From what I can tell, it memoizes the intermediate additions and then uses that to amortize the adds across multiple array elements and achieve speedups.

neither of these comments is correct

Mind offering an explanation?

there is a perfectly clear explanation in the paper