Blind separation of noisy Gaussian stationary sources. Application to cosmic microwave background imaging

We present a new source separation method which maximizes the likelihood of a model of noisy mixtures of stationary, possibly Gaussian, independent components. The method has been devised to address an astronomical imaging problem. It works in the spectral domain where, thanks to two simple approximations, the likelihood assumes a simple form which is easy to handle (low dimensional sufficient statistics) and to maximize (via the EM algorithm).