No strict linearized descent at a local minimizer
ProvedVectorSpaceOpt.kkt_no_strict_linearized_descentkktlinearizationlocal-minimumstrict-feasibility
Let x₀ be feasible for the cone inequality G x ≤ₚ 0 and a local minimizer of f on that feasible set. Suppose f and G have continuous-linear Gâteaux derivatives f' and G' at x₀. Then there is no direction h that strictly lowers the linearized objective and simultaneously enters the strict interior of the linearized feasible cone:
The assertion is deliberately based on IsLocalMinOn, not a stronger global-minimum hypothesis. It is the nonintersection fact that turns local optimality into a geometric separation problem and isolates the only use of the nonlinear maps in the KKT argument.
Preamble
import Definitions.Def_VectorSpaceOpt_cone_optimization open Set
Formal statement
namespace VectorSpaceOpt
/-- The nonintersection step in Luenberger's proof of §9.4, Theorem 1. -/
theorem kkt_no_strict_linearized_descent
{X Z : Type*}
[NormedAddCommGroup X] [NormedSpace ℝ X]
[NormedAddCommGroup Z] [NormedSpace ℝ Z]
(P : ConvexCone ℝ Z) (f : X → ℝ) (G : X → Z) (x₀ : X)
(f' : X →L[ℝ] ℝ) (G' : X →L[ℝ] Z)
(hf : HasGateauxDerivAt f f' x₀) (hG : HasGateauxDerivAt G G' x₀)
(hmin : IsLocalMinOn f {x | coneLE P (G x) 0} x₀)
(hfeas : coneLE P (G x₀) 0) :
¬ ∃ h : X, f' h < 0 ∧ coneLT P (G x₀ + G' h) 0 := by
sorry
end VectorSpaceOptSource
David G. Luenberger, Optimization by Vector Space Methods (Wiley, 1969), Chapter 9, §9.4, first part of the proof of Theorem 1, printed p. 250 (physical PDF p. 268). Scan: https://sites.science.oregonstate.edu/~show/old/142_Luenberger.pdf