From Schrödinger Bridge to Optimal Transport over Sub-Riemannian Manifolds
D. O. Adu, K. Elamvazhuthi, and B. Gharesifard · 2026
Read From Schrödinger Bridge to Optimal Transport over Sub-Riemannian Manifolds by D. O. Adu, K. Elamvazhuthi, and B. Gharesifard (2026) 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-2605-11429.txt.
Introduction
Adu and two others study the cheapest way to reshape a probability distribution under motion in a horizontal bundle. This is optimal transport over sub-Riemannian spaces, driven by steering density for systems with few controls. Adding small noise along control ways gives a smoothed Schrödinger bridge problem. Under bracket-forming rules they obtain smooth nonzero transition density and a forward-backward account. This leads to a Sinkhorn-type method that regains plain optimal transport as noise fades. A worked example shows the method in action.
Catalog record
| CREATOR | D. O. Adu, K. Elamvazhuthi, and B. Gharesifard |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 70241 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.11429
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
From Schrödinger Bridge to Optimal Transport over Sub-Riemannian Manifolds D. O. Adu, K. Elamvazhuthi, and B. Gharesifard Abstract . We study the least-energy way to reshape a probability distribution when motion is constrained to a horizontal bundle, i.e., optimal transport and distribution steering in sub-Riemannian geometry—motivated by density control over underactuated systems. To obtain a continuous and numerically tractable formulation, we introduce an entropic regularization by adding small noise aligned with the control directions and study the associated Schrödinger bridge problem. The resulting reference process is a degenerate diffusion on the sub-Riemannian manifold. Under bracket-generating hypotheses we obtain smooth, strictly positive transition densities and a forward–backward characterization of the optimal bridge. This leads to a practical Sinkhorn-type algorithm for the Schrödinger potentials and, as the noise level vanishes, a recovery of the deterministic sub-Riemannian optimal transport problem. We demonstrate with a numerical example. 1. Introduction Motivated by underactuated vehicles [ 36 , 32 ] , minimum effort control over Sub-Riemannan manifolds [ 39 , 40 , 28 ] and the need to reshape an entire distribution rather than single trajectories [ 12 , 3 , 14 , 1 , 29 ] , we ask the following question: what is the minimum-effort feedback law that steers the ensemble of trajectories of a control-affine system from an initial to a prescribed terminal distribution? Usually, initial data comprises an initial and a terminal probability measure \mu_{0},\mu_{f}\in\mathcal{P}(\mathbb{R}^{d}) with densities \rho_{0},\rho_{f} on \mathbb{R}^{d} and a time-varying matrix-valued map g:[0,t_{f}]\times\mathbb{R}^{d}\rightarrow\mathbb{R}^{d\times m} . The goal is to find a control u , if it exists, that transports with minimum effort along the control-affine system (1.1) \dot{x}_{t}^{u}=g(t,x_{t}^{u})u(t,x_{t}^{u}),\qquad x_{0}^{u}\sim\mu_{0},\quad t\in[0,t_{f}]. Here, x^{u}_{t}\in\mathbb{R}^{d} denotes the state process influenced by the time-varying feedback control u:[0,t_{f}]\times\mathbb{R}^{m}\rightarrow\mathbb{R}^{d} . The notation x_{0}^{u}\sim\mu_{0} implies that at t=0 the random variable x_{0}^{u} is distribution according to \mu_{0} . We say that the system is underactuated whenever m<d . We measure effort via (1.2) \inf_{u\in\mathcal{U}}\ \mathbb{E}\left[\int_{0}^{t_{f}}\frac{1}{2}\|u(t,x_{t}^{u})\|^{2}dt\right]\quad\text{s.t.}\quad x_{t_{f}}^{u}\sim\mu_{f}, where the expectation is with respect to the law of x^{u} from x_{0}^{u}\sim\mu_{0} to x_{t_{f}}^{u}\sim\mu_{f} . Here a convenient admissible class is \mathcal{U}:=\Big\{u:\ [0,t_{f}]\times\mathbb{R}^{d}\to\mathbb{R}^{m}:\mathbb{E}\left[\int_{0}^{t_{f}}\|u(t,x_{t}^{u})\|^{2}dt\right]<\infty\Big\}. Throughout, we assume that the densities of \rho_{0},\rho_{f} are positive and compactly supported. The latter assumption can be replaced with having an exponential tail on \mathbb{R}^{d} . The structural constraint m<d implies that the admissible velocities \dot{x}_{t} in ( 1.1 ) are constrained to lie in the distribution \dot{x}_{t}\in D_{(t,x_{t})} , where (1.3) \mathcal{D}_{(t,x_{t})}:=\mathrm{span}\{g_{1}(t,x_{t}),\dots,g_{m}(t,x_{t})\}\subset T_{x_{t}}\mathbb{R}^{d}\equiv\mathbb{R}^{d}, with g_{i}(t,x_{t}) the i -th column of g(t,x_{t}) in ( 1.1 ). This distribution induces a sub-Riemannian structure with a metric g_{x}^{\text{SR}}(v,v):=v^{\top}G(x)^{\dagger}v for v\in\mathcal{D}_{(t,x)} , where G^{\dagger} is the pseudoinverse of G:=gg^{\top} . The associated sub-Riemannian distance is d_{SR}(x_{0},x_{f})=\inf\left\{\int_{0}^{t_{f}}\sqrt{g^{SR}_{x_{t}}(\dot{x}_{t},\dot{x}_{t})}dt:\dot{x}_{t}\in\mathcal{D}_{(t,x_{t})},\penalty\ \penalty\ x_{t_{f}}=x_{f}\right\} and using ( 1.2 ), we have that \frac{1}{2}d^{2}_{SR}(x_{0},x_{f})=\inf\left\{\int_{0}^{t_{f}}\frac{1}{2}\|\dot{x}_{t}\|_{G^{\dagger}}^{2}dt:\dot{x}_{t}\in\mathcal{D}_{(t,x_{t})},\penalty\ \penalty\ x_{t_{f}}=x_{f}\right\}, where \|v\|_{G^{\dagger}}^{2}=v^{\top}G^{\dagger}v . Therefore, the optimal affine control ( 1.1 )-( 1.2 ) is a dynamic ensemble control problem over a sub-Riemannian manifold (\mathbb{R}^{d},\mathcal{D},g^{SR}) , which seeks a feedback control law u(t,x) steering \mu_{0} to \mu_{f} and the dynamics of the ensemble are constrained to the distribution \mathcal{D} in ( 1.3 ). This problem has been studied in [ 23 , 24 ] , where the authors established the existence of optimal feedback control using Young measures for general nonlinear control affine-systems [ 23 ] and subsequently, as regular feedback control function under suitable absence-of-abnormality assumptions [ 24 ] leveraging results of [ 3 , 25 ] . At times, we refer to problem ( 1.1 )-( 1.2 ) as an optimal transport problem over a sub-Riemannian manifold, because it induces a static optimal transport (OT) problem over a sub-Riemannian manifold (1.4) \inf_{\gamma\in\Pi(\mu_{0},\mu_{f})}\int_{\mathbb{R}^{d}\times\mathbb{R}^{d}}\frac{1}{2}d_{SR}^{2}(x_{0},x_{f}),d\gamma(x_{0},x_{f}), where (1.5) \Pi(\mu_{0},\mu_{f}):=\{\gamma\in\mathcal{P}(\mathbb{R}^{d}\times\mathbb{R}^{d}):\pi_{0\#}\gamma=\mu_{0}\quad\text{and}\quad\pi_{f\#}\gamma=\mu_{f}\} is the set of couplings of (\mu_{0},\mu_{f}) and \pi_{0}(x_{0},x_{f})=x_{0} and \pi_{f}(x_{0},x_{f})=x_{f} , for all x_{0},x_{f}\in\mathbb{R}^{d} , are the projection maps. This static problem ( 1.4 )-( 1.5 ), which seeks the minimal cost coupling between \mu_{0} and \mu_{f} , has been studied in [ 3 , 25 ] . As stated in the literature [ 3 , 25 , 23 , 24 , 37 ] , the core difficulty of OT over a sub-Riemannian manifold stems from the non-smooth geometry induced by the constraints: the control-to-state map can have singularities, so the regularity and uniqueness of minimizers may fail; even when a Monge map exists, it can be discontinuous or non-unique. The novelty in this work is to establish a framework that mitigates these difficulties. Motivated by the regularizing effects of entropic penalization in optimal transport [ 17 ] , our contribution is of a different nature: we introduce an entropic regularization of sub-Riemannian optimal transport. Specifically, we perturb ( 1.1 ) by a small diffusion acting only along the horizontal distribution : (1.6) dX_{t}^{u}=g(t,X_{t}^{u})u(t,X_{t}^{u})dt+\sqrt{\epsilon}g(t,X_{t}^{u})dW_{t},\qquad X_{0}^{u}\sim\mu_{0}, where W_{t} is the standard \mathbb{R}^{m} -Brownian motion, with \epsilon>0 , and consider the problem (1.7) \inf_{u\in\mathcal{U}}\ \mathbb{E}\!\left[\int_{0}^{t_{f}}\frac{1}{2}\|u(t,X_{t}^{u})\|^{2}dt\right]\quad\text{subject to}\quad X_{t_{f}}^{u}\sim\mu_{f}. This stochastic control problem is equivalent to the Schrödinger bridge problem (SBP) [ 38 ] : among all path measures on \Omega:=C([0,t_{f}];\mathbb{R}^{d}) that are absolutely continuous with respect to a reference law and match the endpoint marginals (\mu_{0},\mu_{f}) , find the one that minimizes the relative entropy with respect the reference law. A difficulty immediately arises: the reference law, if it exists, is degenerate and the setup ( 1.6 )-( 1.7 ) is ill-posed. The associated Kolmogorov operator is not uniformly elliptic, and parabolic regularity theory in [ 18 , 34 ] does not apply directly. To define a well-defined Schrödinger system, one requires a strictly positive transition density function. Unlike the linear stochastic degenerate case [ 2 , 1 , 14 , 13 ] where the presence of drift is helpful, for a non-linear case, establishing the existence of a positive smooth transition density is delicate. While uniformly ellipticity ensures smooth and everywhere positive transition density function, these desirable properties are decoupled in the non-linear non-elliptic diffusion case. For smoothness, one typically imposes the Hörmander condition which leads to hypoellipticity of the generator [ 30 ] . Under Hörmander, positivity is tied to submersion property of the endpoint map [ 31 , 5 ] (Bismut/submersion condition) rather than the Chow-Rashevskii controllability condition. The driftless system is the structural choice that ensures that Chow-Rashevskii controllability is equivalent to Hörmander-Bismut/submersion condition [ 5 ] . Aside the fact that our settings smoothens the geometric singularities present in the OT-SR settings, the SBP setup is strictly convex and hence acts as a selection principle up a to subsequence among potentially many deterministic SR-optimal solutions. Moreover, the dynamic regularity formulation transforms the continuity equation associated to ( 1.1 ) to a parabolic associated to ( 1.6 ). This adds a numerical novelty in the paradigm of OT-SR setup. Aside the fact that the hypoelliptic forward-backward characterization improves numerical stability, it yields a practical algorithmic pipeline that mirrors (and extends) the classical entropic optimal transport/Sinkhorn paradigm to the sub-Riemannian regime. We show that the Schrödinger system associated to our problem can be solved via a Sinkhorn/IPFP fixed-point iteration on the cone of positive functions. Finally, the hypoelliptic degeneracy introduces a genuine discretization challenge absent in uniformly elliptic settings: standard Euclidean finite-difference discretizations may fail to preserve positivity and mass. Our schemes therefore, discretize the generator and its adjoint in a manner that is consistent with the underlying sub-Riemannian geometry. We state