Order estimation for subspace methods (Q5947627)

The author proposes three methods for order estimation in the context of subspace methods. Two of them, both in the cases with observed inputs and without exogenous inputs, are based on the information contained in the estimated singular values (SV) and lower bounds on the (subjective) penalty term in order for the estimates to be (strongly) consistent. The third uses the innovation variance, and it is shown to suffer from severe theoretical disadvantages. The first two can be applied with almost no computational costs to Larimore type and MOESP type procedures, whereas the third is only applied to Larimore type algorithms. Numerical examples and simulation studies illustrate the obtained results.