Let $(X,d_X,\mu_X)$ and $(Y,d_Y,\mu_Y)$ be two metric measure spaces. A probability measure $\mu$ over $X\times Y$ is called a coupling if $(\pi_1)_\sharp \mu=\mu_X$ and $(\pi_2)_\sharp \mu=\mu_Y$. We define the Gromov–Wasserstein distance $d_{\mathcal GW}(X,Y)$ $^{[1]}$ by

$$\frac{1}{2}\inf_\mu(\int \int |d_X(x,x')-d_Y(y,y')|^p\mu(dx\times dy)\mu(dx' \times dy'))^{1/p}$$

where the infimum is taken over all the couplings of $\mu_X$ and $\mu_Y$.

I wonder if there are any computations/estimate for $d_{\mathcal GW}(S^m,S^n)$, where the distances are geodesic distances and measures are uniform measures on spheres.

Reference: [1] Mémoli, F. Found Comput Math (2011) 11: 417. https://doi.org/10.1007/s10208-011-9093-5


1 Answer 1


I'll preface this by saying that this is not a complete answer.

First, there is a very nice Python package called POT which has calculation of Wasserstein-Gromov distance included (for discrete measures).

I calculated quite a few examples using a quantization method to discretize the uniform distributions on the spheres. Numerical results suggest that the following coupling is optimal (for all $p\geq 1$), even though I cannot prove it:

Let $m > n$ and $\mu_X$ be the uniform distribution on $S_m$ and $\mu_Y$ the uniform distribution on $S_n$. Let $\mu = \mu_X \circ (id, T)^{-1}$, where $T$ is simply a projection of the spherical coordinates. So $(\varphi_1, ..., \varphi_{m-1})$ is mapped to $(\varphi_{m-n+1}, ..., \varphi_{m-1})$.

Notably, it is not even clear to me that the second marginal of $\mu$ again gives the uniform distribution on $S_n$, even though I guess people more familiar with spherical coordinates might see this immediately.

Initially I had hope to prove that this coupling is optimal by showing that any optimizer has to satisfy certain symmetries, which in the end leads to a coupling equivalent to this projection (a unique optimizer is of course impossible).


Your Answer

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

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