New tests of random numbers for simulations in physical systems

Abstract

The aim of this Thesis is to present five new tests for random numbers, which are widely used {\em e.g.} in computer simulations in physics applications. The first two tests, the cluster test and the autocorrelation test, are based on analogies to the two-dimensional Ising model. The next two, the random walk test and the $n$-block test, are based on studies of random walks, and the condition number test presented last uses some results of Gaussian distributed random matrices. Studies with several commonly used pseudorandom number generators reveal that the cluster test is particularly powerful in finding periodic correlations on bit level, and that the autocorrelation test, the random walk test, and the $n$-block test are very effective in detecting short-ranged correlations. The results of the condition number test are mostly inconclusive, however. By means of the tests presented in this work, two important results are found. First, we show quantitatively that the reason for erroneous results in some recent high precision Monte Carlo simulations for some commonly used pseudorandom number generators are the so called triple correlations in the sequences. Then, we show that the properties of such a sequence may be considerably improved, if only a certain portion of it is used.