I love this quote from a follow up paper: > Interestingly, hash algorithms with collision probabilities equal to JP have already been unintentionally presented

I love this quote from a follow up paper: > Interestingly, hash algorithms with collision probabilities equal to JP have already been unintentionally presented before JP was actually discovered and thoroughly analyzed in [8]. In [7] a data structure called HistoSketch was proposed to calculate signatures for JN ... after some simplifications and thanks to a nonequivalent transformation that eliminated the scale dependence, the final HistoSketch algorithm had a collision probability equal to JP instead of the originally desired JN. It's none the less a very interesting measure. Thanks for sharing! I recently worked on a project trying to determine "the best" locality sensitive hashing amoung all measures of similarity for sets: https://arxiv.org/abs/1904.04045 https://arxiv.org/abs/1904.04045 I wonder if something similar could be done for probability distributions. It seems hard.