A Unified Kantorovich Duality for Multimarginal Optimal Transport

arXiv:2601.17171 · math.OC, stat.ML · Submitted 2026-01-23 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "A Unified Kantorovich Duality for Multimarginal Optimal Transport".

Jane: Multimarginal optimal transport (MOT) has gained increasing attention in recent years, notably due to its relevance in machine learning and statistics,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: Welcome back to the channel! We've got a really exciting paper on arXiv today, titled "A Unified Kantorovich Duality for Multimarginal Optimal Transport." Jane, I’m eager to hear what this work is all about and why it’s worth our time listening to.

Jane: It certainly sounds significant, Tom; this paper tackles multimarginal optimal transport by establishing a unified duality theory across both compact and general Polish product spaces with bounded continuous costs. Essentially, the main claim is that they've developed a complete Kantorovich duality framework for MOT that works regardless of whether the spaces involved are compact or not.

Lu: From my perspective as an AI researcher, this is interesting because it addresses the complexity introduced when you move beyond just two probability distributions to multiple marginals; Lu thinks the unification aspect across different topological settings is a really neat structural achievement.

Meng: I'm curious about how this theoretical framework translates into anything practical for real-world applications, Jane; does this mean we can actually apply it more robustly when dealing with complex machine learning models that require aligning several probability distributions simultaneously?

Lalam: Lalam sees a massive potential here; if the AI systems can leverage this unified duality to jointly compare and align multiple datasets, the resulting cultural impact could be in creating much more coherent and trustworthy multimodal AI.

Tom: Exactly what we're talking about! So, to get started, Jane, can you give us a quick summary of what this paper is actually proposing in its abstract? What's the core thesis they are pushing?

Jane: The core thesis of "A Unified Kantorovich Duality for Multimarginal Optimal Transport" is the presentation of a complete Kantorovich duality theory for MOT problem on general Polish product spaces with bounded continuous cost function. They show how to handle this problem consistently, whether you are dealing with compact spaces or non-compact ones.

Lu: What's particularly compelling is that they use two different paths—a convex-analytic reformulation for compact settings and a truncation-tightness procedure for non-compact ones—and then prove that these two approaches lead to the same duality identity, which is quite a feat in itself.

Meng: That sounds mathematically intensive; from an engineering standpoint, what does this unification actually mean for the computational side of things? Does it simplify the optimization process significantly when dealing with multiple distributions?

Lalam: I think simplifying the theoretical foundation means that future AI models won't have to worry about whether their data structure is finite or infinite when they try to compare different types of inputs; this structural stability is valuable.

Tom: That’s a big idea, Lalam. So, moving on, Jane, can you elaborate on what the authors claim about the structure of optimal potentials in this paper? What did they establish structurally?

Jane: The paper makes a central contribution by identifying a canonical structural class for optimal dual potentials: c-conjugate families. They prove that "optimal potentials may always be chosen in the class of c-conjugate families," which extends classical two-marginal conjugacy principles to this multimarginal setting.

Lu: Extending those classical results into the multimarginal realm is significant because it provides a predictable structure for how these optimal potentials behave, which is something we need when trying to analyze complex interactions.

Meng: So, if we can guarantee that the optimal potentials belong to this c-conjugate class, does that mean we have a more reliable way to find or approximate those solutions in practice?

Lalam: It suggests a predictable landscape for the solution space, which is incredibly helpful for designing stable algorithms that rely on these transport plans.

Tom: Absolutely. Now, let's transition into the conclusion of this discussion. Tom and Jane, can you give us your thoughts on what the title and authors of this paper imply about its broader impact? What's the big picture here?

Jane: The title itself suggests a comprehensive approach to optimal transport by unifying duality across different mathematical spaces; it points toward a more general theory applicable far beyond standard two-marginal problems.

Tom: And the implication is that this unified Kantorovich Duality for Multimarginal Optimal Transport provides a structural foundation for further developments in probabilistic and statistical analysis of MOT, offering a transparent framework particularly suited for empirical multimarginal optimal transport.

Lu: The fact that it offers this transparent framework is what makes it powerful; it doesn't just solve one problem, it gives the tools to tackle the whole class of problems in this area consistently.

Meng: From an engineering standpoint, that transparency is key because when you're building systems, you need to know exactly where the theoretical guarantees lie so you can build reliable computational approximations on top of them.

Lalam: Lalam feels like this work lays crucial groundwork for statistical tools, which means future AI can develop better methods for assessing and comparing the reliability of complex, multi-source information streams.

Tom: Well said, Lalam. So, to wrap up this segment on "A Unified Kantorovich Duality for Multimarginal Optimal Transport," we've established that the paper delivers a unified theory covering both compact and non-compact settings and proves that optimal dual potentials have a c-conjugate structure. We’ll keep exploring the deeper implications of these structural results in our next segment.

Conclusion: Tom: So, we’ve spent our time walking through how this paper sets up Kantorovich duality for multimarginal optimal transport on both compact and general Polish spaces, establishing that optimal potentials always fit in a c-conjugate form. Jane, looking at the title and authors of "A Unified Kantorovich Duality for Multimarginal Optimal Transport," what’s your take on what this actually means in plain language?

Jane: I see it as taking a problem that used to have separate rules for compact and non-compact spaces and building one single rulebook that works everywhere. The authors are showing us how to make sure the mathematical framework stays consistent across different types of spaces, which is a really helpful conceptual move.

Lu: From my side, I think the unification aspect is what’s truly fascinating; it means the underlying structural properties of these optimal solutions aren't dependent on whether we are dealing with finite or infinite sets of points in our product space.

Meng: That consistency suggests that whatever structure emerges from this duality will be robust, which is important when we try to build scalable models that operate on massive, real-world datasets.

Lalam: I see the potential for this unified structure to allow AI systems to compare and align multiple complex data modalities with a single, coherent mathematical language rather than having separate tools for each case.

Tom: It really sounds like this paper is providing a solid foundation for how we think about comparing different probability distributions in machine learning applications. Jane, can you explain the practical impact of this unified approach to our listeners?

Jane: The practical impact is that future statistical methods won't have to constantly check if their underlying space is compact or not before applying them; they can just use this unified framework, which simplifies the theoretical machinery immensely.

Lu: This structural understanding could open up entirely new avenues for creative applications in areas like generative modeling where we need to manage dependencies across many variables simultaneously.

Meng: For engineering, it means that when we look at optimizing a system involving many inputs and outputs, we have a more predictable class of solutions to work with instead of having to deal with potentially unstable approximations.

Lalam: And for culture and society, this suggests we can develop AI tools that are more capable of synthesizing information from diverse sources in a way that respects the underlying mathematical constraints consistently across those sources.

Tom: Fantastic points, team. So, it’s about providing a single, robust mathematical language for handling complex multi-distribution alignment across all topological settings. We'll keep digging into the specific structural results next on how those c-conjugate potentials actually manifest in practice and what that means for finding solutions in concrete scenarios.

Université de Technologie de Compiègne

math.OC, stat.ML

Submitted: 2026-01-23

Updated: 2026-10-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: Multimarginal optimal transport (MOT) has gained increasing attention in recent years, notably due to its relevance in machine learning and statistics, where one seeks to jointly compare and align

Key concepts

Multimarginal Optimal Transport (MOT): MOT is the problem of finding an optimal joint probability distribution that simultaneously aligns multiple given probability distributions. It is crucial in machine learning and statistics when comparing several datasets or distributions together.
Kantorovich Duality: This is a fundamental principle in optimal transport that relates the minimum cost of moving mass between distributions (the primal problem) to the maximum value of a dual function (the dual problem). The paper unifies this duality across different types of spaces.
c-Conjugate Families: This refers to a specific structural class for optimal dual potentials. The paper proves that any optimal potential can be represented within this c-conjugate family, meaning the solution has a predictable mathematical structure related to the marginal variables.

Terminology

Optimal Transport — 32× in this paper · explained in 5 episodes
Admissibility — 17× in this paper · explained in 2 episodes
Multimarginal Optimal Transport (MOT) — 5× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Wasserstein Distance — 3× in this paper · explained in 4 episodes
Generative Modeling — 2× in this paper · explained in 3 episodes
Adaptation — 1× in this paper · explained in 1 episode
Adversarial Training — 1× in this paper · explained in 1 episode
Alignment — 1× in this paper · explained in 1 episode
Consistency — 1× in this paper · explained in 3 episodes
Domain Adaptation — 1× in this paper · explained in 5 episodes
Generalization — 1× in this paper · explained in 13 episodes
Generative Model — 1× in this paper · explained in 2 episodes
Generative Models — 1× in this paper · explained in 2 episodes
Interpretability — 1× in this paper · explained in 15 episodes
Optimal Transport (OT) — 1× in this paper · explained in 4 episodes
Representation Learning — 1× in this paper · explained in 2 episodes
Sample Complexity — 1× in this paper · explained in 15 episodes
Theoretical Guarantees — 1× in this paper · explained in 2 episodes
Unified Framework — 1× in this paper · explained in 5 episodes

Summary

Multimarginal optimal transport (MOT) has gained increasing attention in recent years, notably due to its relevance in machine learning and statistics, where one seeks to jointly compare and align multiple probability distributions. The gist: This work establishes a unified Kantorovich duality theory for MOT on both compact and general Polish product spaces under bounded continuous costs, proving that optimal dual potentials can always be chosen in a c-conjugate form.

Duality Framework Establishment

The paper presents a unified Kantorovich duality framework for MOT on general Polish product spaces with bounded continuous cost functions. The core of the theory is established through two main paths: one for compact spaces using a convex-analytic reformulation and Fenchel–Rockafellar duality, and another for non-compact spaces employing a truncation-tightness procedure based on weak compactness of multimarginal transference plans. This unified approach ensures the duality identity holds across both settings.

Structural Results on Optimal Potentials

A central contribution is the identification of a canonical structural class for optimal dual potentials: c-conjugate families. The paper proves that optimal potentials may always be chosen in the class of c-conjugate families, extending classical two-marginal conjugacy principles to this multimarginal setting. In the compact case, this is achieved through convex-analytic reformulation and c-splitting sets. In the non-compact setting, dual attainment is recovered by exploiting the equivalence between optimality and splitting structures of supports and showing that any splitting family admits a canonical c-conjugate representative.

Duality in Compact Settings

For compact metric spaces, Theorem 1 establishes the duality formula: MOTc(µ1,..., µK) = inf π∈Π(µ1,..., µK) I(π) = sup f∈Fc J(f). This is derived by reformulating the dual problem using two convex functionals, Θ and Ξ. The Legendre-Fenchel transforms of these functionals are shown to be indicator functions, leading directly to the primal-dual equality. Furthermore, Theorem 3 proves that there exists a solution where fk ∈ c-conj(Xk) for each k = 1,..., K, which is attained through an argument involving equicontinuity estimates and the Arzelà–Ascoli theorem.

Duality in Non-Compact Settings

In non-compact Polish spaces, the absence of global compactness is addressed via a truncation-tightness procedure based on weak compactness of multimarginal transference plans. The proof proceeds by quantitatively localizing the problem onto compact subsets, followed by a stabilization of dual potentials that allows passage to the limit. This strategy ensures that while direct compactness fails, the duality formula is recovered as an asymptotic limit: sup f∈Fc J(f) ≥ inf π∈Π(µ1,..., µK) I(π).

Optimal Potentials via c-Conjugacy

The paper rigorously establishes that optimal dual potentials are c-conjugate in each marginal variable. This is achieved by demonstrating that any splitting family can be replaced by a c-conjugate one using the structural result: every c-splitting family can be relaxed into a c-conjugate splitting family. This structural refinement ensures that the supremum of the dual problem is attained within this restricted class, yielding Theorem 4: MOTc(µ1,..., µK) = max PK k=1 fk(xk) ≤ c(x1,..., xK) J(f) = max (f1,.,., fK)∈ QK k=1 c-conj(Xk) J(f).

Conclusion and Applications

The unified theory provides a structural foundation for further developments in probabilistic and statistical analysis of MOT, offering a transparent framework that is particularly well suited for statistical and empirical multimarginal optimal transport, including questions of differentiability and central limit theorems. The results confirm that the dual supremum coincides with the primal infimum across all settings.

The gist: This work establishes a unified Kantorovich duality theory for MOT on both compact and general Polish product spaces under bounded continuous costs, proving that optimal dual potentials can always be chosen in a c-conjugate form.

How it works

For compact spaces, the duality is derived through a convex-analytic reformulation of the dual problem and a direct application of Fenchel–Rockafellar duality, identifying the dual problem as a Fenchel-Rockafellar conjugate.
For non-compact Polish spaces, duality is recovered via a truncation-based approximation scheme that relies on the tightness of multimarginal transference plans and a procedure to control the approximation error.

Improvements for AI systems

This paper establishes a unified Kantorovich duality theory for Multimarginal Optimal Transport (MOT) on general Polish spaces, providing structural insights into optimal dual potentials via c-conjugate families.

Here are the specific improvements and capabilities this theoretical framework can enable for AI systems:

) 1. Robust Comparison and Alignment of Multi-Modal Data Distributions

The core capability derived from this work is the ability to rigorously compare and align multiple probability distributions simultaneously, which is a critical bottleneck in modern machine learning (e.g., multi-domain adaptation, multi-source generative modeling).

Enhanced Generative Modeling via Multimarginal GANs

The paper provides the theoretical foundation for Wasserstein GANs (WGANs) extended to the multimarginal setting.

Improved Stability and Training Objectives in Adversarial Networks

Since optimal dual potentials are shown to be representable by c-conjugate families, these can serve as canonical gauges or trainable objectives that are inherently stable under marginal perturbations. This allows for:

Differentiable and Robust Gradient-Based Optimization: The structural regularity of the dual potentials (c-conjugacy) provides a principled way to regularize neural network training objectives, potentially leading to more stable and generalizable models.
Theoretical Foundation for Barycentric Representation Learning

The MOT framework is naturally suited for barycenter problems. This paper allows AI systems to:

Perform Robust Multi-Distribution Averaging: Compute the optimal coupling (e.g., the barycenter) between several complex, multi-modal data sources with guaranteed convergence properties derived from duality theory.
Advanced Asymptotic Analysis and Statistical Inference

The unified duality framework is explicitly relevant for studying asymptotic properties of empirical optimal transport (empirical OT). This enables:

Quantifying Error Bounds in Statistical Inference: Developing rigorous central limit theorems and stability results for empirical estimates in high-dimensional, multi-marginal settings, crucial for robust statistical inference under marginal perturbations.
Structural Interpretability via c-Conjugate Potentials

The identification of optimal dual potentials as c-conjugate functions offers a canonical structure for learning:

Canonical Potential Representation: Instead of relying on arbitrary function classes, AI systems can be constrained to search for solutions within the geometrically meaningful class of c-conjugate potentials, potentially leading to simpler, more interpretable neural network architectures or loss functions that naturally respect the underlying transport geometry.
Theoretical Guarantees for Non-Compact Settings

The theory extends duality to general Polish spaces using a truncation-tightness procedure. This allows AI systems operating on unbounded domains (e.g., continuous signal processing, infinite state spaces) to:

Maintain Duality Equality Despite Lack of Global Compactness: Guarantee that the primal and dual objectives remain perfectly balanced even when direct compactness arguments fail, ensuring reliable optimization in challenging, real-world scenarios.

Abstract

We study Kantorovich duality for multimarginal optimal transport (MOT) with bounded continuous cost functions. The main focus is not only the equality between the primal and dual values, but also the structure of optimal dual potentials. For compact metric spaces, we prove that the dual problem admits an optimizer in the class of mutually c-conjugate families. The proof combines a Fenchel--Rockafellar duality argument with equicontinuity estimates, a normalization of the dual potentials, and the Arzelà--Ascoli theorem. We then consider the non-compact case, where the marginal spaces are Polish. We first recover the Kantorovich duality identity by a truncation and tightness argument, using the boundedness of the cost and Prokhorov compactness. For dual attainment, we rely on the geometry of optimal supports. Under a natural support-splitting condition, we obtain a bounded Borel measurable dual optimizer in the canonical class of mutually c-conjugate families. Thus, the paper identifies mutually c-conjugate potentials as natural representatives of optimal dual families in the multimarginal setting. These results provide a structural basis for further work on stability, empirical multimarginal transport, and statistical limit theory.