Multivariate time series discrimination in the spectral domain

Multivariate time series discrimination for evoked potentials from two cognitively different populations is discussed. A method for constructing linear and quadratic discriminant functions in the spectral domain using a stepwise frequency selection procedure is described and applied. Further, the effects of rank transformations on the estimated nonerror rates are examined. Finally, the validity of the signal plus stationary noise model (SSN) is examined. The performance of the SSN model is compared with the performance of a model (SNN) which consists of a signal plus nonstationary noise and a model (Mix) which consists of mixtures of subpopulations.