Non-uniform BigInt distributions

[tabatkins]

The current proposal has Random.bigint(), which generates a uniform BigInt in a range. All the rest of the proposed number-returning functions, particularly the non-uniform distributions like Random.normal(), return Numbers.

In #8, tho, @michaelficarra mentions wanting non-uniform distributions over BigInts. How can we best address this?

• Should we have normalBigInt()/etc variants?
• The alternative is just saying "eh, BigInt(Math.round(Random.foo(...))) is good enough".

Are the use-cases for non-uniform BigInt distributions outside of the ±2^53 range large enough to need an API solution? Is it sufficient to just shift your distribution to within the SAFE_INTEGER range, generate a Number, convert to BigInt, then shift back? The only weakness is if you need integer-accurate values where the whole meaningful range is wider than 2^54.

[kgryte]

Nit: a normal distribution is continuous, meaning sampling should return real-valued numbers. For the discrete case, the most common equivalent distribution is the binomial distribution.

[tabatkins]

Yes, the implication of this is that you'd sample from a distribution and round/floor/etc to a BigInt.

[kgryte]

I am not sure sampling from, e.g., a continuous distribution and then rounding makes practical sense. If you are wanting discrete samples, you'd sample from a discrete distribution which typically will have a different parameterization than an analogous continuous distribution. Normal vs binomial is but one example.

[tabatkins]

An example was given in #8 (comment), where they're requesting basically an exponential distribution but just over the integers.

[kgryte]

Splitting hairs here, but that is not an exponential distribution, but a geometric distribution, which, as I mentioned above, is parameterized differently than an exponential.