I think the point is that a correct analytical result needs to be computed for all shapes at once. You can't composite overlapping shapes individually, since the alpha value (subpixel coverage), even if analytically correct, is lossy; you lose the information about the exact subpixel geometry, which is needed to do correct compositing (eg. does this shape completely occlude this other one within this pixel?). The standard Porter-Duff blend function assumes the subpixel geometry is a uniform random mist, which is not necessarily a great assumption for AA.