Riemannian ODE Framework: Theory and Applications

• Riemannian ODEs are defined on curved manifolds using intrinsic metrics and connections to capture dynamics beyond traditional Euclidean settings.
• They enable reduction of complex geometric PDEs to manageable ODEs, facilitating analysis of solitons, geodesics, and variational flows.
• This framework underpins advanced methods in optimization, stochastic processes, and deep learning by preserving manifold structure and stability.

A Riemannian ODE framework refers to any structure in which ordinary differential equations are formulated, analyzed, or solved intrinsically on a Riemannian manifold, instead of on a Euclidean space. This generalization is fundamental for geometric analysis, optimization theory, dynamical systems, geometric numerical integration, deep learning, and stochastic processes with underlying manifold constraints. Riemannian ODEs encode dynamics compatible with the manifold’s metric and connection, yielding fundamentally different (and often richer) behavior compared to classical ODEs in flat spaces.

1. Foundations: ODEs on Riemannian Manifolds

Let $(\mathcal{M},g)$ be a smooth, connected, $n$-dimensional Riemannian manifold, with metric $g$ and Levi-Civita connection $\nabla$. The natural ODE on $\mathcal{M}$ is the tangent bundle-valued equation

$\dot x = H(x),\quad x(0)=x_0\in\mathcal{M}$

where $H:\mathcal{M}\to T\mathcal{M}$ is a smooth vector field, with each $H(x)\in T_x\mathcal{M}$—the tangent space at $x$ (Shah, 2017). The solution flow $\Phi^H(t,x_0)$ yields an integral curve that lies entirely on the manifold.

The theory extends to higher-order ODEs. For instance, the geodesic equation,

$\nabla_{\dot{\gamma}} {\dot{\gamma}} = 0$

defines geodesics—curves that locally minimize path length.

Unlike in Euclidean space, intrinsic definitions of addition and scaling do not generally exist on $\mathcal{M}$. This necessitates the use of retractions, exponential/logarithmic maps, and parallel transport for the discrete-time simulation or analysis of Riemannian ODEs.

2. Riemannian ODEs in Geometry: PDE to ODE Reductions

The geometric structure of a manifold, particularly the presence of symmetries (Killing or conformal-Killing fields) and Riemannian submersions, enables drastic reductions of otherwise intractable geometric PDEs to low-dimensional ODEs. A prominent instantiation is provided by the reduction framework for mean curvature solitons on warped products with a nowhere-vanishing Killing field (Artacho et al., 22 Jan 2025). The general procedure is:

1. Warped Product Structure: If $(N,g)$ has a nowhere-vanishing Killing field $K$ with integrable orthogonal distribution, then $N \cong M \times_c I$, $g = g_M + c(x) dr^2$ with $K = \partial_r$.
2. Soliton PDE: The mean curvature soliton condition for graphs $F(x) = (x,u(x))$ on $N$ is $H = \langle K,\nu \rangle$, which yields the quasilinear PDE

$$\text{div}\!\left(\frac{\nabla u}{W}\right) =\frac{1}{W} - \frac{1}{2 W c}\langle\nabla c, \nabla u\rangle,\qquad W = \sqrt{\|\nabla u\|_{g_M}^2 + 1/c}$$

for $u:M\to I$.

1. Riemannian Submersion Ansatz: If $M$ admits a Riemannian submersion $\pi:M\to J\subset\mathbb{R}$ with $c = \hat{c}\circ\pi$ and each fiber $\pi^{-1}(s)$ of constant mean curvature $h(s)$, restricting to $u = f\circ\pi$ reduces the PDE to a scalar ODE for $f$

$$f''(s) = [\alpha(s) + \hat{c}(s) f'(s)^2] \left(1 - \frac{\hat{c}'(s) f'(s)}{2\hat{c}(s)\alpha(s)} - \frac{f'(s) h(s)}{\sqrt{\alpha(s)}} \right) + \frac{f'(s)}{2}\left(\ln \frac{\alpha(s)}{\hat{c}(s)} \right)'$$

The result is a rigorous reduction of an $n$-dimensional geometric PDE to a 1-dimensional ODE, preserving all the geometric invariance of the problem and allowing for existence, uniqueness, and qualitative classification via ODE theory.

Applications include the explicit construction of new classes of soliton "rotator" examples in hyperbolic space, characterized by a cubic ODE with data determined by hyperbolic geometry (Artacho et al., 22 Jan 2025).

3. Variational and Optimizing Flows: Accelerated Dynamics

A critical domain for Riemannian ODEs is continuous-time optimization where the objective is to minimize a geodesically convex function $f:\mathcal{M}\to\mathbb{R}$. Here, ODE-based frameworks extend Nesterov-style accelerations and allow geometric discretizations that are symmetry and structure-preserving.

Variational Riemannian ODEs

A general family of variational ODEs on $(\mathcal{M},g)$ is given by the Euler–Lagrange flow for a time-dependent Bregman Lagrangian,

$$\mathcal{L}_{\alpha,\beta,\gamma}(X,V,t) = \frac{1}{2} e^{\lambda^{-1}\zeta\gamma(t) - \alpha(t)} g_X(V,V) - e^{\alpha(t) + \beta(t) + \lambda^{-1}\zeta\gamma(t)} f(X)$$

with time-scaling functions $\alpha$, $\beta$, $\gamma$ and curvature parameter $\zeta$ (Duruisseaux et al., 2021, Duruisseaux et al., 2021).

The associated ODE,

$$\nabla_{\dot X}\dot X + (\lambda^{-1}\zeta e^{\alpha(t)} - \dot\alpha(t))\dot X + e^{2\alpha(t)+\beta(t)}\nabla f(X) = 0$$

features a curvature-modulated friction coefficient and accelerates the convergence rate of $f(X(t))$ to the minimizer $x^*$ at arbitrarily prescribed rates $\mathcal{O}(1/t^p)$ by the choice of scaling functions (Duruisseaux et al., 2021).

These frameworks admit Hamiltonian formulations, are invariant under time reparameterizations, and are compatible with manifold-preserving symplectic integrators. This theoretical foundation supports robust, structure-preserving geometric algorithms for Riemannian accelerated optimization (Duruisseaux et al., 2021, Alimisis et al., 2019).

Discrete and Symplectic Integrators

Geometric discretizations, such as discrete Euler–Lagrange equations with holonomic constraints or adaptive Hamiltonian Taylor variational integrators, preserve the symplectic form and constraint manifold. Time-adaptive schemes are possible by treating the time coordinate as an extra variable, discretizing the reparameterization invariance directly. Empirical results on spheres and Stiefel manifolds confirm enhanced stability and accelerated convergence relative to first-order methods (Duruisseaux et al., 2021).

4. Riemannian ODEs in Stochastic Processes and Learning

The Riemannian ODE framework is fundamental in analyzing stochastic approximation (SA) schemes and optimization algorithms constrained to manifolds (Shah, 2017). The canonical update is

$x_{k+1} = R_{x_k}(a_k [H(x_k) + M_{k+1}])$

where $R_x$ is a chosen retraction, $a_k$ is the step size, $H$ is a tangent vector field, and $M_{k+1}$ is martingale noise.

Under classical learning rates and regularity assumptions, the interpolated trajectory of iterates converges almost surely to invariant sets of the Riemannian ODE $\dot{x} = H(x)$. When $H = -\nabla \Psi$ for a smooth $\Psi$, convergence to the set of critical points is ensured. For embedded submanifolds, convergence to the projected ODE $\dot{x} = P_{T_x S}[H(x)]$ is guaranteed, where $P_{T_x S}$ denotes tangent bundle projection. For arbitrary closed sets, stochastic processes converge to solutions of differential inclusions involving the tangent and normal cones (Shah, 2017).

This theory directly underpins Riemannian stochastic gradient descent and manifold-constrained learning algorithms.

5. Riemannian ODEs in Applied, Data-Driven, and Graph Dynamics

Beyond geometric analysis and optimization, Riemannian ODE frameworks are core to numerical computation and modern machine learning with manifold-valued data.

Geodesics and Data-Driven Manifolds

Shortest paths on learned Riemannian manifolds are computed by solving boundary value problems for the geodesic ODE

$$\ddot c^k(t) + \Gamma^k_{ij}(c(t))\dot c^i(t)\dot c^j(t) = 0$$

with boundary conditions $c(0) = x_a$, $c(1) = x_b$. Challenges posed by ill-conditioned Jacobians in learned metrics are overcome by fixed-point, Jacobian-free solvers based on Gaussian process interpolation and Mann-iteration (Arvanitidis et al., 2019). This approach enhances computational speed, robustness, and scalability in deep generative models and metric learning.

Graph Neural ODEs on Riemannian Manifolds

Recent developments generalize continuous-time graph neural networks to non-Euclidean settings by defining graph ODEs on constant-curvature Riemannian manifolds. Examples include the Riemannian Liquid Spatio-Temporal Graph Network (RLSTG) (Lu et al., 20 Jan 2026), which models node and edge evolution via manifold ODEs, and the Pioneer framework (Sun et al., 5 Feb 2025), which enforces entropy non-decreasing Ricci-flow edge dynamics, manifold-valued node embeddings, and manifold preserving updates via gyro-transforms.

These frameworks provide theoretical guarantees on stability and expressivity (e.g., bounds on trajectory lengths), with empirical superiority over Euclidean models in structured spatio-temporal and hierarchical data domains.

6. Riemannian Geometric Encodings of ODEs

The geometric approach to scalar ODEs associates to each equation a Riemannian structure such that the original ODE's solutions correspond to geodesics or flows.

For a first-order ODE $u'(x) = \phi(x, u)$, one constructs the metric

$ds^2 = (1+\phi^2)dx^2 - 2\phi dx\,du + du^2$

so the ODE integral curves coincide with unit-speed geodesics (Pan-Collantes et al., 2023). Lie point symmetries correspond to Jacobi fields, and metric flatness is equivalent to integration by quadratures.

Autonomous second-order scalar ODEs $\ddot u = \phi(u, \dot u)$ are encoded as geodesic flows on their first jet bundle, endowed with a constructed metric dependent on $(u, \dot u)$ and $\phi$ (Pan-Collantes et al., 2024). This Riemannian encoding produces a minimal surface foliation corresponding to energy level sets, and recovers variational formulations—including for dissipative systems via a non-standard Lagrangian.

This geometric treatment unifies qualitative, variational, and analytic aspects of ODEs and enables generalization to higher-order or higher-dimensional systems.

7. Outlook and Extensions

The Riemannian ODE framework synthesizes intrinsic geometry, analysis, and computation, bridging differential geometry with applied mathematics, optimization, numerical analysis, and machine learning. Key ongoing directions include:

• Extension to non-differentiable and non-smooth (e.g., Finslerian) settings.
• Integration with differential inclusions for optimization with inequality constraints and projection operators (Shah, 2017).
• Geometric variational approaches to time-adaptive, symplectic, and constraint-preserving integration (Duruisseaux et al., 2021).
• Manifold-aware graph neural networks with physical and thermodynamic constraints (Lu et al., 20 Jan 2026, Sun et al., 5 Feb 2025).
• Data-driven adaptive geodesic solvers on high-dimensional, nonparametric manifolds (Arvanitidis et al., 2019).
• Further exploration of energy foliations, curvature–integrability relations, and geometric flows for both theoretical and practical models (Pan-Collantes et al., 2024, Pan-Collantes et al., 2023).

The Riemannian ODE paradigm thus operates as a unifying backbone for manifold-driven analysis, optimization, stochastic processes, and neural computation.