Sorry for taking so long to answer, I didn't see notifications somehow.

Raymond, my use case is in general something that happens when I'm doing analytics on sequences of events (e.g., URLs visited by a browser) or paths in a graph. I look at pairs and do something based on the pair of events (e.g., did the user likely clicked an advertising link? did they go to a potentially risky webpage, possibly by clicking a link?)

I see the argument for generalizing to a sliding window, although that may lead people to choosing inefficient algorithms for sliding average or median.