\documentclass[ALCO,ThmDefs,epreuves,Unicode,published]{cedram}
\OneNumberAllTheorems
\usepackage{tikz}
\usetikzlibrary{patterns, matrix}
\tikzstyle{NE-lines}=[pattern=north east lines, pattern color=black!45]
% \pattern{ options to tikzpicture }{ |\pi| }{ pi }{ R }
\newcommand{\pattern}[4]{
\raisebox{0.6ex}{
\begin{tikzpicture}[scale=0.3, baseline=(current bounding box.center), #1]
\foreach \x/\y in {#4}
\fill[NE-lines] (\x,\y) rectangle +(1,1);
\draw (0.01,0.01) grid (#2+0.99,#2+0.99);
\foreach \x/\y in {#3}
\filldraw (\x,\y) circle (5pt);
\end{tikzpicture}}
}
\newcommand{\hpii}[2]{
\pattern{}{2}{1/#1,2/#2}
{ 0/1,
1/0,1/1,1/2,
2/1}
}
\newcommand{\hpiv}[4]{
\pattern{}{4}{1/#1,2/#2,3/#3,4/#4}
{ 0/1,0/2,0/3,
1/0,1/1,1/2,1/3,1/4,
2/0,2/1,2/2,2/3,2/4,
3/0,3/1,3/2,3/3,3/4,
4/1,4/2,4/3}
}
\newcommand{\eps}{\epsilon}
\newcommand{\NN}{\mathbb{N}}
\newcommand{\ZZ}{\mathbb{Z}}
\newcommand{\QQ}{\mathbb{Q}}
\newcommand{\etal}{et~al.}
\newcommand{\Sym}{\mathcal{S}}
\newcommand{\Mesh}{\mathcal{M}}
\newcommand{\Pow}{\mathcal{P}}
\newcommand{\id}{\mathrm{id}}
\newcommand{\rid}{\overline{\id}}
\newcommand{\from}{\leftarrow}
\newcommand{\tostar}{\stackrel{*}{\to}}
\newcommand{\toplus}{\stackrel{+}{\to}}
\newcommand{\fromstar}{\stackrel{*}{\from}}
\newcommand{\nf}[1]{{#1\!\downarrow}}
\newcommand{\floor}[1]{\lfloor#1\rfloor}
\newcommand{\bracket}[1]{[\mskip1mu #1\mskip1mu]}
\newcommand{\fsum}[1]{\sum_{m\geq 0}m!\left(#1\right)^{m}}
\newcommand{\cC}{\mathcal{C}}
\newcommand{\uu}{\boldsymbol{u}}
\newcommand{\one}{\boldsymbol{1}}
\DeclareMathOperator{\olap}{\mathcal{O}}
\DeclareMathOperator{\dom}{dom}
\DeclareMathOperator{\st}{st}
\DeclareMathOperator{\dd}{d}
%%%%%%%%% a placer avant \begin{document}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\newcommand*{\mk}{\mkern -1mu}
\newcommand*{\Mk}{\mkern -2mu}
\newcommand*{\mK}{\mkern 1mu}
\newcommand*{\MK}{\mkern 2mu}
\hypersetup{urlcolor=purple, linkcolor=blue, citecolor=red}
\newcommand*{\romanenumi}{\renewcommand*{\theenumi}{\roman{enumi}}}
\newcommand*{\Romanenumi}{\renewcommand*{\theenumi}{\Roman{enumi}}}
\newcommand*{\alphenumi}{\renewcommand*{\theenumi}{\alph{enumi}}}
\newcommand*{\Alphenumi}{\renewcommand*{\theenumi}{\Alph{enumi}}}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%% Auteur
\author{\firstname{Anders} \lastname{Claesson}}
\address{Science Institute\\
University of Iceland\\
Iceland}
\email{akc@hi.is}
%%%%% Sujet
\keywords{Hertzsprung's problem, cluster method,
pattern, permutation, rewriting system.}
\subjclass{05A05, 05A15, 68R05, 68R15, 68Q42}
%%%%% Gestion
%\DOI{10.5802/alco.202}
%\datereceived{2021-04-07}
%\daterevised{2021-10-27}
%\dateaccepted{2021-10-28}
%%%%% Titre et résumé
\title[On Hertzsprung's problem]{From Hertzsprung's problem to pattern-rewriting systems}
\begin{abstract}
Drawing on a problem posed by Hertzsprung in 1887, we say that a given
permutation $\pi\in\mathcal{S}_n$ contains the Hertzsprung pattern
$\sigma\in\mathcal{S}_k$ if there is a factor $\pi(d+1)\pi(d+2)\cdots\pi(d+k)$
of $\pi$ such that $\pi(d+1)-\sigma(1) =\cdots =
\pi(d+k)-\sigma(k)$. Using a combination of the Goulden--Jackson
cluster method and the transfer-matrix method we determine the joint
distribution of occurrences of any set of (incomparable) Hertzsprung
patterns, thus substantially generalizing earlier results by
Jackson~et~al. on the distribution of ascending and descending runs in
permutations. We apply our results to the problem of counting
permutations up to pattern-replacement equivalences, and using
pattern-rewriting systems -- a new formalism similar to the much
studied string-rewriting systems -- we solve a couple of open problems
raised by Linton~et~al. in 2012.
\end{abstract}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\datepublished{2022-12-19}
\begin{document}
\maketitle
\section{Introduction}
Severin Carl Ludvig Hertzsprung (1839--1893) was a Danish senior civil
servant with a graduate degree in astronomy from the University of
Copenhagen. In 1887 a letter authored by him titled ``En
kombinationsopgave'' appeared in Tidsskrift for
mathematik~\cite{Hertzsprung1887}. The first paragraph reads
\begin{quote}
\emph{At bestemme Antallet af Maader, hvorpaa Tallene
$1$, $2$, $3$, $4$,\dots, $n$ kunne opstilles i Række saaledes, at
Differensen mellem to ved Siden af hinanden staaende Tal overalt skal
være numerisk forskjellig fra 1.}
\end{quote}
Or, translated, the problem is to determine the number of ways in which
the numbers $1$, $2$, $3$, $4$, \dots, $n$ can be arranged such that the
difference between two adjacent numbers is different from~1. For $n=4$
there are only two such arrangements, namely $2413$ and $3142$. An
alternative interpretation is that we are asked for the number of ways
$n$ kings can be placed on an $n$ by $n$ chessboard, one on each row and
column, so that no two attack each other. Indeed,
Kaplansky~\cite{Kaplansky1944} -- who rediscovered Hertzsprung's problem
in 1944 -- called it the $n$-kings problem.
In Hertzsprung's solution -- which in his own words is only slightly
elegant -- the problem is first generalized by defining $u_{n,k}$ as the
number of arrangements such that the numbers $1$, $2$, \dots, $k$ satisfy
the requirement that two numbers whose difference is $1$ cannot be
placed next to each other. Hertzsprung derives a recurrence relation for
the numbers $u_{n,k}$ to which he guesses a solution and subsequently
verifies that it indeed satisfies the recurrence. The solution to the
original problem is then given by
\[
u_{n,n} \,=\, n! + \sum_{k=1}^n (-1)^k \sum_{i=1}^k
\binom{k-1}{i-1}\binom{n-k}{i} 2^i (n-k)!\,.
\]
The first few numbers of this sequence are $1$, $1$, $0$, $0$, $2$,
$14$, $90$, $646$, $5242$, $47622$, which is A002464 in the
OEIS~\cite{OEIS}. A more elegant, but less explicit, solution (see \eg
p.~737 of Flajolet and Sedgewick~\cite{Flajolet2009}) is obtained using
generating functions: $u_{n,n}$ is the coefficient of $x^n$ in
the expansion of
\begin{equation}\label{Hertzsprung-gf}
\fsum{\frac{x-x^2}{1+x}}.
\end{equation}
A reader familiar with mesh patterns~\cite{BrCl2011} may have noticed
that Hertzsprung's problem is to count permutations in $\Sym_n$ that
avoid the two patterns
\[
\hpii{1}{2}\quad\text{and}\quad\hpii{2}{1}\,.
\]
We will use the following terminology. Let $\pi$ be a word (\eg a
permutation written in one line notation). A word $\beta$ is said to be
a \emph{factor} of $\pi$ if there are words $\alpha$ and $\gamma$ such
that $\pi=\alpha\beta\gamma$. If $\alpha$ is empty we also say that
$\beta$ is a \emph{prefix} of $\pi$. If, in addition, $\gamma$ is
nonempty then $\beta$ is said to be a \emph{proper prefix} of $\pi$.
Similarly, if $\gamma$ is empty we say that $\beta$ is a \emph{suffix}
of $\pi$, and that $\beta$ is a \emph{proper suffix} if, in addition,
$\alpha$ is nonempty.
Let $\tau\in \Sym_k$ and $\pi\in\Sym_n$. We call $\tau$ a
\emph{Hertzsprung factor} of $\pi$ if there is an integer $c$ and a
factor $\beta=b_1b_2\dots b_k$ of $\pi$ such that $\tau(i)-b_i=c$ for
each $i\in [k]$. In this context we also say that $\beta$ is an
\emph{occurrence} of $\tau$ and sometimes we write $\beta\simeq\tau$.
As an example, $213$ is a Hertzsprung factor of $\pi=1546372$; indeed,
$546\simeq 213$ is an occurrence of $213$ in $\pi$. We similarly define
\emph{Hertzsprung prefix} and \emph{Hertzsprung suffix}.
Another way to say that $\tau$ is a Hertzsprung factor of $\pi$ is that
$\tau$ appears inside $\pi$ via a subsequence $\beta$ which forms a
consecutive set both horizontally (the positions of~$\beta$) and
vertically (the values of~$\beta$). Thus, $\tau$ is a Hertzsprung
factor of $\pi$ if $\pi$ contains the mesh pattern $(\tau,H_k)$, where
$H_k$ is the mesh $\{0,1,\dots,k\}^2\setminus\{0,k\}^2$. For instance,
$2341$ is a Hertzsprung factor of $\pi$ if and only if $\pi$ contains
the mesh pattern
\[
(2341, H_4) = \hpiv{2}{3}{4}{1}\,.
\]
Mesh patterns of the form $(\tau,H_k)$ will be referred to as
\emph{Hertzsprung patterns}. Since all mesh patterns in this paper will
be of this form we will, by slight abuse of notation, identity the
pattern $(\tau,H_k)$ with $\tau$. While the Hertzsprung terminology is
original with this paper, there are a number of results on these
patterns under different names in the literature. Monotone Hertzsprung
patterns have the longest history and are usually called (increasing or
decreasing) runs in the
literature~\cite{Abramson1967,Jackson1977,Jackson1978,Jackson1976,Riordan1945}.
Myers~\cite{Myers2002} seem to have been the first to study non-monotone Hertzsprung
patterns. She calls them rigid patterns. Bóna~\cite{Bona2008} talks of
permutations very tightly avoiding or containing a pattern. Let us now
detail some of these results.
The identity permutation (as well as the corresponding Hertzsprung
pattern) will be denoted by $\id_k=12\dots k$. Its reverse will be
denoted $\rid_k=k\dots 21$. Jackson and Read~\cite{Jackson1978} showed
that the generating functions for respectively $\id_k$-avoiding and
$\{\id_k,\rid_k\}$-avoiding permutations are
\[
\fsum{\frac{x-x^k}{1-x^k}}\quad\text{and}\quad
\fsum{\frac{x-2x^k+x^{k+1}}{1-x^k}}.
\]
The special case $k=2$ in the latter formula is the generating
function~\eqref{Hertzsprung-gf}.
Some Hertzsprung patterns $\tau\in\Sym_k$ can overlap with themselves in
the sense that there is a permutation $\sigma$ of length less than $2k$
such that $\tau$ is both a proper Hertzsprung prefix and a proper
Hertzsprung suffix of $\sigma$. Myers~\cite{Myers2002} calls such patterns
extendible, we will call them \emph{self-overlapping}. The monotone patterns,
$\id_k$ and $\rid_k$, are both self-overlapping. The pattern $2143$ is also
self-overlapping, while $132$ is not. Myers derives a formula for
counting the number of permutations of $[n]$ with a prescribed number
$m$ of occurrences of any single non-self-overlapping pattern
$\tau\in\Sym_k$:
\begin{equation}\label{Myers-formula}
\sum_i(-1)^{m-i}\binom{i}{m}\binom{n-(k-1)i\mskip1mu}{i}\bigl(n-(k-1)i\bigr)!\,.
\end{equation}
Let $\Sym_n(\tau)$ denote the set of permutations in $\Sym_n$ that avoid
the Hertzsprung pattern $\tau$. Bóna~\cite{Bona2008} showed that
$|\Sym_n(\tau)|\leq |\Sym_n(\id_k)|$ for each non-self-overlapping pattern
$\tau$. He also explains why most patterns are non-self-overlapping. Both
Myers and Bóna ask if it is possible to determine $|\Sym_n(\tau)|$ for
any self-overlapping pattern $\tau$ other than $\id_k$ and $\rid_k$.
Let $T$ be an \emph{antichain} of Hertzsprung patterns. That is, every
pair of distinct patterns in $T$ are incomparable in the sense that one
is not a Hertzsprung factor of the other. Let
$\Sym_n(T)=\cap_{\tau\in T}\Sym(\tau)$ be the set of \emph{$T$-avoiding
permutations} in $\Sym_n$. In Section~\ref{enumeration} we show that
there is a rational function $R(x)\in \QQ(x)$ such that
\[
\sum_{n\geq 0} |\Sym_n(T)|\mskip1mu x^n = \sum_{m\geq 0}m!R(x)^{m}.
\]
Moreover, there is an efficient way of determining $R(x)$ from the set
$T$. We in fact show something stronger.
It is known~\cite{Jackson1977, Jackson1976} that the number of
permutations of $[n]$ with exactly $\ell$ occurrences of $\id_k$ is the
coefficient of $u^{\ell}x^n$ in
\begin{equation}\label{id-jackson}
\sum_{m\geq 0}m!x^m\left(\frac{1-ux-(1-u)x^{k-1}}{1-ux-(1-u)x^{k}} \right)^m.
\end{equation}
Suppose that $T=\{\tau_1,\tau_2,\dots,\tau_k\}$ is an antichain of
Hertzsprung patterns. We show that there is a rational function
$R(u_1,\dots,u_k; x)$ in $\QQ(u_1,\dots,u_k, x)$ such that
\[
\sum_{\pi\in\Sym} u_1^{\tau_1(\pi)}u_2^{\tau_2(\pi)}\cdots\mskip1mu u_k^{\tau_k(\pi)} x^{|\pi|}
= \sum_{m\geq 0}m!R(u_1,u_1,\dots,u_k; x)^{m},
\]
where $\Sym=\cup_n\Sym_n$ and $\tau_i(\pi)$ denotes the number of
occurrences of $\tau_i$ in $\pi$.
In Section~\ref{PRS} we apply the machinery developed in
Section~\ref{enumeration} to the problem of counting permutations up to
certain pattern-replacement equivalences introduced by Linton, Propp,
Roby and West~\cite{Linton2012}. We provide a new formalism called
pattern-rewriting systems, which are similar to the much studied
string-rewriting systems. Using these we solve a couple of open problems
posed by Linton~\etal
\section{Enumeration}\label{enumeration}
Let $T$ be an antichain of Hertzsprung patterns. We will apply the
Goulden--Jackson cluster method~\cite{Goulden1979} to count permutations
with respect to the number of occurrences of patterns in $T$. The
context of the original formulation of this method is the free monoid
over a finite set. The method has, however, been successfully adopted
and applied to permutations by Goulden and Jackson themselves and more
recently by Dotsenko and Khoroshkin~\cite{Dotsenko2013} and Elizalde and
Noy~\cite{Elizalde2012}.
A \emph{marked permutation} is a pair $(\pi,M)$ where $\pi$ is a
permutation and $M$ is a subset of all occurrences in $\pi$ of patterns
from $T$. The members of $M$ are called \emph{marked occurrences} and
$\pi$ is called the \emph{underlying permutation} of $(\pi,M)$. As an
example, let $T=\{123\}$, $\pi=1234567\in \Sym_7$ and $M=\{123,234,567\}$. The
marked permutation $(\pi,M)$ is then depicted below.
\[
\begin{tikzpicture}[xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (2) {$2$};
\node at (3,0) (3) {$3$};
\node at (4,0) (4) {$4$};
\node at (5,0) (5) {$5$};
\node at (6,0) (6) {$6$};
\node at (7,0) (7) {$7$};
\draw (2) ellipse (1.43 and 0.3);
\draw (3) ellipse (1.43 and 0.3);
\draw (6) ellipse (1.43 and 0.3);
\end{tikzpicture}
\]
Let $(\pi,M)$ be a marked permutation such that $|\pi|\geq 1$ and $M$
covers $\pi$ in the sense that each letter of $\pi$ is in some
occurrence. We say that $\alpha,\beta\in M$ \emph{overlap} if they have
at least one letter in common. Clearly, being overlapping is a symmetric
relation and we can view $M$ as an undirected graph with edge set
consisting of all $(\alpha,\beta)$ such that $\alpha$ and $\beta$
overlap. If this graph is connected, then we say that $(\pi, M)$ is a
\emph{$T$-cluster}, or simply a \emph{cluster} when $T$ is known from
context. The marked permutation in the example above is not a
cluster, but \hspace{-5pt}
\begin{tikzpicture}[baseline=(1.base), xscale=0.25, yscale=0.5]
\node at (1,0) (1) {$\scriptstyle 1$};
\node at (2,0) (2) {$\scriptstyle 2$};
\node at (3,0) (3) {$\scriptstyle 3$};
\node at (4,0) (4) {$\scriptstyle 4$};
\draw (2) ellipse (1.6 and 0.4);
\draw (3) ellipse (1.6 and 0.4);
\end{tikzpicture}\hspace{-1pt} is.
Let $\pi$ be a permutation of $[k]$ and let
$\sigma_1,\sigma_2,\dots,\sigma_k$ be nonempty permutations. The
\emph{inflation} (see Albert and Atkinson~\cite{Albert2005}) of $\pi$ by
$\sigma_1,\sigma_2,\dots,\sigma_k$ is
\[
\pi[\sigma_1,\sigma_2,\dots,\sigma_k] =
\sigma_1'\sigma_2'\dots\sigma_k',\]
where $\sigma_{\pi^{-1}(i)}'$ is
obtained from $\sigma_{\pi^{-1}(i)}$ by adding the constant
$|\sigma_{\pi^{-1}(1)}|+|\sigma_{\pi^{-1}(2)}|+\dots+|\sigma_{\pi^{-1}(i-1)}|$
to each of its letters. An example should make this clear,
\begin{equation*}
231[1,213,21] \,=\, 3\; 546\; 21.
\end{equation*}
Note that $\sigma_i'$ is an occurrence of $\sigma_i$ in
$\pi[\sigma_1,\sigma_2,\dots,\sigma_k]$ by construction. The inflation
operation naturally generalizes to marked permutations. Here is a
permutation with some marked occurrences of $123$ and $132$:
\begin{multline}\label{big-marked-perm}
\qquad 213456\bigl[
\begin{tikzpicture}[baseline=(1.base), xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (2) {$2$};
\node at (3,0) (3) {$3$};
\node at (4,0) (4) {$4$};
\node at (5,0) (5) {$5$};
\draw (2) ellipse (1.44 and 0.3);
\draw (3) ellipse (1.44 and 0.3);
\draw (4) ellipse (1.44 and 0.3);
\end{tikzpicture},
\begin{tikzpicture}[baseline=(1.base), xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (2) {$2$};
\node at (3,0) (3) {$3$};
\node at (4,0) (5) {$5$};
\node at (5,0) (4) {$4$};
\draw (2) ellipse (1.44 and 0.3);
\draw (5) ellipse (1.44 and 0.3);
\end{tikzpicture}
,\,1,\,1,
\begin{tikzpicture}[baseline=(1.base), xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (3) {$3$};
\node at (3,0) (2) {$2$};
\draw (3) ellipse (1.44 and 0.3);
\end{tikzpicture}
,1\,\bigr]\\[-0.3ex]
\,=\,
\begin{tikzpicture}[baseline=(6.base), xscale=0.5]
% 12345
\node at (1,0) (6) {$6$};
\node at (2,0) (7) {$7$};
\node at (3,0) (8) {$8$};
\node at (4,0) (9) {$9$};
\node[xshift=-1pt] at (5,0) (10) {$10$};
\draw (7) ellipse (1.5 and 0.3);
\draw (8) ellipse (1.5 and 0.3);
\draw[xshift=1pt] (9) ellipse (1.5 and 0.3);
\end{tikzpicture}
\begin{tikzpicture}[baseline=(1.base), xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (2) {$2$};
\node at (3,0) (3) {$3$};
\node at (4,0) (5) {$5$};
\node at (5,0) (4) {$4$};
\draw (2) ellipse (1.5 and 0.3);
\draw (5) ellipse (1.5 and 0.3);
\end{tikzpicture}
\,11\hspace{3pt}12\,
\begin{tikzpicture}[baseline=(1.base), xscale=0.53]
\node at (1,0) (13) {$13$};
\node at (2,0) (15) {$15$};
\node at (3,0) (14) {$14$};
\draw (15) ellipse (1.55 and 0.33);
\end{tikzpicture}
\,16.\qquad
\end{multline}
It is clear that a marked permutation $(\pi,M)$ is a cluster if and only if its
underlying permutation $\pi$ is nonempty and $(\pi,M)$ cannot be expressed
as the inflation of two or more nonempty marked permutations.
Let $(\tau, M)$ be any nonempty marked permutation with respect to some
antichain $T$ of Hertzsprung patterns. As described above, view $M$ as
an undirected graph whose edge set is
$\{(\alpha,\beta)\in M\times M: \text{$\alpha$ and $\beta$
overlap}\}$. Consider the connected components of this graph. For
instance, with $(\tau, M)$ as in~\eqref{big-marked-perm} there are six
connected components:
\[
\begin{tikzpicture}[baseline=(6.base), xscale=0.5]
% 12345
\node at (1,0) (6) {$6$};
\node at (2,0) (7) {$7$};
\node at (3,0) (8) {$8$};
\node at (4,0) (9) {$9$};
\node[xshift=-1pt] at (5,0) (10) {$10$};
\draw (7) ellipse (1.5 and 0.3);
\draw (8) ellipse (1.5 and 0.3);
\draw[xshift=1pt] (9) ellipse (1.5 and 0.3);
\end{tikzpicture},\;
\begin{tikzpicture}[baseline=(1.base), xscale=0.4]
\node at (1,0) (1) {$1$};
\node at (2,0) (2) {$2$};
\node at (3,0) (3) {$3$};
\node at (4,0) (5) {$5$};
\node at (5,0) (4) {$4$};
\draw (2) ellipse (1.5 and 0.3);
\draw (5) ellipse (1.5 and 0.3);
\end{tikzpicture},\; 11,\; 12,\;
\begin{tikzpicture}[baseline=(1.base), xscale=0.53]
\node at (1,0) (13) {$13$};
\node at (2,0) (15) {$15$};
\node at (3,0) (14) {$14$};
\draw (15) ellipse (1.55 and 0.33);
\end{tikzpicture},\; 16.
\]
Note that the values in any component form a set of consecutive
numbers. Construct a word by picking a single value from each
component (\eg the smallest) and let $\pi$ be its
standardization. By \emph{standardization} we mean the result of
replacing the smallest letter by $1$, the next smallest letter by $2$,
etc. In our example we arrive at $6,1,11,12,13,16$, whose
standardization is $\pi = 213456$. On standardizing the connected
components we arrive at a list $(\sigma_1,M_1)$, $(\sigma_2,M_2)$,
\dots, $(\sigma_k,M_k)$ each part of which is either a $T$-cluster or
the (trivially) marked permutation $(1,\emptyset)$. Returning to our
example, we have $\sigma_1=12345$, $M_1=\{123,234,345\}$,
$\sigma_2=12354$, $M_2=\{123,354\}$, $\sigma_3=1$, $M_3=\emptyset$,
etc. Note that $(\tau,M) = \pi[(\sigma_1,M_1),\dots,(\sigma_k,M_k)]$.
Let us summarize this in a theorem.
\begin{theorem}\label{thm:unique-representation}
Any marked permutation, with respect to an antichain $T$ of
~Hertzsprung patterns, can be uniquely written
\[
\pi\bigl[(\sigma_1,M_1),(\sigma_2,M_2),\dots,(\sigma_k,M_k)\bigr],
\]
where $\pi\in\Sym_k$ and each $(\sigma_i,M_i)$ is either
$(1,\emptyset)$ or a $T$-cluster.
\end{theorem}
For each nonempty permutation $\pi$ let there be an associated
indeterminate $u_\pi$ and let
$\uu=(u_1,u_{12},u_{21},u_{123},u_{132},\dots)$ be the sequence
of all such indeterminates. Let $T$ be an antichain of Hertzsprung
patterns. For any marked permutation $(\pi,M)$, define the monomial
\[
\uu_M = \prod_{\alpha\in M}u_{\st(\alpha)},
\]
where $\st(\alpha)$ denotes the standardization of $\alpha$.
If $M$ is the set of marked occurrences
in~\eqref{big-marked-perm} then $\uu_M=u_{123}^4u_{132}^2$. The
\emph{cluster generating function} associated with $T$ is defined by
\[
C(\uu; x) =
\sum_{(\pi,M)}\mskip-1mu \uu_M\mskip1mu x^{|\pi|},
\]
where the sum is over all $T$-clusters. Similarly, let
$F(\uu; x) = \sum_{(\pi,M)}\mskip-1mu \uu_M\mskip1mu x^{|\pi|}$, where
the sum is over \textit{all} marked permutations. Note that the only difference
between the summations defining $C(\uu; x)$ and $F(\uu; x)$ is which sets are being summed over.
Further, let
\[
\uu^{T(\pi)} = \prod_{\tau\in T}u_{\tau}^{\tau(\pi)},
\]
where $\tau(\pi)$ denotes the number of occurrences of $\tau$ in
$\pi$. Then
\[
\sum_{\pi\in\Sym}x^{|\pi|} (\one+\uu)^{T(\pi)}
= \fsum{x + C(\uu; x)}.
\]
Indeed, the left-hand side is the generating function $F(\uu; x)$ of
marked permutations: for each $\tau\in T$ and for each of the
$\tau(\pi)$ occurrences of $\tau$ in $\pi$ there is a choice to be made,
mark it with a $u_\tau$ or leave it unmarked. That the right-hand side
equals $F(\uu; x)$ follows from Theorem~\ref{thm:unique-representation}.
On replacing $\uu$ by $\uu-\one$ we have the following
result.
\begin{theorem}\label{cluster-thm}
Let $T$ be an antichain of Hertzsprung patterns. Then
\[
\sum_{\pi\in\Sym}\uu^{T(\pi)}x^{|\pi|}
= \fsum{x + C(\uu - \one; x)}.
\]
In particular,
\[
\sum_{n\geq 0}|\Sym_n(T)|\mskip1mu x^n = \fsum{x + C(\mathbf{-1}; x)}.
\]
\end{theorem}
If $T=\{\tau\}$ and $\tau$ is a non-self-overlapping pattern, then the
only $T$-cluster is $(\tau, \{\tau\})$ and $C(u,x)=ux^{|\tau|}$, where
$u = u_{\tau}$. Hence we arrive at the following generating function
version of Myers's formula~\eqref{Myers-formula}.
\begin{coro}
For any non-self-overlapping pattern $\tau$,\smallskip
\[
\sum_{\pi\in\Sym}u^{\tau(\pi)}x^{|\pi|}
\;=\; \fsum{x + (u-1)x^{|\tau|}}.
\]
\end{coro}
\begin{exam}\label{Ex132}
The distribution of the number of
occurrences of $132$ is given by
\begin{align*}
\sum_{\pi\in\Sym}u^{132(\pi)}x^{|\pi|}
&= \fsum{x + (u-1)x^3}.
\end{align*}
Letting $u=0$ we get a generating function for $132$-avoiding
permutations:
\[
\sum_{n\geq 0}|\Sym_n(132)|\mskip1mu x^n =
\fsum{x-x^3}.
\]
It follows that
\begin{equation}\label{Myers132}
|\Sym_n(132)| = \sum_{i=0}^{\floor{n/3}}(-1)^i(n-2i)!\binom{n-2i}{i},
\end{equation}
which is the case $\tau=132$, $m=0$ and $k=3$ of Myers's
formula~\eqref{Myers-formula}.
\end{exam}
\begin{exam}\label{Ex123}
For $T=\{123\}$ the underlying permutation of any cluster is $\id_k$ for
some $k\geq 3$. A cluster of length $k$ is built from clusters of length
$k-2$ or length $k-1$ by marking the suffix $(k-2,k-1,k)$ of
$\id_k$. Formally, if $\cC_k$ denotes the set of clusters whose
underlying permutation is $\id_k$ then $\cC_2=\emptyset$,
$\cC_3=\{(123,\{123\})\}$, and $\cC_k$ consists of all clusters
\[
\bigl(\,\id_k,\, M\cup \{(k-2,k-1,k)\}\,\bigr),
\]
where $(\id_{k-1},M)\in \cC_{k-1}$ or $(\id_{k-2},M)\in \cC_{k-2}$. It follows that
\[
C(u, x) = ux^3 + uxC(u, x) +ux^2C(u, x),
\]
in which $u=u_{123}$ tracks marked factors and $x$ tracks the length
of the cluster. Thus the number of clusters of fixed length is a
Fibonacci number, the generating function for clusters is
$C(u, x) = ux^3/(1-u(x+x^2))$, and we have rediscovered a special
case of~\eqref{id-jackson}:
\begin{align*}
\sum_{\pi\in\Sym}u^{132(\pi)}x^{|\pi|}
&= \fsum{x + C(u-1, x)} \\
&= \fsum{x + \frac{(u-1)x^3}{1-(u-1)(x+x^2)}}.
\end{align*}
\end{exam}
Having seen a couple of examples we now return to the general case. Let
$T$ be any antichain of Hertzsprung patterns. We will show that the
cluster generating function associated with $T$ is counting walks in a
digraph and thus it is rational. This can be seen as an application of
the transfer-matrix method~\cite{EC1}.
For $\sigma,\tau\in T$ define the set of permutations
$\olap(\sigma,\tau)$ by stipulating that $\pi\in \olap(\sigma,\tau)$ if
and only if
\begin{itemize}
\item $\sigma$ is a proper Hertzsprung prefix of $\pi$,
\item $\tau$ is a proper Hertzsprung suffix of $\pi$ and
\item $|\pi| < |\sigma| + |\tau|$.
\end{itemize}
If $\olap(\sigma,\tau)$ is nonempty, then we say that $\sigma$ and
$\tau$ \emph{overlap}. In particular, $\olap(\tau,\tau)$ is nonempty if
and only if $\tau$ is self-overlapping. Let
\[
\Omega(\sigma,\tau) = \sum_{\pi\in \olap(\sigma,\tau)} x^{|\pi|-|\sigma|}.
\]
Construct an edge-weighted digraph $D_T$ with vertices
$V = \{\eps\} \cup T$, edge set $V\times V$ and weight function
$w:V\times V\to \QQ[\uu, x]$ defined by
\[
w(\sigma,\tau) =
\begin{cases}
0 & \text{if $\tau = \eps$,}\\
u_{\tau}x^{|\tau|} & \text{if $\sigma = \eps$,}\\
u_{\tau}\Omega(\sigma,\tau) &\text{otherwise.}
\end{cases}
\]
\begin{exam}\label{Ex123_132}
Let $T = \{123, 132\}$. We have $\olap(132,123) =\olap(132,132)=\emptyset$,
\[\olap(123,123) =\{1234,12345\}\quad\text{and}\quad\olap(123,132) =\{12354\}.
\]
For ease of notation, let $u=u_{123}$ and $v=u_{132}$. The digraph
$D_T$ together with its adjacency matrix are depicted below.
\[
\begin{tikzpicture}[baseline=2.7em]
\node at (0,2) (123) {$123$};
\node at (3,2) (132) {$132$};
\node at (1.5,0) (eps) {$\eps$};
\draw[->] (eps) -- (123) node[midway, left=2pt, yshift=-2pt] {$ux^3$};
\draw[->] (eps) -- (132) node[midway, right=1pt, yshift=-2pt] {$vx^3$};
\draw[->] (123) -- (132) node[midway, above=3pt] {$vx^2$};
\path (123) edge[->,in=160,out=200,looseness=6] node[left] {$ux+ux^2$} (123);
\end{tikzpicture}\qquad
\begin{bmatrix}
\;0 & u x^3 & v x^3\; \\
\;0 & ux+ux^2 & vx^2\; \\
\;0 & 0 & 0\;
\end{bmatrix}
\]
Here, and in what follows, edges that are known to have weight 0 are omitted.
\end{exam}
\begin{theorem}\label{transfer-thm}
Let $T$ be an antichain of Hertzsprung patterns. Let $A$ be the
adjacency matrix of the digraph $D_T$ defined above. Then
\[
C(\uu; x) \,=\,
\frac{1}{\det(1-A)}
\sum_{i=1}^{|T|}(-1)^{i}\det\bigl(1-A: i+1, 1 \bigr),
\]
where $(B:i,j)$ denotes the matrix obtained by removing the $i$th row
and $j$th column of $B$. In particular, $C(\uu; x)$ belongs to the
field of rational functions $\QQ(x)(u_\tau: \tau\in T)$.
\end{theorem}
\begin{proof}
The generating function for (weighted) directed walks in $D_T$
starting at $\eps$ and ending in any other vertex equals $C(\uu;x)$ by
construction of $D_T$. Thus $C(\uu;x)$ is given by summing all but the
first entry of the first row of
\[
(1-A)^{-1} = 1 + A + A^2 + A^3 + \cdots
\]
If $B$ is any invertible matrix, then Cramer's rule gives
\[
(B^{-1})_{ij}=(-1)^{i+j}\det(B;j,i)/\det(B),
\] from which the result follows.
\end{proof}
By applying Theorem~\ref{transfer-thm} to $T=\{132\}$ and $T=\{123\}$ we
find that the cluster generating functions are respectively $ux^3$ and
$ux^3/(1-u(x+x^2))$, which agree with examples~\ref{Ex132}
and~\ref{Ex123}. More generally, we have the following corollary.
\begin{coro}\label{single-pattern}
For any Hertzsprung pattern $\tau$,
\[
\sum_{\pi\in\Sym}u^{\tau(\pi)}x^{|\pi|} =
\fsum{x +
\frac
{(u-1)x^{|\tau|}}
{1-(u-1)\Omega(\tau,\tau)}
}.
\]
\end{coro}
\begin{exam}
Hertzsprung's problem generalized to counting permutations with
respect to the joint distribution of the Hertzsprung
patterns $12$ and $21$ has the solution
\[
\sum_{\pi\in\Sym}u^{12(\pi)}v^{21(\pi)}x^{|\pi|} \;=\;
\fsum{\frac
{x - (u-1)(v-1)x^3}
{(1+x-ux)(1+x-vx)}}.
\]
More generally, the joint distribution of $\id_k$ and $\rid_{\ell}$
is given by
\[
\fsum{x \,+\,
\frac{(u-1)x^k}{1 - (u-1)(\bracket{k}_x-1)} \,+\,
\frac{(v-1)x^\ell}{1 - (v-1)(\bracket{\ell}_x-1)}\,
},
\]
where $\bracket{k}_x=(1-x^{k})/(1-x) = 1+x+\cdots+x^{k-1}$.
Generalizing Hertzsprung's problem in another direction we find that the
generating function for permutations avoiding all patterns in $\Sym_3$
is
\begin{multline*}
\fsum{
\frac
{-2x^6 + 3x^5 - 3x^4 - 5x^3 + x^2 + x}
{x^4 + x^3 - x^2 - x - 1}
} \\
= 1 + x + 2x^{2} + 4x^{4} + 34x^{5} + 298x^{6} + 2434x^{7} + 21374x^{8}\\
+ 205300x^{9} + 2161442x^{10}+ 24804386x^{11} + \cdots
\end{multline*}
\end{exam}
Theorem~\ref{cluster-thm} together with Theorem~\ref{transfer-thm}
solves the problem of determining the distribution of a set of
Hertzsprung patterns. Or, at least they solve the problem as long as the
patterns are short. The weight function is defined in terms of the sets
$\olap(\sigma,\tau)$ and those sets are a priori expensive to compute.
We shall now detail why, in fact, the weight function can be efficiently
computed.
The polynomials $\Omega(\sigma,\tau)$ are akin to the correlation polynomials
of Guibas and Odlyzko~\cite{Guibas1981}. Consider
$\pi\in \olap(\sigma,\tau)$. Let $\delta$ be the factor of $\pi$ in
which its prefix corresponding to $\sigma$ intersects with its suffix
corresponding to $\tau$. The permutation $978563421$ is, for instance, a
member of $\olap(53412, 563421)$ and $\delta=56$. Let
$i=|\delta|$. Note that the smallest element in $\delta$ is either
$|\sigma|-i+1$ or $|\tau|-i+1$. Indeed, either $\sigma$ is a (literal)
prefix of $\pi$ or $\tau$ is a (literal) suffix of $\pi$. In the first
case, $\min\delta=|\sigma|-i+1$. In the second case,
$\min\delta=|\tau|-i+1$. Thus we arrive at the formula
\[
\Omega(\sigma,\tau)
= \sum_{i\geq 1}\,\chi_i(\sigma,\tau)x^{|\tau|-i},
\]
where $\chi_i(\sigma,\tau)\in \{0,1\}$ is computed as follows. Consider
the last $i$ letters $a_1a_2\dots a_i$ of $\sigma$ and the first $i$
letters $b_1b_2\dots b_i$ of $\tau$. If the differences $a_1-b_1$,
$a_2-b_2$, \dots, $a_i-b_i$ are all equal and
their common difference is $|\sigma|-i$ or $-(|\tau|-i)$, then
$\chi_i(\sigma,\tau)=1$; otherwise, $\chi_i(\sigma,\tau)=0$. We can
organize this computation in a table as in the following example.
\newcommand{\z}{\phantom{0}}
\[
\begin{array}{lll}
\sigma: & 53412 & \\
\tau: & \z 563421 & 1 \\
& \z\z 563421 & 0 \\
& \z\z\z 563421 & 1 \\
& \z\z\z\z 563421 & 0 \\
\end{array}
\]
Writing $\chi_j$ for $\chi_j(\sigma,\tau)$ we have
$(\chi_1,\chi_2,\chi_3,\chi_4)=(0,1,0,1)$ and
$\Omega(\sigma,\tau)=x^2+x^4$.
\begin{exam}
Penney's game~\cite{Penney1969, Guibas1981} is a well-known coin
tossing game played by two players. Player~I picks a sequence of
heads and tails, say $\alpha\in \{H,T\}^k$. Player~I shows $\alpha$
to Player~II, who subsequently picks a word $\beta\in \{H,T\}^k$. A
fair coin is tossed until either $\alpha$ or $\beta$ appears as a
consecutive subsequence of outcomes. The player whose sequence
appears first wins. Thus a winning sequence of outcomes for Player I
is a word $\omega\in \{H,T\}^n$ such that $\alpha$ is a suffix of
$\omega$ and $\omega$ otherwise avoids both $\alpha$ and
$\beta$. There is a simple formula~\cite[Formula (1.5)]{Guibas1981},
due to Conway~\cite{Gardner1974}, expressing the odds that Player~II
will win in terms of correlation polynomials. Further, this game has
the curious property of being nontransitive: Assuming $k\geq 3$, no
matter what $\alpha$ Player~I picks, Player~II can pick $\beta$ so
that his probability of winning exceeds $1/2$. Consider the
following related question for permutations. Let $T$ be an antichain
of Hertzsprung patterns and fix a pattern $\alpha\in T$. How many
permutations in $\Sym_n$ avoid $T$ except for a single occurrence of
$\alpha$ at the end of the permutation? Let us call this number
$f^{\alpha}(n)$. Let $F^{\alpha}(\uu; x)$ be the generating function
for marked permutations that end in an unmarked occurrence of
$\alpha$. Then
\[
F^{\alpha}(\uu;x) =
C^{\alpha}(\uu;x)\sum_{m\geq 1}m!\left(x + C(\uu; x)\right)^{m-1},
\]
where $C(\uu;x)$ is the usual cluster generating function and
$C^{\alpha}(\uu;x)$ is the generating function for marked
permutations $(\pi,M)$ such that the last $k$ letters
$\gamma=\pi(n-k+1)\ldots\pi(n-1)\pi(n)$ of $\pi$ form an unmarked
occurrence of $\alpha$ and $(\pi,M\cup\{\gamma\})$ is a
$T$-cluster. By construction,
\[
F^{\alpha}(\mathbf{-1}; x) = \sum_{n\geq 0}f^{\alpha}(n)x^n.
\]
This formula is, however, only useful if we can efficiently compute
$C^{\alpha}(\uu;x)$. Define the digraph $D_T^{\alpha}$ with
vertices $V=\{\eps, \hat\alpha\}\cup T$, in which $\hat\alpha$ is a
distinct copy of $\alpha$, and weight function
$\hat w: V\times V\to \QQ[\uu, x]$ given by
\[
{\hat w}(\sigma, \tau) =
\begin{cases}
w(\sigma, \tau) & \text{if $\sigma\neq\hat\alpha$ and $\tau\neq\hat\alpha$}, \\
0 & \text{if $\sigma=\hat\alpha$},\\
x^{|\alpha|} & \text{if $\sigma=\eps$ and $\tau=\hat\alpha$}, \\
\Omega(\sigma, \alpha) & \text{if $\sigma\notin\{\eps,\hat\alpha\}$ and $\tau=\hat\alpha$}.\\
\end{cases}
\]
Here, $w$ -- in the first clause -- is the weight function of the
digraph $D_T$. Let $A$ denote the adjacency matrix of
$D_T^{\alpha}$. In particular, if $T=\{\alpha, \beta\}$ then
$V=\{\eps,\alpha,\beta,\hat\alpha\}$ and the adjacency matrix is
\[A =
\begin{bmatrix}
\;0 & ux^{|\alpha|} & v x^{|\beta|} & x^{|\alpha|}\; \\
\;0 & u\Omega(\alpha,\alpha) & v\Omega(\alpha,\beta) & \Omega(\alpha,\alpha) \; \\
\;0 & u\Omega(\beta,\alpha) & v\Omega(\beta,\beta) & \Omega(\beta,\alpha) \; \\
\;0 & 0 & 0 & 0 \\
\end{bmatrix}.
\]
Now, $C^{\alpha}(\uu;x)$ is the generating function for walks in
$D_T^{\alpha}$ that start at $\eps$ and end in $\hat\alpha$. By the
transfer-matrix method and Cramer's rule,
\[
C^{\alpha}(\uu;x) =
(-1)^{|T|+1}\det\bigl(1-A: |T|+2, 1 \bigr)/\det(1-A).
\]
\end{exam}
\section{Pattern-rewriting systems}\label{PRS}
The Robinson--Schensted--Knuth (RSK) algorithm associates a permutation
$\pi\in\Sym_n$ with a pair $\bigl(P(\pi),Q(\pi)\bigr)$ of standard Young
tableaux having the same shape. For $\pi,\sigma\in\Sym_n$, let us write
$\pi\equiv \sigma$ if $P(\pi)=P(\sigma)$. In particular,
\begin{equation*}
132\equiv 312\quad\text{and}\quad 213\equiv 231.
\end{equation*}
Knuth~\cite{Knuth1970} showed that, when appropriately extended, these
two elementary transformations suffice to determine whether
$\pi\equiv \sigma$. Indeed, $\pi\equiv \sigma$ if and only if one can be
obtained from the other through a sequence of transformations of the
form
\[
\ldots acb \ldots \,\equiv\, \ldots cab \ldots
\quad\text{and}\quad
\ldots bac \ldots \,\equiv\, \ldots bca \ldots
\]
where $a**] (a) -- (b);
\draw[->] (a) -- (c);
\draw[->] (b) -- (d);
\draw[->] (c) -- (d);
\end{tikzpicture}
\]
in which $\hat\beta_1\simeq\beta_1$ and $\hat\beta_2\simeq\beta_2$.
Thus, to prove local confluence, we can restrict attention to
permutations of the form $\pi=\pi_1\gamma\pi_2$ and pairs of
rules $(\alpha_1\to\beta_1, \alpha_2\to\beta_2)$ in $R\times R$
such that $\alpha_1$ is a proper Hertzsprung
prefix of $\gamma$, $\alpha_2$ is a proper Hertzsprung suffix of
$\gamma$ and $|\gamma| < |\alpha_1|+|\alpha_2|$. The situation is
summarized in the following lemma.
\begin{lemma}\label{prs-confluence}
Let $R$ be a terminating pattern-rewriting system. Let
\[
\olap(R) = \bigcup_{(\alpha_1,\alpha_2)}\!\olap(\alpha_1,\alpha_2),
\]
where the union is over all
pairs $(\alpha_1,\alpha_2)\in\dom(R)\times \dom(R)$. If, for all
$\pi\in\olap(R)$,
\[
\rho_1 \from \pi \to \rho_2 \implies
\exists\sigma.\ \rho_1 \tostar \sigma \fromstar \rho_2,
\]
then $\to$ is confluent.
\end{lemma}
As a corollary to Lemma~\ref{prs-confluence}, confluence of a finite and
terminating pattern-rewriting system is decidable.
A relation $\to$ is called \emph{globally finite} if for each element
$x$ there are finitely many elements $y$ such that $x\toplus y$, where
$\toplus$ denotes the transitive closure of $\to$. A relation $\to$ is
called \emph{acyclic} if there is no element $x$ such that $x\toplus
x$. It is well known that any acyclic and globally finite relation is
terminating. Let $R$ be a pattern-rewriting system. Since each rule in
$R$ preserves the length of a permutation, it is clear that $\to_R$ is
globally finite. Thus to prove that $\to_R$ terminates it suffices to
prove that that it is acyclic. Let us call $f:\Sym\to\NN$ an
\emph{increasing statistic} with respect to $R$ if $\pi\to_R\sigma$
implies $f(\pi) < f(\sigma)$. Clearly, if there exists such a function
$f$ then $\to_R$ is acyclic.
\begin{lemma}
Let $R$ be a pattern-rewriting system. If there exists an increasing
statistic with respect to $R$, then $\to_R$ is terminating.
\end{lemma}
Having established the general framework we will now turn to
applications. A summary of the cases we shall consider can be found in
Table~\ref{EQ-Table}; \hyperlink{EQ2}{EQ2} through~\hyperlink{EQ6}{EQ6} were introduced by
Linton~\etal~\cite{Linton2012} and in the case of Hertzsprung patterns
they left the enumeration of equivalence
classes as open problems. Kuszmaul and Zhou~\cite{Kuszmaul2020} have
characterized the equivalence classes induced by the three separate
relations $123\equiv 132$, $123\equiv 321$ and
$123\equiv 132\equiv 321$. These are referred to as~\hyperlink{EQ2}{EQ2}, \hyperlink{EQ3}{EQ3} and~\hyperlink{EQ4}{EQ4} in
Table~\ref{EQ-Table}. Kuszmaul and Zhou provided explicit formulas for
the number of equivalence classes modulo~\hyperlink{EQ2}{EQ2} and~\hyperlink{EQ3}{EQ3} and noted that the
formula is the same in both cases. Indeed, it is
formula~\eqref{Myers132}, which is a special case of Myers's
formula~\eqref{Myers-formula}. For each of~\hyperlink{EQ1}{EQ1} through~\hyperlink{EQ7}{EQ7} we compute
generating functions for the number of equivalence classes in
$\Sym_n$. In particular we solve the two remaining open problems, \hyperlink{EQ5}{EQ5}
and~\hyperlink{EQ6}{EQ6}, posed by Linton~\etal{} The generating function for~\hyperlink{EQ7}{EQ7}
is also new. Our approach is quite general and could be applied to other
pattern-replacement equivalences of the Hertzsprung type. The reason for
choosing~\hyperlink{EQ1}{EQ1}--\hyperlink{EQ6}{EQ6} is that those equivalences have a history, and that we need
to limit the scope of our investigation. We choose~\hyperlink{EQ7}{EQ7} for two reasons.
It is a natural generalization of Stanley's equivalence (\hyperlink{EQ1}{EQ1}) and it
illustrates that large systems are within reach of our methods.
\begin{table}[htb]\centering
\vspace*{-3mm}
\begin{alignat*}{5}
&\qquad\;\;&
\text{Equivalences}&&&
\text{Rules}&&
\text{OEIS}&&
\text{Reference}\\[1.5ex]
&\text{\hypertarget{EQ1}{EQ1}} & 12\equiv 21 &\qquad&&
\begin{aligned}[t]
& 21 \to 12 \\
& 231 \to 312
\end{aligned}
&&\text{A013999} &\qquad&\text{Stanley~\cite{Stanley2012}} \\[1.5ex]
&\text{\hypertarget{EQ2}{EQ2}} & 123 \equiv 132 &&& 132\to 123 && \text{A212580} &&
\begin{aligned}[t]
&\text{Kuszmaul \&}\\[-0.6ex]
&\text{Zhou~\cite{Kuszmaul2020}}
\end{aligned}\\[1ex]
&\text{\hypertarget{EQ3}{EQ3}} & 123 \equiv 321 &&&
\begin{aligned}[t]
& 321 \to 123 \\
& 2341 \to 4123
\end{aligned}
&&\text{A212580} &&
\begin{aligned}[t]
&\text{Kuszmaul \&}\\[-0.6ex]
&\text{Zhou~\cite{Kuszmaul2020}}
\end{aligned}\\[1.5ex]
&\text{\hypertarget{EQ4}{EQ4}} &
\begin{aligned}[t]
123 &\equiv 132 \\
&\equiv 213
\end{aligned}
&&&
\begin{aligned}[t]
& 132 \to 123\\
& 213 \to 123
\end{aligned}
&& \text{A212581} &&
\begin{aligned}[t]
&\text{Kuszmaul \&}\\[-0.6ex]
&\text{Zhou~\cite{Kuszmaul2020}}
\end{aligned}\\[1.5ex]
&\text{\hypertarget{EQ5}{EQ5}} &
\begin{aligned}[t]
123 & \equiv 132 \\
& \equiv 321
\end{aligned}&&&
\begin{aligned}[t]
& 132 \to 123 \\
& 321 \to 123 \\
& 2341 \to 4123
\end{aligned}
&& \text{A212432} &&\text{Example~\ref{EQ5}} \\[1.5ex]
&\text{\hypertarget{EQ6}{EQ6}} &
\begin{aligned}[t]
123 &\equiv 132 \\
&\equiv 213 \\
&\equiv 321
\end{aligned}
&&&
\begin{aligned}[t]
& 132 \to 123 \\
& 213 \to 123 \\
& 321 \to 123 \\
& 2341 \to 4123
\end{aligned}
&\qquad& \text{A212433} &&\text{Example~\ref{EQ6}}\\[1.5ex]
&\text{\hypertarget{EQ7}{EQ7}} &
\begin{aligned}[t]
123 &\equiv 132 \\
&\equiv 213 \\
&\equiv 231 \\
&\equiv 312 \\
&\equiv 321
\end{aligned}
&&&
\makebox[4em][l]{%
$\begin{aligned}[t]
& 132 \to 123 \\
& 213 \to 123 \\
& 231 \to 123 \\
& 312 \to 123 \\
& 321 \to 123 \\
& 2341 \to 4123 \\
& 34512 \to 45123 \\
& 54123 \to 45123 \\
& 6745123 \to 7456123
\end{aligned}$}
&\qquad& &&\text{Example~\ref{EQ7}}
\end{alignat*}
\caption{Equivalences and corresponding pattern-rewriting systems}\label{EQ-Table}
\end{table}
For reference, numerical data on the number of equivalence classes
modulo each of~\hyperlink{EQ2}{EQ2} through~\hyperlink{EQ7}{EQ7} is given in Table~\ref{SEQ-Table} of the
appendix.
\begin{exam}[\hyperlink{EQ1}{EQ1}]\label{EQ1}
Stanley~\cite{Stanley2012} considered the equivalence relation on
$\Sym_n$ generated by the interchange of any two adjacent elements
$\pi(i)$ and $\pi(i+1)$ such that $|\pi(i)-\pi(i+1)|=1$. In our
terminology it is the equivalence $12\equiv 21$. As an example,
for $n=3$ there are two equivalence classes, $\{123,132,213\}$ and
$\{231,312,321\}$. Let us express
Stanley's equivalence relation using a rewrite system. Consider
$R = \{21 \to 12\}$. Note that $231\from 321\to 312$ in which both $231$
and $312$ are in normal form. Hence this system is not confluent. We
can however make it into a confluent system by adding a rule, namely
$231\to 312$. In other words, we redefine $R$ by setting
$R=\{21 \to 12, 231\to 312\}$. Since $231$ and $312$ are
equivalent under $12\equiv 21$, the equivalence relation
$\equiv_R$ induced by $R$ still coincides with that of Stanley.
To see that the rewrite rule induced by $R$ terminates consider the
following permutation statistic. Let $\Sigma_{12}(\pi)$ be the sum
of positions of occurrences of the Hertzsprung pattern $12$. In
other words, $\Sigma_{12}(\pi)$ is the sum of all $i\in[n-1]$ such
that $\pi(i+1)=\pi(i)+1$. Clearly, $\Sigma_{12}$ is increasing with
respect to $R$ and hence $\to_R$ terminates.
To show confluence we will use
Lemma~\ref{prs-confluence}. We have $\olap(21,21)=\{321\}$,
$\olap(21,231)=\olap(231,231)=\emptyset$, $\olap(231,21)=\{3421\}$ and
hence $\olap(R) = \{321, 3421\}$. Confluence thus follows from the
two diagrams below.
\begin{align*}
\begin{tikzpicture}[baseline=0]
\node (3421) {$3421$};
\node (4231) [below left of=3421, node distance=3.3em, xshift=-0.7em] {$4231$};
\node (3412) [below of=3421, node distance=7.1em] {$3412$};
\node (4312) [below of=4231, node distance=2.5em] {$4312$};
\draw[->] (3421) -- (4231);
\draw[->] (3421) -- (3412);
\draw[->] (4231) -- (4312);
\draw[->] (4312) -- (3412);
\end{tikzpicture}\qquad\quad
\begin{tikzpicture}[baseline=2em]
\node at (0,0) (321) {$321$};
\node at (0,-1.3) (231) {$231$};
\node at (1.8,-1.3) (312) {$312$};
\draw[->] (321) -- (231);
\draw[->] (321) -- (312);
\draw[->] (231) -- (312);
\end{tikzpicture}
\end{align*}
By Corollary~\ref{eq-classes}, each equivalence class
under $12\equiv 21$ contains exactly one permutation avoiding
$\{21,231\}$. Hence the generating function for the number of
equivalence classes can be computed using
Theorem~\ref{cluster-thm} together with Theorem~\ref{transfer-thm}.
Stanley defines a permutation $\pi=a_1a_2\cdots a_n\in\Sym_n$ as
\emph{salient} if we never have $a_i=a_{i+1}+1$ ($1\leq i\leq n-1$) or
$a_i=a_{i+1}+2= a_{i+2}+1$ ($1\leq i\leq n-2$). In terms of
Hertzsprung patterns, $\pi$ is salient if and only if it is
$\{21,312\}$-avoiding. Stanley shows that each equivalence class under
$12\equiv 21$ contains exactly one salient permutation. On account of
the preceding paragraph, this implies that the two sets $\{21,231\}$
and $\{21,312\}$ are Wilf-equivalent (their sets of avoiders are
equipotent).
Stanley counts the salient permutations by a direct
inclusion-exclusion argument. The resulting generating function is
\[
\fsum{x(1-x)}.
\]
We, of course, obtain the same answer. In fact,
Theorems~\ref{cluster-thm} and~\ref{transfer-thm} show that the joint
distribution of $21$ and $231$ is the same as the joint distribution
of $21$ and $312$. In both cases the cluster generating function is
$x^2(vx + u)/(1-ux)$, where $u=u_{21}$ and $v=u_{231}$ or $v=u_{312}$.
\end{exam}
\begin{exam}[\hyperlink{EQ2}{EQ2}]\label{EQ2}
Consider the pattern-rewriting system
$R = \bigl\{ 132\to 123 \bigr\}$. As usual, let $\to$ also denote the
rewrite relation induced by $R$. The equivalence closure of $\to$ is
equal to the equivalence relation induced by $123\equiv 132$. Note
that $\to$ is terminating. Indeed, the number of occurrences of $123$
is an increasing statistic ($\Sigma_{12}$ is also an
increasing statistic). Moreover, since $\olap(132,132)$ is empty,
$\to$ is trivially confluent. By Corollary~\ref{eq-classes}, the
number of equivalence classes is $|\Sym_n(132)|$ for which we derived
a formula in Example~\ref{Ex132}.
\end{exam}
\begin{exam}[\hyperlink{EQ3}{EQ3}]\label{EQ3}
The relation induced by $123\equiv 321$ is the equivalence closure
of the rewrite relation $\to$ induced by
$R=\{321\to 123, 2341\to 4123\}$. This relation is terminating
since $\Sigma_{123}$, the sum of positions of occurrences of $123$,
is an increasing statistic.
We shall show that it is confluent as well. The relation induced by
$\{321\to 123\}$ is not confluent, that is why we include the rule
$2341\to 4123$. We have
$\olap(321,2341) = \olap(2341,2341) = \emptyset$,
\begin{equation*}
\olap(321,321) = \{4321, 54321\}\quad\text{and}\quad
\olap(2341,321) = \{456321\}.
\end{equation*}
Confluence thus follows from Lemma~\ref{prs-confluence} and the
three diagrams below.
\begin{align*}
&\begin{tikzpicture}[baseline=-5.78em]
\node (456321) {$456321$};
\node[below left of=456321, node distance=3.9em, xshift=-0.6em] (634521) {$634521$};
\node[below of=456321, node distance=10.7em] (456123) {$456123$};
\node[below of=634521, node distance=2.65em] (652341) {$652341$};
\node[below of=652341, node distance=2.65em] (654123) {$654123$};
\draw[->] (456321) -- (456123);
\draw[->] (456321) -- (634521);
\draw[->] (634521) -- (652341);
\draw[->] (652341) -- (654123);
\draw[->] (654123) -- (456123);
\end{tikzpicture}\qquad\quad
\begin{aligned}
&\begin{tikzpicture}[scale=0.75]
\node at (0,2) (4321) {$4321$};
\node at (0,0) (2341) {$2341$};
\node at (3,0) (4123) {$4123$};
\draw[->] (4321) -- (2341);
\draw[->] (4321) -- (4123);
\draw[->] (2341) -- (4123);
\end{tikzpicture}\\[0.8em]
&\begin{tikzpicture}[scale=0.75]
\node at (0,2) (54321){$54321$};
\node at (0,0) (34521){$34521$};
\node at (3,0) (52341){$52341$};
\node at (6,0) (54123){$54123$};
\draw[->] (54321) -- (34521);
\draw[->] (54321) -- (52341);
\draw[->] (54321) -- (54123);
\draw[->] (34521) -- (52341);
\draw[->] (52341) -- (54123);
\end{tikzpicture}
\end{aligned}
\end{align*}
Linton \etal~\cite{Linton2012} noted that the number of equivalence
classes under $123\equiv 132$ and $123\equiv 321$ appears to be the
same. This was later proved by Kuszmaul and Zhou~\cite{Kuszmaul2020}.
The $\{321,2341\}$-cluster generating function is
\[
\frac
{u_{321}u_{2341}x^5 - u_{2341}x^4 - u_{321}x^3}
{u_{321}x^2 + u_{321}x - 1}.
\]
Recall that the generating function for $132$-clusters is
$u_{132}x^3$. While these generating functions are quite different, on
specializing $u_{2341}$, $u_{321}$ and $u_{132}$ to $-1$ we get $-x^3$
in both instances, resulting in an alternative proof of the observation
made by Linton \etal{} In particular, the two sets of patterns
$\{132\}$ and $\{321, 2341\}$ are Wilf-equivalent. Finding a
combinatorial proof of this fact remains an open problem.
\end{exam}
\begin{exam}[\hyperlink{EQ4}{EQ4}]\label{EQ4}
Let $R=\{132\to 123,213\to 123\}$. Its equivalence closure coincides
with the equivalence relation induced by $123\equiv 132\equiv 213$.
That $\to$ terminates follows from $\Sigma_{123}$ being an increasing
statistic. Note that
\[
\olap(R) = \olap(132,213)\cup\olap(213,132) = \{1324, 21354\}.
\]
Using either $132\to 123$ or $213\to 123$ we have $1324\to 1234$ in
which $1234$ is in normal form. The diagram for $21354$ is rather
simple too:
\[
\begin{tikzpicture}[node distance=4em,baseline=0]
\node (a) {$21354$};
\node (b) [below left of=3421] {$12354$};
\node (c) [below right of=3421]{$21345\,.$};
\node (d) [below of=3421,node distance=5.6em] {$12345$};
\draw[->] (a) -- (b);
\draw[->] (a) -- (c);
\draw[->] (b) -- (d);
\draw[->] (c) -- (d);
\end{tikzpicture}
\]
\goodbreak
\noindent
Thus, by Corollary~\ref{eq-classes}, the number of equivalence classes
under $123\equiv 132\equiv 213$ is the same as the number of
permutations of $[n]$ that avoid $132$ and $213$, and the
corresponding generating function is
\[
\sum_{n\geq 0}|\Sym_n(132,213)|\mskip1mu x^n =
\fsum{\frac{(1+x)^2(1-x)x}{x^2 + x + 1}}.
\]
The coefficient of $x^n$ is given by Kuszmaul and
Zhou's formula~\cite[Theorem 2.32]{Kuszmaul2020}
\[
\sum_{i=0}^{\floor{n/3}}(-2)^i\binom{n-2i}{i}(n-2i)!
\]
\end{exam}
\begin{exam}[\hyperlink{EQ5}{EQ5}]\label{EQ5}
Consider the pattern-rewriting system
\[
R = \{132\to 123, 321\to 123, 2341\to 4123\}.
\]
The equivalence closure of $\to$ is induced by
$123\equiv 132\equiv 321$. There is no overlap of the pattern $132$
with $123$ or $2341$. Hence
\[
\olap(R) = \olap(321,321) \cup \olap(2341,321) = \{4321,54321,456321\}
\]
and the proofs of confluence and termination are the same as in
Example~\ref{EQ3} (EQ3). Thus the number of equivalence classes under
$\equiv$ equals the number of permutations avoiding the patterns
in $\dom(R)=\{132,321,2341\}$. To count such permutations we construct
the corresponding digraph. Letting $u=u_{132}$, $v=u_{321}$ and
$s=u_{2341}$ it is depicted below.
\vspace{-1.5ex}
\[
\begin{tikzpicture}[scale=0.8]
\node at (0,3) (132) {$132$};
\node at (3,3) (321) {$321$};
\node at (6,3) (2341) {$2341$};
\node at (3,0) (eps) {$\eps$};
\draw[->] (eps) -- (132) node[midway, left=3pt, yshift=-1pt] {$ux^3$};
\draw[->] (eps) -- (321) node[midway, right=2pt, yshift=7pt] {$vx^3$};
\draw[->] (eps) -- (2341) node[midway, right=3pt, yshift=-1pt] {$sx^4$};
\path (321) edge[loop] node[above] {$sx+sx^2$} (321);
\path[->] (2341) edge node[above=3pt, xshift=3pt] {$sx^2$} (321);
\end{tikzpicture}
\vspace{-1ex}
\]
The cluster generating function is
\[
C(u,v,s; x) =
\frac
{(uv + vs)x^5 + (uv -s)x^4 - (u + v)x^3}
{vx^2 + vx - 1}
\]
and on applying Theorem~\ref{cluster-thm} we find that
\[
\sum_{n\geq 0}|\Sym_n(132,321,2341)|\mskip1mu x^n = \fsum{x(1-2x^2)}.
\]
% The corresponding OEIS sequence is A212432.
\end{exam}
\begin{exam}[\hyperlink{EQ6}{EQ6}]\label{EQ6}
Consider the pattern-rewriting system
\[
R=\{132\to 123, 213\to 123, 321\to 123,2341\to 4123\}.
\]
Its equivalence closure coincides with the equivalence relation
induced by $123\equiv 132\equiv 213\equiv 321$. That $\to$ terminates
follows from $\Sigma_{123}$ being an increasing statistic with respect
to $R$. Note that
$\olap(R) = \{1324, 4321, 21354, 54321, 456321\}$. In
Example~\ref{EQ3} (\hyperlink{EQ3}{EQ3}) we have seen diagrams for $4321$, $54321$ and
$456321$. This leaves $1324$ and $21354$, but we have seen diagrams
for those permutations in Example~\ref{EQ4} (\hyperlink{EQ4}{EQ4}). As before, we use
Theorem~\ref{transfer-thm} to compute the cluster generating
function. In the end we find that the generating function for the
number of equivalence classes under
$123\equiv 132\equiv 213\equiv 321$ is
\[
\sum_{n\geq 0}|\Sym_n(132,213,321,2341)|\mskip1mu x^n =
\fsum{\frac{-x^5 - 2x^4 - 2x^3 + x^2 + x}{x^2 + x + 1}}.
\]
\end{exam}
\begin{exam}[\hyperlink{EQ7}{EQ7}]\label{EQ7}
As detailed in Example~\ref{EQ1} (\hyperlink{EQ1}{EQ1}), Stanley characterized and
counted the equivalence classes induced by $12\equiv 21$. In other
words, the equivalences induced by considering all permutations in
$\Sym_2$ as equivalent. We shall consider the corresponding problem
for $\Sym_3$. That is, the equivalence induced by
\[123\equiv
132\equiv
213\equiv
231\equiv
312\equiv
321.
\]
Let $R$ consist of the rules $\alpha\to 123$, for
$\alpha\in\Sym_3\setminus\{123\}$, together with the following four rules
that are needed to make the relation confluent:
\begin{align*}
2341 \to 4123,\;\,
34512 \to 45123,\;\,
54123 \to 45123,\;\,
6745123 \to 7456123.
\end{align*}
One can check that $\Sigma_{12}$ is an increasing statistic and hence
$\to$ terminates. Confluence follows from verifying local confluence
for each permutation in $\olap(R)$, which is the union of the
sets
\begin{gather*}
\olap(123,213) = \{1324\},\mskip6mu
\olap(213,132) = \{21354\},\mskip6mu
\olap(231,312) = \{45312\}, \\
\olap(231,321) = \{45321\},\mskip6mu
\olap(231,54123) = \{6754123\},\mskip6mu
\olap(312,231) = \{4231\}, \\
\olap(312,6745123) = \{86745123\},\mskip6mu
\olap(321,312) = \{54312\}, \\
\olap(321,321) = \{4321, 54321\} ,\mskip6mu
\olap(321,54123) = \{654123, 7654123\} \\
\olap(2341,312) = \{456312\} ,\mskip6mu
\olap(2341,321) = \{456321\}, \\
\olap(2341,54123) = \{67854123\} ,\mskip6mu
\olap(34512,231) = \{456231\}, \\
\olap(34512,6745123) = \{8\,9\mskip2mu 1\mskip-1mu0\,6\,7\,4\,5\,1\,2\,3\} ,\mskip6mu
\olap(54123,2341) = \{652341\}, \\
\olap(54123,34512) = \{7634512\} ,\mskip6mu
\olap(6745123,2341) = \{78562341\}, \\
\olap(6745123,34512) = \{896734512\}.
\end{gather*}
We omit the details. Thus, the two cardinalities $|\Sym_n/{\equiv}|$
and $|\Sym_n(\dom(R))|$ are equal. The adjacency matrix of $D_T$ (with $T=\dom(R)$) is
\[
\begin{bmatrix}
0& u_0x^3& u_1x^3& u_2x^3& u_3x^3& u_4x^3& u_5x^4& u_6x^5& u_7x^5& u_8x^7\,\\
0& 0& u_1x& 0& 0& 0& 0& 0& 0& 0 \\
0& u_0x^2& 0& 0& 0& 0& 0& 0& 0& 0 \\
0& 0& 0& 0& u_3x^2& u_4x^2& 0& 0& u_7x^4& 0 \\
0& 0& 0& u_2x& 0& 0& 0& 0& 0& u_8x^5 \\
0& 0& 0& 0& u_3x^2& u_4x^2+u_4x& 0& 0& u_7x^4+u_7x^3& 0 \\
0& 0& 0& 0& u_3x^2& u_4x^2& 0& 0& u_7x^4& 0 \\
0& 0& 0& u_2x& 0& 0& 0& 0& 0& u_8x^5 \\
0& 0& 0& 0& 0& 0& u_5x& u_6x^2& 0& 0 \\
0& 0& 0& 0& 0& 0& u_5x& u_6x^2& 0& 0
\end{bmatrix}
\]
and the corresponding generating function is
\[
\fsum{\frac{-x^5 - 3x^4 - 4x^3 + x^2 + x}{x^2 + x + 1}}.
\]
\end{exam}
\goodbreak
\section*{Questions, conjectures and remarks}
Consider two Hertzsprung patterns $\sigma,\tau\in\Sym_k$ as equivalent
if $|\Sym_n(\sigma)|=|\Sym_n(\tau)|$ for all $n\geq 0$. How many
equivalence classes are there? In other words, what is the number of
Wilf-classes? By Corollary~\ref{single-pattern} it is the same as the
number of distinct ``autocorrelation polynomials''
$\Omega(\sigma,\sigma)$ with $\sigma\in\Sym_k$. For $k=1,2,\dots,7$ we find
\begin{gather*}
\{\,0 \,\} \\
\{\,x \,\} \\
\{\,0,\, x^2 + x \,\} \\
\{\,0,\, x^2,\, x^3,\, x^3 + x^2 + x \,\} \\
\{\,0,\, x^3,\, x^4,\, x^4 + x^3 + x^2 + x \,\} \\
\{\,0,\, x^3,\, x^4,\, x^4 + x^2,\, x^5,\, x^5 + x^4,\, x^5 + x^4 + x^3 + x^2 + x \,\}\\
\{\,
0,\,
x^4,\,
x^5,\,
x^6,\,
x^6 + x^3,\,
x^6 + x^5,\,
x^6 + x^5 + x^4 + x^3 + x^2 + x
\,\}
\end{gather*}
and the sequence for the number of Wilf-classes starts
\[
1, 1, 2, 4, 4, 7, 7, 11, 12, 18, 17, 25, 27, 38, 38\quad(k = 1,2,\dots,15).
\]
For $k\geq 3$ these numbers coincide with those of sequence A304178 in
the OEIS, which leads us to the conjecture below. For a palindrome
$w=c_1\ldots c_n\in\{0,1\}^n$ let
$P(w) = \{ i\in [n]: \text{$c_1\ldots c_i$ is a palindrome}\}$ be the
set of palindrome prefix lengths of $w$; \eg
$P(0100010) = \{1,3,7\}$.
\begin{conjecture}
Let $a_k=|\{\Omega(\sigma,\sigma): \sigma\in\Sym_k\}|$ be the number
of Wilf-classes of Hertzsprung patterns of length $k$. Let $b_k$ be
the number of distinct sets $P(w)$ for palindromes
$w\in\{0,1\}^k$. Then $a_k = b_{k+1}$ for $k\geq 3$.
\end{conjecture}
Recall that Bóna~\cite{Bona2008} showed that
$|\Sym_n(\tau)|\leq |\Sym_n(\id_k)|$ for each non-self-overlapping
pattern $\tau\in\Sym_k$. Data collected using
Corollary~\ref{single-pattern} suggest that the inequality holds
regardless of whether $\tau$ is self-overlapping or not.
\begin{conjecture}
We have $|\Sym_n(\tau)|\leq |\Sym_n(\id_k)|$ for
all Hertzsprung patterns $\tau\in\Sym_k$ and all natural numbers $n$.
\end{conjecture}
In view of Theorem~\ref{cluster-thm} it is natural to ask if there are
nontrivial mesh-patterns $p$ other than the Hertzsprung patterns for
which there is a rational function $R(x)$ such that
$\sum_{n\geq 0}|\Sym(p)|x^n = \sum_{m\geq 0}m!R(x)^{m}$? It appears
that the answer is yes.
\begin{conjecture}
We have
\[\sum_{n\geq 0}|\Sym_n(p)|\mskip1mu x^n
= \fsum{\frac{x}{1+x^2}},
\;\text{ where }\,
p = \pattern{}{3}{1/1,2/3/,3/2}{0/2,0/3,1/2,1/3,2/0,2/1,2/2,2/3,3/1,3/2}\,.
\]
\end{conjecture}
This conjecture is based on computing the numbers $|\Sym_n(p)|$ for
$n\leq 14$. They are $1$, $1$, $2$, $5$, $20$, $103$, $630$, $4475$,
$36232$, $329341$, $3320890$, $36787889$, $444125628$, $5803850515$,
and $81625106990$. At the time of writing this sequence is not in the
OEIS~\cite{OEIS}. If the conjecture is true then there is, however, a
close connection with a sequence in the OEIS, namely A177249. It is
defined by letting $a_n$ be number of permutations of $[n]$ whose
disjoint cycle decompositions have no adjacent transpositions, that
is, no cycles of the form $(i,i+1)$. Let
$F(x)=\sum_{n\geq 0}|\Sym_n(p)|x^n$ be the sought series and let
$A(x)=\sum_{n\geq 0}a_nx^n$. Brualdi and Deutsch~\cite{Brualdi2012}
have shown that $(1+x^2)A(x)=F(x)$. Our conjecture is thus equivalent
to $|\Sym_n(p)|=a_n+a_{n-2}$ for $n\geq 2$. Alternatively, one may
note that the compositional inverse of $x/(1+x^2)$ is $xC(x^2)$, where
$C(x)$ is the generating function for the Catalan numbers. Thus, our
conjecture is also equivalent to $[x^n]F(xC(x^2))=n!$, which could
lead to a novel decomposition of permutations.
As defined, pattern-rewriting systems are limited to transforming
Hertzsprung patterns. This is an artificial limitation though. Let
$\Mesh_k = \Pow\bigl([0,k]\times [0,k]\bigr)$ be the set of
$(k+1)\times(k+1)$ meshes. Define a \emph{pattern-rewriting system} as a
set of rules
$R \subseteq \bigcup_{k\geq 0}(\Sym_k\times\Mesh_k\times\Sym_k)$. Each
rule $(\alpha, M, \beta)$ is extended to $\Sym = \cup_n \Sym_n$ by
allowing that any occurrence of the mesh pattern $(\alpha, M)$ is
rewritten as an occurrence of $\beta$. In particular, this encompasses
the three types of equivalences studied by Linton~\etal, $M=\emptyset$,
$M = [1,k-1]\times [0,k]$ and $M=H_k$.
Using Lemma~\ref{prs-confluence} we are able to automate the proof of
confluence for any finite and terminating pattern-rewriting system. Our
current strategy to prove that a pattern-rewriting system terminates
requires us to manually come up with an increasing statistic. This was
easy for the small systems that we have considered but may not be easy
in general. Can there be a mechanical test? Or is this property
undecidable? It is known that termination of string-rewriting systems is
undecidable, even assuming that the system is
length-preserving~\cite{Caron1991}.
The starting point of this article was Hertzsprung's problem and we have
implicitly considered it the simplest nontrivial instance of a larger
class of problems. There is arguably a simpler, yet nontrivial, instance
though: permutations avoiding the single Hertzsprung pattern 12. The
study of the corresponding counting sequence appears to have started
with Euler. We have
\[
\sum_{\pi\in\Sym}u^{12(\pi)}x^{|\pi|} = \fsum{\frac{x}{1-(u-1)x}}.
\]
One may also note that any permutation $\pi$ can be uniquely written as
an inflation
\[
\pi=\sigma[\id_{k_1},\id_{k_2},\dots,\id_{k_m}],
\]
where $\sigma$ avoids the Hertzsprung pattern 12. Allowing ourselves to
use the terminology of $L$-species~\cite{species-book} this amounts to
the isomorphism $S'=F'\cdot E$, where $E$, $S$ and $F$ are the
$L$-species of sets, permutations and 12-avoiding permutations,
respectively. Indeed,
suppose $n\geq 1$ and write $\pi\in S_n$ as
$\pi=\sigma[\id_{k_1},\id_{k_2},\dots,\id_{k_m}]=\pi_1\pi_2\cdots\pi_m$. Let
$\hat\sigma$ be the permutation obtained by taking the first element of
each $\pi_i$ and let the set $A$ consist of all the remaining
elements. Then $\pi\mapsto (\hat\sigma, A)$ is the sought
isomorphism. Clearly $\hat\sigma\simeq\sigma$ and $12(\pi)=|A|$. For
instance,
\[
3142[12,1,1234,12] = 451678923
\;\mapsto\; \bigl(\mskip1mu 4162,\, \{3,5,7,8,9\} \mskip1mu\bigr).
\]
Since $S(x)=(1-x)^{-1}$ and $E(x)=e^x$ it follows from $S'(x)=F'(x)E(x)$
that $F'(x) = e^{-x}/(1-x)^2$. Replacing the species $E$ with the
$\ZZ[u]$-weighted species $E_u$ in which each set $A$ has weight $u^{|A|}$ and,
similarly, replacing the species $S$ with $S_w$ where the weight
function is $w(\pi)=u^{12(\pi)}$, we have $S_w'=F'\cdot E_u$ and hence
\[
\sum_{n\geq 0}\bigg(\sum_{\pi\in\Sym_n}u^{12(\pi)}\bigg)\frac{x^n}{n!}
\;=\; 1 + \int_0^x \frac{e^{(u-1)t}}{(1-t)^2} \dd t.
\]
Even simple Hertzsprung patterns have interesting properties!
\section*{Appendix}
Table~\ref{SEQ-Table} gives the number of equivalence classes modulo
\hyperlink{EQ2}{EQ2}--\hyperlink{EQ7}{EQ7} for $1\leq n \leq 20$.
\begin{table}
\[
\hspace{-0.5em}
\renewcommand{\arraystretch}{1.12}
\begin{array}{r||l|l|l}
& \text{\hyperlink{EQ2}{EQ2}}\;\&\;\text{\hyperlink{EQ3}{EQ3}}\;(\text{A212580}) &
\text{\hyperlink{EQ4}{EQ4}}\;(\text{A212581}) &
\text{\hyperlink{EQ5}{EQ5}}\;(\text{A212432}) \\%[0.1ex]
\hline &&\\[-2ex]
\begin{array}{rl}
1 \\
2 \\
3 \\
4 \\
5 \\
6 \\
7 \\
8 \\
9 \\
10 \\
11 \\
12 \\
13 \\
14 \\
15 \\
16 \\
17 \\
18 \\
19 \\
20 \\%[1ex]
\end{array}&
% EQ2--EQ3
\begin{array}{l}
1 \\
2 \\
5 \\
20 \\
102 \\
626 \\
4458 \\
36144 \\
328794 \\
3316944 \\
36755520 \\
443828184 \\
5800823880 \\
81591320880 \\
1228888215960 \\
19733475278880 \\
336551479543440 \\
6075437671458000 \\
115733952138747600 \\
2320138519554562560 \\
\end{array}&
% EQ4
\begin{array}{l}
1 \\
2 \\
4 \\
17 \\
89 \\
556 \\
4011 \\
32843 \\
301210 \\
3059625 \\
34104275 \\
413919214 \\
5434093341 \\
76734218273 \\
1159776006262 \\
18681894258591 \\
319512224705645 \\
5782488507020050 \\
110407313135273127 \\
2218005876646727423 \\
\end{array}&
% EQ5
\begin{array}{l}
1 \\
2 \\
4 \\
16 \\
84 \\
536 \\
3912 \\
32256 \\
297072 \\
3026112 \\
33798720 \\
410826624 \\
5399704320 \\
76317546240 \\
1154312486400 \\
18604815528960 \\
318348065548800 \\
5763746405053440 \\
110086912964367360 \\
2212209395234979840 \\
\end{array}\\
\hline
\hline &&\\[-2.2ex]
&\text{\hyperlink{EQ6}{EQ6}}\;(\text{A212433}) &
\text{\hyperlink{EQ7}{EQ7}} & \\%[0.1ex]
\hline &&\\[-2ex]
\begin{array}{rl}
1 \\
2 \\
3 \\
4 \\
5 \\
6 \\
7 \\
8 \\
9 \\
10 \\
11 \\
12 \\
13 \\
14 \\
15 \\
16 \\
17 \\
18 \\
19 \\
20 \\%[1ex]
\end{array}&
% EQ6
\begin{array}{l}
1 \\
2 \\
3 \\
13 \\
71 \\
470 \\
3497 \\
29203 \\
271500 \\
2786711 \\
31322803 \\
382794114 \\
5054810585 \\
71735226535 \\
1088920362030 \\
17607174571553 \\
302143065676513 \\
5484510055766118 \\
104999034898520903 \\
2114467256458136473 \\
\end{array}&
% EQ7
\begin{array}{l}
1 \\
2 \\
1 \\
6 \\
40 \\
330 \\
2664 \\
23258 \\
222154 \\
2326410 \\
26568950 \\
328995136 \\
4392819522 \\
62935547966 \\
963253101304 \\
15688298164890 \\
270944692450742 \\
4946387077324072 \\
95184319122508074 \\
1925732716758497918 \\
\end{array}
&
\begin{array}{ll}
\end{array}
\end{array}
\]
\caption{Enumeration of equivalence classes EQ2--EQ7}\label{SEQ-Table}
\end{table}
\bibliographystyle{amsplain-ac}
\bibliography{ALCO_Claesson_580}
\end{document}
**