A new empirical test for parallel pseudo-random number generators (Q5938373): Difference between revisions

For parallel simulations, parallel random number generators (PRNG) are needed. While single random number sequences (RNS) often are checked whether they represent pure randomness, the unwanted presence of correlation between the individual RNS produced by PRNG rarely had been investigated. Some known whole-period number-theoretic tests don't give information about the correlation behavior of subsets so that it is necessary to consider statistical tests. From two (or more) RNS, a new binary sequence \((Y_j)_{j=1,\ldots,l}\) is formed with \(Y_j=1\) if an arbitrary fixed bit pattern of length \(s\) coincides in the two RNS, \(Y_j=0\) otherwise. Based on the run length distribution of 1's in the \(Y_j\)'s given by \textit{O. E. Percus} [Testing for correlations between independent parallel random number generators, New York University (1995)], a goodness-of-fit test is constructed as follows: For a sequence of the PRNG of sufficient length, a \(\chi^2\) value is computed, finally these values are tested by the Kolmogorov-Smirnov criterion. The structure of the test program written in C is described in greater detail. The test results are less encouraging: Even famous PRNG as R250, nested Weyl, shuffled Weyl and inversive congruential show a poor behavior with respect to uncorrelatedness. As a sign of hope, the combination of two different linear congruential generators as organized by \textit{P. L'Ecuyer} and \textit{S. Tezuka} [Math. Comput. 57, No.~196, 735-746 (1991; Zbl 0748.65007)] seems to be acceptable. In a longer part of Section 2, it is not clear what \(p=P\) \((Y_j=1)\) is. Formula (8) is doubtful: What is \(j\)? In formula (11), one \(l\) is superfluous, another is missing. The symbols \(p\) and \(q\) are used to denote two completely different magnitudes, respectively.