A comparison of some GMRES-like methods (Q1183135)

The paper investigates a new iterative method for solving a linear system with nonsymmetric matrix proposed by \textit{T. Eirola} and \textit{O. Nevanlinna} [(EN-method); ibid. 121, 511--520 (1989; Zbl 0683.65018)] and compares this method with the GMRES method. This investigation is performed by both a very careful analysis of numerical experiments and a theoretical study. For example, numerical experiments reveal a relation between the convergence behaviour of EN and GMRES which is then confirmed by the theoretical result -- the inequality \(\| r_k^{\text{EN}}\|_2 \geq \| r_{2k}^{\text{GMRES}}\|_2\) for the residuals of EN and GMRES methods. Later some specific properties of the EN method are discussed, e.g. it is shown that the convergence behaviour of the EN method and also its stability depends on scalar scaling of the solved system. As a consequence, a new scaling invariant version of the EN method is suggested which removes these drawbacks and which seems to be more efficient than the original EN method. Further, another formulation of the GMRES method equivalent to the EN method and the relations between the above methods and \textit{C. G. Broyden's} iterative method [Comput. J. 12, 94--99 (1969; Zbl 0164.45101)] are discussed. The numerical results show that EN may take less work than GMRES, but the theoretical investigation shows that the efficiency of EN may be only slightly better than that of GMRES.