Group Greedy RLS Sparsity Estimation via Information Theoretic Criteria

This work introduces a group sparse adaptive greedy algorithm that uses information theoretic criteria (ITC) to estimate online the sparsity level. The algorithm selects a set of candidate groups using group neighbor permutations and maintains a partial QR decomposition to compute the solution. It contains a mechanism that allows group joining which, complementing the splitting of groups, produces a robust algorithm. We focus here on a study of the ITC use, namely the predictive least squares (PLS) and Bayesian information criterion (BIC), in conjunction with the group sparse algorithm. We propose several forms of group oriented ITC and evaluate them with extensive simulations for a time-varying channel identification problem. Compared to the non group aware counterparts, the performance is improved at the cost of higher complexity. The best results are given by a group PLS criterion directly generalizing the standard PLS.