Proof of Theorem 3.5 — at a dual optimal , a minimizer of has and maximizes in
ProvedPenaltyLag.Exact.stationary_at_dual_optimalLet be a nonempty convex subset of a real vector space and convex functions on , defining the convex program (P). Let , let be an optimal solution to the dual problem (maximize over ), and let minimize over . Then the function is differentiable at with
and maximizes over all .
In the paper this is the step of the proof of Theorem 3.5 that follows from Theorem 3.2: , and the concavity of turns stationarity into maximality, so that is a saddle point.
Formalization Note The standing assumption of p. 358 is a hypothesis. "Optimal solution to " means . The gradient statement is HasGradientAt … 0 ȳ in EuclideanSpace ℝ (Fin m), which asserts differentiability as well as the value of the gradient. No sign restriction is placed on .
import Mathlib import Definitions.Def_PenaltyLag_Exact_Basic
namespace PenaltyLag.Exact
/-- Proof of Theorem 3.5, pp. 362–363: if ȳ is an optimal solution to (D_r), r > 0, and
x̄ ∈ X minimizes L_r(·, ȳ) over X, then ∇_y L_r(x̄, ȳ) = 0 and ȳ maximizes L_r(x̄, ·) over ℝ^m. -/
theorem stationary_at_dual_optimal {E : Type*} [AddCommGroup E] [Module ℝ E] {m : ℕ}
(X : Set E) (hX : Convex ℝ X) (hXne : X.Nonempty)
(f₀ : E → ℝ) (f : Fin m → E → ℝ) (hf₀ : ConvexOn ℝ X f₀) (hf : ∀ i, ConvexOn ℝ X (f i))
(r : ℝ) (hr : 0 < r) (ybar : PenaltyLag.Asymptotic.Mult m) (hybar : IsDualOptimal X f₀ f r ybar)
(xbar : E) (hxbar : xbar ∈ X) (hmin : ∀ x ∈ X, PenaltyLag.Asymptotic.Lr f₀ f r xbar ybar ≤ PenaltyLag.Asymptotic.Lr f₀ f r x ybar) :
HasGradientAt (fun y => PenaltyLag.Asymptotic.Lr f₀ f r xbar y) 0 ybar ∧ ∀ y, PenaltyLag.Asymptotic.Lr f₀ f r xbar y ≤ PenaltyLag.Asymptotic.Lr f₀ f r xbar ybar := by sorry
end PenaltyLag.Exact
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.