Searching for efficient Markov chain Monte Carlo proposal kernels.

Area of Science:

• Computational statistics
• Bayesian inference
• Algorithm development

Background:

• Markov chain Monte Carlo (MCMC) methods, particularly the Metropolis-Hastings algorithm, are fundamental to modern Bayesian statistical inference.
• The efficiency of various Metropolis-Hastings proposal kernels is crucial for MCMC performance but remains understudied, with limited focus beyond the Gaussian proposal.

Purpose of the Study:

• To propose and evaluate a novel class of Bactrian kernels for Metropolis-Hastings algorithms.
• To compare the efficiency of Bactrian kernels against existing proposals, including the Gaussian kernel, for simulating diverse target distributions.
• To assess the practical applicability of these kernels in real-world Bayesian analyses.

Main Methods:

• Development of a unique class of Bactrian kernels designed to avoid proposing values too close to the current state.
• Comparative analysis of proposal kernel efficiency using asymptotic variance of parameter estimates.
• Implementation and testing of the proposed kernels within a Bayesian program for molecular clock dating.

Main Results:

• The uniform kernel demonstrated higher efficiency than the standard Gaussian kernel.
• The proposed Bactrian kernels were found to be even more efficient than the uniform kernel.
• Optimally scaled Bactrian kernels achieved at least 50% greater efficiency compared to optimally scaled Gaussian kernels.
• Results were validated through implementation in a molecular clock dating application, confirming general applicability.

Conclusions:

• Bactrian kernels represent a significant advancement in MCMC proposal efficiency, outperforming traditional Gaussian and uniform kernels.
• The findings challenge previous assertions of similar performance across different MCMC proposals.
• This research encourages further investigation into developing and utilizing efficient MCMC proposal mechanisms for enhanced Bayesian inference.