I'm not arguing that this approximator is necessary (not sufficient) for this class of networks. I've proposed some conjectures on what we might expect to see, but there are certainly other salient ingredients and common principles that we haven't discovered, and I think it's important to hunt for them.