The gradient of the augmented-Lagrangian inequality term is locally Lipschitz
ProvedNonsmoothNewton.AugLagrangian.eta_fderiv_locallyLipschitzaugmented-lagrangiancalculusnonsmooth-analysissource-faithful-child
Let and be , and assume is . Then the map , valued in the continuous linear functionals on and measured in operator norm, is locally Lipschitz on . On the two open regions separated by the surface this map equals the two smooth expressions and ; these agree on the surface because there, and both are smooth there because is continuous. This is the local Lipschitz continuity of asserted in item 3 of Theorem 4.1, the regularity that makes the semismooth Newton method applicable.
Formal statement
import Mathlib
import Definitions.Def_NonsmoothNewton_Shared_clarkeJac
import Definitions.Def_NonsmoothNewton_AugLagrangian_SemismoothAt
import Definitions.Def_NonsmoothNewton_AugLagrangian_augLagrangian
namespace NonsmoothNewton.AugLagrangian
/-- Child of `eta_C1_grad_locLip`, third step of Qi-Sun (1993), Section 4, proof of Theorem 4.1,
pp. 363-364: $\nabla\eta$ is locally Lipschitz (item 3 of the theorem, the sentence following the
display of the gradient).
The gradient is viewed as a linear functional and bounded in operator norm. Since $g$ is $C^2$,
$\nabla g=fderiv\ \mathbb{R}\ g$ is $C^1$ and therefore locally Lipschitz, and the gradient of $\eta$
differs from the continuous function $(x,s)\mapsto g(x)$ by the term $(s+rg(x))\cdot\Psi$ whose
factor $\Psi$ is bounded on bounded sets. -/
theorem eta_fderiv_locallyLipschitz {n : ℕ} (r : ℝ) (hr : 0 < r)
(g : EuclideanSpace ℝ (Fin n) → ℝ) (hg : ContDiff ℝ 2 g)
(hη : ContDiff ℝ 1 (eta r g)) :
LocallyLipschitz (fun z => fderiv ℝ (eta r g) z) := by sorry
end NonsmoothNewton.AugLagrangianSource
Qi, Sun, A nonsmooth version of Newton's method, Math. Programming 58 (1993), pp. 363-364, Section 4, proof of Theorem 4.1, the sentence following the display of (item 3 of the theorem).