Transient performance degradation of the LMS, RLS, sign, signed regressor, and sign-sign algorithms with data correlation

Abstract

The transient performance degradation, for correlated input data, is studied for various adaptive algorithms. The algorithms are least mean square (LMS), recursive least squares (RLS), sign (SA), signed regressor (SRA), and sign-sign (SSA). Analysis is performed for adaptive plant identification with stationary Gaussian inputs. A closed-form expression for the mean convergence time is derived for each algorithm. The degradation measure used is the ratio of convergence time for correlated data to the convergence time for white data. The LMS and SRA degradations are the same. The SA and SSA degradations are also the same. The smallest degradation occurs for RLS, the largest for LMS and SRA. The SRA, RLS, and LMS degradations are independent of plant-noise variance. The SA and SSA degradations increase with increased noise variance. The LMS and SRA degradations do not depend upon the weight initialization, RLS (SA and SSA) depends weakly (significantly) upon the weight initialization. The analytical results are supported by simulations.