Multi-marginal optimal transport The notion of Wasserstein distance between two probability measures is well-studied and well-motivated in many different branches of math and stat. 
Let $\mu$ and $\nu$ be any two probability measures over a space $\mathcal{X}$ and let $\mathcal{M}(\mu,\nu)$  be the set of all couplings of $\mu$ and $\nu$. For a given metric $d$ over $\mathcal{X},$ the Wasserstein distance is defined as:
$$W(\mu, \nu):=\min_{(X,Y)\sim \pi\in \mathcal{M}(\mu,\nu)} \mathbb{E}d(X,Y)~.$$
The well-known Kantorovich duality states that 
$$W(\mu, \nu)=\sup_{Lip(f)\leq 1} \mathbb{E}_{\mu}f(X)-\mathbb{E}_\nu[f(X)],$$
where supremum is taked over all 1-Lipschitz functions, i.e. $f:\mathcal{X}\to \mathbb{R}$ such that $|f(x)-f(y)|\leq d(x, y)$ for all $x, y\in \mathcal{X}$.
In many different applications, it is favorable to find a similar quantity given $m$ probability measures. Let $\mu_1, \mu_2, \dots, \mu_m$ be $m$ probability measures on $\mathcal{X}$. The following is defined (for example see this)  as the multi-marginal version of the above
$$W_{(m)}(\mu_1, \mu_2, \dots, \mu_m):=\min \mathbb{E}_{\pi}c(X_1, X_2, \dots, X_m)~,$$
where $c$ is a continuous "cost" function and the infimum is taken over all joint measures $\pi$ on $\mathcal{X}^m$ whose marginals are $\mu_1, \mu_2, \dots, \mu_m$.
I have been trying to find a Kantorovich-like dual formula for $W_{(m)}$.
Has it been studied before? Any idea?  
 A: In general, under regularity conditions such as compactness, one can use standard minimax duality argument ($\min_y\max_x F(x,y)=\max_x\min_y F(x,y)$ -- for bilinear or, more generally, concave-convex functions $F$) to quickly show that 
\begin{multline}
 W(\mu_1, \mu_2, \dots, \mu_m):=\inf_\gamma\int c d\gamma \\
 =\inf_\gamma^*\sup_{f_1,\dots,f_m}^*\Big(
 \int c(x) \gamma(dx)
 +\sum_{j=1}^m \int f_i d\mu_i-\sum_{j=1}^m \int f_i(x_i) \gamma(dx)
 \Big) \\ 
 =\sup_{f_1,\dots,f_m}^*\inf_\gamma^*\Big(
 \int c(x) \gamma(dx)
 +\sum_{j=1}^m \int f_i d\mu_i-\sum_{j=1}^m \int f_i(x_i) \gamma(dx)
 \Big) \\ 
 =\sup
 \sum_{j=1}^m \int f_i d\mu_i,  
\end{multline}
where $\inf_\gamma$ is taken over all probability measures $\gamma$ with marginals $\mu_1, \mu_2, \dots, \mu_m$, $\inf\limits_\gamma^*$ is taken over all (nonnegative) measures $\gamma$ on $\mathcal{X}^m$, 
$\sup\limits_{f_1,\dots,f_m}^*$ is taken over all appropriately integrable real-valued measurable functions $f_1,\dots,f_m$ on $\mathcal{X}$, 
$\sup$ is taken over all such functions $f_1,\dots,f_m$ on $\mathcal{X}$ that satisfy the condition 
\begin{equation}
 c(x)\ge \sum_{j=1}^m f_i(x_i)\text{ for all }x\in\mathcal{X}^m, 
\end{equation}
$x=(x_1,\dots,x_m)$, $dx=dx_1\times\dots\times dx_m$.
The second equality above follows because 
\begin{multline}
 \sup_{f_1,\dots,f_m}^*\Big(
 \sum_{j=1}^m \int f_i d\mu_i-\sum_{j=1}^m \int f_i(x_i) \gamma(dx)
 \Big) \\
 =\begin{cases}
 0&\text{ if the marginals of $\gamma$ are $\mu_1, \mu_2, \dots, \mu_m$}, \\
 \infty&\text{ otherwise.}
 \end{cases}
\end{multline}
The third equality above is the minimax duality. 
The last equality there
follows because 
\begin{multline}
\inf_\gamma^*\Big(
 \int c(x) \gamma(dx)
-\sum_{j=1}^m \int f_i(x_i) \gamma(dx)
 \Big)  \\
 =\begin{cases}
 0&\text{ if $c(x)\ge \sum_{j=1}^m f_i(x_i)\text{ for all }x\in\mathcal{X}^m$}, \\
 -\infty&\text{ otherwise.}
 \end{cases}
\end{multline}

So, in general $\sum_{j=1}^m \int f_i d\mu_i$ and the condition $c(x)\ge \sum_{j=1}^m f_i(x_i)$ for all $x\in\mathcal{X}^m$ replace, respectively, 
  $\mathbb{E}_{\mu}f(X)-\mathbb{E}_\nu f(X)=\int f d\mu-\int f d\nu$ 
  and the condition $|f(x)-f(y)|\leq d(x, y)$ for all $x, y\in \mathcal{X}$ 
  in the definition of the Wasserstein distance. 

Even in the case $m=2$, the reduction from two functions $f_1$ and $f_2$ in the condition $c(x)\ge \sum_{j=1}^2 f_i(x_i)$ for all $x\in\mathcal{X}^2$ to one function $f$ in the condition $|f(x)-f(y)|\leq d(x, y)$ for all $x, y\in \mathcal{X}$ is based on the special properties of a metric, available only if $c=d$, a metric. This is how it is done: Suppose $d(x_1,x_2)\ge \sum_{j=1}^2 f_i(x_i)$ for all $x\in\mathcal{X}^2$, that is, $f(s):=\inf\limits_t[d(s,t)-f_2(t)]\ge f_1(s)$ for all $s$. Also, $f(s)\le d(s,s)-f_2(s)=-f_2(s)$ for all $s$, so that $f_1\le f$ and $f_2\le-f$ and hence 
\begin{equation}
 \sum_{j=1}^2 \int f_i\, d\mu_i\le \int f\, d\mu_1-\int f\, d\mu_2. 
\end{equation}
Also,
$$d(s,t)+f(t)=\inf\limits_u[d(s,t)+d(t,u)-f_2(u)]
\ge\inf\limits_u[d(s,u)-f_2(u)]=f(s)$$ 
and hence $d(s,t)\ge f(s)-f(t)$ for all $s,t$ or, equivalently, 
$|f(t)-f(s)|\le d(s,t)$ for all $s,t$. 
A: Actually, the answer is almost in the paper you linked. There the author refers to 

Duality theorems for marginal problems, Hans G. Kellerer, Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete, November 1984, Volume 67, Issue 4, pp 399–432, https://link.springer.com/article/10.1007/BF00532047

One dual problem is already stated on page one. If you are looking for a "Kantorovich-Rubinstein" dual problem  with Lipschitz-functions: I am not sure if such a thing exists - one uses the metric in an essential ways and I don't see anything that replaces this here…
