Abstract
In this paper, we consider the Benamou-Brenier formulation of optimal transport for nonlinear control-affine systems on $\mathbb{R}^d$, removing the compactness assumption of the underlying manifold required in previous work. By using Bernard's Young measure-based weak formulation of optimal transport, results are established for cases not covered by previous treatments using the Monge problem. Particularly, no assumptions are made regarding the non-existence of singular minimizing controls or the Lipschitz continuity of the cost function. Consequently, the existence of solutions to the dynamical formulation is established for general Sub-Riemannian energy costs. The results also establish the controllability of the continuity equation whenever the corresponding Kantorovich problem admits a feasible solution, leveraging the equivalence between the Kantorovich and Benamou-Brenier formulations. Furthermore, when the cost function does not admit singular minimizing curves, we demonstrate that the Benamou-Brenier problem is equivalent to its convexified formulation in momentum and measure coordinates. In this regular setting, we show that the constructed transport solutions possess sufficient regularity: for the feedback control laws that achieve transport, the associated continuity equation admits a unique weak solution. These findings apply specifically to linear-quadratic costs for controllable linear time-invariant (LTI) systems and certain classes of driftless nonlinear systems.
Introduction
Optimal Transport (OT) is a fundamental mathematical framework for determining the most efficient way to transform one probability distribution into another. While the classical static formulations—the Monge and Kantorovich problems—focus on the initial and final states, many applications in robotics, aerospace, and biology require a "dynamic" perspective. This dynamic view, known as the Benamou-Brenier formulation, treats OT as a fluid-flow problem where mass evolves over time according to a continuity equation.
In control theory, this problem becomes more complex because the movement of mass is restricted by the system's dynamics. For instance, a robotic vehicle cannot move in any arbitrary direction; it must follow specific "control-affine" constraints. Previous research in this domain often relied on the assumption that the state space is a compact manifold (like a torus or sphere) and that the dynamics are sufficiently "nice" to avoid singular behaviors. This paper extends the Benamou-Brenier framework to unbounded domains ($\mathbb{R}^d$) and provides a robust mathematical foundation for systems with more general cost functions and singular dynamics.
Methodology
Young Measures and Weak Formulations
To handle the complexities of nonlinear control, the author employs "Young measures." Instead of searching for a single deterministic control law at every point in space and time, the problem is relaxed by representing the control as a probability distribution over possible actions. This approach allows the mathematical framework to "average out" behaviors that would otherwise be non-differentiable or singular, ensuring that a solution exists even when traditional methods fail.
Equivalence and Convexification
A primary contribution of this work is the proof of equivalence between the static Kantorovich cost and the dynamic Benamou-Brenier cost for systems on $\mathbb{R}^d$. This bridge allows researchers to solve static transport plans and be certain that a corresponding dynamic trajectory exists. Furthermore, for specific costs—such as the standard minimum energy $|u|^2$ cost—the author demonstrates how to transform the non-convex control problem into a convex optimization problem using momentum and measure coordinates. This transformation is vital for making these high-dimensional transport problems numerically tractable.
Hamilton-Jacobi-Bellman (HJB) Analysis
To address the uniqueness of the resulting mass flow, the paper analyzes the regularity of the value function associated with the HJB equation. Typically, ensuring a unique solution to a transport PDE requires the velocity field to be Lipschitz smooth. However, optimal controls are rarely that smooth. By leveraging the underlying structure of the control problem, the author proves that the mass still moves in a predictable, unique way, provided the system does not admit "singular minimizing curves."
Key Results
Controllability of the Continuity Equation
The research establishes a powerful link between feasibility and controllability: if a valid static transport plan exists between two distributions $\mu_0$ and $\mu_T$, then the continuity equation is controllable. This means there exists a dynamic feedback law capable of steering the entire distribution from the start to the target.
Linear Time-Invariant (LTI) Systems
The paper provides a concrete corollary for LTI systems with quadratic costs. It proves that a solution to the Benamou-Brenier problem always exists if the system satisfies the standard Kalman rank condition for controllability. This result bridges classical control theory with modern optimal transport.
Regularity of Feedback Laws
The author demonstrates that for a wide class of systems (including "2-generating" driftless systems), the constructed feedback control laws possess sufficient regularity—specifically, they are in the space of Bounded Variation ($BV_{loc}$). This level of regularity is enough to guarantee that the transport process is well-defined and that the weak solution to the continuity equation is unique.
Discussion and Future Work
By removing the compactness requirement and handling singular dynamics, this work significantly broadens the applicability of Optimal Transport to real-world control systems. However, the reliance on the absence of singular minimizing curves for uniqueness results remains a limitation for certain highly nonlinear systems.
Future research may focus on developing numerical algorithms tailored for the $\mathbb{R}^d$ formulation and investigating whether uniqueness can be maintained in the presence of singular minimizers. Additionally, applying this framework to Mean Field Games—where multiple agents optimize their paths simultaneously—presents a promising avenue for extending these results to multi-agent systems.
Summary updated