Closed restricted convex functions equal their biconjugates
ProvedVectorSpaceOpt.fenchel_biconjugate_onLet be nonempty, let be convex on , and suppose the restricted epigraph is closed in . Then, at every ,
The supremum ranges only over continuous linear functionals whose restricted conjugate is finite. This is the pointwise-on- content of Luenberger's equality between a closed epigraph and its biconjugate epigraph; it avoids assigning an arbitrary real value outside . The theorem is a reusable representation of a closed convex function by continuous affine minorants and a central separation milestone for the mission.
import Definitions.Def_VectorSpaceOpt_restrictedConvexConjugate
namespace VectorSpaceOpt
/-- Luenberger, Chapter 7, §7.10, Proposition 2, restricted to points of `C`. -/
theorem fenchel_biconjugate_on
{X : Type*} [NormedAddCommGroup X] [NormedSpace ℝ X]
(C : Set X) (f : X → ℝ)
(hC : C.Nonempty) (hf : ConvexOn ℝ C f)
(hclosed : IsClosed {p : X × ℝ | p.1 ∈ C ∧ f p.1 ≤ p.2}) :
∀ x ∈ C,
f x = sSup
((fun ℓ : X →L[ℝ] ℝ ↦
ℓ x - restrictedConvexConjugate C f ℓ) ''
{ℓ : X →L[ℝ] ℝ |
BddAbove ((fun y : X ↦ ℓ y - f y) '' C)}) := by sorry
end VectorSpaceOptRead-back
What the Lean code literally says, in plain math · gpt-5
For every real normed vector space , nonempty set , and function , assume is convex on and the restricted epigraph is closed. Then for every , equals the real supremum, over all continuous real-linear functionals for which is bounded above, of . Each inner supremum is the usual finite real supremum because is nonempty and the indexing condition supplies boundedness. The outer operation is nevertheless the total real : the statement gives no separate nonemptiness or bounded-above hypothesis for its set of displayed values, so an empty or unbounded outer value set would receive the totalized value , rather than an extended-real infinity. The equality is asserted only for .
Confirmed by the mission captain (proposal self-audit).