Towards robust separation of the cosmological 21-cm signal: structured modeling, regularization, and Bayesian inference
Abstract
Separating the cosmological 21-cm signal is particularly sensitive to the statistical modeling of contaminants and to model misspecification. ML-GPR makes it possible to introduce more flexible covariance models learned from simulations. However, separation remains fragile when these simulations are not sufficiently representative or when the statistical structures of the signal and contaminants become too similar. The research question is whether additional structural information can improve separation robustness and the assessment of its reliability.
This document presents an independent research proposal developed from an analysis of several studies on 21-cm signal separation using GPR and ML-GPR. Two complementary methodological directions are considered: structured modeling of contaminants, particularly through learned covariances and spatial-spectral information, and regularization of covariance-model learning. An evaluation strategy is proposed, first using simulations and then, where feasible, real observations. Evaluation would use probabilistic and Bayesian methods to assess separation robustness, signal loss, biases in the power spectrum, and the reliability of the resulting estimates.
About this document
This independent research proposal is based on an analysis of several studies of cosmological 21-cm signal separation using Gaussian process regression (GPR and ML-GPR).
It considers two complementary methodological directions: structured contaminant modeling and regularization of learned covariance models, together with a probabilistic and Bayesian strategy for evaluating the separation.
Initial version: September 30, 2026. Revised version: November 9, 2026.
