A high-performance method for the biharmonic dirichlet problem on rectangles

Co-authored by: C. C. Christara J. Zhang

We propose a fast solver for the linear system resulting from the application of a sixth-order Bi-Quartic Spline Collocation method to the biharmonic Dirichlet problem on a rectangle. The fast solver is based on Fast Fourier Transforms (FFTs) and preconditioned GMRES (PGMRES) and has complexity O(n2 log(n)) on an n n uniform partition. The FFTs are applied to two auxiliary problems with dierent boundary conditions on the two vertical sides of the domain, while the PGMRES is applied to a problem related to the two vertical boundaries. We show that the number of PGMRES iterations required to reduce the relative residual to a certain tolerance is independent of the gridsize n. Numerical experiments verify the eectiveness of the solver. The biharmonic Dirichlet problem on a rectangle is a two-dimensional fourth-order partial dierential equation (PDE) problem given by
2

u = g

in , on , on ,

(1) (2) (3)

u = g1 u = g2 n

where denotes the Laplacian, is a rectangular domain, is the boundary of , /n is the outer normal derivative on , and g, g1 and g2 are given functions of variables x and y. This problem arises in several scientic, engineering and industrial applications. For example, the biharmonic problem for the bending of a clamped rectangular plate has been considered as one of the classical problems in the Theory of Elasticity [4]. Various numerical methods have been developed for problem (1)-(3). These methods consist of a discretization strategy that converts the continuous problem into a discrete set of algebraic equations, and an associated solver for the resulting linear system. The eectiveness of a numerical method for a problem such as (1)-(3) depends primarily on the accuracy of the discretization strategy and the eciency of the solver. We rst introduce a Bi-Quartic Spline Collocation (BQSC) discretization method for general two-dimensional linear fourth-order Boundary Value Problems involving variable coecients and any derivatives of the unknown function up to fourth order in the PDE operator and up to third order in the boundary conditions operator. The discretization error for this method turns out to be sixth order locally on the gridpoints and midpoints of a uniform rectangular partition and fth order globally in the uniform norm.
1 2

Department of Computer Science, University of Toronto Department of Computer Science, University of Toronto and Scotiabank, Toronto

We then present two instances of biharmonic problems: (A) the problem (1)-(3), with g = g1 = g2 = 0 on the unit square; (B) the problem consisting of PDE (1) with boundary conditions u = 0, 2u x2 u u = 0, y = 0 = 0 on x = 0, x = 1, on y = 0, y = 1, for 0 y 1, for 0 x 1. (4) (5)

Notice that Problems (A) and (B) dier only in the boundary conditions along the vertical boundaries. Our goal is the ecient solution of Problem (A). The main ingredients of the solution technique are a fast solver for Problem (B) and a preconditioned iterative solution of a Schur-complement problem related to the two vertical boundaries. Below, we summarize the main results of this paper. We give the tensor product form of the matrices arising from the application of the BQSC method to Problems (A) and (B). We develop explicit formulae for the eigenvalues and eigenvectors of some of the constituent matrices. Based on these formulae, an FFT (direct) solver with complexity O(n2 log n) is developed for the matrix of Problem (B). A solution technique for the matrix arising from the application of the BQSC method to Problem (A) is then developed, which consists of two applications of the FFT solver to Problem (B) and the solution of an one-dimensional problem along the two opposite boundaries of the domain, which can be viewed as a Schur-complement problem. This problem is solved by GMRES and appropriate preconditioners. The analysis of PGMRES for the matrix A arising from the one-dimensional Schur-complement T +A problem is based on nding a lower bound above 0 for the minimum eigenvalue of A 2 and an upper bound for the maximum eigenvalue of AT A. Both bounds are shown to be independent of n, the size of the partition in one dimension. Using these bounds and a well-known result for the convergence of the GMRES method ([5], pages 134 and 193), we show that the PGMRES method converges to a specied tolerance in a number of iterations independent of n. With the cost of each iteration being O(n2 ), the complete solver for Problem (A) runs at O(n2 log n) ops. The solver can be easily extended to problem (1)-(3) with non-homogeneous boundary conditions. Experimental results demonstrate the accuracy of the BQSC method, and verify the theoretical analysis of the convergence of PGMRES and the eciency of the algorithm applied to problem (1)-(3). Among the fast solvers for the biharmonic Dirichlet problem found in the literature, the one that is closest to ours is the method in [2]. However, in [2], the discretization is obtained by standard second order nite dierences, the arising matrices are symmetric and the PCG method is used for the solution of the problem related to the two opposite boundaries. Consequently, the convergence analysis is simpler. Some other relevant methods found in the literature are the methods in [3, 1] which also solve a Schur complement problem related to the two opposite 2

boundaries. However, the methods in [3, 1] reduce the biharmonic to a coupled system of second order PDEs, and apply fourth order discretizations to it, while we apply a sixth order discretization directly to the fourth order derivatives.

References
[1] B. Bialecki. A fast solver for the orthogonal spline collocation solution of the biharmonic Dirichlet problem on rectangles. Journal of Computational Physics, 191:601621, 2003. [2] P. Bjrstad. Fast numerical solution of the biharmonic Dirichlet problem on rectangles. SIAM Journal on Numerical Analysis, 20:5971, 1983. [3] D. B. Knudson. A piecewise Hermite bicubic nite element Galerkin method for the biharmonic Dirichlet problem. PhD thesis, Colorado School of Mines, Golden, Colorado, U.S.A., 1997. [4] V. V. Meleshko. Selected topics in the history of the two-dimensional biharmonic problem. Applied Mechanics Reviews, 56(1):3385, 2003. [5] Y. Saad. Iterative methods for sparse linear systems. PWS, 1996.