Theorem 3.7 — randomized estimation-to-testing information bound
ProvedInventoryBounds.two_point_estimation_informationConsider M IID zero/unit observations with zero probabilities rho0 or rho1, both strictly between zero and one. The same arbitrary Markov kernel maps the archive to a randomized real estimate. If it lies within epsilon of v0 or v1 with probability at least 1-eta respectively, where v1-v0>2epsilon and 0<eta<1/2, then M kl(rho1,rho0) is at least (1-2eta) log((1-eta)/eta).
import Definitions.Def_InventoryBounds_StationaryValuation set_option autoImplicit false open MeasureTheory ProbabilityTheory open scoped BigOperators
namespace InventoryBounds
theorem two_point_estimation_information (M : ℕ)
(κ : Kernel (Fin M → Bool) ℝ) [IsMarkovKernel κ]
(ρ₀ ρ₁ v₀ v₁ ε η : ℝ)
(hρ₀ : 0 < ρ₀) (hρ₀1 : ρ₀ < 1)
(hρ₁ : 0 < ρ₁) (hρ₁1 : ρ₁ < 1)
(hε : 0 < ε) (hη : 0 < η) (hη2 : η < 1 / 2)
(hsep : 2 * ε < v₁ - v₀)
(hs₀ : 1 - η ≤ estimatorSuccess κ ρ₀ v₀ ε)
(hs₁ : 1 - η ≤ estimatorSuccess κ ρ₁ v₁ ε) :
(1 - 2 * η) * Real.log ((1 - η) / η) ≤ (M : ℝ) * binaryKL ρ₁ ρ₀ := by sorry
end InventoryBoundsRead-back
What the Lean code literally says, in plain math · gpt-6
For every natural number , every Markov kernel from the finite binary archive space to the real line with its Borel measurable structure, and every six reals satisfying , , , , and , assume that, for each , , where and . Being a Markov kernel means that every is a probability measure and the dependence on is measurable; its interval masses in this expression are interpreted as real numbers. Then . The same kernel is used at both parameter values, and the conclusion uses the displayed order . There is no ordering or distinctness assumption on and no sign restriction on beyond their strict ordered separation. The quantifier allows : there is then a single empty archive and the empty product is , although the two success assumptions cannot both hold because the two closed success intervals are disjoint and each would have probability greater than . Likewise, if , the separated success assumptions cannot both hold. Natural is interpreted as real in the conclusion.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.