\ket{1}$ and~$|4\rangle$.
Detuning~$\delta_\text{cen}$ equals~$\delta_\text{c}$ at the first window and equals~$\delta_\text{s}$ at the second window.
The partial derivative of the dispersion in the denominator is determined by
\begin{equation}
\left.\frac{\partial \text{Re}[\chi^{(1)}_\text{p}]}{\partial\delta_\text{p}}\right\vert_{\delta_\text{cen}}
=\lim_{\delta_\text{p}\to \delta_\text{cen}}
\frac{\text{Re}[\chi^{(1)}_\text{p}(\delta_\text{p})]
-\text{Re}[\chi^{(1)}_\text{p}(\delta_\text{p}=\delta_\text{cen})]}{\delta_\text{p}-\delta_\text{cen}}
\label{eq:slope}
\end{equation}
Therefore, the partial derivative of the dispersion
\begin{equation}
\left.\frac{\partial \text{Re}[\chi^{(1)}_\text{p}]}{\partial\delta_\text{p}}\right\vert_{\delta_\text{c}}=\frac{\eta\left|\Omega_\text{c}\right|^2}{\left(\gamma_2\gamma_4+\left|\Omega_\text{c}\right|^2\right)^2}
\label{l1}
\end{equation}
at the center of the first window and
\begin{equation}
\left.\frac{\partial \text{Re}[\chi^{(1)}_\text{p}]}{\partial\delta_\text{p}}\right\vert_{\delta_\text{s}}=\frac{\eta\left|\Omega_\text{s}\right|^2}{(\gamma_3\gamma_4+\left|\Omega_\text{s}\right|^2)^2}
\label{l2}
\end{equation}
at the center of the second window.
Equations.~(\ref{l1}) and~(\ref{l2}) yield the slope of the tangent line to points
$\delta_\text{p}=\delta_\text{cen}$ as shown in Fig.~\ref{fig:Re0k}(b).
In Fig.~\ref{fig:Re0k}(b),
the group velocity is shown to be approximately constant in each of the two EIT windows,
which can be seen by the straight-line tangents.
The group velocity scales inversely with slope
so the ratio of group velocities for each EIT window is the inverse of the
ratio of the slopes for each window.
From Eqs.~(\ref{l1}) and~(\ref{l2}) and from Fig.~\ref{fig:Re0k}(b),
the group velocity at the first window evidently exceeds the group velocity at the second window for the given parameters.\
Under conditions~(\ref{eq:hom1}) and~(\ref{eq:hom2}),
the group velocity reduces to
\begin{equation}
\label{eq:groupvelocity1}
v_{\text{g}_1}=\frac{2c}{\eta}\frac{\left|\Omega_\text{c}\right|^2}{\omega_{14}}
\end{equation}
at the first window
and to
\begin{equation}
\label{eq:groupvelocity2}
v_{\text{g}_2}=\frac{2c}{\eta}\frac{\left|\Omega_\text{s}\right|^2}{\omega_{34}}
\end{equation}
at the second window.
Hence, for stationary atoms, the group velocities in both windows are linearly proportional to the intensities of the respective driving fields.
\section{Doppler-broadened optical susceptibility}
\label{sec:Doppler}
At non-zero temperature atoms move randomly due to thermal energy.
Thermal atomic motion leads to a spreading of the absorbed frequency due to the Doppler effect, which broadens the optical line profile and is known as Doppler broadening.
In this section we solve susceptibility numerically and also derive approximate analytical expressions
as a function of temperature.
These results are used to find the widths of transparency windows and also group velocities of the
probe field in each of the two DDEIT windows.
Our approximate analytical technique is based on approximating the Maxwell-Boltzmann velocity
distribution for atoms by Lorentzian distributions over the narrow but relevant domain of small atomic velocities~\cite{Javan2002}.
This approximation is valid as large velocities are sufficiently detuned so as not to affect the optics.
In our scheme, the electromagnetic field passes through a gas of atoms at temperature~$T$.
Each atom of mass~$m$ has a velocity~$\bm v$ obeying the Gaussian Maxwell-Boltzmann distribution
\begin{equation}
\label{eq:fv}
f(v)=\frac{1}{u\sqrt{\pi}}\exp\left(-\frac{v^2}{u^2}\right),\;
u=\sqrt{\frac{2k T}{m}}
\end{equation}
with~$v$ the component of velocity~$\bm v$ in the direction of the three co-propagating
signal, probe, and coupling fields.
One effect of moving atoms is detuning of resonant frequencies due to the Doppler shift,
which results in a velocity-dependent probe-field susceptibility~$\chi_\text{p}(v)$.
For our Doppler-broadened system,
the susceptibility is thus averaged over the entire velocity distribution
according to~\cite{Julio1995}
\begin{equation}
\bar{\chi}_\text{p}:=\int^{\infty}_{-\infty}{\chi_\text{p}(v) f(v) \text{d}v}.
\label{eq:Dopp}
\end{equation}
The velocity-dependent expression for susceptibility is obtained from Eq.~(\ref{eq:suscep})
by the replacement
\begin{equation}
\delta_\text{x}\mapsto\delta_\text{x}+\frac{v\omega_\text{x}}{c},\;
\text{x}\in\{\text{p,c,s}\},
\end{equation}
for
\begin{equation}
\label{eq:omegax}
\omega_\text{x}
=\left\{
\begin{array}{lr}
\omega_{14}\equiv\omega_0,&\text{x}=\text{p},\\
\omega_{24},&\text{x}=\text{c},\\
\omega_{34},&\text{x}=\text{s},
\end{array}
\right.
\end{equation}
the atomic frequencies and~$c$ the speed of light $in \text{ }vacuo$.
Our scheme relies on neglecting Doppler effect on two-photon detuning $\delta_\text{xy}$~(\ref{eq:deltaxy}),
which is achieved for the co-propagating fields driving
approximately equal transition frequencies:
\begin{equation}
\omega_0\equiv\omega_{14}\approx\omega_{24}\approx\omega_{34}.
\end{equation}
This choice is commensurate with our case of a~$^{87}$Rb gas.
For this atom,
we assign~$\ket{1}$,~$\ket{2}$, and~$\ket{3}$ to the~$5S_{1/2}$ level
with~$F=1$,~$m_F=0$ and~$F=2$,~$m_F=\{-2,0\}$ respectively.
Level~$|4\rangle$ corresponds to the~$5P_{1/2}$ level with~$F=2$ and~$m_F=-1$.
Therefore, the quantities~$\{\delta_\text{xy}\}$ in Eqs.~(\ref{eq:suscep})
do not change under Doppler broadening.
Integration of Eq.~(\ref{eq:Dopp}) corresponds to a convolution of Lorentzian~$\chi_\text{p}$
with the Gaussian profile,
which is known as the Voigt profile~\cite{Pag08}. The Voigt profile can be solved numerically
but is hard to solve analytically~\cite{Julio1995,VA96,Ye2002,Li2007}.
The lack of an exact analytical solution inhibits finding a simple expression relating
the group velocity or width of each EIT window to Doppler broadening.
Instead, we approximate the Maxwell-Boltzmann distribution by a Lorentzian function over a narrow velocity domain~\cite{Javan2002} to obtain an approximate analytical expression for the optical susceptibility.
This approximation is valid insofar as we are interested in the optical response near the spectral center.
\subsection{Lorentzian line-shape approximation}
\label{sec:Lor}
In this subsection, we determine an analytical approximation to the optical susceptibility
for a Doppler-broadened system.
Our approximation uses a Lorentzian fit to the Maxwell-Boltzmann velocity distribution over a narrow range of velocity.
We use this approximation to show that the first probe transparency window
is independent of the signal-field Rabi frequency and the second transparency window is nonlinear in the signal-field Rabi frequency. Furthermore, we derive the connection between the transparency window and the Doppler broadening width, which is directly dependent on the temperature.
The Lorentzian line-shape function~\cite{Yuri2002}
\begin{equation}
L\left(\frac{v\omega_0}{c}\right)=\frac{1}{\sqrt{\pi}}\frac{W_\text{L}}{W^2_\text{L}+\left(\frac{v\omega_0}{c}\right)^2}
\label{Lore}
\end{equation}
is a function of the atomic velocity
with~$W_\text{L}$ is the HWHM of the Lorentzian profile.
To see that the Lorentzian~(\ref{Lore}) approximates the Gaussian~(\ref{eq:fv}) well over
a narrow domain, we first write both functions as Maclaurin series.
The Gaussian~(\ref{eq:fv}) is approximated by
\begin{align}
\label{eq:fvapprox}
f\left(\frac{v\omega_0}{c}\right)=&\frac{\sqrt{\ln{2}}}{\sqrt{\pi}W_\text{G}}-\frac{\omega^2_0(\sqrt{\ln{2}})^3}{c^2\sqrt{\pi}W^3_\text{G}}v^2\nonumber\\&
+\frac{\omega^4_0(\sqrt{\ln{2}})^5}{c^4\sqrt{\pi}W^4_\text{G}}v^5-\cdots
\end{align}
with
\begin{equation}
\label{eq:WG}
W_\text{G}:=\frac{\omega_0}{c}\sqrt{\frac{2k T \ln2}{m}}
\end{equation}
the HWHM of the Gaussian profile
and
\begin{align}
L\left(\frac{v\omega_0}{c}\right)=&\frac{1}{\sqrt{\pi}W_\text{L}}-\frac{\omega^2_0}{c^2\sqrt{\pi}W^3_\text{L}}v^2\nonumber\\&
+\frac{\omega^4_0}{c^4\sqrt{\pi}W^5_\text{L}}v^5-\cdots,
\label{expaLor}
\end{align}
for
\begin{equation}
-1<\frac{\omega_0v}{c W_\text{L}}<1.
\end{equation}
The two expansions~(\ref{eq:fvapprox}) and~(\ref{expaLor})
are approximately equal under the conditions that
\begin{equation}
\label{eq:WLWG}
W_\text{L}=\frac{1}{\sqrt{\ln2}}W_\text{G},
\end{equation}
for
\begin{equation}
-1\ll\frac{\omega_0v}{c W_\text{G}}\sqrt{\ln{2}}\ll1.
\end{equation}
Combining Eqs.~(\ref{eq:WG}) and~(\ref{eq:WLWG}) yields the connection between the Lorentzian linewidth and the temperature.
These conditions are satisfied near the center of both function profiles as shown in Fig.~\ref{LoGa}, where the higher-order terms of Eqs.~(\ref{eq:fvapprox}) and~(\ref{expaLor}) have insignificant influence.
\begin{figure}
\includegraphics[width=0.8\columnwidth]{FLorGau.eps}
\caption{
Plot of Lorentzian function (dashed line) and Gaussian function (dotted line) vs. normalized atomic velocity.}
\label{LoGa}
\end{figure}
Integration of Eq.~(\ref{eq:Dopp}) using $L(v)$ instead of~$f(v)$ has two terms evaluated with the contour integral using the residue theorem.
The final optical susceptibility,
including the Doppler broadening effect, is
\begin{equation}
\bar{\chi}_\text{p}(\delta_\text{p})
=I_1(\delta_\text{p})+I_2(\delta_\text{p})
\label{eq:susplorent}
\end{equation}
with~$\delta_\text{p}$ the detuning~(\ref{eq:deltap}).
The terms on the right-hand side of Eq.~(\ref{eq:susplorent}) are
\begin{equation}
\label{eq:I1}
I_1=\frac{\text{i}\eta}{2}\frac{\sqrt{\pi }}{B_1+2\text{i}B_2+W_\text{L}}
\end{equation}
and
\begin{align}
I_2
=&-\frac{\text{i}\eta}{2}\frac{\sqrt{\pi }}{B_1
+W_L+2\text{i}B_2}\frac{C_1+\text{i}C_2}{A_1-W_L-2\text{i}A_2}
\nonumber\\&
-\frac{C_1+\text{i}C_2}{A_1+B_1+2\text{i}(B_2-A_2)}
\nonumber\\&\times
\frac{\text{i}\eta W_\text{L}\sqrt{\pi }}{W^2_L+4A^2_2-A^2_1+4\text{i}A_1A_2}.
\label{eq:I2}
\end{align}
The HWHM~$\bar{\daleth}_1$ of the first transparency window,
and the group velocity for this window,
depend on~$I_1(\delta_\text{p})$
but not on~$I_2(\delta_\text{p})$ over the domain of~$\delta_\text{p}$ pertaining to the first window.
In the case of the second transparency window for the probe field,
both $I_1(\delta_\text{p})$ and $I_2(\delta_\text{p})$ are non negligible
for calculating the HWHM~$\bar{\daleth}_2$ and group velocity.
\begin{figure}
\centering
\subfloat[\label{subfig-absorption1}]{%
\includegraphics[width=0.5\columnwidth]{FImLG01kP.eps}
}
\subfloat[\label{subfig-dispersion1}]{%
\includegraphics[width=0.5\columnwidth]{FReLG01kP.eps}
}
\hfill
\subfloat[\label{subfig-absorption100}]{%
\includegraphics[width=0.5\columnwidth]{FImLG10kP.eps}
}
\subfloat[\label{subfig-dispersion100}]{%
\includegraphics[width=0.5\columnwidth]{FReLG10kP.eps}
}
\hfill
\subfloat[\label{subfig-absorption1000}]{%
\includegraphics[width=0.48\columnwidth]{FImLG100kP.eps}
}
\hfill
\subfloat[\label{subfig-dispersion1000}]{%
\includegraphics[width=0.48\columnwidth]{FReLG100kP.eps}
}
\caption{%
Plots of Im$[\chi^{(1)}_\text{p}]$ and Re$[\chi^{(1)}_\text{p}]$
vs probe detuning $\delta_\text{p}$ at different temperatures
for $\gamma_4=18 $MHz, $\gamma_3=10$ kHz,
$\gamma_2=40 $kHz, $\Omega_\text{c}=\gamma_4$,
$\Omega_\text{s}=0.35\gamma_4$, $\delta_s=9$MHz, and $\delta_c=0$. (a), (c) and (e) are Im$[\bar{\chi}_\text{p}]$ and~(b), (d), and (f) are Re$[\bar{\chi}_\text{p}]$. We set $T$= (1, 10, 100) K for (a),(b), (c),(d) and (e),(f) respectively, which is equivalent to $W_\text{L}=(34.8, 110, 348) $MHz, respectively. The dotted line-line corresponds to the analytical solution using the Lorentzian line-shape function, whereas the dashed line line is the numerical solution using the Maxwell-Boltzmann distribution function.}
\label{fig:LGd}
\end{figure}
In Fig.~\ref{fig:LGd}, we plot the imaginary and real terms of the susceptibility~$\chi_\text{p}^{(1)}$
as a function of the probe-field detuning~$\delta_\text{p}$
at various temperature values based on the average susceptibility~(\ref{eq:Dopp})
for the Maxwell-Boltzmann distribution function~$f(v)$ and for the approximation
using the Lorentzian function~$L(v)$.
At Low temperatures,
for which the broadening is low,
there is a discrepancy between the two functions.
At higher temperatures,
for which
\begin{equation}
\label{eq:ValAppro}
W^2_\text{L}\gg\gamma^2_4,
\end{equation}
the numerical data agree with the analytical data near the center as seen by comparing the two plots.
The plots differ at the tail,
which describes far-off-resonant atoms whose contribution is negligible.
This numerical result validates the Lorentzian approximation for condition~(\ref{eq:ValAppro}) near the center, which leads to a rather simple form of the inhomogeneously broadened susceptibility.
Analyzing the numerical result reveals that condition~(\ref{eq:hom1})
and condition
\begin{equation}
\label{eq:inhom1}
\left|\Omega_\text{c}\right|^2>\gamma_2W_\text{L}
\end{equation}
are required to observe the first transparency window.
These conditions~(\ref{eq:hom1}) and~(\ref{eq:inhom1}) eliminate the homogeneous broadening and reduce the effect of inhomogeneous broadening, respectively.
At a temperature for which the Doppler broadening satisfies condition~(\ref{eq:ValAppro}), and Eqs.~(\ref{eq:susplorent})-(\ref{eq:I2}) are a valid approximation,
satisfying condition~(\ref{eq:inhom1}) certainly implies satisfying condition~(\ref{eq:hom1}).
As shown in Fig.~\ref{fig:LGd},
the width~$\daleth_2$ of the second transparency window is not noticeably affected
by varying the Doppler width~$W_\text{L}$.
The reason for the robustness of~$\daleth_2$ is that the nonlinear interaction in~$I_2$,
but not in~$I_1$,
protects the second window from deleterious temperature effects.
Therefore, the strong-signal-field condition is not required to overcome Doppler broadening damaging the second transparency window.
In other words, condition
\begin{equation}
\label{eq:inhom2}
\left|\Omega_\text{s}\right|^2>\gamma_3 W_\text{L}
\end{equation}
is no longer mandatory to observe the second window.
Condition~(\ref{eq:hom2})
is still required to eliminate the homogeneous broadening for significant transparency at the second window; see Appendix~\ref{subsec:stationaryatom} for detailed mathematical proofs of the conditions
required to observe the transparency windows.
Furthermore, the relaxation of condition~(\ref{eq:inhom2}) leads to further reduction
of group velocity in Doppler-broadened media, which was limited by the Doppler width
appearing in the right-hand side of condition~(\ref{eq:inhom2}).
The two terms $\gamma_2W_\text{L}$ in Eq.~(\ref{eq:inhom1})
and $\gamma_3W_\text{L}$ in Eq.~(\ref{eq:inhom2})
quantify the inhomogeneous broadening of the two EIT windows.
In other words, the Doppler broadening alone is not the whole story;
rather, the products~$\gamma_{2,3}W_\text{L}$
incorporating the rates~$\gamma_2$ and~$\gamma_3$ are the key quantities.
In Sec.~\ref{subsec:width}
we derive the linewidth and the group velocity
for which the requisite conditions~(\ref{eq:hom1}) and~(\ref{eq:hom2})
for eliminating homogeneous broadening, are always satisfied for both windows.
\subsection{Width of the transparency windows}
\label{subsec:width}
The EIT width in a three-level Doppler-broadened~$\Lambda$ system can be maintained by keeping the temperature of the system constant while changing the driving field~\cite{Taichenachev2000, Yuri2002, Javan2002, Ye2002, Figueroa2006}.
Here, we follow a different approach by studying the dependence of the linewidth on temperature while fixing the intensity of the driving fields. The intensities of the driving fields are chosen such to eliminate the homogeneous broadening.\
The HWHM of the first window for the Doppler-broadened system is equal to
\begin{widetext}
\begin{equation}
\bar{\daleth}_1
=\frac{\left|\Omega_\text{c}\right|^2}{2}
\Bigg[\frac{(2\gamma_2W_\text{L}+\left|\Omega_\text{c}\right|^2)}{2(\gamma_4+W_\text{L})^2(\gamma_2W_\text{L}+\left|\Omega_\text{c}\right|^2)-W_\text{L}(W_\text{L}+2\gamma_4)(2\gamma_2W_\text{L}+\left|\Omega_\text{c}\right|^2)}\Bigg]^{1/2}.
\label{Gamma1}
\end{equation}
\end{widetext}
The width decreases nonlinearly as the Doppler width~$W_\text{L}$ increases
as shown in Fig.~\ref{fig:Wtemp}.
The condition $\left|\Omega_\text{c}\right|^2\gg\gamma_2W_\text{L}$ is valid for all~$W_\text{L}$ values in the figure.
For a high-intensity coupling field~(\ref{eq:inhom1}), the width of the first window reduces to
\begin{equation}
\bar{\daleth}_1=\frac{\left|\Omega_\text{c}\right|^2}{2\sqrt{W_\text{L}(2\gamma_4+W_\text{L})}}.
\end{equation}
The formula for HWHM can be further simplified if $W_\text{L}\gg\gamma_4$,
thereby yielding
\begin{equation}
\label{eq:width1}
\bar{\daleth}_1=\frac{\left|\Omega_\text{c}\right|^2}{2W_\text{L}}.
\end{equation}
This result is consistent with the previous result for a three-level~$\Lambda$ atom,
subject to a high-intensity driving field,
for which the linewidth is proportional to the intensity of the driving field and inversely proportional to the Doppler width~\cite{Javan2002}.
\begin{figure}
\setbox1=\hbox{3cm}{\includegraphics[width=0.95\columnwidth]{FWidthT.eps}}\llap{\raisebox{2.6cm}{\includegraphics[height=2.5cm]{W2nogain}}}
\caption{%
Numerical (dashed line) and analytical (dotted line) solutions of the HWHM ($\bar\daleth$)
for the (a)~first and~(b)~second EIT transparency windows vs Doppler width~$W_\text{L}$
for $\gamma_4=18$ MHz, $\gamma_3=10$ kHz,
$\gamma_2=40$ kHz, $\Omega_\text{c}=\gamma_4$,
$\Omega_\text{s}=0.35\gamma_4$, $\delta_s=9 $MHz, and $\delta_c=0$.
Inset:
Numerical (dashed line) and analytical (dotted line)
HWHM of the second EIT window evaluated for the gain term eliminated.%
}
\label{fig:Wtemp}
\end{figure}
The HWHM of the second window of the Doppler-broadened system
has a more complicated form than for the first window:
\begin{equation}
\begin{split}
\bar{\daleth}_2=\frac{\left|\Omega_\text{s}\right|^2}{2}\sqrt{\frac{(\gamma_4+W_\text{L})+W_\text{L}\left[\bar{\varkappa}_2(\gamma_4+W_\text{L})-\frac{1}{2}\right]}{4W_\text{L}\gamma^2_4\left[\frac{1}{2}-\bar{\varkappa}_2(\gamma_4+W_\text{L})\right]+\left|\Omega_\text{s}\right|^2(\gamma_4+W_\text{L})}}\\
\label{eq:Gamma2}
\end{split}
\end{equation}
where
\begin{align}
\bar{\varkappa}_2
=&\frac{\eta\sqrt{\pi}}{4}
\Bigg\{\frac{2\gamma_3W_\text{L}+\Omega^2_\text{s}}{(\gamma_4+W_L)(\gamma_3W_\text{L}+\Omega^2_\text{s})}
\nonumber\\&
+\frac{2\gamma_3 W_\text{L}(\gamma_4-W_\text{L})+\Omega^2_\text{s}(2\gamma_4-W_\text{L})}
{W_\text{L}\left[\gamma_3 W^2_\text{L}+(W_\text{L}-\gamma_4)\Omega^2_\text{s}\right]}
\Bigg\}
\label{vaarkappa2}
\end{align}
is the half-maximum value of~Im$\bar{\chi}_{\text p}$ of the second window.
The dependence of the HWHM of the second window on Doppler width is shown in Fig.~\ref{fig:Wtemp}. The width of the second window slightly decreases as the Doppler width increases.
For large Doppler broadening,
$W_L\gg\gamma_4$,
$\bar{\varkappa}_2$ depends on the population difference $\rho_{11}-\rho_{33}$.
As we set $\rho_{11}\approx\rho_{33}\approx0.5$, $\bar{\varkappa}_2$
is always located at Im$\bar{\chi}_\text{p}\approx0$,
i.\ e., where absorption vanishes.
Consequently, the width of the second window remains approximately constant
with respect to Doppler width
\begin{equation}
\bar{\daleth}_2=\frac{\left|\Omega_\text{s}\right|^2}{2\sqrt{2}\sqrt{\gamma^2_4+2\left|\Omega_\text{s}\right|^2}}.
\label{eq:width2}
\end{equation}
This independence Doppler broadening width response of the second window is
due to the gain described by Im$I_2$ of Eq.~(\ref{eq:susplorent}).
Expression~(\ref{eq:width2}) reveals that further reduction of the group velocity can be achieved by reducing the intensity of the signal field without losing the transparency window due to Doppler broadening.
The two EIT windows have the same width at the intercept point between the two curves
as shown in Fig.~\ref{fig:Wtemp}.
For all values of Doppler width the signal field has lower intensity than the coupling field. The inset to Fig.~\ref{fig:Wtemp} shows how the second window would behave as a function of
$W_\text{L}$ if the nonlinear contribution~$I_2$ were suppressed.
This inset makes clear how important the optical nonlinearity is for achieving quite different
temperature sensitivities of the two transparency windows for the probe field.
Mathematically, an effect of forcing $I_2\equiv 0$ is that the HWHM of the second transparency
window is given by a modification of the HWHM of the first window~(\ref{Gamma1})
with the proviso that~$\Omega_\text{c}$ is replaced by~$\Omega_\text{s}$
and~$\gamma_2$ is replaced by~$\gamma_3$.
For atoms copropagating with the probe field, the gain term suppresses the narrowing of the width results from Doppler broadening.
Generalizing the choice of atomic propagation direction relative to the direction of
the three driving fields
would of course lead to different results~\cite{VA96}.
In summary, Eq.~(\ref{eq:Gamma2}) is the full expression of the HWHM
of the second transparency window and accounts for the nonlinear interaction between the probe and the signal field.
Its behavior is depicted in Fig.~\ref{fig:Wtemp} and
shows the insensitivity of the second transparency window on
temperature,
which is represented by width~$W_\text{L}$.
Contrariwise, the first window is sensitive to~$W_\text{L}$.
\subsection{Group velocities at the transparency windows}
\label{sec:Gvel}
From Sec.~\ref{sec:Lor}, we have approximate analytical expressions for susceptibilities at the two transparency windows.
In this subsection we determine the derivative of the susceptibility with respect to the detuning~$\delta_\text{p}$ and use these partial derivatives of dispersion~(\ref{eq:slope})
to calculate the group velocities for the probe field in each of the two transparency windows.
The response of the partial derivative of dispersion with respect to Doppler broadening system
is shown in Fig.~\ref{fig:Dtemp}.
\begin{figure}
\centering
\setbox1=\hbox{5cm}{\includegraphics[width=1\columnwidth]{FDerWL.eps}}\llap{\raisebox{2.5cm}{\includegraphics[height=2.3cm]{FSDerNOGain}}}
\caption{%
Plots of the numerical (dashed line) and analytical (dotted) results for the
partial derivative of dispersion with respect to Doppler width for
(a)~the first window and~(b)~the second window
for $\gamma_4=18$ MHz, $\gamma_3=10$ kHz,
$\gamma_2=40$ kHz, $\Omega_\text{c}=1.5\gamma_4$,
$\Omega_\text{s}=0.5\gamma_4$, $\delta_s=13.5$ MHz, and $\delta_c=0$.\
Inset: numerical (dashed line) and analytical (dotted line) results
for the partial derivative of dispersion vs Doppler width at the second window for $I_2\equiv 0$.%
}
\label{fig:Dtemp}
\end{figure}
In this figure, numerical calculations show constant group velocity at the first window and a sharply increased group velocity at the second window.
The analytical expression for the group velocity of the Doppler-broadened system is evaluated using Eq.~(\ref{eq:vg}) but with the Doppler-broadened susceptibility~(\ref{eq:susplorent}) replaced the free Doppler-broadened susceptibility $\chi^{(1)}_\text{p}$~\cite{Kash1999}.
The partial derivative of Re$[\bar{\chi}_\text{p}]$ at the center of the first window is
\begin{equation}
\label{eq:DLF}
\left.\frac{\partial \text{Re}[\bar{\chi}_\text{p}]}{\partial\delta_\text{p}}\right|_{\delta_\text{c}}
=\frac{\eta\sqrt{\pi}|\Omega_\text{c}|^2}
{\left(\gamma_2W_\text{L}+{\left|\Omega_\text{c}\right|^2}\right)^2}
\end{equation}
and at the center of the second window is
\begin{align}
\left.\frac{\partial \text{Re}[\bar{\chi}_\text{p}]}{\partial\delta_\text{p}}\right|_{\delta_\text{s}}
=&\frac{2\eta\sqrt{\pi}\left|\Omega_\text{s}\right|^2\gamma_4}{\left(\gamma_4-W_\text{L}\right)
\left(\gamma_3W_\text{L}+\left|\Omega_\text{s}\right|^2\right)^2}
\nonumber\\&
+\frac{4\eta\sqrt{\pi}\left|\Omega_\text{s}\right|^2\gamma_4}{W_\text{L}\left|\Omega_\text{s}\right|^4}.
\label{eq:DLS2}
\end{align}
For the first transparency window and for a strong coupling field~(\ref{eq:inhom1}),
the group velocity of the probe field at the center of the first window
has the same group velocity as for the Doppler-free case~(\ref{eq:groupvelocity1}).
The negligibility of the Doppler broadening effect is due to the intensity of the coupling field being large, as can be explained from the analytical expression~(\ref{eq:DLF}).
\begin{figure}
\includegraphics[width=1\columnwidth]{FDerCoup}
\caption{%
Plot of the numerical (dotted) and analytical (dashed line) results for the partial derivative of dispersion
vs coupling field at the first window for Doppler width $W_\text{L}=409$ MHz,
$\gamma_4=18$ MHz, $\gamma_3=10$ kHz, $\gamma_2=40$ kHz.
}
\label{DerCoup}
\end{figure}
Figure~\ref{fig:Dtemp} shows
agreement between the analytical expression~(\ref{eq:DLF}) and the full numerical result
applicable for small~$W_\text{L}$.
This agreement diminishes slightly as~$W_\text{L}$ increases.
Therefore, the Lorentzian function can be used to study
the Doppler-broadened dispersion response of the~$\Lambda$ configuration comprising the three states~$\ket{1}$,~$\ket{2}$, and~$|4\rangle$
provided that condition~(\ref{eq:inhom1}) is satisfied.
Our analytical expression is reliable in practical parameter regimes.
This agreement between the analytical Lorentzian approximation
and the full numerical result under condition~(\ref{eq:inhom1})
is presented in Fig.~\ref{DerCoup} for varying coupling-field Rabi frequency.
We establish reliability of our approximation by comparing to an approximate
Lorentzian expression derived for a~$\Lambda$ EIT system~\cite{Kash1999}.
In our notation,
their result for group index is
\begin{equation}
\label{eq:kash}
n_\text{g}\propto\frac{\gamma_4\left|\Omega_\text{c}\right|^2}
{\left[\gamma_2\left(\gamma_4+W_\text{L}\right)
+\left|\Omega_\text{c}\right|^2\right]^2}
\end{equation}
with the relation between group index and derivative of dispersion (\ref{eq:DLF})
given by~Eq.~(\ref{eq:vg}).
We can neglect~$\gamma_4$ from~(\ref{eq:kash})
according to the approximation~(\ref{eq:hom1}).
Although result~(\ref{eq:kash})
is derived for a~$\Lambda$ system and our result~(\ref{eq:DLF}) for a~$\pitchfork$ system,
both results pertain to an EIT window in a strong-coupling regime,
and the two Lorentzian-based approximations agree.
At the second window, the analytical calculation fits the numerical solution for all chosen Doppler widths in the figure.
Eliminating~$I_2$,~(\ref{eq:I2}) leads to an equation for group velocity at the center of the second window
being similar equation to Eq.~(\ref{eq:DLF}) but with~$\Omega_\text{c}$ replaced by~$\Omega_\text{s}$ and~$\gamma_2$ replaced by~$\gamma_3$.
Similar dependence on Doppler width is shown in the inset of Fig.~\ref{fig:Dtemp}.
To achieve matched group velocity for the probe pulse propagating through the first and through the second window,
a non-zero nonlinearity is required.
The nonlinearity~$I_2$ is zero only if the condition $\rho_{44}=\rho_{33}=0$ is met.
This case for nonlinearity is depicted in the inset of Fig.~\ref{fig:Dtemp}. By fixing $\rho_{44}=\rho_{33}=0$ we have the unwanted additional effect of violating condition~(\ref{eq:hom2}) and thereby destroying the second window.
The intercept point between the two curves shown in Fig.~\ref{fig:Dtemp} reveals the operating temperature for group-velocity matching.
At temperatures exceeding the matched group-velocity case,
the group velocity in the first window is lower than the group velocity for the second window
and vice versa for temperatures lower than the condition for matched group velocity.
In summary, we demonstrate three important points in this subsection.
First, the Lorentzian approximation is a useful and valid approximation for studying the dispersion response of the probe field
as long as the conditions~(\ref{eq:inhom1}) and~(\ref{eq:inhom2}) for $I_2\equiv 0$
[Eq.~(\ref{eq:I2})] hold.
Second, the second term of Eq.~(\ref{eq:susplorent}) modifies the optical dispersion at the second window, which leads to a capacity for group velocity control through manipulating the temperature.
Finally, due to nonlinearity,
a signal-field intensity less than the coupling-field intensity does not necessarily imply
that the probe field has lower group velocity at the second EIT window than at the first window.
\subsection{Group-velocity reduction}
\label{sec:group}
In the previous Sec. \ref{subsec:width} and \ref{sec:Gvel} we have studied the behavior of the width and the group velocity for both EIT windows of the probe field in Doppler-broadening media. We have shown that a high-intensity coupling field is required to overcome inhomogeneous broadening, which represents an obstacle for group-velocity reduction.
The width of the second EIT window is independent of temperature,
which means that the enhanced group-velocity reduction is superior to the case that would hold
if the width did depend on temperature as temperature dependence could only worsen this effect.
In this section, we derive two expressions that relate the signal-field Rabi frequency~$\Omega_\text{s}$
to the coupling-field Rabi~$\Omega_\text{c}$ and Doppler width~$W_\text{L}$.
Satisfying the first expression guarantees that the probe field has the same group velocity in each transparency window.
Satisfying the second expression guarantees the same HWHM for the two EIT windows.
The relation between~$\Omega_\text{s}$ and~$\Omega_\text{c}$ can be satisfied for a wide
range of temperatures bounded above and below by the requirements for the analytical approximations to be valid according to Eqs.~(\ref{eq:ValAppro}) and~(\ref{eq:inhom1}).
We then use these two expressions to divide the signal-field intensity to three regimes:
a low-strength regime where the group velocity and EIT width are lower than the first window,
a high-strength regime where both group velocity and width of EIT window are greater than for the first window,
and a middle regime where the group velocity is lower and the width is higher than for the first window.
\begin{figure}
\includegraphics[width=1\columnwidth]{FVgreduction.eps}
\caption{%
Plot of the partial derivative of dispersion (dotted-dashed line) and HWHM (dashed line) for the second EIT window
and HWHM (upper horizontal dotted line line)
and partial derivative of dispersion (lower horizontal dotted line line) for the first EIT window
vs normalized signal-field Rabi frequency
with $\Omega_\text{c}=\gamma_4$, $W_\text{L}=700$ MHz,
$\gamma_4=18$ MHz, $\gamma_3=10$ kHz, and $\gamma_2=40$ kHz.%
}
\label{fig:Vreduction}
\end{figure}
In Fig.~\ref{fig:Vreduction},
we plot the HWHM and partial derivative of dispersion for the second EIT window
using Eqs.~(\ref{eq:width2}) and~(\ref{eq:DLS2}),
respectively.
We also plot the HWHM and partial derivative of dispersion for the first EIT window.
Intercepts between lines show which signal-field Rabi frequencies yield matched
HWHM or group-velocity conditions.
Matched HWHM occurs at $\Omega_{\text{s}_\text{l}}$
and matched group velocity occurs at $\Omega_{\text{s}_\text{h}}$
with $\Omega_{\text{s}_\text{l}}$ lower than $\Omega_{\text{s}_\text{h}}$,
and~l and~h refer to lower and higher values, respectively.
We can choose values of~$\Omega_\text{s}$ to control which of the two windows
has higher HWHM and group velocity.
We exploit our analytical expressions for the HWHM and the group velocity at the center of each window
to find the lower and higher boundary values of the signal field.
Equating Eqs.~(\ref{eq:width1}) and~(\ref{eq:width2}) for real values
of~$\Omega_\text{c}$ and~$\Omega_\text{s}$ gives us the lower boundary value of~$\Omega_\text{s}$:
\begin{equation}
\Omega_{\text{s}_\text{l}}=2^{\frac{3}{4}}\Omega_\text{c}\sqrt{\frac{\gamma_4}{W_\text{L}}}.
\label{eq:lowB}
\end{equation}
Similarly equating Eqs.~(\ref{eq:DLF}) and~(\ref{eq:DLS2})
gives us the higher boundary value of~$\Omega_\text{s}$:
\begin{equation}
\Omega_{\text{s}_\text{h}}=\frac{2}{3}\sqrt{\frac{9}{2}\gamma_3W_\text{L}+\frac{\gamma_4\Omega_\text{c}^2}{3W_\text{L}}\left(19+\frac{2\gamma_4\Omega_\text{c}^2}{W_\text{L}^2\gamma_3}\right)}.
\label{eq:higB}
\end{equation}
Equations~(\ref{eq:lowB}) and~(\ref{eq:higB})
reveal which signal-field strength should be selected to achieve either matched width or matched group velocity, respectively.
At certain Doppler width, the boundary values of~$\Omega_\text{s}$ in Eqs.~(\ref{eq:lowB}) and~(\ref{eq:higB}) can be tuned by varying the coupling-field strength~$\Omega_\text{c}$.
Both the matched group velocity and the matched HWHM have lower value
as~$\Omega_\text{c}$ is reduced.
In summary our four-level atom optical system can be operated at the second window
in three different regimes depending on the signal-field strength.
In the low-strength regime,
the second window has very low group velocity compared to the first window
but also has a lower EIT width. However, we can operate in this regime for lower group velocity as long as the width is resolvable experimentally.
Alternatively, in the high-strength regime,
the second window has a higher group velocity than for the first window.
which makes this high-strength regime less desirable for low group-velocity experiments.
\section{Conclusion}
\label{sec:summary}
We have achieved our objective of showing that the second DDEIT window
has advantages over first window with respect to obtaining an enhanced reduction of group velocity.
The presence of a nonlinear interaction between the probe and signal fields in optical susceptibility plays a crucial role in enabling temperature-controlled modification of the optical response.
At the second window, this term signifies the ability to reduce the narrowing of width and thereby yields increases of the group-velocity as the Doppler width increases.
The modified optical response due to nonlinear interaction permits observing the second window for low intense signal field and promises for more group velocities' reduction in the second EIT window.
By identifying the signal-field boundary values~$\Omega_{\text{s}_\text{l}}$ and $\Omega_{\text{s}_\text{h}}$,
we are able to identify the regime of the signal-field strength values
that could result in slower group velocity than for the first window.
The low-strength regime is the best for realizing low group velocity,
but the EIT window could be difficult to resolve.
The middle-strength regime is more robust in that the second EIT window is expected to be resolvable and the group velocity is expected to be low.
The high-strength regime is less interesting as the group velocity is relatively high.
Our approximate analytical calculation succeeds in describing the optical response
of the Doppler-broadened four-level optical system and helps in analyzing the system in the presence or absence of the nonlinear interaction.
These analytical calculations also provide us with intuition of how the width or group velocities
change in a Doppler-broadened system.
Importantly, our analytical expression helps us to study the relation between the coupling and signal fields and to achieve matching of either widths or the group velocities of the two windows.
These conditions are not intuitively clear otherwise, and hence
would be difficult to discern using only numerical calculations.
\acknowledgments
We acknowledge financial support fromAlberta Innovates -Technology Futures (AITF) and Natural Sciences and Engineering Research Council of Canada (NSERC).
|
\section{\label{app:dft_details}Supplementary Material A: \\
Details on \textit{ab-initio} electronic structure
calculations}
For the electronic structure calculations with SOC and on-site Coulomb
interaction, OPENMX code\cite{openmx,*openmx2}, which is based on the
linear-combination-of-pseudo-atomic-orbital basis formalism, was
used. A non-collinear DFT scheme and a fully relativistic
$j$-dependent pseudopotential were used to treat SOC, and
Perdew-Burke-Ernzerhof parametrization of the generalized
gradient approximation (GGA) was chosen for the exchange-correlation
functional\cite{PBE}, which was compared and found to be
almost identical with the results with the Perdew and Zunger local
density approximation functional\cite{CA,*PZ}. 400 Ry of energy
cutoff was used for the real-space sampling, and $9\times9\times9$
$k$-grid was adopted for the primitive unit cell. Electron
interactions are treated as on-site Coulomb interactions via a
simplified LDA+$U$ formalism implemented in OPENMX
code\cite{han2006n}, and up to 3.0~eV of
$U_{\rm eff} \equiv U-J_{\rm H}$ parameter ($J_{\rm H}$ is
Hund's coupling) was used for Ir $d$
orbital in our GGA+SOC+$U$ calculations. Maximally-localized Wannier
orbital method\cite{marzari1997maximally}, which is implemented in
OPENMX code\cite{weng2009revisiting}, were used to obtain the
tight-binding Hamiltonian for Ir $t_{\rm 2g}$ atoms.
\section{\label{app:wann}Supplementary Material B: \\
Ir $t_{\rm 2g}$ and $j_{\rm eff}$=1/2 Wannier orbitals}
In order to obtain the hopping integrals between the Ir $t_{\rm 2g}$ states,
the Wannier orbitals for the Ir $t_{\rm 2g}$ bands were calculated in the presence
of SOC. Fig. \ref{figA:Wannier} shows the results, where the three $t_{\rm 2g}$
Wannier orbitals at an Ir site with spin up component are shown. Other 21 orbitals
at the four Ir sites in the unit cell are obtained by translating them and applying
time-reversal operation $i\sigma_y \mathcal{K}$, where $\sigma_y$ and $\mathcal{K}$ are
the Pauli matrix acting on the spin sector and the complex conjugation, respectively.
The oxygen $p$-orbital components hybridized into the Ir
$t_{\rm 2g}$ bands are manifested as the oxygen $p$-orbital tails shown in the figure.
Size of oxygen hybridization is slightly enhanced as the value of $U_{\rm eff}$ is increased,
which contribute to the $t_2$ hopping term dominated by the oxygen-mediated channel.
The values of the effective SOC strength $\lambda_{t_{\rm 2g}}$ mentioned in the main text
were estimated from the on-site energies between the $t_{\rm 2g}$ Wannier orbitals.
The on-site energy matrix can be approximately expressed as
$H_{\rm on} \approx \lambda_{t_{\rm 2g}} \boldsymbol{l}_{t_{\rm 2g}} \cdot \boldsymbol{s} + \Delta_t$,
where $\Delta_t$ is the minor trigonal crystal field terms. The value of $\lambda_{t_{\rm 2g}}$
for each $U_{\rm eff}$ value
was obtained by taking the average of the matrix elements corresponding to the SOC term,
with their standard deviation smaller than 10\% of their average.
Fig. \ref{figA:Wannier}(b) shows the $j_{\rm eff}$=1/2-like Wannier orbital,
obtained from the low-energy window dominated by the $j_{\rm eff}$=1/2 character as
illustrated in Fig. 3(c) in the main text,
in a GGA+SOC+$U$ calculation with $U_{\rm eff}$ = 3 eV. Like the $t_{\rm 2g}$ orbitals,
other seven $j_{\rm eff}$=1/2-like Wannier orbitals in the unit cell can be obtained by
translation and time reversal operations.
Decomposing the Wannier orbital in terms of the local Ir $t_{\rm 2g}$ basis reveals
the dominant $j_{\rm eff}$=1/2 character at the center with
the $j_{\rm eff}$=3/2-dominated tails on the three next-neighboring Ir sites, as
schematically shown in Fig. 3(c) in the main text.
The tail components are gradually reduced as the $U_{\rm eff}$ value is increased,
so that the Wannier orbital becomes more localized in the presence of higher $U_{\rm eff}$.
\begin{figure*}[htb!]
\centering
\includegraphics[width=0.9 \textwidth]{FigA2.pdf}
\caption{(Color online) (a) Ir sites projected onto $xy$-plane
defined in terms of local cubic axes. ${\bf a}_1$, ${\bf a}_2$
and ${\bf a}_3$ are the Bravais lattice vector for the primitive cell,
where four sublattices within the primitive cell are labeled as Ir1 to Ir4.
Centers for $C^{a,b,z}_2$ rotations and inversion are depicted as
red square and black dots on the Z and X (X') bonds, respectively.
Note that $d$ becomes same with the Ir-O bond length in the absence of
trigonal distortion. Inset shows the three major hopping channels
between NN Ir sites.
(b,c) All NNN and (d) third NN neighbors
for an Ir site, Ir4, are depicted as colored arrows.
Note that different colors (blue and green) are used to distinguish
Ir sites and bonds belonging to one zigzag chain to another.
(a) shows NNN hopping paths that can be reached through intermediate NN
Ir sites. Intra- and inter-chain bonds are colored as cyan and purple,
respectively.
(b) shows NNN paths that cannot be reached through one NN bonds.
Yellow and grey arrows represent paths connecting same (Ir4 to Ir4)
and different sublattices (Ir4 to Ir1 in the figure), respectively.
(c) shows all third NN hopping paths. Purple and red arrows show bonds
connecting different and same sublattices, respectively.
Note that, bonds in (b) and red bonds in (c) does not have
any counterparts in the 2D honeycomb lattice.
}
\label{figA:NNNN}
\end{figure*}
\section{\label{app:hops}Supplementary Material C: \\
$t_{\rm 2g}$ hopping terms and tight-binding bands from Wannier orbitals}
Table \ref{tabA:hops} shows a partial list of Ir $t_{\rm 2g}$ hopping terms (up to third NN)
from the Wannier orbitals, where the convention for the coordinate system and the
illustration of NNN and third NN hopping terms are in Fig. \ref{figA:NNNN}.
Full list of hopping terms can be restored by applying the $C_2$ rotations and
inversion operations at the centers of Z and X (X') bonds, respectively.
Three $C_2$ rotations --- $C^{a,b,z}_2$ --- are allowed,
where $a \equiv \hat{x} + \hat{y}$ and $b \equiv \hat{y} - \hat{x}$.
Contrary to the relatively simple NN hopping channels as shown in the inset of
Fig. \ref{figA:NNNN}(a), a number of distinct NNN hopping
terms are introduced due to the three-dimensional twisting of the honeycomb
lattice\cite{Eric2014band}.
The NNN hopping channels can be classified into two kinds, depending on whether
they are analogous to the NNN hopping in the 2D honeycomb lattice or not.
Fig. \ref{figA:NNNN}(b) shows the 2D-like NNN hoping channels, which can be
reached through one intermediate NN Ir site. Depending on whether they belong to same zigzag
chain composed of only X (or X') bonds or connect different chains, they are divided into two different classes
$t^{\rm I}_{\rm NNN}$ and $t^{\rm II}_{\rm NNN}$. Hopping amplitudes in these channels are
larger than the other channels, $t^{\rm III}_{\rm NNN}$ and $t^{\rm IV}_{\rm NNN}$,
which cannot be reached through one Ir site as shown in Fig. \ref{figA:NNNN}(c).
There are also non-negligible third NN hopping terms, $t^{\rm I}_{\rm 3NN}$
and $t^{\rm II}_{\rm 3NN}$, which can be seen in Fig. \ref{figA:NNNN}(d) and
Table \ref{tabA:hops}. Like NNN hopping channels, third NN channels can be classified
depending on whether they have their 2D counterparts or not. $t^{\rm I}_{\rm 3NN}$
resembles the third NN hopping channel in the 2D honeycomb lattice, while
$t^{\rm II}_{\rm 3NN}$ is similar to the interlayer hopping terms in \aaio{$A$} series.
\begin{figure*}[htb!]
\centering
\includegraphics[width=0.7\textwidth]{FigA3.pdf}
\caption{(Color online)
Band structures (solid red lines) calculated from tight-binding calculations including up to
(a) NN, (b) NNN, and (c) third-NN hopping terms from the Wannier orbital calcualtions
tabulated in Table. \ref{tabA:hops}, compared to those from {\it ab-initio} results
(dashed grey lines). SOC is included in the calculations.
}
\label{figA:bands}
\end{figure*}
The role of further-neighbor hopping amplitudes in the band structure are shown in Fig. \ref{figA:bands},
where the evolution of band structure as we include NNN and third NN terms are presented.
Fig. \ref{figA:bands} shows the change of the band structures as further-neighbor
hopping terms are included with the presence of SOC; Fig. \ref{figA:bands}(a), (b) and (c)
show the bands with hopping terms up to NN, NNN, and third NN terms, respectively, with the
presence of SOC.
One can see that, the large SOC in Ir tends to make the $j_{\text{eff}}=1/2${} bands to be flatter in this
locally honeycomb-like lattice, and including NNN and third NN terms does not change the overall
behavior. Comparing Fig. \ref{figA:bands}(a) and Fig. \ref{figA:bands}(b),
one can notice that the dispersion inside the $j_{\text{eff}}=1/2${} subbands is affected by the NNN terms,
but the semi-metallic character is left unchanged. Inclusion of third NN terms,
as can be seen in Fig. \ref{figA:bands}(c), makes the dispersion slightly closer to the {\it ab-initio} bands.
\section{\label{app:exchanges}Supplementary Material D:\\
NN exchange interactions}
The exchanges $J$, $K$, and $\Gamma$ are given by (suppressing the
bond label $\alpha$)
\begin{align}
J\!&=\! \frac{4}{27}\left[
\frac{(2t_1+t_3)^2(4J_{H}+3U)}{U^2} - \frac{16J_{H}(t_1-t_3)^2}{(2U+3\lambda)^2}\right]\nonumber \\
K\!&=\! \frac{32J_H}{9}\!\!\left[
\frac{(t_1-t_3)^2\!-\!3t^2_2}{(2U+3\lambda)^2}\right]\!,\quad \!\!\!\!\!\! \Gamma= \frac{64J_H}{9}
\frac{t_2(t_1-t_3)}{(2U+3\lambda)^2},
\label{eq:exch}
\end{align}
where $t_i~(i=1,2,3)$, $J_H$, $U$, and $\lambda$ are the NN hopping
amplitudes, Hund's coupling, on-site Coulomb repulsion, and SOC
respectively\cite{rau2014generic}. $t_i$ is illustrated in Fig. \ref{figA:NNNN}(a).
Note that, the small amount of NN Heisenberg interaction is attributed to the cancelation between the
$2t_1$ and $t_3$ in the antiferromagnetic contrubution to $J$ in Eq. 1.
Since $t_2$ is the largest term, as mentioned in the main text, ferromagnetic $K$ becomes
the most dominant contribution in the exchange interactions.
\renewcommand*{\arraystretch}{1.2}
\begin{longtable*}{@{\extracolsep{\fill}}llllrrrrrrrrr}
\caption{A subset of Ir $t_{\rm 2g}$ hopping terms ${\bf T}_{ij}$ as representatives of each hopping channels up to third NN,
where $\mathcal{H}_{\rm hop} = \sum_{ij} {\bf C}^{\dag}_{i} \cdot {\bf T}_{ij} \cdot {\bf C}_j$
and ${\bf C}^{\dag}$ and ${\bf C}^{\dag}$ being the creation and annihilation operator for $t_{\rm 2g}$ states, respectively.
$d$ is approximate distance between Ir and O. Other hopping terms can be recovered by applying ${\bf T}_{ji} = {\bf T}^{\dag}_{ij}$,
$C^{a,b,z}_{2}$ rotations, and inversion operations.} \\
\hline\hline
Kind &${\bf r}_{ij}$ (in Cartesian coord.) & Sublattice & & \multicolumn{3}{r}{$U_{\rm eff}$ = 0.0~eV} & \multicolumn{3}{r}{$U_{\rm eff}$ = 1.5~eV} & \multicolumn{3}{r}{$U_{\rm eff}$ = 3.0~eV} \\ [5pt]
\hline
\endfirsthead
\multicolumn{13}{r}{{Continued from previous page...}} \\ \hline\hline
Kind &${\bf r}_{ij}$ (in Cartesian coord.) & Sublattice & & \multicolumn{3}{r}{$U_{\rm eff}$ = 0.0~eV} & \multicolumn{3}{r}{$U_{\rm eff}$ = 1.5~eV} & \multicolumn{3}{r}{$U_{\rm eff}$ = 3.0~eV} \\ [5pt] \hline
\endhead
\hline \multicolumn{13}{r}{{Continued in next page...}} \\ \hline
\endfoot
\hline\hline
\endlastfoot
$t_{\rm NN}$ &&&& ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
X,X'&(-$d$,~0,+$d$) & $1 \rightarrow 4$ & $d_{xy}$ & +0.088 & +0.018 & +0.260 & +0.080 & +0.019 & +0.276 & +0.064 & +0.021 & +0.289 \\
& &&$d_{xz}$& +0.018 & -0.152 & +0.013 & +0.020 & -0.110 & +0.013 & +0.021 & -0.051 & ~0.005 \\
& &&$d_{yz}$& +0.259 & +0.013 & +0.078 & +0.276 & +0.013 & +0.067 & +0.288 & ~0.003 & +0.052 \\[5pt]
$t_{\rm NN}$ &&&& ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
Z &(+$d$,+$d$,~0) & $1 \rightarrow 2$ & $d_{xy}$ & -0.162 & -0.022 & +0.021 & -0.119 & -0.024 & +0.023 & -0.059 & -0.031 & +0.030 \\
& &&$d_{xz}$& +0.016 & +0.087 & -0.239 & +0.017 & +0.078 & -0.255 & +0.025 & +0.072 & -0.269 \\
& &&$d_{yz}$& -0.016 & -0.239 & +0.086 & -0.017 & -0.254 & +0.077 & -0.024 & -0.271 & +0.056 \\[5pt]
$t^{\rm I}_{\rm NNN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(+$d$,+$2d$,-$d$) & $1 \rightarrow 3$ & $d_{xy}$ & ~0.002 & -0.012 & +0.039 & ~0.001 & -0.015 & +0.044 & ~0.001 & -0.018 & +0.047 \\
& &&$d_{xz}$& +0.013 & ~0.001 & +0.011 & +0.018 & ~0.001 & +0.014 & +0.024 & ~0.001 & +0.017 \\
& &&$d_{yz}$& +0.063 & ~0.004 & ~0.002 & +0.075 & ~0.007 & ~0.000 & +0.089 & -0.010 & ~0.001 \\[5pt]
$t^{\rm II}_{\rm NNN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(-$d$,+$d$,+$2d$) & $1 \rightarrow 1$ & $d_{xy}$ & ~0.002 & ~0.008 & -0.014 & ~0.003 & -0.011 & -0.017 & ~0.003 & -0.014 & -0.020 \\
& &&$d_{xz}$& +0.014 & ~0.001 & +0.039 & +0.017 & ~0.000 & +0.045 & +0.020 & ~0.002 & +0.050 \\
& &&$d_{yz}$& ~0.008 & +0.075 & ~0.001 & +0.011 & +0.089 & ~0.000 & +0.014 & +0.103 & ~0.001 \\[5pt]
$t^{\rm III}_{\rm NNN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(-$d$,+$d$,-$2d$) & $1 \rightarrow 1$ & $d_{xy}$ & ~0.001 & +0.011 & ~0.001 & ~0.000 & +0.013 & ~0.000 & ~0.002 & +0.015 & ~0.003 \\
& &&$d_{xz}$& ~0.001 & ~0.007 & +0.038 & ~0.000 & ~0.005 & +0.045 & ~0.003 & ~0.002 & +0.051 \\
& &&$d_{yz}$& -0.011 & +0.030 & ~0.008 & -0.013 & +0.036 & ~0.006 & -0.015 & +0.047 & ~0.004 \\[5pt]
$t^{\rm IV}_{\rm NNN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(+$d$,-$2d$,+$d$) & $1 \rightarrow 4$ & $d_{xy}$ & +0.012 & ~0.007 & -0.030 & ~0.009 & ~0.008 & -0.035 & ~0.005 & ~0.008 & -0.041 \\
& &&$d_{xz}$& ~0.007 & ~0.008 & ~0.006 & ~0.008 & ~0.009 & ~0.008 & ~0.008 & +0.011 & ~0.008 \\
& &&$d_{yz}$& -0.030 & ~0.007 & ~0.003 & -0.035 & ~0.008 & ~0.007 & -0.041 & ~0.008 & -0.012 \\[5pt]
$t^{\rm I}_{\rm 3NN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(~0,+$2d$,-$2d$) & $1 \rightarrow 2$ & $d_{xy}$ & ~0.007 & -0.013 & -0.014 & ~0.007 & -0.014 & -0.015 & ~0.008 & -0.015 & -0.017 \\
& &&$d_{xz}$& ~0.008 & ~0.006 & ~0.006 & ~0.009 & ~0.008 & ~0.007 & -0.011 & +0.011 & ~0.009 \\
& &&$d_{yz}$& -0.014 & -0.016 & -0.034 & -0.015 & -0.018 & -0.030 & -0.017 & -0.020 & -0.022 \\[5pt]
$t^{\rm II}_{\rm 3NN}$&& & & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ & ~~~~~~~~~~ $d_{xy}$ & $d_{xz}$ & $d_{yz}$ \\
&(-$d$,+$d$,~0) & $1 \rightarrow 1$ & $d_{xy}$ & -0.045 & ~0.006 & -0.012 & -0.046 & ~0.007 & -0.013 & -0.046 & -0.010 & -0.013 \\
& &&$d_{xz}$& +0.012 & ~0.008 & ~0.001 & +0.013 & ~0.009 & ~0.001 & +0.013 & +0.012 & ~0.002 \\
& &&$d_{yz}$& ~0.006 & -0.015 & ~0.008 & ~0.008 & -0.017 & ~0.009 & +0.010 & -0.018 & +0.012
\label{tabA:hops}
\end{longtable*}
|
\section{Introduction}
The first work on lens modelling
\citep{1981ApJ...244..723Y,1981ApJ...244..736Y} was developed after
the discovery of the first two gravitational lenses
\citep{1979Natur.279..381W,1980Natur.285..641W}, where a massive
galaxy causes a background quasar to appear as two or four
images. For the second lens to be discovered (PG1115+080), mass
models scored an early success with the prediction that one of the
lensed images seen would split further into two at higher resolution.
That galaxies must sometimes cause multiple images had long been
expected \citep{1937ApJ....86..217Z}, and it had even been argued that
the phenomenon could help measure cosmological parameters
\citep{1964MNRAS.128..307R,1966MNRAS.132..101R}, but apparently nobody
was expecting that lenses would need detailed modelling. The first
observations, however, immediately stimulated models. The reason for
that lies in the image separation. Recall that image separations are
of order of the angular Einstein radius
\begin{equation}
\Theta_{\rm E}
\sim \left(\frac{4GM}{c^2 D_L}\right)^{1/2}
\simeq 0.1'' \left(\frac{M}{M_\odot}\right)^{1/2}
\left(\frac{D_L}{\rm pc}\right)^{-1/2},
\end{equation}
where $D_L$ is the distance to the lens, and $M$ its mass. A lensing
galaxy with $M\sim10^{11}\rm\,M_\odot$ at $\sim1\rm\,Gpc$ would cause image
separations of $\sim1''$, which is comparable to the size of the
galaxy; typically the lensed images are seen through the galaxy halo.
Hence, the lensed images depend on the detailed mass distribution of
the lensing galaxy. Galaxy lenses therefore require models of their
mass distributions.
Since those early discoveries, more than 400 secure lenses are now
known. Modelling of the mass distribution is part of any research
using lenses, but so far no modelling study has spanned all known
lenses. The largest single one \citep{2009ApJ...703L..51K} models 58
separate lenses to infer the distribution of dark matter around
galaxies. In other work, \cite{2011ApJ...740...97L} combined lens
models of 21 galaxies with models of their stellar populations, to
find the relation between stars and dark matter, and
\cite{2014MNRAS.437..600S} modelled 18 time-delay lenses together to
infer cosmological parameters.
Imaging surveys now under way aim to increase the inventory of lenses
another ten or a hundred fold \citep[see
e.g.][]{Marshall2005,OguriMarshall2010}, with both automated and
visual search techniques proposed
\citep[e.g.][]{Marshall2009,More2012ApJ,Gavazzi2014}. For example, Space Warps\xspace
(Marshall et al, in prep; More et al, in prep) is a citizen science
project\footnote{\url{http://www.spacewarps.org}} in which volunteers
are presented sky-survey images and are invited to identify lens
candidates, by eye. Simulated lenses are mixed in with the data,
both to help train volunteers on what to look for, and to provide
measures of the effectiveness of the search. The motivation for Space Warps\xspace
is to enable volunteers, some of whom had previously serendipitously
identified lens candidates on earlier citizen-science surveys,
either to make discoveries missed in automatic searches by
software robots, or to perform the necessary inspection of an
automatically-generated sample, for quality control. Robots can be
built to be good at detecting lensing system in clean, uncrowded fields
with high signal-to-noise, but in more general test situations, robots
miss lenses (low completeness) or contaminate the results with
non-lenses (low purity) \citep{Marshall2009}.
The encouraging early results from the first Space Warps\xspace lens search, carried
out on the $\simeq172$ square degree Canada-France-Hawaii Telescope
Legacy Survey (CFHTLS) imaging by over 30,000 volunteers (Marshall et
al, in prep; More et al, in prep) prompted the question: could the
modelling of the lenses also be done by the volunteers? If so,
modelling could help prioritise lens candidates at an early stage,
which would be very useful with new wide-field and sensitive surveys,
which will yield thousands of lens candidates. There are several
software tools for lens modelling available, and work has been done on
generic interfaces \citep[e.g.][]{2014A&C.....5...28L}. Some early
designs for Space Warps\xspace included a prototype lens modelling tool
\citep{2010AAS...21543527N}. Moreover, some Space Warps\xspace volunteers are quite
experienced from earlier projects, having individually spent a
thousand hours or more with data, and are very interested in more
demanding projects. The interests of citizen-science communities are
just beginning to be studied \citep[e.g.,][]{2013AEdRv..12a0106J}, but
it is clear that some volunteers welcome open-ended challenges, and
sometimes these have led to new scientific results: one example is the
discovery of an exceptional extra-solar planet
\citep{2013ApJ...768..127S}; another is the development of new
algorithms for protein folding \citep{Khatib22112011}. All these are
grounds for optimism. There is, however, a basic difficulty in strong
gravitational lensing. Lensed images do not look much like their
source, and still less do they resemble the lensing-mass distribution.
To model a lens, one needs either to do a lot of random guessing, or
to have a good intuition for what works.
In this paper, we propose a way around the difficulty, and report on a
modelling test on Space Warps\xspace using simulated lenses. The three following
sections are devoted to the concept, the implementation, and tests
respectively.
In \secref{Fermat} we introduce a markup system for lensed images,
which we call a ``spaghetti diagram.'' A spaghetti diagram resembles
the visible image system, in a cartoon-like way, and at the same time
it encodes the basis of a mass model. This supplies an intuitive link
between the image system and the mass distribution, which look
frustratingly different from each other. Spaghetti diagrams are
essentially the saddle-point contours originally introduced to
gravitational lensing by \cite{1986ApJ...310..568B} as a way of
classifying lensed images. They are sometimes shown as part of the
output of lens models \citep[for
example][]{2001ApJ...557..594R,2003ApJ...590...39K,Lubini2012}. In
the present work, however, spaghetti diagrams are the {\em input\/}
through which the modeller tells the SpaghettiLens\xspace program what to do.
In \secref{SpaghettiLens} we describe the SpaghettiLens\xspace program, which
implements the above scheme. SpaghettiLens\xspace is an interface to and
extension of the GLASS framework for modelling lenses
\citep{2014arXiv1401.7990C}. We will not go into software details in
this paper, instead concentrate on lens modelling per se, but we
remark that SpaghettiLens\xspace is designed to be friendly to the forum style of
citizen-science projects, and enables incremental collaborative
model refinement by different people, without sacrificing any of the
technical features of GLASS.
In \secref{mod_challenge} we describe a modelling challenge where a
diverse sample of 29 simulated lenses was modelled multiple times
by a small number of Space Warps\xspace volunteers using SpaghettiLens\xspace. The models
were then examined in two ways. One was whether the spaghetti diagram
was correct. The other was the recovery of the Einstein radius of the
lens. In addition, we show some visual comparisons of the actual and
recovered lens shape and radial profile, and identify some areas to
improve. Profile and shape recovery with GLASS has been studied in
more detail in \citep{2014arXiv1401.7990C}.
\secref{todo} gives the general outlook and next steps.
\section{Fermat's principle and spaghetti diagrams} \label{sec:Fermat}
We first explain the lensing theory relevant to SpaghettiLens\xspace, following the
formulation of gravitational lensing in terms of Fermat's principle by
\cite{1986ApJ...310..568B}.
\subsection{Geometrical and gravitational time delays}
Consider a lens at some redshift $z_L$ and let $(x,y)$ be planar
coordinates at the lens, transverse to the line of sight. Let
$\Sigma(x,y)$ be the mass distribution. It is a mass per unit area,
i.e., density projected along the line of sight. The mass
distribution is often given in a dimensionless form
\begin{equation} \label{eq:kappa}
\kappa(x,y) \propto \Sigma(x,y)
\end{equation}
called the convergence. Let there be light, in the form of a more
distant source, at redshift $z_S$, behind point $(x_s,y_s)$ on the
lens.
We now imagine a virtual photon flying from the source to some $(x,y)$
on the lens, then changing direction and coming to the observer. Such
a direction change would increase the light travel time compared to
coming through $(x_s,y_s)$. The increased light travel time from the
geometry of deflection would be
\begin{equation} \label{eq:tgeom}
t_{\rm geom}(x,y) \propto (x-x_s)^2 + (y-y_s)^2 ,
\end{equation}
assuming the delay is small compared to the total light travel time.
An additional delay of the photon comes from travelling through the
curved spacetime at the lens. This gravitational time delay $t_{\rm grav}$
is related to the mass distribution of the lens. The relation is
generally written as a two-dimensional Poisson equation, but an
alternative expression, avoiding calculus, is as follows. The value
of $t_{\rm grav}$ through $(x,y)$ equals its average value on the
circumference of a small circle centred at $(x,y)$, plus a constant
times the mass within that circle. The constant is $2G/c^3$ times the
cosmological expansion factor $(1+z_L)$. Thus
\begin{equation} \label{eq:tgrav}
t_{\rm grav}(x,y) = \left\langle t_{\rm grav}(x_{\lower.33ex\hbox{$\circ$}},y_{\lower.33ex\hbox{$\circ$}}) \right\rangle
+ (1+z_L) \frac{2G}{c^3} M(x_{\lower.33ex\hbox{$\bullet$}},y_{\lower.33ex\hbox{$\bullet$}}) \,.
\end{equation}
We have used $(x_{\lower.33ex\hbox{$\circ$}},y_{\lower.33ex\hbox{$\circ$}})$ to denote the circumference of
a circle, and $(x_{\lower.33ex\hbox{$\bullet$}},y_{\lower.33ex\hbox{$\bullet$}})$ to indicate the integrated
mass within the circle. Appendix~\ref{more-theory} relates this
expression to the better-known explicit form for the gravitational
time delay.
The light travel time of a virtual photon is therefore longer by
\begin{equation} \label{eq:tarriv}
t(x,y) = t_{\rm geom} + t_{\rm grav}
\end{equation}
than it would have been with no lens present. Real photons take paths
that make $t(x,y)$ extremal, that is, having a minimum, maximum or
saddle point (Fermat's principle).
The proportionality factors in \eqref{eq:kappa} and \eqref{eq:tgeom}
depend on the redshifts and cosmological parameters, and are given in
Appendix~\ref{more-theory}.
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{fig/arriv_0}
\includegraphics[width=0.95\columnwidth]{fig/arriv_1}
\includegraphics[width=0.95\columnwidth]{fig/arriv_2}
\caption{Perspective views and contour maps of example arrival time
surfaces. Contours are coloured in rainbow order (red: least delay,
violet: highest delay). The special contours that self-cross at
saddle points are the basis of spaphetti diagrams. {\em Upper
panel:\/} No lens, hence showing the parabolic shape of the
geometrical time delay. The image would be at the bottom,
coinciding with the source. {\em Middle panel:\/} A circular lensing
mass (offset from the source) has been added, which has pushed the
minimum to one side and introduced a maximum and saddle point, each
corresponding to an image. The saddle-point is characterized by a
self-crossing ``spaghetti'' contour. {\em Lower panel:\/} An
elongated lensing mass has been added. There are now two minima,
two saddle points, and a maximum, each corresponding to an image.}
\label{fig:arriv}
\end{figure}
\subsection{Arrival-time contours} \label{sec:arriv}
The full function $t(x,y)$, also known as the arrival-time surface,
applies to virtual photons. In other words, it is an abstract
construct and not itself observable. But observable image positions
can be derived from the arrival-time surface, so visualising the
surface is useful. \figref{arriv} does so. In this figure, a
maximum, if present, is easy to see. To locate mimima and
saddle-points, however, one needs to examine the contours of equal
arrival time. A saddle point is characterised by a contour crossing
itself, forming an~X. Mimima, on the other hand, have contours
looping around them, as do maxima.
The saddle-point contours which form an X are especially interesting,
because they set the overall topography of the arrival-time surface.
They obviously give the locations of the saddle points, and roughly
localise the minima and maxima as well. If more precise locations of
the minima and maxima are added, the whole arrival-time surface is
already approximately known. Since the arrival-time surface has an
exact relation to the lens-mass distribution and the source position,
in effect the mass distribution is also automatically approximately
specified. In other words, a simple sketch of saddle-point contours
along with locations of minima and maxima ---which we call a
``spaghetti diagram''--- is implicitly already an approximation to a
lens-mass distribution.
The preceding assumes a point source. To get an idea of what an
extended source would do, let us imagine moving the original source
slightly. The contours of constant arrival time will naturally move
slightly, and so will the images. The movement of the contours will
be most noticeable where the contours are far apart, that is where the
arrival-time surface is nearly flat. As is evident from
\figref{arriv}, this is the region where the minimum and saddle points
lie, or near the images. In this region, points on the source that
are close together produce images that are comparatively far apart.
In other words, the image is highly magnified. In summary, lower
curvature in the arrival-time surface for a point source implies
larger magnification of an extended source. Conversely, where the
arrival-time surface is strongly curved, the image will be
demagnified. We see from \figref{arriv} that the arrival-time surface
tends to be highly curved near the maximum. Hence maximum tend to be
demagnified. In practice, maxima of the arrival time are nearly
always too faint to see. The minima and saddle points dominate.
\section{A lens-modelling program}
\label{sec:SpaghettiLens}
SpaghettiLens\xspace is a mass modelling program that makes use of the Space Warps\xspace
infrastructure, in particular, the image database and the discussion
forum.\footnote{\url{ http://talk.spacewarps.org}} The forum is
essential for establishing contact between interested members of the
Space Warps\xspace community and the project science team, and then for enabling
collaboration between them. We were able to collaborate together on
modelling objects from Space Warps\xspace in the usual style of medium-sized
astronomical collaborations, with video-conferencing and in-person
meetings where possible. Preliminary results were immediately
summarized on modelling threads on the forum, and anyone interested
was made welcome to join at any time.
Modelling with SpaghettiLens\xspace involves three stages, (1)~markup of the image,
followed by (2)~intensive numerical computation carried out on a
server in the background, followed by (3)~review of diagnostics and
possible discussion. Human interaction is essential to the first and
third stages, while stage~2 is completely automated. We now describe
the three stages.
\subsection{Image markup}
One begins by going to the SpaghettiLens\xspace web application\footnote{\url{
http://mite.physik.uzh.ch}} and entering the number of a
Space Warps\xspace image tile. SpaghettiLens\xspace then presents the image, along with zoom and pan
options and a markup tool to construct a spaghetti diagram. The human
modeller now has to make an educated guess for the topography of the
arrival-time surface, and input the locations time-ordering of the
maxima, minima, and saddle-points. The markup tool \citep[which is
inspired by Figure~6 of][and is like that figure made interactive and
overlaid on data]{1986ApJ...310..568B} lets the modeller enter the
information by sketching saddle-point contours. Examples can be seen in Figure~\ref{fig:screenshot} and the upper-left
panels of Figures \ref{fig:6941} to \ref{fig:7022}. The loops in the
markup tool were the origin of the ``spaghetti'' metaphor.
\begin{figure}
\centering
\includegraphics[width=0.90\linewidth]{fig/screenshot}
\caption{Screen grab of SpaghettiLens\xspace in action. A Space Warps\xspace image has been loaded
in, re-centred and zoomed. Five images and the associated
``spaghetti'' contours have then been suggested, using the marking
tools associated with the buttons along the top of the panel. The
mass model is generated server-side when the right-most button is
pressed.}
\label{fig:screenshot}
\end{figure}
The markup tool allows only valid lensing configurations to be
entered. The user does not need to think explicitly about the image
parities (though the markup tool provides this information using
colour codes) or about time-ordering, or worry about the odd-image
theorem. The exact placement of the loops in a spaghetti diagram has
no significance. Only the hierarchy of which loop is inside which is
relevant. The loops are there simply to help modeller's intuition.
As implemented so far, SpaghettiLens\xspace assumes that the lens is dominated by a
single galaxy. Accordingly, only one maximum in the arrival-time surface
is permitted, and it is taken to be the centre of the main lensing
galaxy. The user can, however, mark additional minor galaxies: these
are modelled as point masses, the mass being fitted by the program along
with the rest of the mass distribution.
\subsection{Numerics}
Having sketched a spaghetti diagram, the user presses a button to
initiate the next stage. SpaghettiLens\xspace then translates the spaghetti diagram
into input for GLASS, and forwards this input. The task of GLASS,
which runs server-side as it is compute-intensive, is to find a mass
distribution $\kappa(x,y)$ that exactly reproduces the given locations
of the maximum, minima and saddle points. This criterion by itself is
extremely under-determined --- there are infinitely many mass
distributions that will reproduce a given set of maxima, minima and
saddle points, but typically they (a)~produce lots of extra images,
and (b)~look very unlike galaxies. Additional assumptions (a prior)
are necessary. GLASS uses the following priors
\citep[cf.][]{1997MNRAS.292..148S,2008ApJ...679...17C}.
\begin{enumerate}
\item The mass distribution is built out of non-negative tiles of
mass. (Sometimes these tiles are called mass pixels, but we should
emphasize that they are unrelated to image pixels, and are much
larger.)
\item There is a notional lens centre, say $(x_0,y_0)$ which is
identified with the maximum of the arrival time. The source can
have an arbitrary offset with respect to the lens centre.
\item The mass distribution must be centrally concentrated, in two
respects. First, the circularly averaged density must fall away
like $$ \left[(x-x_0)^2+(y-y_0)^2\right]^{-1/2}$$ or more steeply.
Second, the direction of increasing density at any $(x,y)$ can point
at most $45^\circ$ away from $(x_0,y_0)$.
\item The lens must be symmetrical with respect to $180^\circ$ rotations
about $(x_0,y_0)$. This symmetry assumption can be relaxed if the
user wishes.
\end{enumerate}
There are still infinitely many models that satisfy both data and
prior constraints, but now they are more credible as galaxy lenses.
It is then possible to generate an ensemble of models. The sampling
technique used by GLASS is described in \citep{Lubini2012}.
Typically, ensembles of 200 models are used. That is to say, what we
call a SpaghettiLens\xspace model is really the mean of an ensemble of 200 models, and
its estimated uncertainty is the range covered by the whole ensemble.
\subsection{Diagnostics}
\label{sec:diag}
After the model ensemble has been generated, SpaghettiLens\xspace post-processes it to
present results and diagnostics to the user for inspection. This takes
the form of three figures.
\begin{enumerate}
\item A synthetic image of the lensed features.
\item A contour map of the arrival-time surface $t(x,y)$.
\item A gray scale plus contour map of the mass distribution.
\end{enumerate}
The synthetic image generated by SpaghettiLens\xspace assumes a simple
circularly symmetric source with linearly decreasing
surface brightness profile. The user can change the contrast level
on the image, which (though it is not saved) amounts to adjusting the
size of the source. These synthetic images are still very
crude, and not always useful for assessing models. The best
indicator, in practice, of whether the modelling was successful is
contour map of $t(x,y)$, with saddle-point contours highlighted. It
is, in effect, the computer's refinement of the spaghetti diagram
input by the user. If the arrival-time surface looks qualitatively
similar to the spaghetti diagram, that generally indicates a
successful model. The mass distribution also provides indications;
successful models generally lead to smooth-looking mass distribution,
whereas an irregular or checkboard pattern in the mass map signals a
bad model.
After examining this feedback, the user can choose to save the
model to the SpaghettiLens\xspace archive, at which point it is assigned an unique
URL. They can also modify the input and try again, or discard the
attempt altogether. After archiving, there can be discussion among
modellers, through the Space Warps\xspace forum or by any other means, and revision
of the model. This is achieved simply by sharing the model's
URL; following its hyperlink takes one to the SpaghettiLens\xspace app, pre-loaded
with the correct data image and input spaghetti. Any archived model
can be revised by any user: they can modify the spaghetti
configuration slightly or drastically, or change options like the size
of the mass tiles. Particularly interesting lens candidates lead to
trees of models in this way. Discussion among modellers tends to
prune a model tree, focusing attention on the most interesting
models.\footnote{See ``Collaborative gravitational lens
modelling\dots'' in {\tt http://letters.zooniverse.org} for an example.}
\section{A lens modelling challenge} \label{sec:mod_challenge}
We now describe a test of the lens-modelling system, under conditions
that mimic as closely as possible the modelling of real lens
discoveries. The lenses to be modelled were the simulated lenses
(known as ``sims'') already sprinkled onto the Space Warps\xspace field. Once a
small user base had grown around SpaghettiLens\xspace, a modelling challenge was
announced through the Space Warps\xspace forum. The challenge set consisted of 29
sims, chosen to represent the different visual morphologies of Space Warps\xspace
sims. Modellers then contributed 119 models for these sims (at least
two for each sim). Models were reported on the same forum used to
model real candidate lenses. Modellers were free to consult and
refine each other's models, but had no information on how the sims
were generated.
Once the modelling was complete, the models were compared with the
originals. There were two main tests: a check of whether the
spaghetti diagrams were correct for the lens in question, and a
comparison of the effective Einstein radii of the sims and the models.
\subsection{The simulated lenses} \label{sec:sims}
The Space Warps\xspace sims are described in detail in More et al (in prep), but
relevant here is that the sims were of three kinds, as follows.
\begin{enumerate}
\item Lensed quasars: The lens is modelled as a singular isothermal
ellipsoid (SIE) and a constant external shear whereas the quasar is
represented with a circular Gaussian source whose size is given by
the point spread function (PSF) in each imaging band.
\item Galaxy-scale lenses: The lens model is the same as above whereas
the background galaxy is modelled as an elliptical de~Vaucouleurs.
\item Group-scale lenses: The lens model includes SIE models for the
central galaxy and the inner group members, plus a circular NFW
\citep{1996ApJ...462..563N,1997ApJ...490..493N} to represent the
underlying dark matter distribution and the background galaxy model
stays the same as galaxy-scale lenses.
\end{enumerate}
\FloatBarrier
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006941_input}
\includegraphics[width=\myplotswidth]{fig/006941_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW000102p_006941_arriv}
\includegraphics[width=\myplotswidth]{fig/006941_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW000102p_006941_kappa}
\includegraphics[width=\myplotswidth]{fig/006941_mass}
\caption[result 6941 (ASW000102p)]{A simulated lens that mimics a
lensed quasar, and model results. The left panels derive from the
simulation, and the right panels are SpaghettiLens\xspace output. Details of
individual panels are in \secref{example_models}.}
\label{fig:6941}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006990_input}
\includegraphics[width=\myplotswidth]{fig/006990_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0004oux_006990_arriv}
\includegraphics[width=\myplotswidth]{fig/006990_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0004oux_006990_kappa}
\includegraphics[width=\myplotswidth]{fig/006990_mass}
\caption[result 6990 (ASW0004oux)]{Results from a system with an arc
plus a counter-image, typical of lensed galaxies. (See Section
\ref{sec:example_models} for details.)}
\label{fig:6990}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006919_input}
\includegraphics[width=\myplotswidth]{fig/006919_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0002z6f_006919_arriv}
\includegraphics[width=\myplotswidth]{fig/006919_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0002z6f_006919_kappa}
\includegraphics[width=\myplotswidth]{fig/006919_mass}
\caption[result 6919 (ASW0002z6f)]{Another configuration of arc plus
counter-image: an arc and counter image where the arc is closer to
the lensing galaxy than the counter image. (See
\secref{example_models} for details.)
\vspace{1em}}
\label{fig:6919}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006915_input}
\includegraphics[width=\myplotswidth]{fig/006915_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0001hpf_006915_arriv}
\includegraphics[width=\myplotswidth]{fig/006915_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0001hpf_006915_kappa}
\includegraphics[width=\myplotswidth]{fig/006915_mass}
\caption[result 6915 (ASW0001hpf)]{A four-image configuration
typical of lensed quasars. (See \secref{example_models} for
details.)}
\label{fig:6915}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006975_input}
\includegraphics[width=\myplotswidth]{fig/006975_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW000195x_006975_arriv}
\includegraphics[width=\myplotswidth]{fig/006975_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW000195x_006975_kappa}
\includegraphics[width=\myplotswidth]{fig/006975_mass}
\caption[result 6975 (ASW000195x)]{A lens with unrecovered mass
substructure. (See \secref{example_models} for details.)
\vspace{1em}
}
\label{fig:6975}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/006937_input}
\includegraphics[width=\myplotswidth]{fig/006937_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000vqg_006937_arriv}
\includegraphics[width=\myplotswidth]{fig/006937_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000vqg_006937_kappa}
\includegraphics[width=\myplotswidth]{fig/006937_mass}
\caption[result 6937 (ASW0000vqg)]{A sim with unrecovered
substructure, resulting in a poor mass model. (See Section
\ref{sec:example_models} for details.)}
\label{fig:6937}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/007025_input}
\includegraphics[width=\myplotswidth]{fig/007025_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000h2m_007025_arriv}
\includegraphics[width=\myplotswidth]{fig/007025_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000h2m_007025_kappa}
\includegraphics[width=\myplotswidth]{fig/007025_mass}
\caption[result 7025 (ASW0000h2m)]{A four-image system with image
parities incorrectly identified. The model is poor, but the
estimated Einstein radius is not bad. (See Section
\ref{sec:example_models} for details.)}
\label{fig:7025}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\myplotswidth]{fig/007022_input}
\includegraphics[width=\myplotswidth]{fig/007022_arr_time_ipol} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000h2m_007022_arriv}
\includegraphics[width=\myplotswidth]{fig/007022_spaghetti} \\
\includegraphics[width=\myplotswidth]{fig/ASW0000h2m_007022_kappa}
\includegraphics[width=\myplotswidth]{fig/007022_mass}
\caption[result 7022 (ASW0000h2m)]{The same system as in Figure
\ref{fig:7025}, this time with image parities correctly
identified. (See \secref{example_models} for details.)}
\label{fig:7022}
\end{figure}
\FloatBarrier
The {\tt gravlens} program \citep{2001astro.ph..2340K} was used.
Formulas for the lenses appear in \cite{2001astro.ph..2341K}. The SIE
lenses follow equations (33--35) of that work, with core radius set to
zero. The NFW lens is in equations (48) and (50), while shear is the
$\gamma$ term in equation (76).
The information in this section was not revealed to the main developer
of SpaghettiLens\xspace (RK, who also chose the challenge set) or to the modellers
(EB, CC, CM, JO, PS and JW) while modelling was in progress. That is,
the modellers had no advance knowledge of what kind of
parameterisation had been used to make the sims. After the modelling
stage, AM released the details of the sims for post-modelling
analysis. Results from the latter now follow.
\subsection{Some example models} \label{sec:example_models}
Of the 119 models proposed, we now discuss eight examples in some
detail. Results from these are shown in Figures
\ref{fig:6941}--\ref{fig:7022}. The first four of these show the most
common image morphologies, the other four explain some problem cases.
Each of Figures \ref{fig:6941}--\ref{fig:7022} figures has the
following layout.
$$ \begin{matrix}
\hbox{marked-up CFHTLS image} \qquad &\hbox{model synthetic image} \\
t(x,y) &\hbox{model } t(x,y) \\
\kappa(x,y) &\hbox{model } \kappa(x,y)
\end{matrix} $$
The model synthetic image presented in this paper is not the original one,
but an interpolated version generated by an updated version of SpaghettiLens\xspace. The image
presented during the original experiment was of lower resolution. The two plots in
the middle showing $t(x,y)$ have uniform, but arbitrary spaced contour lines.
The $\kappa(x,y)$ plots in the bottom row show solid lines for $\kappa>1$
and dashed lines for $\kappa<1$. The spacing is logarithmic, with 10 contours
for every decade --- that means contour spacing is a quarter magnitude in
optical terms.
Let us now consider these cases in turn.
\begin{itemize}
\item Figure \ref{fig:6941} shows the simplest case, with two clear
images produced by a nearly-circular lens. The center of the
lensing galaxy is a maximum, the image nearer to the galaxy is a
saddle point, and the image further away is a minimum. All these
were correctly identified. As note above, in
\secref{SpaghettiLens}, the precise shape of the loops in the
spaghetti diagrams is unimportant, only the implied image locations,
parities and time-ordering matters. $\kappa(x,y)$ shows, that
the model has a more shallow mass distribution than the simulation.
This is a persistent issue throughout all models and is discussed in
\secref{tests.t2}.
\item Figure \ref{fig:6990} shows an example of an arc that has split
into three images. This kind of configuration, with a counter-image
close to the lensing galaxy and a more distant arc/triplet on the
other side, generically arises from an elongated mass distribution
when the source is displaced along the elongated direction. The
spaghetti diagram in this case has another markup element, a grey
point and circle overlaid on a probable secondary lensing galaxy.
This is an instruction to SpaghettiLens\xspace to allow a point mass at that
location, distinct from the main mass map.
\item Figure \ref{fig:6919} shows another example of an arc plus
counter-image, but (in contrast to \figref{6990}) the arc is closer
to the lens than the counter-image. This configuration arises if the
source displacement is perpendicular to the long axis of the lensing
mass. Comparing the two panels in the middle row, we see that the
modeller interpreted arc as consisting of three images, whereas the
sim shows a single saddle point associated with the arc. But the
identification is not really erroneous --- we just need to take into
account that the source is extended. In fact, in the sim the
brightest part of the source is only doubly imaged, but the source
extends into a region that produces four images. In the $t(x,y)$ of
the sim, the hairpin-bending contours are typical of double on the
verge of splitting into a quad.
\item Figure \ref{fig:6915} shows another quad. This kind of
configuration arises when the mass is elongated and the source is
displaced at an angle to the elongation. The minima and saddle
points are correctly identified, and the orientation of the
ellipticity of the mass distribution is correctly reproduced.
\item Figure \ref{fig:6975} shows a lens with substructure in the form
of a smaller secondary galaxy. The galaxies in such group or
cluster sims were based on galaxies visible in the images, but the
modelers were not told in advance whether this was the case. The
minimum and saddle point are correctly identified. The mass
distribution misses the substructure, but overall appears
reasonable.
\item Figure \ref{fig:6937} shows a sim with substructre, like
Figure \ref{fig:6937}. In comparison to the above, the resulting mass
model is poor.
\item Figure \ref{fig:7025} shows a quad. In this one, the
identification of the minima and saddle points was incorrect, and
mass distribution comes out elongated East-West instead of
North-South. The mass distribution also appears somewhat jagged and
the saddle-point contours are not as clean as in the previous
examples; these are often indicators of a problem with the model.
The enclosed mass is, however, none the worse --- the reason is
probably that in a relatively symmetrical image configuration, the
Einstein radius is quite well constrained by the images in a fairly
model-independent way.
\item Figure \ref{fig:7022} shows another model of the same system, the
only one done by an expert in this sample. The image parities are
correct. The elongation has the right orientation, but is too shallow.
\end{itemize}
\subsection{Test of image identification} \label{sec:tests.t1}
The first post-modelling test was a qualitative comparison of the
original arrival-time surfaces and the input spaghetti diagrams given
by the modeller. This tested first, for correct identification and
location of the lensed images, and second, for the correct parities and
ordering of the lensed images in respect of the arrival time.
While we expected the identification of lensed images to be trivial,
given the generally clean appearance of the sims in the test, we
expected the parities and time-ordering to be more difficult. While
the SpaghettiLens\xspace tutorials had provided general guidelines, to be consistently
correct with the time-ordering, a modeller needs to develop some
intuition for arrival time surfaces. This is an area where experience
and tutorials training could improve results at a later stage, and
correspondingly, feedback on the difficulties modellers encounter can
help improve the tutorial materials.
\tabref{stats} presents a summary of the test. The evaluation was
done manually, comparing the input to SpaghettiLens\xspace with the actual
arrival-time surface of the sim. This amounts to comparing the
middle-left and middle-right panels in each of
Figures~\ref{fig:6941}--\ref{fig:7022}, and similarly for the other 111
models.
\begin{table}\centering\begin{tabular}{lrr}\hline
Total & 119 & 100\% \\
\hline
errors in image locations & 9 & 8\% \\
errors in image parities or time ordering & 49 & 41\% \\
\hline
inaccurate image placement over an arc & 21 & 18\% \\
identified two images of four & 5 & 4\% \\
identified two nearby images as one & 3 & 3\% \\
missed faint images & 1 & 1\% \\
proposed too many images & 1 & 1\% \\
modelled a three-image arc as one image & 4 & 3\% \\
modelled one image as a three-image arc & 5 & 4\% \\
swapped minimum and saddle in double & 2 & 2\% \\
swapped minima and saddles in quad & 38 & 32\% \\
swapped early and late saddles in a quad & 7 & 6\% \\
\hline
\end{tabular}
\caption{Table of image-identification errors and the number of models
containing each. A model can contain more than one type of error.}
\label{tab:stats}
\end{table}
The images of the system were considered to have been located
correctly, if all the images were identified and were approximated
within about 5\% of SpaghettiLens\xspace frame used to draw the spaghetti diagram.
That frame size is adjustable by the user, but in practice it is
somewhat larger than the spaghetti diagram. Such image-placement
errors were found in only 9 models. That does not include inaccurate
image placement over an arc, which was considered a separate category
of error.
In addition to simple image-placement errors, ten types of errors were
recognised and are listed in \tabref{stats}. Most of the problems
were due to unclear arc-like structures. Critical errors like the
failure to identify all five images in a five images system, or to
include too many images, were rare.
The assignment of the parity of the images was a more difficult task,
and was successful in only about 60\% of the cases. The most common
error was swapping of minima and saddle-points in a quad;
\figref{7025} shows an example. Another, less common, error was
flipping the spaghetti diagram, thus swapping the time-ordering of the
two saddle points.
Incorrect image parities and time orderings tended to produce
poorer-looking models, such as the checker board patterns in the mass
map in \figref{7025}. Interestingly, however, the enclosed-mass
profiles were quite robust. We will consider this aspect in the next
section.
\FloatBarrier
\begin{figure}
\centering
\subfigure{\includegraphics[width=0.8\linewidth]
{fig/007025_kappa_encl}}
\subfigure{\includegraphics[width=0.8\linewidth]
{fig/007022_kappa_encl}}
\subfigure{\includegraphics[width=0.8\linewidth]
{fig/007021_kappa_encl}} \\
\caption{Mean $\kappa$ inside a circle around the lens centre, as a
function of the radius of the circle. (See \secref{tests.t2} for
details.) The upper panel corresponds to the model shown in
\figref{7025}, in which the minima and saddle-points have been
incorrectly swapped. The middle corresponds to \figref{7022},
where the image parities were correct. The lower panel
corresponds to another model, where the image parities were
correct but the time-ordering was incorrect.}
\label{fig:kapenc_compare_faulty}
\end{figure}
\subsection{Test of mass-profile recovery} \label{sec:tests.t2}
The second test was to compare the mass distributions $\kappa(x,y)$ of
the sims and of the SpaghettiLens\xspace models. A visual comparison is presented for
the eight models in Figures~\ref{fig:6941}--\ref{fig:7022}, in the
lower-left versus lower-right panels. We will summarize the mass
distributions drastically in a single number, the effective Einstein
radius. Other measures for comparison of free-form lensing mass
distributions appear in \cite{2014arXiv1401.7990C}, but comparing
Einstein radii is already useful.
There is no standard way of defining the Einstein radius of a general
non-circular lens. We adopt the simple definition
\begin{equation} \label{eq:effRE}
\langle\kappa\rangle_{\Theta_{\rm E}} = 1,
\end{equation}
that is, the effective Einstein radius \ERf is such that the
mean $\kappa$ is unity inside a circle of radius \ERf centered at
the lens center.
\begin{figure*}
\centering
\includegraphics[width=0.90\linewidth]{fig/eR_5}
\caption{Model-recovered versus actual Einstein radii \ERf[,rec] and
\ERf[,act].
Plus signs indicate models flagged by the modeller as
failures by commenting negatively about it in the forum. Light and dark grey
bands show standard deviation of volunteers (15\%) and expert (10\%).}
\label{fig:ER_per_sim}
\end{figure*}
To illustrate, \figref{kapenc_compare_faulty} compares the circularly
averaged mass profiles of three different models of one particular lens;
two of the models are shown in Figures~\ref{fig:7025} and
\ref{fig:7022}. Each panel in \figref{kapenc_compare_faulty} shows the
mean $\kappa$ within a circle of given radius. The red curve is the
correct profile for the sim. The two blue curves are the minimal and
maximal mean enclosed $\kappa$ from the internal ensemble in SpaghettiLens\xspace.
Radial locations of the images are marked, along with the image
parities. The region between the blue curves is shaded between the
radii of the innermost and the outermost images: this is the
confidence region from the modelling. The definition \eqref{eq:effRE}
for \ERf corresponds to crossing the dashed horizontal line at $1$:
the red curve crosses the dashed line at the actual Einstein radius
\ERf[,act]; the recovered Einstein radius \ERf[,rec] and its
uncertainty are given by the blue curves crossing the dashed line. We
see that in all three panels, the blue curves are shallower than the
red curve and \ERf[,rec] is more than \ERf[,act], by more than the
model uncertainties. Now, steeper mass profiles tend to give wider
image separations --- recall that the image separation for a circular
isothermal lens is 2\ERf, whereas for a point mass it is more
\citep[see, e.g.,][]{2002LNP...608....1C} --- so \ERf[,rec] being too
high is really a consequence of the GLASS models being too shallow for
the sims.
\Figref{ER_per_sim} shows that \ERf[,rec] of the models tend to be too
high. However, this is entirely due to the GLASS model density
profiles being too shallow, as illustrated above. We can separate out
the performance of the SpaghettiLens\xspace interface and its users by comparing their
results with the Einstein radii of SpaghettiLens\xspace models made by an expert
(PS). Discounting the models which were flagged by the volunteers as
poor, the mean Einstein radius overestimate was 10\%, with a 15\%
standard deviation (shown by the light grey band in
\Figref{ER_per_sim}). The expert models show a similar bias, with
standard deviation 10\% (the dark grey band in \Figref{ER_per_sim}).
One source of this systematic error is that it is difficult to centre
the lens accurately: an offset leads to a flatter mass profile for the
model compared to the simulation.
\section{Outlook} \label{sec:todo}
This work has developed the concept of saddle-point contours in the
travel time of virtual photons, originally introduced by
\cite{1986ApJ...310..568B} for understanding image structure in strong
gravitational lenses, into a technique for mass-mapping lenses.
Despite being highly abstract, saddle-point contours look like
schematic arcs, and hence lend themselves to an intuitive markup tool
for lenses or lens candidates, which we call a spaghetti diagram. At
the same time, saddle-point contours encode information about possible
mass distributions, which can be translated into input for an existing
lens-modelling engine \citep[GLASS, by][]{2014arXiv1401.7990C}.
SpaghettiLens\xspace is an implementation of these ideas, enabling experienced but
non-professional lens enthusiasts to model newly-discovered lens
candidates from the Space Warps\xspace citizen-science platform. The tests in this
paper indicate that such modelling would be both feasible and
scientifically interesting: given a suitable modelling tool, and
appropriate guidance, a small team of non-professional volunteers was
able to model a sample of 29 test lenses, and measure their Einstein
radii with comparable accuracy to a professional expert.
There is, however, plenty of room for improvement:
\begin{enumerate}
\item SpaghettiLens\xspace tends to overestimate the Einstein radius (evident from
\Figref{ER_per_sim}), and the model density profiles tend to be
shallower than the SIE model used for generating the sims. The
likely explanation is that while the sims are steeply peaked at the
centre, the pixellated mass model fixes a comparatively large area
near the central at constant density. Allowing smaller pixels in the
centre region, thus enabling a steeper centre \citep[similar to the
``high resolution'' feature implemented in][]{2014arXiv1401.7990C}
may remove this bias. The use of simply parameterised,
appropriately steeply-profiled models would also avoid the problem.
\item Currently, SpaghettiLens\xspace does not attempt to model the source shape; the
user identifies the brightest points on the image, and these are
taken as images of a point-like source, whose positions must be
reproduced exactly. For generating a synthetic image, a conical
source profile is assumed. Fitting for the source profile to
optimize resemblance to the observed lensed image after the lens
model has been generated, is algorithmically straightforward
\citep[cf.][]{2003ApJ...590..673W,2006MNRAS.371..983S} and planned
to be implemented. This would alleviate another problem with SpaghettiLens\xspace,
which is that there is as yet no quantitative figure of merit for
any given model: assessment of each model is a judgment call based
on the synthetic image, and on whether the mass distribution and the
arrival-time surface show suspicious features. Another possibility
would be use the SpaghettiLens\xspace models as a feeder to a different
lens-modelling program that already implements source-profile
fitting.
\item Another limitation so far in SpaghettiLens\xspace is that the lens is assumed to
be dominated by one galaxy, which puts most galaxy-group lenses
beyond the reach of the modeler. Since complicated group lenses are
some of the most interesting candidates present, removing this
limitation is most desirable. From the users' point of view, it
would mean that spaghetti contours with more than one maximum can be
allowed. For examples, see Fig. 5c in \citep{2001ApJ...557..594R}
and Fig. 4b in \cite{2003ApJ...590...39K}.
\item At present, a single false-color composite is used as the data.
An option could be added to use all available filters, individually
or in combination, at the user wishes.
\item As mentioned above, the option of revising an already-archived
model is already available. Desired now are tools for comparing
different models of a given system, both visually and through
different statistical measures. As evidenced by a current
collaborative modelling effort, a particularly interesting candidate
can lead to an extended discussion and dozens of models, that
in some way sample the high likelihood region of model parameter
space.
\item Better tutorial materials are also needed, and this would
address some of the problem areas found in the modelling challenge.
For example, we saw in \secref{tests.t1} that volunteers are most
prone to making errors in two situations: when in identifying an arc-
like structure while placing the points, and in identifying the
correct ordering of the points in nearly-symmetric configurations.
Better and more detailed introductory materials would also allow
the community of modellers to grow faster and without individual
instructions by experts or experienced volunteers.
\end{enumerate}
The SpaghettiLens\xspace program was developed by Küng, with design suggestions from
Coles, Cornen and Saha, and feedback from all co-authors. The
simulations were created by A.~More, in consultation with Marshall,
S.~More and Verma. Modelling was done by Baeten, Cornen, Macmillan,
Odermatt, Saha and Wilcox, with post-modelling analysis by Küng and
Saha. All authors participated in writing and editing the manuscript.
\section*{Acknowledgments}
We thank the Swiss Society for Astrophysics and Astronomy and the
Swiss Academy of Sciences. The work of AM and SM was supported by
World Premier International Research Center Initiative (WPI
Initiative), MEXT, Japan. AV is supported by a research fellowship
from the Leverhulme Trust. RK is supported by the Swiss National
Science Foundation. This work was supported in part by the U.S.
Department of Energy under contract number DE-AC02-76SF00515.
|
\section{Introduction}\label{Introduction}
Consider the following convex optimization problem:
\begin{equation}\label{opt}\begin{array}{ll}
\displaystyle\min_{u,v}& \theta(u,v): = p(u) + q(v) + \phi(u,v),\\ [8pt]
\text{s.t.} & \mathcal{A}^*u + \mathcal{B}^*v = c,
\end{array}
\end{equation}
where $p: \mathcal{U}\to(-\infty,\infty]$, $q: \mathcal{V}\to(-\infty,\infty]$ are two closed proper convex functions (possibly nonsmooth), $\phi:\mathcal{U}\times \mathcal{V}\to (-\infty, \infty)$ is a smooth convex function whose gradient mapping is Lipschitz continuous,
$\mathcal{A}: \mathcal{X}\to\mathcal{U}$ and
$\mathcal{B}:\mathcal{X}\to\mathcal{V}$ are two given linear operators, $c\in\mathcal{X}$ is a given vector, and $\mathcal{U},\mathcal{V}$ and $\mathcal{X}$ are three real finite dimensional Euclidean spaces each equipped with an inner product $\langle \cdot,\cdot\rangle$ and its induced norm $\|\cdot\|$.
Many interesting optimization problems are of the form (\ref{opt}).
One particular case is the following problem whose objective is the sum of a quadratic function and a squared distance function to a closed convex set:
\begin{equation}\label{ex}\begin{array}{ll}
\min & \displaystyle\frac{1}{2}\bigg\langle \left(\begin{array}{c} u\\v\end{array}\right), \widetilde{\mathcal{Q}}\left(\begin{array}{c} u\\v\end{array}\right)\bigg\rangle + \frac{\rho}{2}\bigg\|\left(\begin{array}{c} u\\v\end{array}\right) - \Pi_{{\mathcal{K}_1}}\left(\begin{array}{c} u\\v\end{array}\right)\bigg\|^2, \\[8pt]
\text{s.t.} & \mathcal{A}^*u + \mathcal{B}^*v = c, \\[8pt]
&
u\in \mathcal{K}_2, ~v\in{\mathcal{K}_3},
\end{array}
\end{equation}
where $\rho>0$ is a penalty parameter, $\widetilde{\mathcal{Q}}:\mathcal{U}\times\mathcal{V}\to\mathcal{U}\times \mathcal{V}$ is a self-ajoint positive semidefinite linear operator, $\mathcal{K}_1\subseteq \mathcal{U}\times\mathcal{V}$, $\mathcal{K}_2\subseteq \mathcal{U}$ and $\mathcal{K}_3\subseteq \mathcal{V}$ are closed convex sets and $\Pi_{\mathcal{K}_1}(\cdot, \cdot)$ denotes the metric projection onto $\mathcal{K}_1$.
One popular way to solve problem (\ref{opt}) is the augmented Lagrangian method (ALM).
Given the Lagrangian multiplier $x\in\mathcal{X}$ of the linear constraint in (\ref{opt}), the augmented Lagrangian function associated with the parameter $\sigma >0$ is defined as
\begin{equation}
\mathcal{L}_\sigma(u,v;x) = \theta(u,v) + \langle x, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle +\frac{\sigma}{2}\|\mathcal{A}^*u + \mathcal{B}^*v -c\|^2, \quad (u,v)\in\mathcal{U}\times\mathcal{V}.
\end{equation}
The ALM minimizes $\mathcal{L}_\sigma(u,v;x)$ with respect to $(u,v)$ simultaneously regardless of whether the objective function is coupled or not before updating the Lagrangian multiplier $x$ along the gradient ascent direction. Numerically, however, to minimize $\mathcal{L}_\sigma(u,v;x)$ with respect to $(u,v)$ jointly may be a difficult task due to the non-separable structure of $\theta(\cdot,\cdot)$ combined with the nonsmoothness of $p(\cdot)$ and $q(\cdot)$.
When the objective function in (\ref{opt}) is separable for $u$ and $v$, one can alleviate the numerical difficulty in the ALM by directly applying the alternating direction method of multipliers (ADMM). The iteration scheme of the ADMM works as follows:
\begin{equation}\label{alm}
\left\{\begin{array}{ll}
u^{k+1} =\displaystyle\arg\min_{u} \mathcal{L}_\sigma(u,v^k; x^k),\\[8pt]
v^{k+1} = \displaystyle\arg\min_{v} \mathcal{L}_\sigma(u^{k+1},v; x^k), \\[8pt]
x^{k+1} = x^k + \tau\sigma(\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c),
\end{array}\right.
\end{equation}
where $\tau>0$ is the step length. The global convergence of the ADMM with $\tau\in(0,\frac{1+\sqrt{5}}{2})$ and a separable objective function has been extensively studied in the literature, see, for examples, ~\cite{Gabay83, Gabay1976, Glowinski1980lectures, Glowinski1975, Eckstein_1992}. For a recent survey, see Eckstein and Yao~\cite{Eckstein2014}. Although it is possible to apply the ADMM directly to problem (\ref{opt}) even if $\phi(\cdot,\cdot)$ is not separable, its convergence analysis is largely non-existent. One way to deal with the non-separablity of $\phi(\cdot,\cdot)$ is to introduce a new variable $w = \left(\begin{array}{c}u\\ v\end{array}\right)$. By letting $\widetilde{\mathcal{A}} = \left(\begin{array}{c}\mathcal{A}\\ {{\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}}_1\\0\end{array}\right)$, $\widetilde{\mathcal{B}} = \left(\begin{array}{c}\mathcal{B}\\ 0\\{{\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}}_2\end{array}\right)$, $\widetilde{\mathcal{C}} = \left(\begin{array}{cc}0 & 0\\ {{\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}}_1 & 0\\0 &{{\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}}_2\end{array}\right)$ and $\tilde{c} = \left(\begin{array}{c} c\\0 \\0 \end{array}\right)$ {with identity maps ${\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}_1:\cU \to \cU$ and ${\cal I}} \def\cX{{\cal X}} \def\cY{{\cal Y}} \def\cZ{{\cal Z}_2:\cV\to\cV$}, we can rewrite the optimization problem (\ref{ex}) equivalently as
\begin{equation}\label{opt_3}\begin{array}{ll}
\displaystyle\min_{u,v,w} & \tilde{\theta}(u,v,w): = p(u) + q(v) + \phi(w),\\[8pt]
\text{s.t.} & \widetilde{\mathcal{A}}^*u + \widetilde{\mathcal{B}}^*v + \widetilde{\mathcal{C}}^*w = \tilde{c}.
\end{array}
\end{equation}
For given $\sigma>0$, the corresponding augmented Lagrangian function for problem (\ref{opt_3}) is
\begin{equation*}
\widetilde{L}_\sigma(u,v,w; x) = \tilde{\theta}(u,v,w) + \langle x, \widetilde{\mathcal{A}}^*u + \widetilde{\mathcal{B}}^*v + \widetilde{\mathcal{C}}^*w - \tilde{c}\rangle + \displaystyle\frac{\sigma}{2}\| \widetilde{\mathcal{A}}^*u + \widetilde{\mathcal{B}}^*v + \widetilde{\mathcal{C}}^*w - \tilde{c}\|^2,
\end{equation*}
where $(u,v,w)\in\mathcal{U}\times\mathcal{V}\times(\mathcal{U}\times\mathcal{V})$ and $x\in\mathcal{X}$.
Directly applying the 3-Block ADMM yields the following framework:
\begin{equation*}
\left\{\begin{array}{ll}
u^{k+1} =\displaystyle\arg\min_{u} \widetilde{\mathcal{L}}_\sigma(u,v^k,w^k; x^k),\\[8pt]
v^{k+1} =\displaystyle\arg\min_{v} \widetilde{\mathcal{L}}_\sigma(u^{k+1},v,w^k; x^k),\\[8pt]
w^{k+1} =\displaystyle\arg\min_{w} \widetilde{\mathcal{L}}_\sigma(u^{k+1},v^{k+1},w; x^k),\\[8pt]
x^{k+1} = x^k + \tau\sigma(\widetilde{\mathcal{A}}^*u^{k+1} + \widetilde{\mathcal{B}}^*v^{k+1} + \widetilde{\mathcal{C}}^*w^{k+1} - \tilde{c}),\end{array}\right.
\end{equation*}
where $\tau >0$ is the step length.
Even though numerically the 3-block ADMM works well for many applications, generally it is not a convergent algorithm even if $\tau$ is as small as $10^{-8}$ as shown in the
counterexamples given by Chen et al.~\cite{Chen_divergence2013}.
In this paper, we will conduct a thorough convergence analysis about the 2-block ADMM when it is applied to problem (\ref{opt}) with non-separable objective functions. Unlike the case with separable objective functions, there are very few papers on the ADMM targeting the problem (\ref{opt}) except for the work of Hong et al.~\cite{Hong2014}, where the authors studied a majorized multi-block ADMM for linearly constrained optimization problems with non-separable objectives. When specialized to the 2-block case for problem (\ref{opt}), their algorithm works as follows:
\begin{equation}\label{Hong_iter}\left\{\begin{array}{ll}
u^{k+1} =\displaystyle\arg\min_{u}\{ p(u) +\langle x^k,\mathcal{A}^*u \rangle +
\hat{h}_1(u; u^k, v^k)\},\\[8pt]
v^{k+1} = \displaystyle\arg\min_{v} \{q(v) + \langle x^k, \mathcal{B}^*v\rangle +\hat{h}_2(v; u^{k+1}, v^k)\}, \\[8pt]
x^{k+1} = x^k + \alpha_k\sigma(\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c),
\end{array}\right.\end{equation}
where $\hat{h}_{1}(u; u^k,v^k)$ and $\hat{h}_{2}(v; u^{k+1},v^k)$ are majorization functions of $\phi(u,v) + \frac{\sigma}{2}\|\mathcal{A}^*u + \mathcal{B}^*v - c\|^2$ at $(u^k,v^k)$
and $(u^{k+1}, v^{k})$, respectively and $\alpha_k >0$ is the step length.
Hong et al.~\cite{Hong2014} provided a very general convergence analysis of their majorized ADMM assuming that the step length $\alpha_k$
is a sufficiently small fixed number or converging to zero,
among other conditions. Since a large step length is almost always desired in practice, one needs to develop a new convergence theorem beyond the one in~\cite{Hong2014}.
{Similar to} Hong et al.'s work \cite{Hong2014},
our approach also relies on the majorization technique applied to the smooth coupled function $\phi(\cdot,\cdot)$. One difference is that we majorize $\phi(u,v)$ at $(u^k,v^k)$ before the $(k+1)$th iteration instead of changing the majorization function based on $(u^{k+1},v^k)$ when updating $v^{k+1}$ as in (\ref{Hong_iter}). Interestingly, if $\phi(\cdot,\cdot)$ merely consists of quadratically coupled functions and separable smooth functions,
our majorized ADMM is exactly the same as the one proposed by Hong et al. under a proper choice of the majorization functions. Moreover, for applications like (\ref{ex}), a potential advantage of our method is that we only need to compute
the projection $\Pi_{\mathcal{K}_1}(\cdot,\cdot)$ once in order to compute
$\nabla \phi(\cdot, \cdot)$ as a part of the majorization function within one iteration, while {the procedure} (\ref{Hong_iter}) needs
to {compute $\Pi_{\mathcal{K}_1}(\cdot,\cdot)$ at}
two different points $(u^k,v^k)$ and $(u^{k+1}, v^k)$. In the subsequent discussions one can see that
by making use of nonsmooth analysis, especially the generalized Mean-Value Theorem, we are able to establish the global convergence and the iteration complexity for {our majorized ADMM} with the step length $\tau\in(0,\frac{1+\sqrt{5}}{2})$. To the best of our knowledge, this is the first paper providing the convergence {properties} of {the} majorized ADMM {with a large step length} for solving {linearly constrained convex optimization problems} with coupled smooth objective functions.
The remaining parts of our paper are organized as follows. In the next section, we provide some preliminary results. Section 3 focuses on our framework of a majorized ADMM and two important inequalities for the convergence analysis.
In Section 4, we prove the global convergence and several iteration complexity results of the proposed algorithm. We conclude our paper in the last section.
\section{Preliminaries}
In this section, we shall provide some preliminary results that will be used in our subsequent discussions.
Denote $w:= \left(\begin{array}{c} u\\ v \end{array}\right)$. Since $\phi(\cdot)$ is assumed to be a convex function with a Lipschitz continuous gradient,
$\nabla \phi(\cdot)$ is globally Lipschitz continuous and $\nabla^2 \phi(\cdot)$ exists almost everywhere.
Thus, the following Clarke's generalized Hessian at given $w\in\mathcal{U}\times\mathcal{V}$ is well defined~\cite{Clarke}:
\begin{equation}\label{gen_Hessian}
\partial^2 \phi(w) = \text{conv}\{\displaystyle\lim_{w^k\to w} \nabla^2 \phi(w^k), \nabla^2 \phi(w^k) \,\text{exists}\},
\end{equation}
where ``conv$\{S\}$'' denotes the convex hall of a given set $S$.
Note that $\mathcal{W}$ is self-adjoint and positive semidefinite, i.e.,
$\mathcal{W}\succeq 0$, for any $\mathcal{W}\in\partial^2\phi(w)$, $w\in\mathcal{U}\times\mathcal{V}$.
In~\cite{JBHU_1984}, Hiriart-Urruty and Nguyen provide a second order Mean-Value Theorem for $\phi$, which states that
for any $w'$ and $w$ in $\mathcal{U}\times\mathcal{V}$, {there exists $z\in[w',w]$ and $\mathcal{W}\in\partial^2\phi(z)$ such that}
\begin{equation*}
\phi(w) = \phi(w') + \langle \nabla\phi(w') , w-w'\rangle +\frac{1}{2}\langle w-w', \mathcal{W}(w-w')\rangle,
\end{equation*}
where $[w',w]$ denotes the line segment connecting $w'$ and $w$.
Since $\nabla\phi$ is globally Lipschitz continuous, there exist two self-adjoint positive semidefinite linear operators $\mathcal{Q}$ and $\mathcal{H}: \mathcal{U}\times\mathcal{V}\to\mathcal{U}\times\mathcal{V}$
such that for any $w\in\mathcal{U}\times\mathcal{V}$,
\begin{equation}\label{bound}
\mathcal{Q} \preceq \mathcal{W} \preceq \mathcal{Q}+\mathcal{H} \quad\forall \,\mathcal{W}\in\partial^2 \phi(w).
\end{equation}
Thus, for any $w, w'\in\mathcal{U}\times\mathcal{V}$, we have
\begin{equation}\label{phi_convex}
\phi(w)\geq \phi(w') + \langle \nabla \phi (w'), w - w' \rangle+\frac{1}{2}\|w'-w\|^2_{\mathcal{Q}}
\end{equation}
and
\begin{equation}\label{phi_majorize}
\phi(w) \leq \hat{\phi}(w; w'): = \phi(w') + \langle \nabla \phi (w'), w - w' \rangle+\frac{1}{2}\|w'-w\|^2_{\mathcal{Q}+\mathcal{H}}.\end{equation}
In this paper we further assume that
\begin{equation}\label{diag_H}
\mathcal{H} = \text{Diag}\,(\mathcal{D}_1, \mathcal{D}_2),
\end{equation}
where $\mathcal{D}_1: \mathcal{U}\to\mathcal{U}$ and $\mathcal{D}_2: \mathcal{V}\to\mathcal{V}$ are two self-adjoint positive semidefinite linear operators. In fact, this kind of structure naturally appears in applications like (\ref{ex}), where
the best possible lower bound of the generalized Hessian is $\widetilde{\mathcal{Q}}$ and the best possible upper bound of the generalized Hessian is $\widetilde{\mathcal{Q}}+\mathcal{I}$, where $\mathcal{I}: \mathcal{U}\times\mathcal{V}\to\mathcal{U}\times\mathcal{V}$ is the identity operator. For this case, the tightest estimation of $\mathcal{H}$ is $\mathcal{I}$, which is block diagonal.
Since the coupled function $\phi(u,v)$ consists of two block variables $u$ and $v$, the operators $\mathcal{Q}$ and $\mathcal{W}$ can be decomposed accordingly as $\mathcal{Q} = \begin{pmatrix} \mathcal{Q}_{11} & \mathcal{Q}_{12}\\ \mathcal{Q}_{12}^* & \mathcal{Q}_{22}\end{pmatrix}$ and $\mathcal{W} = \begin{pmatrix} \mathcal{W}_{11} & \mathcal{W}_{12}\\ \mathcal{W}_{12}^* & \mathcal{W}_{22}\end{pmatrix}$, where $\mathcal{W}_{11},~\mathcal{Q}_{11}: \mathcal{U}\to\mathcal{U}$ and $\mathcal{W}_{22}, ~\mathcal{Q}_{22}: \mathcal{V}\to\mathcal{V}$ are self-adjoint positive semidefinite linear operators, and $\mathcal{W}_{12}, ~\mathcal{Q}_{12}:\mathcal{V}\to\mathcal{U}$ are two linear mappings whose adjoints are given by $\mathcal{W}_{12}^*$ and $\mathcal{Q}_{12}^*$, respectively.
Denote $\eta \in [0,1]$ as a constant that satisfies
\begin{equation}\label{cross_control}
|\langle u, (\mathcal{W}_{12}-\mathcal{Q}_{12}) v\rangle |\leq \frac{\eta}{2}(\|u\|^2_{\mathcal{D}_1} + \|v\|^2_{\mathcal{D}_2})\quad\forall\,\mathcal{W}\in \partial^2 \phi(u,v), ~u\in\mathcal{U}, ~v\in\mathcal{V}.
\end{equation}
Note that (\ref{cross_control}) always holds true for $\eta = 1$ according to the Cauchy-Schwarz inequality.
In order to prove the convergence of the proposed majorized ADMM, the following constraint qualification is needed:
\begin{assumption}\label{cq}
There exists $(\hat{u},\hat{v})\in\textup{ri}\;(\textup{dom}(p)\times\textup{dom}(q))$ such that $\mathcal{A}^*\hat{u} + \mathcal{B}^*\hat{v} = c$.
\end{assumption}
Let $\partial p$ and $\partial q$ be the subdifferential mappings of $p$ and $q$, respectively. Define the set-valued mapping $\mathcal{F}$ by
\begin{equation*}
\mathcal{F}(u,v,x): = \nabla \phi(w) + \left(\begin{array}{cc}\partial p(u) + \mathcal{A}x\\[5pt] \partial q(v) + \mathcal{B}x\end{array}\right), \quad (u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}.
\end{equation*}
Under Assumption \ref{cq}, $(\bar{u}, \bar{v})$ is optimal to (\ref{opt}) if and only if there exists $\bar{x}\in\mathcal{X}$ such that the following Karush-Kuhn-Tucker (KKT) condition holds:
\begin{equation}\label{kkt}\left\{\begin{array}{ll}
0\in F(\bar{u},\bar{v},\bar{x}),\\[8pt]
\mathcal{A}^*\bar{u} + \mathcal{B}^*\bar{v} = c,
\end{array}\right.
\end{equation}
which is equivalent to the following variational inequality:
\begin{equation}\label{VI}\begin{array}{rr}
&(p({u}) + q({v}) ) -( p(\bar u) + q(\bar v) ) + \langle {w}-\bar{w},\nabla\phi(\bar{w})\rangle+ \langle u - \bar{u}, \mathcal{A}\bar{x}\rangle + \langle v - \bar v , \mathcal{B}\bar{x}\rangle\\[8pt]
&
- \langle x - \bar x, \mathcal{A}^*\bar u + \mathcal{B}^*\bar v - c\rangle\geq 0
\quad\forall (u,v,x)\in \mathcal{U}\times\mathcal{V}\times\mathcal{X}.
\end{array}
\end{equation}
Motivated by Nesterov's definition of an $\varepsilon$-approximation solution based on the first order optimality condition \cite[Definition 1]{Nesterov_2013} , we say that $(\tilde{u}, \tilde{v}, \tilde{x} ̃)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}$ is an $\varepsilon$-approximation solution to problem (\ref{opt}) if
\begin{equation}\label{def_approxiamteVI}\begin{array}{ll}
& (p(\tilde{u}) + q(\tilde{v}) ) -( p(u) + q(v) )+\langle \tilde{w}-w,\nabla\phi(w)\rangle+ \langle \tilde{u} - u, \mathcal{A}x\rangle + \langle \tilde{v} - v , \mathcal{B}x\rangle \\[8pt]
&
- \langle \tilde x - x, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle\leq \varepsilon \quad\forall (u,v,x)\in{{B}}(\tilde{u}, \tilde{v}, \tilde{x} ̃),
\end{array}
\end{equation}
where ${
{B}(\tilde{u}, \tilde{v}, \tilde{x} ̃)}= \{(u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}|\|(u,v,x) - (\tilde u, \tilde v, \tilde x)\| \leq 1\}
$.
Furthermore, since $p$ and $q$ are convex functions, $\partial p(\cdot)$ and $\partial q(\cdot)$ are maximal monotone operators. Then, for any $u,\hat{u}\in \textup{dom}(p)$, $\xi\in\partial p(u)$, and $\hat{\xi}\in\partial p(\hat{u})$, we have
\begin{equation}\label{p_convex}\begin{array}{ll}
&\langle u - \hat{u},\xi - \hat{\xi}\rangle \geq 0,
\end{array}
\end{equation}
and similarly for any $v,\hat{v} \in\textup{dom}(q)$, $\zeta\in\partial q(v)$, and $\hat{\zeta}\in\partial q(\hat{v})$, we have
\begin{equation}\label{q_convex}\begin{array}{ll}
&\langle v - \hat{v}, \zeta- \hat{\zeta}\rangle \geq 0.
\end{array}
\end{equation}
\section{A majorized ADMM with coupled objective functions}
In this section, we will first present the framework of our majorized ADMM and then prove two important inequalities that play an essential role for our convergence analysis.
Let $\sigma>0$.
For given $w' = (u',v')\in\mathcal{U}\times \mathcal{V}$,
define the following majorized augmented Lagrangian function associated with (\ref{opt}):
\begin{equation*}\begin{array}{ll}
\widehat{\mathcal{L}}_\sigma(w; (x,w')): =& p(u) + q(v) +\hat{\phi}(w; w') + \langle x,\mathcal{A}^*u + \mathcal{B}^*v - c\rangle +\displaystyle\frac{\sigma}{2}\|\mathcal{A}^*u + \mathcal{B}^*v - c\|^2,
\end{array}
\end{equation*}
where $(w,x) = (u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}$ and
the majorized function $\hat\phi$ is given by (\ref{phi_majorize}).
Then our proposed algorithm works as follows:
\bigskip
\centerline{\fbox{\parbox{\textwidth}{
\textbf{Majorized ADMM:
A majorized ADMM with coupled objective functions} \\[3mm]
Choose an initial point $(u^0, v^0, x^0)\in\text{dom}(p)\times\text{dom}(q)\times\mathcal{X}$ and parameters $\tau >0$. Let $\mathcal{S}$ and $\mathcal{T}$ be given self-adjoint positive semidefinite linear operators. Set $k:=0.$ Iterate until convergence:\\[3mm]
\textbf{Step 1.} Compute $u^{k+1} = \displaystyle\arg\min_{u\in\mathcal{U}}\{ \widehat{\mathcal{L}}_\sigma(u,v^k; (x^k,w^k)) + \frac{1}{2}\|u-u^k\|^2_\mathcal{S}\} .$ \\ [8pt]
\textbf{Step 2.} Compute $v^{k+1} = \displaystyle\arg\min_{v\in\mathcal{V}}\{\widehat{\mathcal{L}}_\sigma(u^{k+1},v;(x^k,w^k)) + \frac{1}{2}\|v-v^k\|^2_\mathcal{T}\}$.\\ [8pt]
\textbf{Step 3.} Compute $x^{k+1}=x^k + \tau\sigma(\mathcal{A}^*u^{k+1}+\mathcal{B}^*v^{k+1}-c).$
}}}
\bigskip
In order to simplify subsequent discussions, for $k = 0,1,2,\cdots$, define
\begin{equation}\label{notation1}\left\{
\begin{array}{ll}
\tilde{x}^{k+1} := x^k + \sigma(\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c),\\[8pt]
\Xi_{k+1} := \|v^{k+1} - v^k\|^2_{\mathcal{D}_2+ \mathcal{T}}
+ \eta\|u^{k+1} - u^k\|^2_{\mathcal{D}_1},\\[8pt]
\Theta_{k+1} := \|u^{k+1} - u^k\|^2_{\mathcal{S} } + \|v^{k+1} - v^k\|^2_{\mathcal{T}} + \frac{1}{4}\|w^{k+1}-w^k\|^2_{\mathcal{Q}},\\[8pt]
\Gamma_{k+1} := \Theta_{k+1}+ \min(\tau, 1+\tau - \tau^2)\|v^{k+1} - v^k\|^2_{\sigma\mathcal{B}\mathcal{B}^* } - \|u^{k+1} - u^k\|^2_{\eta\mathcal{D}_1} - \|v^{k+1} - v^k\|^2_{\eta\mathcal{D}_2 }
\end{array}\right.
\end{equation}
and denote for $(u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}$,
\begin{equation}\label{notation2}\left\{\begin{array}{ccl}
\Phi_{k}(u,v,x)& := &(\tau\sigma)^{-1} \|x^{k}-x\|^2 + \|u^{k}-u\|^2_{\mathcal{D}_1+ \mathcal{S}} + \|v^{k}-v\|^2_{\mathcal{Q}_{22}+ \mathcal{D}_2 + \mathcal{T}} + \frac{1}{2}\|w^{k}-w\|^2_\mathcal{Q}\\[8pt]
&&
+ \sigma\|\mathcal{A}^*u+\mathcal{B}^*v^k-c\|^2,\\[8pt]
\Psi_{k}(u,v,x) &:= & \Phi_{k}(u,v,x)+\|w^{k}-w\|^2_{\mathcal{Q}}+ \max(1-\tau,1-\tau^{-1})\sigma\|\mathcal{A}^*u^{k} + \mathcal{B}^*v^{k} - c\|^2.
\end{array}\right.
\end{equation}
\begin{prop}\label{decrease}
Suppose that the solution set of problem (\ref{opt}) is nonempty and Assumption \ref{cq} holds. Assume that $\mathcal{S}$ and $\mathcal{T}$ are chosen such that the sequence $\{(u^k,v^k,x^k)\}$ is well defined. Then the following conclusions hold:\\[8pt]
(i) For $\tau\in(0,1]$, we have that for any $k\geq 0$ and $(u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}$,
\begin{equation}\label{decrease_11}
\begin{array}{ll}
&\displaystyle(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&\displaystyle
- \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle +
\frac{1}{2}(\Phi_{k+1}({u}, {v}, {x}) - \Phi_k({u}, {v}, {x}))\\[8pt]
\leq & \displaystyle-\frac{1}{2}(\Theta_{k+1}+ \sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2
+(1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2 ).
\end{array}
\end{equation}
(ii) For $\tau \geq 0$, we have that for any $k\geq 1$ and $(u,v,x)\in\mathcal{U}\times\mathcal{V}\times\mathcal{X}$,
\begin{equation}\label{decrease_22}
\begin{array}{ll}
&(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&\displaystyle
- \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle +
\frac{1}{2}(\Psi_{k+1}({u}, {v}, {x}) + \Xi_{k+1} - (\Psi_k({u}, {v}, {x}) +\Xi_k))\\[8pt]
\leq &\displaystyle
-\frac{1}{2}(\Gamma_{k+1} +\min(1,1+\tau^{-1} - \tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2).
\end{array}
\end{equation}
\end{prop}
{\bf Proof.}
In the majorized ADMM iteration scheme, the optimality condition for $(u^{k+1},v^{k+1})$ is
\begin{equation}\label{opt_1}\left\{\begin{array}{ll}
0\in&\partial p(u^{k+1}) + \nabla_u \phi(w^k) + \mathcal{A}x^k + \sigma\mathcal{A}(\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c) + (\mathcal{Q}_{11} + \mathcal{D}_1+ \mathcal{S})(u^{k+1} - u^k),\\[8pt]
0\in&\partial q(v^{k+1})+ \nabla_v \phi(w^k) + \mathcal{B}x^k + \sigma\mathcal{B}(\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c) + (\mathcal{Q}_{22} + \mathcal{D}_2+ \mathcal{T})(v^{k+1} - v^k) \\[8pt]
&
+ \mathcal{Q}_{12}^*(u^{k+1} - u^k).
\end{array}\right.
\end{equation}
Denote
\begin{equation*}\left\{\begin{array}{ll}
a^{k+1} &= -x^{k+1} - (1-\tau)\sigma(\mathcal{A}^*u^{k+1}+\mathcal{B}^*v^{k+1}-c) - \sigma\mathcal{B}^*(v^k - v^{k+1}),\\[8pt]
b^{k+1} & = -x^{k+1} - (1-\tau)\sigma (\mathcal{A}^*u^{k+1}+\mathcal{B}^*v^{k+1}-c).
\end{array}\right.
\end{equation*}
Then by noting that
$$x^{k+1}= x^k + \tau\sigma(\mathcal{A}^*u^{k+1}+\mathcal{B}^*v^{k+1}-c),$$
we can rewrite (\ref{opt_1}) as
\begin{equation}\label{opt_2}\left\{\begin{array}{ll}
&\mathcal{A}a^{k+1} -\nabla_u \phi(w^k) - (\mathcal{Q}_{11} + \mathcal{D}_1+ \mathcal{S})(u^{k+1} - u^k)\in\partial p(u^{k+1}),\\[8pt]
&\mathcal{B}b^{k+1} - \nabla_v \phi(w^k)-(\mathcal{Q}_{22} + \mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k) - \mathcal{Q}_{12}^*(u^{k+1} - u^k) \in\partial q(v^{k+1}).
\end{array}\right.
\end{equation}
Therefore, by the convexity of $p$ and $q$, we have that for any $u\in\mathcal{U}$ and $v\in\mathcal{V}$,
\begin{equation}\label{pqineq}\left\{\begin{array}{l}
p(u) \geq p(u^{k+1}) + \langle u - u^{k+1},\mathcal{A}a^{k+1} -\nabla_u \phi(w^k) - (\mathcal{Q}_{11}+\mathcal{D}_1+ \mathcal{S})(u^{k+1} - u^k)\rangle,\\[8pt]
q(v) \geq q(v^{k+1}) + \langle v - v^{k+1}, \mathcal{B}b^{k+1} - \nabla_v \phi(w^k)-(\mathcal{Q}_{22} + \mathcal{D}_2 + \mathcal{T})( v^{k+1}-v^k)-\mathcal{Q}^*_{12}(u^{k+1}- u^{k})\rangle.
\end{array}\right.
\end{equation}
By noting the relationship between $x^{k+1}, \tilde{x}^{k+1}$ and $x^k$, we obtain from the above inequalities that for any $(u,v)\in\mathcal{U}\times\mathcal{V}$,
\begin{equation}\label{main_ineq}\begin{array}{ll}
&(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle \\[8pt]
\leq & \sigma\langle \mathcal{B}^*(v^{k+1} - v^k),\mathcal{A}^*(u^{k+1}-u)\rangle
-\langle w^{k+1} - w, \nabla\phi(w^{k}) - \nabla\phi(w)\rangle - \langle \mathcal{Q}_{12}^*(u^{k+1} - u^k), v^{k+1}-v\rangle\\[8pt]
&
-\langle (\mathcal{Q}_{11}+\mathcal{D}_1 + \mathcal{S})(u^{k+1}-u^k),u^{k+1} - u\rangle
- \langle (\mathcal{Q}_{22}+\mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k), v^{k+1}-v\rangle\\[8pt]
&
- (\tau\sigma)^{-1}\langle x^{k+1} - x^k, x^{k+1} - x\rangle - (1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2.
\end{array}
\end{equation}
By taking $(w, w') = (\bar{w}, w^k)$ and $(w^{k+1}, \bar{w})$ in (\ref{phi_convex}), we know that
\begin{equation*}\begin{array}{rll}
\phi({w}) &\geq & \phi(w^k) + \langle \nabla \phi(w^k), {w} - w^k\rangle + \frac{1}{2}\|{w} - w^k\|^2_{\mathcal{Q}} ,\\ [8pt]
\phi(w^{k+1}) & \geq & \phi({w}) + \langle \nabla \phi({w}),w^{k+1} - {w}\rangle + \frac{1}{2}\|w^{k+1} - {w}\|^2_{\mathcal{Q}} .
\end{array}
\end{equation*}
By taking $(w, w') = (w^{k+1}, w^k)$ in (\ref{phi_majorize}), we can also get that
\begin{equation*}
\phi(w^{k+1}) \leq \phi(w^{k}) + \langle \nabla \phi(w^{k}), w^{k+1} - w^k\rangle +\frac{1}{2}\|w^{k+1} - w^k\|^2_{\mathcal{Q} +\mathcal{H}}.
\end{equation*}
Putting the above three inequalities together, we get
\begin{equation}\label{fg_estimate}
\langle \nabla \phi(w^k) - \nabla \phi({w}), w^{k+1}-w\rangle
\geq
\frac{1}{2}(\|w^k-w\|^2_{\mathcal{Q}}+\|w^{k+1}-w\|^2_{\mathcal{Q}}) -\frac{1}{2}\|w^{k+1} - w^k\|^2_{\mathcal{Q} +\mathcal{H}}.
\end{equation}
Note that
\begin{equation}\label{equiv}
\begin{array}{ll}
&\displaystyle\frac{1}{2}(\|w^{k+1} - w^k\|^2_{\mathcal{Q}} - \|w^{k+1}-w\|^2_\mathcal{Q} - \|w^k-w\|^2_\mathcal{Q}) - \langle \mathcal{Q}_{11}(u^{k+1} - u^k), u^{k+1}-u\rangle\\[8pt]
& - \langle \mathcal{Q}_{22}(v^{k+1} - v^k), v^{k+1}-v\rangle
- \langle \mathcal{Q}^*_{12}( u^{k+1} - u^k), v^{k+1}-v\rangle\\[8pt]
= &\displaystyle \frac{1}{2}(\|w^{k+1} - w^k\|^2_{\mathcal{Q}} - \|w^{k+1}-w\|^2_{\mathcal{Q}} - \|w^k-w\|^2_{\mathcal{Q}}) - \langle w^{k+1} - w^k,\mathcal{Q}(w^{k+1}-w)\rangle \\[8pt]
&
+ \langle \mathcal{Q}_{12}(v^{k+1} - v^k), u^{k+1}-u\rangle\\[8pt]
= &-\|w^{k+1}-w\|^2_{\mathcal{Q}}+ \langle \mathcal{Q}_{12}(v^{k+1} - v^k), u^{k+1}-u\rangle.
\end{array}
\end{equation}
Substituting (\ref{fg_estimate}) and (\ref{equiv}) into (\ref{main_ineq}) and by the assumption (\ref{diag_H}), we can further obtain that
\begin{equation}\label{main_ineq_2}\begin{array}{ll}
&(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle \\[8pt]
\leq
&
\sigma\langle \mathcal{B}^*(v^{k+1} - v^k),\mathcal{A}^*(u^{k+1}-u)\rangle- \|w^{k+1}-w\|^2_{\mathcal{Q}}+\frac{1}{2}\|w^{k+1} - w^k\|^2_{\text{Diag}\,(\mathcal{D}_1, \mathcal{D}_2)}\\[8pt]
&
+\langle v^{k+1}-v^k, \mathcal{Q}_{12}^*(u^{k+1} - u)\rangle
-\langle (\mathcal{D}_1 + \mathcal{S})(u^{k+1}-u^k),u^{k+1} - u\rangle\\[8pt]
&
- \langle (\mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k), v^{k+1}-v\rangle
- (\tau\sigma)^{-1}\langle x^{k+1} - x^k, x^{k+1} - x\rangle\\[8pt]
& - (1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2.
\end{array}
\end{equation}
Recall that for any $\xi$, $\zeta$ in the same space and a self-adjoint positive semidefinite operator $\mathcal{G}$, it always holds that
\begin{equation}\label{inner_product_eq}
\langle \xi, \mathcal{G}\zeta\rangle = \frac{1}{2}(\|\xi\|_\mathcal{G}^2 + \|\zeta\|_\mathcal{G}^2 - \|\xi-\zeta\|_\mathcal{G}^2) = \frac{1}{2}(\|\xi+\zeta\|_\mathcal{G}^2 - \|\xi\|_\mathcal{G}^2 - \|\zeta\|_\mathcal{G}^2).
\end{equation}
Then we can get that
\begin{equation}\label{cross_st}\begin{array}{rcl}
\langle (\mathcal{D}_1 + \mathcal{S})(u^{k+1} - u^k), u^{k+1} - u\rangle &=& \displaystyle\frac{1}{2}(\|u^{k+1} - u^k\|_{\mathcal{D}_1+\mathcal{S}}^2 + \|u^{k+1} - u\|_{\mathcal{D}_1+\mathcal{S}}^2 - \|u^k - u\|_{\mathcal{D}_1+\mathcal{S}}^2),\\[8pt]
\langle (\mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k), v^{k+1} - v\rangle &=& \displaystyle\frac{1}{2}(\|v^{k+1} - v^k\|_{\mathcal{D}_2 + \mathcal{T}}^2 + \|v^{k+1} - v\|_{\mathcal{D}_2 + \mathcal{T}}^2 - \|v^k - v\|^2_{\mathcal{D}_2 + \mathcal{T}}),\\[8pt]
\langle x^{k+1} - x^k, x^{k+1} - x\rangle &=& \displaystyle \frac{1}{2}(\|x^{k+1} - x^k\|^2 + \|x^{k+1} - x\|^2 - \|x^k - x\|^2),\\[8pt]
\langle \mathcal{Q}_{22}(v^{k+1} - v^{k}), v^{k+1}-v \rangle & = &\displaystyle \frac{1}{2}(\|v^{k+1} - v^k\|_{\mathcal{Q}_{22}}^2 + \|v^{k+1} - v\|_{\mathcal{Q}_{22}}^2 - \|v^k - v\|_{\mathcal{Q}_{22}}^2).
\end{array}
\end{equation}
{\bf (i)} Assume that $\tau\in(0,1]$.
By using the last equation in (\ref{cross_st}), we can obtain that
\begin{equation}\label{cross_1}\begin{array}{lll}
\inprod{v^{k+1} - v^{k}}{\mathcal{Q}_{12}^*(u^{k+1}-u) }
&= & \Inprod{ \left(\begin{array}{c} 0\\ v^{k+1} - v^{k}\end{array}\right)}{\mathcal{Q}( w^{k+1}-w )}- \langle \mathcal{Q}_{22}(v^{k+1} - v^{k}), v^{k+1}-v \rangle \\[8pt]
&\leq & \displaystyle\frac{1}{2}(\|v^{k+1} - v^k\|^2_{\mathcal{Q}_{22}} + \|w^{k+1}-w\|^2_{\mathcal{Q}}) - \frac{1}{2}(\|v^{k+1} - v^k\|^2_{\mathcal{Q}_{22}} \\[8pt]
&&
+ \|v^{k+1}-v\|^2_{\mathcal{Q}_{22}} - \|v^k-v\|^2_{\mathcal{Q}_{22}})\\[8pt]
&= & \displaystyle\frac{1}{2}\|w^{k+1}-w\|^2_{\mathcal{Q}} + \frac{1}{2}(\|v^{k}-v\|^2_{\mathcal{Q}_{22}} - \|v^{k+1}-v\|^2_{\mathcal{Q}_{22}}),
\end{array}
\end{equation}
where the inequality is obtained by the Cauchy-Schwarz inequality.
By some simple manipulations we can also see that
\begin{equation}\label{cross_2}\begin{array}{ll}
\sigma\langle \mathcal{B}^*(v^{k+1} - v^k), \mathcal{A}^*(u^{k+1}-u)\rangle
& = \displaystyle\frac{\sigma}{2}( \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2 - \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2)\\[8pt]
& \displaystyle\quad +\frac{\sigma}{2}(\|\mathcal{A}^*u + \mathcal{B}^*v^k-c\|^2 -\|\mathcal{A}^*u + \mathcal{B}^*v^{k+1}-c\|^2).
\end{array}
\end{equation}
Finally, by substituting (\ref{cross_st}), (\ref{cross_1}) and (\ref{cross_2}) into (\ref{main_ineq_2}) and recalling the definition of $\Phi_{k+1}(\cdot, \cdot, \cdot)$ and $\Theta_{k+1}$ in (\ref{notation1}) and (\ref{notation2}), we have that
\begin{equation*}\begin{array}{ll}
&(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&
\displaystyle - \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle +
\frac{1}{2}(\Phi_{k+1}({u}, {v}, {x}) - \Phi_k({u}, {v}, {x}))\\[8pt]
& \leq \displaystyle - \frac{1}{2} ( \|u^{k+1} - u^k\|^2_{\mathcal{S} } + \|v^{k+1} - v^k\|^2_{\mathcal{T}} + \frac{1}{2}\|w^{k+1}-w\|^2_{\mathcal{Q}}+\frac{1}{2} \|w^k-w\|^2_{\mathcal{Q}}
+ \sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2 \\[8pt]
&
\quad\quad\quad+(1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2)\\[8pt]
& \leq \displaystyle - \frac{1}{2}(\Theta_{k+1} + \sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2+ (1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2),
\end{array}
\end{equation*}
where the last inequality comes from the fact that
\begin{equation*}
\frac{1}{2}\|w^{k+1}-w\|^2_{\mathcal{Q}} + \frac{1}{2}\|w^{k}-w\|_{\mathcal{Q}}^2\geq \frac{1}{4}\|w^{k+1} - w^k\|^2_{\mathcal{Q}}.
\end{equation*}
This completes the proof of part (i).\\[8pt]
{\bf (ii)} Assume that $\tau\geq 0$. In this part, we first reformulate (\ref{main_ineq_2}) as
\begin{equation}\label{main_ineq_3}\begin{array}{ll}
&(p(u^{k+1}) + q(v^{k+1}) ) -(p(u) + q(v) ) +\langle w^{k+1}-w,\nabla\phi(w)\rangle+ \langle u^{k+1} - u, \mathcal{A}x\rangle + \langle v^{k+1} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \tilde{x}^{k+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle \\[8pt]
\leq
&
\displaystyle\sigma\langle \mathcal{B}^*(v^{k+1} - v^k),\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\rangle+\frac{\sigma}{2}(\|\mathcal{A}^*u + \mathcal{B}^*v^k-c\|^2 -\|\mathcal{A}^*u + \mathcal{B}^*v^{k+1}-c\|^2) \\[8pt]
&
\displaystyle- \frac{1}{2}\|v^{k+1} - v^k\|^2_{\sigma\mathcal{B}\mathcal{B}^*}
- \|w^{k+1}-w\|^2_{\mathcal{Q}}+\frac{1}{2}\|w^{k+1} - w^k\|^2_{\text{Diag}\,(\mathcal{D}_1, \mathcal{D}_2)} - \langle \mathcal{Q}_{22}(v^{k+1} - v^k), v^{k+1}-v\rangle\\[8pt]
&
+ \langle v^{k+1}-v^k, \mathcal{Q}_{12}^*(u^{k+1} - u) + \mathcal{Q}_{22}(v^{k+1} - v)\rangle
-\langle (\mathcal{D}_1 + \mathcal{S})(u^{k+1}-u^k),u^{k+1} - u\rangle\\[8pt]
&
- \langle (\mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k), v^{k+1}-v\rangle
- (\tau\sigma)^{-1}\langle x^{k+1} - x^k, x^{k+1} - x\rangle \\[8pt]
&
- (1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2.
\end{array}
\end{equation}
Next we shall estimate the following cross term
$$\sigma\langle \mathcal{B}^*(v^{k+1} - v^k), \mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\rangle + \langle v^{k+1} - v^k, \mathcal{Q}_{12}^*(u^{k+1}-u) + \mathcal{Q}_{22}(v^{k+1}-v)\rangle.$$
It follows from (\ref{opt_2}) that
\begin{equation}\label{second_opt}
\left\{\begin{array}{ll}
\mathcal{B}b^{k+1} - \nabla_v \phi(w^k)-(\mathcal{Q}_{22}+\mathcal{D}_2 + \mathcal{T})(v^{k+1} - v^k) - \mathcal{Q}_{12}^*(u^{k+1} - u^k) \in\partial q(v^{k+1}), \\[8pt]
\mathcal{B}b^{k} - \nabla_v \phi(w^{k-1})-(\mathcal{Q}_{22}+\mathcal{D}_2 + \mathcal{T})(v^{k} - v^{k-1})-\mathcal{Q}_{12}^*(u^k-u^{k-1}) \in\partial q(v^{k}).
\end{array}\right.
\end{equation}
Since $\nabla \phi$ is globally Lipschitz continuous, it is known from Clarke's Mean-Value Theorem \cite[Proposition 2.6.5]{Clarke} that there exists a self-adjoint and positive semidefinite operator \\
$\mathcal{W}^k \in \text{conv}\{\partial^2\phi([w^{k-1}, w^k])\} $ such that
\begin{equation*}
\nabla\phi(w^k) - \nabla\phi(w^{k-1})= \mathcal{W}^k(w^k-w^{k-1}),
\end{equation*}
{where the set $\text{conv}\{\partial^2\phi[w^{k-1},w^k]\}$ denotes the convex hull of all points
$\cW\in \partial^2\phi(z)$ for any $z\in [w^{k-1},w^k]$.}
Denote $\mathcal{W}^k: = \begin{pmatrix} \mathcal{W}_{11}^k & \mathcal{W}_{12}^k \\ (\mathcal{W}_{12}^k)^* & \mathcal{W}_{22}^k\end{pmatrix}$, where $\mathcal{W}_{11}^k: \mathcal{U}\to\mathcal{U}$, $\mathcal{W}_{22}^k: \mathcal{V}\to\mathcal{V}$ are self-adjoint positive semidefinite operators and $\mathcal{W}_{12}^k: \mathcal{U}\to\mathcal{V}$ is a linear operator. {Substituting (\ref{second_opt}) into (\ref{q_convex}) at $ v = v^{k+1}$ and $\hat{v} = v^k$, we obtain that
\begin{equation*}\begin{array}{ll}
&\langle \mathcal{B}(b^{k+1} - b^k), v^{k+1} - v^k\rangle -\langle \mathcal{Q}_{22}(v^{k+1} - v^{k}) +\mathcal{Q}_{12}^* (u^{k+1} - u^{k}), v^{k+1} - v^k\rangle \\[8pt]
\geq &\langle \nabla_v\phi(w^k) - \nabla_v \phi(w^{k-1}), v^{k+1} -v^k\rangle -\langle (\mathcal{Q}_{22} + \mathcal{D}_2+ \mathcal{T})(v^{k} - v^{k-1}), v^{k+1} - v^k\rangle \\[8pt]
&
\quad + \|v^{k+1} - v^k\|^2_{\mathcal{T}+\mathcal{D}_2} -\langle u^k-u^{k-1}, \mathcal{Q}_{12}(v^{k+1} - v^k)\rangle\\[8pt]
= & \langle u^k-u^{k-1},(\mathcal{W}_{12}^k-\mathcal{Q}_{12})( v^{k+1} - v^k)\rangle -\langle ( \mathcal{Q}_{22}+ \mathcal{D}_2+\mathcal{T} - \mathcal{W}_{22}^k)(v^{k} - v^{k-1}), v^{k+1} - v^k\rangle\\[8pt]
&
\quad + \|v^{k+1} - v^k\|^2_{\mathcal{T}+\mathcal{D}_2} \\[8pt]
\geq & \displaystyle-\frac{\eta}{2}(\|u^k - u^{k-1}\|^2_{\mathcal{D}_1} + \|v^{k+1} - v^k\|^2_{\mathcal{D}_2}) - \frac{1}{2}(\|v^{k+1} - v^k\|^2_{\mathcal{T}+ \mathcal{D}_2}
+ \|v^k-v^{k-1}\|^2_{\mathcal{T}+ \mathcal{D}_2})\\[8pt]
&
\quad+\|v^{k+1} - v^k\|^2_{\mathcal{T}+\mathcal{D}_2}\\[8pt]
= & \displaystyle\frac{1}{2}\|v^{k+1} - v^k\|^2_{\mathcal{T} +(1-\eta)\mathcal{D}_2} - \frac{1}{2}\|v^k - v^{k-1}\|^2_{\mathcal{T} +\mathcal{D}_2} - \frac{\eta}{2}\|u^k-u^{k-1}\|^2_{\mathcal{D}_1},
\end{array}
\end{equation*}
where the second inequality is obtained from (\ref{cross_control}) and the fact that $\mathcal{W}_{22}^k\succeq \mathcal{Q}_{22}$.
Therefore, with $\mu_{k+1} = (1-\tau)\sigma\langle \mathcal{B}^*(v^{k+1} - v^k), \mathcal{A}^*u^{k} + \mathcal{B}^*v^{k} - c\rangle$, the cross term can be estimated as
\begin{equation}\label{lim_10}\begin{array}{ll}
&\sigma\langle \mathcal{B}^*(v^{k+1} - v^k), \mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\rangle + \langle \mathcal{Q}_{12}^*(u^{k+1}-u) + \mathcal{Q}_{22}(v^{k+1}-v), v^{k+1} - v^k\rangle\\[8pt]
=& (1-\tau)\sigma\langle \mathcal{B}^*(v^{k+1} - v^k), \mathcal{A}^*u^{k} + \mathcal{B}^*v^{k} - c\rangle - \langle
\mathcal{B}^*(v^{k+1} - v^k),b^{k+1}-b^{k}\rangle\\ [8pt]
& + \langle \mathcal{Q}_{12}^*(u^k-u) + \mathcal{Q}_{22}(v^k-v), v^{k+1} - v^k\rangle
+ \langle \mathcal{Q}_{12}^*(u^{k+1} - u^k) + \mathcal{Q}_{22}(v^{k+1} - v^k), v^{k+1} - v^k\rangle\\[8pt]
\leq &\displaystyle\mu_{k+1} + \frac{1}{2}(\|w^k-w\|^2_{\mathcal{Q}} + \|v^{k+1} - v^k\|^2_{\mathcal{Q}_{22}})
- \frac{1}{2}\|v^{k+1} - v^k\|^2_{\mathcal{T} +(1-\eta)\mathcal{D}_2} + \frac{1}{2}\|v^k - v^{k-1}\|^2_{\mathcal{T} + \mathcal{D}_2}\\[8pt]
&
+\displaystyle \frac{\eta}{2}\|u^k-u^{k-1}\|^2_{\mathcal{D}_1}.
\end{array}
\end{equation}
Finally, by the Cauchy-Schwarz inequality we know that
\begin{equation}\label{lim_11}
\mu_{k+1} \leq\left\{\begin{array}{ll}
\displaystyle \frac{1}{2}(1-\tau)\sigma(\|\mathcal{B}^*(v^{k+1} - v^k)\|^2 + \|\mathcal{A}^*u^k + \mathcal{B}^*v^k - c\|^2), \quad\tau \in(0,1],
\\[8pt]
\displaystyle\frac{1}{2}(\tau-1)\sigma(\tau\|\mathcal{B}^*(v^{k+1} - v^k)\|^2 +\tau^{-1} \|\mathcal{A}^*u^k + \mathcal{B}^*v^k - c\|^2), \quad \tau >1.
\end{array}\right.
\end{equation}
Substituting (\ref{cross_st}), (\ref{lim_10}) and (\ref{lim_11}) into (\ref{main_ineq_3}), we can obtain (\ref{decrease_22}). This completes the proof of part (ii).
\hfill $ \mathchoice\sqr74\sqr74\sqr{2.1}3\sqr{1.5}3 $
\section{Convergence analysis}
With all the preparations given in the previous sections, we can now discuss the main convergence results of our paper.
\subsection{The global convergence}
First we prove that under mild conditions, the iteration sequence $\{(u^k, v^k, x^k)\}$ generated by the majorized ADMM with $\tau\in(0,\frac{1 +\sqrt{5}}{2})$ converges to an optimal solution of problem (\ref{opt}) and its dual.
Let $\bar{w} = (\bar{u}, \bar{v})\in\mathcal{U}\times\mathcal{V}$ be an optimal solution of (\ref{opt}) and $\bar{x}\in\mathcal{X}$ be the corresponding optimal multiplier.
For $k = 0,1,2,\cdots$, define
\begin{equation*}\begin{array}{cc}
u_e^k = u^k - \bar{u}, \quad v_e^k = v^k - \bar{v}, \quad w_e^k = w^k - \bar{w}, \quad x_e^k = x^k - \bar{x}.
\end{array}
\end{equation*}
\begin{theorem}\label{global_convergence}
Suppose that the solution set of (\ref{opt}) is nonempty and Assumption \ref{cq} holds. Assume that $\mathcal{S}$ and $\mathcal{T}$ are chosen such that \[\mathcal{Q}_{11} + \sigma\mathcal{A}\mathcal{A}^*+\mathcal{S}\succ 0, \quad \mathcal{Q}_{22} + \sigma\mathcal{B}\mathcal{B}^*+\mathcal{T}\succ 0.\]
(i) Assume that $\tau\in(0,1]$. If for any $w = \left(\begin{array}{c}u\\v\end{array}\right) \in\mathcal{U}\times\mathcal{V}$, it holds that
\begin{equation}\label{case1_assumption}
\langle w, [\mathcal{Q}+\text{Diag}\,(\mathcal{S}+(1-\tau)\sigma\mathcal{A}\mathcal{A}^*, \mathcal{T} + (1-\tau)\sigma\mathcal{B}\mathcal{B}^*)]w\rangle = 0 \Rightarrow \|u\|\|v\| = 0,
\end{equation}
then the generated sequence $\{(u^k, v^k)\}$ converges to an optimal solution of (\ref{opt}) and $\{x^k\}$ converges to the corresponding optimal multiplier.\\[8pt]
(ii) Assume that $\tau\in (0, \frac{1+\sqrt{5}}{2})$. Under the conditions that
\begin{equation}\label{convergence_2}\begin{array}{cc}
\displaystyle\mathcal{M}: = \frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}-\eta\mathcal{D}_1, ~\mathcal{T} -\eta\mathcal{D}_2 )\succeq 0,\\[8pt]
\displaystyle\frac{1}{4} \mathcal{Q}_{11}+\mathcal{S} + \sigma\mathcal{A}\mathcal{A}^* - \eta\mathcal{D}_1\succ 0,\quad
\frac{1}{4}\mathcal{Q}_{22}+\mathcal{T} + \sigma\mathcal{B}\mathcal{B}^* - \eta\mathcal{D}_2\succ 0
\end{array}
\end{equation}
and for any $w = \left(\begin{array}{c}u\\v\end{array}\right) \in\mathcal{U}\times\mathcal{V}$, it holds that
\begin{equation}\label{case2_assumption}
\langle w, [\mathcal{M}+\sigma\text{Diag}\,(\mathcal{A}\mathcal{A}^*, \mathcal{B}\mathcal{B}^* )]w\rangle = 0 \Rightarrow \|u\|\|v\| = 0,
\end{equation}
the generated sequence $\{(u^k, v^k)\}$ converges to an optimal solution of (\ref{opt}) and $\{x^k\}$ converges to the corresponding optimal multiplier.
\end{theorem}
{\bf Proof.}
{\bf (i)} Let $\tau\in(0,1]$. By letting $(u,v,x) = (\bar{u}, \bar{v}, \bar{x})$ in inequality (\ref{decrease_11}) and the optimality condition (\ref{VI}),
we can obtain that for any $k\geq 0$,
\begin{equation}\label{convergence_1}
\begin{array}{ll}
&\Phi_{k+1}(\bar{u}, \bar{v}, \bar{x}) - \Phi_k(\bar{u}, \bar{v}, \bar{x})\\[8pt]
\leq & -(\Theta_{k+1}+ \sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2
+(1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2 ).
\end{array}
\end{equation}
The above inequality shows that $\{\Phi_{k+1}(\bar{u}, \bar{v}, \bar{x})\}$ is bounded,
which implies that $\{\|x^{k+1}\|\}$, $\{\|w_e^{k+1}\|_{\mathcal{Q}}\}$, $\{\|u_e^{k+1}\|_{S}\}$ and $\{\|v_e^{k+1}\|_{\mathcal{Q}_{22} + \sigma\mathcal{B}\mathcal{B}^*+\mathcal{T} }\}$ are all bounded. From the positive definiteness of $\mathcal{Q}_{22} + \sigma\mathcal{B}\mathcal{B}^*+\mathcal{T}$, we can see that $\{\|v_e^{k+1}\|\}$ is bounded. By using the inequalities
\begin{equation*}\begin{array}{ll}
\|\mathcal{A}^*u_e^{k+1}\|&\leq \|\mathcal{A}^*u_e^{k+1} + \mathcal{B}^*v_e^{k+1}\| + \|\mathcal{B}^*v_e^{k+1}\|\\[8pt]
& \leq \tau\sigma(\|x_e^{k+1}\| + \|x_e^k\|) + \|\mathcal{B}^*v_e^{k+1}\|,\\[8pt]
\|u^{k+1}_e\|_{\mathcal{Q}_{11}} & \leq \|w_e^{k+1}\|_{\mathcal{Q}} + \|v_e^{k+1}\|_{\mathcal{Q}_{22}},
\end{array}
\end{equation*}
we know that the sequence $\{\|u_e^{k+1}\|_{\sigma\mathcal{A}\mathcal{A}^* + \mathcal{Q}_{11}}\}$ is also bounded. Therefore, $\{\|u_e^{k+1}\|_{\mathcal{Q}_{11} + \sigma\mathcal{A}\mathcal{A}^*+\mathcal{S}} \}$ is bounded. By the positive definiteness of $\mathcal{Q}_{11} + \sigma\mathcal{A}\mathcal{A}^*+\mathcal{S}$, we know that $\{\|u_e^{k+1}\|\}$ is bounded. On the whole, the sequence $\{(u^k,v^k,x^k)\}$ is bounded. Thus, there exists a subsequence $\{(u^{k_i},v^{k_i}, x^{k_i})\}$ converging to a cluster point, say $(u^{\infty},v^{\infty},x^{\infty})$.
Next we will prove that $(u^{\infty},v^{\infty})$ is optimal to (\ref{opt}) and $x^\infty$ is the corresponding optimal multiplier. The inequality (\ref{convergence_1}) also implies that $$\displaystyle \lim_{k\to\infty} (\Theta_{k+1} + \sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^k - c\|^2
+(1-\tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2) = 0,$$ which is equivalent to
\begin{equation}\label{lim_1}\begin{array}{ll}
\displaystyle \lim_{k\to\infty} \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k} - c\|= 0, \\[8pt]
\displaystyle\lim_{k\to\infty} (1-\tau)\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|= 0, \\[8pt]
\displaystyle \lim_{k\to\infty} \|w^{k+1} - w^k\|_{\mathcal{Q}+\text{Diag}\,(\mathcal{S}, \mathcal{T})} = 0.\\[8pt]
\end{array}
\end{equation}
For $\tau\in(0,1)$, since $\displaystyle \lim_{k\to\infty} \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|=0$, by using (\ref{lim_1}) we see that
\begin{equation*}\begin{array}{ll}
\displaystyle \lim_{k\to\infty} \|\mathcal{A}^*(u^{k+1} - u^k)\|\leq
\displaystyle \lim_{k\to\infty} (\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k} - c\| + \|\mathcal{A}^*u^{k} + \mathcal{B}^*v^{k} - c\| ) = 0,\\[8pt]
\displaystyle \lim_{k\to\infty} \|\mathcal{B}^*(v^{k+1} - v^k)\|\leq
\displaystyle \lim_{k\to\infty}( \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\| + \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k} - c\|) = 0,
\end{array}
\end{equation*}
which implies $\displaystyle \lim_{k\to\infty} \|w^{k+1} - w^k\|_{\mathcal{Q}+\text{Diag}\,(\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*, \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*)} = 0$.
Therefore, for $\tau \in (0,1]$, we know that $\displaystyle \lim_{k\to\infty} \|w^{k+1} - w^k\|_{\mathcal{Q}+\text{Diag}\,(\mathcal{S}+(1-\tau)\sigma\mathcal{A}\mathcal{A}^*, \mathcal{T} + (1-\tau)\sigma\mathcal{B}\mathcal{B}^*)} = 0$. By condition (\ref{case1_assumption}) we can see that this implies either $\displaystyle \lim_{k\to\infty}\|u^{k+1} - u^k\| = 0$ or $\displaystyle \lim_{k\to\infty}\|v^{k+1} - v^k\| = 0$.
Without loss of generality we assume that
$\displaystyle \lim_{k\to\infty} \|v^{k+1} - v^k\| = 0$. Thus,
\begin{equation}\label{Fineq}
\begin{array}{ll}
\displaystyle \lim_{k\to\infty} \|\mathcal{A}^*(u^{k+1} - u^k)\|& \leq
\displaystyle \lim_{k\to\infty}( \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k} - c\| + \|\mathcal{A}^*u^{k} + \mathcal{B}^*v^{k-1} - c\| + \|\mathcal{B}^*(v^k-v^{k-1})\| )\\[8pt]
& = 0, \\[8pt]
\displaystyle \lim_{k\to\infty} \|u^{k+1} - u^k\|_{\mathcal{Q}_{11}}& \leq
\displaystyle \lim_{k\to\infty}( \|w^{k+1} - w^k\|_{\mathcal{Q}} + \|v^{k+1} - v^k\|_{\mathcal{Q}_{22}} ) = 0.
\end{array}
\end{equation}
Therefore, $\displaystyle \lim_{k\to\infty} \|u^{k+1} - u^k\|_{\mathcal{Q}_{11}+\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*} = 0$. This implies $\displaystyle \lim_{k\to\infty} \|u^{k+1} - u^k\|=0$ by the positive definiteness of $\mathcal{Q}_{11}+\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*$.
Now taking limits on both sides of (\ref{opt_1}) along the subsequence $\{(u^{k_i},v^{k_i},x^{k_i})\}$, and by using the closedness of the graphs of $\partial p$, $\partial q$ and the continuity of $\nabla \phi$, we obtain
\begin{equation*}\left\{\begin{array}{ll}
0\in F(u^\infty, v^\infty,x^\infty),\\[8pt]
\mathcal{A}^*u^\infty + \mathcal{B}^*v^\infty = c.
\end{array}\right.
\end{equation*}
This indicates that $(u^\infty,v^\infty)$ is an optimal solution to (\ref{opt}) and $x^\infty$ is the corresponding optimal multiplier.
Since $(u^\infty, v^\infty, x^{\infty})$ satisfies (\ref{kkt}), all the above arguments involving $(\bar{u},\bar{v},\bar{x})$ can be replaced by
$(u^\infty, v^\infty, x^{\infty})$. Thus the subsequence $\{\Phi_{k_i}(u^\infty, v^\infty, x^{\infty}) \}$ converges to $0$ as $k_i\to \infty$. Since $\{\Phi_{k_i}(u^\infty, v^\infty, x^{\infty}) \}$ is non-increasing, we obtain that
\begin{equation}\label{lim_phi}\begin{array}{ll}
\displaystyle\lim_{k\to\infty} \Phi_{k+1}(u^\infty, v^\infty, x^{\infty})= &\displaystyle\lim_{k\to\infty}~ (\tau\sigma)^{-1}\|x^{k+1} - x^\infty\|^2 + \|v^{k+1} - v^{\infty}\|^2_{\sigma\mathcal{B}\mathcal{B}^* + \mathcal{T}+\mathcal{Q}_{22}}+ \|u^{k+1} - u^{\infty}\|^2_{\mathcal{S}}\\[8pt]
& \quad\quad\quad +\|w^{k+1} - w^{\infty}\|_{\mathcal{Q}}^2 = 0.
\end{array}
\end{equation}
From this we can immediately get $\displaystyle\lim_{k\to\infty} x^{k+1} = x^\infty$ and $\displaystyle\lim_{k\to\infty} v^{k+1} = v^\infty$. Similar to inequality (\ref{Fineq}) we have that
$\displaystyle \lim_{k\to\infty} \sigma\|\mathcal{A}^*(u^{k+1} - u^{\infty})\| = 0$ and $\displaystyle \lim_{k\to\infty} \|u^{k+1} - u^{\infty}\|_{\mathcal{Q}_{11}} = 0$, which, together with (\ref{lim_phi}), imply that $\displaystyle \lim_{k\to\infty} \|u^{k+1} - u^{\infty}\| = 0$ by the positive definiteness of $\mathcal{Q}_{11}+\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*$.
Therefore, the whole sequence $\{(u^k,v^k,x^k)\}$ converges to $(u^{\infty},v^\infty, x^\infty)$, the unique limit of the sequence. This completes the proof for the first case.
\bigskip
\noindent
{\bf (ii)} From the inequality (\ref{decrease_22}) and the optimality condition (\ref{VI})
we know that for any $k\geq 1$,
\begin{equation}\label{decrease_2}
\begin{array}{ll}
&
(\Psi_{k+1}(\bar{u}, \bar{v}, \bar{x}) + \Xi_{k+1}) - (\Psi_k(\bar{u}, \bar{v}, \bar{x}) +\Xi_k)\\[8pt]
\leq &
-(\Gamma_{k+1} +\min(1,1+\tau^{-1} - \tau)\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2).
\end{array}
\end{equation}
By the assumptions $\tau\in (0, \frac{1+\sqrt{5}}{2})$ and $\mathcal{M}\succeq 0$, we can obtain that $\Gamma_{k+1}\geq 0$ and $\min(1,1+\tau^{-1} - \tau)\geq 0$. Then both $\{\Psi_{k+1}(\bar{u}, \bar{v}, \bar{x})\}$ and $\{\Xi_{k+1}\}$ are bounded. Thus, by a similar approach to case (i), we see that the sequence $\{(u^k,v^k,x^k)\}$ is bounded. Therefore, there exists a subsequence $\{(u^{k_i},v^{k_i}, x^{k_i})\}$ that converges to a cluster point, say $(u^{\infty},v^{\infty},x^{\infty})$. Next we will prove that
$(u^{\infty},v^{\infty})$ is optimal to (\ref{opt}) and $x^\infty$ is the corresponding optimal multiplier. The inequality (\ref{decrease_2}) also implies that
\begin{equation*}\begin{array}{ll}
\displaystyle \lim_{k\to\infty} \|x^{k+1} - x^k\| = \displaystyle\lim_{k\to\infty}(\tau\sigma)^{-1}\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|= 0, \\[8pt]
\displaystyle \lim_{k\to\infty} \|w^{k+1} - w^k\|_{\mathcal{M}} = 0,\quad
\displaystyle \lim_{k\to\infty}\|\mathcal{B}(v^{k+1} - v^k)\| = 0.
\end{array}
\end{equation*}
By the relationship
\begin{equation*}\begin{array}{ll}
\displaystyle\lim_{k\to\infty} \|\mathcal{A}^*(u^{k+1}-u^k)\| & \leq \displaystyle\lim_{k\to\infty} (\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1}-c\| + \|\mathcal{A}^*u^k + \mathcal{B}^*v^k - c\| + \|\mathcal{B}^*(v^{k+1} - v^k)\|)\\[8pt]
& = 0,
\end{array}
\end{equation*}
we can further get $\displaystyle \lim_{k\to\infty} \|w^{k+1} - w^k\|_{\mathcal{M} + \text{Diag}\,(\sigma\mathcal{A}\mathcal{A}^*, \sigma\mathcal{B}\mathcal{B}^*)}= 0$. Thus, by the condition (\ref{case2_assumption}), we can get that either $\displaystyle\lim_{k\to\infty} \|u^{k+1} - u^k\| = 0$
or $\displaystyle\lim_{k\to\infty} \|v^{k+1} - v^k\| = 0$. Again similar to case (i), we can see that in fact both of them would hold by the positive definiteness of $\frac{1}{4} \mathcal{Q}_{11}+\mathcal{S} + \sigma\mathcal{A}\mathcal{A}^* - \eta\mathcal{D}_1$ and $\frac{1}{4} \mathcal{Q}_{22}+\mathcal{T} + \sigma\mathcal{B}\mathcal{B}^* - \eta\mathcal{D}_{22}$. The remaining proof about the convergence of the whole sequence $\{(u^k,v^k,x^k)\}$ follows exactly the same as in case (i).
This completes the proof for the second case.
\hfill $ \mathchoice\sqr74\sqr74\sqr{2.1}3\sqr{1.5}3 $
\begin{remark}
In Theorem \ref{global_convergence}, for $\tau\in(0,1]$, a sufficient condition for the convergence is
\[\mathcal{Q}+\text{Diag}\,(\mathcal{S}+(1-\tau)\sigma\mathcal{A}\mathcal{A}^*,
\mathcal{T} + (1-\tau)\sigma\mathcal{B}\mathcal{B}^*)\succ 0,\]
and for $\tau\in [1, \frac{1+\sqrt{5}}{2})$, a sufficient condition for the convergence is
\begin{equation*}\begin{array}{ll}
\displaystyle\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}-\eta\mathcal{D}_1, ~\mathcal{T} -\eta\mathcal{D}_2 )\succeq 0,\quad
\frac{1}{4}\mathcal{Q}+\text{Diag}\,(\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*-\eta\mathcal{D}_1, \mathcal{T} +\sigma\mathcal{B}\mathcal{B}^* -\eta\mathcal{D}_2)\succ 0.
\end{array}
\end{equation*}
\end{remark}
\begin{remark}\label{Q_plus_separable}
An interesting application of Theorem \ref{global_convergence} is for the linearly constrained convex optimization problem with a quadratically coupled objective function of the form \[\phi(w) = \frac{1}{2}\langle w, \widetilde{\mathcal{Q}}w\rangle + f(u) + g(v),\] where $\widetilde{\mathcal{Q}}:\mathcal{U}\times\mathcal{V}\to\mathcal{U}\times\mathcal{V}$ is a self-adjoint positive semidefinite linear operator, $f: \mathcal{U}\to (-\infty, \infty)$ and $g: \mathcal{V}\to (-\infty, \infty)$ are two convex smooth functions with Lipschitz continuous gradients.
In this case, there exist four self-adjoint positive semidefinite operators $\Sigma_f, \widehat{\Sigma}_f: \mathcal{U}\to\mathcal{U}$ and $\Sigma_g, \widehat{\Sigma}_g:\mathcal{V}\to\mathcal{V}$
such that
\begin{equation*}
\Sigma_f\preceq \xi\preceq \widehat{\Sigma}_f \quad\forall \xi\in\partial^2 f(u), \,u\in\mathcal{U}
\end{equation*}
and
\begin{equation*}
\Sigma_g\preceq \zeta\preceq \widehat{\Sigma}_g \quad\forall \zeta\in\partial^2 g(v), \, v\in\mathcal{V},
\end{equation*}
where $\partial^2 f$ and $\partial^2 g$ are defined in (\ref{gen_Hessian}).
Then by letting $\mathcal{Q} = \widetilde{\mathcal{Q}} + \text{Diag}\,(\Sigma_f, \Sigma_g)$ in (\ref{phi_convex}) and $\mathcal{Q} + \mathcal{H} = \widetilde{\mathcal{Q}} + \text{Diag}\,(\widehat{\Sigma}_f, \widehat{\Sigma}_g)$ in (\ref{phi_majorize}), we have $\eta = 0$ in (\ref{cross_control}). This implies that $\mathcal{M}\succeq 0$ always holds in (\ref{convergence_2}). Therefore, for $\tau\in(0,\frac{1+\sqrt{5}}{2})$,
the conditions for the convergence can be equivalently written as
\begin{equation}\label{QP_assumption_1}\begin{array}{ll}
\widetilde{\mathcal{Q}}_{11} +\Sigma_f+ \mathcal{S} + \sigma\mathcal{A}\mathcal{A}^*\succ 0, ~\widetilde{\mathcal{Q}}_{22}+\Sigma_g + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*\succ 0
\end{array}
\end{equation}
and
\begin{equation}\label{QP_assumption_2}
\langle w, [\widetilde{\mathcal{Q}}+\text{Diag}\,(\Sigma_f + \mathcal{S}+\sigma\mathcal{A}\mathcal{A}^*, \Sigma_g + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*)]w\rangle = 0 \Rightarrow \|u\|\|v\| = 0.
\end{equation}
A sufficient condition for ensuring (\ref{QP_assumption_1}) and (\ref{QP_assumption_2}) to hold is
\begin{equation}
\widetilde{\mathcal{Q}} + \text{Diag}\,(\Sigma_f + \mathcal{S} + \sigma\mathcal{A}\mathcal{A}^*, \Sigma_g + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*)\succ 0.
\end{equation}
If $\widetilde{\mathcal{Q}} = 0$, i.e., if the objective function of the original problem (\ref{opt}) is separable, we will recover the convergence conditions given in \cite{LiSunToh_indefinite2014} for a majorized ADMM with semi-proximal terms.
\end{remark}
\subsection{The non-ergodic iteration complexity for general coupled objective functions }
In this section, we will present the non-ergodic iteration complexity for the majorized ADMM in terms of the KKT optimality condition.
\begin{theorem}\label{nonerg_complexity}
Suppose that the solution set of (\ref{opt}) is nonempty and Assumption \ref{cq} holds. Assume that
one of the following conditions holds:\\[8pt]
(i) $\tau\in(0,1]$ and
$ \mathcal{O}_1: = \displaystyle\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}+(1-\tau)\sigma\mathcal{A}\mathcal{A}^*, \mathcal{T} + (1-\tau)\sigma\mathcal{B}\mathcal{B}^*)\succ 0$;\\[8pt]
(ii) $\tau\in (0,\frac{1+\sqrt{5}}{2})$, $\displaystyle\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}- \eta\mathcal{D}_1, \mathcal{T} - \eta\mathcal{D}_2)\succeq 0$ and
$\mathcal{O}_2: = \displaystyle\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}+\sigma\mathcal{A}\mathcal{A}^* - \eta\mathcal{D}_1, \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^* - \eta\mathcal{D}_2)\succ 0.$\\[8pt]
Then there exists a constant $C$ only depending on the initial point and the optimal solution set such that
the sequence $\{(u^k, v^k,x^k)\}$ generated by the majorized ADMM satisfies that for $k\geq 1$,
\begin{equation}\label{nonerg_main}
\displaystyle\min_{1\leq i\leq k} \{\text{dist}^2(0, F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2\} \leq C/k.
\end{equation}
{
Furthermore, for the limiting case we have that
\begin{equation}\label{nonerg_limit}
\displaystyle\lim_{k\to\infty} k(\min_{1\leq i\leq k} \{\text{dist}^2(0, F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2 \})= 0.
\end{equation}
}
\end{theorem}
{\bf Proof.}
From the optimality condition for $(u^{k+1}, v^{k+1})$, we know that
\begin{equation*}\begin{array}{ll}
&\left(\begin{array}{cc}
-(1-\tau)\sigma\mathcal{A}(\mathcal{A}^*u^{k+1} +\mathcal{B}^*v^{k+1} - c) - \sigma\mathcal{A}\mathcal{B}^*(v^{k} - v^{k+1}) - \mathcal{S}(u^{k+1} - u^k) + \mathcal{Q}_{12}(v^{k+1}- v^k)\\
-(1-\tau)\sigma\mathcal{B}(\mathcal{A}^*u^{k+1} +\mathcal{B}^*v^{k+1} - c) - \mathcal{T}(v^{k+1} - v^k)
\end{array}\right) \\[10pt]
&
- (\mathcal{Q}+\mathcal{H})(w^{k+1} - w^k)+ \nabla \phi(w^{k+1}) - \nabla \phi(w^k)\\[8pt]
\in & F(u^{k+1}, v^{k+1}, x^{k+1}).
\end{array}
\end{equation*}
Therefore, we can obtain that
\begin{equation}\label{nonerg_1}\begin{array}{ll}
&\text{dist}^2(0, F(u^{k+1}, v^{k+1}, x^{k+1}))
+ \|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1}-c\|^2\\[8pt]
\leq & 5\|\sigma\mathcal{A}\mathcal{B}^*(v^{k+1} - v^k)\|^2 + 5(1-\tau)^2\sigma^2(\|\mathcal{A}\|^2+\|\mathcal{B}\|^2)\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2\\[8pt]
&
+ 5\| (\mathcal{Q} + \mathcal{H})(w^{k+1} - w^k)-\nabla\phi(w^{k+1}) +\nabla \phi(w^k)\|^2 + 5\|\mathcal{Q}_{12}(v^{k+1} - v^k)\|^2
+ 5\|\mathcal{T}(v^{k+1} - v^k)\|^2 \\[8pt]
&
+ 5\|\mathcal{S}(u^{k+1} - u^k)\|^2+\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2\\[8pt]
\leq & 5\sigma\|\mathcal{A}\|^2\|v^{k+1} - v^k\|_{\sigma\mathcal{B}\mathcal{B}^*}^2
+ (5(1-\tau)^2\sigma^2(\|\mathcal{A}\|^2 + \|\mathcal{B}\|^2)+1)\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1}-c\|^2 \\[8pt]
& + 5\|\sqrt{\mathcal{Q}_{12}^*\mathcal{Q}_{12}}\|\|v^{k+1} - v^k\|_{\sqrt{\mathcal{Q}_{12}^*\mathcal{Q}_{12}}}^2
+ 5\|\mathcal{H}\|\|w^{k+1} - w^k\|_{\mathcal{H}}^2 + 5\|\mathcal{S}\|\|u^{k+1} - u^k\|^2_{\mathcal{S}}\\[8pt]
&
+ 5\|\mathcal{T}\|\|v^{k+1} - v^k\|^2_{\mathcal{T}}\\[8pt]
\leq & C_1\|w^{k+1} - w^k\|^2_{\widehat{\mathcal{O}}} + C_2\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1}-c\|^2,
\end{array}
\end{equation}
where $$C_1 = 5\max(\sigma\|\mathcal{A}\|^2,\|\sqrt{\mathcal{Q}_{12}^*\mathcal{Q}_{12}}\|, \|\mathcal{H}\|,\|\mathcal{S}\|,\|\mathcal{T}\|), \quad C_2 = 5(1-\tau)^2\sigma^2(\|\mathcal{A}\|^2 + \|\mathcal{B}\|^2)+1,$$
$$\widehat{\mathcal{O}} = \mathcal{H} + \text{Diag}\,(\mathcal{S}, \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^* + \sqrt{\mathcal{Q}_{12}^*\mathcal{Q}_{12}})$$ and the second inequality comes from the fact that there exists some $\mathcal{W}^k\in\text{conv}\{\partial^2\phi([w^{k-1}, w^k])\}$ such that
\begin{equation*}\begin{array}{ll}
& \|(\mathcal{Q} + \mathcal{H})(w^{k+1} - w^k)-\nabla \phi(w^{k+1}) + \nabla \phi(w^k)\|^2\\[8pt]
= & \|(\mathcal{Q} + \mathcal{H}-\mathcal{W}^k)(w^{k+1} - w^k)\|^2
\leq \|\mathcal{H}\|\|w^{k+1} - w^k\|^2_{\mathcal{H}}.
\end{array}
\end{equation*}
Next we will estimate the upper bounds for $\|w^{k+1} - w^k\|^2_{\widehat{\mathcal{O}}}$ and $\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2$ by only involving the initial point and the optimal solution set under the two different conditions.
First, assume condition (i) holds. For $\tau\in(0,1]$, by using (\ref{convergence_1}) in the proof of Theorem \ref{global_convergence},
we have that for $i\geq 1$,
\begin{equation*}\begin{array}{ll}
&\|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}, \mathcal{T})} + \sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^i - c\|^2 + (1-\tau)\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} -c\|^2 \\[8pt]
\leq & \Phi_i(\bar{u}, \bar{v}, \bar{x}) - \Phi_{i+1}(\bar{u}, \bar{v}, \bar{x}),
\end{array}
\end{equation*}
which, implies that,
\begin{equation*}\begin{array}{ll}
& \displaystyle\sum_{i=1}^k (\|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}, \mathcal{T})} + \sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^i - c\|^2 + (1-\tau)\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} -c\|^2 \\[8pt]
\leq & \Phi_1(\bar{u}, \bar{v}, \bar{x}) - \Phi_{k+1}(\bar{u}, \bar{v}, \bar{x})\leq \Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{array}
\end{equation*}
This shows that
\begin{equation}\label{nonerg_7}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{Diag}(\mathcal{S}, \mathcal{T})} \leq \Phi_1(\bar{u}, \bar{v}, \bar{x}),\\[8pt]
\displaystyle\sum_{i=1}^k \sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i} -c\|^2 \leq \Phi_1(\bar{u}, \bar{v}, \bar{x}),\\[8pt]
\displaystyle\sum_{i=1}^k (1-\tau)\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 \leq \Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{array}
\end{equation}
From the above three inequalities we can also get that
\begin{equation*}\begin{array}{ll}
(1-\tau)\displaystyle\sum_{i=1}^k \|u^{i+1} - u^i\|^2_{\sigma\mathcal{A}\mathcal{A}^*}& \leq (1-\tau)\displaystyle\sum_{i=1}^k (2\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^i - c\|^2 + 2\sigma\|\mathcal{A}^*u^{i} + \mathcal{B}^*v^{i} - c\|^2) \\[8pt]
&\leq 2(2-\tau)\Phi_1(\bar{u}, \bar{v}, \bar{x}),\\[8pt]
(1-\tau)\displaystyle\sum_{i=1}^k \|v^{i+1} - v^i\|^2_{\sigma\mathcal{B}\mathcal{B}^*}& \leq (1-\tau)\displaystyle\sum_{i=1}^k (2\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^i - c\|^2 + 2\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2) \\[8pt]
&
\leq 2(2-\tau)\Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{array}
\end{equation*}
With the notation of operator $\mathcal{O}_1$ we have that
\begin{equation}\label{nonerg_3}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\mathcal{O}_1} &=\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{Diag}(\mathcal{S}, \mathcal{T})} + \displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{(1-\tau)\text{Diag}(\sigma\mathcal{A}\mathcal{A}^*, \mathcal{B}\mathcal{B}^*)}\\[8pt]
& \leq (9-4\tau)\Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{array}
\end{equation}
If $\tau\in(0,1)$, we further have that
\begin{equation}\label{nonerg_8}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 \leq (1-\tau)^{-1}\sigma^{-1}\Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{array}
\end{equation}
If $\tau = 1$, by the condition that $\mathcal{O}_1 = \displaystyle\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}, \mathcal{T})\succ 0$, we have that
\begin{equation}\label{nonerg_9}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 & \leq \displaystyle\sum_{i=1}^k (2\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i} - c\|^2 + 2\|v^{i+1} - v^i\|^2_{\mathcal{B}\mathcal{B}^*})\\[8pt]
& \leq \displaystyle\sum_{i=1}^k (2\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i} - c\|^2 \\[8pt]
&\quad \quad\quad+ 2\|\mathcal{O}_1^{-\frac{1}{2}}\text{Diag}\,(0,\mathcal{\mathcal{B}\mathcal{B}^*})\mathcal{O}_1^{-\frac{1}{2}}\|\|w^{i+1} - w^i\|^2_{\mathcal{O}_1})\\[8pt]
&
\leq (2\sigma^{-1} + (18-8\tau)\|\mathcal{O}_1^{-\frac{1}{2}}\text{Diag}\,(0,\mathcal{\mathcal{B}\mathcal{B}^*})\mathcal{O}_1^{-\frac{1}{2}}\|)\Phi_1(\bar{u}, \bar{v}, \bar{x}),
\end{array}
\end{equation}
where the second inequality is obtained by the fact that for any $\xi$, a self-adjoint positive definite operator $\mathcal{G}$ with square root $\mathcal{G}^{\frac{1}{2}}$ and a self-adjoint positive semidefinite operator $\widehat{G}$ defined in the same Hilbert space, it always holds that
\begin{equation*}\begin{array}{ll}
\|\xi\|^2_{\widehat{\mathcal{G}}} = \langle \xi, \widehat{\mathcal{G}}\xi\rangle = \langle \xi, (\mathcal{G}^{\frac{1}{2}}\mathcal{G}^{-\frac{1}{2}})\widehat{\mathcal{G}}(\mathcal{G}^{-\frac{1}{2}}\mathcal{G}^{\frac{1}{2}})\xi\rangle
& = \langle \mathcal{G}^{\frac{1}{2}}\xi, (\mathcal{G}^{-\frac{1}{2}}\widehat{\mathcal{G}}\mathcal{G}^{-\frac{1}{2}})\mathcal{G}^{\frac{1}{2}}\xi\rangle \\[8pt]
&
\leq \|\mathcal{G}^{-\frac{1}{2}}\widehat{\mathcal{G}}\mathcal{G}^{-\frac{1}{2}}\|\|\xi\|^2_{\mathcal{G}}.
\end{array}
\end{equation*}
Therefore, by using (\ref{nonerg_1}), (\ref{nonerg_3}) and the positive definiteness of operator $\mathcal{O}_1$, we know that
\begin{equation*}\begin{array}{ll}
&\displaystyle\min_{1\leq i\leq k} \{\text{dist}^2(0, F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2\}\\[8pt]
\leq & (\displaystyle\sum_{i=1}^k(\text{dist}^2(0, F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2))/k
\leq C\Phi_1(\bar{u}, \bar{v}, \bar{x})/k,
\end{array}
\end{equation*}
where
\begin{equation*}C =\left\{\begin{array}{ll}
C_1(9-4\tau)\|\mathcal{O}_1^{-\frac{1}{2}}\widehat{\mathcal{O}}\mathcal{O}_1^{-\frac{1}{2}}\| + C_2(1-\tau)^{-1}\sigma^{-1}, \quad\tau\in(0,1),\\[8pt]
C_1(9-4\tau)\|\mathcal{O}_1^{-\frac{1}{2}}\widehat{\mathcal{O}}\mathcal{O}_1^{-\frac{1}{2}}\| + C_2(2\sigma^{-1} + (18-8\tau)\|\mathcal{O}_1^{-\frac{1}{2}}\text{Diag}\,(0,\mathcal{\mathcal{B}\mathcal{B}^*})\mathcal{O}_1^{-\frac{1}{2}}\|)
,\quad \tau = 1.
\end{array}\right.
\end{equation*}
To prove the limiting case (\ref{nonerg_limit}), by using inequalities (\ref{nonerg_3}), (\ref{nonerg_8}), (\ref{nonerg_9}) and \cite[Lemma 2.1]{LiSunToh_indefinite2014}, we have that
\begin{equation*}
\min_{1\leq i\leq k}\|w^{i+1} - w^i\|^2_{\mathcal{O}_1} = o(1/k),\quad
\min_{1\leq i\leq k}\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 = o(1/k),
\end{equation*}
which, together with (\ref{nonerg_1}), imply that
\begin{equation*}\begin{array}{ll}
&\displaystyle\lim_{k\to\infty} k (\min_{1\leq i\leq k} \{\text{dist}^2(0, F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2\})\\[8pt]
\leq &\displaystyle\lim_{k\to\infty} k (\min_{1\leq i\leq k} \{C_1\|\mathcal{O}_1^{-\frac{1}{2}}\widehat{\mathcal{O}}\mathcal{O}_1^{-\frac{1}{2}}\|\|w^{i+1} - w^i\|^2_{\mathcal{O}_1} + C_2\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2\}) = 0.
\end{array}
\end{equation*}
Now, we complete the proof of the conclusions under condition (i). Next, we consider the case under condition (ii). The proof of this part is similar to the case under condition (i). For $\tau\in(0,\frac{1+\sqrt{5}}{2})$,
let $\rho(\tau) = \min(\tau,1+\tau-\tau^2)$.
We know
from (\ref{convergence_2}) that for $\tau\in(0,\frac{1+\sqrt{5}}{2})$ and any $k\geq 1$,
\begin{eqnarray*}
&&\hspace{-0.7cm}
\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{diag}\,(\mathcal{S} - \eta\mathcal{D}_1,\mathcal{T} - \eta\mathcal{D}_2)}
+ \rho(\tau) \|v^{i+1} - v^i\|^2_{\sigma
\mathcal{B}\mathcal{B}^*}+ \frac{\rho(\tau)}{\tau}\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2
\\[5pt]
&\leq & (\Psi_1(\bar u, \bar v, \bar x) +\Xi_1) - (\Psi_{k+1}(\bar u, \bar v, \bar x) + \Xi_{k+1})\leq \Psi_1(\bar u, \bar v, \bar x) + \Xi_1.
\end{eqnarray*}
Thus by the positive semidefiniteness of $\frac{1}{4}\mathcal{Q} + \text{Diag}\,(\mathcal{S}- \eta\mathcal{D}_1, \mathcal{T} - \eta\mathcal{D}_2)$, we can get that
\begin{equation}\label{nonerg_4}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{diag}\,(\mathcal{S} - \eta\mathcal{D}_1,\mathcal{T} - \eta\mathcal{D}_2)} \leq \Psi_1(\bar u, \bar v, \bar x) + \Xi_1,\\[8pt]
\displaystyle\sum_{i=1}^k \|v^{i+1} - v^i\|^2_{\sigma\mathcal{B}\mathcal{B}^*} \leq (\Psi_1(\bar u, \bar v, \bar x)+\Xi_1)/\rho(\tau),\\[8pt]
\displaystyle\sum_{i=1}^k \sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 \leq \tau (\Psi_1(\bar u, \bar v, \bar x)+\Xi_1)/\rho(\tau),
\end{array}
\end{equation}
which, implies that,
\begin{equation}\label{nonerg_5}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|u^{i+1} - u^i\|^2_{\sigma\mathcal{A}\mathcal{A}^*} & \leq \displaystyle\sum_{i=1}^k(3\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2 + 3\sigma\|\mathcal{A}^*u^i + \mathcal{B}^*v^i - c\|^2 + 3\|v^{i+1} - v^i\|_{\sigma\mathcal{B}\mathcal{B}^*}^2)\\[8pt]
&
\leq (6\tau+2)(\Psi_1(\bar u, \bar v, \bar x)+\Xi_1)/\rho(\tau).
\end{array}
\end{equation}
Combining (\ref{nonerg_4}) and (\ref{nonerg_5}) one can find that
\begin{equation}\label{nonerg_6}\begin{array}{ll}
\displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\mathcal{O}_2}& = \displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\frac{1}{4}\mathcal{Q} + \text{diag}\,(\mathcal{S} - \eta\mathcal{D}_1,\mathcal{T} - \eta\mathcal{D}_2)} + \displaystyle\sum_{i=1}^k \|w^{i+1} - w^i\|^2_{\text{Diag}\,(\sigma\mathcal{A}\mathcal{A}^*, \sigma\mathcal{B}\mathcal{B}^*)}\\[8pt]
& \leq (1+ (6\tau+3)/\rho(\tau))(\Psi_1(\bar u, \bar v, \bar x) +\Xi_1).
\end{array}
\end{equation}
Therefore, by using (\ref{nonerg_1}), (\ref{nonerg_4}), (\ref{nonerg_6}), and recalling the positive definiteness of operator $\mathcal{O}_2$, we finally have that
\begin{equation*}\begin{array}{ll}
& \displaystyle\min_{1\leq i\leq k} \{\text{dist}^2(0,F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2\}\\[8pt]
\leq & \big(\displaystyle\sum_{i=1}^k (\text{dist}^2(0,F(u^{i+1}, v^{i+1}, x^{i+1}))
+ \|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1}-c\|^2\big)/k\\[8pt]
\leq & C'(\Psi_1(\bar u, \bar v, \bar x)+\Xi_1)/k,
\end{array}
\end{equation*}
where $C' = C_1\|\mathcal{O}_2^{-\frac{1}{2}}\widehat{\mathcal{O}}\mathcal{O}_2^{-\frac{1}{2}}\|(1+ (6\tau+3)/\rho(\tau))+ C_2\sigma^{-1}\tau/\rho(\tau)$.
The limiting property (\ref{nonerg_limit}) can be derived in the same way as for the case under condition (i).
This completes the proof of Theorem \ref{nonerg_complexity}.
\hfill $ \mathchoice\sqr74\sqr74\sqr{2.1}3\sqr{1.5}3 $
\begin{remark}\label{nonerg_remark}
Theorem \ref{nonerg_complexity} gives the non-ergodic complexity of the
KKT optimality condition, which
does not seem to be known even for the classic ADMM with separable objective functions. For the latter, Davis and Yin~\cite{Yin2014} provided related non-ergodic iteration complexity results for the primal feasibility and the objective functions and constructed an interesting example to show
for any $\alpha>1/2$, there exists an initial point such that the sequence $\{(u^k,v^k)\}$ generated by the classic ADMM $(\tau = 1)$
satisfies $\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\| \geq 1/k^{\alpha}$.
This implies that the non-ergodic iteration complexity results presented in
Theorem \ref{nonerg_complexity}
in terms of the KKT optimality condition may be optimal.
\end{remark}
\subsection{The ergodic iteration complexity for general coupled objective functions}
In this section, we will discuss the ergodic iteration complexity of the majorized ADMM for solving problem (\ref{opt}). For $k = 1,2,\cdots,$ denote
\begin{equation*}\begin{array}{ll}
\hat{x}^k = \displaystyle\frac{1}{k}\displaystyle\sum_{i=1}^k \tilde{x}^{i+1}, \quad
\hat{u}^k = \frac{1}{k}\displaystyle\sum_{i=1}^k u^{i+1}, \quad \hat{v}^k = \frac{1}{k}\displaystyle\sum_{i=1}^k v^{i+1}, \quad \hat{w}^k = (\hat{u}^k, \hat{v}^k)
\end{array}
\end{equation*}
and
\begin{equation*}\left\{\begin{array}{ll}
\Lambda_{k+1} = \|u^{k+1}_e\|^2_{\mathcal{D}_1 + \mathcal{S}} +\|v^{k+1}_e\|^2_{\mathcal{D}_2 +\mathcal{T}+\mathcal{Q}_{22}+\sigma\mathcal{B}\mathcal{B}^*} +(\tau\sigma)^{-1}\|x^{k+1}\|^2, \\[8pt]
\overline\Lambda_{k+1} = \Lambda_{k+1} + \Xi_{k+1} + \|w^{k+1}_e\|^2_{\mathcal{Q}} + \max(1-\tau, 1-\tau^{-1})\sigma\|\mathcal{A}^*u^{k+1} + \mathcal{B}^*v^{k+1} - c\|^2.
\end{array}\right.
\end{equation*}
\begin{theorem}\label{erg_main}
Suppose that
$\mathcal{S}$ and $\mathcal{T}$ are chosen such that \[\mathcal{Q}_{11} + \sigma\mathcal{A}\mathcal{A}^*+\mathcal{S}\succ 0, \quad \mathcal{Q}_{22} + \sigma\mathcal{B}\mathcal{B}^*+\mathcal{T}\succ 0.\]
Assume that either (a) $\tau\in(0,1]$ and (\ref{case1_assumption}) holds or (b) $\tau\in(0,\frac{1+\sqrt{5}}{2})$ and (\ref{convergence_2}) and (\ref{case2_assumption}) hold.
Then there exist constants $D_1$, $D_2$ and $D_3$ that only depending on the initial point and the optimal solution set such that for $k\geq 1$, the following conclusions hold:
\noindent
(i) \begin{equation}\label{erg_pri}
\|\mathcal{A}^*\hat{u}^{k} + \mathcal{B}^*\hat{v}^{k} - c\|\leq D_1/k.
\end{equation}
\noindent
(ii) For any $(u,v,x)\in {B}_k: = \{(u,v,x)\in \mathcal{U}\times\mathcal{V}\times\mathcal{X}|\|(u,v,x) - (\hat{u}^k, \hat{v}^k, \hat{x}^k)\| \leq 1\}$,
\begin{equation}\label{complexity_decrease_1}\begin{array}{ll}
&(p(\hat{u}_k) + q(\hat{v}_k) ) -( p(u) + q(v) )+\langle \hat{w}^{k}-w,\nabla\phi(w)\rangle+ \langle \hat{u}^{k} - u, \mathcal{A}x\rangle + \langle \hat{v}^{k} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \hat{x}^{k} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle \leq D_2/k.
\end{array}
\end{equation}
\noindent
(iii) For case (b), if we further assume
that $\mathcal{S}- \eta\mathcal{D}_1\succeq 0$ and $\mathcal{T}- \eta\mathcal{D}_2\succeq 0$, then
\begin{equation}\label{erg_obj}
|\theta(\hat{u}^{k} ,\hat{v}^{k}) - \theta(\bar{u}, \bar{v}) | \leq D_3/k.
\end{equation}
The inequality (\ref{erg_obj}) holds for case (a) without additional assumptions.
\end{theorem}
{\bf Proof.}
{\bf (i)} Under the conditions for case (a), the inequality (\ref{convergence_1}) indicates that $\{\Phi_{k+1}(\bar{u}, \bar{v}, \bar{x})\}$ is a non-increasing sequence, which implies that
\begin{equation*}
(\tau\sigma)^{-1}\|x^{k+1} - \bar{x}\|^2 \leq \Phi_{k+1}(\bar{u}, \bar{v}, \bar{x}) \leq \Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{equation*}
Similarly under the conditions for case (b), we can get from (\ref{decrease_2}) that
\begin{equation*}
(\tau\sigma)^{-1}\|x^{k+1} - \bar{x}\|^2\leq \Psi_{k+1}(\bar{u}, \bar{v}, \bar{x})+\Xi_{k+1}\leq \Psi_1(\bar{u}, \bar{v}, \bar{x}) + \Xi_1.
\end{equation*}
Therefore, in terms of the ergodic primal feasibility, we have that
\begin{equation}\label{erg_1}\begin{array}{ll}
\|\mathcal{A}^*\hat{u}^{k} + \mathcal{B}^*\hat{v}^{k} - c\|^2 & = \|\frac{1}{k}\displaystyle\sum_{i=1}^k (\mathcal{A}^*{u}^{i+1} + \mathcal{B}^*{v}^{i+1} - c)\|^2 \\[8pt]
&
= \|(\tau\sigma)^{-1}\displaystyle\sum_{i=1}^k (x^{i+1} - x^i)\|^2/k^2\\[8pt]
&
= \|(\tau\sigma)^{-1}(x^{k+1} - x^1)\|^2/k^2\\[8pt]
&
\leq 2\|(\tau\sigma)^{-1}(x^{k+1} - \bar x)\|^2/k^2 + 2\|(\tau\sigma)^{-1}(x^1 - \bar x)\|^2/k^2
\leq C_3/k^2,
\end{array}
\end{equation}
where for case (a), $C_3 = 2(\tau\sigma)^{-1}\Phi_1(\bar{u}, \bar{v}, \bar{x}) +2\|(\tau\sigma)^{-1}(x^1 - \bar x)\|^2$ and for case (b), $C_3 = 2(\tau\sigma)^{-1}(\Psi_1(\bar{u}, \bar{v}, \bar{x}) + \Xi_1) +2\|(\tau\sigma)^{-1}(x^1 - \bar x)\|^2$. Then by taking the square root on inequality (\ref{erg_1}),
we can obtain (\ref{erg_pri}).
\bigskip
\noindent
{\bf (ii)} First, assume that the conditions for case (a) hold. Then the
inequality (\ref{decrease_11}) implies that for $i\geq 1$,
\begin{equation*}
\begin{array}{ll}
&(p(u^{i+1}) + q(v^{i+1}) ) -(p(u) + q(v) ) +\langle w^{i+1}-w,\nabla\phi(w)\rangle+ \langle u^{i+1} - u, \mathcal{A}x\rangle + \langle v^{i+1} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \tilde{x}^{i+1} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle \\[8pt]
\leq & -\frac{1}{2}(\Phi_{i+1}({u}, {v}, {x}) - \Phi_i({u}, {v}, {x})).
\end{array}
\end{equation*}
Thus,
summing up the above inequalities over $i = 1, \cdots,k$ and by using the convexity of functions $p$ and $q$, we can obtain that
\begin{equation}\label{vi_complexity}\begin{array}{ll}
p(\hat{u}_k) + q(\hat{v}_k) - (p(u) + q(v)) +\langle \hat{w}^{k}-w,\nabla\phi(w)\rangle+ \langle \hat{u}^{k} - u, \mathcal{A}x\rangle + \langle \hat{v}^{k} - v , \mathcal{B}x\rangle\\[8pt]
- \langle \hat{x}^k -x, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle
\leq (\Phi_{1}(u,v,x) - \Phi_{k+1}(u,v,x))/2k \leq \Phi_1(u,v,x)/2k.
\end{array}
\end{equation}
Next, we will provide an explicit bound of $\Phi_1(u,v,x)$ for $(u,v,x)\in {B}_k$
that only depends on the initial point and the optimal solution set.
Since $\{\Phi_{i+1}(\bar u,\bar v,\bar x)\}$ is non-increasing with $i$, for $i\geq 1$ we have
\begin{equation*}
(\tau \sigma)^{-1} \|\bar x - x^{i+1}
\|^2 + \|\bar u - u^{i+1} \|^2_{\mathcal{D}_1 + \mathcal{S} } +
\|\bar v - v^{i+1}
\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*} + \frac{1}{2}\|\bar w - w^{i+1}\|^2_{\mathcal{Q}}
\leq \Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{equation*}
Thus, summing up the above inequalities from $i = 1$ to $k$ and by applying Jensen's inequality to the convex function $\|\cdot\|^2$, we can get that
\begin{equation}\label{ergodic_2}
(\tau \sigma)^{-1} \|\bar x - \frac{1}{k}\displaystyle\sum_{i=1}^k x^{i+1}
\|^2 + \|\bar u - \hat u^k \|^2_{\mathcal{D}_1 + \mathcal{S} } +
\|\bar v - \hat v^k
\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*} + \frac{1}{2}\|\bar w - \hat w^k\|^2_{\mathcal{Q}}
\leq \Phi_1(\bar{u}, \bar{v}, \bar{x}).
\end{equation}
Recall that $\tilde x^{i+1} = x^i + \sigma(\mathcal{A}^*u^{i+1} +\mathcal{B}^*v^{i+1} - c) = x^i +\tau^{-1}(x^{i+1} - x^i)$. Then we also have
\begin{equation*}
\sum_{i=1}^k\tilde x^{i+1} = \sum_{i=1}^k x^i + \tau^{-1}(x^{k+1} - x^1) = \sum_{i=1}^k x^{i+1} + (\tau^{-1} -1)(x^{k+1} - x^1),
\end{equation*}
which, implies that for $k\geq 1$,
\begin{equation}\label{ergodic_3}\begin{array}{ll}
\|\bar{x} - \hat{x}^k\|^2 &= \|\bar{x} - \displaystyle\frac{1}{k}(\displaystyle\sum_{i=1}^k x^{i+1} + (\tau^{-1} -1)(x^{k+1} - x^1))\|^2\\[8pt]
&\leq 2\|\bar{x} - \displaystyle\frac{1}{k}\displaystyle\sum_{i=1}^k x^{i+1}
\|^2 + 2(\tau^{-1} - 1)^2\|\frac{1}{k}(x^{k+1} - x^1)\|^2\\[8pt]
&\leq 2\|\bar{x} - \displaystyle\frac{1}{k}\displaystyle\sum_{i=1}^k x^{i+1}
\|^2 + 4(\tau^{-1} - 1)^2\|x^{k+1} -\bar{x}\|^2 + 4(\tau^{-1} - 1)^2\|x^{1} -\bar{x}\|^2\\[8pt]
&\leq 2\|\bar{x} - \displaystyle\frac{1}{k}\displaystyle\sum_{i=1}^k x^{i+1}
\|^2 + 8(\tau^{-1} - 1)^2\Phi_1(\bar{u},\bar{v}, \bar{x}).
\end{array}
\end{equation}
Therefore, by (\ref{ergodic_2}) and (\ref{ergodic_3}), we can obtain that
\begin{equation}\label{ergodic_4}\begin{array}{ll}
&(\tau \sigma)^{-1} \|\bar x - \hat{x}^k
\|^2 + \|\bar u - \hat{u}^k \|^2_{\mathcal{D}_1 + \mathcal{S} } + \|\bar v - \hat{v}^k\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*} + \displaystyle\frac{1}{2}\| \bar w - \hat{w}^k\|^2_{\mathcal{Q}}\\[8pt]
\leq & 2(\tau\sigma)^{-1}\|\bar{x} - \displaystyle\frac{1}{k}\displaystyle\sum_{i=1}^k x^{i+1}
\|^2 + \|\bar u - \hat{u}^k \|^2_{\mathcal{D}_1 + \mathcal{S} } + \|\bar v - \hat{v}^k\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*} + \frac{1}{2}\| \bar w - \hat{w}^k\|^2_{\mathcal{Q}}\\[8pt]
& +8(\tau^{-1} - 1)^2\Phi_1(\bar{u},\bar{v}, \bar{x})\\[8pt]
\leq & (2 + 8(\tau^{-1}-1)^2)\Phi_1(\bar{u},\bar{v}, \bar{x}).
\end{array}
\end{equation}
In addition, we also know from (\ref{erg_1}) that for $k\geq 1$,
\begin{equation}\label{ergodic_5}\begin{array}{ll}
\|\mathcal{A}^*\hat{u}^{k} + \mathcal{B}^*\hat{v}^{k} -c\|^2 \leq C_3/k^2\leq C_3.
\end{array}
\end{equation}
Now putting inequalities (\ref{ergodic_4}) and (\ref{ergodic_5}) together, we see that for any {$(u,v,x)\in {B}_k,$}
\begin{equation*}
\begin{array}{ll}
& \displaystyle(\tau \sigma)^{-1} \|x^1 - x\|^2+ \|u^1 -
u\|^2_{\mathcal{D}_1 + \mathcal{S}} + \|v^1-v\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} +\mathcal{T} } + \frac{1}{2}\|w^{1} - w\|^2_{\mathcal{Q}} + \sigma \|\mathcal{A}^*u + \mathcal{B}^*v^1 - c\|^2 \\[8pt]
= & (\tau \sigma)^{-1} \|(x^1 - \bar{x}) + (\bar{x} - \hat{x}^k) + (\hat{x}^k - x)\|^2+ \|(u^1 -\bar{u}) + (\bar{u} - \hat{u}^k) + (\hat{u}^k - u)\|^2_{\mathcal{D}_1 + \mathcal{S}} \\[8pt]
&
+ \displaystyle\|(v^1-\bar v) + (\bar{v} - \hat{v}^k) + (\hat{v}^k - v)\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} +\mathcal{T} } + \frac{1}{2}\|(w^{1} - \bar{w}) +( \bar{w} - \hat{w}^k) + (\hat{w}^k - w)\|^2_{\mathcal{Q}} \\[8pt]
&
\displaystyle+ \sigma \|\mathcal{A}^*(u - \hat{u}^k) + \mathcal{B}^*(v^1 -\bar{v}) + \mathcal{B}^*(\bar{v}- \hat{v}^k)+( \mathcal{A}^*\hat{u}^k + \mathcal{B}^*\hat{v}^k - c)\|^2\\[8pt]
\leq & 3[(\tau \sigma)^{-1} \|x^1 - \bar x
\|^2 + \|u^1 - \bar u \|^2_{\mathcal{D}_1 + \mathcal{S} } + \|v^1 - \bar v\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + 2\sigma\mathcal{B}\mathcal{B}^*} + \frac{1}{2}\| w^1 - \bar w\|^2_{\mathcal{Q}}] \\[8pt]
&\displaystyle + 3[(\tau \sigma)^{-1} \|\bar x - \hat{x}^k
\|^2 + \|\bar u - \hat{u}^k \|^2_{\mathcal{D}_1 + \mathcal{S} } + \|\bar v - \hat{v}^k\|^2_{\mathcal{Q}_{22} + \mathcal{D}_{2} + \mathcal{T} + \sigma\mathcal{B}\mathcal{B}^*} + \frac{1}{2}\| \bar w - \hat{w}^k\|^2_{\mathcal{Q}}]\\[8pt]
& + 3[(\tau \sigma)^{-1} \|\hat{x}^k - x\|^2 + \|
\hat{u}^k - u\|^2_{\mathcal{D}_1 + \mathcal{S}+2\sigma \mathcal{A}\mathcal{A}^*} + \|
\hat{v}^k - v\|^2_{\mathcal{Q}_{22} + \mathcal{D}_2 + \mathcal{T}} + \frac{1}{2}\|\hat{w}^k - w\|^2_{\mathcal{Q}}] \\[8pt]
&
\displaystyle+ 3\|\mathcal{A}^*\hat{u}^k + \mathcal{B}^*\hat{v}^k - c\|^2\\[8pt]
\leq & 6\Phi_1(\bar{u}, \bar{v}, \bar{x}) + 3(2 + 8(\tau^{-1}-1)^2)\Phi_1(\bar{u},\bar{v}, \bar{x})+ 3C_4 + 3C_3,
\end{array}
\end{equation*}
where
$C_4 = \max((\tau\sigma)^{-1}, \|\frac{1}{2}\mathcal{Q}+ \text{Diag}\,(\mathcal{D}_1 + \mathcal{S}+2\sigma \mathcal{A}\mathcal{A}^*, \mathcal{Q}_{22} + \mathcal{D}_2 + \mathcal{T})\|)$.
Then we can get that for any $(u,v,x)\in {B}_k$, $\Phi_1(u,v,x) $ is bounded by a positive constant $$C_5 = (12 + 24(\tau^{-1}-1)^2)\Phi_1(\bar{u},\bar{v}, \bar{x})+ 3C_4 + 3C_3.$$ This, together with (\ref{vi_complexity}), implies that
\begin{equation*}\begin{array}{ll}
&(p(\hat{u}^k) + q(\hat{v}^k) ) -(p(u) + q(v) )+\langle \hat{w}^{k}-w,\nabla\phi(w)\rangle+ \langle \hat{u}^{k} - u, \mathcal{A}x\rangle + \langle \hat{v}^{k} - v , \mathcal{B}x\rangle\\[8pt]
&
- \langle \hat{x}^{k} - x , \mathcal{A}^*u + \mathcal{B}^*v - c\rangle
\leq C_5/2k.
\end{array}
\end{equation*}
Under the conditions for case (b),
by a very similar approach we can get an upper bound $\overline{C}_5$ for $(\Psi_1(u,v,x) +\Xi_1)$
such that
\begin{equation*}\begin{array}{ll}
(p(\hat{u}^k) + q(\hat{v}^k)) - (p(u) + q(v)) +\langle \hat{w}^{k}-w,\nabla\phi(w)\rangle+ \langle \hat{u}^{k} - u, \mathcal{A}x\rangle + \langle \hat{v}^{k} - v , \mathcal{B}x\rangle\\[8pt]
- \langle \hat{x}^k -x, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle
\leq (\Psi_1(u,v,x)+\Xi_1)/2k\leq \overline{C}_5/2k.
\end{array}
\end{equation*}
The above arguments show that for both cases, the property (\ref{complexity_decrease_1}) holds.
\noindent
{\bf (iii)} For the complexity of primal objective functions, first, we know from (\ref{kkt}) that
\begin{equation*}\begin{array}{ll}
p(u)\geq p(\bar{u}) + \langle -\mathcal{A}\bar{x} - \nabla_u\phi(\bar{w}), u-\bar{u}\rangle \quad \forall u\in\mathcal{U},\\[8pt]
q(v)\geq q(\bar{v}) + \langle -\mathcal{B}\bar{x} - \nabla_v\phi(\bar{w}), v - \bar{v}\rangle \quad\forall v\in\mathcal{V}.
\end{array}
\end{equation*}
Therefore, summing them up and by noting $\mathcal{A}^*\bar{u} + \mathcal{B}^*\bar{v} = c$ and the convexity of function $\phi$, we have that
\begin{equation*}\begin{array}{ll}
\theta(u,v) - \theta(\bar{u}, \bar{v}) & \geq -\langle \bar{x}, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle
+ \phi(w) - \phi(\bar{w}) - \langle \nabla\phi(\bar{w}), w-\bar{w}\rangle\\[8pt]
&\geq -\langle \bar{x}, \mathcal{A}^*u + \mathcal{B}^*v - c\rangle
\quad\forall u\in\mathcal{U}, v\in\mathcal{V}.
\end{array}
\end{equation*}
Thus, with $(u,v) = (\hat{u}^{k},\hat{v}^{k})$, it holds that
\begin{equation}\label{ergfun_1}\begin{array}{ll}
\theta(\hat{u}^{k},\hat{v}^{k}) - \theta(\bar{u}, \bar{v}) & \geq -\langle \bar{x}, \mathcal{A}^*\hat{u}^{k} + \mathcal{B}^*\hat{v}^{k} - c\rangle
\\[8pt]
&
\geq \displaystyle-\frac{1}{2}(\frac{1}{k}\|\bar{x}\|^2 + k\|\mathcal{A}^*\hat{u}^{k} + \mathcal{B}^*\hat{v}^{k} - c\|^2)\\[8pt]
&
\geq \displaystyle-\frac{1}{2}(\|\bar{x}\|^2 + C_3)/k,
\end{array}
\end{equation}
where $C_3$ is the same constant as in (\ref{erg_1}).
For the reverse part,
by (\ref{phi_convex}) and (\ref{phi_majorize}) we can obtain that for any $i\geq 1$,
\begin{equation*}\begin{array}{ll}
\phi(w^{i+1}) \leq \displaystyle\phi(w^{i}) + \langle \nabla \phi(w^{i}),w^{i+1} - w^i\rangle + \frac{1}{2}\|w^{i+1} - w^i\|^2_{\mathcal{Q} + \mathcal{H}},\\[8pt]
\phi(\bar{w}) \geq \displaystyle\phi(w^{i}) + \langle \nabla \phi(w^{i}),\bar{w}-w^i\rangle + \frac{1}{2}\|\bar{w}-w^i\|^2_{\mathcal{Q}},
\end{array}
\end{equation*}
which, indicate, that
\begin{equation}\label{erg_10}
\phi(w^{i+1}) - \phi(\bar{w})\leq \langle \nabla\phi(w^i), w^{i+1} - \bar{w}\rangle + \frac{1}{2}\|w^{i+1} - w^i\|^2_{\mathcal{Q} + \mathcal{H}} - \frac{1}{2}\|w^i - \bar{w}\|^2_{\mathcal{Q}}.
\end{equation}
Thus, (\ref{pqineq}) and (\ref{erg_10}) imply that for $\tau\in(0,1]$ and any $i\geq 1$,
\begin{equation}\label{ergfun_3}
\begin{array}{ll}
&\theta(u^{i+1}, v^{i+1}) -\theta(\bar u,\bar v)\\[8pt]
\leq & \displaystyle\frac{1}{2}\|w^{i+1} - w^i\|^2_{\mathcal{Q}+\mathcal{H}} - \frac{1}{2}\|w^i-\bar{w}\|^2_{\mathcal{Q}} +\langle\bar{w} - w^{i+1}, \mathcal{Q}(w^{i+1}- w^i)\rangle + \langle \bar{u} - u^{i+1}, \mathcal{A}\tilde{x}^{i+1}\rangle \\[8pt]
&
\displaystyle + \langle \bar{v} - v^{i+1}, \mathcal{B}\tilde{x}^{i+1}\rangle - \langle v^{i+1} - v^i, \mathcal{Q}_{12}^*(\bar{u} - u^{i+1})\rangle
+\sigma \langle \mathcal{A}^*( u^{i+1}-\bar{u}), \mathcal{B}^*(v^{i+1}-v^i)\rangle\\[8pt]
&
\displaystyle+\langle \bar{u} - u^{i+1}, ( \mathcal{D}_1 + \mathcal{S})(u^{i+1} - u^i)\rangle + \langle \bar{v} -v^{i+1}, ( \mathcal{D}_2+\mathcal{T})(v^{i+1} - v^i)\rangle\\[8pt]
\leq &\displaystyle\frac{1}{2}(\Lambda_{i} - \Lambda_{i+1})
- \frac{1}{2}(\|u^{i+1} - u^i\|^2_{\mathcal{S}} +\|v^{i+1} - v^i\|^2_{\mathcal{T}} + \sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^i - c\|^2 \\[8pt]
&
\displaystyle+\sigma(1-\tau)\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2) \\[8pt]
\leq & \displaystyle\frac{1}{2}(\Lambda_{i} - \Lambda_{i+1}).
\end{array}
\end{equation}
Therefore, summing up the above inequalities over $i = 1, \cdots k$ and by using the convexity of function $\theta$ we can obtain that
\begin{equation}\label{ergfun_2}
\theta(\hat{u}_k,\hat{v}_k) - \theta(\bar{u}, \bar{v})
\leq (\Lambda_{1}(u,v,x) - \Lambda_{k+1}(u,v,x))/2k
\leq \Lambda_1/2k.
\end{equation}
The inequalities (\ref{ergfun_1}) and (\ref{ergfun_2}) indicate that (\ref{erg_obj}) holds for case (a).
Next, assume that the conditions for case (b) hold. Similar to (\ref{ergfun_3}), we have that
\begin{equation*}
\begin{array}{ll}
&\theta(u^{i+1}, v^{i+1}) -\theta(\bar u,\bar v)\\[8pt]
\leq &\displaystyle\frac{1}{2}(\overline{\Lambda}_{i} - \overline{\Lambda}_{i+1})
- \frac{1}{2}(\|u^{i+1} - u^i\|^2_{\mathcal{S}-\eta\mathcal{D}_1} +\|v^{i+1} - v^i\|^2_{\mathcal{T} + \min(\tau, 1+\tau - \tau^2)\sigma\mathcal{B}\mathcal{B}^* - \eta\mathcal{D}_2} \\[8pt]
&
+\min(1,1 + \tau^{-1}-\tau)\sigma\|\mathcal{A}^*u^{i+1} + \mathcal{B}^*v^{i+1} - c\|^2).
\end{array}
\end{equation*}
By the assumptions that $\mathcal{S}-\eta\mathcal{D}_1\succeq 0$ and $\mathcal{T} - \eta\mathcal{D}_2\succeq 0$, we can obtain that
\begin{equation}\label{ergfun_5}
\theta(\hat{u}^{k},\hat{v}^{k}) - \theta(\bar{u}, \bar{v})\leq (\overline{\Lambda}_1 - \overline{\Lambda}_{k+1})/2k\leq \overline{\Lambda}_1/2k.
\end{equation}
Thus, by (\ref{ergfun_2}) and (\ref{ergfun_5}) we can obtain the inequality (\ref{erg_obj}).
\hfill $ \mathchoice\sqr74\sqr74\sqr{2.1}3\sqr{1.5}3 $
\bigskip
Below we make a couple of remarks about the results in Theorem \ref{erg_main}.
\begin{remark}
The result in part (ii) can be regarded as an ergodic version to (\ref{nonerg_main}) on the KKT optimality condition, though less explicit. Note that if one takes the square root on (\ref{nonerg_main}), the right hand side will become $O(1/\sqrt{k})$.
The inequality (\ref{complexity_decrease_1})
indicates that the majorized ADMM requires no more than $O(1/\varepsilon)$ iterations to obtain an $\varepsilon$-approximation solution in the sense of (\ref{def_approxiamteVI}).
When the objective function is separable,
this kind of results
has been studied for the (proximal) classic ADMM with separable objective functions, for examples,~\cite{Monteiro2013,HeYuan2012} and in a recent work by Li et al.~\cite{LiSunToh_indefinite2014}
for a majorized ADMM
with indefinite proximal terms.
\end{remark}
\begin{remark}
The results in parts (i) and (iii), which are on the ergodic complexity of the primal feasibility and the objective function, respectively, are extended from the work of Davis and Yin~\cite{Yin2014} on the classic ADMM with separable objective functions. These results are more explicit than the one in part (ii). However, there is no corresponding result available on the dual problem. Therefore, it will be very interesting to see if one can develop a more explicit ergodic complexity result containing all the three parts in the KKT condition.
\end{remark}
\section{Conclusions}
In this paper, we establish the convergence properties for the majorized ADMM with a large step length to solve linearly constrained convex programming whose {objective function} includes a coupled smooth function. From Theorem \ref{global_convergence}, one can see the influence of the coupled objective on the convergence condition. For $\tau\in(0,\frac{1+\sqrt{5}}{2})$, a joint condition like (\ref{case1_assumption}) or (\ref{case2_assumption}) is needed
to analyze the behaviour of the iteration sequence. One can further observe that the parameter $\eta$, which controls the off-diagonal term of the generalized Hessian, also affects the choice of proximal operators $\mathcal{S}$ and $\mathcal{T}$. However, as is pointed out in Remark \ref{Q_plus_separable}, when the {coupled function is convex quadratic}, $\eta = 0$ and the corresponding influence would disappear. Although, in this paper we focus on the 2-block case, it is not hard to see that, with the help of the Schur complement technique introduced in~\cite{LiSunToh_indefinite2014},
one can apply our majorized ADMM to solve large scale convex optimization problems with many smooth blocks.
\bigskip
{\bf Acknowledgements.} The authors would like to thank Dr.~Caihua Chen at the Nanjing University for discussions on the iteration complexity described in the paper.
|
\section{introduction}
The multifractal analysis of chaotic dynamical systems consists in the study of fine geometric structures of invariant sets.
One considers the so-called multifractal decompositions
of an invariant set, and the associated multifractal spectra which encodes this decomposition.
By connecting the spectra to other characteristics of the system,
such as entropy and Lyapunov exponents of invariant measures,
one tries to get more refined description of the underlying dynamics than purely stochastic considerations.
In this paper we treat the Birkhoff averages of continuous functions.
Although this type of problem is well-understood for uniformly hyperbolic systems,
much less is known for non hyperbolic ones.
We treat certain non-hyperbolic two-dimensional maps at the boundary of uniform hyperbolicity,
having quadratic tangencies
between invariant manifolds.
We are concerned with a family of H\'enon-like diffeomorphisms
\begin{equation*}\label{henon}
f_a\colon(x,y)\in\mathbb R^2\mapsto(1-ax^2,0)+b\cdot\Phi(a,b,x,y),\quad a\in\mathbb R, \ 0<b\ll1.\end{equation*}
Here, $\Phi$ is bounded continuous in $(a,b,x,y)$ and
$C^2$ in $(a,x,y)$.
We assume there exists a constant $C>0$ such that for all $a$ near $2$
and small $b$,
\begin{equation*}\label{jacobian}\|D\log|\det Df_{a}|\|\leq C.\end{equation*}
This family describes the transition from uniformly hyperbolic to non hyperbolic regimes.
It is known \cite{BedSmi06,CLR08,DevNit79,Tak13} that there is
a {\it first bifurcation parameter} $a^*=a^*(b)$ with the following properties:
the non wandering
set of $f_a$ is a uniformly hyperbolic horseshoe for $a>a^*$;
for $a=a^*$ there is a unique orbit of homoclinic or heteroclinic tangency,
and the tangency is quadratic.
The aim of this paper is to perform the multifractal analysis of $f_{a^*}$.
Although the dynamics of $f_{a^*}$ resembles that of the horseshoe before the first bifurcation,
the presence of tangency presents novel obstructions
for understanding the global dynamics.
We state our settings and goals in more precise terms.
Write $f$ for $f_{a^*}$.
Let $P$, $Q$ denote the fixed saddles of $f$ near $(1/2,0)$, $(-1,0)$ respectively.
The orbit of tangency intersects a small neighborhood of the origin exactly at one point, denoted by $\zeta_0$ (FIGURE 1).
If $f$ preserves orientation, then $\zeta_0\in W^s(Q)\cap W^u(Q)$.
If $f$ reverses orientation, then $\zeta_0\in W^s(Q)\cap W^u(P)$.
The sole obstruction to uniform hyperbolicity is the orbit of the tangency $\zeta_0$.
Let $\Omega$ denote
the non wandering set of $f$, which is a compact set.
If $f$ preserves orientation, let $W^u=W^u(Q)$. Otherwise,
let $W^u=W^u(P)$.
The (non-uniform) expansion along $W^u$ is responsible for the chaotic behavior of $f$.
Hence, a good deal of multifractal information of $\Omega$
is contained in its unstable slice
$$\Omega^u=\Omega\cap W^u.$$
\begin{figure}
\begin{center}
\includegraphics[height=6.5cm,width=14cm]
{figure1.eps}
\caption{Organization of invariant manifolds of the
fixed saddles $P$, $Q$ of $f=f_{a^*}$ near $(1/2,0)$, $(-1,0)$
respectively. In the case $\det Df>0$ (left),
the stable and unstable manifolds of $Q$ meet each other tangentially. In the case $\det Df<0$
(right), the stable manifold of $Q$ meets the unstable manifold of $P$ tangentially. The shaded
regions represent the region $R$ containing the non wandering set $\Omega$ (see Sect.\ref{family}).}
\end{center}
\end{figure}
Given a continuous function $\varphi\colon \Omega\to \mathbb R$ consider {\it level sets} of the Birkhoff averages of $\varphi$:
$$\Omega_\varphi^u(\beta)=\left\{x\in \Omega^u\colon\lim_{n\to\infty}
\frac{1}{n}S_n\varphi(x)=\beta\right\},\ \ \beta\in\mathbb R,$$
where $S_n\varphi=\sum_{i=0}^{n-1}\varphi\circ f^i$.
Define
$$c_\varphi=\inf_{x\in \Omega}\liminf_{n\to\infty}\frac{1}{n}S_n\varphi(x)
\ \text{ and } \ d_\varphi=\sup_{x\in\Omega}\limsup_{n\to\infty}\frac{1}{n}S_n\varphi(x).$$
In what follows we assume
$c_\varphi<d_\varphi$.
For otherwise the Birkhoff averages of $\varphi$ along
all orbits are equal.
Set $I_\varphi=[c_\varphi,d_\varphi]$. Consider the multifractal decomposition
\begin{equation*}\label{decomposition}\Omega^u
=\left(\bigcup_{\beta\in I_\varphi}
\Omega_\varphi^u(\beta)\right)\cup \hat \Omega_\varphi^u,\end{equation*}
where
$\hat \Omega_\varphi^u$ denotes the set of points in $\Omega^u$ for which
$(1/n)S_n\varphi(x)$ does not converge.
This decomposition has extremely complicated topological structures:
each $\Omega_\varphi^u(\beta)$ is nonempty (See Sect.\ref{lowest} for details);
one can show that each set appearing in the decomposition is dense in $\Omega^u$;
namely, a decomposition into an uncountable number of dense subsets.
Let
$\mathcal M(f)$ denote the set of $f$-invariant
Borel probability measures. The entropy of $\mu\in\mathcal M(f)$ is denoted by $h(\mu)$.
An {\it unstable Lyapunov exponent} of $\mu\in\mathcal M(f)$ is the number $\lambda^u(\mu)$
defined by
$$\lambda^u(\mu)=\int\log J^u(x) d\mu(x).$$
Here, $J^u(x)=\Vert D_xf|E^u_x\Vert $, and
$E^u_x$ is a one-dimensional subspace of $T_x\mathbb R^2$ called
an \emph{unstable direction} at $x\in\Omega$ that is characterized by
the following backward contraction property \cite{SenTak1}:
\begin{equation*}
\limsup_{n\to\infty}\frac{1}{n}\log\|D_xf^{-n}|E^u_x\|<0.\end{equation*}
By a result of \cite{CLR08}, $\inf\{\lambda^u(\mu)\colon\mu\in\mathcal M(f)\}>0$.
Relationships between entropy, unstable Lyapunov exponents and dimension of
invariant probability measures were established in \cite{LedYou85}.
Our main result connects these characteristics to the Hausdorff dimension of $\Omega_\varphi^u(\beta)$
that is defined as follows.
Given $p\in(0,1]$
the unstable Hausdorff
$p$-measure of a set $A\subset W^u$ is defined by
$$m_p^u(A)=\lim_{\delta\to0}\left(\inf\sum_{U\in\mathcal
U } {\rm length}(U)^p \right),$$ where ${\rm length}(\cdot)$ denotes the length on $W^u$ with respect to the induced Riemannian metric,
and the infimum is taken over all
coverings $\mathcal U$ of $A$ by open sets of $W^u$ with
length $\leq\delta$. The unstable Hausdorff dimension of $A$,
denoted by $\dim_H^u$, is the unique number in $[0,1]$ such that
$$\dim_H^u(A)=\sup\{p\colon
m_p^u(A)=\infty\}=\inf\{p\colon m_p^u(A)=0\}.$$
Now, set
$$B_\varphi^u(\beta)=\dim_H^u(\Omega_\varphi^u(\beta)),$$
and
$$I_\varphi'=\{\beta\in I_\varphi
\colon B_\varphi^u(\beta)\leq 2/\log(1/ b)\}.$$
Our main result is stated as follows.
\begin{theorema}
Let $b>0$ be sufficiently small and $f=f_{a^*(b)}$ as above.
For any continuous function $\varphi\colon \Omega\to\mathbb R$ and
all $\beta\in I_\varphi\setminus I_\varphi'$,
$$B_\varphi^u(\beta)=\lim_{\varepsilon\to0}\sup\left\{\frac{h(\mu)}{\lambda^u(\mu)}\colon \mu\in\mathcal M(f),\left|\int\varphi d\mu-\beta\right|<\varepsilon\right\}.$$
\end{theorema}
This type of formula has been proved under different settings and
assumptions on the hyperbolicity of the systems:
uniformly hyperbolic ones
\cite{BarSau01,Ols03, PesWei01, Wei99} (a more complete list of previous results
can be found in \cite{Pes97});
maps with parabolic fixed points \cite{Chu10,JohJorObePol10};
certain non-uniformly expanding quadratic maps on the interval \cite{Chu10,ChuTak13}.
Up to present,
many of the known results for non hyperbolic systems are limited to the Lyapunov spectrum
\cite{GelPrzRam10,GelRam09,IomTod11,KesStr04,Tak13'}.
In \cite{Chu10}, Chung established the formula for a class of one-dimensional maps
admitting ``nice" induced Markov maps.
A strategy for a proof of our theorem is to
use a (locally defined) stable foliation to identify points on the same leaf
(called {\it long stable leaves} in \cite[Sect.2.8]{Tak13'}), and to extend the one-dimensional argument
in \cite{Chu10}.
The same strategy has been taken in
\cite{Tak13'} in which the non continuous function $\log J^u$ was treated instead of $\varphi$.
Since the stable foliation is not globally defined,
it is not possible to tell whether such a leaf through a given point exist.
The argument in \cite{Tak13'} to handle this difficulty consists of three steps:
(i) introduce {\it dynamically critical points} in the spirit of Benedicks and Carleson \cite{BenCar91},
and define a {\it bad set} in terms of the recurrence to the critical points;
(ii) show that long stable leaves exist for points outside of the bad set;
(iii) show that the Birkhoff averages of $\log J^u$ do not converge on the bad set.
A novel obstruction in dealing with continuous $\varphi$ is that the Birkhoff averages of $\varphi$ can converge,
for points in the bad set (denoted by
$\Omega_*^u$ in Sect.\ref{outline}).
What we can do at best is to show that the dimension of $\Omega_*^u$ is small, and
establish the formula for those $\beta$ for which $B_\varphi(\beta)$ is not too small.
This is the reason for the restriction on $\beta$ in the theorem.
To clarify the range of $\beta$ for which the formula in the theorem holds,
let us recall the thermodynamic formalism of $f$
developed in \cite{SenTak1,SenTak2}.
For $t\in\mathbb R$ define
$$P(t)=\sup\left\{h(\mu)-t\lambda^u(\mu)\colon\mu\in\mathcal M(f)\right\}.$$
A measure which attains this supremum is called an {\it equilibrium measure} for $-t\log J^u$.
The function $t\mapsto P(t)$ is convex. One has $P(0)>0$, and Ruelle's inequality \cite{Rue78} gives $P(1)\leq0$.
Since $f$ has no SRB measure \cite{Tak12}, $P(1)<0$ holds. Hence the equation $P(t)=0$ has a unique solution in $(0,1)$,
denoted by $t^u$.
There exists a unique equilibrium measure for $-t^u\log J^u$
(\cite[Theorem A]{SenTak2}),
denoted by $\mu_{t^u}$, and $t^u=\dim_H^u(\Omega^u)$, $t^u\to1$ as $b\to0$ (\cite[Theorem B]{SenTak2}).
From the theorem and the Ergodic Theorem,
$B_\varphi^u(\int\varphi d\mu_{t^u})\geq h(\mu_{t^u})/\lambda(\mu_{t^u})=t^u$.
It follows that $B_\varphi^u$ takes its maximum at $\beta=\int\varphi d\mu_{t^u}$.
Similarly to the proof of \cite[Theorem C]{Tak13'}
one can show that $B_\varphi^u$ is continuous on $I_\varphi\setminus I_\varphi'$,
increasing on $\{\beta\in I_\varphi\setminus I_\varphi'\colon
\beta<\int\varphi d\mu_{t^u}\}$ and decreasing on $\{\beta\in I_\varphi
\setminus I_\varphi'\colon
\beta>\int\varphi d\mu_{t^u}\}$, so that
the set $I_\varphi\setminus I_\varphi'$ is an interval containing $t^u$.
The rest of this paper consists of two sections.
Sect.2 is a preliminary, and the theorem is proved
in Sect.3.
\section{Preliminaries}
The main reference of this section is \cite{Tak13'}. We collect several results and constructions, and
prove two lemmas needed for the proof of the theorem.
Throughout this paper we shall be concerned with positive constants $\delta$, $b$,
chosen in this order.
The letter $C$ is used to denote any positive constant which is independent of
$\delta$ or $b$.
\subsection{The non wandering set}\label{family}
By a {\it rectangle} we mean any
compact domain bordered by two compact curves in $W^u$ and two in the
stable manifolds of $P$ or $Q$. By an {\it unstable side} of a
rectangle we mean any of the two boundary curves in $W^u$. A {\it
stable side} is defined similarly.
By the result of \cite[Lemma 3.2]{Tak13} there exists a rectangle $R$ contained in the set
$\{(x,y)\in \mathbb R^2\colon |x|<2, |y|< \sqrt{b}\}$ with the following properties (See FIGURE 1):
\begin{itemize}
\item $\displaystyle{\Omega=\{x\in R\colon f^nx\in R\ \text{ for every }n\in\mathbb Z\}}$;
\item one of the unstable sides of $R$ contains $\zeta_0$;
\item one of the stable sides of $R$ contains $f\zeta_0$. This side is denoted by $\alpha_0^+$. The other side, denoted by $\alpha_0^-$, contains $Q$;
\item $f\alpha_0^+\subset\alpha_0^-$.
\end{itemize}
\subsection{Critical points}\label{critical}
Set
$$I(\delta)=\{(x,y)\in R\colon |x|<\delta\}.$$
The derivatives grow exponentially, as long as the orbit is outside of $I(\delta)$.
To treat returns to $I(\delta)$ we mimic the strategy of Benedicks $\&$ Carleson \cite{BenCar91} and
introduce the notion of critical points.
The reference for the contents in this subsection is \cite[Sect.2.4 $\&$ Sect.2.5]{Tak13'}.
From
the hyperbolicity of the saddle $Q$,
there exist two mutually disjoint connected open sets $U^-$, $U^+$ independent of $b$ such that
$\alpha_0^-\subset U^-$, $\alpha_0^+\subset U^+$, $U^+\cap fU^+=\emptyset=U^+\cap fU^-$ and
a foliation $\mathcal F^s$ of $U=U^-\cup U^+$ by one-dimensional leaves such
that:
\begin{itemize}
\item $\mathcal F^s(Q)$, the leaf of $\mathcal F^s$ containing $Q$,
contains $\alpha_0^-$;
\item if $x,fx\in U$, then $f(\mathcal F^s(x))
\subset\mathcal F^s(fx)$;
\item Let $e^s(x)$ denote the unit vector in $T_x\mathcal F^s(x)$ whose second component is positive.
Then $x\mapsto e^s(x)$ is $C^{1}$, $\|D_xfe^s(x)\|\leq Cb$ and $\|D_xe^s(x)\|\leq C$;
\item If $x,fx\in U$, then $s(e^s(x))\geq
C/\sqrt{b}.$
\end{itemize}
\begin{definition}
{\rm We say $\zeta\in W^u\cap I(\delta)$ is a {\it critical point} if $f\zeta\in U^+$ and
$T_{f\zeta}W^u=T_{f\zeta}\mathcal F^s(f\zeta)$.}
\end{definition}
From the first two conditions on $\mathcal F^s$ and $f\alpha_0^+\subset\alpha_0^-$, there is a leaf of $\mathcal F^s$ which
contains $\alpha_0^+$. Since $f\zeta_0\in\alpha_0^+$ we have
$f\zeta_0\in U^+$ and
$T_{f\zeta_0}W^u=T_{f\zeta_0}\mathcal F^s(f\zeta_0)$,
namely, $\zeta_0$ is a critical point.
To locate all other critical points we need some preliminary considerations.
Let $\alpha_1^+$ denote the (connected) component of $W^s(P)\cap R$ containing $P$,
and $\alpha_1^-$ the component of $f^{-1}\alpha_1^+\cap R$ not containing $P$.
Let $\Theta$ denote the rectangle bordered by $\alpha_1^-$, $\alpha_1^+$ and the unstable sides of $R$.
Let $\tilde\Gamma^u$ denote the collection of
components of $\Theta\cap W^u$.
By a \emph{$C^2(b)$-curve} we mean a compact, nearly horizontal $C^2$ curve in $R$ such that the slopes of its tangent directions are $\leq\sqrt{b}$ and the
curvature is everywhere $\leq\sqrt{b}$.
Let $S$ denote the compact lenticular domain
which is bounded by the parabola $f^{-1}\alpha_0^+\cap R$ and the unstable side of $R$ containing $\zeta_0$.
Then the following holds \cite[Lemma 2.5 $\&$ Lemma 2.8]{Tak13'}:
\begin{itemize}
\item {\bf (Location)} any element of $\tilde\Gamma^u$
is a $C^2(b)$-curve with endpoints in $\alpha_1^-$, $\alpha_1^+$,
and contains a unique critical point;
\item {\bf (Non recurrence)}
all critical points are contained in $S$.
\end{itemize}
In particular, all critical points never return to the interior of $R$ under forward iteration.
The dynamics of $f$ is amenable to analysis primarily due to this non-recurrence of critical points.
To recover the loss of derivatives suffered from the return to $I(\delta)$, we bind the point
to a suitable critical point \cite[Lemma 2.9]{Tak13'}, and let it copy the exponential derivative growth along the critical orbit
\cite[Lemma 2.6]{Tak13'}.
\subsection{Inducing}\label{induced map}
We introduce an inducing scheme associated with the first return map to $\Theta$.
The reference for the contents in this subsection is \cite[Sect.2.10]{Tak13'}.
Define a sequence $\{\tilde\alpha_n\}_{n=0}^\infty$ of compact curves in $R\cap W^s(P)$ inductively as follows.
First, set $\tilde\alpha_0=\alpha_1^+$. Given $\tilde\alpha_{n-1}$, define $\tilde\alpha_n$ to be one of the two connected components of $f^{-1}\tilde\alpha_{n-1}\cap R$ which is at the left of $\zeta_0$. Observe that $\tilde\alpha_1=\alpha_1^-$. By the Inclination Lemma,
the Hausdorff distance between $\tilde\alpha_n$ and $\alpha_0^-$ converges to $0$ as $n\to\infty$.
For each $n\geq0$ let $\alpha_n$ denote the connected component of $R\cap f^{-1}\tilde\alpha_n$
which is not $\tilde\alpha_{n+1}$.
The set $R\cap f^{-1}\alpha_n$ consists of two curves, one at the left
of $\zeta_0$ and the other at the right. They are denoted by $\alpha_{n+1}^-$,
$\alpha_{n+1}^+$ respectively.
By definition, these curves obey the following diagram
\begin{equation*}
\{\alpha_{n+1}^-,\alpha_{n+1}^+\}\stackrel{f^2}{\to}\tilde\alpha_n
\stackrel{f}{\to}\tilde\alpha_{n-1}
\stackrel{f}{\to}\tilde\alpha_{n-2}\stackrel{f}{\to}\cdots
\stackrel{f}{\to}\tilde\alpha_1=\alpha_1^-\stackrel{f}{\to}
\tilde\alpha_0=\alpha_1^+.\end{equation*}
Define $r\colon \Theta\to \mathbb N\cup\{\infty\}$ by
$$r(x)=\inf(\{n>0\colon f^nx\in\Theta\}\cup\{\infty\}),$$
which is the first return time of $x$ to $\Theta$.
Note that:
\begin{itemize}
\item $r(x)=1$ if and only if $x\in\alpha_1^-\cup\alpha_1^+$;
$r(x)=n+1$ $(n\geq1)$ if and only if $x$ is sandwiched by $\alpha_{n}^+$ and $\alpha_{n+1}^+$, or
by $\alpha_{n}^-$ and $\alpha_{n+1}^-$;
$r(x)=\infty$ if and only if $x\in S$;
\item each level set of $r$ except $S$ has exactly two connected components.
\end{itemize}
Let $\mathcal P$ denote the partition of the set $\Theta\setminus(S\cup\alpha_1^-\cup\alpha_1^+)$ into connected components of the level sets of
the function $r$.
The $\mathcal P$ is well-defined because the Hausdorff distance between
$\alpha_n$ and $\alpha_0^+$ converges to $0$ as $n\to\infty$.
Set $\mathcal P_1=\{\omega=\overline{\eta}\colon\eta\in\mathcal P\}$,
where the bar denotes the closure operation.
For each $n\geq2$ define
$$\mathcal P_n=\left\{\omega_0\cap \bigcap_{i=1}^{n-1} f^{-r(\omega_0)}\circ f^{-r(\omega_1)}
\circ\cdots\circ f^{-r(\omega_{i-1})}\omega_i
\colon\omega_0,\omega_1,\ldots,\omega_{n-1}\in \mathcal P_1\right\}.$$
Elements of $\bigcup_{n\geq0}\mathcal P_n$ are called \emph{proper rectangles.}
The unstable sides of a proper rectangle are formed by two curves contained in the unstable sides of $\Theta$.
Its stable sides are formed by two curves contained in $W^s(P)$.
On the interior of each $\omega\in\mathcal P_1$, the value of $r$ is constant. This value is denoted by $r(\omega)$.
For each $\omega\in\mathcal P_n$
define its \emph{inducing time} $\tau(\omega)$ by
\begin{equation*}\label{tau}\tau(\omega)=\sum_{i=0}^{n-1}r(\omega_i).\end{equation*}
Clearly, the unstable sides of $f^{\tau(\omega)}\omega$ are formed by two curves in $\tilde\Gamma^u$.
Its stable sides are formed by two curves contained in the stable sides of $\Theta$ (See FIGURE 2).
The next bounded distortion result is contained in \cite[Lemma 2.15]{Tak13'}.
\begin{lemma}\label{global}
For any $\gamma^u\in\tilde\Gamma^u$ and any proper rectangle $\omega$,
$\gamma^u\cap\omega$ is a compact curve joining the stable sides of $\omega$. In addition,
$$\sup_{x,y\in\gamma^u\cap\omega}
\frac{\|D_yf^{\tau(\omega)}|E_y^u\|}{\|D_xf^{\tau(\omega)}|E_x^u\|} \leq
C|f^{\tau(\omega)}x-f^{\tau(\omega)}y|.$$
\end{lemma}
\begin{figure}
\begin{center}
\includegraphics[height=4cm,width=10cm]{multifigure1.eps}
\caption{The proper rectangles (shaded) in $\mathcal P_1$ with inducing time $n$ and their $f^n$-images}
\end{center}
\end{figure}
\subsection{Horseshoes}\label{horse}
We introduce a horseshoe structure which naturally comes from the inducing scheme in
Sect.\ref{induced map}. The reference for the contents in this subsection is \cite[Sect.2.10]{Tak13'}.
Let $\mathcal A$ be a finite collection of
proper rectangles contained in the interior of $\Theta$,
labeled with $1,2,\ldots,\ell=\#\mathcal A$. We assume any two elements of $\mathcal A$ are either disjoint,
or intersect each other only at their stable sides.
Endow $\Sigma_{\ell}=\{1,\ldots,\ell\}^{\mathbb Z}$ with the product topology
of the discrete topology,
and let $\sigma\colon\Sigma_{\ell}\circlearrowleft$ denote the left shift.
Define a coding map $\pi\colon\Sigma_{\ell}\to\mathbb R^2$ by
$\pi(\{x_i\}_{i\in\mathbb Z})=y$, where
$$\{y\}=\left(\bigcap_{k=1}^{\infty}\omega_k^s\right)
\cap\left(\bigcap_{k=1}^{\infty}\omega_k^u\right)$$
and
$$\omega^s_k=\omega_{x_0}\cap\left(\bigcap_{i=1}^k
f^{-\tau(\omega_{x_0})}\circ\cdots\circ f^{-\tau(\omega_{x_{i-1}})}\omega_{x_i}\right)\text{ and }
\omega^u_k=\bigcap_{i=1}^{k}f^{\tau(\omega_{x_{-1}})}\circ\cdots\circ f^{\tau(\omega_{x_{-i}})}\omega_{x_{-i}}.$$
\begin{lemma}\cite[Lemma 2.19]{Tak13'}\label{hyp}
The map $\pi$ is well-defined, continuous, injective, and satisfies $\pi(\Sigma_{\ell})\subset\Omega$.
\end{lemma}
\subsection{Bounded distortions}\label{bdddist}
We need two more distortion results for points which are slow recurrent to the critical set.
For $x\in\Omega^u$ define
$$d_{\rm crit}^u(x)=\begin{cases}|\zeta(x)-x|&\text{ if $x\in I(\delta)$};\\
1&\text{ otherwise,}\end{cases}$$
where $\zeta(x)$ is the critical point
on the $C^2(b)$-curve in $\tilde\Gamma^u$ containing $x$.
The function $d_{\rm crit}^u$ is a ``distance to the critical set".
For each $m\geq0$ define
$$G_m^u=
\{x\in\Omega^u\colon \text{$d_{\rm crit}^u(f^{n}x)> b^{\frac{n}{9}}$ for every $n\geq m$}\}.$$
The next lemma, the proof of which is a slight modification of \cite[Lemma 2.20]{Tak13'} and hence omitted here,
gives a distortion bound for derivatives along the unstable direction.
\begin{lemma}\label{lyapbdd}
For every $m\geq0$ there exists a constant $D_m>0$ such that for
any proper rectangle $\omega$ intersecting $G_m^u$ and $\tau(\omega)>m$,
$$\sup_{x,y\in\Omega\cap\omega}\frac{\|D_yf^{\tau(\omega)}|E_y^u\|}{ \|D_xf^{\tau(\omega)}|E_x^u\| } \leq D_m.$$
\end{lemma}
The next lemma gives a distortion bound for Birkhoff averages of H\"older continuous functions.
\begin{lemma}\label{variation}
If $\varphi\colon\Omega\to\mathbb R$ is H\"older continuous, then
for every $m\geq0$ there exists
$K_{m,\varphi}>0$ such that
for
any proper rectangle $\omega$ intersecting $G_m^u$ and $\tau(\omega)>m$,
$$\sup_{x,y\in\Omega\cap\omega}|S_{\tau(\omega)}\varphi(x)-S_{\tau(\omega)}\varphi(y)|\leq K_{m,\varphi}.$$
\end{lemma}
\begin{proof}
Let $x\in\Omega\cap\omega$. There exists a nearly horizontal $C^1$-curve, denoted by $\gamma^u(f^{\tau(\omega)}x)$ (called a {\it long unstable leaf through $f^{\tau(\omega)}x$} in \cite{Tak13'}),
which is contained in $f^{\tau(\omega)}\omega$, joins the stable sides of $\Theta$ and
satisfies
${\rm length}(f^{-n}\gamma)\leq C\rho_0^n$ for some $C>0$, $\rho_0\in(0,1)$ and all $n\geq0$.
Let $\hat x$ denote the point of intersection between $\gamma^u(f^{\tau(\omega)}x)$ and $\alpha_1^+$,
and set $x'=f^{-\tau(\omega)}\hat x$.
Let $y\in\Omega\cap\omega$ and define $y'$ in the same way.
Let $\theta\in(0,1]$ be a H\"older exponent of $\varphi$.
We have
\begin{equation*}
\sum_{n=0}^{\tau(\omega)-1}|f^nx-f^ny|^\theta\leq
\sum_{n=0}^{\tau(\omega)-1}|f^nx-f^nx'|^\theta+\sum_{n=0}^{\tau(\omega)-1}|f^nx'-f^ny'|^\theta+
\sum_{n=0}^{\tau(\omega)-1}|f^ny'-f^ny|^\theta.\end{equation*}
From the backward contraction, the first and the third summands are uniformly bounded.
For the second one,
by \cite[Lemma 2.18]{Tak13'} there exists $k\in[0,m]$ such that
the curve $f^{k-\tau(\omega)}\alpha_1^+$ is contracted exponentially by a factor $b^{\frac{1}{2}}$
under forward iteration (i.e., $f^{k-\tau(\omega)}\alpha_1^+$ is contained in a {\it long stable leaf} \cite[Sect.2.8]{Tak13'}).
Hence
\begin{equation*}
\sum_{n=0}^{\tau(\omega)-1}|f^nx-f^ny|^\theta\leq
2C^\theta\sum_{n=0}^{\infty}\rho_0^{\theta n}+\sum_{n=0}^{k-1}|f^nx'-f^ny'|^\theta+
C^\theta \sum_{n=0}^{\infty}b^{\frac{\theta n}{2}},\end{equation*}
which is bounded by
a uniform constant depending only on $b$, $m$, $\theta$.
Since $\omega$ and $x$, $y$ are arbitrary, the desired inequality follows.
\end{proof}
\noindent{\it Remark.
{\rm The function $x\in\Omega\mapsto \log \|D_xf|E_x^u\|$ is not covered by
Lemma \ref{variation} since it is not continuous at $Q$.}}
\subsection{Approximation of non ergodic measures by ergodic ones}\label{nonergod}
Let $\mathcal M^e(f)$ denote the set of $f$-invariant ergodic Borel probability measures.
\begin{lemma}\label{approximate}
For any continuous $\varphi\colon\Omega\to\mathbb R$, $\mu\in\mathcal M(f)$ and $\varepsilon>0$
there exists $\nu\in\mathcal M^e(f)$ such that
$h(\nu)>0$, $|h(\mu)- h(\nu)|<\varepsilon$, $|\lambda^u(\mu)-\lambda^u(\nu)|<\varepsilon$
and $|\int\varphi d\mu-\int\varphi d\nu|<\varepsilon$.
\end{lemma}
\begin{proof}
From \cite[Lemma 2.23]{Tak13'} there exists $\nu\in\mathcal M(f)$ such that
$h(\nu)>0$, $\nu\{Q\}=0$,
$|h(\mu)-h(\nu)|<\varepsilon$ and $|\lambda^u(\mu)-\lambda^u(\nu)|<\varepsilon.$
We note that
$f|\Omega$ is a factor of the full shift on two symbols
\cite[Proposition 3.1]{SenTak2}, and therefore has the specification
\cite[Proposition 1(b)]{Sig74}.
Hence, ergodic measures are entropy-dense
\cite{EizKifWei94}:
there exists a sequence
$\{\xi_n\}_n$ in $\mathcal M^e(f)$ such that $\xi_n\to\nu$ and $h(\xi_n)\to h(\nu)$ as $n\to\infty$.
By \cite[Lemma 4.4]{SenTak1} and $\nu\{Q\}=0$, we obtain $\lambda^u(\xi_n)\to\lambda^u(\nu)$.
\end{proof}
\section{Proof of the theorem}
In this section we complete the proof of the theorem.
\subsection{Outline}\label{outline}
For the rest of this paper we assume $\varphi\colon \Omega\to\mathbb R$ is continuous
and $\beta\in I_\varphi\setminus I_\varphi'$.
We prove the theorem by estimating $B_\varphi^u(\beta)$ from both sides.
This is done along the line of \cite{Tak13'},
but there is one key difference.
In Sect.\ref{lowest} we estimate $B_\varphi^u(\beta)$ from below,
by constructing a large subset of $\Omega_\varphi^u(\beta)$.
For the upper estimate, define
$$\Omega_*^u=\Omega^u\setminus\bigcup_{m=0}^\infty G_m^u,$$
and split $\Omega_\varphi^u(\beta)=\Pi_1\cup\Pi_2,$
where $$\Pi_1=\bigcup_{m=0}^\infty \Omega^u(\beta)\cap G_m^u\ \text{ and }\ \Pi_2=\Omega^u(\beta)\cap\Omega_*^u.$$
The upper estimate of $\dim_H^u(\Pi_1)$ is done in Sect.\ref{up1}
in much the same way as in \cite{Tak13'}.
The key difference from \cite{Tak13'} is that $\Pi_2$ can be nonempty.
To bypass this problem, in Sect.\ref{dimens} we estimate the dimension of the larger set $\Omega_*^u$
from above.
\subsection{Lower estimate of $B_\varphi^u(\beta)$}\label{lowest}
Define
\begin{equation}\label{Feps}
d^{u}_\varepsilon={\sup}\left\{\frac{h(\mu)}{\lambda^u(\mu)}\colon\mu\in\mathcal
M(f),\ \left|\int\varphi d\mu
-\beta\right|<\varepsilon\right\}.\end{equation}
We also define $d^{u,e}_\varepsilon$ by restricting the range of the supremum to $\mathcal M^e(f)$.
We shall show
\begin{equation}\label{lowe}
B_\varphi^u(\beta)\geq \displaystyle{\lim_{\varepsilon\to0}d^{u,e}_\varepsilon}.\end{equation}
Since
$\displaystyle{\lim_{\varepsilon\to0}d^u_\varepsilon=
\lim_{\varepsilon\to0}d^{u,e}_\varepsilon}$ from Lemma \ref{approximate},
the desired lower estimate of $B_\varphi^u(\beta)$ follows.
The idea is to construct a sequence of horseshoes in the sense of Sect.\ref{horse} with Birkhoff averages arbitrarily close to $\beta$,
and then glue these horseshoes together to construct a set of points whose Birkhoff averages are precisely $\beta$.
Let $\{\mu_n\}_n$ be a sequence in $\mathcal M^e(f)$ such that
$|\int\varphi d\mu_n-\beta|\to0$
and $h(\mu_n)/\lambda^u(\mu_n)$ converges as $n\to\infty$.
Since $\varphi$ is continuous and $\mathcal M(f)$ is compact with respect to the topology of weak convergence,
$c_\varphi=\min\{\int\varphi d\mu\colon\mu\in\mathcal M(f)\}$ and
$d_\varphi=\max\{\int\varphi d\mu\colon\mu\in\mathcal M(f)\}$.
Since
both $c_\varphi$ and $d_\varphi$ are attained by elements of $\mathcal M^e(f)$,
considering linear combinations of them and then
using Lemma \ref{approximate} one can show there indeed exists such a sequence.
If $h(\mu_n)\to0$, then $h(\mu_n)/\lambda^u(\mu_n)\to0$ from
$\inf\{\lambda^u(\mu)\colon\mu\in\mathcal M(f)\}>0$, and so \eqref{lowe} is obvious.
Hence we assume $h(\mu_n)>0$.
In what follows we first assume $\varphi$ is H\"older continuous, and prove \eqref{lowe}.
Lastly we indicate necessary minor modifications to treat merely continuous $\varphi$.
If $\varphi$ is H\"older continuous, then
slightly modifying the proof of \cite[Lemma 2.21]{Tak13'} one can prove a variant of well-known Katok's theorem \cite[Theorem S.5.9]{KatHas95}:
for each $n$ there exist a positive integer $q_n$ and a family $\mathcal R_n$
of proper rectangles with the following properties:
\begin{itemize}
\item[(i)] for each $\omega\in\mathcal R_n$,
$\tau(\omega)=q_n$;
\item[(ii)] $|(1/q_n)\log \#\mathcal R_n-h(\mu_n)|<1/n;$
\item[(iii)] for any $x\in \bigcup_{\omega\in\mathcal R_n} \Omega\cap\omega$,
$\left|(1/q_n)S_{q_n}\log J^u(x)-\lambda^u(\mu_n)\right|<1/n$;
\item[(iv)] for any $x\in \bigcup_{\omega\in\mathcal R_n} \Omega\cap\omega$,
$\left|(1/q_n)S_{q_n}\varphi(x)-\int \varphi d\mu_n\right|<1/n$.
\end{itemize}
The only one difference from \cite[Lemma 2.21]{Tak13'} is (iv), which follows from Lemma \ref{variation}.
The rest of the proof proceeds much in parallel to that of \cite{Tak13'}, and so we only give a sketch of the proof.
For an integer $\kappa\geq1$ let
$$\mathcal R_n(\kappa)=\{\omega_0\cap f^{-q_n}\omega_1\cap\cdots\cap f^{-(\kappa-1)q_n}\omega_{\kappa-1}
\colon\omega_1,\ldots,\omega_{\kappa-1}\in\mathcal R_n\}.$$
Let $\{\kappa_n\}_{n=1}^\infty$ be a sequence of positive integers.
For each $k\geq1$ let $(N,s)$ be a pair of integers such that
$
k=\kappa_1+\kappa_2+\cdots+\kappa_{N-1}+s \ \text{and}\ \ 0\leq s< \kappa_{N}.$
Define $\mathcal S(k)$ to be the collection of all proper rectangles of the form
$$\omega_0\cap f^{-\kappa_1q_1}\omega_1\cap
\cdots\cap f^{-\kappa_1q_1-\cdots-\kappa_{N-1}q_{N-1}}\omega_{N},$$
where $\omega_n\in\mathcal R_{n}(\kappa_{n+1})$ $(n=0,\ldots,N-1)$ and
$\omega_{N}\in\mathcal R_{N}(s)$.
The set $\bigcup_{\omega\in\mathcal S(k)}\omega$ is compact, and
decreasing in $k$.
Set
$$Z=\gamma^u(\zeta_0)\cap \bigcap_{k=1}^\infty\bigcup_{\omega\in\mathcal S(k)}\omega,$$
where $\gamma^u(\zeta_0)$ denotes the unstable side of $\Theta$ containing $\zeta_0$.
By appropriately choosing $\{\kappa_n\}_{n=1}^\infty$ so that the orbits of points in $Z$ spend longer and longer times
around the horseshoes as $n$ increases,
one can make sure that
$Z\subset \Omega_\varphi^u(\beta)$ and
\begin{equation*}
\dim_H^u(Z)\geq \lim_{n\to\infty}\frac{h(\mu_n)}{\lambda^u(\mu_n)}.\end{equation*}
Since $\{\mu_n\}_n$ is arbitrary, \eqref{lowe} holds.
If $\varphi$ is merely continuous, then take a sequence $\{\varphi_n\}_n$ of real-valued H\"older continuous
functions on $\Omega$
such that $\sup\{|\varphi(x)-\varphi_n(x)|\colon x\in\Omega\}<1/n$.
Find $q_n$ and $\mathcal R_n$ as above, satisfying (i) (ii) (iii) and (iv)
with $\varphi_n$ in the place of $\varphi$.
Then
\begin{align*}\left|\frac{1}{q_n}S_{q_n}\varphi(x)-\int \varphi d\mu_n\right|<&\left|\frac{1}{q_n}(S_{q_n}\varphi(x)- S_{q_n}\varphi_n(x)) \right|\\
& +\left|\frac{1}{q_n}S_{q_n}\varphi_n(x)-\int \varphi_n d\mu_n\right|+\left|\int\varphi_n -\varphi d\mu_n\right|\\
<&\frac{1}{n}+\frac{1}{n}+\frac{1}{n},\end{align*}
and so the same argument prevails. \qed
\subsection{Upper estimate of $\dim_H^u(\Pi_1)$.}\label{up1}
From the next proposition and the countable stability of $\dim_H^u$,
we obtain $\dim_H^u(\Pi_1)\leq \displaystyle{\lim_{\varepsilon\to0}d^u_\varepsilon}$.
\begin{prop}\label{c}
For every $m\geq0$,
$\dim_H^u(\Omega^u(\beta)\cap G_m^u)\leq \displaystyle{\lim_{\varepsilon\to0}d^u_\varepsilon}.$
\end{prop}
\begin{proof}
Recall that $\gamma^u(\zeta_0)$ denotes the unstable side of $\Theta$ containing $\zeta_0$.
Since $\gamma^u(\zeta_0)$ contains a fundamental domain in $W^u$,
for any $x\in \Omega_\varphi^u(\beta)$ which is not the fixed point in $W^u$ there exists $n\in\mathbb Z$ such that
$f^nx\in\gamma^u(\zeta_0)$.
From the countable stability and the $f$-invariance
of $\dim_H^u$,
$B_\varphi^u(\beta)=
\dim_H ^u(\Omega_\varphi^u(\beta)\cap\gamma^u(\zeta_0))$.
Set
$$\tilde\Omega_\varphi^u(\beta)=\{x\in\Omega_\varphi^u(\beta)\cap \gamma^u(\zeta_0)\colon f^nx\in\Theta\text{ for infinitely many $n>0$}\}.$$
Since points in $\Omega_\varphi^u(\beta)\cap\gamma^u(\zeta_0)$ which return to $\Theta$ under forward iteration
only finitely many times form a countable subset,
we have $B_\varphi^u(\beta)=\dim_H^u (\tilde\Omega_\varphi^u(\beta))$.
From now on we restrict ourselves to $\tilde\Omega_\varphi^u(\beta)$.
For $c>0$ let $D_c(\zeta_0)$ denote the closed ball in $\gamma^u(\zeta_0)$ of
radius $c$ about $\zeta_0$. Define
$$\mathcal A_{n,\varepsilon}=\left\{
\omega\in\mathcal P_{n}\colon \omega\cap G_m^u\neq\emptyset,\ \omega\cap
D_c(\zeta_0)=\emptyset,\
\inf_{x\in\omega\cap\gamma^u(\zeta_0)}\left|\frac{1}{\tau(\omega)}S_{\tau(\omega)}\varphi(x)-\beta\right|<\frac{\varepsilon}{2}\right\}.$$
Observe that $\mathcal A_{n,\varepsilon}$ is a finite set, because its elements
do not intersect $D_c(\zeta_0)$.
For each $\omega\in\mathcal A_{n,\varepsilon}$
write $\omega^u=\omega\cap\gamma^u(\zeta_0)$ and set
$\mathcal A^u_{n,\varepsilon}=\{\omega^u\colon\omega\in\mathcal A_{n,\varepsilon}\}$.
Clearly we have
$$(\tilde\Omega^u(\beta)\cap G_m^u)\setminus D_c(\zeta_0)\subset\limsup_{n\to\infty}\bigcup_{
\omega^u\in\mathcal A_{n,\varepsilon}^u}\omega^u,$$
and there exist $C>0$ and $\rho_1\in(0,1)$ such that
for each $\omega\in\mathcal A_{n,\varepsilon}^u$,
$${\rm length}(\omega^u)\leq C\rho_1^n.$$
It is enough to show
\begin{equation}\label{pperd}
\limsup_{n\to\infty}\frac{1}{n}\log\sum_{\omega^u\in\mathcal A^u_{n,\varepsilon}}{\rm length}(\omega^u)^{ d^u_\varepsilon}\leq 0\ \text{ for any $\varepsilon>0$.}
\end{equation}
Indeed, if this holds, then
for any $d>0$ we have
$$\limsup_{n\to\infty}\frac{1}{n}\log\sum_{A\in\mathcal A^u_{n,\varepsilon}}{\rm length}(\omega^u)^{ d^u_\varepsilon+d}\leq-d\log\rho_1.$$
It follows
that $\sum_{A\in\mathcal A^u_{n,\varepsilon}}{\rm length}(\omega^u)^{ d^u_\varepsilon+d}$
has a negative growth rate as $n$ increases.
Therefore
the Hausdorff $( d^u_\varepsilon+d)$-measure of the set $(\tilde\Omega^u(\beta)\cap G_m^u)\setminus
D_c(\zeta_0)$
is $0$. Since $d>0$ is arbitrary,
$\dim_H^u((\tilde\Omega^u(\beta)\cap G_m^u)\setminus D_c(\zeta_0))\leq d^u_\varepsilon$, and
by the countable stability of $\dim_H^u$ we obtain
$\dim_H^u(\tilde\Omega^u(\beta)\cap G_m^u)\leq d^u_\varepsilon$.
Letting
$\varepsilon\to0$ yields the desired inequality in Proposition \ref{c}.
\medskip
It is left to prove \eqref{pperd}.
Set $\ell=\#\mathcal A_{n,\varepsilon}$ and
Write
$\mathcal A_{n,\varepsilon}=\{\omega{(1)},\omega{(2)},\ldots,\omega{(\ell)}\}$
so that
\begin{equation}\label{align}
\tau(\omega{(1)})\geq\tau(\omega{(s)})>m\ \text{ for every }s\in \{1,2,\ldots,t\}.\end{equation}
Let $\pi_{\ell}\colon\Sigma_\ell\to\bigcup_{\omega\in\mathcal A_{n,\varepsilon}}\omega$ denote the coding map defined in Sect.\ref{horse}
and $\sigma\colon\Sigma_\ell\circlearrowleft$ the left shift.
Define
$$B=\{\underline{a}\in\Sigma_\ell\colon\pi\underline{a}\subset W^s(P)\setminus\{P\}\}.$$
Proper rectangles can intersect each other only at their stable sides, and there is only one proper rectangle containing $P$
in its stable side. Hence,
for any $\underline{a}\in \Sigma_\ell\setminus B$ there exists a unique element
of $\mathcal A_{n,\varepsilon}$ containing $\pi\underline{a}$ which we denote by
$\omega(\underline{a})$.
Define
$\Psi\colon \Sigma_\ell\setminus B\to\mathbb R$ by
$$\Psi(\underline{a})=- d^u_\varepsilon\sum_{i=0}^{\tau(\omega(\underline{a}))-1}\log J^u(f^i(\pi\underline{a})).$$
Since $\pi(\Sigma_\ell)\subset\Omega\setminus\{Q\}$ and $\log J^u$ is continuous except at $Q$,
$\Psi$ is continuous.
Let
$\mathcal M(\sigma)$ denote the space of $\sigma$-invariant Borel probability measures on $\Sigma_\ell$
endowed with the topology of weak convergence.
For each $k\geq 1$ define an atomic probability measure $\nu_k\in\mathcal M(\sigma)$ concentrated on
the set $E_k=\{\underline{a}\in \Sigma_\ell\colon \sigma^{k}\underline{a}=\underline{a}\}$ by
$$\nu_k=
\left(\sum_{\underline{b}\in E_k}\exp\left( S_{k}\Psi(\underline{b}) \right)\right)^{-1}\sum_{\underline{a}\in E_k}
\exp\left(S_{k}\Psi(\underline{a})\right)
\delta_{\underline{a}},$$
where $S_{k}\Psi=\sum_{i=0}^{k-1}\Psi\circ\sigma^i$ and
$\delta_{\underline{a}}$ denotes the Dirac measure at $\underline{a}$.
Let $\nu_0$ denote an accumulation point of the
sequence $\{\nu_k\}_k$ in $\mathcal M(\sigma)$. Taking
a subsequence if necessary we may assume $\nu_k\to\nu_0$.
We have $\nu_0\in\mathcal M(\sigma)$.
Define a Borel probability measure $\overline{\mu}$ on $\pi(\Sigma_\ell)$ by
$$\overline{\mu}=\sum_{\omega\in\mathcal A_{n,\varepsilon}}\nu_0|_{\pi^{-1}\omega}.$$
By \cite[Sublemma 3.5]{Tak13'},
$\nu_0(B)=0$ and so
$\overline{\mu}$ is indeed a probability.
Define $\mu\in\mathcal M(f)$ by
$$\mu=\left(\sum_{\omega\in\mathcal A_{n,\varepsilon}}
\tau(\omega)\overline{\mu}(\omega)\right)^{-1}\sum_{\omega\in\mathcal A_{n,\varepsilon}}\sum_{i=0}^{\tau(\omega)-1}(f^i)_*(\overline{\mu}|_{\omega}).$$
We show
\begin{equation}\label{minus}
h(\mu)- d^u_\varepsilon\lambda^u(\mu)\leq0.
\end{equation}
To show this,
let $\omega\in\mathcal A_{n,\varepsilon}$ and
$x\in \omega$. Choose
$y\in \omega\cap\gamma^u(\zeta_0)$ such that
$\left|(1/\tau(\omega))S_{\tau(\omega)}\varphi(y)-\beta\right|<\varepsilon/2.$
If $\varphi$ is H\"older continuous, then by Lemma \ref{lyapbdd}
and $\tau(\omega)\geq 2n$
we have
\begin{align*}\left|\frac{1}{\tau(\omega)}S_{\tau(\omega)}\varphi(x)-\beta\right|\leq&
\left| \frac{1}{\tau(\omega)}(S_{\tau(\omega)}\varphi(x)-S_{\tau(\omega)}\varphi(y))\right|+
\left|\frac{1}{\tau(\omega)}S_{\tau(\omega)}\varphi(y)-\beta\right|\\
\leq& \frac{K_{m,\varphi}}{\tau(\omega)}+\frac{\varepsilon}{2}\leq \frac{K_{m,\varphi}}{2n}+\frac{\varepsilon}{2} <\varepsilon.\end{align*}
If $\varphi$ is merely continuous, then approximating $\varphi$ by a H\"older continuous
function we get the same inequality for sufficiently large $n$.
Since $\omega\in\mathcal A_{n,\varepsilon}$ and $x\in\omega$ are arbitrary, this implies
$ |\int \varphi d\mu-\beta|<\varepsilon$.
Then \eqref{minus} follows from the definition of $d^u_\varepsilon$ in \eqref{Feps}.
Observe that
\begin{align*}
\log\sum_{\underline{a}\in E_k}\exp(S_{k}\Psi(\underline{a}))
=-\sum_{\underline{a}\in E_k}\nu_k(\{\underline{a}\})
\log\nu_k(\{\underline{a}\})+k\int\Psi d\nu_k.\end{align*}
A slight modification of the argument in \cite[pp.220]{Wal82} shows that for
any integer $p$ with $1\leq p<k$,
\begin{equation}\label{x}
\frac{1}{k}\log\sum_{\underline{a}\in E_k}\exp(S_{k}\Psi(\underline{a}))\leq-\frac{1}{p}\sum_{\underline{a}\in E_p}\nu_k(\{\underline{a}\})
\log\nu_k(\{\underline{a}\})+\int\Psi d\nu_k
+\frac{2p\log\#E_p}{k}.\end{equation}
Similarly to the proof of \cite[Sublemma 3.7]{Tak13'} one can show that
$\int\Psi d\nu_k\to\int \Psi d\nu_0$ as $k\to\infty$.
Letting $k\to\infty$ in \eqref{x},
\begin{equation*}
\limsup_{k\to\infty}\frac{1}{k}\log\sum_{\underline{a}\in E_k}\exp(S_{k}\Psi(\underline{a}))\leq
-\frac{1}{p}\sum_{\underline{a}\in E_p}\nu_0(\{\underline{a}\})
\log\nu_0(\{\underline{a}\})+\int\Psi d\nu_0.
\end{equation*}
Letting $p\to\infty$ we get
\begin{equation}\label{lem2}
\limsup_{k\to\infty}\frac{1}{k}\log\sum_{\underline{a}\in E_k}\exp(S_{k}\Psi(\underline{a}))\leq
h(\sigma;\nu_0)+\int\Psi d\nu_0,
\end{equation}
where $h(\sigma;\nu_0)$ denote the entropy of $\nu_0\in\mathcal M(\sigma)$.
To estimate the left-hand-side of \eqref{lem2} from below, set $E_k'=\{\underline{a}\in E_k\colon a_0=1\}.$
Let
$\underline{a}\in E_k'$, $\underline{b}\in E_{k-1}'$ be such that
$a_i=b_i$ for every $0\leq i< k-1$.
Slightly modifying the proof of \cite[Sublemma 3.8]{Tak13'} one can show that
$\pi\underline{a}$ and $\pi\underline{b}$ are contained in the same proper rectangle
with inducing time $>m$ and intersecting $G_m^u$.
Lemma \ref{lyapbdd} gives
$$\frac{\exp(S_{k}\Psi(\underline{a}))}{ \exp(S_{k-1}\Psi(\underline{b})) }=\frac{\exp(S_{k-1}\Psi(\underline{a}))}{ \exp(S_{k-1}\Psi(\underline{b})) }\cdot\exp(S_{0}\Psi(\sigma^{k-1}\underline{a}))\geq D_m^{-d^u_\varepsilon}\cdot D_m^{-2d^u_\varepsilon}{\rm length}(\omega^u(a_{k-1}))^{d^u_\varepsilon}.$$
Using this inequality repeatedly gives
\begin{align*}
\sum_{\underline{a}\in E_k}\exp(S_k\Psi(\underline{a}))&>
\sum_{\underline{a}\in E_k'} \exp(S_{k}\Psi(\underline{a}))=
\sum_{\underline{b}\in E_{k-1}'} \exp(S_{k-1}\Psi(\underline{b}))\sum_{\stackrel{\underline{a}\in E_k'}
{a_i=b_i\ 0\leq \forall i< k-1}}\frac{ \exp(S_{k}\Psi(\underline{a}))}
{ \exp(S_{k-1}\Psi(\underline{b}))}\\
&\geq\sum_{\underline{b}\in E_{k-1}'} \exp(S_{k-1}\Psi(\underline{b}))\cdot D_m^{-3d^u_\varepsilon}\sum_{\omega\in\mathcal A_{n,\varepsilon}}
{\rm length}(\omega^u)^{ d^u_\varepsilon}\\
&\geq\cdots\geq
\sum_{\underline{b}\in E_{1}'} \exp(S_{0}\Psi(\underline{b}))\left(D_m^{-3d^u_\varepsilon}\sum_{\omega\in\mathcal A_{n,\varepsilon}}
{\rm length}(\omega^u)^{d^u_\varepsilon}\right)^{k-1}\\
&\geq \left(D_m^{-3d^u_\varepsilon}\sum_{\omega\in\mathcal A_{n,\varepsilon}}
{\rm length}(\omega^u)^{d^u_\varepsilon}\right)^{k}.\end{align*}
Hence
\begin{equation}\label{lem1}\liminf_{k\to\infty}\frac{1}{k}
\log\sum_{\underline{a}\in E_k}\exp(S_k\Psi(\underline{a}))\geq\log\sum_{\omega^u\in\mathcal A^u_{n,\varepsilon}}
{\rm length}(\omega^u)^{d^u_\varepsilon}- 3d^u_\varepsilon\log D_m.\end{equation}
Putting \eqref{lem2} \eqref{lem1} together and then using \eqref{minus} yield
\begin{align*}
\frac{1}{n}\log\sum_{\omega\in\mathcal A^u_{n,\varepsilon}}
{\rm length}(\omega^u)^{d^u_\varepsilon}&\leq
\frac{1}{n}\left(h(\sigma;\nu_0)+\int \Psi d\nu_0\right)+\frac{3}{n} d^u_\varepsilon\log D_m\\
&=\frac{1}{n}(h(\mu)- d^u_\varepsilon\lambda^u(\mu))\sum_{\omega\in\mathcal A_{n,\varepsilon}}\tau(\omega)\overline{\mu}(\omega)+ \frac{3}{n} d^u_\varepsilon\log D_m\\
& \leq \frac{3}{n} d^u_\varepsilon\log D_m.\end{align*}
This implies \eqref{pperd}, and hence finishes the proof of Proposition \ref{c}.
\end{proof}
\subsection{Upper estimate of $\dim_H^u(\Omega_*^u)$}\label{dimens}
We finish by proving the next
\begin{prop}\label{new}
$\dim_H^u(\Omega_*^u)\leq 2/\log (1/b)$.
\end{prop}
\begin{proof}
If $x\in\Omega_*^u$, then there exist infinitely many $n\geq0$ such that
$d_{\rm crit}(f^nx)\leq b^{\frac{n}{9}}.$
Define a sequence $k_i=k_i(x)$ $(i=1,2,\ldots)$ of positive integers inductively as follows:
$k_1=\min\{n>0\colon d_{\rm crit}(f^nx)\leq b^{\frac{n}{9}}\}.$
Given $k_1,\ldots,k_{i}$ with $d_{\rm crit}(f^{k_1+\cdots+k_i}x)\leq b^{\frac{k_i}{9}}$,
define $k_{i+1}=\min\{n>0\colon d_{\rm crit}(f^{k_1+\cdots+k_i+n}x)\leq b^{\frac{n}{9}}\}.$
Define $\eta=\eta(b)\gg1$ by
\begin{equation}
\eta=\left[-\frac{1}{20}\log b\right],
\end{equation}
where $[$ $\cdot$ $]$ denotes the integer part.
Since $b^{\frac{k_i}{9}}\cdot \|Df^{2\eta k_i}\|< b^{\frac{k_i}{9}}\cdot 5^{2\eta k_i} \ll1$,
$f^{k_1+\cdots+k_i}x$ shadows the forward orbit of the binding critical point
at least up to time $2\eta k_i$, namely
\begin{equation}\label{d1}
k_{i+1}(x)\geq 2\eta k_i(x).\end{equation}
From \eqref{d1} and $k_1(x)>1$
we get $k_{i}(x)\geq(2\eta)^{i-1}$, and
\begin{equation}\label{s2}k_{1}(x)+k_{2}(x)+\cdots+k_{i}(x)\geq
\eta^{i-1}.\end{equation}
Now, given a sequence $\{l_i\}_{i=1}^\infty$ of positive integers, define a collection
$\mathcal Q(l_1,l_2,\ldots,l_i)$
of pairwise disjoint compact curves in $\gamma^u(\zeta_0)$ inductively as follows. Start with
$$\mathcal Q(l_1)= \{\gamma_{1}\subset\gamma^u(\zeta_0)\colon f^{l_1}\gamma_1\in\tilde\Gamma^u\text{ and $k_1(x)=l_1$ for some
$x\in\gamma_1\cap\Omega_*^u$}\}.$$
Given $\mathcal Q(l_1,\ldots,l_{i})$,
for
each $\gamma_{i}\in\mathcal Q(l_1,\ldots,l_{i})$ set
$$\mathcal Q(\gamma_{i},l_{i+1})= \{\gamma_{i+1}\subset\gamma_{i}\colon
f^{l_1+\cdots+l_{i}+l_{i+1}}\gamma_{i+1}\in\tilde\Gamma^u\text{ and $k_{i+1}(x)=l_{i+1}$ for some
$x\in\gamma_{i+1}\cap\Omega_*^u$}\},$$
and define
$$\mathcal Q(l_1,\ldots,l_{i+1})=
\bigcup_{\gamma_{i}\in \mathcal Q(l_1,\ldots,l_{i}) }
\mathcal Q(\gamma_{i},l_{i+1}).$$
Obviously,
\begin{equation}\label{cardinal}
\mathcal Q(\gamma_{i},l_{i+1})<2^{l_{i+1}}.
\end{equation}
If $x\in\Omega_*^u$, then
for each $i=1,2,\ldots$ there exists a unique element of $\mathcal Q(k_1(x),k_2(x),\ldots,k_i(x))$
containing $x$.
Hence
$$\Omega_*^u\subset
\bigcup_{L=\eta^{i-1}}^\infty\bigcup_{l_1+\cdots+l_i=L}\bigcup_{\gamma_i\in \mathcal Q(l_1,\ldots,l_i)}\gamma_{i}.$$
Now, let $1\leq p<q$ and define
$$\Omega_*^{(p)}=\{f^{k_1(x)+k_2(x)+\cdots+k_p(x)}x\colon x\in\Omega_*^u\}.$$
If $x\in\Omega_*^{(p)}$, then
$k_{p+1}(x)+k_{p+2}(x)+\cdots+k_{q}(x)\geq
\eta^{q-1}.$
Hence
$$\Omega_*^{(p)}\subset
\bigcup_{\gamma_{q}\in \mathcal Q(l_1,\ldots, l_q)}f^{l_1+\cdots+l_p}\gamma_{q}=
\bigcup_{L=\eta^{q-1}}^\infty\bigcup_{l_{p+1}+\cdots+l_q=L}\bigcup_{\gamma_{q}\in \mathcal Q(l_1,\ldots,l_q)}f^{l_1+\cdots+l_p}\gamma_{q}.$$
From the countable stability and the $f$-invariance
of $\dim_H^u$,
$\dim_H^u(\Omega_*^u)=\dim_H^u(\Omega_*^{(p)}).$
To get a better estimate, we shall work with large $p$.
Let $d\in(2/\log (1/b),1)$. For each $i\geq p$ we have
\begin{align*}
\sum_{\gamma_{i+1}\in\mathcal Q(l_1,\ldots,l_{i+1})}
{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i+1})^d&=\sum_{\gamma_{i}\in
\mathcal Q(l_1,\ldots,l_{i})}
{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i})^d\\
&\times\sum_{\gamma_{i+1}\in\mathcal
Q(\gamma_{i},l_{i+1}) } \frac{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i+1})^d}
{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i})^d}.\end{align*}
On the second sum of the fractions,
we have $f^{k_1+\cdots+k_{i}}\gamma_{i}\in\tilde\Gamma^u$ and
${\rm length}(f^{k_1+\cdots+k_{i}}\gamma_{i+1})< 2b^{\frac{k_{i+1}}{9}}.$
From this and the bounded distortion in Lemma \ref{global},
\begin{equation}\label{d4}\frac{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i+1})}{{\rm length}(f^{l_1+\cdots+l_p}\gamma_i)}\leq
C\cdot\frac{{\rm length}(f^{l_1+\cdots+l_{i}}\gamma_{i+1})}{{\rm length}(f^{l_1+\cdots+l_{i}}\gamma_i)}\leq
3b^{\frac{l_{i+1}}{9}}.\end{equation}
Using \eqref{cardinal} \eqref{d4} and $d\in(2/\log (1/b),1)$,
$$\sum_{\gamma_{i+1}\in\mathcal Q(\gamma_{i},l_{i+1})
}\frac{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i+1})^d}{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i})^d} \leq
\#\mathcal Q(\gamma_{i},l_{i+1}) 3^db^{\frac{dl_{i+1}}{10}}<b^{\frac{dl_{i+1}}{20}}.$$
Plugging this
into the right-hand-side of the above equality we get
\begin{equation}\label{d-1}
\sum_{\gamma_i\in\mathcal Q(l_1,\ldots,l_{i+1})}{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i+1})^d
\leq b^{\frac{d l_{i+1}}{20}}
\sum_{\gamma_{i}\in\mathcal Q(l_1,\ldots,l_{i})}{\rm length}(f^{l_1+\cdots+l_p}\gamma_{i})^d
.\end{equation}
Using \eqref{d-1} inductively
yields
\begin{equation*}
\sum_{\gamma_q\in\mathcal Q(l_1,\ldots,l_q)}{\rm length}(f^{l_1+\cdots+l_p}\gamma_q)^d
\leq L_pb^{\frac{d}{20}(l_{p+1}+\cdots+l_q)},\end{equation*}
where $$L_p=\sum_{\gamma_{p+1}\in\mathcal Q(\gamma_p,l_{p+1})
}\frac{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{p+1})^d}{{\rm length}(f^{l_1+\cdots+l_p}\gamma_{p})^d}.$$
Hence
\begin{align*}
\sum_{L=\eta^{q-1}}^\infty\sum_{l_{p+1}+\cdots+l_q=L}\sum_{\gamma_q\in\mathcal
Q(l_1,\ldots,l_q)}{\rm length}(f^{l_1+\cdots+l_p}\gamma_q)^d&\leq
L_p\sum_{L=\eta^{q-1}}^\infty
b^{\frac{dL}{20}}\#\left\{
(l_{p+1},\ldots,l_{q})\colon \sum_{i=p+1}^q l_i=L\right\}.
\end{align*}
To estimate the right-hand side we use the following from Stirling's formula for factorials: for sufficiently small $\chi>0$ there exist $c(\chi)>0$ with $c(\chi)\to0$ as $\chi\to0$ such that for any two positive integers $p$, $q$ with $q/p\leq\chi$ one has
$\left(\begin{smallmatrix}p+q\\q\end{smallmatrix}\right)\leq e^{c(\chi)p}$.
The number of all
feasible $(l_{p+1},\ldots,l_{q})$ with $\sum_{i=p+1}^{q} l_{i}=L$ is bounded by the
number of ways of dividing $L$ objects into $q-p$ groups, which is
$\left(\begin{smallmatrix}L+q-p\\q-p\end{smallmatrix}\right)$.
Since $l_i\geq\eta^p$ for $i=p+1,\ldots,q$, we have $(q-p)/L\leq \eta^{-p}$,
which goes to $0$ as $p\to\infty$.
In particular,
there exists $p_0$ such that for all $p$, $q$ with $p_0\leq p<q$,
$$\#\left\{
(l_{p+1},\ldots,l_{q})\colon \sum_{i=p+1}^q l_{i}=L\right\}\leq \begin{pmatrix}L+q-p\\q-p\end{pmatrix}
\leq b^{-\frac{dL}{30}}.$$
\begin{align*}
\sum_{L=\eta^{q-1}}^\infty\sum_{l_{p+1}+\cdots+l_{q}=L}\sum_{\omega_q\in\mathcal
Q(l_1,\ldots,l_q)}{\rm length}(f^{l_1+\cdots+l_p}\gamma_q)^d
&\leq
L_p\sum_{L=\eta^{q-1}}^\infty b^{\frac{bdL}{60}}.
\end{align*}
The summand of the right-hand-side decays exponentially
in $q$, and so the
Hausdorff $d$-measure of $\Omega_*^{(p)}$ is zero.\end{proof}
From the lower estimate in Sect.\ref{lowest}, Proposition \ref{c} and Proposition \ref{new} we obtain
$$\lim_{\varepsilon\to0}d^u_\varepsilon\leq B_\varphi^u(\beta)=\max\{\dim_H^u(\Pi_1),\dim_H^u(\Pi_2)\}\leq\max\{\lim_{\varepsilon\to0}d^u_\varepsilon,2/\log(1/b)\}.$$
Since $I_\varphi\setminus I_\varphi'$ we have $B_\varphi^u(\beta)>2/\log(1/b)$, and so the above two inequalities are equalities.
This completes the proof of the theorem. $\qed$
\subsection*{Acknowledgments}
Partially supported by the Grant-in-Aid for Young Scientists (B) of the JSPS, Grant No.23740121.
\bibliographystyle{amsplain}
|
\section{Introduction}
There has been an explosion of papers recently dealing with pattern containment and avoidance in various combinatorial structures. And the study of statistics on combinatorial objects has a long and venerable history. By comparison, there are relatively few papers which study a variety of statistics on a number of different avoidance classes. The focus of the present work is pattern avoidance in set partitions combined with four important statistics defined by Wachs and White~\cite{ww:pqs}. It is the first comprehensive study of these statistics on avoidance classes. In particular, we consider the distribution of these statistics over every class avoiding a set of partitions of $\{1,2,3\}$.
We will start by providing the necessary definitions and setting notation.
A {\em set partition} of a set $S$ is a collection $\sigma$ of nonempty subsets whose disjoint union is $S$. We write $\sigma=B_1/\dots/B_k\vdash S$ where the subsets $B_i$ are called {\em blocks}. When no confusion will result, we often drop the curly braces and commas in the $B_i$.
For $[n]=\{1,\dots,n\}$, we use the notation
$$
\Pi_n=\{\sigma\ :\ \sigma\vdash[n]\}.
$$
To define pattern avoidance in this setting, suppose $\sigma=B_1/\dots/B_k\in\Pi_n$ and $S\subseteq[n]$. Then $\sigma$ has a corresponding {\em subpartition} $\sigma'$ whose blocks are the nonempty intersections $B_i\cap S$. For example, if $\sigma=14/236/5\vdash[6]$ and $S=\{2,4,6\}$ then $\sigma'=26/4$. We {\em standardize} a set partition with integral elements by replacing the smallest element by $1$, the next smallest by $2$, and so forth. So the standardization of $\sigma'$ above is $13/2$. Given two set partitions $\sigma$ and $\pi$, we say that $\sigma$ {\em contains $\pi$ as a pattern} if there is a subpartition of $\sigma$ which standardizes to $\pi$. Otherwise we say that $\sigma$ {\em avoids} $\pi$. Continuing our example, we have already shown that $\sigma=14/236/5$ contains $13/2$. But $\sigma$ avoids $123/4$ because the only block of $\sigma$ containing three elements also contains the largest element in $\sigma$, so there can be no larger element in a separate block.
We let
$$
\Pi_n(\pi)=\{\sigma\in\Pi_n\ :\ \text{$\sigma$ avoids $\pi$}\}.
$$
In order to connect set partitions with the statistics of Wachs and White, we will have to convert them into restricted growth functions. A {\em restricted growth function} (RGF) is a sequence $w=a_1\dots a_n$ of positive integers subject to the restrictions
\begin{enumerate}
\item $a_1=1$, and
\item for $i\ge2$ we have
\begin{equation}
\label{RGF}
a_i\le 1+\max\{a_1,\dots,a_{i-1}\}.
\end{equation}
\end{enumerate}
The number of elements in $w$ is called its {\em length} and we let
$$
R_n=\{w\ :\ \text{$w$ is an RGF of length $n$}\}.
$$
There is a simple bijection $\Pi_n\rightarrow R_n$. We say $\sigma=B_1/\dots/B_k\in\Pi_n$ is in {\em standard form} if $\min B_1<\dots<\min B_k$. Note that this forces $\min B_1=1$. We henceforth assume all partitions in $\Pi_n$ are written in standard form. Associate with $\sigma$ the word $w(\sigma)=a_1\dots a_n$ where
$$
\text{$a_i = j$ if and only if $i\in B_j$.}
$$
Using the example from the previous paragraph $w(\sigma)=122132$. It is easy to see that $w(\sigma)$ is a restricted growth function and that the map $\sigma\mapsto w(\sigma)$ is the desired bijection. It will be useful to have a notation for the RGFs of partitions avoiding a given pattern $\pi$, namely
$$
R_n(\pi)=\{w(\sigma)\ :\ \sigma\in\Pi_n(\pi)\}.
$$
Sagan~\cite{sag:pas} described the set partitions in $\Pi_n(\pi)$ for each $\pi\in\Pi_3$. Although it is not difficult to translate his work into the language of restricted growth functions, we include the proof of the following result for completeness and since it will be used many times subsequently. Define the {\em initial run} of an RGF $w$ to be the longest prefix of the form $12\dots m$. Also, we will use the notation $a^l$ to indicate a string of $l$ consecutive copies of the letter $a$ in a word.
Finally, say that $w$ is {\em layered} if $w=1^{n_1} 2^{n_2}\dots m^{n_m}$ for positive integers $n_1,n_2,\dots,n_m$.
\begin{thm}[\cite{sag:pas}]
\label{avoidance}
We have the following characterizations.
\begin{enumerate}
\item $ R_n(1/2/3) = \{w\in R_n\ :\ \text{$w$ consists of only $1$s and $2$s}\}$.
\item $ R_n(1/23)=\left\{w\in R_n\ :\ \text{$w$ is obtained by inserting a single $1$ into a word}\right.$
\hspace{160pt}$\left.\text{of the form $1^l 23\dots m$ for some $l\ge0$ and $m\ge1$}\right\}.$
\item $ R_n(13/2)=\{w\in R_n\ :\ \text{$w$ is layered}\}$.
\item $ R_n(12/3)=\{w\in R_n\ :\ \text{$w$ has initial run $1\dots m$ and $a_{m+1}=\dots=a_n\le m$}\}$.
\item $ R_n(123)=\{w\in R_n\ :\ \text{$w$ has no element repeated more than twice}\}$.
\end{enumerate}
\eth
\begin{proof} In all cases it is easy to see that $ R_n(\pi)$ is contained in the right-hand side. So we will concentrate on proving the other inclusions.
1. If $w(\sigma)$ can have only $1$s and $2$s then $\sigma$ has at most two blocks and so avoids $1/2/3$.
2. Suppose $\sigma\in\Pi_n(1/23)$ and let $m=\max\sigma$. We assume $m>1$ since otherwise $w$ is clearly of the desired form. Then no number $1<a\le m$ can be repeated since, if it were, the initial $1$ and two $a$'s in $w$ would correspond to a copy of $1/23$ in $\sigma$. Similarly, there can not be two $1$'s in $w$ appearing after the $2$. These two restrictions are equivalent to the description in the theorem.
3. It suffices to show that if $a_i=r$ then $a_{i+1}=r$ or $r+1$ whenever $\sigma$ avoids $13/2$. If not, then $a_{i+1}=s$ where $s<r$ or $s>r+1$. In the former case, $s$ must appear in $w$ in a position to the left of $a_i$ and the two copies of $s$ together with the copy of $r$ form a $13/2$ in $\sigma$. If $s>r+1$, then $a_i$ can not be the first copy of $r$ in $w$. Now these two copies of $r$ together with the $s$ give another contradiction.
4. Suppose $\sigma$ avoids $12/3$ and consider $a_{m+1}$. Condition~\ree{RGF} implies that $a_{m+1}=a_i\le m$ for some $i\le m$. So if $a_{m+2}\neq a_{m+1}$, then the subpartition $i,m+1/m+2$ would be a copy of $12/3$ in $\sigma$ which is a contradiction. Continuing in this way we obtain $a_{m+1}=\dots=a_n$.
5. The given condition on $w$ implies that the blocks of $\sigma$ all have one or two elements. Thus $\sigma$ avoids $123$.
\end{proof}
Using these characterizations, it is a simple matter to find the cardinalities of the avoidance classes.
\begin{cor}[\cite{sag:pas}]
\label{avoidancecor}
We have the following cardinalities.
$$
\begin{array}{l}
\#\Pi_n(1/2/3)=\#\Pi_n(13/2)=2^{n-1},\\[10pt]
\#\Pi_n(1/23)=\#\Pi_n(12/3)=\displaystyle 1+\binom{n}{2},\\[10pt]
\#\Pi_n(123)=\displaystyle\sum_{k\ge 0} {n\choose 2k} (2k)!!
\end{array}
$$
where $(2k)!!=(1)(3)(5)\dots(2k-1)$.\hfill \qed
\end{cor}
The four statistics of Wachs and White are denoted $\left\{$, $\ls$, $\right\}$, and $\rs$ where ``l" stands for ``left," ``r" stands for ``right," ``b" stands for ``bigger," and ``s" stands for ``smaller." We will describe the left-bigger statistic and the other three should become clear by analogy. Given a word $w=a_1\dots a_n$ we define
$$
\left\{(a_j)=\#\{a_i\ :\ \text{$i<j$ and $a_i>a_j$}\}.
$$
In words, we count the set of integers occuring before $a_j$ and bigger than $a_j$. It is important to note that we are taking the cardinality of a set, so if there are multiple copies of such an integer then it is only counted once. Also, clearly $\left\{(a_j)$ depends on the word containing $a_j$, not just $a_j$ itself even though, for simplicity, our notation does not reflect that. By way of example, if $w=1223142$ then
$\left\{(a_5)=2$ since there is a $2$ and a $3$ to the left of $a_5=1$. Finally, define
$$
\left\{(w)=\left\{(a_1)+\dots+\left\{(a_n).
$$
Continuing our example,
$$
\left\{(1223142)=0+0+0+0+2+0+2=4.
$$
To simplify notation, we will write $\left\{(\sigma)$ for the more cumbersome $\left\{(w(\sigma))$. Our main objects of study will be the generating functions
$$
\LB_n(\pi)=\LB_n(\pi;q)=\sum_{\sigma\in\Pi_n(\pi)} q^{\left\{(\sigma)}
$$
and the three analogous polynomials for the other statistics. Often, we will even be able to compute the multivariate generating function
$$
F_n(\pi)=F_n(\pi;q,r,s,t)=\sum_{\sigma\in\Pi_n(\pi)} q^{\left\{(\sigma)} r^{\ls(\sigma)} s^{\right\}(\sigma)} t^{\rs(\sigma)}.
$$
The rest of this article is structured as follows. In the next four sections we will compute $F_n(\pi)$ for $\pi=1/2/3,\ 1/23,\ 13/2$, and $12/3$, respectively. In Section~\ref{sec123} we study $123$ which is a more difficult pattern to analyze.
One can also consider avoiding more than one pattern at once and this is the goal of Section~\ref{secmpa}. Various results which did not fit earlier into the paper will be found in Section~\ref{secmr}. Finally we end with some open problems and areas for future research.
\section{The pattern $1/2/3$}
\label{sec1/2/3}
We first consider the set partition $1/2/3$. We begin by presenting the four-variable generating function from which we derive the generating functions associated with each individual statistic.
\begin{thm}
We have
$$F_n(1/2/3) = 1 + \sum \limits_{l=1}^{n-1}r^{n-l}s^l + \sum \limits_{l=2}^{n-1}\sum \limits_{k=0}^{n-l-1}\sum \limits_{i, j \geq 1} \binom{n-i-j-k-2}{l-i-j} q^{l-i}r^{n-l}s^{l-\delta_{k,0} j}t^{n-l-k}$$
where $\delta_{k, 0}$ is the Kronecker delta function.
\eth\label{F1/2/3}
\begin{proof}
By Theorem~\ref{avoidance}, any word $w \in R_n(1/2/3)$ is composed solely of ones and twos. Let $l$ denote the number of ones in $w$. If such a word is weakly increasing, it is easy to see that these words contribute
$$1 + \sum \limits_{l=1}^{n-1}r^{n-l}s^{l}$$
to the generating function.
Otherwise, let $w$ have at least one descent and $l$ ones. We can see that the word $w$ has the form $1^{i}w'1^{j}2^k$, where $i, j \geq 1$, the subword $w'$ begins and ends with a two, and $0\leq k \leq n-l-1$.
For such $w$ the lb statistic is given by the number of ones after the first two, that is, by the number of ones not in $1^i$. Thus, $\left\{(w) = l-i$. The ls statistic is given by the total number of twos in $w$, namely $n-l$. For the rb statistic, if $k$ is non-zero, then each one in $w$ contributes to the statistic. Otherwise, only the ones that are not in $1^j$ contribute. Combining the two cases gives $\right\}(w) = l-\delta_{k,0} j$. Finally, the rs statistic is given by the number of twos in $w'$, namely $n-l-k$. Putting all four statistics together produces
$$q^{\left\{(w)}r^{\ls(w)}s^{\right\}(w)}t^{\rs(w)} = q^{l-i}r^{n-l}s^{l-\delta_{k,0} j}t^{n-l-k}.$$
Choosing the number of ways of arranging the ones in $w'$ gives a coefficient of $$\binom{n-i-j-k-2}{l-i-j}.$$
Summing over $i, j, k, l$ and combining the cases gives our desired polynomial.
\end{proof}
The equations in the following corollary can be derived either by specialization of the four-variable generating function \ree{F1/2/3} and standard hypergeometric series techniques or by using the ideas in the proof of the previous result and ignoring the other three statistics.
\begin{cor}
\label{1/2/3cor}
We have
$$\LB_n(1/2/3) = \RS_n(1/2/3) = 1 + \sum \limits_{k = 0}^{n-2} \binom{n-1}{k+1}q^k,$$
and
\eqed{\LS_n(1/2/3) = \RB_n(1/2/3) = (r+1)^{n-1}.}
\end{cor}
In view of the preceeding corollary, it would be nice to find explicit bijections $\phi: R_n(1/2/3) \to R_n(1/2/3)$ and $\psi: R_n(1/2/3) \to R_n(1/2/3)$ such that $\phi$ takes $\left\{$ to $\rs$ and $\psi$ takes $\ls$ to $\right\}$. In the next two propositions, we present such bijections.
\begin{prop} \label{bijection1/2/3}
There exists an explicit bijection $\phi: R_n(1/2/3) \to R_n(1/2/3)$ such that for $v \in R_n(1/2/3)$,
$$\left\{(v) = \rs(\phi(v)).$$
\end{prop}
\begin{proof}
Let $v = a_1a_2\dots a_n \in R_n(1/2/3)$. Define
$$\phi(v) = a_1(3-a_n)(3-a_{n-1})\dots(3-a_3)(3-a_2).$$
Because $v \in R_n(1/2/3)$, by Theorem~\ref{avoidance}, it must be composed of only ones and twos and begin with a one. It is clear that $\phi(v)$ has the same form, so $\phi$ is well defined. Also, $\phi$ is its own inverse and is therefore a bijection.
If $\left\{(v) = k$, then $v$ must contain a subword $v' = 21^k$ and no subword of the form $21^l$, with $l>k$. In fact, this condition is clearly equivalent to $\left\{(v)=k$. It follows that $\phi(v') = 2^k1$ is a subword of $\phi(v)$ and $\phi(v)$ has no subword $2^l1$ with $l>k$. Therefore, $\rs(\phi(v)) = k = \left\{(v)$, as desired.
\end{proof}
\begin{prop}
There exists an explicit bijection $\psi: R_n(1/2/3) \to R_n(1/2/3)$ such that for $v \in R_n(1/2/3)$,
$$\ls(v) = \right\}(\psi(v)).$$
\end{prop}
\begin{proof}
Let $v \in R_n(1/2/3)$. If $v = 1^n$, then define $\psi(v) = v$. Clearly in this case $\ls(v) = 0 = \right\}(v)$.
Otherwise, let $v = a_1a_2 \dots a_{i-1}a_i 1^{n-i}$ where $a_i = 2$ and $n-i \geq 0$. Define
$$\psi(v) = (3 - a_i)(3-a_{i-1}) \dots (3-a_2)(3-a_1)1^{n-i}.$$
The proof is now similar to that of Proposition~\ref{bijection1/2/3}, using the fact that the $1^{n-i}$ at the end of $v$ contributes to neither $\ls$ or $\right\}$.
\end{proof}
\section{The pattern $1/23$}
\label{sec1/23}
In this section we will determine $F_n(1/23)$, and thus the generating functions for all four statistics. We will find that $\left\{$ and $\rs$ are equal for any $w\in R_n(1/23)$.
\begin{thm} \label{F1/23} We have \begin{equation}\label{F1/23eq} F_n(1/23)=(rs)^{\binom{n}{2}}+\sum_{m=1}^{n-1}\sum_{j=1}^{m}(qt)^{j-1}r^{\binom{m}{2}}s^{(n-m)(m-1)+m-j+\binom{m-1}{2}}.\end{equation}
\end{thm}
\begin{proof}
If $\sigma$ avoids $1/23$ we know from Theorem~\ref{avoidance} that the associated RGF is obtained by inserting a single $1$ into a word of the form $1^l23\dots m$ for some $l\geq 0$ and $m\geq 1$. If $l=0$ then the inserted $1$ must be at the beginning of the word in order for $w$ to be a RGF, so $w=12\dots n$. If $l>0$ then the inserted $1$ can be inserted after $j$ for any $1\leq j \leq m$, and the maximal letter $m$ satisfies $1\leq m \leq n-1$. If $w$ has maximal letter $m$ and we insert the $1$ after $j$ then $w$ is completely determined to be $1^{n-m}23\dots j1\dots m$.
In summary, either $w=12\dots n$ or $w$ is determined by the choice of $1\leq j \leq m$ and $1\leq m \leq n-1$. If $w=12\dots n$ then $\right\}(w)=\ls(w)=\binom{n}{2}$ and $\left\{(w)=\rs(w)=0$. For all other $w$
we have the following:
\begin{enumerate}
\item $\left\{(w)=j-1$,
\item $\ls(w)=\binom{m}{2}$,
\item $\right\}(w)=(n-m)(m-1)+m-j+\binom{m-1}{2}$
\item $\rs(w)=j-1$.
\end{enumerate}
1. Only the inserted $1$ has elements which are left and bigger which are the numbers $2$ through $j$. So $\left\{(w)=j-1$.
2. Since $w$ is an RGF every letter $i$ contributes $i-1$ to the $\ls$ giving a total of $\ls(w)=1+\dots +(m-1)=\binom{m}{2}$.
3. The first $n-m$ ones of $w$ each have $m-1$ elements which are right and bigger, so they contribute $(n-m)(m-1)$ to the $\right\}$. The inserted $1$ has $m-j$ letters which are right and bigger. Any element $i$ such that $2\leq i \leq m$ appears only once and contributes $m-i$ to the $\right\}$. This means we have an additional $(m-2)+\dots +0 =\binom{m-1}{2}$. Hence $\right\}(w)=(n-m)(m-1)+m-j+\binom{m-1}{2}$.
4. The only elements which have a number right and smaller are the elements $2$ through $j$, and the only number which is right and smaller of these elements is the inserted $1$. Hence $\rs(w)=j-1$.
Summing over all the valid values for $m$ and $j$ gives us our equality.
\end{proof}
The following result can be quickly seen by specializing Theorem~\ref{F1/23} or its demonstration,
so we have omitted the proofs.
\begin{cor}
\label{1/23cor}
We have $\left\{(w)=\rs(w)$ for all words $w\in R_n(1/23)$ and $$\LB_n(1/23)=\RS_n(1/23)=1+\sum_{j=1}^{n-1}(n-j)q^{j-1}.$$
Also
$$\LS_n(1/23)=r^{\binom{n}{2}}+\sum_{m=1}^{n-1}m r^{\binom{m}{2}},$$
and
\eqed{\RB_n(1/23)=s^{\binom{n}{2}}+\sum_{m=1}^{n-1}\sum_{j=1}^{m}s^{(n-m)(m-1)+m-j+\binom{m-1}{2}}.}
\end{cor}
\section{The pattern 13/2 }
\label{sec13/2}
In this section, we begin by evaluating the four-variable generating function $F_n(13/2)$. Goyt and Sagan~\cite{gs:sps} have previously proven a theorem regarding the single-variable generating functions for the $\ls$ and $\right\}$ statistics, and we will adapt their map and proof to obtain the multi-variate generating function for $13/2$. This generating function is closely related to integer partitions. A {\em partition $\lambda = (\lambda_1, \lambda_2, \dots, \lambda_k)$ of an integer $t$} is a weakly increasing sequence of positive integers such that $\sum_{i=1}^k \lambda_i = t$. The $\lambda_i$ are called \emph{parts}. Additionally, we will define an integer partition $n-\lambda = (n-\lambda_k,\dots n-\lambda_2, n-\lambda_1)$. Let $|\lambda| = \sum_{i=1}^k \lambda_i$. We will denote by $D_{n-1}$ the set of integer partitions with distinct parts of size at most $n-1$.
\begin{thm}
\label{13/2}
We have
$$F_n(13/2) = \prod_{i=1}^{n-1} (1+r^{n-i} s^i).$$
\eth
\begin{proof}
Suppose $w \in R_n(13/2)$. By Theorem~\ref{avoidance}, $w$ is layered and so $\left\{$ and $\rs$ are zero, resulting in no contribution to the generating function. For the other two statistics, since $w$ is layered it has the form
$w=1^{n_1}2^{n_2} \dots m^{n_m}$ where $m$ is the maximum element of $w$. Define $\phi: R_n(13/2) \to D_{n-1}$ by
$$ \phi(w) = (\lambda_1, \lambda_2, \dots, \lambda_{m-1})$$
where $\lambda_j = \sum_{i=1}^{j}n_i$ for $1 \leq j \leq m-1$. Note that since the $n_j$ are positive, the $\lambda_j$ are distinct, increasing, and less than $n$ since the sum never includes $n_m$. Thus the map is well defined.
We now show that $\phi$ is a bijection by constructing its inverse. Given $\lambda = (\lambda_1, \lambda_2, \dots, \lambda_{m-1})$, consider for $1 \leq j \leq m$, the differences $n_j = \lambda_j - \lambda_{j-1}$, where we define $\lambda_0 = 0$ and $\lambda_m = n$. It is easy to see that sending $\lambda$ to $w=1^{n_1}2^{n_2} \dots m^{n_m}$ is a well-defined inverse for $\phi$.
We next claim that if $\phi(w)=\lambda$ then $\right\}(w)=|\lambda|$. Indeed, from the form of $w$ and $\lambda$ we see that
$$
\right\}(w)=\sum_{i=1}^{m-1} n_i(m-i)=\sum_{j=1}^{m-1} \sum_{i=1}^{j}n_i = |\lambda|.
$$
Similarly we obtain $\ls(w)=|n-\lambda|$. It follows that
$$
F_n(13/2)=\sum_{\lambda \in D_{n-1}}r^{|n-\lambda|}s^{|\lambda|}=\prod_{i=1}^{n-1} (1+r^{n-i} s^i)
$$
as desired.
\end{proof}
The generating function of each individual statistic is easy to obtain by specialization of Theorem~\ref{13/2} so we have omitted the proofs.
\begin{cor}[\cite{gs:sps}]
\label{13/2cor}
We have
$$\LB_n(13/2) = 2^{n-1} = \RS_n(13/2)$$
and
\eqed{\LS_n(13/2) = \prod_{i=1}^{n-1} (1 + q^i) = \RB_n(13/2).}
\end{cor}
\section{The pattern $12/3$}
\label{sec12/3}
In this section, we determine $F_n(12/3)$. The other polynomials associated with $12/3$ are obtained as corollaries. We find this avoidance class interesting because it leads to a connection with number theory.
\begin{thm}\label{F12/3} We have
\begin{equation}\label{F12/3eq}
F_n(12/3) = r^{\binom{n}{2}}s^{\binom{n}{2}}+\sum_{m=1}^{n-1}\sum_{i=1}^{m}q^{(n-m)(m-i)}r^{\binom{m}{2}+(n-m)(i-1)}s^{\binom{m}{2}}t^{m-i}.
\end{equation}
\eth
\begin{proof} By Theorem~\ref{avoidance}, the elements of $ R_n(12/3)$ are the words of the form
$$
w = 123\dots m i^{n-m}
$$
where $i\le m$. If $w=123\dots n$ then $\ls(w)=\right\}(w)=\binom{n}{2}$ and $\left\{(w)=\rs(w)=0$. Otherwise $m<n$. In this case, we will show the following:
\begin{enumerate}
\item $\left\{(w) = (n-m)(m-i)$,
\item $\ls(w) = \binom{m}{2} + (n-m)(i-1)$,
\item $\right\}(w) = \binom{m}{2}$,
\item $\rs(w) = m-i$.
\end{enumerate}
1. There are $n-m$ copies of $i$ in $w$ and these are the only elements contributing to $\left\{$. Each $i$ has the elements $(i+1)(i+2)\dots m$ to its left that are bigger than it. So $\left\{(i) = m-i$ for all $i$ and $\left\{(w) = (n-m)(m-i)$.
2. Each element $w_j$ of $w$ has $\ls(w_j)=w_j-1$ by condition \ree{RGF}. Using this and the form of $w$ easily yields the desired equality.
3. This is similar to the previous case, noting that only the initial run of $w$ contributes to $\right\}$.
4. We can see that the only elements $w_j$ with $\rs(w_j)>0$ will be those in the initial run such that $w_j>i$. These are precisely the elements $(i+1)(i+2)\dots m$ and each element has exactly one element to its right that is smaller than it. So $\rs(w) = m-i$.
Summing over the valid values of $m$ and $i$, we have \ree{F12/3eq}.
\end{proof}
The next corollary follows easily by specialization of \ree{F12/3}.
\begin{cor}\label{ls12/3} We have
$$
\LS_n(12/3) = r^{\binom{n}{2}} + \sum_{m=1}^{n-1}\sum_{i=1}^{m}r^{\binom{m}{2}+ (n-m)(i-1)},
$$
and
$$
\RB_n(12/3)=s^{\binom{n}{2}}+ \sum_{m=1}^{n-1} ms^{\binom{m}{2}},
$$
as well as
\eqed{\RS_n(12/3)=1 + \sum_{k=0}^{n-2}(n-k-1)t^k.}
\end{cor}
The coefficients of $\LB_n(12/3)$ have an interesting interpretation.
\begin{prop} We have
\begin{equation}\label{lb12/3}
\LB_n(12/3)=\sum_{k=0}^{\left\lfloor (n-1)^2/4 \right\rfloor} D_kq^k,
\end{equation}
where $D_k = \#\{d\ge1 : d\mid k \text{ and } d+ \frac{k}{d} +1 \leq n\}$.
\end{prop}
\begin{proof} Set $r=s=t=1$ in~\ree{F12/3eq}. We begin by showing the degree of $\LB_n(12/3)$ is $\left\lfloor (n-1)^2/4\right\rfloor$.
By \ree{avoidance} we can let $w= 123\dots m i^{n-m}$ be a word in $ R_n(12/3)$.
In order to maximize the $\left\{(w)$, we can assume $i=1$.
So, using the formula for $\left\{(w)$ derived in the proof of Theorem~\ref{F12/3}, we must maximize $(n-m)(m-1)$. We take the derivative with respect to $m$ and set the equation equal to zero to obtain $n-2m+1=0$
and $m=\frac{n+1}{2}$. To get integer values of $m$, we obtain
\begin{equation}
\begin{cases}
m = \frac{n+1}{2} \mbox{ if } n \text{ is odd}, \\
m = \left\lceil \frac{n+1}{2}\right\rceil \mbox{ or } \left\lfloor\frac{n+1}{2}\right\rfloor \mbox{ if } n \text{ is even.}
\end{cases}
\end{equation}
In either case, the maximum value of $\left\{$ is $\left\lfloor (n-1)^2/4\right\rfloor$.
We now show the coefficient of $q^k$ is $D_k$. As before, let $w= 123\dots m i^{n-m}$ be a word associated with a set partition that avoids $12/3$ and let $\left\{(w)=k$. If we let $d=n-m$ be the number of $i$'s, it is clear that $\left\{(w) = d(m-i) = k$ and therefore, $m-i = \frac{k}{d}$. Because $w$ must be of length $n$, we now must determine which divisors $d$ of $k$ are valid. Each of the $d$ trailing $i$'s has $\frac{k}{d}$ elements to its left and bigger. Because $i\geq 1$, the leading one cannot be such an element. Thus in order for $w$ to be of length $n$ we must have $d+\frac{k}{d}+1 \leq n$.
\end{proof}
The above formulation of $\LB_n(12/3)$ leads to the following corollary, showing a connection to number theory.
\begin{cor} When $k\leq n-2$, we have $D_k=\tau(k)$, the number-theoretic function which counts the divisors of $k$.
\end{cor}
\begin{proof} We show that if $k\leq n-2$ then all positive divisors $d$ of $k$ are valid. We know that $d+\frac{k}{d}\leq k+1$ because $d=1$ and $d=k$ are the divisors of $k$ which maximize $d+\frac{k}{d}$. Thus, we have $d+\frac{k}{d}+1 \leq k+2 \leq n$. Therefore every positive divisor of $k$ satisfies the inequality in the definition of $D_k$, and this implies $D_k=\tau(k)$.
\end{proof}
\section{The pattern $123$}
\label{sec123}
The reader will have noticed that for the other four set partitions of [3], we provided a $4$-variable generating function describing all four statistics on the avoidance class of those partitions.
The pattern $123$, however, is much more difficult to deal with and so we will content ourselves with results about the individual statistics. We will start with the left-smaller statistic.
\begin{thm}\label{thm:ls123} We have
\begin{equation}\label{eq:ls123}
\LS_n(123) = \sum_{m = \ce{n/2}}^n \left [ \sum_{L} \left ( \prod_{g=1}^{n-m} (m-\ell_g + g) \right )q^{\binom{m}{2} + \sum\limits_{\ell \in L} (\ell - 1)} \right ]
\end{equation}
where the inner sum is over all subsets $L = \{\ell_1, \ell_2, \dots, \ell_{n-m} \}$ of $[m]$
with $\ell_1 > \dots > \ell_{n-m}$.
\eth
\begin{proof}
We start by noting that if a word has a maximum element $m$, then there must be $n-m$ repeated elements in the word, i.e., elements $i$ that appear after the initial occurrence of $i$. The bounds on our outer sum are given by the largest possible value of $m$ being $n$, and the smallest possible value of $m$ being $\ce{n/2}$, since we can repeat each element a maximum of two times.
We will now build our word $w$ by starting with a base sequence $12\dots m$ and adding in repeated elements. The base sequence will contribute $1 + 2 + \dots + (m-1) = \binom{m}{2}$ to $\ls(w)$. Let $L$ be the set of repeated elements we want to add to $w$. Then $L$ must contain $n-m$ elements from $[m]$, and since $w$ can have no element appear more than twice, $L$ can have no element appear more than once.
For each element $\ell \in L$ that we add to our base sequence, we will increase $\ls(w)$ by $\ell - 1$.
So for any word $w$ with maximum $m$ formed in this way, we have $\ls(w) = \binom{m}{2} + \sum_{\ell \in L} (\ell - 1)$.
To find how many possible words can be so created, we start with our base sequence $12\dots m$, and build up our word by placing in the repeated elements from $L$ one at a time. There are $m - (\ell_1 - 1)$ spots where we can place the largest repeated element, $\ell_1$: anywhere after the original occurrence of $\ell_1$. Then when we place our second repeated element, $\ell_2$, we will have $m - (\ell_2 - 1) + 1$ spots, where the plus one comes from the extra space the first repeated element added in front of $\ell_2$. In general, when we place $\ell_g$ we will have $m - (\ell_g - 1) + (g-1) = m - \ell_g + g$ places to put it. The condition $\ell_1 > \dots > \ell_{n-m}$ is used since it implies that regardless of where $\ell_i$ is placed, one will have the same number of choices for the placement of $\ell_{i+1}$.
Multiplying all these terms together and then summing over all possible subsets $L$ of $[m]$ gives us the coefficient of $q$. Finally, summing over all possible maximums of the words in the avoidance class gives us equation~\eqref{eq:ls123}.
\end{proof}\medskip
We were only able to find explicit expressions for certain coefficients of the polynomials generated from other statistics. We will now look at the left-bigger statistic.
\begin{thm}\label{lb123degree} We have the following.
\begin{enumerate}
\item The degree of $\LB_n(123)$ is $$\left\fl{\frac{n(n-1)}{6}\right}.$$
\item The leading coefficient of $\LB_n(123)$ is
$$\begin{cases}
k! & \text{if } n = 3k \text{ or } 3k+1,\\
(k+2)k! & \text{if } n = 3k+2,\\
\end{cases}$$
for some nonnegative integer $k$.
\end{enumerate}
\eth
\begin{proof}
We will show that a word of the form $w = 12\dots iw_{i+1}\dots w_n$ with $w_{i+1},\dots, w_n$ being a permutation of the interval $[1,n-i]$ will provide a maximum $\left\{$ which is $\fl{(n(n-1))/6}$.
First we will prove that the elements after the initial run $12\dots i$ must be less than or equal to $i$. Note that, by definition of the initial run, $w_{i+1} \le i$. Now suppose, towards a contradiction, that for some $j \in [i+2,n]$, there was some element $w_j > i$. Then, since $w$ is an RGF, we must have $w_k = i+1$ for some $k \in [i+2,j]$. But by switching $w_k$ and $w_{i+1}$, we would increase $\left\{$ by at least one since $w_{i+1} \le i$. So if any element after the initial run is greater than $i$, $\left\{$ is not maximum.
Next we will show that the elements after the initial run have to be exactly those in the interval $[1,n-i]$, up to reordering. Suppose towards contradiction there was some element $t \in [1,n-i]$ that did not appear in the sequence after the initial run, and instead there appeared some element $s \in [n-i+1,i]$. Then $\left\{(s) = i-s$. But $\left\{(t) = i-t$, and since $s > t$, it follows that $\left\{(t) > \left\{(s)$. Therefore, if we want to maximize $\left\{$, we must have the sequence after the initial run being exactly the interval $[1,n-i]$, up to reordering.
Now that we've established that our word is of the form $w = 12\dots iw_{i+1}\dots w_n$ with $w_{i+1},\dots ,w_n$ being exactly those elements in the interval $[1,n-i]$, we simply need to maximize $\left\{$ using some elementary calculus.
\begin{align}
\left\{(w) &= (i - w_{i+1}) + (i - w_{i+2}) + \dots + (i - w_n) \nonumber\\
&= (i - 1) + (i - 2) + \dots + (2i- n)\nonumber \\
&= \frac{(4n+1)i - 3i^2 - n^2 - n}{2}\label{lbeqn}
\end{align}
Considering $i$ as a real variable and differentiating gives us a maximum value of $\left\{(w)$ when $i = (4n+1)/6$. We must modify this slightly since we want $i$ to be integral. Rounding $i$ to the closet integer gives
$$i = \begin{cases}
\left \lfloor \frac{4n+1}{6}\right\rfloor & \text{if } n = 3k,\\
\left \lceil \frac{4n+1}{6}\right\rceil & \text{if } n = 3k+1,\\
\left \lfloor \frac{4n+1}{6}\right\rfloor \text{or }\left \lceil \frac{4n+1}{6}\right\rceil & \text{if } n = 3k+2,\\
\end{cases}$$
for some nonnegative integer $k$.
Plugging each value of $n$ and $i$ back into equation~\eqref{lbeqn} gives us an $\left\{$ of $\fl{(n(n-1))/6}$ in all cases. As we've mentioned before, the elements $w_{i+1}, \dots, w_n$ must be exactly those in the interval $[1,n-i]$, but the ordering doesn't matter. This means the leading coefficient of $\LB_n(123)$ will be precisely the number of ways to permute the $n-i$ elements after the initial run. This gives us our second result.
\end{proof}\medskip
Our next theorem will involve the Fibonacci numbers. Recall that the $n$th Fibonacci number $F_n$ is defined recursively as
\begin{equation}\label{fib}
F_n = F_{n-1} + F_{n-2}
\end{equation}
with initial conditions $F_0 = 1$ and $F_1 = 1$.
\begin{thm}\label{lbcoeffs} We have the following coefficients.
\begin{enumerate}
\item The constant term of $\LB_n(123)$ is $F_n$.
\item The coefficient of $q$ in $\LB_n(123)$ is $(n-2)F_{n-2}$.
\end{enumerate}
\eth
\begin{proof}
If $\left\{(\sigma) = 0$, then $w = w(\sigma)$ must be layered. Let $L(n)$ be the set of layered words $w(\sigma)$ with $\sigma \in \Pi_n(123)$. It follows that the constant term of $\LB_n(123)$ is $\#L(n)$. Define $L_i(n) = \{w \in L(n)\ |\ w \text{ starts with } i \text{ ones}\}$.
Then $\#L(n) = \#L_1(n) + \#L_2(n)$. But $\#L_i(n) = \#L(n-i)$ for $i = 1,2$, since if $w$ begins with $i$ ones then the rest of the word is essentially a layered word with $n-i$ elements. Therefore, $\#L(n) = \#L(n-1) + \#L(n-2)$. Since $\#L(0) = 1$ and $\#L(1) = 1$, we have $\#L(n) = F_{n}$.
To prove the second claim, let $w \in R_n(123)$ with $\left\{(w) = 1$. Then there must be exactly one descent in $w$ and it must be of the form $w_{j+1}=w_j-1$ for some $2\le j\le n- 1$. Removing $w_j$ and $w_{j+1}$ from $w$ and then subtracting one from all $w_k$ with $k>j+1$ gives an element $w'\in R_{n-2}$ which is layered. So, from the previous paragraph, there are $F_{n-2}$ choices for $w'$. Further, there were $n-2$ choices for $j$ and so the total number of $w$ is $(n-2)F_{n-2}$.
\end{proof}\medskip
We will now look at the right-smaller statistic.
\begin{thm}\label{RSdegree/con} We have the following.
\begin{enumerate}
\item The degree of $\RS_n(123)$ is $$\left\fl{\frac{(n-1)^2}{4}\right}.$$
\item The leading coefficient of $\RS_n(123)$ is $1$ when $n$ is odd, and $2$ when $n$ is even.
\item The constant term of $\RS_n(123)$ is $F_n$.
\end{enumerate}
\eth
\begin{proof}
The proof of the first result is very similar to the proof of the degree of $\LB_n(123)$. When looking at the right-smaller statistic, the word that maximizes $\rs$ is of the form $w = 12\dots i (n-i) \dots 21$, where $12\dots i$ is the initial run. Calculating $\rs(w)$ gives
\begin{equation}\label{maxrs}
\rs(w) = (n-i)(i-1),
\end{equation}
and differentiating with respect to the real variable $i$ and maximizing gives $i = (n+1)/2$. Since we want $i$ to be integral, we have
$$i = \begin{cases}
\frac{n+1}{2} & \text{if } n \text{ is odd,}\\
\left \lfloor \frac{n+1}{2}\right\rfloor \text{or }\left \lceil \frac{n+1}{2}\right\rceil & \text{if } n \text{ is even.}\\
\end{cases}$$
Plugging each value of $i$ and $n$ into \eqref{maxrs} gives $\fl{(n-1)^2/4}$ in both cases. Also, the number of choices for $i$ gives the leading coefficient of $\RS_n(123)$.
The proof for the constant term of $\RS_n(123)$ is the same as for $\LB_n(123)$ since for any $w$ we have $\rs(w) = 0$ if and only if $\left\{(w) = 0$.
\end{proof}\medskip
Our final result of this section gives the degree of $\RB_n(123)$. It follows immediately from the easily proved fact that the word which maximizes $\right\}$ is $w = 12\dots n$.
\begin{thm}\label{rbdegree}
$\RB_n(123)$ is monic and has degree $\binom{n}{2}.$\hfill \qed
\eth
\section{Multiple pattern avoidance}
\label{secmpa}
\begin{table}
\begin{center}
\scalebox{1.1}{
\begin{tabular}{ |c|c| }
\hline
\multicolumn{1}{|c|}{Avoidance Class} & \multicolumn{1}{|c|}{Associated RGFs} \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 1/23)$ & $1^n,\ 1^{n-1}2,\ 1^{n-2}21$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 13/2)$ & $1^m2^{n-m}$ for all $1 \leq m \leq n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 12/3)$ & $1^n,\ 12^{n-1},\ 121^{n-2}$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 13/2)$ & $1^{n-m+1}23 \dots m$ for all $1 \leq m \leq n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 12/3)$ & $1^n,\ 12 \dots (n-1)1,\ 12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 123)$ & $12 \dots n,\ 12 \dots (n-1)$ with an additional $1$ inserted \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(13/2, 12/3)$ & $12 \dots m^{n-m+1}$ for all $1 \leq m \leq n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(13/2, 123)$ & layered RGFs with at most two elements in each layer \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(12/3, 123)$ & $12 \dots (n-1)m$ for all $1 \leq m \leq n$ \rule{0pt}{20pt}\\ [5pt]
\hline
\end{tabular}}\\
\caption{Avoidance classes avoiding two partitions of $[3]$ and associated RGFs}
\label{table:avoidance elements}
\end{center}
\end{table}
Rather than avoiding a single pattern, one can avoid multiple patterns. Define, for any set $P$ of set partitions
$$
\Pi_n(P)=\{\sigma\in\Pi_n\ :\ \text{$\sigma$ avoids every $\pi\in P$}\}.
$$
Similarly adapt the other notations we have been using.
Goyt~\cite{goy:apt} characterized that cardinalities of $\Pi_n(P)$ for any $P\subseteq{\mathfrak S}_3$.
Our goal in this section is to do the same for $F_n(P)$. We will not include those $P$ containing both $1/2/3$ and $123$ since it is easy to see from Theorem~\ref{avoidance} that there are no such partitions for $n\ge5$.
Table~\ref{table:avoidance elements} shows the avoidance classes and the resulting restricted growth functions that arise from avoiding two patterns of length $3$. These as well as the entries in Table~\ref{table:threefour patterns} also appear in Goyt's work, but we include them here for completeness.
For ease of references, we give a total order to $\Pi_3$ as follows
\begin{equation}
\label{order}
1/2/3,\ 1/23,\ 13/2,\ 12/3,\ 123
\end{equation}
and list the elements of any set $P$ in lexicographic order with respect to~\ree{order}. Finally, for any $P\subseteq \Pi_3$ we have $\Pi_n(P)=\Pi_n$ for $n<3$. So we assume for the rest of this section that $n\ge3$.
The next result translates this table into generating functions. This is routine and only uses techniques we have seen in earlier sections so the proof is omitted.
\begin{thm} \label{F(1/2/3,13/2)}
For $n\ge3$ we have
\begin{align*}
F_n(1/2/3, 1/23) &= 1 + rs^{n-1} + qrs^{n-2}t, \\
F_n(1/2/3, 13/2) &= 1+ \sum_{i=1}^{n-1}r^is^{n-i}, \\
F_n(1/2/3, 12/3) &= 1+rs^{n-1} + q^{n-2}rst, \\
F_n(1/23, 13/2) &= 1 + \sum_{i=1}^{n-1} r^{\binom{n-i+1}{2}}s^{\binom{n}{2}-\binom{i}{2}}, \\
F_n(1/23, 12/3) &= 1 + (qt)^{n-2}(rs)^{\binom{n-1}{2}} + (rs)^{\binom{n}{2}}, \\
F_n(1/23, 123) &= (rs)^{\binom{n}{2}} + r^{\binom{n-1}{2}} \sum_{i=0}^{n-2}(qt)^is^{\binom{n}{2}-i-1}, \\
F_n(13/2, 12/3) &= 1 + \sum_{i=1}^{n-1}r^{\binom{n}{2}-\binom{i}{2}}s^{\binom{n-i+1}{2}}, \\
F_n(13/2, 123) &= \sum_{m = \ce{n/2}}^n \left [ \sum_{L} r^{\binom{m}{2} + \sum\limits_{\ell \in L} (\ell - 1)} s^{\binom{m}{2} + \sum\limits_{\ell \in L} (m - \ell)} \right ], \\
F_n(12/3, 123) &= ( rs)^{\binom{n}{2}} + s^{\binom{n-1}{2}} \sum_{i=0}^{n-2} (qt)^ir^{\binom{n}{2}-i-1},
\end{align*}
where $L$ and $m$ in $F_n(13/2, 123)$ are defined as in Theorem~\ref{thm:ls123}.\hfill \qed
\eth
Note that from this theorem we immediately get the following nice equidistribution results.
\begin{cor}
\label{multicor}
Consider the generating function $F_n(P)$ where $P\subseteq\Pi_3$.
\begin{enumerate}
\item We have $F_n(P)$ invariant under switching $q$ and $t$ if $13/2\in P$ or $P$ is one of
$$
\{1/2/3, 1/23\}; \ \{1/23, 12/3\}; \ \{1/23, 123\}; \ \{12/3, 123\}.
$$
\item We have $F_n(P)$ invariant under switching $r$ and $s$ if $P$ is one of
$$
\{1/2/3, 13/2\}; \ \{1/23, 12/3\}.
$$
\item We have the following equalities between generating functions for different $P$:
$$
F_n(1/23, 13/2;q,r,s,t)=F_n(13/2, 12/3;q,s,r,t)
$$
and
\eqed{F_n(1/23, 123;q,r,s,t)=F_n(12/3, 123;q,s,r,t).}
\end{enumerate}
\end{cor}
Next, we will examine the outcome of avoiding three and four partitions of $[3]$. We can see the avoidance classes and the resulting restricted growth functions in Table~\ref{table:threefour patterns}.
The entries in this table can easily be turned into a polynomial by the reader if desired. Avoiding all five partitions of $[3]$ is not included because it would contain both $1/2/3$ and $123$.
\begin{table}[h]
\begin{center}
\scalebox{1.1}{
\begin{tabular}{ |c|c| }
\hline
\multicolumn{1}{|c|}{Avoidance Class} & \multicolumn{1}{|c|}{Associated RGFs} \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 1/23, 13/2)$ & $1^n$, \ $1^{n-1}2$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 1/23, 12/3)$ & $1^n$, \ $121$ when $n=3$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/2/3, 13/2, 12/3)$ & $1^n$, \ $12^{n-1}$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 13/2, 12/3)$ & $1^n, \ 12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 13/2, 123)$ & $1^22 \dots (n-1), \ 12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 12/3, 123)$ & $12 \dots (n-1)1, \ 12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(13/2, 12/3, 123)$ & $12 \dots (n-2)(n-1)^2$, \ $12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline\hline
$\Pi_n(1/2/3, 1/23, 13/2, 12/3)$ & $1^n$ \rule{0pt}{20pt}\\ [5pt]
\hline
$\Pi_n(1/23, 13/2, 12/3, 123)$ & $12 \dots n$ \rule{0pt}{20pt}\\ [5pt]
\hline
\end{tabular}}\\
\caption{Avoidance classes and associated RGFs avoiding three and four partitions of $[3]$ }
\label{table:threefour patterns}
\end{center}
\end{table}
\section{Miscellaneous Results}
\label{secmr}
In this section we present other interesting results we have found. These include theorems regarding longer patterns and several bijections.
We will start with a sequence of results concerning the pattern $14/2/3$.
Our first theorem concerns applying the $\left\{$ statistic to the avoidance class of $14/2/3$, from which a connection arises between $14/2/3$ avoiding set partitions and integer compositions. First, we characterize $ R_n(14/2/3)$. We define the index $i$ to be a \emph{dale of height $a$} in $w$ if $a_i=a$ and
$$
a_i=\max\{a_1,\dots,a_{i-1}\}-1.
$$
\begin{lem}
\label{14/2/3_avoid}
For an RGF $w$, $w$ is contained in $ R_n(14/2/3)$ if and only if $w$ meets the following restrictions:
\begin{itemize}
\item for $i\geq 2$ we have $a_i\geq \max\{a_1,\dots,a_{i-1}\}-1$, and
\item if $w$ has a dale of height $a$, then $w$ does not have a dale of height $a+1$.
\end{itemize}
\end{lem}
\begin{proof}
Let $\sigma$ avoid $14/2/3$. Assume, towards contradiction, that there existed an $a_i$ in $w=w(\sigma)$ with $a_i<\max\{a_1,\dots,a_{i-1}\}-1$ and let $a=a_i$. By the structure of restricted growth functions, this implies that $a(a+1)(a+2)a$ exists as a subword in $w$. But then these four elements give rise to an occurence of $14/2/3$ in $\sigma$, which is a contradiction. This shows the first inequality.
Now assume that there existed dales of height $a$ and height $a+1$ in $w$. This would require $w$ to contain $(a+1)a(a+2)(a+1)$ as a subword, which again implies an occurance of $14/2/3$ in $\sigma$. This shows the height requirement for dales.
Now assume that $\sigma$ is a partition with $w=w(\sigma)$ meeting the listed requirements. If $\sigma$ contained $14/2/3$ as a pattern, then $abca$ must occur as a subword in $w$, with $a\neq b\neq c$. If $a$ was the minimum value in this subword, then either $a<b-1$ or $a<c-1$, which contradicts the first restriction put on $w$ in view of the second $a$ in the subword. Further, if $a$ was the maximum value in this subword, then either $b<a-1$ or $c<a-1$, raising the same contradiction in view of the second $a$. Similarly, we can rule out $c<a<b$. Thus the only remaining possibility is that $b<a<c$. By the first condition in the lemma, it then must be that the subword is exactly $a(a-1)(a+1)a$, which contradicts the restriction on dales. Thus $\sigma$ avoids $14/2/3$, showing the reverse implication.
\end{proof}\medskip
Note that a dale in a word $w$ contributes exactly one to $\left\{(w)$.
And by the previous lemma, dales are the only source of $\left\{$ for words in $ R_n(14/2/3)$. For the proof of our theorem about $\LB(14/2/3)$ we will also need the following notion: call $i$ a \emph{left-right maximum of value $a$} in $w$ if $a_i=a$ and
$$
a_i>\max\{a_1,\dots,a_{i-1}\}.
$$
Being an RGF is equivalent to having left-right maxima of values $1,2,\dots,m$ for some $m$.
\begin{thm}
\label{LB(14/2/3)}
For $n\geq1$, we have
$$
\LB_n(14/2/3)=2^{n-1}+\sum_{k=1}^{n-2}\left[\sum_{m\geq2}\binom{n-1}{k+m-1}\sum_{j\geq1}\binom{k-1}{j-1}\binom{m-j}{j}\right]q^k.
$$
\eth
\begin{proof}
It is easy to see that the constant term in this polynomial comes from the layered partitions of $[n]$, all of which avoid $14/2/3$. Now consider the coefficient of $q^k$ for $k\geq 1$. From the discussion before the statement of the theorem, for a word in $ R_n(14/2/3)$ to have an $\left\{$ of $k$, it must have $k$ dales. Further, we know that $i=1$ is always a left-right maximum of value $1$ in any RGF, and that $i=1$ is never a dale. It follows by Lemma~\ref{14/2/3_avoid} that, to completely characterize an RGF of $\left\{$ equal to $k$ and maximum value $m$ in $ R_n(14/2/3)$, it suffices to specify the remaining $m-1$ left-right maxima and the $k$ dale indices. As such, there are $\binom{n-1}{m+k-1}$ ways to choose a set $I$ which is the union of these two index sets.
Let $I=\{i_1<i_2<\dots<i_{m+k-1}\}$ be such a set. We will indicate indices chosen for dales by coloring them blue, and left-right maxima by coloring them red. We define a \emph{run} to be a maximal sequence of indices $i_c,i_{c+1},\dots,i_d$ which is monochromatic. Let $j$ be the number of blue runs, and let $b_s$ be the number of indices in the $s$th blue run, for $1\leq s\leq j$. As these numbers count the dales in $w$, we must have
$$
b_1+b_2+\dots+b_j=k,
$$
or equivalently that $b_1,\dots,b_j$ form an integer composition of $k$. Thus there are $\binom{k-1}{j-1}$ ways of choosing $j$ blue runs.
Now note that $I$ must start with a red run, and can end with either a red or blue run. Thus there are $j$ or $j+1$ red runs. Let $r_t$ be the length of the $t$th red run, for $1\leq t\leq j+1$, where we set $r_{j+1}=0$ if there are $j$ red runs. Furthermore, by the dale height restriction in Lemma~\ref{14/2/3_avoid}, we have $r_t \geq 2$ for $2\leq t\leq j$. Now as before, we have
$$
r_1+r_2+\dots+r_{j+1}=m-1,
$$
subject to $r_1\geq 1$, $r_2,\dots,r_j\geq 2$, and $r_{j+1}\geq 0$. Using a standard composition manipulation, we can put this sum in correspondence with a composition of $m-j+1$ into $j+1$ parts, which gives $\binom{m-j}{j}$ ways to choose the red runs. Putting everything together and summing over the possible values of $m$ and $j$ gives the coefficient of $q^k$ as
$$
\sum_{m\geq2}\binom{n-1}{k+m-1}\sum_{j\geq1}\binom{k-1}{j-1}\binom{m-j}{j}.
$$
All that is left is to give appropriate bounds for $k$. It follows by Lemma~\ref{14/2/3_avoid} that $w=121^{n-2}$ is in $ R_n(14/2/3)$ and that $w$ gives a maximizing $\left\{$ of $n-2$. This gives $1\leq k\leq n-2$, and provides the correct parameters for the polynomial.
\end{proof}\medskip
From the previous theorem, and from the characterization of $ R_n(14/2/3)$, several corollaries follow.
\begin{cor}
\label{LB(14/2/3)=RS(14/2/3)}
We have
$$
\LB_n(14/2/3)=\RS_n(14/2/3).
$$
\end{cor}
\begin{proof}
We proceed by finding a bijection $\phi$ that takes $ R_n(14/2/3)$ to itself, and that takes the $\left\{$ statistic to the $\rs$ statistic. Let $w$ be a member of $ R_n(14/2/3)$. From Lemma~\ref{14/2/3_avoid}, we can partition $w$ into sections based on the dales of $w$. Specifically, let $a_i$ be a letter in $w$, and let $a=a_i$. If there is no dale of height $a$ or $a-1$ in $w$, then it follows that every copy of $a$ is adjacent in $w$. That is to say, we can break $w$ into
$$
w=w_1 a^l w_2,
$$
with $a_j<a$ for all $a_j$ in $w_1$, and $a_k>a$ for all $a_k$ in $w_2$. Call such a string a {\em plateau} of $w$. It follows that plateaus in $w$ contribute nothing to $\left\{(w)$ or $\rs(w)$. We will let $\phi$ act trivially on the plateaus of $w$.
If this is not the case, then there is a dale of height $a$ or $a-1$ in $w$.
By Lemma~\ref{14/2/3_avoid} again, both $a$ and $a-1$ can not be dale heights. So suppose $a-1$ is a dale height.
It follows that the occurances of $a$ and $a-1$ in $w$ are adjacent and we have
$$
w=w_1 (a-1)^{l_0}a^{j_1}(a-1)^{l_1}\dots a^{j_t}(a-1)^{l_t} w_2,
$$
with $l_0,\dots,l_{t-1}>0$, $l_t\geq 0$, and $j_1,\dots,j_t>0$. Further, we have $a_j<a-1$ for all $a_j$ in $w_1$, and $a_k>a$ for all $a_k$ in $w_2$. Such a string will be called a {\em dale section} of $w$. Breaking up $w$ in this manner shows that such a dale section contributes $l_1+\dots+l_t$ to $\left\{(w)$, and either $j_1+\dots+j_{t-1}$ or $j_1+\dots+j_{t}$ to $\rs(w)$, depending on whether or not $l_t=0$. As such, if
$$
d=(a-1)^{l_0}a^{j_1}(a-1)^{l_1}\dots a^{j_t}(a-1)^{l_t}
$$
is a dale section in $w$, we let
$$
\phi(d)=
\begin{cases}
\begin{array}{ll}
(a-1)^{l_0}a^{l_1}(a-1)^{j_1}\dots a^{l_t}(a-1)^{j_t} &\text{ if }l_t>0, \\
(a-1)^{l_0}a^{l_1}(a-1)^{j_1}\dots a^{l_{t-1}}(a-1)^{j_{t-1}}a^{j_t} &\text{ if }l_t=0.
\end{array}
\end{cases}
$$
It follows that $\phi$ exchanges $\left\{$ and $\rs$ for a dale section.
Now by the nature of $ R_n(14/2/3)$, we know that $w$ is merely a concatenation of plateaus and dale sections. Having defined $\phi$ on these parts of $w$, we define $\phi(w)$ by applying $\phi$ to the plateaus and dale sections of $w$ in a piecewise manner. It follows that $\phi$ is a bijection, since it is an involution. Finally, since $\left\{(w)$ and $\rs(\phi(w))$ are sums over the dale sections of $w$ and $\phi(w)$, and since $\phi$ exchanges the two statistics on each dale section, it follows that we have $\left\{(w)=\rs(\phi(w))$.
\end{proof}\medskip
\begin{cor}
\label{14/2/3cor}
For $t\geq2$, we have
$$
\LB_n(14/2/3,1/2/\dots/t)=\sum_{i=0}^{t-2}\binom{n}{i}+ \sum_{k=1}^{n-2} \left[\sum_{m=2}^{t-1} \binom{n-1}{k+m-1}\sum_{j\geq1}\binom{k-1}{j-1}\binom{m-j}{j}\right]q^k
$$
and the equality
$$
\LB_n(14/2/3,1/2/\dots/t)=\RS_n(14/2/3,1/2/\dots/t).
$$
\end{cor}
\begin{proof}
Avoiding $1/2/\dots/t$ as well as $14/2/3$ adds the restriction that words must have maximum value less than or equal to $t-1$. Following the proof of Theorem~\ref{LB(14/2/3)} with this additional restriction gives the generating function $\LB_n(14/2/3,1/2/\dots/t)$.
Next, we note that the same bijection from Corollary~\ref{LB(14/2/3)=RS(14/2/3)} also provides a bijection from $ R_n(14/2/3,1/2\dots/t)$ to itself, since $\phi$ preserves maximum values. The same map then ensures the second equality.
\end{proof}\medskip
\begin{cor}
The polynomial $\LB_n(14/2/3,123)$ has degree $\lfloor n/3\rfloor$ and leading coefficient equal to
$$
\begin{cases}
\begin{array}{ll}
1& \text{ if } n=3k, \\
n& \text{ if } n=3k+1, \\
\frac{3n^2-7n+14}{6}& \text{ if } n=3k+2,
\end{array}
\end{cases}
$$
for some integer $k$.
\end{cor}
\begin{proof}
Avoiding the pattern $123$ as well as $14/2/3$ adds the restriction that letters can be repeated at most twice in a word. Adapting the notation used in the proof of Corollary~\ref{LB(14/2/3)=RS(14/2/3)}, this implies that, for $w\in R_n(14/2/3,123)$, the dale sections of $w$ must have length equal to $3$ or $4$. Further, these dale sections can only contribute $1$ to $\left\{(w)$. Thus to maximize $\left\{(w)$, we maximize the number of dale sections contained in $w$. It follows from the restrictions on $w$ that this leads to a maximum of $\lfloor n/3\rfloor$.
We now move to the leading coefficient. If $n=3k$ for some integer $k$, then it is clear that the only RGF $w$ in $ R_n(14/2/3,123)$ that achieves this maximum is
$$
w=121343\dots(2k-1)2k(2k-1),
$$
giving a leading coefficient of $1$.
Now let $w\in R_n(14/2/3,123)$ for $n=3k+1$. It follows that $w$ either has one dale section of length $4$, or one plateau of length $1$. In the first case, we note that a dale section of length $4$ has the form $a(a+1)(a+1)a$ or $a(a+1)a(a+1)$. As there will be $k$ total dales in $w$, we have $k$ choices for which dale section to extend, and $2$ choices for how to extend it. This gives $2k$ possible words of the first form. Now assume $w$ has a plateau of length $1$. Note that, once the index of this plateau has been chosen, the rest of the word is uniquely determined. As such, we can choose to place the plateau directly in front of any of the $k$ dale sections, or after the last dale section in $w$. This gives $k+1$ possible words of the second form. Summing over both possibilities now gives a leading coefficient of $n=3k+1$.
Finally, we have $w\in R_n(14/2/3,123)$ for $n=3k+2$. There are four distinct possibilities for $w$ in this case. First, $w$ could contain one plateau of length $2$. This gives $k+1$ possibilities as in the previous paragraph. The second possibility is that $w$ contains two plateaus of length $1$. If these plateaus are adjacent, then as in the previous case we have $k+1$ possibilities. Otherwise, we choose $2$ distinct places from these options, giving $\binom{k+1}{2}$ more words. In the third case, $w$ contains one plateau of length $1$ and one dale section of length $4$. We have $k+1$ choices for the plateau, and $2k$ possibilities for the dale section, giving $2k(k+1)$ words of this form. Finally, $w$ could contain two dale sections of length $4$. In this case, we choose two dale sections to extend. As there are two distinct ways to extend each dale section, this gives $4\binom{k}{2}$ such words. Summing over these four cases and using the substitution $n=3k+2$ gives the final result.
\end{proof}\medskip
Our last corollary regarding the pattern $14/2/3$ involves multiple pattern avoidance with two partitions of $[4]$. First, we need a lemma.
\begin{lem}
\label{RGF(14/2/3,13/2/4)}
For an RGF $w$, $w$ is contained in $ R_n(14/2/3,13/2/4)$ if and only if $w$ meets the following restrictions:
\begin{itemize}
\item For $n\geq2$ we have $a_i\geq \max\{a_1,\dots, a_{i-1}\}-1$, and
\item If $i$ is a dale of height $a$, then $a_j=a$ or $a_j=a+1$ for all $j>i$.
\end{itemize}
\end{lem}
\begin{proof}
First, let $\sigma$ avoid $14/2/3$ and $13/2/4$, and let $w=w(\sigma)$. Since $ R_n(14/2/3,13/2/4)$ is a subset of $ R_n(14/2/3)$, the first inequality follows from Lemma~\ref{14/2/3_avoid}. Now assume that $i$ is a dale of height $a$ in $w$, and assume towards a contradiction that there exists $a_j$ in $w$ with $j>i$, $a_j\neq a$, and $a_j\neq a+1$. From the first inequality, it must be that $a_j>a+1$. Because $w$ is an RGF, it follows that $a(a+1)a(a+2)$ exists as a subword in $w$. But now these four elements will cause an occurance of $13/2/4$ in $\sigma$, which is a contradiction.
For the reverse implication, let $\sigma$ be a partition with $w=w(\sigma)$ satisfying the above restrictions. From Lemma~\ref{14/2/3_avoid}, it follows that $\sigma$ will avoid $14/2/3$. To see that $\sigma$ will also avoid $13/2/4$, note that if $\sigma$ contained $13/2/4$, then the subword $abac$ would exist in $w$, with $a\neq b\neq c$. Using the first inequality, we can rule out all cases except $b<a<c$. But, as this implies a dale of height $b$ in $w$, this would lead to a contradiction with respect to the second restriction put on $w$ by the lemma. Thus $\sigma$ must also avoid $13/2/4$.
\end{proof}\medskip
\begin{cor}
We have
$$
\LB_n(14/2/3,13/2/4)=2^{n-1}+\sum_{k=1}^{n-2}\left[\sum_{m\geq 2}\binom{n-1}{k+m-1}\right]q^k
$$
and
$$
\LB_n(14/2/3,13/2/4)=\RS_n(14/2/3,13/2/4).
$$
\end{cor}
\begin{proof}
Following the proof of Theorem~\ref{LB(14/2/3)}, we note that the constant term in this polynomial comes from the layered partitions of $[n]$. Now consider a word $w$ in $ R_n(14/2/3,13/2/4)$ with $\left\{$ equal to $k$ and maximum value $m$, for $k\geq 1$. From the previous lemma, it follows that the $k$ dales in $w$ must come to the right of the $m$ left to right maxima in $w$. As the leading one in $w$ provides the first left to right maximum, it suffices to choose $k+m-1$ other indices where we place the remaining left to right maxima in the left-most $m-1$ indices, and the $k$ dales afterwards. This gives $\binom{n-1}{k+m-1}$ such words, and summing over all possible values of $m$ gives the coefficient of $q^k$ for $k\geq1$.
Finally, we note that the bijection from Corollary~\ref{LB(14/2/3)=RS(14/2/3)} also takes $ R_n(14/2/3,13/2/4)$ to itself. This gives the second equality.
\end{proof}\medskip
For our final result,
we provide two interesting relationships between the avoidance classes $\Pi(1/23)$ and $\Pi(12/3)$.
\begin{prop}
\label{mpprop}
For $n\geq0$, we have the following equalities:
\begin{align*}
\LB_n(1/23)&=\RS_n(12/3),\\
\LS_n(1/23)&=\RB_n(12/3).
\end{align*}
\end{prop}
\begin{proof}
We will prove this theorem by providing a bijection that maps from $ R_n(1/23)$ to $ R_n(12/3)$. This bijection will interchange the $\left\{$ and $\rs$ statistics, as well as the $\ls$ and $\right\}$ statistics. Let $w$ be an element of $ R_n(1/23)$. By Theorem~\ref{avoidance}, we know that $w$ is of the form $1^l23\dots m$, with possibly a single one inserted.
Let $j$ be the number of ones in $w$, and let $i$ be the index of the rightmost one in $w$. We define $\phi: R_n(1/23)\mapsto R_n(12/3)$ as
$$
\phi(w)=123\dots(n-j+1)(n-i+1)^{j-1}.
$$
From the characterization of $ R_n(12/3)$ provided in Theorem~\ref{avoidance}, it follows that $\phi(w)$ is indeed contained in $ R_n(12/3)$. Furthermore, by Corollary~\ref{avoidancecor} we know that $\# R_n(1/23)=\# R_n(12/3)$. It is also immediate that $\phi$ is injective, which then gives that $\phi$ is a bijection.
Now we show that $\phi$ takes the $\left\{$ statistic to the $\rs$ statistic. First, note that if $w$ is a member of $ R_n(1/23)$ with $\left\{(w)=0$, then $w$ must be of the form
$$
w=1^l23\dots (n-l+1),
$$
for some $l$ with $1\leq l\leq n$. In this case $i=j=l$. Therefore when we apply $\phi$, we are left with
$$
\phi(w)=123\dots(n-l+1)(n-l+1)^{l-1},
$$
and it follows that $\rs(\phi(w))=0$. Now consider the case where $\left\{(w)=k$, for $k>0$. In this instance, $w$ must be of the form
$$
w=1^l23\dots(k+1)1(k+2)\dots (n-l).
$$
It follows that the rightmost one in $w$ has index $l+k+1$, and that there are $l+1$ ones in $w$. Thus when we apply $\phi$, we get
$$
\phi(w)=123\dots (n-l)(n-l-k)^l,
$$
which satisfies $\rs(\phi(w))=k$.
Finally, we show that $\phi$ takes the $\ls$ statistic to the $\right\}$ statistic.
From the proof of Theorem~\ref{F1/23}, we know that if $w\in R_n(1/23)$ with maximum value $m$, then $\ls(w)=\binom{m}{2}$. Similarly, from the proof of Theorem~\ref{F12/3}, if $w'\in R_n(12/3)$ with maximum value $m'$, then $\right\}(w)=\binom{m'}{2}$. Since $\phi$ preserves maximum values, it follows that $\ls(w)=\right\}(\phi(w))$.
\end{proof}\medskip
\section{Open problems and future research}
\label{secopfr}
We have far from exhausted the possible avenues of research concerning these statistics on avoidance classes. Here are some open problems and indications of future avenues to pursue.
\medskip
1. {\bf Partions of larger sets.} As we saw in Section~\ref{secmr}, there are interesting results about the pattern $\pi=14/2/3$. It is natural to consider other partitions of $[n]$ for $n\ge4$. For example, $\Pi_n(13/24)$ is the set of noncrossing partitions introduced by Kreweras~\cite{kre:pnc}. We will be considering the noncrossing case in a future paper~\cite{ddggprs:rgf}.
\medskip
2. {\bf Equidistribution.} In their original paper, Wachs and White showed that $\left\{$ and $\rs$ are equidistributed (have the same generating function) over all RGFs of length $n$ and maximum $m$.
They also showed that the same holds for $\ls$ and $\right\}$. We have seen that these pairs of statistics are equidistributed over various avoidance classes as well in Corollaries~\ref{1/2/3cor}, \ref{1/23cor}, \ref{13/2cor}, \ref{multicor}, and~\ref{14/2/3cor}. Is there some more general theorem about equidistribution which will have some (or even all) of these results as special cases?
\medskip
3. {\bf Mahonian pairs.} It is well known that the permutation statistics $\inv$ and $\maj$ are equidistributed over the symmetric group ${\mathfrak S}_n$. See Stanley's book~\cite{sta:ec1} for details.
Any statistic on ${\mathfrak S}_n$ which has this same distribution is said to be {\em Mahonian}.
In~\cite{ss:mp}, Sagan and Savage defined a pair of subsets $(S,T)$ of ${\mathfrak S}_n$ to be a {\em Mahonian pair} if the distribution of $\maj$ over $S$ is the same as the distribution of $\inv$ over $T$. They give connections of this concept with the Rogers-Ramanujan identities, the Catalan
triangle, and the Greene-Kleitman decomposition of a Boolean algebra into symmetric chains. Again, we have seen similar examples in Corollary~\ref{multicor} and Proposition~\ref{mpprop}. This indicates that exploring the analogous concept for the Wachs and White statistics and avoidance classes should yield interesting results.
\medskip
4. {\bf RGF avoidance.} There is a second notion of avoidance for set partitions which we have not touched on in this article. It is easiest to explain directly in terms of RGFs. We standardize a sequence of integers by replacing all copies of the smallest element of the sequence by $1$, all copies of the next smallest by $2$, and so on. Say that an RGF $w$ contains another one $v$ if there is a subsequence of $w$ which standardizes to $v$. Avoidance is defined in the obvious manner. If $\sigma$ avoids $\pi$ then $w(\sigma)$ avoids $w(\pi)$, but the converse is not always true. We will be investigating this less restrictive notion of pattern avoidance in a forthcoming article~\cite{ddggprs:rgf}.
\nocite{*}
\bibliographystyle{alpha}
\newcommand{\etalchar}[1]{$^{#1}$}
|
\section{Introduction}
Divalent-Eu intermetallic compounds order magnetically due to the indirect RKKY exchange interaction \cite{freeman} between the Eu 4f-spins. Several Eu-based compounds with composition EuTX$_3$, where T is a $d$-transition element and X~=~Si or Ge, crystallizing in the non-centrosymmetric BaNiSn$_3$-type structure are known. Of these, the magnetic properties of EuPtSi$_3$ \cite{Neeraj2010EuPtSi3}, EuPtGe$_3$ \cite{Neeraj2012EuPtGe3}, EuPdGe$_3$ \cite{EuPdGe3} and EuNiGe$_3$ \cite{Johnston_EuNiGe3,Arvind_EuNiGe3} have recently been reported in the literature in single crystal samples. In these materials, an anisotropic behaviour of the 2\,K magnetisation seems to be the prerequisite for the existence of a cascade of close transitions, around 15\,K: a transition from the paramagnetic to an incommensurate, moment modulated antiferromagnetic (AF) state occurs first, followed by another one to a single moment regular AF state, a few K below. This is the case for EuPtSi$_3$ and EuNiGe$_3$, whereas EuPtGe$_3$ shows a unique transition and an isotropic behaviour of the magnetisation.
Here, we report on the magnetic properties of iso-structural EuTSi$_3$ (T=Rh and Ir) compounds, with a single crystal sample for EuRhSi$_3$\ only. An early M\"ossbauer spectroscopy study of these two compounds was performed in Ref.\cite{Chevalier}. We show that these two materials belong to the ``transition cascade'' type, with an anisotropic behaviour of the magnetisation documented for EuRhSi$_3$. We present in addition specific heat, transport and $^{151}$Eu M\"ossbauer spectroscopy data. We find that the magnetic and transport properties in these two materials are notably enhanced with respect to those in other members of the series, and we show that this enhancement can be attributed to an unusually large density of conduction electron states at the Fermi level $n(E_F)$. As a remarkable result, the values of the spin-wave gap derived from such different techniques as resistivity, specific heat and single crystal magnetisation measurements are in good agreement.
\section{Experimental}
Polycrystalline samples of EuIrSi$_3$\ and EuRhSi$_3$\ were prepared by melting Eu (99.9\% purity), Ir/Rh (99.99\%) and Si (99.999\%) in an arc furnace under an inert argon atmosphere.
Single crystal growth of the two Eu compounds was tried using Sn and In as flux and following the same protocol as reported in Refs.\cite{Neeraj2010EuPtSi3,Neeraj2012EuPtGe3}. Powder-diffraction spectra were recorded on a Phillips Pan-analytical set up using Cu-$K_\alpha$ radiation. The magnetisation as a function of field (up to 16\,T) and temperature (1.8 to 300\,K) was measured using Quantum Design MPMS and VSM magnetometers. The electrical resistivity between 1.8 and 300\,K in zero and applied fields, and the heat capacity were measured in a Quantum Design PPMS set-up. $^{151}$Eu M\"ossbauer spectra were recorded at various temperatures using a commercial $^{151}$SmF$_3$ source mounted on a constant acceleration spectrometer.
\section{Results and Discussion}
\subsection{Structure}
Our attempt to grow single crystals succeeded only for EuRhSi$_3$, with In as flux. Powder diffraction spectra of EuIrSi$_3$\ and EuRhSi$_3$\ could be indexed on the basis of the BaNiSn$_3$ type structure (space group $I4mm$). The lattice parameters obtained by the Rietveld analysis of the powder diffraction spectra are in good agreement with the previously reported values \cite{Chevalier}.
\subsection{Susceptibility and isothermal magnetisation}
\begin{figure}[h]
\includegraphics[width=0.40\textwidth]{Chi_Tn.pdf}
\caption{\label{Chi_T} Magnetic susceptibility $\chi(T)$ of single crystalline EuRhSi$_3$\ with a field of 0.1\,T along [001] and [100]. The slight increase of $\chi$ below 8\,K is due to some parasitic phase, though no unidentified line appears in the x-ray diffraction spectrum. Inset: close-up of $\chi(T)$ for EuRhSi$_3$\ with $H~\parallel$~[100] and for polycrystalline EuIrSi$_3$.}
\end{figure}
The susceptibility of EuRhSi$_3$\ measured in a field of 0.1\,T applied along [001] and [100] is shown below 60\,K in Fig.\ref{Chi_T}. It is strongly anisotropic and a clear peak near 49\,K shows the onset of the AF phase. The close-up in the inset of Fig.\ref{Chi_T} shows that this peak is split (arrows). In agreement with the heat capacity and $^{151}$Eu M\"{o}ssbauer data ({\it vide infra}), these two peaks correspond to closely spaced magnetic transitions near 48 and 46\,K. The susceptibility at high temperature (not shown) is nearly isotropic and a fit of the 1/$\chi$ data to a Curie-Weiss law furnishes effective moments $\mu_{\rm eff}$=7.39 and 7.52\,$\mu_{\rm B}$\ and paramagnetic Curie temperature $\theta_{\rm p}=-$11 and $-$14\,K for H along [001] and [100] respectively. These effective moments are lower than the free ion value of 7.94\,$\mu_{\rm B}$\ expected for Eu$^{2+}$ ($g=2$, $S=7/2$), which is due either to the presence of residual In-flux or to a slight Eu off-stoichiometry, corresponding to about 10 at.\% Eu deficit. As to polycrystalline EuIrSi$_3$, its susceptibility is shown between 30 and 60\,K in the inset of Fig.\ref{Chi_T}: two anomalies (arrows) are also present, near 52 and 43\,K, witnessing the same phenomenon as in EuRhSi$_3$. The $\theta_{\rm p}$ value is $-$17\,K, while $\mu_{\rm eff}$=7.7\,$\mu_{\rm B}$\ is closer to the Eu$^{2+}$free ion value.
\begin{figure}[h]
\centerline{\includegraphics[width=0.37\textwidth]{MH2n.pdf}}
\vspace{0.1cm}
\centerline{\includegraphics[width=0.36\textwidth]{eurh_mH.pdf}}
\caption{\label{MH} (a) Isothermal magnetisation $M(H)$ of EuRhSi$_3$\ at 2\,K along major crystallographic directions. Inset: $M(H)$ curves at higher temperatures. (b) Simulation of the $M(H)$ curves in EuRhSi$_3$\ according to the model described in the text. The calculated values are 10\% higher than the data due to the assumed Eu mass deficit in the single crystal sample.}
\end{figure}
The N\'eel temperature in intermetallic compounds is the result of the indirect RKKY exchange between $4f$ spins {\bf S} mediated by the $4f$-conduction electron coupling with constant $J_{kf}$:
\begin{equation}
\label{kf}
{\cal H}_{kf} = -J_{kf}\ {\bf s}.{\bf S},
\end{equation}
where {\bf s} is the conduction electron spin density, and it has the form: $T_{\rm N} \propto J_{kf}^2\ n(E_F)\ S(S+1)$ \cite{freeman}. Its value in these materials, near 50\,K, is much larger than in other members of the series ($\simeq 15$\,K), pointing to an enhanced value of $n(E_F)$.
The isothermal magnetisation versus field scan at 2\,K in EuRhSi$_3$, shown in Fig.\ref{MH}(a), is a textbook example of an antiferromagnet with the tetragonal [001] c-axis as the easy axis of magnetisation and the (001) $ab$-plane as the hard plane, in line with the susceptibility data of Fig.\ref{Chi_T}. For $H~\parallel$~[001], an unusually large spin-flop field of 13\,T is observed. A linear extension of the magnetisation curves for both field directions up to a saturation moment value $m_0$= 7\,$\mu_{\rm B}$\ yields large spin-flip fields H$_{sf}^c \simeq$28\,T and H$_{sf}^a \simeq$41\,T. The latter is larger since the field aligns the moments in the hard magnetic plane. In the standard molecular field theory \cite{Herpin}, one has: H$_{sf}^c$=2(H$_e -$H$_a$) and H$_{sf}^a$=2(H$_e +$H$_a$), where H$_e$ and H$_a$ are respectively the exchange and the ``anisotropy'' field, the latter being defined as H$_a=K/m_0$ where $K$ is the anisotropy energy density. Then we obtain H$_e$=17\,T and H$_a$=3.35\,T, and the critical spin-flop field H$_{cr}$ = 2$\sqrt{{\rm H}_a({\rm H}_e-{\rm H}_a)}$ = 13.5\,T, in very good agreement with experiment. The spin-flop field in EuRhSi$_3$\ is much larger than in EuNiGe$_3$ \cite{Arvind_EuNiGe3} (2-3\,T), a logical consequence of a large H$_e$ linked to the high $T_{\rm N}$ value.
In the inset of Fig.\ref{MH}(a) are plotted the magnetisation curves along the easy axis at higher temperature. On heating, the spin-flop field decreases and the jump at the spin-flop broadens.
\subsection{Modeling}
However, such a simple model as described above cannot account for the quite different values of $T_{\rm N}$ and $\vert \theta_p \vert$, and the oscillatory nature of the RKKY exchange compels one to introduce at least two different exchange integrals. Therefore, we use a numerical self-consistent calculation which has been described thoroughly in Ref.\cite{Arvind_EuNiGe3}: i) the infinite range dipolar interaction is added, ii) two exchange integrals $J_1$ (intra-plane first neighbor) and $J_2$ (interplane first neighbor) for the centered tetragonal structure are considered, as well as exchange anisotropy, and iii) the single ion crystalline anisotropy is described by a term $D S_z^2$, where $Oz$=c. We also assume a magnetic structure made of ferromagnetic (ab) planes ($J_1 > 0$) coupled antiferromagnetically along c ($J_2 < 0$), i.e. a propagation vector {\bf k}=[001]. To obtain a first estimation of the exchange constants, the molecular field equations linking $T_{\rm N}$ and $\theta_p$ with $J_a$ and $J_c$ are used \cite{Johnston_EuNiGe3,Arvind_EuNiGe3}, yielding: $J_a$=1.1\,K and $J_c=-$0.88\,K. We have taken $T_{\rm N}$=60\,K (see the section about M\"ossbauer spectroscopy) and a mean value $\theta_p=-$13\,K. Figure \ref{MH}(b) shows the curves which reproduce best the experimental data, with a small exchange anisotropy: $J_a^{\parallel}$=0.8\,K, $J_a^\perp$=1.1\,K, $J_c^{\parallel}=-$0.7\,K, $J_c^\perp=-$0.9\,K and a crystalline anisotropy parameter $D=-$0.85\,K. The model yields $T_{\rm N}$=57\,K and $\theta_p$=$-$6 and $-$17\,K for $H~\parallel$~[001] and $H~\parallel$~[100] respectively (experimental values $-$11 and $-$14\,K resp.). At 30\,K and above, the model cannot exactly reproduce the smoothing of the spin-flop transition (see inset of Fig.\ref{MH}(b)).
\subsection{Electrical resistivity}
The electrical resistivity of EuIrSi$_3$\ and EuRhSi$_3$\ (and of non-magnetic LaIrSi$_3$) is shown in Figs.\ref{RT}(a) and (b). Above T$_N$ it varies almost linearly up to 300\,K in both compounds due to phonon scattering. It shows a small upturn at the magnetic transitions, likely caused by antiferromagnetic fluctuations just above $T_{\rm N1}$ and decreases rapidly on cooling due to a strong depletion in the spin-disorder scattering in the magnetically ordered state. One striking feature is that the resistivity of polycrystalline EuIrSi$_3$\ is one order of magnitude larger than that of single crystal EuRhSi$_3$. We reckon that this is probably due to strong grain boundary scattering which enhances the resistivity of the polycrystalline sample. The residual resistivity $\rho_0=\rho(2\,K)$ is lower than 10\,$\mu$Ohm.cm in both materials (see Table \ref{Table1}), indicating they are chemically well ordered.
\begin{figure}[h]
\centerline{\includegraphics[width=0.35\textwidth]{Rho_T3.pdf}}
\caption{\label{RT} (a) Electrical resistivity as a function of temperature $\rho(T)$ of EuIrSi$_3$ in zero field and in a 14\,T field. (b) $\rho(T)$ data of EuRhSi$_3$ when current density $J~\parallel$~[100] and [001]. The insets show the low temperature data together with their fit to Eq.(\ref{rho_gap}).}
\end{figure}
The total spin disorder resistivity $\rho_m$, estimated as $\rho(T_{\rm N}) - \rho_0$, amounts to about 35\,$\mu$Ohm.cm in EuRhSi$_3$\ while it is of the order of a few $\mu$Ohm.cm in the other compounds of the series \cite{Neeraj2010EuPtSi3,Neeraj2012EuPtGe3,Arvind_EuNiGe3}. Since $\rho_m \propto J_{kf}^2\ n(E_F)\ S(S+1)$ \cite{freeman}, its enhanced value in EuRhSi$_3$\ is in line with a stronger $n(E_F)$.
In the AF phase below 20\,K, magnon scattering is the dominant mechanism for resistivity. With an antiferromagnetic spin wave dispersion relation:
\begin{equation}
\label{disp}
E(k)=\sqrt{\Delta^2+\sigma k^2}
\end{equation}
where $\Delta$ is the anisotropy gap and $\sigma$ the spin-wave stiffness, the electrical resistivity in zero field for $T < \Delta$ is given by \cite{fontes}:
\begin{eqnarray}
\label{rho_gap}
\rho (T)=\rho_0 &+& A \Delta^2 {\left(\frac{T}{\Delta}\right)}^{1/2}e^{- \Delta/T} \\ \nonumber
& \times &\left[1+2/3 \left(\frac{T}{\Delta}\right)+2/15\left(\frac{T}{\Delta}\right)^2\right],
\end{eqnarray}
\begin{table}[!]
\caption{\label{Table1} Parameters obtained after fitting Eq.(\ref{rho_gap}) to the zero field resistivity data of EuRhSi$_3$\, with $J \parallel [001]$, and of EuIrSi$_3$.}
\begin{tabular}{|c|c|c|c|} \hline \hline
& $\rho_0(\mu\Omega~cm$) & $A(\mu\Omega~cm/K^2$) & $\Delta (K)$ \\ \hline
EuRhSi$_3$ \ & 5.1 & 0.04 & 23.3 \\ \hline
EuIrSi$_3$ & 5.27 & 0.43 & 25.2 \\ \hline
\end{tabular}
\end{table}
where $A$ a material dependent constant. Equation (\ref{rho_gap}) provides a good fit to the zero field data in both compounds between 1.8 and 15\,K (and also for a 14\,T field), as shown by the solid lines in the insets of Fig.\ref{RT}. In EuIrSi$_3$, the zero field spin wave gap is 25.2\,K and it is 23.5\,K in EuRhSi$_3$. The AF spin wave gap is linked to the exchange and anisotropy fields H$_e$ and H$_a$ by the relation \cite{tqs}:
\begin{equation}
\label{gap_h}
\Delta = 2g \mu_B \sqrt{{\rm H}_a\ ({\rm H}_a+{\rm H}_e)}.
\end{equation}
Using the values determined above in EuRhSi$_3$: H$_e$=17\,T, H$_a$=3.35\,T and $g$=2, one obtains $\Delta$ = 22.1\,K, in excellent agreement with experiment in this material and close to the value in EuIrSi$_3$.
At a field of 14\,T, $\rho_0$ increases to 35.9\,$\mu\Omega$cm in EuIrSi$_3$\ and to 9\,$\mu\Omega$cm in EuRhSi$_3$\ for $H~\parallel$~[100] and $J~\parallel$~[001].
\begin{figure}[h]
\centerline{\includegraphics[width=0.40\textwidth]{MR2.pdf}}
\caption{\label{MR} Magnetoresistance ($MR$) of EuIrSi$_3$ at different temperatures. (b) $MR$ of EuRhSi$_3$ for $H~\parallel$~[100] and $H~\parallel$~[001] with $J~\parallel$~[100] in both the cases.}
\end{figure}
The significant increase of $\rho_0$ with field corresponds to a positive magnetoresistance ratio $MR$, defined as $MR=[\rho(H)-\rho(H=0)]/\rho(H=0)$, as expected for antiferromagnets \cite{yamada,mcewen}. The plots of $MR$ at selected temperatures in EuIrSi$_3$\ are shown in Fig.\ref{MR}(a). It reaches giant values: at 14\,T, it exceeds 100\% at 10\,K and it reaches 600\% at 2\,K. At higher temperatures (not shown), $MR$ decreases and becomes negative near 47\,K, shows an absolute minimum near 50\,K and has a small residual value of $-$3\% at 200\,K, far above $T_{\rm N1}$.
In EuRhSi$_3$, the $MR$ is lower, as shown in Fig.\ref{MR}(b) at 2\,K for $J~\parallel$~[100] and $H~\parallel$~[001] and [100]. A small positive jump in the $MR$ occurs for $H~\parallel$~[001] at the spin-flop field (13\,T), although the $MR$ is expected to drop to a small (positive) value above the spin-flop field, according to the molecular field calculation in Ref.\cite{yamada}.
\subsection{Heat capacity}
The main panels of Fig.\ref{HC} show the heat capacity of EuIrSi$_3$\ and EuRhSi$_3$\ together with that of the non-magnetic La-reference. The two plots show two major peaks at 51.8 and 43.1\,K in EuIrSi$_3$\ and at 48.3 and 45.8\,K in EuRhSi$_3$\ in close correspondence with the anomalies seen in the susceptibility data, thus confirming the occurrence of two magnetic transitions in these two compounds. The jump in the heat capacity at the higher transition temperature $T_{\rm N1}$ is approximately 9.5 and 12\,J/mol.K in the Ir and Rh-compound, respectively, which is lower than the value $\delta C_{7/2}$=20.14\,J/mol.K for an equal moment antiferromagnetic transition for $S$=7/2 ions in the mean field model, and closer to that predicted for an amplitude modulated structure (2/3 $\delta C_{7/2}$) \cite{Neeraj2010EuPtSi3,gignoux}. This suggests that at $T_{\rm N1}$ occurs a transition to a modulated moment structure. The transition from this intermediate structure to an equal moment structure takes place at $T_{\rm N2}$. This behaviour is confirmed by the $^{151}$Eu M\"osbauer spectra recorded at few selected temperatures.
Deep in the AF phase, the specific heat should be the sum of 3 terms \cite{conti}:
\begin{eqnarray}
\label{cp}
C(T)~=~\gamma T &+& \beta T^3 + B \Delta^4 {\left(\frac{T} {\Delta}\right)}^{1/2} e^{- \Delta/T} \\ \nonumber &\times& \left[1+39/20 \left(\frac{T}{\Delta}\right)+51/32\left(\frac{T}{\Delta}\right)^2\right],
\end{eqnarray}
where the first linear term is the conduction electron heat capacity, the second is the phonon contribution and the third the magnon heat capacity corresponding to a dispersion law given by Eq.(\ref{disp}) with a gap $\Delta$.
Equation (\ref{cp}) provides a good fit of the data between 1.8 and 10\,K, shown by the solid line in the insets of Figs.\ref{HC}(a) and (b). For EuIrSi$_3$, the best fit estimates of $\gamma$ and $\Delta$ are 30.5\,mJ/mol.K$^2$ and 22.9\,K respectively. The gap value is very close to that inferred from the resistivity data (25.2\,K), which lends credibility to our analysis. In the case of EuRhSi$_3$, one obtains: $\gamma$= 40.3\,mJ/mol.K$^2$ and $\Delta$=17.6\,K, the latter value somewhat lower than the resistivity derived value (23.3\,K). It may be noted that the Sommerfeld coefficient $\gamma$ is an order of magnitude larger than in $sp$-metals, and even larger than in many $d$-metal alloys. The Sommerfeld coefficient is expressed as \cite{tqs}: $\gamma = \frac{2}{3} \pi^2 k_B^2\ n(E_F)$, and its large value must be related to an enhanced $n(E_F)$.
\begin{figure}[h]
\centerline{\includegraphics[width=0.37\textwidth]{HC_2.pdf}}
\caption{\label{HC} (a) Variation of heat capacity with temperature $C(T)$ of EuIrSi$_3$\ and of its non magnetic analogue LaIrSi$_3$. The solid green line represents the calculated entropy S$_{4f}(T)$. Inset shows the fitted Eq.(\ref{cp}) to the the low temperature data (b) $C(T)$ of EuRhSi$_3$\ and LaRhSi$_3$ with a similar legend as for the upper panel.}
\end{figure}
The entropy $S_{4f}$ associated with the magnetic ordering was estimated by integrating $C_{4f}/T$, where $C_{4f}$ was obtained by subtraction from $C(T)$ of both the normalized heat capacity of LaTSi$_3$, taking into account the slight difference in the atomic masses of Eu and La, and the large conduction electron contribution $\gamma T$. It is seen that $S_{4f}$ attains the expected value of $R \ln8$ (for $S$~=~7/2 Eu$^{2+}$ ions) close to $T_{\rm N1}$ in both compounds.
\subsection{$^{151}$Eu M\"{o}ssbauer spectra}
The M\"ossbauer spectra on the isotope $^{151}$Eu in EuIrSi$_3$\ are shown in Fig.\ref{Moss1} in the two temperature ranges defined as phase I ($T <$43.1\,K) and phase II (43.1\,K $<$ T $<$ 51.8\,K). Spectra in EuRhSi$_3$\ are similar. The spectrum at 4.2\,K is a standard hyperfine field pattern characteristic of Eu$^{2+}$ ($S$=7/2, $L$=0), with an isomer shift relative to EuF$_3$ of $-$7.92(5)\,mm/s which matches well with the value reported in Ref.\cite{Chevalier}. This spectrum presents the peculiarity of a very high hyperfine field H$_{hf}$ (4.2\,K) = 52.6(3)\,T (51.6\,T in EuRhSi$_3$). This high value can be explained by the large $n(E_F)$ prevailing in these materials. Indeed, in magnetically ordered intermetallic materials with the $L$=0 ion Eu$^{2+}$, the hyperfine field is solely due to the spin polarisation of the $s$-type electrons at the nucleus site. It can be expressed as \cite{bleaney}: $
{\rm H}_{hf}(T) = A \ m(T) + {\rm H}_{ce}$, where the first term is the core polarisation field proportional to the Eu$^{2+}$ moment $m(T)=-g \mu_B \langle S \rangle_T$ and worth $\simeq 34$\,T at saturation \cite{nowik}. The second term H$_{ce}$ is due to the conduction electron spin polarisation $\langle s \rangle_T$ induced by the $4f$ shell through $J_{kf}$ exchange and is given approximately by \cite{bleaney}: $
{\rm H}_{ce} \simeq A_{ce} \ \mu_B \ \langle s \rangle_T $, where $A_{ce}$ is a hyperfine constant. Acording to Eqn.(\ref{kf}), the effective field on {\bf s} is: $\frac{J_{kf} {\bf S}}{g\mu_B}$, and introducing the Pauli susceptibility: $\chi_P = 2 \ \mu_B^2\ n(e_F)$, one obtains:
\begin{equation}
\label{hce}
{\rm H}_{ce} \simeq A_{ce} \ J_{kf} n(E_F)\ m(T).
\end{equation}
In most Eu$^{2+}$ intermetallic materials and in the other members of the series, the saturated hyperfine field amounts to about 30\,T, i.e. H$_{ce}$ is negative and worth a few T. In EuIrSi$_3$\ and EuRhSi$_3$, due to the enhanced $n(E_F)$ value, H$_{ce}$ is much larger ($\simeq$18\,T) and happens to be positive. The same situation holds in EuFe$_4$P$_{12}$, which presents the largest hyperfine field ever measured with $^{151}$Eu (67\,T), implying H$_{ce} \simeq$33\,T \cite{gerard}.
\begin{figure}[h]
\centerline{\includegraphics[width=0.47\textwidth]{Moss_1.pdf}}
\caption{\label{Moss1} $^{151}$Eu M\"ossbauer spectra in EuIrSi$_3$\ at selected temperatures: in the commensurate phase (4.2 and 40\,K), and in the incommensurate phase (45 and 52\,K). The spectrum at 52\,K is typical of the ``coexistence'' region, with a contribution from the paramagnetic phase of some part of the sample (red subspectrum) occurring in case of a first order transition. The green subspectrum represents an impurity phase containing Eu$^{3+}$ (relative intensity 2\%).}
\end{figure}
Whereas the spectra in phase I (below 43\,K) show the presence of a unique hyperfine field, and hence of a unique Eu$^{2+}$ magnetic moment, there occurs a sudden change of the spectral shape above 43\,K, i.e. when entering phase II, where the spectra can be fitted to an incommensurate modulation of hyperfine fields \cite{bonville}. The latter is well described by the first 3 odd harmonics of a Fourier series:
\begin{equation}
\label{eq.4}
H_{hf}(kx)=h_1 \sin(kx)+h_3 \sin(3kx)+h_5 \sin(5kx)
\end{equation}
where $k$ is the propagation vector of the modulation and $x$ the distance along $k$. The three coefficients $h_1$, $h_3$ and $h_5$ were fitted to the spectral shape at each temperature to obtain the modulation profiles shown in Fig.\ref{Moss2}. The modulation becomes more ``squared'' as temperature decreases and approaches the transition to the incommensurate phase.
\begin{figure}[h]
\centerline{\includegraphics[width=0.37\textwidth]{Moss_2.pdf}}
\caption{\label{Moss2} Hyperfine field modulation over a period in EuIrSi$_3$\ in the incommensurate modulated phase II at selected temperatures.}
\end{figure}
\begin{figure}[h]
\centerline{\includegraphics[width=0.37\textwidth]{Moss_3.pdf}}
\caption{\label{Moss3} In EuIrSi$_3$, thermal variation of the hyperfine field in Phase I (black dots) and of the maximum of the hyperfine field modulation in Phase II (blue disks). The dashed line is the $S$=7/2 mean field law with $T_{\rm N}$=60\,K, extrapolated in phase II above 43\,K (in red).}
\end{figure}
The thermal variation of the hyperfine field is plotted in Fig.\ref{Moss3}. In phase I, the hyperfine field values approximately follow a mean field law for $S$=7/2, in line with its proportionality to the Eu$^{2+}$ moment derived above. The transition temperature of this mean field law (60\,K) does not correspond to the actual N\'eel temperature because of the presence of the commensurate - incommensurate transition at 43.1\,K and of the first-order character of the transition to the paramagnetic phase at 51.8\,K (see spectrum at 52\,K in Fig.\ref{Moss1}).
\section{Conclusion}.
The whole set of our thermodynamic and spectroscopic measurements in the two divalent Eu intermetallics EuIrSi$_3$\ and EuRhSi$_3$\ can be coherently and qualitatively interpreted by assuming a high density of electronic band states at the Fermi energy, which sets them apart from the other members of the EuTX$_3$ family. We attribute the observed enhanced values to a large $n(E_F)$ rather than to an anomalously large $4f$-conduction electron coupling $J_{kf}$ since the Sommerfeld coefficient does not involve $J_{kf}$ and there is {\it a priori} no reason for the stable Eu$^{2+}$ $4f$ shell to be prone to strong hybridisation with the conduction band. The two compounds present a cascade of magnetic transitions near 50\,K, from a paramagnetic to an incommensurate modulated, then to a commensurate antiferromagnetic phase. In the EuRhSi$_3$\ single crystal sample, we could evidence an important anisotropy of the magnetisation, confirming the link between these two phenomena.
|
\section{Introduction}
\textit{What are the best possible lower and upper bounds on the
$k$\textit{'th largest eigenvalue }$\lambda_{k}\left( G\right) $\textit{ of
a graph }$G$\textit{ of order }$n?$
Yuan Hong raised this fundamental question in 1993, in his paper \cite{Hon93}.
Apparently he was unaware that five years earlier Powers \cite{Pow89}, p. 5,
had published the following result:\medskip
\emph{If }$G$\emph{ is a connected graph of order }$n$\emph{, then
\begin{equation}
\lambda_{k}\left( G\right) \leq\left\lfloor n/k\right\rfloor . \label{Powbo
\end{equation}
It is not hard to realize that if inequality (\ref{Powbo}) were true, it would
essentially answer Hong's question. Alas, it is not. Its proof is flawed, and
it fails for all $k\geq5.$ However, in fairness, the inequality certainly
holds for $k=1,2,$ while for $k=3,4,$ it is a challenging open problem. For
that matter, except for $k=2,$ connectedness is an irrelevant premise in these questions.
Other than this unsuccessful attempt, the general problem of Hong has never
been tackled seriously. This is all the more inexplicable, as the problem is
indeed challenging, easier for some values of $k,$ and well beyond reach for
other. What is more, Hong's problem is not a backyard puzzle that is of
interest only to spectral graph theorists; it is related to other fundamental
areas of combinatorics and analysis, like existence of symmetric Hadamard
matrices, Ramsey's theorem, and extremal norms of graphs. We feel that the
appeal and the importance of Hong's problem should attract the attention of
many a researcher, and to this effect we take a few steps in the present paper.
We shall extend Hong's problem to the $k$'th largest singular value
$\lambda_{k}^{\ast}(G)$ of $G,$ and shall give upper and lower bounds on
$\lambda_{k}\left( G\right) $ and $\lambda_{k}^{\ast}(G)$, including a few
exact results and asymptotics for the general cases. Many open problems and
questions will be raised to outline directions for further study.
Two fundamental results underpin our constructions: first, the strongly
regular graphs of Taylor \cite{Tay71,Tay77}; and second, Kharaghani's method
for constructing Hadamard matrices \cite{Kha85}. These two topics deserve to
be known better in spectral graph theory, as their potential uses seem indeed unlimited.
We shall apply the same constructions to other open problems, like, e.g.,
Nordhaus-Gaddum problems of the following kind:\medskip\
\emph{If }$G$\emph{ is a graph of order }$n$ \emph{and }$\overline{G}$\emph{
is its complement,} \emph{how large can }$\lambda_{k}\left( G\right)
+\lambda_{k}(\overline{G})$\emph{ be}$?\medskip$\emph{ }
Such problems have been raised in \cite{Nik07}, and some recent progress has
been given in \cite{NiYu14}. We shall exhibit a new infinite family of
solutions, and derive general asymptotics.\medskip
We make also progress with an open problem about maximum Ky Fan norms of
graphs. Recall that the \emph{Ky Fan }$k$\emph{-norm} of a graph $G$ is
defined as $\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{k}^{\ast
}\left( G\right) $. In \cite{Ni11}, the following problem has been
raised:\medskip
\emph{If }$G$\emph{ is a graph of order }$n,$\emph{ how large can the Ky Fan
}$k$\emph{-norm of }$G$\emph{ be? \medskip}
Note that the Ky Fan $n$-norm is also known as the \emph{trace norm} of $G,$
and has been extensively studied under the name \emph{graph energy,} a concept
introduced by Gutman in \cite{Gut78}. There is vast research on graph energy,
but Ky Fan norms can open even larger horizons. Here we shall solve the above
problem whenever $k$ is an even square.\medskip
The structure of the paper is as follows: In Section \ref{Hp} we present
results on Hong's problem. Section \ref{NGs} is dedicated to spectral
Nordhaus-Gaddum problems, and in Section \ref{KFs} we present results on Ky
Fan norms of graphs. Section \ref{NBS} is for reader's convenience: it
contains references, notation and basics on Weyl's inequalities, blow-ups of
graphs, Taylor's strongly regular graphs, and symmetric Latin squares. Section
\ref{ps} contains the proofs of several theorems, which are either too
involved of would have disrupted the exposition. Finally, Section
\ref{RM}contains a selection of the presented results and problems, translated
from graphs to symmetric $\left( -1,1\right) $-matrices. It becomes obvious
that such matrices provide a more balanced and natural setup for such
research. In particular, the strongly regular graphs of Taylor are translated
into a $\left( -1,1\right) $-matrix with a rather peculiar spectrum.
\section{\label{Hp}Hong's problem and its variations}
Let $G$ be a graph of order $n.$ The eigenvalues $\lambda_{1}\left( G\right)
,\ldots,\lambda_{n}\left( G\right) $ of $G$ are the eigenvalues of its
adjacency matrix $A\left( G\right) ,$ ordered as $\lambda_{1}\left(
G\right) \geq\cdots\geq\lambda_{n}\left( G\right) .$ The singular values
$\lambda_{1}^{\ast}\left( G\right) ,\ldots,\lambda_{n}^{\ast}\left(
G\right) $ of $G\ $are the absolute values of $\lambda_{1}\left( G\right)
,\ldots,\lambda_{n}\left( G\right) ,$ ordered as $\lambda_{1}^{\ast}\left(
G\right) \geq\cdots\geq\lambda_{n}^{\ast}\left( G\right) .$ In particular,
$\lambda_{1}^{\ast}\left( G\right) =\lambda_{1}\left( G\right) ,$ and
\[
\left\{ \lambda_{1}^{\ast}\left( G\right) ,\lambda_{2}^{\ast}\left(
G\right) ,\ldots,\lambda_{n}^{\ast}\left( G\right) \right\} =\left\{
\lambda_{1}\left( G\right) ,\left\vert \lambda_{2}\left( G\right)
\right\vert ,\ldots,\left\vert \lambda_{n}\left( G\right) \right\vert
\right\} .
\]
Note that, in general, graph singular values cannot be reduced to graph
eigenvalues, as the two multisets may be ordered very differently.
We shall extend the original problem of Hong to the largest singular values of
graphs, and shall bring to the fore the study of the smallest eigenvalues.
These changes correspond to the present day interest in these spectral
parameters.\medskip
Let $n\geq k\geq1.$ Define the functions $\lambda_{k}\left( n\right) ,$
$\lambda_{-k}\left( n\right) ,$ and $\lambda_{k}^{\ast}\left( n\right) $
as\ \
\begin{align*}
\lambda_{k}\left( n\right) & =\max_{v\left( G\right) =n}\text{
\lambda_{k}\left( G\right) ,\\
\lambda_{-k}\left( n\right) & =\max_{v\left( G\right) =n}\text{
}\left\vert \lambda_{n-k+1}\left( G\right) \right\vert ,\\
\lambda_{k}^{\ast}\left( n\right) & =\max_{v\left( G\right) =n
\text{\ }\lambda_{k}^{\ast}\left( G\right) .
\end{align*}
Note that if $n\geq\binom{2k-1}{k-1},$ then $\lambda_{n-k+1}\left( G\right)
$ is always nonpositive (see Theorem \ref{lob} below), and so, $\min_{v\left(
G\right) =n}$ $\lambda_{n-k+1}\left( G\right) =-\lambda_{-k}\left(
n\right) ;$ thus, the use of the absolute value in the definition of
$\lambda_{-k}\left( n\right) $ is just to make the setup more
uniform.\medskip
Now, we restate Hong's problem into two separate problems:
\begin{problem}
\label{mpro}For any $k\geq1,$ find $\lambda_{k}\left( n\right) ,$
$\lambda_{-k}\left( n\right) ,$ and \ $\lambda_{k}^{\ast}\left( n\right)
.$
\end{problem}
\begin{problem}
\label{minpro} For any $k\geq1,$ find $\min\limits_{v\left( G\right) =n
$\ $\lambda_{k}\left( G\right) ,$ $\max\limits_{v\left( G\right) =n}$
$\lambda_{n-k+1}\left( G\right) ,$ and $\min\limits_{v\left( G\right) =n
$\ $\lambda_{k}^{\ast}\left( G\right) .$
\end{problem}
This separation is justified, as the two problems are of incomparable
difficulty: indeed, presently the full solution of Problem \ref{mpro} is
beyond reach, while we shall dispose of Problem \ref{minpro} right away.
Indeed, if $n\geq k,$ the complete graph $K_{n}$ of order $n$ satisfies
$\lambda_{k}\left( K_{n}\right) =-1;$ and likewise, the edgeless graph
$\overline{K}_{n}$ of order $n$ satisfies $\lambda_{n-k+1}(\overline{K
_{n})=0$ and $\lambda_{k}^{\ast}(\overline{K}_{n})=0.$ These bounds are also
best possible: indeed, obviously $\lambda_{k}^{\ast}\left( G\right) \geq0$
for any graph $G$ of order $n\geq k$; and for $\lambda_{k}$ and\ $\lambda
_{n-k+1}$ this fact is true in view of the following theorem:
\begin{theorem}
\label{lob}If $n\geq\binom{2k-1}{k-1}$ and $G\ $is a graph of order $n$, then
\begin{equation}
\lambda_{k}\left( G\right) \geq-1\text{ \ \ and \ \ }\lambda_{n-k+1}\left(
G\right) \leq0. \label{bo1
\end{equation}
\end{theorem}
\begin{proof}
We shall use Ramsey's theorem, whose application in graph spectra has been
pioneered only recently, in \cite{NiYu14} and \cite{ZhCh14}.
The classical bound of Erd\H{o}s and Szekeres implies that every graph of
order at least $\binom{2k-1}{k-1}$ contains either a complete graph on $k+1$
vertices or an independent set on $k$ vertices. If $G$ contains a complete
graph on $k+1$ vertices, then Cauchy's interlacing theorem implies that
\[
\lambda_{k}\left( G\right) \geq\lambda_{k}\left( K_{k+1}\right) =-1\text{
\ and \ }\lambda_{n-k+1}\left( G\right) \leq\lambda_{2}\left(
K_{k+1}\right) =-1,
\]
so (\ref{bo1}) follows. If $G$ contains an independent set on $k$ vertices,
then Cauchy's interlacing theorem implies that
\[
\lambda_{k}\left( G\right) \geq\lambda_{k}(\overline{K}_{k})=0\text{ \ and
\ \ \ }\lambda_{n-k+1}\left( G\right) \leq\lambda_{1}(\overline{K}_{k})=0,
\]
and (\ref{bo1}) follows again.
\end{proof}
\medskip
Now, let us turn to Problem \ref{mpro}. We start with an observation, which
exhibits some dependencies between the functions $\lambda_{k}\left( n\right)
,$ $\lambda_{-k}\left( n\right) ,$ and $\lambda_{k}^{\ast}\left( n\right)
.$
\begin{proposition}
\label{spro}If $k\geq2,$ then
\[
\lambda_{k}\left( n\right) \leq\lambda_{k}^{\ast}\left( n\right) ,\text{
\ \ }\lambda_{-k+1}\left( n\right) \leq\lambda_{k}^{\ast}\left( n\right)
,\text{ \ \ and \ \ \ }\lambda_{k}\left( n\right) +1\leq\lambda
_{-k+1}\left( n\right) .
\]
\end{proposition}
The first two inequalities follow from the definition of $\lambda_{k}^{\ast
}\left( n\right) .$ For the last inequality recall that Weyl's inequalities
(see \ref{WS}) imply that if $G$ is a graph of order $n$ and $2\leq k\leq n,$
then $\lambda_{k}\left( G\right) +\lambda_{n-k+2}(\overline{G})\leq-1,$ and
so, $\lambda_{k}\left( G\right) +1\leq|\lambda_{-k+1}(\overline{G
)|.$\medskip\
Our first goal is to establish concise asymptotics of $\lambda_{k}\left(
n\right) ,$ $\lambda_{-k}\left( n\right) ,$ and $\lambda_{k}^{\ast}\left(
n\right) $. To this effect, for any integer $k\geq1,$ define the real numbers
$c_{k},$ $c_{-k},$ and $c_{k}^{\ast}$ as
\begin{align*}
c_{k} & =\sup\left\{ \text{ \ \ }\lambda_{k}\left( G\right)
/n\text{\ \ \ \ \ \ }:\text{ }G\text{ is a graph of order }n\geq k\right\}
,\\
c_{-k} & =\sup\left\{ \left\vert \lambda_{n-k+1}\left( G\right)
\right\vert /n\text{ }:\text{ }G\text{ is a graph of order }n\geq k\right\}
,\\
c_{k}^{\ast} & =\sup\left\{ \text{$\ $\ \ $\ \lambda_{k}^{\ast}\left(
G\right) /n$ \ \ \ \ }:\text{ }G\text{ is a graph of order }n\geq k\right\}
.
\end{align*}
Clearly, these definitions imply that if $G$ is a graph of order $n,$ the
\[
\lambda_{k}\left( G\right) \leq c_{k}n,\text{ \ \ }\lambda_{n-k+1}\left(
G\right) \geq-c_{-k}n,\text{ \ \ \ and \ \ \ }\lambda_{k}^{\ast}\left(
G\right) \leq c_{k}^{\ast}n.
\]
The above bounds are handy, and fortunately they are also tight, as shown by
the following theorem, which can be proved with the methods of \cite{Nik06}:
\begin{theorem}
\label{limth}For every $k\geq1,
\[
\lim_{n\rightarrow\infty}\text{ }\lambda_{k}\left( n\right) /n=c_{k},\text{
\ }\lim_{n\rightarrow\infty}\text{ }\lambda_{-k}\left( n\right)
/n=c_{-k},\text{ \ \ and \ \ }\lim_{n\rightarrow\infty}\text{ }\lambda
_{k}^{\ast}\left( n\right) /n\text{ \ }=c_{k}^{\ast}.
\]
\end{theorem}
Therefore, a good deal of information about $\lambda_{k}\left( G\right)
,$\ $\lambda_{n-k+1}\left( G\right) ,$\ and $\lambda_{k}^{\ast}\left(
G\right) $ can be obtained if we knew the constants $c_{k},$ $c_{-k},$ and
$c_{k}^{\ast}$ or some good estimates thereof.
\subsection{\textbf{Upper bounds on }$\lambda_{k}\left( n\right) ,$
$\lambda_{-k}\left( n\right) ,$ and $\lambda_{k}^{\ast}\left( n\right) $}
Next, we give an easy upper bound on $\lambda_{k}^{\ast}\left( n\right) ;$
later, by much harder work, we shall show that this bound is almost as good as
one can get.
Note that if $G$ is a graph of order $n,$ with $e\left( G\right) $ edges and
adjacency matrix $A,$ the
\[
\lambda_{1}^{2}\left( G\right) +\lambda_{2}^{\ast2}\left( G\right)
+\cdots+\lambda_{n}^{\ast2}\left( G\right) =\mathrm{tr}\text{
A^{2}=2e\left( G\right) .
\]
Hence, using the inequality $\lambda_{1}\left( G\right) \geq2e\left(
G\right) /n$ and the AM-GM inequality, one finds that
\[
\lambda_{2}^{\ast2}\left( G\right) +\cdots+\lambda_{k}^{\ast2}\left(
G\right) \leq2e\left( G\right) -\lambda_{1}^{2}\left( G\right)
\leq2e\left( G\right) -\left( \frac{2e\left( G\right) }{n}\right)
^{2}\leq\frac{n^{2}}{4}.
\]
Therefore, $\left( k-1\right) \lambda_{k}^{\ast2}\left( G\right) \leq
n^{2}/4,$ and Proposition \ref{spro} implies the following bounds:
\begin{theorem}
\label{th1}If $n\geq k\geq2$ and $G$ is a graph of order $n,$ then
\[
\lambda_{k}\left( G\right) \leq\lambda_{k}^{\ast}\left( G\right) \leq
\frac{n}{2\sqrt{k-1}},
\]
and
\[
\left\vert \lambda_{n-k+2}\left( G\right) \right\vert \leq\lambda_{k}^{\ast
}\left( G\right) \leq\frac{n}{2\sqrt{k-1}}.
\]
\end{theorem}
Further, letting $n\rightarrow\infty$, we get the bound
\begin{equation}
c_{k}\leq\frac{1}{2\sqrt{k-1}},\text{ \ \ }c_{-k+1}\leq\frac{1}{2\sqrt{k-1
},\text{ \ \ and \ \ \ }c_{k}^{\ast}\leq\frac{1}{2\sqrt{k-1}}.\label{ubc
\end{equation}
Simple as they are, bounds (\ref{ubc}) give the correct rate of growth of
$c_{k},$ $c_{-k}$ and $c_{k}^{\ast}$ in $k;$ in particular, the bound on
$c_{k}^{\ast}$ is quite tight. Note also that if $k=2,$ then equality holds in
each of the bounds (\ref{ubc}); on the other hand, if $k\geq3,$ the bound on
$c_{k}^{\ast}$ is attained for infinitely many $k,$ but the bounds on $c_{k}$
and $c_{-k+1}$ are never attained.
\begin{theorem}
\label{thub}If $k\geq3,$ then there is an $\varepsilon_{k}>0$ such that
\[
c_{k}<\frac{1}{2\sqrt{k-1}}-\varepsilon_{k}\text{ \ \ and\ \ \ }c_{-k+1
<\frac{1}{2\sqrt{k-1}}-\varepsilon_{k}.
\]
\end{theorem}
Our proof of Theorem \ref{thub} is quite complicated and uses the Removal
Lemma of Alon, Fischer, Krivelevich, and Szegedy \cite{AFKS00}, together with
other tools of analytic graph theory. Due to its length, it will not be given
in the present paper.
Although Theorem \ref{thub} may cast doubts as to the tightness of the bounds
(\ref{ubc}), we shall show that they can be matched by close lower bounds.
\subsection{Lower bounds on $c_{k}$ and $c_{-k}$}
In \cite{Tay71,Tay77}, Taylor came up with a remarkable class of strongly
regular graphs, which we shall use in several ways to give lower bounds on
$\lambda_{k}\left( n\right) $ and $\lambda_{-k}\left( n\right) .$ It seems
that Taylor's strongly regular graphs are a cornerstone in spectral graph
theory, and need to be known better. For reader's sake, in Section \ref{NBS}
gives a brief discussion on Taylor's graphs and their complements.
Note that Taylor's graphs contain roughly half of the total number of edges,
and the same holds for their complements. At the same time, almost all
eigenvalues of a Taylor graph are positive, and almost all eigenvalues of its
complement are negative. This combination of properties makes Taylor's graphs
and their complements very suitable for lower bounds on $c_{k}$ and $c_{-k}.$
To begin with, in the following theorem, we shall use Taylor's graphs to show
the tightness of the bounds (\ref{ubc}) for infinitely many, albeit handpicked
values of $k.$ The proof of Theorem \ref{th2} is in Section \ref{ps}.
\begin{theorem}
\label{th2}If $q$ is an odd prime power and $k$ $=q^{2}-q+1,$ the
\begin{equation}
c_{k}>\frac{1}{2\sqrt{k-1}+1}\text{ \ and \ \ \ }c_{-k+1}>\frac{1}{2\sqrt
{k-1}+1}. \label{lbc
\end{equation}
\end{theorem}
Unfortunately, the bounds (\ref{lbc}) seem to be close to the best ones that
Taylor graphs can provide. Nevertheless, in the following theorem, we shall
use Taylor graphs to provide general asymptotics of $c_{k}$ and $c_{-k}$ for
any $k:$
\begin{theorem}
\label{thglb}There exists $k_{0}$ such that if $k>k_{0}$, the
\begin{equation}
c_{k}>\frac{1}{2\sqrt{k-1}+\sqrt[3]{k}},\text{ \ \ and \ \ }c_{-k+1}>\frac
{1}{2\sqrt{k-1}+\sqrt[3]{k}}. \label{glb
\end{equation}
\textbf{ }
\end{theorem}
Theorem \ref{thglb}, whose proof is in Section \ref{ps}, shows that the upper
bounds (\ref{ubc}) on $c_{k}$ and $c_{-k}$ are asymptotically tight, although
there is a lot to improve. A particularly weak point of bounds (\ref{glb}) is
the fact that $k_{0}$ is not known explicitly, due to a number-theoretic
result used in the proof. Below we provide a weaker theorem, with explicit
bounds. It also shows that the bound (\ref{Powbo}) fails for any $k\geq5$ and
$n$ sufficiently large.
\begin{theorem}
\label{thSB}If $5\leq k\leq15,$ then
\[
c_{k}\geq\frac{1}{k-1/2}\text{ \ \ and \ \ }c_{-k+1}\geq\frac{1}{k-1/2}.
\]
If $k\geq16,$ then
\[
c_{k}\geq\frac{1}{4\sqrt{k-1}}\text{ \ \ and \ \ }c_{-k+1}\geq\frac{1
{4\sqrt{k-1}}.
\]
\end{theorem}
Theorem \ref{thSB}, whose proof is also in Section \ref{ps}, leaves the
following two questions open:$\medskip$
\begin{question}
Is it true that $c_{3}=1/3?$ \medskip
\end{question}
\begin{question}
Is it true that $c_{4}=1/4?$
\end{question}
\subsection{Bounds on $c_{k}^{\ast}$}
It turns out that $\lambda_{k}^{\ast}\left( n\right) $ and $c_{k}^{\ast}$
can be estimated with greater precision than $\lambda_{k}\left( n\right) $
and $c_{k}$. We start by establishing a crucial connection between
$c_{k}^{\ast}$ and the existence of certain symmetric $\left( -1,1\right)
$-matrices. Thus, write $\mathbb{U}_{n}$ for the set of symmetric $\left(
-1,1\right) $-matrices or order $n.$\medskip
\begin{theorem}
\label{thMx}If $A$ is a symmetric $\left( -1,1\right) $-matrix of order $n$,
with $\lambda_{k}^{\ast}\left( A\right) =n/\sqrt{k},$ then
\[
c_{k+1}^{\ast}=\frac{1}{2\sqrt{k}}.
\]
\end{theorem}
\begin{proof}
Let $A=\left[ a_{i,j}\right] \in\mathbb{U}_{n},$ with $\lambda_{k}^{\ast
}\left( A\right) =n/\sqrt{k}.$ Sinc
\[
\sum_{i=1}^{k}\left( \lambda_{i}^{\ast}\left( A\right) \right) ^{2}\geq
k\left( n/\sqrt{k}\right) ^{2}=n^{2}=\sum_{i=1}^{n}\sum_{l=1}^{n}a_{i,j
^{2}=\sum_{i=1}^{n}\left( \lambda_{i}^{\ast}\left( A\right) \right) ^{2},
\]
we see that
\[
\lambda_{1}^{\ast}\left( A\right) =\cdots=\lambda_{k}^{\ast}\left(
A\right) =n/\sqrt{k}\text{ \ \ and \ \ \ }\lambda_{i}^{\ast}\left( A\right)
=0\text{ for }k<i\leq n\text{.
\]
Define a matrix $A^{\prime}\in\mathbb{U}_{2n}$ b
\[
A^{\prime}=\left[
\begin{array}
[c]{cc
A & -A\\
-A & A
\end{array}
\right] ,
\]
and note that all rowsums of $A^{\prime}$ are zero; therefore $0$ is an
eigenvalue of $A^{\prime}$\ with eigenvector\ $\mathbf{j}_{2n}.$ Now, define a
symmetric $\left( 0,1\right) $-matrix $B$ by
\[
B=\frac{1}{2}\left( A^{\prime}\otimes J_{t}+J_{2nt}\right) .
\]
The order of $B$ is $2nt.$ Obviously $\lambda_{1}^{\ast}\left( B\right) =nt$
and
\begin{align*}
\lambda_{2}^{\ast}\left( B\right) & =\cdots=\lambda_{k+1}^{\ast}\left(
B\right) =nt/\sqrt{k},\\
\lambda_{i}^{\ast}\left( B\right) & =0,\text{ \ \ }i=k+1,\ldots,nt.
\end{align*}
Now, zero the diagonal of $B$ and write $A$ for the resulting matrix. Clearly
$A$ is the adjacency matrix of some graph $G$ of order $2nt.$ Using Weyl's
inequalities (see Proposition \ref{proW0}), we find that
\begin{align*}
\lambda_{k+1}^{\ast}\left( G\right) & =\lambda_{k+1}^{\ast}\left(
B-\left( B-A\right) \right) \geq\lambda_{k+1}^{\ast}\left( B\right)
-\lambda_{1}^{\ast}\left( B-A\right) \\
& \geq\lambda_{k+1}^{\ast}\left( B\right) -\lambda_{1}^{\ast}\left(
I_{2nt}\right) \\
& =\frac{2nt}{2\sqrt{k}}-1.
\end{align*}
Letting $t\rightarrow\infty,$ we ge
\[
c_{k+1}^{\ast}=\frac{1}{2\sqrt{k}},
\]
completing the proof.
\end{proof}
Theorem \ref{thMx} motivates the introduction of a class of symmetric $\left(
-1,1\right) $-matrices, which extend symmetric Hadamard matrices in a natural way.
\subsection{An extension of symmetric Hadamard matrices}
Write $n\left( A\right) $ for the order of a square matrix $A,$ and let
$\mathbb{S}_{k}$ be the set of symmetric $\left( -1,1\right) $-matrices with
$\lambda_{k}^{\ast}\left( A\right) =n\left( A\right) /\sqrt{k}.$ Note that
$\mathbb{S}_{k}$ can contain matrices of different order; for that matter, if
$\mathbb{S}_{k}$ is nonempty, then it is infinite. Note also that if $A\in$
$\mathbb{S}_{k},$ and a matrix $B$ can be obtained by permutations or
negations performed simultaneously on rows and columns of $A,$ then $B\in$
$\mathbb{S}_{k}$ as well; the reason is that singular values are not affected
by such operations.
As in the proof of Theorem \ref{thMx}, one can check the validity of the
following statement.
\begin{proposition}
\label{proma}A symmetric $\left( -1,1\right) $-matrix $A$ belongs to
$\mathbb{S}_{k}$ if and only if
\[
\lambda_{1}^{\ast}\left( A\right) =\cdots=\lambda_{k}^{\ast}\left(
A\right) =n\left( A\right) /\sqrt{k},\text{ \ \ and \ \ }\lambda_{i}^{\ast
}\left( A\right) =0\text{\ for }k<i\leq n\left( A\right) \text{.
\]
\end{proposition}
Here is a summary of some properties of $\mathbb{S}_{k}$:\medskip
\qquad(1) If $A\in\mathbb{S}_{k}$ then $-A\in\mathbb{S}_{k};$
\qquad(2) If $A$ is a symmetric $\left( -1,1\right) $-matrix of rank 1, then
$A\in\mathbb{S}_{1};$ thus, $J_{n}\in\mathbb{S}_{1};$
\qquad(3) If $H$ is a symmetric Hadamard matrix of order $k,$ then
$H\in\mathbb{S}_{k};$
\qquad(4) If $A\in\mathbb{S}_{k}$ and $B\in\mathbb{S}_{l},$ then $A\otimes
B\in\mathbb{S}_{kl};$ hence, if $\mathbb{S}_{k}\neq\varnothing,$ then
$\mathbb{S}_{2k}\neq\varnothing;$
\qquad(5) If $A\in\mathbb{S}_{k}$, then $A\otimes J_{n}$ $\in\mathbb{S}_{k}$
for any $n\geq1;$ hence $\mathbb{S}_{k}$ is infinite;
\qquad(6) If $\mathbb{S}_{k}\neq\varnothing$, then $\mathbb{S}_{k}$ contains
matrices with all their rowsums equal to zero.\medskip
We omit the proofs of (1)-(5), but here is sketch of a proof of (6):\emph{
}the rank of the matri
\[
K=\left[
\begin{array}
[c]{cc
1 & -1\\
-1 & 1
\end{array}
\right]
\]
is $1$, so $K\in S_{1}.$ Now, if $A\in S_{k},$ then $K\otimes A\in S_{k},$ and
the rowsums of $K\otimes A$ are zero.\emph{\medskip}
Properties (1)-(6) allow to show that $\mathbb{S}_{k}$ contains matrices with
some special properties, as in the following proposition:
\begin{proposition}
\label{props}If $A\in\mathbb{S}_{k},$ then there is a $B\in\mathbb{S}_{2k}$
such that:
(i) $B$ has exactly $k$ positive and exactly $k$ negative eigenvalues;
(ii) the rowsums of $B$ are equal to $0.$
\end{proposition}
To check this proposition, set
\[
H_{2}=\left[
\begin{array}
[c]{cc
1 & 1\\
1 & -1
\end{array}
\right] ,
\]
and let $B=K\otimes\left( H_{2}\otimes A\right) ;$ obviously $B\in$
$\mathbb{S}_{2k},$ and $B$ satisfies \emph{(i) }and\emph{ (ii)}.\bigskip
The principal question about $\mathbb{S}_{k}$ is the following one:
\begin{problem}
\label{probS}For which $k$ is $\mathbb{S}_{k}$ nonempty?
\end{problem}
Below we shall show that $\mathbb{S}_{k}$ is empty if $k$ is odd and is not a
square. On the positive side, Property (3) implies that there are infinitely
many $k$ for which $\mathbb{S}_{k}$ is nonempty. In particular, Paley's
construction of symmetric Hadamard matrices (see, e.g., \cite{CrKh07}) implies
the following fact:
\begin{proposition}
If $p$ is a prime power and $p=1$ $\operatorname{mod}$ $4,$ then
$\mathbb{S}_{2\left( p+1\right) }$ is not empty.
\end{proposition}
We shall prove more definite assertions about $\mathbb{S}_{k}$ by using the
fact that the singular values of a real symmetric matrix are the absolute
values of its eigenvalues.
\begin{proposition}
\label{proq}If $A\in\mathbb{S}_{k},$ then either $k$ is an exact square, or
$A$ has the same number of positive and negative eigenvalues.
\end{proposition}
\begin{proof}
Let $A=\left[ a_{i,j}\right] \in\mathbb{S}_{k}.$ Setting $l=\lambda
_{1}^{\ast}\left( A\right) $ and $n=n\left( A\right) ,$ we have
\[
kl^{2}=\sum_{\lambda_{i}\left( A\right) >0}\lambda_{i}^{2}\left( A\right)
+\sum_{\lambda_{i}\left( A\right) <0}\lambda_{i}^{2}\left( A\right)
=\mathrm{tr}\text{ }A^{2}=\sum_{i=1}^{n}\sum_{l=1}^{n}a_{i,j}^{2}=n^{2}.
\]
On the other hand, writing $n_{+}$ and $n_{-}$ for the number of positive and
negative eigenvalues of $A,$ we see tha
\[
\left( n_{+}-n_{-}\right) l=\sum_{\lambda_{i}\left( A\right) >0
\lambda_{i}\left( A\right) +\sum_{\lambda_{i}\left( A\right) <0
\lambda_{i}\left( A\right) =\mathrm{tr}\text{ }A.
\]
Since $\mathrm{tr}$ $A$ is an integer, either $l$ is rational or $n_{+}=n_{-
$. Since $l=n/\sqrt{s}$, it may be rational only if $\sqrt{s}$ is an integer,
completing the proof.
\end{proof}
\begin{corollary}
If $k$ is odd and $\mathbb{S}_{k}$ is nonempty, then $k$ is an exact square.
\end{corollary}
It is interesting to see what we can say about $\mathbb{S}_{k}$ for small $k,$
say for $k\leq10.$ First, the above corollary implies that $\mathbb{S}_{k}$ is
empty for $k=3,5,$ and $7.$ Below we shall show that the converse of
Proposition \ref{proq} is partially true as well, that is to say: \emph{if
}$k$\emph{ is an exact square, then }$S_{k}$\emph{ is nonempty.} Thus, in view
of properties (1)-(5), we see that $\mathbb{S}_{k}$ is nonempty for
$k=1,2,4,8,$ and $9.$ The first unknown cases are $k=6$ and $k=10.$
\begin{question}
Is $\mathbb{S}_{6}$ empty?
\end{question}
\medskip
\subsection{\label{cons}Constructions of matrices in \ $\mathbb{S}_{s^{2}}$}
In this subsection we shall construct classes of symmetric $\left(
-1,1\right) $-matrices, which show that $\mathbb{S}_{k}$ is nonempty if $k$
is an exact square or is the double of an exact square. Except for this
primary goal, our matrices must have other specific properties, which are
necessary for subsequent applications; thus the statements are somewhat involved.
The proofs of Theorems \ref{thKHN}, \ref{thj}, and \ref{thj1} are variations
of Kharaghani's method \cite{Kha85} for constructing Hadamard matrices. Our
approach suggests that this method is generic and can be used to construct
$\left( -1,1\right) $-matrices that are more general than Hadamard matrices.
The proofs of Theorems \ref{thKHN}, \ref{thj}, and \ref{thj1} are in Section
\ref{ps}.
\begin{theorem}
\label{thKHN}For any integer $s\geq2$, there are an integer $n\geq s$ and a
symmetric $\left( -1,1\right) $-matrix $B$ of order $ns$ such that:
(i) $B$ has exactly $s^{2}$ nonzero eigenvalues, of which $\binom{s-1}{2}$ are
equal to $n,$ and $\binom{s+1}{2}$ are equal to $-n$;
(ii) the rowsums of $B$ are equal to $0;$
(iii) the diagonal entries of $B$ are equal to $-1.$
\end{theorem}
Note that clause \emph{(i)} of Theorem \ref{thKHN} implies that $B\in
\mathbb{S}_{s^{2}}.$ Further, negating the matrix $B,$ we get the following variation.
\begin{corollary}
For any integer $s\geq2$, there are an integer $n\geq s$ and a symmetric
$\left( -1,1\right) $-matrix $B$ of order $ns$ such that:
(i) $B$ has exactly $s^{2}$ nonzero eigenvalues, of which $\binom{s+1}{2}$ are
equal to $n,$ and $\binom{s-1}{2}$ are equal to $-n$;
(ii) the rowsums of $B$ are equal to $0;$
(iii) the diagonal entries of $B$ are equal to $1.$
\end{corollary}
Theorem \ref{thKHN} will be used to give some answers to Problem \ref{mpro}.
For other purposes we shall need two other theorems.
\begin{theorem}
\label{thj}For any integer $s\geq2$, there are an integer $n\geq s$ and a
symmetric $\left( -1,1\right) $-matrix $B$ of order $ns$ such that:
(i) $B$ has exactly $s^{2}$ nonzero eigenvalues, of which $\binom{s+1}{2}$ are
equal to $n,$ and $\binom{s-1}{2}$ are equal to $-n$;
(ii) the vector $\mathbf{j}_{ns}$ is an eigenvector of $B$ to the eigenvalue
$-n$;
(iii) the diagonal entries of $B$ are equal to $1.$
\end{theorem}
\begin{theorem}
\label{thj1}For any integer $s\geq2$, there are an integer $n\geq s$ and a
symmetric $\left( -1,1\right) $-matrix $B$ of order $ns$ such that:
(i) $B$ has exactly $s^{2}$ nonzero eigenvalues, of which $\binom{s+1}{2}-1$
are equal to $n,$ and $\binom{s-1}{2}+1$ are equal to $-n$;
(ii) all rowsums of $B$ are equal to $-n$.
\end{theorem}
Note that Theorem \ref{thKHN} and Property (2) imply that for any natural
number $s$, the classes $\mathbb{S}_{s^{2}}$ and $\mathbb{S}_{2s^{2}}$ are
nonempty. Before continuing, let us state some explicit solutions to Problem
\ref{probS}:
\begin{proposition}
\label{propE}Let $s$ be a natural number and $p$ be a prime power, with $p=1$
$\operatorname{mod}$ $4.$ Then the classes $\mathbb{S}_{s^{2}},$
$\mathbb{S}_{2s^{2}},$ $\mathbb{S}_{2s^{2}\left( p+1\right) }$ and
$\mathbb{S}_{4s^{2}\left( p+1\right) }$ are nonempty.
\end{proposition}
\subsection{Finding $\lambda_{k}^{\ast}\left( n\right) $ and $c_{k
^{\ast\text{ }}$for infinitely many $k$ and $n$}
Observe that using the classes $\mathbb{S}_{k},$ Theorem \ref{thMx} can be
restated as:\emph{ if }$\mathbb{S}_{k}$\emph{ is nonempty, then
$c_{k+1}^{\ast}=1/(2\sqrt{k}).$ Proposition \ref{propE} and properties (1)-(6)
give many values of $k$ for which this equality holds. Moreover, using the
finer properties outlined in Theorem \ref{thKHN}, we can determine the exact
value of $\lambda_{s^{2}+1}^{\ast}\left( n\right) $ for infinitely many $n.$
\begin{theorem}
\label{thp}If $s\geq1,$ then there is an integer $n>s,$ such that for every
integer $t\geq1,$ there is a graph $G$ of order $snt$ with\emph{
\[
\lambda_{s^{2}+1}^{\ast}\left( G\right) =\frac{nt}{2}.
\]
\end{theorem}
\begin{proof}
Let $n\geq s,$ and let $B$ be the matrix of order $sn$ constructed in Theorem
\ref{thKHN}. Set
\[
C=\frac{1}{2}\left( B+J_{sn}\right) ,
\]
and note that $C$ is a symmetric $\left( 0,1\right) $-matrix with zero
diagonal. Since all rowsums of $B$ are zero, the vector $\mathbf{j}_{sn}$ is
an eigenvector of $B$ to the eigenvalue $0;$ thus all eigenvectors to nonzero
eigenvalues of $B$ are orthogonal to $\mathbf{j}_{sn}.$ Obviously every
eigenvector of $B$ is an eigenvector to $C.$ Thus, $sn/2$ is an eigenvalue of
$C;$ also, $C$ has $\binom{s-1}{2}$ eigenvalues equal to $n/2$ and
$\binom{s+1}{2}$ eigenvalues equal to $-n/2;$ the remaining eigenvalues of $C$
are zero. Now, letting $A=C\otimes J_{t},$ one sees that $A$ is the adjacency
matrix of a graph $G$ of order $snt$ with
\[
\lambda_{s^{2}+1}^{\ast}\left( G\right) =\frac{1}{2}\lambda_{s^{2}}^{\ast
}\left( B\right) t=\frac{snt}{2s},
\]
completing the proof of Theorem \ref{thp}.
\end{proof}
Theorem \ref{thKHN} helps also to find concise asymptotics of $c_{k}^{\ast}.$
Indeed, since $c_{k}^{\ast}$ is nonincreasing in $k,$ letting $s$ to be the
smallest integer such that $s^{2}+1\geq k,$ we see that $c_{k}^{\ast
\geq1/\left( 2s\right) ;$ since $\left( s-1\right) ^{2}+1<k,$ we get the
following theorem:
\begin{theorem}
\label{thga}For any $k\geq3,
\[
\frac{1}{2\sqrt{k-1}}\geq c_{k}^{\ast}>\frac{1}{2\sqrt{k-1}+2}=\frac{1
{2\sqrt{k-1}}+O\left( k^{-1}\right) .
\]
\end{theorem}
It seems quite clear that the lower bound on $c_{k}^{\ast}$ can be improved,
so we raise the following problem.
\begin{problem}
Is there a positive constant $C$ such that for any $k\geq3,$
\[
c_{k}^{\ast}>\frac{1}{2\sqrt{k+C}}?
\]
\end{problem}
\section{\label{NGs}Spectral Nordhaus-Gaddum problems}
It turns out that the classes $\mathbb{S}_{k}$ help to find infinitely many
solutions to a general spectral Nordhaus-Gaddum problem. Nordhaus-Gaddum
problems in general, form a notable part of extremal graph theory, see, e.g.,
the recent survey \cite{AoHa13} for their numerous variations. In particular,
spectral Nordhaus-Gaddum problems have been studied first in 1970, by Nosal
\cite{Nos70}, and have attracted a lot of attention since then.
Thus, let $\overline{G}$ denote the complement of a graph $G.$ Given $n\geq
k\geq1,$ define the functions $f_{k}\left( n\right) ,$ $f_{-k}\left(
n\right) ,$ and $f_{k}^{\ast}\left( n\right) $ as
\begin{align*}
f_{k}\left( n\right) & =\max_{v\left( G\right) =n}\lambda_{k}\left(
G\right) +\lambda_{k}(\overline{G}),\\
f_{-k}\left( n\right) & =\max_{v\left( G\right) =n}\left\vert
\lambda_{n-k+1}\left( G\right) \right\vert +|\lambda_{n-k+1}(\overline
{G})|,\\
f_{k}^{\ast}\left( n\right) & =\max_{v\left( G\right) =n}\lambda
_{k}^{\ast}\left( G\right) +\lambda_{k}^{\ast}(\overline{G}).
\end{align*}
Clearly, we define $f_{k}\left( n\right) ,$ $f_{-k}\left( n\right) ,$ and
$f_{k}^{\ast}\left( n\right) $ similarly to $\lambda_{k}\left( n\right) ,$
$\lambda_{-k}\left( n\right) ,$ and $\lambda_{k}^{\ast}\left( n\right) .$
Note that $f_{k}^{\ast}\left( n\right) $ is a new function, but
$f_{k}\left( n\right) $ and $f_{-k}\left( n\right) $ have been introduced
in \cite{Nik07} with different, albeit essentially equivalent definitions.
Now, let us reiterate and extend a problem raised in \cite{Nik07}:
\begin{problem}
\label{NGpro}For any $k\geq1,$ find $f_{k}\left( n\right) ,$\ $f_{-k}\left(
n\right) ,$ and $f_{k}^{\ast}\left( n\right) .$
\end{problem}
The function $f_{1}\left( n\right) $ has been studied by Nosal \cite{Nos70}:
finding it have turned out to be a hard problem, which has been resolved only
recently, in \cite{Csi09} and \cite{Ter11}. This case sticks out from the
rest, both with its particular extremal graphs, as with its particular
methods. The first "mainstream" case is $f_{2}\left( n\right) ,$ which has
been determined in \cite{Nik07}.
Recently, in \cite{NiYu14}, two tight upper bounds have been given:
\emph{If }$k\geq2$\emph{ and }$n\geq15\left( k-1\right) ,$\emph{ then
\begin{equation}
f_{k}\left( n\right) \,\leq\frac{n}{\sqrt{2\left( k-1\right)
}-1.\label{bo2
\end{equation}
\emph{Likewise, if }$k\geq1$ \emph{and} $n\geq4^{k},$\emph{ then
\begin{equation}
f_{-k}\left( n\right) \,\leq\frac{n}{\sqrt{2k}}+1.\label{bo3
\end{equation}
In turns out that the bounds (\ref{bo2}) and (\ref{bo3}) capture the rate of
growth of $f_{k}\left( n\right) $\ and $f_{-k}\left( n\right) $ pretty
tightly. In \cite{NiYu14}, it has been shown that (\ref{bo2}) and (\ref{bo3})
are essentially best possible if $k=2^{s-1}+1$ and $s=2,3,\ldots.$ What is
more, for such $k$ it has proved that
\[
f_{k}\left( n\right) \geq\frac{n}{\sqrt{2\left( k-1\right) }}-2\text{
\ \ and \ \ \ }f_{-k}\left( n\right) \geq\frac{n}{\sqrt{2k}}.
\]
for infinitely many $n.\medskip$
The study of $f_{k}\left( n\right) ,$ $f_{-k}\left( n\right) ,$ and
$f_{k}^{\ast}\left( n\right) $ can be put on the same ground as $\lambda
_{k}\left( n\right) ,$ $\lambda_{-k}\left( n\right) ,$ and $\lambda
_{k}^{\ast}\left( n\right) .$ First, using the methods of \cite{Nik06}, one
can show that the limits
\[
f_{k}=\lim_{n\rightarrow\infty}\frac{f_{k}\left( n\right) }{n
,\text{\ \ \ \ }f_{-k}=\lim_{n\rightarrow\infty}\frac{f_{-k}\left( n\right)
}{n},\text{ \ \ and \ \ \ }f_{k}^{\ast}=\lim_{n\rightarrow\infty}\frac
{f_{k}^{\ast}\left( n\right) }{n
\]
exist, and the following inequalities hold for every $n:
\[
f_{k}\left( n\right) \leq f_{k}n-1,\text{ \ \ }f_{-k}\left( n\right) \leq
f_{-k}n+1,\text{\ \ \ \ and \ \ \ \ }f_{k}^{\ast}\left( n\right) \leq
f_{k}^{\ast}n+1.
\]
Now, the constructions developed in Section \ref{cons} allow to resolve
Problem \ref{NGpro} for infinitely many values of $k.$ Indeed, it turns out
that if \ $\mathbb{S}_{k}$ is nonempty, then
\[
f_{k+1}=\frac{1}{\sqrt{2k}}\text{ \ \ and \ \ \ }f_{-k}=\frac{1}{\sqrt{2k}}.
\]
These equalities follow from the theorem below, where we prove a more precise
result, involving $n$ as well.
\begin{theorem}
\label{thNG}If $\mathbb{S}_{k}$ is nonempty, then there is an integer $n>k,$
such that for every integer $t\geq1,$ there is a graph $G$ of order $nt$ with
\begin{equation}
f_{k+1}\left( n\right) \geq\frac{nt}{\sqrt{2k}}-2 \label{in1
\end{equation}
an
\begin{equation}
f_{-k}\left( n\right) \geq\frac{nt}{\sqrt{2k}}. \label{in2
\end{equation}
\end{theorem}
\begin{proof}
If $\mathbb{S}_{k}$ is nonempty, Proposition \ref{props} implies that there
exists $B\in\mathbb{S}_{2k},$ say of order $n,$ such that,
\begin{align*}
\lambda_{1}\left( B\right) & =\cdots=\lambda_{k}\left( B\right)
=\frac{n}{\sqrt{2k}},\\
\lambda_{n-k+1}\left( B\right) & =\cdots=\lambda_{n}\left( B\right)
=-\frac{n}{\sqrt{2k}},\\
\lambda_{i}^{\ast}\left( B\right) & =0\text{ for }k<i\leq n-k\text{.
\end{align*}
Define a matrix $A^{\prime}$ b
\[
A^{\prime}=\frac{1}{2}\left( B\otimes J_{t}\right) +J_{nt}.
\]
Now, zero the diagonal of $A^{\prime}$ and write $A$ for the resulting matrix.
Note that $A$ is a symmetric $\left( 0,1\right) $ matrix with zero diagonal,
so it is the adjacency matrix of graph $G$ of order $nt.$ As in the proof of
Theorem \ref{thMx}, we see that
\[
\lambda_{k+1}\left( G\right) \geq\frac{nt}{2\sqrt{2k}}-1.
\]
On the other hand, taking the matrix
\[
\overline{A}^{\prime}=\frac{1}{2}\left( -B\otimes J_{t}\right) +J_{nt},
\]
and zeroing its main diagonal, we obtain the matrix $\overline{A},$ which is
obviously the adjacency matrix of the complement of $G.$ Like above we have,
\[
\lambda_{k+1}(\overline{G})\geq\frac{nt}{2\sqrt{2k}}-1,
\]
and inequality (\ref{in1}) follows. The proof of (\ref{in2}) is similar and is omitted.
\end{proof}
In view of Theorem \ref{thNG} and properties (1)-(6), we get numerous examples
for which bounds (\ref{bo2}) and (\ref{bo3}) are essentially best; we refer to
Proposition \ref{propE} for some explicit values.
We finish the discussion of $f_{k}\left( n\right) $ and $f_{-k}\left(
n\right) $ with general asymptotics of $f_{k}$ and $f_{-k}.$ Since $f_{k}$
and $f_{-k}\left( n\right) $ are nonincreasing in $k$, letting $s$ to be the
smallest integer such that $s^{2}\geq k,$ we see that
\[
f_{k+1}\geq\frac{1}{\sqrt{2}s}\text{ \ \ and \ \ }f_{-k}>\frac{1}{\sqrt{2}s};
\]
since $\left( s-1\right) ^{2}<k,$ we get the following theorem:
\begin{theorem}
For any $k\geq2,
\[
\frac{1}{\sqrt{2\left( k-1\right) }}\geq f_{k}>\frac{1}{\sqrt{2\left(
k-1\right) }+\sqrt{2}}=\frac{1}{\sqrt{2\left( k-1\right) }}+O\left(
k^{-1}\right) ,
\]
an
\[
\frac{1}{\sqrt{2\left( k-1\right) }}\geq f_{-k+1}>\frac{1}{\sqrt{2\left(
k-1\right) }+\sqrt{2}}=\frac{1}{\sqrt{2\left( k-1\right) }}+O\left(
k^{-1}\right) .
\]
\end{theorem}
Finally, let us briefly discuss the function $f_{k}^{\ast}\left( n\right) ,$
which is somewhat easier to deal with, and can be derived mainly from
$\lambda_{k}^{\ast}\left( n\right) $. First, Theorem \ref{th1} implies
immediately that
\begin{equation}
f_{k}^{\ast}\left( n\right) \leq2\lambda_{k}^{\ast}\left( n\right)
\leq\frac{n}{\sqrt{k-1}}.\label{bo4
\end{equation}
This easy bound is rather different from (\ref{bo2}) and (\ref{bo3}), whose
proofs are much subtler anyway. Nonetheless, bound (\ref{bo4}) gives the
correct rate of growth of $f_{k}^{\ast}\left( n\right) .$ Using Theorems
\ref{thMx} and \ref{thga}, we immediately come up with the following statements:
\begin{theorem}
If $\mathbb{S}_{k}$ is nonempty, then there is an integer $n>k,$ such that for
every integer $t\geq1,$ there is a graph $G$ of order $nt$ with
\[
f_{k+1}^{\ast}\left( n\right) \geq\frac{nt}{\sqrt{k}}-2
\]
\end{theorem}
\begin{theorem}
For any $k\geq2,
\[
\frac{1}{\sqrt{k-1}}\geq f_{k}^{\ast}>\frac{1}{\sqrt{k-1}+1}=\frac{1
{\sqrt{k-1}}+O\left( k^{-1}\right) .
\]
\end{theorem}
\section{\label{KFs}Sums of eigenvalues and sums of singular values}
In addition to individual eigenvalues and singular values of graphs, it is of
interest to consider certain sums thereof. In particular, let
\begin{align*}
\tau_{k}\left( n\right) & =\max_{v\left( G\right) =n}\lambda_{1}\left(
G\right) +\cdots+\lambda_{k}\left( G\right) ,\\
\xi_{k}\left( n\right) & =\max_{v\left( G\right) =n}\lambda_{1}^{\ast
}\left( G\right) +\cdots+\lambda_{k}^{\ast}\left( G\right) .
\end{align*}
Note that $\xi_{k}\left( n\right) $ is the maximal \emph{Ky Fan
$k$\emph{-norm} of a graph of order $n$. In particular, the Ky Fan $n$-norm is
known as the \emph{trace norm} of $G,$ and has been extensively studied under
the name \emph{graph energy,} a concept introduced by Gutman in \cite{Gut78};
see also \cite{GLS12} for the current state of this research. Note that,
$\xi_{n}\left( n\right) $ is just the maximum energy of a graph of order
$n,$ which also has been studied, see, e.g., \cite{Hae08}, \cite{KoMo01}, and
\cite{Nik07j}.
The research on graph energy is truly monumental, but with the flexibility of
the parameter $k,$ the Ky Fan $k$-norms offer a considerably vaster
playground. Here we shall focus only on the following principal question,
raised in \cite{Ni11}:
\begin{problem}
\label{MNpro}For any $k\geq1,$ find $\tau_{k}\left( n\right) $\ and $\xi
_{k}\left( n\right) .$
\end{problem}
For a start, note that the inequality $\lambda_{i}\left( G\right)
\leq\lambda_{i}^{\ast}\left( G\right) $ implies that $\tau_{k}\left(
n\right) \leq\xi_{k}\left( n\right) $ for any $k$ and $n.$ For general
$k>2,$ the function $\tau_{k}\left( n\right) $ has been studied by Mohar in
\cite{Moh09}; and in turn, $\xi_{k}\left( n\right) $ has been studied by the
author in \cite{Ni11}. The exact values of $\tau_{k}\left( n\right) $ and
$\xi_{k}\left( n\right) $ are unknown for most values of $k;$ in particular,
neither $\tau_{2}\left( n\right) $ nor $\xi_{2}\left( n\right) $ are known
yet: see \cite{EMNA08} for $\tau_{2}\left( n\right) ,$ and \cite{GHK01} for
$\xi_{2}\left( n\right) .$ However, estimating $\tau_{k}\left( n\right) $
and $\xi_{k}\left( n\right) $ is possible for large $k$. Indeed, Mohar
\cite{Moh09} proved the asymptotic
\begin{equation}
\frac{1}{2}\left( \frac{1}{2}+\sqrt{k}-o\left( k^{-2/5}\right) \right)
<\frac{\tau_{k}\left( n\right) }{n}\leq\frac{1}{2}\left( 1+\sqrt{k}\right)
.\label{Mohbo
\end{equation}
Note the gap $1/2+o\left( 1\right) $ between the upper and lower bounds in
(\ref{Mohbo}), which is very challenging to close. In general, finding
$\tau_{k}\left( n\right) $ seems a hard problem, a lot harder than finding
$\xi_{k}\left( n\right) $. In particular, it is easy to show that the limit
$\tau_{k}=\lim\limits_{n\rightarrow\infty}\tau_{k}\left( n\right) /n$ exists
for any fixed $k\geq1,$ but this limit is not known for any $k\geq2.$ Thus, we
suggest the following concrete conjecture:
\begin{conjecture}
For any $k\geq2,$ there is an $\varepsilon_{k}>0$ such that
\[
\tau_{k}<\frac{1}{2}\left( 1+\sqrt{k}-\varepsilon_{k}\right) .
\]
\end{conjecture}
In contrast to $\tau_{k}\left( n\right) $, we shall find $\xi_{k}\left(
n\right) $ for infinitely many values of $k$ and $n.$ To begin with, in
\cite{Ni11} it was shown that if $n\geq k\geq1,$ then
\begin{equation}
\xi_{k}\left( n\right) \leq\frac{1}{2}\left( 1+\sqrt{k}\right)
n,\label{Nikbo
\end{equation}
which strengthens the upper bound (\ref{Mohbo}). In fact, unlike $\tau
_{k}\left( n\right) ,$ the function $\xi_{k}\left( n\right) $ attains the
upper bound (\ref{Nikbo}) for infinitely many $k$ and $n.$ Indeed, let $H$ be
a symmetric regular Hadamard matrix of order $k,$ with positive rowsums, and
with $-1$ along the main diagonal. Then the matrix
\[
A=\frac{1}{2}\left( H\otimes J_{n}\right) +J_{kn
\]
is the adjacency matrix of a graph $G$ of order $kn,$ with
\[
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{k}^{\ast}\left(
G\right) =\frac{1}{2}\left( 1+\sqrt{k}\right) n,
\]
and so $\xi_{k}\left( n\right) $ attains the upper bound (\ref{Nikbo}). It
is known that symmetric regular Hadamard matrix with equal rowsums and with
$-1$ along the main diagonal exist for $k=4m^{4}$ and any $m=1,2,\ldots$, see
\cite{HaXi09} for details. In fact, there are many more cases of $k$ for which
the upper bound (\ref{Nikbo}) is attained.\medskip
First, we shall show that if the bound (\ref{Nikbo}) is attained, then $k$ is
an exact square:
\begin{theorem}
\label{thk}If $G$ is a graph of order $n$ such that
\[
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{k}^{\ast}\left(
G\right) =\frac{1}{2}\left( 1+\sqrt{k}\right) n,
\]
then $k$ is an exact square.
\end{theorem}
\begin{proof}
Write $\left\Vert A\right\Vert _{\ast k}$ the sum of the $k$ largest singular
values of $A$. Suppose that $G$ is a graph that satisfies the hypothesis and
write $A$ for its adjacency matrix.
Note that $J_{n}-2A$ is a symmetric $\left( -1,1\right) $-matrix and so, in
view of the AM-QM inequality, we find tha
\[
\sum_{i=1}^{k}\lambda_{i}^{\ast}\left( J_{n}-2A\right) \leq\sqrt{k\sum
_{i=1}^{k}\lambda_{i}^{\ast}{}^{2}\left( J_{n}-2A\right) }\leq\sqrt
{k\sum_{i=1}^{n}\lambda_{i}^{\ast}{}^{2}\left( J_{n}-2A\right) }=\sqrt{k}n.
\]
Therefore, using the the triangle inequality for the Ky Fan $k$-norm
$\left\Vert X+Y\right\Vert _{\ast k}\leq$ $\left\Vert X\right\Vert _{\ast
k}+\left\Vert Y\right\Vert _{\ast k}$ (see \cite{HoJo94}, p.196), we find
that
\[
\left( 1+\sqrt{k}\right) n=2\left\Vert A\right\Vert _{\ast k}=\left\Vert
2A\right\Vert _{\ast k}\leq\left\Vert 2A-J_{n}\right\Vert _{\ast k}+\left\Vert
J_{n}\right\Vert _{\ast k}\leq\sqrt{k}n+n.
\]
Thus, equalities hold throughout the above line, and so, $2A-J_{n}$ has $k$
nonzero singular values, which are equal. We get
\[
\lambda_{k}^{\ast}\left( J_{n}-2A\right) =n/\sqrt{k},
\]
implying that $J_{n}-2A\in\mathbb{S}_{k}.$ On the other hand, $\mathrm{tr
\left( J_{n}-2A\right) =n\neq0,$ so $J_{n}-2A$ cannot have the same number
of positive and negative eigenvalues, and Proposition \ref{proq} implies that
$k$ is an exact square.
\end{proof}
The matrix built in Theorem \ref{thj} helps to prove that the converse of the
above theorem is partially true as well.
\begin{theorem}
\label{thck}Let $s$ be an even positive integer. There exists a positive
integer $n,$ such that for every positive integer $t,$ there is a graph $G$ of
order $snt,$ with
\[
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{s^{2}}^{\ast}\left(
G\right) =\frac{1}{2}\left( 1+s\right) snt.
\]
\end{theorem}
Theorem \ref{thck} is proved in Section \ref{ps}. It is as good as one can
get, but we can prove it only if $s$ is even. If $s$ is odd, we can do just
slightly worse, showing that $\xi_{k}\left( n\right) $ is just below the
upper bound. To this effect, we shall prove a more general theorem, and deduce
this fact as a corollary. The proof of the theorem is in Section \ref{ps}.
\begin{theorem}
\label{thck1}Suppose that $\mathbb{S}_{k}$ contains a regular matrix $B$ with
nonzero rowsums, say of order $n$. Then for any positive integer $t,$ there is
a graph $G$ of order $nt$ with
\begin{equation}
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{k}^{\ast}\left(
G\right) \geq\frac{1}{2}\left( 1+\sqrt{k}\right) nt-k. \label{in3
\end{equation}
\end{theorem}
Dividing both sides of (\ref{in3}) by $nt$ and letting $t\rightarrow\infty,$
we obtain the following corollary.
\begin{corollary}
If \ $\mathbb{S}_{k}$ contains a regular matrix $B$ with nonzero rowsums,
then
\[
\lim\limits_{n\rightarrow\infty}\frac{\xi_{k}\left( n\right) }{n
=\frac{1+\sqrt{k}}{2}.
\]
\end{corollary}
Let us note that the premise that $\mathbb{S}_{k}$ contains a regular matrix,
with nonzero rowsums is not difficult to satisfy. Indeed, Theorem \ref{thj1}
implies that for any integer $s\geq2,$ the set $\mathbb{S}_{s^{2}}$ contains a
regular matrix with nonzero rowsums.\medskip
Theorem \ref{thk} does not shed any light on the case when $k$ is not an exact
square, so we suggest the following concrete conjecture.
\begin{conjecture}
There exist infinitely many integers $k$ such that
\[
\lim\limits_{n\rightarrow\infty}\frac{\xi_{k}\left( n\right) }{n
<\frac{1+\sqrt{k}}{2}.
\]
\end{conjecture}
We end up this section with the easy asymptotic
\[
\frac{\sqrt{k}}{2}\leq\lim\limits_{n\rightarrow\infty}\frac{\xi_{k}\left(
n\right) }{n}\leq\frac{1+\sqrt{k}}{2},
\]
whose proof is omitted.
\section{\label{NBS}Notation, background, and support}
For graph notation and concepts undefined here, the reader is referred to
\cite{Bol98}. For general reference on graph spectra, see \cite{CRS10}; for
reference on Hadamard matrices and symmetric Latin squares, see \cite{CrKh07}
and \cite{IoSh06}; for reference on strongly regular graphs and their
eigenvalues, see \cite{GoRo01}.\medskip
We write $I_{n}$ and $J_{n}$ for the identity and the all ones matrix of order
$n.$ The $n$-dimensional vector of all ones is denoted by $\mathbf{j}_{n}.$ As
usual, the Kronecker product of two matrices $A$ and $B$ is denoted by
$A\otimes B.$ We recall that if $A$ and $B$ are square, then the spectrum of
$A\otimes B$ consists are all products of eigenvalues of $A$ and eigenvalues
of $B,$ with multiplicities counted. Also, the Kronecker product of symmetric
matrices is symmetric.\medskip
In this paper \emph{regular matrix} means a matrix whose rowsums are
equal.\medskip
Next, we shall give necessary details on Weyl's inequalities, graphs blowups,
Taylor strongly regular graphs, and symmetric Latin squares.\medskip
\subsection{\label{WS}Weyl's inequalities}
If $A$ is a Hermitian matrix of order $n,$ write $\lambda_{1}\left( A\right)
,\ldots,\lambda_{n}\left( A\right) $ for its eigenvalues ordered as
$\lambda_{1}\left( A\right) \geq\cdots\geq\lambda_{n}\left( A\right) .$
Weyl proved the following useful inequalities for the eigenvalues of sums of
Hermitian matrices, (see, e.g. \cite{HoJo88}, p. 181):
Let $A$ and $B$ be Hermitian matrices of order $n,$ and let $1\leq i\leq n$
and $1\leq j\leq n.$ Then
\[
\lambda_{i}(A)+\lambda_{j}(B)\leq\lambda_{i+j-n}(A+B),\text{ if }i+j\geq n+1.
\]
\medskip
The following two immediate corollaries are used throughout the paper.
\begin{proposition}
\label{proW0}Suppose that $A^{\prime}$ is a symmetric $\left( 0,1\right)
$-matrix of order $n.$ If $A$ is the matrix obtained by zeroing the main
diagonal of $A^{\prime}$ and $1\leq k\leq n,$ the
\[
\lambda_{k}\left( A\right) \geq\lambda_{k}(A^{\prime})-1.
\]
\end{proposition}
Indeed, $X=A^{\prime}-A$ is a $\left( 0,1\right) $-diagonal matrix, and so
$\lambda_{1}\left( A\right) \leq1$. Therefore,
\[
\lambda_{k}\left( A\right) +1\geq\lambda_{k}\left( A\right) +\lambda
_{1}\left( A\right) \geq\lambda_{k}(A^{\prime}).
\]
\begin{proposition}
\label{proW}If $G$ is a graph of order $n$ and $2\leq k\leq n,$ then
\[
\lambda_{k}\left( G\right) +\lambda_{n-k+2}(\overline{G})\leq-1.
\]
\end{proposition}
Indeed, if $A$ and $\overline{A}$ are the adjacency matrices of $G$ and
$\overline{G},$ then $A+\overline{A}$ is the adjacency matrix of the complete
graph $K_{n}.$ Hence $\lambda_{k}\left( G\right) +\lambda_{n-k+2
(\overline{G})\leq\lambda_{k}\left( K_{n}\right) =-1.$\medskip
\subsection{\textbf{Blowups of graphs and their eigenvalues}}
Given a graph $G$ and an integer $t\geq1,$ replace each vertex of $G$ by an
independent set on $t$ vertices and each edge of $G$ by a complete bipartite
graph $K_{t,t}.$ Write $G^{\left( t\right) }$ for the resulting graph and
call it a \emph{blowup} of $G$.
If $G$\ is a graph of order $n,$ then $G^{\left( t\right) }$ is a graph of
order $nt$ and its adjacency matrix $A\left( G^{\left( t\right) }\right) $
is given by the equation
\[
A\left( G^{\left( t\right) }\right) =A\left( G\right) \otimes J_{t}.
\]
This algebraic representation of $A\left( G^{\left( t\right) }\right) $
gives a key to its spectrum:
\begin{proposition}
\label{prob}If $t\geq1$ and $G$ is a graph of order $n,$ with eigenvalues
$\lambda_{1}\left( G\right) ,\ldots,\lambda_{n}\left( G\right) ,$ then the
eigenvalues of $G^{\left( t\right) }$ are $\lambda_{1}\left( G\right)
t,\ldots,\lambda_{n}\left( G\right) t,$ together with $\left( t-1\right)
n$ additional zeros.
\end{proposition}
Most often we shall use the following variation of the blow-up operation:
given a graph $G$ and an integer $t\geq1,$ replace each vertex of $G$ by a
complete graph on $t$ vertices and each edge of $G$ by a complete bipartite
graph $K_{t,t}.$ Write $G^{\left[ t\right] }$ for the resulting graph and
call it a \emph{closed blowup} of $G$.
If $G$\ is a graph of order $n,$ then $G^{\left[ t\right] }$ is a graph of
order $nt$ and its adjacency matrix $A\left( G^{\left[ t\right] }\right) $
is given by the equation
\[
A\left( G^{\left[ t\right] }\right) =\left( A\left( G\right)
+I_{n}\right) \otimes J_{t}-I_{nt}.
\]
This algebraic representation of $A\left( G^{\left[ t\right] }\right) $
can be used to find the spectrum of $G^{\left[ t\right] }$:
\begin{proposition}
\label{probu}If $t\geq1$ and $G$ is a graph of order $n,$ with eigenvalues
$\lambda_{1}\left( G\right) ,\ldots,\lambda_{n}\left( G\right) ,$ then the
eigenvalues of $G^{\left[ t\right] }$ are $\lambda_{1}\left( G\right)
t+t-1,\ldots,\lambda_{n}\left( G\right) t+t-1,$ together with $\left(
t-1\right) n$ additional $-1$'s.
\end{proposition}
\medskip
\subsection{\label{TS}Taylor's strongly regular graphs and their complements}
In \cite{Tay71,Tay77} Taylor came up with a remarkable family of strongly
regular graphs $T\left( q\right) ,$ defined for every odd prime power $q,$
and with parameter
\[
v=q^{3},\text{ \ }k=\frac{1}{2}\left( q-1\right) \left( q^{2}+1\right)
,\text{ \ }a=\frac{1}{4}\left( q-1\right) ^{3}-1,\text{ \ \ }c=\frac{1
{4}\left( q-1\right) \left( q^{2}+1\right) ,
\]
Following the general rules, one finds that the eigenvalues of $T\left(
q\right) $ ar
\begin{align*}
\lambda_{1}\left( T\left( q\right) \right) & =\text{ \ }\frac{1
{2}\left( q-1\right) \left( q^{2}+1\right) \text{ with multiplicity 1};\\
\lambda_{2}\left( T\left( q\right) \right) & =\text{ \ }\frac{1
{2}\left( q-1\right) ,\text{\ with multiplicity }\left( q-1\right) \left(
q^{2}+1\right) ;\\
\lambda_{n}\left( T\left( q\right) \right) & =-\frac{1}{2}\left(
q^{2}+1\right) ,\text{ with multiplicity \ }q\left( q-1\right) .
\end{align*}
The complement $\overline{T\left( q\right) }$ is a strongly regular graph
with parameter
\[
v=q^{3},\text{ \ }k=\frac{1}{2}\left( q+1\right) \left( q^{2}-1\right)
,\text{ \ }a=\frac{1}{4}\left( q+3\right) \left( q^{2}-3\right) ,\text{
\ \ }c=\frac{1}{4}\left( q+1\right) \left( q^{2}-1\right) .
\]
For the eigenvalues of $\overline{T\left( q\right) }$ one finds that
\begin{align}
\lambda_{1}(\overline{T\left( q\right) }) & =\text{ \ }\frac{1}{2}\left(
q+1\right) \left( q^{2}-1\right) \text{ with multiplicity 1};\nonumber\\
\lambda_{2}(\overline{T\left( q\right) }) & =\text{ \ }\frac{1}{2}\left(
q^{2}-1\right) ,\text{ with multiplicity \ }q\left( q-1\right)
;\label{ref}\\
\lambda_{n}(\overline{T\left( q\right) }) & =-\frac{1}{2}\left(
q+1\right) ,\text{\ with multiplicity }\left( q-1\right) \left(
q^{2}+1\right) .\nonumber
\end{align}
\subsubsection{Some analytic properties of Taylor graphs}
Below we focus on certain properties of the Taylor graphs that may be of
interest to researchers in spectral and extremal graph theory, as well as in
quasi-random (\cite{CGW89}) and pseudo-random (\cite{KrSu06}, \cite{Tho1},
\cite{Tho2}) graphs. To simplify the view on the graph $T\left( q\right) $
for sufficiently large $q,$ we let $q^{3}=n,$ and disregard low order terms
when needed. Then $T\left( q\right) $ is a $\left( n/2\right) $-regular
graph $G$ of order $n,$ with the following properties:
\begin{enumerate}
\item Every two distinct vertices of $G$ have $\thickapprox n/4$ common
neighbors, and the same holds for $\overline{G}$. Therefore, both $G$ and
$\overline{G}$ are quasi-random (pseudo-random) graphs of density $1/2$;
\item For the spectrum of $G$ one finds that
\begin{align*}
\lambda_{1}\left( G\right) & \thickapprox\text{ \ }n/2\text{ with
multiplicity }1;\\
\lambda_{2}\left( G\right) & \thickapprox\text{ \ }n^{1/3}/2,\text{\ with
multiplicity}\thickapprox n;\\
\lambda_{n}\left( G\right) & \thickapprox-n^{2/3}/2,\text{\ with
multiplicity\ }n^{2/3}.
\end{align*}
Therefore, almost all eigenvalues of $G$ are positive.
\item Nonetheless, the sum of squares of the non-principal positive
eigenvalues of $G$ is a vanishing proportion of the sum of squares of all
eigenvalues:
\[
\sum_{\lambda_{i}\left( G\right) >0,i>1}\lambda_{i}^{2}\left( G\right)
\thickapprox\frac{1}{2}n^{5/3}=o\left( 1\right) \sum_{i=1}^{n}\lambda
_{i}^{2}\left( G\right) =o\left( 1\right) e\left( G\right) .
\]
\item For the spectrum of $\overline{G}$ one finds that
\begin{align*}
\lambda_{1}(\overline{G}) & \thickapprox\text{ \ }n/2\text{ with multiplicity
}1;\\
\lambda_{2}(\overline{G}) & \thickapprox\text{ \ }n^{2/3}/2,\text{ \ with
multiplicity}\thickapprox n^{2/3};\\
\lambda_{n}(\overline{G}) & \thickapprox-n^{1/3}/2,\text{\ with
multiplicity}\thickapprox n.
\end{align*}
Therefore, almost all eigenvalues of $\overline{G}$ are negative.
\item Nonetheless, the sum of squares of the negative eigenvalues of
$\overline{G}$ is a vanishing proportion of the sum of squares of all
eigenvalues
\[
\sum_{\lambda_{i}(\overline{G})<0}\lambda_{i}^{2}(\overline{G})\thickapprox
\frac{1}{2}n^{5/3}=o\left( 1\right) \sum_{i=1}^{n}\lambda_{i}^{2
(\overline{G})=o\left( 1\right) e(\overline{G}).
\]
\end{enumerate}
\medskip
\subsection{\label{LS}Some symmetric Latin squares}
In the proofs of Theorems \ref{thKHN}, \ref{thj}, and \ref{thj1} we shall use
two types of symmetric Latin squares: back-circulant Latin square and
symmetric Latin square with constant diagonal. These constructions are simple
and well-known, but for reader's sake we shall describe them below.
Let $s$ be a positive integer. The back-circulant Latin square of size $s$
with symbol set $\left\{ 1,\ldots,s\right\} $ is an $s\times s$ square
matrix $L=\left[ l_{i,j}\right] ,$ with $l_{i,j}$ given by
\[
l_{i,j}=(\left( i+j\right) \text{ }\operatorname{mod}\text{ }s)+1,\text{
}1\leq i,j\leq s.
\]
Obviously $L$ is a symmetric Latin square and its entries belong to $\left\{
1,\ldots,s\right\} .$
Note that if $L$ is a symmetric Latin square of odd order with symbol set $S$,
then every symbol $s\in S$ occurs above the main diagonal as many times as
below it; hence, $s$ also occurs on the main diagonal, as the total number of
occurrences of $s$ is odd. Therefore, the main diagonal of $L$ contains each
symbol exactly once.
Next we want to construct symmetric Latin squares with constant diagonals. By
the above observation, the order of such Latin square cannot be odd, and for
any even $s,$ we shall give a construction, which seems well-known: we borrow
it from \cite{IoSh06}. Thus, let $s$ be an even positive integer, and define
an $s\times s$ square matrix $L=\left[ l_{i,j}\right] ,$ with entries given
by
\[
l_{i,j}=\left\{
\begin{array}
[c]{ll
s, & \text{if \ }1\leq i\leq s\text{ \ and }i=j;\\
(\left( i+j\right) \operatorname{mod}\text{ }\left( s-1\right) )+1, &
\text{if \ }1\leq i<s,\text{ }1\leq j<s,\text{ and }i\neq j;\\
(\text{ }2j\text{ }\operatorname{mod}\text{ }\left( s-1\right) )+1, &
\text{if \ }i=s\text{ and\ \ }1\leq j<s;\\
(\text{ }2i\text{ }\operatorname{mod}\text{ }\left( s-1\right) )+1, &
\text{if \ }1\leq i<s\text{ \ and }j=s.
\end{array}
\right.
\]
The matrix $L$ is a symmetric Latin square with symbol set $\left\{
1,\ldots,s\right\} $ and the symbol $s$ along the main diagonal. For example,
for $s=2,4,$ and $6,$ this construction give
\[
L=\left[
\begin{array}
[c]{cc
2 & 1\\
1 & 2
\end{array}
\right] ,\text{ \ }L=\left[
\begin{array}
[c]{cccc
4 & 1 & 2 & 3\\
1 & 4 & 3 & 2\\
2 & 3 & 4 & 1\\
3 & 2 & 1 & 4
\end{array}
\right] \text{ \ }L=\left[
\begin{array}
[c]{cccccc
6 & 4 & 5 & 1 & 2 & 3\\
4 & 6 & 1 & 2 & 3 & 5\\
5 & 1 & 6 & 3 & 5 & 2\\
1 & 2 & 3 & 6 & 2 & 4\\
2 & 3 & 5 & 2 & 6 & 1\\
3 & 5 & 2 & 4 & 1 & 6
\end{array}
\right] .
\]
\medskip
\section{\label{ps}Proofs of some theorems}
\bigskip
\subsection{Proofs of Theorems \ref{th2}, \ref{thglb}, and \ref{thSB}}
\medskip
\begin{proof}
[\textbf{Proof of Theorem \ref{th2}}]We shall prove only the first bound, as
the other one follows by Proposition \ref{spro}. Let $\overline{T\left(
q\right) }$ be the complement of the Taylor strongly regular graph $T\left(
q\right) $ of order $q^{3}$ (see \ref{TS} for details). Let $G$ be a closed
blowup of $\overline{T\left( q\right) },$ i.e., $G=\overline{T\left(
q\right) }^{\left[ t\right] },$ and let $n=tq^{3}=v\left( G\right) .$
Then
\[
c_{k}\geq\sup\frac{\lambda_{k}\left( G\right) }{n}=\sup\frac{\lambda
_{q\left( q-1\right) +1}(\overline{T\left( q\right) })t+t-1}{q^{3}t
=\sup\frac{q^{2}+1}{2q^{3}}-\frac{1}{q^{3}t}=\frac{q^{2}+1}{2q^{3}}.
\]
To finish the proof we need to show that
\[
\frac{q^{2}+1}{2q^{3}}\geq\frac{1}{2\sqrt{q^{2}-q}+1}=\frac{1}{2\sqrt{k-1
+1}.
\]
This inequality follows fro
\[
\frac{q^{2}+1}{2q^{3}}>\frac{2q-1}{4q^{2}+2q-1}>\frac{1}{2\sqrt{q^{2}-q}+1
\]
after some simple algebra, which we omit.
\end{proof}
\bigskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thglb}}]Fix a sufficiently large integer $k,$
and let $q$ be the smallest prime such tha
\[
q\left( q-1\right) +1\geq k.
\]
A result of Baker, Harman, and Pintz \cite{BHP01} on the distribution of
primes implies that if $k$ is sufficiently large, the
\[
q\leq\sqrt{k}+1/2+\left( \sqrt{k}+1/2\right) ^{21/40}.
\]
It is not hard to see that if $k$ is sufficiently large, then $q<\sqrt
{k-1}+\sqrt[3]{k}/2.$ Let $\overline{T\left( q\right) }$ be the complement
of the Taylor graph $T\left( q\right) ,$ let $G$ be a closed blowup of
$\overline{T\left( q\right) }$, i.e., $G=$ $\overline{T\left( q\right)
}^{\left[ t\right] },$ and set $n=tq^{3}=v\left( G\right) .$ We see that
\[
\lambda_{k}\left( G\right) =\lambda_{2}(\overline{T\left( q\right)
})t+t-1=\frac{1}{2}\left( q^{2}-1\right) t+t-1>\frac{1}{2}q^{2}t.
\]
Hence,
\[
c_{k}\geq\frac{\lambda_{k}\left( G\right) }{n}>\frac{q^{2}t}{2q^{3}t
=\frac{1}{2q}\geq\frac{1}{2\sqrt{k-1}+\sqrt[3]{k}}.
\]
Now, the bound on $c_{-k+1}$ follows in view of Proposition \ref{spro}.
\end{proof}
\bigskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thSB}}]In view of Proposition \ref{spro}, if
$k\geq2$, we always have $c_{-k+1}\geq c_{k},$ so our main goal is to prove
the bounds on $c_{k}.$ This proof naturally splits into two cases: $5\leq
k\leq15$ and $k\geq16.$
If $5\leq k\leq15,$ we just take an appropriate strongly regular graph $H$
with parameters $(v,k,a,c),$ and let $G$ be a closed blowup of $H\ $of order
$n,$ i.e., $G=H^{\left[ n/v\right] },$ where $n$ is a multiple of $v.$ In
view of Proposition \ref{probu},
\[
\lambda_{k}\left( G\right) =\left( \lambda_{k}\left( H\right) +1\right)
\frac{n}{v}-1.
\]
Using well known sources, say the home page of A. Brouwer, we obtain the
following table
\
\begin{tabular}
[t]{ccc
\ $H(v,k,a,c)$ & \ Eigenvalue of $H$ \ & Eigenvalue of $G=H^{\left[
n/v\right] }\medskip$\\
(9, 4, 1, 2) & $\lambda_{5}\left( H\right) =1$ & $\lambda_{5}\left(
G\right) =\frac{2}{9}n-1\medskip$\\
(10, 3, 0, 1) & $\lambda_{6}\left( H\right) =1$ & $\lambda_{6}\left(
G\right) =\frac{1}{5}n-1\medskip$\\
(13, 6, 2, 3) & \ \ \ \ \ \ $\lambda_{7}\left( H\right) =\frac{\sqrt{13
-1}{2}$ & \ \ \ \ \ $\lambda_{7}\left( G\right) =\frac{\sqrt{13}+1
{26}n-1\medskip$\\
(15, 6, 1, 3) & $\lambda_{8}\left( H\right) =1$ & $\lambda_{8}\left(
G\right) =\frac{2}{15}n-1\medskip$\\
(15, 6, 1, 3) & $\lambda_{9}\left( H\right) =1$ & $\lambda_{9}\left(
G\right) =\frac{2}{15}n-1\medskip$\\
(15, 6, 1, 3) & $\lambda_{10}\left( H\right) =1$ & $\lambda_{10}\left(
G\right) =\frac{2}{15}n-1\medskip$\\
(21,10,3,6) & $\lambda_{11}\left( H\right) =1$ & $\lambda_{11}\left(
G\right) =\frac{2}{21}n-1\medskip$\\
(21,10,3,6) & $\lambda_{12}\left( H\right) =1$ & $\lambda_{12}\left(
G\right) =\frac{2}{21}n-1\medskip$\\
(21,10,3,6) & $\lambda_{13}\left( H\right) =1$ & $\lambda_{13}\left(
G\right) =\frac{2}{21}n-1\medskip$\\
(21,10,3,6) & $\lambda_{14}\left( H\right) =1$ & $\lambda_{14}\left(
G\right) =\frac{2}{21}n-1\medskip$\\
(21,10,3,6) & $\lambda_{15}\left( H\right) =1$ & $\lambda_{15}\left(
G\right) =\frac{2}{21}n-1\medskip
\end{tabular}
\
\]
Now, letting $n\rightarrow\infty,$ we obtain
\[
c_{5}\geq2/9,\text{ \ }c_{6}\geq1/5,\text{ \ }c_{7}\geq\sqrt{13}/2+1/2,\text{
\ }c_{8}\geq2/15,\text{ \ }c_{9}\geq2/15,\text{ \ }c_{10}\geq2/15.
\]
Likewise, if $11\leq k\leq15,$ we obtain $c_{k}\geq2/21.$ These inequalities
obviously imply that
\[
c_{k}\geq\frac{1}{k-1/2
\]
whenever $5\leq k\leq15.$
Now, let $k\geq16,$ and let $q$ be the smallest prime $q$ such that
\begin{equation}
q\geq1/2+\sqrt{k-3/4}.\label{ink
\end{equation}
Bertrand's postulate guarantees that for any real $x>3,$ there is a prime $q$
such tha
\[
\left\lceil x\right\rceil <q\leq2\left\lceil x\right\rceil -3.
\]
Since $2\left\lceil x\right\rceil -3<2x-1,$ in our case this implies that
\begin{equation}
q<2\left( 1/2+\sqrt{k-1}\right) -1=\sqrt{4k-3}<2\sqrt{k-1}.\label{ink1
\end{equation}
Let $\overline{T\left( q\right) }$ be the complement of the Taylor graph
$T\left( q\right) ,$ and let $G=$ $\overline{T\left( q\right) }^{\left[
t\right] }.$ Since inequality (\ref{ink}) implies that $k\leq q\left(
q-1\right) +1,$ in view of (\ref{ref}), we see tha
\[
\lambda_{k}(\overline{T\left( q\right) })=\frac{1}{2}\left( q^{2}-1\right)
,
\]
and therefore
\[
\lambda_{k}\left( G\right) =\lambda_{k}(\overline{T\left( q\right)
})t+t-1=\frac{1}{2}\left( q^{2}-1\right) t+t-1>\frac{1}{2}q^{2}t.
\]
Now, inequality (\ref{ink1}) implies that
\[
c_{k}^{\ast}\geq\frac{\lambda_{k}\left( G\right) }{q^{3}t}>\frac{1
{2q}>\frac{1}{4\sqrt{k-1}},
\]
completing the proof of Theorem \ref{thSB}.
\end{proof}
\bigskip
\subsection{Proofs of Theorems \ref{thKHN}, \ref{thj}, and \ref{thj1}}
The proofs of Theorems \ref{thKHN}, \ref{thj}, and \ref{thj1} are very close,
but for reader's sake we give them separately. All three proofs exploit the
construction of Hadamard matrices due to Kharaghani \cite{Kha85}, see also
\cite{IoSh06}, Theorem 4.4.16. We shall vary both the blocks and the
underlying Latin square, so the reader is referred to \ref{LS} for necessary
details about Latin squares. The idea of using symmetric Latin squares with
constant diagonal is borrowed from Haemers \cite{Hae08}, Theorem 2, and Ionin
and Shrikhande \cite{IoSh06}, Corollary 5.3.17.\medskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thKHN}}]Suppose that $L=\left[ l_{i,j}\right]
$ is a back-circulant Latin square of size $s,$ with symbol set $\left\{
1,\ldots,s\right\} .$ Let $\mathbf{x}_{1},\ldots,\mathbf{x}_{s}$ be
orthogonal $\left( -1,1\right) $-vectors of dimension $n\geq s$ that are
also orthogonal to the all ones vector $\mathbf{j}_{n}.$ An easy choice is to
take the last $k$ rows of a normalized Hadamard matrix of order $n>k.$ For
each $s=1,\ldots,s,$ define a square matrix $A_{s}$ by $A_{s}=-\mathbf{x
_{s}\otimes\mathbf{x}_{s}.$ Obviously $A_{1},\ldots,A_{s}$ are symmetric
$\left( -1,1\right) $-matrices of size $n$ and rank $1,$ with diagonal
entries equal to $-1.$
Now, let $B$ be the block matrix obtained by replacing each entry $l_{i,j}$ of
$L$ by the matrix $A_{l_{i,j}}.$ Note that $B$ is a symmetric $\left(
-1,1\right) $-matrix of size $sn.$ Obviously the diagonal entries of $B$ are
equal to $-1,$ thus \emph{(iii) }holds. Note also that
\[
A_{p}A_{q}=\left\{
\begin{array}
[c]{ll
0 & \text{if }q\neq p;\\
-nA_{p} & \text{if }q=p.
\end{array}
\right.
\]
Hence, $B^{2}$ is block diagonal with each diagonal block equal to
$-nA_{1}-\cdots-nA_{s}.$ But $-nA_{1}-\cdots-nA_{s}$ is of rank $s,$ so it has
exactly $s$ nonzero eigenvalues, each equal to $n^{2}.$ Thus, $B$ has $s^{2}$
nonzero eigenvalues, and their absolute value is equal to $n.$
Writing $n_{+}$ and $n_{-}$ for the number of positive and negative
eigenvalues of $B,$ we hav
\[
\left( n_{+}-n_{-}\right) n=\mathrm{tr}\text{ }B=-sn,
\]
and so, $n_{+}-n_{-}=-s,$ implying that
\[
n_{+}=\binom{s-1}{2}\text{ and }n_{-}=\binom{s+1}{2},
\]
completing the proof of \emph{(i)}.
To prove \emph{(ii)} note that each rowsum of each matrix $A_{i}$ is zero, so
the rowsums of $B$ are zero as well. This completes the proof of Theorem
\ref{thKHN}.
\end{proof}
\bigskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thj}}]Suppose that $L=\left[ l_{i,j}\right] $
is a symmetric Latin square of order $s$ with constant diagonal. Let $\left\{
1,\ldots,s\right\} $ be the symbol set of $L$ and $s$ be the diagonal symbol.
Next, select $s$ vectors $\mathbf{x}_{1},\ldots,\mathbf{x}_{s}$ of dimension
$n\geq s$ such that $\mathbf{x}_{1}=\mathbf{j}_{n}$ and every two of the
vectors $\mathbf{x}_{1},\ldots,\mathbf{x}_{s}$ are orthogonal$.$ An easy
choice is to take the first $k$ rows of a normalized Hadamard matrix of order
$n\geq k.$ Let $A_{1}=-J_{n},$ and for each $i=2,\ldots,s,$ define a square
matrix $A_{i}$ by
\[
A_{i}=\mathbf{x}_{i}\otimes\mathbf{x}_{i}.
\]
Obviously $A_{1},\ldots,A_{s}$ are symmetric $\left( -1,1\right) $-matrices
of size $n$ and rank $1.$ Note also that the diagonal entries of $A_{s}$ are
equal to $1.$
Now, let $B$ be the block matrix obtained by replacing each entry $l_{i,j}$ of
$L$ by the matrix $A_{l_{i,j}}.$ Note that $B$ is a symmetric $\left(
-1,1\right) $-matrix of size $sn.$ The diagonal entries of $B$ are equal to
$1,$ thus \emph{(iii) }holds. Note also tha
\[
A_{p}A_{q}=\left\{
\begin{array}
[c]{ll
0 & \text{if }q\neq p;\\
nJ_{n} & \text{if }q=p=1;\\
nA_{p} & \text{if }q=p\neq1.
\end{array}
\right.
\]
Hence, $B^{2}$ is block diagonal with each diagonal block equal to
$nJ_{n}+\cdots+nA_{s}.$ Since $nJ_{n}+\cdots+nA_{s}$ has exactly $s$ nonzero
eigenvalues, each equal to $n^{2},$ we see that $B$ has $s^{2}$ nonzero
eigenvalues, and their absolute value is equal to $n.$
Writing $n_{+}$ and $n_{-}$ for the number of positive and negative
eigenvalues of $B,$ we hav
\[
\left( n_{+}-n_{-}\right) n=\mathrm{tr}\text{ }B=sn,
\]
and so, $n_{+}-n_{-}=s,$ implying that
\[
n_{+}=\binom{s+1}{2}\text{ and }n_{-}=\binom{s-1}{2},
\]
completing the proof of \emph{(i)}.
To prove \emph{(ii)} note that\emph{
\[
A_{p}\mathbf{j}_{n}=\left\{
\begin{array}
[c]{cc
0 & \text{if }p\neq1;\\
-n\mathbf{j}_{n} & \text{if }p=1.
\end{array}
\right.
\]
Therefore,
\[
B\mathbf{j}_{sn}=\left(
\begin{array}
[c]{c
A_{1}\mathbf{j}_{n}\\
A_{1}\mathbf{j}_{n}\\
\vdots\\
A_{1}\mathbf{j}_{n
\end{array}
\right) =-n\mathbf{j}_{sn}.
\]
Thus, $\mathbf{j}_{kn}$ is an eigenvector of $B$ to the eigenvalue $-n,$
completing the proof of Theorem \ref{thj}.
\end{proof}
\bigskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thj1}}]Our proof combines the proofs of
Theorems \ref{thKHN} and \ref{thj}. Suppose that $L=\left[ l_{i,j}\right] $
is a back-circulant Latin square of size $s,$ with symbol set $\left\{
1,\ldots,s\right\} .$ Next, select $s$ vectors $\mathbf{x}_{1},\ldots
,\mathbf{x}_{s}$ of dimension $n\geq s$ such that $\mathbf{x}_{1
=\mathbf{j}_{n}$ and every two of the vectors $\mathbf{x}_{1},\ldots
,\mathbf{x}_{s}$ are orthogonal$.$ Let $A_{1}=-J_{n},$ and for each
$i=2,\ldots,s,$ define a square matrix $A_{i}$ by $A_{i}=\mathbf{x}_{i
\otimes\mathbf{x}_{i}.$ Let $B$ be the block matrix obtained by replacing each
entry $l_{i,j}$ of $L$ by the matrix $A_{l_{i,j}}.$ Note that $B$ is a
symmetric $\left( -1,1\right) $-matrix of size $sn,$ with $\left(
s-1\right) n$ diagonal entries equal to $1$ and $n$ diagonal entries equal to
$-1.$ Note also tha
\[
A_{p}A_{q}=\left\{
\begin{array}
[c]{ll
0 & \text{if }q\neq p;\\
nJ_{n} & \text{if }q=p=1;\\
nA_{p} & \text{if }q=p\neq1.
\end{array}
\right.
\]
Hence, $B$ has $s^{2}$ nonzero eigenvalues, and their absolute value is equal
to $n.$ Writing $n_{+}$ and $n_{-}$ for the number of positive and negative
eigenvalues of $B,$ we hav
\[
\left( n_{+}-n_{-}\right) n=\mathrm{tr}\text{ }B=\left( s-2\right) n,
\]
and so, $n_{+}-n_{-}=s-2,$ implying that
\[
n_{+}=\binom{s+1}{2}-1\text{ and }n_{-}=\binom{s-1}{2}+1,
\]
completing the proof of \emph{(i)}.
To prove \emph{(ii)} let us note that if $2\leq i\leq s$, then all rowsums of
$A_{p}$ are zero. So the rowsums of $B$ are equal to the rowsums of $-J_{n},$
which are equal to $-n.$
\end{proof}
\medskip
\subsection{Proofs of Theorems \ref{thck} and \ref{thck1} \emph{ }}
\medskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thck}}]Let $B$ be a matrix constructed by
Theorem \ref{thj} and let $C=B\otimes J_{t}.$ The properties of $B$ given by
Theorem \ref{thj} imply that $C$ is a symmetric $\left( -1,1\right) $-matrix
of order $snt$ such that:
- $C$ has exactly $s^{2}$ nonzero eigenvalues, of which $\binom{s-1}{2}$ are
equal to $nt$ and $\binom{s+1}{2}$ are equal to $-nt$;
- the diagonal entries of $C$ are equal to $-1$;
- the vector $\mathbf{j}_{snt}$ is an eigenvector of $C$ to the eigenvalue
$-nt.\medskip$
Now, let
\[
A=\frac{1}{2}\left( J_{snt}-C\right) .
\]
Clearly $A$ is a symmetric $\left( 0,1\right) $-matrix of order $snt$, with
zero diagonal; hence, $A$ is the adjacency matrix of some graph $G$ of order
$snt$. Let $\mathbf{x}_{1}=\mathbf{j}_{snt},\mathbf{x}_{2},\ldots
,\mathbf{x}_{snt}$ be orthogonal eigenvectors to $C.$ Note that
\[
\frac{1}{2}\left( J_{snt}-C\right) \mathbf{j}_{snt}=\left( \frac{snt
{2}+\frac{nt}{2}\right) \mathbf{j}_{snt},
\]
so $snt/2+nt/2$ is an eigenvalue to $G.$ Also, for any $i=2,\ldots,snt,$ we
see that
\[
A\mathbf{x}_{i}=\frac{1}{2}\left( J_{snt}-C\right) \mathbf{x}_{i}=-\frac
{1}{2}C\mathbf{x}_{i},
\]
so $G$ has $\binom{s+1}{2}$ eigenvalues equal to $-nt/2$ and $\binom{s-1
{2}-1$ eigenvalues equal to $nt/2.$ Therefore
\[
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{s^{2}}^{\ast}\left(
G\right) =\frac{snt}{2}+\frac{nt}{2}+\left( s^{2}-1\right) \frac{nt
{2}=\frac{1}{2}\left( 1+s\right) \frac{snt}{2},
\]
completing the proof of Theorem \ref{thck}.
\end{proof}
\medskip
\begin{proof}
[\textbf{Proof of Theorem \ref{thck1}}]Let $\lambda$ be the rowsum of $B,$
which clearly is a nonzero eigenvalue of $B$ with eigenvector $\mathbf{j}_{n
$. Since $B\in\mathbb{S}_{k},$ either $\lambda=n/\sqrt{k}$ or $\lambda
=-n/\sqrt{k}.$ We shall assume that $\lambda=-n/\sqrt{k},$ for otherwise we
just take $-B$ for $B.$ Now, for any positive integer $t,$ define a symmetric
$\left( 0,1\right) $-matrix $A^{\prime}$ by
\[
A^{\prime}=\frac{1}{2}\left( J_{nt}-B\otimes J_{t}\right) ,
\]
and note tha
\[
\lambda_{1}\left( A^{\prime}\right) =\frac{nt}{2}+\frac{nt}{2\sqrt{k
},\text{ \ \ and \ \ }\lambda_{i}^{\ast}\left( A^{\prime}\right) =\frac
{nt}{2\sqrt{k}}\text{ for }1<i\leq k.
\]
Next, zero the diagonal of $A^{\prime}$ and write $A$ for the resulting
matrix. Clearly $A$ is a symmetric $\left( 0,1\right) $-matrix with zero
diagonal, so $A$ is the adjacency matrix of some graph $G\ $of order $nt.$
Using Weyl's inequalities (Proposition \ref{proW0}), we see that
\[
\lambda_{1}\left( G\right) \geq\frac{nt}{2}+\frac{nt}{2\sqrt{k}}-1,\text{
\ \ and \ \ }\lambda_{i}^{\ast}\left( G\right) \geq\frac{nt}{2\sqrt{k
}-1\text{ for }1<i\leq k.
\]
Therefore,
\[
\lambda_{1}^{\ast}\left( G\right) +\cdots+\lambda_{k}^{\ast}\left(
G\right) \geq\left( \frac{nt}{2}+k\frac{nt}{2\sqrt{k}}\right) -k=\frac
{1}{2}\left( 1+\sqrt{k}\right) nt-k,
\]
completing the proof of Theorem \ref{thck1}.
\end{proof}
\medskip
\section{\label{RM}A recap for symmetric $\left( -1,1\right) $-matrices}
Many solutions in this paper come from $\left( -1,1\right) $-matrices. This
is not incidental, for if $G$ is a regular graph, then its adjacency spectrum
is linearly equivalent to the spectrum of its Seidel's matrix, which is a
$\left( 0,-1,1\right) $-matrix. But there is more to that: if $G$ is a
$\left( n/2\right) $-regular graph, the Seidel matrix effectively eliminates
the largest eigenvalue of $G$, which may be nuisance in certain spectral
problems, like most of the problems discussed in this paper. One cannot but
agree that many of the questions raised above for graphs seem more balanced
and natural if translated for $\left( -1,1\right) $-matrices. In this
section we explore such translations.
Thus, for any $k\geq1,$ let us introduce the function
\[
\Lambda_{k}\left( n\right) =\max_{A\in\mathbb{U}_{n}}\lambda_{k}\left(
A\right) \text{ \ \ and \ \ \ \ }\Lambda_{k}^{\ast}\left( n\right)
=\max_{A\in\mathbb{U}_{n}}\lambda_{k}^{\ast}\left( A\right) .
\]
Obviously $\Lambda_{k}\left( n\right) $ and $\Lambda_{k}^{\ast}\left(
n\right) $ are the matrix analogs of $\lambda_{k}\left( n\right) $ and
$\lambda_{k}^{\ast}\left( n\right) ;$ we do not need an analog to
$\lambda_{-k}\left( n\right) ,$ as $\mathbb{U}_{n}$ is closed under
negation. Next, in the general spirit of the paper, we raise the problem:
\begin{problem}
For any $k\geq2,$ find $\Lambda_{k}\left( n\right) $ and $\Lambda_{k}^{\ast
}\left( n\right) $.
\end{problem}
Much of what we have achieved for graphs applies to symmetric $\left(
-1,1\right) $-matrices as well. First, obviousl
\begin{equation}
\Lambda_{k}\left( n\right) \leq\Lambda_{k}^{\ast}\left( n\right) \leq
n/\sqrt{k}.\label{umc
\end{equation}
Note that for $\Lambda_{k}^{\ast}\left( n\right) ,$ bound (\ref{umc}) is
precise for infinitely many $k$ and $n.$ Indeed, if \ $\mathbb{S}_{k
\neq\varnothing,$ for arbitrary large $n,$\ we have $\Lambda_{k}^{\ast}\left(
n\right) =n/\sqrt{k}$.
Further, in analogy to $c_{k}$ and $c_{k}^{\ast},$ let
\[
d_{k}=\sup_{n\geq1}\frac{\Lambda_{k}\left( n\right) }{n}\text{ \ \ \ and
\ \ \ }d_{k}^{\ast}=\sup_{n\geq1}\frac{\Lambda_{k}^{\ast}\left( n\right)
}{n}.
\]
The constants $d_{k}$ and $d_{k}^{\ast}$ are handy, as for any $n$ and any
matrix $A\in\mathbb{U}_{n},$ we have
\[
\lambda_{k}\left( A\right) \leq d_{k}n\text{ \ \ and \ \ }\lambda_{k}^{\ast
}\left( A\right) \leq d_{k}^{\ast}n.
\]
In turns out that these inequalities are best possible, for one can show tha
\[
\lim_{n\rightarrow\infty}\frac{\Lambda_{k}\left( n\right) }{n}=d_{k}\text{
\ \ \ \ and \ \ \ \ }\lim_{n\rightarrow\infty}\frac{\Lambda_{k}^{\ast}\left(
n\right) }{n}=d_{k}^{\ast}.
\]
Thus, much about $\Lambda_{k}\left( n\right) $ and $\Lambda_{k}^{\ast
}\left( n\right) $ would be known if we knew $d_{k}$ and $d_{k}^{\ast}$ or
estimates thereof$.$ First, from (\ref{umc}) we immediately get an upper boun
\[
d_{k}\leq d_{k}^{\ast}\leq1/\sqrt{k},
\]
so the difficulty is to find matching lower bounds.
As one may expect, $d_{k}^{\ast}$ is easier to tackle than $d_{k}.$ Indeed,
since $d_{k}^{\ast}$ is nonincreasing in $k,$ letting $s$ to be the smallest
positive integer such that $s^{2}\geq k,$ in view of $\mathbb{S}_{s^{2}
\neq\varnothing,$ we see that $d_{k}^{\ast}\geq1/s.$ But $\left( s-1\right)
^{2}<k,$ and so
\[
\frac{1}{\sqrt{k}}\geq d_{k}^{\ast}\geq\frac{1}{s}>\frac{1}{\sqrt{k}+1
=\frac{1}{\sqrt{k}}+O\left( k^{-1}\right) .
\]
This argument does not fit to bound $d_{k}$, so we need another idea. Since
Taylor's graphs have been useful for $c_{k}$, we can hope to use them for
$d_{k}$ as well. Thus, let $A(\overline{T\left( q\right) })$ be the
adjacency matrix of the complement of the Taylor graph $T\left( q\right) $
of order $q^{3}.$ Define the matrix $T\in\mathbb{U}_{q^{3}}$, by setting
\[
T=2A(\overline{T\left( q\right) })-J_{q^{3}}.
\]
It is not hard to see that the matrix $T$ has three distinct eigenvalues
\begin{align*}
\lambda_{1}(T) & =q^{2}-1,\text{ \ \ \ \ \ with multiplicity \ }q\left(
q-1\right) ;\\
\lambda_{2}(T) & =q^{2}-q-1,\text{ with multiplicity }1;\\
\lambda_{3}(T) & =-q-1,\text{ \ \ \ \ with multiplicity }\left( q-1\right)
\left( q^{2}+1\right) .
\end{align*}
Note in passing that the mapping $A(\overline{T\left( q\right) }\rightarrow
T$ preserves all eigenvalues except $\lambda_{1}(\overline{T\left( q\right)
}),$ whose magnitude is reduced essentially to $\lambda_{2}(\overline{T\left(
q\right) }).$
Now, using the Baker, Harman, and Pintz result \cite{BHP01} again, for
sufficiently large $k,$ we get the asymptotics
\[
\frac{1}{\sqrt{k}}\geq d_{k}\geq\frac{1}{\sqrt{k}+\sqrt[3]{k}}=\frac{1
{\sqrt{k}}+O\left( k^{-2/3}\right) .
\]
\medskip We end up with a question about the maximum Ky Fan $k$-norm of
matrices in $\mathbb{U}_{n}.$
\begin{problem}
For any $k\geq2,$ find $\max\limits_{A\in\mathbb{U}_{n}}\left\Vert
A\right\Vert _{\ast k}.$
\end{problem}
Without a proof, let us mention the bound
\[
(\sqrt{k}-1)n\leq\max_{A\in\mathbb{U}_{n}}\left\Vert A\right\Vert _{\ast
k}\leq n\sqrt{k}.
\]
\bigskip
\textbf{Acknowledgement. }Part of this paper has been prepared for a talk at
the Algebraic Combinatorics Workshop held in the Fall of 2014, at the
University of Science and Technology of China, Hefei. I am grateful for the
hospitality of the organizers, in particular to prof. Jack Koolen.\bigskip
|
\section{Introduction}
At present, graphene is still an object of highly actual interest, especially for its optical properties and plasmonics. There is an enormous number of papers on, for recent reviews see \cite{blud13-27-1341001} and \cite{abal14-1-135}. One of the approaches to the theoretical description of its electronic excitations rests on the Random Phase Approximation (RPA) for the density-density correlation function and for the conductivity. Usually, these quantities are calculated in some approximation, the non-dispersive limit for example. Another approximation rests on the smallness of the Fermi velocity, $\v$, which is 300 times smaller than the speed of light. As shown below, this smallness allows formally to neglect one of the two form factors entering the polarization tensor.
The electronic properties of graphene are equally well described by the Dirac model, consisting of a relativistic spinor in a (2+1)-dimensional space-time (the plane of graphene). It has its own Lorentz group with $\v$ in place of $c$. The coupling to the (3+1)-dimensional electromagnetic field is the usual one. In this way, one has a modified Quantum Electrodynamics (QED), to which the well-known formalisms can be applied. For instance, one calculates the photon polarization tensor $\Pi^{\mu\nu}$. In the language of Quantum Field Theory (QFT) it consists of a fermion loop (if restricting to one loop approximation). This has been done in quite a number of papers, for instance in \cite{gusy07-21-4611} and related papers (mainly in application to external magnetic field). For a kind of review we refer to \cite{fial12-27-1260007}. In fact, this approach is in principle equivalent to RPA, where, however, one makes some approximations for the very beginning.
The formulas obtained using the polarization tensor allow to calculate the Casimir force between graphene and, say a conducting wall, ideal \cite{bord09-80-245406} or real \cite{bord12-86-165429}, and to investigate surface plasmons, e.g. in the TE mode \cite{bord14-89-035421}. Surface plasmons were considered earlier using expressions for the conductivity obtained in RPA. For instance, in \cite{mikh07-99-016803}, using a high frequency approximation for the conductivity, a TE mode plasmon was predicted.
In the present paper, the approach from QFT is taken. The complete polarization tensor is calculated for both, mass $m$ (gap parameter) and chemical potential $\mu$, non-zero. In this case, like with finite temperature, the polarization tensor, which is transverse for gauge invariance, has two form factors, and both are accounted for. Also, the complete frequency and momentum dependence is kept. All approximations, usually done within the RPA approach can be obtained afterwards as special cases.
The polarization tensor was previously calculated in a number of special cases. First of all one needs to mention that the calculations in the RPA approach deliver special cases for components of the polarization tensor. So in \cite{wuns06-8-318,hwan07-75-205418}, in fact the (00)-component of the polarization tensor, which is responsible for static screening and for the TE mode plasmon, was calculated in the massless case. In the QFT approach, this component was calculated with chemical potential and mass in \cite{pyat09-21-025506}.
In application to graphene, the second form factor was calculated in \cite{fial11-84-035446} with mass, temperature and chemical potential, however restricted to Matsubara values of its argument, i.e., to discrete imaginary frequencies. It was applied to the calculation of the Casimir effect. Recently, it was calculated for finite temperature and mass for all, including non-Matsubara and real, frequencies \cite{bord15-91-045037}, but without chemical potential. From these papers, a representation with mass and chemical potential for real frequencies, as it is needed for plasmons, cannot be derived in any simple way. Here we fill this gap and calculate the complete polarization tensor for both non-zero, mass and chemical potential, for real frequencies.
We apply this polarization tensor to the investigation of surface plasmons. For this we start from the generic formulas for the electromagnetic field in (3+1)-dimensions and insert the polarization tensor resulting from the (2+1)-dimensional fermions. We show that this polarization tensor, which involves the Fermi speed $\v$, can be expressed in terms of the same polarization tensor calculated with unit speed. The actual calculation is done in the Appendix. Further we (re-)derive the equations for the reflection coefficients in terms of the two form factors, Eq. \Ref{1.27}, and, which is equivalent, in terms of $\Pi_{00}$ and $\Pi_{\rm tr}$, Eq. \Ref{2.6}. Further, we solve the equations for the plasmon dispersions numerically and represent the results graphically for the TE and TM-modes.
Throughout the paper we use units with $\al=e^2/(4\pi)$ for the coupling and $v=v_F/c$ for the Fermi speed and refer to
\begin{equation}}\newcommand{\ee}{\end{equation} \al\sim\frac{1}{137}, \ v\sim\frac{1}{300},
\label{0.1}\ee
as the physical values of the parameters.
\section{Electrodynamics with polarization tensor from graphene}
By the Dirac equation model, the interaction of the electromagnetic field with the long wavelength electronic excitations in graphene can be described by relativistic quantum electrodynamics. In this formalism, the Dirac equation is
\begin{equation}}\newcommand{\ee}{\end{equation} (i {\slashed \partial}-e\slashed A -m)\psi=0,
\label{1.1}\ee
where
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} &&\slashed \partial=\tilde{\gamma}^\mu\frac{\partial}{\partial x^\mu},\quad \slashed A=\tilde{\gamma}^\mu A_\mu,
\quad \tilde{\gamma}^\mu=\eta^\mu_{\mu'}\gamma^{\mu'},
\nonumber\\[4pt]&&\eta^\mu_{\mu'}={\rm diag}(1,v,v),
\label{1.2}\eea
with the usual gamma matrices $\gamma^{\mu}$ and the Fermi speed $v$. The resulting polarization tensor is
\begin{equation}}\newcommand{\ee}{\end{equation} \tilde{\Pi}^{\mu\nu}(p)=
ie^2\int\frac{dq_0d^2q}{(2\pi)^3}\, tr \frac{1}{i\slashed q-m}\tilde{\gamma}^\mu \frac{1}{i\slashed q-i\slashed p-m}\tilde{\gamma}^\nu
\label{1.3}\ee
where we used momentum representation. Assuming the graphene sheet located at $z=0$, the momenta are in the directions $\mu=0,1,2$. It can be reduced to the polarization tensor $\Pi^{\mu\nu}(p)$, defined by the same formula but with $v=1$, by the substitution $q_i\to q_i/v$ ($i=1,2$) in the integration,
\begin{equation}}\newcommand{\ee}{\end{equation} \tilde{\Pi}^{\mu\nu}(p)=
\frac{1}{v^2}\,\eta^\mu_{\mu'}\Pi^{\mu'\nu'}(\tilde{p})\,\eta^\nu_{\nu'},
\label{1.4}\ee
where the vector $p$ is substituted by $\tilde{p}$. This vector is defined by
\begin{equation}}\newcommand{\ee}{\end{equation}
\tilde{p}^\mu=\eta^\mu_{\mu'} p^{\mu'}.
\label{1.5}\ee
We will use the notation with a tilde for vectors with the Fermi speed $v$ in the spatial components and for the polarization tensor \Ref{1.3} throughout the paper.
In the four dimensional formulation of electrodynamics, the Maxwell equations with polarization tensor (which we denote by a 'hat'), read
\begin{equation}}\newcommand{\ee}{\end{equation} \partial_\mu F^{\mu\nu}+\hat{\Pi}^{\nu\mu}A_\mu=0.
\label{1.6}\ee
These can be viewed as effective Maxwell equations, which appear, e.g., integrating out the spinor fields in a functional integral representation.
Rewritten in terms of induced current,
\begin{equation}}\newcommand{\ee}{\end{equation} j^\mu=\frac{-c}{4\pi}\hat{\Pi}^{\mu\nu}A_\nu,
\label{1.7}\ee
these equations are $\partial_\mu F^{\mu\nu}=(4\pi/c)j^\nu$.
Switching to 3-dimensional notations with $\mu=(0,k)$ ($k=1,2,3$),
\begin{equation}}\newcommand{\ee}{\end{equation} j^\mu=(c\rho,\vec{j}),\quad A^\mu=(\Phi,\vec{A}),
\quad \vec{E}=-\vec{\nabla} \Phi-\partial_0\vec{A},
\label{1.8}\ee
we get from the transversality of the polarization tensor
\begin{equation}}\newcommand{\ee}{\end{equation} \hat{\Pi}^{\mu\nu}A_\mu=\hat{\Pi}^{\nu k}\partial_0^{-1}E_k,
\label{1.9}\ee
and the induced charge density and current,
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} \rho &=& \frac{-1}{4\pi}\,\hat{\Pi}^{0 k}\partial_0^{-1}E_k,
\nonumber\\[4pt] j_k &=& \frac{-c}{4\pi}\,\hat{\Pi}^{k l}\partial_0^{-1}E_{l}.
\label{1.10}\eea
We would like to mention that from this formula the conductivity tensor can be defined as
$\sigma_{kl}=(-c/4\pi)\hat{\Pi}^{kl}\partial_0^{-1}$.
The Maxwell equations \Ref{1.4} read now
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} {\rm div} \vec{E} &=& 4\pi\rho,
\nonumber\\[4pt] \left(-\partial_0^2+\Delta-\nabla\circ \nabla \right)\vec{E} &=& \frac{4\pi}{c}\partial_0\vec{j},
\label{1.11}\eea
or, substituting Gauss's law,
\begin{equation}}\newcommand{\ee}{\end{equation} \left(-\partial_0^2+\Delta \right)\vec{E} =
\frac{4\pi}{c}\left(\nabla c\rho+\partial_0\vec{j}\right).
\label{1.12}\ee
Inserting from \Ref{1.10}, the right hand side can be expressed in terms of the polarization tensor,
\begin{equation}}\newcommand{\ee}{\end{equation} \left(-\partial_0^2+\Delta \right)E_k =
-\left(\nabla_k\hat{\Pi}^{0 l}\partial_0^{-1}+\hat{\Pi}^{kl}\right)E_l.
\label{1.13}\ee
These are general formulas. For graphene we have to insert
\begin{equation}}\newcommand{\ee}{\end{equation} \hat{\Pi}^{\mu\nu}=\left\{
{ \delta(z)\tilde{{\Pi}}^{\mu\nu}(p)\quad{\rm for} \ \mu,\nu=0,1,2,
\atop 0 ~~~~~~~~{\rm for} \ \mu=3 \ {\rm or} \ \nu=3,
} \right.
\label{1.14}\ee
with $\tilde{\Pi}^{\mu\nu}(p)$ from \Ref{1.3}.
In momentum representation we assume all relevant quantities $\sim\exp\left(-i\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha t+i\k \mathbf{x} \right)$
and define the vectors in the plane of graphene by
\begin{equation}}\newcommand{\ee}{\end{equation} \k=\left(k_1\atop k_2\right),\quad \mathbf{x}=\left(x_1\atop x_2\right),
\label{1.15}\ee
and similar for other vectors. Further we need to define the vector
\begin{equation}}\newcommand{\ee}{\end{equation}\label{1.15a} p^\mu=(\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha,\k,0) ,\quad p=\sqrt{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2-\k^2}.
\ee
Now we split equations \Ref{1.12} into two with $n,m=1,2$, i.e., parallel to the plane of graphene, and and a third in the perpendicular direction,
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} \left(p^2+\partial_z^2\right) E_n(p,z) &=& \delta(z) \Xi^{nm}E_{m}(p,0),
\nonumber\\[4pt] \left(p^2+\partial_z^2\right) E_3(p,z) &=& \partial_z \delta(z)\frac{1}{i\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}\tilde{\Pi}^{0n}(p)E_n(p,0),
\label{1.16}\eea
where we introduced the notation
\begin{equation}}\newcommand{\ee}{\end{equation} \Xi^{nm}=\frac{p_n}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}\tilde{\Pi}^{0m}(p)-\tilde{\Pi}^{nm}(p),
\label{1.17}\ee
appearing in the right side in \Ref{1.13}. In \Ref{1.16}, we have to solve the equations for the components $E_n(p,z)$ of the electric field parallel to the plane, whereas $E_3(p,z)$ follows from $E_n(p,z)$ by integration.
Now we need a more specific expression for the polarization tensor for graphene.
From \Ref{1.4} we get
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray}\label{1.17a} \tilde{\Pi}^{00}(p)&=&\frac{1}{v^2}\Pi^{00}(\tilde{p}),
\nonumber\\ \tilde{\Pi}_{\rm tr}&=&\frac{1-v^2}{v^2} \Pi^{00}(\tilde{p})
+\Pi_{\rm tr}(\tilde{p}),
\eea
where we defined
\begin{equation}}\newcommand{\ee}{\end{equation}\label{1.17b} \Pi_{\rm tr}(\tilde{p})=g_{\mu\nu}\Pi^{\mu\nu}(\tilde{p}),\quad
\tilde{\Pi}(p)_{\rm tr}=g_{\mu\nu}\tilde{\Pi}(p)
\ee
for the traces.
The polarization tensor is calculated in the Appendix for $v=1$. There it is represented in terms of form factors $A(p)$ and $B(p)$, Eqs. \Ref{A.6}, where we have to insert with \Ref{1.5}
\begin{equation}}\newcommand{\ee}{\end{equation} \tilde{p}^\mu=(\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha,v\k,0),\quad \tilde{p}=\sqrt{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2-(v\k)^2}
\label{1.18}\ee
in the form factors and in the tensor structures, \Ref{A.7}. For these we note
\begin{equation}}\newcommand{\ee}{\end{equation}\begin{array}{rclrcl}
\tilde{P}^{0m}(\tilde{p})&=& -\frac{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha k_m}{\tilde{p}^2}, &
\tilde{P}^{nm}(\tilde{p}) & =& -\delta_{nm}-\frac{v^2k_n k_m}{\tilde{p}^2},
\\[4pt] \tilde{M}^{0m}(\tilde{p}) &=& \frac{v^2 k^2 k_m}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha\tilde{p}^2},&
\tilde{M}^{nm}(\tilde{p}) &=& \frac{v^2 k_n k_m}{\tilde{p}^2} . \end{array}
\label{1.19}\ee
Accounting also for the $\eta^\mu_{\mu'}$ we get from \Ref{1.4}
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} \tilde{\Pi}^{0m} &=& \frac{-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha k_m}{\tilde{p}^2}
\left(A(\tilde{p})-\frac{v^2k^2}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}B(\tilde{p})\right),
\nonumber\\[4pt] \tilde{\Pi}^{nm} &=& -\left(\delta_{nm}+
\frac{v^2 k_n k_m}{\tilde{p}^2}\right)A(\tilde{p})
+\frac{v^2 k_n k_m}{\tilde{p}^2}B(\tilde{p}).
\label{1.20}\eea
This allows to rewrite \Ref{1.17} in the form
\begin{equation}}\newcommand{\ee}{\end{equation} \Xi^{nm}=\left(\delta_{nm}-(1-v^2)\frac{ k_n k_m}{\tilde{p}^2}\right)A(\tilde{p})
-\frac{v^2 p^2\,k_nk_m}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2\tilde{p}^2}\,B(\tilde{p}).
\label{1.20M}\ee
The form factors $A({p})$ and $B({p})$ are calculated in the Appendix, Eqs. \Ref{A.9} and \Ref{A.50}, \Ref{A.61}.
Next we introduce the polarizations for the electric field. We consider only the components $E_n$ with $n=1,2$,
\begin{equation}}\newcommand{\ee}{\end{equation} E_{n}(p,z)=\left({-k_2\atop k_1}\right)\Phi_{\rm TE}
+\left({k_1\atop k_2}\right)\Phi_{\rm TM},
\label{1.21}\ee
where, up to a normalization, we have $E_{\rm TE}=\Phi_{\rm TE}$ and $E_{\rm TM}=\partial_z\Phi_{\rm TM}$.
We do not need to consider here the third component, $E_3$, since it follows with \Ref{1.16} from $E_n$. Now, Eq. \Ref{1.16} is diagonal in the polarizations, i.e., the graphene does not mix these. So we get with \Ref{1.21} from \Ref{1.16}
\begin{equation}}\newcommand{\ee}{\end{equation} (p^2+\partial_z^2)\Phi_{\rm TX}(p,z)=\delta(z) \Xi_{\rm TX} \Phi_{\rm TX}(p,0).
\label{1.22}\ee
Here the subscript ${\rm 'TX'}$ stands for one of the polarizations and
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} \Xi_{\rm TE} &=& \frac{1}{k^2}\left({-k_2\atop k_1}\right)_nM^{nm}
\left({-k_2\atop k_1}\right)_m= A(\tilde{p}),
\nonumber\\[4pt] \Xi_{\rm TM} &=& \frac{1}{k^2}\left({k_1\atop k_2}\right)_nM^{nm}
\left({k_1\atop k_2}\right)_m,
\nonumber\\[4pt]&=&\frac{p^2}{\tilde{p}^2}\left(A(\tilde{p})
-v^2\frac{k^2}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}B(\tilde{p})\right).
\label{1.23}\eea
Finally, we rewrite the equation \Ref{1.22} for $z=0$ in terms of matching conditions,
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} \Phi_{\rm TX}(z=+0)- \Phi_{\rm TX}(z=-0) &=&0,
\nonumber\\[4pt] \partial_z \Phi_{\rm TX}(z=+0)-\partial_z \Phi_{\rm TX}(z=+0)&=&\Xi_{\rm TX} \Phi_{\rm TX}(z=0),
\label{1.24}\eea
for both polarizations, i.e., the functions are continuous and their derivatives jump. In a standard scattering setup, the solutions are
\begin{equation}}\newcommand{\ee}{\end{equation} \Phi_{\rm TX}(z)=\left(e^{ipz}+r_{\rm TX}e^{-ipz}\right) \Theta(-z)
+t_{\rm TX}e^{ipz}\Theta(z)
\label{1.25}\ee
and the reflection and transmission coefficients are
\begin{equation}}\newcommand{\ee}{\end{equation} r_{\rm TX}=\frac{-1}{1+Q_{\rm TX}^{-1}},\quad t_{\rm TX}=\frac{1}{1+Q_{\rm TX}},
\label{1.26}\ee
with
\begin{equation}}\newcommand{\ee}{\end{equation} Q_{\rm TE}=\frac{-1}{2ip}{A}(\tilde{p}),\quad
Q_{\rm TM}=\frac{-p}{2i\tilde{p}^2}\left(A(\tilde{p})
-v^2\frac{k^2}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}B(\tilde{p})\right).
\label{1.27}\ee
These are the final formulas for the coefficients. Similar formulas were derived earlier, e.g., Eq. (23) in \cite{fial11-84-035446}.
\section{Surface plasmons}
\subsection{General formulas}
Surface plasmons appear if the reflection and transmission coefficients \Ref{1.26} have a pole. So these are solutions of the equations
\begin{equation}}\newcommand{\ee}{\end{equation} 1+ Q_{\rm TX}=0.
\label{2.1}\ee
We look for them in the frequency region
\begin{equation}}\newcommand{\ee}{\end{equation} vk < \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha < k .
\label{2.2}\ee
The upper bound implies a frequency below the continuous spectrum in order to get a wave function \Ref{1.25} decreasing to both sides of the graphene sheet. In this region the momentum $p$ is imaginary and we use the notation
\begin{equation}}\newcommand{\ee}{\end{equation} p=i\eta\equiv i\sqrt{-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2+k^2}
\label{2.3}\ee
and get from \Ref{1.27}
\begin{equation}}\newcommand{\ee}{\end{equation} Q_{\rm TE}=\frac{ 1}{2\eta}{A}(\tilde{p}),\quad
Q_{\rm TM}=\frac{-\eta}{2\tilde{p}^2}\left(A(\tilde{p})
-v^2\frac{k^2}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}B(\tilde{p})\right).
\label{2.4}\ee
Here we insert from \Ref{A.9},
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} A(\tilde{p}) &=& \frac{\tilde{p}^2}{v^2k^2}\Pi^{00}(\tilde{p})
+\Pi_{\rm tr}(\tilde{p}),
\nonumber\\[4pt] B(\tilde{p}) &=& \frac{2\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2\tilde{p}^2}{v^4k^4}\Pi^{00}(\tilde{p})
+\frac{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}{v^2k^2}\Pi_{\rm tr}(\tilde{p}).
\label{2.5}\eea
These form factors can be inserted into \Ref{2.4} and we get
\begin{equation}}\newcommand{\ee}{\end{equation}\begin{array}{rclrl}
Q_{\rm TE} &=& \frac{ 1}{2\eta}\left(\frac{\tilde{p}^2}{v^2k^2}\Pi^{00}(\tilde{p})
+\Pi_{\rm tr}(\tilde{p})\right)
&=&\frac{ 1}{2\eta}\left(\frac{{p}^2}{k^2}\tilde{\Pi}^{00}({p})
+\tilde{\Pi}_{\rm tr}({p})\right),
\\[4pt] Q_{\rm TM} &=& \frac{\eta}{2v^2k^2}\,\Pi^{00}(\tilde{p})
&=& \frac{\eta}{2 k^2} \, \tilde{\Pi}^{00}({p}),
\end{array}\label{2.6}\ee
where we also displayed the expressions in terms of the $\tilde{\Pi}^{\mu\nu}(p)$, Eq. \Ref{1.3}, using \Ref{1.17a}.
Eqs. \Ref{2.6} are the final formulas for the investigation of the equations \Ref{2.1}. The components of the polarization tensor $\Pi^{\mu\nu}(p)$ are given by Eqs. \Ref{A.51a} and \Ref{A.61}. These coincide with Eq.(24) in \cite{fial11-84-035446}, or Eq. (12) in \cite{bord15-91-045037}, where, however, different notations are used.
It is meaningful to check the case $\mu\le m$, which we considered in \cite{bord14-89-035421}. In that case we have $B=0$ for the second form factor and from \Ref{A.15} the relation
\begin{equation}}\newcommand{\ee}{\end{equation} \Pi_{\rm tr}(\tilde{p})=\frac{-2\tilde{p}^2}{v^2k^2}\Pi^{00}(\tilde{p})
\label{2.7}\ee
holds. From \Ref{2.5} we get
\begin{equation}}\newcommand{\ee}{\end{equation} A(\tilde{p})=-\frac{\tilde{p}^2}{v^2k^2}\Pi^{00}(\tilde{p}).
\label{2.8}\ee
In the notations of \cite{bord14-89-035421} the polarization tensor is expressed in terms of the function
\begin{equation}}\newcommand{\ee}{\end{equation} \Phi(\tilde{p})=
\frac{2}{\tilde{p}} \left( 2m\tilde{p}
-(\tilde{p}^2+4m^2){\rm arctanh}\frac{\tilde{p}}{2m}\right)
\label{2.9}\ee
by
\begin{equation}}\newcommand{\ee}{\end{equation} \Pi^{00}(\tilde{p})=-\frac{e^2}{4\pi}\ \frac{v^2k^2}{2\tilde{p}^2} \ \Phi(\tilde{p})
\label{2.10}\ee
delivering
\begin{equation}}\newcommand{\ee}{\end{equation} Q_{\rm TE}=\frac{\alpha}{2\eta}\ \frac12 \ \Phi(\tilde{p}),
\quad Q_{\rm TM}=-\frac{\alpha\eta}{2\tilde{p}^2}\ \frac12 \ \Phi(\tilde{p}).
\label{2.11}\ee
This coincides with Eq. (18) in \cite{bord14-89-035421}.
\subsection{Formulas for the surface plasmon in the massless case}
Frequently, graphene is considered with zero gap widths, which seems justified given the small value of the gap widths. In the Dirac model this translates into the massless case. For $m=0$, the formulas \Ref{1.26} and \Ref{2.6} remain unchanged whereas those for the polarization tensor simplify. We collected them in the Appendix B. With the substitution $p\to\tilde{p}$ we get from \Ref{B.2} and \Ref{2.6}
\begin{eqnarray}}\newcommand{\eea}{\end{eqnarray} Q_{\rm TE}&=& \frac{4\al}{\eta}
\left[\frac{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2\,\mu}{(vk)^2}-\frac{\tilde{p}}{4}\ \frac12\sum_{\lambda_1=\pm1}{\rm sign}(Q)
\left(x\sqrt{x^2-1}+{\rm arccosh}(x)\right)\right],
\nonumber\\
Q_{\rm TM}&=& \frac{4\al \eta}{(vk)^2}\left[\mu
-
\frac{(vk)^2}{4\tilde{p}} \ \frac12\sum_{\lambda_1=\pm1}{\rm sign}(Q)
\left(x\sqrt{x^2-1}-{\rm arccosh}(x)\right)\right].
\label{3.1}\eea%
The variables are
\begin{equation}}\newcommand{\ee}{\end{equation} \eta=\sqrt{k^2-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2},\ \tilde{p}=\sqrt{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2-(vk)^2}, \ x=\frac{2\mu-\lambda_1\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}{vk}, \ Q=\tilde{p}^2+2\lambda_1\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha\mu.
\label{3.2}\ee
As said above, surface plasmons appear as solutions of the equation \Ref{2.1}.
Now, inserting numbers using \Ref{4.1} shows that for the TE polarizations the solutions start with $k>0$, whereas for the TM case the solutions start from $k=0$. In this case, the solution can be considered for small $k$. A direct expansion of \Ref{3.1} in powers of $k$ gives
\begin{equation}}\newcommand{\ee}{\end{equation} Q_{\rm TM}=
-\al\frac{\sqrt{k^2-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}}{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}\ \ln\frac{2\mu+\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}{2\mu-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}+O(k).
\label{3.4}\ee
This expression is in agreement with Eq. (8) in \cite{mikh07-99-016803}, which turns out to be the approximation for small $k$ and $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha$. Eq. \Ref{2.1} with \Ref{3.4} can be solved for small $k$ by iteration. First we rewrite the equation in the form
\begin{equation}}\newcommand{\ee}{\end{equation} \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=\al {\sqrt{k^2-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}}\ \ln\frac{2\mu+\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}{2\mu-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha}+O(k).
\label{3.5}\ee
Next we insert $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=0$ in the right side and get $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_{\rm sf}(k)$ in the left side with
\begin{equation}}\newcommand{\ee}{\end{equation} \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_{\rm sf}(k)=\al\,k+O(k^2).
\label{3.6}\ee
This solution confirms that we get small $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_{\rm sf}(k)$ for small $k$ as anticipated in the derivation of \Ref{4.4}. It is interesting to note, that this solution, because of $v k<\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_{\rm sf}(k)< k$, \Ref{2.2}, is consistent for
\begin{equation}}\newcommand{\ee}{\end{equation} v<\al<1
\label{3.7}\ee
only, which is fulfilled for the physical values of the parameters.
\subsection{Regions in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane}
As said in Eq. \Ref{2.2}, plasmons may exist in a region in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane below the continuous spectrum. In addition, the polarization tensor must be real. Since we are in the Minkowskian region, it is real below the threshold at
\begin{equation}}\newcommand{\ee}{\end{equation} \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s=\sqrt{(v k)^2+(2m)^2},
\label{4.1}\ee
i.e., for $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha<\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$. In the case without chemical potential, i.e., for $\mu=0$, this is the only region where the polarization tensor is real. With chemical potential there is also above the threshold a region of realness. It is bounded from above by the condition \Ref{A.63a}, which reads
\begin{equation}}\newcommand{\ee}{\end{equation} \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha<\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-
\label{4.2}\ee
with
\begin{equation}}\newcommand{\ee}{\end{equation} \omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-=\mu+\sqrt{\mu^2+(vk)^2-2vk\,k_F}
\label{4.3}\ee
written in the notations used here. Combining \Ref{2.2} and \Ref{4.2}, we get
\begin{equation}}\newcommand{\ee}{\end{equation} vk<\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha<\min(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-),
\label{4.4}\ee
which defines the region in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane, where plasmons may exist.
This region is shown in Fig.~\ref{regions}\footnote{All plots are made for $\mu=1$.}. It looks different for different ratios $m/\mu$.
For $m=0$ it is shown in Fig.~\ref{regions}(a). Here $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$ degenerates, $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-=\mu+|\mu-vk|$, and the region \Ref{4.4} becomes a triangle. Since here $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s=vk$ holds, all solutions are automatically above the threshold.
For $m>0$ we introduce the following notations,
\begin{equation}}\newcommand{\ee}{\end{equation} k_s=\frac{2m}{\sqrt{1-v^2}},\ k_m=\frac{2(\mu-vk_F)}{1-v^2}, \
k_t=\frac{2k_F}{v} ,
\label{4.5}\ee
where $k_F=\sqrt{\mu^2-m^2}$ is the Fermi momentum. In \Ref{4.5}, $k_s$ is the intersection $k=\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$, $k_m$ is the intersection $k=\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$, and $k_t$ is the touching point $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-=\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$. For $m=m_t$, with
\begin{equation}}\newcommand{\ee}{\end{equation} m_t=\sqrt{1-v^2}\mu,
\label{4.6}\ee
these coincide,
\begin{equation}}\newcommand{\ee}{\end{equation} k_s=k_m=k_t=2\mu \ \ \ \ \ (m=m_t).
\label{4.7}\ee
The case $m<m_t$ is shown in Fig. \ref{regions} (b) and $m_t<m<\mu$ is shown in Fig. \ref{regions} (d). It must be mentioned that the last region is very narrow,
\begin{equation}}\newcommand{\ee}{\end{equation} \frac{\mu-m_t}{\mu}\sim 5\cdot 10^{-6},
\label{4.8}\ee
and $m_t=0.999994\mu$ for physical values of the parameters.
\begin{figure}[h]
\epsfig{file=Regions_m0.eps,width=6cm}
\epsfig{file=Regions_06.eps,width=6cm}\\
\epsfig{file=Regions_ms.eps,width=6cm}
\epsfig{file=Regions_mt.eps,width=6cm}
\caption{The $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane with the curves for $v k$, $k$, $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$ and $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$ for (a) $m=0$, (b) $m=0.6$, (c) $m=0.999984\mu$, i.e., $m<m_t$, and (d) $m=0.999997$, i.e., $m_t<m<\mu$, for physical values of the parameters. The region \Ref{4.4}, where a plasmon may exist, is shaded.
In panel (b), $k_t\gg k_m$ and outside the graph. The curves $\omega_p$ and $\omega_s$ appear as straight lines and the line $v k$ nearly coincides with the $k$-axis.
The inset shows the same picture for larger $k$ such that $k_t$ is seen. Now the line $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=k$ nearly coincides with the $\omega$ axis.
Panel (d) corresponds to the narrow region $m_t<m<\mu$, where the touching point $k_t$ is below $k_s$ and $k_m$.}
\label{regions}
\end{figure}%
\subsection{TM mode surface plasmons}
TM mode surface plasmons are solutions of the equation \Ref{2.1},%
\begin{equation}}\newcommand{\ee}{\end{equation} 1+Q_{\rm TM}=0,
\label{5.1}\ee
where $Q_{\rm TM}$ is given by \Ref{A.51a} and $\Pi^{00}(\tilde{p})$ by \Ref{A.54} with $\tilde{p}$ inserted for $p$.
We start with the case $m=0$. Here, $Q_{\rm TM}$ is given by \Ref{4.1}. As said above, the solution exists in the triangular region shown in Fig.~\ref{regions}(a). Examples are shown in Fig. \ref{TM1}, left panel, for several values of $\al$ and $v=1/300$. These solutions start in $k=0$ and terminates on the line $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=2\mu-vk$.
Since these solutions are all restricted to $k<v/\mu$, the spatial extend of the solutions in the sense of $\exp(-\kappa |z|)$ is determined by $\kappa=\sqrt{k^2-\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha^2}$. It is shown in Fig. \ref{TM1} in the inset.
For the physical value $\al=1/137$, the TM solution bends closer to $v k$. A similar picture may be found in \cite{stau14-26-123201}.
\begin{figure}[h]
\epsfig{file=TM_plasm_M0_mu_v3.eps,width=8cm}
\epsfig{file=TE_plasm_M0_mu_v3.eps,width=8cm}
\caption{Left: Transverse magnetic plasmon with $m=0$, $\mu\ne0$, and $v=1/300$, for different values of $\alpha$. The inset shows $\varkappa=\sqrt{k^2-\omega_{TM}^2}$ as a function
of $v k/\mu$ for $\alpha=1/137$.
Right: Transverse electric plasmon. These plots are very close one to the other differing significantly only by their endpoints. The inset (a) is the zoom of TE curves. The nearly horizontal dashed line is $\omega=2\mu- v k$. The inset (b) shows $\sqrt{k^2-\omega_{TE}^2}$ as a function of $ v k/\mu$.}
\label{TM1}
\end{figure}%
For $m>0$, the plasmon solution exists in the regions shown in Fig.~\ref{regions}, (b)-(d). The curve $\omega_{+}^{-}$ separates from $\omega= v k$ and in the gap between them a new branch of the solution appears
for large $k$, as shown in Fig.~\ref{TM2}, panel (a).
Obviously this branch and the lower one are parts of a single solution which, however in the gap between these branches is not real. When further increasing $m$, starting from $m=m_0$, these branches merge, see Fig.~\ref{TM2}, panel (b).
The mass, for which the solution touches the curve $\omega_{+}^{-}$ at its minimum, $k_{min}=\sqrt{\mu^2-m^2}/2$, is denoted by $m_0$. It can be found by substituting $k_{min}$ and $\omega_{+}^{-}(k_{min})=m+\mu$,
into Eq.~\Ref{5.1}. With physical values of the parameters, this equation yields $m_0=0.23235\mu$.
\begin{figure}[h]
\epsfig{file=TM_m_v300_al137_m01.eps,width=7cm}
\epsfig{file=TM_m_v300_al137_m0232347.eps,width=7cm}
\epsfig{file=TM_m_v300_al137_m06.eps,width=7cm}
\caption{TM-plasmon dispersion law for physical values of the parameters.
Panel (a) shows the solution with $m=0.1<m_0$ consisting of two branches.
Panel (b) presents the plot for $m=m_0=0.232347$, corresponding to the mass when two branches of the solution merge. For this mass $k_{\rm max}=597.863$. Panel (c) with $m=0.6$ demonstrates that with increasing mass the endpoint moves toward the origin.
The curves for $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$, $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$ and $vk$ (from top to bottom) are shown as dashed lines.}
\label{TM2}
\end{figure}
The solution has an endpoint,
\begin{equation}}\newcommand{\ee}{\end{equation}
k_{max}=\frac{\sqrt{3} \sqrt{8 \alpha^2 \mu (\mu-m) \left(1-v^2\right)+3 m^2 v^2}}{2 \alpha v \sqrt{1-v^2}}-\frac{3 m}{2 \alpha \sqrt{1-v^2}},
\label{5.4}
\ee
which moves down to smaller $k$ when further increasing $m$, see
Fig.~\ref{TM2}, panel (c). Finally, for $m\to\mu$, i.e. when the chemical potential disappears, this endpoint goes down to $k=0$, and the solution disappears.
For small $k$ it is possible to find the solution explicitly. For this, we rewrite Eq.~\Ref{5.1} in the form
\begin{equation}}\newcommand{\ee}{\end{equation} \omega=\sqrt{k^2-4 v^2 k^4 \Pi_{00}^{-2}},
\label{5.4a}\ee
which allows for iteration with inserting $\omega=k$ in the right side,
\begin{equation}}\newcommand{\ee}{\end{equation} \omega=k -\frac{v^4}{8 \alpha^2 \mu^2}
\frac{1}{(1 - \frac{\mu}{\sqrt{m^2 v^2 + \mu^2 (1 - v^2)}})^2}\,k^3+O(k^4).
\label{5.5}\ee
The coefficient in front of $k^3$ is negative, showing that the solution goes indeed below $\omega=k$, i.e., below the
border of the continuous spectrum.
This coefficient is small, proportional to $(v^2/\al)^2\sim 2.3\cdot10^{-6}$, unless $m\to\mu$, where it becomes infinite and the expansion breaks down (expansion \Ref{3.6} holds instead).
\subsection{TE mode surface plasmon}
TE mode surface plasmons are solutions of the equation \Ref{2.1},
\begin{equation}}\newcommand{\ee}{\end{equation} 1+Q_{\rm TE}=0,
\label{6.1}\ee
where $Q_{\rm TE}$ is given by \Ref{2.6} with \Ref{A.51a} and \Ref{A.61} with $\tilde{p}$ inserted for $p$.
The TE mode surface plasmon solution has a nonzero staring point, which we denote by $k_0$. It can be found as a solution of the equation $1+{Q_{\rm TE}}_{|_{\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=k}}=0$. For $k<k_0$, the solution goes into the continuous spectrum. It is shown in Fig. \ref{TE1}, together with $k_s$ and $k_m$, Eq. \Ref{4.5}, as a function of $m$. The intersection of these curves in the left panel is denoted by $m_0$. This means, for $m>m_0$, that the starting point in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane, see Fig. \ref{regions}, is to the {left} of $k_s$ and for $m<m_0$ to the {right}. The latter implies that the starting point is {above} the threshold.
In the case $m=0$, as mentioned above, the solution exists in the triangular region shown in Fig. \ref{regions}, panel (a). Examples are shown in Fig.~\ref{TM1}, right panel, for several values of $\al$ and $v=1/300$.
The TE solution goes close to $\omega=k$. Its frequency lies in the band $1.667<\omega/\mu<1.9934$. This is in agreement with~\cite{mikh07-99-016803}. The starting point lies at $\omega=k$ and does not depend on $\alpha$. The endpoint is situated on the line $\omega=2\mu- v k$ and depends on $\alpha$, see the inset. With decreasing $\alpha$ it tends to $k_{m}=2\mu/(1+v)$.
For $m<m_t$, the solution ends on the line $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$, thus above the threshold. Such solutions are shown in Fig. \ref{TE2} for several $m<m_t$. In the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane, all appear nearly on one and the same line. Therefore we represented them in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha/k_F)$-plane. Close to the end, these lines have a knee. This is shown in Fig. \ref{TE2} in the right panel for a particular value of $m$ using smaller values of $\al$ enlarging this part.
Further increasing $m$, for $m>m_t$, the endpoint appears below the threshold and it goes to larger $k$ the closer $m$ comes to $\mu$. At the same time the starting point goes to zero, see Fig. \ref{TE1}. For $m\to\mu$, which is the transition to the case without chemical potential, the starting point is zero and the endpoint goes to infinity. The picture turns into that shown on Fig. 2 in \cite{bord14-89-035421}.
\begin{figure}[h]
\epsfig{file=TE_starting_point1.eps,width=8cm}
\caption{The starting point $k_0$ of the TE-plasmon as a function of $m$ for physical values of $\al$ and $v$, together with $k_s$ and $k_m$.}
\label{TE1}
\end{figure}
\begin{figure}[h]
\epsfig{file=TE_plasm_around_m0_v2.eps,width=6cm}
\epsfig{file=TE_plasm_m09_alpha.eps,width=6cm}
\caption{Left panel: TE-plasmon dispersion laws for physical values of the parameters (solid lines) for several values of the mass, $m=0, 0.4, 0.6, 0.8, 0.9$, as function of $k/k_F$. The uppermost curve corresponds to zero mass and is the same as in Fig. \ref{TM1}, right panel.
Right panel: the same for $m=0.9$, for several values of the coupling, $\alpha=1, 0.5, 0.2, 1/137$. The knee at the end of the curves in the left panel becomes visible only for coupling stronger than its physical value. The curves for $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_+^-$ and $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha_s$ are shown as dashed lines. These are nearly horizontal. Also the straight lines for $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha=k$ are shown as dashed lines. The corresponding line for $vk$ coincides in the plot with the $k$-axis.}
\label{TE2}
\end{figure}
\section{Conclusions}
In the foregoing sections we calculated the polarization tensor for graphene with mass and chemical potential using formalism and notations of quantum field theory.
The actual calculation of the polarization tensor is carried out in the Appendix, for unit speed of light. In the case considered here, all integrations can be carried out and explicit formulas for the complete polarization tensor emerge. It is seen that it does not have an ultraviolet divergence, which is the expected result from the interplay of gauge invariance and dimensionality.
The polarization tensor with $m$ and $\mu$ has 2 form factors. The result is given in terms of these. Formulas of this kind for the polarization tensor were obtained previously, but never in such completeness as here. Especially, in \cite{pyat09-21-025506}, only one form factor, $\Pi^{00}(\tilde{p})$, was calculated.
It should be mentioned, that the second form factor, $\Pi_{\rm tr}(\tilde{p})$, Eq.\Ref{A.9tr}, frequently gives only a small contribution. This can be seen from the upper line in Eq. \Ref{2.6}, where $\Pi_{\rm tr}(\tilde{p})$ enters with a relative weight proportional to $v\sim 1/300$ as compared to $\Pi^{00}(\tilde{p})$. In other cases, its contribution is essential. For example, for $\mu\le m$, from \Ref{2.7}, $\Pi_{\rm tr}(\tilde{p})$ gives twice the contribution as $\Pi^{00}(\tilde{p})$ and has opposite sign. Thus dropping $\Pi_{\rm tr}(\tilde{p})$ in $Q_{\rm TE}$, Eq. \Ref{2.6}, changes even the sign, which results in the disappearing of the TE plasmon in this case.
The explicit form of the final formulas looks different in different regions in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane. All these are related by corresponding analytic continuation, which can be best understood in terms of the left side of Eq. \Ref{A.36}, viewed as an integral representation of the result. For the regions relevant for plasmons, i.e, where the polarization tensor is real, we gave explicit expressions in terms of real functions, Eqs. \Ref{A.51a}, \Ref{A.61} for frequencies below the threshold, and Eqs. \Ref{A.65}, \Ref{A.66} for above. Also, we gave in Appendix B explicit formulas for the massless case.
In section II we started from the general 4-dimensional notations of QED, used the effective Maxwell equations and related these to the (3+1)-dimensional formulation with (2+1)-dimensional polarization tensor. This allowed, after separation of the polarizations into TE and TM, to formulate a scattering setup for the electromagnetic~ field and to relate the scattering coefficients with the form factors of the polarization tensor. We represented the whole derivation in necessary detail to make the paper self contained. Also, at the end of the section, we discussed the restoration of $v\ne1$, including the mixing of the from factors, Eq. \Ref{1.17a}.
Finally, in Sec. III.A, from the reflection coefficients, we derived the equations for the plasmons, including the simplifications appearing for $m=0$, Sec. III.B.
In the remaining part of Sec. III, we investigated the plasmons. There, a quite sophisticated number of cases appeared. First of all, the regions in the $(k,\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha)$-plane, where plasmons exist, quite strongly depend on $m$ in the interval $0\le m\le\mu$, see Fig. \ref{regions}. In the massless case, we reproduce known results, see Fig. \ref{TM1}. When the mass starts to grow, these pictures deform, see Figs. \ref{TM2} and \ref{TE2}. Finally, for $m=\mu$, i.e., when the chemical potential disappears, we turn into the case considered in \cite{bord14-89-035421}.
A general property of these solutions is that there are always (except for $\mu=0$) solutions for both polarizations present. Thereby, for small $m$, the TM solution is present starting from $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha\sim k\ge0$ and the TE solution only for larger $k$ and $\omega}\newcommand{\Om}{\omega_p}\newcommand{\al}{\alpha$, see Fig.~\ref{TM1}. For $m$ closer to $\mu$, the range of $k$, where the TM solution exists, shrinks, Fig. \ref{TM2}, panel (c), whereas the TE solution extends to larger $k$-intervals, see Fig. \ref{TE2}, right panel. Roughly speaking, for $m=0$, $\mu\ne0$, there is a TM solution and for $m\ne0$, $\mu=0$, there is a TE solution.
It must be mentioned that, possibly, not all solutions considered here, are physical in the sense that some, e.g., Fig. \ref{TM2}, left panel, exist only for $k$ exceeding the range of validity of the Dirac model for graphene. Another observation is that the smallness of the physical parameters, enters in different way. Besides the general proportionality to $\al$ of the polarization tensor, and to the Fermi speed $v$ in the range \Ref{2.2}, the combination $v^2/\al$ enters, Eq. \Ref{5.5}. The combination $\al/v$, which is of order of unity, did not play any pronounced role.
It would be interesting to carry on a similar analysis of the plasmons for finite temperature without and with chemical potential.
\acknowledgements
We acknowledge a partial support from the Heisenberg-Landau Programme.
|
\section{Introduction}
Noble metal nanoclusters are currently attracting a great deal of attention within both
experimental and theoretical communities, due to their interesting structural
and electronic properties and promising technological applications in nano electronics,
nano optics, biological sensing, catalysis,
and biomedicine \cite{Udayabhaskararao,biological,Burda,Murphy,Xia}.
While transition metal ions may induce health and environmental
problems \cite{Bioinorgan,Aragay},
noble metal nanoclusters exhibit lower toxicity and hence more biocompatibility.
Noble metal nanoclusters are alloyed with palladium atoms to improve their
catalytic activity \cite{Neerga,WL}.
In recent years, extensive computational studies have been devoted to
properties of pure noble metal clusters \cite{Berkahem,7,8,9},
although less information is available about structural and
electronic properties of the corresponding bimetallics \cite{10,11,Wissam}.
Among noble metal based bimetallic nanoclusters, copper and silver based nano-alloys
have been studied considerably less than gold based bimetallic nanoclusters.
Wang et al. \cite{Wang} and Romanowski et al.\cite{Roman} used
density functional (DFT) computations to study AgPd and CuPd dimers
and their interaction with H$_2$.
Other calculations have been performed on Cu-Pd trimers
and their interaction with molecular and atomic oxygen \cite{Gobal}.
The stable geometry of small Ag$_n$Pd$_m$ clusters up to five atoms
were found to transform from two-dimensional
to three-dimensional as the Pd content increases \cite{Papageo}.
Zhao et al. performed first-principles calculations to study
the effects of Pd doping on structural and electronic properties
of Ag$_n$Pd ($n\leq5$) clusters and their mono hydrides \cite{Zhao}.
Efremenko et al.\cite{32} have studied the geometric structures and electronic
properties of the Pd$_n$Cu$_m$ ($n+m\leq6$) clusters using DFT.
The obtained results show that stability of nanoclusters of the same shape
and composition increases linearly with increasing number of Pd–Cu bonds.
Up to our knowledge, there is no report on magic number, vibrational spectra,
and many body corrected electronic structure of small Pd doped Ag and Cu clusters.
In this paper, quantum mechanical calculations are employed to study
pure M$_n$ (M = Ag, Cu) and doped M$_{n-m}$Pd$_m$ ($n\leq9, m=1,2$) clusters
and investigate the behavior of their structural, electronic, and vibrational
properties as a function of size.
The rest of the paper is organized as follows.
In section II, we give a brief introduction of the computational method
used in this work.
Then in section III, atomic dimers are investigated with several
exchange-correlation functional to select the proper one for our calculations.
Next section involves the results of our search for lower energy structures
of the selected nano-clusters.
In section IV, the structural, electronic, magnetic and vibrational properties
of the most stable clusters are presented.
Finally, we will summarized our findings in the last section.
\section{Method}
All presented calculations are performed in the framework of
the spin-polarized Kohn-Sham density functional theory by using the all electron
full-potential code {\sc FHI-aims} \cite{FHI-aims}.
This package employs basis sets consisting of atom-centered
numerical orbitals of the form:
\begin{equation}
R(r)=\dfrac{u_{i}(r)}{r}
\end{equation}
As the name implies, the radial free atom like orbitals $u_i(r)$
are numerically tabulated and therefore exhibit very high flexibility.
The strict localized nature of these basis functions lead to much slower
scaling of the computational time versus the system size.
The calculations reported here are done with the "tier2+spd", "tier1+spd",
and "tier2" basis sets which contain 116, 152, and 45 basis functions
for Cu, Ag, and Pd atoms, respectively.
All the calculations were carried out in the scalar relativistic limit,
while the spin-orbit correction was neglected.
A recent computational study on the atomic structure of transition metal clusters
concluded that this relativistic term has negligible effect on the relative energy
differences of the 4d systems \cite{17a}.
Geometry optimization is performed by the standard Broyden-Fletcher-Goldfarb-Shanno
(BFGS) algorithm \cite{19} with a force accuracy of about $10^{-2} eV/\AA$.
Harmonic frequencies are calculated using
the finite displacement of all atomic positions by $10^{-3} \AA$.
The lowest energy structures are confirmed to be the true minima by calculating
their vibrational frequencies.
The FHI-aims package enables us to describe electronic single-quasiparticle
excitations in molecules by using many-body correction GW self-energy \cite{gw}.
\section{Dimers}
First we focus on the atomic dimers to select a reliable exchange-correlation
(XC) functional for our computations and moreover to gain some insights
about interatomic interaction in our systems.
The equilibrium bond length, binding energy, and
vibrational frequency of the Cu$_2$ and Ag$_2$ dimers were calculated
by using different XC approximations, including PBE \cite{20},
revPBE \cite{21} and BLYP \cite{blyp} generalized gradient (GGA) functionals
and PW \cite{pw} local density functional.
The obtained results are compared with available experimental data
in table \ref{dimer}.
It is seen that PBE, revPBE, and BLYP give significantly more accurate
binding energy and vibrational frequency while accuracy of
equilibrium bond length is qualitatively the same within LDA and
GGA based functionals.
In the case of Cu$_2$, the overall accuracy of revPBE seems to be slightly
better than PBE, while the parameters of Ag$_2$ within PBE is overall
closer to experiment.
Although, BLYP gives better binding energy for Ag dimer, for larger silver clusters,
other theoretical calculations \cite{27} argue that the BLYP functional
fails to predict the correct stable structure of the system.
Therefore, we adapt the PBE and revPBE functionals for
calculation of Ag and Cu based clusters, respectively.
\begin{table}
\newcommand{\cite{22}}{\cite{22}}
\newcommand{\cite{23}}{\cite{23}}
\newcommand{\cite{24}}{\cite{24}}
\newcommand{\cite{25}}{\cite{25}}
\newcommand{\cite{28}}{\cite{28}}
\newcommand{\cite{29}}{\cite{29}}
\newcommand{\cite{30}}{\cite{30}}
\caption{\label{dimer}
Calculated equilibrium bond length $r_e$ (\AA),
binding energy $E_b$ (eV/atom) and harmonic vibrational frequency
$\omega$ (cm$^{-1}$) of the Cu$_2$, Ag$_2$, Pd$_2$, CuPd, and AgPd dimers.
The Cu$_2$ and Ag$_2$ dimers are calculated in four different
exchange-correlation (XC) functionals.
The corresponding experimental values (Expt.) are
also given for comparison.
}
\begin{ruledtabular}
\begin{tabular}{lllll}
& XC & $r_e$ & $E_b$ & $\omega$ \\
\hline
Cu$_2$ & PBE & 2.21 & -1.13 & 269 \\
& BLYP & 2.23 & -1.11 & 261 \\
& revPBE & 2.24 & -1.02 & 259 \\
& PW & 2.15 & -1.39 & 299 \\
& Expt. & 2.21~\cite{22} & -1.04~\cite{23} & 265~\cite{23} \\
Ag$_2$ & PBE & 2.57 & -0.91 & 180 \\
& BLYP & 2.60 & -0.83 & 172 \\
& revPBE & 2.60 & -0.79 & 171 \\
& PW & 2.48 & -1.45 & 270 \\
& Expt. & 2.53~\cite{24} & -0.83~\cite{25} & 192~\cite{24} \\
CuPd & revPBE & 2.32 & -0.79 & 232 \\
AgPd & PBE & 2.56 & -0.70 & 179 \\
Pd$_2$ & PBE & 2.49 & -0.49 & 195 \\
& Expt. & 2.48~\cite{28} & -0.51~\cite{29} & 210~\cite{30} \\
\end{tabular}
\end{ruledtabular}
\end{table}
For better understanding of the interatomic bonds in our target systems,
we calculated the Pd$_2$, CuPd, and AgPd dimers and
presented their results in table \ref{dimer}.
It is observed that the bond length is increasing in
the Cu$_2$ - CuPd - Pd$_2$ series while the absolute binding energy and
vibrational frequency is decreasing.
Hence, Cu-Cu bond has a larger strength and stiffness compared
with the Cu-Pd and Pd-Pd bonds.
Bond length and normal mode frequencies of AgPd and Ag$_2$ dimers
are almost equal while these are larger than those of Pd$_2$ dimer.
The larger bond length of Ag$_2$ may be due to the
large atomic radius of Ag relative to Pd and Cu atoms.
Binding energy of Pd$_2$ is smaller than that of AgPd
which in turn is smaller than Ag$_2$ dimer.
The Mulliken population analysis of valence orbitals shows that
Cu and Ag have a similar electronic configuration of $s^{0.94}p^{0.05}d^{10.01}$
in the Cu$_2$ and Ag$_2$ dimers.
Comparing this configuration with the free Cu and Ag atoms
electronic configuration ($s^{1}p^{0}d^{10}$) evidences
a small $spd$ hybridization in these dimers.
On the other hand, comparing the electronic configuration of Pd in
the Pd$_2$ dimer ($s^{0.58}p^{0.05}d^{9.37}$) with
the free Pd atom ($s^{0}p^{0}d^{10}$) indicates
a significant $spd$ hybridization in the palladium dimer.
The reason is that the valence shell of the free Pd atom is composed
of fully occupied and fully unoccupied orbitals and hence interatomic
bonding between palladium atoms requires significant promotion
of electrons from the occupied $d$ states to the unoccupied $s$ shell.
While, the valence shell of the free Cu and Ag atoms has a half filled
$s$ orbital which is used for interatomic bonding.
It is also found that the charge transfer in the Pd-Cu and Pd-Ag bonds
happens from Cu and Ag to the more electronegative Pd atom.
\section{Stable isomers}
In order to investigate properties of atomic clusters, the first essential step
is identification of the lowest energy structure of the clusters.
Therefore, we performed a careful search for the stable structure of the pure
and doped clusters.
In the case of the pure clusters, all probable atomic configurations of
the clusters were included in our search.
After accurate atomic relaxation of the relevant configurations and comparing
their minimized total energies, the most stable isomers of the pure
Cu$_n$ and Ag$_n$ clusters ($n\le9$) were identified and sketched in Fig.~\ref{isomer}.
These findings are in agreement with a recent study on silver clusters \cite{31}.
Because of the presence of two different atoms, the process of
finding the global minimum energy structure of the doped clusters
is more complicated than the pure clusters.
Therefore, for finding the most stable structures of the palladium doped clusters,
we tried to limit our search to the more stable configurations
proposed by previous studies on the Cu$_n$Pd$_m$ ($n+m\leq6$)\cite{32}
and Ag$_n$Pd$_m$ ($n+m\leq5$)\cite{33,Papageo} clusters.
The resulting most stable isomer of the doped clusters are
presented in Fig.~\ref{isomer}.
It is seen that the most stable M$_{n-2}$Pd$_2$ clusters are
made of the maximum number of pyramids and moreover
the two Pd atoms in these systems tend to bond together.
These observations were used for obtaining the most stable
structure of the M$_6$Pd$_2$ and M$_7$Pd$_2$ clusters.
\begin{figure}
\includegraphics*[scale=0.5]{fig1-str}
\caption{\label{isomer}
Obtained lowest energy isomers of the pure Ag$_n$ and Cu$_n$
and doped Cu$_{n-1}$Pd, Ag$_{n-1}$Pd, Cu$_{n-1}$Pd$_2$,
and Ag$_{n-1}$Pd$_2$ clusters for $3\leq n\leq9$.
The Pd atoms are shown by light balls while the Cu/Ag atoms
are indicated by black balls.
}
\end{figure}
The results show that up to the size of 6 (heptamer), pure clusters prefer
2 dimensional (2D) planar structures while larger clusters
stabilize in 3D geometries.
In order to understand the origin of this 2D-3D cross over,
we analyzed the average bond length ($d$) and the average
coordination number of atoms ($n_c$) in the most stable 2D and 3D
isomers of the Cu$_n$ and Ag$_n$ clusters (Fig.~\ref{cross}).
It is seen that the 2D isomers have lower average coordination and
shorter average bond length, because the valence charge density of
these systems is distributed among lower number of bonds,
compared with the 3D isomers.
As a result of that, individual bonds in the planner isomers are stronger and shorter,
while 3D isomers have more number of bonds.
The 2D-3D structural transformation reflects the competition between
the individual bond strength and the total number of bonds.
It is observed that (Fig.~\ref{cross}) in the smaller sizes ($n\le6$),
the more pronounced difference in the individual bond strength leads to
the stability of the 2D isomers, while for the larger clusters
the difference in the average coordination wins the competition and
stabilizes the 3D isomers.
\begin{figure}
\includegraphics*[scale=0.95]{fig2-cross}
\caption{\label{cross}
Average bond length $d$ and coordination number of atoms $n_c$
in the most stable 2D and 3D structures of Cu$_n$ (left) and
Ag$_n$ (right) clusters.
}
\end{figure}
The results (Fig.~\ref{isomer}) indicate that the most stable 2D isomers of
Cu$_n$ and Ag$_n$ clusters are made of equilateral triangles.
It is also seen that the 2D-3D cross over in the Ag$_{n-1}$Pd
clusters occurs at $n=6$ while the Cu$_{n-1}$Pd clusters,
follow the same structural transformation as the pure Cu cluster.
As a general trend, we observe that the Pd atom prefers the high coordinated
sites of the Cu$_{n-1}$Pd and Ag$_{n-1}$Pd clusters.
The obtained results indicate high tendency of the M$_{n-2}$Pd$_2$
clusters toward 3D configurations.
Considering various 2D and 3D structures for the M$_2$Pd$_2$, it was found that
the 3D structures have lower energies, in agreement with the previous works \cite{32,33}.
Based on these results, the theoretical search for the stable isomers of
the M$_{n-2}$Pd$_2$ clusters were limited to all 3D isomers of the pure Ag and Cu clusters.
It is worth while to mention that other authors have not investigated
Cu$_{n-2}$Pd$_2$ and Ag$_{n-2}$Pd$_2$ clusters larger
than five and six atoms, respectively.
\section{Structural properties}
In order to address stability and structural behaviour of the clusters,
the binding energy per atom, average coordination number of atoms,
and the average bond length of the most stable structures of all clusters
under study as a function of the cluster size are presented in Fig. \ref{bond}.
The harmonic vibrational frequencies of the clusters were also calculated
to address dynamical stability and bond stiffness of the systems.
The maximum vibrational frequency of the clusters are shown in Fig. \ref{bond}.
The calculated binding energy of the copper clusters are in well agreement
with the measured data.
It can be seen that by increasing the cluster size,
the binding energy of the pure clusters increases gradually toward
the binding energy of bulk copper (3.50 eV)\cite{35} and silver (2.95 eV)\cite{35},
although these clusters are clearly far from the bulk limit.
Notably, the eight atom pure clusters occurs in a local maximum
of the absolute binding energy.
It is clearly seen that the 2D to 3D structural transformation in
the pure and mono-doped clusters is reasonably accompanied by a sharp
increases in the average coordination number of atoms.
Observation of no imaginary frequency indicates the dynamical stability of
the lowest energy isomers shown in Fig. \ref{isomer}.
It is generally seen that the Ag based clusters have lower
vibrational frequency than Cu based clusters, which is clearly
related to the higher atomic mass of silver.
Considering the trend of maximum frequencies along with the geometry of the most stable
isomers of the clusters (Fig. \ref{isomer}) clarifies that the structural cross over
from planner to 3D geometries in the pure Ag and Cu clusters and doped Cu$_{n-1}$Pd cluster
is accompanied with a significant mode softening in the vibrational spectra.
Comparing the properties of the pure and doped clusters show that
Pd doping in the copper clusters increases the average
bond length of the system, because of the larger atomic radius
of Pd compared with Cu.
This increase is expected to be accompanied by a bond softening effect,
which is generally visible in the calculated harmonic vibrational
frequency of the doped Cu clusters.
In contrast to this bond elongating and softening effects,
we observe that Pd doping has a week effect on the binding energy of Cu clusters.
It seems that the doped Pd atom in the Cu cluster, inserts effective
amount of electrons in the neighboring bonds and hence compensate
the observed bond softening effects.
It is reasonable, because Pd has much more electrons than Cu.
On the other hand, the average bond length of the doped Ag clusters
is slightly smaller than those of the pure clusters.
The reason is that atomic size of silver is larger than that of palladium
and hence Ag-Pd and Pd-Pd bonds are shorter than Ag-Ag one (table~\ref{dimer}).
As a result of that, a bond hardening effect is visible in the calculated
vibrational spectra of the doped Ag clusters, compared with the pure ones.
As a result of these bond shortening and hardening effects,
Pd doping enhances the absolute binding energy of Ag clusters (Fig.~\ref{bond}).
\begin{figure}
\includegraphics*[scale=0.95]{fig3-bond}
\caption{\label{bond}
Calculated maximum vibrational frequency ($\omega$),
Second difference of the energy $\Delta_{2}E(n)$,
absolute binding energy per atom $E_b$,
average coordination number of atoms $n_c$,
and the average bond length $d$ of the pure,
mono-doped, bi-doped Ag and Cu clusters
as a function of size.
The empty circles show the measured binding energy
of the pure Cu clusters \cite{Eb-Cu}.
}
\end{figure}
In cluster science, the second-order energy difference ($\Delta_{2}E$)
is a conventional quantity to address the relative stability of clusters
as a function of size and reproduce the experimental magic numbers
observed in mass spectra measurements.
Large values of $\Delta_{2}E(n)$ indicate that the $n$-atoms cluster
is more stable than $n-1$ and $n+1$ atoms clusters.
This parameter is defined as follows:
\begin{equation}
\Delta_{2}E(n) = E_{tot}(n+1) + E_{tot}(n-1) - 2 E_{tot}(n)
\end{equation}
$E_{tot}(n)$ is the minimized energy of the M$_{n-m}$Pd$_m$ cluster
(M = Ag, Cu) and $E_{tot}(n+1)$ ($E_{tot}(n-1)$) is the minimized
energy of the cluster with one more (less) M atom.
The calculated second-order energy difference of the pure and
doped Cu and Ag clusters is plotted in Fig. \ref{bond}.
It is seen that pure clusters show strong peaks at $n=2,8$ which
indicate the highest relative stability of these systems and
propose $n=2,8$ as the first magic numbers of small pure Ag and Cu clusters,
in agreement with the experimental observations
\cite{exper1,exper2,exper3,exper4}
and electronic shell jellium model \cite{jellium1,jellium2}.
The predicted magic numbers for mono-doped clusters are 3 and 7,
while in the case of bi-doped clusters, the main peaks are observed at $n=3,4,8$.
Furthermore, a clear odd-even oscillation is seen in the pure
and mono-doped clusters, although bi-doped clusters
do not display such clear oscillations.
It is worthwhile to mention that electron pairing in HOMO can explain this
oscillations and the more stability of the even pure and odd mono-doped clusters.
The valence electrons of the pure M$_n$ and mono-doped M$_{n-1}$Pd clusters
are mainly composed of the M $s$ electrons,
hence clusters with odd number of M atoms have one unpaired electron
and consequently exhibit lower stability.
The pure and mono-doped clusters with even number of M atoms have
a close electronic shell and hence are more stable.
On the other hand, the behaviour of the $\Delta_{2}E$ curve of the bi-doped clusters
is more complicated and needs to further arguments.
We attribute this more complicated behaviour to the effective contribution
of $d$ electrons to the valence shell of the bi-doped clusters.
As it was argued in section III,
the direct bonding between Pd atoms in the bi-doped clusters
give rises to effective promotion of the Pd d electrons to
the valence shell of the system.
\begin{figure}
\includegraphics*[scale=0.93]{fig4-band}
\caption{\label{band}
Calculated HOMO-LUMO gap, Ionization potential (IP),
and electron affinity (EA)
of the pure, mono-doped, and bi-doped
Ag (left column) and Cu clusters (right column) after the GW correction.
The empty circles show the available measured data
for pure clusters, collected from references
\cite{Ip-Ag, EA-Ag, EA-Cu1, EA-Cu2, IP-Cu1,IP-Cu2}.
}
\end{figure}
\section{Electronic properties}
Since the single particle Kohn-Sham eigenvalues have no clear physical meaning,
the reliable energy level of the highest occupied (HOMO) and the lowest
unoccupied molecular orbital (LUMO) of the clusters are determined
after applying the many-body GW perturbation theory \cite{gw}.
The GW corrected HOMO and LUMO energies are expected to be
comparable to the experimental ionization potential (IP)
and electron affinity (EA) of the clusters.
These values along with the computed HOMO-LUMO gaps are presented in Fig.~\ref{band}.
The available measured data for pure clusters are also presented in this figure.
The calculated values of IP and Gap exhibit good agreement with
the experimental data which is mainly due to the GW correction,
while the calculated EA values in some cases are considerably overestimated.
In order to show the effect of this many body correction,
we have compared in Fig. \ref{gap} the experimental values of the HOMO-LUMO gap
of the pure clusters with the calculated ones before and after
application of the GW correction.
It is seen that the Kohn-Sham gaps are significantly underestimated
while the GW corrected gaps are generally much closer to experiment.
The observed odd-even oscillations in the IP and EA of the pure clusters are
attributed to the electron pairing in these systems, discussed in the previous section.
The lower value of EA for even clusters indicates less tendency of these closed shell systems
to receive an extra electron, compared with the open shell clusters.
A consistent reversed trend is visible in the obtained ionization potentials.
\begin{figure}
\includegraphics*[scale=0.93]{fig5-gap}
\caption{\label{gap}
Calculated HOMO-LUMO gap of the pure Ag and Cu clusters
before and after the GW correction
along with the available experimental data,
collected from references
\cite{Ip-Ag, EA-Ag, EA-Cu1, EA-Cu2, IP-Cu1,IP-Cu2}.
}
\end{figure}
In the same way, the ionization potential and electron affinity of the most
stable structure of the doped clusters were calculated and displayed in Fig.~\ref{band}.
The higher value of IP for the odd M$_{n-1}$Pd clusters indicates their more
chemical stability compared to the neighboring even clusters.
In the bi-doped clusters, the IP and EA curves do not show clear oscillations
with respect to cluster sizes, which is likely due to the effective
contribution of $d$ electrons to the Pd-Pd bonding in these clusters.
The trend of the calculated HOMO-LUMO gap is generally consistent with
the $\Delta_{2}E$, indicating that the systems with higher gap are
generally more stable.
It is consistent with the chemical intuition that a higher value of
the HOMO-LUMO gap indicates less chemical activity and hence more stability.
Considering the spin polarized electronic structure of the most stable isomers,
we found out that consistent with the simple spin pairing rule,
all studied clusters prefer the lowest possible spin multiplicity.
Hence, the pure and doped clusters with even number of Ag and Cu atoms
are nonmagnetic while others exhibit a total spin moment of 1 $\mu_B$.
\section{Conclusions}
In this paper, spin-polarized all-electron calculations were performed
to investigate structural and electronic properties of the pure
and Pd doped Cu$_n$ and Ag$_n$ clusters ($n\le9$).
It was argued that a 2D to 3D structural cross over occurs in the size
of seven of the pure clusters. The magic numbers of the pure clusters
were found to be 2 and 8.
The obtained results indicate that the doped Pd atoms
generally prefer the high coordinated sites of the pure clusters.
We found that the magic numbers of the mono-doped clusters are
3 and 7 while bi-doped clusters exhibit magic numbers 3, 4, and 8.
It was argued that doping with one Pd atom slightly reduces the onset
of the 2D to 3D structural crossover of the pure Ag clusters,
while doping with two Pd atoms strongly enhances stability
of the 3D isomers and completely destroys this cross over.
Moreover, it was observed that the bond stiffness of the Cu clusters
are decreased after doping with Pd atoms while their
average bond strength have lower sensitivity to doping.
On the other hand, Pd doping increases the bond stiffness
and bond strength of the Ag clusters.
It was discussed that electron pairing in the HOMO,
induces a significant odd-even fluctuation in the ionization potential,
electron affinity, and HOMO-LUMO gap of the pure Ag and Cu clusters,
while doping with Pd atoms considerably weakens this fluctuation.
\section*{Acknowledgments}
This work was jointly supported by the Vice Chancellor
for Research Affairs of Isfahan University of Technology (IUT),
Centre of Excellence for Applied Nanotechnology,
and ICTP Affiliated Centre at IUT.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.