7
$\begingroup$

Fix a positive integer $n$, let $\mathbb{H}^n$ be the $n$-dimensional hyperbolic space, $r>0$, $x\in \mathbb{H}^n$ and consider the closed (compact) geodesic ball $B_{\mathbb{H}^n}(x,r)$. Are there known estimates on the minimum distortion of a bi-Lipschitz embedding of $B_{\mathbb{H}^n}(x,r)$ into $\mathcal{W}_2(\mathbb{R})$.

Let me just note that a bi-Lipschitz embedding into $\mathcal{W}_2(\mathbb{R})$ must exist by this paper.

$\endgroup$

3 Answers 3

2
$\begingroup$

For global distortion.

Choose a 4 point set --- vertices of an equilateral triangle in $B(x,r)_{\mathbb{H}^n}$ and its center and observe that it cannot be emebdded isometrically in nonnegatively curved space in the sense of Alexandrov, which includes $\mathcal{W}_2(\mathbb{R})$.

In is straightforward to get some explicit bounds this way, most likely they are far from being optimal.

$\endgroup$
4
  • 1
    $\begingroup$ Just apply the (3+1)-point comparison, you will get that for any $\varepsilon>0$ and all sufficiently large $R$ the map cannot be $[\sqrt3,2-\varepsilon]$-bilipschitz. $\endgroup$ Apr 3, 2022 at 16:48
  • 1
    $\begingroup$ By the way, if you apply (n+1)-point comparison you could get that distortion should be at least $\sqrt2-\varepsilon$ for large $R$. Most likely the bound should approach infinity, but I do not see an argument to prove it. $\endgroup$ Apr 3, 2022 at 16:55
  • $\begingroup$ May I ask you to add the details for this, I'm really not an expert in this type of argument :/ $\endgroup$ Apr 8, 2022 at 17:33
  • 1
    $\begingroup$ @Carlos_Petterson, I guess it is about (3+1)-point comparison --- check our book arxiv.org/abs/1903.08539 $\endgroup$ Apr 8, 2022 at 18:08
1
$\begingroup$

$\mathcal{W}_2(\mathbb{R})$ is isometric to a convex subset of a Hilbert space (and embedding a hyperbolic space into Euclidean/Hilbert one has a lot of history).

The argument can find an argument in S.S. Vallender "Calculation of the Wasserstein distance between probability distributions on the line". I will do an informal sketch.

I want to see probability measures in $\mathbb{R}$ as piles of sand with sand grains enumerated from left to right and indexed by $[0,1]$. For a measure $\mu$ define function $\Phi_{\mu}:[0,1]\rightarrow \mathbb{R}$ by $\Phi_\mu(t)$ be a position of $t$-th grain in $\mathbb{R}$. I claim that for every $\mu, \nu$ $$\mathcal{W}_2(\mu,\nu) = ||\Phi(\mu)-\Phi(\nu)||_{L_2}.$$

Before proving this let say how the optimal transport looks on $\mathbb{R}$.

Lemma 1: for the measures $\mu$ and $\nu$ the optimal transport plan between them sends $t$-th sand grain of $\mu$ into $t$-th sand grain of $\nu$.

I give the proof of the lemma for the case when $\mu = \frac{1}{n}(\delta(\mu_1) + \dots + \delta(\mu_n)),$ for some real $\mu_1 < \dots < \mu_n$ and $\nu = \frac{1}{n}(\delta(\nu_1) + \dots + \delta(\nu_n))$, for some real $\nu_1 < \dots < \nu_n$. I need to show that optimal transport moves $\frac{1}{n} \delta(\mu_1)$ into $\frac{1}{n}\delta(\nu_1)$ and then the rest will follow by induction. Okay, lets assume its not and (a part of) $\frac{1}{n} \delta(\mu_1)$ is moved to $\frac{1}{n} \delta(\nu_k)$ and (a part of) $\frac{1}{n}\delta(\mu_j)$ is moved to $\frac{1}{n} \delta(\nu_1)$ for some $j,k > 1$. I claim that if we change our transport by moving (the part of) $\frac{1}{n} \delta(\mu_1)$ into $\frac{1}{n}\delta(\nu_1)$ and (the part of) $\frac{1}{n}\delta(\mu_j)$ into $\frac{1}{n} \delta(\nu_k)$ it will get cheaper. Which follows from the following trivial lemma.

Lemma 2: Suppose that $a,b,c > 0$ then $a^2 + (b + c)^2 < (a + c)^2 + b^2$ iff $b < a$.

Now it's really easy to proof our main statement $$\mathcal{W}_2(\mu,\nu) = ||\Phi(\mu)-\Phi(\nu)||_{L_2}.$$
for the case when $\mu = \frac{1}{n}(\delta(\mu_1) + \dots + \delta(\mu_n)),$ for some real $\mu_1 < \dots < \mu_n$ and $\nu = \frac{1}{n}(\delta(\nu_1) + \dots + \delta(\nu_n))$, for some real $\nu_1 < \dots < \nu_n$. Indeed, squares of both sides are equal to $$\frac{1}{n}\sum_{i=1}^{n}|\mu_i - \nu_i|^2.$$

PS: note that measures of the type $\mu = \frac{1}{n}(\delta(\mu_1) + \dots + \delta(\mu_n)),$ for some real $\mu_1 < \dots < \mu_n$ are dense in Wasserstein space see Proposition 2.10 from Karl-Theodor Sturm. "On the geometry of metric measure spaces.". So all the above can be formalized with some suffering.

$\endgroup$
-1
$\begingroup$

It admits an isometric embedding.

The Nash--Kuiper theorem provides an isometric embedding of $\mathbb{H}^n\to \mathbb{R}^{n+1}$; moreover, its image might be in a bounded set. Now observe that $\mathcal{W}_2(\mathbb{R})$ contains a subset isometric to a ball in $\mathbb{R}^{n+1}$.

$\endgroup$
2
  • 2
    $\begingroup$ The Nash-Kuiper theorem provides an isometric embedding (in the Riemannian sense) and not in the metric sense (as YCor pointed out to me a couple of days ago). So then $f:\mathbb{H}^n \rightarrow \mathbb{R}^{2n+1}$ is only locally Lipschitz with constant $\max_{x\in ...}\,\|\nabla f(x)\|\neq 1$. $\endgroup$ Apr 2, 2022 at 12:15
  • $\begingroup$ @Carlos_Petterson then you should say "global" somewhere in your question, otherwise it is unclear. $\endgroup$ Apr 2, 2022 at 15:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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