It's a great insight, but not that hard to prove (or at least "get" the reasoning behind it): anything that distinguishes the data from randomness is redundant information, and that redundancy can be used to further compress it.

(Obviously, that doesn't count as a formal proof, &c.)