Discrete Tiling for Numerically Stable Hyperbolic Representations

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 113 papers26 conceptsfrozen 2026-10-04

Hyperbolic geometry is the natural language of hierarchy. Because the "volume" of hyperbolic space grows exponentially with its radius, it can embed complex trees and social networks with almost no distortion. However, this exponential growth creates a fundamental computational hurdle: numerical instability.

In standard continuous models like the Poincaré disk, points near the boundary require extreme floating-point precision. As you move away from the origin, the distance between two distinct points can become smaller than the smallest representable difference in a 64-bit float. This leads to the "NaN problem" in machine learning and rendering artifacts in visualization.

To solve this, researchers are moving away from continuous coordinates toward discrete hyperbolic tilings. By partitioning the space into a regular grid of "tiles," we can represent positions using discrete combinatorial addresses rather than fragile floating-point numbers.

The Precision Wall in Continuous Space

In Euclidean space, if you double the radius of a circle, the area quadruples. In hyperbolic space, doubling the radius causes the area to increase exponentially. This property is what makes hyperbolic space so useful for data—it has "enough room" for everything.

But as Eryk Kopczyński and Dorota Celińska-Kopczyńska point out in Generating Regular Hyperbolic Honeycombs, this growth is a nightmare for standard hardware. When representing a large-scale network, nodes at the periphery become numerically indistinguishable. In deep learning, this manifests as gradients that explode or vanish, causing optimization to fail.

Discretization as a Numerical Scaffold

The core insight of this research direction is to treat hyperbolic space like a giant, infinite puzzle of interlocking tiles (a tessellation). Instead of a single coordinate $(x, y)$, a point is represented by:

  1. A discrete tile ID: An integer or a path in a tree representing which "cell" the point lives in.
  2. A local offset: A small floating-point value representing the position within that cell.

Tao Yu and Christopher De Sa demonstrated in Numerically Accurate Hyperbolic Embeddings Using Tiling-Based Models that this approach keeps numerical error bounded and constant, regardless of how far a point is from the origin. By "snapping" coordinates to a structured grid, the model avoids the ill-conditioned metrics that lead to overflow.

Because hyperbolic space is "tree-like," its tilings can be mapped to geodesic regular tree structures (GRTS). In Generating Tree Structures for Hyperbolic Tessellations, Celińska-Kopczyńska and Kopczyński show that navigating these spaces doesn't require complex trigonometry. Instead, one can move from tile to tile using discrete graph transitions.

This combinatorial approach has three major advantages:

  • Speed: In Bridging ML and algorithms, the authors found that discrete algorithmic approaches can be 100 times faster than continuous ML-based embeddings while maintaining equal quality.
  • Stability: In Real-Time Visualization in Non-Isotropic Geometries, the world is partitioned into a honeycomb. Objects are rendered relative to the observer's local tile, preventing floating-point drift even in infinite manifolds.
  • Exactness: Discrete distances in a hyperbolic grid are highly accurate approximations of continuous distances. Kopczyński shows in Dynamic Distances in Hyperbolic Graphs that this enables $O(1)$ query times for distances in massive hierarchical graphs.

From Theory to Physical Circuits

The utility of discrete hyperbolic grids extends beyond software. In Simulating hyperbolic space on a circuit board, Patrick M. Lenggenhager and colleagues physically realized a hyperbolic lattice using a network of capacitors and inductors. By discretizing the plane into a regular tessellation where more polygons meet at each vertex than in Euclidean geometry, they were able to observe hyperbolic physical phenomena—like unique signal propagation—on a flat, finite circuit board.

Summary

The shift from continuous to discrete hyperbolic models replaces fragile floating-point math with robust combinatorial logic. Whether through the Discrete Hyperbolic Random Graph (DHRG) model for network analysis or tiling-based embeddings for machine learning, discretization provides a "numerical scaffold" that allows us to exploit the vast capacity of hyperbolic space without falling off its mathematical edge.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_discrete_tiling_for_numerically_stable_h_a85916ed,
  title = {Discrete Tiling for Numerically Stable Hyperbolic Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/discrete-tiling-for-numerically-stable-hyperbolic-representations/rdr_a85916ed004d42709417c20076e93422},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_discrete_tiling_for_numerically_stable_h_a85916ed_v1,
  title = {Discrete Tiling for Numerically Stable Hyperbolic Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/discrete-tiling-for-numerically-stable-hyperbolic-representations/rdr_a85916ed004d42709417c20076e93422/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}