Comparison of Two Methods of Modeling Stationary EEG Signals

This paper compares the performances—when applied to stationary EEG signals—of two methods of modeling stochastic time-series; viz, a maximum-likelihood search method for a mixed autoregressive and moving-average time-series, and the well-known method of least-squares identification of a prefiltered autoregressive series. Computing effort per step is derived for different-order search strategies, expressed in the number of a set of basic operations and also in equivalent number of likelihood function evaluations. Power spectra and total computing effort are evaluated and compared for four representative EEG samples. It appears that the least-squares method is generally superior because iteration is not needed, and in spite of the fact that higher orders are needed instead.