---
title: Monge–Kantorovich Optimal Transport
url: https://www.emergentmind.com/topics/monge-kantorovich-optimal-transport-problem
type: topic
---

# Monge–Kantorovich Optimal Transport

The Monge–Kantorovich optimal transport problem is a foundational concept in mathematical analysis, probability, and optimization, addressing the task of transporting mass or distributions between spaces in a cost-minimizing way. This theory underpins numerous advances in geometry, partial differential equations, economics, machine learning, quantum information, and signal processing.

## 1. Definitions and Primal–Dual Structure

Let \((X,d_X)\) and \((Y,d_Y)\) be Polish spaces and \(\mu\in\mathcal{P}(X)\), \(\nu\in\mathcal{P}(Y)\) Borel probability measures. Consider a continuous cost \(c:X\times Y\to\mathbb{R}\).

**Monge’s problem** seeks a measurable map \(T:X\to Y\) (the transport map) such that \(T_\#\mu=\nu\) and
\[
\inf_{T_\#\mu=\nu}\int_X c\big(x,T(x)\big)\,d\mu(x).
\]

**Kantorovich’s relaxation** allows for transport plans \(\pi\in\mathcal{P}(X\times Y)\) with marginals \(\mu,\nu\):
\[
\inf_{\pi\in\Pi(\mu,\nu)}\int_{X\times Y}c(x,y)\,d\pi(x,y).
\]
Here, \(\Pi(\mu,\nu)\) is the set of Borel probability measures on \(X\times Y\) with specified marginals.

**Dual formulation:** Introduce potentials \((\varphi,\psi)\in L^1(\mu)\times L^1(\nu)\) with
\[
\varphi(x)+\psi(y)\leq c(x,y) \quad \forall (x,y)\in X\times Y.
\]
Kantorovich duality yields:
\[
\sup_{\varphi,\psi\,:\,\varphi(x)+\psi(y)\leq c(x,y)} \left\{\int_X\varphi\,d\mu + \int_Y\psi\,d\nu\right\} = \inf_{\pi\in\Pi(\mu,\nu)}\int_{X\times Y}c\,d\pi.
\]
For bounded below, lower semicontinuous costs, the supremum is attained for \(c\)-concave functions \(\varphi\) [1403.3103], [1011.2911].

## 2. Structural Results: Twist Conditions and Graph Decomposition

### m-Twist and Generalized-Twist

Assume \(c\) is \(C^1\) in \(x\). For each \(x_0\in X\), define
\[
\text{m-twist:} \qquad \#\{y \in Y : D_x c(x_0, y) = D_x c(x_0, y_0)\} \leq m,\ \forall y_0.
\]
\[
\text{generalized-twist:} \qquad \{y : D_x c(x_0, y) = D_x c(x_0, y_0)\} \text{ finite } \forall y_0.
\]

**Theorem 1.3 (finite-graph decomposition under m-twist):**  
Let \(X\) be a separable Riemannian manifold, \(\mu\) non-atomic, \(c\) bounded continuous and m-twist, every \(c\)-concave function differentiable \(\mu\)-a.e. Then every optimal plan admits a decomposition:
\[
\pi = \sum_{i=1}^k [a_i(x)\,(\mathrm{Id} \times G_i)_\# \mu],\quad 1\leq k\leq m, \ \sum_{i=1}^k a_i(x)=1\ \mu\text{-a.e.}
\]
for measurable maps \(G_i:X\to Y\), nonnegative weights \(a_i\in L^1(\mu)\) [1403.3103].  

**Theorem 1.4 (countable-graph under generalized-twist):**  
The same holds with a countable decomposition for generalized-twist.

This result extends Brenier–McCann’s uniqueness theorem for strictly twisted costs.

## 3. Uniqueness Criteria and Support

When the support of an optimal plan is a finite union of graphs, refined uniqueness results are available.

**Theorem 4.1 (uniqueness for finitely many graphs):**  
Given measurable maps \(T_i:X\to Y\) with injectivity and disjointness of ranges (except possibly for \(T_1\)), and a separation function \(\theta:Y\to\mathbb{R}\) such that
\[
\theta(T_1(x))-\theta(T_i(x))\geq 0\ \forall x, \quad\,(\text{equality } \Leftrightarrow T_1(x)=T_i(x)),
\]
there is at most one \(\pi\in\Pi(\mu,\nu)\) supported on \(\bigcup_{i=1}^k\mathrm{Graph}(T_i)\) [1403.3103].

Consequently, under strict twist (m=1), or when the convex decomposition collapses to one map, the solution is a unique Monge map; for m-twist, at most m maps appear in the decomposition.

## 4. Measure-Theoretic and Proof Perspectives

A central tool is the measure-theoretic selection principle and Choquet representation on the map \((T_1, T_2):(X,\mu) \to (X\times Y, \pi)\). The convex set of preimages of \(T_1\) admits an extreme-point decomposition (von Weizsäcker–Winkler theory):

- Each extreme section corresponds to a measurable selection, and the m-twist condition limits the number of these to m [1403.3103].
- By Kantorovich duality, optimal plans relate to \(c\)-concave potentials, whose gradients relate to the support: \(D_x c(T_1(x), T_2(x)) = \nabla\varphi(T_1(x))\) \(\mu\)-a.e.
- For the uniqueness theorem, the argument partitions the support into aperiodic unions and reduces uniqueness to that for doubly stochastic measures on such unions.

## 5. Illustrative Examples and Applications

- **Circle cost:** \(c(x,y)=1-\cos(x-y)\) on \(S^1\) is 2-twist; for \(\mu\neq\nu\), the optimizer is supported on two graphs.
- **Smooth manifolds:** A \(C^2\) cost \(c(x,y)\) with everywhere non-singular mixed Hessian \(D^2_{xy}c\) is locally 1-twisted, which recovers classical results that optimal plans are supported on Lipschitz submanifolds.
- **Classical case:** Under 1-twist and absolute continuity of \(\mu\), the unique solution is a Monge map, matching the Brenier–McCann theorem [1403.3103].

### Applications

- **Boundary Matching in Computer Vision:** Stratified Monge–Kantorovich problems address cases where sources and/or targets live on lower-dimensional sets, such as the interior/boundary decomposition relevant to shape recognition [2404.13616].
- **Extensions to Multi-Marginal Settings:** Recent theory generalizes the above characterization to multi-marginal optimal transport problems, replacing classical twist with twistedness on c-splitting sets, and yield analogous finite or countable graph representations of optimizers [1403.3389].

## 6. Generalizations and Future Directions

The measure-theoretic and duality-based framework for the Monge–Kantorovich problem allows for significant extensions:

- **Stratified OT:** Existence and uniqueness even under singular marginals supported on stratified or multi-layered sets, combining absolutely continuous and singular parts [2404.13616].
- **Graph Decomposition:** For non-degenerate costs (having locally finite pre-images of the gradient in \(y\)), the generalized-twist condition applies and the support decomposes accordingly.
- **Martingale and Vector-valued OT:** The abstract duality approach supports further generalizations, such as vector measure versions or problems with martingale constraints [2501.13557].
- **Numerical Methods:** Discretizations and adaptive schemes—such as monotone wide-stencil finite-difference and entropic-regularized linear programming—allow practical computation of optimal maps and plans, with convergence guarantees to the continuum problem [1710.05594], [1801.07745].
- **Quantum and Matrix-valued OT:** Quantum analogues define transport between density matrices with cost operators and semidefinite constraints; matrix-valued formulations allow spectral transport with rotation cost terms [2102.07787], [1304.3931].

## 7. References to Key Results

| Topic            | Main Theorems/Conditions                                 | Reference          |
|------------------|---------------------------------------------------------|--------------------|
| m-twist & decomposition | Theorems 1.3, 1.4: Graph decomposition, support structure | [1403.3103]        |
| Uniqueness      | Theorem 4.1: Uniqueness for support on union of graphs   | [1403.3103]        |
| Stratified, Lower-dim | Existence and decomposition for layers/mixed support | [2404.13616]       |
| Generalized-twist | Link to non-degenerate cost, support, existence         | [1403.3103]        |

This measure-theoretic characterization, leveraged with duality, provides a unified description of structure and uniqueness for optimal transport plans in a broad range of settings, encompassing and extending classical results. The support of optimal plans, governed by twist-type conditions, may comprise multiple graphs, but is always sharply controlled by the regularity—and degeneracy—properties of the cost.

Source: https://www.emergentmind.com/topics/monge-kantorovich-optimal-transport-problem