Comment by FabHK

Thanks. I've looked at the code, and it does not seem to try to estimate the empirical distribution (doesn't appear to be using dictionaries, for examples).

Then the discrepancy maybe comes from the number of glyphs within certain categories, or their repetition?