Abstract:This work presents a new block-sparse least mean square (BS-LMS) algorithm with adaptive and nonuniform group sizes. Motivated by the fact that large and zero coefficients in block-sparse systems assemble in clusters of variant sizes, we divide the adaptive filter tap-weights into groups of nonuniform size. A predetermined threshold value is utilized for real-time group partitioning Large coefficient groups consist of clusters of coefficients whose magnitudes are above the threshold value, while zero coefficient groups consist of clusters of coefficients whose magnitudes are below the threshold value. Then, coefficients in large coefficient groups are updated by following the standard LMS algorithm, while those in zero coefficient groups suffer from an attractive force towards zero. A practical choice guideline is provided for the threshold value to facilitate applications. The proposed algorithm is tested in the identification of single- and multiple-block-sparse systems. Simulation results show that the proposed scheme effectively locates and tracks the non-zero blocks in the unknown systems, outperforming existing block-sparsity-inspired LMS algorithms with uniform group sizes.

A New Block-Sparse Lms Algorithm With Adaptive And Nonuniform Group Sizes