Kernel-based identification with frequency domain side-information (Q6157806)

In identification of linear dynamic systems it is possible to incorporate some side (or background) information available about the system, e.g., its low complexity, low order, positivity, location of eigenvalues, etc. These features can be forced through various penalty criteria (such as the sparsity promoting regularization criterion) that are minimized along with the underlying measure of fit. Unfortunately, inclusion of some frequency-domain side information has not been satisfactorily investigated yet, although quite often features such that the boundedness of the \(\mathcal{H}_\infty\)-norm of the plant or plant dissipativity are \textit{a priori} known for many systems either due to their inherent physical nature (e.g., in electric circuits energy is dissipated by resistors) or owing to some data-driven analysis. This paper by Khosravi and Smith aims at filling this gap. Primarily, the authors deal with the case where the \(\mathcal{H}_\infty\)-norm of the system is bounded by a given scalar. This setting includes various forms of frequency-domain information, such as system dissipativity. The authors propose a non-parametric identification approach for estimating the impulse response of the system under the given side information. They heavily exploit the kernel-based approach and formulate the identification problem as a constrained regularized regression problem in a stable reproducing kernel Hilbert space (RKHS) subject to suitable constraints to impose the frequency-domain side information on the estimation. The RKHS framework is employed for both discrete-time and continuous-time dynamics. The resulting infinite-dimensional optimization problem with an infinite number of constraints turns out to be convex and possess a unique solution. To make its solution tractable, a finite set of frequencies is selected and an approximate estimation problem with the same objective function but with constraints imposed according to these frequencies is further considered. This new problem yields a unique solution in a specific parametric form. It can be constructed numerically by solving an equivalent convex quadratically constrained quadratic programming problem. Proper bounds on the tightness of this approximation are derived and theoretical guarantees on the convergence of the approximate solution to that of the original problem are provided. Several numerical examples nicely illustrate the efficiency of the proposed method. The results presented are original and interesting, and the article is well written. Overall, the paper is a good piece of work in a subject that attracts considerable attention.