Constrained optimal estimation and control (Q1388123)

The classical theories of the linear quadratic regulator and of the linear Gaussian estimator define the full gain matrix completely. However, many control and estimation problems would benefit from prescribing a different structure to the gain matrix. Typical is the case of the output feedback control, object of a wide literature since the 1970s. The present paper analyzes first of all the problem of a constrained estimator. In this problem, the minimization of the covariance matrix \(P\) assumes its role in the most natural way, and no additional complexities arise from initial conditions. The linear constraints are treated by making explicit their dependence upon the ``free'' gain components, thus reducing the optimized control variables to their minimum necessary number. The optimization is performed with reference to the implicit definition of \(P\) given by a ``modified Riccati equation''. A stable initial guess is required, and this is obtained by positively forcing stability by means of a stability constraint in the first step of the procedure. A sequence of stable estimators is then obtained by means of an iterative scheme with a sufficiently small step size. Once the estimator has been obtained, the controller is deduced by assuming that the duality between the two extends also to this application. This assumption eliminates concern about the initial conditions.