Corollary 2 — Moment bound
ProvedFBMSP.corollary2_moment_boundAssume nonnegative integrable predictions with finite second moment, independent integrable mean-zero error, and nonnegative realized valuations almost surely. If prediction variance is and , then for every ,
Author-confirmed correction. The factor two was confirmed on October 30, 2026. This result uses the same nonnegative realized-valuation setting as corrected Theorem 3. Minimizing the proof's expression yields this constant. The formal proof remains open and must handle conditional truncation, boundary atoms, and the zero-variance case.
import Definitions.Def_FBMSP_Model open MeasureTheory
namespace FBMSP
theorem corollary2_moment_bound (ν η : Measure ℝ)
(hν : ValidPredictions ν) (hη : ValidNoise η)
(hV : NonnegativeValuations ν η)
(hMoment : Integrable (fun x : ℝ => x ^ 2) ν)
(σ : ℝ) (hσ : 0 ≤ σ) (hVar : predictionVariance ν = σ ^ 2)
(k : ℕ) (hk : 1 ≤ k) :
segmentationLoss ν η k ≤
2 * Real.sqrt (2 * σ * (predictionMean ν + σ) / (k : ℝ)) := by sorry
end FBMSPRead-back
What the Lean code literally says, in plain math · gpt-6-astra
For every pair of probability measures and on , assume that is integrable with respect to , that for -almost every , that is integrable with respect to , that , and that for -almost every , the inequality holds for -almost every . Assume additionally that is integrable with respect to . Write . For every real number satisfying , and every natural number with , let be the real number obtained from , let be the family of all finite sets with , and define for every and every real . The asserted inequality is , with interpreted as a real number in the quotient. The first supremum optimizes a nonnegative price separately for each , while each menu supremum chooses one fixed finite price set used for all . The interval defining is closed at its lower endpoint and therefore counts any mass at the posted-price threshold. Empty menus are allowed and yield ; the zero-revenue option in the maximum does not consume a menu element, and fewer than distinct prices may be used. The conclusion does not assert that an optimal price or menu is attained. Zero variance is included: makes the stated upper bound . The case is included and is excluded, so division by zero does not arise. The square root is the total real square-root function, which returns on negative inputs; here its argument is nonnegative because , , and . No second-moment assumption on , nor any density, continuity, or hazard-rate assumption on either measure, is imposed. All displayed integrals are real Bochner integrals, whose total definition gives for a nonintegrable integrand. All displayed suprema are real suprema, whose total definition gives for an empty or unbounded-above set; the optimization sets here are nonempty because price and the empty menu are available. Conversion of a measure value to a real sends to , but the probability-measure hypotheses make every survival value finite.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.