Quantitative Stability of Optimal Transport Plans and Potentials

53 concepts29 papers1 proposal

Imagine you have a collection of data points and you want to find their average. If you nudge one point by a millimeter, the average moves by a tiny fraction of a millimeter. This predictable, well-behaved response to small changes is the essence of stability.

Now imagine doing this with entire probability distributions using Optimal Transport (OT). If you slightly perturb an input distribution—perhaps due to measurement noise or finite sampling—does the optimal plan to reshape it change smoothly, or does it snap into a completely different configuration?

Historically, proving that OT solutions are stable has been difficult because the geometry of transport maps can be highly rigid. However, a robust body of mathematical research has established quantitative stability guarantees for OT plans, potentials, maps, and barycenters. Understanding why these solutions remain stable is crucial for everything from proving statistical limit theorems to ensuring that generative models converge reliably.

The Dual Lever: Why Potentials and Plans Coalesce

To understand why transport plans do not wildly fluctuate under small perturbations, we must look at the dual formulation of optimal transport. Instead of looking directly at the coupling (the transport plan), we can analyze the potentials—scalar functions that act like optimal price vectors at the source and target locations.

In Central limit theorems for general transportation costs, Eustasio del Barrio et al. demonstrate that the stability of these optimal transport potentials under weak convergence is a fundamental prerequisite for statistical inference. When we only have access to empirical data (finite samples approximating a true distribution), we need to know that our empirical cost estimates won't explode. If the target distribution has a positive density, the optimal potentials are unique and stable, preventing discontinuous jumps in the transport map (Central limit theorems for empirical transportation cost in general dimension). This stability is precisely what allows the empirical transportation cost to converge to a standard Gaussian distribution, enabling the use of Central Limit Theorems in high-dimensional data analysis.

[Perturbed Marginals] ---> [Stable Dual Potentials] ---> [Predictable Transport Plan]
                                                                  |
                                                                  v
                                                      [Gaussian Convergence (CLT)]

When we introduce entropic regularization—which blurs the transport plan to make it computationally tractable via the Sinkhorn algorithm—this stability becomes even more structured. Marcel Nutz et al. proved in Stability of Schrödinger potentials and convergence of Sinkhorn’s algorithm that small perturbations in the input marginal distributions lead to correspondingly small, predictable changes in the computed optimal couplings and their associated Schrödinger potentials.

To quantify how small these changes are, Stephan Eckstein et al. developed novel geometric proof techniques, such as "shadow couplings" and "coordinate transformations," in Quantitative Stability of Regularized Optimal Transport and Convergence of Sinkhorn's Algorithm. They established that the optimal transport plan exhibits Hölder continuity relative to changes in the marginal distributions, even when dealing with unbounded supports and quadratic costs. This provides a strict mathematical bound on how much a transport coupling can shift, ensuring that algorithms like Sinkhorn converge predictably in the physically meaningful Wasserstein distance.

The Convexity of Averages: Stabilizing Wasserstein Barycenters

What happens when we want to compute the geometric average of multiple distributions—known as a Wasserstein barycenter? If the input distributions are noisy or evolving over time, does the calculated barycenter drift erratically?

In Existence and Consistency of Wasserstein Barycenters, Jean-Michel Loubes et al. established that finding a distributional mean is a continuous operation, meaning that small changes in the input distributions result in corresponding convergences in their barycenters. To turn this qualitative continuity into a quantitative guarantee, Alex Delalande et al. leveraged the dual formulation of the barycenter problem in Quantitative stability of barycenters in the Wasserstein space. They demonstrated that the stability of Wasserstein barycenters is fundamentally linked to the strong convexity of the barycenter functional.

Intuition helps clarify why this works:

  • Flat Landscape (Unstable): If an optimization landscape is flat, a tiny nudge to the inputs can cause the optimal solution to slide far away.
  • Steep Bowl (Stable): If the landscape is strongly convex—shaped like a sharp, steep bowl—the minimum is tightly locked in place. A nudge to the inputs might shift the bowl slightly, but the minimum cannot escape the bottom of the well.

By exploiting this strong convexity, the authors proved that Wasserstein barycenters are Hölder-continuous with respect to their input measures, providing explicit bounds on how much a barycenter can shift based on the dimensionality of the underlying space.

The challenge intensifies when we add structural constraints to the transport problem. For instance, in financial mathematics and robust risk management, Martingale Optimal Transport (MOT) requires that the transport plan preserves conditional expectations (the martingale property). This constraint introduces immense structural rigidity; you cannot easily perturb a distribution without risking violating the martingale constraint entirely.

Despite this rigidity, Mathias Beiglböck et al. demonstrated in Approximation of martingale couplings on the line in the adapted weak topology that martingale couplings are stable on the real line, provided one uses the "adapted" version of the weak topology. This specific topology accounts for the temporal flow of information. Because the couplings are stable, the optimal costs associated with these transport problems vary continuously with their marginal distributions, ensuring that financial models derived from market option prices remain robust to minor data inaccuracies (An extension of martingale transport and stability in robust finance).

Similarly, when projecting a probability measure onto a constrained set—such as the set of measures in a specific "convex order"—one might fear numerical instability. Lattice operations like taking the minimum or maximum in convex order are notoriously unstable. However, Benjamin Jourdain et al. proved in Lipschitz continuity of the Wasserstein projections in the convex order on the line that Wasserstein projections onto these sets are Lipschitz continuous in one dimension. This guarantee ensures that small perturbations in the input distributions result in strictly bounded changes to the projected output, preserving model reliability.

Bridging the Stability Gap in Machine Learning

Why do these abstract geometric bounds matter for practical machine learning? Consider measure-transport algorithms like Normalizing Flows, Flow-Matching, or GANs, which train a neural network to learn a map that transforms a simple reference measure into a complex target distribution.

As identified by Ricardo Baptista et al. in An Approximation Theory Framework for Measure-Transport Sampling Algorithms, a major theoretical hurdle is the "stability gap": the difficulty in relating the distance between two transport maps in function space to the statistical divergence between their resulting pushforward measures. If a learned map is slightly imperfect, does the resulting distribution completely break down? Their framework demonstrates that while metrics like the Kullback-Leibler (KL) divergence are highly sensitive to map variations, Wasserstein distances and Maximum Mean Discrepancy (MMD) are remarkably stable, providing a reliable foundation for measure transport.

Map Space Distance ||T_1 - T_2|| 
       |
       +---> (KL Divergence) --------> Highly Sensitive (Unstable Gap)
       |
       +---> (Wasserstein / MMD) ----> Bounded & Predictable (Stable Bedrock)

Furthermore, this stability allows researchers to bound estimation errors directly without relying on overly restrictive geometric assumptions. Shubo Li et al. developed map stability inequalities in Stability and Oracle Inequalities for Optimal Transport Maps between General Distributions that link estimation errors directly to the objective function without requiring bounds on the Hessian or curvature of the transport map.

This mathematical bridge allows generative modeling frameworks to maintain provable convergence guarantees even when dealing with heavy-tailed, complex, or unbounded real-world data (Provable Mixed-Noise Learning with Flow-Matching). By proving that optimal transport potentials and plans vary continuously with their inputs, these works collectively show why optimal transport functions as a robust, noise-tolerant framework across modern statistics and machine learning.

Key Foundational Work

Existence and consistency of Wasserstein barycenters (2016) — contributed 1 concept elements to this direction.
Dynamics of Optimal Partial Transport (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_stability_of_optimal_transport_potential_8a8009c1,
  title = {stability of optimal transport potentials},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stability-of-optimal-transport-potentials/rdr_8a8009c1485f4eaab04340d491364f00},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_stability_of_optimal_transport_potential_8a8009c1_v1,
  title = {stability of optimal transport potentials},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stability-of-optimal-transport-potentials/rdr_8a8009c1485f4eaab04340d491364f00/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake