On the impossibility of basing trapdoor functions on trapdoor predicates

We prove that, somewhat surprisingly, there is no black-box reduction of (poly-to-one) trapdoor functions to trapdoor predicates (equivalently, to public-key encryption schemes). Our proof follows the methodology that was introduced by R. Impagliazzo and S. Rudich (1989), although we use a new, weaker model of separation.