On the divergence between linear processes

Abstract

In this paper we consider the problem of discriminating between stationary Gaussian processes. We examine the Kullback-Liebler divergence, which has been widely applied to classify time series with respect to the dominant frequency bands. Moreover, we analyse the connections between that approach based on the comparison of spectral densities and the alternative form in terms of the parameters or ARMA models. Finally, the relations between the J-divergence and other discrimination measures are discussed.