Lower semicontinuity of the Wasserstein distance under weak convergence #
On a Polish ground space the p-Wasserstein distance is a lower semicontinuous function of the
pair of laws for the weak topology on ProbabilityMeasure X × ProbabilityMeasure X: if
μₙ ⇀ μ and νₙ ⇀ ν, then
W_p (μ, ν) ≤ liminf W_p (μₙ, νₙ).
The inequality can be strict — for μₙ = (1 - 1 / n) δ₀ + (1 / n) δₙ on ℝ one has μₙ ⇀ δ₀
while W₁ (μₙ, δ₀) = 1 — so weak convergence does not make W_p continuous; that needs the
convergence of p-moments as well. Lower semicontinuity is what the direct method consumes
whenever a Wasserstein distance appears in a functional to be minimised over weakly compact sets
of laws.
For a positive finite exponent the distance is a continuous increasing function of the optimal
transport cost of edist ^ p (TauCeti.wassersteinEDist_rpow_eq_transportCost), and that cost is
weakly lower semicontinuous in the marginals by TauCeti.lowerSemicontinuous_transportCost; at
exponent zero the distance between probability measures vanishes. These arguments need only an
extended pseudometric whose topology is Polish. In this same regime the endpoint p = ∞ follows
from TauCeti.wassersteinEDist_top_eq_iSup, since a supremum of lower semicontinuous
functions is lower semicontinuous.
Main statements #
TauCeti.lowerSemicontinuous_wassersteinEDist— lower semicontinuity for every exponent, includingp = ∞, on an extended pseudometric space with Polish topology;TauCeti.wassersteinEDist_le_liminf— its form along weakly convergent families of laws;TauCeti.WassersteinSpace.edist_le_liminf— the same inequality for the distance from a fixed law of the Wasserstein spaceP_p (X), along weakly convergent families ofP_p (X).
References #
- C. Villani, Optimal Transport: Old and New, Grundlehren 338, Springer 2009, Chapter 6, where
W_pis shown to be lower semicontinuous under weak convergence. - L. Ambrosio, N. Gigli and G. Savaré, Gradient Flows in Metric Spaces and in the Space of Probability Measures, 2nd edition, Birkhäuser 2008, §7.1.
The Wasserstein distance is weakly lower semicontinuous, for every exponent including
p = ∞: on an extended pseudometric space with Polish topology, (μ, ν) ↦ W_p (μ, ν) is lower
semicontinuous on ProbabilityMeasure X × ProbabilityMeasure X.
Weak limits do not increase the Wasserstein distance: on an extended pseudometric space
with Polish topology, if
μₙ ⇀ μ and νₙ ⇀ ν along a filter l, then W_p (μ, ν) ≤ liminf W_p (μₙ, νₙ) for every
exponent p, including p = ∞.
Weak limits do not increase the distance from a fixed law of P_p (X): on a pseudometric
ground space with Polish topology, if the laws of a family v of P_p (X) converge weakly to the
law of ν along a filter
l, then W_p (μ, ν) ≤ liminf W_p (μ, v i) for every law μ of P_p (X).