Minimax Rates for Hyperbolic Hierarchical Learning
Divit Rawal, Sriram Vishwanath · 2026
Read Minimax Rates for Hyperbolic Hierarchical Learning by Divit Rawal, Sriram Vishwanath (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-20047.txt.
Description
Divit Rawal and Sriram Vishwanath prove that curved space learns hierarchies exponentially more efficiently than flat space. Hierarchical data, like trees of categories, are common in language and biology. Euclidean embeddings of deep trees suffer volumetric collapse: too many distant leaves are squeezed close together. Hyperbolic embeddings avoid this by expanding outward.
The authors prove an exponential separation in sample complexity between the two geometries. Their setting uses hierarchies of given depth and branching under standard Lipschitz regularity. For Euclidean models, bounded radius forces the collapse; hyperbolic models escape it. The theorem gives theory to a practice already popular in representation learning. It tells practitioners when geometry choice is not a detail but the whole game.
Catalog record
| CREATOR | Divit Rawal, Sriram Vishwanath |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-20047