Balancing an arbitrary matrix using computations with Stieltjes matrices (Q2563382)

This paper considers the problem of matrix scaling of the type \(CAC^{-1}\), where \(A\) is a square matrix, \(C\) is diagonal, and \(C\) minimizes the quantity \(\Phi (d) = \| DAD^{-1}\|_E\). Here \(d = (d_1,\dots,d_n)^T\) are the diagonal entries of matrix \(D\). Instead of minimizing \(\Phi(d)\) directly the author proposes to solve the nonlinear system \(D \text{grad\,}\Phi(d) = 0\) by Newton's iterations. In this way the linear system solved at each iteration has a Stieltjes matrix which is weakly diagonally dominant. This improves the convergence significantly. Convergence results are presented for the \(2\times 2\) case, and numerical examples show the effectiveness of the approach.