14
7

Learning algorithms versus automatability of Frege systems

Abstract

We connect learning algorithms and algorithms automating proof search in propositional proof systems: for every sufficiently strong, well-behaved propositional proof system PP, we prove that the following statements are equivalent, 1. Provable learning: PP proves efficiently that p-size circuits are learnable by subexponential-size circuits over the uniform distribution with membership queries. 2. Provable automatability: PP proves efficiently that PP is automatable by non-uniform circuits on propositional formulas expressing p-size circuit lower bounds. Here, PP is sufficiently strong and well-behaved if I.-III. holds: I. PP p-simulates Je\v{r}\ábek's system WFWF (which strengthens the Extended Frege system EFEF by a surjective weak pigeonhole principle); II. PP satisfies some basic properties of standard proof systems which p-simulate WFWF; III. PP proves efficiently for some Boolean function hh that hh is hard on average for circuits of subexponential size. For example, if III. holds for P=WFP=WF, then Items 1 and 2 are equivalent for P=WFP=WF. If there is a function hNEcoNEh\in NE\cap coNE which is hard on average for circuits of size 2n/42^{n/4}, for each sufficiently big nn, then there is an explicit propositional proof system PP satisfying properties I.-III., i.e. the equivalence of Items 1 and 2 holds for PP.

View on arXiv
Comments on this paper