Faulty values calling BinomialDistribution's RNG

[jaltevogt]

For carefully selected success rate p ~=0.999000999000999 the binomial distributions random number generator seems bugged. Maybe some numerical instabilities?

Code snippet:

int n=1000;
List<Double> pList = new ArrayList<>();
pList.add(0.999);
pList.add(0.99900099900099);
pList.add(0.999000999000999);
pList.add(0.9990009999);
pList.add(0.999001);
pList.add(0.999002);
pList.add(0.999006);
pList.add(0.9991);
pList.add(0.9992);
for (double p : pList) {
    BinomialDistribution distb = new BinomialDistribution(n,p);
    double rn = distb.rand();
    if (rn<900 || rn>1000) {
        System.out.println(rn + " for n="+n+ " and p="+p);
    }
}

results in:

-2.147482648E9 for n=1000 and p=0.999000999000999
469955.0 for n=1000 and p=0.9990009999
289762.0 for n=1000 and p=0.999001
1286.0 for n=1000 and p=0.999002
1153.0 for n=1000 and p=0.999006
1002.0 for n=1000 and p=0.9991

Can you reproduce this?

[haifengl]

It is caused by underflow/overflow. Those p are too large (very close to 1). Although we can harden the code to throw exceptions in such case, it may not be very useful for random number generations.

[haifengl]

A workaround is to use a large n (e.g. 10x) and then scale down the random number by dividing 10.

[jaltevogt]

Thanks, scaling down works. I'm not familiar with the inner workings of the rng, but I just realized BinomialDistribution.rand() uses PoissonDistribution.tinyLambdaRand() as a workaround in case n*p < 1.E-6 . Should this workaroun be used in case p/n<1.E-6 or (1-p)/n < 1.E-6 .

[haifengl]

I made some enhancement. It works fine with your examples. Please try master branch.

[jaltevogt]

Thanks for the fix!