Riemannian Flow Maps

• Riemannian flow maps are geometric objects that describe the evolution of data, functions, or measures on manifolds through gradient flows, optimal transport, and neural generative models.
• They play a pivotal role in geometric analysis, optimization, and generative modeling with applications from minimal surface mappings to stochastic gradient descent on curved spaces.
• Analytical foundations such as gradient and Ricci–Yamabe flows, along with mean flow approaches, provide unique solution properties under curvature and topological constraints.

A Riemannian flow map is a geometric object—either a time-indexed diffeomorphism, stochastic process, or neural transport mechanism—describing the evolution, transformation, or generative mapping of data, functions, or measures on a Riemannian manifold. It encompasses both classical continuous flows (e.g., gradient flows of geometric energies, optimal transport maps) and modern machine-learning architectures for few-step generative modeling, all underpinned by the structure of the manifold and its metric geometry.

1. Analytical Foundations: Gradient and Geometric Flows

A prototypical class of Riemannian flow maps are gradient flows associated with variational energies or geometric PDEs for maps between manifolds:

• 1-Harmonic Flow:

Let $\Omega \subset \mathbb{R}^m$ or a compact Riemannian domain, and $M \subset \mathbb{R}^N$ a smooth embedded submanifold. The total variation functional,

$TV[u] = \int_\Omega |Du(x)|\,dx,$

where $|Du|^2 = g_{u(x)}(\nabla u(x), \nabla u(x))$, yields the $L^2$-steepest descent flow:

$$\partial_t u = \pi_u\left( \operatorname{div}\left( \frac{\nabla u}{|\nabla u|} \right) \right),$$

with Neumann boundary conditions and orthogonal projection $\pi_u$ onto $T_u M$. Regular solutions require Lipschitz bounds and satisfy uniqueness, local or global existence depending on curvature, and extinction properties (the flow can become constant in finite time under suitable conditions) (Giacomelli et al., 2017, Giacomelli et al., 11 Nov 2025).

• $p$-Harmonic Map Flow:

For $p \geq 2$,

$\partial_t u = \operatorname{div}(|du|^{p-2} du),$

with critical points being $p$-harmonic maps. Regularized flows and bootstrap regularity produce unique global solutions under convexity or nonpositive curvature, culminating in convergence to stationary $p$-harmonic maps (Dawoud, 2023).

• Ricci–Yamabe Flow:

For a 1-parameter family of Riemannian metrics $g(t)$, the Ricci–Yamabe flow is

$$\partial_t g_{ij} = -2\alpha\, \operatorname{Ric}_{ij} + \beta\, R\, g_{ij},$$

interpolating between Ricci and Yamabe flows; the associated flow maps act on the space of metrics and metric-induced objects, with explicit volume and entropy variation formulae (Crasmareanu et al., 2017).

• Teichmüller–Harmonic Map Flow:

On a closed surface $\Sigma$, the coupled flow evolves $(u, g)$ by the joint gradient of the Dirichlet energy with respect to both map and metric, finding branched minimal immersions as stationary points (Rupflin et al., 2012).

Analytically, Riemannian flow maps appear as solution operators (flow maps) $\Phi_{s,t}: M \to M$ for time-dependent vector fields on $M$ or on related bundles, preserving the geometric constraint of the underlying manifold.

2. Deterministic and Stochastic Flow Maps in Optimization

In geometric optimization, flow maps arise as continuous or discrete-time transformations approximating algorithms such as stochastic gradient descent. On a Riemannian manifold $(\mathcal{M},g)$, the deterministic gradient flow,

$\frac{d}{dt}x_t = -\nabla_g f(x_t),\quad x_0 = x,$

gives rise to the flow map $\varphi_t^0(x) = x_t$, which enjoys the semigroup property $\varphi_{t+s}^0 = \varphi_t^0 \circ \varphi_s^0$. For Riemannian stochastic gradient descent (RSGD) with noisy updates and retraction maps,

$$Z_{n+1}(x) = \operatorname{Retr}_{Z_n(x)}\left( -\eta f(Z_n(x), \zeta_{n+1}) \right),$$

the law of $\{Z_n\}$ converges, as $\eta \to 0$, to that of the deterministic gradient flow map, and with higher-order approximation, to a Riemannian stochastic modified flow governed by a Stratonovich SDE. The corresponding flow maps $\varphi_t^0$ (deterministic) and $\varphi_t^\eta$ (stochastic, order-2 weak error) are constructed using the Riemannian connection, metric, and suitable retractions, with precise error estimates between discrete and continuous dynamics (Gess et al., 2024).

3. Modern Normalizing Flows and Generative Map Frameworks

Riemannian flow maps underpin contemporary geometric generative modeling on manifolds, extending normalizing flows to non-Euclidean settings:

• Riemannian Convex and Concave Potential Maps:

For a compact manifold $(M, g)$, flows are constructed as

$F(x) = \exp_x[\nabla^M \phi(x)],$

where $\phi$ is a $c$-concave potential for the quadratic cost $c(x, y) = \tfrac{1}{2}d(x, y)^2$. Universality of such maps is ensured by McCann’s theorem: any $\mu \ll \mathrm{vol}$, $\nu$ admits optimal transport realized as a flow map of this form. Discrete parameterizations and neural architectures approximate $\phi$ (Cohen et al., 2021), and extensions to implicit layers (IRCPMs) allow symmetry constraints and optimal transport-theoretic invertibility guarantees (Rezende et al., 2021).

• Multi-chart Flows:

For manifolds embedded in high-dimensional spaces with nontrivial topology, a collection of local flows $f_i: U_i \to \mathbb{R}^d$ forms a global flow map via chart-wise density modeling and partition of unity. Geodesic computations, metric corrections, and neural responsibility assignment are essential components. This framework is robust to topology and geometric structure, facilitating accurate sample generation and density estimation (Yu et al., 30 May 2025).

4. Few-Step and Mean Flow Approaches: Algorithms and Characterizations

Recent advances include explicit parameterization of the manifold flow map for efficient, few-step generative sampling:

• MeanFlow and Generalised Flow Maps:

The time-dependent vector field $v_t : \mathcal{M} \to T\mathcal{M}$ generates flow maps $\Phi_{s,t}$ by the ODE

$$\frac{d}{dt} \Phi_{s,t}(x) = v_t(\Phi_{s,t}(x)),\quad \Phi_{s,s}(x) = x,$$

together with the semigroup property $\Phi_{r,t} \circ \Phi_{s,r} = \Phi_{s,t}$ (Woo et al., 8 Feb 2026, Davis et al., 24 Oct 2025).

The average velocity is defined via the log map:

$u_{s,t}(x_s) = \frac{1}{t-s} \log_{x_s}(x_t),$

where $\log_{x_s}(x_t)$ is the initial velocity of the geodesic from $x_s$ to $x_t$. Equivalent integral representations—Eulerian, Lagrangian, semigroup—are used both for analysis and as training objectives in neural implementations. Key stabilization strategies include stop-gradient targets, adaptive loss weighting, and carefully structured time sampling (Woo et al., 8 Feb 2026).

• Training and Sampling Algorithms:

Neural nets may parameterize $u_{s,t}(x_s)$ (the average velocity) or $x_t$ directly, using projections to the tangent space or manifold exponential/logarithmic maps to respect manifold constraints. Algorithmic skeletons common to RMF/GFM/consistency model frameworks support efficient low-step inference and enable reward-guided generation via look-ahead gradients in sequential design tasks.

5. Geometric Implications, Existence, and Uniqueness

The existence, uniqueness, and qualitative behavior of Riemannian flow maps depend critically on curvature, topology, and boundary data:

• Non-positive sectional curvature ($K_M \leq 0$) implies global well-posedness, uniqueness, and contractivity for flows such as the regular $1$-harmonic and $p$-harmonic map flows, with extinction in finite time under smallness assumptions (Giacomelli et al., 2017, Giacomelli et al., 11 Nov 2025, Dawoud, 2023).
• Positive curvature can lead to finite-time blow-up for degenerate flows, bounding maximal existence time by geometric invariants.
• For total-variation flows of curves, every $BV$ initial datum leads to global strong solutions and eventual extinction to constant maps under suitable geometric constraints (Giacomelli et al., 11 Nov 2025).
• In the flow of maps to minimal surfaces with dynamic conformal structure, the coupled system avoids degeneration under topological non-compressibility, yielding existence of branched minimal immersions in each incompressible homotopy class (Rupflin et al., 2012).

6. Applications, Numerical Schemes, and Future Directions

Riemannian flow maps are fundamental in fields including:

• Geometric analysis: finite-time extinction, bubble tree decomposition, and geometric flows towards singular or minimal structures.
• Generative modeling: few-step sampling in protein design, DNA sequence generation, geospatial density estimation, and scientific ML, with state-of-the-art effective sample size and likelihood matching using one- or few-step RMF/GFM frameworks (Woo et al., 8 Feb 2026, Davis et al., 24 Oct 2025, Yu et al., 30 May 2025).
• Stochastic optimization: continuous approximations of discrete RSGD, diffusion approximations, and weak error bounds are critical for analysis and algorithm engineering in manifold-valued optimization (Gess et al., 2024).
• Numerical Riemannian geometry: efficient geodesic shooting and path optimization in multi-chart covers enable accurate computations on data manifolds with complex topology (Yu et al., 30 May 2025).

Key open problems remain in fully BV-based weak solution theories, fine singularity structure for positive curvature, algorithmic handling of anisotropic and fidelity-augmented flows, and scalable learning of geometric flows in very high dimensions (Giacomelli et al., 2017, Giacomelli et al., 11 Nov 2025).