No label defined (Q4511248)

A new method is presented in this stimulating paper, in order to solve singular consistent or inconsistent linear systems \(Ax=b\). No restriction is placed on \(A\), i.e. it is not necessarily Hermitian or Hermitian positive definite. Nor is any condition placed upon the spectrum of \(A\). As to its index, i.e. the size of the largest Jordan block corresponding to the eigenvalue \(0\), it is only required that it be known in advance. The method calculates the solution provided by the Drazin generalized inverse of \(A\) by means of a Krylov subspace method, which is based on a biorthogonalization process that computes simultaneously two Krylov subspaces, one for \(A\) and another for \(A^*\), as in the bi-conjugate gradient method, starting with two arbitrary initial vectors. That the method has the property of finite termination follows from a more general result recently obtained by the first author. The paper includes a detailed error analysis of the method for the cases when the system is singular inconsistent Hermitian and when it is a general singular non-Hermitian. The method is illustrated with an application to a standard discretization of the Poisson equation, which produces a singular non-symmetric matrix \(A\) with index \(1\). A column in \(A\) is perturbed by a vector in its null space and the result is taken as \(b\) in the system. In dimension \(4096\) and with a standard stopping criterium, a satisfactory solution is attained after 230 iterations. It is interesting to point out in connection with this example that the Drazin inverse vanishes on the null space of \(A\), which is not the case for the Moore-Penrose pseudoinverse if that space and the range space are not orthogonal, something likely to happen in the non-Hermitian case. Thus the example illustrates at the same time the accuracy of the method proposed in the paper and the usefulness of the Drazin pseudoinverse for non-symmetric inconsistent systems that may arise in common problems.