Unstable Optimal Transport Maps

Basic Information

• Paper ID: 2510.13265
• Title: Unstable optimal transport maps
• Author: Cyril Letrouit (Université Paris-Saclay, CNRS)
• Classification: math.OC (Mathematical Optimization and Control)
• Publication Date: October 15, 2025
• Paper Link: https://arxiv.org/abs/2510.13265

Abstract

The stability of optimal transport maps with respect to perturbations of marginal distributions is an important problem with implications for the validity of the linearized optimal transport framework, numerical analysis, and statistics. Under various assumptions on the source measure, it is known that the variation of optimal transport maps with respect to changes in the target measure is stable. This paper focuses on the opposite mechanism—factors that lead to instability. The author identifies two mechanisms and illustrates the instability, even severe instability, of optimal transport maps through examples of absolutely continuous source measures $\rho$ in $\mathbb{R}^d$. First, it is proven that unboundedness of the density can lead to instability: a source density on the unit ball in $\mathbb{R}^d$ is constructed that exhibits super-polynomial blow-up at two boundary points, with the corresponding optimal transport map being severely unstable. Then it is shown that even for uniform densities on bounded open sets, optimal transport maps can be quite unstable when approaching configurations where the uniqueness of the optimal plan is lost.

Research Background and Motivation

Problem Description

A central question in optimal transport theory is to study the stability of the optimal transport map $T_\mu$ (from a fixed source measure $\rho$ to a target measure $\mu$) with respect to perturbations of the target measure. Specifically, one investigates whether there exist constants $C, \alpha > 0$ and $p \geq 1$ such that for any probability measures $\mu, \nu$ on a compact set $Y \subset \mathbb{R}^d$: $\|\|T_\mu - T_\nu\|\|_{L^2(\rho)} \leq C W_p(\mu, \nu)^\alpha$

Significance

The importance of this problem manifests in multiple aspects:

1. Validity of the linearized optimal transport framework: The map $\mu \mapsto T_\mu$ provides an embedding from the Wasserstein space $(P(Y), W_2)$ into the Hilbert space $L^2(\rho, \mathbb{R}^d)$
2. Numerical analysis: When the target measure $\mu$ is only available through an approximation $\hat{\mu}$, one needs to ensure that the computed $T_{\hat{\mu}}$ is close to the true $T_\mu$
3. Statistical estimation: In statistical optimal transport, one needs to estimate optimal transport maps from samples

Limitations of Existing Research

Although quantitative stability inequalities have been established under various assumptions on the source measure $\rho$, it has never been proven that optimal transport maps can be unstable. Existing results primarily focus on positive stability results, lacking a deep understanding of the mechanisms of instability.

Core Contributions

1. First construction of severely unstable optimal transport maps: Proves the existence of absolutely continuous source measures for which no form of quantitative stability inequality holds
2. Identification of two instability mechanisms:
   • Super-polynomial blow-up of the density
   • Proximity to configurations where optimal plan uniqueness is lost
3. Provision of concrete counterexample constructions: Demonstrates instability phenomena through precise mathematical constructions
4. Revelation of regularity limitations of the map $\mu \mapsto T_\mu$: Proves that in certain cases this map is not $\frac{1}{2}$-Hölder continuous

Detailed Methodology

Task Definition

Given $\rho, \mu \in P_2(\mathbb{R}^d)$, where $\rho$ is absolutely continuous, Brenier's theorem guarantees the existence of a unique optimal transport map $T_\mu \in L^2(\rho)$ from $\rho$ to $\mu$. The study concerns the stability of $T_\mu$ with respect to perturbations of $\mu$.

Construction Method for Theorem 1.1

Source Measure Construction

Let $A = (1, 0, \ldots, 0)$, $A' = (-1, 0, \ldots, 0)$, $E = \{A, A'\}$. Define the function: $f(r) = r^{-d} \min(1, (\log r)^{-2})$

Construct the density: $\rho(x) = c_0 f(\text{dist}(x, E))$ where $c_0 > 0$ is chosen so that $\rho$ is a probability density on the unit ball $B_{\mathbb{R}^d}(0,1)$.

Target Measure Sequence

Consider $B_\theta = (R\sin\theta, R\cos\theta, 0, \ldots, 0)$ and $B'_\theta = (-R\sin\theta, -R\cos\theta, 0, \ldots, 0)$, and define: $\mu_\theta = \frac{1}{2}(\delta_{B_\theta} + \delta_{B'_\theta})$

Key Estimates

Through geometric analysis, it is proven that: $\|\|T_{\mu_\theta} - T_{\mu_0}\|\|^2_{L^2(\rho)} \geq \frac{c_0 c_1 \sigma_{d-1} R^2}{|\log(\theta/4)|}$

While $W_p(\mu_0, \mu_\theta) \sim R\theta/2$, therefore the stability inequality fails for any $\alpha > 0$.

Construction Method for Theorem 1.2

Support Set Construction

Construct a union of infinitely many pairs of rectangular parallelepipeds: $X = \bigcup_{i=1}^{+\infty} S_i$ where $S_i = T^+(A_i^+, \ell_i, r_i) \cup T^-(A_i^-, \ell_i, r_i)$.

Parameter Selection

Choose sequences satisfying:

• $r_i = c_0 2^{-i}$ (super-polynomial decay)
• $\ell_i = w_i = c_0 c_1 i^{-2}$
• Inter-cell distances far exceed intra-cell distances

Instability Proof

By analyzing the change in optimal transport maps under specific perturbations, it is proven that: $\frac{\|\|T_\mu - T_{\nu_i}\|\|^2_{L^2(\rho)}}{W_p(\mu, \nu_i)^{2\alpha}} \geq C w_i^{3-\frac{2\alpha}{p}} r_i^{1-\frac{2\alpha}{p}-2\alpha}$

When $\alpha > \frac{p}{2(p+1)}$, the right-hand side tends to infinity.

Experimental Setup

Theoretical Verification

This is a purely theoretical work, with results verified through rigorous mathematical proofs:

1. Geometric analysis: Detailed analysis of the geometric structure of optimal transport maps
2. Measure estimation: Precise calculation of measures in various regions
3. Distance computation: Accurate estimation of Wasserstein distances and $L^2$ distances

Construction Verification

The correctness of the construction is verified through the following steps:

1. Proving that the constructed density is indeed a probability density
2. Verifying the explicit expressions for optimal transport maps
3. Confirming that the geometric configuration satisfies the required separation conditions

Experimental Results

Main Results

Theorem 1.1 Results

For the constructed source density $\rho$ (exhibiting super-polynomial blow-up at two boundary points of the unit ball), it is proven that for any ball $Y = B_{\mathbb{R}^d}(0,R)$, any $C, \alpha > 0$ and $p \geq 1$, the inequality: $\forall \mu, \nu \in P(Y), \|\|T_\mu - T_\nu\|\|_{L^2(\rho)} \leq C W_p(\mu, \nu)^\alpha$ does not hold.

Theorem 1.2 Results

For uniform probability measures $\rho$ on bounded open sets, it is proven that for any $C > 0$, $p \geq 1$ and $\alpha > \frac{p}{2(p+1)}$ (in particular, $\alpha = \frac{1}{2}$), the stability inequality fails.

Technical Findings

1. Criticality of density blow-up: Polynomial blow-up $r^{-d+\delta}$ still maintains stability, but super-polynomial blow-up leads to complete instability
2. Hölder regularity limitations: The map $\mu \mapsto T_\mu$ is not $\frac{1}{2}$-Hölder continuous in certain cases, and may not even be $\frac{1}{3}$-Hölder continuous
3. Precursors of non-uniqueness: Stability deterioration occurs before optimal plan uniqueness is lost

Related Work

Positive Stability Results

• Gigli (2011): Theory of $\frac{1}{2}$-Hölder continuity of the map $\mu \mapsto T_\mu$
• Letrouit & Mérigot (2024): Stability inequalities for bounded densities on John domains
• Other works: Stability results for log-concave densities, polynomial blow-up densities, etc.

Instability Research

Previously, only instability examples of Kantorovich potential functions existed; this paper is the first to construct unstable transport maps.

Conclusions and Discussion

Main Conclusions

1. Optimal transport maps can be severely unstable, and any quantitative stability inequality may fail
2. Super-polynomial blow-up of the density is a sufficient condition for instability
3. Even for uniform densities, instability occurs when approaching non-uniqueness configurations

Limitations

1. Theorem 1.1 requires unbounded densities, which may be unrealistic in practical applications
2. The construction in Theorem 1.2 requires infinitely many connected components, resulting in complex geometry
3. It remains undetermined whether bounded densities can achieve the strong instability conclusion of Theorem 1.1

Future Directions

1. Study the strongest instability for bounded densities
2. Develop statistical estimation methods for unstable source measures
3. Explore deeper connections between instability and geometric structure

In-Depth Evaluation

Strengths

1. Theoretical breakthrough: First proof of instability of optimal transport maps, filling a theoretical gap
2. Ingenious construction: Achieves theoretical objectives through clever geometric construction and parameter selection
3. Rigorous analysis: Complete mathematical proofs with careful handling of technical details
4. Deep insights: Reveals intrinsic connections between stability, uniqueness, and density regularity

Weaknesses

1. Limited practical applicability: The constructed examples may be uncommon in practical applications
2. Open problems: Certain natural questions (e.g., strong instability for bounded densities) remain unresolved
3. Absence of numerical verification: As a theoretical work, it lacks numerical experiments for verification

Impact

1. Theoretical value: Provides important counterexamples and insights for optimal transport theory
2. Practical warning: Alerts practitioners to instability risks in certain configurations
3. Research inspiration: Provides new directions and tools for subsequent research

Applicable Scenarios

1. In-depth research in optimal transport theory
2. Theoretical analysis of the linearized optimal transport framework
3. Stability analysis in statistical optimal transport
4. Stability assessment of numerical optimal transport algorithms

References

The paper cites important literature in the optimal transport field, including foundational work by Brenier, regularity results by Gigli, and recent quantitative stability research, providing a solid theoretical foundation for the research.