Proof of Theorem A.1, p. 26 (corrected) — if R(λ̂) > α then R̂⁺(λ†) < α at λ† = sup{λ ∈ Λ : R(λ) > α}
ProvedRiskControl.UCB.ucb_lt_at_failure_pointcalibrationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1risk-controlupper-confidence-bound
Let be closed, let be continuous on , let , and let be any function (one realization of the upper confidence bound ). Let
as in (4). If this calibration set is nonempty and , then the set is nonempty and, with ,
This is a purely deterministic statement: whenever UCB calibration selects a parameter with risk above , the upper confidence bound is already below at the fixed point . Together with , the failure of calibration is contained in the failure of the pointwise bound (3) at .
Formalization Note The printed proof argues "" and then "" with ; that route needs , false for non-interval (see the failure-point milestone). This statement uses instead. No monotonicity of is assumed.
Preamble
import Mathlib import Definitions.Def_RiskControl_UCB_Setting
Formal statement
namespace RiskControl.UCB
/-- Proof of Theorem A.1, p. 26, corrected: deterministically, for one realization `r` of
`λ ↦ R̂⁺(λ)`, if the calibration set of (4) is nonempty and `R(λ̂) > α`, then the set
`{λ ∈ Λ : R(λ) > α}` is nonempty and `R̂⁺(λ†) < α` at its supremum `λ†`. -/
theorem ucb_lt_at_failure_point (Λ : Set EReal) (hΛ : IsClosed Λ) (R : EReal → ℝ)
(hcont : ContinuousOn R Λ) (r : EReal → ℝ) (α : ℝ)
(hS : (calSet Λ r α).Nonempty) (hfail : α < R (lambdaHat Λ r α)) :
{l ∈ Λ | α < R l}.Nonempty ∧ r (sSup {l ∈ Λ | α < R l}) < α := by sorry
end RiskControl.UCB
Source
Bates, Angelopoulos, Lei, Malik & Jordan, arXiv:2101.02703v3, Proof of Theorem A.1, p. 26, second and third sentences, corrected (λ† for λ*)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.