Lemma 18 — an explicit solution of
ProvedOptimalBAI.TrackStop.lambert_inversioninequalitylambert-wp2o-batch-p200bp2o-gran-per-chapterp2o-plan-paperp2o-v1real-analysis
Let and let be constants with . If
is positive, then
The lemma inverts the inequality up to the constant , as bounds on the Lambert function do; it converts the threshold into an explicit time after which Chernoff's rule has stopped.
Formalization Note The paper states the lemma for all . Two hypotheses are added so that the printed expressions are defined: (otherwise is the logarithm of a non-positive number) and (otherwise is undefined; happens for slightly above ). Every use in the paper has .
Preamble
import Mathlib
Formal statement
namespace OptimalBAI.TrackStop
/-- Lemma 18 (Garivier–Kaufmann, arXiv:1602.04589v2, p. 27). For `α ∈ [1, e/2]` and constants
`c₁, c₂ > 0`, the number
`x = (α/c₁) [log(c₂ e / c₁^α) + log log(c₂ / c₁^α)]` satisfies `c₁ x ≥ log(c₂ x^α)`.
Two hypotheses are added so that the printed expressions are defined: `c₂ / c₁^α > 1` (so that
`log log(c₂/c₁^α)` is the logarithm of a positive number) and `x > 0` (so that `x^α` is defined). -/
theorem lambert_inversion (α : ℝ) (hα : α ∈ Set.Icc (1 : ℝ) (Real.exp 1 / 2))
(c₁ c₂ : ℝ) (hc₁ : 0 < c₁) (hc₂ : 0 < c₂) (hy : c₁ ^ α < c₂)
(x : ℝ)
(hx : x = α / c₁ * (Real.log (c₂ * Real.exp 1 / c₁ ^ α) + Real.log (Real.log (c₂ / c₁ ^ α))))
(hx_pos : 0 < x) :
Real.log (c₂ * x ^ α) ≤ c₁ * x := by sorry
end OptimalBAI.TrackStop
Source
Garivier, Kaufmann, Optimal Best Arm Identification with Fixed Confidence, arXiv:1602.04589v2, p. 27, Lemma 18
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.