The kernel size (smoothing width) should be chosen based on something like leave-on-out cross validation. At the moment the sampling is too sparse and everythin

The kernel size (smoothing width) should be chosen based on something like leave-on-out cross validation. At the moment the sampling is too sparse and everything is independent and not smooth at all. This is described in the intro of "Assessment of probability density estimation methods: Parzen window and Finite Gaussian Mixtures" http://www0.cs.ucl.ac.uk/staff/c.archambeau/publ/iscas_aa06.pdf http://www0.cs.ucl.ac.uk/staff/c.archambeau/publ/iscas_aa06.... (as well as many other places that just happened to be the first google result) email me if you want more details, I am ex-informatics.