Proof of Theorem 9, p. 20 — h(x̄) = 0 and 0 ∈ M(x̄) give λ ∈ [0,1], ḡ ∈ ∂f(x̄), ĝ ∈ ∂h(x̄) with λḡ + (1−λ)ĝ = 0
ProvedMifflinSemismooth.Optimality.exists_convex_combination_zerofritz-johngeneralized-gradientnonsmooth-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let both be Lipschitz on a ball about , suppose and . Then there exist , and such that
This rewrites the stationarity condition on the boundary of the feasible set as a Fritz John type multiplier condition, which the rest of the proof of Theorem 9 splits into the cases and .
Formalization Note The two Lipschitz hypotheses are the part of the semiconvexity hypothesis of Theorem 9 that this step uses: they make and nonempty (Proposition 1(a)). Without them, with would not produce .
Preamble
import Mathlib import Definitions.Def_MifflinSemismooth_Optimality_Setting
Formal statement
namespace MifflinSemismooth.Optimality
/-- Mifflin (1976), §5, proof of Theorem 9, p. 20: if `h(x̄) = 0` and `0 ∈ M(x̄)`, there exist
`λ ∈ [0, 1]`, `ḡ ∈ ∂f(x̄)` and `ĝ ∈ ∂h(x̄)` with `λ ḡ + (1 - λ) ĝ = 0`. -/
theorem exists_convex_combination_zero {n : ℕ} (f h : EuclideanSpace ℝ (Fin n) → ℝ)
(xbar : EuclideanSpace ℝ (Fin n)) (hf : MifflinSemismooth.Extremal.LipschitzNear f xbar) (hh : MifflinSemismooth.Extremal.LipschitzNear h xbar)
(hzero : h xbar = 0) (h0 : (0 : EuclideanSpace ℝ (Fin n)) ∈ Mmap f h xbar) :
∃ lam ∈ Set.Icc (0 : ℝ) 1, ∃ gbar ∈ MifflinSemismooth.Extremal.genGrad f xbar, ∃ ghat ∈ MifflinSemismooth.Extremal.genGrad h xbar,
lam • gbar + (1 - lam) • ghat = 0 := by sorry
end MifflinSemismooth.Optimality
Source
Mifflin, Semismooth and semiconvex functions in constrained optimization, IIASA Research Report RR-76-21 (December 1976), p. 20, §5, proof of Theorem 9, case h(x̄) = 0
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.