Sparse optimal control in the Wasserstein space

Enrico Sartor, Jeremy Coulson, Nicolas LanzettiarXiv 2026 · DOI · arXiv

Abstract

We study sparse optimal control of a non-local continuity equation, where the goal is to steer a distribution via finitely many controllable agents or actuators. This model arises naturally in mean-field multi-agent systems and takes the form of a coupled PDE-ODE system where the PDE describes the evolution of the distribution and the controlled ODE captures the dynamics of the controllable agents. A natural objective is distribution steering via terminal costs based on optimal transport, such as the squared Wasserstein distance. These costs are problematic for finite-agent formulations due to non-smoothness at empirical measures and they fall outside common expected-value-type cost classes. We address these challenges by studying the resulting optimal control problem in the Wasserstein space. Under suitable assumptions on the system dynamics and Wasserstein differentiability of the terminal cost, we prove first-order sensitivity of the control-to-state map, derive an adjoint system and an explicit formula for the gradient of the cost functional, and obtain Pontryagin-type necessary conditions. To illustrate the resulting adjoint-based method, we present numerical experiments on a representative distribution-splitting task.

Introduction

Coordinating large-scale systems, such as drone swarms or pedestrian crowds, often requires "sparse" control—influencing a large population through a small number of dedicated actuators or "leaders." This scenario is mathematically modeled as a mean-field system where a Partial Differential Equation (PDE) governs the density of the crowd, while an Ordinary Differential Equation (ODE) tracks the trajectories of the controllers. A primary challenge in this domain is steering the crowd toward a target distribution using metrics like the Wasserstein distance, which measures the effort required to transport one distribution to another.

The Challenge of Non-Smoothness

While the Wasserstein distance is a natural choice for distribution steering, it presents significant optimization hurdles. In discrete agent models, Wasserstein costs are often non-smooth and "pointy," causing standard gradient-based methods to fail. This paper addresses this by performing the analysis at the continuum level—within the Wasserstein space $P_2(\mathbb{R}^d)$. By lifting the problem from discrete particles to continuous measures, the authors recover the smoothness necessary for classical control theory.

Methodology

The system is modeled as a coupled PDE-ODE system. The PDE is a non-local continuity equation representing the crowd, and the ODE represents the actuated agents. The authors' approach involves:

  • Variational Analysis in $P_2(\mathbb{R}^d)$: Sensitivity analysis is performed directly on the space of probability measures.
  • Linearization via Pull-backs: To handle the evolving tangent spaces in Wasserstein geometry, the system's flow is used to pull vectors back to a common reference measure.
  • Adjoint-Based Gradient: A "backward adjoint equation" is derived, providing an explicit formula for the gradient of the cost functional. This enables the use of standard gradient descent for optimization.
  • Optimize-then-Discretize: The continuous equations are solved using a forward Euler scheme in time and a conservative finite-volume discretization in space.

Numerical Experiments: Distribution Splitting

To validate the framework, the authors introduce a "distribution-splitting" benchmark. The task is to take a single Gaussian-like distribution at the origin and use controlled leaders to split it into two equal-mass clusters at specific target coordinates.

Figure 1: Illustration of the distribution-splitting task. A single initial distribution $\mu(0)$ is steered by controllers $y_1, y_2$ toward two distinct targets, resulting in a split terminal distribution $\mu(T)$.

Key Results and Insights

The theoretical framework successfully yields a Pontryagin Maximum Principle (PMP) for sparse control under minimal regularity assumptions. Key findings include:

  • Continuum Smoothness: A major insight is that while Wasserstein distances are non-smooth for discrete points, they are often differentiable for continuous distributions. Lifting the problem to the Wasserstein space recovers the "smoothness" needed for optimization.
  • Implicit Assignment: The Wasserstein gradient naturally handles the partitioning of the crowd. The geometry of the optimal transport map automatically determines which part of the distribution is pushed toward which target.
  • Particle Consistency: In 2D experiments, the "fluid-like" mean-field controls were tested on a discrete system of 500 agents, yielding a terminal cost (0.050) very close to the ideal mean-field cost (0.023). This demonstrates that the continuum approximation is highly effective for large but finite groups.
  • Efficiency: The adjoint-based method proved computationally efficient, with iterations taking less than 60 seconds on standard consumer hardware.

Limitations and Future Work

The current numerical implementation faces the "curse of dimensionality" due to its reliance on grid-based discretization, making it less suitable for high-dimensional state spaces. Additionally, the performance guarantees are best suited for short-to-medium steering horizons. Future work will focus on overcoming dimensionality constraints and establishing formal convergence proofs for the proposed algorithm.

Summary updated

Abstract

We study sparse optimal control of a non-local continuity equation, where the goal is to steer a distribution via finitely many controllable agents or actuators. This model arises naturally in mean-field multi-agent systems and takes the form of a coupled PDE-ODE system where the PDE describes the evolution of the distribution and the controlled ODE captures the dynamics of the controllable agents. A natural objective is distribution steering via terminal costs based on optimal transport, such as the squared Wasserstein distance. These costs are problematic for finite-agent formulations due to non-smoothness at empirical measures and they fall outside common expected-value-type cost classes. We address these challenges by studying the resulting optimal control problem in the Wasserstein space. Under suitable assumptions on the system dynamics and Wasserstein differentiability of the terminal cost (with no smoothness requirement on the associated Wasserstein gradient), we prove first-order sensitivity of the control-to-state map, derive an adjoint system and an explicit formula for the gradient of the cost functional, and obtain Pontryagin-type necessary conditions. To illustrate the resulting adjoint-based method, we present numerical experiments on a representative distribution-splitting task.
BibTeX
@misc{sartor2026sparse,
  author = {Enrico Sartor and Jeremy Coulson and Nicolas Lanzetti},
  title = {{Sparse optimal control in the Wasserstein space}},
  year = {2026},
  eprint = {2603.00373},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.00373},
}

Loading related content...

Figure