Modern Bayesian Sampling Methods for Cosmological Inference: A Comparative Study

TL;DR

This paper compares various Bayesian sampling methods for cosmological inference, highlighting their strengths and limitations in different scenarios, especially in high-dimensional and complex distributions.

Contribution

It provides a comprehensive evaluation of MCMC, HMC, slice sampling, and nested sampling methods on test problems and real cosmological data, revealing their relative efficiencies.

Findings

HMC and nested sampling outperform traditional MCMC in complex distributions.

Nested sampling maintains accuracy but requires more computational resources.

All samplers perform well on simple Gaussian distributions.

Abstract

We present a comprehensive comparison of different Markov Chain Monte Carlo (MCMC) sampling methods, evaluating their performance on both standard test problems and cosmological parameter estimation. Our analysis includes traditional Metropolis-Hastings MCMC, Hamiltonian Monte Carlo (HMC), slice sampling, nested sampling as implemented in dynesty, and PolyChord. We examine samplers through multiple metrics including runtime, memory usage, effective sample size, and parameter accuracy, testing their scaling with dimension and response to different probability distributions. While all samplers perform well with simple Gaussian distributions, we find that HMC and nested sampling show advantages for more complex distributions typical of cosmological problems. Traditional MCMC and slice sampling become less efficient in higher dimensions, while nested methods maintain accuracy but at higher computational cost. In cosmological applications using BAO data, we observe similar patterns, with particular challenges arising from parameter degeneracies and poorly constrained parameters.