Block-Sparsity-Induced System Identification Using Efficient Adaptive Filtering

In this paper, we propose an efficient proportionate type block sparse LMS algorithm with a group zero-point attraction (GZA) penalty term for clustered sparse system identification. The proposed algorithm is based on the combination of a mechanism for proportionate gain control, and a mixed
$l_{2},0$
norm regularization, and outperforms the existing class of block proportionate sparsity-induced algorithms. The performance analysis of the proposed algorithm is then carried out, providing limits to the mean deviation from the original system. We also propose an improved proportionate type block sparse adaptive filtering algorithm with modified gain control mechanism. This one is more robust to the varying degrees of sparsity in the system to be identified than the former. Numerical simulations to identify single and two clustered sparse systems using white, correlated, and speech signals manifest the superiority of the proposed algorithms.