Documentation

LeanPool.Superdiffusion.Algsuperdiff.Section3.Provider.Corrector.MollifierConvergence

MollifierConvergence #

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

Provider: mollification converges in L²(Ω) as the radius shrinks #

Algsuperdiff/Section3/Provider/Corrector/Mollification.lean builds the smoothing operator mollify ρ X ω = ∫ ρ(y) • X((-y) +ᵥ ω) along the stationary translation action and shows that it stays inside the stationary L² layer. This file supplies the quantitative half needed by part (i) of l.approximation.stationary.by.local: mollification is an approximate identity on L²(Ω),

‖mollify ρ X − X‖²_{L²(Ω)} ≤ sup_{|y| < radius ρ} ‖X((−y) + ·) − X‖²_{L²(Ω)},

and the right-hand side tends to 0 with the radius because the Koopman orbit y ↦ X((−y) + ·) is L²-continuous (Algsuperdiff.Probability.Stationary.continuous_koopman_orbit).

Along the way the file records the elementary algebra of mollify (linearity and equivariance under the translation action) and the identity ‖f‖²_{L²} = ∫ ‖f‖², both of which are also consumed by the gradient identification of Algsuperdiff/Section3/Provider/Corrector/GradientIdentification.lean.

The squared L² norm as an integral #

The squared norm of an L² class is the integral of the squared pointwise norms of any of its representatives.

The L² energy of a square-integrable function is the squared norm of its class.

Elementary algebra of the mollification #

theorem Algsuperdiff.Section3.Provider.Corrector.mollify_comp_vadd {d : ℕ} {Ω : Type u_1} [AddAction (HCPolySupport.Vec d) Ω] {E : Type u_2} [NormedAddCommGroup E] [NormedSpace ℝ E] (κ : HCPolySupport.Vec d → ℝ) (X : Ω → E) (x : HCPolySupport.Vec d) (ω : Ω) :
mollify κ (fun (ω' : Ω) => X (x +ᵥ ω')) ω = mollify κ X (x +ᵥ ω)

Mollification commutes with the translation action: this is the pointwise form of the fact that the Koopman operators commute with the smoothing.

theorem Algsuperdiff.Section3.Provider.Corrector.mollify_sub {d : ℕ} {Ω : Type u_1} [AddAction (HCPolySupport.Vec d) Ω] {E : Type u_2} [NormedAddCommGroup E] [NormedSpace ℝ E] {κ : HCPolySupport.Vec d → ℝ} {X Y : Ω → E} {ω : Ω} (hX : MeasureTheory.Integrable (fun (y : HCPolySupport.Vec d) => κ y • X (-y +ᵥ ω)) MeasureTheory.volume) (hY : MeasureTheory.Integrable (fun (y : HCPolySupport.Vec d) => κ y • Y (-y +ᵥ ω)) MeasureTheory.volume) :
mollify κ (fun (ω' : Ω) => X ω' - Y ω') ω = mollify κ X ω - mollify κ Y ω

Mollification is additive, given the integrability of the two pieces.

theorem Algsuperdiff.Section3.Provider.Corrector.mollify_smul {d : ℕ} {Ω : Type u_1} [AddAction (HCPolySupport.Vec d) Ω] {E : Type u_2} [NormedAddCommGroup E] [NormedSpace ℝ E] (κ : HCPolySupport.Vec d → ℝ) (c : ℝ) (X : Ω → E) (ω : Ω) :
mollify κ (fun (ω' : Ω) => c • X ω') ω = c • mollify κ X ω

Mollification is homogeneous.

Integrability of the mollification kernel #

Joint integrability of the weighted squared increment of a stationary L² field.

The weighted increment integrability facts used by the approximate-identity estimate, for almost every sample.

The approximate-identity estimate #

The pointwise Jensen bound behind the approximate-identity estimate.

The approximate-identity estimate. If the L² increment of the field under every translation of size below the radius of the mollifier is at most ε, then the mollification differs from the field by at most ε in squared L² norm.

L² continuity of the translation action #

The Koopman translate of an L² class is represented by the translated representative.

The L² increment of a stationary field under a translation is the squared distance travelled by the Koopman orbit.

The L² modulus of continuity of the translation action is continuous. This is the only place where the ContinuousAction-style carrier hypotheses of Algsuperdiff/Probability/StationaryProjection.lean are used.

Mollification is an approximate identity on L²(Ω). For every tolerance there is a radius at which the mollification of a stationary L² field differs from it by at most that tolerance in squared L² norm.