On the concept of estimation memory in adaptive filtering (Q1876953)

The paper studies the exponentially weighted least squares (EWLS), least mean squares (LMS) and Kalman filtering (KF) algorithms from the point of view of their equivalent estimation memory, i.e. the number of steps required by the least squares (LS) algorithm applied to a time-invariant system to reach the same steady-state excess mean square prediction error of the algorithm under consideration. It is shown that the LMS algorithms are more robust, with respect to variations of the covariance structure of the regressors, than the EWLS algorithms, while the KF algorithm has an intermediate performance in terms of predictive ability.