Benamou–Brenier and Kantorovich are equivalent on sub-Riemannian manifolds with no abnormal geodesics
Giovanna Citti, Mattia Galeotti, Andrea Pinamonti · 2025
Read Benamou–Brenier and Kantorovich are equivalent on sub-Riemannian manifolds with no abnormal geodesics by Giovanna Citti, Mattia Galeotti, Andrea Pinamonti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-20959.txt.
Introduction
Giovanna Citti, Mattia Galeotti, and Andrea Pinamonti prove two classic views of optimal transport agree in a curved, constrained geometry in a 2025 mathematics study. The Kantorovich view plans transport as a coupling between distributions, while the Benamou-Brenier view watches mass flow over time. On sub-Riemannian manifolds, where motion is allowed only in certain directions, their equivalence was not known in general.
With no boundary, completeness, and no abnormal shortest paths, the authors prove the two problems match for measures with finite second moments. They also prove a flowing minimizer exists and ties to the optimal plan.
Catalog record
| CREATOR | Giovanna Citti, Mattia Galeotti, Andrea Pinamonti |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-20959
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Benamou–Brenier and Kantorovich are equivalent on sub-Riemannian manifolds with no abnormal geodesics Giovanna Citti, Mattia Galeotti, Andrea Pinamonti Abstract . We prove that the Benamou–Brenier formulation of the Optimal Transport problem and the Kantorovich formulation are equivalent on a sub-Riemannian connected and complete manifold M M without boundary and with no abnormal geodesics, when the problems are considered between two measures of compact supports. Furthermore, we prove the existence of a minimizer for the Benamou–Brenier formulation and link it to the optimal transport plan. 2020 Mathematics Subject Classification 49Q22, 53C17 1. intro The Benamou–Brenier “dynamical” formulation of Optimal Transport is the minimization problem for the functional (1.1) J_{\BB}(\mu_{t},v_{t})=\int_{0}^{1}||v_{t}(x)||^{2}\mathrm{d}\mu_{t}(x)\mathrm{d}t where v_{t} is a Borel family of vector fields and \mu_{t} a narrowly continuous family of probabilities starting at a given \mu_{0} and finishing at a given \mu_{1} . Both \mu_{t} and v_{t} are defined on a differentiable manifold M and constrained by the Continuity Equation (1.2) \dot{\mu}_{t}+\Div(\mu_{t}v_{t})=0. For a focus on the Continuity Equation see [ 5 , Chapter 8] and [ 3 ] , we point out that this equation links the path \mu_{t} to the flow of the ODE associated to v_{t} , expressing the conservation of mass for the action ( 1.1 ). All this gives a time dependent approach to the transport problem. By contrast, in the classical Kantorovich formulation the dynamic part is encoded in the squared distance function, d^{2}(x,y):=\inf\left\{\left.\int_{0}^{1}||\dot{\omega}_{t}||^{2}\mathrm{d}t\right|\,\omega\colon[0,1]\to M\ \mbox{absolutely continuous,}\,\omega_{0}=x,\,\omega_{1}=y\right\}. This is called “cost” function, and the minimization problem regards the coupling of points. In particular, in this “static” formulation we want to minimize (1.3) J_{\Kan}(\gamma)=\int_{M\times M}d^{2}(x,y)\mathrm{d}\gamma(x,y) where \gamma is taken among admissible plans, i.e. probabilities on M\times M with marginals equal to \mu_{0} and \mu_{1} respectively on the two coordinate projections. The Kantorovich and Benamou–Brenier formulations have been proven equivalent on bounded Euclidian domains [ 8 ] and more generally on complete Riemannian manifolds without boundary, see for instance [ 4 ] . In this work we are going to prove their equivalence on a sub-Riemannian manifold M under some assumptions that we state precisely below. The transport problem was introduced by Monge at the end of the 18th century, it consists in the the minimization problem for the functional J_{\Mon}(T)=\int_{M}c(x,T(x))\mathrm{d}x for some cost functon c\colon M\times M\to\mathbb{R}_{\geq 0}\cup\{+\infty\} with T\colon M\to M a function constrained by the transport condition T_{\#}\mu_{0}=\mu_{1} . However, this formulation is ill-suited for a general resolution, because the transport condition lacks closedness with respect to the principal weak topologies. The Kantorovich formulation overcomes this obstructions. In the Kantorovich setting the existence of a minimizer, an optimal plan, is established for any Polish space (M,d) if the cost function is lower semi-continuous and bounded from below, see [ 4 , Theorem 1.5] , but also [ 25 , Theorem 5.10] for a wider treatment of the problem. Furthermore, in the Riemannian setting, when the cost is the square distance, we also have the existence and uniqueness of an optimal map i.e. a map inducing an optimal plan. This was proved, under opportune assumptions, by Brenier in the Euclidian case [ 11 ] and by McCann in the general Riemannian case [ 20 ] . The first decade of the 20th century saw a renewd effort for Optimal Transport in the setting of sub-Riemannian manifolds, that is on differentiable manifolds admitting a totally non-integrable distribution HM\subset TM and a positive definite quadratic form g_{SR} on HM . For an introduction see [ 22 ] . In particular, some existence, uniqueness and regularity results for the optimal map were proved for Heiseneberg groups [ 6 ] , 2 -generating distributions [ 2 ] and also for more general cases [ 16 , 17 ] , but the problem in full generality is still open. While in the Riemannian context there exists only one kind of geodesics, the so called normal geodesics, on sub-Riemannian manifolds a second kind, the abormal ones, may exist. Because of the existence of abnormal geodesics, the sub-Riemannian distance fails to be local convex, and this is the main obstacle to the existence of the optimal map. To treat the dynamical problem in sub-Riemannian context, we consider a relaxed version of the problem. In particular, we minimize the integral of ||v||^{2} along a probability \eta on [0,1]\times HM , instead of considering just one vector field v_{t} for every time t , integrated with respect to \mu_{t} . This approach was already used in [ 15 ] and [ 14 ] in the case M=\mathbb{R}^{n} endowed with a sub-Riemannian structure, and fits into the setting of control relaxation techniques. For some examples see [ 13 , 12 , 19 ] and more recently [ 21 , 23 ] . For a proper formulation of the relaxed problem, we consider the class of Young measures. For any Polish space (X,d) , a Young measure \eta is a (positive) measure defined on [0,1]\times X with the property that the first coordinate projection \pi^{(t)}_{\#}\eta is Lebesgue measurable. Therefore, a disintegration \eta_{t} exists with respect to the Lebesgue measure on [0,1] . If we denote by \pi\colon TM\to M the canonical projection from the tangent bundle, then by pushforwarding a Young probability on [0,1]\times HM we obtain a narrowly continuous path of probabilities \mu_{t}=\pi_{\#}\eta_{t}\in\mathcal{P}(M) . The Continuity Equation can be adapated to this framework, see Equation ( 3.2 ), and used as the constraint for the minimization of the relaxed functional (1.4) J_{\BB}^{\star}(\eta)=\int_{[0,1]\times HM}||v||^{2}\mathrm{d}\eta(t,v), where we also ask that \pi_{\#}\eta_{0}=\mu_{0} and \pi_{\#}\eta_{1}=\mu_{1} . We first prove the equivalence of the original (sub-Riemannian) Benamou–Brenier and the relaxed version. Then, by proving the compactness of the J^{\star}_{BB} sub-levels in Lemma A.10 , we obtain the existence of a minimizer. For the last part, we prove the equivalence with the Kantorovich formulation. To do this, we use a superposition principle that allows to link the value of J_{\Kan} on admissible plans to the value of J_{\BB}^{\star} on some admissible Young measures. We resume. Theorem 1.1 . Let M be a sub-Riemannian, connected and complete manifold without boundary and with no abnormal geodesics. Let \mu_{0},\mu_{1} be two probabilities in \mathcal{P}_{2}(M) with compact supports. Then, the infima of the Kantorovich, Benamou–Brenier and relaxed Benamou–Brenier transport problems are realized and equal. Moreover, if \gamma is a minimizer for Kantorovich, then a minimizer for Benamou–Brenier is obtained by pushforwarding \mathcal{L}\otimes\gamma along the F map defined at ( 3.10 ). Viceversa, if (\mu_{t},v_{t}) is a minimizer for Benamou–Brenier, then a minimizer for Kantorovich is obtained by pushforwarding an opportune measure associated to \mu_{t} along the E map defined at ( 3.14 ). With respect to [ 14 ] , our treatment of the problem is developed outside the control theory setting. Furthermore, with respect to the case with underlying manifold equipped with an Euclidean structure, we solve two subtle additional issues. First, we lack the well behaved Euclidean distance, and therefore we prove the compactness of the J^{\star}_{\BB} sub-levels by using the intrinsic sub-Riemannian distance of the manifold. Second, the use of the superposition principle relies on the construction of a measurable map S that for any pair (x,y)\in M\times M gives a geodesics from x to y . The construction of S poses non-trivial technical problems that we