12
$\begingroup$

Let $(\Omega,d)$ be a compact metric space and $\mathcal P(\Omega)$ its space of Borel probability measures. Let $D=\{ d_p\mid p\in\Omega\}$ where $d_p(x)=d(p,x)$ be the set of all "distance functionals". As usual, we can think of $D$ acting on $\mathcal P(\Omega)$ (or vice versa) via integration i.e. $\langle d_p,\mu\rangle = \int_\Omega d_p(x)\,\mathrm d\mu(x)$.

Title Question

Does $D$ acting on $\mathcal P(\Omega)$ via integration separate points?

Or equivalently,

If $\mu,\nu \in \mathcal P(\Omega)$ and $\langle d_p,\mu\rangle = \langle d_p,\nu\rangle$ for all $p\in \Omega$, then must $\mu=\nu$?

Alternative Formulations

There are a few other ways to frame the question as well.

Probabilistic Formulation

Rewriting all integrals as expectations the question becomes,

If $\mathbb E_{X\sim\mu}[d_p(X)] = \mathbb E_{Y\sim\nu}[d_p(Y)]$ for all $p\in \Omega$, then must $\mu=\nu$?

In other words, does knowing the expected distance to a point for all points determine the measure?

Geometric Formulation

Recall that the 1-Wasserstein distance is given by $W_1(\mu,\nu) = \inf_{\gamma\in\Gamma(\mu,\nu)} \int_{\Omega\times\Omega} d(x,y) \,\mathrm d\gamma(x,y)$ where $\Gamma(\mu,\nu)$ is the set of couplings between $\mu$ and $\nu$ i.e. Borel probability measures on $\Omega\times\Omega$ with marginals $\mu$ and $\nu$ respectively. Since the product measure $\delta_p\otimes\mu$ is the unique coupling between a Dirac delta measure $\delta_p$ and $\mu$, we have that

$$W_1(\delta_p,\mu)=\int_{\Omega\times\Omega} d(x,y)\,\mathrm d(\delta_p\otimes\mu)(x,y)=\int_\Omega d(p,y)\,\mathrm d\mu(y)=\langle d_p,\mu\rangle$$

Now the question can be stated geometrically as

If $W_1(\delta_p,\mu)=W_1(\delta_p,\nu)$ for all $p\in \Omega$, then must $\mu=\nu$?

In other words, does knowing the $W_1$ distance to the extreme points of $\mathcal P(\Omega)$ completely determine the probability measure?

Integral Transform Forumlation

Define the distance transform of $\mu\in\mathcal P(\Omega)$ as the function $\phi_\mu:\Omega\to\mathbb R$ given by $\phi_\mu(p) = \int_\Omega d(p,x)\,\mathrm d\mu(x)$. The question can now be restated as,

Is the distance transform injective on $\mathcal P(\Omega)$?

Moreover, by the geometric formulation we have $\phi_\mu(p) = W_1(\delta_p,\mu)$. We will use the weak-$*$ topology for $\mathcal P(\Omega)$ (which coincides with the $W_1$ topology). Since the map $p\mapsto \delta_p$ is an embedding of $\Omega$ into $\mathcal P(\Omega)$, it follows that $\phi_\mu:\Omega\to\mathbb R$ is continuous. Denote the distance transform by $\Phi(\mu)=\phi_\mu$. Since $\mathcal P(\Omega)$ is compact Hausdorff and $C(\Omega)$ is Hausdorff we can restate the question as

If $\Phi:\mathcal P(\Omega)\to C(\Omega)$ is continuous, is it an embedding?

Final Thoughts

Are any of these equivalent statements true? I have unfortunately only been able to reformulate the question and have not identified any clear proof, though I wouldn't be surprised if there is an easy one I'm overlooking. The geometric formulation of the problem leads me to believe that $D$ does indeed separate points in $\mathcal P(\Omega)$. However, if the answer is affirmative then I feel the resulting nice properties of $\Phi$ would make it something that would be easy to look up. Any insight would be appreciated.

Update: In light of George Lowther's elegant 4-point counter-example and Pietro Majer's affirmative answer for $\Omega=[0,1]$, it would be interesting to better understand what factors determine whether the underlying metric space yields an affirmative answer.

George's counter-example can be extended to counter-examples where $\Omega$ is a sphere (with intrinsic metric). Thus, requiring $\Omega$ to be positive-dimensional, a manifold, connected, path-connected, simply-connected, etc, will not make the issue go away. On the other hand, Pietro suspects that the answer is again affirmative in the case when $\Omega$ is a compact convex subset of Euclidean space.

$\endgroup$
  • 1
    $\begingroup$ I don't know the answer, but one can reduce the problem to the case that $\Omega$ is finite, since probability measures with finite support are weak*-dense. $\endgroup$ – Michael Greinecker Aug 17 at 8:24
  • $\begingroup$ Isn't this a consequence of the Riesz representation theorem and the Stone-Weierstraß approximation theorem? $\endgroup$ – Jochen Glueck Aug 17 at 8:45
  • 1
    $\begingroup$ But Stone-Weierstraß only implies that the algebra generated by the distance functionals is dense. One needs the weak$^*$ density of the generated subspace. $\endgroup$ – Jochen Wengenroth Aug 17 at 9:15
  • $\begingroup$ @JochenWengenroth: You're right, of course. I confused a few arguments and thus thought this would not be a problem; but of course, it actually is a problem... $\endgroup$ – Jochen Glueck Aug 17 at 9:29
15
$\begingroup$

No. Suppose that $\Omega$ consists of four points arranged in a square, where adjacent points have distance 1 between them and opposite points have distance 2. Specifically, if the points are labeled A,B,C,D then \begin{align} & d(A,C)=d(B,D)=2,\\ & d(A,B)=d(B,C)=d(C,D)=d(D,A)=1. \end{align} For example, A,B,C,D could be equally spaced around a circle, using the internal circle metric.

There are precisely two probability measures assigning probability 1/2 to each of two opposite points and probability zero to the remaining two points. \begin{align} & \mu(\{A\})=\mu(\{C\}) = 1/2,\ \mu(\{B,D\})=0,\\ & \nu(\{B\})=\nu(\{D\})=1/2,\ \nu(\{A,C\})=0. \end{align} You can check that these two measures give the same integral for all `distance functions'. The average distance from every point is equal to 1 under both of these.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Elegant counter-example.This idea can be pushed a bit further as well though it was perhaps implicit: The empirical measure of a pair of antipodal points on the unit circle $\delta_p / 2+\delta_{-p} / 2$ produces an expected intrinsic distance of $\pi/2$ to any point on the circle. Thus, any two "empirical measure of antipodal points" cannot be separated by distance functionals. This shows the obstruction isn't simply due to finiteness of $\Omega$ and that the obstruction can't be avoided by requiring the underlying space to be nonzero dimensional, a manifold, connected, path-connected, etc. $\endgroup$ – Christian Bueno Aug 18 at 8:29
  • $\begingroup$ Actually, this idea using antipodal points works for spheres as well. Thus being simply-connected is not a sufficient condition to ensure an affirmative answer either. $\endgroup$ – Christian Bueno Aug 18 at 8:48
7
$\begingroup$

On the positive side, the answer is affirmative if $\Omega$ is the unit interval $[0,1]$ with its standard distance. In this case $\phi_\mu$ is a convex $1$-Lipschitz function (in fact, it is also defined for all $p\in\mathbb{R}$, with $\phi'(p)=\mathrm{sgn} p$ for $p\notin[0,1]$), with left and right derivatives $$\phi_-'(p)=\mu[0,p)-\mu[p,1]= 2\mu[0,p)-1$$ $$\phi_+'(p)=\mu[0,p] -\mu (p,1]= 2\mu[0,p] -1=1-2\mu(p,1]$$ so that $\mu$ is determined on all intervals, hence on all Borel subsets.

Conversely, note that any convex function $\phi$ as above
may be written in the form $\phi(p)=\int_{[0,1]}|t-p|dm(t)$ for some Borel probability measure $m$ on $[0,1]$. This because $g:= \frac{1}{2}\big(1-\phi_+'\big) $ is a nonnegative bounded cadlag function, so there is a Borel probability function $m$ such that $g(p)=m(p,1]$, whence $\phi(p)=\int_{[0,1]}|t-p|dm(t)$ follows easily from the above relations.

I'd guess the answer is also affirmative for $\Omega$ a convex compact set of $\mathbb{R}^n$ with the Euclidean distance.

| cite | improve this answer | |
$\endgroup$

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.