Consistency and relative efficiency of subspace methods (Q1911279)

Some identification procedures for linear stochastic systems are analyzed and compared. The systems are assumed to be stable. Strictly, minimum phase and having a known order. The first algorithm was proposed by Akaike (1976), the second one was presented by Larimore (1983). The algorithms use different schemes of state estimation but after that preliminary stage, the system matrices are computed by ordinary least-squares in both cases. The authors show that the matrices estimates are strongly consistent. The third algorithm under consideration was given by Hannan, Risanen, Astrom and Mayne (1982) and it is a variant of the iterative least-squares procedure. It is shown that the first algorithm and the third one are asymptotically equivalent. A relative efficiency of the second algorithm was analyzed by a Monte Carlo simulation. The results suggest that it is asymptotically efficient too and it can be recommended practically as the most numerically simple scheme.