Truncated KL estimator for mutual information

[kahaaga]

In this paper, Zhao and Lai propose to use a truncated version of the KL-estimator in cases where some samples are far away from other samples.

The truncation involves setting a maximum allowed distance to other points, and if some points have neighbors with distances exceeding this threshold, then we replace the distance with a sample size-dependent truncation value.

In practice, whether or not to truncate should be a keyword to the KozachenkoLeonenko estimator.