v21 Lemmas 4–6: continuation, geometry and interpolation

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

Source header · Attribution source

public sectionopen Filter DFP.Polaropen scoped Topology InnerProduct BigOperators ContDiffnamespace DFPWolfe.V21

Helper for lem:polar-continuation: the fast error is exactly the source displacement from at the state's own first coordinate.

theorem polarFastErrorFormula (x : ℝ × ℝ × ℝ) : FastContraction.fastError x = ![x.2.1 - x.1 * cycleW (TwoCycle.inverseParameter x.1), x.2.2 - (1 + x.1 * cycleW (TwoCycle.inverseParameter x.1) ^ 2)] := by rw [FastContraction.fastError_def, FastContraction.fastCoordinates_apply, FastContraction.fastCoordinates_apply, TwoCycle.evenState_apply] ext i fin_cases i all_goals rfl

lem:polar-continuation. For every fixed cycle parameter sufficiently close to DFP.Polar.baseAngle : ℝbaseAngle from above, choose the fast width Unknown identifier `N`N, then the slow width Unknown identifier `M`M, then a positive natural threshold Unknown identifier `J`J. Every freshly initialized original trajectory with Unknown identifier `j₀`sorry ≥ sorry : Propj₀≥Unknown identifier `J`J obeys the slow and fast tube bounds at every cycle. The fast error is the original curve displacement displayed in Unknown identifier `polarFastErrorFormula`polarFastErrorFormula; both its fixed source norm and ordinary norm are bounded. The original matrix recurrence holds at every index, and every raw matrix is positive definite with positive off-diagonal entry, secant scalar Unknown identifier `σ`σ and factor Unknown identifier `W`W. All widths and thresholds precede every later choice of Unknown identifier `j₀`j₀ (main.tex:649–739).

theorem polarContinuation : ∃ η > 0, ∀ νStar ∈ Set.Ioo baseAngle (baseAngle + η), νStar ∈ TwoCycle.parameterSet ∧ ∃ N > 0, ∃ M > 0, ∃ J : ℕ, 0 < J ∧ ∀ j₀ ≥ J, let p := Parameters.of j₀ νStar (∀ j : ℕ, |(p.normalizedEvenState j).1 - p.kStar| ≤ M * p.scale j ∧ FastContraction.fixedNorm (FastContraction.fastError (p.normalizedEvenState j)) ≤ N * p.scale j ^ 3 ∧ ‖FastContraction.fastError (p.normalizedEvenState j)‖ ≤ (‖FastContraction.matrixOperator FastContraction.basisMatrix‖ + 1) * N * p.scale j ^ 3 ∧ (p.normalizedEvenState j).1 ∈ TwoCycle.normalizationSet ∧ 0 < (p.normalizedEvenState j).1) ∧ (∀ k, p.matrix (k + 1) = matrixStep (p.matrix k) (p.angleIncrement k) (p.zeta k) (p.zeta (k + 1))) ∧ ∀ k, (p.matrix k).PosDef ∧ 0 < p.matrix k 0 1 ∧ 0 < p.sigma k ∧ 0 < p.radiusRatio k := by obtain ⟨η, hη, hcontinuation⟩ := Continuation.polarContinuation refine ⟨η, hη, ?_⟩ intro νStar hnear obtain ⟨hparameter, N, hN, M, hM, J, hJ, hall⟩ := hcontinuation νStar hnear refine ⟨hparameter, N, hN, M, hM, J, hJ, ?_⟩ intro j₀ hj₀ obtain ⟨htube, hpositive⟩ := hall j₀ hj₀ refine ⟨?_, Parameters.matrix_succ _, hpositive⟩ intro j have hfast := (htube j).2.1 have hnorm := FastContraction.norm_le_basisOperator_mul_fixedNorm (FastContraction.fastError ((Parameters.of j₀ νStar).normalizedEvenState j)) have hnonnegative := FastContraction.fixedNorm_nonneg (FastContraction.fastError ((Parameters.of j₀ νStar).normalizedEvenState j)) have hconstant : 0 ≤ ‖FastContraction.matrixOperator FastContraction.basisMatrix‖ + 1 := by positivity have hordinary : ‖FastContraction.fastError ((Parameters.of j₀ νStar).normalizedEvenState j)‖ ≤ (‖FastContraction.matrixOperator FastContraction.basisMatrix‖ + 1) * N * (Parameters.of j₀ νStar).scale j ^ 3 := by calc _ ≤ (‖FastContraction.matrixOperator FastContraction.basisMatrix‖ + 1) * FastContraction.fixedNorm (FastContraction.fastError ((Parameters.of j₀ νStar).normalizedEvenState j)) := by nlinarith _ ≤ (‖FastContraction.matrixOperator FastContraction.basisMatrix‖ + 1) * (N * (Parameters.of j₀ νStar).scale j ^ 3) := mul_le_mul_of_nonneg_left hfast hconstant _ = _ := by ring exact ⟨(htube j).1, hfast, hordinary, (htube j).2.2⟩

Helper for lem:polar-geometry: the radial expansion uses precisely the source coefficients and , with the original cycle slopes.

theorem polarRadiusCoefficients (νStar : ℝ) : RadiusExpansions.radiusCoefficient νStar 0 = 2 * baseAngle * (cycleZ νStar + 2 * baseAngle) - 2 * baseAngle ^ 2 ∧ RadiusExpansions.radiusCoefficient νStar 1 = baseAngle * (cycleW νStar + baseAngle) - baseAngle ^ 2 / 2 := by constructor · rw [RadiusExpansions.radiusCoefficient_even, RadiusExpansions.referenceSlope_even] · rw [RadiusExpansions.radiusCoefficient_odd, RadiusExpansions.referenceSlope_odd]

lem:polar-geometry. The original points have radial expansion , radial decrements comparable to , and successive distances comparable to . Here Unknown identifier `D`sorry = sorry : PropD=Unknown identifier `radiusCoefficient`radiusCoefficient νStar is the original coefficient from the even and odd factor expansions. The radii strictly decrease to one and satisfy the exact quotient recurrence. The full accumulation set is the unit circle, and every distinct pair has the source maximum-scale separation. All positive constants are fixed before every Unknown identifier `j₀`sorry ≥ sorry : Propj₀≥Unknown identifier `J`J (main.tex:761–815).

theorem polarGeometry : ∃ η > 0, ∀ νStar ∈ Set.Ioo baseAngle (baseAngle + η), νStar ∈ TwoCycle.parameterSet ∧ ∃ K > 0, ∃ c > 0, ∃ B > 0, ∃ lower > 0, ∃ upper > 0, ∃ separation > 0, ∃ J : ℕ, 0 < J ∧ ∀ j₀ ≥ J, let p := Parameters.of j₀ νStar StrictAnti p.radius ∧ Tendsto p.radius atTop (𝓝 1) ∧ (∀ k, 1 < p.radius k ∧ 1 < p.radiusRatio k ∧ p.radius (k + 1) = p.radius k / p.radiusRatio k) ∧ (∀ k, |p.radius k - 1 - 3 * (RadiusExpansions.radiusCoefficient νStar 0 + RadiusExpansions.radiusCoefficient νStar 1) * p.scale (k / 2)| ≤ K * p.scale (k / 2) ^ 2 ∧ (c * p.scale (k / 2) ^ 4 ≤ p.radius k - p.radius (k + 1) ∧ p.radius k - p.radius (k + 1) ≤ B * p.scale (k / 2) ^ 4) ∧ lower * p.scale (k / 2) ^ 2 ≤ ‖p.point (k + 1) - p.point k‖ ∧ ‖p.point (k + 1) - p.point k‖ ≤ upper * p.scale (k / 2) ^ 2) ∧ (∀ k m, ‖p.point m - p.point k‖ ^ 2 = (p.radius m - p.radius k) ^ 2 + 4 * p.radius m * p.radius k * Real.sin ((p.angle m - p.angle k) / 2) ^ 2) ∧ {x | MapClusterPt x atTop p.point} = Metric.sphere 0 1 ∧ ∀ k m, k ≠ m → separation * max (p.scale (k / 2) ^ 2) (p.scale (m / 2) ^ 2) ≤ ‖p.point m - p.point k‖ := by obtain ⟨η, hη, hgeometry⟩ := PointGeometry.polarPointGeometry refine ⟨η, hη, ?_⟩ intro νStar hnear obtain ⟨hparameter, K, hK, c, hc, B, hB, lower, hlower, upper, hupper, separation, hseparation, J, hJ, hall⟩ := hgeometry νStar hnear refine ⟨hparameter, K, hK, c, hc, B, hB, lower, hlower, upper, hupper, separation, hseparation, J, hJ, ?_⟩ intro j₀ hj₀ obtain ⟨hradius, hpoints⟩ := hall j₀ hj₀ have hcluster : {x | MapClusterPt x atTop (Parameters.of j₀ νStar).point} = Metric.sphere 0 1 := by ext x simpa only [Set.mem_setOf_eq, Metric.mem_sphere, dist_zero_right] using hpoints.cluster x refine ⟨hradius.strictAnti, hradius.tendsto_one, ?_, ?_, hpoints.chord, hcluster, hpoints.separate⟩ · intro k exact ⟨hradius.above_one k, hradius.factor_above_one k, hradius.recurrence k⟩ · intro k exact ⟨hradius.expansion k, hradius.drop k, hpoints.step k⟩

Helper for lem:polar-interpolation: the unchanged exponentially flat η and its exact source quotient χ are smooth, lie between zero and one, equal one on the inner third-ball, and vanish on the full closed exterior of the unit ball.

theorem polarCutoffFormula : (∀ t : ℝ, ExplicitCutoff.eta t = if 0 < t then Real.exp (-1 / t) else 0) ∧ ContDiff ℝ ∞ (ExplicitCutoff.chi (E := EuclideanSpace ℝ (Fin 2))) ∧ ∀ z : EuclideanSpace ℝ (Fin 2), ExplicitCutoff.chi z = ExplicitCutoff.eta (1 - ‖z‖ ^ 2) / (ExplicitCutoff.eta (1 - ‖z‖ ^ 2) + ExplicitCutoff.eta (‖z‖ ^ 2 - 1 / 9)) ∧ 0 ≤ ExplicitCutoff.chi z ∧ ExplicitCutoff.chi z ≤ 1 ∧ (‖z‖ ≤ 1 / 3 → ExplicitCutoff.chi z = 1) ∧ (1 ≤ ‖z‖ → ExplicitCutoff.chi z = 0) := by refine ⟨ExplicitCutoff.eta_apply, ExplicitCutoff.contDiff_chi, ?_⟩ intro z exact ⟨ExplicitCutoff.chi_apply z, ExplicitCutoff.chi_nonneg z, ExplicitCutoff.chi_le_one z, ExplicitCutoff.chi_eq_one, ExplicitCutoff.chi_eq_zero⟩

Helper for lem:polar-interpolation: one constant depending only on the original cutoff controls the global half-Hölder Hessian of each actual smooth correction, with exactly the source factor sorry ^ (-3 / 2) : ?m.17‖Unknown identifier `aₖ`aₖ‖ ρₖ^(-3/2).

theorem polarBumpRegularity : ∃ Cχ : ℝ, 0 ≤ Cχ ∧ ∀ (p : Parameters) (κ : ℝ), 0 < p.startIndex → 0 < κ → ∀ k, ContDiff ℝ ∞ (InterpolationSum.polarBump p κ k) ∧ ∀ x y, ‖hessian (InterpolationSum.polarBump p κ k) x - hessian (InterpolationSum.polarBump p κ k) y‖ ≤ Cχ * ‖InterpolationSum.polarCoefficient p k‖ * (InterpolationSum.polarSupportRadius p κ k) ^ (-3 / 2 : ℝ) * ‖x - y‖ ^ (1 / 2 : ℝ) := by obtain ⟨Cχ, hCχ, hholder⟩ := ActualObjective.exists_polarBump_holder refine ⟨Cχ, hCχ, ?_⟩ intro p κ hstart hκ k refine ⟨ActualObjective.contDiff_polarBump p κ k, ?_⟩ intro x y rw [hessian_def, hessian_def, ActualObjective.norm_hessian_sub_eq_secondFDeriv_sub] exact hholder p κ hstart hκ k x y

lem:polar-interpolation. Fix the original cutoff's constant Unknown identifier `Cχ`Cχ, then a sufficiently small right parameter neighborhood. For each fixed Unknown identifier `νStar`νStar, choose the support coefficient Unknown identifier `κ`κ and uniform constants before the accuracy and every later initialized Unknown identifier `j₀`sorry ≥ sorry : Propj₀≥Unknown identifier `J`J. Each smooth original bump has the source scaled half-Hölder Hessian bound. The accepted Unknown identifier `ObjectiveEstimates`ObjectiveEstimates records the disjoint open support balls and their closed support bounds, the exact infinite correction sum, C² regularity across the circle, vanishing two-jet on the circle and exact center two-jets. It also records the uniform Unknown identifier `Ct₀`Ct₀ Hessian bound, global half-Hölder modulus Unknown identifier `H`H, and the actual objective's values, original gradients and identity Hessians at every endpoint. The objective is quadratic on the closed unit disk and uniformly convex. Unknown identifier `polarInterpolationConclusions`polarInterpolationConclusions exposes these fields directly. All objects use the original Unknown identifier `eta`eta, Unknown identifier `chi`chi and (main.tex:910–1040).

theorem polarInterpolation : ∃ Cχ : ℝ, 0 ≤ Cχ ∧ ∃ η > 0, ∀ νStar ∈ Set.Ioo baseAngle (baseAngle + η), ∃ κ > 0, ∃ C > 0, ∃ H > 0, ∀ ε : ℝ, 0 < ε → ε < 1 → ∃ J : ℕ, 0 < J ∧ ∀ j₀ ≥ J, let p := Parameters.of j₀ νStar ActualObjective.ObjectiveEstimates p κ ε C H ∧ ∀ k, ContDiff ℝ ∞ (InterpolationSum.polarBump p κ k) ∧ ∀ x y, ‖hessian (InterpolationSum.polarBump p κ k) x - hessian (InterpolationSum.polarBump p κ k) y‖ ≤ Cχ * ‖InterpolationSum.polarCoefficient p k‖ * (InterpolationSum.polarSupportRadius p κ k) ^ (-3 / 2 : ℝ) * ‖x - y‖ ^ (1 / 2 : ℝ) := by obtain ⟨Cχ, hCχ, hbumps⟩ := polarBumpRegularity obtain ⟨η, hη, hobjective⟩ := ActualObjective.polarObjectiveEstimates refine ⟨Cχ, hCχ, η, hη, ?_⟩ intro νStar hnear obtain ⟨κ, hκ, C, hC, H, hH, haccuracy⟩ := hobjective νStar hnear refine ⟨κ, hκ, C, hC, H, hH, ?_⟩ intro ε hε hεone obtain ⟨J, hJ, hall⟩ := haccuracy ε hε hεone refine ⟨J, hJ, ?_⟩ intro j₀ hj₀ have hstart : 0 < (Parameters.of j₀ νStar).startIndex := by simpa only [Parameters.of_startIndex] using hJ.trans_le hj₀ exact ⟨hall j₀ hj₀, hbumps _ κ hstart hκ⟩

Helper for lem:polar-interpolation: direct mathematical conclusions of the proved actual objective certificate. The original cutoff formula, exact support radii, disjoint function supports, closed support exterior, true infinite sum, C² regularity, vanishing circle and center two-jets, exact endpoint Hessian Unknown identifier `I`I, uniform Unknown identifier `Ct₀`Ct₀ Hessian bound and both global half-Hölder bounds are explicit. The certificate is supplied for every later Unknown identifier `j₀`j₀ by Unknown identifier `polarInterpolation`polarInterpolation, rather than assumed as a geometric or regularity premise of that source theorem.

theorem polarInterpolationConclusions {p : Parameters} {κ ε C H : ℝ} (h : ActualObjective.ObjectiveEstimates p κ ε C H) : Set.univ.PairwiseDisjoint (fun k ↦ Metric.ball (p.point k) (InterpolationSum.polarSupportRadius p κ k)) ∧ Set.univ.PairwiseDisjoint (fun k ↦ Function.support (InterpolationSum.polarBump p κ k)) ∧ (∀ k, 0 < InterpolationSum.polarSupportRadius p κ k ∧ InterpolationSum.polarSupportRadius p κ k = κ * p.scale (k / 2) ^ 2 ∧ tsupport (InterpolationSum.polarBump p κ k) ⊆ Metric.closedBall (p.point k) (InterpolationSum.polarSupportRadius p κ k) ∧ ∀ x ∈ Metric.closedBall (p.point k) (InterpolationSum.polarSupportRadius p κ k), 1 < ‖x‖) ∧ (∀ k x, InterpolationSum.polarBump p κ k x = ExplicitCutoff.chi ((κ * p.scale (k / 2) ^ 2)⁻¹ • (x - p.point k)) * inner ℝ (p.gradient k - p.point k) (x - p.point k)) ∧ (∀ x, InterpolationSum.polarPerturbation p κ x = ∑' k, InterpolationSum.polarBump p κ k x) ∧ (∀ x, InterpolationSum.polarObjective p κ x = (1 / 2 : ℝ) * ‖x‖ ^ 2 + InterpolationSum.polarPerturbation p κ x) ∧ ContDiff ℝ 2 (InterpolationSum.polarPerturbation p κ) ∧ ContDiff ℝ 2 (InterpolationSum.polarObjective p κ) ∧ (∀ x, ‖x‖ = 1 → InterpolationSum.polarPerturbation p κ x = 0 ∧ gradient (InterpolationSum.polarPerturbation p κ) x = 0 ∧ hessian (InterpolationSum.polarPerturbation p κ) x = 0) ∧ (∀ k, InterpolationSum.polarPerturbation p κ (p.point k) = 0 ∧ gradient (InterpolationSum.polarPerturbation p κ) (p.point k) = p.gradient k - p.point k ∧ hessian (InterpolationSum.polarPerturbation p κ) (p.point k) = 0) ∧ (∀ k, InterpolationSum.polarObjective p κ (p.point k) = (1 / 2 : ℝ) * p.radius k ^ 2 ∧ gradient (InterpolationSum.polarObjective p κ) (p.point k) = p.gradient k ∧ hessian (InterpolationSum.polarObjective p κ) (p.point k) = 1) ∧ (∀ x, ‖hessian (InterpolationSum.polarPerturbation p κ) x‖ ≤ C * p.scale 0) ∧ (∀ x y, ‖hessian (InterpolationSum.polarPerturbation p κ) x - hessian (InterpolationSum.polarPerturbation p κ) y‖ ≤ H * ‖x - y‖ ^ (1 / 2 : ℝ)) ∧ (∀ x y, ‖hessian (InterpolationSum.polarObjective p κ) x - hessian (InterpolationSum.polarObjective p κ) y‖ ≤ H * ‖x - y‖ ^ (1 / 2 : ℝ)) ∧ (∀ x, ‖x‖ ≤ 1 → InterpolationSum.polarObjective p κ x = (1 / 2 : ℝ) * ‖x‖ ^ 2) ∧ HasHessianBounds (1 - ε) (1 + ε) (InterpolationSum.polarObjective p κ) ∧ StrongConvexOn Set.univ (1 - ε) (InterpolationSum.polarObjective p κ) ∧ InterpolationSum.polarObjective p κ 0 = 0 ∧ ∀ x, (∀ y, InterpolationSum.polarObjective p κ x ≤ InterpolationSum.polarObjective p κ y) ↔ x = 0 := by have hopen : ∀ k, Function.support (InterpolationSum.polarBump p κ k) ⊆ Metric.ball (p.point k) (InterpolationSum.polarSupportRadius p κ k) := by intro k rw [ActualObjective.polarBump_eq_bump] exact InterpolationSum.support_bump_subset_ball h.geometry k have hsupport := h.geometry.disjoint_open.mono hopen refine ⟨h.geometry.disjoint_open, hsupport, ?_, InterpolationSum.polarBump_apply p κ, h.sum_formula, InterpolationSum.polarObjective_apply p κ, h.perturbation_contDiff, h.objective_contDiff, ?_, ?_, ?_, ?_, ?_, ?_, h.disk, h.hessian_bounds, h.strongConvex, h.origin, h.unique_minimizer⟩ · intro k refine ⟨h.geometry.radius_pos k, InterpolationSum.polarSupportRadius_apply p κ k, h.closed_support k, ?_⟩ intro x hx have hmargin := h.geometry.closedBall_norm_margin k hx have hradius := h.geometry.radius_pos k linarith · intro x hx simpa only [hessian_def] using h.sphere_jets x hx · intro k simpa only [hessian_def, InterpolationSum.polarCoefficient_apply] using h.center_jets k · intro k simpa only [hessian_def] using h.endpoints k · intro x rw [hessian_def, norm_fderiv_gradient_eq_norm_fderiv_fderiv] exact h.perturbation_bound x · intro x y rw [hessian_def, hessian_def, ActualObjective.norm_hessian_sub_eq_secondFDeriv_sub] exact h.perturbation_holder x y · intro x y simpa only [hessian_def] using h.hessian_holder x y
end DFPWolfe.V21