Learning algorithms versus automatability of Frege systems

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