Hyperbolic Embeddings

• Hyperbolic embeddings are techniques that map structured, hierarchical, or scale-free data into spaces of constant negative curvature, minimizing distortion.
• They employ models like the Poincaré ball and Lorentz model with Riemannian optimization to efficiently learn low-dimensional representations.
• Applications span NLP, graph analysis, and multilayer networks, delivering high performance in tasks such as link prediction and classification.

Hyperbolic embeddings are representation learning methods that map data—typically from structured or hierarchical domains—into spaces of constant negative curvature, such as the Poincaré ball or Lorentz (hyperboloid) model. These geometries enable the faithful encoding of trees, taxonomies, and scale-free graphs with minimal distortion in low-dimensional spaces by leveraging the exponential growth of volume with respect to radius, a property fundamentally distinct from polynomial volume growth in Euclidean space. Hyperbolic geometry provides a natural coordinate system for data with hierarchical or power-law structure, such as language ontologies, knowledge graphs, biological taxonomies, and complex networked systems.

1. Foundations of Hyperbolic Geometry and Embedding Models

Hyperbolic space is characterized by constant negative curvature, enabling exponential scaling of metric volume. The two most widely used models are:

• Poincaré Ball Model: The $d$‐dimensional Poincaré ball is $$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$, with metric tensor $g_x = (2/(1 - \|x\|^2))^2 I_d$. The geodesic distance between $u, v \in \mathbb{B}^d$ is

$$d_\mathbb{B}(u, v) = \operatorname{arcosh}\left(1 + 2\frac{\|u - v\|^2}{(1 - \|u\|^2)(1 - \|v\|^2)}\right)$$

• Lorentz (Hyperboloid) Model: Hyperbolic space $\mathbb{H}^d$ is embedded as the upper sheet

$$\mathcal{L}^d = \{ x \in \mathbb{R}^{d+1} : [x, x] = -1,~ x_0 > 0 \}$$

under the Minkowski inner product $[x, y] = -x_0 y_0 + \sum_{i=1}^d x_i y_i$, with distance

$d_\mathbb{H}(x, y) = \arccosh(-[x, y])$

Mappings between these models (e.g., stereographic projection) enable interchangeable use in optimization and analysis (Tabaghi et al., 2020, Casulo et al., 30 Apr 2026).

Hyperbolic geometry allows hierarchical or scale-free structures to be embedded with low distortion in a fixed low-dimensional space, while Euclidean embeddings require dimensionality to grow polynomially with the hierarchy's depth to avoid distortion (Song et al., 2019, Sa et al., 2018, Casulo et al., 30 Apr 2026).

2. Algorithmic Frameworks and Riemannian Optimization

Designing hyperbolic embeddings involves Riemannian optimization due to the curvature and non-Euclidean metric:

• Gradient Mapping: The Riemannian gradient is derived from the Euclidean gradient by rescaling: $\nabla^R f(u) = ((1-\|u\|^2)/2)^2 \nabla^E f(u)$ in the Poincaré ball (Song et al., 2019).
• Retraction: Updates move points along the geodesic via the exponential map or by a simplified addition plus projection, ensuring $$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$0 (Song et al., 2019, Casulo et al., 30 Apr 2026).
• Loss Functions: Typical objectives include negative sampling for link prediction, hinge loss for structural balance in signed networks, or margin-based/ranking losses.

In the Lorentz model, updates are projected to the tangent space via the Minkowski inner product, and points are retracted via $$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$1 (Casulo et al., 30 Apr 2026, Jawanpuria et al., 2019).

Frameworks such as HypeGRL unify a range of embedding methods (Hydra, (D-)Mercator, Poincaré Maps) under a Riemannian optimization module, exposing consistent APIs for fitting, transforming, and evaluating on downstream tasks (Casulo et al., 30 Apr 2026).

3. Embedding Hierarchical Structures: Trees, Graphs, and Networks

Hyperbolic embeddings are well-suited for trees, scale-free, and hierarchical networks due to their negative curvature:

• Deterministic Embedding: Trees can be embedded in $$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$2 with arbitrarily low distortion and mean average precision (MAP) close to 1 in $$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$3 dimensions using combinatorial constructions (Sa et al., 2018).
• Multidimensional Scaling (h-MDS): Given pairwise hyperbolic distances, classical MDS generalizes to the hyperboloid via eigendecomposition of the hyperbolic Gram matrix, yielding exact recovery up to isometry, with quantitative perturbation and dimensionality reduction analysis (Tabaghi et al., 2020, Sa et al., 2018).
• Low-Rank Factorization: Hyperbolic embeddings can be approximated in a low-rank subspace, scaling to large graphs or wordnets while retaining most of the embedding fidelity—enabling compression and tractable computation (Jawanpuria et al., 2019).

Empirical evaluations confirm that hyperbolic embeddings in 2–4 dimensions match or outperform high-dimensional Euclidean baselines for tasks such as link prediction and node classification in complex, hierarchical graphs (Casulo et al., 30 Apr 2026, Sa et al., 2018, Song et al., 2019).

4. Embedding Learning for Signed and Multilayer Networks

Structured networks with semantics beyond unsigned edges, such as signed or multilayer networks, benefit from geometric inductive biases of hyperbolic space.

Signed Networks:

• Hyperbolic signed network embedding (HSNE) is based on the Poincaré ball and structural balance theory. Margin-based hinge losses are enforced for node triplets to encode that friends are closer than enemies. Riemannian SGD on the Poincaré ball yields embeddings that preserve both signed connectivity and latent network hierarchy (Song et al., 2019).
• On a suite of signed networks (Wiki-editor, Epinions, Slashdot, CoW), HSNE outperforms or matches Euclidean methods in Macro/Micro-F1 and AUC. Hyperbolic radius encodes hierarchy and hostilities, with “neutral” nodes near the center, “aggressive” nodes near the boundary, reflecting latent social/geopolitical structure (Song et al., 2019).

Multilayer Networks:

• Jointly embedding multilayer networks in a global Poincaré disk (or ball) by constructing a block connectivity matrix capturing intra- and inter-layer edges preserves both community and global coupling structure. Parameters controlling inter-layer coupling (μ) and radial assignment (β) enable flexible adjustment between community separation and interlayer alignment (Guillemaud et al., 26 May 2025).
• On synthetic and real-world multilayer data (e.g., brain connectomics), simultaneous multilayer hyperbolic embedding yields superior community recovery, alignment metrics, and disease marker clustering compared to independent-layer approaches (Guillemaud et al., 26 May 2025).

5. Hyperbolic Embeddings in Natural Language Processing

Hyperbolic geometry has become prominent in NLP for encoding lexical and semantic hierarchies:

• Word and Concept Embeddings: Poincaré GloVe and Skip-gram variants replace Euclidean similarity with hyperbolic distance or merged Riemannian distance functions. Embeddings are optimized using Riemannian SGD or Adagrad to capture hierarchical or taxonomic relationships (Tifrea et al., 2018, Leimeister et al., 2018, Saxena et al., 2022).
• Analogy and Hypernymy Tasks: Parallel transport and Möbius gyrovector arithmetic generalize vector arithmetic to hyperbolic settings for analogy-solving (Tifrea et al., 2018).
• Centroid and Aggregation Methods: Standard vector averaging is not meaningful; instead, document or phrase embeddings are formed as Fréchet means (or efficient approximations) of sets of hyperbolic points, using Möbius addition and gyrovector calculus (Gerek et al., 2022).
• Empirical Performance: Hyperbolic embeddings attain or exceed SOTA in unsupervised hypernymy detection, semantic similarity, and cross-lingual analogy tasks, maintaining greater efficiency and hierarchical fidelity at low dimensionality (Tifrea et al., 2018, Saxena et al., 2022, Gerek et al., 2022).
• Text Classification: For classification tasks (e.g., 20 Newsgroups, Turkish corpora), k-NN using Poincaré distances on hyperbolic centroids matches or outperforms Euclidean baselines, especially in morphologically rich languages (Gerek et al., 2022).

6. Practical and Theoretical Trade-offs in Hyperbolic Embedding

Key trade-offs inherent in hyperbolic embedding include:

• Dimension vs. Precision: Theory establishes that high-quality hierarchical (tree) embeddings can be achieved in low dimension ($$\mathbb{B}^d = \{ x \in \mathbb{R}^d : \|x\| < 1 \}$$4), but require increased numerical precision as the hierarchy deepens ( coordinates approach the ball boundary, increasing required bits per dimension) (Sa et al., 2018).
• Optimization Complexity: Modern Riemannian optimization frameworks mitigate the difficulty of learning in non-Euclidean space. Efficient initialization schemes, such as spectral methods (Hydra) followed by Riemannian refinement, accelerate convergence (Casulo et al., 30 Apr 2026).
• Low-Rank Approximations: Memory and computational cost can be reduced by constraining hyperbolic embeddings to lie in low-rank product submanifolds, at negligible loss of global embedding quality except under aggressive compression (Jawanpuria et al., 2019).
• Model Compatibility: By mapping hyperbolic embeddings via random Laplacian feature maps into Euclidean space, isometry-invariant kernels enable seamless integration with standard Euclidean GNNs and classifiers at improved efficiency and stability (Yu et al., 2022).

7. Applications, Limitations, and Future Directions

Applications:

Hyperbolic embeddings power a spectrum of downstream tasks:

• Link and sign prediction, node classification, and community detection in graphs with hierarchical or scale-free structure (Casulo et al., 30 Apr 2026, Song et al., 2019, Guillemaud et al., 26 May 2025).
• Hierarchical out-of-distribution (OOD) detection and prototype-based classification in deep learning, where hyperbolic prototypes yield improved OOD discrimination and hierarchical awareness (Kasarla et al., 11 Jun 2025).
• Unsupervised and cross-lingual semantic representation in NLP, where hierarchical linguistic phenomena are naturally encoded (Tifrea et al., 2018, Saxena et al., 2022).

Limitations and Open Problems:

• Availability of explicit hierarchy or tree structure is assumed in many algorithms; unsupervised hierarchy inference remains an open challenge (Kasarla et al., 11 Jun 2025).
• Flat or highly imbalanced hierarchies may reduce the geometric benefit, and extreme precision demands near the boundary complicate scaling for very deep hierarchies (Sa et al., 2018).
• The optimal choice of curvature, dimension, and aggregation strategies is data-dependent and remains an active research area (Kasarla et al., 11 Jun 2025, Gerek et al., 2022).

Research Directions:

• Integration of curvature learning and automatic dimensionality selection for adaptive embeddings.
• Unified frameworks (e.g., HypeGRL) for reproducible benchmarking and deployment across methods (Casulo et al., 30 Apr 2026).
• Extension to streaming, multilayer, and dynamic networks, online inference, and hybrid hyperbolic-Euclidean models for heterogeneous data (Guillemaud et al., 26 May 2025, Yu et al., 2022).

Hyperbolic embeddings thus constitute a mature, theoretically founded and practically effective framework for geometric representation learning, with ongoing advances improving scalability, interpretability, and domain applicability.