Bregman Divergence Based Approach for Adaptive System Identification and Line Enhancement

This article puts forth a novel class of least mean square (LMS) techniques, including beta divergence-inspired LMS (BLMS), Itakura-Saito divergence-inspired LMS (ISBLMS), and Kullback-Leibler divergence-inspired LMS (KLLMS). The mentioned divergence measures are part of the Bregman divergence category of information-theoretic divergence. Mean and mean-square analysis for the proffered class of algorithms is derived to find the bound on the learning-rate for stable convergence. The effectiveness of the introduced class of algorithms is showcased for the application of time-varying system identification and adaptive line enhancement.