Reduced storage matrix methods in stiff ODE systems (Q1123558): Difference between revisions

In a previous paper [SIAM J. Numer. Anal. 23, 610-638 (1986; Zbl 0615.65078)] the authors considered the use of Krylov-subspace projection methods in solving large stiff systems of ODE's that arise from PDE systems by the method of lines. In this long paper (50 pages) the investigation is continued with particular emphasis on the importance of preconditioning the linear systems solved by the Krylov methods. Section 2 summarizes the Newton and Newton like iteration for nonlinear systems of equations. The Krylov subspace projection methods and then three iterative linear solvers: Arnoldi's algorithm, the generalized minimum residual method (GMRES) and the conjugate gradient (CG) method are defined. All that is required is the ability to perform coefficient- matrix vector multiplications. Useful theorems and examples are presented. In section 3 finite difference versions of Arnoldi's algorithm and GMRES are summarized and a similar theory for a finite-difference version of the CG method is presented (in the main theorem 3.6). The core of this paper are sections 4 and 5. Scaled preconditioned methods are discussed in general and specific preconditioners suitable for ODE systems arising from certain PDE system are given. The authors construct several preconditioner combinations for reactor- diffusion problems, using the reactor and transport operators individually or in combination as in operator splitting, and using grouping of blocks of Jacobian elements involved to achieve both storage and computational economics. A numerical test with a solver LSODPK (based on LSODE and previous results) is described. All tests were done on Cray- 1 and various cumulative performance statistics are given.