On joint subtree distributions under two evolutionary models
Extracted source text · arXiv 1508.03139
\documentclass[smallextended]{svjour3}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{mathrsfs}
%\usepackage{graphicx}
\usepackage{comment}
\usepackage{color}
\usepackage{wrapfig}
\usepackage{floatrow}
%\usepackage{subfigure}
\usepackage{graphicx,epic,tikz,pifont}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{natbib}
%\usepackage{lineno}
%\pagewiselinenumbers
%\newtheorem{theorem}{Theorem}[section]
%\newtheorem{definition}[theorem]{Definition}
%\newtheorem{lemma}[theorem]{Lemma}
%\newtheorem{proposition}[theorem]{Proposition}
%\newtheorem{corollary}[theorem]{Corollary}
\newcommand{\nbt}{\varphi}
\newcommand{\prt}{\mathbb{P}}
\newcommand{\pp}{\mathbb{P}}
\newcommand{\ee}{\mathbb{E}}
\newcommand{\ind}{\mathbb{I}}
\newcommand{\tsp}{\mathcal{T}}
%\newcommand{\ch}{C} %number of cherry
%\newcommand{\pf}{F} %number of pitchfork
\newcommand{\pf}{A}
\newcommand{\ch}{C}
\newcommand{\var}{\mathbb{V}}
\newcommand{\cov}{Cov}
\newcommand{\corr}{\rho}%mathop{Corr}}
%\newcommand{\puni}{\mathbb{P}_{\text{PDA}}}
%\newcommand{\YHK}{\textnormal{\tiny \textsc{Y}}}
%\newcommand{\YHK}{\textnormal{\tiny \textsc{Y}}}
%\newcommand{\PDA}{\textnormal{\tiny \textsc{U}}}
\newcommand{\YHK}{y}
\newcommand{\PDA}{u}
\newcommand{\epf}{\hfill $\square$}
\newcommand{\pyule}{\mathbb{P}_{\YHK}}
\newcommand{\puni}{\mathbb{P}_{\PDA}}
\newcommand{\rev}[1]{#1}%\textcolor{red}{#1}}
\newcommand{\otype}{t}
\newcommand{\tw }{ \bf}
\usepackage{marvosym}
\newcommand{\envelope}{(\raisebox{-.5pt}{\scalebox{1.45}{\Letter}}\kern-1.7pt)}
\usepackage[text={6.6in,9.6in},centering]{geometry}
\linespread{1.5}
%\usepackage{hyperref}
\sloppy
\graphicspath{{./figs/}}
%\bibliographystyle{natbib}
%\bibliographystyle{spmpsci}
%\usepackage[round]{natbib}
%\bibliographysyle{apalike}
% Use natbib for \citet, \citep, etc.
% (JMB: numbered citations that are sorted and compressed)
%\usepackage[numbers,sort&compress]{natbib}
%\bibliographystyle{spmpsci}
\bibliographystyle{spbasic}
\begin{document}
\title{On joint subtree distributions under two evolutionary models
}%: cherries and pitchforks}
\author{Taoyang Wu \and Kwok Pui Choi}
\institute{
T. Wu %\envelope
\at School of Computing Sciences, University of East Anglia, Norwich, United Kingdom \\\email{taoyang.wu@uea.ac.uk, taoyang.wu@gmail.com}
\and
K.P. Choi
\at Department of Statistics and Applied Probability, and Department of Mathematics, National University of Singapore, Singapore 117546, \\\email{stackp@nus.edu.sg}
}
%\thanks{todo.}
%\subjclass{}
%\keywords{tree rearrangement, \tbr, phylogenetic tree}
\date{\today}
\maketitle
\begin{abstract}
In population and evolutionary biology, hypotheses about micro-evolutionary and macro-evolutionary processes are commonly tested by comparing the shape indices of empirical evolutionary trees with those predicted by neutral models. A key ingredient in this approach is the ability to compute and quantify distributions of various tree shape indices under random models of interest. As a step to meet this challenge, in this paper we investigate the joint distribution of cherries and pitchforks (that is, subtrees with two and three leaves) under two widely used null models: the Yule-Harding-Kingman (YHK) model and the proportional to distinguishable arrangements (PDA) model. Based on two novel recursive formulae, we propose a dynamic approach to numerically compute the exact joint distribution (and hence the marginal distributions) for trees of any size.
We also obtained insights into the statistical properties of trees generated under these two models, including a constant correlation between the cherry and the pitchfork distributions under the YHK model, the log-concavity and unimodality of cherry distributions under both models. In particular, we show the existence of a unique change point for cherry distribution between the two models, that is, there exists a critical value $\tau_n$ for each $n\geq 4$ such that the probability that a random tree with $n$ leaves generated under the YHK model contains $k$ cherries is lower than that under the PDA model if $1<k< \tau_n$, and higher if $\tau_n<k\le n/2$.
%, that is, there exists a critical value $\tau_n$ for each $n\geq 4$ such that a random tree with $n$ leaves generated under the YHK model is more likely to contain $k$ cherries than under the PDA model if and only if $1\le k< \tau_n$.
\bigskip
\keywords{ phylogenetic tree $\cdot$ subtree distribution $\cdot$ Yule-Harding-Kingman model $\cdot$ PDA model $\cdot$ tree indices
$\cdot$ joint distribution
}
%\keywords{Phylogenetic trees \and Null models \and Clade \and Clan \and Log-convexity}
% \subclass{MSC code1 \and MSC code2 \and more}
% 05C05 Trees
% 05C80 Random graphs
% 92B10 Taxonomy, cladistics, statistics
% 92D15 Problems related to evolution
\end{abstract}
%\noindent{\bf Keywords:}
\newpage
\section{Introduction}
In population and evolutionary biology, hypotheses about micro-evolutionary processes (e.g. population genetics) and macro-evolutionary processes (e.g. speciation and extinction) are commonly tested by comparing frequencies of the shapes of empirical evolutionary trees with those predicted by null models~\citep[see, e.g.][]{mooers97a, nordborg01a,blum2006random,purvis2011shape,hagen2015age}. One challenge in this approach is the ability to compute the distributions of various tree shape indices under the models of interest, which is needed in statistical testing for calculating the $p$-value of the empirical shape statistics or constructing a confidential interval. Even for some relatively simple null models, this can still be a challenging task. Many current approaches are based on approximating techniques, such as Monte Carlo sampling~\citep[see, e.g.][]{blum2006random} or Gaussian approximation~\citep[see, e.g.][]{McKenzie2000}, which could be computationally intensive or restricting the tests to the trees above a certain size. Therefore it is desirable to explore efficient ways of computing these distributions exactly.
%However, Monte Carlo sate first approach could be computational intensive and the second one may limit the
%The Yule model and coalescent model are two natural stochastic models for
Two widely used null models for generating random trees in population and evolutionary biology are the Yule-Harding-Kingman (YHK) model~\citep{harding71a,yule25a,Kingman1982} and the proportional to different arrangements (PDA) model~\citep{Aldous2001}. %The YHK model is the tree
Under the PDA model all rooted binary trees of the same size are chosen with the same probability \citep{Aldous2001} whilst under the YHK model each tree is chosen with a probability proportional to the number of total orderings that can be assigned to its internal nodes so that the relative partial ordering derived from the tree topology is preserved.
%More precisely, the YHK model assumes that the speciation process develops with a constant pure-birth rate \citep{Blum2006, Pinelis2003}, and is equivalent to Kingman's coalescent process \citep{Kingman1982} when the branch length is not taken into account.
%Both models are commonly used to generate prior tree distributions in Bayesian phylogenetic analyses ~\citep[see e.g.][]{Li2000, Rannala1996}.
In this paper, we are interested in the exact computation of the joint distribution for the number of subtrees under the YHK and PDA model. Here a subtree, also known as a fringe subtree in ~\citet{aldous1991asymptotic}, consists of a node and all its descendants. More specifically, we study the distributions of the number of cherries, subtrees with two leaves, and that of pitchforks, subtrees with three leaves. Note that this is equivalent to study the joint distributions of $2$-pronged and $3$-pronged nodes as defined in~\citep{rosenberg06a}, as well as the joint distributions of clades of size two and three as defined in~\citep{zhu11a}.
We now describe the contents of the rest of this paper.
In the next section we gather some necessary notation and background.
In particular, we present a random tree generating process for realising both the YHK and PDA models as described in~\cite{McKenzie2000}. In contrast to the splitting model that were used in several previous studies concerning the asymptotical distributions of subtrees~\citep[see, e.g.][]{chang2010limit}, the process used here is based on iteratively attaching leaves. We therefore also collect some observations on the change of the numbers of cherries and pitchforks in a tree when an additional leaf is attached.
In Sections~\ref{sec:YHK} and~\ref{sec:PDA} we study subtree distributions under the YHK and the PDA models, respectively.
Our main results include two novel recursive formulae on the joint distributions of cherries and pitchforks; see Theorem~\ref{thm:yule:pf} for the one under the YHK model and Theorem~\ref{thm:pda:pf} for the one under the PDA model. These recursions enable us to develop a dynamic approach to numerically compute the joint distributions, and hence also their marginal distributions, for trees of any size. As an example, in Fig.~\ref{fig:jd} we illustrate the distributions for trees with 200 leaves, which suggests the joint distributions under both models can be well-approximated by bivariate normal distributions. %In addition,
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{joint_YHK_200}
% \caption{A subfigure}
% \label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=1.1\linewidth]{joint_PDA_200}
% \caption{A subfigure}
% \label{fig:sub2}
\end{subfigure}
\caption{Probability density function for the joint distribution of cherries and pitchforks on phylogenetic trees with 200 leaves.}
% with $n=200$.}
\label{fig:jd}
\end{figure}
Rewritten in functional forms, the recursions also provide a way to compute the covariance and correlation of the joint distributions under these two models. Somewhat surprisingly, we find that under the YHK model the correlation between the cherry and the pitchfork distributions is a constant $-\sqrt{14/69}$, which is independent of the number of leaves (see, e.g., Corollary~\ref{cor:yule:cor}). In addition, the recursions also lead to an alternative, and arguably more unified approach to computing the moments of these two distributions, and we demonstrate this by reaffirming several results obtained in previous studies. % (see
Using the recursions on cherry distribution derived from the joint distribution, we obtain in Theorem~\ref{ch:pda} the exact formula for cherry distribution under the PDA model, and derive some interesting properties concerning cherry distribution in general, including that this distribution is log-concave and hence unimodal under both models (see Theorems~\ref{thm:lc:yule} and~\ref{thm:lc:pda}).
In Section~\ref{sec:comp} we present a comparative study of cherry and pitchfork distributions under the YHK and PDA models.
We first compare the mean and the variance of these two distributions under these two models. Then we show in Theorem~\ref{thm:change} that there exists a unique change point when comparing cherry distribution, that is, there exists a critical value $\tau_n$ for each $n\geq 4$ such that the probability that a random tree with $n$ leaves generated under the YHK model contains $k$ cherries is lower than that under the PDA model if $1<k< \tau_n$, and higher if $\tau_n<k\le n/2$. %or equal otherwise. %if $\tau_n\le k \le n$.
Finally, we conclude in Section~\ref{sec:discussion} with discussions and some open problems.
%In this paper we will be mainly interested in More specifically, we consider two subtree families that of special interest, that is, those with two and three leaves, also known as cherry and pitchfork, respectively.
%In this paper will will study the number of cherry and pitchfork. Note that the number of cherry in a tree is the same as the number of $2$-pronged nodes, as well as that of $2$-caterpillars or clades of size two. Similarly, the number of pitchforks in a tree is the same as the number of $3$-pronged nodes, as well as $3$-caterpillars and clades of size three.
%The remainder of the paper is organized as follows. In Section~\ref{sec:pre} we gather some necessary notation and background used in the paper and a brief review of the YHK and the PDA models. We then collect in Section~\ref{sec:subtree} some technical results concerning the change of the numbers of cherries and pitchforks when a new leaf is attached to a phylogenetic tree. In Sections~\ref{sec:YHK} and~\ref{sec:PDA} we study the cherry and the pitchfork distributions under the YHK and the PDA models, respectively. Finally, we present a comparative study of the two distributions under the two models in Section~\ref{sec:comp}, and conclude in Section~\ref{sec:discussion} with discussions and some open problems.
\newpage
\section{Preliminaries}
\label{sec:pre}
For later use, we present in this section some basic notation and results concerning phylogenetic trees. Throughout this article, $X$ denotes a finite set with $|X|=n \geq 2$.
%, and $X$ usually represents a set of species or genes.
%used to denote the leaf set, and we assume that $X$ is a finite set of size $n=|X|\geq 3$ unless stated otherwise.
\bigskip
%\noindent
%{\bf Phylogenetic trees}
%\subsection*{Phylogenetic trees}
%Let $X$ be a finite set, which usually represents a collection of given species.
\noindent
{\bf Phylogenetic trees}
A {\em phylogenetic tree} $T=(V(T),E(T))$ on $X$ is a rooted tree with leaf set $L(T)=X$ such that the root has one child whilst all other vertices have either zero or two children (see Fig.~\ref{fig:tree} for an example).
Note that in this paper phylogenetic trees are rooted, with their edges directed away from the root. In addition,
% Although in literature the root is often assumed to have two children, in some studies, including the current one, it is more convenient to
for technical simplicity we assume without loss of generality that the root has one child~(also referred to as planted phylogenetic trees by~\citet{baroni2005bounding}).
Let $E^*(T)$ be the set of pendant edges in $T$, i.e., those edges incident with a leaf. Then we have $|E(T)|=2n-1$ and $|E^*(T)|=n$.
\begin{figure}[h]
\begin{center}
{\includegraphics[scale=0.8]{PhyloTree.pdf}}
\end{center}
\caption{Examples of phylogenetic trees. $T$ is a phylogenetic tree on $X=\{x_1,\dots, x_6\}$, and $T'=T[e_9;x_7]$ is a phylogenetic tree on $\{x_1,\dots,x_7\}$ that is obtained from $T$ by attaching the leaf labelled $x_7$ to edge $e_9$. Here the directions of all edges are directed away from the root $r$, and hence omitted for simplicity.
%is obtained from $T$ by adding the leaf labeling $6$ to the edge incident with the leaf labeling $2$ in $T$.
}
\label{fig:tree}
\end{figure}
Let $e$ be an edge in a phylogenetic tree $T$. The tree consisting of $e$ and all edges below $e$ is called a {\em subtree} of $T$, and is denoted by $T(e)$. In particular, a {\em cherry} is a subtree with two leaves, and a {\em pitchfork} is a subtree with three leaves.
The number of cherries and pitchforks contained in $T$ are denoted by $\ch(T)$ and $\pf(T)$, respectively.
Note first that we always have $1\leq \ch(T)\leq n/2$ and $0\leq \pf(T)\leq n/3$. Moreover, in our definition a cherry contains three edges and a pitchfork contains five edges. As an example, for the tree $T$ depicted in Fig.~\ref{fig:tree} we have $C(T)=2$ and $A(T)=1$. In addition, $T(e_8)$ is a pitchfork with edge set $\{e_1,e_3,e_5,e_7,e_8\}$, and $T(e_7)$ is a cherry with edge set $\{e_1,e_5,e_7\}$. Finally, $\ch(T)$ and $\pf(T)$ are respectively equal to the number of $2$-pronged nodes and $3$-pronged nodes contained in $T$ (see~\cite{rosenberg06a} for the definitions of $r$-pronged nodes).
Given an edge $e$ in a phylogenetic tree $T$ and a taxon $x_0\not \in L(T)$, let $T[e;x_0]$ be the phylogenetic tree obtained from $T$ by attaching a new leaf labelled with $x_0$ to the edge $e$. Formally, let $e=\{u,v\}$ and let $w$ be a vertex not contained in $V(T)$, then $T[e;x_0]$ has vertex set $V(T)\cup \{x_0,w\}$ and edge set $\big(E(T)\setminus \{e\} \big) \cup \{(u,w),(v,w),(w,x_0)\}$ (see Fig.~\ref{fig:tree} for an illustration of this construction).
When the labelling of the new leaf is clear from the context, $T[e;x_0]$ is abbreviated to $T[e]$.
%\subsection{The YHK and the PDA model}
\bigskip
\noindent
{\bf The YHK and the PDA model}
In this subsection, we present a formal definition of the two null models investigated in this paper: the {\it proportional to distinguishable arrangements} (PDA) model and the {\it Yule--Harding--Kingman} (YHK) model. In contrast to the splitting process used by~\cite{Aldous2001} to accommodate the two models, the random process used here is based on iteratively attaching leaves.
%To begin with, recall that the number of rooted phylogenetic trees with leaf set $X$ with $n=|X|$ is $\varphi(|X|) =\varphi(n)= (2n-3)!! = 1\cdot 3 \dotsb (2n-3)$ (see, e.g., \cite{semple03a}). Here we will use the convention that $\varphi(1)=1$.
%Under the Yule--Harding model~\cite{harding71a, yule25a}, a rooted phylogenetic tree on $X$ is generated as follows. Beginning with a two leafed tree, we ``grow'' it by repeatedly splitting a leaf into two new leaves. The splitting leaf is chosen randomly and uniformly among all the present leaves in the current tree. After obtaining an unlabeled tree with $|X|$ leaves, we label each of its leaves with a label sampled randomly uniformly (without replacement) from $X$. When branch lengths are ignored, the Yule--Harding model is shown by~\cite{aldous96a} to be equivalent to the trees generated by coalescent process \cite{Kingman1982}, and so we call it the YHK model.
%
%Now we introduce a stochastic process that randomly generates a tree in $\tsp_n$:
Under the Yule--Harding model~\citep{harding71a, yule25a},
a rooted phylogenetic tree on $X$ is generated as follows.
Beginning with the tree with two leaves, we ``grow'' it by repeatedly uniformly sampling a pendant edge $e$ in the current tree $T_{cur}$ %according to a given distribution $\pp_{cur}$ on $E(T_{cur})$
and replace $T_{cur}$ by $T_{cur}[e]$.
This process continues until a binary tree with $n$ leaves is obtained. Finally, we label each of its leaves with a label sampled randomly uniformly (without replacement) from $\{x_1,\dots,x_n\}$.
When branch lengths are ignored, the Yule--Harding model is shown by~\citet{aldous96a} to be equivalent to the trees generated by the coalescent process in population genetics introduced by \citet{Kingman1982}, and so we call it the YHK model.
The probability of generating a tree $T$ under this model is denoted by $\pyule(T)$.
Let $\tsp_n$ be the set of phylogenetic trees with leaf set $\{x_1,\dots,x_n\}$. It is well known that the number of trees contained in $
\tsp_n$ is $\varphi(n):=(2n-3)!! = 1\times 3 \times \dotsb \times (2n-3)$~\citep[see e.g.][]{semple03a}. Here we adopt the convention that $\varphi(1)=1$.
Under the PDA model, each tree has the same probability, that is,
$1/{\varphi(n)}$, to be generated.
Alternatively, a tree can be generated under the PDA model using a Markov process similar to the one used in the YHK model; the only difference is that the edge $e$ is uniformly sampled from $E(T)$, instead of $E^*(T)$.
We use $\ee _{\YHK}$, $\var _{\YHK}$, $\cov _{\YHK}$ and $\corr _{\YHK}$ to denote respectively the expectation, variance, covariance and correlation %of $\pf_n$
taken with respect to the probability measure $\pp _{\YHK}$ under the YHK model. Similarly, $\ee _{\PDA}$, $\var _{\PDA}$, $\cov _{\PDA}$ and $\corr _{\PDA}$ are defined with respect to the probability $\pp _{\PDA}$ under the PDA model.
For $n\geq 2$, let $\pf_n$ (resp. $\ch_n$) be the random variable
$\pf(T)$ (resp. $\ch(T)$) for a random tree $T$ in $\tsp_n$. In this paper, we are interested in the joint distributions and the marginal properties of $\pf_n$ and $\ch_n$ under the YHK and the PDA models.
\bigskip
\noindent
{\bf Subtree Pattern}
\label{sec:subtree}
For later use, we present in this subsection several technical results concerning the change of the numbers of cherries and pitchforks when a new leaf is attached to a phylogenetic tree.
We begin with the following notation. Given a phylogenetic tree $T$, let $E_{1}(T)$ be the set of pendant edges that are contained in a pitchfork but not a cherry; $E_{2}(T)$ the set of edges in $T$
that are contained in a cherry but not in a pitchfork;
$E_{3}(T)$ the set of pendant edges that are contained in neither a cherry nor a pitchfork; and $E_{4}(T)=E(T)\setminus (E_{1}(T)\cup E_{2}(T) \cup E_{4}(T))$. For instance, for the tree $T$ depicted in Fig.~\ref{fig:tree}, we have $E_1(T)=\{e_3\}$, $E_2(T)=\{e_2,e_4,e_9\}$, $E_3(T)=\{e_6\}$ and $E_4(T)=\{e_0,e_1,e_5,e_7,e_8,e_{10}\}$. In addition, $E(T)$ can be decomposed into the disjoint union of these four sets of edges.
%i.e., $E(T)=E_1(T)\sqcup E_2(T) \sqcup E_3(T) \sqcup E_4(T)$, where $\sqcup$ denotes disjoint union.
The following lemma, whose proof is straightforward and hence omitted here, shows this observation holds for all phylogenetic trees, where $\sqcup$ denotes disjoint union.
\begin{lemma}
\label{lem:edge-set}
Suppose that $T$ is a phylogenetic tree with $n$ leaves. Then we have
\begin{equation}
\label{eq:edge:dec}
E(T)=E_{1}(T)\,\sqcup\,E_{2}(T)\,\sqcup\,E_{3}(T)\,\sqcup\,E_{4}(T).
\end{equation}
In addition, we have
% $|E(T)|=2n-1$,
$|E_{1}(T)|=\pf(T)$,
$|E_{2}(T)|=3(\ch(T)-\pf(T))$,
$|E_{3}(T)|=n-\pf(T)-2\ch(T)$,
and
$|E_{4}(T)|=n-1+3\pf(T)-\ch(T)$.
% \epf
\end{lemma}
The last lemma provides a decomposition for the set of edges in a phylogenetic tree, which is useful to the study of the PDA model. For the YHK model, we need an analogous decomposition for $E^*(T)$, the set of the pendant edges in $T$. To this end, note first that we have $E_1(T)\subseteq E^*(T)$ and $E_3(T)\subseteq E^*(T)$. In addition, let $E^*_i(T):=E_i(T)\cap E^*(T)$ be the set of pendant edges in $E_i(T)$ for $i=2,4$. Then we have the following lemma, whose proof is straightforward and hence omitted.
\begin{lemma}
\label{lem:pen:edge-set}
Suppose that $T$ is a phylogenetic tree with $n$ leaves. Then we have
\begin{equation}
\label{eq:pendant:edge:dec}
E^*(T)=E_{1}(T)\,\sqcup\,E^*_{2}(T)\,\sqcup\,E_{3}(T)\,\sqcup\,E^*_{4}(T).
\end{equation}
In addition, we have
% $|E(T)|=2n-1$,
$|E^*_{2}(T)|=2(\ch(T)-\pf(T))$
and
$|E^*_{4}(T)|=2\pf(T)$.
% \epf
\end{lemma}
We end this section with the following result relating the values $\ch(T[e])-\ch(T)$ and $\pf(T[e])-\pf(T)$ to the choice of $e$.
\begin{proposition}
\label{prop:type}
Suppose that $e$ is an edge in a phylogenetic tree $T$ and $T'=T[e]$. Then
we have
\small{
\begin{equation*}
\pf(T') = \begin{cases}
\pf(T) & \text{if } e\in E_3(T)\cup E_4(T), \\
\pf(T)-1 & \text{if } e\in E_1(T),\\
\pf(T)+1 & \text{if } e\in E_2(T);
\end{cases} %\quad
\mbox{and} \quad
\ch(T') = \begin{cases}
\ch(T) & \text{if } e \in E_2(T)\cup E_4(T), \\
{} & \\
\ch(T)+1 & \text{if } e \in E_1(T)\cup E_3(T).
\end{cases}
\end{equation*}
}
\end{proposition}
\begin{proof}
Let $\{F_1,\dots,F_k\}$ be the set of pitchforks contained in $T$, and let $\{H_1,\dots,H_l\}$ ($l \geq k$) be the set of cherries contained in $T$. Here we may assume that indices are chosen in the way so that $H_1$ is contained in $F_1$. % for $1\leq i \leq k$.
Suppose first that $e=(u,v) \in E_1(T)$. Swapping the labelling of $F_i$ if necessary, we may assume that $e$ is the pendant edge contained in the pitchfork $F_1$ but not in the cherry $H_1$. Let $u_0$ be the parent of $u$, and let $u_1$ be the child of $u$ that is distinct from $v$. In addition, let $w$ be the newly added interior vertex in $T'$. Now consider $e_0=(u_0,u)$ and $e'=(u,w)$ in $T'$. Then $T'(e_0)$ is not a pitchfork as $u_0$ has four leaves as its descendants. On the other hand, $T'(e')$ is a cherry of $T'$ that is not contained in $T$. Therefore, we have $\pf(T')=k-1$ and $\ch(T')=l+1$, as required.
By a similar argument, we can establish the proposition for the other three cases, i.e., $e\in E_i(T)$ for $2\leq i \leq 4$. Since by Lemma~\ref{lem:edge-set} these four cases cover all possible choices of $e$, the proposition follows.
\epf
\end{proof}
One useful consequence of the last proposition is the following corollary, whose proof is straightforward and hence omitted.
\begin{corollary}
\label{cor:range}
Suppose that $e$ is an edge in a phylogenetic tree $T$ with $\pf(T)=a$ and $\ch(T)=b$. Then for the phylogenetic tree $T'=T[e]$, we have
%Suppose that $e$ is an edge in a phylogenetic tree $T$ and $T'=T[e]$. Denote then we have
$$
(\pf(T'),\ch(T'))\in \{(a-1,b+1),(a+1,b),(a,b+1),(a,b)\}
$$
according to the index $i$ ($1\leq i \leq 4$) with $e\in E_i(T)$.
%\epf
\end{corollary}
%\hfill $\square$
\section{Subtree Distributions under the YHK Model}
\label{sec:YHK}
In this section, we study the distributions of the random variables $A_n$ and $C_n$ under the YHK model.
Our starting point is the following recursion on their joint distribution.
% of $A_n$ and $C_n$.
\begin{theorem}
\label{thm:yule:pf}
We have
\begin{equation*}
\begin{split}
\pyule(\pf_{n+1}=a, \ch_{n+1}=b) &=
\frac{2a}{n}\pyule(\pf_{n}=a, \ch_{n}=b)+\frac{(a+1)}{n}\pyule(\pf_{n} =a+1, \ch_n=b-1)\\
& \qquad +\frac{2(b-a+1)}{n}\pyule(\pf_{n}=a-1, \ch_{n} =b) +\frac{(n-a-2b+2)}{n}\pyule(\pf_{n}=a, \ch_{n}=b-1)
\end{split}
\end{equation*}
for $n>3$ and $1<b<n$. Moreover,
$\pyule(\pf_{3}=a, \ch_{3}=b)$ equals to $1$ if $(a,b)=(1,1)$, and $0$ otherwise.
\end{theorem}
\begin{proof}
Fix $n> 3$, and let $T_2,\dots,T_n,T_{n+1}$ be a sequence of random trees generated by the YHK process, that is, $T_2$ contains two leaves and $T_{i+1}=T_i[e_i]$ for a uniformly chosen pendant edge $e_i$ in $T_i$ for $2\leq i \leq n$. In particular, we have $|E^*(T_i)|=i$ for
$2\leq i \leq n+1$. Then %by Lemma~\ref{lem:leaf-set}
we have
%\small{
\begin{align}
\label{eq:total:yule}
\pyule&(\pf_{\scriptstyle n+1}=a, \ch_{n+1}=b) =\pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b) \notag \\
&=\sum_{p,q} \pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b\,|\,\pf(T_n)=p,\ch(T_n)=q) \pp(\pf(T_n)=p,\ch(T_n)=q) \notag \\
&=\sum_{p,q} \pp (\pf(T_{n+1})=a, \ch(T_{n+1})=b\,|\,\pf(T_n)=p,\ch(T_n)=q) \pp _{\YHK}(\pf_n=p,\ch_n=q),
\end{align}
where the first and second equalities follow from the law of total probability, and the definition of random variables $A_n$ and $C_n$.
Let $e_n$ be the pendant edge in $T_n$ chosen in the above YHK process for generating $T_{n+1}$, that is, $T_{n+1}=T_n[e_n]$. Since Corollary~\ref{cor:range} implies that
\begin{equation}
\label{eq:yule:5}
%\small{
\pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b~|~\pf(T_n)=p,\ch(T_n)=q)=0 %}
\end{equation}
for $(p,q) \not \in\{(a,b),(a+1,b-1),(a-1,b),(a,b-1)\}$,
it suffices to consider the following four cases in the summation in (\ref{eq:total:yule}): case (i): $p=a, q=b$; case (ii): $p=a+1, q=b-1$; case (iii): $p=a-1, q=b$; and case (iv): $p=a, q=b-1$.
Firstly, Proposition~\ref{prop:type} implies that case (i) occurs if and only if $e_n\in E_4(T_n)\cap E^*(T_n)=E_4^*(T_n)$.
Together with Lemma~\ref{lem:pen:edge-set}, we have
\begin{align}
\label{eq:yule:1}
\pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b~|~\pf(T_n)=a,\ch(T_n)=b) &=
\frac{|E^*_4(T_n)|}{|E^*(T_n)|}
=\frac{2\pf(T_n)}{n}=\frac{2a}{n}.
\end{align}
%\begin{equation}
%\label{eq:yule:1}
%\small{ \pp(\pf(T_{n+1})=a, \ch(T_{n+1}=b)|\pf(T_n)=a,\ch(T_n)=b)=
%\frac{|L_1(T_n)\cap L_2(T_n)|}{|L(T_n)|}=\frac{2\pf(T_n)}{n}=\frac{2a}{n}.}
%\end{equation}
Similarly, Proposition~\ref{prop:type} implies that case (ii)
occurs if and only if $e_n\in E_1(T_n)\cap E^*(T_n)=E_1(T_n)$.
Hence by Lemma~\ref{lem:edge-set} we have
\begin{align}
\label{eq:yule:2}
%\small{
\pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b~|~\pf(T_n)=a+1,\ch(T_n)=b-1)=
\frac{|E_1(T_n)|}{|E^*(T_n)|}
=\frac{a+1}{n}.% }
\end{align}
Next, Proposition~\ref{prop:type} implies case (iii) occurs if and only if $e_n\in E_2(T_n)\cap E^*(T_n)=E_2(T_n)$. Hence using Lemma~\ref{lem:edge-set} we have
\begin{align}
\label{eq:yule:3}
%\small{
\pp(\pf(T_{n+1})=a, \ch(T_{n+1})=b~|~\pf(T_n)=a+1,\ch(T_n)=b) =
\frac{|E_2(T_n)|}{|E^*(T_n)|}
=\frac{2(b-a-1)}{n}.%}
\end{align}
Finally, by Proposition~\ref{prop:type} case (iv) % Moreover, $\pf(T_{n+1})=\pf(T_{n})$ and $\ch(T_{n+1})=\ch(T_n)+1$
occurs if and only if $e_n$ is contained in $E_3(T_n)\cap E^*(T_n)=E^*_3(T_n)$. Hence by Lemma~\ref{lem:pen:edge-set} it follows that
\begin{equation}
\label{eq:yule:4}
%\small{
\pp(\pf(T_{n+1})=a, \ch(T_{n+1}=b)~|~\pf(T_n)=a,\ch(T_n)=b-1)=
\frac{|E^*_3(T_n)|}{|E^*(T_n)|}=\frac{n-a-2b+2}{n}. %}
\end{equation}
%Finally, because all possible choices of $e$ have been covered by the above four cases, we have
%\begin{equation}
%\label{eq:yule:5}
%\small{ \pp(\pf(T_{n+1})=a, \ch(T_{n+1}=b)|\pf(T_n)=p,\ch(T_n)=q)=0 }
%\end{equation}
%for $(p,q)\not \in\{(a,b),(a+1,b-1),(a-1,b),(a,b-1)\}$.
Now substituting Eq.~(\ref{eq:yule:1})--(\ref{eq:yule:4}) into Eq.~(\ref{eq:total:yule})
%and using the fact that $$ \pp(l_1(T_n)=p,l_2(T_n)=q)=\pp _{\YHK}(\pf_n=p,\ch_n=q)$$ holds for all $(p,q)$
completes the proof of the theorem.
\epf
\end{proof}
%Note first that the recursion in the last theorem provides a way of computing the joint distribution, and hence also the marginal distributions, of $\pf_n$ and $\ch_n$ under the YHK model. For example, their joint distribution with $n=200$ is depicted in Fig.~\ref{fig:jd:yhk} with $n=200$, which suggests the joint distribution can be well-approximated by a bivariate normal distribution.
%\begin{figure}[h]
%\begin{center}
%{\includegraphics[scale=0.7]{Fig/joint_YHK_200.pdf}}
%\end{center}
%\caption{The joint distribution of $\pf_n$ and $\ch_n$ under the YHK model with $n=200$.}
%\label{fig:jd:yhk}
%\end{figure}
The recursion in the last theorem can be used for a dynamic approach to numerically compute the joint distribution of $A_n$ and $C_n$ . More precisely, let $M_m$ ($m\geq 3$) be the $(m+1)\times (m+1)$ matrix whose $(i,j)$-entry is $\pyule(\pf_{m}=i-1, \ch_{m}=j-1)$. Then $M_3$ contains a unique non-zero entry, which is at position $(2,2)$ and has a value of $1$ . Next, starting with $m=4$ and assuming that $M_{m-1}$ is already constructed, each entry in $M_{m}$ can be computed using time $O(1)$, and hence $M_m$ can be constructed in time $O(m^2)$ with $M_{m-1}$ given. In other words, $M_n$, which specifies the joint distribution of cherry and pitchfork under the YHK model, can be computed in $O(n^3)$. Note that an alternative way of computing the joint distribution of $A_n$ and $C_n$ under the YHK model is proposed in~\cite{dw13}, which is based on integrating and differentiating generating functions.
For later use, we rewrite the recursion in Theorem~\ref{thm:yule:pf} in the following functional form.
%, which will be useful to study the moments of $\pf_n$ and $\ch_n$.
\begin{theorem}
\label{thm:yule:fun}
Let $\varphi: \mathbb{R} \times \mathbb{R} \to \mathbb{R}$ be an arbitrary function. Then, under the YHK model, we have
\begin{align*}
\ee _{\YHK} \varphi(\pf_{n+1}, \ch_{n+1}) &= \frac{2}{n} \ee _{\YHK}[\pf_{n} \: \varphi (\pf_{n}, \ch_{n})] + \frac{1}{n} \ee _{\YHK}[\pf_{n} \: \varphi (\pf_{n}-1, \ch_{n}+1)] \\
& \quad + \frac{2}{n} \ee _{\YHK}[ (\ch_{n}-\pf_n) \: \varphi (\pf_{n}+1, \ch_{n})] + \frac{1}{n} \ee _{\YHK}[(n-\pf_{n}-2\ch_{n}) \: \varphi (\pf_{n}, \ch_{n}+1)]
\end{align*}
for $n>2$.
\end{theorem}
\begin{proof}
Consider the indicator function $I_{(a, b)}$ on $\mathbb{R} \times \mathbb{R}$ defined as
$$I_{(a, b)}(x, y)=
\begin{cases}
1 & \mbox{ if $x=a$ and $y=b$,}\\
0 & \mbox{ otherwise.}
\end{cases}
$$
We multiply the equation in Theorem~\ref{thm:yule:pf} by $\varphi(a,b)$ and rewrite them as follows
{
\begin{align*}
&\ee _{\YHK}[\varphi(\pf_{n+1}, \ch_{n+1}) I_{(a, b)}(\pf_{n+1}, \ch_{n+1})]
= \frac{1}{n} {\big \{} 2 \ee _{\YHK}[\pf_n \varphi(\pf_{n}, \ch_{n}) I_{(a, b)}(\pf_{n}, \ch_{n})] \\
& \quad+ \ee _{\YHK}[\pf_n \varphi(\pf_{n}-1, \ch_{n}+1) I_{(a, b)}(\pf_{n}-1, \ch_{n}+1)]
+ 2 \ee _{\YHK}[(\ch_n-\pf_n) \varphi(\pf_{n}+1, \ch_{n}) I_{(a, b)}(\pf_{n}+1, \ch_{n})] \\
& \quad + \ee _{\YHK}[(n-\pf_n-2\ch_n) \varphi(\pf_{n}, \ch_{n}+1) I_{(a, b)}(\pf_{n}, \ch_{n}+1)] {\big \} } .
\end{align*}
}
Summing over all $a$ and $b$ completes the proof.
\epf
\end{proof}
In the remainder of this section we study cherry and pitchfork distributions using Theorem~\ref{thm:yule:fun}. We begin with a functional recursion on cherry distribution $\ch_n$. This enables us to show that the cherry distribution is log-concave under the YHK model, and obtain an alternative approach to computing the central moments of cherry distribution. \begin{proposition}
\label{prop:yule:ch:fun}
Let $\psi: \mathbb{R} \to \mathbb{R} $ be an arbitrary function. Then %under the YHK model
we have
\begin{align}
\label{eq:ch:yule:functional}
\ee _{\YHK} &\psi(\ch_{n+1}) = \frac{1}{n} \ee _{\YHK}[ 2\ch_n\:\psi (\ch_{n}) + (n-2\ch_{n}) \: \psi (\ch_{n}+1)]
\end{align}
for $n>2$. In particular, we have $\pp _{\YHK}(\ch_2=1)=1$, $\pp _{\YHK}(\ch_2=k)=0$ for $k\not =1$, and
\begin{equation}
\label{eq:cherry:yule}
\pp _{\YHK}(\ch_{n+1}=k) = \frac{2k}{n}\pp _{\YHK}(\ch_{n}=k)+\frac{n-2k+2}{n}\pp _{\YHK}(\ch_{n}=k-1)
\end{equation}
for $n>2$ and $1<k<n$.
\end{proposition}
\begin{proof}
For $\psi: \mathbb{R} \to \mathbb{R} $ given in the statement of the proposition, we define $\varphi^*(x,y)=\psi(y)$, a function on $\mathbb{R} \times \mathbb{R} $. Applying Theorem~\ref{thm:yule:fun} to the function $\varphi^*$ leads to Eq.~(\ref{eq:ch:yule:functional}). Eq.~(\ref{eq:cherry:yule}) follows from Eq.~(\ref{eq:ch:yule:functional}) by taking $\psi(x)=I_k(x)$, where $I_k(x)$ equals $1$ if $x=k$, and $0$ otherwise.
\epf
\end{proof}
Using the last proposition, the mean and the variance of cherry distribution
%Since similar techniques will be used to study pitchfork distribution under this model, a completed proof is included here.
can be obtained by substituting $\psi(x)=x$ and $\psi(x)=x^2$, respectively, in the recursive equation Eq.~(\ref{eq:ch:yule:functional}) in Proposition~\ref{prop:yule:ch:fun}. % show that $\ee _{\YHK}(\ch_n)
\begin{corollary}
\label{cor:yule:ch}
~\citep{heard92a,McKenzie2000}
%~(see, e.g. \cite[Theorem 2]{McKenzie2000})
We have $\ee _{\YHK}(\ch_n)=n/3$ for $n> 2$ and $\var _{\YHK}(\ch_n)=2n/45$ for $n\geq 5$.
%\epf
\end{corollary}
Recall that a sequence of numbers, $\{y_1,\dots,y_m\}$, is said to be {\em positive} if each
number in the sequence is greater than zero.
It is called {\em log-concave} if
$y_{k-1}y_{k+1}\leq y_k^2$ holds for $2\leq k \leq m-1$. Clearly, a positive sequence $\{y_k\}_{1\leq k \leq m}$ is log-concave if and only if the sequence $\{y_{k}/y_{k+1}\}_{1\leq k \leq m-1}$ is increasing. Therefore, a log-concave sequence is necessarily {\em unimodal}, that is, there exists an index $1\leq k \leq m$ such that
\begin{equation}
\label{def:unimodal}
y_1\leq y_2 \leq \dots \leq y_k \geq y_{k+1} \geq \cdots \geq y_m
\end{equation}
holds.
%Recall that a sequence $\{y_i\}_{1\leq i \leq m}$ is also called unimodal if $y_1\geq y_2 \geq \dots \geq y_k \leq y_{k+1} \leq \cdots \leq y_m$ holds for some $1\leq k\leq m$. However, in this paper, unimodal is always referred to the situation specified in Eq.~(\ref{def:unimodal}) unless stated otherwise.
Finally, a non-negative integer valued
random variable $Y$ with probability mass function $\{p_k: k\geq 0\}$ is log-concave if $\{p_k\}_{k\geq 0}$ is a log-concave sequence.
%with P° x=0 px = 1 (discrete distribution is log-concave)
To show that the probability density function of $\ch_n$ is log-concave, we need the following lemma.
\begin{lemma}
\label{lem:lc:four}
Let $z_1,z_2,z_3,z_4$ be four positive numbers with $z_2^2\geq z_1z_3$ and $z_3^2\geq z_2z_4$. Then we have
\begin{align*}
z_2z_3\geq z_1z_4~~~~~\mbox{and}~~~~~z_1z_3+z_2z_4\geq 2z_1z_4.
\end{align*}
% {\rm (i)} $z_2z_3\geq z_1z_4$ and {\rm (ii)} $z_1z_3+z_2z_4\geq 2z_1z_4$.
%\begin{itemize}
%\item[{\rm (i)}] $z_2z_3\geq z_1z_4$, and \\
%\item[{\rm (ii)}] $z_1z_3+z_2z_4\geq 2z_1z_4$
%\end{itemize}
\end{lemma}
\begin{proof}
Since $z_i$ are positive for $1\leq i \leq 4$, from $z_2^2\geq z_1z_3$ and $z_3^2\geq z_2z_4$ if follows that
$$
\frac{z_2}{z_1}\geq \frac{z_3}{z_2} \geq \frac{z_4}{z_3}.
$$
Hence we have
\begin{align}
\label{eq:first}
z_2z_3\geq z_1z_4,
\end{align}
which completes the proof of the first inequality in the lemma.\\
To prove the second inequality in the lemma, we consider the following two cases.
\noindent
{\bf Case 1:} $z_1\geq z_2$. Together with $z_2^2\geq z_1z_3$, this implies $z_2\geq z_3$, and hence $z_3\geq z_4$ in view of $z_3^2\geq z_2z_4$. Therefore, we have
$$
(z_1-z_2)(z_3-z_4)\geq 0.
$$
This leads to $z_1z_3+z_2z_4\geq z_1z_4+z_2z_3\geq 2z_1z_4$,
where the last inequality follows from Eq.~(\ref{eq:first}).
\noindent
{\bf Case 2:} $z_1< z_2$. If $z_3\leq z_4$, then we have $
(z_1-z_2)(z_3-z_4)\geq 0$, and hence $z_1z_3+z_2z_4\geq z_1z_4+z_2z_3\geq 2z_1z_4$, as required. Therefore, we may assume that $z_3>z_4$. This implies $z_1z_3\geq z_1z_4$ and $z_2z_4\geq z_1z_4$, and hence $z_1z_3+z_2z_4\geq 2z_1z_4$, as required.
\epf
\end{proof}
Using the last lemma, we present the following theorem concerning the log-concavity of the cherry distribution under the YHK model.
\begin{theorem}
\label{thm:lc:yule}
Under the YHK model, we have
\begin{equation}
\label{eq:yhk:lc}
\pp _{\YHK}(\ch_n=k)^2\geq \pp _{\YHK}(\ch_n=k+1)\pp _{\YHK}(\ch_n=k-1)
\end{equation}
for $n>2$ and $1<k<n$.
\end{theorem}
\begin{proof}
For simplicity, we put $a_{n,k}:=\pp _{\YHK}(\ch_n=k)$. We prove this theorem by induction; the basic case $n=3$ is straight-forward. Now assuming that $n\ge 3$ and Eq.~(\ref{eq:yhk:lc}) holds for all $1<k<n$, it suffices to show that
\begin{align}
\label{eq:log-concave:pf:yhk}
a_{n+1,k}^2\geq a_{n+1,k-1} a_{n+1,k+1}
\end{align}
for all $1<k\leq n$. Using the recursion described in Eq.~(\ref{eq:cherry:yule}), we have
$$
a_{n+1,k}^2= 4k^2a_{n,k}^2+(n+2-2k)^2a_{n,k-1}^2+4k(n+2-2k)a_{n,k}a_{n,k-1} $$
and $a_{n+1,k-1} a_{n+1,k+1}$ is equal to
\begin{align*}
(2k+2)(2k-2)&a_{n,k-1}a_{n,k+1}+(2k+2)(n-2k+4)a_{n,k-2}a_{n,k+1} \\
&+(n-2k)(2k-2)a_{n,k}a_{n,k-1}+(n-2k)(n-2k+4)a_{n,k}a_{n,k-2}.
\end{align*}
Therefore, by the inductive assumption and Lemma~\ref{lem:lc:four}
we have
%\small{
\begin{align*}
& a_{n+1,k}^2-a_{n+1,k-1} a_{n+1,k+1} \\
&= 2[k(n-2k)+(n+2k)](a_{n,k}a_{n,k-1}-a_{n,k-2}a_{n,k+1}) +4k^2(a^2_{n,k}-a_{n,k-1}a_{n,k+1}) \\
&\quad +4a_{n,k-1}(a_{n,k-1}-a_{n,k-2}) +4(n+2-2k)^2(a_{n,k+1}^2-a_{n,k}a_{n,k-2})+4a_{n,k-2}(a_{n,k}-a_{n,k+1})\\
&\geq 4a_{n,k+1}(a_{n,k-1}-a_{n,k-2})+4a_{n,k-2}(a_{n,k}-a_{n,k+1}) \\
&=4(a_{n,k-1}a_{n,k+1}+a_{n,k}a_{n,k-2}-2a_{n,k-2}a_{n,k+1}) \geq 0,
\end{align*}
from which Eq.~(\ref{eq:log-concave:pf:yhk}) follows, as required.
\epf
\end{proof}
\bigskip
In the next result we compute the mean and the variance of pitchfork distribution $\pf_n$ under the YHK model, and calculate the covariance and correlation of $\pf_n$ and $\ch_n$. Note that the mean and the variance of $\pf_n$ was also obtained by~\citet[Theorem 4.4]{rosenberg06a}. Since the proof is similar to that of Corollary~\ref{cor:yule:ch}, we only outline the main step here.
\begin{proposition}
\label{prop:yule:distr}
For $n\geq 7$ we have
\begin{equation}
\ee _{\YHK}(\pf_n)=\frac{n}{6},
\quad
\cov_{\YHK}(\pf_n,\ch_n)=-\frac{n}{45},
\quad
\mbox{and}
\quad
\var _{\YHK}(\pf_n)=\frac{23n}{420}.
\end{equation}
\end{proposition}
\begin{proof}
Applying Theorem~\ref{thm:yule:fun} to $\varphi(x,y)=x$ and using Corollary~\ref{cor:yule:ch}, it follows that
\begin{align*}
\ee _{\YHK}(\pf_{n+1})&=\frac{1}{n}\ee _{\YHK}[2\pf_n^2+\pf_n(\pf_n-1)+2(\ch_n-\pf_n)(\pf_n+1)+(n-\pf_n-2\ch_n)\pf_n] \\
%&=\frac{n-3}{n}\ee _{\YHK}(\pf_n)+\frac{2}{n} \ee _{\YHK}(\ch_n) \\
&=\frac{2}{3}+\frac{n-3}{n}\ee _{\YHK}(A_n)
\end{align*}
holds for $n>2$. Together with $\ee _{\YHK}(A_3)=1$, we have $\ee _{\YHK}(A_n)=n/6$ for $n\geq 4$, as required.\\
Next, applying Theorem~\ref{thm:yule:fun} to the function $\varphi(x,y)=xy$
shows that
%implies that
\begin{align*}
\ee _{\YHK}(\pf_{n+1}\ch_{n+1})
%=\frac{1}{n} \ee _{\YHK} \, \big [\, 2\pf^2_n\ch_n+\pf_n(\pf_n-1)(\ch_n+1)+2(\ch_n-\pf_n)(\pf_n+1)\ch_n+\\ &\quad \quad\quad (n-\pf_n-2\ch_n)\pf_n(\ch_n+1) \big] \\
=\frac{n-5}{n} \ee _{\YHK} (\pf_n\ch_n)+\frac{n-1}{n} \ee _{\YHK}(\pf_n)+\frac{2}{n}\ee _{\YHK}(\ch_n^2)
\end{align*}
holds for $n>2$. By Corollary~\ref{cor:yule:ch} and $ \ee _{\YHK}(\pf_n)=n/6$ it follows that
\begin{align*}
\cov_{\YHK}(\pf_{n+1},\ch_{n+1})
&=\ee _{\YHK}(\pf_{n+1}\ch_{n+1})-\frac{(n+1)^2}{18} \\
&=\frac{n-5}{n} \ee _{\YHK} (\pf_n\ch_n)+\frac{n-1}{6}+\frac{2(5n+2)}{45} -\frac{(n+1)^2}{18}\\
%&=\frac{n-5}{n}\big[\ee(\pf_n\ch_n)-\frac{n^2}{18}\big]+\frac{n(n-5)}{18}+\frac{n-1}{6}+\frac{2(5n+2)}{45} -\frac{(n+1)^2}{18} \\
&=\frac{n-5}{n}\cov_{\YHK}(\pf_n,\ch_n)-\frac{2}{15}
\end{align*}
holds for $n\geq 5$. Solving the last recursion equation, % on $\cov_{\YHK}(\pf_n,\ch_n)$,
we obtain $\cov_{\YHK}(\pf_n,\ch_n) = -n/45$ for $n\geq 6$, as required.
Now the formula on $\var _{\YHK}(\pf_n)$ can be established by an argument similar to that for $\cov_{\YHK}(\pf_n,\ch_n) $ by applying Theorem~\ref{thm:yule:fun} to the function $\varphi(x,y)=x^2$.
\epf
\end{proof}
Interestingly, the last proposition imply that the correlation coefficient between the cherry and pitchfork distribution under the YHK model is a negative constant for $n\ge 7$. Note that negative correlation is to be expected as the more cherries are found in a tree, the more likely that there are fewer pitchforks in that tree.
%$\pf_n$ and $\ch_n$ is a constant for $n \ge 7$. (This shows....todo)
\begin{corollary}
\label{cor:yule:cor}
Under the YHK model, the correlation coefficient $\corr _{\YHK}(\pf_n,\ch_n)$ between $\pf_n$ and $\ch_n$ is $-\sqrt{14/69}$, which is independent of $n$ for $n \ge 7$.
\end{corollary}
\begin{proof}
The proposition follows directly from Corollary~\ref{cor:yule:ch} and Proposition~\ref{prop:yule:distr}.
% using $$\corr _{\YHK}(\pf_n,\ch_n)=\frac{\cov _{\YHK}(\pf_n,\ch_n)}{\sqrt{\var _{\YHK}(\pf_n)\var _{\YHK}(\ch_n)}}.$$
\epf
%, the correlation coefficient between $\pf_n$ and $\ch_n$ is
%$$
%\frac{\cov(\pf_n,\ch_n)}{\sqrt{\var(\pf_n)\var(\ch_n)}}
%=\frac{-\frac{n}{45}}{\sqrt{\frac{23\times 2 n^2}{420\times 45}}}
%=-\sqrt{\frac{420}{45\times 2 \times 23}}=
%-\sqrt{\frac{14}{69}}
%$$
%for $n\geq 7$, as required.
\end{proof}
\section{Subtree Distributions under the PDA model}
\label{sec:PDA}
In this section, we shall investigate the cherry and pitchfork distributions under the PDA model. Similar to the study on the YHK model in Section~\ref{sec:YHK}, our starting point is the following recursion relating the joint distribution of cherries and pitchforks.
%{\bf Good to describe the model here?:} Fix $n> 2$, and let $T_2,\dots,T_n,T_{n+1}$ be a random tree sequence generated by the PDA process, that is, $T_2$ contains two leaves and $T_{i+1}$ is obtained from $T_i$ by uniformly choosing an edge $e=(u,v)$ in $T_i$ and add a new leaf to it for $3\leq i \leq n+1$.
\begin{theorem}
\label{thm:pda:pf}
We have
{
\begin{align*}
\puni&(\pf_{n+1}=a, \ch_{n+1}=b)
=\frac{n+3a-b-1}{2n-1}\puni(\pf_{n}=a,\ch_n=b)+\frac{a+1}{2n-1}\puni(\pf_{n}=a+1, \ch_n =b-1) \\
&\quad +\frac{3(b-a+1)}{2n-1}\puni(\pf_{n}=a-1, \ch_n=b)+\frac{n-a-2b+2}{2n-1}\puni(\pf_{n}=a, \ch_n=b-1)
\end{align*}
}
for $n>3$ and $1<b<n$.
Moreover, $\puni(\pf_3=a,\ch_3=b)$ equals to $1$ if $(a,b)=(1,1)$ and $0$ otherwise. %, $\puni(\pf_n=0)=0$, and
%In particular, we have $\puni(\pf_{n}=a, \ch_{n}=b)=0$ for $a>b$ or $a+2b>n$.
\end{theorem}
\begin{proof}
We give a sketch of the proof as it is similar to the proof of Theorem \ref{thm:yule:pf}.
The only modifications needed are the conditional
probabilities in the four cases there. For case (i),
by Proposition~\ref{prop:type} this case occurs if and only if $e_n\in E_4(T_n)$, and hence the conditional probability is
$|E_4(T_n)|/|E(T_n)|= (n+3a-b-1)/(2n-1)$
by Lemma~\ref{lem:edge-set}. Using similar arguments, for case (ii), the conditional probability is $|E_1(T_n)|/|E(T_n)|=
(a+1)/(2n-1)$. For case (iii), the conditional probability is $|E_2(T_n)|/|E(T_n)|= 3(b-a+1)(2n-1).$ Finally, for case (iv), the conditional probability is
$|E_3(T_n)/|E(T_n)|= (n-a-2b+2)/(2n-1).$ The rest of the proof proceeds as in the proof of Theorem \ref{thm:yule:pf}.
\epf
\end{proof}
%In particular, we have $|L(T_i)|=i$ for
%$2\leq i \leq n+1$. Then by Lemma~\ref{lem:edge-set} we have
%\small{
%\begin{align}
%\label{eq:total:pda}
%\puni&(\pf_{\scriptstyle n+1}=a, \ch_{n+1}=b) =\pp(l_1(T_{n+1})=a, l_2(T_{n+1})=b) \\
%&=\sum_{p,q} \pp _{\PDA}(l_1(T_{n+1})=a, l_2(T_{n+1})=b\,|\,l_1(T_n)=p,l_2(T_n)=q) \pp _{\PDA}(l_1(T_n)=p,l_2(T_n)=q) \notag \\
%&=\sum_{p,q} \pp _{\PDA}(l_1(T_{n+1})=a, l_2(T_{n+1})=b\,|\,l_1(T_n)=p,l_2(T_n)=q) \puni(\pf_n=p,\ch_n=q). \notag
%\end{align}
%}
%Let $e$ be the edge in $T_n$ chosen in the PDA process for generating $T_{n+1}$. %Then we have the following five cases to consider.\\
%%\noindent {\bf Case 1:}
%First note that $l_1(T_{n+1})=l_1(T_{n})$ and $l_2(T_{n+1})=l_2(T_n)$
% happens if and only if $e\in E_4(T)$. By Lemma~\ref{lem:edge-set} it follows that
%\begin{equation}
%\label{eq:pda:1}
%\small{ \pp(l_1(T_{n+1})=a, l_2(T_{n+1}=b)|l_1(T_n)=a,l_2(T_n)=b)
%=\frac{|E_4(T_n)|}{|E(T_n)|}=
%\frac{n+3a-b-1}{2n-1}.}
%\end{equation}
%Secondly, noting that $l_1(T_{n+1})=l_1(T_{n})-1$ and $l_2(T_{n+1})=l_2(T_n)+1$
% happens if and only if $e\in E_1(T)$, by Lemma~\ref{lem:edge-set} it follows that
%\begin{equation}
%\label{eq:pda:2}
%\small{ \pp(l_1(T_{n+1})=a, l_2(T_{n+1}=b)|l_1(T_n)=a+1,l_2(T_n)=b-1)
%=\frac{|E_1(T_n)|}{|E(T_n)|}=
%\frac{a+1}{2n-1}.}
%\end{equation}
%Similarly, as $l_1(T_{n+1})=l_1(T_{n})+1$ and $l_2(T_{n+1})=l_2(T_n)$
%occurs if and only if $e\in E_2(T)$, by Lemma~\ref{lem:edge-set} it follows that
%\begin{equation}
%\label{eq:pda:3}
%\small{ \pp(l_1(T_{n+1})=a, l_2(T_{n+1}=b)|l_1(T_n)=a-1,l_2(T_n)=b)
%=\frac{|E_2(T_n)|}{|E(T_n)|}=
%\frac{3(b-a+1)}{2n-1}.}
%\end{equation}
%Next, as $l_1(T_{n+1})=l_1(T_{n})$ and $l_2(T_{n+1})=l_2(T_n)+1$
%occurs if and only if $e\in E_3(T)$, by Lemma~\ref{lem:edge-set} it follows that
%\begin{equation}
%\label{eq:pda:4}
%\small{ \pp(l_1(T_{n+1})=a, l_2(T_{n+1}=b)|l_1(T_n)=a-1,l_2(T_n)=b)
%=\frac{|E_3(T_n)|}{|E(T_n)|}=
%\frac{n-a-2b+2}{2n-1}.}
%\end{equation}
%Finally, by Lemma~\ref{lem:edge-set} the edge $e$ is contained in $E_1(T_n)\cup E_2(T_n)\cup E_3(T_n) \cup E_4(T_n)$. Therefore we have
%\begin{equation}
%\label{eq:pda:5}
%\small{ \pp(l_1(T_{n+1})=a, l_2(T_{n+1}=b)|l_1(T_n)=p,l_2(T_n)=q)=0 }
%\end{equation}
%for $(p,q)\not \in\{(a,b),(a+1,b-1),(a-1,b),(a,b-1)\}$.
%Now substituting Eq.~(\ref{eq:pda:1}) - (\ref{eq:pda:4}) into Eq.~(\ref{eq:total:pda})
% completes the proof of the theorem.
%\end{proof}
%\begin{figure}[h]
%\begin{center}
%{\includegraphics[scale=0.7]{Fig/joint_PDA_200.pdf}}
%\end{center}
%\caption{Joint distribution of cherry and pitchfork under the PDA model with $n=200$.}
%\label{fig:jd:pda}
%\end{figure}
%Note that the recursion in the last theorem provides a way of computing the exact joint distributions of cherry and pitchfork under the PDA model.
% For example, the joint distribution with $n=200$ is depicted in Fig.~\ref{fig:jd:pda} for an example with $n=200$, which suggests that the joint distribution contains only one peak. Compared to the joint distribution under the YHK model (see, Fig.~\ref{fig:jd:yhk}), it is clearer the coordinates for the peak under the PDA model is smaller than those under the YHK model. This observation is in line with the comparison results on the means of cherry and pitchfork under these two models (see Corollary~\ref{cor:ch:comp} and ~\ref{cor:pf:comp}).
Using an approach similar to the remark after Theorem~\ref{thm:yule:pf}, the last theorem leads to a dynamic programming approach to compute the joint distribution of cherry and pitchfork. In addition, we present the following result which
will enable us to study the moments of $\pf_n$ and $\ch_n$, whose proof
is similar to that of Theorem~\ref{thm:yule:fun} and hence omitted.
\begin{theorem}
\label{thm:pda:fun}
Let $\varphi: \mathbb{R} \times \mathbb{R} \to \mathbb{R} $ be an arbitrary function. For $n>3$ we have
\begin{align*}
& \ee _{\PDA} \varphi(\pf_{n+1}, \ch_{n+1}) \\ = & \frac{1}{2n-1} \ee _{\PDA}[(n+3\pf_{n}-\ch_n-1) \: \varphi (\pf_{n}, \ch_{n})] + \frac{1}{2n-1} \ee _{\PDA}[\pf_{n} \: \varphi (\pf_{n}-1, \ch_{n}+1)] \\
& \quad + \frac{3}{2n-1} \ee _{\PDA}[ (\ch_{n}-\pf_n) \: \varphi (\pf_{n}+1, \ch_{n})] + \frac{1}{2n-1} \ee _{\PDA}[(n-\pf_{n}-2\ch_{n}) \: \varphi (\pf_{n}, \ch_{n}+1)].
\end{align*}
\end{theorem}
%$\hfill \square$
%\begin{proof}
%The proof is similar to that of Theorem~\ref{thm:yule:fun}.
% Consider the indicator function $I_{(a, b)}(x, y)$ on $\mathbb{R} \times \mathbb{R}$ defined as
%$$I_{(a, b)}(x, y)=
%\begin{cases}
%1 & \mbox{ if $x=a$ and $y=b$,}\\
%0 & \mbox{ otherwise.}
%\end{cases}
%$$
%We multiply the equation in Theorem~\ref{thm:pda:pf} by $\varphi(a,b)$ and rewrite them as follows
%\small{
%\begin{align*}
% &\ee[\varphi(\pf_{n+1}, \ch_{n+1}) I_{(a, b)}(\pf_{n+1}, \ch_{n+1})]
%= \frac{1}{2n-1} {\big \{} \ee[(n+3\pf_n-\ch_n+1) \varphi(\pf_{n}, \ch_{n}) I_{(a, b)}(\pf_{n}, \ch_{n})] \\
%& \quad+ \ee[\pf_n \varphi(\pf_{n}-1, \ch_{n}+1) I_{(a, b)}(\pf_{n}-1, \ch_{n}+1)]
% + 3 \ee[(\ch_n-\pf_n) \varphi(\pf_{n}+1, \ch_{n}) I_{(a, b)}(\pf_{n}+1, \ch_{n})] \\
% & \quad + \ee[(n-\pf_n-2\ch_n) \varphi(\pf_{n}, \ch_{n}+1) I_{(a, b)}(\pf_{n}, \ch_{n}+1)] {\big \} } .
%\end{align*}
%}
%Summing over all $a$ and $b$ completes the proof.
%\end{proof}
\bigskip
In the remainder of this section we shall apply Theorem~\ref{thm:pda:fun}
to study cherry and pitchfork distributions under the PDA model. To begin with, we present the following functional recursion between cherry distribution, which will enable us to obtain the exact formula for cherry distribution and show that cherry distribution is log-concave under this model.
\begin{proposition}
\label{prop:pda:ch:fun}
Let $\psi: \mathbb{R} \to \mathbb{R}$ be an arbitrary function. Then for $n>2$ we have
\begin{align}
\label{eq:ch:pda:fun}
\ee _{\PDA} &\psi(\ch_{n+1}) = \frac{1}{2n-1} \ee _{\PDA}[(n+2\ch_n-1) \: \psi (\ch_{n})] + \frac{1}{2n-1} \ee _{\PDA}[(n-2\ch_{n}) \: \psi (\ch_{n}+1)]
\end{align}
%for $n>2$. Moroever, we have $\pp _{\PDA}(\ch_2=k)=1$ if $k=0$, and $0$ otherwise. Moreover, we have
and
\begin{equation}
\label{eq:cherry:pda}
\pp _{\PDA}(\ch_{n+1}=k) = \frac{n+2k-1}{2n-1}\pp _{\PDA}(\ch_{n}=k)+\frac{n-2k+2}{2n-1}\pp _{\PDA}(\ch_{n}=k-1), \quad { 1\leq k<n}.
\end{equation}
\end{proposition}
\begin{proof}
Let $\psi: \mathbb{R} \to \mathbb{R}$ be an arbitrary function as in the statement of the proposition. Then $\varphi^*(x,y)=\psi(y)$ is an function on $\mathbb{R} \times \mathbb{R}$. Now applying Theorem~\ref{thm:pda:fun} to the function $\varphi^*$ leads to Eq.~(\ref{eq:ch:pda:fun}). Finally, Eq.~(\ref{eq:cherry:pda}) follows from Eq.~(\ref{eq:ch:pda:fun}) by taking $\psi(x)=I_k(x)$, where $I_k(x)$ equals $1$ if $x=k$, and $0$ otherwise.
\epf
\end{proof}
Note that the recursion presented in the last proposition enables us to study the moments of cherry distribution under the PDA model. As an example, we present below an alternative computation for the mean and the variance of $\ch_n$. Since the techniques used to solve difference equations under this model is rather different from that used under the YHK model (i.e., Corollary~\ref{cor:yule:ch}), a complete proof is included here.
Note that in the proof we will use the following well-known Faulhaber's formulae (also known as Bernoulli's formulae) concerning the sum of powers of integers ~\citep[see e.g.][]{con96}.
%, which we may refer to without explicitly mentioning.
\begin{align*}
%\sum_{i=1}^n i &=& \frac{n(n+1)}{2} \\
\sum_{i=1}^n i^2 &= \frac{n(n+1)(2n+1)}{6},
\hspace{.8in}
&\sum_{i=1}^n i^3 &= \frac{n^2(n+1)^2}{4}, \\
\sum_{i=1}^n i^4 &= \frac{n(n+1)(2n+1)(3n^2 + 3n-1)}{30},
%\hspace{.2in}
&\sum_{i=1}^n i^5 &= \frac{n^2 (n+1)^2 (2n^2 + 2n-1)}{12}.
\end{align*}
\begin{corollary}
\label{cor:pda:ch}
~\citep[Proposition 5]{chang2010limit}
For $n\geq 2$ we have
%$\ee _{\PDA}(\ch_n)=\frac{n(n-1)}{2(2n-3)}$ for $n\geq 2$. In addition, we have
$$
\ee _{\PDA}(\ch_n)=\frac{n(n-1)}{2(2n-3)} \sim \frac{n}{4}
~~~\mbox{and}~~~
\var _{\PDA}(\ch_n)=\frac{n(n-1)(n-2)(n-3)}{2(2n-3)^2(2n-5)}
\sim \frac{n}{16}.
$$
%$\var(\ch_3)=0$, $\var(\ch_4)=2/9$, and $\var(\ch_n)=2n/45$ for $n\geq 5$.
\end{corollary}
\begin{proof}
We may assume that $n\ge 3$ in the remainder of the proof as the case $n=2$ clearly holds. Substituting $\psi(x)=x$ in the recursive equation Eq.~(\ref{eq:ch:pda:fun}) in Proposition~\ref{prop:pda:ch:fun} leads to that
\begin{align*}
\ee _{\PDA}(\ch_{n+1})&=\frac{1}{2n-1}\ee _{\PDA}\big[ (n+2\ch_n-1)\ch_n+(n-2\ch_n)(\ch_n+1)\big] =\frac{n}{2n-1}+\frac{2n-3}{2n-1}\ee _{\PDA}(\ch_n)
\end{align*}
holds for $n>2$. Together with the initial condition $\ee _{\PDA}(\ch_2)=1$, multiplying the both sides of the last difference equation on $\ee _{\PDA}(\ch_n)$ by $2n-1$ and solving it leads to
$$(2n-3)\ee _{\PDA}(\ch_n)=1+\cdots+(n-1)=\frac{n(n-1)}{2}$$ for $n>2$, as required.
\bigskip
For simplicity, let $f(n)=(2n-3)(2n-5)$ and $g(n)=(n-1)(n^2-2n-1)$.
Then
$$
\sum_{k=1}^n g(k)=\sum_{k=1}^n (k^3-3k^2+k+1)=\frac{n(n-1)(n^2-n-4)}{4}.
$$
Next, applying Proposition~\ref{prop:pda:ch:fun} to the function $\psi(x)=x^2$ implies that
\begin{align*}
\ee _{\PDA}(\ch_{n+1}^2) &=\frac{1}{2n-1} \ee _{\PDA} \, \big [\, (n+2\ch_n-1)\ch_n^2+(n-2\ch_n)(\ch_n+1)^2 \big] \\
&=\frac{n}{2n-1}+\frac{2n-2}{2n-1}\ee _{\PDA}(\ch_n)+\frac{2n-5}{2n-1} \ee _{\PDA}(\ch^2_n) \\
%&=\frac{n(n^2-2)}{(2n-1)(2n-3)}+\frac{2n-5}{2n-1} \ee(\ch^2_n)
&=\frac{g(n+1)}{(2n-1)(2n-3)}+\frac{2n-5}{2n-1} \ee _{\PDA}(\ch^2_n)
\end{align*}
holds for $n>2$. Now multiplying $f(n+1)$ on both sides of the above recursion leads to
$$
f(n)\ee _{\PDA}(\ch_{n}^2)-f(n-1)\ee _{\PDA}(\ch_{n-1}^2)=g(n)
$$
for $n\geq 3$. Since $\ee _{\PDA}(C_2^2)=1=-g(2)$ and $g(1)=0$, we have
$$
(2n-3)(2n-5)\ee _{\PDA}(\ch_n^2)=f(n)\ee _{\PDA}(\ch_n^2)=\sum_{k=1}^n g(k)
=\frac{n(n-1)(n^2-n-4)}{4}
$$
for $n\geq 3$, from which we have
\begin{align}
\label{eq:ch:u:sec}
\ee _{\PDA}(\ch^2_n)=\frac{n(n-1)(n^2-n-4)}{4(2n-3)(2n-5)}
\end{align}
and hence $\var _{\YHK}(\ch_n)$ follows.
\epf
\end{proof}
Another consequence of the recursion in Proposition~\ref{prop:pda:ch:fun}
is the following exact formula on cherry distribution for the PDA model, whose proof is a straightforward application of induction and hence omitted here.
\begin{theorem}
\label{ch:pda}
For $n\geq 2$ and $1\leq k\leq n/2$ we have
\begin{equation}
\label{eq:chDistr:pda}
\pp _{\PDA}(\ch_n=k)=
\frac{n!(n-1)!(n-2)!2^{n-2k}}{(n-2k)!(2n-2)!k!(k-1)!}~.
%\frac{n!\,\nbt(k)\,{n-2 \choose 2k-2}}{k!\, \nbt(n) \,2^k}
%\frac{(n-2k)!{n \choose 2k}{n-2 \choose 2k-2}}{\nbt(n)}.
\end{equation}
%\begin{equation}
%\label{eq:chDistr:pda}
%\pp _{\PDA}(\ch_n=k)=
%\frac{(2k)!\nbt(k)}{k!2^k}
%\frac{(n-2k)!{n \choose 2k}{n-2 \choose 2k-2}}{\nbt(n)}.
%\end{equation}
\end{theorem}
%\begin{proof}
%There are two approaches to establish the above theorem. One is a conceptual proof, which is based on counting the number of trees with exactly $k$-cherries. The other is by induction, which is based on the recursion formula. Here we shall proceed with the second one, as it is relatively easy.
%We shall establish Eq.~(\ref{eq:chDistr:pda}) by induction. The basic case, $n=2$ and $k=1$, clearly holds. To establish the induction step, assuming that it holds for some integer $n\geq 2$ and we shall show that it also holds for $n+1$. Since the case $k=1$ can be established in a similar manner, we may further assume that $k>1$ for simplicity.
%Now by Eq.~(\ref{eq:cherry:pda}) in Proposition~\ref{prop:pda:ch:fun} and the induction assumption we have
%\begin{align*}
%\pp _{\PDA}(\ch_{n+1}=k)
%&=\frac{n+2k-1}{2n-1}
%\frac{n!(n-1)!(n-2)!2^{n-2k}}{(n-2k)!(2n-2)!k!(k-1)!}+\frac{n-2k+2}{2n-1}
%\frac{n!(n-1)!(n-2)!2^{n-2k+2}}{(n-2k+2)!(2n-2)!(k-1)!(k-2)!} \\
%&=\frac{n!(n-1)!(n-2)!2^{n-2k}}{(2n-1)!(n-2k+1)!k!(k-1)!}
%\Big[(n+2k-1)(n-2k+1)+4k(k-1) \Big] \\
%&= \frac{(n+1)!n!(n-1)!2^{n+1-2k}}{(n-2k+1)!(2n-2)!k!(k-1)!},
%\end{align*}
%as required.
%\epf
%\end{proof}
Interestingly, a similar formula for unrooted trees was obtained by~\citet{hendy1982branch}, that is, the probability that a random tree generated by the PDA model contains exactly $k$ cherries is
\begin{equation}
\label{eq:chDistr:unrooted:pda}
\frac{n!(n-2)!(n-4)!2^{n-2k}}{(n-2k)!(2n-4)!k!(k-2)!}
\end{equation}
for $2\leq k \leq n/2$~\citep[see, also][Theorem 4]{McKenzie2000}.
% for a related result on unrooted phylogenetic trees.
A direct consequence of Theorem \ref{ch:pda} is that the cherry distribution under the PDA model is log-concave, and hence also unimodal.
%Compared to Theorem~\ref{thm:lc:yule} on the YHK model, this result
\begin{theorem}
\label{thm:lc:pda}
For $n\geq 2$ and $1<k<n$ we have
\begin{equation}
\label{eq:chDistr:pda:convex}
\pp _{\PDA}(\ch_n=k)^2\geq \pp _{\PDA}(\ch_n=k+1)\pp _{\PDA}(\ch_n=k-1).
\end{equation}
Moreover, let
$
\Delta(n)=\frac{(n+1)(n+2)}{2(2n+1)}.
$
Then
$$
\pp _{\PDA}(\ch_n=k-1) < \pp _{\PDA}(\ch_n=k)
~\mbox{for $1<k < \Delta(n)$, and }
\pp _{\PDA}(\ch_n=k) > \pp _{\PDA}(\ch_n=k+1)
~\mbox{for $\Delta(n)\leq k <n/2$}.
$$
\end{theorem}
\begin{proof}
Since $\pp _{\PDA}(\ch_n=k)=0$ for $k>n/2$, the theorem clearly holds for $k\geq n/2-1$. Hence in the remainder of the proof we may assume $k< (n-2)/2$. Now by Theorem~\ref{ch:pda} we have
\begin{align}
\frac{\pp _{\PDA}(\ch_n=k-1)}{\pp _{\PDA}(\ch_n=k)}
%&=\frac{(2k-2)!\nbt(k-1)(n-2k+2)!{n \choose 2k-2}{n-2 \choose 2k-4} k! 2^k}{(2k)!\nbt(k)(n-2k)!{n\choose 2k}{n-2 \choose 2k-2}(k-1)!2^{k-1}} \notag\\
%&=\frac{\nbt(k-1)k!2^k{n-2 \choose 2k-4}}{\nbt(k)(k-1)!2^{k-1}{n-2 \choose 2k-2}}
=\frac{4k(k-1)}{(n-2k+1)(n-2k+2)} :=g(k,n). \label{eq:ch:pda:inc}
\end{align}
Considering the function %$g(k,n):=\frac{4k(k-1)}{(n-2k+1)(n-2k+2)}$,
$g(k,n)$ defined in Eq.~(\ref{eq:ch:pda:inc}),
then $g(k+1,n)> g(k,n)$ holds for $1<k<(n-2)/2$. This, together with Eq.~(\ref{eq:ch:pda:inc}), completes the proof of Eq.~(\ref{eq:chDistr:pda:convex}).
The second part of the theorem follows from the observation that $g(k,n)> 1$ if and only if $k\geq \Delta(n)$.
\epf
\end{proof}
Now we apply Theorem~\ref{thm:pda:fun} to study pitchfork distribution, and the joint distribution between pitchforks and cherries under the PDA model.
Note that the mean and the variance of pitchfork distributions under this model were also derived by~\citet[Proposition 5]{chang2010limit}.
Since the proof is similar to that in Corollary~\ref{cor:pda:ch}, we only outline the main steps used here.
%We begin with an alternative approach to computing the mean of $\pf_n$. % which was initially obtained by....
\begin{proposition}
\label{prop:pda:distr}
For $n\geq 3$ we have
\begin{align}
\ee _{\PDA}(\pf_n)&=\frac{n(n-1)(n-2)}{2(2n-3)(2n-5)} \sim\frac{n}{8}, \label{eq:pda:ch:mean}\\
{\cov_{\PDA}}(\pf_n,\ch_n)&=\frac{-n(n-1)(n-2)(n-3)}{2(2n-3)^2(2n-5)(2n-7)}
\sim -\frac{n}{32},\label{eq:pda:ch:pf:cov}\\
\var_{\PDA}(\pf_n)&=\frac{3n(n-1)(n-2)(n-3)(4n^3-40n^2+123n-110)}{4(2n-3)^2(2n-5)^2(2n-7)(2n-9)}
\sim
\frac{3n}{64}.
\label{eq:pda:pf:mean}
\end{align}
\end{proposition}
\begin{proof}
Applying Theorem~\ref{thm:pda:fun} to the function $\varphi(x,y)=x$ and using Corollary~\ref{cor:pda:ch}, we have
\begin{align*}
\ee _{\PDA}(\pf_{n+1})
% &=\frac{\ee _{\PDA}[(n+3\pf_n-\ch_n-1)\pf_n+\pf_n(\pf_n-1)+3(\ch_n-\pf_n)(\pf_n+1)+(n-\pf_n-2\ch_n)\pf_n]}{2n-1} \\
%&=\frac{2n-5}{2n-1}\ee _{\PDA}(\pf_n)+\frac{3}{2n-1} \ee _{\PDA}(\ch_n) \\
=\frac{3n(n-1)}{2(2n-1)(2n-3)}+\frac{2n-5}{2n-1}\ee _{\PDA}(A_n)
\end{align*}
for $n>2$. Now Eq.~(\ref{eq:pda:ch:mean}) follows by solving the last recursion with an approach similar to that in Corollary~\ref{cor:pda:ch}.\\
To this end, applying Theorem~\ref{thm:pda:fun} to the function $\varphi(x,y)=xy$ implies that
\begin{align*}
\ee _{\PDA}(\pf_{n+1}\ch_{n+1}) =\frac{2n-7}{2n-1} \ee _{\PDA} (\pf_n\ch_n)+\frac{n(n-1)(5n^2-9n-8)}{4(2n-1)(2n-3)(2n-5)}
\end{align*}
holds for $n>2$. Solving this recursion we have\begin{align}
\label{eq:u:mix}
\ee _{\PDA}(\pf_n\ch_n)=\frac{n(n-1)(n-2)(n^2-3n-2)}{4(2n-3)(2n-5)(2n-7)},
\end{align}
from which Eq.~(\ref{eq:pda:ch:pf:cov}) follows.
Finally, applying Theorem~\ref{thm:pda:fun} to the function $\varphi(x,y)=x^2$ shows that
\begin{align*}
\ee _{\PDA} (\pf^2_{n+1}) =\frac{2n-9}{2n-1} \ee _{\PDA} (\pf^2_n)+\frac{g(n+1)}{4(2n-1)(2n-3)(2n-5)(2n-7)} \\
\end{align*}
holds for $n>2$.
Solving the above recursion leads to
$$
\ee _{\PDA}\pf_{n}^2 = \frac{n(n-1)(n-2)(n^3-4n^2-17n+66)}{4(2n-3)(2n-5)(2n-7)(2n-9)},
$$
from which Eq.~(\ref{eq:pda:pf:mean}) follows.
\epf
\end{proof}
We end this section with the following correlation result for the PDA model.
\begin{corollary}
\label{cor:pda:cor}
For $n \geq 4$ we have
\begin{align}
\label{eq:corr:yule}
\corr _{\PDA}(\pf_n,\ch_n)=-\sqrt{\frac{2(2n-5)(2n-9)}{3(2n-7)(4n^3-40n^2+123n-110)}}
\sim -\frac{1}{\sqrt{3}\,n}.
\end{align}
In addition, $\{|\corr _{\PDA}(\pf_n,\ch_n)|\}_{n\geq 4}$ is a decreasing sequence converging to $0$.
\end{corollary}
\begin{proof}
Note first that Eq.~(\ref{eq:corr:yule}) follows from Corollary~\ref{cor:pda:ch} and Proposition~\ref{prop:pda:distr}. Since the sequence $\{|\corr _{\PDA}(\pf_n,\ch_n)|\}_{n\geq 4}$ clearly approaches $0$, it remains to show that this sequence is decreasing. To this end, it suffices to show that the ratio
$$
R(n)=\frac{\corr _{\PDA}(\pf_n,\ch_n)^2}{\corr _{\PDA}(\pf_{n+1},\ch_{n+1})^2}
$$
is greater than 1 for $n\geq 4$.
Using Eq.~(\ref{eq:corr:yule}), we have
$$
R(n)=\frac{(2n-5)^2(2n-9)(4(n+1)^3-40(n+1)^2+123(n+1)-110}{(2n-7)^2(2n-3)(4n^3-40n^2+123n-110)}.
$$
By numerical computation, we can check that $R(n)>1$ for $4\leq n \leq 15$, therefore we may assume that $n>15$ in the remainder of the proof. Now denoting the numerator and denominator of $R(n)$ by $R_1(n)$ and $R_2(n)$, respectively, then we have
\begin{align*}
R_1(n)-R_2(n)&=64n^5-944n^4+5408n^3-15048n^2+20436n-10995 \\
&> 64n^4(n-15)+5408n^3(n-15)+20436(n-15)
> 0
\end{align*}
for $n>15$. This implies $R(n)=R_1(n)/R_2(n)>1$ for $n>15$, as required.
\epf
\end{proof}
\section{A comparative study of two models}
\label{sec:comp}
%Based on the results from the previous two sections,
%we have studied the cherry and pitchfork distributions under the YHK and the PDA model, respectively. Based on these results,
In this section, we compare and contrast the distributional properties of the number of cherries and the pitchforks in random trees generated under the YHK and the PDA models. %In this section
To begin with, note that the recursions in Theorems~\ref{thm:yule:pf} and~\ref{thm:pda:pf} provide us exact computation of the joint distribution, and hence also the marginal distributions, of $\pf_n$ and $\ch_n$ under the two models. For example, the joint distributions with $n=200$ for the two models are depicted in Fig.~\ref{fig:jd}, which suggests the joint distributions under both models can be well-approximated by bivariate normal distributions.
In addition, when $n=200$, the maximum for the joint distribution under the YHK model is $0.177...$, which is attained at $(\pf_n,\ch_n)=(33,67)$. On the other hand, under the PDA model the maximum is $0.145...$, which is attained at $(\pf_n,\ch_n)=(25,50)$.
%This suggests
%that in general the maximum for the joint distribution under the YHK model is greater than that under the PDA model. Moreover, the numbers of cherries and pitchforks at which the maximum is attained under the YHK model are greater than those under the PDA model, respectively, which is in line with the following comparison results on the means of the cherry and the pitchfork distributions under these two models.
% (see Corollary~\ref{cor:ch:comp} and ~\ref{cor:pf:comp}).
On average, the result below shows that trees generated by the YHK model contain more cherries and more pitchforks than trees of the same size generated by the PDA model.
%Note that the recursion in the last theorem provides a way of computing the exact joint distributions of cherry and pitchfork under the PDA model. For example, the joint distribution with $n=200$ is depicted in Fig.~\ref{fig:jd:pda} for an example with $n=200$, which suggests that the joint distribution contains only one peak. Compared to the joint distribution under the YHK model (see, Fig.~\ref{fig:jd:yhk}), it is clearer the coordinates for the peak under the PDA model is smaller than those under the YHK model. This observation is in line with the comparison results on the means of cherry and pitchfork under these two models (see Corollary~\ref{cor:ch:comp} and ~\ref{cor:pf:comp}).
\begin{proposition}
\label{prop:ch:comp}
%We have $\ee _{\YHK}(C_3)=\ee _{\PDA}(C_3)=1$.
For $n> 3$, we have
\begin{align}
\label{eq:ch:mean:comp}
\ee_{\PDA}(C_n)<\ee_{\YHK}(C_n) <\frac{4}{3}\ee_{\PDA}(C_n)
%\ee_{\YHK}(C_n) = \left[1 + \frac{n-3}{3(n-1)} \right ] \ee_{\PDA}(C_n) >\ee_{\PDA}(C_n)
\end{align}
and
\begin{align}
\label{eq:pf:mean:comp}
\ee _{\PDA}(A_n)<\ee _{\YHK}(A_n)<\frac{4}{3}\ee _{\PDA}(A_n).
%\ee _{\YHK}(A_n) =\left[1+ \frac{n^2-7n +9}{3(n-1)(n-2)}\right] \ee _{\PDA}(A_n) >\ee _{\PDA}(A_n).
\end{align}
%Moreover, $\big\{\frac{\ee _{\YHK}(C_n)}{\ee _{\PDA}(C_n)} \big\}_{n>3}$ and $\big\{\frac{\ee _{\YHK}(A_n)}{\ee _{\PDA}(A_n)} \big\}_{n>3}$ are monotonically increasing sequences with a common limit $4/3$.
\end{proposition}
\begin{proof}
By Corollaries~\ref{cor:yule:ch} and~\ref{cor:pda:ch} we have
$$
\ee_{\YHK}(C_n) = \left[1 + \frac{n-3}{3(n-1)} \right ] \ee_{\PDA}(C_n),
$$
from which Eq.~(\ref{eq:ch:mean:comp}) follows.
Similarly, by Propositions~\ref{prop:yule:distr} and~\ref{prop:pda:distr} we have
$$
\ee _{\YHK}(A_n) =\left[1+ \frac{n^2-7n +9}{3(n-1)(n-2)}\right] \ee _{\PDA}(A_n),
$$
from which Eq.~(\ref{eq:pf:mean:comp}) follows.
\epf
\end{proof}
Next, we study the variances of cherry and pitchfork distributions under the two models.
\begin{proposition}
For $n> 5$, we have
$$\frac{32}{45}\var _{\YHK}(C_n)<\var _{\PDA}(C_n)< \frac{49}{54} \var _{\YHK}(C_n).$$
%Moreover, $\big\{\frac{\var _{\YHK}(C_n)}{\var _{\PDA}(C_n)} \big\}_{n>4}$ is a monotonically decreasing sequence with limit $32/45$.
\end{proposition}
\begin{proof}
Let $R_n = \var _{\YHK}(C_n)/\var _{\PDA}(C_n)$. Then by Corollaries~\ref{cor:yule:ch} and~\ref{cor:pda:ch} we have
$$
R_n = \frac{4(2n-3)^2(2n-5)}{45(n-1)(n-2)(n-3)}. $$
This implies
$$\frac{R_n}{R_{n+1}} = \frac{n(2n-3)(2n-5)}{(n-3)(2n-1)^2} = 1 + \frac{2n+3}{(n-3)(2n-1)^2}>1,$$
and hence that $R_n$ is decreasing in $n$. Noting that
$\lim_{n\to \infty} R_n = \frac{32}{45}$, we have
$$ \frac{32}{45} < R_n < R_5 = \frac{49}{54} <1,
$$
from which the proposition follows.
\epf
\end{proof}
%The last result shows that the variance of cherry distribution under the YHK model is smaller than that under the PDA model. However, the opposite is true for pitchfork distribution.
\begin{proposition}
\label{prop:pf:comp}
For $n\geq 7$, we have
$$
1.168\,{\var _{\PDA}(A_n)}<\frac{368}{315}{\var _{\PDA}(A_n)}<\var _{\YHK}(A_n)<\frac{8349}{6520}{\var _{\PDA}(A_n)}<1.281\, {\var _{\PDA}(A_n)} .
$$
%Moreover, $\big\{\frac{\var _{\YHK}(A_n)}{\var _{\PDA}(A_n)} \big\}_{n>7}$ is a monotonically decreasing sequence with limit $368/315$.
\end{proposition}
\begin{proof}
Let $R_n = \var _{\YHK}(A_n)/\var _{\PDA}(A_n) $ for $n\geq 7$. Then by Proposition~\ref{prop:yule:distr} and~\ref{prop:pda:distr} we have
$$ R_n = \frac{23(2n-3)^2(2n-5)^2(2n-7)(2n-9)}{315(n-1)(n-2)(n-3)(4n^3-40n^2+123n -110)}, $$
%This implies $\lim_{n \to \infty} =\frac{368}{315}$. Moreover, we have
and hence
\begin{align*}
\frac{R_n}{R_{n+1}} &= \frac{n(2n-5)(2n-9)(4n^3-28n^2+55n-23)}{(n-3)(2n-1)^2(4n^3-40n^2+123n-110)} \\
&= 1+ \frac{2(24n^3-180n^2+382n-165)}{(n-3)(2n-1)^2(4n^3-40n^2+123n-110)}>1.
\end{align*}
Therefore, $R_n$ is strictly decreasing in $n$ and we have
$$
\frac{\var _{\YHK}(A_n)}{\var _{\PDA}(A_n)}
=R_n > \lim_{m \to \infty} R_m
= \frac{23 \cdot 4 \cdot 4\cdot 4 }{315 \cdot 4} =\frac{368}{315}>1.168\,.
$$
This, together with $R_7=\frac{8349}{6520}<1.281$, completes the proof.
\epf
\end{proof}
Proposition~\ref{prop:ch:comp} shows that trees generated by the YHK model have smaller variation in the number of cherries than trees of the same size generated by the PDA model. On the contrary, Proposition~\ref{prop:pf:comp} shows that YHK model generates trees with larger variation in the number of pitchforks than the PDA model does. This is not unexpected as the covariances of cherries and pitchforks are found to be negative by Propositions~\ref{prop:yule:distr} and~\ref{prop:pda:distr}.
%We end this section with
Now we present a result concerning the correlation coefficients between the cherry and the pitchfork distributions under the two models. Since these two distributions are negatively correlated, we will focus on their absolute values.
%the absolute value of the coefficients.
\begin{proposition}
For $n\geq 7$, we have
$|\corr _{\YHK}(A_n, C_n)|\geq {|\corr _{\PDA}(A_n, C_n)|}.
$
Moreover, $\big\{\frac{\corr _{\PDA}(A_n, C_n)}{\corr _{\YHK}(A_n, C_n)} \big\}_{n\ge 7}$ is a monotonically decreasing sequence with limit $0$.
\end{proposition}
\begin{proof}
This follows from Corollary~\ref{cor:yule:cor} and~\ref{cor:pda:cor} and the observation that \\
$$
|{\corr _{\PDA}}(A_7,C_7)|=\sqrt{\frac{30}{7\times 163}}\leq {|\corr _{\YHK}(A_7,C_7)|}=
\sqrt{\frac{14}{69}}.
$$
\epf
\end{proof}
\begin{figure}[h]
\begin{center}
{\includegraphics[scale=0.5]{yuratio_200}}
\end{center}
\caption{Plot of the ratio ${\pyule(C_n=k)}/{\puni(C_n=k)}$ for $n=200$ and $1\le k \le 100$. Here the $y$-axis is in the logarithmic scale; the dotted line $y=0$ indicates the ratio being $1$.
% $\pyule(C_n=k)=\puni(C_n=k)$.
}
\label{fig:ratio}
\end{figure}
\bigskip
Proposition~\ref{prop:ch:comp} states that the mean of $\ch_n$ is greater under the YHK model than the PDA model.
In the remainder of this section, we shall present a more detailed study on $\ch_n$. Intuitively, it is easy to see that the number of cherries contained in a random tree generated by the YHK model is likely to be greater than that by the PDA model: firstly,
by Proposition~\ref{prop:type} we know that the number of cherries in $T[e]$, the phylogenetic tree obtained from $T$ by attaching a new leaf to edge $e$ in $T$, is strictly greater than that in $T$ precisely when $e$ is a pendant edge of $T$;
secondly, in the YHK process the edge to which the new leaf is attached is sampled only from the pendant edges while in the PDA model that edge is sampled from all possible edges. Indeed, this intuition can also be corroborated by numerical results. As an example, considering the ratio of
$\pyule(C_n=k)/\puni(C_n=k)$ with $n=200$ as depicted in Fig.~\ref{fig:ratio} using a logarithmic scale, then it is clear that the ratio is strictly increasing and is greater than 1 when $k$ is greater than a certain value.
The following theorem establishes the existence of a unique change point between the two models for $n\geq 4$. Note that a similar phenomenon is shown to hold for clade sizes by~\citet[Theorem 5]{zhu2014clades}.
\begin{theorem}
\label{thm:change}
%Given $n>2$,
Suppose $n\ge 3$. The ratio $\pyule(C_n=k)/\puni(C_n=k)$ is strictly increasing for $1\leq k \leq n/2$. In particular, there exists a number $\tau_n$ with $1\leq \tau_n\leq n/2$ such that
$$
\pyule(C_n=k)< \puni(C_n=k)
~~~\mbox{for $1\le k< \tau_n$, and~~}
\pyule(C_n=k)> \puni(C_n=k)
~~~\mbox{for $\tau_n<k\le n/2$.}
$$
\end{theorem}
\begin{proof}
For simplicity, put $a^k_n=\pyule(C_n=k)$. By Eq.~(\ref{eq:ch:pda:inc}), it suffices to show that
\begin{align}
\label{eq:ch:ratio}
f(k,n)=:\frac{a^{k-1}_n}{a^k_n} \leq \frac{\puni(C_n=k-1)}{\puni(C_n=k)}= \frac{4k(k-1)}{(n-2k+1)(n-2k+2)}:=g(k,n)
\end{align}
holds for $1<k\leq n/2$. To this end, we shall use induction on $n$. The base case $n=3$ is clear because $\pyule(C_3=1)=\puni(C_3=1)=1$.
For induction step, assuming that $f(k,m)< g(k,m)$ holds for a given $m>3$ and all $1\leq k \leq m/2$, it remains to show that $f(k,m+1)\leq g(k,m+1)$ for all $1< k\leq (m+1)/2$.
Note first that $f(1,m+1)=g(1,m+1)=0$. Now fix $2\leq k\leq (m+1)/2$, then $a_m^{k-1}>0$. Since Proposition~\ref{prop:yule:ch:fun} implies
\begin{align*}
a_{m+1}^{k-1}=\frac{2k-2}{m}a_m^{k-1}+\frac{m-2k+4}{m}a_m^{k-2}
~~\mbox{and}~~
a_{m+1}^{k}=\frac{2k}{m}a_m^{k}+\frac{m-2k+2}{m}a_m^{k-1},
\end{align*}
it follows that Eq.~(\ref{eq:ch:ratio}) is equivalent to
\begin{align*}
(2k-2)a_m^{k-1}+(m-2k+4)a_m^{k-2} < \big(2ka_m^k+(m-2k+2)a_m^{k-1}\big)\,g(k,m+1).
\end{align*}
Since $a_m^{k-1}>0$, dividing both sides of the last inequality by $a_m^{k-1}$ leads to
\begin{align}
\label{eq:ch:ratio:alt}
(2k-2)+(m-2k+4)f(k-1,m) < 2k\frac{g(k,m+1)}{f(k,m)}+(m-2k+2)g(k,m+1).
\end{align}
By induction assumption we have $f(k-1,m)\leq g(k-1,m)$ and $0<f(k,m)< g(k,m)$, hence it remains to show that
\begin{align}
\label{eq:ch:ratio:mid}
(2k-2)+(m-2k+4)g(k-1,m) < 2k\frac{g(k,m+1)}{g(k,m)}+(m-2k+2)g(k,m+1).
\end{align}
To this end, denoting the left and the right side of Inequality~(\ref{eq:ch:ratio:mid}) by $L(k,m)$ and $R(k,m)$, respectively, then we need to show that $R(k,m)-L(k,m)>0$. Note first that
$$
L(k,m)=(2k-2)+\frac{4(k-1)(k-2)}{m-2k+3}.
$$
On the other hand, since
$$
\frac{g(k,m+1)}{g(k,m)}=\frac{m-2k+1}{m-2k+3}=1-\frac{2}{m-2k+3},
$$
we have
$$
R(k,m)=2k-\frac{4k}{m-2k+3}+\frac{4k(k-1)}{m-2k+3}.
$$
Therefore, we have
\begin{align*}
R(k,m)-L(k,m) &=2+\frac{4k(k-1)-4k-4(k-1)(k-2)}{m-2k+3} \\
&=2+\frac{4(k-2)}{m-2k+3}>0,
\end{align*}
as required. Here the last inequality follows from $m>3$ and $2\leq k \leq (m+1)/2$.
\epf
\end{proof}
\section{Discussion and Conclusion}
\label{sec:discussion}
Tree shape indices are summary statistics of some aspect of the shape of a phylogenetic tree, particularly the `balance' of a tree. Since the introduction of the first tree shape index by~\citet{Sackin1972}, many such indices have been proposed (see~\citet{mooers97a} for an excellent review and ~\citet{mir2013new} for some recent development).
%Tree indices based on subtree pattern, such as the number of cherries or pitchforks, are an important family of genealogical features in the study of phylogenetics and population genetics. Various studies have been devoted to deriving their moments and limiting distributions under the YHK and the PDA models, two null models which are commonly used in ecology and evolutionary biology.
%Using a bottom-up interpretation of the YHK and the PDA models,
In this paper we present several results concerning the distributions of cherries and pitchforks under the YHK and PDA models. Our main results include two novel recursive formulae on the joint distributions of cherries and pitchforks under these two models, which enable us to numerically compute their joint probability density functions (and hence also the marginal distributions) for trees of any size numerically. This is relevant because one of the main applications of tree indices is their use as test statistics to discriminate stochastic models of evolution~\citep[see e.g.][and the references therein]{blum2006random}, %which is pursuing in a follow up project.
and we will pursue this application more thoroughly elsewhere.
\begin{figure}[h]
\begin{center}
{\includegraphics[scale=0.5]{jointAppro}}
\end{center}
\caption{Plot of the total variation distance between the joint distributions of cherry and pitchfork and discretised bivariate normal distributions for $10\leq n \leq 200$ under the YHK model (solid line) and PDA model (dotted line).
}
\label{fig:app}
\end{figure}
Our numerical results (e.g. Fig.~\ref{fig:app}) indicate
that the limiting joint distributions of cherries and pitchforks can be well approximated by bivariate normal distributions. For the YHK model, this was recently confirmed by~\citet{Janson2014} and it remains open to establish the analogous result for the PDA model.
In addition, it would be interesting to study the rate of the convergence.
In this paper we concentrate on rooted trees, but it is of interest to investigate to what extent the results obtained in this paper for rooted trees can be carried over to unrooted trees.
For example, using Eq.~(\ref{eq:chDistr:unrooted:pda}) and an argument similar to the proof in Theorem~\ref{thm:lc:pda}, it follows that the cherry distribution of unrooted trees under the PDA model is also log-concave, and hence unimodal. However, whether the same property holds for the cherry distribution of unrooted trees under the YHK model remains open. One challenge is to derive the recursions for unrooted trees as in
Theorem~\ref{thm:yule:pf} or an exact formula as in Theorem~\ref{ch:pda}.
To end this article, we mention several additional questions. For instance, is pitchfork distribution, or other subtree distribution, log-concave? Our numerical calculation suggests the pitchfork distribution is log-concave.
A related question is whether there also exists a unique change point
for other subtree distribution. Finally, cherry pattern and pitchfork pattern are closely related to instances of recursive shape index (in the sense of~\citet{matsen2007optimization}), therefore it would also be of interest to see whether some of the properties obtained here can be carried over to some other tree indices as well.
\bigskip
\noindent
{\bf Acknowledgements}
%Kwok Pui Choi acknowledges
K.P. Choi acknowledges the support of Singapore Ministry of Education Academic Research Fund R-155-000-147-112.
%The research reported in this paper has been partially supported by the Singapore Ministry of Education Academic Research Fund R-155-000-147-112.
%\bibliographystyle{abbrv}%plain}
%\bibliography{StatPhyloTree.bib}
\newcommand{\noop}[1]{}
\begin{thebibliography}{25}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{{#1}}
\providecommand{\urlprefix}{URL }
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{DOI~\discretionary{}{}{}#1}\else
\providecommand{\doi}{DOI~\discretionary{}{}{}\begingroup
\urlstyle{rm}\Url}\fi
\providecommand{\eprint}[2][]{\url{#2}}
\bibitem[{Aldous(1991)}]{aldous1991asymptotic}
Aldous D (1991) Asymptotic fringe distributions for general families of random
trees. The Annals of Applied Probability 1:228--266
\bibitem[{Aldous(1996)}]{aldous96a}
Aldous D (1996) Probability distributions on cladograms. In: Aldous D, Pemantle
R (eds) Random Discrete Structures, The IMA Volumes in Mathematics and its
Applications, vol~76, Springer-Verlag, pp 1--18
\bibitem[{Aldous(2001)}]{Aldous2001}
Aldous D (2001) Stochastic models and descriptive statistics for phylogenetic
trees, from {Yule} to today. Statistical Science 16(1):23--34
\bibitem[{Baroni et~al(2005)Baroni, Gr{\"u}newald, Moulton, and
Semple}]{baroni2005bounding}
Baroni M, Gr{\"u}newald S, Moulton V, Semple C (2005) Bounding the number of
hybridisation events for a consistent evolutionary history. Journal of
mathematical biology 51(2):171--182
\bibitem[{Blum and Fran{\c{c}}ois(2006)}]{blum2006random}
Blum MGB, Fran{\c{c}}ois O (2006) Which random processes describe the tree of
life? {A} large-scale study of phylogenetic tree imbalance. Systematic
Biology 55(4):685--691
\bibitem[{Chang and Fuchs(2010)}]{chang2010limit}
Chang H, Fuchs M (2010) Limit theorems for patterns in phylogenetic trees.
Journal of Mathematical Biology 60(4):481--512
\bibitem[{Conway and Guy(1996)}]{con96}
Conway JH, Guy R (1996) The Book of Numbers. Springer
\bibitem[{Disanto and Wiehe(2013)}]{dw13}
Filippo D, Wiehe T (2013) Exact enumeration of cherries and pitchforks in ranked trees under the coalescent model. Mathematical Biosciences 242(2):195--200.
\bibitem[{Hagen et~al(2015)Hagen, Hartmann, Steel, and Stadler}]{hagen2015age}
Hagen O, Hartmann K, Steel M, Stadler T (2015) Age-dependent speciation can
explain the shape of empirical phylogenies. Systematic Biology 64(3):432--440
\bibitem[{Harding(1971)}]{harding71a}
Harding EF (1971) The probabilities of rooted tree-shapes generated by random
bifurcation. Advances in Applied Probability 3(1):44--77
\bibitem[{Heard(1992)}]{heard92a}
Heard SB (1992) Patterns in tree balance among cladistic, phenetic, and
randomly generated phylogenetic trees. Evolution 46(6):1818--1826
\bibitem[{Hendy and Penny(1982)}]{hendy1982branch}
Hendy M, Penny D (1982) Branch and bound algorithms to determine minimal
evolutionary trees. Mathematical Biosciences 59(2):277--290
\bibitem[{Holmgren and Janson(2015)}]{Janson2014}
Holmgren C, Janson S (2015) Limit laws for functions of fringe trees for binary
search trees and recursive trees. Electronic Journal of Probability 20:1--51
\bibitem[{Kingman(1982)}]{Kingman1982}
Kingman J (1982) On the genealogy of large populations. Journal of Applied
Probability 19:27--43
\bibitem[{Matsen(2007)}]{matsen2007optimization}
Matsen F (2007) Optimization over a class of tree shape statistics. IEEE/ACM
Transactions on Computational Biology and Bioinformatics (TCBB) 4(3):506--512
\bibitem[{McKenzie and Steel(2000)}]{McKenzie2000}
McKenzie A, Steel MA (2000) Distributions of cherries for two models of trees.
Mathematical Biosciences 164:81--92
\bibitem[{Mir et~al(2013)Mir, Rossell{\'o}, and Rotger}]{mir2013new}
Mir A, Rossell{\'o} F, Rotger L (2013) A new balance index for phylogenetic
trees. Mathematical Biosciences 241(1):125--136
\bibitem[{Mooers and Heard(1997)}]{mooers97a}
Mooers AO, Heard SB (1997) Evolutionary process from phylogenetic tree shape.
Quarterly Review of Biology 72:31--54
\bibitem[{Nordborg(2001)}]{nordborg01a}
Nordborg M (2001) Coalescent theory. In: Balding DJ, Bishop M, Cannings C (eds)
Handbook of Statistical Genetics, Wiley, Chichester, UK, chap~7, pp 179--212
\bibitem[{Purvis et~al(2011)Purvis, Fritz, Rodr{\'\i}guez, Harvey, and
Grenyer}]{purvis2011shape}
Purvis A, Fritz SA, Rodr{\'\i}guez J, Harvey PH, Grenyer R (2011) The shape of
mammalian phylogeny: patterns, processes and scales. Philosophical
Transactions of the Royal Society B: Biological Sciences 366(1577):2462--2477
\bibitem[{Rosenberg(2006)}]{rosenberg06a}
Rosenberg NA (2006) The mean and variance of the numbers of r-pronged nodes and
r-caterpillars in {Yule}-generated genealogical trees. Annals of
Combinatorics 10:129--146
\bibitem[{Sackin(1972)}]{Sackin1972}
Sackin MJ (1972) ``{Good}" and ``{bad}" phenograms. Systematic Zoology
21(2):225--226
\bibitem[{Semple and Steel(2003)}]{semple03a}
Semple C, Steel MA (2003) Phylogenetics. Oxford University Press, Oxford, UK
\bibitem[{Yule(1925)}]{yule25a}
Yule GU (1925) A mathematical theory of evolution: based on the conclusions of
{Dr. J.C. Willis, F.R.S}. In: Philosophical Transactions of the Royal Society
of London. Series B, Containing Papers of a Biological Character, vol 213,
The Royal Society, pp 21--87
\bibitem[{Zhu et~al(2011)Zhu, Degnan, and Steel}]{zhu11a}
Zhu S, Degnan JH, Steel MA (2011) Clades, clans and reciprocal monophyly under
neutral evolutionary models. Theoretical Population Biology 79:220--227
\bibitem[{Zhu et~al(2015)Zhu, Than, and Wu}]{zhu2014clades}
Zhu S, Than C, Wu T (2015) Clades and clans: A comparison study of two
evolutionary models. Journal of Mathematical Biology 71:99--124
\end{thebibliography}
\end{document}