Strong-Wolfe counterexamples with a Hölder Hessian

Lean authors: Zichen Wang.

Source header

The same realization satisfies the Wolfe conditions, the exact uniform Hessian bounds, nonconvergence, and global one-half Hölder continuity of its Hessian.

public sectionopen Filteropen scoped Topologynamespace DFP

Helper for thm:main: the planar strong-Wolfe construction has a globally one-half Hölder Hessian, with all properties attached to the same objective.

theorem existsPlanarStrongWolfeCounterexampleHolder {c₁ c₂ : ℝ} (hc₁_pos : 0 < c₁) (hc₁_lt_two_thirds : c₁ < 2 / 3) (hc₂_ge_two_thirds : (2 / 3 : ℝ) ≤ c₂) (hc₂_lt_one : c₂ < 1) : ∃ c : StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂, ∃ C > 0, ∀ x y, ‖hessian c.iteration.objective x - hessian c.iteration.objective y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ) := by apply existsPlanarStrongWolfeCounterexampleWithProperty (fun f _ ↦ ∃ C > 0, ∀ x y, ‖hessian f x - hessian f y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ)) _ hc₁_pos hc₁_lt_two_thirds hc₂_ge_two_thirds hc₂_lt_one intro curve obtain ⟨η, hη, C, hC, hHolder⟩ := PlanarConvergence.realizedObjectiveHessianHolder curve refine ⟨η, hη, ?_⟩ intro ε₀ hε₀ dsimp only intro Clim hClim Glim hGlim hGlimLimit refine ⟨C, hC, ?_⟩ simpa only [hessian_def, EuclideanPlane.hessian_def] using hHolder ε₀ hε₀ Clim hClim Glim hGlim hGlimLimit

Helper for cor:identity-initialization: a continuous linear change of coordinates preserves existence of a positive global one-half Hölder constant.

theorem existsHessianHolderPullback {E F : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E] [NormedAddCommGroup F] [InnerProductSpace ℝ F] [CompleteSpace F] {f : F → ℝ} (hf : ContDiff ℝ 2 f) (L : E ≃L[ℝ] F) (h : ∃ C > 0, ∀ x y, ‖hessian f x - hessian f y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ)) : ∃ C > 0, ∀ x y, ‖hessian (f ∘ L) x - hessian (f ∘ L) y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ) := by obtain ⟨C, hC, h⟩ := h let D := ‖L.toContinuousLinearMap‖ ^ 2 * C * ‖L.toContinuousLinearMap‖ ^ (1 / 2 : ℝ) have hD : 0 ≤ D := by positivity have hDpos : 0 < D + 1 := by positivity have hhalf : (0 : ℝ) ≤ 1 / 2 := by norm_num refine ⟨D + 1, hDpos, ?_⟩ intro x y apply (hessianHolder_comp_continuousLinearEquiv hf L hC.le hhalf h x y).trans exact mul_le_mul_of_nonneg_right (le_add_of_nonneg_right zero_le_one) (Real.rpow_nonneg (norm_nonneg _) _)

Helper for thm:main: in every dimension at least two, one classical strong-Wolfe counterexample has the exact Hessian bounds and a globally one-half Hölder Hessian.

theorem existsStrongWolfeCounterexampleHolder_of_dimension_ge_two (n : ℕ) (hn : 2 ≤ n) {c₁ c₂ : ℝ} (hc₁_pos : 0 < c₁) (hc₁_lt_two_thirds : c₁ < 2 / 3) (hc₂_ge_two_thirds : (2 / 3 : ℝ) ≤ c₂) (hc₂_lt_one : c₂ < 1) : ∃ c : StrongWolfeCounterexample (Fin n) (1 / 2) (3 / 2) c₁ c₂, ∃ C > 0, ∀ x y, ‖hessian c.iteration.objective x - hessian c.iteration.objective y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ) := by classical obtain ⟨c, C, hC, hHolder⟩ := existsPlanarStrongWolfeCounterexampleHolder hc₁_pos hc₁_lt_two_thirds hc₂_ge_two_thirds hc₂_lt_one have hlower : (1 / 2 : ℝ) ≤ 1 := by norm_num have hupper : (1 : ℝ) ≤ 3 / 2 := by norm_num obtain ⟨d, hd, _⟩ := c.orthogonalSumWithObjective (κ := Fin (n - 2)) hlower hupper have hhalf : (0 : ℝ) ≤ 1 / 2 := by norm_num have hDHolder : ∀ x y, ‖hessian d.iteration.objective x - hessian d.iteration.objective y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ) := by rw [hd, DFP.OrthogonalSum.objective_eq] exact EuclideanSpace.OrthogonalSum.Gradient.hessianHolder_objective c.objectiveContDiff hC.le hhalf hHolder have hdim : 2 + (n - 2) = n := by omega let e : Fin n ≃ Fin 2 ⊕ Fin (n - 2) := (finCongr hdim.symm).trans finSumFinEquiv.symm let Q : EuclideanSpace ℝ (Fin n) ≃ₗᵢ[ℝ] EuclideanSpace ℝ (Fin 2 ⊕ Fin (n - 2)) := LinearIsometryEquiv.piLpCongrLeft 2 ℝ ℝ e have hupperPos : (0 : ℝ) ≤ 3 / 2 := by norm_num obtain ⟨result, hresult, _⟩ := d.pullbackWithObjective Q hhalf hupperPos refine ⟨result, ?_⟩ rw [hresult] exact existsHessianHolderPullback d.objectiveContDiff Q.toContinuousLinearEquiv ⟨C, hC, hDHolder⟩

Corollary 2 (): identity-initialized classical DFP fails to converge in every dimension at least two even with a global one-half Hölder Hessian and positive ordered uniform Hessian bounds.

theorem existsMatrixIdentityLiminfStrongWolfeHolder (n : ℕ) (hn : 2 ≤ n) {c₁ c₂ : ℝ} (hc₁_pos : 0 < c₁) (hc₁_lt_two_thirds : c₁ < 2 / 3) (hc₂_ge_two_thirds : (2 / 3 : ℝ) ≤ c₂) (hc₂_lt_one : c₂ < 1) : ∃ m M : ℝ, ∃ c : WolfeCounterexample.MatrixIdentityLiminfStrongWolfeCertificate n m M c₁ c₂, ∃ C > 0, ∀ x y, ‖hessian c.iteration.objective x - hessian c.iteration.objective y‖ ≤ C * ‖x - y‖ ^ (1 / 2 : ℝ) := by obtain ⟨c, hHolder⟩ := existsStrongWolfeCounterexampleHolder_of_dimension_ge_two n hn hc₁_pos hc₁_lt_two_thirds hc₂_ge_two_thirds hc₂_lt_one obtain ⟨L, a, b, q, ha, hb, hq, hfactor, hlower, hupper, hgradient⟩ := WolfeCounterexample.factorAndBounds_of_initialPosDef c.initialInverseHessianPosDef have hc₁₂ : c₁ < c₂ := hc₁_lt_two_thirds.trans_le hc₂_ge_two_thirds obtain ⟨d, hd⟩ := WolfeCounterexample.matrixIdentityLiminfStrongWolfeWithObjective hn c L hfactor ha hb hq hc₁_pos hc₁₂ hc₂_lt_one hlower hupper hgradient refine ⟨(1 / 2 : ℝ) * a, (3 / 2 : ℝ) * b, d, ?_⟩ rw [hd] exact existsHessianHolderPullback c.objectiveContDiff L hHolder
end DFP