Variable additive preconditioning procedures (Q1293472): Difference between revisions

The authors propose four new (conjugate) gradient-like algorithms, in order to solve large scale linear problems with a symmetric positive definite matrix. Their common features are that all of them are based on additive symmetric positive definite preconditioners and that the parameter coefficients in the linear combination producing the inverse of the preconditioner vary at each iteration. This variation is determined by solving a minimization problem. For all the algorithms it is proved that they converge at least as fast as the Richardson iterative method with the corresponding additive preconditioner with optimal scaling factors. The algorithms are numerically tested with the Poisson equation and with a magnetic field problem, using finite element discretizations. The results are compared with those obtained by applying the preconditioned conjugate gradient (PCG) method. For all the methods, the additive preconditioners are produced by a non-overlapping domain decomposition. From the comparison it follows that two of the algorithms have a similar computational cost to that of PCG, when PCG is applied with optimal factors, excluding the cost of obtainig them. If standard factors are used in the PCG method, then those two methods produce uniformly better results than PCG, for grids with the number of nodes varying from the few hundreds to the hundreds of thousands. An open problem regarding a low cost implementation of an optimization problem is stated, whose satisfactory solution would increase the interest that the present algorithms deserve.