Optimal transport

Optimal transport is the mathematical theory of moving mass from one distribution to another with minimal cost. Originating in 1781 with Gaspard Monge's inquiry into the most efficient way to move a pile of soil into a ditch, the theory was revolutionized in the 1940s by Leonid Kantorovich, who reframed it as a linear programming problem and thereby made it computationally tractable. What began as a problem in civil engineering has become one of the most influential mathematical frameworks of the twenty-first century, providing the geometric foundation for synthetic Ricci curvature, the rigorous basis for Wasserstein geometry, and a powerful tool in economics, machine learning, and partial differential equations.

The Monge formulation asks: given two probability measures μ and ν on a metric space X, find a transport map T: X → X that pushes μ forward to ν while minimizing the total cost ∫_X c(x, T(x)) dμ(x), where c(x, y) is the cost of moving a unit of mass from x to y. The difficulty is that such a map may not exist — if μ contains atoms and ν does not, no map can push μ to ν. Kantorovich's relaxation allows mass to be split: instead of a map, one seeks a coupling γ ∈ Γ(μ, ν) that minimizes the total cost ∫_{X×X} c(x, y) dγ(x, y). This reformulation always has a solution and connects optimal transport to linear programming, convex analysis, and duality theory.

The Kantorovich Duality and Its Consequences

The central structural result of optimal transport is the Kantorovich duality theorem. For a cost function c(x, y) = d(x, y)^p, the optimal transport cost equals the supremum:

sup_{φ, ψ} ( ∫_X φ dμ + ∫_X ψ dν )

over pairs of functions satisfying φ(x) + ψ(y) ≤ c(x, y). The maximizing potentials φ and ψ are called Kantorovich potentials, and their existence and regularity have profound geometric implications. When the cost is quadratic (p = 2) and the measures are absolutely continuous, the optimal coupling is concentrated on the graph of a gradient map — a result known as Brenier's theorem. The transport map is the gradient of a convex function, and this convexity encodes deep geometric information about the measures and the ambient space.

The duality theory connects optimal transport to the Monge-Ampère equation, a fully nonlinear partial differential equation that arises when one expresses the pushforward condition T_#μ = ν in terms of the Jacobian of the transport map. The regularity theory of the Monge-Ampère equation — developed by Caffarelli, Urbas, and others — determines when the optimal map is smooth, when it develops singularities, and how these singularities propagate. This PDE connection makes optimal transport a bridge between geometry and analysis.

Displacement Interpolation and Curvature

A revolutionary insight, due to Robert McCann, is that optimal transport induces a natural notion of interpolation between probability measures. The displacement interpolation between μ and ν is the path of measures μ_t obtained by transporting mass along geodesics in the base space. Unlike linear interpolation (1-t)μ + tν, which mixes mass at each point, displacement interpolation moves the mass itself, preserving the geometric structure of the distributions.

In the space of probability measures equipped with the Wasserstein metric, displacement interpolations are geodesics. This observation, developed by Felix Otto into the Otto calculus, reveals that the Wasserstein space possesses a formal Riemannian structure. Gradient flows in this geometry correspond to dissipative evolution equations: the heat equation is the gradient flow of entropy, the porous medium equation is the gradient flow of a different functional, and the Fokker-Planck equation is the gradient flow of the free energy.

The connection to curvature is even more profound. John Lott and Cédric Villani, and independently Karl-Theodor Sturm, proved that Ricci curvature lower bounds in a Riemannian manifold are equivalent to convexity properties of the entropy functional along Wasserstein geodesics. This equivalence permits the definition of synthetic Ricci curvature on arbitrary metric measure spaces — spaces that need not be manifolds at all. Optimal transport is not merely a technique for comparing distributions; it is the lens through which curvature itself is visible in its most general form.

Optimal transport began as a question about moving dirt. It has become the language in which geometry, probability, and physics speak to one another. The Kantorovich relaxation — allowing mass to split, to flow, to redistribute continuously — is not a computational convenience. It is a conceptual necessity that reveals the continuous structure underlying discrete allocation. The fact that this same structure detects curvature, drives PDEs, and optimizes economic matching suggests that optimal transport is not a specialized tool but a universal language — the geometry of how things move, and the physics of how systems equilibrate. Every field that has adopted it has found that its problems were already written in this language, waiting to be translated.