Documentation

TauCeti.MeasureTheory.OptimalTransport.GromovWasserstein.Space

The metric space of measured kernels #

GromovWassersteinPresentation p Z bundles a standard Borel probability carrier and an almost-everywhere strongly measurable Z-valued kernel with finite p-size. This is the finite-distance category, not a replacement for the unrestricted raw-measure definitions in GromovWasserstein.Basic. The exponent is a parameter, and the moment condition existentially quantifies its basepoint, so no chosen basepoint is part of a presentation.

For 1 ≤ p, presentations carry the Gromov–Wasserstein pseudometric. GromovWassersteinSpace p Z uses Mathlib's SeparationQuotient of this pseudometric. Two presentations have the same class exactly when their kernels are weakly isomorphic, not when their carriers or kernels are equal. The quotient has a genuine metric, with the original coupling formula as its extended distance. This includes p = ∞, where finite size means essential boundedness.

One-point kernels give an isometric embedding of Z into this space. Comparing with these kernels proves finiteness of all distances without requiring a globally separable or complete target. Common-parametrization bounds and filter convergence criteria transfer L^p convergence of pulled-back kernels on an auxiliary space directly to convergence of presentations and weak isomorphism classes. Carrier universes are fixed independently of the target universe.

References #

structure TauCeti.GromovWassersteinPresentation (p : ENNReal) (Z : Type v) [MetricSpace Z] :
Type (max (u + 1) v)

A finite-p measured-kernel presentation on a standard Borel probability carrier. The target needs no separability or completeness. Presentations themselves are not identified under null changes or measure-preserving parametrizations; GromovWassersteinSpace makes exactly those identifications.

Instances For
    @[implicit_reducible]

    The one-point presentation with constant kernel value z.

    Instances For
      @[simp]

      Distance to the constant one-point kernel is exactly the size about that point.

      Between constant one-point kernels, GW recovers the target metric at nonzero exponents.

      Finite size at possibly different basepoints suffices for finite GW distance.

      @[instance_reducible]

      The GW pseudometric on finite-size standard Borel presentations. Distinct presentations may have distance zero; only the subsequent quotient is a metric space.

      @[simp]

      The pseudometric on presentations is the original infimum over couplings.

      theorem TauCeti.GromovWassersteinPresentation.edist_le_eLpNorm_of_hasLaw {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {Ω : Type u_1} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {K L : GromovWassersteinPresentation p Z} {f : Ω → K.Carrier} {g : Ω → L.Carrier} (hf : ProbabilityTheory.HasLaw f K.measure γ) (hg : ProbabilityTheory.HasLaw g L.measure γ) :
      edist K L ≤ MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist (K.kernel (f q.1, f q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)

      The common-parametrization bound for presentations along a.e.-measurable maps: if f and g push an s-finite measure γ forward to K.measure and L.measure, the GW distance of K and L is at most the L^p (γ ⊗ γ) discrepancy of their pulled-back kernels.

      theorem TauCeti.GromovWassersteinPresentation.edist_le_eLpNorm {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {Ω : Type u_1} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {K L : GromovWassersteinPresentation p Z} {f : Ω → K.Carrier} {g : Ω → L.Carrier} (hf : MeasureTheory.MeasurePreserving f γ K.measure) (hg : MeasureTheory.MeasurePreserving g γ L.measure) :
      edist K L ≤ MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist (K.kernel (f q.1, f q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)

      The common-parametrization bound for presentations along measure-preserving maps.

      theorem TauCeti.GromovWassersteinPresentation.tendsto_of_tendsto_eLpNorm_of_hasLaw {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {ι : Type u_1} {Ω : Type u_2} {l : Filter ι} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {Ks : ι → GromovWassersteinPresentation p Z} {L : GromovWassersteinPresentation p Z} {f : (i : ι) → Ω → (Ks i).Carrier} {g : Ω → L.Carrier} (hf : ∀ᶠ (i : ι) in l, ProbabilityTheory.HasLaw (f i) (Ks i).measure γ) (hg : ProbabilityTheory.HasLaw g L.measure γ) (hlim : Filter.Tendsto (fun (i : ι) => MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist ((Ks i).kernel (f i q.1, f i q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)) l (nhds 0)) :

      Convergence of presentations from a common parametrization. Along any filter l, if a.e.-measurable maps f i (eventually along l) and g push an s-finite measure γ forward to (Ks i).measure and L.measure, and the L^p (γ ⊗ γ) discrepancy between the pulled-back kernels tends to zero, then Ks tends to L in the GW pseudometric.

      theorem TauCeti.GromovWassersteinPresentation.tendsto_of_tendsto_eLpNorm {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {ι : Type u_1} {Ω : Type u_2} {l : Filter ι} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {Ks : ι → GromovWassersteinPresentation p Z} {L : GromovWassersteinPresentation p Z} {f : (i : ι) → Ω → (Ks i).Carrier} {g : Ω → L.Carrier} (hf : ∀ᶠ (i : ι) in l, MeasureTheory.MeasurePreserving (f i) γ (Ks i).measure) (hg : MeasureTheory.MeasurePreserving g γ L.measure) (hlim : Filter.Tendsto (fun (i : ι) => MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist ((Ks i).kernel (f i q.1, f i q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)) l (nhds 0)) :

      Convergence of presentations from measure-preserving common parametrizations.

      def TauCeti.GromovWassersteinSpace (p : ENNReal) (Z : Type v) [MetricSpace Z] [Fact (1 ≤ p)] :
      Type (max (u + 1) v)

      Finite-size standard Borel measured kernels modulo weak isomorphism, with the full Gromov–Wasserstein metric. The carrier universe is fixed, but is independent of the target.

      Instances For
        @[instance_reducible]

        The separation quotient of the presentation pseudometric is a genuine metric space.

        Pass from a presentation to its weak-isomorphism class.

        Instances For

          Every GW class has a presentation in the fixed carrier universe.

          @[simp]

          The quotient extended distance is exactly the raw GW coupling infimum.

          Passing from presentations to weak-isomorphism classes is an isometry.

          @[simp]

          The real metric is the finite raw GW value read in ℝ.

          @[simp]

          Equality of classes is precisely weak isomorphism, witnessed by an exact coupling. It does not assert equality of presentations, even after null modifications.

          theorem TauCeti.GromovWassersteinSpace.tendsto_mk_of_tendsto_eLpNorm_of_hasLaw {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {ι : Type u_1} {Ω : Type u_2} {l : Filter ι} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {Ks : ι → GromovWassersteinPresentation p Z} {L : GromovWassersteinPresentation p Z} {f : (i : ι) → Ω → (Ks i).Carrier} {g : Ω → L.Carrier} (hf : ∀ᶠ (i : ι) in l, ProbabilityTheory.HasLaw (f i) (Ks i).measure γ) (hg : ProbabilityTheory.HasLaw g L.measure γ) (hlim : Filter.Tendsto (fun (i : ι) => MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist ((Ks i).kernel (f i q.1, f i q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)) l (nhds 0)) :
          Filter.Tendsto (fun (i : ι) => mk (Ks i)) l (nhds (mk L))

          Convergence in GW space from a common parametrization. Along any filter l, if a.e.-measurable maps f i (eventually along l) and g push an s-finite measure γ forward to (Ks i).measure and L.measure, and the L^p (γ ⊗ γ) discrepancy between the pulled-back kernels tends to zero, then mk (Ks i) tends to mk L.

          theorem TauCeti.GromovWassersteinSpace.tendsto_mk_of_tendsto_eLpNorm {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {ι : Type u_1} {Ω : Type u_2} {l : Filter ι} [MeasurableSpace Ω] {γ : MeasureTheory.Measure Ω} [MeasureTheory.SFinite γ] {Ks : ι → GromovWassersteinPresentation p Z} {L : GromovWassersteinPresentation p Z} {f : (i : ι) → Ω → (Ks i).Carrier} {g : Ω → L.Carrier} (hf : ∀ᶠ (i : ι) in l, MeasureTheory.MeasurePreserving (f i) γ (Ks i).measure) (hg : MeasureTheory.MeasurePreserving g γ L.measure) (hlim : Filter.Tendsto (fun (i : ι) => MeasureTheory.eLpNorm (fun (q : Ω × Ω) => edist ((Ks i).kernel (f i q.1, f i q.2)) (L.kernel (g q.1, g q.2))) p (γ.prod γ)) l (nhds 0)) :
          Filter.Tendsto (fun (i : ι) => mk (Ks i)) l (nhds (mk L))

          Convergence in GW space from measure-preserving common parametrizations.

          noncomputable def TauCeti.GromovWassersteinSpace.lift {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {β : Type u_1} (f : GromovWassersteinPresentation p Z → β) (hf : ∀ (K L : GromovWassersteinPresentation p Z), AreWeaklyIsomorphicKernels K.measure K.kernel L.measure L.kernel → f K = f L) :

          Descend any weak-isomorphism-invariant function on presentations to GW classes.

          Instances For
            @[simp]
            theorem TauCeti.GromovWassersteinSpace.lift_mk {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {β : Type u_1} (f : GromovWassersteinPresentation p Z → β) (hf : ∀ (K L : GromovWassersteinPresentation p Z), AreWeaklyIsomorphicKernels K.measure K.kernel L.measure L.kernel → f K = f L) (K : GromovWassersteinPresentation p Z) :
            lift f hf (mk K) = f K

            A descended function takes its prescribed value on every presentation.

            noncomputable def TauCeti.GromovWassersteinSpace.point {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] (z : Z) :

            The class of the one-point kernel with value z.

            Instances For
              @[simp]

              Distance from a presented kernel to a one-point class is its size about that point.

              @[simp]
              theorem TauCeti.GromovWassersteinSpace.mk_eq_point_iff {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] {K : GromovWassersteinPresentation p Z} {z : Z} :
              mk K = point z ↔ K.kernel =ᵐ[K.measure.prod K.measure] fun (x : K.Carrier × K.Carrier) => z

              A measured-kernel presentation is the one-point class at z exactly when its kernel is almost everywhere equal to z.

              @[simp]
              theorem TauCeti.GromovWassersteinSpace.edist_point {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] (z w : Z) :
              edist (point z) (point w) = edist z w

              One-point classes recover the ground extended distance, without a normalization factor.

              The target embeds isometrically as constant one-point kernels in GW space.

              @[simp]
              theorem TauCeti.GromovWassersteinSpace.dist_point {p : ENNReal} {Z : Type v} [MetricSpace Z] [Fact (1 ≤ p)] (z w : Z) :
              dist (point z) (point w) = dist z w

              The real distance of one-point classes is exactly their ground distance.