poisotlab / fauxcurrences.jl Goto Github PK

The way to generate points at random now is to pick an angle, and a distance, to have a point in a circle. This clusters points too close to the center.

The fix is to have the square root of the radius, properly ranged. This isn't likely to change the results, but this might speed up the code because less biases proposals will be required.

KL-d goes to 0 only when p=0, q>0, which can be prevented with additive smoothing using a value of alpha > 0

This requires to solve #1 before, but also demands a deeper refactor as the problem becomes asymmetric, so all pairs needs to calculated twice.