Sparse component selection with application to MEG source localization

In several applications, the observed signal can be modeled as the projection of a sparse signal with constant support over time plus additive noise. In this paper, we develop a sparse component selection method which models the latent signal to be sparse and to be composed of a number unknown basis signals. The estimation is based on the maximization of the penalized log-likelihood, for which we develop an efficient minorization-maximization (MM) algorithm. We use simulations with synthetic data and real data from a magnetoencephalography (MEG) experiment to demonstrate the performance of the method.