Hmmm... I've been trying to get my head around that one too.

Well, let's see. The way I'd do it (I think) is...

Phase-invert one of the channels and subtract it from the other channel. This will give you a track that's only the outside two channels. Then take THAT and subtract it from a mono-summed version of the original signal. The result is just the stuff that's common to both channels.

Note that the first half of the equation (phase-invert and subtract) is how most of the "vocal eliminators" work. They assume that the voice is going to be mono center, and the instruments will be panned L and R.