🚀 go-pugleaf

On 05.07.2016 10:13, M.M. wrote:

>> Ciekawe, moze cos dobrego masz :-)
> Nic rewelacyjnego, mam złożenie dwóch generatorów fibbonaciego. Jest
> bardzo szybki, ma duży stan, a że przechodzi testy kryptograficzne - to
> jestem zaskoczony.

Zwiększ parametry w dieharder.

Z wspominanego manuala (analiza RANDU)
Even this RNG, well-known to be terrible for most purposes, is capable
of producing sequences that are “random enough” to pass some of
Diehard’s difficult tests at the default Dieharder values, which in turn
are invariably more challenging than those of Diehard.
What happens if one cranks up the tests to to where they make a RNG die
harder? In the case of randu, mostly it dies. For example the sums test
fails without any doubt at 500 p-values. runs fails badly at 1000. The
squeeze test fails at 300. Even parking lot fails at 2000 p-values. With
randu, “passing” Diehard tests appears to be a matter more of how hard
you look, not whether or not the result is truly random
according to the test measure.
In contrast, mt19937 1999 still passes the runs test at 10000 p-samples,
although at that point the pass gets to be visibly marginal. We
reiterate the previous observation concerning the true purpose of
Dieharder with its variable controls: The point isn’t “pass-
ing” any given test, the point is determining where it fails in a
quantitative way! When this is done one can compare the performance of
different RNGs in a meaningful way on a test by test basis.

pzdr
bartekltg