Introduction to Hyperbolic Embeddings

Published:

Why Hyperbolic Geometry?

Euclidean space has zero curvature — it is flat. But many real-world structures are intrinsically hierarchical and tree-like:

  • Taxonomies (WordNet, biological classification)
  • Social networks (power-law degree distributions)
  • Natural language syntax (parse trees)

These structures exhibit exponential growth — the number of nodes at distance \(r\) from a root grows as \(\exp(r)\). In Euclidean space, volume grows only polynomially (\(\sim r^d\)). This mismatch causes exponential distortion when embedding trees in flat space.

Hyperbolic space, with its constant negative curvature, has volume growth \(\sim \exp(r)\) — perfectly matching tree growth. Hence, trees embed with zero distortion in hyperbolic space.

The Poincaré Disk Model

The Poincaré disk \(\mathbb{D}^n\) is the unit ball \(\{x \in \mathbb{R}^n : \|x\| < 1\}\) equipped with the metric:

\[ds^2 = \frac{4 \sum_{i=1}^n dx_i^2}{(1 - \sum_{i=1}^n x_i^2)^2}\]

Key properties:

  • Geodesics are circular arcs orthogonal to the boundary
  • Distances grow exponentially as you approach the boundary
  • The boundary \(\partial \mathbb{D}^n\) represents “infinity”