Inseparability and Strong Hypotheses for Disjoint NP Pairs

This paper investigates the existence of inseparable disjoint pairs of NP languages and related strong hypotheses in computational complexity. Our main theorem says that, if NP does not have measure 0 in EXP, then there exist disjoint pairs of NP languages that are P-inseparable, in fact TIME(2^(n^k))-inseparable. We also relate these conditions to strong hypotheses concerning randomness and genericity of disjoint pairs.

💡 Research Summary

The paper addresses a long‑standing question in computational complexity: whether there exist disjoint pairs of languages in NP that cannot be separated by any efficient algorithm. A pair (L₁, L₂) is called inseparable if no algorithm running within a given resource bound can correctly decide, for every input, whether it belongs to L₁ or to L₂. While earlier work showed that under certain unproven assumptions such P‑inseparable NP pairs must exist, a concrete sufficient condition that guarantees their existence had been missing.

The authors introduce a measure‑theoretic assumption about the size of NP inside EXP. In resource‑bounded measure theory, a class C has measure zero inside a larger class D if, roughly speaking, the proportion of D‑machines that compute languages in C is negligible. The central hypothesis of the paper is that NP does not have measure 0 in EXP. Intuitively, this means that NP is not “tiny” relative to EXP; it occupies a non‑negligible fraction of the exponential‑time world.

Under this hypothesis the authors prove two main theorems.

1. P‑inseparability: There exist disjoint languages A, B ∈ NP such that no polynomial‑time deterministic algorithm can separate them.
2. Stronger time‑inseparability: For every integer k ≥ 1, there are disjoint NP languages Aₖ, Bₖ such that no algorithm running in time 2^{n^{k}} can separate the pair. In other words, the inseparability holds even against sub‑exponential time bounds that grow as a fixed exponential of a polynomial.

The proof proceeds in two conceptual stages.
Stage 1 – Measure to Randomness: The authors adapt the resource‑bounded measure framework to construct a resource‑bounded randomness test (a variant of Martin‑Löf tests tailored to exponential‑time machines). They show that if NP has non‑zero measure in EXP, then there exist NP languages that pass this randomness test with positive probability. Such languages can be viewed as “random” or “generic” within EXP.

Stage 2 – Multi‑Mask Construction: Using the random NP languages from Stage 1, the authors define a family of multi‑mask languages. Each language in the family is specified by a collection of masks (binary strings) that dictate which inputs are accepted. The masks are chosen to be of length roughly 2^{n^{k}} for the k‑th theorem, ensuring that any algorithm limited to time 2^{n^{k}} can only inspect a tiny fraction of the mask bits. By a counting argument based on the non‑zero measure assumption, any such algorithm fails to correctly identify the mask configuration for a non‑negligible fraction of inputs, and therefore cannot separate the two languages.

Beyond the core technical results, the paper formulates two strong hypotheses that are shown to be equivalent to the non‑zero‑measure assumption.
Randomness Hypothesis: “NP contains languages that are random with respect to the resource‑bounded measure in EXP.”
Genericity Hypothesis: “A generic (i.e., dense and open) set of disjoint NP pairs is inseparable under the same time bounds.”
The authors prove that each of these statements implies the other and both imply the existence of the TIME(2^{n^{k}})‑inseparable pairs. This equivalence ties together three perspectives—measure‑theoretic size, algorithmic randomness, and topological genericity—offering a richer conceptual picture of why inseparability should be expected.

The implications of the work are multifold. First, it provides a concrete sufficient condition (NP ⊄ measure‑zero in EXP) that yields inseparable NP pairs, strengthening earlier conditional results that required unproven circuit lower bounds or derandomization assumptions. Second, the time‑inseparability result shows that even algorithms with super‑polynomial but sub‑exponential resources cannot resolve certain NP disjointness problems, suggesting a robust barrier that persists across a wide range of computational budgets. Third, the connection to randomness and genericity hints that similar inseparability phenomena may appear in higher complexity classes (e.g., NEXP vs. EXP) if analogous measure‑theoretic statements hold.

From a practical standpoint, the existence of such hard‑to‑separate NP pairs can be leveraged in cryptographic constructions: a hard core of an NP language that remains indistinguishable from its complement even for powerful adversaries can serve as a building block for one‑way functions or pseudorandom generators.

In summary, the paper establishes that if NP is not negligible inside EXP, then disjoint NP pairs exist that are inseparable by any algorithm running in time 2^{n^{k}} for any fixed k. It further demonstrates that this condition is equivalent to natural randomness and genericity hypotheses about NP. The results deepen our understanding of the structural relationship between NP and EXP, and open new avenues for exploring inseparability, hardness, and randomness across complexity hierarchies.