curvature as a tool for evaluating dimensionality reduction and estimating intrinsic dimension

Charlotte Beylier, Parvaneh Joharinad, Jürgen Jost, Nahid Torbati · 2025

Read curvature as a tool for evaluating dimensionality reduction and estimating intrinsic dimension by Charlotte Beylier, Parvaneh Joharinad, Jürgen Jost, Nahid Torbati (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-2509-13385.txt.

Introduction

Charlotte Beylier, Parvaneh Joharinad and two co-authors use sectional curvature in 2025 to grade shrink maps and gauge true data dimension. Curvature here reads metric links among triples and further points. A profile sums those links for a discrete space in full.

A quantitative score grades representation quality from the profile. Tests show the profile also estimates intrinsic dimension well. Large empirical networks are explored with the tool. Shrink methods are graded on real sets. Geometry thus gives a shared yardstick where error alone misleads. Dimension calls gain a check that is apart from the shrink method itself.

Catalog record

CREATORCharlotte Beylier, Parvaneh Joharinad, Jürgen Jost, Nahid Torbati
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-13385

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

curvature as a tool for evaluating dimensionality reduction and estimating intrinsic dimension Charlotte Beylier, Parvaneh Joharinad, Jürgen Jost, Nahid Torbati Abstract Utilizing recently developed abstract notions of sectional curvature, we introduce a method for constructing a curvature-based geometric profile of discrete metric spaces. The curvature concept that we use here captures the metric relations between triples of points and other points. More significantly, based on this curvature profile, we introduce a quantitative measure to evaluate the effectiveness of data representations, such as those produced by dimensionality reduction techniques. Furthermore, Our experiments demonstrate that this curvature-based analysis can be employed to estimate the intrinsic dimensionality of datasets. We use this to explore the large-scale geometry of empirical networks and to evaluate the effectiveness of dimensionality reduction techniques. 1 Introduction Often, the geometric structure of datasets can be captured by graphs, and they play a central role in modern machine learning methods such as Convolutional Neural Networks (CNNs) LeCun et al., 1998 and Graph Neural Networks (GNNs) Scarselli et al., 2008 . It is then natural, if not mandatory, to ask how well graphs preserve or reveal the underlying geometry. We therefore need principled tools that can infer geometric and topological features within these models. Graphs themselves being viewed as metric spaces, require tools evolving from geometry. However, many of such tools were initially developed for particular smooth structures, namely the Riemannian manifolds. For example, if S\subset\mathbb{R}^{3} is an oriented surface, the Gaussian curvature quantifies the deviation of S at each point from flatness. Following Riemann’s foundational work, it is well-established that Gaussian curvature can be computed using the intrinsic geometrical structure induced by the Euclidean geometry of \mathbb{R}^{3} . This established the foundation for the development of Riemannian geometry by extending the concept of Gaussian curvature to sectional curvature. The latter is defined as the measure of deviation from a flat geometry, providing a comprehensive set of invariants for the characterization of the geometry of a manifold Jost, 2017 . Sectional curvature, however, is derived from the second derivative of the metric tensor and thus cannot be directly applied to general metric spaces. Still, upper and lower bounds on sectional curvature can impose restrictions on the metric structure of manifolds, providing a qualitative measure of how general metric spaces (i.e., geodesic spaces) deviate from Riemannian manifolds of constant sectional curvature (i.e., model spaces). For such general metric spaces, there exist various synthetic curvature notions that replace the infinitesimal computations. Here, we use the notion introduced in Joharinad and Jost, 2019 , which applies even to discrete metric spaces, for analyzing the large-scale geometry of graphs. Our curvature profile is a distribution function evaluated on equilateral triples of nodes, organized according to their perimeter. This profile acts as a geometric signature, capturing curvature characteristics of the graphs as deviations from tree-like structure at different scales. For a metric dataset (X,d) ( d might record pairwise (dis)similarities), we first transform the metric structure into a combinatorial model by constructing a neighborhood graph and evaluating its curvature profile as a geometric fingerprint. For quantifying how well low dimensional embedding methods preserve the intrinsic geometry of the data, we compute the 1-Wasserstein distance ( W_{1} ) Villani et al., 2008 between the curvature profiles of the initial dataset and its corresponding embedding. In particular, varying the target dimension, we can identify the embedding whose curvature distribution best matches that of the original dataset, providing an effective estimate of the intrinsic dimension. To demonstrate the effectiveness of the curvature profile in capturing and inferring underlying geometric structure, we first apply our method to metric datasets with known geometries, the circle, the plane, and a metric tree, and then to empirical data. Our full code is available at 1 1 1 https://anonymous.4open.science/r/curvature-D6FE . 2 Graphs and curvatures Curvature is a concept that originated in differential geometry and that plays a fundamental role in Riemannian geometry. It measures the local deviation of a space from being Euclidean. In fact, in 1854 Bernhard Riemann showed that the sectional curvatures of a Riemannian manifold provide a complete set of local invariants of the Riemannian metric Jost, 2025 . That is, if you know the curvatures, you know the intrinsic geometry of a Riemannian manifold. Here, intrinsic refers to measurements that are taken inside the manifold and do not use any relations in an ambient space into which the manifold is embedded. And a contraction of the sectional curvatures, the Ricci curvatures became fundamental, for instance, in Einstein’s theory of general relativity. In the 20th century, mathematicians developed notions of synthetic curvature for more general metric spaces. When the metric space is Riemannian, they should of course reduce to the classical notions. The most useful such generalized curvature notions were introduced by Alexandrov Alexandrov, 1957 and Busemann Busemann, 1955 . They provide inequalities that encode important local and global properties of the space. They have become powerful tools for probing the geometry of metric space. Graphs can also be seen as metric spaces. In the simplest case, when the graph is unweighted, the distance between two vertices is the minimal number of edges needed to get from one to the other. Therefore, synthetic curvature notions also apply to graphs, and they have become useful tools for the analysis of empirical networks represented as graphs. As indicated, there are two important notions of curvature, sectional and Ricci. While the latter is a contraction of the former, in Riemannian geometry they play somewhat opposite roles, see e.g. Jost, 2017 . From a positive lower bound on the Ricci curvature, one can derive many consequences for the local geometry, like diameter bounds, restrictions on the growth of the volume of balls as function of their radius, average divergence of geodesics, eigenvalue bounds and so on. In contrast, negative Ricci curvature does not imply any restriction at all. For sectional curvature, it is the opposite. Here, nonpositive or even better, negative sectional curvature implies many global restrictions, that is, restrictions on the large scale geometry. Synthetic versions of either curvature have been applied in graph theory and network analysis. Since the Ricci curvature of a Riemannian manifold is evaluated on tangent vectors, the Ricci curvature of a graph is naturally computed for edges. Since the sectional curvature is evaluated on tangent planes and synthetic notions are usually formulated in terms of distance relations in geodesic triangles, the sectional curvature of a graph is naturally computed on triangles of vertices. Here, a triangle is a triple of vertices connected by shortest edge paths. These triangles can be large, in line with the observation that Riemannian sectional curvatures encode global properties. There are several versions of synthetic Ricci curvature. While for Riemannian manifolds, they all coincide, as they should, for graphs they are very different. The most prominent notions were developed by Bakry-Émery, Forman, and Ollivier. The statistics of such Ricci curvatures on empirical networks has become a useful and versatile to detect significant properties and to distinguish different types of networks, see for instance, as samples of a very rich literature, Samal et al., 2018 , Saucan et al., 2018 , Mondal et al., 2024 . But since in the present paper, we are rather interested

Read the complete text