Yeah I get your point. But I guess for this model you can kinda have a concept of the "ideal" training set, where all high frequency features appear at the same rate as in real world.