Proof of Theorem 4.1, pp. 14–15 — u(p*; j) = Σᵢ wᵢ D_G(p* ∥ pᵢ) + µⱼ
ProvedQAPooling.MaxMin.profit_identitybregman-divergencep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1proper-scoring-rulequasi-arithmetic-pooling
Let be a scoring rule with exposure function , let with weights adding to , and let . Suppose that
for some and with whenever . Then for every outcome , the agent's profit satisfies
This is the computation at the heart of the proof of Theorem 4.1: the profit at the QA pool is the same weighted Bregman sum for every outcome, plus a non-negative correction that vanishes on the support of .
Formalization Note The identity is algebraic: it needs neither properness nor , only Equation (2) at and at each and .
Preamble
import Mathlib import Definitions.Def_QAPooling_MaxMin_Model
Formal statement
namespace QAPooling.MaxMin
/-- The profit identity in the proof of Theorem 4.1 (pp. 14–15): if `p* ∈ D` and
`g(p*) = Σ_i w_i g(p_i) + λ 1_n + μ` with `μ_j = 0` whenever `p*_j > 0`, then for every outcome `j`,
`u(p*; j) = Σ_i w_i D_G(p* ∥ p_i) + μ_j`. -/
theorem profit_identity {n m : ℕ} (s : ScoringRule n) (g : (Fin n → ℝ) → Fin n → ℝ)
(hg : IsExposure s g)
(p : Fin m → Fin n → ℝ) (hp : ∀ i, p i ∈ forecastDomain s)
(w : Fin m → ℝ) (hw1 : ∑ i, w i = 1)
(pstar : Fin n → ℝ) (hD : pstar ∈ forecastDomain s)
(lam : ℝ) (mu : Fin n → ℝ)
(hkkt : ∀ j, g pstar j = ∑ i, w i * g (p i) j + lam + mu j)
(hcs : ∀ j, 0 < pstar j → mu j = 0) :
∀ j, profit s p w pstar j = ((∑ i, w i * bregman s g pstar (p i) + mu j : ℝ) : EReal) := by sorry
end QAPooling.MaxMin
Source
Neyman & Roughgarden, From Proper Scoring Rules to Max-Min Optimal Forecast Aggregation, arXiv:2102.07081v2, pp. 14–15, proof of Theorem 4.1 (chain of equalities for u(p*; j))
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.