multivariate MI for under-determined data

[Oswald Berthold]

Dear group,

does anyone know (something about) what is the behavior of continuous
MutualInfo estimators (kraskov1 / 2) on under-determined data, that is
(much) less examples than dimensions per example?

my observation is, that the source and destination entropies (each
computed as their respective self-information) is somehow tightly
bounded by the number of examples (it is the same for source and
destination and increases with number of examples). the "real" MI
between source and destination does seem plausible, though.

thanks for any pointers.

bst, oswald