In machine learning, we often treat the "space" where our data lives as a flat, Euclidean void. However, the structural properties of data—such as the branching hierarchies of a family tree or the cyclical nature of a knowledge graph—often conflict with the rigid geometry of Euclidean space. This has led to a fundamental realization in recent research: there is a direct trade-off between the dimensionality of a representation and the curvature of the geometry it inhabits.
The Roominess of Curved Geometry
To understand why researchers are moving away from flat spaces, consider a tree structure. In a tree, the number of nodes grows exponentially as you move away from the root. Euclidean space, however, only grows polynomially with respect to its radius. This creates "crowding," where nodes at the bottom of a hierarchy are forced too close together, destroying the model's ability to distinguish between them.
Hyperbolic geometry offers a solution. As Ines Chami and colleagues explain in Low-Dimensional Hyperbolic Knowledge Graph Embeddings, hyperbolic space is "roomier" than Euclidean space. Its volume grows exponentially with the radius, mirroring the growth of a tree. This allows models like ATTH to pack complex hierarchies into very few dimensions—often achieving better results with 32 dimensions than a Euclidean model achieves with 512.
This efficiency is particularly vital for Knowledge Graphs (KGs). Hidetoshi Shimodaira and Yihua Zhu demonstrate in 3D Rotation and Translation for Hyperbolic Knowledge Graph Embedding that by combining hyperbolic geometry with 3D rotations (quaternions), models can simultaneously capture hierarchy, symmetry, and composition in a compact vector space.
The Convergence of High Dimensions
A recurring pattern across these papers is that the "hyperbolic advantage" is not absolute; it is a function of resource constraints. As Wei Peng notes in Hyperbolic Deep Neural Networks: A Survey, the performance gap between hyperbolic and Euclidean models narrows as dimensionality increases.
Why does this happen? In high-dimensional settings, Euclidean space acquires enough "capacity" to approximate the properties of curved manifolds. Peng Wang suggests in Hyperbolic Hierarchy-Aware Knowledge Graph Embedding for Link Prediction that as dimensions grow, the learned curvature of hyperbolic models often approaches zero, effectively reverting to Euclidean behavior. Essentially, you can either use a complex geometry in low dimensions or a simple geometry in high dimensions to achieve the same structural fidelity.
The Hub Highway Hypothesis
The most surprising evidence for this trade-off comes from the world of vector search. The Hierarchical Navigable Small World (HNSW) algorithm has long been the gold standard for finding nearest neighbors by organizing data into a multi-layered hierarchy.
However, Blaise Munyampirwa and Benjamin Coleman challenge this in Down with the Hierarchy: The 'H' in HNSW Stands for "Hubs". They propose the Hub Highway Hypothesis: in high-dimensional Euclidean spaces, certain points naturally emerge as "hubs" with high connectivity. These hubs act as a built-in routing system, making explicit hierarchical layers redundant. Their "FlatNav" approach matches HNSW's performance by simply leveraging the natural geometric properties of high-dimensional space, proving that hierarchy is often an emergent property of dimensionality itself.
Simulating Curvature in Flat Space
If high-dimensional Euclidean space can approximate curved geometry, can we get the best of both worlds? Some researchers argue we can enjoy the benefits of hierarchical awareness without the mathematical complexity of hyperbolic manifolds.
Ran Iwamoto and the team at IBM Research introduced Polar Embedding in Polar Embedding, which uses standard Euclidean polar coordinates to separate a word's "generality" (radius) from its "meaning" (angle). Similarly, Kai Wang shows in High-efficiency Euclidean-based Models for Low-dimensional Knowledge Graph Embeddings that by using "Flexible Addition" and stacked transformations, Euclidean models can match hyperbolic performance while training twice as fast.
Choosing the Right Signature
The research suggests that the choice of representation space should be driven by the "geometric signature" of the data. As Kirill Shevkunov argues in Overlapping Spaces for Compact Graph Representations, real-world data is rarely purely hierarchical or purely cyclical. It is often a mix.
The current frontier involves Product Spaces or Overlapping Spaces, where a single embedding is split across multiple geometries—part hyperbolic for hierarchies, part spherical for cycles (as explored by Xuefei Zhe in Directional Statistics-based Deep Metric Learning for Image Classification and Retrieval), and part Euclidean for everything else. This allows the model to learn the optimal geometric "signature" for the data, ensuring that the representation is as compact and accurate as possible.