Theorems 6 and 7, p. 14 — I-ELF is strictly incentive compatible under belief independence and limit accurate for independent events
ProvedForecastComp.IELF.ielf_ic_and_limit_accurateLet and let be a strictly proper scoring rule with values in on reports in , measurable in the report. Then the Independent-Event Lotteries mechanism — run an ELF lottery (1) on every event, then select the forecaster with the most event wins, breaking ties uniformly at random — has both of the following properties.
- Strict incentive compatibility under belief independence (Theorem 6). For every number of events such that, if , for all and : for every forecaster , every belief-independent joint law of the outcomes and the others' reports with , and every with ,
- Limit accuracy for independent events (Theorem 7). For every and there is such that for all , all independent outcome laws with marginals , all reports and every forecaster whose expected score exceeds every other forecaster's by more than ,
This is the paper's main result: a forecasting competition with a single prize that is both truthful and, as the number of events grows, selects the most accurate forecaster with probability tending to one.
Formalization Note The two theorems are stated as one conjunction. The condition for is part of the mechanism's definition (p. 14, footnote 10). Belief independence is a product of per-event joint laws, and independent outcome laws are given by their marginals.
import Mathlib import Definitions.Def_ForecastComp_IELF_Model
namespace ForecastComp.IELF
/-- Theorems 6 and 7, p. 14: for a strictly proper scoring rule `R` with values in `[0,1]`
(measurable in the report), I-ELF is strictly incentive compatible under belief independence for
every `m ≥ 1` (when `R < 1` on reports in `[0,1]` for `m ≥ 2`, as the mechanism requires), and it
is limit accurate for independent events. -/
theorem ielf_ic_and_limit_accurate {n : ℕ} (hn : 2 ≤ n) (R : ℝ → Bool → ℝ)
(hprop : IsStrictlyProper R) (hbdd : IsUnitBounded R)
(hmeas : ∀ x, Measurable (fun y => R y x)) :
(∀ m : ℕ, 1 ≤ m → (2 ≤ m → ∀ y ∈ Set.Icc (0:ℝ) 1, ∀ x, R y x < 1) →
StrictlyICUnderBeliefIndep (ielfSelProb (m := m) (n := n) R)) ∧
IsLimitAccurateIndep R (fun m θ y i => ielfSelProbIndep (m := m) (n := n) R θ y i) := by sorry
end ForecastComp.IELF
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.