On Apr 7, 2014 3:59 AM, "Yaroslav Halchenko" <li...@onerussian.com> wrote: > so I would assume that the devil is indeed in R post-processing and would look > into it (if/when get a chance).
The devil here is the pigeon and the holes problem. Mersenne Twister generates random integers in a certain range. The output is guaranteed to be deterministic, uniform, and reproducible. But when you want to cast those m possible input in n possible outputs, you need to do magic (or maths) to keep the new distribution truly uniform. Simply getting random bytes and viewing them as ints will give you low quality random numbers. The algorithm that casts MT output to a random integer is probably what is different, and perhaps you could find it documented somewhere. As a side note, C++11 includes in the standard a MT RNG guaranteed to be the same across implementations, but there is no promise about the distributions -not even across versions of the same implementation.
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
