More on the asymptotic behaviour of moments of branching Markov processes

Christopher B. C. Dean, Emma Horton * · 2025

Read More on the asymptotic behaviour of moments of branching Markov processes by Christopher B. C. Dean, Emma Horton * (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-19268.txt.

Introduction

Christopher Dean and Emma Horton extend results on the long run behavior of averages in branching Markov processes. Such processes model populations in which each particle moves by a random rule and spawns offspring. Their averages matter because they reveal the family structure hidden in the population over time.

Earlier work treated the case where the leading growth rate is simple. Dean and Horton allow that rate to be non-simple and study higher order wobbles around the average as well. The generalizations set up proofs of central limit theorems and broader laws of large numbers. The paper is a technical step for probability theory, aimed at specialists.

Catalog record

CREATORChristopher B. C. Dean, Emma Horton *
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-19268

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

More on the asymptotic behaviour of moments of branching Markov processes Christopher B. C. Dean, Emma Horton * Abstract Consider a branching Markov process, X = ( X t , t ≥ 0 ) X=(X_{t},t\geq 0) , with non-local branching mechanism. Studying the asymptotic behaviour of the moments of X X has recently received attention in the literature [ 6 , 7 ] due to the importance of these results in understanding the underlying genealogical structure of X X . In this article, we generalise the results of [ 7 ] to allow for a non-simple leading eigenvalue and to also study the higher order fluctuations of the moments of X X . These results will be useful for proving central limit theorems and extending well-known LLN results. Keywords : Branching process, fluctuations, moments, non-local branching. MSC: 60J80, 60F05. 1 Introduction Let E be a Polish space and \dagger\notin E a cemetery (absorbing) state. We consider a measure-valued stochastic process X=(X_{t},t\geq 0) given by X_{t}:=\sum_{i=1}^{N_{t}}\delta_{x_{i}(t)},\qquad t\geq 0, whose atoms \{x_{i}(t):i=1,\dots,N_{t}\} evolve in E\cup\{\dagger\} according to the following dynamics. From an initial position x\in E , particles evolve independently in E according to a Markov process (\xi,\mathbf{P}_{x}) . When at y\in E , particles branch at rate \gamma(y) , which we assume to be bounded, at which point the particle is replaced by a random number, N , of offspring at positions x_{1},\dots,x_{N} in E . The law of the offspring locations and their number is denoted by \mathcal{P}_{y} , where y denotes the position of the parent particle at the branch time. We will also often use the notation \mathcal{E}_{y} for the associated expectation operator. We refer to X as a (\mathbf{P},\gamma,\mathcal{P}) -branching Markov process (BMP). We will let \mathbb{P}_{\mu} denote its law when initiated from \mu\in\mathcal{M}(E) , the space of finite measures on E . We also let B(E) denote the space of bounded complex-valued functions, B^{+}(E) denote the space of non-negative real bounded functions and B_{1}(E) , (resp. B_{1}^{+}(E) ) denote those functions in B(E) , (resp. B^{+}(E) ) that are bounded by unity. One way to characterise the behaviour of X is via its linear expectation semigroup, \psi_{t}[f](x):=\mathbb{E}_{\delta_{x}}\left[X_{t}[f]\right],\quad t\geq 0,\,x\in E,\,f\in B(E), (1.1) where \mathbb{E}_{\delta_{x}} denotes the expectation conditional on X_{0} consisting of a single particle at x , and, for a measure \mu and a function f , we have used the notation \mu[f]:=\int_{E}f(x)\mu({\rm d}x). In many cases, it is well known that the semigroup \psi satisfies the following Perron Frobenius-type decomposition. (PF) There exists \lambda_{1}\in\mathbb{R} , \varphi_{1}\in B^{+}(E) and a probability measure \tilde{\varphi}_{1} on E such that \psi_{t}[\varphi_{1}]={\rm e}^{\lambda_{1}t}\varphi_{1},\qquad\tilde{\varphi}_{1}[\psi_{t}[f]]={\rm e}^{\lambda_{1}t}\tilde{\varphi}_{1}[f], and \sup_{x\in E,f\in B_{1}(E)}\left|{\rm e}^{-\lambda_{1}t}\psi_{t}[f](x)-\tilde{\varphi}_{1}[f]\varphi_{1}(x)\right|\to 0,\qquad\text{as }t\to\infty. When this decomposition holds, along with moment conditions on the offspring distribution, it has been shown in [ 7 ] that similar asymptotics hold for the moments \mathbb{E}_{\delta_{x}}[X_{t}[f]^{k}] . We also mention the previous works [ 3 , 5 , 11 , 12 ] , where asymptotics for moments of branching processes were studied under varying assumptions. Results of this kind are particularly useful in proving laws of large numbers [ 1 , 4 , 9 ] , for setting benchmarks for Monte Carlo codes [ 2 ] and for understanding the genealogical structure of branching Markov processes [ 6 , 8 , 10 ] . However, these results do not give a complete description of the asymptotic behaviour of X . Indeed, they do not allow us to say anything about the fluctuations of the process via a CLT, for example. Moreover, there are many cases where (PF) is not satisfied, for example when the first eigenvalue is not simple. The aim of this article is to study the asymptotic moments of X in more depth by considering what happens when we have a more general version of (PF) . These results will play a key role in proving CLTs for X and will also allow us to provide new proofs for a LLN result. 1.1 Assumptions Before stating our main results, we first spend some time discussing the assumptions. Our main assumption generalises (PF) to the case when \lambda_{1} is not a simple eigenvalue and where we may have more points in the discrete spectrum of \psi . Here and throughout, we will use the notation {\rm ker}(\lambda_{\ell}):=\{f\in B(E):\tilde{\varphi}_{i,j}^{(k)}[f]=0,1\leq i\leq\ell,1\leq j\leq p_{i},1\leq k\leq k_{i,j}\} for each 1\leq\ell\leq m , where the \tilde{\varphi}_{i,j}^{(k)} are defined in the following assumption. (H1) There exists an integer m\geq 1 , a real number \lambda_{1}>0 , \lambda_{2},\dots,\lambda_{m}\in\mathbb{C} , that satisfy 1 1 1 If m=1 , the list \lambda_{2}\dots,\lambda_{m} is an empty list. \lambda_{1}>\mathrm{Re}\lambda_{2}\geq\dots\geq\mathrm{Re}\lambda_{m} , such that, for 1\leq i\leq m , there exist bounded functions \varphi_{i,1}^{(1)},\dots,\varphi_{i,1}^{(k_{i,1})},\dots,\varphi_{i,p_{i}}^{(1)},\dots,\varphi_{i,p_{i}}^{(k_{i,p_{i}})}\in B(E), bounded linear functionals \tilde{\varphi}_{i,1}^{(1)},\dots,\tilde{\varphi}_{i,1}^{(k_{i,1})},\dots,\tilde{\varphi}_{i,p_{i}}^{(1)},\dots,\tilde{\varphi}_{i,p_{i}}^{(k_{i,p_{i}})}:B(E)\rightarrow\mathbb{C}, such that \tilde{\varphi}_{i,j}^{(k)}[\varphi_{i,j}^{(k)}]=1 , \tilde{\varphi}_{i,j}^{(k)}[\varphi_{\ell,n}^{(q)}]=0 for each (i,j,k)\neq(\ell,n,q) . There also exists and a linear operator \mathcal{N} , such that • for each 1\leq i\leq m,\,1\leq k\leq k_{i,1} , \mathcal{N}\varphi_{i,1}^{(k)}=0 ; • for each 1\leq i\leq m,\,2\leq j\leq p_{i},\,1\leq k\leq k_{i,j} there exists 2 2 2 If k_{1}\neq k_{2} then k^{*}(i,j,k_{1})\neq k^{*}(i,j,k_{2}) k^{*}\in\{1,\dots,k_{i,j-1}\} such that \mathcal{N}\varphi_{i,j}^{(k)}=\varphi_{i,j-1}^{(k^{*})} ; • for f\in\mathrm{ker}(\lambda_{m}) , \mathcal{N}f=0 , and, for any x\in E , t\geq 0 and f\in B(E) , \psi_{t}[\varphi_{i,j}^{(k)}](x)={\rm e}^{(\lambda_{i}+\mathcal{N})t}\varphi_{i,j}^{(k)}(x),\quad\text{and}\quad\tilde{\varphi}_{i,j}^{(k)}[\psi_{t}[f]]={\rm e}^{\lambda_{i}t}\tilde{\varphi}_{i,j}^{(k)}[{\rm e}^{\mathcal{N}t}f]. (1.2) Moreover, \sup_{x\in E,f\in B_{1}(E)}\mathrm{e}^{-\mathrm{Re}\lambda_{m}t}\bigg|\psi_{t}[f](x)-\sum_{j=1}^{p_{i}}\mathrm{e}^{(\lambda_{i}+\mathcal{N})t}\Phi_{i}[f](x)\bigg|\to 0,\quad\text{as }t\to\infty, (1.3) where \Phi_{i,j}[f](x)=\sum_{k=1}^{k_{i,j}}\tilde{\varphi}_{i,j}^{(k)}[f]\varphi_{i,j}^{(k)}(x),\quad\Phi_{i}[f](x)=\sum_{j=1}^{p_{i}}\Phi_{i,j}[f](x). Let us spend some time unpacking the above assumption. First note that, as mentioned above, this allows for the case that the leading eigenvalue associated with \psi is not simple. For simplicity, assume that m=1 in (H1) . In the case that p_{1}=k_{1,1}=1 , we recover (PF) . In the case that p_{1}=1 but k_{1,1}>1 , we extend (PF) to the case where \lambda_{1} is not simple. This case implies that k_{1,1} is both the algebraic and geometric multiplicity of \lambda_{1} , with corresponding eigenfunctions given by \varphi^{(1)},\dots,\varphi^{(k_{1,1})} (note, we have dropped the subscript for simplicity). The fact that these are eigenfunctions can be deduced from ( 1.2 ) and the first property of \mathcal{N} . The case that p_{1}=2 , say, the functions \varphi_{1,1}^{(1)},\dots,\varphi_{1,1}^{(k_{1,1})} are eigenfunctions, but the functions \varphi_{1,2}^{(1)},\dots,\varphi_{1,2}^{(k_{1,2})} are generalised eigenfunctions of rank 2 , meaning \mathcal{N}\varphi_{1,2}^{(k)}\neq 0 but \mathcal{N}^{2}\varphi_{1,2}^{(k)}=0 , which can be deduced from the first two properties of \mathcal{N} . In general, if p_{1}\geq 3 , then \varphi_{1,j}^{(k)} is of rank j . Note that, in the case that m=1 and p_{1}\geq 2 , the geometric multiplicity of \lambda_{1} is k_{1,1} and the algebraic multiplicity is \sum_{j=1}^{p_{1}}k_{1,j} . Allowing m\geq 2 generalises this to m eigenvalues \mathrm{e}^{\lambda_{1}t},\dots,\mathrm{e}^{\lambda_{m}t} , each with eigenfunctions \varphi_{i,1}^{(1)},\dots,\varphi_{i,1}^{(k_{i,1})} and generalised eigenfunctions \varphi_{i,j}^{(1)},\dots,\varphi_{i,j}^{(k_{i,j})} of rank j . The asymptotic given in ( 1.3 ) stipulates that, when normalised by the smallest growth rate {\rm e}^{\mathrm{Re}\lambda_{m}t} , the expectation semigroup converges. 1.2 Notation We are interested in seeing how (H1) propagates to higher moments of X_{t} . We will shortly see that it will obey one of three possible dynamics depending on the growth rate of \psi_{t}[f] . However, we first introduce some notation that will be used throughout the rest of the article. For k\geq 1 , A\subseteq\mathbb{[}k]:=\{1,\dots,k\} , and \boldsymbol{f}\in B(E)^{k} , we write \psi_{t}^{(A)}[\boldsymbol{f}](x)=\mathbb{E}_{\delta_{x}}\left[\prod_{i\in A}X_{t}[f_{i}]\right],\quad t\geq 0. If A=[k] , we will write \psi_{t}^{(k)} instead. Similarly to [ 7 ] , we will use an inductive argument to prove

Read the complete text