A statistical perspective on state-space modeling using subspace methods

Abstract

The authors investigate aspects of subspace-based state-space identification techniques from a statistical perspective. They concentrate their efforts on a simple approach which is based on finding the range-space of the observability matrix of a state-space representation. The system description is then found using the shift-invariance property of the observability matrix. It is shown that this results in a consistent system description for multivariable output-error models if the measurement noise is white in time and independent from output to output. The asymptotic covariance of the estimated poles of the system is also derived. In the test case studied, the subspace technique performs comparably with the statistically efficient PE (prediction error) method, whereas the instrumental variable method does notably worse. Hence, the subspace technique may be a strong candidate for determining initial values for the optimization in the efficient PE method.<>