Comparison of sampling techniques for Bayesian parameter estimation

Abstract

The posterior probability distribution for a set of model parameters encodes all that the data have to tell us in the context of a given model; it is the fundamental quantity for Bayesian parameter estimation. In order to infer the posterior probability distribution we have to decide how to explore parameter space. Here we compare three prescriptions for how parameter space is navigated, discussing their relative merits. We consider Metropolis-Hasting sampling, nested sampling and affine-invariant ensemble MCMC sampling. We focus on their performance on toy-model Gaussian likelihoods and on a real-world cosmological data set. We outline the sampling algorithms themselves and elaborate on performance diagnostics such as convergence time, scope for parallelisation, dimensional scaling, requisite tunings and suitability for non-Gaussian distributions. We find that nested sampling delivers high-fidelity estimates for posterior statistics at low computational cost, and should be adopted in favour of Metropolis-Hastings in many cases. Affine-invariant MCMC is competitive when computing clusters can be utilised for massive parallelisation. Affine-invariant MCMC and existing extensions to nested sampling naturally probe multi-modal and curving distributions.

Probing spatial homogeneity with LTB models: a detailed discussion

“…Convergence of the samples was ensured by letting the random walks proceed for multiple integrated auto-correlation times τ int after removing all samples drawn during the initial burn-in period. This procedure was advocated by Akeret et al (2013) and Allison & Dunkley (2014), who provided a detailed discussion of convergence diagnostics with the auto-correlation times described above.…”

A semicoherent glitch-robust continuous-gravitational-wave search method

“…In future, we intend to extend the functionality to include nested sampling [27] which will improve the robustness of the evidence calculation (see, e.g., Ref. [28] for a comparison).…”

Optimization of statistical methods for HpGe gamma-ray spectrometer used in wide count rate ranges

“…Ellipsoidal sampling [3] replaces the iso-likelihood surface L ¼ L n by a hyper-ellipsoid given by the covariance matrix of the living samples and centered in their mean value, and L new 4 L n is sampled within the intersection of the domain of integration and this ellipsoidal boundary [7]. When the integrand in (3) is a multimodal function, a possible strategy is partitioning the set of all living points in clusters, and then enclose each cluster in a "small" hyper-ellipsoid.…”