1-Wasserstein Metric and Generalizations

Author

Seyit Emre Duzoylum

Let \(\mathcal{M}\) denote the space of all Radon measures on \(\mathbb{R}^d\) with finite mass. Moreover, let \(\mathcal{M}^p\) denote the space of Radon measures with finite \(p^{th}\) moment, that is, \(\int_{\mathbb{R}^d} |x|^p d\mu(x) < \infty\). Then, for \(\mu,\nu \in \mathcal{M}^p\) with \(|\mu|=|\nu|\) the \(p\)-Wasserstein distance is defined as :\(W_p(\mu,\nu) = ( |\mu| \underset{ \pi \in \Gamma(\mu,\nu) }{ \inf } \int_{\mathbb{R}^d \times \mathbb{R}^d } |x-y|^p d\pi(x,y))^{1/p}\) where \(\Gamma(\mu,\nu)\) denotes the set of all transport plans from \(\mu\) to \(\nu\). Note that, if we restrict \(\mu\) and \(\nu\) to be probability measures, then the \(p\)-Wasserstein distance can be seen as a special case of the Kantorovich Problem, where \(c(x_1,x_2) = |x_1-x_2|^p\). Furthermore, if we let \(\mathcal{P}^p\) denote the space of probability measures on \(\mathbb{R}^d\) with finite \(p^{th}\) moment, then \(W_p\) defines a metric on \(\mathcal{P}^p\).

Measures with unequal mass and signed measures

The classical Wasserstein distance can be generalized for measures with unequal mass by allowing the addition and removal of mass to \(\mu\) and \(\nu\). In this case, there is unit cost \(a >0\) for the addition or removal of mass from both \(\mu\) and \(\nu\), whereas the transport cost of mass between \(\mu\) and \(\nu\) stays the same with the classical Kantorovich Problem; multiplied with some rate \(b >0\) .

Definition. For some \(a,b \in \mathbb{R}_{++}\) and \(p \geq 1\), the generalized Wasserstein distance \({W}_p^{a,b}\) is given by :\({W}_p^{a,b}(\mu,\nu) = \underset{\underset{ |\bar{\mu}| = |\bar{\nu}| }{\bar{\mu},\bar{\nu} \in \mathcal{M}^p}}{\inf} (a(|\mu-\bar{\mu}|) + a(|\nu-\bar{\nu}|) + b {W}_p(\bar{\mu},\bar{\nu}) )\) where \({W}_p(\bar{\mu},\bar{\nu})\) is the classical \(p\)-Wasserstein distance for measures with equal mass.

Note that, under this definition \({W}_p^{a,b}\) defines a metric on \(\mathcal{M}\), and \((\mathcal{M},{W}_p^{a,b})\) is a complete metric space.. Furthermore, this generalization allows one to extend the \(1\)-Wasserstein distance to the signed Radon measures as well. Let us denote the space of all signed Radon measures on \(\mathbb{R}^d\) with finite mass with \(\mathcal{M}^s(\mathbb{R}^d)\).

Definition. For some \(a,b \in \mathbb{R}_{++}\), the generalized Wasserstein distance for signed measures \(\mathbb{W}_1^{a,b}\) is given by :\(\mathbb{W}_1^{a,b}(\mu,\nu) = W_1^{a,b}(\mu_+ + \nu_- , \mu_- + \nu_+)\) where \(\mu_+,\nu_-,\mu_-,\nu_+\) are any measures in \(\mathcal{M}(\mathbb{R}^d)\) such that \(\mu = \mu_+ - \mu_-\) and \(\nu = \nu_+ - \nu_-\).

Moreover, if we let \(||\mu||^{a,b} = \mathbb{W}_1^{a,b}(\mu,0) = W_1^{a,b}(\mu_+ , \mu_- )\), then \((\mathcal{M}^s,||\mu||^{a,b})\) is a normed vector space. However, as opposed to the completeness of \((\mathcal{M},{W}_p^{a,b})\) , \((\mathcal{M}^s,||\mu||^{a,b})\) fails to be a Banach space.

Duality

As a special case of the Kantorovich Dual Problem when \(c(x_1,x_2) = |x_1-x_2|\), \(1\)-Wasserstein metric has the following dual characterization.

Theorem. Let Lip\((\mathbb{R}^d)\) denote the space of all Lipschitz functions on \(\mathbb{R}^d\), and let \(||\varphi||_{\text{Lip}} = \underset{x\neq y}{\sup} \frac{\varphi(x)-\varphi(y)}{|x-y|}\). Then,

\(W_1(\mu,\nu) = \sup \{ \int_{\mathbb{R}^d} \varphi d(\mu-\nu) : \varphi \in \mathcal{C}^0_c, \text{ } ||\varphi||_{\text{Lip}} \leq 1 \}\). In a similar spirit, this duality result can be extended for generalized \(1\)-Wasserstein distances \(W^{1,1}_1\) and \(\mathbb{W}^{1,1}_1\) as well, where \(a\) and \(b\) are taken as 1. For measures with unequal mass, when the additional constraint \(||\varphi||_\infty \leq 1\) is imposed on the test functions, it holds that

:\(W^{1,1}_1(\mu,\nu) = \sup \{ \int_{\mathbb{R}^d} \varphi d(\mu-\nu) : \varphi \in \mathcal{C}^0_c, \text{ } ||\varphi||_\infty \leq 1, \text{ } ||\varphi||_{\text{Lip}} \leq 1 \}\).

In terms of the generalized 1-Wasserstein metric for signed measures, when we relax the compact support condition on the test functions we obtain the equivalent duality characterization as follows.

:\(\mathbb{W}^{1,1}_1(\mu,\nu) = \sup \{ \int_{\mathbb{R}^d} \varphi d(\mu-\nu) : \varphi \in \mathcal{C}^0_b, \text{ } ||\varphi||_\infty \leq 1, \text{ } ||\varphi||_{\text{Lip}} \leq 1 \}\).

References

Piccoli, B., Rossi, F. Generalized Wasserstein Distance and its Application to Transport Equations with Source. Arch Rational Mech Anal 211, 335–358 (2014). https://doi.org/10.1007/s00205-013-0669-x

Piccoli, B., Rossi, F., Tournus M. A Wasserstein norm for signed measures, with application to nonlocal transport equation with source term. arXiv:1910.05105 (2019).

Piccoli, B., Rossi, F. Piccoli, B., Rossi, F. On Properties of the Generalized Wasserstein Distance. Arch Rational Mech Anal 222, 1339–1365 (2016). https://doi.org/10.1007/s00205-016-1026-7

C. Villani, Topics in Optimal Transportation, Chapter 1.