math / data /lean /OAI /Analysis /Brenier /CellMeasure.lean
introvoyz045's picture
Migrated from GitHub
4a56cd6 verified
Raw History Blame Contribute Delete
21.3 kB
import OAI.Analysis.Brenier.CellGeometry
namespace OAI
noncomputable section
open MeasureTheory
open scoped ENNReal
namespace Problem358.NormalizedCell
variable {X : Type*} [MeasurableSpace X]
/-- The probability law obtained by normalizing the restriction to a positive finite cell. -/
def measure (ρ : Measure X) (P : Set X) : Measure X :=
(ρ P)⁻¹ • ρ.restrict P
variable {ρ : Measure X} {P : Set X}
/-- Positive finite extended mass is positive in the real-mass convention. -/
theorem real_mass_pos (hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) :
0 < ρ.real P :=
ENNReal.toReal_pos hpos.ne' hfinite
/-- Normalization is a probability whenever the cell has positive finite mass. -/
theorem probability (hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) :
IsProbabilityMeasure (measure ρ P) := by
constructor
simp only [measure, Measure.smul_apply, Measure.restrict_apply_univ, smul_eq_mul]
exact ENNReal.inv_mul_cancel hpos.ne' hfinite
/-- Multiplying the normalized law by its cell mass recovers the raw restriction. -/
theorem mass_smul (hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) :
(ρ P) • measure ρ P = ρ.restrict P := by
rw [measure, smul_smul, ENNReal.mul_inv_cancel hpos.ne' hfinite, one_smul]
/-- The same cancellation in the real-mass notation used by cell derivatives. -/
theorem ofReal_mass_smul (hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) :
ENNReal.ofReal (ρ.real P) • measure ρ P = ρ.restrict P := by
rw [Measure.real, ENNReal.ofReal_toReal hfinite]
exact mass_smul hpos hfinite
/-- The normalized restriction remains almost surely inside its measurable cell. -/
theorem ae_mem (hP : MeasurableSet P) : ∀ᵐ x ∂measure ρ P, x ∈ P := by
rw [ae_iff]
change measure ρ P Pᶜ = 0
simp [measure, Measure.smul_apply, Measure.restrict_apply hP.compl]
/-- No integrability is lost or gained by a positive finite cell normalization. -/
theorem integrable_iff {V : Type*} [NormedAddCommGroup V] {f : X → V}
(hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) :
Integrable f (measure ρ P) ↔ IntegrableOn f P ρ := by
exact integrable_smul_measure (ENNReal.inv_ne_zero.mpr hfinite)
(ENNReal.inv_ne_top.mpr hpos.ne')
/-- Weighted averages under the normalized cell law recover raw cell integrals.
The identity is valid for real or vector valued functions, even without integrability. -/
theorem mass_smul_integral {V : Type*} [NormedAddCommGroup V] [NormedSpace ℝ V]
(hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) (f : X → V) :
ρ.real P • (∫ x, f x ∂measure ρ P) = ∫ x in P, f x ∂ρ := by
change (ρ P).toReal • (∫ x, f x ∂measure ρ P) = _
rw [← integral_smul_measure f (ρ P), mass_smul hpos hfinite]
/-- Scalar form of the weighted-average identity. -/
theorem mass_mul_integral (hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) (f : X → ℝ) :
ρ.real P * (∫ x, f x ∂measure ρ P) = ∫ x in P, f x ∂ρ :=
mass_smul_integral hpos hfinite f
/-- Any random variable with the normalized cell law has the same weighted moments. -/
theorem mass_smul_integral_of_map_eq
{Z : Type*} [MeasurableSpace Z] {π : Measure Z} {g : Z → X}
{V : Type*} [NormedAddCommGroup V] [NormedSpace ℝ V]
(hpos : 0 < ρ P) (hfinite : ρ P ≠ ∞) (hg : Measurable g)
(hmap : Measure.map g π = measure ρ P) {f : X → V}
(hf : AEStronglyMeasurable f (measure ρ P)) :
ρ.real P • (∫ z, f (g z) ∂π) = ∫ x in P, f x ∂ρ := by
have hf' : AEStronglyMeasurable f (Measure.map g π) := hmap.symm ▸ hf
rw [← integral_map hg.aemeasurable hf', hmap]
exact mass_smul_integral hpos hfinite f
/-- Normalizing a cell removes any positive finite scalar prefactor in the source. -/
theorem measure_smul (c : ℝ≥0∞) (hc0 : c ≠ 0) (hctop : c ≠ ∞) :
measure (c • ρ) P = measure ρ P := by
unfold measure
rw [Measure.smul_apply, smul_eq_mul, Measure.restrict_smul, smul_smul,
ENNReal.mul_inv (Or.inl hc0) (Or.inl hctop), mul_right_comm,
ENNReal.inv_mul_cancel hc0 hctop, one_mul]
/-- Restricting the source to a larger set does not change the normalized cell law. -/
theorem measure_restrict {K : Set X} (hP : MeasurableSet P) (hPK : P ⊆ K) :
measure (ρ.restrict K) P = measure ρ P := by
unfold measure
rw [Measure.restrict_apply hP, Set.inter_eq_left.mpr hPK,
Measure.restrict_restrict_of_subset hPK]
end Problem358.NormalizedCell
namespace Problem358
/-- The probability law obtained by conditioning a measure on a positive finite cell. -/
def normalizedCellLaw {X : Type*} [MeasurableSpace X]
(ρ : Measure X) (P : Set X) : Measure X :=
(ρ P)⁻¹ • ρ.restrict P
variable {X V : Type*} [MeasurableSpace X]
[NormedAddCommGroup V] [NormedSpace ℝ V]
{ρ : Measure X} {P : Set X}
lemma normalizedCellLaw_isProbabilityMeasure (hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞) :
IsProbabilityMeasure (normalizedCellLaw ρ P) := by
constructor
rw [normalizedCellLaw, Measure.smul_apply, Measure.restrict_apply_univ, smul_eq_mul]
exact ENNReal.inv_mul_cancel hpos hfin
lemma normalizedCellLaw_ae_mem (hP : MeasurableSet P) :
∀ᵐ x ∂normalizedCellLaw ρ P, x ∈ P := by
exact Measure.ae_smul_measure (ae_restrict_mem hP) _
/-- Multiplying a normalized cell law by its original mass recovers the restricted measure. -/
lemma mass_smul_normalizedCellLaw (hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞) :
ENNReal.ofReal (ρ.real P) • normalizedCellLaw ρ P = ρ.restrict P := by
rw [Measure.real_def, ENNReal.ofReal_toReal hfin, normalizedCellLaw, smul_smul,
ENNReal.mul_inv_cancel hpos hfin, one_smul]
lemma integral_normalizedCellLaw (f : X → V) :
(∫ x, f x ∂normalizedCellLaw ρ P) = (ρ.real P)⁻¹ • ∫ x in P, f x ∂ρ := by
rw [normalizedCellLaw, integral_smul_measure, ENNReal.toReal_inv, Measure.real_def]
/-- Raw scalar moments and vector barycentres are the mass times their normalized integrals. -/
lemma mass_smul_integral_normalizedCellLaw (f : X → V)
(hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞) :
(ρ.real P) • (∫ x, f x ∂normalizedCellLaw ρ P) = ∫ x in P, f x ∂ρ := by
rw [← mass_smul_normalizedCellLaw hpos hfin, integral_smul_measure,
ENNReal.toReal_ofReal (measureReal_nonneg)]
omit [NormedSpace ℝ V] in
lemma integrable_normalizedCellLaw_iff (f : X → V)
(hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞) :
Integrable f (normalizedCellLaw ρ P) ↔ IntegrableOn f P ρ := by
exact integrable_smul_measure (ENNReal.inv_ne_zero.2 hfin) (ENNReal.inv_ne_top.2 hpos)
lemma lintegral_normalizedCellLaw (f : X → ℝ≥0∞) :
(∫⁻ x, f x ∂normalizedCellLaw ρ P) = (ρ P)⁻¹ * ∫⁻ x in P, f x ∂ρ := by
exact lintegral_smul_measure _ f
/-- A first marginal equal to a normalized cell law recovers the cell's raw moment. -/
lemma mass_smul_integral_fst_of_normalizedCellLaw
{Y : Type*} [MeasurableSpace Y] (π : Measure (X × Y)) (f : X → V)
(hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞)
(hmarg : Measure.map Prod.fst π = normalizedCellLaw ρ P)
(hf : IntegrableOn f P ρ) :
(ρ.real P) • (∫ p, f p.1 ∂π) = ∫ x in P, f x ∂ρ := by
have hi : Integrable f (Measure.map Prod.fst π) := by
rw [hmarg]
exact (integrable_normalizedCellLaw_iff f hpos hfin).2 hf
rw [← integral_map measurable_fst.aemeasurable hi.aestronglyMeasurable, hmarg]
exact mass_smul_integral_normalizedCellLaw f hpos hfin
/-- The corresponding second-marginal formula, used for the other coupled cell. -/
lemma mass_smul_integral_snd_of_normalizedCellLaw
{Y : Type*} [MeasurableSpace Y] (π : Measure (Y × X)) (f : X → V)
(hpos : ρ P ≠ 0) (hfin : ρ P ≠ ∞)
(hmarg : Measure.map Prod.snd π = normalizedCellLaw ρ P)
(hf : IntegrableOn f P ρ) :
(ρ.real P) • (∫ p, f p.2 ∂π) = ∫ x in P, f x ∂ρ := by
have hi : Integrable f (Measure.map Prod.snd π) := by
rw [hmarg]
exact (integrable_normalizedCellLaw_iff f hpos hfin).2 hf
rw [← integral_map measurable_snd.aemeasurable hi.aestronglyMeasurable, hmarg]
exact mass_smul_integral_normalizedCellLaw f hpos hfin
/-- Normalized laws of a finite almost-everywhere partition reconstruct the source. -/
lemma sum_mass_smul_normalizedCellLaw {ι : Type*} [Fintype ι]
(P : ι → Set X) (hP : ∀ i, MeasurableSet (P i))
(hdis : Pairwise (fun i j => AEDisjoint ρ (P i) (P j)))
(hcover : ∀ᵐ x ∂ρ, x ∈ ⋃ i, P i)
(hpos : ∀ i, ρ (P i) ≠ 0) (hfin : ∀ i, ρ (P i) ≠ ∞) :
(∑ i, ENNReal.ofReal (ρ.real (P i)) • normalizedCellLaw ρ (P i)) = ρ := by
simp_rw [mass_smul_normalizedCellLaw (hpos _) (hfin _)]
rw [← Measure.sum_fintype, ← Measure.restrict_iUnion_ae hdis
(fun i => (hP i).nullMeasurableSet)]
exact Measure.restrict_eq_self_of_ae_mem hcover
/-- Scalar moments and vector barycentres sum over normalized cell laws with raw masses. -/
lemma sum_mass_smul_integral_normalizedCellLaw {ι : Type*} [Fintype ι]
(P : ι → Set X) (hP : ∀ i, MeasurableSet (P i))
(hdis : Pairwise (fun i j => AEDisjoint ρ (P i) (P j)))
(hcover : ∀ᵐ x ∂ρ, x ∈ ⋃ i, P i)
(hpos : ∀ i, ρ (P i) ≠ 0) (hfin : ∀ i, ρ (P i) ≠ ∞)
(f : X → V) (hf : Integrable f ρ) :
(∑ i, (ρ.real (P i)) • ∫ x, f x ∂normalizedCellLaw ρ (P i)) = ∫ x, f x ∂ρ := by
simp_rw [mass_smul_integral_normalizedCellLaw f (hpos _) (hfin _)]
have h := integral_iUnion_ae (fun i => (hP i).nullMeasurableSet) hdis hf.restrict
rw [Measure.restrict_eq_self_of_ae_mem hcover, tsum_fintype] at h
exact h.symm
end Problem358
namespace Problem358
/-- The normalized restriction of a uniform source to a cell is exactly the
uniform law of that cell; this identifies the laws used by midpoint couplings. -/
theorem normalizedCell_uniformMeasure {d : ℕ} {K P : Set (E d)}
(hKpos : 0 < volume K) (hKfinite : volume K ≠ ∞)
(hP : MeasurableSet P) (hPK : P ⊆ K) :
NormalizedCell.measure (uniformMeasure K) P = uniformMeasure P := by
rw [uniformMeasure, NormalizedCell.measure_smul _
(ENNReal.inv_ne_zero.mpr hKfinite) (ENNReal.inv_ne_top.mpr hKpos.ne'),
NormalizedCell.measure_restrict hP hPK]
rfl
/-- Cell mass for a uniform source is the ratio of cell and source volume. -/
theorem uniformMeasure_real_cell {d : ℕ} {K P : Set (E d)}
(hP : MeasurableSet P) (hPK : P ⊆ K) :
(uniformMeasure K).real P = (volume P).toReal / (volume K).toReal := by
unfold Measure.real uniformMeasure
rw [Measure.smul_apply, smul_eq_mul, Measure.restrict_apply hP,
Set.inter_eq_left.mpr hPK, ENNReal.toReal_mul, ENNReal.toReal_inv]
ring
end Problem358
end
noncomputable section
open MeasureTheory Set
open scoped BigOperators
namespace Problem358.CellGeometry
variable {d : ℕ} {ι : Type*}
/-- A closed-cell mass is the normalized scalar volume integral used by the
moving-cut calculus. No positivity or distinct-site hypothesis is needed. -/
theorem uniformMass_eq_normalized_integral_one
{K : Set (E d)} (hK : IsClosed K)
(y : ι → E d) (c : ι → ℝ) (i : ι) :
(uniformMeasure K).real (closedCell K y c i) =
(volume K).toReal⁻¹ * ∫ _x in closedCell K y c i, (1 : ℝ) := by
rw [uniformMeasure_real_cell (isClosed_closedCell hK y c i).measurableSet
(closedCell_subset K y c i)]
simp only [integral_const, Measure.real, Measure.restrict_apply_univ,
smul_eq_mul, mul_one]
ring
/-- The uniform restriction on a cell rescales its volume integral. This also
holds for nonscalar moments and for degenerate source volume. -/
theorem integral_uniform_closedCell
{V : Type*} [NormedAddCommGroup V] [NormedSpace ℝ V]
(K : Set (E d)) (y : ι → E d) (c : ι → ℝ) (i : ι)
(f : E d → V) :
(∫ x in closedCell K y c i, f x ∂uniformMeasure K) =
(volume K).toReal⁻¹ • ∫ x in closedCell K y c i, f x := by
rw [uniformMeasure, Measure.restrict_smul,
Measure.restrict_restrict_of_subset (closedCell_subset K y c i),
integral_smul_measure, ENNReal.toReal_inv]
/-- A scalar moving-volume derivative immediately gives the actual normalized
real cell-mass derivative. -/
theorem hasDerivAt_uniformMass_of_integral_one
{K : Set (E d)} (hK : IsClosed K)
(y : ℝ → ι → E d) (c : ℝ → ι → ℝ) (i : ι) {t D : ℝ}
(hD : HasDerivAt (fun s => ∫ _x in closedCell K (y s) (c s) i, (1 : ℝ)) D t) :
HasDerivAt (fun s => (uniformMeasure K).real (closedCell K (y s) (c s) i))
((volume K).toReal⁻¹ * D) t := by
have h := hD.const_mul (volume K).toReal⁻¹
simpa only [uniformMass_eq_normalized_integral_one hK] using h
/-- The same normalization bridge for a Fréchet derivative in an arbitrary
real normed parameter space. -/
theorem hasFDerivAt_uniformMass_of_integral_one
{P : Type*} [NormedAddCommGroup P] [NormedSpace ℝ P]
{K : Set (E d)} (hK : IsClosed K)
(y : P → ι → E d) (c : P → ι → ℝ) (i : ι) {p : P} {D : P →L[ℝ] ℝ}
(hD : HasFDerivAt (fun q => ∫ _x in closedCell K (y q) (c q) i, (1 : ℝ)) D p) :
HasFDerivAt (fun q => (uniformMeasure K).real (closedCell K (y q) (c q) i))
((volume K).toReal⁻¹ • D) p := by
have h := hD.const_smul (volume K).toReal⁻¹
simpa only [uniformMass_eq_normalized_integral_one hK, Pi.smul_def, smul_eq_mul] using h
/-- Strict differentiability is preserved by the fixed source normalization,
as required by the implicit-function theorem. -/
theorem hasStrictFDerivAt_uniformMass_of_integral_one
{P : Type*} [NormedAddCommGroup P] [NormedSpace ℝ P]
{K : Set (E d)} (hK : IsClosed K)
(y : P → ι → E d) (c : P → ι → ℝ) (i : ι) {p : P} {D : P →L[ℝ] ℝ}
(hD : HasStrictFDerivAt
(fun q => ∫ _x in closedCell K (y q) (c q) i, (1 : ℝ)) D p) :
HasStrictFDerivAt
(fun q => (uniformMeasure K).real (closedCell K (y q) (c q) i))
((volume K).toReal⁻¹ • D) p := by
have h := hD.const_smul (volume K).toReal⁻¹
simpa only [uniformMass_eq_normalized_integral_one hK, Pi.smul_def, smul_eq_mul] using h
end Problem358.CellGeometry
end
/-! Coordinate assembly and source normalization for genuine vector cell moments.
The scalar moving-cell calculus is an explicit input; no differentiability of a
cell or of its coefficients is assumed implicitly. -/
noncomputable section
open MeasureTheory Set
open scoped ENNReal Topology BigOperators
namespace Problem358.CellVectorDerivative
variable {d : ℕ}
/-- A coordinate of a Bochner integral in the problem's Euclidean space. -/
theorem integral_coordinate {X : Type*} [MeasurableSpace X]
{μ : Measure X} {f : X → E d} (hf : Integrable f μ) (k : Fin d) :
(∫ x, f x ∂μ) k = ∫ x, f x k ∂μ :=
eval_integral_piLp (fun k => hf.eval_piLp k) k
/-- Assemble scalar coordinate derivatives in the Euclidean, rather than sup, norm. -/
theorem hasDerivAt_of_coordinates {f : ℝ → E d} {b : E d} {t : ℝ}
(h : ∀ k : Fin d, HasDerivAt (fun s => f s k) (b k) t) :
HasDerivAt f b t := by
have hp : HasDerivAt (fun s k => f s k) (fun k => b k) t := hasDerivAt_pi.mpr h
have hh := (EuclideanSpace.equiv (Fin d) ℝ).symm.toContinuousLinearMap.hasFDerivAt
|>.comp_hasDerivAt t hp
convert hh using 1 <;> simp [Function.comp_def]
/-- Restriction to a subset of the source removes the extra source indicator.
This identity is valid even in the zero-volume degenerate case. -/
theorem setIntegral_uniformMeasure {V : Type*} [NormedAddCommGroup V]
[NormedSpace ℝ V] {K P : Set (E d)} (hPK : P ⊆ K) (f : E d → V) :
(∫ x in P, f x ∂uniformMeasure K) =
(volume K).toReal⁻¹ • ∫ x in P, f x ∂volume := by
rw [uniformMeasure, Measure.restrict_smul,
Measure.restrict_restrict_of_subset hPK, integral_smul_measure, ENNReal.toReal_inv]
/-- Scalar volume derivatives of the coordinate weights give the actual
uniform-source vector-moment derivative, with the source normalization explicit. -/
theorem hasDerivAt_uniform_cellMoment_of_coordinates
{ι : Type*} {K : Set (E d)} (hK : IsCompact K)
(y : ℝ → ι → E d) (c : ℝ → ι → ℝ) (i : ι) {t : ℝ} {b : E d}
(h : ∀ k : Fin d, HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K (y s) (c s) i, x k ∂volume)
(b k) t) :
HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K (y s) (c s) i, x ∂uniformMeasure K)
((volume K).toReal⁻¹ • b) t := by
have hi (s : ℝ) : IntegrableOn (fun x : E d => x)
(CellGeometry.closedCell K (y s) (c s) i) volume :=
ContinuousOn.integrableOn_compact
(CellGeometry.isCompact_closedCell hK (y s) (c s) i) continuousOn_id
have hv : HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K (y s) (c s) i, x ∂volume) b t := by
apply hasDerivAt_of_coordinates
intro k
simpa only [integral_coordinate (hi _) k] using h k
have hn := hv.const_smul (volume K).toReal⁻¹
apply hn.congr_of_eventuallyEq
exact Filter.Eventually.of_forall fun s =>
setIntegral_uniformMeasure (CellGeometry.closedCell_subset K (y s) (c s) i) id
/-- Normalize the scalar cell-mass derivative using the same source factor. -/
theorem hasDerivAt_uniform_cellMass
{ι : Type*} {K : Set (E d)} (hK : IsCompact K)
(y : ℝ → ι → E d) (c : ℝ → ι → ℝ) (i : ι) {t a : ℝ}
(h : HasDerivAt
(fun s => (volume (CellGeometry.closedCell K (y s) (c s) i)).toReal) a t) :
HasDerivAt
(fun s => (uniformMeasure K).real (CellGeometry.closedCell K (y s) (c s) i))
(a / (volume K).toReal) t := by
simpa only [uniformMeasure_real_cell
(CellGeometry.isClosed_closedCell hK.isClosed _ _ _).measurableSet
(CellGeometry.closedCell_subset K _ _ _)] using h.div_const (volume K).toReal
/-- Coordinate formula for a finite sum of weighted vector surface moments. -/
theorem coordinate_sum_integral_smul {ι : Type*} [Fintype ι] {X : ι → Type*} [∀ j, MeasurableSpace (X j)]
(μ : ∀ j, Measure (X j)) (z : ∀ j, X j → E d) (w : ∀ j, X j → ℝ)
(hw : ∀ j, Integrable (fun x => w j x • z j x) (μ j)) (k : Fin d) :
(∑ j, ∫ x, w j x • z j x ∂μ j) k =
∑ j, ∫ x, w j x * z j x k ∂μ j := by
change (EuclideanSpace.proj k) (∑ j, ∫ x, w j x • z j x ∂μ j) = _
rw [map_sum]
congr 1
funext j
exact integral_coordinate (hw j) k
/-- The scalar face formulas for every coordinate assemble into the vector face
formula; integrability of the actual surface integrands is the only extra input. -/
theorem hasDerivAt_uniform_cellMoment_of_faces
{ι κ : Type*} [Fintype κ] {X : κ → Type*} [∀ j, MeasurableSpace (X j)]
{K : Set (E d)} (hK : IsCompact K)
(y : ℝ → ι → E d) (c : ℝ → ι → ℝ) (i : ι) {t : ℝ}
(μ : ∀ j, Measure (X j)) (z : ∀ j, X j → E d) (w : ∀ j, X j → ℝ)
(hw : ∀ j, Integrable (fun x => w j x • z j x) (μ j))
(h : ∀ k : Fin d, HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K (y s) (c s) i, x k ∂volume)
(∑ j, ∫ x, w j x * z j x k ∂μ j) t) :
HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K (y s) (c s) i, x ∂uniformMeasure K)
((volume K).toReal⁻¹ • ∑ j, ∫ x, w j x • z j x ∂μ j) t := by
apply hasDerivAt_uniform_cellMoment_of_coordinates hK y c i
intro k
rw [coordinate_sum_integral_smul μ z w hw k]
exact h k
/-- Separate the stationary intercept speed from normalized vector face moments.
No nonzero denominator assumptions are needed for this algebraic identity. -/
theorem normalized_stationary_face_sum
{κ : Type*} [Fintype κ] {X : κ → Type*} [∀ j, MeasurableSpace (X j)]
(μ : ∀ j, Measure (X j)) (z : ∀ j, X j → E d)
(M : ℝ) (a q : κ → ℝ) :
M⁻¹ • (∑ j, ∫ x, (q j / a j) • z j x ∂μ j) =
∑ j, q j • ((M * a j)⁻¹ • ∫ x, z j x ∂μ j) := by
simp_rw [integral_smul, Finset.smul_sum, smul_smul]
apply Finset.sum_congr rfl
intro j _
congr 1
simp only [div_eq_mul_inv, mul_inv_rev]
ring
/-- Convenient stationary-direction form, ready for the canonical face coefficient
`b_ij = (volume(K) * normalLength_ij)⁻¹ • firstMoment_ij`. -/
theorem hasDerivAt_uniform_cellMoment_stationary
{ι κ : Type*} [Fintype κ] {X : κ → Type*} [∀ j, MeasurableSpace (X j)]
{K : Set (E d)} (hK : IsCompact K)
(y : ι → E d) (c q : ι → ℝ) (i : ι)
(μ : ∀ j, Measure (X j)) (z : ∀ j, X j → E d)
(a r : κ → ℝ) (hz : ∀ j, Integrable (z j) (μ j))
(h : ∀ k : Fin d, HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K y (fun j => c j + s * q j) i,
x k ∂volume)
(∑ j, ∫ x, (r j / a j) * z j x k ∂μ j) 0) :
HasDerivAt
(fun s => ∫ x in CellGeometry.closedCell K y (fun j => c j + s * q j) i,
x ∂uniformMeasure K)
(∑ j, r j • (((volume K).toReal * a j)⁻¹ • ∫ x, z j x ∂μ j)) 0 := by
have hd := hasDerivAt_uniform_cellMoment_of_faces hK (fun _ => y)
(fun s j => c j + s * q j) i μ z (fun j _ => r j / a j)
(fun j => (hz j).smul (r j / a j)) h
rw [normalized_stationary_face_sum] at hd
exact hd
end Problem358.CellVectorDerivative
end
end OAI