Modified parameterized inexact Uzawa method for singular saddle-point problems (Q297552)

The following system of linear equations, called a saddle-point linear system, is considered: \(Mx=b\), where \(M\) has the block form \(M=\left(\begin{smallmatrix} A & B\\ -B^T & 0 \end{smallmatrix}\right)\), where \(A\) is a symmetric positive definite matrix, \(B\) is a rectangular matrix with at least as many columns as there are rows and \(b\) is a given input vector. This equation models a variety of problems arising for instance, from constrained optimization and electronic networks, among others. When the columns of \(B\) are linearly independent, \(M\) is invertible. In the case when \(B\) is rank-deficient, the system may be inconsistent. A modified parametrized inexact Usawa method, to solve the system \(Mx=b\), is proposed. Theoretical results demonstrate that in the presence of certain assumptions on the choice of the parameters, the iterative sequence generated by the new method converges to the least squares solution of the equation. It is shown that the convergence is independent of the choice of the initial guess. The corresponding quasi-optimal parameter and the convergence factor are determined. Numerical examples are presented as illustrations.