Proposition 3.4, p. 11, first bullet, for G the expected reward of a proper scoring rule and q ∈ D — Bregman divergence is non-negative
ProvedQAPooling.MaxMin.proposition_3_4Let be a regular proper scoring rule with expected reward function , forecast domain and exposure function , and write for the Bregman divergence. For every and every ,
with equality only when .
By Remark 3.3 of the paper, is the expected reward an expert with belief loses by reporting . Non-negativity and strict positivity off the diagonal are what make the agent's worst-case profit in Theorem 4.1 non-negative, and positive unless all experts with positive weight agree.
Formalization Note The paper states this for a general differentiable strictly convex ; here is the expected reward of a proper scoring rule and an exposure function plays the role of the gradient, with the second argument in (where is a finite subgradient). The second bullet of Proposition 3.4 (strict convexity of ) is not formalized.
import Mathlib import Definitions.Def_QAPooling_MaxMin_Model
namespace QAPooling.MaxMin
/-- Proposition 3.4, first bullet (p. 11), for `G` the expected reward of a regular proper scoring
rule and `g` an exposure function: for `q ∈ D` and `x ∈ Δⁿ`, `D_G(x ∥ q) ≥ 0`, with equality only
when `x = q`. -/
theorem proposition_3_4 {n : ℕ} (s : ScoringRule n) (g : (Fin n → ℝ) → Fin n → ℝ)
(hreg : IsRegular s) (hprop : IsProper s) (hg : IsExposure s g)
(q : Fin n → ℝ) (hq : q ∈ forecastDomain s) (x : Fin n → ℝ) (hx : x ∈ stdSimplex ℝ (Fin n)) :
0 ≤ bregman s g x q ∧ (bregman s g x q = 0 → x = q) := by sorry
end QAPooling.MaxMin
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.