Spectral Analysis of the MIXMAX Random Number Generators

Abstract

We study the lattice structure of random number generators of the MIXMAX family, a class of matrix linear congruential generators that produces a vector of random numbers at each step. The design of these generators was inspired by Kolmogorov K-systems over the unit torus in the real space, for which the transition function is measure preserving and produces a chaotic behavior. In actual implementations, however, the state space is a finite set of rational vectors, and the MIXMAX has a lattice structure just like linear congruential and multiple recursive generators. Its matrix entries were also selected in a special way to allow a fast implementation, and this has an impact on the lattice structure. We study this lattice structure for vectors of successive and nonsuccessive output values in various dimensions. We show in particular that for coordinates at specific lags not too far apart, in three dimensions, or if we construct points of k+2 or more successive values from the beginning of an output vector of size k, all the nonzero points lie in only two hyperplanes. This is reminiscent of the behavior of lagged-Fibonacci and add-with-carry/subtract-with-borrow generators. And even if we skip the output coordinates involved in this bad structure, other highly structured projections often remain, depending on the choice of parameters. We show that empirical statistical tests can easily detect this structure.