Measure-valued P\'olya processes

TL;DR

This paper introduces a measure-valued generalization of Polya urn processes, allowing for infinite color spaces and analyzing their asymptotic behavior, including convergence results related to branching random walks and recursive trees.

Contribution

It extends Polya urn models to measure-valued processes on Polish spaces, providing new convergence conditions and almost sure results for these generalized processes.

Findings

Convergence in distribution of measure-valued urns under certain conditions

Almost sure convergence in models related to branching random walks

Gaussian limit for the profile of random recursive trees

Abstract

A P\'olya urn process is a Markov chain that models the evolution of an urn containing some coloured balls, the set of possible colours being $\{1,\ldots,d\}$ for $d\in \mathbb{N}$. At each time step, a random ball is chosen uniformly in the urn. It is replaced in the urn and, if its colour is $c$, $R_{c,j}$ balls of colour $j$ are also added (for all $1\leq j\leq d$). We introduce a model of measure-valued processes that generalises this construction. This generalisation includes the case when the space of colours is a (possibly infinite) Polish space $\mathcal P$. We see the urn composition at any time step $n$ as a measure ${\mathcal M}_n$ -- possibly non atomic -- on $\mathcal P$. In this generalisation, we choose a random colour $c$ according to the probability distribution proportional to ${\mathcal M}_n$, and add a measure ${\mathcal R}_c$ in the urn, where the quantity ${\mathcal R}_c(B)$ of a Borel set $B$ models the added weight of "balls" with colour in $B$. We study the asymptotic behaviour of these measure-valued P\'olya urn processes, and give some conditions on the replacements measures $({\mathcal R}_c, c\in \mathcal P)$ for the sequence of measures $({\mathcal M}_n, n\geq 0)$ to converge in distribution, possibly after rescaling. For certain models, related to branching random walks, $({\mathcal M}_n, n\geq 0)$ is shown to converge almost surely under some moment hypothesis; a particular case of this last result gives the almost sure convergence of the (renormalised) profile of the random recursive tree to a standard Gaussian.