Documentation

LeanPool.EllipticPDE.Embedding.HolderOfGradClosed

Hölder regularity of finite order from a bounded supply of weak derivatives #

Guo's Sobolev embedding (Guo, Partial Differential Equations, Theorem IV.2.3) has two cases. The first raises the exponent, and EllipticPdes.Embedding.memLp_of_gradClosed_fullStep runs it. The second reads a bounded supply of weak derivatives as classical ones: for u ∈ W^{m,p}(Ω) with m > n/p, the conclusion is u ∈ C^{m-1-⌊n/p⌋, γ}(Ω). This file proves the second case at p = 2, locally, for a family closed under weak differentiation as far as m.

Order the supply pays for #

The supply is spent in three places. Morrey asks for the weak gradient, so one order goes there; the ladder takes ⌊d/2⌋ more raising that gradient from L² to L^{2d}; and reading the n-th classical derivative asks the same of every index n levels up. So an index of depth dep i reaches C^n while dep i + n + 1 + ⌊d/2⌋ ≤ m, and at dep i = 0 that is n = m - 1 - ⌊d/2⌋, which is Guo's order exactly.

Hölder exponent #

The ladder lands on L^{2d} and EllipticPdes.Embedding.morrey_ball reads off 1 - d/(2d), so the exponent is 1/2 in every dimension. For d odd this is Guo's ⌊d/2⌋ + 1 - d/2 on the nose. For d even d/2 is an integer, Guo's statement leaves γ free in (0, 1), and 1/2 is one admissible choice: the ladder's landing reciprocal is 0 there, so every finite exponent is reached and 2d is the one this file fixes.

Main declarations #

References #

Guo, Partial Differential Equations, Theorem IV.2.3. Evans, Partial Differential Equations (2nd ed.), §5.6.3.

theorem EllipticPdes.Embedding.morreyExponent_two_mul {d : ℕ} (hd : 0 < d) :
morreyExponent d (2 * ↑d) = 1 / 2

At the exponent 2d the Morrey exponent is 1/2, whatever the dimension.

theorem EllipticPdes.Embedding.contDiffOn_holder_of_gradClosed {d : ℕ} (hd : 0 < d) (c : EuclideanSpace ℝ (Fin d)) {r R : ℝ} (hr : 0 < r) (hrR : r < R) {ι : Type u_1} {F : ι → EuclideanSpace ℝ (Fin d) → ℝ} {nxt : ι → Fin d → ι} {dep : ι → ℕ} {m : ℕ} (hdep : ∀ (i : ι) (k : Fin d), dep (nxt i k) ≤ dep i + 1) (hgrad : ∀ (i : ι), dep i < m → HasWeakGradOn (Metric.ball c R) (F i) fun (k : Fin d) => F (nxt i k)) (hmem : ∀ (i : ι), dep i ≤ m → MeasureTheory.MemLp (F i) 2 (MeasureTheory.volume.restrict (Metric.ball c R))) :
∃ (v : ι → EuclideanSpace ℝ (Fin d) → ℝ), (∀ (i : ι), dep i + 1 + d / 2 ≤ m → v i =ᵐ[MeasureTheory.volume.restrict (Metric.ball c r)] F i) ∧ (∀ (i : ι), dep i + 1 + d / 2 ≤ m → ∃ (M : NNReal), HolderOnWith M (1 / 2) (v i) (Metric.ball c r)) ∧ (∀ (n : ℕ) (i : ι), dep i + n + 1 + d / 2 ≤ m → ContDiffOn ℝ (↑n) (v i) (Metric.ball c r)) ∧ ∀ (i : ι), dep i + 2 + d / 2 ≤ m → ∀ y ∈ Metric.ball c r, HasFDerivAt (v i) (gradCLM (fun (k : Fin d) => v (nxt i k)) y) y

Classical derivatives of finite order from a bounded supply of weak ones. Let F assign a function to each index of ι, let nxt i k name a weak k-derivative of F i on Metric.ball c R, and let dep record how far an index sits above the root. If every index of depth at most m lies in L² there and every index of depth below m has its weak gradient in the family, then on any smaller concentric ball each index has a representative of class C^n whenever dep i + n + 1 + ⌊d/2⌋ ≤ m, and that representative is Hölder-1/2 as soon as dep i + 1 + ⌊d/2⌋ ≤ m.

The proof is the one EllipticPdes.Embedding.contDiffOn_of_gradClosed takes, run against a supply that runs out. The ladder raises the weak gradient of a qualifying index to L^{2d}, Morrey converts that into a Hölder representative, and a continuous function with a continuous weak gradient is classically differentiable, so an induction on the order reads the family as C^n with no further shrinking of the ball. Each induction step spends one order, which is what the depth condition records.

The last component states what the family means: on the inner ball the classical derivative of v i is the tuple of the v (nxt i k), so an entry of depth n is the order-n classical partial derivative of the root and the Hölder bound above is a bound on that derivative.

Hölder exponent Guo leaves open #

theorem EllipticPdes.Embedding.exists_holderOnWith_of_gradClosed {d : ℕ} (hd : 0 < d) (c : EuclideanSpace ℝ (Fin d)) {r R : ℝ} (hr : 0 < r) (hrR : r < R) {ι : Type u_1} {F : ι → EuclideanSpace ℝ (Fin d) → ℝ} {nxt : ι → Fin d → ι} {dep : ι → ℕ} {m : ℕ} (hdep : ∀ (i : ι) (k : Fin d), dep (nxt i k) ≤ dep i + 1) (hgrad : ∀ (i : ι), dep i < m → HasWeakGradOn (Metric.ball c R) (F i) fun (k : Fin d) => F (nxt i k)) (hmem : ∀ (i : ι), dep i ≤ m → MeasureTheory.MemLp (F i) 2 (MeasureTheory.volume.restrict (Metric.ball c R))) {s : ℕ} {P : NNReal} (hsd : 2 * s ≤ d) (hP2 : 2 ≤ P) (hPd : ↑d < ↑P) (hPs : 2⁻¹ - ↑s * (↑d)⁻¹ ≤ (↑P)⁻¹) (i : ι) (hi : dep i + 1 + s ≤ m) :
∃ (w : EuclideanSpace ℝ (Fin d) → ℝ), w =ᵐ[MeasureTheory.volume.restrict (Metric.ball c r)] F i ∧ ∃ (M : NNReal), HolderOnWith M (morreyExponent d ↑P) w (Metric.ball c r)

Hölder clause at a general exponent. The ladder run for s rungs lands at any P the reciprocal relation 1/2 - s/d ≤ 1/P admits, and Morrey at P > d reads off the exponent 1 - d/P. The fixed exponent 1/2 of contDiffOn_holder_of_gradClosed is this at s = ⌊d/2⌋ and P = 2d.

theorem EllipticPdes.Embedding.exists_holderOnWith_of_gradClosed_even {d : ℕ} (hd : 0 < d) (hdeven : 2 * (d / 2) = d) (c : EuclideanSpace ℝ (Fin d)) {r R : ℝ} (hr : 0 < r) (hrR : r < R) {ι : Type u_1} {F : ι → EuclideanSpace ℝ (Fin d) → ℝ} {nxt : ι → Fin d → ι} {dep : ι → ℕ} {m : ℕ} (hdep : ∀ (i : ι) (k : Fin d), dep (nxt i k) ≤ dep i + 1) (hgrad : ∀ (i : ι), dep i < m → HasWeakGradOn (Metric.ball c R) (F i) fun (k : Fin d) => F (nxt i k)) (hmem : ∀ (i : ι), dep i ≤ m → MeasureTheory.MemLp (F i) 2 (MeasureTheory.volume.restrict (Metric.ball c R))) {γ : NNReal} (hγ0 : 0 < γ) (hγ1 : γ < 1) (i : ι) (hi : dep i + 1 + d / 2 ≤ m) :
∃ (w : EuclideanSpace ℝ (Fin d) → ℝ), w =ᵐ[MeasureTheory.volume.restrict (Metric.ball c r)] F i ∧ ∃ (M : NNReal), HolderOnWith M γ w (Metric.ball c r)

Guo's free Hölder exponent in even dimension. When d/2 is an integer, which at p = 2 is Guo's case n/p ∈ ℕ, the ladder reaches every finite exponent, so the Hölder exponent may be any value in (0,1). In odd dimension the reciprocal 1/2 - ⌊d/2⌋/d = 1/(2d) caps the exponent at 2d and the Hölder exponent at 1/2, which is Guo's other case.

theorem EllipticPdes.Embedding.exists_const_holderOnWith_of_gradClosed {d : ℕ} (hd : 0 < d) (c : EuclideanSpace ℝ (Fin d)) {r R : ℝ} (hr : 0 < r) (hrR : r < R) (ι : Type u_1) {s : ℕ} {P : NNReal} (hsd : 2 * s ≤ d) (hP2 : 2 ≤ P) (hPd : ↑d < ↑P) (hPs : 2⁻¹ - ↑s * (↑d)⁻¹ ≤ (↑P)⁻¹) :
∃ (C : NNReal), ∀ {F : ι → EuclideanSpace ℝ (Fin d) → ℝ} {nxt : ι → Fin d → ι} {dep : ι → ℕ} {m : ℕ}, (∀ (i : ι) (k : Fin d), dep (nxt i k) ≤ dep i + 1) → (∀ (i : ι), dep i < m → HasWeakGradOn (Metric.ball c R) (F i) fun (k : Fin d) => F (nxt i k)) → (∀ (i : ι), dep i ≤ m → MeasureTheory.MemLp (F i) 2 (MeasureTheory.volume.restrict (Metric.ball c R))) → ∀ (i : ι), dep i + 1 + s ≤ m → ∃ (w : EuclideanSpace ℝ (Fin d) → ℝ), w =ᵐ[MeasureTheory.volume.restrict (Metric.ball c r)] F i ∧ HolderOnWith (C * ∑ k : Fin d, (MeasureTheory.eLpNorm (F (nxt i k)) (ENNReal.ofReal ↑P) (MeasureTheory.volume.restrict (Metric.ball c r))).toNNReal) (morreyExponent d ↑P) w (Metric.ball c r)

Hölder estimate with its constant, in the shape Guo states: one constant, depending on the dimension, the exponent and the ball alone, bounding the Hölder seminorm of every member of every family by the L^P norms of that member's first derivatives.

exists_holderOnWith_of_gradClosed is this with the constant discarded. The remaining step to Guo's ‖u‖_{C^{k-1-⌊n/p⌋,γ}} ≤ C‖u‖_{W^{k,p}} is the ladder's own constant, which takes the L^P norms here back to the L² data.

theorem EllipticPdes.Embedding.exists_const_holderOnWith_of_gradClosed_of_bound {d : ℕ} (hd : 0 < d) (c : EuclideanSpace ℝ (Fin d)) {r R : ℝ} (hr : 0 < r) (hrR : r < R) (ι : Type u_1) {s : ℕ} {P : NNReal} (hsd : 2 * s ≤ d) (hP2 : 2 ≤ P) (hPd : ↑d < ↑P) (hPs : 2⁻¹ - ↑s * (↑d)⁻¹ ≤ (↑P)⁻¹) :
∃ (C : NNReal), ∀ {F : ι → EuclideanSpace ℝ (Fin d) → ℝ} {nxt : ι → Fin d → ι} {dep : ι → ℕ} {m : ℕ}, (∀ (i : ι) (k : Fin d), dep (nxt i k) ≤ dep i + 1) → (∀ (i : ι), dep i < m → HasWeakGradOn (Metric.ball c R) (F i) fun (k : Fin d) => F (nxt i k)) → (∀ (i : ι), dep i ≤ m → MeasureTheory.MemLp (F i) 2 (MeasureTheory.volume.restrict (Metric.ball c R))) → ∀ (M : NNReal), (∀ (j : ι), dep j ≤ m → MeasureTheory.eLpNorm (F j) 2 (MeasureTheory.volume.restrict (Metric.ball c R)) ≤ ↑M) → ∀ (i : ι), dep i + 1 + s ≤ m → ∃ (w : EuclideanSpace ℝ (Fin d) → ℝ), w =ᵐ[MeasureTheory.volume.restrict (Metric.ball c r)] F i ∧ HolderOnWith (C * M) (morreyExponent d ↑P) w (Metric.ball c r)

Guo's clause (ii) with its constant against the L² data. Composing the ladder's constant with Morrey's gives one constant, depending on the dimension, the rung count, the exponent and the two radii, bounding the Hölder seminorm of every member by a uniform L² bound on the family over the outer ball. This is the shape of Guo's ‖u‖_{C^{k-1-⌊n/p⌋,γ}} ≤ C‖u‖_{W^{k,p}} at p = 2.