[Python-ideas] Should our default random number generator be secure?

Sturla Molden sturla.molden at gmail.com
Tue Sep 15 14:09:07 CEST 2015

On 15/09/15 10:45, M.-A. Lemburg wrote:

> k-dim equidistribution is a way to measure how well your
> PRNG behaves, because it describes in analytical terms how
> far you can get with increasing the linear complexity of your
> RNG output.

Yes and no. Conceptually it means that k subsequent samples will have 
exactly zero correlation. But any PRNG that produces detectable 
correlation between samples 623 steps apart is junk anyway. The MT have 
proven equidistribution for k=623, but many have measured 
equidistribution for far longer periods than that. Numerical 
computations are subject to rounding error and truncation error whatever 
you do. The question is whether the deviation from k-dim 
equidistribution will show up in your simulation result or drown in the 
error terms.


More information about the Python-ideas mailing list