Regularity of Solutions to Beckmann's Parametric Optimal Transport
| Authors | Hanno Gottschalk & Tobias J. Riedlinger |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2603.19755 |
| Download | |
| Categories | stat.ML |
Abstract
Beckmann's problem in optimal transport minimizes the total squared flux in a continuous transport problem from a source to a target distribution. In this article, the regularity theory for solutions to Beckmann's problem in optimal transport is developed utilizing an unconstrained Lagrangian formulation and solving the variational first order optimality conditions. It turns out that the Lagrangian multiplier that enforces Beckmann's divergence constraint fulfills a Poisson equation and the flux vector field is obtained as the potential's gradient. Utilizing Schauder estimates from elliptic regularity theory, the exact Hölder regularity of the potential, the flux and the flow generating is derived on the basis of Hölder regularity of source and target densities on a bounded, regular domain. If the target distribution depends on parameters, as is the case in conditional (``promptable'') generative learning, we provide sufficient conditions for separate and joint Hölder continuity of the resulting vector field in the parameter and the data dimension. Following a recent result by Belomnestny et al., one can thus approximate such vector fields with deep ReQu neural networks in C^(k,alpha)-Hölder norm. We also show that this approach generalizes to other probability paths, like Fisher-Rao gradient flows.
Engineering Breakdown
Plain English
This paper develops mathematical regularity theory for Beckmann's optimal transport problem, which minimizes the total squared flux when moving mass from a source distribution to a target distribution. The authors solve this as an unconstrained Lagrangian optimization problem and prove that the Lagrangian multiplier enforcing the divergence constraint satisfies a Poisson equation, allowing them to recover the flux as a potential gradient. Using Schauder estimates from elliptic regularity theory, they derive exact Hölder continuity bounds for the potential, flux, and flow fields based on the regularity of input densities. The key result is a complete characterization of solution smoothness on bounded domains with regular boundaries.
Core Technical Contribution
The core novelty is establishing rigorous regularity theory for Beckmann's optimal transport by reformulating it as an unconstrained Lagrangian problem with first-order optimality conditions. Prior work on optimal transport focused on Monge-Kantorovich problems or empirical convergence; this paper provides the first systematic analysis of how smoothness propagates from input densities to output flux fields through the Poisson equation structure. The authors prove that if source and target densities have Hölder regularity α, then the flux field inherits the same regularity—a quantitative result that was previously unknown. This connects infinite-dimensional optimization theory with classical elliptic PDE regularity results in a novel way.
How It Works
The paper starts with Beckmann's problem: minimize the total squared norm of a flux vector field subject to a divergence constraint matching source and target distributions. Instead of solving this constrained problem directly, they introduce a Lagrangian multiplier λ that enforces the divergence constraint without penalty parameters. The first-order optimality conditions yield that the flux is proportional to the gradient of λ, and λ itself satisfies a Poisson equation with the density difference (target minus source) as the right-hand side. The authors then apply Schauder estimates—classical regularity results for elliptic PDEs—to the Poisson equation to control the Hölder norm of λ and hence its gradient (the flux). The final step bounds the regularity of the optimal transport flow by composing these estimates, showing that Hölder-α densities produce Hölder-α flux.
Production Impact
For production optimal transport systems (used in domain adaptation, generative modeling, and computational fluid dynamics), this work provides theoretical guarantees on solution smoothness that can guide algorithm design and discretization. Engineers building numerical solvers can use the Hölder regularity bounds to set appropriate tolerance thresholds and mesh refinement strategies—coarse meshes suffice for non-smooth densities, but finer discretization is justified when inputs are smooth. The Poisson equation formulation is computationally attractive because Poisson solvers are mature, stable, and scalable; practitioners can leverage existing fast solvers (multigrid, FFT-based) instead of developing custom transport algorithms. The regularity results validate that classical discretization schemes (finite elements, finite differences) will converge at predictable rates based on input smoothness, reducing tuning overhead. However, the theory assumes bounded regular domains and smooth densities, so it doesn't directly apply to unbounded spaces or empirical distributions with point masses.
Limitations and When Not to Use This
The theory assumes densities are smooth (Hölder continuous) on a bounded domain with regular boundary—real production data often violates this (empirical distributions are discrete, domains are unbounded, densities have singularities). The paper does not provide computational complexity analysis or convergence rates for actual solvers; knowing that a solution is Hölder-α doesn't tell you how fast an algorithm reaches it or how many iterations it needs. The Poisson equation approach works well for the squared flux formulation of Beckmann's problem but may not extend to other optimal transport distances (e.g., Wasserstein-1) or regularized variants popular in modern machine learning. The analysis is purely theoretical with no experiments validating the predicted regularity or demonstrating practical speedups on real datasets, making it unclear how tight the bounds are in practice.
Research Context
This work builds on decades of optimal transport theory starting with Monge and Kantorovich, but specifically extends recent regularity results for optimal transport maps and the Monge-Ampère equation. Beckmann's problem itself dates to the 1950s in operations research but has received renewed attention in computational optimal transport and machine learning; this paper fills a gap by rigorously analyzing the smoothness of its solutions rather than just proving existence. The regularity results complement empirical studies on convergence rates for discrete optimal transport solvers and justify the use of classical PDE numerical methods in transport-based algorithms. The paper likely influences future work on regularization schemes and adaptive discretization for optimal transport in machine learning, as well as broader connections between convex optimization theory and elliptic PDE theory.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::