A New Recursive Approach to Sparse Representation

This work presents a new approach to online sparse representation. Online sparse representation concerns the estimation of non-redundant structures using data that are sequentially and continuously collected. A novel sparse regularized recursive least squares algorithm, named SP-R-RLS, is proposed. SP-R-RLS combines a reweighting technique to approximate the L0 norm (a measure of sparsity) with a smooth approximation to address lack of differentiability. When compared to state-of-the-art algorithms, it is shown that SP-R-RLS has better performance in terms of sparsity of the estimated system and accuracy of the estimate.