Proof of Proposition 5, p. 26 (corrected) — if R̂⁺_WSR < R then maxᵢ 𝒦ᵢ(R) ≥ 1/δ
ProvedRiskControl.WSR.ucb_lt_imp_max_gebetting-confidence-boundp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1risk-controlupper-confidence-bound
Let , , let be a sample with every , and let . If , then
that is, for some (see the definition RiskControl.WSR.Process for and ).
Applied at the true mean , this places the failure event inside the event controlled by Ville's inequality.
Formalization Note The paper writes "if , then ", with a stray ""; the intended claim, stated here, is the event . The comparison is in the extended reals. The hypothesis holds for the paper's , the mean of a nonnegative loss.
Preamble
import Mathlib import Definitions.Def_RiskControl_WSR_Process
Formal statement
namespace RiskControl.WSR
/-- Proof of Proposition 5, p. 26 (arXiv:2101.02703v3), with the page's stray `P(` removed: for
a sample of losses in `[0, 1]` and `R ≥ 0`, if `R̂⁺_WSR < R` then `max_{i=1,…,n} 𝒦_i(R; λ) ≥ 1/δ`. -/
theorem ucb_lt_imp_max_ge (n : ℕ) (δ : ℝ) (ω : Fin n → ℝ)
(hω : ∀ j, ω j ∈ Set.Icc (0 : ℝ) 1) (R : ℝ) (hR : 0 ≤ R) :
wsrUCB n δ ω < (R : EReal) → ∃ i ∈ Finset.Icc 1 n, 1 / δ ≤ K n δ ω i R := by sorry
end RiskControl.WSR
Source
Bates, Angelopoulos, Lei, Malik & Jordan, arXiv:2101.02703v3, Proof of Proposition 5, p. 26, last sentence before the third display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.