All Sky Modelling Requirements for Bayesian 21 cm Power Spectrum Estimation with BayesEoR

TL;DR

This paper evaluates the BayesEoR code for 21 cm power spectrum recovery, addressing challenges of foreground contamination and computational constraints, and proposes a modified approach for all-sky modeling within feasible computational limits.

Contribution

It introduces a modified BayesEoR method enabling all-sky foreground modeling with limited 21 cm signal modeling, improving power spectrum recovery under realistic observational conditions.

Findings

BayesEoR accurately recovers the 21 cm power spectrum with limited sky modeling.

Including all-sky foregrounds without modification leads to contamination.

The modified method enables feasible all-sky analysis on large compute clusters.

Abstract

We present a comprehensive simulation-based study of the BayesEoR code for 21 cm power spectrum recovery when analytically marginalizing over foreground parameters. To account for covariance between the 21 cm signal and contaminating foreground emission, BayesEoR jointly constructs models for both signals within a Bayesian framework. Due to computational constraints, the forward model is constructed using a restricted field-of-view (FoV) in the image domain. When the only EoR contaminants are noise and foregrounds, we demonstrate that BayesEoR can accurately recover the 21 cm power spectrum when the component of sky emission outside this forward-modelled region is downweighted by the beam at the level of the dynamic range between the foreground and 21 cm signals. However, when all-sky foreground emission is included along with a realistic instrument primary beam with sidelobes above this threshold extending to the horizon, the recovered power spectrum is contaminated by unmodelled sky emission outside the restricted FoV model. Expanding the combined cosmological and foreground model to cover the whole sky is computationally prohibitive. To address this, we present a modified version of BayesEoR that allows for an all-sky foreground model, while the modelled 21 cm signal remains only within the primary FoV of the telescope. With this modification, it will be feasible to run an all-sky BayesEoR analysis on a sizeable compute cluster. We also discuss several future directions for further reducing the need to model all-sky foregrounds, including wide-field foreground subtraction, an image-domain likelihood utilizing a tapering function, and instrument primary beam design.