A Frontal Computation Scheme for the Schur Algorithm to Efficiently Solve Large Boundary-El.pdf

A Frontal Computation Scheme for the Schur Algorithm to Efficiently Solve Large Boundary-Element Problems A.J. van Genderen, N.P. van der Meijs Delft University of Technology Department of Electrical Engineering Mekelweg 4, 2628 CD Delft, The Netherlands ABSTRACT The Schur algorithm provides an approximate inverse for partially specified (on a staircase band) matrices. In this paper, we develop a frontal computation scheme for the Schur algorithm that achieves an O(b 2 ) space bound, where b is the width of the staircase band. This property allows certain classes of boundary element problems, such as for example occur with VLSI capacitance modeling, to be solved in linear time and using constant memory. We describe such an application and give experimental results that illustrate the efficiency and practicality of the new algorithm. 1. Introduction The boundary-element method is often applied to solve scientific and engineering problems that are expressed in the form of an integral equation. As an example, the capacitance problem to compute the charge distribution over the surfaces of a set of conductors, given a potential distribution, may be expressed as [1] φ(p) = ? ∫ G(p, q) ρ(q) dq, (1) where φ(p) is the potential at a point p, ρ(q) is the charge distribution function on the surfaces of the conductors, G(p, q) is the Green’s function for the domain in which the conductors are located, and ? is the domain consisting of the conductor surfaces. To solve the above integral equation, the boundary- element method discretizes the surfaces of the conductors into elements and sets up a set of simultaneous equations φ h = G ρ h , (2) where φ h is the vector that collects the potentials of the elements, ρ h is the vector that collects the charges on the elements, and G is an elastance matrix that describes the Green’s influences between the different element pairs. From the matrix G in (2), the short-circuit capacitance matrix C s is found as C s = F T G ?1 F, (3) where F