6

Consider the space $M$ of Borel measures on the real unit interval $[0,1]$, equipped with the 1-Wasserstein metric $d_W$ (or "Earth mover's distance"). The expected value is then a map $M\to [0,1]$ given by: $$ p \mapsto E_p := \int_{[0,1]} x \, dp(x)\;. $$

Is it true that this map is Lipschitz? That is, can we find a constant $C$ such that for every $p,q\in M$, $|E_p-E_q|\le C\,d_W(p,q)$ ?

geodude
  • 8,357

2 Answers2

9

Let $p$ and $q$ be two arbritrary probability measures on $[0,1]$. By definition,

$$d_W(p,q) = \inf_{\gamma \in \Gamma_{p,q}} \int |x-y| \, d\gamma(x,y)$$

where $\Gamma_{p,q}$ is the set of Borel measures on $[0,1] \times [0,1]$ with marginals $p$ and $q$. As $p$ and $q$ are probability measures, we have for any $\gamma \in \Gamma_{p,q}$

$$\int_{[0,1]} x \, dp(x) = \int_{[0,1] \times [0,1]} x \, d\gamma(x,y) \quad \text{and} \quad \int_{[0,1]} y \, dq(y) = \int_{[0,1] \times [0,1]} y \, d\gamma(x,y).$$

Hence,

$$\begin{align*} |E_p-E_q| &= \left| \int_{[0,1] \times [0,1]} (x-y) \, d\gamma(x,y) \right|\leq \int_{[0,1] \times [0,1]} |x-y| \, d\gamma(x,y). \end{align*}$$

Since this inequality holds for any $\gamma \in \Gamma_{p,q}$ we can take the infimum over all $\gamma \in \Gamma_{p,q}$ to conclude

$$|E_p-E_q| \leq d_W(p,q).$$

saz
  • 123,507
  • Thanks for this. For (my own) reference, this result holds for Borel probability measures on $\mathbb{R}$ and for any order $p \geq 1$ of the Wasserstein distance (due to Jensen's inequality). – ttt Dec 02 '19 at 05:55
3

For $(M,d)$ compact, let $I_f(\mu):=\int_M f\,d\mu$. The Kantorovich-Rubinstein formula states that,

$$ W_1(\mu,\nu) =\!\!\!\sup_{f\in\operatorname{Lip}_1(M,\mathbb R)} |I_f(\mu)-I_f(\nu)|, $$

Immediately we see that if $f$ is 1-Lipschitz then so is $I_f$ because $|I_f(\mu)-I_f(\nu)|\leq W_1(\mu,\nu)$. Moreover, if $f$ is $K$-Lipschitz then so is $I_f$ (simply substitute $g=\tfrac{1}{K}f\in\operatorname{Lip}_1(M,\mathbb R)$).

For your question, let $M=[0,1]$ and $f(x)=x$. Then $E_p=I_f(p)$ is 1-Lipschitz because $f$ is.

  • If you don't mind me asking, may I now if the duality formula for $W_1(\mu,\nu)$ remains valid when $M = \mathbb{R}^d$ (where $d \geq 1$)? – Fei Cao Jul 30 '23 at 22:30