Anisotropy-based optimal filtering in linear discrete time invariant systems
Provides a unified robust filtering framework for control engineers dealing with colored or non-Gaussian disturbances.
The paper develops a robust filter for linear discrete-time systems that minimizes the worst-case RMS estimation error under bounded mean anisotropy of disturbances, bridging Kalman and H∞ filters.
This paper is concerned with a problem of robust filtering for a finite-dimensional linear discrete time invariant system with two output signals, one of which is directly observed while the other has to be estimated. The system is assumed to be driven by a random disturbance produced from the Gaussian white noise sequence by an unknown shaping filter. The worst-case performance of an estimator is quantified by the maximum ratio of the root-mean-square (RMS) value of the estimation error to that of the disturbance over stationary Gaussian disturbances whose mean anisotropy is bounded from above by a given parameter $a \ge 0$. The mean anisotropy is a combined entropy theoretic measure of temporal colouredness and spatial "nonroundness" of a signal. We construct an $a$-anisotropic estimator which minimizes the worst-case error-to-noise RMS ratio. The estimator retains the general structure of the Kalman filter, though with modified state-space matrices. Computing the latter is reduced to solving a set of two coupled algebraic Riccati equations and an equation involving the determinant of a matrix. In two limiting cases, where $a = 0$ or $a \to +\infty$, the $a$-anisotropic estimator leads to the standard steady-state Kalman filter or the $H_{\infty}$-optimal estimator, respectively.