Understanding posterior projection effects with normalizing flows

TL;DR

This paper introduces a method using normalizing flows with an evidence loss to accurately learn and interpret high-dimensional posterior distributions, especially in cosmology, reducing projection effects and enabling reliable model comparison.

Contribution

The authors develop a robust approach to learn smooth posterior representations with normalizing flows, improving interpretation and evidence estimation in high-dimensional Bayesian inference.

Findings

Accurately estimates Bayesian evidence with 0.2 log error.

Reduces projection effects by optimizing over parameters.

Effective on multi-modal Gaussian mixtures up to 32 dimensions.

Abstract

Many modern applications of Bayesian inference, such as in cosmology, are based on complicated forward models with high-dimensional parameter spaces. This considerably limits the sampling of posterior distributions conditioned on observed data. In turn, this reduces the interpretability of posteriors to their one- and two-dimensional marginal distributions, when more information is available in the full dimensional distributions. We show how to learn smooth and differentiable representations of posterior distributions from their samples using normalizing flows, which we train with an added evidence error loss term, to improve accuracy in multiple ways. Motivated by problems from cosmology, we implement a robust method to obtain one and two-dimensional posterior profiles. These are obtained by optimizing, instead of integrating, over other parameters, and are thus less prone than marginals to so-called projection effects. We also demonstrate how this representation provides an accurate estimator of the Bayesian evidence, with log error at the 0.2 level, allowing accurate model comparison. We test our method on multi-modal mixtures of Gaussians up to dimension 32 before applying it to simulated cosmology examples. Our code is publicly available at https://github.com/mraveri/tensiometer.