If we could somehow guarantee that every split were good...
We used the (provable, but unproven) fact that the maximum value of over the range is attained at k=3n/4 (i.e., the least balanced split).