Large entropy consumption for fractions not in lowest form

[mjdemedeiros]

When arguments to DiscreteGaussian are not in lowest form, its mean entropy usage increases substantially.

Below are some histograms profiling the arguments given to probUniformP2 for different values of DiscreteGaussian(num, den, mix) (see this fork for the script).

Up to (90, 90, 1) there is minimal degradation in performance; I suspect this is because even the excessively large samples fit within 4 bytes. Nevertheless, reducing the fractions to lowest form at the start of the DiscreteGaussian function could reduce the amount of entropy consumed.

There is, however, steep degradation in performance after (90, 90, 1). In fact, the (91, 91, 1) benchmark does not complete on my system. This might be a separate issue, since the (91, 90, 1) benchmark also gets stuck before after around 20 trials but the (90, 91, 1) does not. I am still uncertain about where the root cause.