Consistency of system identification by global total least squares (Q1301449)

The global total least squares (GTLS) approach to the identification of multiple-input multiple-output (MIMO) discrete-time linear systems within a stochastic framework is considered. The aim is to find models with minimal error for a given a priori bound on the model complexity. It is assumed that the observed time-series data come from a multivariable stationary stochastic process and statistical properties of the method are investigated. Sufficient conditions for the almost sure consistency of the GTLS estimate for the class of factor (suitably decomposed) models are established. Under these conditions the identified model converges, as the number of observations tends to infinity, to an optimal approximation of the data generating process. The theoretical results are illustrated by computer simulations.