Lemma 17 — directional derivative of an optimal value
ProvedGenEmpLik.Coverage.lemma_17hadamard-differentiabilityoptimal-valuep2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a lower semicontinuous real objective on a nonempty closed decision set . Suppose some nonempty near-minimizer set is compact. Let be continuous and bounded on , let each be bounded there, and suppose . For positive , writing , one has
This is the optimal-value directional derivative used to identify the influence function.
Formalization Note Signed measures are represented by their action on the loss class. Positive follows the directional reading used in Appendix C.2.
Preamble
import Mathlib open Filter Topology
Formal statement
namespace GenEmpLik.Coverage
/-- Lemma 17, pp. 38–40, in the action of signed measures on the loss class. -/
theorem lemma_17 {E : Type*} [MetricSpace E]
(X : Set E) (F H : E → ℝ) (Hn : ℕ → E → ℝ) (t : ℕ → ℝ)
(hXclosed : IsClosed X) (hXnonempty : X.Nonempty)
(hFlsc : LowerSemicontinuousOn F X)
(hsublevel : ∃ ε₀ : ℝ, 0 < ε₀ ∧
IsCompact {x | x ∈ X ∧ F x ≤ sInf (F '' X) + ε₀} ∧
Set.Nonempty {x | x ∈ X ∧ F x ≤ sInf (F '' X) + ε₀})
(hHcont : ContinuousOn H X)
(hHbdd : BddAbove ((fun x => |H x|) '' X))
(hHnbdd : ∀ n, BddAbove ((fun x => |Hn n x|) '' X))
(htpos : ∀ n, 0 < t n) (htzero : Tendsto t atTop (𝓝 0))
(hHn : Tendsto (fun n => ⨆ x : X, |Hn n x - H x|) atTop (𝓝 0)) :
Tendsto (fun n =>
(sInf ((fun x => F x + t n * Hn n x) '' X) - sInf (F '' X)) / t n)
atTop (𝓝 (sInf (H '' {x | x ∈ X ∧ F x = sInf (F '' X)}))) := by sorry
end GenEmpLik.Coverage
Source
Duchi, Glynn & Namkoong, Statistics of Robust Optimization: A Generalized Empirical Likelihood Approach, arXiv:1610.03425v3, p. 38, Lemma 17; proof pp. 39–40
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.