Proposition 6(b) — uniqueness of the SPO+ risk minimizer
ProvedSmartPTO.Fisher.prop6b_unique_minimizerbayes-riskp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1spo-plus-lossuniqueness
Let be an integrable probability law on , centrally symmetric about its mean and continuous on all of . Let be nonempty, compact, convex, and have nonempty interior, with an arbitrary optimal-decision oracle . If a prediction has SPO+ risk no greater than that of , then
Together with Proposition 6(a), this says the mean is the unique minimizer of the surrogate risk.
Formalization Note The risk is an extended nonnegative integral. “Continuous on all” is pinned to absolute continuity plus full support: an absolutely continuous law supported inside a small ball can have many minimizers, contrary to the printed uniqueness claim.
Preamble
import Mathlib import Definitions.Def_SPOBounds_Natarajan_Model import Definitions.Def_SmartPTO_Fisher_Setting open scoped InnerProductSpace open MeasureTheory ProbabilityTheory
Formal statement
namespace SmartPTO.Fisher
/-- Proposition 6(b), p. 23: under nonempty interior the mean is the unique minimizer. -/
theorem prop6b_unique_minimizer {d : ℕ} (S : Set (EuclideanSpace ℝ (Fin d)))
(hSne : S.Nonempty) (hScpt : IsCompact S) (hScvx : Convex ℝ S)
(wstar : EuclideanSpace ℝ (Fin d) → EuclideanSpace ℝ (Fin d))
(hw : SPOBounds.Natarajan.IsOracle S wstar)
(P : Measure (EuclideanSpace ℝ (Fin d))) [IsProbabilityMeasure P]
(hInt : Integrable id P) (hCont : ContinuousOnAll P)
(hSymm : CentrallySymmetric P) (hSinner : (interior S).Nonempty)
(chat : EuclideanSpace ℝ (Fin d))
(hMin : spoPlusRisk S wstar P chat ≤ spoPlusRisk S wstar P (∫ c, c ∂P)) :
chat = ∫ c, c ∂P := by sorry
end SmartPTO.Fisher
Source
Elmachtoub & Grigas, Smart "Predict, then Optimize", arXiv:1710.08005v5, p. 23, Proposition 6(b); §4.1 p. 22
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.