B.2, p. 27 — B(D(λ), D_tr) ≤ λB(D, D_tr) + (1−λ)B(D_tr, D_tr) < r along D(λ) = λD + (1−λ)D_tr
ProvedBootRobust.Optimality.bootDist_segment_leconvexityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1relative-entropy
Let with having all coordinates positive, let and . Then
- ;
- ;
- if moreover with and , then
This is the convexity step of the proof of Proposition 1: moving from towards strictly lowers the bootstrap distance below .
Formalization Note is extended-real valued; with every value here is finite. The condition in item 3 is needed for the strict inequality and holds in Proposition 1.
Preamble
import Mathlib import Definitions.Def_BootRobust_Optimality_Setting
Formal statement
namespace BootRobust.Optimality
/-- B.2, p. 27, the convexity step: along the segment `D(λ) = λ D + (1 - λ) D_tr`, `λ ∈ [0, 1]`,
`B(D(λ), D_tr) ≤ λ B(D, D_tr) + (1 - λ) B(D_tr, D_tr)`, and `B(D_tr, D_tr) = 0`; hence, if
`B(D, D_tr) = r` with `r > 0` and `λ ∈ (0, 1)`, then `B(D(λ), D_tr) < r`. -/
theorem bootDist_segment_le {ι : Type*} [Fintype ι] [DecidableEq ι]
(D Dtr : ι → ℝ) (hD : D ∈ stdSimplex ℝ ι) (hDtr : Dtr ∈ stdSimplex ℝ ι)
(hDtrpos : ∀ i, 0 < Dtr i) (lam : ℝ) (hlam : lam ∈ Set.Icc (0 : ℝ) 1) :
BootRobust.Perf.bootDist (lam • D + (1 - lam) • Dtr) Dtr ≤
(lam : EReal) * BootRobust.Perf.bootDist D Dtr + ((1 - lam : ℝ) : EReal) * BootRobust.Perf.bootDist Dtr Dtr ∧
BootRobust.Perf.bootDist Dtr Dtr = 0 ∧
∀ r : ℝ, BootRobust.Perf.bootDist D Dtr = (r : EReal) → 0 < r → lam ∈ Set.Ioo (0 : ℝ) 1 →
BootRobust.Perf.bootDist (lam • D + (1 - lam) • Dtr) Dtr < (r : EReal) := by sorry
end BootRobust.Optimality
Source
Bertsimas and Van Parys, Bootstrap robust prescriptive analytics, arXiv:1711.09974v2, B.2 (proof of Proposition 1), p. 27, "From convexity of B we have …"
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.