Lemma 3.6, p. 123 — strictly convex, bounded below, ργ < 1: f*ₙ = f*ₙ₊₁, and (n ≥ 2) xₙ changes infinitely often
ProvedNelderMeadLD.Dim2.lemma_3_6Let () be strictly convex and bounded below, let the coefficients satisfy (2.1) and in addition
and let be a run of Algorithm NM on from a nondegenerate initial simplex. Then
- the next-worst and worst vertex values have a common limit: ;
- if , there are infinitely many iterations with .
The lemma is the input to Lemma 5.1, where it guarantees that the next-worst vertex of a triangle keeps moving.
Formalization Note is 0-based index (nextWorst n), is Fin.last n. Part 1 asserts the existence of a common limit of and . The paper's statement does not name the start; its proof uses Lemma 3.4, which assumes a nondegenerate start, so that hypothesis is kept. Part 2 is restricted to : for the next-worst vertex is the best vertex, and the printed claim is false (take , , , start : every iteration is an accepted inside contraction and for all ); the proof's contradiction uses the strict inequality , which the paper notes holds only for . The paper applies the lemma with .
import Mathlib import Definitions.Def_NelderMeadLD_Dim2_Algorithm open Filter Topology
namespace NelderMeadLD.Dim2
theorem lemma_3_6 {n : ℕ} [NeZero n] (f : E n → ℝ) (ρ χ γ σ : ℝ) (hpar : NelderMeadLD.Conv1D.ParamsOK ρ χ γ σ)
(Δ : ℕ → Fin (n + 1) → E n) (hrun : IsNMRun f ρ χ γ σ Δ) (hnd : Nondegenerate (Δ 0))
(hf : StrictConvexOn ℝ Set.univ f) (hbdd : BddBelow (Set.range f)) (hργ : ρ * γ < 1) :
(∃ L : ℝ, Tendsto (fun k => f (Δ k (nextWorst n))) atTop (𝓝 L) ∧
Tendsto (fun k => f (Δ k (Fin.last n))) atTop (𝓝 L)) ∧
(2 ≤ n → ∀ N : ℕ, ∃ k ≥ N, Δ (k + 1) (nextWorst n) ≠ Δ k (nextWorst n)) := by sorry
end NelderMeadLD.Dim2
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.