[split] (lack of) scribal mistakes / corrections

[Jorge_Stolfi]

Cosine similarity analysis between folios shows the transition from Currier A to B "descends smoothly, almost linearly, with increasing rank" with "the complete absence of any sudden change in the slope" (You are not allowed to view links. Register or Login to view., p. 307).

That argument is flawed.  For one thing, the most common word, daiin, has similar frequencies on A and B: (~9% and ~4%) 
   
The word count vector of each page is likely to be dominated by the "AB-words" -- daiin and few other words that are common in both languages.  Whereas the big difference between A abd B is a larger set of words -- the "A-words"  and "B-words" -- that occur in only one language, but  more rarely. So the vector is a few largish numbers, always in the same place, a few dozen 1s and a bunch of zeros.  

When you compute the cosine of the two vectors, the coordinates are implicitly normalized to have norm 1.  That makes the vectors nearly parallell to the AB-words general direction.  The sampling noise in the AB-word counts will swamp the differences due to the presence of A-words or B-words.  So it is no wonder that the distribution of all page-page cosines is like that or random vectors.

To check your claim, you should choose two sets of presumed A-words and B-words, and map each page to two numbers a,b which are the relative frequencies in the page of words from each set, and plot each page as a color dot.  Being combined counts, those numbers will be less affected by sampling noise.  

Then, if your claim is true, you should see a single cloud, as A-words get replaced gradually by B-words.  But many previous tests have shown the opposite: A-pages and B-pages form two clusters with few if any dots in between.

All the best, --stolfi