A commonly occurring problem is that of maximizing a multinomial likelihood over a restricted region. We show that if the region is convex, then a dual problem always exists which is frequently more tractable. A solution to the dual problem leads directly to a solution for the original problem...