2
$\begingroup$

I'm trying to find a bound for the multivariate Kolmogorov distance in terms of the Wasserstein distance. Denoting by $F$ and $G$ two cumulative distribution functions (cdf) on $\mathbb{R}^n$ the Kolmogorov and Wasserstein distances between $F$ and $G$ are given by \begin{align} d_{K}(F,G) &:= \sup_{\mathbf{x} \in \mathbb{R}^n} \left| F(\mathbf{x})- G(\mathbf{x}) \right|,\\ d_{W_1}(F,G) &:= \inf_{M_{F,G}} \left\{ \int_{\mathbb{R}^{n}\times\mathbb{R}^{n}} \Vert \mathbf{x}-\mathbf{y} \Vert dM_{F,G}(\mathbf{x},\mathbf{y}); M_{F,G} \text{ is a cdf with margins } F \text{ and } G \right\}, \end{align} where $\Vert \mathbf{x} \Vert$ is a vector norm on $\mathbb{R}^n$.

If $G$ has a density $g$ bounded by $C := \sup_{x\in\mathbb{R}} g(x)$ we have in the univariate case the following bound (see Lemma 1 in https://statweb.stanford.edu/~souravc/Lecture2.pdf): \begin{align} d_{K}(F,G) \leq 2 \sqrt{C} \sqrt{d_{W_1}(F,G)}. \end{align}

In the multivariate ($n$-dimensional) case I used the same approach to get the following result. If $G$ has a density $g$ bounded by $C := \sup_{\mathbf{x}\in\mathbb{R}^n} g(x)$ and $G$ is compactly supported on a set $D \subsetneq \mathbb{R}^n$ where $K = \max_{i=1,\ldots,n} \{ \sup_{\mathbf{x,y}\in D}|x_i-y_i| \}$ is the longest possible distance in one direction in $D$ then \begin{align} d_{K}(F,G) \leq 2 \sqrt{CnK^{n-1}} \sqrt{d_{W_1}(F,G)}. \end{align}

However, I would like to get rid of the bounded support assumption for $G$ (density, smoothness or moment assumptions for $F$ and $G$ would be fine) and I was wondering if anyone knows of a less restrictive result of the same flavor?

$\endgroup$
1
+50
$\begingroup$

The following result for the two-dimensional case can be straightforwardly generalised to higher dimensions.

Let $Q=\displaystyle\sup_{\bf{x}\in \mathbb{R}^2} \left\{ \int_0^x g(t,y)\mathrm{d}t+\int_0^y g(x,t) \mathrm{d}t\right\}$

Then $d_K(F,G)\le \sqrt{2Q}\sqrt{d_{W_1}}$

Proof:

Let the point $(x,y)$ be chosen to maximise $F(x,y)-G(x,y)$. Imagine that $f$ and $g$ are two different ways of distributing soil over the plane. Then the distance $d_K(F,G)$ is the difference between $f$ and $g$ in the amount of soil which is contained in the region $ S=\{(t_1,t_2):t_1 \le x, t_2\le y\}$.

The Wasserstein distance $d_{W_1}$ is the amount of soil that needs to be moved to transform $f$ into $g$ multiplied by the average distance that the soil needs to be moved.

If all of the soil which needs to be moved across the boundary of $S$ is right next to the boundary than the Wasserstein distance can be small. However, soil will have to be removed from $g$ on one or the other side of the boundary of $S$.

$Q$ is an upper bound on the 'cross-section' of $g$ near the boundary of $S$ and it can be deduced that a volume $\frac{d_K}2$ of earth must be spread out in such a way that it needs to be moved an average distance of at least $\frac{d_K}{2Q}$ to cross the boundary of $S$.

Therefore, $d_{W_1}\ge \frac{d_K^2}{4Q}$.

$\endgroup$
1
  • $\begingroup$ Thank you for taking the time to answer! If you know by any chance of a reference that I can cite I would be thankful for it, but for sure I will try to figure out the generalization. $\endgroup$ May 5 '19 at 21:50

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.