On the state estimation problem for subspace identification

In this paper the focus is on those subspace methods that estimate the state sequence as an intermediate step, i.e. CVA and N4SID. To estimate the state sequence, it is proposed to use a statistically sound matrix multiplication, rather than computing the SVD in a routine fashion. In a simulation study it is especially considered how certain design parameters can be tuned to affect the bias distribution of the identified model. Additionally, it is discussed how the proposed approach can be applied to frequency weighted model reduction