Documentation

LeanPool.Superdiffusion.Algsuperdiff.Section5.Support.CutoffFieldLimit

CutoffFieldLimit #

Ported from scottnarmstrong/Superdiffusion at 7f20476385d4af418f110303085ad0740c9eba46. Module imports, support namespace and local option directives adapted for Lean Pool.

The Dirichlet solutions of the cutoff fields converge to those of the full field #

Three steps, on the cube y + □_n.

The shift is invisible. The cutoff coefficient a_L and its origin normalization ã_L = a_L - k_L(0) differ by a constant skew matrix, so they have exactly the same divergence-form weak solutions (isDirichletSolutionAt_normalizedCoefficientCutoff_iff). This is the printed step "the equation is unchanged if we replace a_L by a_L - (k_L - k_m)".

The coefficient distance is geometric. On the cube the entrywise distance between ã_L and the full field a = ν I + k is at most Field.cutoffLimitGap, which vanishes geometrically in L.

The solution map is Lipschitz in the coefficient. The energy estimate sqrt_energy_grad_sub_le gives

  ‖∇u_L - ∇u‖_{L²(y+□_n)} ≤ ν⁻¹ d² ‖ã_L - a‖_{L^∞(y+□_n)} ‖∇u‖_{L²(y+□_n)} ,

and the zero-trace Poincaré inequality upgrades it to the L² distance of the values. Both solutions carry the same forcing field g; ν is the symmetric part of both coefficient fields.

Main results #

References #

1. Ellipticity of the two coefficient families on the cube #

The normalized cutoff field is elliptic on y + □_n, with lower constant the molecular diffusivity: its symmetric part is ν I and its skew part is continuous, so compactness of the closed cube supplies the upper constant.

The full stream coefficient is elliptic on y + □_n, with the same lower constant.

2. The constant skew shift does not change the Dirichlet problem #

The constant skew shift is invisible to the Dirichlet problem. This is the printed remark that the equation for a_L is unchanged when a_L is replaced by a_L minus a constant skew matrix.

3. The energy estimate with the explicit geometric distance #

theorem Algsuperdiff.Section5.Support.sqrt_energy_grad_sub_normalizedCutoff_le {d : ℕ} {nu gamma : ℝ} (hnu : 0 < nu) (hgamma : gamma < 1) {C : ℕ} {omega : Field.FullSample d gamma} (hC : Field.SharpTailBounded gamma C ↑omega) {y : HCPolySupport.Vec d} {n ell L : ℤ} (hcube : cubeSetAt y n ⊆ HCPolySupport.openCubeSet (HCPolySupport.originCube d ell)) {u v : HCPolySupport.H1Function (cubeSetAt y n)} {g : HCPolySupport.Vec d → HCPolySupport.Vec d} (hu : IsDirichletSolutionAt (Section3.Cutoff.coefficientCutoff nu L ↑omega).toCoeffField y n u g) (hv : IsDirichletSolutionAt (Field.streamCoefficient nu omega) y n v g) :
√(∫ (x : HCPolySupport.Vec d) in cubeSetAt y n, ‖u.grad x - v.grad x‖ ^ 2) ≤ nu⁻¹ * (↑d * ↑d * Field.cutoffLimitGap d gamma C ell L) * √(∫ (x : HCPolySupport.Vec d) in cubeSetAt y n, ‖v.grad x‖ ^ 2)

Continuity of the solution map in the coefficient, along the cutoff family. For the same forcing field g, the gradient of the cutoff solution u_L differs from that of the full-field solution u by at most ν⁻¹ d² · gap(L) · ‖∇u‖_{L²}, with gap(L) = Field.cutoffLimitGap the explicit geometric coefficient distance.

4. From the gradients to the values #

Two solutions of Dirichlet problems with the same forcing field differ by a zero-trace function.

The L² distance of two solutions with the same forcing field is controlled by the L² distance of their gradients, through the zero-trace Poincaré inequality on the cube. The constant depends only on the cube.

5. L² convergence of the cutoff solutions #

theorem Algsuperdiff.Section5.Support.exists_l2_bound_cutoff_stream {d : ℕ} [NeZero d] {nu gamma : ℝ} (hnu : 0 < nu) (hgamma : gamma < 1) {C : ℕ} {omega : Field.FullSample d gamma} (hC : Field.SharpTailBounded gamma C ↑omega) (y : HCPolySupport.Vec d) {n ell : ℤ} (hcube : cubeSetAt y n ⊆ HCPolySupport.openCubeSet (HCPolySupport.originCube d ell)) :
∃ (CP : ℝ), 0 ≤ CP ∧ ∀ (L : ℤ) (u v : HCPolySupport.H1Function (cubeSetAt y n)) (g : HCPolySupport.Vec d → HCPolySupport.Vec d), IsDirichletSolutionAt (Section3.Cutoff.coefficientCutoff nu L ↑omega).toCoeffField y n u g → IsDirichletSolutionAt (Field.streamCoefficient nu omega) y n v g → √(∫ (x : HCPolySupport.Vec d) in cubeSetAt y n, (u.toFun x - v.toFun x) ^ 2) ≤ CP * (nu⁻¹ * (↑d * ↑d * Field.cutoffLimitGap d gamma C ell L) * √(∫ (x : HCPolySupport.Vec d) in cubeSetAt y n, ‖v.grad x‖ ^ 2))

The L² bound on the distance between the cutoff solution and the full-field solution, uniform in L up to the explicit geometric factor.

theorem Algsuperdiff.Section5.Support.tendsto_sqrt_integral_sq_sub_cutoff_atTop_zero {d : ℕ} [NeZero d] {nu gamma : ℝ} (hnu : 0 < nu) (hgamma : gamma < 1) {C : ℕ} {omega : Field.FullSample d gamma} (hC : Field.SharpTailBounded gamma C ↑omega) {y : HCPolySupport.Vec d} {n ell : ℤ} (hcube : cubeSetAt y n ⊆ HCPolySupport.openCubeSet (HCPolySupport.originCube d ell)) {u : ℤ → HCPolySupport.H1Function (cubeSetAt y n)} {v : HCPolySupport.H1Function (cubeSetAt y n)} {g : HCPolySupport.Vec d → HCPolySupport.Vec d} (hu : ∀ (L : ℤ), IsDirichletSolutionAt (Section3.Cutoff.coefficientCutoff nu L ↑omega).toCoeffField y n (u L) g) (hv : IsDirichletSolutionAt (Field.streamCoefficient nu omega) y n v g) :
Filter.Tendsto (fun (L : ℤ) => √(∫ (x : HCPolySupport.Vec d) in cubeSetAt y n, ((u L).toFun x - v.toFun x) ^ 2)) Filter.atTop (nhds 0)

u_L → u in L²(y + □_n). The forcing field is the same at every scale; only the coefficient field moves.