A model reduction method based on the least squares algorithm is deriv
ed. This method calculates a low order autoregressive moving average (
ARMA) predictor equation from a high order ARMA equation. The low orde
r ARMA equation minimizes the sum of the squares of the prediction err
ors when the input is white noise. This is almost equivalent to minimi
zing the sum of the squares of the error in the impulse response funct
ion. Transfer function models can also be used and a steady-state gain
constraint can be incorporated into the procedure. The merits of this
method of model order reduction are shown with three examples. In the
se examples, the proposed method produced results that compared favora
bly with highly regarded existing model reduction techniques which req
uire many more computations.