Stick-breaking processes with exchangeable length variables

TL;DR

This paper explores a broad class of Bayesian non-parametric models called stick-breaking processes with exchangeable length variables, providing conditions for their validity, support, and practical inference methods.

Contribution

It introduces a general framework for these processes, establishes conditions for properness and support, and offers a tunable parameter to control weight ordering, along with an MCMC inference algorithm.

Findings

Conditions for properness and full support of the processes

A tunable parameter modulates stochastic ordering of weights

An MCMC algorithm for density estimation is proposed

Abstract

Our object of study is the general class of stick-breaking processes with exchangeable length variables. These generalize well-known Bayesian non-parametric priors in an unexplored direction. We give conditions to assure the respective species sampling process is proper and the corresponding prior has full support. For a rich sub-class we explain how, by tuning a single $[0,1]$-valued parameter, the stochastic ordering of the weights can be modulated, and Dirichlet and Geometric priors can be recovered. A general formula for the distribution of the latent allocation variables is derived and an MCMC algorithm is proposed for density estimation purposes.