Further analysis of the statistical independence of the NIST SP 800-22 randomness tests

Abstract

In multiple applications, from Statistics to Particle Physics and notably in Cryptography and Computer Security, it is necessary to obtain long sequences of random numbers. In order to verify the properties of these sequences, different statistical tests are commonly applied, which are usually included in the so-called test batteries or test suites. The batteries need to be both effective and efficient. Their effectiveness relates to how well they can spot non-randomness behaviour, the efficiency is related to the computational time they require. It is therefore essential for tests included in batteries to measure their independence features: Test independence is important for good effectiveness, as high correlations between tests could lead to a decreased efficiency (testing for the same features multiple times) and effectiveness (missing an opportunity to test for an orthogonal randomness property when we essentially measure the same twice). Moreover, the related study of test coverage is often based on the assumption that tests are independent. This paper describes a series of experiments aimed at scrutinizing dependencies among the statistical tests in the NIST SP 800-22 suite. In order to do so, sequences of varying lengths from sources of varying entropy have been generated and tested. Afterwards, an inferential study was carried out to find whether significant correlations exist and to present our findings in a statistically sound way.