v21 Theorem 1: complete polar counterexample

Lean contributors (project-level attribution): Zichen Wang.

Source header · Attribution source

public sectionopen DFP.Polarnamespace DFPWolfe.V21

thm:main. For every Unknown identifier `n`sorry ≥ 2 : Propn≥2, , , and , the original v21 polar construction gives a canonical matrix-valued inverse DFP iteration with positive step lengths and strong Wolfe at every index. The standard certificate contains its C² objective and the fixed global Hessian bounds Unknown identifier `I`sorry / 2 : ℕI/2 and .

Unknown identifier `CounterexampleProperties`CounterexampleProperties records gradient-norm limit exactly one, operator distance of the Hessian from identity at most ε, global half-Hölder Hessian continuity, and the unique origin minimizer. It also records failure of every β-Hölder Hessian bound on the original initial level set for , and the exact unit-circle cluster set under Unknown identifier `planeEmbedding`planeEmbedding. This is a linear isometric embedding of the source plane; Unknown identifier `planeEmbedding_range_finrank`planeEmbedding_range_finrank proves its range has dimension exactly two.

The source identities retain the concrete original polar objective, points, step lengths, gradients, and inverse Hessians, followed by the source orthogonal extension.

theorem main (n : ℕ) (hn : 2 ≤ n) (c₁ c₂ ε : ℝ) (hc₁ : 0 < c₁) (hc₁lim : c₁ < 2 / 3) (hc₂lim : 2 / 3 ≤ c₂) (hc₂ : c₂ < 1) (hε : ε ∈ Set.Ioo 0 1) : ∃ w : DFP.StrongWolfeCounterexample (Fin n) (1 / 2) (3 / 2) c₁ c₂, OrthogonalExtension.CounterexampleProperties ε w (OrthogonalExtension.planeEmbedding n hn) ∧ ∃ νStar > baseAngle, ∃ κ > 0, ∃ j₀ : ℕ, 0 < j₀ ∧ ∃ c : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂, PlanarCounterexample.SourceIdentity (Parameters.of j₀ νStar) κ c ∧ OrthogonalExtension.FinExtensionIdentity n hn c w := by exact OrthogonalExtension.polarCounterexample n hn c₁ c₂ ε hc₁ hc₁lim hc₂lim hc₂ hε.1

Helper for thm:main and eq:open-question: the new v21 polar witness disproves global weak-Wolfe convergence for every pair in the source parameter range. The convergence predicate quantifies over ordinary admissible inverse DFP iterations.

theorem not_globalWeakWolfeConvergenceAt (c₁ c₂ : ℝ) (hc₁ : 0 < c₁) (hc₁lim : c₁ < 2 / 3) (hc₂lim : 2 / 3 ≤ c₂) (hc₂ : c₂ < 1) : ¬ DFP.GlobalWeakWolfeConvergenceAt c₁ c₂ := by have hhalf : (1 / 2 : ℝ) ∈ Set.Ioo 0 1 := by norm_num have hbounds : (1 / 2 : ℝ) ≤ 3 / 2 := by norm_num obtain ⟨w, _, _⟩ := main 2 le_rfl c₁ c₂ (1 / 2) hc₁ hc₁lim hc₂lim hc₂ hhalf exact DFP.not_globalWeakWolfeConvergenceAt_of_counterexample w.toWolfeCounterexample le_rfl hhalf.1 hbounds

Helper for thm:main and eq:open-question: the same new polar construction disproves the initial-level-set formulation of ordinary weak-Wolfe DFP convergence, with the source Wolfe parameter range unchanged.

theorem not_levelSetGlobalWeakWolfeConvergenceAt (c₁ c₂ : ℝ) (hc₁ : 0 < c₁) (hc₁lim : c₁ < 2 / 3) (hc₂lim : 2 / 3 ≤ c₂) (hc₂ : c₂ < 1) : ¬ DFP.LevelSetGlobalWeakWolfeConvergenceAt c₁ c₂ := by have hhalf : (1 / 2 : ℝ) ∈ Set.Ioo 0 1 := by norm_num have hbounds : (1 / 2 : ℝ) ≤ 3 / 2 := by norm_num obtain ⟨w, _, _⟩ := main 2 le_rfl c₁ c₂ (1 / 2) hc₁ hc₁lim hc₂lim hc₂ hhalf exact DFP.not_levelSetGlobalWeakWolfeConvergenceAt_of_counterexample w.toWolfeCounterexample le_rfl hhalf.1 hbounds
end DFPWolfe.V21