Weak Optimal Transport (WOT)

• Weak Optimal Transport is a generalization of classical OT that replaces pairwise costs with nonlinear cost functionals on entire conditional distributions.
• It establishes a robust duality framework by linking a primal formulation with convex and order-restricted dual potentials under structural theorems.
• Computational methods like mirror descent, Sinkhorn schemes, and neural adversarial training enable its applications in economics, risk measures, and data science.

Weak Optimal Transport (WOT) generalizes the classical optimal transport framework by allowing nonlinear cost functionals that depend on entire conditional distributions, rather than just pairwise matches. It unifies, extends, and regularizes numerous transport and stochastic order problems found in mathematics, statistics, economics, and data science. The problem admits a robust duality theory, supports general existence and structural theorems, and underpins both theoretical and computational advances in convex geometry, risk measurement, learning, and economic modeling.

1. Static Formulation and Primal Problem

In the general setting, let $X$ and $Y$ be Polish spaces, and $\mu\in\mathcal{P}(X)$, $\nu\in\mathcal{P}(Y)$ be Borel probability measures. The set of transport plans $\Pi(\mu,\nu)$ consists of couplings $\pi\in\mathcal{P}(X\times Y)$ with marginals $\mu$, $\nu$. Any $\pi$ admits a disintegration $\pi(dx,dy) = \mu(dx)\,\pi_x(dy)$, with each $Y$0 a regular conditional probability.

A weak cost is given by a map $Y$1, jointly lower semicontinuous and convex in the measure variable. The primal weak optimal transport problem is:

$Y$2

For $Y$3, the classical Kantorovich problem is recovered as a special case.

A core instance is the barycentric weak transport: for $Y$4 and convex $Y$5, set $Y$6. This permits mass-splitting and aggregation, and replaces strict point-to-point transport by barycentric projections, as in the Brenier–Strassen and convex order–constrained transport problems (Beiglböck et al., 27 Jan 2025, Backhoff-Veraguas et al., 2020, Cazelles et al., 2021).

2. Duality and Structural Theorems

The dual problem involves maximizing over potential functions, often under monotonicity and convexity restrictions. For admissible test functions $Y$7, define the $Y$8–transform

$Y$9

The fundamental duality theorem asserts (Beiglböck et al., 27 Jan 2025):

$\mu\in\mathcal{P}(X)$0

Under additional $\mu\in\mathcal{P}(X)$1-decreasing or order-monotonicity constraints (i.e., $\mu\in\mathcal{P}(X)$2 decreases along convex order), the dual can be restricted to convex (or increasing–convex) test potentials, unifying the classical Brenier, Kantorovich–Rubinstein, Strassen, mechanism design, and martingale Benamou–Brenier dualities (Pramenković, 9 Jul 2025).

For "barycentric" WOT, the dual takes the form

$\mu\in\mathcal{P}(X)$3

where

$\mu\in\mathcal{P}(X)$4

generalizing the $\mu\in\mathcal{P}(X)$5-transform (Pramenković, 9 Jul 2025, Beiglböck et al., 27 Jan 2025).

Complementary slackness holds: the pair $\mu\in\mathcal{P}(X)$6 is optimal if and only if, $\mu\in\mathcal{P}(X)$7-a.e.,

$\mu\in\mathcal{P}(X)$8

3. Existence, Monotonicity, and Stability

The direct method of calculus of variations yields primal attainment provided $\mu\in\mathcal{P}(X)$9 has appropriate convexity, joint lower semicontinuity, and quadratic (or appropriate $\nu\in\mathcal{P}(Y)$0-power) growth (Beiglböck et al., 27 Jan 2025, Backhoff-Veraguas et al., 2020). Existence and strong duality follow for any $\nu\in\mathcal{P}(Y)$1 convex and l.s.c. in the measure variable, and lower bounded.

A key property is $\nu\in\mathcal{P}(Y)$2-monotonicity: a plan $\nu\in\mathcal{P}(Y)$3 is optimal iff it is supported on the $\nu\in\mathcal{P}(Y)$4-monotone set—generalizing $\nu\in\mathcal{P}(Y)$5-cyclical monotonicity:

A set $\nu\in\mathcal{P}(Y)$6 is $\nu\in\mathcal{P}(Y)$7-monotone if, for tuples $\nu\in\mathcal{P}(Y)$8 with $\nu\in\mathcal{P}(Y)$9,

$\Pi(\mu,\nu)$0

Stability holds: $\Pi(\mu,\nu)$1-monotonicity is closed under weak convergence of couplings and cost perturbations (Backhoff-Veraguas et al., 2019).

4. Examples and Special Cases

Representative cost functions and order constraints recover diverse classical and modern transport and stochastic order problems:

• Classical OT: $\Pi(\mu,\nu)$2.
• Barycentric WOT (Brenier–Strassen, convex-order projections): $\Pi(\mu,\nu)$3.
• Martingale/Martingale Benamou–Brenier: $\Pi(\mu,\nu)$4 infinity unless $\Pi(\mu,\nu)$5; otherwise a functional of $\Pi(\mu,\nu)$6.
• Mechanism Design: $\Pi(\mu,\nu)$7.
• Entropic/Schrödinger Problems: $\Pi(\mu,\nu)$8, or includes regularization terms (Zou et al., 16 Jan 2025).

Table: Core Weak Optimal Transport Costs

Cost $\Pi(\mu,\nu)$9	Induced Order	Applications / Theorems
$\pi\in\mathcal{P}(X\times Y)$0	None	OT, Kantorovich–Rubinstein, Brenier
$\pi\in\mathcal{P}(X\times Y)$1	Convex order	Brenier–Strassen, weak barycenters
$\pi\in\mathcal{P}(X\times Y)$2	None	Schrödinger bridge, entropy OT
$\pi\in\mathcal{P}(X\times Y)$3 if $\pi\in\mathcal{P}(X\times Y)$4, $\pi\in\mathcal{P}(X\times Y)$5 else	Convex order	Strassen's theorem
$\pi\in\mathcal{P}(X\times Y)$6	Increasing–convex order	Monopoly revenue, mechanism design

5. Dynamic Formulations and PDE Links

Dynamic equivalents of WOT are established using generalizations of the Benamou–Brenier and Fokker–Planck equations. For convex, 1-homogeneous costs $\pi\in\mathcal{P}(X\times Y)$7 on measure-valued symmetric matrix flows $\pi\in\mathcal{P}(X\times Y)$8 solving generalized Fokker–Planck evolutions, one writes (Bulanyi, 2023):

$\pi\in\mathcal{P}(X\times Y)$9

and cost

$\mu$0

Static and dynamic problems are equivalent: $\mu$1, with strong duality and uniqueness under regularity assumptions (Bulanyi, 2023).

6. Computational Methods and Algorithms

Discrete and continuous algorithms have been developed including mirror descent, Sinkhorn-type methods, and neural parameterized saddle-point approaches (Paty et al., 2022, Korotin et al., 2022).

• Mirror Descent: For entropic regularization, mirror descent in dual or primal variables with KL divergence yields $\mu$2 convergence in the duality gap (Paty et al., 2022).
• Neural Adversarial Training: Parameterizes primal maps (plans) and dual potentials with neural networks, enabling scalable computation in high dimensions and non-convex landscapes (Korotin et al., 2022).
• Sinkhorn-type Schemes: Regularized dynamic WOT admits iterative improvement analogously to classical entropic transport (Hasenbichler et al., 2 Apr 2026).

For unnormalized-kernel variants (WOTUK), mass can be split non-proportionally, supporting unbalanced scenarios and flexible economic matching frameworks (Choné et al., 2022, Paty et al., 2022).

7. Applications and Extensions

WOT underpins a wide array of modern applications:

• Barycenters and Aggregation: WOT barycenters extract shared latent structures, are robust to outliers, and admit efficient fixed-point or streaming algorithms (Cazelles et al., 2021, Chung et al., 2019).
• Risk Measures: Penalty-based risk functionals via WOT robustly account for worst-case scenarios in insurance and finance (Kupper et al., 2023).
• Mechanism Design: WOT duals deliver sharp pricing characterizations in revenue-maximizing mechanisms through convex or order-restricted potentials (Pramenković, 9 Jul 2025, Backhoff-Veraguas et al., 2020).
• Stochastic Orderings: Characterizations via kernels and convex (or positively 1-homogeneous) order functions generalize Strassen's theorem, extending to martingales and unbalanced case (Choné et al., 2022, Pramenković, 9 Jul 2025).
• Numerics and Economics: Labor market matching, production models, outlier-robust clustering, and Wasserstein barycenters in high dimensions all benefit from the WOT calculus (Paty et al., 2022, Korotin et al., 2022, Cazelles et al., 2021).

---

In summary, Weak Optimal Transport forms a fundamental bridge between classical OT and a host of sophisticated, nonlinear and stochastic optimization problems. WOT theory unifies convex geometric, informational, and probabilistic paradigms; its duality, order, and regularity structures are central to modern transport, risk, and economic modeling (Beiglböck et al., 27 Jan 2025, Bulanyi, 2023, Paty et al., 2022, Pramenković, 9 Jul 2025, Backhoff-Veraguas et al., 2020).