Choosing the right MCMC sampler: a systematic benchmark of gradient-free methods

Abstract

We present a set of metrics and methods for testing and comparing a range of modern gradient-free Markov Chain Monte Carlo (MCMC) samplers against the commonly used Metropolis-Hastings (MH) algorithm. The goal is to quantify key performance metrics, including sampler ergodicity, robustness and overall likelihood performance. To provide a controlled and interpretable testbed, we use the Rosenbrock function and Neal's funnel as representative unimodal cases, while Gaussian random likelihood landscapes in three, five, and eight dimensions serve as multimodal test scenarios. The samplers considered include affine-invariant moves from the literature, such as the stretch and walk moves, the differential evolution move, and the snooker move. We additionally introduce two novel variations: a modified stretch move that incorporates a Principal Component Analysis (PCA) transformation, and a hybrid blend move that combines features of both differential evolution and stretch dynamics. Beyond sampler evaluation, we demonstrate reconstructing likelihood landscapes from sampled points using a quadtree algorithm. Additionally, we explore the use of optimisation algorithms to refine the best parameter set in terms of its likelihood, and find consistent improvements in log-likelihood values, with the post-sampling gain becoming more significant in higher-dimensional problems. Our comparative results of sampler testing show that the differential evolution algorithm, when tuned to a target acceptance fraction of 25%, consistently outperforms all other samplers in terms of ergodicity, robustness, and likelihood performance.