Direct Sampling with a Step Function

TL;DR

This paper introduces a step function-based variant of the direct sampling method for univariate weighted distributions, improving reliability and reducing manual tuning in generating exact samples from complex distributions.

Contribution

It proposes a new step function approximation technique within the direct sampling framework, enhancing sampling reliability and efficiency for univariate distributions.

Findings

Reliable sampling from complex distributions achieved

Reduced manual tuning and rejection rates

Successful applications to diverse statistical models

Abstract

The direct sampling method proposed by Walker et al. (JCGS 2011) can generate draws from weighted distributions possibly having intractable normalizing constants. The method may be of interest as a useful tool in situations which require drawing from an unfamiliar distribution. However, the original algorithm can have difficulty producing draws in some situations. The present work restricts attention to a univariate setting where the weight function and base distribution of the weighted target density meet certain criteria. Here, a variant of the direct sampler is proposed which uses a step function to approximate the density of a particular augmented random variable on which the method is based. Knots for the step function can be placed strategically to ensure the approximation is close to the underlying density. Variates may then be generated reliably while largely avoiding the need for manual tuning or rejections. A rejection sampler based on the step function allows exact draws to be generated from the target with lower rejection probability in exchange for increased computation. Several applications of the proposed sampler illustrate the method: generating draws from the Conway-Maxwell Poisson distribution, a Gibbs sampler which draws the dependence parameter in a random effects model with conditional autoregression structure, and a Gibbs sampler which draws the degrees-of-freedom parameter in a regression with t-distributed errors.