Cosmological Parameters from CMB Maps without Likelihood Approximation

TL;DR

This paper introduces a novel Bayesian MCMC algorithm for estimating cosmological parameters from CMB maps that avoids likelihood approximations, improving efficiency across different signal-to-noise regimes.

Contribution

It presents a new joint move in the MCMC algorithm that effectively handles high and low signal-to-noise data by rescaling the signal map after subtracting the Wiener filter mean.

Findings

Accurately recovers the exact posterior with less than 0.006 sigma deviation.

Achieves good mixing with Markov Chain correlation lengths comparable to existing methods.

Demonstrates improved efficiency in high signal-to-noise regimes.

Abstract

We propose an efficient Bayesian MCMC algorithm for estimating cosmological parameters from CMB data without use of likelihood approximations. It builds on a previously developed Gibbs sampling framework that allows for exploration of the joint CMB sky signal and power spectrum posterior, P(s,Cl|d), and addresses a long-standing problem of efficient parameter estimation simultaneously in high and low signal-to-noise regimes. To achieve this, our new algorithm introduces a joint Markov Chain move in which both the signal map and power spectrum are synchronously modified, by rescaling the map according to the proposed power spectrum before evaluating the Metropolis-Hastings accept probability. Such a move was already introduced by Jewell et al. (2009), who used it to explore low signal-to-noise posteriors. However, they also found that the same algorithm is inefficient in the high signal-to-noise regime, since a brute-force rescaling operation does not account for phase information. This problem is mitigated in the new algorithm by subtracting the Wiener filter mean field from the proposed map prior to rescaling, leaving high signal-to-noise information invariant in the joint step, and effectively only rescaling the low signal-to-noise component. To explore the full posterior, the new joint move is then interleaved with a standard conditional Gibbs sky map move. We apply our new algorithm to simplified simulations for which we can evaluate the exact posterior to study both its accuracy and performance, and find good agreement with the exact posterior; marginal means agree to less than 0.006 sigma, and standard deviations to better than 3%. The Markov Chain correlation length is of the same order of magnitude as those obtained by other standard samplers in the field.