A theoretical and empirical comparison of mainframe, microcomputer, and pocket calculator pseudorandom number generators

Abstract

This article presents an extensive theoretical and empirical analysis of the pseudorandom number generators provided by subroutine libraries (NAG, CERN, IMSL, and ESSL), statistical and simulation software packages (GLIM, SAS, SPSS, DATASIM, ESSP, and LLRANDOMII), builtin functions of programming languages (APL, Turbo Pascal, Advanced BASIC, GW-BASIC, and QBASIC), and autoimplemented algorithms (Fishman & Moore, 1986; Wichmann & Hill, 1982; Van Es, Gill, & Van Putten, 1983). On the basis of these analyses, it is concluded that most of the built-in functions of the software packages that were tested can be used safely. In addition, it is concluded that the Wichmann and Hill algorithm is a good choice if only single-precision arithmetic is available, but that a prime-modulus multiplicative congruential generator with modulus 231 −1 and multiplier 16,807 is a better choice if double-precision arithmetic is available, and that the same generator with multiplier 62,089,911 or 742,938,285 is the best choice if extended-precision arithmetic is available. A Turbo Pascal and a VS FORTRAN program for the latter are given in the Appendixes.