Moment inequality implies a power-saving recursive cost
ProvedIntMul.Kappa.moment_recurrence_boundMoment inequality implies a power-saving cost. Let and be integers, and let be a child-width histogram: a finite list of pairs with . Let and suppose the moment inequality
holds. Let be nondecreasing and satisfy, for a constant and every ,
Then there is a constant with for every .
This is the recursive saving of the OpenAI manuscript in histogram form: a network on roles of full width makes child calls of width on streams of volume , so the normalised cost of a call obeys the displayed recursion, and the strict moment inequality ( in the uniform case) yields cost exponent . It is the analytic step behind every moment certificate of the -improvement rounds.
Formalization Note The moment is histMoment from IntMul_KappaFramework. The recursion is required only at multiples ; monotonicity of covers the other arguments.
import Definitions.Def_IntMul_KappaFramework
namespace IntMul.Kappa
theorem moment_recurrence_bound (h : List (ℕ × ℕ)) (m W : ℕ) (τ K : ℝ) (c : ℕ → ℝ)
(hm : 2 ≤ m) (hτ0 : 0 < τ) (hτ1 : τ ≤ 1)
(hw : ∀ p ∈ h, 0 < p.1 ∧ p.1 < m) (hW : 0 < W)
(hmom : histMoment h m τ < W) (hmono : Monotone c)
(hrec : ∀ f : ℕ, 1 ≤ f →
c (m * f) ≤ (h.map fun p => (p.2 : ℝ) * c (p.1 * f)).sum / W + K) :
∃ A : ℝ, ∀ x : ℕ, 1 ≤ x → c x ≤ A * (x : ℝ) ^ τ := by sorry
end IntMul.Kappa