Title:Generalized Golub-Kahan bidiagonalization for generalized saddle point systems

Abstract:We consider the iterative solution of generalized saddle point systems. When the right bottom block is zero, Arioli [SIAM J. Matrix Anal. Appl., 34 (2013), pp. 571--592] proposed a CRAIG algorithm based on generalized Golub-Kahan Bidiagonalization (GKB) for the augmented systems with the leading block being symmetric and positive definite (SPD), and then Dumitrasc et al. [SIAM J. Matrix Anal. Appl., 46 (2025), pp. 370--392] extended the GKB for the case where the symmetry condition of the leading block no longer holds and then proposed nonsymmetric version of the CRAIG (nsCRAIG) algorithm. The CRAIG and nsCRAIG algorithms are theoretically equivalent to the Schur complement reduction (SCR) methods where the Conjugate Gradient (CG) method and the Full Orthogonalization Method (FOM) are applied to the associated Schur-complement equation, respectively. We extend the GKB and its nonsymmetric counterpart used separately in CRAIG and nsCRAIG algorithms for the case where the right bottom block of saddle point system is nonzero. On this basis, we propose CRAIG and nsCRAIG algorithms for the solution of the generalized saddle point problems with the leading block being SPD and nonsymmetric positive definite (NSPD), respectively. They are also theoretically equivalent to the SCR methods with inner CG and FOM iterations for the associated Schur-complement equation, respectively. Moreover, we give algorithm steps of the two new solvers and propose appropriate stopping criteria based on an estimate of the energy norm for the error and the residual norm. Numerical comparison with MINRES or GMRES highlights the advantages of our proposed strategies regarding its high computational efficiency and/or low memory requirements and the associated implications.