Proposition 3.2 — convexity of the SPO+ loss in the prediction
ProvedSmartPTO.Fisher.prop3_convexconvexityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1spo-plus-losssurrogate-loss
Let be nonempty, compact, and convex, with an arbitrary optimal-decision oracle . For each fixed realized cost vector ,
This is the surrogate's structural property used in the paper's population-risk analysis.
Formalization Note Convexity is stated on the entire Euclidean prediction space, with no restriction on the oracle's tie-breaking.
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 3.2, p. 18: SPO+ is convex in the prediction for fixed realized cost. -/
theorem prop3_convex {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 : EuclideanSpace ℝ (Fin d)) :
ConvexOn ℝ Set.univ (fun chat => spoPlusLoss S wstar chat c) := by sorry
end SmartPTO.Fisher
Source
Elmachtoub & Grigas, Smart "Predict, then Optimize", arXiv:1710.08005v5, p. 18, Proposition 3.2; §2 standing assumptions pp. 8–9
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.