Improved convergence of gradient algorithms for adaptive IIR filters

Abstract

The introduction of a simple correction term in gradient algorithms for adaptive IIR filtering is shown to improve their convergence and robustness. The error surface, i.e. the mean-squared value of the output error as a function of adaptive coefficients, is quadratic in adaptive FIR (finite-impulse-response) filtering, and thus a simple gradient search works well. This is not the case in adaptive IIR (infinite-impulse-response) filtering, where the error surface is far from being quadratic. Nevertheless, a simple modification of the updating equation of lattice-form algorithms can reduce this problem, resulting in faster adaptive IIR filters