Theorem 4.1, p. 14 — the QA pool uniquely maximizes the worst-case profit min_j u(p; j), attained at every j with p*_j > 0, non-negative and positive unless all reports agree
ProvedQAPooling.MaxMin.theorem_4_1Let be a regular proper scoring rule on outcomes such that each and the expected reward are continuous on , and let be its exposure function. Fix forecasts with non-negative weights adding to , and define the agent's profit
Let be the quasi-arithmetic pool. Then:
- is uniquely maximized at : every report with has ;
- the minimum over outcomes is attained simultaneously by all with : for each such ;
- ;
- unless all reports with positive weights are equal.
Read as a contract: an agent paid by who sub-contracts experts, paying expert the amount , guarantees itself the largest possible worst-case profit by reporting the QA pool of the experts' forecasts, and no other report guarantees as much. This is the paper's max-min justification of quasi-arithmetic pooling.
Formalization Note The QA pool is a hypothesis: with a subgradient of at relative to ; its existence and uniqueness is the theorem qaPool_existsUnique. Regularity and the two continuity assumptions are the paper's standing assumptions of pp. 9–10. Scores live in the extended reals, so when ; the minimum over outcomes is an infimum over the finitely many outcomes, and follows from . "Unless all positive-weight reports are equal" is read as: if two reports with positive weights differ, the minimum is positive.
import Mathlib import Definitions.Def_QAPooling_MaxMin_Model
namespace QAPooling.MaxMin
/-- Theorem 4.1 (p. 14). For a regular, proper, continuous scoring rule `s` with exposure function
`g`, forecasts `p_i ∈ D` and non-negative weights `w_i` adding to 1, let `p*` be their QA pool and
`u(q; j) = s(q; j) − Σ_i w_i s(p_i; j)`. Then:
1. every report `q ∈ Δⁿ` other than `p*` has `min_j u(q; j) < min_j u(p*; j)`;
2. `u(p*; j) = min_j u(p*; j)` for every `j` with `p*_j > 0`;
3. `min_j u(p*; j) ≥ 0`;
4. `min_j u(p*; j) > 0` unless all reports with positive weight are equal. -/
theorem theorem_4_1 {n m : ℕ} (s : ScoringRule n) (g : (Fin n → ℝ) → Fin n → ℝ)
(hreg : IsRegular s) (hprop : IsProper s)
(hcont : ∀ j, ContinuousOn (fun p => s p j) (stdSimplex ℝ (Fin n)))
(hGcont : ContinuousOn (expectedReward s) (stdSimplex ℝ (Fin n)))
(hg : IsExposure s g)
(p : Fin m → Fin n → ℝ) (hp : ∀ i, p i ∈ forecastDomain s)
(w : Fin m → ℝ) (hw0 : ∀ i, 0 ≤ w i) (hw1 : ∑ i, w i = 1)
(pstar : Fin n → ℝ) (hpstar : IsQAPool s g p w pstar) :
(∀ q ∈ stdSimplex ℝ (Fin n), q ≠ pstar → minProfit s p w q < minProfit s p w pstar) ∧
(∀ j, 0 < pstar j → profit s p w pstar j = minProfit s p w pstar) ∧
0 ≤ minProfit s p w pstar ∧
((∃ i i', 0 < w i ∧ 0 < w i' ∧ p i ≠ p i') → 0 < minProfit s p w pstar) := by sorry
end QAPooling.MaxMin
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.