I wrote:
That said, looking around briefly, I came up with this paper by L'Ecuyer et al that does seem to describe a decent random generator with properties of split worked out:
Derek Elkins wrote:
According to the documentation That -is- what we have.
No, we have a much older algorithm of L'Ecuyer. It has a much smaller period - 2e18 vs. 3e51. Much less was known about testing generator randomness at that time. More importantly, very little was said in that paper about splitting. The newer algorithm includes all the computational details about splitting. Sadly, even the newer paper does not propose any tests for the independence properties of streams after splitting. The assumption seems to be that the voluminous testing of the underlying generator is sufficient for that also. Regards, Yitz