Conditioning Bienaym{\'e}-Galton-Watson trees to have large sub-populations
Extracted source text · arXiv 2311.17716
\documentclass[11pt]{amsart}
\usepackage{}
\usepackage{amssymb}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{amsmath,amssymb,amsthm}
%\usepackage[latin1]{inputenc}
\usepackage{version, tabularx, multicol}
\usepackage{graphicx,float}
\usepackage{color,latexsym,amsfonts}
\usepackage{extarrows}
\usepackage[shortlabels]{enumitem}
\usepackage{bbm}
\usepackage{bm}
\textheight=624pt
\oddsidemargin=18pt
\topmargin=-0pt
\textwidth=15,5cm
\evensidemargin=18pt
\theoremstyle{plain}
\newtheorem{theo*}{Theorem}
\newtheorem{theo}{Theorem}[section]
\newtheorem{defi}[theo]{Definition}
\newtheorem{cor}[theo]{Corollary}
\newtheorem{prop}[theo]{Proposition}
\newtheorem{lem}[theo]{Lemma}
\newtheorem{conj}[theo]{Conjecture}
\theoremstyle{remark}
\newtheorem{rem}[theo]{Remark}
\newtheorem{rem*}{Remark}
\newlist{propenum}{enumerate}{1}
\setlist[propenum]{label=(\roman*)}
%%%%special letter
\newcommand{\cvxhull}{\mathcal O}
\newcommand{\co}{\mathcal O}
\newcommand{\AAke}{A_{S_k^\varepsilon}}
\newcommand{\ddelta}{\Delta}
\newcommand{\ca}{{\mathcal A}}
\newcommand{\cb}{{\mathcal B}}
\newcommand{\cc}{{\mathcal C}}
\newcommand{\cd}{{\mathcal D}}
\newcommand{\ce}{{\mathcal E}}
\newcommand{\supercE}{{\mathbb E}}
\newcommand{\cf}{{\mathcal F}}
\newcommand{\cg}{{\mathcal G}}
\newcommand{\ch}{{\mathcal H}}
\newcommand{\ci}{{\mathcal I}}
\newcommand{\cj}{{\mathcal J}}
\newcommand{\ck}{{\mathcal K}}
\newcommand{\cl}{{\mathcal L}}
\newcommand{\cn}{{\mathcal N}}
\newcommand{\cm}{{\mathcal M}}
\newcommand{\cp}{{\mathcal P}}
\newcommand{\cq}{{\mathcal Q}}
\newcommand{\cs}{{\mathcal S}}
%\newcommand{\Tau}{{\mathcal T}}
\newcommand{\cR}{{\mathcal R}}
\newcommand{\ct}{{\mathcal T}}
\newcommand{\cu}{{\mathcal U}}
\newcommand{\cw}{{\mathcal W}}
\newcommand{\cx}{{\mathcal X}}
\newcommand{\cy}{{\mathcal Y}}
\newcommand{\cz}{{\mathcal Z}}
\newcommand{\A}{{\mathbb A}}
\newcommand{\C}{{\mathbb C}}
\newcommand{\D}{{\mathbb D}}
\newcommand{\E}{{\mathbb E}}
\newcommand{\superE}{{\mathbb E}}
\newcommand{\Eb}{{\bar{\rm E}}}
\newcommand{\F}{{\mathbb F}}
\newcommand{\M}{{\mathbb M}}
\newcommand{\N}{{\mathbb N}}
\renewcommand{\P}{{\mathbb P}}
\newcommand{\Pb}{{\bar{\rm P}}}
\newcommand{\superP}{{\mathbb P}}
\newcommand{\Q}{{\mathbb Q}}
\newcommand{\R}{{\mathbb R}}
\renewcommand{\S}{{\mathbb S}}
\newcommand{\T}{{\mathbb T}}
\newcommand{\Z}{{\mathbb Z}}
\newcommand{\rA}{{\rm A}}
\newcommand{\w}{{\rm w}}
\newcommand{\W}{{\rm W}}
\newcommand{\rN}{{\rm N}}
\newcommand{\rP}{{\rm P}}
\newcommand{\rE}{{\rm E}}
\newcommand{\rS}{{\rm S}}
\newcommand{\ba}{\mathbf a}
\newcommand{\bb}{\mathbf b}
\newcommand{\bd}{\mathbf d}
\newcommand{\be}{\mathbf e}
\newcommand{\bk}{\mathbf k}
\newcommand{\bj}{\mathbf j}
\newcommand{\bn}{\mathbf n}
\newcommand{\bM}{\mathbf M}
\newcommand{\bN}{\mathbf N}
\newcommand{\bt}{{\mathbf t}}
\newcommand{\bff}{{\mathbf f}}
\newcommand{\br}{\mathbf r}
\newcommand{\bs}{{\mathbf s}}
\newcommand{\bu}{{\mathbf u}}
\newcommand{\bv}{{\mathbf v}}
\newcommand{\bV}{{\mathbf V}}
\newcommand{\bx}{\mathbf x}
\newcommand{\bX}{\mathbf X}
\newcommand{\ind}{{\mathbbm 1}}
\newcommand{\un}{\bm{1}}
\newcommand{\zero}{\bm{0}}
\newcommand{\roott}{\emptyset}
\newcommand{\bp}{\mathbf{p}}
\newcommand{\qc}{\theta_{\alpha}}
\newcommand{\anc}{\mathrm{An}}
\newcommand{\fa}{\mathrm{Pa}}
\newcommand{\dom}{{\rm dom}}
\newcommand{\spa}{{\rm span}}
\newcommand{\ri}{{\rm ri}\;}
\newcommand{\aff}{{\rm aff}\;}
\newcommand{\cv}{{\rm cv}\;}
\newcommand{\clo}{{\rm cl}\;}
\newcommand{\inter}{{\rm int}\;}
\newcommand{\Span}{{\rm span}\;}
\newcommand{\range}{{\mathcal R}}
\newcommand{\argmin}{{\rm argmin}\;}
\newcommand{\supp}{{\rm supp}\,}
\newcommand{\Card}{{\rm Card}\;}
\newcommand{\diam}{{\rm diam}\;}
\newcommand{\diag}{{\rm diag}}
\newcommand{\dist}{{\rm dist}\;}
\newcommand{\norm}[1]{\mathop{\parallel #1 \parallel}\nolimits}
\newcommand{\val}[1]{\mathop{\left| #1 \right|}\nolimits}
\newcommand{\inv}[1]{\mathop{\frac{1}{ #1}}\nolimits}
\newcommand{\expp}[1]{\mathop {\mathrm{e}^{ #1}}}
\newcommand{\capa}[1]{\mathop{\mathrm{cap}( #1 )}\nolimits}
\newcommand{\Var}{{\rm Var}\;}
\newcommand{\Cov}{{\rm Cov}\;}
\newcommand{\sgn}{{\rm sign}\;}
\newcommand{\param}{{\rm Pa}}
%intervals for integers
\newcommand{\II}[1] {[\![#1]\!]}
\newcommand{\lb}{[\![}
\newcommand{\rb}{]\!]}
%interior
\DeclareFontFamily{U}{mathb}{\hyphenchar\font45}
\DeclareFontShape{U}{mathb}{m}{n}{ <-6> matha5 <6-7> matha6 <7-8>
mathb7 <8-9> mathb8 <9-10> mathb9 <10-12> mathb10 <12-> mathb12 }{}
\DeclareSymbolFont{mathb}{U}{mathb}{m}{n}
\DeclareMathAccent{\abxring}{0}{mathb}{"38}
\DeclareFontFamily{U}{mathb}{\hyphenchar\font45}
\DeclareFontShape{U}{mathb}{m}{n}{ <-6> matha5 <6-7> matha6 <7-8>
mathb7 <8-9> mathb8 <9-10> mathb9 <10-12> mathb10 <12-> mathb12 }{}
\DeclareSymbolFont{mathb}{U}{mathb}{m}{n}
\newcommand{\interior}[1]{\abxring{#1}}
\newcommand{\ra}[1]{\textcolor{blue}{\textbf{RA:} #1}}
\newcommand{\JFD}[1]{\textcolor{red}{\textbf{JFD:} #1}}
\newcommand{\hb}[1]{{\color{cyan}{\textbf{HB:} #1}}}
\title[Conditioning BGW trees to have large
sub-populations]{Conditioning Bienaym\'{e}-Galton-Watson trees to have
large sub-populations}
\date{\today}
\begin{document}
\author{Romain Abraham}
\address{Romain Abraham, Institut Denis Poisson,
Universit\'{e} d'Orl\'{e}ans,
Universit\'e de Tours, CNRS, France}
\email{romain.abraham@univ-orleans.fr}
\author{Hongwei Bi}
\address{Hongwei Bi, University of international
business and economics, China}
\email{bihw@uibe.edu.cn}
\author{Jean-Fran\c{C}ois Delmas}
\address{Jean-Fran\c{c}ois Delmas, CERMICS, \'{E}cole des Ponts, France}
\email{jean-francois.delmas@enpc.fr}
\subjclass[2010]{60J80; 60B10}
\keywords{Bienaym\'{e}-Galton-Watson tree, generic probability distribution, local limit}
\begin{abstract}
We study the local limit in distribution of
Bienaym\'{e}-Galton-Watson trees conditioned on having large
sub-populations. Assuming a generic and aperiodic condition on the offspring
distribution, we prove the existence of a limit given by a Kesten's tree
associated with a certain critical offspring distribution.
\end{abstract}
\maketitle
\section{Introduction}
Local limit of large
Bienayme-Galton-Watson (in short, BGW) trees have been extensively
studied in recent years.
The classical result of Kesten \cite{k86} describes the local limit
of a critical or subcritical BGW tree conditioned on reaching at least
height $h$ locally converges in distribution as $h$ goes to infinity to the
so-called size-biased tree or Kesten's tree, which is a tree with an
infinite spine. We refer to Section~\ref{sec:kesten} for a precise
description of the Kesten's tree.
Over the years, motivated by various point of view from theoretical
probability, combinatorics, biology or physics, other conditionings have
been considered, such as large total progeny \cite{k75, gk04}, large
number of leaves \cite{js11, ck14}, large number of protected nodes
\cite{abd17}, existence of an individual with a large number of
out-degree or children \cite{he17, he22}. Janson \cite{j12} surveyed
the local limit of BGW trees when conditioned on a large total
population size, and Abraham and Delmas \cite{ad14a, ad14b} provided a
general framework, which describes in full generality the local limit of
critical or subcritical BGW trees conditioned on having a large
sub-population. Notice that in \cite{js11,j12,ad14b, s19} the local
limit may exhibit a condensation phenomenon, as one node of the limiting
tree has an infinite number of children. With other conditioning, such
as a large size of a late generation \cite{abd20, ad19}, or with
exponential weight given by the total height of the tree among tree with
given large size \cite{du23}, the local limit is a tree with an infinite
backbone. Local limit of large multi-type Galton-Watson trees has also
been considered in \cite{p16, adg18}, and also in~\cite{s18,thevenin23}
when conditioning on a linear combination of the sizes of the sub-populations with a given type.
\medskip
One can also consider scaling limits of BGW trees (seen as metric
space), the so called L\'evy continuum trees, as initiated by Aldous
\cite{a91} and generalized by Duquesne and Le Gall \cite{dlg02}. Let us
mention that there is a large recent literature on this subject. In
particular Marzouk \cite{marzouk22} considered the scaling limit of
random trees with a prescribed degree sequence, and Kargin \cite{k23}
and Kortchemski and Marzouk \cite{km23} the scaling limit
of BGW trees conditioned on its total progeny and the number of leaves.
Motivated by those last works, we shall investigate the local limit of
BGW tree when conditioning on its total progeny and the number of leaves
being large. More generally, let $\ca=(A_i)_{i\in \II{1,J}}$, where
$\II{a,b}=[a,b]\cap \N$, be a
finite collection of pairwise subsets of $\N$ and
$\alpha=(\alpha_i)_{i\in \II{1, J}}$ a probability measure on
$\II{1, J}$. We denote by $A_0$ the complementary of
$\bigcup_{i\in \II{1 ,J}} A_i$ in $\N$, which might be empty or
not. We shall then generalize \cite{ad14a} and consider the local
limit of BGW trees conditioned to have $\lfloor \alpha_i n \rfloor $
nodes with out degree in $A_i$ for all $i\in \II{1, J}$ as $n$ goes to
infinity. We stress that there is no condition on the nodes with out-degree in $A_0$.
More precisely let $p=(p(n))_{n\in \N}$ be a probability distribution on
$\N$; its support is $\supp(p)=\{n\in \N\, \colon\, p(n)>0\}$. We assume
that $p$ is non-trivial in the sense that $p(0)>0$ and
$p(\{0, 1\})<1$, where $p(A)=\sum_{n\in A} p(n)$ for $A\subset \N$. We
denote by $\mu(p)\in (0, +\infty ]$ its mean. We denote by $\ct_p$ a
BGW tree with offspring distribution $p$. Notice we don't assume that
$\mu(p)$ is even finite, however since $p(0)$ is positive the tree
$\ct_p$ is finite with positive probability. Set
$\bar \N=\N \cup\{\infty \}$. For a tree $\bt$ we denote by $L_A(\bt)$
the number of its nodes with out-degree (or number of children) in
$A$. We simply set:
\[
L_\ca(\bt)=(L_{A_i}(\bt))_{i\in \II{1, J}}\in \bar \N^J.
\]
In Theorem~\ref{th:p,a-compatibility} we completely characterize the
non-trivial probability distributions $p'$ on $\N$ such that
$\supp(p')\subset \supp(p)$ and for all $\bn=(n_i)_{i\in \II{1, J}} \in \N^J$
such that
$\P( L_{\ca}(\ct_{p'}) = \bn)>0$ (and thus
$\P( L_{\ca}(\ct_p) = \bn)>0$), we have:
\[
\dist\left(\ct_p\, \big|\, L_{\ca}(\ct_p) = \bn\right)
= \dist\left(\ct_{p'}\, \big|\, L_{\ca}(\ct_{p'}) = \bn\right).
\]
Such probability distributions are called \emph{$(p, \ca)$-compatible}.
The $(p, \ca)$-compatible probability distributions can be continuously
parametrized by a parameter $(\theta, \beta)$ in a subset of $[0,
+\infty ]\times \R_+^J$. When the parameter $\theta$ is positive and
finite, then the $(p, \ca)$-compatible probability distribution $\tilde p_{\theta,
\beta}$ associated with the parameter $(\theta, \beta)$ is given by:
\[
\tilde p_{\theta, \beta}(n) = \beta_i \theta^n \, p(n)
\quad\text{for}\quad n\in A_i
\quad\text{and}\quad i\in \II{0, J}
,
\quad\text{where}\quad
\beta_0=\theta^{-1}.
\]
The fact that such exponentially tilted probability distributions are
$(p, \ca)$-compatible was already observed in~\cite{ad14a} for $J=1$ and
in Th\'evenin~\cite{thevenin23} for the multi-type BGW tree setting.
In comparison with those two papers, we give here an exhaustive description of the $(p, \ca)$-compatible probability distributions.
In particular, it
is possible to observe degenerate cases when the parameter $\theta$ can
take the values $0$ and $\infty $, see~\eqref{eq:def-pt0}
and~\eqref{eq:def-pt-infty}. In both cases, when possible, we get
that $0\not \in A_0$ and for the latter that
$A_0\cap \supp(p)$ is either empty or reduced to $\{1\}$. As suggested
by this remark, we shall indeed distinguish in most of the proofs
according to $0\not \in A_0$ (the leaves are directly involved in the
conditioning) or not.
\medskip
For $\bx=(x_i)_{i\in \II{1, J}}\in \R^J$, we set
$|\bx|=\sum_{i\in \II{1, J}} |x_i|$ the $L^1$ norm of $\bx$.
As in~\cite{p16}, it is
interesting to have a fixed (asymptotic) proportion of sub-populations,
and thus consider the local limit of $\ct_p$ conditionally on
$\{ L_{\ca}(\ct_p) = \bn\}$ when $\bn/|\bn|$ converges to some
$\alpha\in \R_+ ^J$ such that $|\alpha|=1$, as $|\bn|$
goes to infinity.
For $p$ non-trivial, intuitively we have that $L_A(\ct_p)$ is of order
$p(A)$ times the total size of the population
$\sharp \ct_p=L_\N(\ct_p)$ when the tree is large, see
\cite{janson16,thevenin20} for precise statement.
Thus, it is natural to consider among the $(p, \ca)$-compatible
probability distribution those which are in the direction $\alpha=(\alpha_i)_{i\in \II{1, J}}$, that
is:
\[
\tilde p_{\theta, \beta} (A_i) \propto\,
\alpha_i
\quad\text{for all}\quad
i\in \II{1, J}.
\]
From this, we can write $\alpha$ as a function of $(\theta, \beta)$,
provided that $\tilde p_{\theta, \beta} (A_0^c)>0$ or equivalently that
$\beta\neq \zero$ (see Remark~\ref{rem:theta=0} for details), and
similarly $\beta$ as a function of $(\theta, \alpha)$. This gives an
elementary reparametrization $(p_{\theta, \alpha})$ of the family
$(\tilde p_{\theta, \beta})$ by the parameter $\theta$ and its direction
$\alpha$, provided $\beta\neq \zero$. The family
$(\tilde p_{\theta, \beta})$ is explicitely given in
Lemma~\ref{lem:p-qa} and the possible directions in
Proposition~\ref{prop:admiss}. In particular, $\alpha$ is not a
possible direction if and only if there exists $j\in \II{1, J}$ such
that $\alpha_j=0$ and $0\in A_j$ or $\alpha_j=1$ and
$A_0\cup A_j\subset \{0, 1\}$ (those two latter conditions correspond to
$p_{\theta, \alpha}$ being non-trivial). In particular, if the entries
of $\alpha$ are all positive then $\alpha$ is a possible
direction. Furthermore, if $\alpha$ is a possible direction, then the
set $I_\alpha\subset [0, +\infty ]$ of possible value for the parameter
$\theta$ is an interval, see Section~\ref{sec:mean-alpha}. As observed
in previous works, the existence of a critical parameter $\qc$ such that
$\mu(p_{\qc, \alpha})=1$, is a key point to obtain the local limit of
the conditioned BGW tree. Proposition~\ref{prop:uniq} asserts that the
mean function $\theta \mapsto \mu(p_{\theta, \alpha})$ is increasing
when $\mu(p_{\theta, \alpha})\leq 1$, and thus there is at most one such
critical parameter $\qc$. Let us stress this result is not obvious as
the map $\theta \mapsto \mu(p_{\theta, \alpha})$ is not monotone in
general (see an example in Remark~\ref{rem:mu-monotone}).
When the critical parameter $\qc$ exists, then the distribution $p$ is
called \emph{generic for the direction $\alpha$}, and we set:
\begin{equation}
\label{eq:def-pa-intro}
p_\alpha=p_{\qc, \alpha} \quad\text{(and thus $\mu(p_\alpha)=1$).}
\end{equation}
We provide necessary and sufficient conditions for $p$ to be generic in
the direction $\alpha$ in Theorem~\ref{thm:generic} which are similar to
those obtained in~\cite{ad14a} when $J=1$.
We don't study further the
relation between the sets $\ca=(A_i)_{i\in \II{1, J}}$ and the fact
that $p$ is generic, and refer to~\cite{ad14a} again to appreciate the
complexity already when $J=1$.
\medskip
Let $\ct_p^*$
denote the Kesten tree associated with the probability distribution $p$
when $\mu(p)=1$, that is, the local limit in distribution of $\ct_p$
conditioned to have height at least $h$, as $h$ goes to infinity.
Before giving the main result of the paper, we recall the hypothesis:
\begin{enumerate}[(H1)]
\item\label{it:H1}
$p$ is a \emph{non-trivial} probability distribution on $\N$ (there is no
moment condition). Without loss of generality we assume that the sets
$\ca=(A_i)_{i\in \II{1, J}}$ are pairwise disjoint
subsets of the support of $p$.
\item \label{it:Hdir-poss}
$\alpha\in \R^J_+$ with $|\alpha|=1$, is a
\emph{possible direction} (this condition is always satisfied if all the entries of
$\alpha$ are positive).
\item $p$ is \emph{generic} in the direction $\alpha$.
\item \label{it:Hlast} $p$ is \emph{aperiodic} in the sense of
Definition~\ref{defi:period}. (In particular being aperiodic depends
on the sets $\ca$ and the
direction $\alpha$, see also Remark~\ref{rem:period0}.)
\end{enumerate}
We are now ready to state the main result of the paper.
Recall $p_\alpha$ in~\eqref{eq:def-pa-intro}.
\begin{theo*}
\label{th:main-intro}
Assume that Hypothesis~\ref{it:H1}-\ref{it:Hlast} hold.
We have the following local limit in distribution:
\[
\dist(\ct_p\,|\, L_{\ca}(\ct_p)=\bn)
\xrightarrow[|\bn|\to\infty]{} \dist(\ct^*_{p_{\alpha}}),
\]
along any sequence $(\bn)$ in $\N^J$ such that: the sub-populations are
large, that is, $\lim |\bn|=\infty$; the conditioning is legit, that is,
$\P(L_\ca(\ct_p)=\bn)>0$, and the direction is strictly $\alpha$, that
is, $ \lim_{|\bn|\rightarrow \infty } \bn/|\bn|=\alpha$ and, with
$\alpha=(\alpha_i)_{i\in \II{1, J}}$ and $\bn=(n_i)_{i\in \II{1, J}}$,
for all $j\in \II{1, J}$:
\begin{equation}
\label{eq:cond-zero-intro}
\alpha_j=0 \implies n_j= 0.
\end{equation}
\end{theo*}
To be complete, we also refer to Lemma~\ref{lem:a-admiss} on the
existence of a sequences of
$\bn$ in $\N^J$ satisfying the hypothesis above.
The proof of Theorem~\ref{th:main-intro} relies on two ingredients. The
first one is the use of Rizzolo's transformation from~\cite{r15} to
reduce the problem to the case $A_0$ empty. The second is the existence
of a local limit for multi-type BGW tree conditioned to the
sub-populations of each type to be large (with proportion given by the
positive left eigenvector of the mean matrix) obtained
in~\cite{ad14b}. One could also use the results of
P\'enisson~\cite{p16}, but this would require stronger hypothesis, see
Remark~\ref{rem:related} for further comments. Let us mention that
Corollary~3.5 in Abraham, Delmas and Guo~\cite{adg18} gives Theorem~\ref{th:main-intro} in the very specific case
where $\mu(p)=1$, $p(A_0)=0$, and the
direction $\alpha$ is the one naturally given by $p$: $\alpha_i=p(A_i)$.
\begin{rem*}[Conditioning on the total size and the number of leaves]
\label{rem:leaves-intro}
Motivated by the scaling limits of BGW trees conditioned on its total
progeny and the number of leaves to be both large considered
in~\cite{k23} and~\cite{km23}, we give as a consequence of the Theorem
above the corresponding local limit of such BGW trees in
Remark~\ref{rem:restriction}. Conditioning on the total progeny and
the number of leaves amount to consider $A_1=\N^* \cap \supp(p)$ and
$A_2=\{0\}$. Notice the directions $\alpha$ can be written as
$(a, 1-a)$. As explained in Remark~\ref{rem:restriction}, we have to
consider two cases.
If the
support of $p$ is reduced to two elements say $0$ and $k$ (with $k\geq
2$ as $p$ is non-trivial), then
Hypothesis~\ref{it:Hlast} is not satisfied. However in this case the
conditioning is equivalent to conditioning on the total size and the
existence of the local limit is then given by~\cite{j12}
and~\cite{ad14a}. Furthermore there is only one possible direction for
which $p$ is generic; it is given by $a=1/k$.
If the
support of $p$ is not reduced to two elements, then provided the
smallest subgroup in $\Z$ containing $\{x-y\, \colon\, x, y\in
\supp(p)\cap \N^*\}$ is $\Z$ itself, then Hypothesis~\ref{it:Hlast} holds. In
this case, Hypothesis~\ref{it:Hdir-poss} is satisfied if and only if
$a\in (0, 1)$. It is
easy to check that $p$ is
generic in the direction $\alpha$ if and only if there exists a positive
finite
root (which is then $\qc$) to the equation:
\[
g(\theta)=p(0)+ a \theta g'(\theta).
\]
The critical probability measure $p_\alpha$ is given by
$p_\alpha(0)= 1-a$ and $p_\alpha(n)= \qc^{n -1} p(n)/ g'(\qc)$
for $n\in \N^*$; and we can apply Theorem~\ref{th:main-intro}.
\end{rem*}
\begin{rem*}[On the strict convergence of the sequence $\bn$ to the
direction $\alpha$]
\label{rem:tbd}
Assume that the offspring distribution $p$ is generic in the
direction $\alpha$ and that $\alpha$ has some zero entries. We
provide in Section~\ref{sec:a=0-n>0} an example where removing
Condition~\eqref{eq:cond-zero-intro} (that is, $n_j=0$ if
$\alpha_j=0$) on the sequence of $\bn=(n_i)_{i\in \II{1, J}}$ such
that $ \lim_{|\bn|\rightarrow \infty } \bn/|\bn|=\alpha$ prevents to
get the local limit of conditioned BGW tree from
Theorem~\ref{th:main-intro}.
\end{rem*}
\begin{rem*}[On non-generic distribution]
\label{rem:gen-intro}
If $p$ is not generic in the possible direction $\alpha$ because of Condition~\ref{item:0inA0-v2} in Theorem~\ref{thm:generic}, then as in
the case $J=1$ studied in~\cite{ad14b}, we conjecture the existence of
a condensation phenomenon at the limit: the existence of a node of the
local limit at finite height with an infinite degree. The first step
to prove this would be considering the condensation for non-generic
multi-type BGW trees. Notice that the others conditions in Theorem~\ref{thm:generic} might happen for probability distributions with bounded supports, see Remark~\ref{rem:admi-a}~\ref{item:non-gene-fin}.
\end{rem*}
\medskip
The rest of the paper is structured as follows. we introduce in
Section~\ref{sec:notation} the general notation and the framework of
discrete trees, BGW trees and Kesten's tree. We define the
$(p,\ca)$-compatible probability distributions in
Section~\ref{sec:def-pqb} and characterize all the $(p,\ca)$-compatible
probability distributions in Section~\ref{sec:compatible} (handling the
degenerate cases $\theta\in \{0, +\infty \}$ and the case $0\not \in A_0$
are delicate). We study in Section~\ref{sec:critical-p} the existence of
the critical parameter $\qc$ and thus the probability distribution
$p_\alpha$. Eventually, we prove the main theorem in Section~\ref{sec:main}
see Theorem~\ref{thm:main-thm} as well as Remark~\ref{rem:tbd}.
\section{Notation}
\label{sec:notation}
\subsection{General notation}
We denote by $\R_+^*=(0,\infty)$ (resp. $\N^*=\{1, 2, \ldots\}$) the
set of positive real numbers (resp. integers) and by
$\R_+=[0,\infty)$ (resp. $\N=\{0, 1, \ldots\}$) the set of
nonnegative real numbers (resp. integers). For $i,j\in \N$ such that
$i\leq j$, note $\II{i, j}=\N\cap [i, j]$. Let $J\in \N^*$. For
$\bx=(x_j)_{j\in \II{1, J}}\in \R^J$, we set $|\bx|=\sum_{j=1}^J |x_j|$. Let:
\[
\Delta_J = \left\{\bx \in \R_+^J\, \colon\,
|\bx|= 1\right\}.
\]
We set $\un\in \R^J$ (resp. $\zero\in \R^J$) the vector of $\R^J$ with
all its coordinates equal to 1 (resp. 0).
\medskip
Let $p = (p(n))_{ n \in \N} $ be a probability distribution on
$\N$ and $\supp(p)=\{n\in\N: p(n)>0\}$ be its support.
For $A \subset \N$, we set:
\[
p(A)=\sum_{n\in A} p(n)
\quad\text{and}\quad
g_A(r)=\sum_{n\in A} r^n p(n)
\quad\text{for $r\geq 0$},
\]
where the sum over an empty set is 0 by convention. In particular, we
have $g_A=0$ for any set $A\subset \N$
such that $p(A)=0$.
We also denote by $\rho_A$ the radius of convergence of $g_A$:
\begin{equation}\label{eq:def-rho-A}
\rho_A=\sup\{r\geq 1\, \colon\, g_A(r)<+\infty \}.
\end{equation}
For simplicity, when $A = \N$, we write $g(r) = g_{\N}(r)$ and
$\rho=\rho_\N$. We write the mean of $p$ by:
\[
\mu(p) = \sum_{n \in \N} np(n).
\]
We say that a probability distribution $p$ is
\emph{critical} (resp. \emph{sub-critical}) if $\mu(p) = 1$
(resp. $\mu(p) < 1$).
The probability distribution $p$ is \emph{non-trivial} if:
\begin{equation}
\label{assumption-p}
0< p(0) \quad\text{and}\quad
p(0) + p(1) < 1.
\end{equation}
\subsection{The set of discrete trees}
We consider ordered rooted trees in the framework of
Neveu \cite{n86}. More precisely, let $\cu=\bigcup_{n\geq 0} (\N^*)^n$
be the set of finite sequences of positive integers with the convention
that $(\N^*)^0=\{\emptyset\}$. Note $H(u)=n$ the generation or the height of $u$ if
$u=(u_1,\ldots, u_n)\in (\N^*)^n$. For $u,v\in \cu$, denote by $uv$
the concatenation of $u$ and $v$, with the convention that $uv = u$
if $v = \emptyset$ and $uv = v$ if $u=\emptyset$.
The set of ancestors of $u$ is the set:
\[
\text{An}(u)=\{v\in\cu: \text{there exists}\ w\in\cu\ \text{such that}\
u=vw\}.
\]
The most recent common ancestor of $\bs\subset\cu$, denoted by $M(\bs)$,
is the unique element $u$ of $\cap_{u\in\bs} \text{An}(u)$ with maximal height $H(u)$.
Let $\prec$ be the usual lexicographic order on $\cu$.
A tree $\bt$ is a subset of $\cu$ that satisfies: $\emptyset \in\bt$; if
$u\in\bt$, then $\text{An}(u)\subset\bt$; for $u\in\bt$, there exists
$k_u(\bt)\in\N^*$, called the out-degree of $u$, such that, for
every $i\in\N^*$, $ui\in\bt$ if and only if $i\in \II{1, k_u(\bt)}$.
The vertex $\emptyset$ is called the root of $\bt$. The vertex
$u\in \bt$ is called a leaf if $k_u(\bt)=0$. We set $k_u(\bt)=-1$ if
$u\not\in \bt$.
Let $L_A(\bt)$ be the number of vertices of the tree $\bt$
whose out-degree belongs to $A \subset \N$:
\[
L_A(\bt)=\Card(\cl_A(\bt))
\quad\text{with}\quad
\cl_A(\bt)=\{u\in\bt:\, k_u(\bt)\in A\}.
\]
We simply write $\sharp \bt=L_\N(\bt)$ for the cardinal of $\bt$, and $L_n(\bt)$ for
$L_{\{n\}}(\bt)$ when $n \in \N$.
Remark that we have:
\begin{equation}
\label{eq:kt=t-1}
\sum_{u\in \bt} k_u(\bt)=\sharp \bt-1.
\end{equation}
Then we get from~\eqref{eq:kt=t-1}:
\begin{equation}
\label{eq:kt=L0-1}
L_{0}(\bt)=1+ \sum_{k\in \N^*} (k-1) L_{k}(\bt) .
\end{equation}
For $u\in\bt$, we define the subtree above $u$ by $\{v\in\cu\, \colon\,
uv\in\bt\}$ and the fringe subtree by:
\begin{equation}
\label{eq:def-fringe}
\bs=\{uv\in \cu\, \colon\, uv\in\bt\}.
\end{equation}
We denote by $\T$ the set of trees, $\T_0$ the subset of finite
trees, and $\T_1$ the set of trees with only one infinite branch:
\[
\T_1=\left\{\bt\in \T\setminus \T_0\, \colon\, \lim_{n\rightarrow
\infty } H\big(M(\{u\in \bt\, \colon\, H(u)=n\})\big) =\infty \right\}.
\]
Let $H(\bt)=\sup\{H(u)\, \colon u\in \bt\}$ be the height of the
tree $\bt$; and for $h\in \N^*$, let
$\T^{(h)}=\{\bt\in \T\, \colon\, H(\bt)\leq h\}$
be the set of trees with height less or equal to $h$.
\subsection{Local convergence of trees}\label{sec:lct}
For $h\in \N$ and a tree $\bt\in \T$, let $r_h(\bt)=\{u\in \bt\,
\colon\, H(u)\leq h\}$ be the tree $\bt$ truncated at level $h$.
Let $(T_n)_{n\in \N}$ and $T$ be $\T$-valued random variables.
We say that the sequence $(T_n)_{n\in \N}$ converges locally in distribution
towards $T$ if:
\begin{equation}\label{eq:cv_law}
\forall h\in\N,\ \forall
\bt\in\T^{(h)},\quad \lim_{n\rightarrow+\infty }
\P\bigl(r_h(T_n)=\bt\bigr)
= \P\bigl(r_h(T)=\bt\bigr),
\end{equation}
and writing $\text{dist}(T)$ for the distribution of
the random variable $T$, we denote it by:
\[
\lim_{n\to\infty} \text{dist}(T_n)=\text{dist}(T).
\]
For $\bt\in\T$ and $x\in\cl_0(\bt)$, we consider a convergence determining
class of trees $\T(\bt,x)=\{\bt\circledast (\tilde{\bt},x): \tilde{\bt}\in \T\}$,
where:
\[
\bt\circledast (\tilde{\bt},x)=\{u\in\bt\}\cup\{xv: v\in\tilde{\bt}\}
\]
is the tree obtained by grafting $\tilde{\bt}$ on the leaf $x$ of $\bt$.
We recall from~\cite[Lemma~2.1]{ad14a}, that if $(T_n)_{n\in \N}$ and $T$ are
$\T_0\cup \T_1$-valued random variables, then the sequence $(T_n)_{n\in \N}$
converges locally in distribution
towards $T$ if and only if for all $\bt\in\T_0$ and $x\in\cl_0(\bt)$:
\[
\lim_{n\rightarrow \infty } \P(T_n\in \T(\bt, x))= \P(T\in \T(\bt,
x))
\quad\text{and}\quad
\lim_{n\rightarrow \infty } \P(T_n=\bt)= \P(T=\bt).
\]
\medskip
\subsection{BGW trees}
Let $p$ be a probability distribution on $\N$. A $\T$-valued random
variable $\tau$ is a BGW tree with offspring distribution $p$ if
$k_\emptyset(\tau)$ is distributed as $p$ and the branching property is
satisfied: for $n\in \N^*$, conditionally on
$\{k_\emptyset(\tau) = n\}$, the subtrees $(S_1(\tau),\ldots,S_n(\tau))$
are independent and distributed as $\tau$. We denote by $\ct_p$ the BGW
tree with offspring distribution $p$. For all finite tree $\bt\in \T_0$,
we have:
\begin{equation}
\label{eq:loi-Tp}
\P(\ct_p = \bt) = \prod_{u\in \bt} p(k_u(\bt))
=\prod_{n\in \N} p(n)^{L_n(\bt)},
\end{equation}
with the convention that $0^0=1$.
When~\eqref{assumption-p} holds and $p$ is critical or sub-critical,
then a.s.\ $\ct_p$ is finite (that is, $\ct_p\in \T_0$) and in this case~\eqref{eq:loi-Tp}
completely characterizes
the distribution of $\ct_p$.
\subsection{Kesten's tree}
\label{sec:kesten}
Let $p$ be a critical probability distribution on $\N$
(and thus $\mu(p)=1$) satisfying~\eqref{assumption-p}. We denote by
$p^*=(p^*(n)=np(n))_{n\in \N}$ the corresponding size-biased
distribution. The so called Kesten's tree, $\ct_p^*$, is a
$\T_1$-valued random tree defined as the local limit in distribution,
when $n$ goes to infinity, of a BGW tree conditioned to have height
larger than $n$:
\[
\lim_{n\to\infty} \text{dist}(\ct_p|\, H(\ct_p)=n)=\text{dist}(\ct^*_p).
\]
Informally, it is the skeleton of a
two-type BGW tree, where: individuals are of type
$\mathrm{s}$ (survivor) or $\mathrm{n}$ (normal);
the root is of type $\mathrm{s}$; each individual
of type $\mathrm{s}$ has a random number of children with offspring
distribution $p^*$, all of them of type $\mathrm{n}$ but for one uniformly chosen
at random which is of type $\mathrm{s}$; each individual
of type $\mathrm{n}$ has a random number of children with offspring
distribution $p$, all of them of type $\mathrm{n}$. Its distribution is
completely characterized by $\P(\ct_p^*\in \T_1)=1$ and:
\begin{equation}
\label{eq:loi-Tp*}
\P(\ct_p^*\in \T(\bt,x))=\frac{\P(\ct_p=\bt)}{p(0)}
\quad\text{for all}\quad
\bt\in\T_0\quad\text{and}\quad
x\in \cl_0(\bt).
\end{equation}
\section{Definition of the distribution $\tilde p_{\theta, \beta}$}
\label{sec:def-pqb}
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{j\in \II{1, J}}$, with
$J\in \N^*$, be pairwise disjoint non-empty subsets of $\supp(p)$.
Note $ A_0$ the complementary of
$\cup _{j\in \II{1, J}} A_j$ in $\supp(p)$. Notice $A_0$ may be empty.
For $\cj\subset \II{0, J}$, we set:
\[
A_{\cj}=\bigcup _{j\in \cj} A_j.
\]
For a probability
distribution $q$ on $\N$, we write:
\[
q(\ca) = (q(A_1), \ldots, q(A_J))\in [0, 1]^J.
\]
For a finite tree $\bt$, we write:
\[
L_\ca(\bt) = (L_{A_1}(\bt), \ldots, L_{A_J}(\bt))\in \N^J.
\]
We extend the definition of generic probability distribution with
respect to a subset of $\N$ in \cite{ad14a} to generic probability
distribution with respect to a family of subsets.
\begin{defi}[Generic probability distribution]
\label{defi:on-p}
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{ j\in \II{1, J}}$, with
$J\in \N^*$, be pairwise disjoint non-empty subsets of $\supp(p)$.
\begin{propenum}
\item\textbf{$(p,\ca)$-compatible probability distribution.}
\label{defi:equiv}
We say that a probability distribution $p'$ is $(p,\ca)$-compatible
if $p'$ satisfies~\eqref{assumption-p}, $\supp(p')\subset
\supp(p)$, and for all $\bn \in \N^J$ such that
$\P( L_{\ca}(\ct_{p'}) = \bn)>0$ (and thus
$\P( L_{\ca}(\ct_p) = \bn)>0$), we have:
\begin{equation}
\label{equi-cond}
\dist\left(\ct_p\, \big|\, L_{\ca}(\ct_p) = \bn\right)
= \dist\left(\ct_{p'}\, \big|\, L_{\ca}(\ct_{p'}) = \bn\right).
\end{equation}
\item\textbf{Generic distribution.}
\label{defi:generic}
We say that $p$ is generic for $\ca$ in the direction
$\alpha\in \Delta_J$ if
there exists a \emph{critical} $(p, \ca)$-compatible probability
distribution $p'$ such that:
\[
p'(\ca)= \left(1-p'( A_0)\right) \, \alpha
\quad \text{with}\quad
p'(A_0)<1.
\]
\end{propenum}
\end{defi}
\begin{rem}[On the one-dimensional case]
\label{rem:dim=1}
When $J=1$, Definition \ref{defi:on-p} \ref{defi:generic} reduces to the definition
of generic probability distribution with respect to the set
$A_1\subset \N$
with $\alpha=1$.
\end{rem}
\begin{rem}[Positivity of $\alpha$]
\label{rem:0inAj}
If $p$ is generic for $\ca$ in the direction $\alpha\in \Delta_J$ with
$0\in A_{j}$ for some $j\in \II{1, J}$, then we have
$\alpha_j>0$. Indeed, the probability distribution $p'$ from
Definition~\ref{defi:on-p}~\ref{defi:generic} is $(p, \ca)$-compatible
and thus $ p'(A_j)= p'(A_0^c) \, \alpha_j$ with $ p'(A^c_0) >0$; since
$p'$ also satisfies~\eqref{assumption-p}, we deduce that
$p'(A_j)\geq p'(0)>0$ which then gives $\alpha_j>0$.
\end{rem}
We now introduce a set of parameters which will allow us to describe all
the compatible probability distributions. We set:
\[
\cj_{\infty }=\{j\in \II{1, J}\, \colon\, \sup A_j <\infty \}.
\]
\begin{defi}[The set of parameters $\param(p, \ca)$]
\label{defi:p-qb}
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{j\in \II{1, J}}$,
with $J\in \N^*$, be pairwise disjoint non-empty subsets of $\supp(p)$.
For $\beta=(\beta_1, \ldots, \beta_J)\in \R_+^J$, we set:
\begin{equation}
\label{eq:def-cj}
\cj^*_\beta=\{j\in \II{1, J}\, \colon\, \beta_j>0\}
\quad\text{and}\quad
\cj_\beta=\{0\}\cup \cj^*_\beta.
\end{equation}
The set $\param(p, \ca)\subset [0, +\infty ]\times \R^J_+$ of
parameters is defined as follows.
\begin{enumerate}[(i)]
\item \textbf{Non degenerate case.}
The parameter
$(\theta, \beta)\in \R_+^* \times \R^J_+$ belongs to $\param(p, \ca)$ if:
\begin{equation}
\label{eq:def-cond-prob}
\sum_{j\in \cj_\beta} \beta_j\, g_{A_j}(\theta)=1
\quad\text{where}\quad
\beta_0=\theta^{-1 }.
\end{equation}
\item \textbf{Degenerate case.}
The parameter $(\theta, \beta)$, with $\theta\in \{0, +\infty \}$
and $\beta\in \R^J_+$, belongs to $\param(p, \ca)$ if:
\begin{equation}
\label{eq:def-cond-prob-0}
\sum_{j\in \cj_\beta} \beta_j=1
\quad\text{where}\quad
\beta_0=p(1)\, \ind_{\{1\in A_0\}},
\end{equation}
and:
\begin{align}
\label{eq:def-cond-struct-0}
&\text{if $\theta=0$, then} \quad 0\not\in A_0 ,\\
\label{eq:def-cond-struct-infini}
& \text{if $\theta=+\infty $, then}\quad
A_0\subset \{0,1\}
\quad\text{and}\quad
\cj^*_\beta\subset \cj_{\infty }.
\end{align}
\end{enumerate}
\end{defi}
For simplicity, we shall write $\cj^*$ and $\cj$ for $\cj^*_\beta$ and
$\cj_\beta$. Notice that $\cj^*$ might be empty in the non degenerate
case, and that $\cj^*$ is non empty in the degenerate cases thanks
to~\eqref{assumption-p}.
We now define the family of probability distribution indexed by the
parameter $(\theta, \beta)$.
\begin{defi}[Probability distribution $\tilde{p}_{\theta,\beta}$]
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{ j\in \II{1, J}}$, with
$J\in \N^*$, be pairwise disjoint non-empty subsets of $\supp(p)$.
For $(\theta, \beta)\in \param(p, \ca)$, we define the probability
distribution
$\tilde{p}_{\theta,\beta}=(\tilde{p}_{\theta, \beta}(n))_{ n\in \N}$ as
follows, where the unspecified probabilities are set to be 0.
\begin{enumerate}[(i)]
\item \textbf{Non degenerate case.} If $\theta\in \R_+^*$, then we set:
\begin{equation}
\label{eq:def-pt}
\boxed{ \tilde{p}_{\theta, \beta} (n)=\beta_j \, \theta^{n}p(n)}
\quad\text{for $j\in \cj$ and $n\in A_j$.}
\end{equation}
\item \textbf{Degenerate case at 0.} If $\theta=0$, then we set:
\begin{equation}
\label{eq:def-pt0}
\boxed{ \tilde{p}_{0, \beta} (n)=\beta_j }
\quad\text{for $j\in \cj$ and $n=\min A_j$.}
\end{equation}
\item \textbf{Degenerate case at infinity.} If $\theta= +\infty $,
then we set:
\begin{equation}
\label{eq:def-pt-infty}
\boxed{ \tilde{p}_{\infty, \beta} (n)=\beta_j }
\quad\text{for $j\in \cj$ and $n=\max A_j$.}
\end{equation}
\end{enumerate}
\end{defi}
Conditions~\eqref{eq:def-cond-prob} and
~\eqref{eq:def-cond-prob-0} insures that
$\tilde{p}_{\theta,\beta}$ is indeed a probability distribution.
In the next remark, we consider some particular
cases of the probability
distributions $\tilde{p}_{\theta, \beta}$. For
convenience, for the BGW trees, we write:
\[
\ct_{\theta, \beta}=\ct_{\tilde{p}_{\theta, \beta}}.
\]
Recall that $g_{A_0}=0$ if $A_0=\emptyset$. We shall consider the
equation $ g_{A_0}(\theta)=\theta$ on $\R_+$ which has at least one root
and at most two. It is elementary to check that:
\begin{equation}
\label{eq:def-qmin0}
\theta_{\min}=\min \{\theta\in \R_+\, \colon\,
g_{A_0}(\theta)=\theta\}\in [0, 1),
\end{equation}
and that the second root, if it exists, belongs to $(1, +\infty )$. It
is elementary to check the following result.
\begin{lem}
\label{lem:q-min0} We have $\theta_{\min}=0$ if and only if
$0\not\in A_0$.
\end{lem}
\begin{rem}[On particular cases]
\label{rem:theta=0}
Recall $\un$ (resp. $\zero$) denotes the vector of $\R^J$ with all its coordinates equal to
1 (resp. 0).
\begin{enumerate}[(a)]
\item \textbf{The case $\theta=1$ and $\beta=\un$.} We trivially have $\tilde
p_{(1, \un)}=p$, and thus $(1, \un)\in \param(p, \ca)$. In
particular, we get
$\ct_p=\ct_{1,\un}$.
\item \label{item:q=0} \textbf{The case $\theta=0$.} The support of
$\tilde{p}_{0,\beta}$ is equal to
$\{1\}\cup\{ \min A_j\, \colon\, j\in \II{1, J}\}$ or to
$\{ \min A_j\, \colon\, j\in \II{1, J}\}$ according to $1$ belonging
to $A_0$ or not.
\item \label{item:q=infty}
\textbf{The case $\theta=\infty $.} The support of
$\tilde{p}_{\infty ,\beta}$ is equal to
$\{1\}\cup\{ \max A_j\, \colon\, j\in \cj_\infty \}$ or
$\{ \max A_j\, \colon\, j\in \cj_\infty \}$ according to $1$ belonging
to $A_0$ or not.
\item\label{item:b=0} \textbf{The case $\beta=\zero$.} If
$(\theta, \beta)\in \param(p, \ca)$, then we have that $\beta=\zero $
is equivalent to $ \tilde{p}_{\theta,\zero}(A_0^c)=0$. In this case,
we have $\theta\in (0, +\infty )$
(as~\eqref{assumption-p}, which
implies $p(1)<1$,
and~\eqref{eq:def-cond-prob-0} rule out the case
$\theta\in \{0, +\infty \}$)
and that $\theta$ is a root of
$g_{A_0}(\theta)=\theta$ by~\eqref{eq:def-cond-prob}. Thus $\theta$ can take the value
$\theta_{\min}$ (only if $\theta_{\min}>0$ and then $0\in A_0$), and
possibly another value, say $\theta_M\in (1, +\infty )$. We also
have that $\mu(\tilde p_{\theta, \zero})=g'_{A_0}(\theta)$, so that
$\tilde{p}_{\theta_{\min} ,\zero}$ is sub-critical and, if $\theta_M$
exists, $\tilde{p}_{\theta_M ,\zero}$ is super-critical.
In conclusion, for $ \tilde{p}_{\theta,\zero}$ not to be
super-critical, we need $0 \in A_0$ and $\theta=\theta_{\min}$.
\end{enumerate}
\end{rem}
In order (partially) remove the particular case $\beta=\zero$, we set:
\begin{align}
\label{eq:def-Pa**}
\param^{**}(p, \ca)
&=\{(\theta, \beta)\in \param(p, \ca)\, \colon\,\beta\neq\zero\},\\
\param ^*(p, \ca)
&=
\begin{cases}
\param^{**}(p, \ca) &\text{if $0\not\in A_0$},\\
\param^{**}(p, \ca)\, \cup\{(\theta_{\min}, \zero)\} & \text{if $0\in
A_0$}.
\end{cases}
\end{align}
The introduction of the set $ \param ^*(p, \ca)$ is
motivated by the following result.
\begin{lem}[Conditional finiteness of $\ct_{\theta, \beta}$]
\label{lem:TinT0}
Let $(\theta, \beta)\in \param(p, \ca)$.
We have:
\[
\P(\ct_{\theta, \beta}\not \in \T_0,\, \sup_{j\in \cj^*}
L_{A_j}(\ct_{\theta,
\beta})<+\infty )=0
\quad\text{if and only if}\quad
(\theta, \beta)\in \param^*(p, \ca).
\]
\end{lem}
\begin{proof}
If $\beta\neq \zero$, then there exists $j\in \II{1, J}$ such that
$\beta_j>0$. This gives that a.s.\ on the event
$\{ \ct_{\theta, \beta}\not \in \T_0\}$ we have
$L_{A_j}(\ct_{\theta, \beta})=+\infty $
and thus
$ \P(\ct_{\theta, \beta}\not \in \T_0,\, \sup_{j\in \cj^*}
L_{A_j}(\ct_{\theta, \beta})<+\infty )=0$.
If $\beta=\zero$, we deduce from
Remark~\ref{rem:theta=0}~\ref{item:b=0} that if $(\theta, \zero)\in
\param(p, \ca)$ then we have either $\theta=\theta_{\min}>0$,
$(\theta, \zero)\in
\param^*(p, \ca)$, $\tilde p_{\theta, \zero}$ is
sub-critical; or
$\theta>\theta_{\min}$, $(\theta, \zero)\not \in
\param^*(p, \ca)$ and $\tilde p_{\theta, \zero}$ is
super-critical. In the former case, we get $\P(\ct_{\theta, \zero}\not
\in \T_0)=0$ and in the latter case:
\[
\P(\ct_{\theta, \zero}\not \in \T_0,\, \sup_{j\in \cj^*}
L_{A_j}(\ct_{\theta,
\zero})<+\infty )= \P(\ct_{\theta, \zero}\not \in \T_0)>0.
\]
This gives the result.
\end{proof}
We shall now restrict our study to the case where $\tilde p_{\theta,
\beta}$ is non trivial.
\begin{defi}[Compatible parameter]
\label{defi:param-admiss}
The parameter $(\theta, \beta)$ is $(p, \ca)$-compatible if $(\theta,
\beta)\in \param^*(p, \ca)$ and the probability distribution
$\tilde{p}_{\theta, \beta}$ satisfies condition~\eqref{assumption-p}.
\end{defi}
We now characterizes the $(p, \ca)$-compatible parameters.
\begin{lem}[Characterization of the compatible parameters]
\label{lem:tp-non-degenere}
The parameter $(\theta, \beta)\in \param^*(p, \ca)$ is
$(p, \ca)$-compatible if and only if:
\begin{enumerate}[(i)]
\item\label{item:def-cond-struct-ndg} For $\theta\in (0, +\infty )$, we have $0\in A_\cj
$ and $
p\left(A_\cj\cap \{0, 1\}^c\right)>0$.
\item\label{item:def-cond-struct-0-ndg}
For $\theta=0$, we have $ 0\in A_{\cj^*}$ and $
\max_{j\in\cj^*}\min (A_j)>1$.
\item \label{item:def-cond-struct-infini-ndg}
For $\theta=+\infty $, we have $A_{j_0}= \{0\}$ for some $j_0\in
\cj^*$ and $ \max_{j\in\cj^*}\max (A_j)>1$ (and $A_0\subset
\{1\}$ and $\cj^*\subset \cj_\infty $).
\end{enumerate}
\end{lem}
\begin{proof}
By construction, we have
$\supp(\tilde p_{\theta, \beta})\subset \supp(p)$. Since $p(0)>0$,
there exists $j_0\in \II{0, J}$ such that $0\in A_{j_0}$.
Let $\theta\in (0, +\infty )$. The condition $0\in A_\cj$ in
Point~\ref{item:def-cond-struct-ndg} is equivalent to
$\beta_{j_0}>0$ and thus to $\tilde p_{\theta, \beta}(0)>0$. The
condition $p\left(A_\cj \cap \{0, 1\}^c\right)>0$ is clearly
equivalent to $\tilde p_{\theta, \beta}(\{0, 1\}^c)>0$. Therefore,
conditions in~\ref{item:def-cond-struct-ndg} are equivalent to
$\tilde p_{\theta, \beta}$ satisfying the non-degeneracy
condition~\eqref{assumption-p}.
Let $\theta=0$ (and thus $0\not \in A_0$). The condition
$0\in A_{\cj^*}$ in Point~\ref{item:def-cond-struct-0-ndg} is
equivalent to $\beta_{j_0}>0$ and to $\tilde p_{\theta, \beta}(0)>0$,
as $\min A_{j_0}=0$. Since
$\{\min A_j\, \colon\, j \in \cj^*\}\subset \supp( \tilde p_{\theta,
\beta}) \subset \{\min A_j\, \colon\, j \in \cj^*\} \cup\{1\}$, we
deduce that $ \max_{j\in\cj^*}\min (A_j)>1$ is equivalent to
$\tilde p_{\theta, \beta}(\{0, 1\}^c)>0$. Thus, conditions
in~\ref{item:def-cond-struct-0-ndg} are equivalent to
$\tilde p_{\theta, \beta}$ satisfying~\eqref{assumption-p}.
Let $\theta=+\infty $. The conditions $j_0\in \cj^*$ and
$A_{j_0}=\{0\}$ in Point~\ref{item:def-cond-struct-infini-ndg} are
equivalent to $\tilde p_{\theta, \beta}(0)>0$ (notice
that, by~\eqref{eq:def-cond-prob-0},
$\tilde p_{\theta, \beta}(A_0)>0$ if and only if $1\in A_0$, and
then $0\in A_0$ implies that $\tilde p_{\theta, \beta}(0)=0$).
Eventually the
condition $ \max_{j\in\cj^*}\max (A_j)>1$ is also clearly equivalent
to $\tilde p_{\theta, \beta}(\{0, 1\}^c)>0$. Thus,
conditions in~\ref{item:def-cond-struct-infini-ndg} are equivalent to
$\tilde p_{\theta, \beta}$ satisfying~\eqref{assumption-p}.
\end{proof}
\section{The $(p,\ca)$-compatible probability
distributions}
\label{sec:compatible}
We identify all the $(p,\ca)$-compatible probability
distributions.
Recall that $p$ is a probability distribution on $\N$
satisfying~\eqref{assumption-p} and $\ca=(A_j)_{j\in \II{1, J}}$, with
$J\in \N^*$, are pairwise disjoint non-empty subsets of $\supp(p)$.
\begin{theo}[Characterization of the compatible probability distributions]
\label{th:p,a-compatibility}
Let $ p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. The probability distribution $p'$ is
$(p,\ca)$-compatible if and only if $p' = \tilde{p}_{\theta, \beta}$,
for some $(p,\ca)$-compatible parameter $(\theta,\beta)$.
\end{theo}
As being $(p,\ca)$-compatible implies by definition that $p$ is non-trivial (that is, Condition~\eqref{assumption-p} holds), we shall only consider
the probability distributions $\tilde p_{\theta, \beta}$ satisfying the
conditions of Lemma~\ref{lem:tp-non-degenere}. \medskip
The end of this section is devoted to the proof of this result. We
first prove that $\tilde p_{\theta, \beta}$ are $(p,\ca)$-compatible,
provided Condition~\eqref{assumption-p} holds: see
Lemma~\ref{lem:pA-non-deg} for the non-degenerate cases and
Lemma~\ref{lem:pA-deg} for the degenerate cases. Then we prove that
all $(p,\ca)$-compatible probability distributions are of the form
$\tilde p_{\theta, \beta}$, distinguishing according to $0\in A_0$ in
Lemma~\ref{lem:pab-A0} or $0\not \in A_0$ in
Lemma~\ref{lem:pab-A1}, where the proof of the latter is more technical.
\medskip
We set $\be^0=\zero$ and for $j\in \II{1, J}$:
\[
\be^j=(e_1^j, \ldots, e_J^j)\in \N^J
\quad\text{with}\quad
e^j_i=\ind_{\{i=j\}}.
\]
In particular, we have $\sum_{j=1}^J \be^j=\un$. Notice that $\P(\ct_p=\bt)>0$
implies $\bt\in \T_0$ and $k_u(\bt)\in A_{\II{0, J}}$ for all
$u\in \bt$.
\begin{lem}
\label{lem:pA-non-deg}
Under the assumptions of Theorem~\ref{th:p,a-compatibility}, if the
parameter $(\theta, \beta)$ is $(p, \ca)$-compatible with
$\theta\in (0, +\infty )$, then the probability distribution
$\tilde p_{\theta, \beta}$ is $(p, \ca)$-compatible.
\end{lem}
\begin{proof}
We suppose that the parameter $(\theta, \beta)$ is $(p,\ca)$-compatible
(which implies that $\tilde p_{\theta, \beta}$ is non-trivial) and that
$\theta\in (0, +\infty )$. We prove that the probability distribution
$\tilde p_{\theta, \beta}$ is $(p,\ca)$-compatible. Recall
$\cj=\{0\}\cup\{j\in \II{1, J}\, \colon\, \beta_j>0\}$.
Notice that $\P(L_{\ca}(\ct_{\theta, \beta}) = \bn)>0$ implies that
$n_j=0$ for all $j\not\in \cj^*$, where $\bn=(n_1, \ldots, n_J)$.
Using the definition of
$\tilde{p}_{\theta, \beta}$, we obtain for
$\bt\in \T_0$ such that $\P(\ct_p=\bt)>0$ and $L_{A_j}(\bt)=0$ for $j\not\in \cj$ that:
\begin{align*}
\P(\ct_{\theta, \beta} = \bt)
&= \prod_{u\in \bt} \tilde{p}_{\theta,\beta}(k_u(\bt))\\
&= \prod_{j\in\cj} \, \prod_{u \in \cl_{A_j}(\bt)} \beta_j
\theta^{k_u(\bt)} p(k_u(\bt)) \\
&= \P(\ct_p= \bt)\, \theta^{\sharp \bt -1}
\prod_{j\in\cj}\beta_j^{L_{A_j}(\bt)}\\
&= \P(\ct_p= \bt)\, \theta^{-1} \prod_{j\in\cj^*}
(\beta_j \theta )^{L_{A_j}(\bt)}.
\end{align*}
where we used~\eqref{eq:kt=t-1}
for the third equality, and $\sum_{j\in\cj} L_{A_j}(\bt)= \sharp \bt$
as well as $\beta_0 \theta=1$ for the fourth equality.
As $\P(\ct_p=\bt)=0$ implies
$\P(\ct_{\theta, \beta}=\bt)=0$, we deduce that for $\bn\in \N^ J$:
\begin{align*}
\P(L_{\ca}(\ct_{\theta, \beta}) = \bn)
&= \sum_{\bt\in \T_0,\, L_{\ca}(\bt) = \bn} \P(\ct_{\theta, \beta}=\bt)
+ \P(\ct_{\theta, \beta}\not \in \T_0, \, L_{\ca}(\ct_{\theta, \beta}) = \bn) \\
&= \sum_{\bt\in \T_0,\, L_{\ca}(\bt) = \bn} \P(\ct_{\theta, \beta}=\bt)\\
&= \P\left(L_{\ca}(\ct_p) = \bn\right)\, \theta^{-1} \prod_{j\in \cj^*}
(\beta_j\theta)^{n_j},
\end{align*}
where we used Lemma~\ref{lem:TinT0}
for the second equality.
So for every $\bn\in \N^J$
such that $\P(L_{\ca}(\ct_{\theta, \beta}) = \bn)>0$, we have
$\P(L_{\ca}(\ct_p) = \bn)>0$, and for every
$\bt\in\T_0$ such that $L_{\ca}(\bt)=\bn$, we get:
\[
\P(\ct_{\theta, \beta}=\bt\,|\, L_{\ca}(\ct_{\theta, \beta})=\bn)
=\P(\ct_p=\bt\, |\, L_{\ca}(\ct_p)=\bn).
\]
Hence, the probability distribution $\tilde p_{\theta, \beta}$ is
$(p, \ca)$-compatible.
\end{proof}
\begin{lem}
\label{lem:pA-deg}
Under the assumptions of Theorem~\ref{th:p,a-compatibility},
if the parameter $(\theta, \beta)$ is $(p,
\ca)$-compatible with $\theta\in \{0, +\infty\}$,
then the probability distribution $\tilde p_{\theta, \beta}$ is $(p,
\ca)$-compatible.
\end{lem}
\begin{proof}
We first consider the case $\theta=+\infty $. For simplicity, write
$p'=\tilde p_{\infty , \beta}$ and $\ct'=\ct_{\infty , \beta}$. We
suppose that the parameter $(\infty , \beta)$ is
$(p,\ca)$-compatible. As $p'$ is non-trivial, we get that
$\beta\in \R^J_+\backslash\{\zero\}$ and thus $\cj^*\neq \emptyset$.
In particular, we have $A_0\subset\{1\}$, thanks to
\eqref{eq:def-cond-struct-infini} and
Lemma~\ref{lem:tp-non-degenere}~\ref{item:def-cond-struct-infini-ndg}.
We prove that the probability distribution $p'$ is
$(p,\ca)$-compatible.
Notice that $\P(L_{\ca}(\ct') = \bn)>0$ implies that
$n_j=0$ for all $j\not\in \cj^*$, where $\bn=(n_1, \ldots, n_J)$, and,
by~\eqref{eq:kt=t-1}, that:
\begin{equation}
\label{dege-infinity}
\sum_{j\in \cj^*}n_j(\max A_j-1)=-1.
\end{equation}
To simplify notations, recall we write $L_k$ for $L_{\{k\}}$ for $k\in \N$.
Fix such $\bn\in \N^ J$ and consider the set $\T_{0, n}=\{ \bt\in \T_0\,
\colon\,L_{\ca}(\bt)=\bn \quad\text{and}\quad
\P(\ct'=\bt)>0\}$, which is clearly not empty. Using~\eqref{eq:kt=t-1},
we get, for $\bt\in \T_{0, n}$, that
$ \sum_{j\in \cj^*} \sum_{k\in A_j} L_{k}(\bt) (k-1)=-1$. We deduce
from~\eqref{dege-infinity} that for $k\not \in A_0$ we have:
$L_{k}(\bt)=n_j$ if $j\in \cj^*$ and $k=\max A_j$, and
$L_{k}(\bt)=0$ otherwise.
\medskip
Let $\bt\in \T_{0, n}$.
We distinguish two cases according to $1$ belonging to $A_0$ or not.
First, we consider the case
$1 \in A_0$, that is $A_0=\{1\}$, elementary computation gives:
\[
\P(\ct'=\bt \, |\, L_{\ca}(\ct')=\bn)
= c^{-1} \, p(1)^{L_{1}(\bt)},
\]
with, as $\cj^*\neq \emptyset$, $c$ positive and finite given by:
\[
c=\sum_{\bt' \in \T_0 \, \text{ s.t.}\
L_\ca(\bt')=\bn} p(1)^{L_{1}(\bt')}
+ \P(\ct'\not \in \T_0, L_{\ca}(\ct')=\bn)
=\sum_{\bt' \in \T_0 \, \text{ s.t.}\
L_\ca(\bt')=\bn} p(1)^{L_{1}(\bt')}.
\]
Similarly, we have:
\[
\P(\ct_p=\bt \, |\, L_{\ca}(\ct_p)=\bn)
= c^{-1} \, p(1)^{L_{1}(\bt)}.
\]
This readily implies that $p'$ is $(p, \ca)$-compatible.
\medskip
Secondly, we consider the case
$1\not\in A_0$, that is $A_0=\emptyset$. We get that the set
$\T_{0, n}$ is finite. Similarly to the first case, we obtain, with
$c=\Card(\T_{0, n})\geq 1$, that:
\[
\P(\ct'=\bt \, |\, L_{\ca}(\ct')=\bn)
= \P(\ct_p=\bt \, |\, L_{\ca}(\ct_p)=\bn)= c^{-1} .
\]
This readily also implies that $p'$ is $(p, \ca)$-compatible.
\medskip
Eventually, the case $\theta=0$ can be handled similarly using
$ \sum_{j\in \cj^*}n_j(\min A_j-1)=-1$ instead of~\eqref{dege-infinity}.
\end{proof}
\begin{lem}
\label{lem:pab-A0}
Under the assumptions of Theorem~\ref{th:p,a-compatibility}, if the
probability distribution $p'$ is $(p, \ca)$-compatible and
$0\in A_0$, then we have $p'=\tilde p_{\theta, \beta}$ for
some $(p, \ca)$-compatible
parameter $(\theta,\beta)$.
\end{lem}
\begin{proof}
For $k\in \N$, let $\bt_k$ denote the
tree with the root having $k$ children, all of them being leaves (that
is: $k_\roott(\bt_k)=k$ and $\sharp \bt_k=k+1$).
We assume that $0\in A_0$ and that $p'$ is $(p, \ca)$-compatible. For
simplicity, we write $\ct'$ for $\ct_{p'}$. We have $p(0)>0$ and $p'(0)>0$
since $p$ and $p'$ satisfy~\eqref{assumption-p}. Let
$j\in \II{0, J}$ be such that $p'(A_j)>0$. There exists $k_j\in A_j$ such
that $p'(k_j)>0$. We have:
\[
\P(L_\ca(\ct')=\be^j)\ge
\P(\ct'=\bt_{k_j})=p'(k_j)p'(0)^{k_j}>0.
\]
This also implies that $ \P(L_\ca(\ct_{p})=\be^j)>0$ as
$\supp(p')\subset \supp(p)$.
In particular, thanks to Equation \eqref{equi-cond}, we have for any
$k\in A_j$:
\[
\P(\ct'=\bt_k\bigm| L_\ca(\ct')=\be^j) =\P(\ct_{p}=\bt _ k\bigm|
L_\ca(\ct_{p})=\be^j).
\]
This gives:
\[
\frac{p'(k)p'(0)^k}{\P(L_\ca(\ct')=\be^j)}
=\frac{p(k)p(0)^k}{\P(L_\ca(\ct_{p})=\be^j)},
\]
that is:
\begin{equation}
\label{eq:p=qbp}
p'(k)=\beta_j \theta^k p(k),
\end{equation}
with $\theta=p(0)/p'(0)>0$ and
$\beta_j=\P(L_\ca(\ct')=\be^j)/\P(L_\ca(\ct_{p})=\be^j)>0$. Notice
that $\theta$ and $\beta_j$ does not depend on $k\in A_j$. For $j=0$,
we have $p'(A_0)>0$ as $p'(0)>0$. For $k=0\in A_0$, we deduce
from~\eqref{eq:p=qbp} that $\beta_0=p'(0)/p(0)=1/\theta$. For $j\in \II{0, J} $
such that $p'(A_j)=0$, Equation~\eqref{eq:p=qbp} also holds with
$\beta_j=0$. Notice that when $\beta=\zero$, we have $\P(L_\ca(\ct')=\zero)=1$
and $\P(L_\ca(\ct_p)=\zero)< 1,$ which entails that
$\beta_0=\P(L_\ca(\ct')=\zero)/\P(L_\ca(\ct_{p})=\zero)>1$, or
equivalently, $\theta\in (0, 1)$.
This proves that $p'=\tilde p_{\theta, \beta}$, the latter
being defined in~\eqref{eq:def-pt} as $\theta\in (0, +\infty )$.
To conclude, notice that Condition~\eqref{eq:def-cond-prob} holds as $p'$
is a probability distribution and $\beta_0=1/\theta$.
\end{proof}
\begin{lem}
\label{lem:pab-A1}
Under the assumptions of Theorem~\ref{th:p,a-compatibility}, if the
probability distribution $p'$ is $(p, \ca)$-compatible and
$0\not \in A_0$, then we have $p'=\tilde p_{\theta, \beta}$ for
some $(p, \ca)$-compatible
parameter $(\theta,\beta)$ .
\end{lem}
The proof of this lemma is more technical and relies on the following
result whose proof is postponed at the end of this section.
We introduce the sets:
\begin{align*}
\cj^* & =\{j\in\lb 1,J\rb\,\colon\, p'(A_j)>0\},\\
\cj^ {**} & =\{j\in \cj^*\, \colon\, \Card(A_j)\geq 2\}.
\end{align*}
Notice that once Lemma~\ref{lem:pab-A1} is proved, then $\cj^*$ coincides with $\cj^*_\beta$ defined in~\eqref{eq:def-cj}.
\begin{lem}
\label{lem:technique}
Assume $0\not\in A_0$ and let $p'$ be a $(p, \ca)$-compatible distribution. We have the
following properties.
\begin{enumerate}[(i)]
\item \label{item:tech-00}
The map $\ell \mapsto \left(p'(\ell)/p(\ell)\right)^{1/(\ell -1)}$ is
constant over $\{\ell\in A_0\, \colon\, \ell \geq 2\}$.
\item \label{item:tech-j0}
Assume there exists $\ell\in A_0$ such that $\ell\geq 2$ and $k', k\in A_j$ with
$j\in\cj^{**}$ such that $k'>k\geq 0$.
Then we have:
\[
p'(k')>0
\quad\Longleftrightarrow\quad
p'(k)>0 \quad\text{and}\quad
p'(\ell)>0.
\]
Furthermore if those conditions hold, then we also have, with $\alpha=
\ell-1$ and $\beta=k'-k$:
\begin{equation}
\label{eq:pp-j0}
\left(\frac{p'(k)}{p'(k')}\right)^\alpha \, p'(\ell) ^\beta
= \left(\frac{p(k)}{p(k')}\right)^\alpha \, p(\ell) ^\beta.
\end{equation}
\item \label{item:tech-jj}
Assume there exist $i, j\in \cj^{**}$,
$\ell , k_i\in A_i$ such that $\ell > k_i\geq 0$ and $p'(k_i)>0$, and
$k_j, k \in A_j$
such that $k_j>k\geq 0$ and
$p'(k_j)>0$. Then we have $p'(\ell)>0$ and $p'(k)>0$ as well as, with
$\alpha=\ell - k_i$ and $\beta= k_j -k$:
\begin{equation}
\label{eq:pp-jj}
p'(k)^\alpha p'(\ell)^\beta =
\left(\frac{p'(k_j)}{p(k_j)}\right)^\alpha
\left(\frac{p'(k_i)}{p(k_i)}\right)^\beta \, p(k)^\alpha
p(\ell)^\beta.
\end{equation}
\end{enumerate}
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lem:pab-A1}]
We assume that $0\not \in A_0$, that is without loss of generality,
$0\in A_1$, and that $p'$ is $(p, \ca)$-compatible.
Since $p$ and $p'$ satisfy~\eqref{assumption-p}, we get $p(0)>0$ and
$p'(0)>0$. So, we can define $\beta_1=p'(0)/p(0)>0$. By considering trees
with vertices having zero or one child, we get:
\begin{equation}
\label{eq:pp'1}
1\in A_0
\, \Longrightarrow \, p'(1)=p(1).
\end{equation}
Recall the set $\cj^*=\{j\in \II{1,
J}\, \colon\, p'(A_j)>0\}$. We set $\beta_j=0$ for $j\in \II{1, J}
\setminus \cj^*$ in accordance with the definition of the $\tilde
p_{\theta, \beta}$. Notice that $\cj^*$ is non empty as $1\in \cj^*$.
For $j\in \cj^*$, let $k_j$ be an element of $A_j$ such that
$p'(k_j)>0$.
For $j\in \cj^*\setminus \cj^{**}$, we have $A_j=\{k_j\}$ and, whatever
the value of $\theta$ which will be defined later on, we set
$\beta_j= \theta^{-k_j} p'(k_j)/ p(k_j)$ if $\theta\in (0, +\infty )$
and $\beta_j= p'(k_j)/ p(k_j)$ otherwise; this is
also in accordance with the
definition of the $\tilde p_{\theta, \beta}$. We now have to consider
the value of $p'(\ell)$ for $\ell\in A_j$ and $j\in \{0\} \cup
\cj^{**}$.
\medskip We first consider the case $\cj^{**}=\emptyset$. In particular
we get that $A_1=\{0\}$. Then we have either that
$p'(A_0)=p'(A_0\cap\{1\})$ or there exists $k_0 \in A_0$
such that $k_0\geq 2$ and $p'(k_0)>0$. In the former case, using~\eqref{eq:pp'1} and
$\beta_0=p(1)\, \ind_{\{1\in A_0\}}$, the probability distribution can
be written as a $\tilde p_{\theta, \beta}$ with $\theta=0$. In the
latter case, using Lemma~\ref{lem:technique}~\ref{item:tech-00}
and~\eqref{eq:pp'1}, we deduce that~\eqref{eq:def-pt} holds for all
$\ell\in A_0$ with a common $\theta\in (0, +\infty )$ (given by the constant value of the map in Lemma~\ref{lem:technique}~\ref{item:tech-00}) and
$\beta_0=1/\theta$; thus the probability distribution $p'$ can be
written as a $\tilde p_{\theta, \beta}$ with $\theta\in (0, +\infty )$.
\medskip We now assume that $\cj^{**}\neq \emptyset$. From
Lemma~\ref{lem:technique}~\ref{item:tech-jj}, only three
cases are possible:
\begin{enumerate}[(a)]
\item\label{item:min}
For all $j\in \cj^{**}$, we have
$k_j=\min A_j$, $p'(k_j)>0$ and $p'(A_j\setminus \{k_j\})=0$.
\item\label{item:max}
For all $j\in \cj^{**}$, we have $k_j=\max A_j$, $p'(k_j)>0$ and $p'(A_j\setminus \{k_j\})=0$.
\item \label{item:all}
For all $j\in \cj^{**}$, we have $p'(k)>0$ for all $k\in A_j$.
\end{enumerate}
We shall investigate each case separately. In case~\ref{item:min}, we
deduce from Lemma~\ref{lem:technique}~\ref{item:tech-j0} that
$p'(\ell)=0$ for all $\ell\in A_0$ with $\ell\geq 2$. Thus,
using~\eqref{eq:pp'1}, we get that the probability distribution $p'$ can be written as a
$\tilde p_{\theta, \beta}$ with $\theta=0$.
\medskip
In case~\ref{item:max}, since $p'(0)>0$, we deduce that $A_1=\{0\}$ and
that if $j$ belongs to $\cj^{**}$ (and thus to $\cj^*$) then $\sup A_j$
is finite. Then use Lemma~\ref{lem:technique}~\ref{item:tech-j0} to
deduce that there is no element $\ell\geq 2$ in $A_0$, that is
$A_0\subset \{1\}$. Thus,
using~\eqref{eq:pp'1}, we get that the probability distribution $p'$ can be written as a
$\tilde p_{\theta, \beta}$ with $\theta=+\infty $.
\medskip
In case~\ref{item:all}, to fix ideas, let $i=\min \cj^{**}$ and consider
$k_i=\min A_i$ and $\ell=\min A_i \setminus \{k_i\}$. This uniquely
determine $\beta_i $ and $\theta\in (0, +\infty )$ solution of, for
$k'\in \{k_i, \ell\}$:
\begin{equation}
\label{eq:pk=bqpk}
p'(k')=
\beta_i \theta^{k'} p(k').
\end{equation}
Then for $k''\in A_i$ larger than $\ell$, use
Lemma~\ref{lem:technique}~\ref{item:tech-jj}, and in
particular~\eqref{eq:pp-jj}, with $j=i$, $k_j=k''$ and $k=\ell$
to deduce that $p'(k'')$ can also be written as in~\eqref{eq:pk=bqpk}.
For $j\in \cj^{**}$ with $j\neq i$, set $k=\min A_j$ and define
$\beta_j$ by:
\begin{equation}
\label{eq:pk=bqpk2}
p'(k)=
\beta_j \theta^{k} p(k).
\end{equation}
Then use Lemma~\ref{lem:technique}~\ref{item:tech-jj}, and in
particular~\eqref{eq:pp-jj}, with $k_j>k$ (and $k_j\in A_j$), to deduce
that~\eqref{eq:pk=bqpk2} holds for $k$ replaced by $k_j$. Then for
$\ell\geq 2$ in $A_0$, use Lemma~\ref{lem:technique}~\ref{item:tech-j0}
to get that $p'(\ell)= \theta^{\ell -1} p(\ell)$ (which is also
consistent with~\eqref{eq:pp'1}). We deduce that the probability
distribution $p'$ can be written as a $\tilde p_{\theta, \beta}$ with
$\theta\in (0,+\infty)$. This concludes the proof.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem:technique}]
Without loss of generality, we assume $0\in A_1$.
In the following description of trees we don't precise the number of
leaves, as it is determined through
Equation~\eqref{eq:kt=L0-1}. The argument is based on considering two
well chosen trees $\bt$ and $\bt'$ such that
$L_\ca(\bt)=L_\ca(\bt')$. Recall that we write $L_k$ for $L_{\{k\}}$
and $k\in \N$.
\medskip
We prove Point~\ref{item:tech-00}.
Assume there exist $k,k'\in A_0$ such that $\min(k, k')\geq 2$. Consider a tree $\bt$
having only leaves and $\alpha=k'-1>0$ vertices
of out-degree $k$ (that is $L_{k}(\bt)=\alpha$) and a tree $\bt'$
having only leaves and $\beta=k-1>0$ vertices of out-degree $k'$ (that is
$L_{k'}(\bt')=\beta$).
Set $a=1+(k'-1)(k-1)> 1$. Thanks to~\eqref{eq:kt=L0-1}, we get $
L_{0}(\bt)= L_{0}(\bt')=a$,
and thus:
\[
L_\ca(\bt)=L_\ca(\bt')= \bn
\quad\text{with}\quad \bn=a\, \be^1.
\]
Assume that $p'(k)> 0$. We have $\P(L_\ca(\ct_{p'})=\bn)\ge
\P(\ct_{p'}=\bt)=p'(k)^\alpha p'(0)^a>0$, and thus
$\P(L_\ca(\ct_{p})=\bn)>0$.
We set:
\[
c=\frac{\P(L_{\ca}(\ct_{p'})=\bn)}{\P(L_{\ca}(\ct_{p})=\bn)}>0.
\]
Then, we apply ~\eqref{equi-cond} to the trees $\bt$ and $\bt'$ to get:
\[
p'(k)^\alpha\, p'(0)^{a} =c p(k)^\alpha\, p(0)^{a}
\quad\text{and}\quad
p'(k')^\beta\, p'(0)^{a} =c p(k')^\beta\, p(0)^{a}.
\]
This readily implies that $p'(k')$ is positive (as $k'\in A_0$ implies
that $p(k')>0$) and that
$\left(p'(k)/p(k)\right)^{1/(k -1)}=\left(p'(k')/p(k')\right)^{1/(k' -1)}$.
This gives Point~\ref{item:tech-00}.
\medskip
We now prove Point~\ref{item:tech-j0}.
Assume there exist $k'>k\geq 0$ which are elements of $A_j$ with
$j\in\cj^{**}$ and $\ell\geq 2$ which is an element of $A_0$.
Notice that $j$ can
possibly take the value $1$ when $1\in \cj^{**}$ (that is, $\Card(A_1)\geq 2$). Consider a tree $\bt$ having
only leaves and $\alpha=\ell-1>0$ vertices of
out-degree $k'$ (that is $L_{k'}(\bt)=\alpha$) and a tree
$\bt'$ having only leaves,
$\beta=k'-k>0$ vertices of out-degree $\ell$ and, if $k>0$, $\alpha$ vertices
of out-degree $k$ (that is $L_{\ell}(\bt')=\beta$ and, if $k>0$,
$L_{k}(\bt')=
\alpha$).
\medskip
We first assume that $k\geq 1$ (which is automatically satisfied if $j\geq 2$).
Thanks to~\eqref{eq:kt=L0-1}, we get with $a=1+\alpha(k' -1)=1+
\alpha(k-1)+ \beta(\ell -1)$ that $ L_{0}(\bt)= L_{0}(\bt')=a$,
and thus:
\begin{equation}
\label{eq:La=La-j0}
L_\ca(\bt)=L_\ca(\bt')= \bn\quad\text{with}\quad \bn=a\, \be^1 + \alpha \be^j.
\end{equation}
Assume that $p'(k')>0$.
We have $\P(L_\ca(\ct_{p'})=\bn)\ge
\P(\ct_{p'}=\bt)=p'(k')^\alpha p'(0)^a>0$, and thus
$\P(L_\ca(\ct_{p})=\bn)>0$.
We set:
\[
c=\frac{\P(L_{\ca}(\ct_{p'})=\bn)}{\P(L_{\ca}(\ct_{p})=\bn)}>0.
\]
Then, we apply ~\eqref{equi-cond} to the trees $\bt$ and $\bt'$ to get:
\[
p'(k')^\alpha\, p'(0)^{a} =c p(k')^\alpha\, p(0)^{a}
\quad\text{and}\quad
p'(k)^\alpha\, p'(\ell)^\beta\, p'(0)^{a} =c p(k)^\alpha\, p(\ell)^\beta\, p(0)^{a}.
\]
This readily implies that $p'(k)$ and $p'(\ell)$ are positive (as
$k\in A_j$ and $\ell\in A_0$ imply
that $p(k)$ and $p(\ell)$ are positive).
Similarly, assuming that $p'(k)$ and $p'(\ell)$ are positive implies
that $p'(k')$ is positive. Notice also that~\eqref{eq:pp-j0} is
obvious.
\medskip We now consider the case $k=0$ and thus $j=1$ (as $0\in A_1$). One has
$L_{A_1}(\bt)=L_{0}(\bt) + L_{k'}(\bt)=1+ \alpha(k'-1)+ \alpha= 1+
(\ell-1) k'$, and
$ L_{A_1}(\bt')=L_{0}(\bt')=1+\beta (\ell -1)= 1+ (\ell-1) k'$,
which implies that~\eqref{eq:La=La-j0} still holds.
We then conclude similarly as in the case $k>0$.
This gives Point~\ref{item:tech-j0}.
\medskip
Eventually, we prove Point~\ref{item:tech-jj}. Assume there exist
$i, j\in \cj^{**}$, $k_j>k\geq 0$ which are elements of $A_j$, $\ell > k_i\geq 0$
which are elements of $A_i$, with $p'(k_j)>0$ and $p'(k_i)>0$. Notice that $i$ and
$j$ can be possibly equal and can possibly take the value $1$ when
$1\in \cj^{**}$. Consider a tree $\bt$ having only leaves,
$\alpha=\ell-k_i>0$ vertices of out-degree $k_j$ and, if $k_i>0$,
$\beta=k_j-k>0$ vertices of out-degree $k_i$ (that is
$L_{k_j}(\bt)=\alpha$ and, if $k_i>0$, $L_{k_i}(\bt)=\beta$) and
a tree $\bt'$ having only leaves, $\beta$ vertices of out-degree $\ell$,
and, if $k>0$, $\alpha$ vertices of out-degree $k$ (that is
$L_{\ell}(\bt')=\beta$ and, if $k>0$, $L_{k}(\bt')=\alpha$).
\medskip
We assume that $k\geq 1$ and $k_i\geq 1$, and leave the cases $k=0$
(and thus $j=1$) and/or $k_i=0$ (and thus $i=1$) to the reader as the
proof can be handled very similarly; see also the end of the proof of
Point~\ref{item:tech-j0}.
Thanks to~\eqref{eq:kt=L0-1}, we get with $a=1+\alpha(k_j -1)+ \beta (k_i-1)=1+
\alpha(k-1)+ \beta(\ell -1)$ that $ L_{0}(\bt)=
L_{0}(\bt')=a$
and thus:
\[
L_\ca(\bt)=L_\ca(\bt')= \bn
\quad\text{with}\quad \bn=a\, \be^1 + \alpha \be^j+ \beta \be^i.
\]
We have $\P(L_\ca(\ct_{p'})=\bn)\ge
\P(\ct_{p'}=\bt)=p'(k_j)^\alpha \, p'(k_i)^\beta\, p'(0)^a>0$, and thus
$\P(L_\ca(\ct_{p})=\bn)>0$.
We set:
\[
c=\frac{\P(L_{\ca}(\ct_{p'})=\bn)}{\P(L_{\ca}(\ct_{p})=\bn)}>0.
\]
Then, we apply ~\eqref{equi-cond} to the trees $\bt$ and $\bt'$ to get:
\[
p'(k_j)^\alpha\, p'(k_i)^\beta\, p'(0)^{a} =c p(k_j)^\alpha\,
p(k_i)^\beta\, p(0)^{a}
\quad\text{and}\quad
p'(k)^\alpha\, p'(\ell)^\beta\, p'(0)^{a} =c p(k)^\alpha\, p(\ell)^\beta\, p(0)^{a}.
\]
This readily implies that $p'(k)$ and $p'(\ell)$ are positive (as
$k\in A_j$ and $\ell\in A_i$ imply
that $p(k)$ and $p(\ell)$ are positive).
Equation~\eqref{eq:pp-jj} is then obvious. This gives Point~\ref{item:tech-jj}.
\end{proof}
\section{Existence of a critical $(p, \ca)$-compatible distribution}
\label{sec:critical-p}
\subsection{Parametrization of the $(p, \ca)$-compatible probability
distributions using their direction}
We define the direction of a
probability distribution $p'$ with respect to $\ca$.
\begin{defi}[Direction of compatible probability distributions]
\label{defi:direction}
The
direction $\alpha\in \Delta^J$ of a
$(p, \ca)$-compatible probability distribution $p'$ such that
$p'(A_0)<1$ is defined by:
\begin{equation}
\label{eq:direction}
\alpha= a^{-1} \, p'(\ca)
\quad\text{with}\quad
a=1-p'(A_0)>0.
\end{equation}
\end{defi}
Because by definition $p'$ is non trivial, see~\eqref{assumption-p},
we shall see below that the set of possible directions is:
\begin{equation}
\label{eq:possible-dir}
\Delta_J^*=\Delta_J\backslash \Delta_J^o,
\end{equation}
where the set of ineligible directions are given by:
\[
\Delta_J^o=\bigcup_{j=1}^J \Big\{\alpha\in\Delta_J: \alpha_j=0
\,\text{ if }\,
0\in A_j
\quad\text{or}\quad
\alpha_j=1 \,\text{ if }\, A_0\cup A_j\subset \{0,1\}\Big\}.
\]
We shall use a parametrization of the probability distribution
$\tilde p_{\theta,\beta}$ using the parameter $\theta$ and its
direction $\alpha\in \Delta_J^*$. Recall that
$\tilde p_{\theta,\beta}(A_0)=1$ if and only if $\beta=\zero$,
see Remark~\ref{rem:theta=0}~\ref{item:b=0}. Recall the set of
parameters $\param^{**}(p,
\ca)$ defined in~\eqref{eq:def-Pa**}, where the cases $\beta=\zero$ are
removed, and the set of $(p,
\ca)$-compatible parameters given in
Definition~\ref{defi:param-admiss}.
The direction map $\cd: (\theta, \beta)\mapsto (\theta, \alpha)$ given
by~\eqref{eq:direction} is defined on the subset of $\param^{**}(p,
\ca)$ of
$(p,
\ca)$-compatible parameters;
and it is clearly injective.
\begin{defi}[Compatible parameters]
\label{defi:direc-compa}
The parameter $(\theta, \alpha)=\cd(\theta, \beta)$ is
$(p, \ca)$-compatible if $(\theta, \beta)$ is $(p, \ca)$-compatible (and
$\beta\neq \zero$).
\end{defi}
We shall write $p_{\theta, \alpha}$ for
$\tilde p_{\theta, \beta}$ when $(\theta,\alpha)=\cd(\theta,\beta)$, and it satisfies~\eqref{assumption-p} when $(\theta, \beta)$ is $(p, \ca)$-compatible.
Notice that for $\alpha\in \Delta_J^*$ the
set:
\[
\cj^*_\beta=\{j\in \II{1, J}\, \colon \, \beta_j > 0\}
=\{j\in \II{1, J}\, \colon \, \alpha_j > 0\}
\]
is by definition non empty, and we shall also denote it by
$\cj^*_\alpha$ or simply $\cj^*$ when there is no ambiguity on the
parameter. For the
convenience of the reader, we give explicit formulas for the probability
distributions $p_{\theta, \alpha}$, using~\eqref{eq:direction} and
Lemma~\ref{lem:tp-non-degenere}. To simplify the expression we set:
\begin{equation}
\label{eq:def-q1}
q_1=1- p(1) \ind_{\{1\in A_0\}}>0.
\end{equation}
\begin{lem}[The probability distributions $p_{\theta, \alpha}$]
\label{lem:p-qa}
Let $\alpha\in \Delta_J^*$ and $\theta\in [0, +\infty ]$ be such that
the parameter $(\theta, \alpha)$ is $(p,
\ca)$-compatible. The
non-zero terms of the probability ditribution $p_{\theta, \alpha} $
are given as follows. (Recall that $0\in A_\cj$ and $
p\left(A_\cj\cap \{0, 1\}^c\right)>0$.)
\begin{enumerate}[(i)]
\item\label{item:p-alpha-q}
If
$\theta\in (0, +\infty )$, then we have
$g_{A_0}(\theta)<\theta$, $g_{A_j}(\theta)<+\infty $ for $j\in \cj^*$ and:
\begin{align*}
p_{\theta, \alpha}(k) & =\theta^{k-1} \, p(k)\quad \text{for $k\in A_0$},\\
p_{\theta, \alpha}(k) & =\alpha_j \frac{\theta -
g_{A_0}(\theta)}{g_{A_j}(\theta)} \, \theta^{k-1} \, p(k)
\quad\text{for $k\in A_j$ and $j\in \cj^*$}.
\end{align*}
\item If $\theta=0$, then we have $0\not \in A_0$, $
\max_{j\in\cj^*}\min (A_j)>1$,
$p_{\theta, \alpha}(1) =p(1)$ if $1\in A_0$ and:
\[
p_{\theta, \alpha}(k)=\alpha_j \, q_1
\quad\text{for $k=\min A_j$ and $j\in \cj^*$}.
\]
\item If $\theta=+\infty $, then we have $A_0\subset \{1\}$, $A_{j_0}= \{0\}$ for some $j_0\in
\cj^*$, $\cj^*\subset \cj_\infty $, $ \max_{j\in\cj^*}\max (A_j)>1$,
$p_{\theta, \alpha}(1) =p(1)$ if $1\in A_0$ and:
\[
p_{\theta, \alpha}(k)=\alpha_j \, q_1
\quad\text{for $k=\max A_j$ and $j\in \cj^*$}.
\]
\end{enumerate}
\end{lem}
For $\alpha\in \Delta^*_J$, we get that $p_{1, \alpha}=\tilde{p}_{1,\beta}$ with
$\beta_j=p(A_0^c)\, \alpha_j/p(A_j)$
satisfies~\eqref{assumption-p}, and thus that $(1, \alpha)$ is
$(p,\ca)$-compatible. So $\Delta^*_J$ is indeed the set of all possible directions
of $(p, \ca)$-compatible probability distributions.
We however complete this picture with the following result.
\begin{prop}[Possible directions]
\label{prop:admiss}
For every
$\alpha\in \Delta_J^*$, there exists $\theta\in (0, 1)$ such that
$(\theta, \alpha)$ is $(p,\ca)$-compatible, that is, such that $(\theta,
\beta)$ is $(p,\ca)$-compatible
for some $\beta\in \R^J_+$ with $\beta\neq \zero$ and
$\alpha$ is the direction of $\tilde p_{\theta, \beta}$.
\end{prop}
\begin{proof}
We recall the convention $g_{A_0}=0$ if $A_0=\emptyset$.
Let us first prove that there exists $\theta\in (0, 1)$ such that
$\theta>g_{A_0}(\theta)$. If $0\in A_0$, then $g_{A_0}(0)=p(0)>0$ and
$g_{A_0}(1)<1$, so there exists $\theta_0\in(0,1)$ such that
$\theta_0=g_{A_0}(\theta_0)$, and then $g'_{A_0}(\theta_0)<1$. This implies
$\theta>g_{A_0}(\theta)$ for $\theta\in (\theta_0,1)$. If
$0\not\in A_0$, then $g_{A_0}(0)=0$ and $g_{A_0}(1)<1$, so $0$ is the
only root of $\theta=g_{A_0}(\theta)$ in $[0, 1]$, and thus
$\theta>g_{A_0}(\theta)$ for $\theta\in (0,1)$.
\medskip
Let us now fix $\theta\in (0, 1)$ such that $\theta > g_{A_0}(\theta)$. We set
for all $j\in \II{1,J}$:
\[
\beta_j
=\alpha_j\frac{\theta-g_{A_0}(\theta)}{\theta g_{A_j}(\theta)}\ge 0,
\]
and $\beta _0=\theta^{-1}$.
Since $\sum_{j=1}^J \alpha _j=1$, we have:
\[
\sum_{j\in \II{0, J}}\beta_jg_{A_j}(\theta)=\theta^{-1}g_{A_0}(\theta)
+\sum_{j\in \II{1, J}}\alpha_j\theta^{-1}(\theta-g_{A_0}(\theta))
=1.
\]
Hence Condition~\eqref{eq:def-cond-prob} is satisfied.
Moreover we have:
\[
\tilde
p_{\theta,\beta}(A_0^c)=1- \tilde
p_{\theta,\beta}(A_0)
=\theta^{-1}\,
(\theta-g_{A_0}(\theta)) >0.
\]
Finally, $\alpha\in \Delta_J^*$ (and thus $\alpha\not\in\Delta_J^o$)
insures that $\tilde p_{\theta,\beta}$ satisfies
Lemma~\ref{lem:tp-non-degenere}~\ref{item:def-cond-struct-ndg}, that is,
$\tilde p_{\theta,\beta}$ is non trivial.
\end{proof}
\subsection{Properties of the mean of $p_{\theta, \alpha}$}
\label{sec:mean-alpha}
For $\alpha\in \Delta_J^*$, we consider the following set:
\begin{equation}
\label{eq:def-I-alpha}
I_\alpha=\left\{\theta\in [0, +\infty ]\, \colon\, (\theta,\alpha)
\text{ is $(p,
\ca)$-compatible}\right\}.
\end{equation}
Notice that $1\in I_\alpha$.
Note $\rho_\cj$ the radius of convergence
of $\sum_{j\in\cj} g_{A_j}$ or equivalently:
\begin{equation}
\label{eq:def-rho-J}
\rho_\cj=\min _{j\in \cj} \rho_{A_j}\in [1, +\infty ].
\end{equation}
We define:
\begin{equation}
\label{eq:def-qmin-max}
\theta_{\min}=\inf I_\alpha \in [0, 1)
\quad\text{and}\quad
\theta_{\max}=\sup I_\alpha \in [1, \rho_\cj ] .
\end{equation}
On the one hand, notice that $\theta_{\min}$ is the only root of
$g_{A_0}(\theta)= \theta$ in $[0, 1)$, so this definition is
consistent with~\eqref{eq:def-qmin0}, and thus $\theta_{\min}$ does
not depend on $\alpha$. On the other hand, we have that
$\theta_{\max} $ depends on the support of $\alpha$ as:
\begin{equation}
\label{eq:q-max}
\theta_{\max} = \max \big(\rho_{\cj^*}\, ,\, \sup\{ \theta\in [1, \rho_{A_0}) \,
\colon\, g_{A_0}(\theta)<\theta\}\big)
\quad\text{with}\quad
\rho_{\cj^*}=\min _{j\in \cj^*} \rho_{A_j} .
\end{equation}
We also have that $(\theta_{\min}, \theta_{\max})\subset
I_\alpha$ and that
$\theta\mapsto p_{\theta, \alpha}$ is continuous on $I_\alpha$ for the
norm of the total variation. The next result is a direct consequence
of Lemma~\ref{lem:p-qa} (the first point is also in
Lemma~\ref{lem:q-min0}).
\begin{lem}
\label{lem:q-min}
We have that:
\begin{enumerate}[(i)]
\item $\theta_{\min}=0$ if and only if $0\not\in A_0$,
\item \label{it:q=0-in-I}
$\theta_{\min} \in I_\alpha$ if and only if
$0\not\in A_0$ and $ \max_{j\in\cj^*}\min (A_j)>1$.
\item If $0\in A_0$, then we have
$I_\alpha\subset (0, +\infty )$.
\end{enumerate}
\end{lem}
\medskip
In order to consider finite means, we set for $\alpha\in \Delta^*_J$:
\begin{equation}
\label{eq:def-q_M}
I^\mathrm{f}_\alpha=\{\theta\in I_\alpha\, \colon\, \mu_{\theta,
\alpha} <+\infty \}
\quad\text{where}\quad
\mu_{\theta, \alpha}=\mu(p_{\theta, \alpha})\in [0, +\infty ]
\end{equation}
is the mean of $p_{\theta, \alpha}$. Notice that
$I_\alpha\subset I^\mathrm{f}_\alpha\cup\{\theta_{\max}\}$.
We are now interested in the existence of a
critical probability distribution among the $(p,\ca)$-compatible
probability distributions $p_{\theta, \alpha}$ with a given direction
$\alpha\in \Delta_J^*$, that is in the existence of $\theta\in I_\alpha$
such that $\mu_{\theta, \alpha}=1$.
For $\alpha\in \Delta^*_J$ and $\theta\geq 0$ such that $\sum_{j\in\cj^*} g_{A_j}(\theta)<+\infty $, we set:
\begin{equation}
\label{eq:def-H}
H_\alpha(\theta)=\sum_{j\in \cj^*} \alpha_j h_j(\theta),
\end{equation}
where, for $j\in \cj^*$, $h_j(0)=\min A_j$ and for $\theta>0$:
\[
h_j(\theta) = \frac{\theta g_{A_j}'(\theta)}{g_{A_j}(\theta)}
= \frac{\E\left[X \theta^X \ind_{\{X \in A_j\}}\right]}
{\E\left[\theta^X \ind_{\{X \in A_j\}}\right]},
\]
where $X$ is distributed according to $p$. For
$\theta\in I_\alpha\cap \R_+^*$, we have using
Lemma~\ref{lem:p-qa}~\ref{item:p-alpha-q}:
\begin{equation}\label{mu-expression}
\mu_{\theta, \alpha}
= g_{A_0}'(\theta) + \frac{\theta - g_{A_0}(\theta)}{\theta}
H_\alpha(\theta).
\end{equation}
Recall $q_1$ from~\eqref{eq:def-q1}. Using Lemma~\ref{lem:p-qa}, if $0\in I_\alpha$ we have that:
\begin{equation}\label{mu-expression-0}
\mu_{0, \alpha}
=(1-q_1) +q_1\sum_{j\in \cj^*}\alpha_j\min A_j.
\end{equation}
and if $+\infty \in I_\alpha$ that:
\begin{equation}\label{mu-expression-infinity}
\mu_{\infty , \alpha}
=(1-q_1) +q_1\sum_{j\in \cj^*}\alpha_j\max A_j.
\end{equation}
We now recall some elementary properties of the function $h_j$, see also
\cite[Lemma 3.1]{j12} for a part of the proof. Notice that if $A_j$ is
a singleton, say $\{k_j\}$, then the function $h_j$ is constant equal to
$k_j$. Recall that $\rho_{A_j}$ is the radius of convergence of
$g_{A_j}$ and that $\lim_{x\rightarrow \rho_{A_j}} g_{A_j}(x)= g_{A_j}(\rho_{A_j})$, with the limit being possibly infinite.
\begin{lem}
\label{lem:hj-monotone}
Let $j\in \II{1, J}$ with $\Card(A_j)\geq 2$. The function $h_j$
defined on $[0, \rho_{A_j})$ is $\cc^1$ and increasing, with
$h_j'>0$ on $(0, \rho_{A_j})$.
If $\rho_{A_j}=+\infty $ or if $g_{A_j}(\rho_{A_j})=+\infty $, then we have $\lim_{\theta \rightarrow
\rho_{A_j}} h_j(\theta )=\sup A_j$.
\end{lem}
As an immediate application, we get the following result (one only needs
to take care of the case $\theta=0$, where
$H_\alpha(0)=\sum_{j\in \cj^*} \alpha_j \min A_j$, and of the case
$\theta=+\infty $, where
$H_\alpha(\infty )=\sum_{j\in \cj^*} \alpha_j \sup A_j$ if
$\min_{j\in \cj^*} \rho_j=+\infty $).
\begin{cor}[Regularity of $\mu_{\theta, \alpha}$]
\label{cor:mu-cont}
The map $\theta \mapsto \mu_{\theta,
\alpha} $ is continuous on $I_\alpha$, finite
on $ I^\mathrm{f}_\alpha$, and $\cc^1$ on $(\theta_{\min},
\theta_{\max})$ and also on $[0, \theta_{\max})$ if $0\in I_\alpha$.
\end{cor}
\subsection{Generic distributions}
Notice that $1 \in I_\alpha$; however we don't assume \emph{a
priori} that $1\in I^\mathrm{f}_\alpha$ as $p_{1, \alpha}$ might have infinite mean.
Recall $\rho_\cj=\min_{j\in \cj}
\rho_{A_j}$, see~\eqref{eq:def-rho-J}.
In the next lemmas we give preliminary results on the existence of
$\theta\in I_\alpha$ for $p_{\theta,\alpha}$ to be sub/super/-critical
according to $0$ belonging to $A_0$ or not.
\begin{lem}\label{lem:0-A-0-sub-sup}
Assume that $0\in A_0$ and let $\alpha\in \Delta_J^*$.
\begin{enumerate}[(i)]
\item\label{item:sub-0-in-min}
There exists $\theta\in I_\alpha$ such that $\mu_{\theta,
\alpha}< 1$.
\item\label{item:sub-0-in-max}
If $\rho_\cj=+\infty$, then there exists $\theta\in I_\alpha$
such that $\mu_{\theta, \alpha}> 1$.
\end{enumerate}
\end{lem}
\begin{proof}
As $0\in A_0$, we get using Lemma~\ref{lem:q-min} that $I_\alpha\cap [0, 1]=
(\theta_{\min}, 1]$.
By
continuity, we deduce from~\eqref{mu-expression} that:
\[
\lim_{\theta\downarrow \theta_{\min}}
\mu_{\theta,\alpha}=g_{A_0}'(\theta_{\min})<1.
\]
This gives Point~\ref{item:sub-0-in-min}.
\medskip
We now prove Point~\ref{item:sub-0-in-max}. On the one hand, if there
exists $k\in A_0$ such that $k\geq 2$, then the function $g_{A_0}$ is
strictly convex and
$\lim_{\theta \rightarrow\infty } g_{A_0}(\theta)/\theta= +\infty $
as
$\rho_{A_0}\geq \rho_\cj=+\infty $. We deduce that
$g_{A_0}(\theta_{\max})=\theta_{\max}$, and, as $\theta_{\max}>1$, that
$g'_{A_0}(\theta_{\max})>1$.
By
continuity, we deduce from~\eqref{mu-expression} that:
\[
\lim_{\theta\uparrow \theta_{\max}}
\mu_{\theta,\alpha}=g_{A_0}'(\theta_{\max})>1.
\]
On the other hand, if $A_0\subset \{0, 1\}$, then we get that $
g_{A_0}(\theta)= p(0) + (1-q_1) \theta$. Thus there is no root of
$g_{A_0}(\theta)=\theta$ on $(1, +\infty )$, but we have:
\[
\lim_{\theta\rightarrow\infty } \frac{\theta -
g_{A_0}(\theta)}{\theta} =q_1>0.
\]
Then, use Lemma~\ref{lem:hj-monotone}
and~\eqref{mu-expression}
to deduce that:
\[
\lim_{\theta\rightarrow +\infty }
\mu_{\theta,\alpha}
=(1-q_1) + q_1 \sum_{j\in \cj^*} \alpha_j \,
\sup A_j > (1-q_1) + q_1 \sum_{j\in \cj^*} \alpha_j = 1,
\]
where for the inequality we used that $\sup_{j\in \cj^*} \, \sup
A_j>1$ as $p_{\theta, \alpha}$ is
non trivial and $\sup A_0 \leq 1$. This gives Point~\ref{item:sub-0-in-max}.
\end{proof}
Recall $ \cj_{\infty }=\{j\in \II{1, J}\, \colon\, \sup A_j <\infty
\}$.
\begin{lem}\label{lem:0-A-1-sub-sup}
Assume that $0\not\in A_0$ and let $\alpha\in \Delta_J^*$.
\begin{enumerate}[(i)]
\item \label{item:sub-0-out-min}
There exists $\theta\in I_\alpha$ such that $\mu_{\theta,
\alpha}\leq 1$ if and only if:
\begin{equation}
\label{eq:cond-min-alpha}
\sum_{j\in\cj^*} \alpha_j \min A_j \leq 1.
\end{equation}
\item \label{item:sub-0-out-max}
If $\rho_\cj=+\infty$, then we have $\mu_{\theta, \alpha}< 1$ for all
$\theta\in I_\alpha$
if and only if:
\begin{equation}
\label{eq:cond-max-alpha}
A_0\subset\{1\}, \quad \cj^*\subset \cj_{\infty} ,
\quad \text{and}\quad \sum_{j\in\cj^*} \alpha_j \max
A_j < 1.
\end{equation}
\end{enumerate}
\end{lem}
\begin{proof}
We prove Point~\ref{item:sub-0-out-min}. We first assume that
$\sum_{j\in \cj^*} \alpha_j \min A_j\leq 1$. By assumption
$0\not\in A_0$ and $0\in A_{\cj^*}$ as $1\in I_\alpha$ and
$p_{1, \alpha}$ satisfies~\eqref{assumption-p}. According to
Lemma~\ref{lem:q-min}, we have
$\theta_{\min}=0$.
If $0\in I_\alpha$, we deduce
from~\eqref{mu-expression-0} that
$\mu_{0, \alpha}= 1-q_1+ q_1 \sum_{j\in \cj^*} \alpha_j \min A_j\leq
1$.
If $0\not \in I_\alpha$, that is
$ \max_{j\in\cj^*}\min (A_j)\leq 1$, we get that
$\sum_{j\in \cj^*} \alpha_j \min A_j <1$ as $0\in A_{\cj^*}$, that is,
$\min A_j=0$ for some $j\in \cj^*$ and $\alpha_j>0$. We then deduce
from~\eqref{mu-expression}, using that $q_1$ defined
in~\eqref{eq:def-q1} is positive, that:
\[
\lim_{\theta \downarrow 0} \mu_{\theta, \alpha}= (1-q_1) +
q_1\sum_{j\in \cj^*}\alpha_j\min A_j <1.
\]
So by Corollary
\ref{cor:mu-cont}, there exists $\theta\in (0,1]$, such that
$\mu_{\theta,\alpha}<1$.
In conclusion,
Condition~\eqref{eq:cond-min-alpha} implies that
$\mu_{\theta,\alpha}\leq 1$ for some $\theta\in I_\alpha$.
\medskip
Let us now assume that $\sum_{j\in \cj^*} \alpha_j \min A_j>1$ (and thus
$0\in I_\alpha$ by Lemma~\ref{lem:q-min}). Using~\eqref{mu-expression},
the fact that the functions $h_j$ are non-decreasing (see
Lemma~\ref{lem:hj-monotone} and the fact that $h_j$ is constant equal to
$k_j$ when $A_j$ is reduced to the singleton $\{k_j\}$), we get that for
$\theta\in I_\alpha\cap(0, +\infty )$:
\[
\mu_{\theta, \alpha}\geq g_{A_0}'(\theta) + \frac{\theta - g_{A_0}(\theta)}{\theta}
\sum_{j\in \cj^*} \alpha_j \min A_j > g_{A_0}'(\theta) + \frac{\theta -
g_{A_0}(\theta)}{\theta} =1+\E[(X-1)\theta^{X-1}\ind_{\{X\in A_0\}}],
\]
where $X$ has distribution $p$. Since $0\not\in A_0$, we deduce that
$\E[(X-1)\theta^{X-1}\ind_{\{X\in A_0\}}]$ is non-negative and thus
$ \mu_{\theta, \alpha}> 1$. Thanks to~\eqref{mu-expression-0}, we also
have $\mu_{0, \alpha}>1$, and thus $\mu_{\theta, \alpha}>1$ for all
$\theta\in I_\alpha$ (notice that for $\theta=+\infty $, if it belongs
to $I_\alpha$, thanks to~\eqref{mu-expression-infinity} one gets
$\mu_{\infty , \alpha}\geq \mu_{0, \alpha} > 1$). This ends the proof
of Point~\ref{item:sub-0-out-min}.
\medskip
We prove Point~\ref{item:sub-0-out-max}.
If $A_0\subset \{1\}$, we get $\theta_{\min}=0$ by Lemma~\ref{lem:q-min}
and $\theta_{\max}=+\infty $ as $g_{A_0}(\theta)=(1-q_1)\theta<\theta$
and $\rho_\cj=+\infty $, see~\eqref{eq:q-max}.
This gives that $(0, +\infty )\subset I_\alpha$ and we have:
\begin{equation}
\label{eq:a0in1}
\mu_{\theta, \alpha}=1- q_1 + q_1 H_\alpha (\theta)
\quad\text{for}\quad
\theta\in I_\alpha.
\end{equation}
So, if~\eqref{eq:cond-max-alpha} holds, we have thanks to
Lemma~\ref{lem:hj-monotone} that $\mu_{\theta, \alpha}<1$ for
$\theta\in (0, +\infty )\subset I_\alpha$. We also get
$\mu_{0, \alpha}<1$, resp. $\mu_{\infty , \alpha}<1$, whenever
$0\in I_\alpha$, resp. $+\infty \in I_\alpha$. This gives that
$\mu_{\theta, \alpha} <1$ for all $\theta\in I_\alpha$.
\medskip
Let us assume that $A_0\subset \{1\}$ and $\sup A_j=+\infty $ for some
$j\in \cj^*$ (that is $\cj^*\not \subset \cj_\infty $).
Since $\rho_\cj=+\infty $, Lemma~\ref{lem:hj-monotone} gives that
$\lim_{\theta\rightarrow \infty } \mu_{\theta, \alpha} =+\infty $.
We now assume that $A_0\subset \{1\}$, $\cj^* \subset \cj_\infty $ and $\sum_{j\in
\cj^*} \alpha_j \max A_j\geq 1$.
Let $j_0\in \cj^*$ such that $0\in A_{j_0}$. If $A_{j_0}=\{0\}$, we get
that $\theta_{\max}=+\infty $ belongs to $I_\alpha$ and~\eqref{eq:a0in1}
implies that $\mu_{\infty , \alpha}\geq 1$. If $\Card(A_{j_0})\geq 2$,
then we get that $\min_{j\in \cj^*} \max A_j\geq 1$ and $\max_{j\in
\cj^*} \max A_j> 1$ (as $p_{\theta, \alpha}$
satisfies~\eqref{assumption-p}), and thus $\sum_{j\in
\cj^*} \alpha_j \max A_j> 1$.
Using~\eqref{eq:a0in1} , we get that
$\lim_{\theta\rightarrow \infty } \mu_{\theta, \alpha} >1$.
In both cases, there exists $\theta\in I_\alpha$ such that
$\mu_{\theta, \alpha}\geq 1$.
We now assume that $A_0\not\subset \{1\}$. Then the function $g_{A_0}$
is increasing and strictly convex. As
$\rho_\cj=\infty $, we deduce from~\eqref{eq:q-max} that
$\theta_{\max}\in (1, +\infty )$ is the maximal root of
$g_{A_0}(\theta)=\theta$, and thus
$g_{A_0}'(\theta_{\max}) > 1$. Then, we get from~\eqref{mu-expression}
that
$ \lim_{\theta \uparrow \theta_{\max}} \mu_{\theta, \alpha} =
g_{A_0}'(\theta_{\max}) > 1$.
In conclusion,
if~\eqref{eq:cond-max-alpha} does not hold then there exists
$\theta\in I_\alpha$ such that $\mu_{\theta, \alpha}\geq 1$.
\end{proof}
Recall $ \cj^ {**}=\{j\in \cj^*\, \colon\, \Card(A_j)\geq 2\}$.
We consider the following condition:
\begin{equation}
\label{eq:patho}
A_0\cap \{1\}^c\neq \emptyset
\quad\text{or}\quad
\cj^{**} \neq \emptyset.
\end{equation}
\begin{prop}[Monotonicity of $\theta\mapsto \mu_{\theta, \alpha}$]
\label{prop:uniq}
Let $\alpha\in\Delta_J^*$.
\begin{enumerate}[(i)]
\item \label{it:mu=cst}
Assume~\eqref{eq:patho} does not hold. Then $I_\alpha=I^\mathrm{f}_\alpha=[0,
+\infty ]$ and the map $\theta\mapsto p_{\theta, \alpha}$ (as well
as the map $\theta\mapsto \mu_{\theta, \alpha}$) is constant.
\item Assume~\eqref{eq:patho} holds. Then
$\partial_\theta \mu_{\theta, \alpha}>0$ on the interval
$\{\theta\in I^\mathrm{f}_\alpha\, \colon\, \mu_{\theta, \alpha}\leq
1\}$. If this set is not empty, then its minimum is
$\theta_{\min}$. Furthermore, there exists at most one element
$\theta\in I_\alpha$ such that $\mu_{\theta, \alpha}=1$.
\end{enumerate}
\end{prop}
\begin{rem}[$\mu_{\theta,\alpha}$ independent of $\theta$]
\label{rem:mu=cst}
Recall $q_1=1- p(1)\ind_{\{1\in A_0\}}$.
\begin{enumerate}[(a)]
\item\label{it:mu=cst-1}
If~\eqref{eq:patho} holds and the map $\theta\mapsto
\mu_{\theta, \alpha}$ is constant, then $\mu_{\theta, \alpha}>1$.
\item
If~\eqref{eq:patho} does not hold, that is $A_0\subset\{1\}$ and $A_j$ is a singleton, say $\{k_j\}$ for
all $j\in \cj^*$ (with one of the $k_j$ equal $0$ and another larger than $1$ in order for
$p_{\theta, \alpha}$ to be non trivial), then
$I_\alpha=[0, +\infty ]$ and $ \mu_{\theta,
\alpha}= (1-q_1)+q_1 \sum_{j\in \cj^*} \alpha_j \, k_j\in (0, +\infty )$.
Thus, we recover
Proposition~\ref{prop:uniq}~\ref{it:mu=cst}.
\item Consider the example: $A_0=\{1, k\}$ with $k\geq 2$,
$ \alpha\in \Delta_J^*$ and $A_j=\{k_j\}$ for $j\in \cj^*$ (and thus
$\cj^{**}= \emptyset$) such that $k=\sum_{j\in
\cj^*} \alpha_j k_j$. Notice that $H_\alpha$ is constant equal to
$k$ and that the mean $\mu_{\theta, \alpha}$ is constant as:
\[
\mu_{\theta, \alpha}= p(1) + k \theta^{k-1} p(k) + (1- p(1) -
\theta^{k-1} p(k)) H_\alpha(\theta)= 1 + (1-p(1))(k-1).
\]
Thus Point~\ref{it:mu=cst-1} of this remark is not void.
\end{enumerate}
\end{rem}
\begin{rem}[Is $\mu_{\theta, \alpha}$ monotone in $\theta$?]
\label{rem:mu-monotone}
Consider the probability $p$ defined by $p(2)=1/2$ and
$p(0)=p(4)=1/4$; $J=1$ (and thus $\alpha=1$) and $A_1=\{0, 4\}$ (and
thus $A_0=\{2\}$). We get that $I_1=(0, 2)$ as by
Lemma~\ref{lem:q-min}~\ref{it:q=0-in-I} $\theta_{\min}=0\not \in I_1$
and $\theta_{\max}=2$ by~\eqref{eq:q-max}. It is elementary to check that
$\lim_{\theta\rightarrow 0+}\mu_{\theta, 1}=0$, $\mu_{\theta, 1}
\simeq 1$ for $\theta\simeq 0.36$ and that $\partial_\theta
\mu_{\theta, 1} <0$ if and only if $\theta\in [\theta_0, \theta_1]$
with $\theta_0\simeq 1.24$ and $\theta_1\simeq 1.92$. This provides
an example where the function $\theta\mapsto \mu_{\theta, \alpha}$ is
not monotone.
\end{rem}
\begin{proof}[Proof of Proposition~\ref{prop:uniq}]
Let $\theta\in I_\alpha^\mathrm{f}\cap \R_+^*$. Let $X_\theta$
be a random variable with distribution $p_{\theta, \alpha}$. We have:
\[
\P(X_\theta\in A_0)=\theta^{-1}\, g_{A_0}(\theta)
\quad\text{and}\quad
g'_{A_0}(\theta)=\E[X_\theta \ind_{\{X_\theta\in A_0\}}].
\]
Set:
\[
f(\theta)= \frac{\theta- g_{A_0}(\theta)}{\theta}=\P(X_\theta\not\in A_0),
\]
so that
$\mu_{\theta , \alpha} = g'_{A_0}(\theta) + f(\theta)
H_\alpha(\theta)$. We get:
\begin{equation}
\label{eq:mu'}
\partial_\theta \mu_{\theta, \alpha}= g''_{A_0}(\theta)+ f(\theta)
H_\alpha'(\theta) + f'(\theta) H_\alpha(\theta).
\end{equation}
(This expression can possibly be equal to $+\infty $ if
$\theta=\theta_{\max}$.) We have $H_\alpha(\theta)\in (0, +\infty ]$ as
$\theta>0$. We have that $H_\alpha'(\theta)= 0$ if $\cj^{**}=\emptyset$
and, by Lemma~\ref{lem:hj-monotone}, that
$H_\alpha'(\theta)\in (0, +\infty ]$ otherwise. We also have
$g''_{A_0}(\theta)=0$ if $A_0\subset \{0, 1\}$ and
$g''_{A_0}(\theta)\in (0, +\infty ]$ otherwise. Eventually we have
$f(\theta)>0$ and:
\begin{equation}
\label{eq:qf'q-1}
f'(\theta)
=\frac{g_{A_0}(\theta)-\theta\, g_{A_0}'(\theta)}{\theta^2}
=\theta^{-1} \E\left[(1-X_\theta)\ind_{\{X_\theta\in A_0\}}\right],
\end{equation}
which is finite as $\theta\in I_\alpha^\mathrm{f}$.
Notice that $\mu_{\theta, \alpha}=\E[X_\theta]$, which is finite, and thus:
\begin{equation}
\label{eq:qf'q}
\theta\, f'(\theta)
= \E\left[(X_\theta-1)\ind_{\{X_\theta\not \in A_0\}}\right] + (1-
\mu_{\theta, \alpha}).
\end{equation}
\medskip
\textbf{Case $0\in A_0$.} We first assume that $0\in A_0$, and thus $I_\alpha\subset \R_+^*$ and
$\theta_{\min}\not\in I_\alpha$, see Lemma~\ref{lem:q-min}. We consider
$\theta\in {I}_\alpha^\mathrm{f}$ such that
$\mu_{\theta, \alpha}\leq 1$. We deduce from~\eqref{eq:qf'q} that
$f'(\theta)=0$ if $\cj^*$ is reduced to a singleton, say $\{j_0\}$, with
$A_{j_0}=\{1\}$ and $\mu_{\theta, \alpha}=1$, and that $f'(\theta)>0$
otherwise (as $0\in A_0$). Notice that it is not possible to have
$A_0\subset\{0, 1\}$, $A_{j_0}=\{1\}$ and $\cj^*=\{j_0\}$ together as
$p_{\theta,\alpha}$ is non trivial, so at least
$g''_{A_0}(\theta)\in (0, +\infty ]$ or
$ \E\left[(X_\theta-1)\ind_{\{X_\theta\not \in A_0\}}\right] >0$. The
latter implies that $f'(\theta)>0$ as $\mu_{\theta, \alpha}\leq 1$.
Since $f(\theta)>0$ and $H_\alpha(\theta)>0$, we deduce
from~\eqref{eq:mu'} that $\partial_\theta \mu_{\theta, \alpha}>0$ on
$\{\theta\in {I}_\alpha^\mathrm{f}\, \colon\, \mu_{\theta, \alpha}\leq
1\}$.
\medskip
\textbf{Case $ 0\not\in A_0$.}
We now assume that $0\not\in A_0$. This implies that $\theta_{\min}=0$.
The function $f$ has a continuous extension at $0$ given by
$f(0)=q_1>0$, see~\eqref{eq:def-q1}. We distinguish according to $A_0$
being empty or reduced to $\{1\}$ and $A_0\cap \{1\}^c\neq \emptyset$.
\textbf{Sub-case $ A_0\subset \{1\}$.} We consider the sub-case $A_0\subset \{1\}$. The function $f$ is constant equal to $q_1$, and
$\mu_{\theta, \alpha}=1- q_1+ q_1 H_\alpha(\theta)$.
If $\cj^{**}=\emptyset$ and
thus~\eqref{eq:patho} does not hold, then
we have $p_{\theta,
\alpha}(k_j)=q_1 \alpha_j$ for $A_j=\{k_j\}$ and $j\in \cj^*$. Thus
Point~\ref{it:mu=cst} is obvious.
If
$\cj^{**}\neq\emptyset$, then we deduce
from Lemma~\ref{lem:hj-monotone} that the functions $H_\alpha$ and
$\theta \mapsto \mu_{\theta, \alpha}$ are increasing on
$I^\mathrm{f}_\alpha$.
\medskip
% \[ *************\]
\medskip
\textbf{Sub-case $0\not\in A_0$ and $ A_0\cap \{1\}^c\neq \emptyset$.}
We get in particular that $\theta_{\max}<+\infty $. We introduce an
auxiliary parameterized function defined on $I^\mathrm{f}_\alpha$ for
$\gamma>0$ by:
\begin{equation}
\label{eq:def-mg}
m_\gamma(\theta)= g_{A_0}'(\theta)
+ \gamma\, f(\theta) .
\end{equation}
Notice that $m_\gamma(0)=(1-q_1) + q_1 \gamma>0$.
On the one hand, direct computation yields:
\begin{equation}
\label{eq:mg}
m_\gamma(\theta)
=m_\gamma(0) + \sum_{k\in A_0\cap\{1\}^c} (k-\gamma)
\theta^{k-1} p(k).
\end{equation}
We deduce that if $\gamma< \min (A_0\cap\{1\}^c)$, and in
particular if
$\gamma<2$, then $m'_\gamma>0$ on $I^\mathrm{f}_\alpha$. (Notice that
$m'_\gamma$ is finite on $I^\mathrm{f}_\alpha$
except possibly at $\theta_{\max}$ in the case
where it
belongs to $I^\mathrm{f}_\alpha$.)
On the other
hand, if there exists $\theta_*\in I^\mathrm{f}_\alpha$ such that
$m_\gamma(\theta_*)\leq 1$, we deduce that $\gamma\leq 1$ if
$\theta_*=0$ and, from~\eqref{eq:def-mg} that:
\[
\gamma \leq \frac{1- g'_{A_0}(\theta_*)}{f(\theta_*)}
= \frac{\theta_*- \theta_*
g'_{A_0}(\theta_*)}{\theta_* - g_{A_0} (\theta_*)}<1
\quad\text{if}\quad \theta_*>0.
\]
In conclusion:
\begin{equation}
\label{eq:mg-concl}
\exists \theta_*\in I^\mathrm{f}_\alpha\quad\text{s.t.}\quad
m_\gamma(\theta_*)\leq 1
\implies m'_\gamma>0.
\end{equation}
Now, we go back to the function $\theta\mapsto \mu_{\theta, \alpha}$.
Assume there exits $\theta_*\in I^\mathrm{f}_\alpha$ such that
$\mu_{\theta_*, \alpha}\leq 1$. Set $\mu_*=\mu_{\theta_*, \alpha}$,
$\gamma_*=H_\alpha(\theta_*)$ and $m_*=m_{\gamma_*}$.
By construction, we have that $m_*(\theta_*)=\mu_*\leq 1$ and, as
$H_\alpha$ is non-decreasing, for $\theta\in I^\mathrm{f}_\alpha$:
\[
(\theta - \theta_*) ( \mu_{\theta, \alpha} - m_*(\theta))
= (\theta - \theta_*) (H_\alpha(\theta) - H_\theta(\theta_*)) f(\theta)
\geq 0.
\]
This implies that $\partial_{\theta=\theta_*} \mu_{\theta, \alpha} \geq
m'_* (\theta_*)$, and thus is positive thanks to~\eqref{eq:mg-concl}. We
have obtained that $\partial_\theta \mu_{\theta, \alpha}>0$ on
$\{\theta\in {I}_\alpha^\mathrm{f}\, \colon\, \mu_{\theta, \alpha}\leq
1\}$.
\medskip
In conclusion, if $A_0\subset \{1\}$ and $\cj^{**}=\emptyset$, then $\theta\mapsto
\mu_{\theta,\alpha}$ is constant; if $A_0\cap\{1\}^c\neq \emptyset$ or
$\cj^{**} \neq \emptyset$,
we have $\partial_\theta \mu_{\theta, \alpha}>0$ on
$\{\theta\in {I}_\alpha^\mathrm{f}\, \colon\, \mu_{\theta, \alpha}\leq
1\}$, this set
is either empty, or equal to $I_\alpha^\mathrm{f}$ or of the form
$I_\alpha^\mathrm{f}\cap [0, \theta']$, and it contains at most one element $\theta$ such that $\mu_{\theta, \alpha}=1$.
\end{proof}
Recall that $\rho_\cj$ is the radius of convergence of the function
$\sum_{j\in\cj} g_{A_j}$ and $I_\alpha$ is an interval of $[0,+\infty]$
defined in~\eqref{eq:def-I-alpha}. We have the following theorem.
Recall that assuming~\eqref{eq:patho} is not very restrictive as
otherwise the map
$\theta\mapsto p_{\theta, \alpha}$ is constant, see
Proposition~\ref{prop:uniq}.
\begin{theo}[Generic distribution]\label{thm:generic}
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{ j\in \II{1, J}}$, with
$J\in \N^*$, be pairwise disjoint non-empty subsets of $\supp(p)$.
Let $\alpha\in\Delta_J^*$ and assume that~\eqref{eq:patho} holds.
The distribution $p$ is not generic
for $\ca$ in the direction $\alpha$ if and only if
one of the following conditions holds.
\begin{enumerate}[(i)]
\item\label{item:0inA0-v2} $\rho_{\cj}< \infty$,
$g_{A_0}'(\rho_{\cj}) \leq 1$, $g_\cj(\rho_{\cj}) < \infty$ and:
\begin{equation}\label{eq:g-rho-less-infty}
H_\alpha (\rho_{\cj})
< \rho_{\cj}\, \frac{1 - g_{A_0}'(\rho_{\cj})}{\rho_{\cj} - g_{A_0}(\rho_{\cj})}\cdot
\end{equation}
\item \label{item:0niA0-v2} $0\not \in A_0$ and
$\sum_{j\in\cj^*}
\alpha_j \min A_j > 1$.
\item \label{item:0niA0-2-v2}
$A_0\subset \{1\}$, $\rho_{\cj} = \infty$ and $\sum_{j\in\cj^*}
\alpha_j \sup A_j < 1$.
\end{enumerate}
\end{theo}
\begin{rem}[Direction and generic distribution]
\label{rem:admi-a}
A probability distribution might not be generic in all the directions;
and it may happen that it is generic only in one direction.
\begin{enumerate}[(a)]
\item \label{item:non-gene-fin} Suppose $\supp(p)=\II{0, 3}$ and
$\ca=(\{0\},\{2,3\})$ and thus $A_0=\{1\}$.
Consider the direction $\alpha=(\alpha_1, \alpha_2)$.
\begin{itemize}
\item
The distribution $p$ is generic for $\ca$
in the direction $\alpha$
if and only if $\alpha_2\in [1/3, 1/2]$.
\item
If $\alpha_2\in (1/2,1)$, then Point~\ref{item:0niA0-v2} of
Theorem~\ref{thm:generic} holds since
$0\not \in A_0$ and $ \sum_{j=1}^2
\alpha_j\min(A_j)=2\alpha_2>1$; and in this case all the $p_{\theta, \alpha}$ are sub-critical.
\item If $\alpha_2\in (0, 1/3)$, then Point~\ref{item:0niA0-2-v2} of
Theorem~\ref{thm:generic} holds since
$A_0\subset\{1\}$ and $ \sum_{j=1}^2
\alpha_j\sup(A_j)=3\alpha_2<1$; and in this case all the $p_{\theta, \alpha}$ are super-critical.
\end{itemize}
\item Suppose $\supp(p)=\{0, 2\}$, and
$\ca=(\{0\}, \{2\})$. Notice that~\eqref{eq:patho}
does not hold. In this example all
probability distribution $p'$ such that $\supp(p')=\supp(p) $ is
$(p, \ca)$-compatible. The direction of $p'$ is given by
$(p'(0), p'(2))$. We recover that for all directions in $\Delta_2^*$
there exists a $(p, \ca)$-compatible probability
distribution. However, the probability distribution $p$ is generic for
$\ca$ only in the direction $(1/2, 1/2)$.
\end{enumerate}
\end{rem}
% \[ ************\]
\begin{proof}
We simply write $M_\alpha$ and $m_\alpha$ for the respective supremum
and infimum of
$\{\mu_{\theta, \alpha}\,\colon\, \theta\in I_\alpha\}$.
Let $\neg$ be the usual logical negation.
\medskip
We first prove that:
\begin{equation}
\label{eq:non-i}
\rho_\cj<+\infty \quad\text{and}\quad
\neg\ref{item:0inA0-v2} \, \implies \, M_\alpha\geq 1
.
\end{equation}
If $\rho_{\cj} < \infty$ and $g_{A_0}'(\rho_{\cj}) > 1$, then we have
$A_0\cap \{0, 1\}^c \neq \emptyset$, the function $g_{A_0}$ is strictly
convex. So there
exists $ \theta^* \in (\theta_{\min}, \rho_\cj)$ such that
$g_{A_0}(\theta^*)<\theta^*$ and $g_{A_0}'(\theta^*)=1$. Notice that
$\theta^*\in I_\alpha$ and use~\eqref{mu-expression} to deduce that
$\mu_{\theta^*, \alpha}\geq 1$.
If $\rho_{\cj} < \infty$, $g_{A_0}'(\rho_{\cj}) \leq 1$ and
$ g_\cj(\rho_{\cj}) = \infty$, then we have $\rho_\cj>1$ and
$\rho_{\cj}\not\in I_\alpha$ (by
Lemma~\ref{lem:p-qa}~\ref{item:p-alpha-q}). Since
$g_{A_0}'(\rho_{\cj}) \leq 1$, we deduce that
$g_{A_0}(\rho_\cj)<\rho_\cj$, and thus $\theta_{\max}= \rho_\cj$
by~\eqref{eq:q-max}. Since $ g_\cj(\rho_{\cj}) = \infty$, we deduce
there exists $j\in\cj^*$ such that $\rho_{A_j}=\rho_\cj$,
$g_{A_j}(\rho_\cj)=\infty$ as well as $\sup A_j=+\infty $. This implies
by Lemma~\ref{lem:hj-monotone} that $H_\alpha(\rho_\cj)=+\infty $ and
hence by continuity
that
$\lim_{\theta\uparrow \theta_{\max}} \mu_{\theta,\alpha}=+\infty $.
If $\rho_{\cj}< \infty$,
$g_{A_0}'(\rho_{\cj}) \leq 1$, $g_\cj(\rho_{\cj}) < \infty$ and:
\begin{equation}\label{eq:g-rho-less-infty-v2}
H_\alpha (\rho_{\cj})
\geq \rho_{\cj}\, \frac{1 - g_{A_0}'(\rho_{\cj})}{\rho_{\cj} -
g_{A_0}(\rho_{\cj})},
\end{equation}
then we have $g_{A_0}(\rho_\cj)< \rho_\cj$ and thus
$\theta_{\max}=\rho_\cj$ belongs to $I_\alpha$.
Use~\eqref{mu-expression} and~\eqref{eq:g-rho-less-infty-v2} to deduce
that $\mu_{\theta_{\max}, \alpha}\geq 1$.
This proves that~\eqref{eq:non-i} holds.
\medskip
We now consider the case $0\in A_0$. Thanks to
Lemma~\ref{lem:0-A-0-sub-sup}~\ref{item:sub-0-in-min}, we have
$m_\alpha< 1$. So $p$ is generic in the direction $\alpha$ if and
only if $M_\alpha\geq 1$. If Point~\ref{item:0inA0-v2} holds, then
$\theta_{\max}=\rho_\cj$ belongs to $I_\alpha$, and by~\eqref{mu-expression}
and~\eqref{eq:g-rho-less-infty} we get that $\mu_{\theta_{\max}, \alpha} <1$,
which by Proposition~\ref{prop:uniq} implies that $M_\alpha< 1$; thus $p$ is not
generic in the direction $\alpha$. Now assume that
Point~\ref{item:0inA0-v2} does not hold. If $\rho_\cj<+\infty$, then
use~\eqref{eq:non-i} to deduce that $M_\alpha\geq 1$ and that $p$ is
generic in the direction $\alpha$. If $\rho_\cj=+\infty$, then use
Lemma~\ref{lem:0-A-0-sub-sup}~\ref{item:sub-0-in-max} to get that
$M_\alpha\geq 1$, and thus $p$ is also generic in the direction
$\alpha$. This proves the theorem in the case $0\in A_0$.
\medskip We now consider the case $0\not \in A_0$. If
Point~\ref{item:0niA0-v2} holds, we deduce from
Lemma~\ref{lem:0-A-1-sub-sup}~\ref{item:sub-0-out-min} that $m_\alpha>1$
and thus $p$ is not generic in the direction $\alpha$.
We now assume that $\sum_{j\in\cj^*} \alpha_j \min A_j \leq 1$. Thanks
to Lemma~\ref{lem:0-A-1-sub-sup}~\ref{item:sub-0-out-min}, we get that
$m_\alpha\leq 1$. So $p$ is generic in the direction $\alpha$ if and
only if $M_\alpha\geq 1$. If $\rho_\cj=\infty $, we deduce from
Lemma~\ref{lem:0-A-1-sub-sup}~\ref{item:sub-0-out-max} that
$A_0\subset \{1\}$ and $\sum_{j\in\cj^*} \alpha_j \sup A_j < 1$ are
equivalent to $M_\alpha<1$, that is, $p$ is not generic in the direction
$\alpha$. We eventually assume that $\rho_\cj<\infty $. If
Point~\ref{item:0inA0-v2} holds,
then $\theta_{\max}=\rho_\cj$ belongs to $I_\alpha$,
$\mu_{\theta_{\max}, \alpha} <1$, and $M_\alpha< 1$ by
Proposition~\ref{prop:uniq}, and thus $p$ is not generic in the
direction $\alpha$. If Point~\ref{item:0inA0-v2} does not hold, then
use~\eqref{eq:non-i} to deduce that $M_\alpha\geq 1$ and that $p$ is
generic in the direction $\alpha$.
\end{proof}
\section{Local limit of large Galton-Watson trees}
\label{sec:main}
We give the main theorem, see Theorem~\ref{thm:main-thm}, on the local
limit of conditioned BGW tree in the next section. Its proof relies on a
transformation of BGW trees from Rizzolo, see Section~\ref{sec:rizzolo},
and a direct application of local limit theorems for multi-type BGW trees
from~\cite{adg18}, see Section~\ref{sec:proof-main}.
\subsection{Main result}
Let $p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}, and
let
$\ca=(A_j)_{j\in \II{1, J}}$, with $J\in \N^*$, be pairwise disjoint
non-empty subsets of $\supp(p)$. Let $\alpha\in \Delta^*_J$ be a
possible direction, see~\eqref{eq:possible-dir}.
We assume the distribution $p$ is
generic for $\ca$ in the direction $\alpha$ (recall Theorem
\ref{thm:generic}). Thus, there exists a (unique) $\qc\in I_\alpha$ such that:
\begin{equation}
\label{eq:def-palpha}
p_{\alpha}:=p_{\qc, \alpha}
\quad\text{is critical.}\quad
\end{equation}
Recall $\ct_q$ denotes a BGW tree with offspring distribution $q$ and
$\ct_q^*$ the corresponding Kesten tree when $\mu(q)\leq 1$.
Recall $|\bn|$ is the $L^1$-norm of $\bn \in \N^J$.
\begin{lem}
\label{lem:a-admiss}
If $p$ is generic for $\ca$ in the direction $\alpha\in \Delta_J^*$,
then there exists a sequence $(\bn^{(m)})_{ m\in \N}$ in $\N^J$
such that:
\begin{equation}\label{eq:cond-n_m2}
\P(L_\ca(\ct_p)=\bn^{(m)})>0,
\end{equation}
\begin{equation}
\label{eq:cond-n_m}
\lim_{m\rightarrow \infty }
|\bn^{(m)}|=\infty, \quad \lim_{m\rightarrow \infty }
\frac{\bn^{(m)}}{ |\bn^{(m)}|}=\alpha,
\end{equation}
and for all $m\in \N$, $j\in \II{1, J}$, with $\bn^{(m)}=(n_1^{(m)},
\ldots, n_J^{(m)})$:
\begin{equation}
\label{eq:cond-zero}
\alpha_j=0 \implies n^{(m)}_j= 0.
\end{equation}
\end{lem}
\begin{proof}
Since $p$ is generic for $\ca$ in the direction
$\alpha\in \Delta_J^*$, there exists a critical $(p, \ca)$-compatible
distribution $p_\alpha$ with direction $\alpha$. Thus, by
Definition~\ref{defi:on-p} of $(p, \ca)$-compatible probability
distribution, it is enough to find a sequence $(\bn^{(m)})_{ m\in \N}$
such that~\eqref{eq:cond-n_m2}-\eqref{eq:cond-zero} hold, with $\ct_p$
replaced by $\ct_{p_{\alpha}}$.
Let $\ct_{p_\alpha}^{(n)}$ be distributed as $\ct_{p_\alpha}$
conditioned to have $n$ vertices. Notice that for all finite $M>0$,
there exists $n>M$ such that the probability of $\ct_{p_ \alpha}$ to
have $n$ vertices is positive. Recall $L_k(\bt)$ denotes the number of
vertices in $\bt$ with out-degree $k$. According to \cite[Theorem
7.11]{j12}, along the sequence
$\{n\in \N\, \colon\, \P(\sharp \ct_{p_ \alpha}=n)>0\}$, the following
convergences hold in probability for all $k\in \N$:
\[
n^{-1} L_k(\ct_{p_\alpha}^{(n)})
\, \xrightarrow[n\to\infty]{\P} \,
p_{\alpha}(k).
\]
Since $(n^{-1} L_{k}(\ct_{p_\alpha}^{(n)}))_{k\in \N}$ is a random
probability distribution on $\N$, we also get that:
\begin{equation}\label{eq:conv-la}
\max_{j\in\cj}\val{n^{-1} L_{A_j}(\ct_{p_\alpha}^{(n)})
- p_{\alpha}(A_j)}\xrightarrow[n\to\infty]{\P} 0.
\end{equation}
\medskip
In particular, for $m\in \N^*$ large enough, we deduce that there exists
a tree $\bt^{[m]}$ such that $\sharp \bt^{[m]}\geq m$;
$\P(\ct_{p_\alpha} = \bt^{[m]})>0$; and, with $\bn^{(m)}=(n_1^{(m)}, \ldots,
n_J^{(m)})= L_\ca(\bt^{[m]})$,
$n^{(m)}_j=0$ if $\alpha_j=0$ and for all $j\in \cj^*$:
\[
\val{ \frac{n^{(m)}_j }{|\bn^{(m)}|} - \frac{p_{ \alpha}
(A_j)}{1-p_{\alpha} (A_0)} }
\leq \frac{1}{m}\cdot
\]
Recall also that $\P(\ct_{p_\alpha}=\bt)>0$ implies that
$\P(\ct_p=\bt)>0$.
This and the fact $\alpha$ is the direction
of $p_{\alpha}$, that is $p_\alpha(A_j)/(1- p_\alpha(A_0))=\alpha_j$, end the proof.
\end{proof}
Recall:
\[
\cj^*=\{j\in \II{1, J}\, \colon\, \alpha_j>0\}
\quad\text{and}\quad
\cj=\{0\}
\cup \cj^*.
\]
For $A\subset \N$, we write $A-1=\{a-1\,
\colon\, a\in A\}$ and $A-A=\{a-b\, \colon\, a,b\in A\}$. We define:
\begin{equation}
\label{eq:Gamma-a}
\Gamma_\alpha= \bigcup_{j\in \cj}\, A'_j,
\quad\text{where}\quad
A'_0=A_0 -1
\quad\text{and}\quad
A'_j=A_j-A_j
\quad\text{for}\quad j\in \cj^*.
\end{equation}
\begin{defi}[Aperiodicity]
\label{defi:period}
Let $p$ be generic for $\ca$ in the direction $\alpha\in \Delta_J^*$.
The probability distribution $p$ is aperiodic for $\ca$ in the direction
$\alpha$ if $\qc\in (0, +\infty )$ and the smallest subgroup of
$\Z$ that contains $\Gamma_\alpha$ is $\Z$.
\end{defi}
\begin{rem}[On aperiodicity]
\label{rem:period0} $ $
\begin{enumerate}[(a)]
\item\label{it:period-0A0}
If $0\in A_0$, then $-1\in A'_0$, and thus the
distribution $p$ is
aperiodic for $\ca$ in the direction~$\alpha$.
\item \label{it:patho-singleton}
If $p$ is aperiodic for $\ca$ in the direction
$\alpha$, then~\eqref{eq:patho} holds.
\item\label{it:period-0infty} Looking carefully at the proof of
Lemma~\ref{p-ca-critical}~\ref{it:period}, it would be more natural to
consider $\gamma'_\alpha$ defined as $\Gamma_\alpha$
in~\eqref{eq:Gamma-a} but with $A_j$ replaced by
$A_j\cap \supp (p_\alpha)$. For $\qc\in (0, +\infty )$ this yields no
modification as $\Gamma'_\alpha=\Gamma_\alpha$. However, for
$\qc\in \{0,\infty\}$ (which is ruled out in
Definition~\ref{defi:period}), the set $\Gamma'_\alpha$ is reduced to
$\{0\}$ (as $A_j\cap \supp(p_\alpha)$ is either a singleton of
empty). So, with the more natural definition that $p$ is aperiodic for
$\ca$ in the direction $\alpha$ if the smallest subgroup of $\Z$ that
contains $\Gamma'_\alpha$ is $\Z$, then we still have
$\qc\in (0,\infty)$.
\end{enumerate}
\end{rem}
We now state the main result. Notice we don't assume that $p$ has a
finite mean.
\begin{theo}[Local limit of conditioned BGW tree]
\label{thm:main-thm}
Let $ p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}. Let $\ca=(A_j)_{j\in \II{1, J}}$,
with $J\in \N^*$, be a family of pairwise
disjoint non-empty subsets of $\supp(p)$ and
consider the direction
$\alpha \in \Delta_J^*$.
We assume that $p$ is generic and aperiodic for $\ca$ in
the direction $\alpha$ (and thus $\qc\in (0, +\infty )$).
If $(\bn^{(m)})_{ m\in \N}$ is a sequence in $\N^J$ satisfying
\eqref{eq:cond-n_m2}-\eqref{eq:cond-zero}, then we have:
\[
\dist(\ct_p\,|\, L_{\ca}(\ct_p)=\bn^{(m)})
\xrightarrow[m\to\infty]{} \dist(\ct^*_{p_{\alpha}}).
\]
\end{theo}
We refer to Remark~\ref{rem:tbd} and the details of its proof given in
Section~\ref{sec:a=0-n>0}
for the usefulness of the condition~\eqref{eq:cond-zero}, that is, $\alpha_j=0
\implies n^{(m)}_j= 0$ for the sequence $(\bn^{(m)})_{ m\in \N}$.
\begin{rem}[Conditioning on the total size and the number of leaves]
\label{rem:restriction}
Notice that conditioning on the total size and the number of leaves,
or equivalently on the number of internal nodes and the number of
leaves, corresponds to $A_1=\N^*\cap \supp(p)$ and $A_2=\{0\}$. Notice
that $A_0=\emptyset$. (By
Remark~\ref{rem:period0}~\ref{it:patho-singleton} notice that the case
$A_1$ reduced to a singleton is excluded from
Theorem~\ref{thm:main-thm}, but then it is equivalent to condition on
the number of leaves; and this is considered in~\cite{ad14a}.)
Provided the assumptions on genericity and aperiodicity are satisfied,
, this case is included in Theorem~\ref{thm:main-thm} whereas it is
excluded \emph{a priori} in Corollary 3.5 in~\cite{adg18}.
\end{rem}
\begin{rem}[On the case $\cj^{**}=\emptyset$ and $A_0=\emptyset$ or $A_0=\{1\}$]
\label{rem:singleton}
Let
$j_0\in \cj^*$ be such that $A_{j_0}=\{0\}$. Notice that $\alpha_{j_0}\in
(0, 1)$. Since $\cj^{**}=\emptyset$ and $A_0\subset \{1\}$,
the aperiodic hypothesis is not satisfied. However the
condition~\eqref{eq:kt=L0-1} implies no choice on $L_0$, so that we can
without loss of generality replace $A_0$ by $A'_0=A_0\cup \{0\}$ and
remove $j_0$ from $\cj^*$, as well as $\alpha$ by $\alpha'\in \R_+^{J-1}$ with
$\alpha'_j=\alpha_j/(1- \alpha_{j_0})$ for $j\in \II{1, J}\setminus\{j_0\}$.
Then the conditioning is the same and the distribution $p$ is aperiodic
for $\ca'=(A_j)_{j\in \II{1, J}\setminus\{j_0\}}$ in the direction $\alpha'$.
Using this trick, we see that the local convergence of Theorem~\ref{thm:main-thm} holds in this case, eventhough $p$ is not aperiodic.
\end{rem}
The next two sections are devoted to the proof of the theorem. In
Section~\ref{sec:rizzolo}, for a tree $\bt$ such that
$\cl_\ca(\bt)\neq \emptyset$, we describe a map from $\cl_\ca(\bt)$ onto
a multi-type tree, which is a direct extension of Rizzolo \cite{r15}.
Then, in Section~\ref{sec:proof-main}, we use \cite{adg18} on local
limit of multi-type BGW trees to conclude.
\medskip
From now on the direction $\alpha\in \Delta_J^*$ is fixed and $p_\alpha$
given by~\eqref{eq:def-palpha}
is the unique critical probability distribution $(p, \ca)$-compatible
with the direction~$\alpha$. In particular, we have
$\mu(p_\alpha)=1$. By construction, we have $p_\alpha(A_j)=0$ if
$\alpha_j=0$ for $j\in \II{1, J}$. Since only the indices $j$ such that
$\alpha_j$ is positive are pertinent, for a sequence
$\bx=(x_j)_{j\in \II{1, J}}$, we shall consider the subsequence:
\begin{equation}
\label{eq:def-restric*}
\bx^*=(x_j)_{j\in \cj^*},
\end{equation}
where, we recall that $\cj^*=\{j\in \II{1,J}\, \colon\,
\alpha_j>0\}$.
For example, we write $L_{\ca^*}(\bt)=(L_{A_j}(\bt))_{j\in \cj^*}$.
\subsection{Extension of Rizzolo's transformation}
\label{sec:rizzolo}
In the following, we use the framework for multi-type trees from
\cite[Section 2]{adg18}. For a tree $\bt$ such that
$\cl_\ca(\bt)\neq \emptyset$, we describe a map from $\cl_\ca(\bt)$ onto
a multi-type tree, which is a direct extension of \cite{r15}.
The vertex $u\in\bt$ is said to have type $j\in\cj$, which we denote by
$e_\bt(u)=j$, if $k_u(\bt)\in A_j$ so that $(\bt,e_\bt)$ can be treated
as a $\cj$-type tree. Note $\bt^{[i]}=\{u\in\bt: e_\bt(u)=i\}$. In
order to remove the $0$-type vertices, following \cite{r15}, we build a
bijection $\phi$ (depending on $(\bt, e_\bt)$) from
$\bt\backslash\bt^{[0]}$ to a tree $\bt^\ca$ with a $\cj^*$-type
$e_{\bt^\ca}$, which preserves the types, that is,
$e_{\bt^\ca} (\phi(u))=e_\bt(u)\in \cj^*$. Furthermore, if $\bt^{[0]}$
is empty (which is automatically the case if $A_0=\emptyset$), then we
shall have $ \bt^\ca=\bt$ and $\phi$ is the identity map (and thus
$e_\bt=e_{\bt^\ca}$).
Let $(\bt, e_\bt)$
be a $\cj$-type tree such that
$\sharp(\bt\backslash\bt^{[0]})=n\geq 1$.
Following \cite{ad14a}, we define recursively a sequence of growing
$\cj^*$-type trees $(\bt_k, e_{\bt_k})_{k\in \II{1, n}}$ and identify
the last one as $(\bt^\ca, e_{\bt^\ca})$. The map $\phi $ is a
by-product of this construction.
Denote $\prec$ the lexicographic order on $\cu$.
Let $u_1\prec\cdots\prec u_n$ be the ordered list of vertices of
$\bt\backslash \bt^{[0]}$. Then, we define recursively:
\begin{itemize}
\item $\phi(u_1)=\emptyset$, $\bt_1=\{\emptyset\}$
and $e_{\bt_1}(\emptyset)=e_{\bt}(u_1)$.
\item For $1<k\leq n$, let $M(u_{k-1},u_k)\in \{u_1, \ldots, u_{k-1}\}$
be the most recent common ancestor of $u_{k-1}$ and $u_k$ and $\bs$ the
fringe subtree of $\bt$ above $M(u_{k-1},u_k)$,
see~\eqref{eq:def-fringe}.
Note $v=\min\{u\in\bs: e_{\bt}(u)\neq 0\}$ (for the lexicographic order).
Then we set $ \phi(u_k)$ as the concatenation of $\phi(v)$ and
$(k_{\phi(v)}(\bt_{k-1})+1)$ and consider the tree:
\[
\bt_k=\bt_{k-1}\cup \{ \phi(u_k)\},
\]
and the type map $e_{\bt_{k}}$ coincide with $e_{\bt_{k-1}}$ on
$\bt_{k-1}$ and $e_{\bt_k}(\phi(u_k)) =e_\bt(u_k)$. (This ensures that
$\phi$ preserves indeed the types.)
\end{itemize}
It is obvious that $(\bt_k,e_{\bt_k})_{k\in \II{1, n}}$ is a sequence of
(increasing) multi-type trees. Let
$(\bt^\ca,e_{\bt^\ca})=(\bt_n,e_{\bt_n})$ and we view $\phi$ as a
bijection from $\bt\backslash\bt^{[0]}$ to $\bt^\ca$ which preserves the
types. See Fig.~\ref{fig:map} for an example of $\bt$ and $\bt^\ca$ and their
types.
Notice that
$L_{A_j}(\bt)=\mathrm{Card}\,\{u\in\bt^\ca\, \colon\,
e_{\bt^\ca}(u)=j\}$ is the total progeny of type $j$ (which is equal to
0 if $j\not\in \cj^*$). For the $\cj^*$-type tree $(\bt^\ca,e_{\bt^\ca})$,
we denote by $\sharp \bt^{\ca}$ the vector of the total progeny of each type
in $\cj^*$ of $\bt^\ca$:
\begin{equation}
\label{eq:def-|tA|}
\sharp \bt^{\ca}=L_{\ca^*}(\bt)\in \N^{\cj^*}.
\end{equation}
\begin{figure}[ht]
\begin{center}
\includegraphics{fig1-1.eps}
\end{center}
\caption{ A tree $\bt$ on the left with $A_0=\{2\}$, $A_1=\{1,3\},$
$A_2=\{0\}$ and the tree $\bt^\ca$ after the map on the right.
We represent type 0 with
$\circ$, type 1 with $\diamond $, and type 2 with
$\bullet$.}
\label{fig:map}
\end{figure}
\medskip
Let $\ct_{\alpha}$ be a BGW tree with critical offspring distribution
$p_{\alpha}$. Let $\ct_{\alpha,*}$ be distributed as $\ct_\alpha$
conditioned to have at least one vertex with out-degree in $A_0^c$, that
is, on $\{L_\ca(\ct_\alpha) \neq \zero\}$, and, with a slight abuse of
notation, we set $(\ct_{\alpha}^\ca,e_{\ct_{\alpha}^\ca})$ the
$\cj^*$-type tree associated with $\ct_{\alpha,*}$ by the previous
construction. The proof of the next result, which is left to the
reader, is an adaptation of the proof of~\cite[Theorem 6]{r15}.
\begin{lem}
\label{lem:J*-BGW}
The random tree $(\ct_{\alpha}^\ca,e_{\ct_{\alpha}^\ca})$ is a
multy-type BGW tree, with types in $\cj^*$.
\end{lem}
The root of $\ct_{\alpha}^\ca$ is of type $e_{\ct_{\alpha}^\ca}(\emptyset)=j$ with probability
$\alpha_j$ for $j\in \cj^*$.
Let $p^{\ca,\alpha}=(p_j^{\ca,\alpha})_{j\in \cj^*}$ be the offspring distribution of the
$\cj^* $-type BGW tree $(\ct_{\alpha}^\ca,e_{\ct_{\alpha}^\ca})$,
where $p^{\ca,\alpha}_j$, a probability distribution on $\N^{\cj^*}$, is the
offspring distribution of an individual of type $j$.
To describe $p^{\ca,\alpha}_j$, we introduce several intermediate
random variables.
\begin{enumerate}[(a)]
\item Let $X$ be a random variable on $\N$ distributed according
to $p_\alpha$.
\item Let $X^j$ be distributed as $X$ conditionally on $\{X\in A_j\}$.
\item Let $(X^0_i)_{i\in\N}$ be independent random
variables distributed as $X-1$ conditionally on $\{X\in A_0\}$.
\item Let $N$ be a geometric random
variable with parameter $p_{\alpha}(A_0^c)$.
\end{enumerate}
We assume that the random variables $X^j$, $(X^0_i)_{i\in\N}$ and $N$
are independent. We adopt the convention $\inf \emptyset=+\infty $.
\begin{enumerate}[resume*]
\item\label{it:def-T} Set $ T=\inf\left\{n\in\N^*\,\colon\, \sum_{i=1}^n
X_i^0=-1\right\}$.
\item\label{it:loi-Y} Set $Y_j= X^j+\sum_{i=1}^{N-1} X^0_i$ on the event $\{N\le T\}$ and $Y_j=0$ otherwise.
\item\label{it:loi-Z} Conditionally on the above random variables, let $Z_j$ be a binomial
random variable with parameters $(Y_j, r)$, where:
\begin{equation}
\label{eq:def-r}
r=\P(N\leq T)\in (0, 1].
\end{equation}
\item\label{it:loi-XA}
Conditionally on the above random variables, let
$X^\ca_j=(X^\ca_j(i))_{i\in \cj^*}$ be a
multinomial random variable with parameter
$(Z_j, \alpha^*)$.
\end{enumerate}
Then, the probability distribution $p_j^{\ca,\alpha}$ is defined as the law of
$X_j^\ca$ conditionally on $\{N\leq
T\}$.
\medskip
Recall that $p^{\ca,\alpha}$ is said to be aperiodic, if the smallest
subgroup of $\Z^{\cj^*}$ that contains $\bigcup_{j\in \cj^*}
\big(\supp(p_j^{\ca,\alpha})-\supp(p_j^{\ca,\alpha})\big) $ is $\Z^{\cj^*}$. The
mean matrix $M=(m_{j\ell})_{ j,\ell\in\cj^*}$ of $p^{\ca,\alpha}$ is defined by:
\begin{equation}
\label{eq:def-mean-matrix}
m_{j\ell}=\E[X^\ca_j(\ell)\, |\, N\leq T].
\end{equation}
The offspring distribution $p^{\ca,\alpha}$ is critical if the spectral radius
of the mean matrix $M$ is one.
We have the following properties.
Recall $\qc$ is the unique $\theta\in [0, +\infty ]$ such that
$p_{\theta, \alpha}$ is critical.
\begin{lem}[Properties of the offspring distribution $p^{\ca,\alpha}$]
\label{p-ca-critical}
Let $ p$ be a probability distribution on $\N$
satisfying~\eqref{assumption-p}, such that it is generic for $\ca$
in the direction $\alpha\in \Delta^*_J$.
\begin{enumerate}[(i)]
\item\label{it:crit}
The offspring distribution $p^{\ca,\alpha}$ is critical and
$\alpha^*$ is the left eigenvector of the mean matrix associated with
the eigenvalue 1.
\item\label{it:primitiv} We have $m_{j\ell}>0$ for all
$j, \ell \in\cj^*$ and $j\neq j_0$, where $j_0\in \cj$ is defined by
$0\in A_{j_0}$. If $j_0\in \cj^*$, then we have that $m_{j_0\ell}=0$
for all $ \ell \in\cj^*$ either if $A_{j_0}= \{0\}$ and
$A_0\subset \{1\}$ or if $\qc=0$, and that $m_{j_0\ell}>0$ for
all $ \ell \in\cj^*$ otherwise.
\item\label{it:period}
The offspring distribution $p^{\ca,\alpha}$ is aperiodic if and only
if $p$ in
aperiodic for $\ca$ is the direction $\alpha$ (and thus $\qc\not\in \{0, +\infty \}$).
\end{enumerate}
\end{lem}
\begin{proof}
We prove Point~\ref{it:crit} on the criticality of $p^{\ca,\alpha}$. By
construction and~\eqref{eq:def-mean-matrix},
the entries of the mean matrix $M$ are given by, for
$j,\ell\in \cj^*$:
\[
m_{j \ell}=\E[X_j^\ca(\ell)\, |\, N\leq T]
=\E[Z_j\, |\, N\leq T]\, \alpha_\ell
=\E[Y_j]\alpha_\ell .
\]
In particular the mean matrix has rank one and
$\alpha^*$ is the left eigenvector
associated with the non-zero eigenvalue, say $\rho$.
Since the mean matrix has nonnegative entries and $\alpha^*$ as positive
entries, we also get that $\rho$ is the Perron-Frobenius eigenvalue and
thus the spectral radius of $M$.
Since $M$ has
rank one, we also get that $\rho$ is the trace of $M$:
\[
\rho=\sum_{j\in \cj^*} \E[Y_j]\, \alpha_j.
\]
We now compute $r\E[Y_j]$ using~\eqref{eq:def-r}:
\[
\E[Y_j]= \E\left[\left(X^j+\sum_{i=1}^{N-1}
X^0_i\right)\ind_{\{N\leq T\}}\right]
=r\E[X^j]+\E\left[\sum_{i=1}^{N-1} X^0_i\right]-
\E\left[\sum_{i=1}^{N-1} X^0_i\ind_{\{N\geq T\}}\right].
\]
Using the strong Markov property of $(X_i^0)_{i\in \N^*}$ at the
stopping time $T$ and its definition (see~\ref{it:def-T}), we get:
\begin{align*}
\E[Y_j]
&=r\E[X^j]+\E\left[\sum_{i=1}^{N-1} X^0_i\right]-
(1-r)\left(-1+\E\left[\sum_{i=1}^{N-1} X^0_i\right]\right)\\
&=1+ r\E\left[X^j-1+\sum_{i=1}^{N-1} X^0_i \right]\\
&=1+r\left(\frac{m_j}{p_{\alpha}(A_j)}-1
+\frac{m_0-p_{\alpha}(A_0)}{p_{\alpha}(A_0^c)}\right)\\
&=1+r\left(\frac{m_j}{p_{\alpha}(A_j)}
+\frac{m_0-1}{p_{\alpha}(A_0^c)}\right),
\end{align*}
where $m_\ell= \sum_{k\in A_\ell} k p_\alpha(k)$ for $\ell\in \cj$. As
$p_\alpha$ is critical, we get that $\sum_{j\in \cj} m_j=1$. Recall
that $\alpha_j=p_\alpha(A_j)/p_\alpha(A_0^c)$ for $j\in \cj^*$.
Therefore, we obtain:
\[
\rho
= \sum_{j\in \cj^*}\alpha_j+r\sum_{j\in
\cj^*}\frac{m_j}{p_\alpha(A_0^c)}
+r\frac{m_0-1}{p_\alpha(A_0^c)}\sum_{j\in \cj^*}\alpha_j
=1+r\frac{1-m_0}{p_\alpha(A_0^c)}+r\frac{m_0-1}{p_\alpha(A_0^c)}=1.
\]
This ensures that $p^{\ca,\alpha}$ is critical.
\medskip We prove Point~\ref{it:primitiv} on the positive entries of the
mean matrix. Let $j\in \cj^*$. We deduce from~\eqref{eq:def-mean-matrix}
and from~\ref{it:loi-XA} (where $\alpha^*$ has positive entries)
and~\ref{it:loi-Z} (where $r>0$) that $(m_{j\ell})_{\ell\in \cj^*}$ are
all positive if $\P(Y_j=0)<1$ and all zero if $\P(Y_j=0)=1$.
Notice that $T=1$ a.s.\ implies $A_0=\{0\}$ and thus
$Y_j>0$ a.s. on $\{N\leq T\}$, so $\P(Y_j=0)=1$ implies that $\P(T\geq 2)>0$. We
deduce that $\P(Y_j=0)=1$ is equivalent to
$\P(X^j=0)=1$ and $\P(N=1)=1$ or $\P(X_i^0=0)=1$. Thus $\P(Y_j=0)=1$ is
equivalent to $0\in A_j$, $p_\alpha(A_j\cap \{0\}^c) =0$ and
$p_\alpha(A_0)=0$ or $p_\alpha(A_0 \cap\{1\}^c)=0$, that is, $0\in A_j$,
$p_\alpha(A_j\cap \{0\}^c) =0$ and $p_\alpha(A_0 \cap\{1\}^c)=0$. To
conclude, notice that those conditions are equivalent to either
$A_j=\{0\}$ and $A_0\subset \{1\}$ or $0\in A_j$ and $\qc=0$.
\medskip
We prove Point~\ref{it:period} on the periodicity of $p^{\ca,\alpha}$.
Thanks to~\ref{it:loi-XA} and the fact that $\alpha^*$ has positive
entries, we deduce that $p^{\ca,\alpha}$ is aperiodic (in $\Z^{\cj^*}$) if and
only if
the smallest
subgroup of $\Z$ that contains $\Gamma=\bigcup_{j\in \cj^*}
\big(\supp(\text{Law}(Z_j\, |\, N\leq T))-\supp(\text{Law}(Z_j\, |\, N\leq T))\big)
$ is $\Z$.
From the definition of the law of $Z_j$ given in~\ref{it:loi-Z}, we
shall consider the two cases $r=1$ and $r<1$.
We also remark that $r=1$ if and only if
$T=+\infty $ a.s.\ or $N=1$ a.s, which corresponds to
$0\not\in A_0\cap \supp(p_\alpha)$, that is, $0\not\in A_0$ as $p_\alpha(0)>0$.
In the easy case $0\in A_0$ (and thus $\qc\in (0, +\infty )$), we get on
the one hand
that $X^j>0$ a.e., and as $\P(N=1)>0$, we deduce that $\P(Y_j>0)>0$ and
thus that $\{0, 1\}\subset \supp(\text{Law}(Z_j\, |\, N\leq T))$. This implies that
$p^{\ca,\alpha}$ is aperiodic.
On the other hand, we also get that $p$ is
$(\ca, \alpha)$-aperiodic, see
Remark~\ref{rem:period0}~\ref{it:period-0A0}.
We now consider the case $0\not\in A_0$, that is $r=1$ and thus
$Z_j=Y_j$ and a.s.\ $T=+\infty $.
If $p_\alpha(A_0)>0$, we have $\P(N=k)>0$ for all $k\in \N$, and if
$p_\alpha(A_0)=0$, we have a.s.\ $N=1$ and $A_0\cap \supp(p_\alpha) \subset \{1\}$. In
both cases, we deduce from~\ref{it:loi-Y} that:
\[
\supp(\text{Law}(Z_j\, |\, N\leq T))=\supp(\text{Law}(Y_j\, |\, N\leq T))=\supp(\text{Law}(X^j\, |\, N\leq T))+ \N
( A_0^\alpha-1),
\]
where $\N B=\{n b\, \colon, n\in \N, b\in B\}$
and $A_0^\alpha=A_0 \cap
\supp(p_\alpha)$.
We set $A_j^\alpha= A_j\cap \supp(p_\alpha)$ for $j\in \cj^*$ and notice that $A_j^\alpha
=\supp(\text{Law}(X^j\, |\, N\leq T))$. We then
get that:
\[
\Gamma=\Big(\bigcup_{j\in \cj^*} (A^\alpha_j- A^\alpha_j)\Big)
\cup \Z (A_0^\alpha-1).
\]
If $\qc\in \{0, +\infty\}$, we get that $A_0^\alpha\subset \{1\}$ and
$A^\alpha_j$ are singletons for $j\in \cj^*$, so that $\Gamma=\{0\}$.
If $\qc\in (0, +\infty)$, we get that $\Gamma=\Gamma_\alpha$ defined
in~\eqref{eq:Gamma-a}.
Thus $p^{\ca,\alpha}$ is aperiodic if and only if $\qc\not\in \{0,
+\infty \}$ and the smallest subgroup in $\Z$ containing $\Gamma_\alpha$
is $\Z$, that is, $p$ is
aperiodic for $\ca$ in the direction $\alpha$.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:main-thm}}
\label{sec:proof-main}
Recall
(see Section \ref{sec:lct}) the set of trees $\T(\bt,x)$ for $\bt\in
\T_0$ and $ x\in \cl_0(\bt)$.
As $p$ is generic for $\ca$ in the direction $\alpha$, there exists by
Lemma \ref{lem:a-admiss} a sequence $(\bn^{(m)})_{ m\in\N}$ satisfying~\eqref{eq:cond-n_m2}-\eqref{eq:cond-zero}.
Since $p_\alpha$ is $(p,\ca)$-compatible, we have for $m\in\N$, $\bt\in\T_0$
and $x\in\cl_{\{0\}}(\bt)$, with $\ct_\alpha=\ct_{p_\alpha}$ and
$\ct^*_\alpha=\ct^*_{p_\alpha}$, that:
\[
\P(\ct_p\in \T(\bt,x)| L_\ca(\ct_p)=\bn^{(m)})
=\P(\ct_\alpha\in \T(\bt,x)| L_\ca(\ct_{\alpha})=\bn^{(m)}).
\]
\medskip
For $j\in \II{1, J}$, recall $\be^j\in \N^ J$ is the vector with all its entries equal to $0$ but the $j$-th which is equal to $1$, and that $\be^0=\zero$.
Set $j_0\in \II{0, J}$ such that $0\in A_{j_0}$ and set:
\[
\bb=\be^{j_0}.
\]
We have:
\begin{align*}
\P(\ct_{\alpha}\in \T(\bt,x), L_\ca(\ct_{\alpha})=\bn^{(m)})
&=\sum_{\tilde{\bt}\in \T_0}\P(\ct_{\alpha}=\bt\circledast (\tilde{\bt},x))
\ind_{\{L_\ca(\bt\circledast (\tilde{\bt},x))=\bn^{(m)}\}}\\
&=\frac{1}{p_\alpha(0)}\sum_{\tilde{\bt}\in \T_0}\P(\ct_{\alpha}=\bt)\,
\P(\ct_{\alpha}=\tilde{\bt})
\ind_{\{L_\ca(\tilde{\bt})=\bn^{(m)}-L_\ca(\bt)+\bb\}}\\
&=\frac{1}{p_\alpha(0)}\P(\ct_{\alpha}=\bt)\,
\P(L_\ca(\ct_{\alpha})=\bn^{(m)}-L_\ca(\bt)+\bb)\\
&=\P(\ct_{\alpha}^*\in\T(\bt,x))\,\P(L_\ca(\ct_{\alpha})
=\bn^{(m)}-L_\ca(\bt)+\bb),
\end{align*}
where we used that $p_\alpha$ is critical (and thus $\ct_\alpha$ is
a.s.\ finite) for first and third equalities
and~\eqref{eq:loi-Tp*}
for the last.
Recall the notation $\bx^*$ from~\eqref{eq:def-restric*}
which is the restriction of the sequence $\bx$ indexed by $\II{1, J}$ to
the indices $\cj^*$, and that $L_{\ca^*}(\bt)=(L_{A_j}(\bt))_{j\in \cj^*}$.
We get:
\begin{align}
\P(\ct_{\alpha}\in \T(\bt,x)\,|\, L_\ca(\ct_{\alpha})=\bn^{(m)})
&=\P(\ct_{\alpha}^*\in \T(\bt,x))\, \frac{\P(L_\ca(\ct_{\alpha})
=\bn^{(m)}-L_\ca(\bt)+\bb)}{\P(L_\ca(\ct_{\alpha})=\bn^{(m)})}\nonumber\\
&=\P(\ct_{\alpha}^*\in \T(\bt,x))
\,\frac{\P(L_{\ca^*}(\ct_{\alpha})=\bn^{(m)*}-L_{\ca^*}(\bt)+\bb^*)}
{\P(L_{\ca^*}(\ct_{\alpha})=\bn^{(m)*})},\label{eq:palpha-L-A}
\end{align}
where we used that
$L_{A_j}(\ct_{\alpha})=0$ for $j\not\in \cj$
(and thus $\P(\ct_{\alpha}\in \T(\bt,x))=\P(\ct^*_{\alpha}\in
\T(\bt,x))=0$ if $L_{A_j}(\bt) \neq 0$ for some $j\not\in \cj$) as well
as that $(\bn^{(m)})_{ m\in
\N}$ satisfies~\eqref{eq:cond-zero} for the second equality.
Now we apply the extension of Rizzolo's transformation
for $\ct_{\alpha}$ to get a $\cj^*$-type BGW tree
$\ct_{\alpha}^\ca$ such that
$\sharp \ct_{{\alpha}}^\ca=L_{\ca^*}(\ct_{{\alpha}})$ (see
definition~\eqref{eq:def-|tA|}).
Hence ~\eqref{eq:palpha-L-A} is equivalent to:
\begin{equation}
\label{eq:pre-19}
\P(\ct_{\alpha}\in \T(\bt,x)\,|\, L_\ca(\ct_{\alpha})=\bn^{(m)})
=\P(\ct_{\alpha}^*\in \T(\bt,x))
\, \frac{\P(\sharp \ct_{{\alpha}}^\ca=\bn^{(m)*}-L_{\ca^*}(\bt)+\bb^*)}
{\P(\sharp \ct_{{\alpha}}^\ca=\bn^{(m)*})}\cdot
\end{equation}
We consider the following condition which appears in
Lemma~\ref{p-ca-critical}~\ref{it:primitiv}:
\begin{equation}
\label{eq:delicat}
A_0\subset \{1\}
\quad\text{and}\quad
A_{j_0}=\{0\}
\quad\text{for some}\quad
j_0\in \cj^*.
\end{equation}
We first assume that~\eqref{eq:delicat} does not hold.
Hypothesis of Theorem~\ref{thm:main-thm} and Lemma~\ref{p-ca-critical}
ensure that Assumptions $(H_1)$ (on the offspring distribution being
critical and the mean matrix primitive) and $(H_2)$ (on the aperiodicity
of the offspring reproduction) hold in~\cite{adg18} and that $\alpha^*$
is the positive normalized left eigenvector of the mean matrix (see
hypothesis in Lemma~3.11 in~\cite{adg18} where $a=\alpha^*$ and use that
$(\bn^{(m)}, \, m\in \N)$ is a sequence in $\N^J$ satisfying
\eqref{eq:cond-n_m2}-\eqref{eq:cond-zero}), so that the strong ratio
theorem or more precisely (19) in~\cite{adg18} holds, which entails
that:
\begin{equation}
\label{eq:19}
\lim_{m\to\infty}\frac{\P(\sharp \ct_{{\alpha}}^\ca=\bn^{(m)*}-L_{\ca^*}(\bt)+\bb^*)}
{\P(\sharp \ct_{{\alpha}}^\ca=\bn^{(m)*})}=1.
\end{equation}
We deduce from~\eqref{eq:pre-19}
and~\eqref{eq:19}
that for all $\bt\in \T_0$ and $x\in \cl_{\{0\}}(\bt)$:
\[
\lim_{m\to\infty}\P(\ct_p\in \T(\bt,x)| L_\ca(\ct_p)=\bn^{(m)})
=\P(\ct_{\alpha}^*\in\T(\bt,x)).
\]
For $\bt\in\T_0$, it is obvious from~\eqref{eq:cond-n_m}
that $
\lim_{m\to\infty}\P(\ct_p=\bt| L_\ca(\ct_p)=\bn^{(m)})
=\P(\ct_{\alpha}^*=\bt)=0.
$
The result thus follows from the fact that the family $\{(\T(\bt,x),\ \bt\in
\T_0,\ x\in \cl_0(\bt)\}\cup\T_0$ is convergence determining for the
local convergence in $\T_0\cup\T_1$.
\medskip We now consider that~\eqref{eq:delicat}
holds. We first check that it is enough to consider the case
$A_0=\emptyset$, where the Rizzolo's transformation is the identity map.
Indeed, if $A_0\neq \emptyset$, that is, $A_0=\{1\}$, then the Rizzolo's
transformation corresponds to discarding individuals with only one
child. This amounts to replace the offspring
distribution $p$ (resp. $p_{\theta,\alpha}$) by $p'$
(resp. $p'_{\theta,\alpha}=(p')_{\theta,\alpha}$) where $p'(k)=p(k)/q_1$
for $k\neq 1$ and $p'(k)=0$ for $k=1$. Then, notice that $\theta$ s.t.\
$p'_{\theta,\alpha}$ is critical is exactly $\qc$, so without
confusion, we can also replace $p_\alpha$ by $p'_\alpha=(p')_\alpha$. In
conclusion, using this modification amounts to only consider the case:
\begin{equation}
\label{eq:delicat2}
A_0=\emptyset
\quad\text{and}\quad
A_{j_0}=\{0\}
\quad\text{for some}\quad
j_0\in \cj^*.
\end{equation}
Notice this case is ruled out in~\cite[Corollary~3.5]{adg18}.
However a
slight modification of the proofs in~\cite{adg18}, which
we sketch in
Section~\ref{sec:appendix} (take $d=\Card(\cj^*)$ and
$d=j_0$ therein) allows to get~\eqref{eq:19}, which we now read as, for
a sequence
$(\bn^{(m)}, \, m\in \N)$ in $\N^J$ satisfying
\eqref{eq:cond-n_m2}-\eqref{eq:cond-zero}:
\begin{equation}
\label{eq:19-A0=0}
\lim_{m\to\infty}\frac{\P(\sharp \ct_{{\alpha}}=\bn^{(m)*}-L_{\ca^*}(\bt)+\bb^*)}
{\P(\sharp \ct_{{\alpha}}=\bn^{(m)*})}=1,
\end{equation}
where $\ct_\alpha$ is seen as a multi-type BGW tree, where a node
$u\in \ct_\alpha$ as type $j\in \cj^*$ if $k_u(\ct_\alpha)\in A_j$.
Notice the corresponding offspring
distribution is
$\bp=(\bp^{(i)})_{i\in \cj^*}$ where
$ \bp^{(i)}=( \bp^{(i)}(\bk))_{\bk\in \N^{\cj^*}}$ is a
probability distribution on $\N^{\cj^*}$ whose non-zero terms are given by:
\[
\bp^{(i)}(\bk)= \frac{ p_\alpha(|\bk|)}{p_\alpha(A_i)}
\, \mathrm{Mult}(\bk, \alpha^*)
\quad\text{for}\quad
|\bk|\in A_i,
\]
and $\mathrm{Mult}(\bk, \alpha^*)$ is the probability that a
multinomial random variable with parameter $(|\bk|, \alpha^*) $
takes the value $\bk$. Furthermore the type of the root is distributed
as $\alpha^*$. With this setting, we emphasize
that~\eqref{eq:19-A0=0} is exactly (19) in~\cite{adg18}, up to a relabeling.
Once~\eqref{eq:19-A0=0} is established, then we finish the proof as
in the case where~\eqref{eq:delicat} does not hold.
We now give the properties of the offspring
distribution $\bp$ and the type of the root (recall
that~\eqref{eq:delicat2} holds and that the Rizollo's
transformation is the identity map); Under assumption of
Theorem~\ref{thm:main-thm}, we have:
\begin{enumerate}[(a)]
\item The type of the root is distributed as $\alpha^*$.
\item $\bp^{(j_0)}(\zero)=1$.
\item By Lemma~\ref{p-ca-critical}~\ref{it:primitiv}, the mean
matrix $M=(m_{ij})_{i,j\in \cj^*}$ is such that for $j\in \cj^*$ we
have $m_{ij}\in (0, +\infty
)$ for $i\neq j_0$ and $m_{ij}=0$ otherwise.
\item By Lemma~\ref{p-ca-critical}~\ref{it:crit}
$\bp$ is critical and $\alpha^*$ is the left eigenvector with
eigenvalue~1.
\item By Lemma~\ref{p-ca-critical}~\ref{it:period}
$\bp$ is aperiodic.
\item Since $p_\alpha$ satisfies~\eqref{assumption-p} and
$A_0=\emptyset$, we deduce from the definition of $\bp$ that there
exists a type $j\in \cj^*$ such that individual of type $j$ has two
children or more with positive probability, that is, $\bp$ is
non-singular.
\end{enumerate}
In particular, the offspring distribution $\bp$ satisfies
hypothesis~\eqref{eq:hyp-p-d}-\eqref{eq:hyp-p-crit-2} from
Section~\ref{sec:appendix}.
To conclude, we refer to
Section~\ref{sec:appendix} on how to get
(19) in~\cite{adg18} under this set of hypothesis.
\begin{rem}[On related work]
\label{rem:related}
The case $A_0=\emptyset$ and $\Card(A_{j_0})\geq 2$ where
$j_0\in \II{1, J}$ is such that $0\in A_{j_0}$ (compare with
condition~\eqref{eq:delicat}) could be handled
using~\cite[Theorem~5.1]{p16} on multi-type BGW processes. (We also
believe that condition (A5) there, which amounts to say that for each
$j\in \II{1, J}$ there is $k\in A_j$ such that $k+1 \in A_j$, could
certainly be relaxed.) Notice that the moments condition considered
there does not allow to consider directions $\alpha$ such that
$\qc=\theta_{\max}$ (this case might indeed exist). The possible vector
$\ba=(a_1, \ldots, a_J)$ considered in~\cite{p16} (which is associated
with the critical BGW multi-type process) corresponds in our framework
to $a_j =\qc \alpha_j/ p(A_j)$ for $j\in \II{1, J}$ and the direction
$\bar \bv$ which appears in~\cite[Eq.~(5.1)]{p16} corresponds to
$\alpha$. In our approach, we first fix the direction $\alpha$, and
then give sufficient (and almost necessary) conditions for the
existence and uniqueness of the critical parameter $\qc$ and thus how
to choose the parameter $\ba$ given the direction $\alpha$.
\end{rem}
\subsection{On the proof of~\eqref{eq:19-A0=0}}
\label{sec:appendix}
In this section we quickly revisit the proof of (19) in~\cite{adg18},
using slightly different assumptions in order to take into account the
particular case~\eqref{eq:delicat} from Section~\ref{sec:proof-main}. In
this section only, we stick to the notations introduced in~\cite{adg18}.
Let $d\geq 2$ and set $[n]=\II{1, n}$ for $n\in \N^*$. Let
$p=(p^{(i)}, i\in [d])$ with $p^{(i)}=(p^{(i)}(\bk), \, \bk\in \N^d)$ being
probability distributions on $\N^d$. We assume
that:
\begin{equation}
\label{eq:hyp-p-d}
\boxed{p^{(d)}(\zero)=1}.
\end{equation}
For $i\in [d]$, let $X_i=(X_i^{(j)}, \, j\in [d])$ be a random variable
on $\N^d$ with probability distribution $p ^{(i)}$. In particular, we
have that a.s.\ $X_d=\zero$. We consider the generating function
$f=(f^{(i)}, i\in [d])$ of $p$ defined by:
\[
f^{(i)} (s)=\E\left[\prod_{j\in [d]} s_j ^{X_i^{(j)}}\right],
\quad\text{where}\quad
s=(s_j, j\in [d])\in [0, 1]^d.
\]
We consider the mean matrix $M=(m_{ij}; \, i,j\in [d])$ with
$m_{ij}=\E[X_i^{(j)}]$. We assume that:
\begin{equation}
\label{eq:hyp-p-mean}
m_{ij}\in (0, +\infty )
\quad\text{for all}\quad
i\in [d-1], \, j\in [d];
\end{equation}
notice that $m_{dj}=0$ for all
$j\in [d]$. In particular \emph{the matrix $M $ is not primitive}, as
there is no $n\in \N^*$ such that $M^n$ has only positive finite
entries; notice that $M$ primitive is part of assumption (H1)
in~\cite{adg18} (this condition is mainly used to apply Perron-Frobenius
theorem on the existence and uniqueness of a left and a right
eigenvector having nonnegative entries, and their corresponding
eigenvalue is in fact the spectral radius of $M$).
We recall that $p$ is critical if the spectral radius of $M$ is $1$; and
that $p$ is non-singular if $f(s)\neq Ms$. We assume that:
\begin{equation}
\label{eq:hyp-p-crit-1}
p \quad\text{is critical and non-singular};
\end{equation}
notice this is the
other part of assumption (H1) in~\cite{adg18}.
We also assume that:
\begin{equation}
\label{eq:hyp-p-crit-2}
p \quad\text{is aperiodic},
\end{equation}
that is, the
smallest subgroup of $\Z^d$ which contains
$\bigcup_{i\in [d]} \big(\supp(p ^{(i)}) - \supp(p ^{(i)})\big)$ is $\Z^d$
itself; this correspond to hypothesis (H2) in~\cite{adg18}.
\medskip
For $i\in [d]$, let $\mathbf{e}_i$ denote the vector of $\R^d$ with all
its entries equal to $0$ but the $i$-th which is equal to $1$. Using
Perron-Frobenius theorem for the matrix $M$ reduced to the first $d-1$
lines and columns and using
that the $d$-th line of $M$ is zero and the other entries are positive, we
deduce that:
\begin{itemize}[-]
\item the eigenvalue $1$ is simple;
\item there exists two left
eigenvectors with non-negative entries: the vector $\mathbf{e}_d$
with eigenvalue $0$, and a vector $a\in \R^d$ having positive
with eigenvalue 1;
\item there exists a unique right
eigenvector $a^*=(a^*(i), i\in [d])$ with eigenvalue 1, and its entries are
positive but for
the $d$-th wich is zero: $a^*(d)=0$.
\end{itemize}
This result is the reason why we can remove the primitive assumption
of $M$.
\medskip
Then the results on the Dwass formula for BGW multi-type trees from
Section~3.2 in~\cite{adg18} also hold, as the expressions therein are
algebraic in the entries of $p$. (For example Lemma~3.8 holds, but notice
that both terms of the equality therein are zero if $r=d$.) Now formula
(19) in~\cite{adg18} is then a direct consequence of the Dwass formula
and the technical Lemma~3.11 therein. This latter result, proved in
Section~3.4, is also a direct consequence of Lemma 3.12, which
asserts that an intermediate random variable $Y$ on $\Z^{2d-1}$ has an
aperiodic distribution, and of Lemma 4.11, which is a variant of the
strong ratio theorem for the random walk with increments distributed as
$Y$. Now looking carefully at the proof of Lemma 3.12, we see that $p$
is assumed to be aperiodic (this is (H2) therein
and~\eqref{eq:hyp-p-crit-2} here) and that hypothesis (H1) is only used at
the end of the proof to get that $\P(X_d=\zero)>0$; but this is clearly
the case if~\eqref{eq:hyp-p-d} holds. To conclude, notice that Lemma
4.11 on the strong ratio theorem requires only that the law of $Y$ is
aperiodic (which is provided by Lemma 3.12) and that $Y$ is
integrable. By the construction of $Y$ given in Section~3.4, we notice
that $Y$ is integrable if and only if the mean matrix $M$ has finite
entries, which is hypothesis~\eqref{eq:hyp-p-mean}. In conclusion, we
obtain that (19) in~\cite{adg18} holds (notice that the root has to be of
type $r\neq d$ otherwise the numerator and denominator are both zero).
\begin{rem}[On the extension to the main result of~\cite{adg18} under
hypothesis~\eqref{eq:hyp-p-d}-\eqref{eq:hyp-p-crit-2}]
\label{rem:Kesten-ADG}
We leave to the interested reader the construction of the
corresponding Kesten tree, see Section~2.6 in~\cite{adg18}, where
here individuals on the infinite spine can not have type $d$ (in
particular, the root has not type $d$). (For example Lemma~2.9
therein holds provided $i, r$ belong to $[d-1]$.) Then, assuming
hypothesis~\eqref{eq:hyp-p-d}-\eqref{eq:hyp-p-crit-2}, we have the
analogue of Theorem~3.1 therein on the local convergence in
distribution, towards the Kesten tree of the BGW multi-type tree
(with the root not being a.s.\ of type $d$ and with offspring
distribution $p$) conditioned to have population of type $i$ equal to
$k(n)_i$ for $i\in [d]$ with
$\lim_{n\rightarrow \infty } k(n)_i/ |k(n)|=a(i)$, where
$|k(n)|=\sum_{j\in [d]} k(n)_j$ and
$\lim_{n\rightarrow \infty } |k(n)|=\infty $.
\end{rem}
\subsection{Details for Remark~\ref{rem:tbd} on the condition $n_j=0$ if
$\alpha_j=0$}
\label{sec:a=0-n>0}
We consider the following example: a probability distribution $p$ such
that $\supp(p)$ containing but not reduced to $\{0, 2\}$,
$1\not \in \supp(p)$, $J=2$, $\ca=(A_1, A_2)$ is a partition of
$\supp(p)$ (that is, $A_0=\emptyset$) with $A_1=\{0, 2\}$ and
$A_2\subset 3+2\N$. Notice that $A_2\neq \emptyset$. We set:
\begin{equation}
\label{eq:def-a}
a= \frac{\sqrt{p(0)}}{\sqrt{p(2)}}\in (0, +\infty ).
\end{equation}
We consider the
direction $\alpha=(1, 0)$. It is elementary to check that $p$ is
generic for $\ca$ in the direction $\alpha$ and that
$p_\alpha=(p_\alpha(n))_{n\in \N}$ is given by
$p_\alpha(0)=p_\alpha(2)=1/2$. The distribution $p$ is however not
aperiodic for $\ca$ in the direction $\alpha$, but thanks to
Remark~\ref{rem:leaves-intro}, we still have the convergence of $\ct$
conditionally on $L_\ca(\ct)=(n,0)$, with $n$ odd going to infinity,
locally in distribution towards the Kesten's tree $\ct^*_\alpha$.
For $n$ odd going to infinity, we shall check that the distribution of
$\ct$ conditionally on $L_\ca(\ct)=(n,1)$ does not converge locally to
the distribution of $\ct^*_\alpha$, and thus
Condition~\eqref{eq:cond-zero} is required in general to get the local
limit of conditioned BGW tree from Theorem~\ref{thm:main-thm}. To do so,
we shall simply check the positivity of the limit, for $n$ odd going to
infinity, of:
\[
\P\left(k_{\roott} (\ct) \neq 2\, |\,
L_{\ca}(\ct)=(n,1)\right)= \frac{B_1(n)}{B_2(n)},
\]
where
\[
B_1(n)= \P\left(k_{\roott} (\ct) \neq 2\,,
L_{\ca}(\ct)=(n,1)\right)
\quad\text{and}\quad
B_2(n)= \P\left(
L_{\ca}(\ct)=(n,1)\right).
\]
Before going further, we recall that the number of planar binary trees
with $n$ leaves is:
\[
f_{1, n}=\frac{1}{n}\, \binom{2n - 2}{n-1},
\]
(in particular $f_{1, n+1}$ is the so called $n$-th Catalan's number) and that
$f_{1,n}=[z^n] zC=[z^{n-1}] C$, where we write simply $C$ for
$C(z)=(1- \sqrt{1-4z})/2z$. Recall also that $zC^2-C+1=0$. We deduce
that the number of planar forests with $k$ binary trees and $n\geq k$
leaves is given by $f_{k,n}=[z^n] z^kC^k=[z^{n-k}] C^k$, and that
according to~\cite[(B.5)]{deutsch}:
\begin{equation}
\label{eq:def-cat-forest}
f_{k, n}=\frac{k}{n}\, \binom{2n - k -1}{n-1}= \binom{2n -k
-1}{n-1} - \binom{2n -k
-1}{n}
\quad\text{for}\quad
n\geq k\geq 1.
\end{equation}
We set $f_{k,n}=0$ if $k>n$.
\medskip
Let $n\geq k$ be odd integers.
On the event $k_\roott(\ct)=k$ and $L_{\ca}(\ct)=(n,1)$,
we get that $\ct$ can be seen as a forest of $k$ trees grafted on the
root and with the forest having $(n+k)/2$ leaves and $(n-k)/2$
internal nodes, all of them binary.
We deduce that:
\[
\P\left(k_{\roott} (\ct)=k\,,
L_{\ca}(\ct)=(n,1)\right)= p(k) p(0)^{(n+k)/2}\, p(2)^{(n-k)/2}
f_{k,(n+k)/2},
\]
and thus:
\[
B_1(n)= (p(0) p(2))^{n/2} \sum_{k\in A_2, \, k\leq n} f_{k, (n+k)/2} \,
p(k)\, a^k.
\]
A tree $\bt$ such that $ L_{\ca}(\bt)=(n,1)$ can be decomposed as a
binary tree with $\ell$ leaves, and on one of those $\ell$ leaves one grafts a
forest with $k\in A_2$ (and $k\leq n$) binary trees with
$(n+k)/2+1 - \ell$ leaves; and in total the tree $\bt$ has $(n+k)/2$
leaves, $(n-k)/2$ binary branching nodes and one node with out-degree
$k$. Thus, we have:
\begin{align}
\label{eq:B2=}
B_2(n)
&= \sum_{k\in A_2, \, k\leq n} p(k)p(0)^{(n+k)/2}
p(2)^{(n-k)/2}\,
\sum_{\ell=1}^{(n-k)/2+1} \ell f_{1,\ell}\, f_{k, (n+k)/2 + 1 - \ell}
\\
\nonumber
&= (p(0) p(2))^{n/2} \sum_{k\in A_2, \, k\leq n} F_{k, (n+k)/2}\, p(k)\, a^k,
\end{align}
where for $n\geq k\geq 1$:
\[
F_{k, n}= \sum_{\ell=1}^{n-k+1} \ell f_{1,\ell}\, f_{k,
n + 1 - \ell}
.
\]
We give an explicit formula of $F_{k, n}$.
\begin{lem}
\label{lem:catalan}
We have:
\[
F_{k,n}= \binom{2n-k}{n}= \frac{2n-k}{k}\, f_{k,n}
\quad\text{for}\quad
n\geq k\geq 1.
\]
\end{lem}
\begin{proof}
We have:
\[ F_{k,n}
=[z^{n}] (zC)' z^{k} C^k
= [z^{n}] \inv{k+1} \left(z^{k+1} C^{k+1}\right)'
= [z^{n+1}] \frac{n+1}{k+1} z^{k+1} C^{k+1}
= \frac{n+1}{k+1} f_{k+1,n+1}.
\]
Then, use~\eqref{eq:def-cat-forest} to conclude.
\end{proof}
We shall now consider that $A_2$ is unbounded. Let $\varepsilon\in (0,
1)$ and write:
\begin{align*}
B_1(n,\varepsilon)
&= \sum_{k\in A_2, \,
k> \varepsilon n} f_{k, (n+k)/2} \, p(k)\, a^k,\\
B_3(n, \varepsilon)
&= \sum_{k\in A_2, \, k\leq \varepsilon n} F_{k,(n+k)/2}\, p(k)\,
a^k,\\
B_4(n, \varepsilon)
&= \sum_{k\in A_2, \, k> \varepsilon n} F_{k,
(n+k)/2}\, p(k)\, a^k,
\end{align*}
so that using the two latter terms, we can rewrite $B_2(n)$ as:
\begin{equation}
\label{eq:B2=B3-4}
B_2(n)=(p(0) p(2))^{n/2}\,\left( B_3(n, \varepsilon) + B_4(n, \varepsilon)\right).
\end{equation}
For $k > \varepsilon n$, we have that:
\[
F_{k, (n+k)/2} = \frac{n}{k} \, f_{k, (n+k)/2}\leq
\inv{\varepsilon}\, f_{k, (n+k)/2} .
\]
This implies that:
\begin{equation}
\label{eq:maj-B4}
B_4(n,\varepsilon)\leq \inv{\varepsilon} B_1(n,\varepsilon).
\end{equation}
We now assume that $A_2=3+2\N$ and there exists $b\in (0, 1)$ and $M\geq 1$ finite such that $M^{-1}\leq p(k) b^{-k} \leq M$ for $k\in A_2$.
Then, we have with $2m=n+3$ and $k=3+2\ell$:
\begin{align*}
B_3(n,\varepsilon)
& \leq M \sum_{k\in 3+2\N, k\leq \varepsilon n} F_{k,
(n+k)/2}\, (ab)^k\\
& \leq M(ab)^3 \sum_{\ell\in \N, \ell\leq \varepsilon m } F_{3+2\ell,
m+\ell}\, (ab)^{2 \ell}\\
& = M(ab)^{-2m+3}\left( 1+(ab)^2\right)^{2m-3} \sum_{\ell\in \N,
\ell\leq \varepsilon m } \binom{2m-
3}{m+\ell} \, r^{m+\ell} (1-r) ^{m-3-\ell}
\end{align*}
with $r/(1-r)=(ab)^2$ and thus $r=(ab)^2/\left( 1+(ab)^2\right)$.
As $r<1$, we deduce that:
\[
\sum_{\ell\in \N, \ell\leq \varepsilon m } \binom{2m- 3}{m+\ell} \,
r^{m+\ell} (1-r) ^{m-3-\ell}= \P(m\leq X \leq (1+\varepsilon) m)\leq
\P(X \leq (1+\varepsilon) m),
\]
where $X$ is binomial with parameter $(2m-3, r)$.
We now assume that $ab>1$ and that $\varepsilon$ is small enough so that
$ab>(1+\varepsilon)/(1-\varepsilon)$. This yields $2r>1+\varepsilon$, so
that for $m$ large enough, we have $(1+\varepsilon)m\leq r (2m -3)$.
We deduce from~\cite[Theorem~2.1]{zzh22} that, with $j=\lfloor (1+\varepsilon)
m\rfloor $ and $x=r(2m-3) -j+1$:
\[
\frac{ \P(X \leq (1+\varepsilon) m)}{ \P(X =j )} \leq 2-x +\sqrt{x^2
+ 4(1-r) j}.
\]
Since $\lim_{m\rightarrow \infty } x=+\infty $ and
$\lim_{m\rightarrow \infty } (x+j)/j=2r/(1+\varepsilon)$, we deduce
that:
\[
\lim_{m\rightarrow \infty } \frac{ \P(X \leq (1+\varepsilon) m)}{
\P(X =j )} =c_0
\quad\text{with}\quad
c_0=\frac{r(1-\varepsilon)}{2r -(1+\varepsilon)}\cdot
\]
Recall that $n=2m-3$. So for $n$ large enough, we have with $k'=3+2\ell'$ and $\ell'=j-m=
\lfloor\varepsilon m\rfloor$, and thus $n+k'=2j$, that:
\begin{align*}
B_3(n,\varepsilon)
&\leq 2c_0 M(ab)^{-2m+3}\left( 1+(ab)^2\right)^{2m-3}
\binom{2m- 3}{j} \, r^{j} (1-r) ^{2m-3-j}\\
% &= 2c_0\, (ab)^{2\ell'} F_{3+2\ell', m+ \ell'}\\
&= 2 c_0M\, F_{k', (n+k')/2}\, (ab)^{k'}.
\end{align*}
Notice that $j\geq (1+\varepsilon)m -1$ and thus $k'\geq \varepsilon n
+1$, so that:
\begin{equation}
\label{eq:majo-B3}
B_3(n,\varepsilon)
\leq 2 c_0 M^2\, B_4(n,\varepsilon).
\end{equation}
Hence, using~\eqref{eq:maj-B4}, we obtain that:
\[
B_3(n,\varepsilon)+ B_4(n,\varepsilon)\leq \frac{1+2c_0M^2}{\varepsilon} B_1(n,\varepsilon).
\]
From the definition of $B_1( n,\varepsilon)$, we get that:
\[
(p(0) p(2))^{n/2}\, B_1(n,\varepsilon)= \P\left(k_{\roott} (\ct) \geq \varepsilon n\,,
L_{\ca}(\ct)=(n,1)\right).
\]
We deduce from~\eqref{eq:B2=B3-4} that:
\[
\liminf_{n\rightarrow \infty }\P\left(k_{\roott} (\ct) \geq
\varepsilon n\, |\,
L_{\ca}(\ct)=(n,1)\right)
= \liminf_{n\rightarrow \infty } \frac{B_1(n,\varepsilon)}{B_2(n)} \geq
\frac{\varepsilon}{1+ 2c_0M^2}>0,
\]
provided that there exists $b\in (0, 1)$ and $M\geq 1$ finite such that
$M^{-1}\leq p(k) b^{-k} \leq M$ for $k\in A_2= 3+2\N$ and
$\varepsilon>0$ is chosen so that $ab>(1+\varepsilon)/(1-\varepsilon)$,
with $a$ defined by~\eqref{eq:def-a}.
\medskip
In conclusion, under the above hypothesis, the distribution of
$\ct$ conditionally on $\{L_\ca(\ct)=(n,1)\}$ does not converge locally to
the
distribution of $\ct^*_\alpha$ as $n$ goes to infinity, whereas
the distribution of
$\ct$ conditionally on $\{L_\ca(\ct)=(n,0)\}$ converges locally to
the
distribution of $\ct^*_\alpha$. Furthermore, conditioning on
$\{L_\ca(\ct)=(n,1)\}$ and letting $n$ goes to infinity gives a condensation at the root with positive
probability.
\bibliographystyle{abbrv}
\bibliography{cgwsub}
\end{document}
\begin{thebibliography}{10}
\bibitem{abd17}
R.~Abraham, A.~Bouaziz, and J.-F. Delmas.
\newblock Local limits of {G}alton-{W}atson trees conditioned on the number of
protected nodes.
\newblock {\em Journal of Applied Probability}, 54(1):55--65, 2017.
\bibitem{abd20}
R.~Abraham, A.~Bouaziz, and J.-F. Delmas.
\newblock Very fat geometric {G}alton-{W}atson trees.
\newblock {\em ESAIM: Probability and Statistics}, 24:294--314, 2020.
\bibitem{ad14b}
R.~Abraham and J.-F. Delmas.
\newblock Local limits of conditioned {G}alton-{W}atson trees: the condensation
case.
\newblock {\em Electron. J. Probab}, 19:1--29, 2014.
\bibitem{ad14a}
R.~Abraham and J.-F. Delmas.
\newblock Local limits of conditioned {G}alton-{W}atson trees: the infinite
spine case.
\newblock {\em Electron. J. Probab}, 19:1--19, 2014.
\bibitem{ad19}
R.~Abraham and J.-F. Delmas.
\newblock Asymptotic properties of expansive {G}alton-{W}atson trees.
\newblock {\em Electronic Journal of Probability}, 24:1--51, 2019.
\bibitem{adg18}
R.~Abraham, J.-F. Delmas, and H.~Guo.
\newblock Critical multi-type {G}alton-{W}atson trees conditioned to be large.
\newblock {\em Journal of Theoretical Probability}, 31:757--788, 2018.
\bibitem{a91}
D.~Aldous.
\newblock {The Continuum Random Tree. I}.
\newblock {\em The Annals of Probability}, 19(1):1 -- 28, 1991.
\bibitem{ck14}
N.~Curien and I.~Kortchemski.
\newblock Random non-crossing plane configurations: A conditioned
{G}alton-{W}atson tree approach.
\newblock {\em Random Structures \& Algorithms}, 45(2):236--260, 2014.
\bibitem{deutsch}
E.~Deutsch.
\newblock Dyck path enumeration.
\newblock {\em Discrete Math.}, 204(1-3):167--202, 1999.
\bibitem{dlg02}
T.~Duquesne and J.-F. Le~Gall.
\newblock Random trees, {L}\'{e}vy processes and spatial branching processes.
\newblock {\em Ast\'{e}risque}, (281):vi+147, 2002.
\bibitem{du23}
B.~Durhuus and M.~\"{U}nel.
\newblock Trees with exponential height dependent weight.
\newblock {\em Probab. Theory Related Fields}, 186(3-4):999--1043, 2023.
\bibitem{gk04}
J.~Geiger and L.~Kauffmann.
\newblock The shape of large {G}alton-{W}atson trees with possibly infinite
variance.
\newblock {\em Random Structures \& Algorithms}, 25(3):311--335, 2004.
\bibitem{he17}
X.~He.
\newblock Conditioning {G}alton-{W}atson trees on large maximal outdegree.
\newblock {\em Journal of Theoretical Probability}, 30:842--851, 2017.
\bibitem{he22}
X.~He.
\newblock Local convergence of critical random trees and continuous-state
branching processes.
\newblock {\em Journal of Theoretical Probability}, 35(2):685--713, 2022.
\bibitem{j12}
S.~Janson.
\newblock Simply generated trees, conditioned {G}alton-{W}atson trees, random
allocations and condensation.
\newblock {\em Probability Surveys}, 9:103--252, 2012.
\bibitem{janson16}
S.~Janson.
\newblock Asymptotic normality of fringe subtrees and additive functionals in
conditioned {G}alton-{W}atson trees.
\newblock {\em Random Structures Algorithms}, 48(1):57--101, 2016.
\bibitem{js11}
T.~Jonsson and S.~{\"O}. Stef{\'a}nsson.
\newblock Condensation in nongeneric trees.
\newblock {\em Journal of Statistical Physics}, 142:277--313, 2011.
\bibitem{k23}
V.~Kargin.
\newblock Scaling limits of slim and fat trees.
\newblock {\em J. Theoret. Probab.}, 36(4):2192--2228, 2023.
\bibitem{k75}
D.~P. Kennedy.
\newblock The {G}alton-{W}atson process conditioned on the total progeny.
\newblock {\em Journal of Applied Probability}, 12(4):800--806, 1975.
\bibitem{k86}
H.~Kesten.
\newblock Subdiffusive behavior of random walk on a random cluster.
\newblock In {\em Annales de l'IHP Probabilit{\'e}s et statistiques},
volume~22, pages 425--487, 1986.
\bibitem{km23}
I.~Kortchemski and C.~Marzouk.
\newblock Large deviation local limit theorems and limits of biconditioned
planar maps.
\newblock {\em Ann. Appl. Probab.}, 33(5):3755--3802, 2023.
\bibitem{marzouk22}
C.~Marzouk.
\newblock On scaling limits of random trees and maps with a prescribed degree
sequence.
\newblock {\em Ann. H. Lebesgue}, 5:317--386, 2022.
\bibitem{n86}
J.~Neveu.
\newblock Arbres et processus de {G}alton-{W}atson.
\newblock In {\em Annales de l'IHP Probabilit{\'e}s et statistiques},
volume~22, pages 199--207, 1986.
\bibitem{p16}
S.~P{\'e}nisson.
\newblock Beyond the {Q}-process: various ways of conditioning the multitype
{G}alton-{W}atson process.
\newblock {\em ALEA}, 13(1):223--237, 2016.
\bibitem{r15}
D.~Rizzolo.
\newblock Scaling limits of {M}arkov branching trees and {G}alton-{W}atson
trees conditioned on the number of vertices with out-degree in a given set.
\newblock {\em Annales de l'IHP Probabilit{\'e}s et statistiques},
51(2):512--532, 2015.
\bibitem{s18}
R.~Stephenson.
\newblock Local convergence of large critical multi-type {G}alton-{W}atson
trees and applications to random maps.
\newblock {\em Journal of Theoretical Probability}, 31:159--205, 2018.
\bibitem{s19}
B.~Stufler.
\newblock Local limits of large {G}alton-{W}atson trees rerooted at a random
vertex.
\newblock {\em Annales de l'Institut Henri Poincaré, Probabilités et
Statistiques}, 55(1):155 -- 183, 2019.
\bibitem{thevenin20}
P.~Th{\'e}venin.
\newblock Vertices with fixed outdegrees in large {G}alton-{W}atson trees.
\newblock {\em Electron. J. Probab}, 25:1--25, 2020.
\bibitem{thevenin23}
P.~Th\'{e}venin.
\newblock Critical exponential tiltings for size-conditioned multitype
bienaym\'e--galton--watson trees, 2023.
\bibitem{zzh22}
H.~Zhu, Z.~Li, and M.~Hayashi.
\newblock Nearly tight universal bounds for the binomial tail probabilities,
2022.
\end{thebibliography}