← All Posts Research

From Loops to Klein Bottles: Uncovering Hidden Topology in High Dimensional Data

March 27, 2025  ·  Gunnar Carlsson
From Loops to Klein Bottles: Uncovering Hidden Topology in High Dimensional Data

This article explores how complex topological structures hidden within high-dimensional datasets can be uncovered and represented without relying on traditional vector embeddings.

Motivation

Dimensionality reduction helps analyze datasets with many features. There is an important situation where data clusters around nonlinear surfaces of low dimension but requiring high-dimensional embedding spaces — understanding the intrinsic shape of data unlocks more powerful representations.

Nonlinear Surfaces

The article discusses representing surfaces through three approaches: polynomial equations, intrinsic coordinates, and triangulation. Triangulation offers an intrinsic description independent of vector embedding, reducing representation to actual surface dimensions.

Figure 1: Triangulations of a sphere and of a torus

The Loop Example

Using climate data from Jena (2009–2016) with four selected features, PCA reveals data organized around a circular loop structure. This suggests a single angular coordinate could replace two-dimensional reduction.

Figure 2: Jena climate data

The Klein Bottle Example

Examining the Mumford dataset of 8 million 3×3 image patches, the analysis reveals concentration around a Klein bottle structure. The authors explain that "Betti numbers are attached to every manifold," with Klein bottles characterized by specific topological properties.

Figure 3: Image Patches
Figure 4: Klein bottle

Topology & Mathematical Representation

The post introduces simplicial complexes as combinatorial descriptions of spaces, using graphs and triangulations to represent manifolds intrinsically rather than through algebraic equations.

Figure 5: Triangulated Jena data set
Figure 6: Tetrahedron
Figure 7: Moebius band
Figure 8: Identification space description of Klein bottle
Figure 9: Image patches laid out on Klein bottle

Applications to Data Science

Three practical questions arise: how topological models are derived, how they generate features, and whether they yield algebraic equations describing datasets. These questions form the foundation of BluelightAI's approach to AI interpretability.