Optimal Transport Duality

• Optimal Transport Duality is the fundamental equivalence between cost-minimizing primal formulations and their convex duals, providing a unified framework for analysis.
• It extends classical Kantorovich duality to multimarginal, martingale, quantum, and robust settings, influencing both theoretical insights and practical algorithms.
• The dual perspective aids in understanding optimal plans, ensuring strong duality via convex analysis tools, while enabling efficient computational methods such as entropic regularization.

Optimal transport duality is the fundamental variational equivalence between the cost-minimizing "primal" formulation of the optimal transport problem and its convex-analytic "dual." This duality provides the theoretical and computational foundation for much of modern optimal transport theory, influences the structure of optimal plans, and determines the analytic properties of associated functionals and equations. The dual perspective extends naturally to multimarginal, martingale, weak, quantum, and robust settings and admits both classical and generalized forms, unifying a range of convex-optimization and measure-theoretic methodologies.

1. Classical Kantorovich Duality

The classical Monge–Kantorovich optimal transport problem seeks, for compact metric spaces $X,Y$, probability measures $\mu$ and $\nu$, and a cost function $c:X\times Y\to\mathbb{R}$, to minimize the total cost

$$\inf_{\pi \in \Pi(\mu, \nu)} \int_{X \times Y} c(x, y)\, d\pi(x, y)$$

where $\Pi(\mu, \nu)$ is the set of all couplings with prescribed marginals. The corresponding dual, as stated in the Kantorovich Duality Theorem, is

$$\sup_{(\varphi, \psi)\in \Phi(c)} \left\{ \int_X \varphi\, d\mu + \int_Y \psi\, d\nu \right\},\quad \Phi(c) = \{ (\varphi, \psi)\in C(X)\times C(Y): \varphi(x)+\psi(y)\le c(x,y)\ \forall x,y\}$$

and strong duality holds; both infimum and supremum are attained (Gover, 23 Jan 2025).

When the cost is bounded below by a sum of integrable functions, dual attainment is guaranteed. Classical proofs use convex analytic tools such as Fenchel–Rockafellar duality (Pennanen et al., 2017), elementary Lagrange multipliers, or measure-theoretic compactness arguments (Korman et al., 2013). In the discrete case, the dual corresponds to a finite-dimensional linear program and strong duality reduces to classical LP duality (Cuturi et al., 2018).

The dual potentials ("Kantorovich potentials") not only characterize optimality but also determine the free boundary and regularity of optimal solutions. With capacity constraints, existence of dual potentials remains valid under mild hypotheses, and complementary slackness conditions at optimality take an explicit form (Korman et al., 2013).

2. Multimarginal and General Cost Duality

Modern generalizations treat multimarginal settings ($m\geq2$), nonlinear cost functionals $c: X_1\times\dots\times X_m \to \mathbb{R}$, and infinite-dimensional marginal spaces. The unified Kantorovich duality for multimarginal optimal transport on Polish product spaces with bounded continuous cost asserts (Cheryala et al., 23 Jan 2026): $$\min_{\pi\in\Pi(\mu_1, \ldots, \mu_m)} \int c\,d\pi = \max_{(f_1,\dots,f_m):\,\sum_k f_k(x_k)\le c(x_1,\dots,x_m)} \sum_{k=1}^m \int f_k\,d\mu_k$$ with dual attainment by $\mu$0-conjugate potential families, even in non-compact settings via truncation-tightness arguments.

Singular repulsive costs, such as multi-particle Coulomb costs and general decreasing kernels, admit complete duality via truncation/approximation and $\mu$1-convergence, circumventing the failure of earlier bounded-cost duality theorems (Pascale, 2015, Gerolin et al., 2018). In the vector-valued case, duality extends to measure-valued linear programming in Banach spaces, encompassing convex-dominance and Strassen-type criteria for existence (Gover, 23 Jan 2025). For applications in martingale optimal transport and capacity constraints, the dual formulation incorporates quasi-sure inequalities and generalized integrals or pathwise stochastic integrals; closedness and capacity arguments provide full strong duality (Beiglböck et al., 2015, Cheridito et al., 2019).

3. Variational Structure and Optimality Conditions

The Kantorovich duality is realized via convex conjugacy: the primal is an infimum over measures; the dual a supremum over continuous potentials. Fenchel–Rockafellar duality underpins this connection in both finite and infinite-dimensional settings (Pennanen et al., 2017). The dual constraint $\mu$2 encodes the subdifferential structure of the cost functional in measure spaces.

At optimality, complementary slackness manifests as equality on the support of optimal plans: $\mu$3 where $\mu$4. In multimarginal settings, the locus of equality defines the $\mu$5-splitting set, and optimal plans are supported on sets of $\mu$6-cyclical monotonicity (Cheryala et al., 23 Jan 2026).

For capacity-constrained transport, the pointwise surplus $\mu$7 determines whether the optimal flow saturates lower, intermediate, or upper bounds, and the dual potentials completely characterize the solution structure (Korman et al., 2013).

In dynamical problems (Benamou–Brenier formulation), the dual variables are value functions or Bellman potentials, as in discrete-time dynamic transport (Wu et al., 2024). In quantum extensions, dual potentials are self-adjoint operators, and the dual domain involves operator inequalities (Bunth et al., 30 Oct 2025).

4. Robust, Regularized, and Weak Formulations

Robust optimal transport duality modifies the classical dual by inclusion of norm-based regularizers; for example, the outlier-robust $\mu$8-Wasserstein distance dual adds a nonnegativity penalty to the potential's range (Nietert et al., 2021). Entropic regularization, vital for computational tractability, induces smooth, strictly convex duals in the form of "soft $\mu$9-transforms" and admits efficient scaling algorithms (Sinkhorn) (Cuturi et al., 2018).

Relative optimal transport duality adapts the Kantorovich–Rubinstein framework for measures defined modulo mass on a “reservoir” set, leading to function spaces of relative Lipschitz functions and reduced metrics (Bubenik et al., 2024).

Weak optimal transport, and its fundamental theorem, extends duality to problems where the cost is a convex functional of the disintegration $\nu$0, yielding dual constraints for test functions $\nu$1 via $\nu$2-transforms and characterizing optimizers through generalized complementary slackness relations (Beiglböck et al., 27 Jan 2025).

5. Geometric and Functional Analytic Interpretations

Optimal transport duality is tightly linked to the geometry of Hessian manifolds: the cost function may be identified with a Legendre pairing, and duality is expressed in terms of convex potentials and Monge–Ampère equations (Hultgren, 2023). The optimal map arises as a gradient of a convex function ($\nu$3), and the solution to the Monge–Ampère PDE reflects the push-forward equation for measures.

This geometric understanding underpins applications in toric geometry and mirror symmetry, for example in the SYZ conjecture, where the transport problem encodes the structure of special Lagrangian torus fibrations and real Monge–Ampère metrics.

Functional analytic frameworks generalize duality beyond measures, embracing convex cones, abstract linear functionals, and sublinear support functionals to encompass linear programming, moment problems, and zero-sum games as special cases (Gover, 23 Jan 2025, Pennanen et al., 2017).

6. Extensions: Martingale, Multimarginal, Quantum, and Beyond

Martingale optimal transport duality incorporates convex-order constraints and requires quasi-sure relaxation of dual inequalities and generalized integrals indexed by concave moderators to avoid false polar sets. Complete duality, existence, and the cyclical monotonicity principle are established for general measurable costs, relying on decomposition into irreducible convex-order components and capacity arguments (Beiglböck et al., 2015).

In the quantum setting, primal variables are quantum states, dual variables are operator-valued, and operator inequalities replace function-level constraints. Strong Kantorovich duality holds for linearized quantum OT problems, and analytic structure is preserved even under noncommutative generalizations (Bunth et al., 30 Oct 2025).

Multimarginal and non-scalar-marginal extensions leverage $\nu$4-conjugacy and truncation-tightness, ensuring that the structural features of the two-marginal dual carry over, including $\nu$5-cyclical monotonicity, envelope formulas, and dual attainment (Cheryala et al., 23 Jan 2026, Pascale, 2015).

7. Algorithmic and Applied Aspects

Optimal transport duality underlies computational methods for Wasserstein distances, barycenters, and density fitting. Dual-based methods, including stochastic gradient and semi-dual formulations, exploit smooth convexity and strong duality for fast convergence (Cuturi et al., 2018). Entropic and robust formulations lend themselves to efficient scaling and robust estimation, and dual modifications are implemented with minimal code changes (Nietert et al., 2021).

Dynamical and adversarial variants, such as discrete-time dynamical OT and revenue-maximization in auction theory, utilize duality structure to reduce PDE-based computations to scalable convex optimization or saddle-point problems (Wu et al., 2024, Gonczarowski, 2017).

The mathematical structure of duality continues to inform fundamental advances in convex analysis, geometry, quantum information, and computational statistics, rendering it intrinsic to contemporary optimal transport theory.