Hard-core distributions for somewhat hard problems

Consider a decision problem that cannot be 1-/spl delta/ approximated by circuits of a given size in the sense that any such circuit fails to give the correct answer on at least a /spl delta/ fraction of instances. We show that for any such problem there is a specific "hard core" set of inputs which is at least a /spl delta/ fraction of all inputs and on which no circuit of a slightly smaller size can get even a small advantage over a random guess. More generally, our argument holds for any non uniform model of computation closed under majorities. We apply this result to get a new proof of the Yao XOR lemma (A.C. Yao, 1982), and to get a related XOR lemma for inputs that are only k wise independent.