Heat Kernel Goes Topological

Maximilian Krahn · 2025

Read Heat Kernel Goes Topological by Maximilian Krahn (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-2507-12380.txt.

Introduction

Maximilian Krahn and Vikas Garg present a 2025 paper on topological neural networks. Such networks extend graph learning to higher-order structures. Their message passing is powerful but costly. The team builds a Laplacian operator on combinatorial complexes, a broad shape frame. Heat kernels from that operator serve as node descriptors. The descriptors carry multiscale shape facts at low cost.

Theory says the method is as expressive as possible: it can tell apart any unlike complexes. Tests show big speed gains over past topological methods. Scores on standard molecule data sets stay competitive with the best descriptors. Harder shape tests favor the new method as well. ML researchers get topology tools that fit transformer-style models without heavy cost. That opens higher-order learning to larger data than past methods could handle.

Catalog record

CREATORMaximilian Krahn
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-12380

Browse the full library · About the Tower of Babel

Full text

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

Heat Kernel Goes Topological Maximilian Krahn Abstract Topological neural networks have emerged as powerful successors of graph neural networks. However, they typically involve higher-order message passing, which incurs significant computational expense. We circumvent this issue with a novel topological framework that introduces a Laplacian operator on combinatorial complexes (CCs), enabling efficient computation of heat kernels that serve as node descriptors. Our approach captures multiscale information and enables permutation-equivariant representations, allowing easy integration into modern transformer-based architectures. Theoretically, the proposed method is maximally expressive because it can distinguish arbitrary non-isomorphic CCs. Empirically, it significantly outperforms existing topological methods in terms of computational efficiency. Besides demonstrating competitive performance with the state-of-the-art descriptors on standard molecular datasets, it exhibits superior capability in distinguishing complex topological structures and avoiding blind spots on topological benchmarks. Overall, this work advances topological deep learning by providing expressive yet scalable representations, thereby opening up exciting avenues for molecular classification and property prediction tasks. 1 Introduction The intersection of topology and deep learning has unlocked new possibilities for encoding complex structural information in data. Traditional graph-based neural networks have proven effective in learning representations for structured data [ 1 , 2 ] , yet they often struggle with capturing higher-order relationships [ 3 ] . In contrast, combinatorial complexes provide a more expressive framework for modelling such intricate structures, extending beyond pairwise relationships to encode higher-order interactions [ 4 ] . Current topological neural networks often rely on higher-order message passing protocols to capture complex structural relationships [ 5 , 6 ] . While effective in theory, these approaches are typically computationally expensive (cf. Section 5 ) and offer limited expressive power in distinguishing non-isomorphic structures [ 6 ] . Recent methods for relational data Method SC R2 IS EQ CIN [ 7 ] ✓ ✗ ✗ ✓ SMCN [ 6 ] ✓ ✓ ✗ ✓ MCN [ 6 ] ✗ ✓ ✓ ✓ TopNet [ 5 ] ✗ ✗ ✗ ✓ iGN [ 8 ] ✓ ✗ ✗ ✓ IEGN [ 9 ] ✓ ✗ ✗ ✓ TopoHKS (ours) ✓ ✓ ✓ ✓ Contributions Section 3: • Defining the Laplacian for CCs • HKS descriptor for CC cells Section 4: • Expressivity of our approach; can distinguish non-isomorphic structures Section 5: • Experiments for expressivity on torus data, scalability, graph and CC classification Table 1: Overview of recent methods for relational data and summary of our contributions . SC: Scalability, R2: Distinguishability for Rank 2, IS: Isomorphism, EQ: Equivariance. Motivated by these limitations, we propose TopoHKS, a scalable and expressive deep learning framework to distinguish between combinatorial complexes and their isomorphisms. Our method leverages Heat Kernel Signatures (HKS) [ 10 ] on combinatorial complexes, yielding continuous, multi-scale descriptors that capture the topological neighbourhood of each cell through a diffusion-based perspective. These descriptors encode rich structural information and serve as input to a transformer network, enabling us to model complex interactions without relying on expensive higher-order message-passing mechanisms. Our contributions can be summarised as follows: 1. Topological Embeddings for Transformers: We introduce a novel combinatorial complex embedding that is permutation-invariant and computationally efficient, overcoming limitations in existing approaches. 2. Theoretical Expressiveness: We establish our framework’s fundamental topological and spectral properties, demonstrating its ability to distinguish non-isomorphic complexes and outperform Weisfeiler-Lehman (WL)-based methods in expressivity. 3. Empirical Performance: Our method achieves state-of-the-art results on molecular prediction benchmarks (MolHIV, Protein) and topological structure datasets, surpassing existing approaches in distinguishing combinatorial complexes. 4. Scalability and Efficiency: Unlike traditional higher-order message passing methods, our approach scales efficiently to large combinatorial complexes. By bridging topological data analysis and deep learning, our work paves the way for more expressive, computationally efficient, and theoretically grounded approaches to learning on structured data. As shown in Table 1 , we formally introduce our framework, establish its theoretical foundations, and validate its performance through extensive experiments. We will publish the code after acceptance. 2 Background Definition 2.1 . A combinatorial complex (CC) is a triple (S,\mathcal{X},\operatorname{rk}) consisting of a set S , a subset \mathcal{X} of \mathcal{P}(S)\setminus\{\emptyset\} , and a function \operatorname{rk}:\mathcal{X}\to\mathbb{Z}_{\geq 0} with the following properties: 1. For all s\in S , \{s\}\in\mathcal{X} , and 2. The function \operatorname{rk} is order-preserving, which means that if x,y\in\mathcal{X} satisfy x\subseteq y , then \operatorname{rk}(x)\leq\operatorname{rk}(y) . Elements of S are called entities or vertices , elements of \mathcal{X} are called relations or cells , and \operatorname{rk} is called the rank function of the CC. This definition was developed in [ 4 ] . When C is a combinatorial complex and we write C^{i} , we mean all cells in C of rank i . We can also interpret graphs as combinatorial complexes. Here, graph nodes are cells of rank 0 and edges are cells of rank 1 . Definition 2.2 . ( Isomorphism ) Two combinatorial complexes C=(S,\mathcal{X},\operatorname{rk}) and C^{\prime}=(S^{\prime},\mathcal{X}^{\prime},\operatorname{rk}^{\prime}) are said to be isomorphic if there exists a bijective map \phi:\mathcal{X}\to\mathcal{X}^{\prime} that preserves both the incidence structure and the rank function. Definition 2.3 . ( Higher order incidence matrix ) The incidence matrix represents the relationships between cells of different ranks. Given a combinatorial complex (S,\mathcal{X},\operatorname{rk}) , the incidence matrix encodes the boundary relationships between cells of consecutive ranks. Formally, let \mathcal{X}_{k} denote the set of cells in \mathcal{X} with rank k . The incidence matrix \delta_{k}\in\mathbb{R}^{|\mathcal{X}_{0}|\times|\mathcal{X}_{k}|} has each entry defined as follows: \delta_{k}(x,y)=\begin{cases}\pm 1,&\text{if }x\subset y,\\ 0,&\text{otherwise}.\end{cases} More specifically let x_{1},\dots x_{n}\subset y be all cells of y , then there exists exactly one x_{i} s.t. \delta_{k}(x_{i},y)=1 and \forall x_{j} with j\neq i \delta_{k}(x_{j},y)=-1 . This matrix can be interpreted as the discrete derivative operator at rank 0 for combinatorial complexes. Hodge Laplacian On cellular complexes, the Laplacian is known as the Hodge Laplacian [ 11 ] . It is defined as follows. For the next paragraph, let C be a cell complex. Let \partial_{k}:C^{k}\to C^{k-1} denote the boundary operator, which maps each k -cell to its (k{-}1) -dimensional faces. Its transpose, \partial_{k}^{\top} , is the coboundary operator —also interpretable as the incidence matrix from rank k to k{-}1 . The k -th Hodge Laplacian is then given by: \displaystyle\Delta_{k}=\partial_{k+1}\partial_{k+1}^{\top}+\partial_{k}^{\top}\partial_{k} (1) This operator acts on k -cochains (real-valued functions defined on k -cells) and captures both upward and downward adjacencies, reflecting how cells are connected through lower- and higher-dimensional neighbours. A more detailed definition is provided in [ 11 ] . While this definition extends naturally to combinatorial complexes , their general structure can limit the boundary operators’ ability to capture higher-order relationships fully. We illustrate this limitation in Corollary 8.2 , showing the restricted expressiveness of classical Laplacians in distinguishing specific structures. 3 Method To address this, we introduce a new approach that defines a Laplacian directly on combinatorial complexes. Next, we construct a heat kernel signature for each node (i.e., a 0-dimensional cell), which serves as the input embedding to our deep learning pipeline. We then describe the whole model setup, while the theoretical and computational benefits of our method are discussed in Section 4.2 . 3.1 Laplacian on Combinatorial complexes Laplacians are well-defined operators on graphs [ 12 ] and geometric shapes [ 13 ] . However, extending this notion to combinatorial complexes introduces several challenges not present in graphs or simplicial complexes: 1. Single-rank connectivity:

Read the complete text