Global 21-cm signal extraction from foreground and instrumental effects I: Pattern recognition framework for separation using training sets

TL;DR

This paper introduces a novel signal extraction method for global 21-cm cosmology experiments using SVD of training sets to effectively separate the cosmological signal from complex foregrounds and systematics, implemented in an open-source Python package.

Contribution

It presents a new SVD-based framework for separating the 21-cm signal from systematics, which does not require precise absolute knowledge of systematics, only their variability.

Findings

The method effectively isolates the 21-cm signal in simulated data.

Minimizing the Deviance Information Criterion yields unbiased signal estimates.

The approach is implemented in the publicly available 'pylinex' Python package.

Abstract

The sky-averaged (global) highly redshifted 21-cm spectrum from neutral hydrogen is expected to appear in the VHF range of $\sim20-200$ MHz and its spectral shape and strength are determined by the heating properties of the first stars and black holes, by the nature and duration of reionization, and by the presence or absence of exotic physics. Measurements of the global signal would therefore provide us with a wealth of astrophysical and cosmological knowledge. However, the signal has not yet been detected because it must be seen through strong foregrounds weighted by a large beam, instrumental calibration errors, and ionospheric, ground and radio-frequency-interference effects, which we collectively refer to as "systematics". Here, we present a signal extraction method for global signal experiments which uses Singular Value Decomposition (SVD) of "training sets" to produce systematics basis functions specifically suited to each observation. Instead of requiring precise absolute knowledge of the systematics, our method effectively requires precise knowledge of how the systematics can vary. After calculating eigenmodes for the signal and systematics, we perform a weighted least square fit of the corresponding coefficients and select the number of modes to include by minimizing an information criterion. We compare the performance of the signal extraction when minimizing various information criteria and find that minimizing the Deviance Information Criterion (DIC) most consistently yields unbiased fits. The methods used here are built into our widely applicable, publicly available Python package, $\texttt{pylinex}$, which analytically calculates constraints on signals and systematics from given data, errors, and training sets.