TheoremBase

The swap of the optimal coupling induced by one map is the optimal coupling of the reversed pair, hence the one induced by the other map; evaluating both on the graph of the second map gives the composition identity.

Proof

Each result cited below is universally quantified over the data in its own statement.

Put π=(id,T)#μ\pi=(\mathrm{id},T)_{\#}\mu and π′=(id,T′)#ν\pi'=(\mathrm{id},T')_{\#}\nu. By Optimal Transport Maps and Uniquely Mapped Pairs of Probability Measures (i) these are optimal couplings of μ\mu and ν\nu, respectively of ν\nu and μ\mu.

1. The swap of π\pi is π′\pi'. Let σ\sigma be the swap of Couplings on Euclidean Space: Product Coupling, Swap, Finiteness of the Cost, Push-Forward Couplings, Modifying One Marginal, Quantisation, Gluing over a Finitely Supported Measure, and the Lipschitz Bound (ii). By that clause σ#π∈Π(ν,μ)\sigma_{\#}\pi\in\Pi(\nu,\mu) and I(σ#π)=I(π)I(\sigma_{\#}\pi)=I(\pi). Since π\pi is optimal, I(π)=W2(μ,ν)2I(\pi)=W_{2}(\mu,\nu)^{2}, and W2(μ,ν)=W2(ν,μ)W_{2}(\mu,\nu)=W_{2}(\nu,\mu) by The Quadratic Wasserstein Distance is a Metric on the Wasserstein Space (i); hence I(σ#π)=W2(ν,μ)2I(\sigma_{\#}\pi)=W_{2}(\nu,\mu)^{2}, so σ#π\sigma_{\#}\pi is an optimal coupling of ν\nu and μ\mu. Because (ν,μ)(\nu,\mu) is uniquely mapped, there is an optimal map RR from ν\nu to μ\mu such that every optimal coupling of ν\nu and μ\mu equals (id,R)#ν(\mathrm{id},R)_{\#}\nu; applying this to the two optimal couplings σ#π\sigma_{\#}\pi and π′\pi' gives

σ#π=(id,R)#ν=π′.\sigma_{\#}\pi=(\mathrm{id},R)_{\#}\nu=\pi' .

2. The first identity. Let

M={z∈Rd+d: pr2(z)=T′(pr1(z))},M=\bigl\{z\in\mathbb{R}^{d+d}:\ \mathrm{pr}_{2}(z)=T'\bigl(\mathrm{pr}_{1}(z)\bigr)\bigr\},

the graph of T′T' in the sense of A Coupling Concentrated on the Graph of a Borel Map is the Push-Forward by That Map, a member of B(Rd+d)\mathcal{B}(\mathbb{R}^{d+d}) as shown there. By Pairs of Euclidean Points: Coordinate Projections, Pairings, the Product Measure on a Euclidean Space, Borel Norm Functions and Finite Sets (i) and Pairs of Euclidean Points: Coordinate Projections, Pairings, the Product Measure on a Euclidean Space, Borel Norm Functions and Finite Sets (ii), (id,T′)(y)=ι(y,T′(y))(\mathrm{id},T')(y)=\iota(y,T'(y)) lies in MM for every y∈Rdy\in\mathbb{R}^{d}, so (id,T′)−1(M)=Rd(\mathrm{id},T')^{-1}(M)=\mathbb{R}^{d} and

π′(M)=ν((id,T′)−1(M))=ν(Rd)=1.\pi'(M)=\nu\bigl((\mathrm{id},T')^{-1}(M)\bigr)=\nu(\mathbb{R}^{d})=1 .

On the other hand σ(ι(x,y))=ι(y,x)\sigma(\iota(x,y))=\iota(y,x) by Pairs of Euclidean Points: Coordinate Projections, Pairings, the Product Measure on a Euclidean Space, Borel Norm Functions and Finite Sets (ii), so, using Pairs of Euclidean Points: Coordinate Projections, Pairings, the Product Measure on a Euclidean Space, Borel Norm Functions and Finite Sets (i),

σ−1(M)={z∈Rd+d: pr1(z)=T′(pr2(z))},\sigma^{-1}(M)=\bigl\{z\in\mathbb{R}^{d+d}:\ \mathrm{pr}_{1}(z)=T'\bigl(\mathrm{pr}_{2}(z)\bigr)\bigr\} ,

and (id,T)−1(σ−1(M))={x∈Rd:x=T′(T(x))}(\mathrm{id},T)^{-1}(\sigma^{-1}(M))=\{x\in\mathbb{R}^{d}:x=T'(T(x))\}, since (id,T)(x)=ι(x,T(x))(\mathrm{id},T)(x)=\iota(x,T(x)) has first coordinate xx and second coordinate T(x)T(x). Step 1 therefore gives

1=π′(M)=σ#π(M)=π(σ−1(M))=μ({x∈Rd: T′(T(x))=x}).1=\pi'(M)=\sigma_{\#}\pi(M)=\pi\bigl(\sigma^{-1}(M)\bigr)=\mu\bigl(\{x\in\mathbb{R}^{d}:\ T'(T(x))=x\}\bigr).

The set {x∈Rd:T′(T(x))≠x}\{x\in\mathbb{R}^{d}:T'(T(x))\ne x\} is its complement and belongs to B(Rd)\mathcal{B}(\mathbb{R}^{d}) as recorded in the statement, so claim 3 of Basic Properties of a Measure gives μ({x∈Rd:T′(T(x))≠x})=1−1=0\mu(\{x\in\mathbb{R}^{d}:T'(T(x))\ne x\})=1-1=0.

3. The second identity. The hypotheses of the statement are symmetric in the two triples (μ,ν,T)(\mu,\nu,T) and (ν,μ,T′)(\nu,\mu,T'): both (μ,ν)(\mu,\nu) and (ν,μ)(\nu,\mu) are uniquely mapped, TT is an optimal map from μ\mu to ν\nu, and T′T' is an optimal map from ν\nu to μ\mu. Exchanging the two triples in steps 1 and 2 yields ν({y∈Rd:T(T′(y))≠y})=0\nu(\{y\in\mathbb{R}^{d}:T(T'(y))\ne y\})=0. This proves claim 1 of the statement.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…