Linear prediction based on higher order statistics by a new criterion

Abstract

This criterion requires only partial Mth-order cumulants C/sub M,e/(0,k/sub 1/, k/sub 1/, . . ., k/sub M/2-1/, k/sub M/ /sub /2-1/) of the prediction error e(k) where M is even. Theoretically, it is shown that the proposed filter associated with a stationary process x(k) is the same as the conventional correlation based (minimum-phase) LPE filter associated with the nonGaussian signal y(k) (noise-free). Simulation results show that when y(k) is an autoregressive process of known order, the proposed filter works well.<>