I believe this is a mischaracterization of the research conclusions.  The actual conclusion was that the maximum value for the blocksize that the network can safely handle (at that time) is some value that is (conservatively) no more than 4MB.  This is because the research only studies one aspect of the effect of blocksize on the network at a time and the true safe value is the minimum of all aspects.  For example, the 4MB doesn't cover the aspect of quadratic hashing for large transactions in large blocks.