A Set-to-Set Distance Measure in Hyperbolic Space
Pengxiang Li Wei Wu Zhi Gao, Xiaomeng Fan Peilin Yu, Yuwei Wu Zhipeng Lu Mehrtash Harandi Yunde Jia · 2025
Read A Set-to-Set Distance Measure in Hyperbolic Space by Pengxiang Li Wei Wu Zhi Gao, Xiaomeng Fan Peilin Yu, Yuwei Wu Zhipeng Lu Mehrtash Harandi Yunde Jia (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-2506-18529.txt.
Introduction
Pengxiang Li and co-authors measure the gap between whole sets of points in curved hyperbolic space in a 2025 paper. Such space suits family tree data, but many jobs compare groups, not lone points, and group shape at two scales carries meaning.
Their distance blends a global term, the geodesic gap between set centers, with a local term from each set's wiring pattern. A finite sequence trick makes the wiring term cheap to compute. Tests on entity matching and image sorting, including few shot rounds, beat prior measures. The gain comes from honoring both scales at once.
Catalog record
| CREATOR | Pengxiang Li Wei Wu Zhi Gao, Xiaomeng Fan Peilin Yu, Yuwei Wu Zhipeng Lu Mehrtash Harandi Yunde Jia |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 88 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-18529
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Set-to-Set Distance Measure in Hyperbolic Space Pengxiang Li Wei Wu Zhi Gao, Xiaomeng Fan Peilin Yu, Yuwei Wu Zhipeng Lu Mehrtash Harandi Yunde Jia Abstract We propose a hyperbolic set-to-set distance measure for computing dissimilarity between sets in hyperbolic space. While point-to-point distances in hyperbolic space effectively capture hierarchical relationships between data points, many real-world applications require comparing sets of hyperbolic data points, where the local structure and the global structure of the sets carry crucial semantic information. The proposed the h yperbolic s et- to - s et d istance measure (HS2SD) integrates both global and local structural information: global structure through geodesic distances between Einstein midpoints of hyperbolic sets, and local structure through topological characteristics of the two sets. To efficiently compute topological differences, we prove that using a finite Thue-Morse sequence of degree and adjacency matrices can serve as a robust approximation to capture the topological structure of a set. In this case, by considering the topological differences, HS2SD provides a more nuanced understanding of the relationships between two hyperbolic sets. Empirical evaluation on entity matching, standard image classification and few-shot image classification demonstrates that our distance measure outperforms existing methods by effectively modeling the hierarchical and complex relationships inherent in hyperbolic sets. 1 Introduction Learning algorithms in hyperbolic spaces with constant negative curvature have achieved impressive performance across multiple domains, including few-shot learning ( gao2021curvature ; ijcai2022p517 ) , clustering ( Lin_2023_ICCV ) , retrieval ( Ermolov_2022_CVPR ) , segmentation ( Atigh_2022_CVPR ) , multi-modal learning ( Long_2023_ICCV ) , and 3D vision ( hsu2021capturing ) . The reason behind their success is mainly due to the exponentially growing space volume in hyperbolic spaces with the radius ( pmlr-v80-sala18a ; balazevic2019multi ) , providing minimal distortion for hierarchical structures of practical data. In the current literature, hyperbolic geometry is usually exploited via point-to-point distance measures: each sample is embedded as a single data point, and downstream tasks such as classification ( ye2020fewshot ) , segmentation ( chen2023hyperbolic ) , and retrieval ( yan2023hyp ) are performed by computing geodesic distances between the data points ( khrulkov2020hyperbolic ; long2020searching ; atigh2022hyperbolic ) . However, many real-world problems naturally involve comparisons between sets of hyperbolic embeddings rather than individual points. For example, in entity matching ( hui2022localized ) (Fig. 1 (a)), each record comprises an unordered set of attribute embeddings, and a hyperbolic set-to-set distance can capture the relational patterns among attributes to decide whether two records refer to the same entity. Similarly, in few-shot image classification ( zhang2022hyperbolic ; hamzaoui2024hyperbolic ) (Fig. 1 (b)), treating the multi-layer feature pyramid of a single image as a set enables richer representations, requiring a robust set-to-set distances. Unfortunately, the set-to-set distance measure in hyperbolic spaces has yet to be explored. Figure 1 : Examples of HS2SD on the entity matching and few-shot learning tasks. In this paper, we propose the h yperbolic s et- to - s et d istance (HS2SD) that leverages both the global and local structures of hyperbolic sets. Given two sets of hyperbolic representations, the global structures are responsible for measuring the relative positioning between them in the entire hyperbolic space, considering their hierarchical relationships. On the other hand, the local structures capture the topological differences between the two sets, allowing the distance measure to discern subtle variations in the sets, which might be overlooked by a purely global perspective. By combining both global and local structures, the hyperbolic set-to-set distance enables a more nuanced and accurate measurement of dissimilarity, capable of capturing both the overall distribution of the sets and their detailed, underlying patterns. This dual consideration enhances the ability to apply set-to-set distance measures in complex tasks where both global relationships and local topological features are important. In doing so, the HS2SD is designed to integrate two key components: geodesic distance and topological distance for the global and local structures, respectively. The geodesic distance aggregates hyperbolic representations of a set into an Einstein midpoint ( gulcehre2018hyperbolic ) and computes the hyperbolic l_{1} distance between two Einstein midpoints for a global difference, thereby quantifying their overall dissimilarity. The topological distance constructs a topology graph of a set and assesses the structural similarities between the topology graphs between two sets for local differences. In this case, unlike existing point-to-point hyperbolic distance measures that focus on relationships between individual points, HS2SD integrates both global and local perspectives to provide a more nuanced understanding of the relationships between data sets. To achieve this, we must solve one critical problem: how to represent the topological structures of the sets and compute their differences. In doing so, we construct a topology graph based on the distances between points and compute its adjacency and degree matrices. Theoretically, through the lens of isomorphism ( McKay ) , we demonstrate that the infinite polynomial combination of degree matrix and adjacency matrix serves as a valid descriptor for evaluating the structural similarity of trees. To make a feasible and efficient computation of the infinite polynomial combinations, we model the combinations of degree and adjacency matrices leveraging Thue-Morse sequences ( morse1921recurrent ) , which exhibit uniform distributions to well approximate the infinite polynomial combination and are computed recursively to reduce the computation burdens ( kuipers2012uniform ; allouche2003automatic ) . By computing the Frobenius norm of the difference between two Thue-Morse sequences, we derive the topological distance. We conduct experiments on the entity matching and few-shot image classification tasks, and our method achieves better performance than compared methods that use a point-to-point distance measure, showing the effectiveness of capturing both global and local structures in hyperbolic spaces. The primary contributions of our work can be summarized as follows: • We propose a hyperbolic set-to-set distance measure that integrates the geodesic distance and topological distance for a comprehensive assessment of both global and local structures in hyperbolic sets. • We introduce an effective and efficient manner to measure the topological distance between two hyperbolic sets, where we use a finite Thue-Morse sequence for robust representations of topological structures of trees, while avoiding computation burdens. 2 Related Work 2.1 Hyperbolic Geometry Hyperbolic geometry, characterized by constant negative curvature, enables efficient embedding of hierarchical data as continuous analogs of trees. This property has driven research in hyperbolic learning across domains including few-shot learning ( Anvekar_2023_CVPR ; hamzaoui2024hyperbolic ) , 3D vision ( lin2023hyperbolic ) , and image segmentation ( chen2022hyperbolic ) . Early approaches to hyperbolic learning employed conventional neural networks for feature extraction in Euclidean space before projecting embeddings into hyperbolic space ( khrulkov2020hyperbolic ; atigh2022hyperbolic ) . For instance, liu2020hyperbolic extracted image features using convolutional networks in Euclidean space and mapped them to hyperbolic space via the exponential map, demonstrating improved robustness in zero-shot recognition tasks. However, these hybrid approaches often introduce data distortion due to the fundamental geometric incompatibility between their Euclidean backbones and hyperbolic target spaces. Recent work addresses this limitation by designing neural architectures that operate natively in hyperbolic spaces, including hyperbolic convolutional networks ( bdeir2024fully ) and hyperbolic residual networks ( van2023poincare ) , which preserve the geometric structure throughout the feature extraction process. Recent work has explored hyperbolic geometry for metric learning, including point-to-set distances for few-shot classification ( ma2022adaptive )