Presumably he instructed mathematica to approximate the roots to an error proportional to the resolution of the image; in that (likely) case the error doesn't have anything to do with the fractals.