7  Concluding Discussion

In this paper a new numerical method for the solution of the Helmholtz eigenvalue problem is described. Essentially, the method is derived from a polynomial approximation in the wavenumber k of the matrix that arises when the boundary element method is applied to the Helmholtz equation. The resulting matrix polynomial eigenvalue problem is recast as a generalised eigenvalue problem with companion matrix structure (11), which is solved through the use of standard methods. An implementation of a particular straightforward form of the method is described and results from the application of the method to test problems are presented. The results show the method working effectively, with clear convergence to the exact solution as the number of elements and the accuracy of the polynomial interpolation increases.
Inspection of the results reveals that the error is characterised by the sum of a component resulting from the approximation of the boundary functions and a component from the polynomial approximation of the matrix. The results indicate that the approximation to the boundary function transmits an error of O(n-2), except for the computation of the eigenvalues of Mkt + [1/2] I on the square, where the error seems to be O(n-1). This reduction in the order of convergence could be attributed to the lack of smoothness of the solution of the indirect formulation (6). The error with respect to the polynomial approximation of the matrix components is more difficult to characterise. It seems that the error is governed by the proximity of the eigenfrequency to the nearest interpolation point as well as the general accuracy of the polynomial approximation.
No particular effort was made to optimise the efficiency of the overall method. Clearly, there is some scope for such improvements. For example the number of quadrature points used to compute the matrix components could be reduced. The efficiency could be improved by adjusting the polynomial approximation method so that the error it contributes is approximately the same as the error resulting from the collocation method. An alternative to the QZ algorithm for solving equation (11), taking advantage of its structure, could also be of benefit. In the tests constant elements have been used, in general higher order elements would result in a more efficient method.
In practical situations, a set of eigenfrequencies and corresponding mode shapes will be required. The most straightforward way of applying the method is to choose a degree for the polynomial approximation and the width of subinterval [kA, kB], then to step through the intervals, computing the successive solutions. It is difficult to compare this method with the existing methods for computing Helmholtz eigenvalues outlined in the introduction on the grounds of computational efficiency. However, the method described in this paper is reasonably straightforward to implement and could prove a useful addition to existing boundary element software libraries.
Acknowledgement. The work in this paper was funded by a SERC/MOD grant (GR/G01416). The Cray Supercomputer at Rutherford Appleton Laboratory was used to obtain the results from the test problems (Supercomputing Grant GR/F61639). The authors are grateful to the reviewers for their helpful suggestions.