Abstract-A new Bayesian approach to linear system identification has been proposed in a series of recent papers. The main idea is to frame linear system identification as predictor estimation in an infinite dimensional space, with the aid of regularization/Bayesian techniques. This approach guarantees the identification of stable predictors based on the prediction error minimization. Unluckily, the stability of the predictors does not guarantee the stability of the impulse response of the system. In this paper we propose and compare various techniques to address this issue. Simulations results comparing these techniques will be provided.
I. INTRODUCTIONRecent approaches for linear system identification describe the unknown system directly in terms of impulse response, thus describing an infinite dimensional model class. Needless to say, this is not entirely free of difficulties, since an alternative way to control the model complexity, i.e., to face the so called-bias variance tradeoff [1], [2], need to be found. It has been shown in the recent literature that the apparatus of Reproducing Kernel Hilbert Spaces (RHKS) or, equivalently, Bayesian Statistics provide powerful tools to face this tradeoff.The paper [3] has shown how these infinite dimensional model classes can be used for identification of linear systems in the framework of prediction error methods, leading naturally to stable predictors. Yet stability of the predictor model does not necessarily guarantee stability of the so called "forward" (or simulation) model. As a matter of fact, we faced this stability issue when performing identification on a real data set from EEG recordings. Physical insight in this case suggests that the transfer function describing the link between potentials in different brain locations are expected to be stable, while the identified models where not.Therefore, motivated by this real-world application, in this paper we shall tackle the problem of identifying stable (simulation) models when nonparametric prediction error methods [3] are used. We shall describe and compare, through an extensive simulation study, four possible solutions to this problem.The paper is structured as follows: Section II formulates the problem. Sections III-V introduce four different approaches to guarantee stability of the identified models. Experimental results are described in Section VI and conclusions are drawn in Section VII.Notation: Given a matrix M , M shall denote its transpose, σ(M ) will be its eigenvalues. If A(z) is a polynomial, σ(A(z)) will denote the set of roots of A(z). Given two discrete time jointly stationary stochastic process y(t) and z(t), the symbol E[y(t)|z(s), s < t] shall denote the linear