v21: comparisons on the same Wolfe witness

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

Source header · Attribution source

public sectionopen Filteropen DFP.Polaropen scoped Topology BigOperators Asymptoticsnamespace DFPWolfe.V21namespace SequencePropertiesopen DFP.Polar.SequenceAsymptotics

Helper for sec:known-results and lem:planar-degeneration: the source sequence conclusions for one fixed original trajectory. The parameters Unknown identifier `Cgrad`Cgrad, Unknown identifier `lower`lower, Unknown identifier `upper`upper, Unknown identifier `productLower`productLower, and Unknown identifier `productUpper`productUpper are fixed comparison constants. The scalar definitions are the accepted original source definitions, with current matrix and previous curvature in Unknown identifier `freeCoefficient`freeCoefficient at every source index Unknown identifier `k`sorry ≥ 1 : Propk≥1.

structure SourceEstimates (p : Parameters) (ν Cgrad lower upper productLower productUpper : ℝ) : Prop where /-- Helper for sec:known-results: the original cubic feedback coefficient is positive. -/ feedback_pos : 0 < p.feedbackRate /-- Helper for sec:known-results: both ratio expansions have the same uniform fourth-order error constant, at every original phase index. -/ gradient_ratio_errors : ∀ j, |gradientRatio p (2 * j) - 1 + p.feedbackRate * p.scale j ^ 3| ≤ Cgrad * p.scale j ^ 4 ∧ |gradientRatio p (2 * j + 1) - 1 - p.feedbackRate * p.scale j ^ 3| ≤ Cgrad * p.scale j ^ 4 /-- Helper for sec:known-results: every even phase decreases and every odd phase increases the actual gradient norm. -/ gradient_oscillation : ∀ j, ‖p.gradient (2 * j + 1)‖ < ‖p.gradient (2 * j)‖ ∧ ‖p.gradient (2 * j + 1)‖ < ‖p.gradient (2 * j + 2)‖ /-- Helper for sec:known-results: the gradient norm is not eventually nonincreasing. -/ gradient_not_nonincreasing : ¬ ∀ᶠ k in atTop, ‖p.gradient (k + 1)‖ ≤ ‖p.gradient k‖ /-- Helper for sec:known-results: the gradient norm is not eventually nondecreasing. -/ gradient_not_nondecreasing : ¬ ∀ᶠ k in atTop, ‖p.gradient k‖ ≤ ‖p.gradient (k + 1)‖ /-- Helper for sec:known-results: the actual physical step norms are comparable to the squared prescribed scales, with the fixed positive comparison constants. -/ step_bounds : ∀ k, lower * p.scale (k / 2) ^ 2 ≤ ‖step p k‖ ∧ ‖step p k‖ ≤ upper * p.scale (k / 2) ^ 2 /-- Helper for sec:known-results: the total physical step length is infinite, both as nonsummability and as partial sums tending to positive infinity. -/ length_diverges : (¬ Summable (fun k ↦ ‖step p k‖)) ∧ Tendsto (fun n ↦ ∑ k ∈ Finset.range n, ‖step p k‖) atTop atTop /-- Helper for sec:known-results: every accepted endpoint directional pairing is nonzero. -/ endpoint_nonzero : ∀ k, WolfeRatios.endSlope p k ≠ 0 /-- Helper for eq:planar-scalar-definition and sec:known-results: the exact physical secant curvature retains the source `σ/(1+ω²)` identity. -/ curvature_law : ∀ k, curvature p k = p.sigma k / (1 + p.omega k ^ 2) /-- Helper for eq:planar-scalar-definition and sec:known-results: the actual normalized secant curvature tends to one. -/ curvature_limit : Tendsto (curvature p) atTop (𝓝 1) /-- Helper for lem:planar-degeneration and sec:known-results: the original physical inverse-Hessian determinants have the exact source quartic equivalent. -/ determinant_order : (fun k ↦ (p.inverseHessian k).det) ~[atTop] (fun k ↦ p.scale (k / 2) ^ 4 / p.kStar) /-- Helper for eq:planar-scalar-definition and sec:known-results: the ORIGINAL `λ_k=κ_(k−1)det H_k` has that same source quartic equivalent. -/ free_coefficient_order : freeCoefficient p ~[atTop] (fun k ↦ p.scale (k / 2) ^ 4 / p.kStar) /-- Helper for lem:planar-degeneration and sec:known-results: the original free coefficient is comparable to consecutive physical step products. -/ free_coefficient_step_bounds : ∀ᶠ k in atTop, productLower * (‖step p (k - 1)‖ * ‖step p k‖) ≤ freeCoefficient p k ∧ freeCoefficient p k ≤ productUpper * (‖step p (k - 1)‖ * ‖step p k‖) /-- Helper for sec:known-results: Xu's ORIGINAL scalar has the even source coefficient `2awStar` and odd source coefficient `azStar`. -/ xu_orders : ((fun j ↦ xuScalar p (2 * j)) ~[atTop] (fun j ↦ (2 * baseAngle * cycleW ν) * p.scale j ^ 4)) ∧ ((fun j ↦ xuScalar p (2 * j + 1)) ~[atTop] (fun j ↦ (baseAngle * cycleZ ν) * p.scale j ^ 4)) /-- Helper for sec:known-results: the NEXT even to current odd Xu ratio tends to the exact source constant, which is strictly larger than four. -/ xu_jump : Tendsto (fun j ↦ xuScalar p (2 * j + 2) / xuScalar p (2 * j + 1)) atTop (𝓝 (2 * cycleW ν / cycleZ ν)) ∧ 4 < 2 * cycleW ν / cycleZ ν /-- Helper for sec:known-results: Xu's original scalar is not eventually nonincreasing. -/ xu_not_nonincreasing : ¬ ∀ᶠ k in atTop, xuScalar p (k + 1) ≤ xuScalar p k

Helper for sec:known-results: the accepted actual sequence theorem yields one source record, with all positive constants fixed before every later start. The product constants use the SAME step bounds and exact source normalization.

theorem sourceSequenceNeighborhood : ∃ η > 0, ∀ ν ∈ Set.Ioo baseAngle (baseAngle + η), ∃ Cgrad > 0, ∃ lower > 0, ∃ upper > 0, ∃ productLower > 0, ∃ productUpper > 0, ∃ J : ℕ, 0 < J ∧ ∀ j₀ ≥ J, SourceEstimates (Parameters.of j₀ ν) ν Cgrad lower upper productLower productUpper := by obtain ⟨η, hη, hactual⟩ := polarSequenceAsymptotics refine ⟨η, hη, ?_⟩ intro ν hnear obtain ⟨hparameter, Cgrad, hCgrad, lower, hlower, upper, hupper, J, hJ, hall⟩ := hactual ν hnear have hK : 0 < cycleK ν := by simpa only [Set.mem_Ioi] using (TwoCycle.cycleFamily_mapsTo_positive hparameter).1 let productLower := 1 / (8 * upper ^ 2 * cycleK ν) let productUpper := 2 / (lower ^ 2 * cycleK ν) have hProductLower : 0 < productLower := by dsimp only [productLower] positivity have hProductUpper : 0 < productUpper := by dsimp only [productUpper] positivity refine ⟨Cgrad, hCgrad, lower, hlower, upper, hupper, productLower, hProductLower, productUpper, hProductUpper, J, hJ, ?_⟩ intro j₀ hj₀ let p := Parameters.of j₀ ν obtain ⟨hL, herrors, _, hsigns, hnotDown, hnotUp, hsteps, hlength, hendpoint, hcurvatureLaw, hcurvatureLimit, hdet, hfree, hEvenXu, hOddXu, hJump, hJumpLarge, hXuNot⟩ := hall j₀ hj₀ have hstart : 0 < p.startIndex := by simpa only [p, Parameters.of_startIndex] using hJ.trans_le hj₀ have hpK : p.kStar = cycleK ν := by rw [Parameters.kStar_def, Parameters.of_cycleParameter] have hKp : 0 < p.kStar := by rw [hpK] exact hK have hproduct : ∀ᶠ k in atTop, productLower * (‖step p (k - 1)‖ * ‖step p k‖) ≤ freeCoefficient p k ∧ freeCoefficient p k ≤ productUpper * (‖step p (k - 1)‖ * ‖step p k‖) := by have h := freeCoefficient_stepProduct_bounds p hstart hKp hlower hupper hsteps hfree simpa only [productLower, productUpper, hpK] using h exact ⟨hL, herrors, hsigns, hnotDown, hnotUp, hsteps, hlength, hendpoint, hcurvatureLaw, hcurvatureLimit, hdet, hfree, hproduct, ⟨hEvenXu, hOddXu⟩, ⟨hJump, hJumpLarge⟩, hXuNot⟩

Helper for sec:known-results: exact source endpoint data turn the actual all-index model Wolfe certificates into the canonical strong Wolfe predicate.

theorem strongWolfeOfSourceModel {p : Parameters} {κ δ C H c₁ c₂ : ℝ} (h : ActualObjective.ObjectiveEstimates p κ δ C H) (hc₁ : 0 < c₁) (horder : c₁ < c₂) (hc₂ : c₂ < 1) (hmodel : ∀ k, WolfeRatios.ModelStrongWolfe p k c₁ c₂) : ∀ k, LineSearch.IsStrongWolfe c₁ c₂ (InterpolationSum.polarObjective p κ) (p.point k) (p.point (k + 1) - p.point k) := by have hendpoints (k : ℕ) := PlanarCounterexample.modelStrongWolfe_radial_endpoints p k (hmodel k) exact PlanarCounterexample.strongWolfe_of_radial_endpoints h hc₁ horder hc₂ (fun k ↦ (hendpoints k).1) (fun k ↦ (hendpoints k).2)

Helper for sec:known-results: the canonical scaled inverse-DFP step is exactly the original physical displacement under the source identity.

theorem canonicalStep_eq_source {p : Parameters} {κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k = step p k := by have hdisplacement : w.iteration.point (k + 1) - w.iteration.point k = DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k := by rw [w.iteration.pointSucc k] abel rw [← hdisplacement, h.point, step_def]

Helper for sec:known-results and eq:planar-scalar-definition: canonical gradient changes equal the original prescribed source secants.

theorem canonicalSecant_eq_source {p : Parameters} {κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : DFP.gradientChanges (DFP.gradients w.iteration.objective w.iteration.point) k = secant p k := by rw [h.gradients, DFP.gradientChanges_apply, secant_def]

Helper for sec:known-results: Xu's scalar formed from the canonical step, gradient change and SAME original step length equals the source scalar exactly.

theorem canonicalXuScalar_eq_source {p : Parameters} {κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : inner ℝ (DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k) (DFP.gradientChanges (DFP.gradients w.iteration.objective w.iteration.point) k) / w.iteration.stepLength k = xuScalar p k := by rw [canonicalStep_eq_source h, canonicalSecant_eq_source h, h.stepLength, xuScalar_def, secantPairing_def]

Helper for sec:known-results and eq:planar-scalar-definition: the canonical normalized secant curvature is exactly the original source Unknown identifier `κ_k`κ_k.

theorem canonicalCurvature_eq_source {p : Parameters} {κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : inner ℝ (DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k) (DFP.gradientChanges (DFP.gradients w.iteration.objective w.iteration.point) k) / ‖DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k‖ ^ 2 = curvature p k := by rw [canonicalStep_eq_source h, canonicalSecant_eq_source h, curvature_def, secantPairing_def]

Helper for sec:known-results and eq:planar-scalar-definition: previous canonical curvature times the CURRENT canonical determinant is the ORIGINAL source , with natural subtraction merely totalizing index zero.

theorem canonicalFreeCoefficient_eq_source {p : Parameters} {κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : (inner ℝ (DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) (k - 1)) (DFP.gradientChanges (DFP.gradients w.iteration.objective w.iteration.point) (k - 1)) / ‖DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) (k - 1)‖ ^ 2) * (w.iteration.inverseHessian k).det = freeCoefficient p k := by rw [canonicalCurvature_eq_source h, h.inverseHessian, freeCoefficient_def]

Helper for sec:known-results: BOTH failures of eventual monotonicity hold for the canonical objective gradients of the SAME standard counterexample.

theorem SourceEstimates.canonicalGradientOscillation {p : Parameters} {ν Cgrad lower upper productLower productUpper κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : SourceEstimates p ν Cgrad lower upper productLower productUpper) (hsource : PlanarCounterexample.SourceIdentity p κ w) : (¬ ∀ᶠ k in atTop, ‖DFP.gradients w.iteration.objective w.iteration.point (k + 1)‖ ≤ ‖DFP.gradients w.iteration.objective w.iteration.point k‖) ∧ (¬ ∀ᶠ k in atTop, ‖DFP.gradients w.iteration.objective w.iteration.point k‖ ≤ ‖DFP.gradients w.iteration.objective w.iteration.point (k + 1)‖) := by rw [hsource.gradients] exact ⟨h.gradient_not_nonincreasing, h.gradient_not_nondecreasing⟩

Helper for sec:known-results: the SAME canonical inverse-DFP scaled steps retain the quadratic comparisons and infinite total physical length.

theorem SourceEstimates.canonicalStepProperties {p : Parameters} {ν Cgrad lower upper productLower productUpper κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : SourceEstimates p ν Cgrad lower upper productLower productUpper) (hsource : PlanarCounterexample.SourceIdentity p κ w) : let s := DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) (∀ k, lower * p.scale (k / 2) ^ 2 ≤ ‖s k‖ ∧ ‖s k‖ ≤ upper * p.scale (k / 2) ^ 2) ∧ (¬ Summable (fun k ↦ ‖s k‖)) ∧ Tendsto (fun n ↦ ∑ k ∈ Finset.range n, ‖s k‖) atTop atTop := by dsimp only simp_rw [canonicalStep_eq_source hsource] exact ⟨h.step_bounds, h.length_diverges⟩

Helper for sec:known-results: the canonical next-endpoint gradient paired with the SAME canonical step is nonzero at every accepted endpoint.

theorem SourceEstimates.canonicalEndpointNonzero {p : Parameters} {ν Cgrad lower upper productLower productUpper κ c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : SourceEstimates p ν Cgrad lower upper productLower productUpper) (hsource : PlanarCounterexample.SourceIdentity p κ w) (k : ℕ) : inner ℝ (DFP.gradients w.iteration.objective w.iteration.point (k + 1)) (DFP.steps w.iteration.stepLength (DFP.directions w.iteration.inverseHessian (DFP.gradients w.iteration.objective w.iteration.point)) k) ≠ 0 := by rw [canonicalStep_eq_source hsource, hsource.gradients, step_def] simpa only [WolfeRatios.endSlope_def] using h.endpoint_nonzero k

Helper for sec:known-results: the exact canonical unit-circle cluster set excludes convergence of the iterates to ANY point, not only to the minimizer.

theorem not_convergent_of_circle_cluster {ε c₁ c₂ : ℝ} {w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂} (h : PlanarCounterexample.CounterexampleProperties ε w) : ¬ ∃ z : EuclideanSpace ℝ (Fin 2), Tendsto w.iteration.point atTop (𝓝 z) := by rintro ⟨z, hz⟩ have hzero := (h.cluster (polarUnit 0)).mpr (PhysicalRealization.polarUnit_norm 0) have hpi := (h.cluster (polarUnit Real.pi)).mpr (PhysicalRealization.polarUnit_norm Real.pi) obtain ⟨φ, hφ, hφlimit⟩ := hzero.tendsto_subseq obtain ⟨ψ, hψ, hψlimit⟩ := hpi.tendsto_subseq have hzeroz := tendsto_nhds_unique hφlimit (hz.comp hφ.tendsto_atTop) have hpiz := tendsto_nhds_unique hψlimit (hz.comp hψ.tendsto_atTop) have hcoordinate := congrArg (fun x : EuclideanSpace ℝ (Fin 2) ↦ x 0) (hzeroz.trans hpiz.symm) norm_num only [polarUnit_apply_zero, Real.cos_zero, Real.cos_pi] at hcoordinate
end SequenceProperties

sec:known-results and the polar paragraph following lem:planar-degeneration: the SAME canonical planar strong-Wolfe counterexample has Theorem 1's analytic, geometric and sharpness conclusions AND all original source sequence properties.

The Wolfe parameter Unknown identifier `νStar`νStar is selected inside the intersection of the source-data and sequence neighborhoods. It is fixed first; the support coefficient and every sequence comparison constant are then fixed before ALL requested accuracies and ALL later starting indices. For each accuracy, one natural threshold combines the source objective, model Wolfe, and sequence thresholds. Unknown identifier `SourceIdentity`SourceIdentity keeps the objective, points, canonical gradients, matrices and positive step lengths equal to the original prescribed data of that SAME Parameters.of sorry sorry : ParametersParameters.of Unknown identifier `j₀`j₀ Unknown identifier `νStar`νStar.

theorem planarCounterexampleFamilyWithSequenceProperties (c₁ c₂ : ℝ) (hc₁ : 0 < c₁) (hc₁lim : c₁ < 2 / 3) (hc₂lim : 2 / 3 ≤ c₂) (hc₂ : c₂ < 1) : ∃ νStar > baseAngle, ∃ κ > 0, ∃ Cgrad > 0, ∃ lower > 0, ∃ upper > 0, ∃ productLower > 0, ∃ productUpper > 0, ∀ ε : ℝ, 0 < ε → ε < 1 → ∃ J : ℕ, 0 < J ∧ ∀ j₀ ≥ J, ∃ w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂, PlanarCounterexample.SourceIdentity (Parameters.of j₀ νStar) κ w ∧ PlanarCounterexample.CounterexampleProperties ε w ∧ SequenceProperties.SourceEstimates (Parameters.of j₀ νStar) νStar Cgrad lower upper productLower productUpper := by obtain ⟨ηs, hηs, hsourceData⟩ := PlanarCounterexample.sourceDataNeighborhood obtain ⟨ηq, hηq, hsequenceData⟩ := SequenceProperties.sourceSequenceNeighborhood obtain ⟨νStar, hνStar, _, Jw, _, hmodels⟩ := WolfeRatios.existsUniformModelWolfe c₁ c₂ (min ηs ηq) hc₁ hc₁lim hc₂lim hc₂ (lt_min hηs hηq) have hnearSource : νStar ∈ Set.Ioo baseAngle (baseAngle + ηs) := Set.Ioo_subset_Ioo_right (add_le_add le_rfl (min_le_left ηs ηq)) hνStar have hnearSequence : νStar ∈ Set.Ioo baseAngle (baseAngle + ηq) := Set.Ioo_subset_Ioo_right (add_le_add le_rfl (min_le_right ηs ηq)) hνStar obtain ⟨κ, hκ, C, _, H, hH, haccuracy⟩ := hsourceData νStar hnearSource obtain ⟨Cgrad, hCgrad, lower, hlower, upper, hupper, productLower, hProductLower, productUpper, hProductUpper, Jq, _, hsequences⟩ := hsequenceData νStar hnearSequence refine ⟨νStar, hνStar.1, κ, hκ, Cgrad, hCgrad, lower, hlower, upper, hupper, productLower, hProductLower, productUpper, hProductUpper, ?_⟩ intro ε hε _ let δ := min ε (1 / 2) have hhalf : (0 : ℝ) < 1 / 2 := by norm_num have hhalfOne : (1 / 2 : ℝ) < 1 := by norm_num have hδ : 0 < δ := lt_min hε hhalf have hδhalf : δ ≤ 1 / 2 := min_le_right _ _ have hδε : δ ≤ ε := min_le_left _ _ have hδone : δ < 1 := hδhalf.trans_lt hhalfOne obtain ⟨Js, hJs, hsource⟩ := haccuracy δ hδ hδone refine ⟨max Js (max Jw Jq), hJs.trans_le (le_max_left _ _), ?_⟩ intro j₀ hj₀ have hjSource : Js ≤ j₀ := (le_max_left _ _).trans hj₀ have hjWolfe : Jw ≤ j₀ := ((le_max_left _ _).trans (le_max_right _ _)).trans hj₀ have hjSequence : Jq ≤ j₀ := ((le_max_right _ _).trans (le_max_right _ _)).trans hj₀ have hactual := hsource j₀ hjSource have hwolfe := SequenceProperties.strongWolfeOfSourceModel hactual.objective hc₁ (hc₁lim.trans_le hc₂lim) hc₂ (hmodels j₀ hjWolfe) obtain ⟨w, hidentity, hproperties⟩ := PlanarCounterexample.certificateOfSourceData hactual hδhalf hδε hH.le hwolfe exact ⟨w, hidentity, hproperties, hsequences j₀ hjSequence⟩

Helper for sec:known-results and lem:planar-degeneration: for any requested accuracy, one canonical planar strong-Wolfe counterexample simultaneously has all Theorem 1 properties and all source sequence refinements, with exact source identity.

theorem planarCounterexampleWithSequenceProperties (c₁ c₂ ε : ℝ) (hc₁ : 0 < c₁) (hc₁lim : c₁ < 2 / 3) (hc₂lim : 2 / 3 ≤ c₂) (hc₂ : c₂ < 1) (hε : 0 < ε) (hεone : ε < 1) : ∃ νStar > baseAngle, ∃ κ > 0, ∃ Cgrad > 0, ∃ lower > 0, ∃ upper > 0, ∃ productLower > 0, ∃ productUpper > 0, ∃ j₀ : ℕ, 0 < j₀ ∧ ∃ w : DFP.StrongWolfeCounterexample (Fin 2) (1 / 2) (3 / 2) c₁ c₂, PlanarCounterexample.SourceIdentity (Parameters.of j₀ νStar) κ w ∧ PlanarCounterexample.CounterexampleProperties ε w ∧ SequenceProperties.SourceEstimates (Parameters.of j₀ νStar) νStar Cgrad lower upper productLower productUpper := by obtain ⟨νStar, hνStar, κ, hκ, Cgrad, hCgrad, lower, hlower, upper, hupper, productLower, hProductLower, productUpper, hProductUpper, haccuracy⟩ := planarCounterexampleFamilyWithSequenceProperties c₁ c₂ hc₁ hc₁lim hc₂lim hc₂ obtain ⟨J, hJ, hfamily⟩ := haccuracy ε hε hεone obtain ⟨w, hidentity, hproperties, hsequences⟩ := hfamily J le_rfl exact ⟨νStar, hνStar, κ, hκ, Cgrad, hCgrad, lower, hlower, upper, hupper, productLower, hProductLower, productUpper, hProductUpper, J, hJ, w, hidentity, hproperties, hsequences⟩
end DFPWolfe.V21