Study of conditional ML estimators in time and frequency-domain system identification

This paper shows that the full covariance matrix can be replaced by its main diagonal in maximum likelihood estimation (MLE) of linear dynamic systems in the time or frequency domain without affecting the asymptotic properties, This is a very important result for frequency-domain identification since it allows to use nonparametric noise models, It also provides a theoretical justification for the common practice in time-domain identification to neglect the initial condition effects of the noise model, (C) 1999 Elsevier Science Ltd, All rights reserved.