Asymptotic variance of subspace methods by data orthogonalization and model decoupling: a comparative analysis (Q705468)

The reviewer agrees with the following (authors' conclusions): We have presented a comparison between two classes of subspace identification methods. The first is based on an orthogonal decomposition of the input-output data, combined with a block-decoupled parametrization of the model while the other methods are subspace methods based on the usual ``joint model parametrization''. Expressions for the asymptotic error covariances of the \((A,B,C,D)\) parameters and also of the transfer function estimates, have been obtained for both classes of subspace methods. In certain cases the formulas show that the errors for the joint parametrization are larger than for the decoupled parametrization. There is still work to be done to describe exactly the situations in which this claim holds true, however there is enough evidence to conclude that (when models with decoupled deterministic-stochastic dynamics are adequate) subspace estimation based on a block-diagonal model parametrization provides in most cases better estimates of the system transfer function than the standard ``joint'' input-output methods. The paper under discussion is a companion of the paper [\textit{A. Chiuso} and \textit{G. Picci}, ``Subspace identification by data orthogonalization and model decoupling'', Automatica 40, No. 10, 1689--1703 (2004; Zbl 1075.93038), reviewed above].