On Oct 4, 2013, at 9:18 AM, Aleksey Shipilev <aleksey.shipi...@oracle.com> wrote:
> On 10/04/2013 03:34 AM, Brian Burkhalter wrote: >> Here is an alternative solution: http://cr.openjdk.java.net/~bpb/7189139.2/. > > Seems OK with me, Same here. > as long as Miller-Rabin test tolerates the > lower-entropy PRNG. > I think so given recent updates to TLR. >> If this seems reasonable, does anyone have suggestions as to testing? > > I would like to see the functional tests. > > We can take the list of prime numbers [1] as set P, and then check: > a) numbers in intersect([2;N), P) return isProbablyPrime=true in more > than >(1 - 1/2^cert) cases; Plus the certainty is capped at 100 if the bit length of the big int < 100 (which is the case for the first 50M primes). For large bit lengths it is even less (capped at 4 for > 1024). (IIUC it is (1 - 1/4^N) for N iterations [1], for the impl N = certainty/2.) Paul. [1] http://mathworld.wolfram.com/Rabin-MillerStrongPseudoprimeTest.html > b) all numbers in difference([2;N), P) return isProbablyPrime=false > > N > 50.000.000 would sound convincing. > > -Aleksey. > > [1] http://primes.utm.edu/lists/small/
