Re-seeding invalidates tests of random number generators

Kim et al. [C. Kim, G.H. Choe, D.H. Kim, Test of randomness by the gambler’s ruin algorithm, Applied Mathematics and Computation 199 (2008) 195–210] recently presented a test of random number generators based on the gambler’s ruin problem and concluded that several generators, including the widely used Mersenne Twister, have hidden defects. We show here that the test by Kim et al. suffers from a subtle, but consequential error: re-seeding the pseudorandom number generator with a fixed seed for each starting point of the gambler’s ruin process induces a random walk of the test statistic as a function of the starting point. The data presented by Kim et al. are thus individual realizations of a random walk and not suited to judge the quality of pseudorandom number generators. When generating or analyzing the gambler’s ruin data properly, we do not find any evidence for weaknesses of the Mersenne Twister and other widely used random number generators.