Proposition 3.1 — the SPO+ loss upper-bounds the SPO loss
ProvedSmartPTO.Fisher.prop3_upper_boundp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1spo-plus-losssurrogate-lossupper-bound
Let be nonempty, compact, and convex, and let return an optimizer of for each cost vector . For every realized cost and prediction ,
Thus the convex surrogate bounds the unambiguous decision loss pointwise.
Formalization Note Both extrema are real infima or suprema, made genuine by the nonempty compact feasible region. The oracle is arbitrary among optimal decisions.
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 3.1, p. 18: the SPO+ loss upper-bounds the unambiguous SPO loss. -/
theorem prop3_upper_bound {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)
(c chat : EuclideanSpace ℝ (Fin d)) :
spoLoss S chat c ≤ spoPlusLoss S wstar chat c := by sorry
end SmartPTO.Fisher
Source
Elmachtoub & Grigas, Smart "Predict, then Optimize", arXiv:1710.08005v5, p. 18, Proposition 3.1; §2 standing assumptions pp. 8–9
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.