Entropy of Pseudo Random Number Generators

Since the work of Ferrenberg et al [PRL 69, (1992)] some pseudo random number generators are known to yield wrong results in cluster Monte Carlo simulations. In this contribution the mechanism behind this failure is revealed. Almost all random number generators calculate a new pseudo random number $x_i$ from preceeding values, $x_i = f(x_{i-1}, x_{i-2},..., x_{i-q})$. Failure of these generators in cluster Monte Carlo simulations and related experiments can be attributed to the low entropy of the production rule $f()$ conditioned on the statistics of the input values $x_{i-1},...,x_{i-q}$. Being a measure only of the arithmetic operations in the generator rule, the conditional entropy is independent of the lag in the recurrence or the period of the sequence. In that sense it measures a more profound quality of a random number generator than empirical tests with their limited horizon.