ON SYSTEM-IDENTIFICATION FOR LINEAR MINIMUM VARIANCE PREDICTION OR CONTROL

We consider system identification of a linear model to be used for linear minimum variance prediction or control. Gaussian noise processes with everywhere positive spectral density may be assumed to be generated via an invertible transfer function. We consider non-Gaussian noise processes whose spectral density is everywhere positive and whose generation involves a non-invertible transfer function. We show that system identification based on least-squares and (incorrectly) assuming invertibility of this transfer function leads to results nonetheless useful for linear minimum variance prediction or control. © 1990.