Gambling tests for pseudorandom number generators (Q5938387)

An important test for randomness of a sequence of pseudo-random numbers is the serial test what, applied to \(s\)-tuples, consists in a partition of the unit cube \([0,1)^s\) in \(m\) nonempty subsets. Although for the cyclic variant, the so-called overlapping serial test, the covariance matrix of the frequency counting vector is easier to compute than for the non-overlapping test, for greater \(s\) and \(m=\tilde m^s\), \(\tilde m\) fixed, the implementation is very time-consuming and almost impossible. The author proposes now the following gambling test: Some feature of a random number will be compared with tossing an ideal coin. Now a gambler can bet whether among \(\lambda\) random numbers at least \(\kappa\) times ``head'' will occur, he can bet on ``tails'' or skip betting. The counting vector is 3-dimensional and its covariance matrix can be computed. This leads to a considerable dimension reduction. \(\lambda=52\) and \(\kappa=26\) are chosen. The usual formed quadratic form is a special case of the \(\phi\)-divergence investigated by \textit{F. Liese} and \textit{I. Vajda} [Convex statistical distances. Teubner (1987; Zbl 0656.62004)], whose parameter as a design parameter of the test can be varied. It is mentioned that the new kind of test at the twisted GFSR generators of \textit{M. Matsumoto} and \textit{Y. Kurita} [ACM Trans. Model. Comput. Simul. 2, No.~3, 179-194 (1992; Zbl 0849.94014); ibid. 4, No.~3, 254-266 (1994; Zbl 0849.94015)] could detect deviations from equidistribution in higher dimensions.