n log n seems reasonable to me (though ill be honest, i had to look it up when you first mentioned it)  Im trying to figure out mongoDB right now to get a full days of posts and simulate what their rewards would have been under a linear system, n log n and a couple of other possibilities (one based on standard deviation also, and a hybrid one with starts out n^2 and has decreasing superlinearity).

One of the great things about changing reward curve is that, iiuc, its already designed to be somewhat motile (kind of like the regeneration rate, its just there defined on one line).  You still need a hard fork, but its as easy as something that requires a harfork can be.