Perfect Sampling of the Posterior in the Hierarchical Pitman-Yor Process

Area of Science:

• Statistics
• Computational Statistics
• Bayesian Inference

Background:

• Hierarchical Pitman-Yor process models are crucial for analyzing complex data structures.
• Current Monte Carlo methods for these models, using the Chinese Restaurant Franchise (CRF) representation, require computationally intensive Markov chain Monte Carlo (MCMC) sampling of auxiliary variables.
• Efficient simulation of the posterior distribution remains a challenge.

Purpose of the Study:

• To develop a perfect sampler for the latent variables in the Chinese Restaurant Franchise (CRF) representation of the hierarchical Pitman-Yor process.
• To evaluate the performance and computational efficiency of the proposed perfect sampler.
• To compare the new algorithm with existing methods like Gibbs sampling and unbiased Monte Carlo estimation.

Main Methods:

• Development of a perfect sampler for CRF latent variables, drawing upon the Propp-Wilson algorithm.
• Extensive simulations to evaluate the average running time of the perfect sampler.
• Comparative analysis against Gibbs sampling and Glynn and Rhee's unbiased Monte Carlo estimation procedure.

Main Results:

• The developed perfect sampler efficiently samples latent variables for the hierarchical Pitman-Yor process.
• Simulation results demonstrate a significant, parameter-dependent running time for the sampler, exhibiting sharp transitions.
• The perfect sampler shows competitive or superior performance compared to existing methods.

Conclusions:

• The perfect sampler provides a valuable tool for accurate and efficient simulation from hierarchical Pitman-Yor process models.
• The algorithm's performance characteristics offer insights into the computational challenges of Bayesian inference.
• The method is applicable to real-world problems, as illustrated by its use in microbial genomics.