Proof of Theorem 6 (Appendix L), p. 34 — increasing E[F_{i,k}] cannot decrease the probability of winning overall
ProvedForecastComp.IELF.win_prob_transfer_monoforecasting-competitionlottery-mechanismp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and let be tables of event-winner probabilities. Fix an event , a forecaster and , and suppose that differs from only on event , where mass moves to uniformly from all others:
and that for every event and every forecaster . Then
where is the probability that has the most event wins (ties broken uniformly at random) when the event winners are drawn independently with .
Combined with the uniform-transfer property of ELF, this says that raising 's expected winning probability on one event cannot lower 's chance of winning overall.
Preamble
import Mathlib import Definitions.Def_ForecastComp_IELF_Model
Formal statement
namespace ForecastComp.IELF
/-- Proof of Theorem 6 (Appendix L), p. 34: with event winners drawn independently across events,
moving probability mass `ε ≥ 0` on event `k₀` toward forecaster `i`, taken uniformly
(`ε/(n−1)` each) from every other forecaster, cannot decrease `i`'s probability of being selected
by the most-wins rule with uniform tie-breaking. -/
theorem win_prob_transfer_mono {m n : ℕ} (hn : 2 ≤ n) (q q' : Fin m → Fin n → ℝ)
(k₀ : Fin m) (i : Fin n) (ε : ℝ) (hε : 0 ≤ ε)
(hother : ∀ l, l ≠ k₀ → q' l = q l)
(hi : q' k₀ i = q k₀ i + ε)
(hj : ∀ j, j ≠ i → q' k₀ j = q k₀ j - ε / ((n:ℝ) - 1))
(hnonneg : ∀ l, l ≠ k₀ → ∀ j, 0 ≤ q l j) :
winProbOf q i ≤ winProbOf q' i := by sorry
end ForecastComp.IELF
Source
Witkowski et al., arXiv:2101.01816v2, p. 34, proof of Theorem 6 (Appendix L), end of second paragraph
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.