Documentation

TauCeti.MeasureTheory.OptimalTransport.GromovWasserstein.Attainment

Optimal Gromov–Wasserstein couplings #

The Gromov–Wasserstein distance TauCeti.gromovWassersteinEDist p μ ωX ν ωY is an infimum of distortions over the couplings of μ and ν. This file proves that the infimum is attained when X and Y are Polish and μ and ν are Borel probability measures: some coupling has distortion equal to the distance. The kernels ωX and ωY are only assumed almost-everywhere strongly measurable, with values in an arbitrary pseudo-extended-metric space Z; no continuity, no lower semicontinuity, and no separability or completeness of Z is needed, and every exponent p : ℝ≥0∞ is allowed, including p = ∞. No integrability is assumed either: the optimal value may be ∞.

The proof is the direct method. The couplings form a nonempty weakly compact set (TauCeti.isCompact_setOfPred_isCoupling), so it suffices that the distortion is weakly lower semicontinuous on it (TauCeti.lowerSemicontinuousOn_gromovWassersteinDistortion). The distortion integrates the discrepancy edist (ωX (x, x')) (ωY (y, y')) against the square π ⊗ π, and this integrand is not lower semicontinuous when the kernels are merely measurable. It is continuous, however, on a closed set carrying all but an arbitrarily small part of the mass of π ⊗ π, uniformly in the coupling π: by Lusin's theorem ωX is continuous on a closed set K ⊆ X × X with (μ ⊗ μ) Kᶜ small, ωY on a closed L ⊆ Y × Y with (ν ⊗ ν) Lᶜ small, and the image of π ⊗ π on X × X is μ ⊗ μ and on Y × Y is ν ⊗ ν whatever π is. That uniformity is exactly the hypothesis of TauCeti.lowerSemicontinuousOn_lintegral_comp. For p = ∞ the same argument is applied to the indicator of {edist > c}, whose integral is positive exactly when the essential supremum exceeds c.

Main statements #

References #

Lower semicontinuity of the distortion. On second-countable pseudometrizable carriers, the p-distortion of a coupling of two fixed finite measures is weakly lower semicontinuous in the coupling, for almost-everywhere strongly measurable kernels and every exponent p.

Existence of an optimal Gromov–Wasserstein coupling for tight marginals. For tight Borel probability measures μ and ν on second-countable metrizable spaces and almost-everywhere strongly measurable kernels ωX and ωY with values in any pseudo-extended-metric space, some coupling of μ and ν has distortion equal to the p-Gromov–Wasserstein distance, for every exponent p. The common value may be ∞.

Existence of an optimal Gromov–Wasserstein coupling. For Borel probability measures μ and ν on Polish spaces and almost-everywhere strongly measurable kernels ωX and ωY with values in any pseudo-extended-metric space, some coupling of μ and ν has distortion equal to the p-Gromov–Wasserstein distance, for every exponent p. The common value may be ∞.

Optimal Gromov–Wasserstein couplings on standard Borel carriers. For probability measures on standard Borel spaces and almost-everywhere strongly measurable kernels with values in any pseudo-extended-metric space, some coupling has distortion equal to the p-Gromov–Wasserstein distance, for every exponent p. The common value may be ∞.

theorem TauCeti.exists_isCoupling_gromovWassersteinDistortion_dist_eq {X : Type u_4} {Y : Type u_5} [PseudoMetricSpace X] [PolishSpace X] [MeasurableSpace X] [BorelSpace X] [PseudoMetricSpace Y] [PolishSpace Y] [MeasurableSpace Y] [BorelSpace Y] (μ : MeasureTheory.Measure X) [MeasureTheory.IsProbabilityMeasure μ] (ν : MeasureTheory.Measure Y) [MeasureTheory.IsProbabilityMeasure ν] (p : ENNReal) :
∃ (π : MeasureTheory.Measure (X × Y)), IsCoupling π μ ν ∧ gromovWassersteinDistortion p (fun (x : X × X) => dist x.1 x.2) (fun (y : Y × Y) => dist y.1 y.2) π = gromovWassersteinEDist p μ (fun (x : X × X) => dist x.1 x.2) ν fun (y : Y × Y) => dist y.1 y.2

Optimal Gromov–Wasserstein couplings for distance kernels. For Borel probability measures on pseudometric spaces with Polish topologies, with the distance functions as kernels, the p-Gromov–Wasserstein distance is attained by some coupling. The chosen distances need not be complete.