Log Weights

[cscherrer]

StatsBase has several options for Weight Vectors. In many applications, we prefer to work in terms of (up-to-additive-constant) log-probabilities, or "LogWeights". Among other benefits, this has the advantage of preventing underflow.

For sampling, there's a well-known "Gumbel-max trick" that is simple, and faster than exponentiation followed by alias sampling. And a "Gumbel-softmax trick" has found applications in inference.

I'll be rolling my own version of this for now, and I'll try to make the interface consistent with StatsBase's. Just leaving this note here in hope of getting this into StatsBase once it's all working.

[nalimilan]

See also #401 for exponential weights.