A fundamental challenge with information maximization methods is that proper definitions of information content require knowing the distribution of vectors, which we do not have; empirical measures are all upper bounds, not lower bounds, so methods use good upper bounds and cross their fingers