The Directional Optimal Transport

Marcel Nutz, Ruodu Wang

Abstract

We introduce a constrained optimal transport problem where origins $x$ can only be transported to destinations $y\geq x$. Our statistical motivation is to describe the sharp upper bound for the variance of the treatment effect $Y-X$ given marginals when the effect is monotone, or $Y\geq X$. We thus focus on supermodular costs (or submodular rewards) and introduce a coupling $P_{*}$ that is optimal for all such costs and yields the sharp bound. This coupling admits manifold characterizations -- geometric, order-theoretic, as optimal transport, through the cdf, and via the transport kernel -- that explain its structure and imply useful bounds. When the first marginal is atomless, $P_{*}$ is concentrated on the graphs of two maps which can be described in terms of the marginals, the second map arising due to the binding constraint.

Introduction

Optimal transport (OT) typically concerns the most efficient way to transform one probability distribution into another. While standard OT allows for arbitrary movement of mass, many real-world scenarios impose directional constraints. In "The Directional Optimal Transport," Marcel Nutz and Ruodu Wang explore a constrained Monge–Kantorovich problem where mass at origin $x$ can only be moved to a destination $y$ such that $y \geq x$. This directional constraint introduces a unique geometric structure and provides solutions to critical problems in causal inference.

Motivation: Monotone Treatment Effects

The primary impetus for this work is the study of monotone treatment effects in statistics. Consider a clinical trial where $X$ represents a baseline measurement (control) and $Y$ represents an outcome after treatment. While marginal distributions of $X$ and $Y$ are observable, their joint distribution—and thus the individual treatment effect $Y-X$—is not.

Researchers often seek the sharp upper bound for $Var(Y - X)$. Under the assumption of a monotone effect ($Y \geq X$), the classical "antitone" coupling (which matches high $X$ with low $Y$) may be infeasible if it violates the directional constraint. The authors define the Optimal Directional Coupling ($P_*$) to find the worst-case variance under this hard constraint.

Geometric and Discrete Characterizations

The coupling $P_*$ is characterized by several equivalent properties. In the discrete case, it follows a "Parenthesis Rule" similar to balancing brackets in computer science: starting from the right, each origin is matched to the leftmost available destination to its right.

Geometrically, $P_*$ satisfies a Constrained Crossing Property. In standard OT, optimal paths generally do not cross. However, in directional OT, paths should cross whenever the $y \geq x$ constraint permits it. If two paths do not cross but could be rearranged to cross without violating the constraint, the coupling is not yet optimal.

Discrete matching and the constrained crossing property.

The Two-Graph Structure and CDF Formula

A significant finding is that when the source distribution $\mu$ is atomless, the optimal coupling $P_*$ is concentrated on the graphs of two maps. Unlike standard OT, which often results in a single deterministic Monge map, directional OT frequently requires mass to be split:

  1. The Identity Map: Some mass remains at $y=x$ because the constraint is binding.
  2. A Transport Map $T(x)$: The remaining mass is moved to a destination $T(x) > x$.

This structure is governed by the difference between the marginal cumulative distribution functions (CDFs), $F(x) = F_\mu(x) - F_\nu(x)$. The joint CDF of $P_*$ is given by: $F_*(x, y) = F_\mu(x) - \inf_{z \in [x,y]} (F_\mu(z) - F_\nu(z))$ for $y > x$.

The transport map T(x) is determined by the behavior of the difference between marginal CDFs.

Key Insights

  • Universal Optimality: $P_*$ is a "master coupling" that is optimal for all supermodular cost functions simultaneously, not just for variance.
  • Randomization: Even with smooth marginals, the optimal solution often involves randomization (splitting mass between $x$ and $T(x)$), a departure from classical Monge solutions.
  • Invariance: The coupling is invariant under identical increasing transformations of both marginals.

Conclusion and Future Directions

The directional optimal transport framework provides sharp bounds for monotone treatment effects and reveals a rich mathematical structure on the real line. While the one-dimensional case is well-characterized, extending these results to higher dimensions ($X, Y \in \mathbb{R}^d$) remains a significant challenge, as the directional constraint evolves into a more complex cone constraint. Future research may also address soft constraints or the impact of discrete atoms in empirical data.

Summary updated

Abstract

We introduce a constrained optimal transport problem where origins $x$ can only be transported to destinations $y\geq x$. Our statistical motivation is to describe the sharp upper bound for the variance of the treatment effect $Y-X$ given marginals when the effect is monotone, or $Y\geq X$. We thus focus on supermodular costs (or submodular rewards) and introduce a coupling $P_{*}$ that is optimal for all such costs and yields the sharp bound. This coupling admits manifold characterizations -- geometric, order-theoretic, as optimal transport, through the cdf, and via the transport kernel -- that explain its structure and imply useful bounds. When the first marginal is atomless, $P_{*}$ is concentrated on the graphs of two maps which can be described in terms of the marginals, the second map arising due to the binding constraint.
BibTeX
@article{nutz2022directional,
  author = {Marcel Nutz and Ruodu Wang},
  title = {{The Directional Optimal Transport}},
  journal = {The Annals of Applied Probability},
  year = {2022},
  doi = {10.1214/21-aap1712},
}

Loading related content...

Figure