Bayesian Nonparametric Estimation of a Unimodal Density via two $\mathbf{S}$-paths

Abstract

A Bayesian nonparametric method for unimodal densities on the real line is provided by considering a class of species sampling mixture models containing random densities that are unimodal and not necessarily symmetric. This class of densities generalize the model considered by Brunner (1992, Statist. Probab. Lett.), in which the Dirichlet process is replaced by a more general class of species sampling models. A novel and explicit characterization of the posterior distribution via a finite mixture of two dependent $\mathbf{S}$-paths is derived. This results in a closed-form and tractable Bayes estimator for any unimodal density in terms of a finite sum over two $\mathbf{S}$-paths. To approximate this class of estimates, we propose a sequential importance sampling algorithm that exploits the idea of the accelerated path sampler, an efficient path-sampling Markov chain Monte Carlo method. Numerical simulations are given to demonstrate the practicality and the effectiveness of our methodology.