Umbrella sampling: a powerful method to sample tails of distributions

TL;DR

Umbrella sampling is an efficient, parallelizable method for sampling extremely low probability regions of complex posterior distributions, improving robustness in multi-modal and high-dimensional Bayesian analyses.

Contribution

The paper introduces umbrella sampling as a versatile, easy-to-implement technique that enhances sampling of rare events and low probability areas in posterior distributions, with a public Python package.

Findings

Efficiently samples low probability regions in complex posteriors.

Outperforms standard MCMC in multi-modal and high-dimensional cases.

Provides accurate sampling down to 15σ credible regions in cosmological data.

Abstract

We present the umbrella sampling (US) technique and show that it can be used to sample extremely low probability areas of the posterior distribution that may be required in statistical analyses of data. In this approach sampling of the target likelihood is split into sampling of multiple biased likelihoods confined within individual umbrella windows. We show that the US algorithm is efficient and highly parallel and that it can be easily used with other existing MCMC samplers. The method allows the user to capitalize on their intuition and define umbrella windows and increase sampling accuracy along specific directions in the parameter space. Alternatively, one can define umbrella windows using an approach similar to parallel tempering. We provide a public code that implements umbrella sampling as a standalone python package. We present a number of tests illustrating the power of the US method in sampling low probability areas of the posterior and show that this ability allows a considerably more robust sampling of multi-modal distributions compared to the standard sampling methods. We also present an application of the method in a real world example of deriving cosmological constraints using the supernova type Ia data. We show that umbrella sampling can sample the posterior accurately down to the $\approx 15\sigma$ credible region in the $\Omega_{\rm m}-\Omega_\Lambda$ plane, while for the same computational work the affine-invariant MCMC sampling implemented in the {\tt emcee} code samples the posterior reliably only to $\approx 3\sigma$.