I haven't tried doing the actual maths, but I would assume that no algorithm can be proven to give you more than 2MPs worth of information *in the general case*. On the other hand, you might probably show that they can give you more than that (up to 6MP) under certain specific conditions. I think the trick to it all is being able to make assumptions about the conditions (i.e. luminance phenomena) that hold most of the time.
That sounds reasonable to me - except that as described, you can get 3MP out of the green channel using just a half-decent interpolation algorithm. The cases where you *can't* get more than that by using the other channels are the cases where the brightness of the image varies significantly between the channels - photos with strong colour casts are the only obvious example that comes to mind. (Indeed, if photographing a scene lit only with red or blue light, you'd only get 1.5MP of information. It's fascinating - analogous to the days of orthochromatic film. :-)
S
