Hardness amplification proofs require majority

Hardness amplification is the fundamental task of converting a δ-hard function f: {0, 1}ⁿ → {0, 1} into a (1/2 - ∈)-hard function Amp(f), where f is γ-hard if small circuits fail to compute f on at least a γ fraction of the inputs. In this paper we study the complexity of black-box proofs of hardness amplification. A class of circuits D proves a hardness amplification result if for any function h that agrees with Amp(f) on a 1/2 + ∈ fraction of the inputs there exists an oracle circuit D ∈ D such that D agrees with f on a 1 - δ fraction of the inputs. We focus on the case where every D ∈ D makes nonadaptive queries to h. This setting captures most hardness amplification techniques. We prove two main results: (1) The circuits in D "can be used" to compute the majority function on 1/∈ bits. In particular, when ∈ ≤ 1/log^ω(1) n, D cannot consist of oracle circuits that have unbounded fan-in, size poly(n), and depth O(1). (2) The circuits in D must make Ω (log(1/δ) ∈²) oracle queries. Both our bounds on the depth and on the number of queries are tight up to constant factors. Our results explain why hardness amplification techniques have failed to transform known lower bounds against constant-depth circuit classes into strong average-case lower bounds. Our results reveal a contrast between Yao's XOR lemma (Amp(f):= f(x₁) ⊕ · · · ⊕ f(x_t) ∈ {0, 1}) and the direct-product lemma (Amp(f):= f(X₁)O · · · of(x_t) ∈ {0, 1}^t; here Amp(f) is non-Boolean). Our results (1) and (2) apply to Yao's XOR lemma, whereas known proofs of the direct-product lemma violate both (1) and (2). One of our contributions is a new technique for handling "nonuniform" reductions, i.e., the case when D contains many circuits.