Documentation

Graphon.Overlay

The overlay theorem: an MP bijection nearly achieves the cut distance (Rokhlin campaign, R3) #

This file proves the last live Rokhlin core, exists_mpEquiv_cutNormDiff_lt_add: on an atomless standard Borel probability space, the cut distance — an infimum over pairs of measure-preserving maps — is achieved up to any ε > 0 by pulling back one graphon along a single measure-preserving bijection, leaving the other bare. This is the classical overlay / bijection-density fact (Borgs–Chayes–Lovász–Sós–Vesztergombi Lemma 3.5, Janson Theorem 6.9), needed by the completeness telescope in Graphon/Compactness.lean.

It lives in its own file because the proof needs the Frieze–Kannan regularity lemma and the atomless carving primitives of Graphon/Regularity.lean, which sit downstream of Graphon/CutDistance.lean in the import graph.

The proof follows docs/overlay-scoping.md §4: no Birkhoff / rational approximation. For step graphons, the value of any coupling (φ, ψ) is realized exactly by an MP bijection: the coupling matrix λ_{ik} = μ(φ⁻¹(S_i) ∩ ψ⁻¹(T_k)) has row and column sums equal to the cell measures, so (atomlessly) both partitions can be refined into cells of mass exactly λ_{ik}, and exists_controlled_cell_alignment — applied twice, once for the answer bijection and once for a proof-only transfer — turns the matched refinements into the required measure-preserving α ≃ᵐ α. The only ε comes from step approximation.

Roadmap (R3) #

O1: prescribed-mass carving and partition assembly #

theorem Graphon.exists_disjoint_subsets_of_measures {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} [MeasureTheory.IsProbabilityMeasure μ] [StandardBorelSpace α] [MeasureTheory.NullSingletonClass μ] {n : } {C : Set α} :
MeasurableSet C∀ (r : Fin nENNReal), i : Fin n, r i μ C∃ (A : Fin nSet α), (∀ (i : Fin n), MeasurableSet (A i)) (∀ (i : Fin n), A iC) (Pairwise fun (i j : Fin n) => Disjoint (A i) (A j)) ∀ (i : Fin n), μ (A i) = r i

Prescribed-mass carving. Out of a measurable set C, carve finitely many pairwise disjoint measurable subsets of any prescribed masses r i with ∑ i, r i ≤ μ C.

This generalizes exists_equal_chunks_inside (equal masses 1/q) to arbitrary prescribed masses; it is the finite-matrix carving step of the overlay construction (docs/overlay-scoping.md §4, step 1).

O1: partition from disjoint cells plus a waste cell #

noncomputable def MeasurablePartition.ofCells {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} {n : } (A : Fin nSet α) (hmeas : ∀ (i : Fin n), MeasurableSet (A i)) (hdisj : Pairwise fun (i j : Fin n) => Disjoint (A i) (A j)) :

Assemble a MeasurablePartition from finitely many pairwise disjoint measurable cells, adding the complement of their union as a waste cell. Every input cell is a part (mem_ofCells_parts).

Equations
Instances For
    theorem MeasurablePartition.mem_ofCells_parts {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} {n : } {A : Fin nSet α} {hmeas : ∀ (i : Fin n), MeasurableSet (A i)} {hdisj : Pairwise fun (i j : Fin n) => Disjoint (A i) (A j)} (i : Fin n) :
    A i (ofCells A hmeas hdisj).parts

    Each input cell is a part of MeasurablePartition.ofCells.

    O2: the exact step-overlay transfer #

    theorem Graphon.sum_measure_inter_preimage_parts {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} (Q : MeasurablePartition α μ) {ψ : αα} ( : MeasureTheory.MeasurePreserving ψ μ μ) {X : Set α} (hX : MeasurableSet X) :
    TQ.parts, μ (X ψ ⁻¹' T) = μ X

    Splitting a measure along the preimages of a partition under a measure-preserving map: ∑_{T ∈ Q.parts} μ (X ∩ ψ⁻¹(T)) = μ X. (With ψ = id and X = univ this recovers that the part measures of a MeasurablePartition sum to one.)

    theorem Graphon.pullback_stepify_congr {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} [MeasureTheory.IsProbabilityMeasure μ] (P : MeasurablePartition α μ) (U : Graphon α μ) {f g : αα} (hf : MeasureTheory.MeasurePreserving f μ μ) (hg : MeasureTheory.MeasurePreserving g μ μ) (h : ∀ᵐ (x : α) μ, SP.parts, f x S g x S) :
    (stepify P U).pullback f hf = (stepify P U).pullback g hg

    Two measure-preserving maps that a.e. land in a common part of P pull a stepified graphon back to the same graphon. This is the transfer engine of the overlay proof: the stepification only sees which part a point lies in, never the point itself.

    theorem Graphon.exists_mpEquiv_pullback_stepify_eq {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} [MeasureTheory.IsProbabilityMeasure μ] [StandardBorelSpace α] [MeasureTheory.NullSingletonClass μ] (P Q : MeasurablePartition α μ) (U W : Graphon α μ) {φ ψ : αα} ( : MeasureTheory.MeasurePreserving φ μ μ) ( : MeasureTheory.MeasurePreserving ψ μ μ) :
    ∃ (σ : α ≃ᵐ α) ( : MeasureTheory.MeasurePreserving (⇑σ) μ μ), ((stepify P U).pullback (⇑σ) ).cutNormDiff (stepify Q W) = ((stepify P U).pullback φ ).cutNormDiff ((stepify Q W).pullback ψ )

    The exact step overlay. For step graphons, the cut-norm value of any coupling (φ, ψ) of measure-preserving maps is achieved exactly by a single measure-preserving bijection σ, leaving the second graphon bare.

    Construction (docs/overlay-scoping.md §4): the coupling matrix λ_{ik} = μ(φ⁻¹(S_i) ∩ ψ⁻¹(T_k)) has row and column sums equal to the part measures, so both partitions refine (atomlessly, exists_disjoint_subsets_of_measures) into cells of mass exactly λ_{ik}. Aligning the positive-mass refined cells twice via exists_controlled_cell_alignmentσ : T_{ik} → S_{ik} (the answer) and τ : T_{ik} → φ⁻¹(S_i) ∩ ψ⁻¹(T_k) (a proof-only transfer) — makes φ ∘ τ and σ land in a common part of P a.e., and ψ ∘ τ and id land in a common part of Q a.e.; then pullback_stepify_congr plus invariance of cutNormDiff under the bijection τ give the equality. No rational approximation or Birkhoff matching is needed.

    O3: the overlay theorem #

    theorem Graphon.exists_mpEquiv_cutNormDiff_lt_add {α : Type u_1} [MeasurableSpace α] {μ : MeasureTheory.Measure α} [MeasureTheory.IsProbabilityMeasure μ] [StandardBorelSpace α] [MeasureTheory.NullSingletonClass μ] (U W : Graphon α μ) {ε : } ( : 0 < ε) :
    ∃ (σ : α ≃ᵐ α) ( : MeasureTheory.MeasurePreserving (⇑σ) μ μ), (U.pullback (⇑σ) ).cutNormDiff W < U.cutDistance W + ε

    An MP bijection nearly achieves the cut distance (corrected Rokhlin consequence 4). On an atomless standard Borel probability space, the cut distance — an infimum over pairs of measure-preserving maps — is achieved up to any ε by pulling back U along a single measure-preserving bijection, leaving W bare.

    This is the honest, true replacement for the false cutDistance_lt_add_of_pos_equiv (Compactness) whose bijection witness was derived from the unprovable exists_common_extension_maps. It is genuinely needed by the completeness/compactness telescope: coherent frames require a bare reference, which forces invertibility (a coupling of maps cannot make a graphon bare without re-pulling the whole history).

    Proof (the overlay argument, docs/overlay-scoping.md): approximate U and W by step graphons at ε/8 (regularity); take ε/8-optimal maps (φ, ψ) for the original pair (cutDistance_lt_add_of_pos); realize the coupling value of the step approximations exactly by an MP bijection (exists_mpEquiv_pullback_stepify_eq); transfer the step approximation errors through the pullback contraction cutNormDiff_pullback_le and the bijection invariance cutNormDiff_pullback_measurableEquiv. Total budget 5ε/8 < ε.