B.8, p. 32 — Fenchel step: sup over λ ≥ 0 of λb − νλ log λ equals ν exp(b/ν − 1)
ProvedBootRobust.NWDual.fenchel_xlogxconvex-analysisconvex-conjugateentropyp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and . Then
with the convention , and the supremum is attained (at ).
This is the Fenchel conjugate of on ; it evaluates the inner maximization over each coordinate of the Lagrangian in the proof of Lemma 2.
Formalization Note The value at uses Lean's Real.log 0 = 0, which is the convention .
Preamble
import Mathlib import Definitions.Def_BootRobust_NWDual_Setting
Formal statement
namespace BootRobust.NWDual
/-- B.8, p. 32 (Fenchel step): for `ν > 0`, `sup_{λ ≥ 0} (λ b − ν λ log λ) = ν exp(b/ν − 1)`,
and the supremum is attained. At `λ = 0` the term `λ log λ` is `0` (Lean's `Real.log 0 = 0`). -/
theorem fenchel_xlogx (ν b : ℝ) (hν : 0 < ν) :
IsGreatest {x : ℝ | ∃ lam : ℝ, 0 ≤ lam ∧ x = lam * b - ν * lam * Real.log lam}
(ν * Real.exp (b / ν - 1)) := by sorry
end BootRobust.NWDual
Source
Bertsimas and Van Parys, Bootstrap robust prescriptive analytics, arXiv:1711.09974v2, B.8 (proof of Lemma 2), p. 32, 'The inner maximization problems over λ can be dealt with using the Fenchel conjugate of the λ ↦ λ · log λ function'
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.