Proposition 3.1, p. 10, only-if direction on the forecast domain D — Savage representation of a proper scoring rule
ProvedQAPooling.MaxMin.proposition_3_1Let be a regular proper scoring rule on outcomes, with expected reward function and forecast domain . Then:
- is convex on ;
- the map is itself an exposure function of : for all and every outcome ,
- for every exposure function of and every , the vector is a subgradient of at relative to :
This is the only-if direction of the Savage representation (Gneiting and Raftery, Theorem 2): a proper scoring rule is the tangent plane of its convex expected reward. Clause 2 shows exposure functions exist, and clause 3 is what makes 's supporting hyperplanes available in the max-min argument.
Formalization Note The statement is restricted to the forecast domain , where all scores are finite; the paper's exposure function may take the value at boundary points outside , which are not covered here. Subgradients are relative to , the domain of (footnote 10). The "if" direction is not formalized.
import Mathlib import Definitions.Def_QAPooling_MaxMin_Model
namespace QAPooling.MaxMin
/-- Proposition 3.1, only-if direction, on the forecast domain `D` (p. 10). A regular proper scoring
rule has a convex expected reward `G` on `Δⁿ`; `s(p; ·)` itself is an exposure function
(Equation (2) on `D`); and every exposure function `g` gives, at every `q ∈ D`, a subgradient
`g(q)` of `G` relative to `Δⁿ`. -/
theorem proposition_3_1 {n : ℕ} (s : ScoringRule n) (hreg : IsRegular s) (hprop : IsProper s) :
ConvexOn ℝ (stdSimplex ℝ (Fin n)) (expectedReward s) ∧
IsExposure s (fun p j => (s p j).toReal) ∧
∀ g : (Fin n → ℝ) → Fin n → ℝ, IsExposure s g →
∀ q ∈ forecastDomain s, IsSubgradientOn (expectedReward s) (g q) q := by sorry
end QAPooling.MaxMin
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.