Pareto genealogies arising from a Poisson branching evolution model with selection (Q2441552): Difference between revisions

Let \(X_1,\dots, X_n\) be i.i.d.\,random variables with \(\operatorname{P}\{X_1>x\}=x^{-\alpha}\) for \(x\geq 1\) and \(\alpha>0\). Also, for \(i\leq n\), let \(U_1,\dots, U_i\) be i.i.d.\,random variables with the uniform \([0,1]\) distribution which are independent of \((X_k)\). Set \(\Sigma_n:=\sum_{i=1}^n X_i\). The quantities \(X_1/\Sigma_n\),\dots, \(X_n/\Sigma_n\) are the lengths of subintervals that constitute a random partition of \([0,1]\). For each \(n\in\mathbb{N}\), denote by \((x_k^{(n)})_{k\in\mathbb{N}_0}\) a nonincreasing Markov chain which takes values in \(\{1,2,\dots,n\}\). Its one-step transition probability \(p_{i,j}^{(n)}\) equals a size-biased by \(\Sigma_n^\beta\), \(\beta\in [0,\alpha)\) probability of the event that the sample \(U_1,\dots, U_i\) hits \(j\) distinct subintervals of \([0,1]\). By proving the convergence of one-step transition probabilities the author obtains the following asymptotic results (also, the author notices that these can be anticipated from the weak convergence of \(\Sigma_n\), properly normalized and centered, to an \(\alpha\)-stable law). If \(\alpha\in (0,1)\), then \(\lim_{n\to\infty}\,x_k^{(n)}\) in distribution exists and is a discrete-time Poisson-Dirichlet \((\alpha,-\beta)\) coalescent with simultaneous multiple collisions. If \(\alpha=1\), then \(p_{2,1}^{(n)}\sim 1/\log n\), and \(\lim_{n\to\infty}\,x_{[t/p_{2,1}^{(n)}]}^{(n)}\) in distribution exists and is a continuous-time beta \((1,1-\beta)\) coalescent with multiple collisions. If \(\alpha\in (1,2)\), then \(p_{2,1}^{(n)}\sim \text{const}\,n^{1-\alpha}\), and \(\lim_{n\to\infty}\,x_{[t/p_{2,1}^{(n)}]}^{(n)}\) in distribution exists and is a continuous-time beta \((2-\alpha,\alpha-\beta)\) coalescent with multiple collisions. If \(\alpha=2\), then \(p_{2,1}^{(n)}\sim \log n/(2n)\), and \(\lim_{n\to\infty}\,x_{[t/p_{2,1}^{(n)}]}^{(n)}\) in distribution exists and is the Kingman coalescent. If \(\alpha>2\), then \(p_{2,1}^{(n)}\sim \text{const}\,n^{-1}\), and \(\lim_{n\to\infty}\,x_{[t/p_{2,1}^{(n)}]}^{(n)}\) in distribution exists and is the Kingman coalescent. As the author writes in the summary, ``this class of coalescent processes\dots may be viewed as the genealogical processes of some forward in time evolving branching population models including selection effects. In such constant-size population models, the reproduction step, which is based on a fitness-dependent Poisson point process with scaling power-law \(\alpha\) intensity, is coupled to a selection step consisting of sorting out the \(n\) fittest individuals issued from the reproduction step''.