Testing the tests: Using random number generators to improve empirical tests

Abstract

The implementer of an empirical test for random number generators is faced with some difficult problems, especially if the test is based on a statistic which is known only approximately: How can the test be tested? How can the approximation be improved? When is it good enough? A number of principles can be applied to these problems. These principles are illustrated using implementations of the overlapping serial "Monkey" tests of Marsaglia and Zaman. © Springer-Verlag Berlin Heidelberg 2009.