Lemma A.3, p. 29 — convex Lipschitz loss against a norm penalty
ProvedRegMT.Classif.lemma_A_3classificationconvex-analysislipschitzp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be convex and Lipschitz continuous. Let be a continuous linear functional on a finite-dimensional real normed space, let be a point of that space, and let . Then
This evaluates each feature supremum in the label-split reformulation. The dual norm is the operator norm induced by the same norm used in the transport cost.
Preamble
import Mathlib import Definitions.Def_RegMT_Classif_Model
Formal statement
namespace RegMT.Classif
/-- Lemma A.3, p. 29: the supremum of a convex Lipschitz loss minus a norm
penalty is either its value at the center or positive infinity. -/
theorem lemma_A_3 {V : Type*} [NormedAddCommGroup V] [NormedSpace ℝ V]
[FiniteDimensional ℝ V] (L : ℝ → ℝ)
(hconv : ConvexOn ℝ Set.univ L)
(hLip : ∃ K : NNReal, LipschitzWith K L)
(β : V →L[ℝ] ℝ) (ζhat : V) (γ : ℝ) (hγ : 0 < γ) :
(⨆ ζ : V, ((L (β ζ) - γ * ‖ζ - ζhat‖ : ℝ) : EReal)) =
if (WassersteinDRO.Duality.lipschitzModulus L).toReal * ‖β‖ ≤ γ
then ((L (β ζhat) : ℝ) : EReal) else ⊤ := by sorry
end RegMT.Classif
Source
Shafieezadeh-Abadeh, Kuhn & Mohajerin Esfahani, Regularization via Mass Transportation, arXiv:1710.10016v3, pp. 29–30, Lemma A.3
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.