On the Necessity of Enumerating All Programs

In a tutorial on average-case complexity that I gave at FOCS 2008, I mentioned four of my favorite open questions in the theory. I have some progress to report on one of the questions (#3 in this post), and on two related problems.

The question is whether a certain exponential loss is necessary in Levin’s proof that there are problems that are “NP-complete on average.” The answer is yes, it is necessary, unless the polynomial hierarchy collapses. The two related problems are whether related exponential losses in the analysis of Levin’s universal search algorithm and of Levin’s universal one-way function are also necessary. The answers are, respectively, yes unless {P=NP}, and yes relative to an oracle.

This is one of those now fashionable papers whose arguments are “obvious in retrospect.” (I would like to claim, however, that the proof of the result for Levin’s average-case completeness, while obvious in hindsight, is not so obvious without hindsight.) In all three cases, the exponential loss is in terms of the code length of certain programs. The idea in each case is to construct programs that contain in their code information about a specific instance of a problem of interest.

Continue reading