Bayesian semi-blind component separation for foreground removal in interferometric 21cm observations

We present a new Bayesian semi-blind approach which is an extension of Independent Component Analysis (ICA) from two-dimensional (2-D) CMB map to the three-dimensional (3-D) 21-cm cosmological signal. This technique provides a fully Bayesian inference of power spectra and maps. Only relying on the statistical independence of the components, this approach can jointly estimate the 3-D power spectrum of the 21-cm signal and, the 2-D angular power spectrum and the frequency dependence of each foreground component, without any prior assumptions about foregrounds. This approach has been tested intensively by applying it to mock data from an interferometric 21-cm intensity mapping observation. Based on the Expectation-Maximization (EM) algorithm, this blind approach provides much better performance in 21-cm power spectrum recovery over all the scales than the commonly used Principal Component Analysis (PCA). This technique could be straightforwardly applied to the epoch of reionization measurements.