The sample_bit_prob function (which is used to sample from a Bernoulli(p) with p \neq 0.5 was indexing into the mantissa incorrectly. The function still gave roughly acceptable answers much of the time, but was wildly incorrect for some values of p. I think this is the cause of some of our utility issues with the Geometric mechanism.
The
sample_bit_prob
function (which is used to sample from aBernoulli(p)
withp \neq 0.5
was indexing into the mantissa incorrectly. The function still gave roughly acceptable answers much of the time, but was wildly incorrect for some values ofp
. I think this is the cause of some of our utility issues with the Geometric mechanism.I have a proposed fix in PR #279