Hybrid methods based on LCG and GMRES (Q291394)

It is first proved that if no breakdown occurs, then the left conjugate gradient (LCG) method [\textit{J. Y. Yuan} et al., BIT 44, No. 1, 189--207 (2004; Zbl 1052.65026)] is in exact arithmetic the same as the full orthogonalization method (FOM) (i.e. Arnoldi) using LU decomposition of the Hessenberg matrix without pivoting. This links LCG and FOM (which are Galerkin based) just like how generalized conjugate gradient (GCR) and GMRES (based on residual norm minimization) are linked. A hybrid algorithm is proposed in which the inner iteration is done by GMRES and the outer by LCG (called LCGR). If the inner residual minimization is maintained in the outer iteration like de Sturler did for GMRES [\textit{E. de Sturler}, J. Comput. Appl. Math. 67, No. 1, 15--41 (1996; Zbl 0854.65026)], the modification of LCGR is called LCGO. Both algorithms are tested numerically and compared with several alternatives to show that the are competitive. They may suffer from breakdown though.