Model Parameter Gradients in Prediction Identification of State-Space Systems

The paper is devoted to the study of the gradient computation related to procedures for identifying state-space systems in the prediction error sense. The knowledge of these gradients is needed when iteratively estimating a state-space model for the system on the basis of data measurements. In classical estimation algorithm, any gradient signal is evaluated by running these data through a state-space dynamics corresponding to the model differentiation with respect to the related parameter. In order to reduce the computation burden of this estimation, the paper put into light the structure of the state-space gradient signals and, as a by product, propose a new method for computing them. The obtained improvement is based on exploiting the properties of matrices that commute with the prediction model state-feedback matrix.