Approximate Best Linear Prediction of a Certain Class of Stationary and Nonstationary Noise-Distorted Signals

Abstract

A technique for approximating the best linear predictor of a noise-distorted signal is presented. The technique is applicable when the signal is stationary with a spectrum which is highly concentrated around known or estimated frequencies, or when the signal is an accumulating sequence whose “spectrum” has poles at these frequencies. The noise must be stationary, but its spectrum need only be known or estimated at the above-mentioned frequencies. The time parameter is assumed to be discrete. Computer-generated data are then used to obtain an idea of the “robustness” of the technique.