Lanczos conjugate-gradient method and pseudoinverse computation on indefinite and singular systems (Q933807): Difference between revisions

The paper is a continuation of the author's previous paper [ibid. 125, No.~3, 523--541 (2005; Zbl 1079.90162)]. He discusses iterative methods for solving systems of linear equations with symmetric matrices, indefinite and possibly singular. Some theoretical properties of the conjugate gradient method are extended. The proposed algorithm, called FLR algorithm, is a general planar method for solving linear systems when the matrix is indefinite, i.e. it avoids the possible pivot breakdown of the conjugate gradient method in the indefinite case, by introducing \(2\times 2\) pivot elements. Also, the author highlights some new properties of the FLR algorithm, in the case when the matrix of the system is singular. It is proved that the sequence of orthogonal residuals from the FLR algorithm yields the sequence of orthogonal vectors of the Lanczos process, for a suitable choise of a parameter. The author concludes that the FLR algorithm seems to be a general tool for the solution of symmetric linear systems i.e. for the search of stationary points of quadratic forms in unconstrained optimization frameworks. In addition, the approximation of the Moore-Penrose pseudoinverse of the matrix of the system provided by the FLR algorithm may be used for the construction of preconditioners.