Preconditioned complex generalized minimal residual algorithm for dense algebraic variational equations in quantum reactive scattering

Variational basis-set formulations of the quantum mechanical reactive scattering problem lead to large, dense sets of equations. In previous work, we showed that the generalized minimal residual (GMRes) algorithm is sometimes competitive in terms of computer time with direct methods for these dense matrices, even when complex-valued boundary conditions are used, leading to non-Hermitian matrices. This paper presents a preconditioning scheme to accelerate convergence and improve performance. We block the potential energy coupling into a series of distortion blocks, and we employ the outgoing wave variational principle with nonorthogonal basis functions, including both dynamically adapted Green’s functions for the distortion blocks and also square integrable functions. The coefficient matrix of the resulting linear system couples the blocks. We have found that preconditioners formed from diagonal blocks of the coefficient matrix corresponding to the distortion blocks and vibrational blocks are effective at accelerating the iterative method in every test case, by factors of 2.9–20, with an average speedup of a factor of 6.5. The storage requirements and computational efficiency of the new scheme compare favorably to those for preconditioners based on banded matrices of variable bandwidth. The new preconditioners yield converged transition probabilities in less computer time than a direct solver even in cases which do not converge in a reasonable amount of time without preconditioning, and the average speedup compared to the direct solution is a factor of 7.6.