Discrete Approximate Circle Bundles
Brad Turow, Jose A. Perea · 2025
Read Discrete Approximate Circle Bundles by Brad Turow, Jose A. Perea (2025) 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-2508-12914.txt.
Introduction
Brad Turow and Jose Perea wrote a 2025 paper giving data scientists a discrete version of circle bundles. In topology, a circle bundle is a space built from circles attached over a base, like a twisted loop structure. Real data, though, arrive as finite samples, not smooth spaces.
The authors define discrete approximate circle bundles and show when such an object can be safely identified with a true bundle's class. Two cohomology invariants pin down that class completely. They give algorithms to compute both from a discrete sample. A further method turns bundle structure into coordinates, reducing dimension for analysis. Tests on synthetic and real computer vision data, including optical flow, show the tools work. An open-source software package with tutorials accompanies the paper.
Catalog record
| CREATOR | Brad Turow, Jose A. Perea |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 143 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-12914
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Discrete Approximate Circle Bundles Brad Turow, Jose A. Perea Abstract In this paper, we introduce discrete approximate circle bundles , a class of objects designed to serve as the data science analog of circle bundles from algebraic topology. We show that, under appropriate conditions, one can meaningfully and stably identify a discrete approximate circle bundle with an isomorphism class of true circle bundles. We also describe two cohomology invariants which uniquely determine the isomorphism class of a circle bundle, and provide algorithms to compute them given a discrete approximate representative. Finally, we propose a novel methodology for coordinatization and dimensionality reduction of circle bundle data. To illustrate the practical utility and viability of our algorithms, we present applications to both real and synthetic datasets from computer vision (e.g., modeling optical flow). 1 Background And Problem Statement 1.1 Motivation And Context In recent years, it has been observed that high-dimensional datasets arising in fields such as computer vision, computational chemistry, and motion tracking often lie near highly non-linear, low-dimensional manifolds [ TSL00 , RS00 , BN03 ] . These manifolds can exhibit complex geometric and topological structure, motivating the development of a range of techniques for identifying and leveraging such structures in tasks including dimensionality reduction, statistical analysis, pattern recognition, and data visualization [ CL06 , Sap01 ] . By modeling the intrinsic manifold geometry, these methods have been shown to improve compression, denoising, and feature extraction in domains where traditional linear techniques fail to capture the underlying structure of the data. In this paper, we focus on a particular class of datasets which can be modeled as circle bundles . Roughly speaking, a circle bundle is a continuously-varying family of (topological) circles parameterized by another space, called the base space of the bundle. The individual circles are called fibers of the bundle, and the union of the fibers is the total space . The parametrization is given by a projection map \pi:E\to B from the total space to the base space – E is locally homeomorphic to the product B\times\mathbb{S}^{1} , but globally the structure may be twisted. For instance, the torus and the Klein bottle can each be characterized as the total space of a circle bundle with base space \mathbb{S}^{1} , but the two are not homeomorphic (see Figure 1 ). Similarly, the unit tangent bundle on the 2-sphere is a circle bundle with base \mathbb{S}^{2} , but one can show that there is no bundle isomorphism between this bundle and the product space \mathbb{S}^{2}\times\mathbb{S}^{1} . K \pi^{-1}(U) \mathbb{S}^{1} U \pi Fiber \mathbb{S}^{1} Figure 1 : The Klein bottle K as a circle bundle over \mathbb{S}^{1} . The highlighted region \pi^{-1}(U) in K shows all fibers above the neighborhood U in the base space and is homeomorphic to U\times\mathbb{S}^{1} . Globally, however, K is not homeomorphic to \mathbb{S}^{1}\times\mathbb{S}^{1} . On the machine learning side, various datasets related to computer vision and graphics have been shown to have circle bundle structures. For instance, in [ Ada+20 ] , Adams et al. use local measurements and a feature map to support a torus model for high-contrast optical flow data. Similarly, in [ Car+08 ] , the authors propose that a certain space of high-contrast natural image patches is concentrated around a Klein bottle embedded in high dimension. Beyond these examples, datasets arising from 3D objects with rotational symmetries — such as projection images of macromolecules in cryo-electron microscopy — naturally give rise to circle bundle structures, where each fiber corresponds to rotations of the object about a fixed axis [ SS14 ] . The possible circle bundle structures over a metric space B are completely described by a pair of discrete invariants called characteristic classes . In particular, a non-vanishing class indicates the bundle is non-trivial, so there is no hope of constructing a coordinatization for the whole space as a product. Additionally, characteristic classes are invariant under continuous perturbations, so one might hope to be able to stably recover them from a noisy finite sampling. Another key feature of characteristic classes is that they can be computed from local measurements alone, which makes them an ideal tool for working with complex, high-dimensional data not amenable to more direct techniques. Lastly, the characteristic classes of a bundle can be used to construct a so-called bundle map from the total space to a more familiar space which respects topological structure. This observation has recently been leveraged in works such as [ Lee+25 ] and [ SP23 ] to develop non-linear dimensionality reduction techniques based on vector bundles. In this paper, we extend this approach to circle bundles. 1.2 Our Contributions Here we summarize our contributions: 1. We define notions of discrete approximate circle bundles and local trivializations and show how these relate to other well-studied objects such as discrete approximate Čech cocycles. Building on results of [ Tin22 ] and [ SP23 ] , we give conditions under which a discrete approximate circle bundle can be meaningfully identified with a true class of circle bundles. 2. We show that every true circle bundle is uniquely characterized by an orientation class and a twisted Euler class, and provide an algorithm for computing these classes from a family of discrete approximate local trivializations. We further introduce a metric on each space of discrete approximate local trivializations and show that two trivializations which are sufficiently close produce the same characteristic classes. In particular, we prove that if a discrete approximate local trivialization is sufficiently close to a true one, the characteristic classes obtained from our algorithm are the same. 3. Given a discrete approximate circle bundle over a base space B and an open cover \mathcal{U} for B , we introduce a weights filtration on the nerve complex \mathcal{N}(\mathcal{U}) and a corresponding notion of persistence for characteristic class representatives. We then propose a scheme for coordinatization and dimensionality reduction which produces a map from the dataset to the total space of the associated bundle V(2,d)\times_{O(2)}\mathbb{S}^{1} , where V(2,d) denotes the Stiefel manifold of 2-planes in \mathbb{R}^{d} . Our pipeline can be tuned to be compatible with any stage of the weights filtration on \mathcal{N}(\mathcal{U}) , yielding a point cloud with a different global topology at each stage. 4. We provide implementations of our algorithms on several real and synthetic datasets related to imaging and optical flow. In Section 7.1 , we use these tools to confirm the torus model proposed in [ Ada+20 ] . A more detailed analysis of our findings will appear in an upcoming companion paper. 1.3 Prior Work A number of approaches to local-to-global inference for analyzing geometrically and topologically complex datasets have recently emerged. The basic principle in each case is to model small pieces of the dataset and analyze how these parts glue together to determine the global shape. Given a dataset X , a common technique is to choose a feature map \pi:X\to S to some metric space S (e.g., \mathbb{R} ) and study how the level sets vary. The Mapper pipeline takes this approach to represent a point cloud as a graph with nodes representing local clusters. Similarly, in [ SP22 ] the authors describe a pipeline called FibeRed which uses local PCA together with a