Proof of Theorem 7 (Appendix M), p. 36 — I-ELF selects i with probability ≥ 1 − 2(n−1)e^{−mΔ²/(2(n−1)²)}
ProvedForecastComp.IELF.ielf_select_lower_boundforecasting-competitionlimit-accuracyp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , let take values in on , let the outcomes be independent with marginals , and let all reports lie in . If forecaster 's expected score exceeds every other forecaster's by at least , i.e. for all , then
In particular, the probability is at least once , which yields limit accuracy.
Formalization Note The paper writes in this display; the mechanism meant (and formalized) is I-ELF.
Preamble
import Mathlib import Definitions.Def_ForecastComp_IELF_Model
Formal statement
namespace ForecastComp.IELF
/-- Proof of Theorem 7 (Appendix M), p. 36: if forecaster `i`'s expected score exceeds every
other forecaster's by at least `Δ > 0`, then under independent outcomes I-ELF selects `i` with
probability at least `1 − 2(n−1) e^{−mΔ²/(2(n−1)²)}`. -/
theorem ielf_select_lower_bound {m n : ℕ} (hn : 2 ≤ n) (R : ℝ → Bool → ℝ)
(hbdd : IsUnitBounded R)
(θ : Fin m → ℝ) (hθ : ∀ k, θ k ∈ Set.Icc (0:ℝ) 1)
(y : Fin n → Fin m → ℝ) (hy : ∀ j k, y j k ∈ Set.Icc (0:ℝ) 1)
(i : Fin n) (Δ : ℝ) (hΔ : 0 < Δ)
(hgap : ∀ j, j ≠ i → Δ ≤ avgScore R θ (y i) - avgScore R θ (y j)) :
1 - 2 * ((n:ℝ) - 1) * Real.exp (-((m:ℝ) * Δ ^ 2) / (2 * ((n:ℝ) - 1) ^ 2)) ≤
ielfSelProbIndep R θ y i := by sorry
end ForecastComp.IELF
Source
Witkowski et al., arXiv:2101.01816v2, p. 36, proof of Theorem 7 (Appendix M), final display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.