coverage

How much of a fresh sample the recommended fit cannot account for at all.

A mixture assembled from contiguous groups has support equal to the union of its components' supports. When a bounded family is chosen the mixture assigns no density outside it, so fresh data can land in a gap and receive density zero — which makes the held-out log-likelihood infinite and therefore unavailable exactly when bounded families were selected, which is not at random.

The share of held-out data with no density is defined whatever happens, and for this method it is the more informative statistic. Treat a share above a percent or two as a warning that the mixture may be describing the sample rather than the process.

Parameters