SIAM Journal on Numerical Analysis · 1999 · 84 citations · 9 references
ComputingComputational EngineeringComputational GeometryFinite Element AnalysisFinite Element MethodsBoundary ConditionsBoundary Element MethodsConvergence AnalysisFinite Element DiscretizationFinite Element FormulationNonlinear Finite Element AnalysisBasis FunctionsBilinear FormsBoundary Element AnalysisBoundary Integral ConditionsBoundary Integral EquationsGalerkin DiscretizationPseudodifferential OperatorsStiffness Matrix Formulation
Consider a boundary integral operator on a bounded,d-dimensional surface in $\mbox{\smallBbb R}^{d+1}$. Suppose that the operator is a pseudodifferential operator of order $2m,$ $m\in\mbox{\smallBbb R},$ and that the associated bilinear form is symmetric and positive-definite. (The surface may be open or closed, and m may be positive or negative.) Let B denote the stiffness matrix arising from a Galerkin boundary element method with standard nodal basis functions. If local mesh refinement is used, then the partition may contain elements of very widely differing sizes, and consequently B may be very badly conditioned. In fact, if the elements are nondegenerate and 2|m| < d, then the $\ell_2$ condition number of B satisfies {\rm cond}({\mbox {\bf {\em B}}})\leq C N^{2|m|/d} \left(\frac{h_{\rm max}}{h_{\rm min}}\right)^{d-2m}, where hmax and h min are the sizes of the largest and smallest elements in the partition, and N is the number of degrees of freedom. However, if B is preconditioned using a simple diagonal scaling, then the condition number is reduced to ${\cal O}(N^{2|m|/d})$. That is, diagonal scaling restores the condition number of the linear system to the same order of growth as that for a uniform partition. The growth in the critical case 2|m|=d is worse by only a logarithmic factor.
9