RiemannGL: Riemannian Geometry Changes Graph Deep Learning

Abstract: Graphs are ubiquitous, and learning on graphs has become a cornerstone in artificial intelligence and data mining communities. Unlike pixel grids in images or sequential structures in language, graphs exhibit a typical non-Euclidean structure with complex interactions among the objects. This paper argues that Riemannian geometry provides a principled and necessary foundation for graph representation learning, and that Riemannian graph learning should be viewed as a unifying paradigm rather than a collection of isolated techniques. While recent studies have explored the integration of graph learning and Riemannian geometry, most existing approaches are limited to a narrow class of manifolds, particularly hyperbolic spaces, and often adopt extrinsic manifold formulations. We contend that the central mission of Riemannian graph learning is to endow graph neural networks with intrinsic manifold structures, which remains underexplored. To advance this perspective, we identify key conceptual and methodological gaps in existing approaches and outline a structured research agenda along three dimensions: manifold type, neural architecture, and learning paradigm. We further discuss open challenges, theoretical foundations, and promising directions that are critical for unlocking the full potential of Riemannian graph learning. This paper aims to provide a coherent viewpoint and to stimulate broader exploration of Riemannian geometry as a foundational framework for future graph learning research.

• The paper argues that Riemannian geometry should guide graph neural network design intrinsically, enabling properties such as symmetry, equivariance, conformality, causality, and quotient structure beyond Euclidean adaptations.
• The paper organizes research across eight manifold families, six neural architectures, and four learning paradigms, showing applications in recommendation, knowledge graphs, molecules, anomaly detection, and information propagation.
• The paper identifies scalability, numerical stability, fragmented benchmarks, inherited over-smoothing and over-squashing, and limited support for directed or heterogeneous graphs as key barriers to practical adoption.

Overview and Position

"RiemannGL: Riemannian Geometry Changes Graph Deep Learning" (2602.10982) is a position survey that argues Riemannian geometry should serve as a foundational framework for graph representation learning rather than a collection of isolated techniques. The authors—Sun, Wan, Zhou, Huang, and Yu—contend that the field has been dominated by hyperbolic-space methods and by extrinsic manifold formulations that merely transplant Euclidean operations onto curved spaces. Their central thesis is that the true mission of Riemannian graph learning is to endow neural networks with intrinsic manifold properties—conformality, equivariance, quotient structure, symmetry—that are difficult to obtain through careful design in flat Euclidean geometry.

The paper defines Riemannian graph learning as a family of neural networks that incorporate the Riemannian metric underlying graph structures and achieve intrinsic geometric properties of manifolds through differentiable neural operations. It explicitly distinguishes this paradigm from metric learning (which lacks geometric priors), geometric deep learning (which typically operates in extrinsic 3D Euclidean space), and hyperbolic deep learning (a strict subset of the broader Riemannian family).

A Three-Dimensional Taxonomy

The organizing contribution is a taxonomy along three axes: manifold type (eight categories), neural architecture (six families), and learning paradigm (four settings).

Dimension	Categories
Manifolds	Hyperbolic, spherical, constant curvature, product/quotient, pseudo-Riemannian, Grassmann, SPD, generic
Architectures	Convolutional networks, VAEs, transformers, neural ODEs, diffusion/SDEs, flow matching
Paradigms	Unsupervised, semi-supervised, self-supervised, foundation model/transfer

The survey is candid that coverage across these cells is highly uneven. Hyperbolic methods dominate the literature; SPD manifolds are "rarely applied in graph learning"; quotient manifolds have seen "limited application." This asymmetry is itself presented as evidence of the field's immaturity and as a research opportunity.

Manifold Families and Their Distinctive Properties

For each manifold, the paper establishes the mathematical framework and identifies what uniquely Riemannian structure it contributes:

• Hyperbolic space: exponential volume growth mirrors tree expansion, so trees embed with bounded distortion in low-dimensional hyperbolic space—a property provably absent in any Euclidean space. The Lorentz group supplies clean matrix representations of boosts and rotations for building layers.
• Hyperspherical space: its link to $SO(n)$ provides intrinsic rotational symmetry for equivariant physical modeling, and angular distance remains discriminative in extremely high dimensions where Euclidean distances collapse.
• Constant curvature space: the $\kappa$-stereographic model unifies Euclidean, hyperbolic, and spherical geometries with learnable curvature $\kappa$, using Möbius addition to keep updates geometrically valid.
• Pseudo-Riemannian manifolds: indefinite metrics introduce causal structure—directed edges can be encoded via lightcone constraints, which standard positive-definite geometry cannot express.
• Grassmann and SPD manifolds: matrix manifolds model entities as subspaces or covariance structures, yielding basis-rotation invariance and natural handling of graph Laplacians (which live on a constrained SPD manifold).
• Generic manifolds: discrete Ricci curvatures (Ollivier–Ricci, Forman–Ricci) treat the graph itself as a discretized manifold; the paper highlights results linking over-smoothing to positive curvature and over-squashing to negative curvature, motivating curvature-based graph rewriting.

A notable cross-cutting observation connects hyperbolic geometry to information geometry: the Fisher-Rao distance between univariate Gaussians is locally coupled to KL divergence ($D_{KL} \approx \tfrac{1}{2}d_{FR}^2$), enabling computationally efficient proxies for hyperbolic distances.

Neural Architectures

The architecture review traces how each deep learning primitive has been ported to manifolds, while flagging where the ports remain shallow:

• Convolutional networks: aggregation proceeds either via tangent-space projection (log/exp maps around each node) or via explicit manifold-valued weighted sums. The paper notes that most existing hyperbolic GNNs rely on tangent aggregation and inherit the over-smoothing and over-squashing pathologies of their Euclidean parents.
• VAEs: latent distributions use the Riemannian Normal (principled but computationally expensive due to intractable normalization) versus the Wrapped Normal (efficient, reparameterizable, but lacking maximum-entropy interpretation)—a clear trade-off the authors state plainly.
• Transformers: attention scores and weighted sums are ill-defined on curved spaces; solutions include $\kappa$-left-multiplication operators and canonical transformations preserving the hyperboloid constraint.
• ODEs and physics-informed dynamics: Hamiltonian formulations encode energy conservation, and constrained Ricci flow enforces entropy-increasing trajectories, connecting dynamics modeling to thermodynamic principles.
• Diffusion/SDEs and flow matching: the central bottleneck is that the heat kernel lacks closed form on general manifolds; remedies progress from geodesic random walks with Varadhan's asymptotics to variational ELBOs, and finally to simulation-free Riemannian flow matching and single-step Schrödinger bridge regression.

Learning Paradigms

Under semi-supervised learning, manifold geometry acts as an inductive-bias regularizer—for example, hyperbolic embedding naturally separates high-degree root nodes from leaves so labels propagate along the intrinsic hierarchy. For unsupervised clustering, the paper emphasizes two mechanisms: alignment of structural entropy with hyperbolic encoding trees, and differentiable Ricci flow, where node distances are trained to match the theoretical flow solution, forcing cluster separation along geodesics. Self-supervised contrastive learning exploits geometric diversity itself: different-curvature views provide augmentation-free complementary perspectives, aligned either via parallel transport to a shared tangent space at the origin or via generalized Lorentz projections mapping one manifold onto another.

On foundation models, the paper's most forward-leaning claim appears: product bundles over hyperbolic and spherical factors (with tangent-bundle attribute encodings) can capture transferable substructures—trees and cycles—as a universal vocabulary for cross-domain pre-training. The authors frame this as a route to structural transferability that text-attributed-graph and LLM-based approaches cannot achieve, since those depend on textual commonality rather than structural commonality.

Applications

Five application domains are surveyed: recommender systems (power-law interaction graphs favor hyperbolic embedding), knowledge graphs (progression from pure hyperbolic to mixed-curvature product spaces for heterogeneous hierarchies-plus-cycles), molecule generation (SE(3)-equivariance handled via torus-manifold torsion-angle flows or direct SE(3) diffusion/flow matching), anomaly detection (geometric outliers identified via curvature), and information propagation (Ricci-guided neural ODEs for time-aware diffusion prediction).

Limitations and Open Problems

The paper is explicit that Riemannian graph learning "remains in its infancy," and it concedes several substantive limitations at the points where they bear on its claims. Repeated exponential/logarithmic map evaluations introduce numerical error and computational cost that hinder scalability to large graphs. Most Riemannian models are formulation-level generalizations, so Euclidean pathologies such as over-smoothing and over-squashing are inherited rather than resolved. Parameterization of Riemannian connections and cohomology in learnable form "largely remains open." Benchmarking is fragmented: models are evaluated on inconsistent dataset choices, and no benchmark exists tailored to diverse topologies and curvatures. Coverage of directed, heterophilic, and heterogeneous graphs is thin. Reinforcement learning and other architecture families remain essentially untouched. Finally, the authors acknowledge there is "no principled approach" yet for activating Riemannian graph learning within AI4S despite the strong historical coupling between science and geometry—an admission that tempers the paper's own advocacy.

Conclusion

This survey consolidates Riemannian graph learning into a coherent three-dimensional framework spanning eight manifold families, six architectures, and four learning paradigms, supported by curated datasets, code resources, and application case studies. Its principal argument—that intrinsic manifold properties, not extrinsic reformulations, should drive architectural design—is well motivated by the demonstrated advantages of hyperbolic hierarchy encoding, pseudo-Riemannian causality, and curvature-based message passing. The most consequential open questions it poses are whether scalable, numerically stable intrinsic operations can replace repeated tangent-space projections, and whether product-bundle constructions can deliver genuinely structural transferability for graph foundation models.