And an absolutely fascinating investigation. It's almost like the famous quote of not being able to define something (pornography), but knowing it when you see it. [1]
Typography, with its rhythms, is almost the purest example of this. It's so easy for a good designer to kern, set linespacing, etc. until it "looks right" to your eye. But I wouldn't even know where to begin to try to translate this into an algorithm -- because it's not just about spacing between letters, but having that spacing be harmonious as well with the structures within letters (like the horizontal rhythm of an "m").
I personally suspect it's something close to trying to equalize the "area" between adjacent letterforms, but the problem is that depending on angles, case, etc. that "area" is a "blob" we feel, not an easy calculation. I wouldn't be suprised if the best approximation involves some kind of fuzzy-bitmap blobs, rather than any hard of hard exact calculations.
I look forward to reading more research results!