I exchanged emails with Greg Maxwell over several aspects of the paper that he questioned.  One point he did make, that I admit is valid but do not personally see as an issue, is that the most profitable "configuration" according to the results from the paper is a single "super pool" made up of ALL the network's hashing power (which would be centralizing).  This would minimize the propagation impedance.  While I agree that this is true, it seems like just another way of looking at the 51% problem.  We already know that if one entity controls a huge amount of hash power they can do nasty things and gain certain advantages.  But it would be nice to find a way to explain why this shouldn't happen with more rigour than the "game theory" or "anti-fragile" fallback positions…