Model reduction by Kautz filters

Abstract

A method is presented for model reduction. It is based on the representation of the original model in an (exact) Kautz series. The Kautz series is an orthonormal model and is non-unique: it depends on the ordering of the poles. The ordering of the poles can be chosen such that the last sections contribute least or the first sections contribute most to the overall impulse response of the originalsystem (in a quadratic sense). Having a specific ordering, the reduced model order, say n, can be chosen by considering the energy contained in a truncated representation. The resulting reduced order model is obtained simply by truncation of the Kautz series at the nth term.