Nonstationary learning characteristics of the LMS algorithm

Upper and lower bounding first-order linear recursions for the mean-squared error realized with the LMS algorithm subjected to a sequence of independent nonstationary training vectors are derived. These bounds coincide to give the exact evolution of mean-squared error for the problem of identification of a nonrecursive time-varying system with white-noise excitation. This leads to an exact formula for time-averaged mean-squared error that is used to study optimization of the step-size parameter for minimum time-average misadjustment. New results on dependence of the minimal step size and the minimum misadjustment on the degree of nonstationarity are obtained.