Concluding Remarks

The main goal in the development of the new algorithm for adaptive multidimensional integration was to improve the reliability of some previous algorithms. Tests [9] of a FORTRAN implementation, DCUHRE, have shown that this goal has been achieved. The new algorithm has been structured to allow its efficient implementation on shared memory parallel computers and good speedups are demonstrated in the test report [9]. In each dimension more than one integration rule is made available to the user because the relative performance of different rules may differ from one problem to another. The other additional feature of the new algorithm are the modifications that allow its application to a vector of similar integrals over a common integration region. This feature should be used with caution, because the algorithm chooses a subdivision of the integration region that may be refined according to the behavior of particular integrands which have certain local regions of difficulty. If the integrands in a vector of integrands are significantly different, the vector should probably be split into smaller vectors of integrands and the algorithm applied separately to each subvector. For integrands that have enough similarity, the new algorithm may save both time and space because of the common subdivision, and will also allow for additional saving in time on many types of parallel computers where the integrand function evaluations can be done in parallel.


Alan C Genz
Tue May 11 09:59:26 PDT 1999