I intend to try to turn it into a more general tool that simply reads its standard input, which is assumed to be the output of any arbitrary PRNG. One just has to specify whether the input is bits or IEEE doubles in the range [0, 1[.

That is what I use with TestU01, Diehard and others. I find it a lot handier than incorporating one generator after another into the code, compiling, etc.

Do you see any problems with that approach?

]]>ive just commited what i found on my disk to

svn://svn.mplayerhq.hu/michael/trunk/randi

I dont know for certain though if this was exactly the code i used or the last version …

]]>Any test even more stringent than BigCrush is of great interest to me – and I daresay to many others as well. ]]>

The Diehard battery of tests:

http://stat.fsu.edu/pub/diehard/

The ENT program:

http://www.fourmilab.ch/random/

About the discussion of the statistical theory of the tests, you can e-mail me: ribeiroalvo(at)sapo.pt

Or if you prefer start a tread at sci.crypt forum.

Thanks

At this moment, the algorithm is in analisys by Peter Helekallek form the Plab project : http://random.mat.sbg.ac.at/

]]>Thanks ]]>

> > parameters (IIRC, Berlekamp provided an algorithms back in 60-70s),

> yes, and berlekamp is faster than the gaussian elimination iam using but later is more

> flexible as it also works for cases where the â€œpredicted fromâ€ bits arent exactly the

> next â€œpredictedâ€ bits.

Btw, TestU01 according to the paper does use berlekamp-massey as one of their tests

]]>> (IIRC, Berlekamp provided an algorithms back in 60-70s),

yes, and berlekamp is faster than the gaussian elimination iam using but later is more flexible, as it also works for cases where the “predicted from” bits arent exactly the next “predicted” bits.

> for LGC itâ€™s even more obvious. Probably the best thing for you is too look at cryptographic

> generators. Even simple (and slow) Blum-Blum-Shub should be fine for you ;).

if i was searching for a PRNG for crypto i surely would not pick one based on TestU01 and some of my own tests ;)

Anyway, another of them (brent-xor4096) failed another linear prediction test, more precissely a 32768×65536 matrix of bits (that being only 1 lsb from each scalar) just has a rank of 5110, while for random bits rank=32768 would be expected.

And just a known fact – from a part of LFSR generator sequence you can restore its parameters (IIRC, Berlekamp provided an algorithms back in 60-70s), for LGC it’s even more obvious. Probably the best thing for you is too look at cryptographic generators. Even simple (and slow) Blum-Blum-Shub should be fine for you ;).

]]>