We use $\mathcal{W}_2(\cdot, \cdot)$ to denote the quadratic Wasserstien distance as defined here. Now, let $X,Y = \mathcal{N}(0,1)$ be two standard normal random variables and for $ a \in[0,1]$ let $P_a$ be a discrete random variable, independent from $X,Y$, such that $$\mathbb{P}(P_a = 0) = 1-a \text{, } \mathbb{P}\left(P_a = \frac{1}{\sqrt{a}}\right) = a.$$ Consider the random variable $P_a\cdot X$, which is a centered mixture between a standard Gaussian, and a Gaussian with variance $\frac{1}{a}$. I'm intersted in estimating $$\mathcal{W}_2(P_a\cdot X, Y),$$ the quadratic Wasserstein distance between the mixture and a standard Gaussian.

Some easy bounds can be given by $$\mathcal{W}^2_2(P_a\cdot X, Y) \leq \mathbb{E}\left[(P_a - 1)^2\right]=a\left(\frac{1}{\sqrt{a}}-1\right)^2,$$ and $$\mathcal{W}^2_2(P_a\cdot X, Y) \geq \left(\sqrt{Var(P_a\cdot X)} - \sqrt{Var(Y)}\right)^2 = \left(\sqrt{2 - a} - 1\right)^2.$$

The two bounds leave a large gap between them, especially when $a$ is small, and I was wondering where the truth actually lies?

A more intricate question would be to ask for the Wasserstein distance between arbitrary mixtures, but that seems less tractable.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.