COMPUTATIONAL INVESTIGATIONS OF LOW-DISCREPANCY SEQUENCES

The Halton, Sobol, and Faure sequences and the Braaten-Weller construc tion of the generalized Halton sequence are studied in order to assess their applicability for the quasi Monte Carlo integration with large number of variates. A modification of the Halton sequence (the Halton sequence leaped) and a new construction of the generalized Halton sequ ence are suggested for unrestricted number of dimensions and are shown to improve considerably on the original Halton sequence. Problems ass ociated with estimation of the error in quasi Monte Carlo integration and with the selection of test functions are identified. Then an estim ate of the maximum error of the quasi Monte Carlo integration of nine test functions is computed for up to 400 dimensions and is used to eva luate the known generators mentioned above and the two new generators. An empirical formula for the error of the quasi Monte Carlo integrati on is suggested.