Hyperbolic Geometry

Hyperbolic space expands exponentially with radius, so trees and hierarchies embed with low distortion where Euclidean layouts run out of room.

Equal hyperbolic steps shrink toward the boundary as room grows exponentially

Why Does This Exist?

Trees grow exponentially: a binary tree of depth 10 has 1024 leaves. Euclidean balls grow polynomially ( $r^d$ ), so shallow embeddings crush deep levels together. Hyperbolic balls grow like $e^{(d-1)r}$ , matching tree branching and giving every subtree its own room.

Word hierarchies, taxonomies, and graphs embed with far less distortion hyperbolically. Parent: curvature. Optimization home: manifold optimization.

The one idea you need first

The Poincaré ball models hyperbolic $d$ -space as the open unit ball with metric scaled by $4/(1-\|x\|^2)^2$ . Center behaves almost Euclidean; near the boundary tiny coordinate moves cover huge distances.

Think of It Like This

A coral reef with endless branches

A coral branches outward with room for every polyp: each fork gets full space because volume explodes with radius. Flatten the coral onto paper and branches pile atop each other.

Hyperbolic embeddings keep the coral 3D. Roots sit centrally, leaves spread toward the boundary with space to spare, and hierarchy depth becomes radius.

Where the analogy stops: numerical precision thins near the boundary, so very deep trees need float64 or tiling tricks.

How It Actually Works

Poincaré distance: $d(x,y) = \text{arcosh}(1 + 2\|x-y\|^2/((1-\|x\|^2)(1-\|y\|^2)))$ . Möbius addition replaces vector sums for gyrovector calculus. Hyperbolic neural layers apply Möbius operations with tangent-space activations via log/exp maps.

Worked example

Poincaré disk, points $a = (0.5, 0)$ , $b = (-0.5, 0)$ :

Euclidean gap: $1.0$ .
Hyperbolic: $\|a-b\|^2 = 1$ , denominators $(1-0.25)^2 = 0.5625$ . Argument $1 + 2/0.5625 \approx 4.556$ . $d = \text{arcosh}(4.556) \approx 2.197$ . Curved distance more than doubles the straight gap.
Root at origin to $a$ : $d(0,a) = 2\,\text{artanh}(0.5) \approx 1.099$ . Two edge hops via root total $2.197$ , equal to direct $a$ - $b$ path here by symmetry. Hierarchy depth reads as radius cleanly.

Watch Out For

Training in float32 near the boundary

Coordinates within $10^{-7}$ of the rim round to the boundary and distances blow to NaN. Deep leaves live exactly there.

Use float64, clip norms below $1 - 10^{-5}$ , or switch to Lorentz coordinates for stability.

Embedding non-hierarchical data hyperbolically

Cycles and grids have no tree core; hyperbolic forcing distorts them worse than Euclidean. Benchmarks on mesh graphs show the damage clearly.

Check Gromov hyperbolicity (thin triangles) before choosing the geometry.

The Quick Version

Hyperbolic volume grows exponentially, matching tree branching rates.
Poincaré ball scales the metric up toward the boundary rim.
Roots embed centrally with leaves fanning outward by depth.
Möbius operations generalize vector arithmetic to the curved space.
Precision and hyperbolicity checks precede any embedding choice.