Convex Reformulation of Minkowski-Sum Outer Approximation for Matrix Ellipsoids in Data-Driven Control

Abstract

Approximating Minkowski sums of matrix ellipsoids by a single matrix ellipsoid provides a tractable description of noise bounds for controller design from noisy data. Existing linear matrix inequality (LMI) conditions can be used to compute such approximations, but directly solving the resulting optimization problems can be computationally expensive. This paper develops an efficient method for finding a minimum-volume outer matrix ellipsoid under these LMI conditions. Although the corresponding problem for ordinary vector-valued ellipsoids is convex, the objective function in the matrix-valued case is nonconvex. We find that the problem admits an equivalent convex reformulation, which resolves this difficulty. We first eliminate the matrix variables using equivalent LMI transformations and objective monotonicity, leaving one positive scalar variable per matrix ellipsoid in the sum. We then show that the objective becomes convex after a logarithmic change of variables. The resulting problem contains neither LMI constraints nor matrix variables and can therefore be solved efficiently. Numerical experiments confirm that the proposed method is substantially faster than the LMI-based method. An application to data-driven control further illustrates the tradeoff between guaranteed control performance and computation time.