Statistical recovery of 21cm visibilities and their power spectra with Gaussian constrained realisations and Gibbs sampling

TL;DR

This paper introduces a Bayesian method combining Gaussian constrained realisations and Gibbs sampling to recover 21cm signals and their power spectra from interferometric data with gaps and uncertain foregrounds, improving analysis robustness.

Contribution

It presents a scalable Bayesian framework that jointly models 21cm signals, foregrounds, and data gaps, advancing previous non-parametric and in-painting techniques.

Findings

Successfully applied to simulated HERA data

Effectively marginalizes over foreground uncertainties

Improves signal recovery in gappy, contaminated data

Abstract

Radio interferometers designed to probe the 21cm signal from Cosmic Dawn and the Epoch of Reionisation must contend with systematic effects that make it difficult to achieve sufficient dynamic range to separate the 21cm signal from foreground emission and other effects. For instance, the instrument's chromatic response modulates the otherwise spectrally smooth foregrounds, making them difficult to model, while a significant fraction of the data must be excised due to the presence of radio frequency interference (RFI), leaving gaps in the data. Errors in modelling the (modulated and gappy) foregrounds can easily generate spurious contamination of what should otherwise be 21cm signal-dominated modes. Various approaches have been developed to mitigate these issues by (e.g.) using non-parametric reconstruction of the foregrounds, in-painting the gaps, and weighting the data to reduce the level of contamination. We present a Bayesian statistical method that combines these approaches, using the coupled techniques of Gaussian constrained realisations (GCR) and Gibbs sampling. This provides a way of drawing samples from the joint posterior distribution of the 21cm signal modes and their power spectrum in the presence of gappy data and an uncertain foreground model in a computationally scalable manner. The data are weighted by an inverse covariance matrix that is estimated as part of the inference, along with a foreground model that can then be marginalised over. We demonstrate the application of this technique on a simulated HERA-like delay spectrum analysis, comparing three different approaches for accounting for the foreground components.