kde to support boundaries and adaptive bandwidth

[cranmer]

Does scikit-learn's normal kde alg support adaptive bandwidth?

Does it support boundary's? The s(x) function has a min and a max. It would be nice if we could enforce the kde has all probability within s_min and s_max

this worked well in the past
http://arxiv.org/abs/hep-ex/0011057

[glouppe]

The current implementation of carl.distributions.KernelDensity is based on scipy.stats.gaussian_kde, which unfortunately does not support either of these features. It would be nice indeed, but this still would require some work to be done upstream in scipy. (or alternatively, to extend sklearn.neighbors.KernelDensity)