Performance of LMS-Newton adaptation algorithms with variable convergence factor in nonstationary environments

An analysis of two LMS (least mean square) Newton adaptive filtering algorithms with variable convergence factor is presented. These algorithms are known to yield zero a posteriori error. The relation between the variable convergence factor used to update R (the input autocorrelation matrix) and a variable forgetting factor that weights the past input samples is addressed. The analysis is carried out for stationary as well as nonstationary environments, and the relations of these algorithms to the RLS (recursive least squares) algorithm are explored. The experimental results obtained agree well with the results predicted by using the derived formulas.<>