Iterative solvers

Most of the solvers provided in Gmm++ come form ITL with slight modifications (gmres has been optimized and adapted for complex matrices). Include the file gmm/gmm_iter_solvers.h to use them.


The iteration object of Gmm++ is a modification of the one in ITL. This is not a template type as in ITL.

The simplest initialization is:

gmm::iteration iter(2.0E-10);

where 2.0E-10 is the (relative) residual to be obtained to have the convergence. Some possibilities:

iter.set_noisy(n) // n = 0 : no output
                  // n = 1 : output of iterations on the standard output
                  // n = 2 : output of iterations and sub-iterations
                  //         on the standard output
                  // ...
iter.get_iteration() // after a computation, gives the number of
                     // iterations made.
iter.converged()     // true if the method converged.
iter.set_maxiter(n)  // Set the maximum of iterations.
                     // A solver stops if the maximum of iteration is
                     // reached, iter.converged() is then false.

Linear solvers

Here is the list of available linear solvers:

gmm::row_matrix< std::vector<double> > A(10, 10);  // The matrix
std::vector<double> B(10); // Right hand side
std::vector<double> X(10); // Unknown
gmm::identity_matrix PS;   // Optional scalar product for cg
gmm::identity_matrix PR;   // Optional preconditioner
gmm::iteration iter(10E-9);// Iteration object with the max residu
size_t restart = 50;       // restart parameter for GMRES

gmm::cg(A, X, B, PS, PR, iter); // Conjugate gradient

gmm::bicgstab(A, X, B, PR, iter); // BICGSTAB BiConjugate Gradient Stabilized

gmm::gmres(A, X, B, PR, restart, iter) // GMRES generalized minimum residual

gmm::qmr(A, X, B, PR, iter) // Quasi-Minimal Residual method.

gmm::least_squares_cg(A, X, B, iter) // unpreconditionned least square CG.

The solver gmm::constrained_cg(A, C, X, B, PS, PR, iter); solve a system with linear constaints, C is a matrix which represents the constraints. But it is still experimental.

(Version 1.7) The solver gmm::bfgs(F, GRAD, X, restart, iter) is a BFGS quasi-Newton algorithm with a Wolfe line search for large scale problems. It minimizes the function F without constraints, be given its gradient GRAD. restart is the max number of stored update vectors.


The following preconditioners, to be used with linear solvers, are available:

gmm::identity_matrix P;   // No preconditioner

gmm::diagonal_precond<matrix_type> P(SM); // diagonal preconditioner

gmm::mr_approx_inverse_precond<matrix_type> P(SM, 10, 10E-17);
                                             // preconditioner based on MR
                                             // iterations

gmm::ildlt_precond<matrix_type> P(SM); // incomplete (level 0) ldlt
                                      // preconditioner. Fast to be
                                      // computed but less efficient than
                                      // gmm::ildltt_precond.

// incomplete ldlt with k fill-in and threshold preconditioner.
// Efficient but could be costly.
gmm::ildltt_precond<matrix_type> P(SM, k, threshold);

gmm::ilu_precond<matrix_type> P(SM);  // incomplete (level 0) ilu
                                      // preconditioner. Very fast to be
                                      // computed but less efficient than
                                      // gmm::ilut_precond.

// incomplete LU with k fill-in and threshold preconditioner.
// Efficient but could be costly.
gmm::ilut_precond<matrix_type> P(SM, k, threshold);

// incomplete LU with k fill-in, threshold and column pivoting preconditioner.
// Try it when ilut encounter too small pivots.
gmm::ilutp_precond<matrix_type> P(SM, k, threshold);

Except ildltt\_precond, all these precontionners come from ITL. ilut_precond has been optimized and simplified and cholesky_precond has been corrected and transformed in an incomplete LDLT preconditionner for stability reasons (similarly, we add choleskyt_precond which is in fact an incomplete LDLT with threshold preconditionner). Of course, ildlt\_precond and ildltt_precond are designed for symmetric real or hermitian complex matrices to be use principaly with cg.

Additive Schwarz method

The additive Schwarz method is a decomposition domain method allowing the resolution of huge linear systems (see [SCHADD] for the principle of the method).

For the moment, the method is not parallelized (this should be done ...). The call is the following:

gmm::sequential_additive_schwarz(A, u, f, P, vB, iter, local_solver, global_solver)

A is the matrix of the linear system. u is the unknown vector. f is the right hand side. P is an eventual preconditioner for the local solver. vB is a vector of rectangular sparse matrices (of type const std::vector<vBMatrix>, where vBMatrix is a sparse matrix type), each of these matrices is of size N \times N_i where N is the size of A and N_i the number of variables in the i^{th} sub-domain ; each column of the matrix is a base vector of the sub-space representing the i^{th} sub-domain. iter is an iteration object. local_solver has to be chosen in the list gmm::using_gmres(), gmm::using_bicgstab(), gmm::using_cg(), gmm::using_qmr() and gmm::using_superlu() if SuperLu is installed. global_solver has to be chosen in the list gmm::using_gmres(), gmm::using_bicgstab(), gmm::using_cg(), gmm::using_qmr().

The test program schwarz_additive.C is the directory tests of GetFEM++ is an example of the resolution with the additive Schwarz method of an elastostatic problem with the use of coarse mesh to make a better preconditioning (i.e. one of the sub-domains represents in fact a coarser mesh).

In the case of multiple solves with the same linear system, it is possible to store the preconditioners or the LU factorisations to save computation time.

A (too) simple program in gmm/gmm_domain_decomp.h allows to build a regular domain decomposition with a certain ratio of overlap. It directly produces the vector of matrices vB for the additive Schwarz method.

Range basis function

The function gmm\_range\_basis(B, columns, EPS=1e-12) defined in gmm/gmm\_range\_basis.h allows to select from the columns of a sparse matrix B a basis of the range of this matrix. The result is returned in columns which should be of type std::set<size_type> and which contains the indices of the selected columns.

The algorithm is specially designed to select independent constraints from a large matrix with linearly dependent columns.

There is four step in the implemented algorithm

  • Elimination of null columns.
  • Selection of a set of already orthogonal columns.
  • Elimination of locally dependent columns by a blockwise Gram-Schmidt algorithm.
  • Computation of vectors of the remaining null space by a global restarted Lanczos algorithm and deduction of some columns to be eliminated.

The algorithm is efficient if after the local Gram-Schmidt algorithm it remains a low dimension null space. The implemented restarted Lanczos algorithm find the null space vectors one by one.

The Global restarted Lanczos algorithm may be improved or replaced by a block Lanczos method (see [ca-re-so1994] for instance), a block Wiedelann method (in order to be parallelized) or simply the computation of more than one vector of the null space at each iteration.