B.2, p. 27 — a linear functional 𝔼_D[G] separates the R-ball from a convex open set on which R > r
ProvedBootRobust.Optimality.separationLet be a distribution distance function on , let and . Let be nonempty, convex and open in , with every point of having all coordinates positive, and suppose for every . Then there exist and with
where .
This is the separation step of the proof of Proposition 1: the -ball is convex because is convex in its first argument, it is disjoint from , and an open convex set can be strictly separated from a disjoint convex set.
Formalization Note The paper takes convex "without loss of generality"; here convexity and nonemptiness are hypotheses. The hypothesis that lies in the relative interior of the simplex is a disclosed restriction: in the proof is a small neighbourhood of a point , which has full support because does.
import Mathlib import Definitions.Def_BootRobust_Optimality_Setting
namespace BootRobust.Optimality
/-- B.2, p. 27: the `R`-ball `{D ∈ 𝒟ₙ : R(D, D_tr) ≤ r}` and a convex, relatively open, nonempty set
`N` on which `R(·, D_tr) > r` are separated by a linear functional `D ↦ 𝔼_D[G] = ∑ i, D i * G i`:
`𝔼_D[G] ≤ a < 𝔼_{D'}[G]` for every `D` in the ball and every `D' ∈ N`. The set `N` lies in the
relative interior of the simplex (all coordinates positive). -/
theorem separation {ι : Type*} [Fintype ι] [DecidableEq ι]
(R : (ι → ℝ) → (ι → ℝ) → EReal) (hR : IsDistributionDistance R)
(Dtr : ι → ℝ) (hDtr : Dtr ∈ stdSimplex ℝ ι) (r : ℝ)
(N : Set (ι → ℝ)) (hN : relOpen N) (hNconv : Convex ℝ N) (hNne : N.Nonempty)
(hNpos : ∀ D' ∈ N, ∀ i, 0 < D' i) (hRN : ∀ D' ∈ N, (r : EReal) < R D' Dtr) :
∃ G : ι → ℝ, ∃ a : ℝ,
(∀ D ∈ stdSimplex ℝ ι, R D Dtr ≤ (r : EReal) → ∑ i, D i * G i ≤ a) ∧
(∀ D' ∈ N, a < ∑ i, D' i * G i) := by sorry
end BootRobust.Optimality
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.