“…In particular, if the algorithm runs in deterministic polynomial-time with access to an NP oracle, then there is a function in E NP that requires exponentially large circuits. Next, in section Section 4.3 we show (in Theorem 4.5), based on ideas from Bshouty, Cleve, Gavaldà, Kannan, and Tamon (1996); Chakaravarthy and Roy (2008) ;Fortnow, Pavan, and Sengupta (2008), that there is a deterministic algorithm that uses an oracle to prAM for learning counterexamples (recall that the result of Fortnow, Pavan, and Sengupta (2008) only gives a randomized algorithm, so it is not good for us). By the hypothesis of Theorem 4.1, we can replace the prAM oracle with an NP oracle and then the lower bound follows in Section 4.4.…”