No Triangulation Without Representation: Generalization in Topological Deep Learning

Johannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck · 2026

Read No Triangulation Without Representation: Generalization in Topological Deep Learning by Johannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck (2026) 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-2605-06467.txt.

Introduction

Johannes S. Schmidt and five co-authors examine how topological deep learning models should be evaluated. These models work on higher-order data such as manifold triangulations, which allow operations that graphs do not. Evaluation practice has lagged, and earlier benchmark claims suggested a clear advantage for higher-order methods.

The team extends the MANTRA benchmark to more manifold types and finds both graph networks and higher-order methods can saturate it, given the right representation and features. Under a stricter protocol using varied representations and refinements, no model shows generalization beyond the data's combinatorial structure. Their protocol gives future work a firmer test for true topological understanding.

Catalog record

CREATORJohannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06467

Browse the full library · About the Tower of Babel

Full text

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

No Triangulation Without Representation: Generalization in Topological Deep Learning Johannes S. Schmidt, Martin Carrasco, Ernst Röell, Guy Wolf, Nello Blaser, Bastian Rieck Abstract Despite an ever-increasing interest in topological deep learning models that target higher-order datasets, there is no consensus on how to evaluate such models. This is exacerbated by the fact that topological objects permit operations, such as structural refinements, that are not appropriate for graph data. In this work, we extend MANTRA , a benchmark dataset containing manifold triangulations, to a larger class of manifolds with more diverse homeomorphism types. We show that, unlike prior claims, both graph neural networks (GNNs) and higher-order message passing (HOMP) methods can saturate the benchmark. However, we find that this is contingent on the right representation and feature assignment, emphasizing their importance in baseline models. We thus provide a novel evaluation protocol based on representational diversity and triangulation refinement. Surprisingly, we find no indication that existing models are capable of generalizing beyond the combinatorial structure of the data. This points towards a research gap in developing models that understand topological structure independent of scale. Our work thus provides the necessary scaffolding to evaluate future models and enable the development of topology-aware inductive biases. 1 Introduction Topological deep learning (TDL) models aim to provide paradigms for learning on higher-order datasets, i.e., datasets that go beyond the dyadic relations captured in graphs [ 30 ] . Such datasets include cell complexes or simplicial complexes, and the last years have seen a plethora of different models and paradigms for learning on such domains [ 5 , 6 , 15 , 18 , 41 , 10 , 22 , 35 , 34 , 1 ] , like higher-order message passing (HOMP). Understanding such data to the point of generalization is not only relevant for topology but also for applied mathematics or physics, where lattice quantum field theory, for example, can be studied via space-time triangulations [ 38 ] . Beyond just being higher-order, topological data differs from graph data in crucial aspects. Most importantly, the same underlying space typically admits many different combinatorial representations; that is, any manifold can be described by different triangulations (cf. Section 2 ). Models that only rely on combinatorial data will generalize poorly in such cases, necessitating different model evaluation strategies for topological data. In this context, TopoBench [ 37 ] already provides a large-scale benchmark and evaluation framework for TDL, but its datasets are predominantly graphs lifted to higher-order representations, a procedure that does not reflect intrinsic properties of higher-order spaces. As a consequence, model evaluations are effectively based on graph data, thus undermining the credibility of TDL models and leading to criticism concerning the necessity of higher-order representations [ 32 ] as well as concerns about the theoretical expressivity of HOMP [ 16 ] . The MANTRA dataset [ 2 ] addresses this gap, being the only dataset consisting of inherently higher-order objects, namely combinatorial manifolds . However, despite providing suitable data , we find that the original MANTRA dataset has several shortcomings in its evaluation methodology, which we address in this work. Our work makes the following contributions : • We extend the MANTRA 1 1 1 Licensed under BSD-3-Clause license. dataset with a fully-characterized set of triangulations in dimensions 2 and 3 , establishing Pachner moves [ 29 ] , i.e., local modifications preserving the homeomorphism type, as principled data augmentation and evaluation tools. • We systematically investigate the impact of different representations and encodings (positional and structural) on predictive performance, demonstrating that standard message-passing algorithms, given the correct input representation, suffice to saturate the benchmark, contrary to prior claims. • We introduce refinement schemes for triangulations as a generalization stress test, revealing the fact that all considered models fail to generalize across refinements, raising doubts about whether they learn genuine topological structure or merely exploit combinatorial artifacts. Taken together, our findings reveal that current TDL models are not truly topological in the sense that they learn combinatorial artifacts rather than topological structure , thus calling for a fundamental reassessment of model design and evaluation. e_{4} e_{5} e_{6} e_{2} e_{3} e_{1} e_{3} e_{1} e_{2} v_{2} v_{3} v_{4} v_{1} v_{1} v_{1} \sigma_{1} \sigma_{3} \sigma_{4} \sigma_{2} (a) Simplicial complex \mathcal{T}\cong S^{2} \sigma_{2} \sigma_{3} \sigma_{4} \sigma_{1} e_{1} e_{2} e_{3} e_{4} e_{5} e_{6} v_{1} v_{2} v_{3} v_{4} (b) Hasse diagram \mathcal{H} \sigma_{1} \sigma_{2} \sigma_{3} \sigma_{4} e_{4} e_{5} e_{6} e_{2} e_{1} e_{3} (c) Dual graph \mathcal{D}\cong K_{4} Figure 1 : Three representations of S^{2} , the 2 -sphere. The simplicial complex and Hasse diagram contain the same information, while the dual graph only represents maximal simplex connectivity. 2 Background: Simplicial complexes and combinatorial manifolds In this paper, we consider combinatorial manifolds , which afford parsimonious representations by means of simplicial complexes . A simplicial complex \mathcal{T} is a finite collection of simplices (finite nonempty sets), that is closed under taking faces (subsets). Given a simplex \sigma=\{v_{0},v_{1},\dotsc,v_{k}\} , we say that \dim\sigma=k and refer to \sigma as a k -simplex . A simplicial complex \mathcal{T} is called pure or homogeneous if all its maximal simplices have the same dimension d . Given a simplex \sigma , its link is the subcomplex \lk(\sigma)\coloneqq\{\tau\in\mathcal{T}\mid\tau\cap\sigma=\emptyset,\tau\cup\sigma\in\mathcal{T}\} . If \sigma=\{v\} is a vertex , i.e., a 0 -simplex , the link can be seen as the boundary of a small neighborhood of v in the simplicial complex. Given a d -dimensional homogeneous simplicial complex \mathcal{T} , we say that \mathcal{T} is a combinatorial d -manifold if \lk(v) is a triangulated ( d-1 )-sphere for every vertex v . 2 2 2 For readers familiar with differential geometry/topology, this mimics the informal definition of a smooth d -manifold “locally looking like” \mathbb{R}^{d} . While combinatorial manifolds are not restricted in terms of their dimension, we focus on d\in\{2,3\} because in this setting, topological manifolds coincide with combinatorial manifolds [ 23 , 25 , 33 ] . This has two implications, namely (i) every combinatorial manifold triangulates some topological manifold, and (ii) the data augmentation procedures we subsequently introduce provide us with a way to represent all topological manifolds in these dimensions as a combinatorial manifold, ensuring that our data is, in principle, a complete combinatorial census of low-dimensional manifolds. Characterizing and representing triangulations. A topological manifold does not have one unique combinatorial manifold, i.e., simplicial complex, assigned to it. To measure whether two simplicial complexes triangulate the same manifold, we calculate topological invariants like the Euler characteristic \chi . For a simplicial complex \mathcal{T} , the Euler characteristic is the alternating sum of simplex counts in each dimension, denoted by \chi(\mathcal{T})\coloneq\sum_{k=0}^{d}(-1)^{k}|\mathcal{T}_{k}| . Another invariant is the orientability of a simplicial complex. We say that \mathcal{T} is orientable if we can assign consistent orientations to all its d -simplices , meaning that for each (d-1) -face shared by two d -simplices , the induced orientations are opposite . 3 3 3 Consider orienting all triangles of a triangulated surface by traversing them either clockwise or counterclockwise. The orientation is consistent if each edge is traversed once in each direction by its two adjacent triangles. As we will describe below, the Euler characteristic and orientability are, taken together, sufficient to fully characterize 2 -manifolds . Finally, notice that a simplicial complex

Read the complete text