Minimizing the condition number of a positive definite matrix by completion (Q1347061)

Let \(W = W(X)\) denote an Hermitian matrix with blocks \(W_{ij}\), \(i = 1,2\). Three of the four blocks are given, viz. \(W_{11} = A\), an Hermitian positive definite \(n \times n\) matrix and \(W_{21} = W_{12}^ H = B\), a \(p \times n\) matrix. The problem is to complete the matrix with \(W_{22} = X\), a \(p \times p\) matrix such that \(\kappa (W)\), the spectral condition number of \(W\), is minimized and such that \(W(X)\) is positive definite. This problem arises in the construction of optimal feedback controlers and optimal preconditioners for the iterative solution of linear systems. The problem is solved by finding the solution of the convex problem \(\min f(t)\) with \(f(t) = | t^ 2 B^ HB + (I - tA)^ 2 |\). If \(f^* = f(t^*)\) is this minimum, then all solutions can be characterized as \(R \leq X \leq S\) where \(R\) and \(S\) are matrices defined in terms of \(A, B, t^*\) and \(f^*\). Similar completion problems have been treated before. Previously \textit{C. Davis}, \textit{W. M. Kahan} and \textit{H. F. Weinberger} [SIAM J. Numer. Anal. 19, 445-469 (1982; Zbl 0491.47003)] minimized the norm of a matrix by completion and \textit{E. Elsner}, \textit{C. He} and \textit{V. Mehrmann} [Van Dooren et al. (eds.) Linear Algebra for Control Theory, (IMA Vol. in Math. Appl. 62, 75-86 (1994; Zbl 0812.93039)] minimized the norm of the inverse by completion.