Alternating convex projection methods for discrete-time covariance control design (Q1908638)

The problem of designing a controller for a linear, discrete-time system is reduced to finding a suitable covariance matrix \(X\) for the closed-loop system. The paper shows that the design problem in terms of \(X\) has some desirable properties. The corresponding assignability and performance constrained sets are convex sets; moreover, expressions for the projections on these sets are derived analytically. This allows the use of alternating convex projection methods to solve the covariance design problem. Improved convergence rates are obtained by using directional information. The algorithms presented can solve the following problems by convex programming: a covariance feasibility problem (covariance assignability subject to output system performance constraints); a covariance optimization problem (covariance assignability subject to output performance constraints such that the covariance is as close as possible to a desired one); and an infeasible covariance optimization problem (the closest assignable covariance to a given performance constraint set is obtained). Also, the case of a reduced-order dynamic controller design is discussed, and an alternating projection approach is suggested to attack this problem.