Stick-breaking Pitman-Yor processes given the species sampling size
share
Random discrete distributions, say F, known as species sampling models, represent a rich class of models for classification and clustering, in Bayesian statistics and machine learning. They also arise in various areas of probability and its applications. Jim Pitman, within the species sampling context, shows that mixed Poisson processes may be interpreted as the sample size up till a given time or in terms of waiting times of appearance of individuals to be classified. He notes connections to some recent work in the Bayesian statistic/machine learning literature, with some more classical results. We let F:=F_α,θ, be a Pitman-Yor process for α∈ (0,1), and θ>-α, with α-diversity equivalent in distribution to S^-α_α,θ, and let (N_S_α,θ(λ),λ> 0) denote a mixed Poisson process with rate S_α,θ. In this paper we derive explicit stick-breaking representations of F_α,θ given N_S_α,θ(λ)=m. More precisely, if (P_ℓ)∼PD(α,θ), denotes a ranked sequence following the two parameter Poisson-Dirichlet distribution, we obtain explicit representations of the sized biased permutation of (P_ℓ)|N_S_α,θ(λ)=m. Due to distributional results we shall develop in a more general context, it suffices to consider the stable case F_α,0|N_S_α(λ)=m. Notably, it follows that F_α,0|N_S_α(λ)=0, is equivalent in distribution to the popular normalized generalized gamma process. Hence, we obtain explicit stick-breaking representations for the generalized gamma class as a special case.
READ FULL TEXT
