Suppose there is an experiment where you have N conditions and in each
conditions there are M possible measurements. The outcomes are
represented by c[N,M] matrix of counts: how many times you observe each
possible measurement for each condition.
We want to ask what's the information gained by taking a measurement.
We assume the probability of each condition is a pc[N] vector, which
defaults to the frequencies in c. (We assume that the choice of
condition is not a random variable, but is made in advance, like most
experiments.)
