A framework for the construction of preconditioners for systems of PDE
- Uppsala Univ. (Sweden)
The authors consider the solution of systems of partial differential equations (PDE) in 2D or 3D using preconditioned CG-like iterative methods. The PDE is discretized using a finite difference scheme with arbitrary order of accuracy. The arising sparse and highly structured system of equations is preconditioned using a discretization of a modified PDE, possibly exploiting a different discretization stencil. The preconditioner corresponds to a separable problem, and the discretization in one space direction is constructed so that the corresponding matrix is diagonalized by a unitary transformation. If this transformation is computable using a fast O(n log{sub 2} n) algorithm, the resulting preconditioner solve is of the same complexity. Also, since the preconditioner solves are based on a dimensional splitting, the intrinsic parallelism is good. Different choices of the unitary transformation are considered, e.g., the discrete Fourier transform, sine transform, and modified sine transform. The preconditioners fully exploit the structure of the original problem, and it is shown how to compute the parameters describing them subject to different optimality constraints. Some of these results recover results derived by e.g. R. Chan, T. Chan, and E. Tyrtyshnikov, but here they are stated in a {open_quotes}PDE context{close_quotes}. Numerical experiments where different preconditioners are exploited are presented. Primarily, high-order accurate discretizations for first-order PDE problems are studied, but also second-order derivatives are considered. The results indicate that utilizing preconditioners based on fast solvers for modified PDE problems yields good solution algorithms. These results extend previously derived theoretical and numerical results for second-order approximations for first-order PDE, exploiting preconditioners based on fast Fourier transforms.
- Research Organization:
- Front Range Scientific Computations, Inc., Boulder, CO (United States); USDOE, Washington, DC (United States); National Science Foundation, Washington, DC (United States)
- OSTI ID:
- 219609
- Report Number(s):
- CONF-9404305--Vol.2; ON: DE96005736
- Country of Publication:
- United States
- Language:
- English
Similar Records
FFT-based preconditioners for Toeplitz-Block least square problems
An Algebraic Multilevel Preconditioner with Low-Rank Corrections for Sparse Symmetric Matrices