On pseudorandom number generators

Statistical sampling and simulations produced by algorithms require fast random number generators; however, true random number generators are often too slow for the purpose, so pseudorandom number generators are usually more suitable. But choosing and using a pseudorandom number generator is no simple task; most pseudorandom number generators fail statistical tests. Default pseudorandom number generators offered by programming languages usually do not offer sufficient statistical properties. Testing random number generators so as to choose one for a project is essential to know its limitations and decide whether the choice fits the project’s objectives. However, this study presents a reproducible experiment that demonstrates that, despite all the contributions it made when it was first published, the popular NIST SP 800-22 statistical test suite as implemented in the software package is inadequate for testing generators.