(21), p. 2325 — 0 ≤ F_β(w) − max_j(w_j − y_j) ≤ β⁻¹ log p, for every β > 0
ProvedHighDimCLT.MultBoot.eq_21log-sum-expp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1smooth-max
Let , and , and define the smooth max
Then for every ,
The smooth max is a differentiable surrogate for the indicator of ; this sandwich controls the error of the substitution.
Formalization Note The page states (21) with , so that ; the proof of Theorem 4.1 (p. 2342) uses it for every , which is what is stated. The maximum over is a supremum over the finite set .
Preamble
import Mathlib import Definitions.Def_HighDimCLT_MultBoot_Setting
Formal statement
namespace HighDimCLT.MultBoot
open MeasureTheory ProbabilityTheory
universe u
/-- **Display (21)**, App. B, p. 2325 (used for every `β > 0` in App. E.2, p. 2342): the smooth
max `F_β(w) = β^{-1} log ∑_j exp(β(w_j − y_j))` satisfies
`0 ≤ F_β(w) − max_j (w_j − y_j) ≤ β^{-1} log p` for all `w ∈ ℝ^p`. -/
theorem eq_21 (p : ℕ) (hp : 3 ≤ p) (β : ℝ) (hβ : 0 < β) (y w : EuclideanSpace ℝ (Fin p)) :
0 ≤ Fβ β y w - (⨆ j : Fin p, (w j - y j)) ∧
Fβ β y w - (⨆ j : Fin p, (w j - y j)) ≤ β⁻¹ * Real.log p := by sorry
end HighDimCLT.MultBoot
Source
Chernozhukov, Chetverikov and Kato, Central limit theorems and bootstrap in high dimensions, Ann. Probab. 45 (2017), p. 2325, App. B, display (21); used for β > 0 at p. 2342, App. E.2
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.