\(L_{2}\) model reduction and variance reduction (Q1614422)

The authors study several variance properties related to model reduction for finite impulse response (FIR) and output error (OE) models. The variances of two models, one deduced directly from data and the other by reducing a high order model by \(L_2\) model reduction, are compared. It is shown that for the FIR case it is always better not to estimate the model from data but to use \(L_2\) model reduction of a higher order FIR model while for the OE, if the reduced model class contains the true system, the reduced model has the same variance as the directly estimated one.