Probing local non-Gaussianities within a Bayesian framework

TL;DR

This paper develops a Bayesian framework for inferring the level of local non-Gaussianity (f_NL) in CMB data, introducing efficient sampling algorithms and demonstrating their effectiveness on a simplified model.

Contribution

It introduces a novel Bayesian inference method for f_NL, including exact Hamiltonian sampling and an approximate Monte Carlo technique for high signal-to-noise data.

Findings

Approximate algorithm outperforms Hamiltonian sampling by two orders of magnitude at high SNR.

The Bayesian approach effectively accounts for instrumental effects and uncertainties.

Bias towards f_NL=0 occurs when data do not strongly constrain non-Gaussianity.

Abstract

Aims: We outline the Bayesian approach to inferring f_NL, the level of non-Gaussianity of local type. Phrasing f_NL inference in a Bayesian framework takes advantage of existing techniques to account for instrumental effects and foreground contamination in CMB data and takes into account uncertainties in the cosmological parameters in an unambiguous way. Methods: We derive closed form expressions for the joint posterior of f_NL and the reconstructed underlying curvature perturbation, Phi, and deduce the conditional probability densities for f_NL and Phi. Completing the inference problem amounts to finding the marginal density for f_NL. For realistic data sets the necessary integrations are intractable. We propose an exact Hamiltonian sampling algorithm to generate correlated samples from the f_NL posterior. For sufficiently high signal-to-noise ratios, we can exploit the assumption of weak non-Gaussianity to find a direct Monte Carlo technique to generate independent samples from the posterior distribution for f_NL. We illustrate our approach using a simplified toy model of CMB data for the simple case of a 1-D sky. Results: When applied to our toy problem, we find that, in the limit of high signal-to-noise, the sampling efficiency of the approximate algorithm outperforms that of Hamiltonian sampling by two orders of magnitude. When f_NL is not significantly constrained by the data, the more efficient, approximate algorithm biases the posterior density towards f_NL = 0.