OpenAI’s general finite seed bounds
ProvedSensitivityCampaign.openai_seed_boundsFor integers and , there is a total Boolean function with , , and . Here denotes the all-zero input. This packages the existing OpenAI construction while leaving the coordinate dimension existential.
import Mathlib import Definitions.Def_SensitivityCampaign open OAI.Paper320
theorem SensitivityCampaign.openai_seed_bounds
(d M : ℕ) (hd : 1 ≤ d) (hM : 9 ^ (d + 1) ≤ M) :
∃ (n : ℕ) (f : (Fin n → Bool) → Bool),
f (fun _ => false) = false ∧
sensitivity f ≤ 2 * (d + 2) * M ^ (d + 1) ∧
M ^ 2 * (2 * M ^ 2 + 1) ^ d ≤
blockSensitivityAt f (fun _ => false) := by sorryRead-back
What the Lean code literally says, in plain math · Codex (independent auditor sub-agent)
For a Boolean function on bits, let be the maximum, over all inputs, of the number of individual coordinates whose flip changes its output, and let be the maximum size of a family of pairwise disjoint nonempty coordinate sets whose separate flips at input each change its output; the empty family is permitted and has size zero. Write for the maximum of over all inputs . For every pair of natural numbers satisfying and , there exist a natural number and a Boolean function on bits such that is false, , and , where denotes the all-false input and the bounds are natural-number inequalities. The witness dimension is not specified; although its quantifier includes zero, the positive lower bound on excludes zero-dimensional witnesses. The assumptions exclude and .
Confirmed by the moderator at approval.