Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs

Cristian BodnarFrancesco Di GiovanniBenjamin Paul ChamberlainPietro LióMichael M. Bronstein

article2022NeurIPS299 citations

Proves that heterophily and oversmoothing in graph neural networks stem from an implicit trivial sheaf assumption, proposing neural sheaf diffusion models that learn custom edge-node linear maps from data to achieve superior node classification on non-homophilic graphs.

Listen

Graph Neural Networks are widely used to analyze networked and relational data across biology, chemistry, and the social sciences. However, standard models suffer from two major performance bottlenecks: severe degradation on heterophilic networks (where connected nodes possess different labels or attributes) and oversmoothing in deeper architectures (where repeated feature aggregation makes node representations indistinguishable). The article addresses these core challenges by demonstrating that both issues stem from a shared root cause: the implicit assumption of an overly simplistic, trivial geometry across graph connections.

The article's main objective is to establish a rigorous mathematical foundation using cellular sheaf theory—a branch of algebraic topology—to explain why standard graph networks fail under heterophily and oversmoothing, and to develop practical neural network models that overcome these constraints by dynamically learning the underlying graph geometry from data.

To achieve this, the authors model graphs using cellular sheaves, which assign dedicated vector spaces (stalks) to nodes and edges and directional linear maps (restriction maps) between them. This construction replaces standard graph diffusion with sheaf diffusion governed by a generalized sheaf Laplacian operator. The researchers established theoretical guarantees regarding the expressive power and asymptotic classification behavior of sheaf diffusion across increasingly expressive sheaf classes (symmetric, non-symmetric, diagonal, and orthogonal). They then designed Neural Sheaf Diffusion models, which parameterize and learn edge transformation matrices end-to-end via neural networks, and evaluated them against extensive baselines on synthetic benchmarks and nine real-world datasets spanning high heterophily to high homophily across standardized evaluation splits.

The theoretical and empirical analyses yielded several key findings. First, standard graph models implicitly rely on trivial, symmetric sheaves, which mathematically guarantees that inter-class node features collapse into indistinguishable representations in heterophilic or multi-class settings. Second, incorporating asymmetric or negative transformation maps enables diffusion processes to linearly separate opposing classes in the infinite time limit, directly preventing oversmoothing. Third, solving classification problems with three or more classes fundamentally requires stalk dimensions greater than one, with diagonal and orthogonal sheaves offering linear separation across multiple classes. Fourth, unlike standard graph convolutions that contract energy and force feature smoothing, non-symmetric sheaf networks retain the mathematical flexibility to increase energy and escape degenerate kernel states. Finally, across real-world benchmarks, the proposed Neural Sheaf Diffusion architectures achieved top performance on five out of six highly heterophilic datasets and ranked among the top three models on eight of the nine evaluated benchmarks, all while remaining within approximately 1% of the top-performing models on homophilic graphs.

These findings carry important operational and architectural implications. Rather than relying on specialized heuristics, ad hoc negative edge weight adjustments, or structural workarounds to counter heterophily, organizations can adopt unified sheaf-based architectures that dynamically discover the proper relational geometry. By tuning the stalk dimension to modest sizes between 1 and 5, sheaf diffusion introduces only a small constant computational overhead relative to standard graph convolutions, avoiding substantial training cost increases while offering superior predictive reliability.

Based on these results, engineering teams deploying graph learning on complex, heterophilic relational data should prioritize adopting neural sheaf architectures—particularly orthogonal bundles, which provide the best balance of regularized parameter complexity, numerical stability, and representational capacity. When implementing these models, practitioners should adjust stalk dimensions according to the expected number of target classes. Future work should focus on exploring how higher-order sheaf Laplacians and advanced topological message-passing frameworks can further improve model expressiveness and scaling on massive enterprise graphs.

The findings are supported by solid mathematical proofs and reproducible empirical benchmarks across multiple random seeds. Nonetheless, users should consider key limitations: the formal separation theorems analyze infinite-time diffusion limits rather than non-asymptotic finite-layer regimes, the general unrestricted matrix parameterization can introduce numerical instabilities during matrix normalizations, and deep learning generalization bounds for learned sheaves remain an open theoretical challenge. Overall, confidence in the demonstrated performance advantages on heterophilic networks remains very high.

Twitter University of Cambridge University of Oxford cellular diffusion equations diffusion probabilistic models Neural Sheaf Diffusion cellular sheaf theory heterophilic graphs

arXiv: 2202.04579 twitter-research/neural-sheaf-diffusion

Paper: Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs, Jiong Zhu et al. (2020). This paper establishes the foundational analysis of why standard GNNs fail under network heterophily and provides the standardized benchmark datasets used to evaluate neural sheaf diffusion.
Paper: Geom-GCN: Geometric Graph Convolutional Networks, Hongbin Pei et al. (2020). It introduces geometric aggregation mechanisms to address disassortative and heterophilic graphs, providing key conceptual motivation and benchmark datasets for sheaf-based geometric learning.
Paper: Deeper Insights into Graph Convolutional Networks for Semi-Supervised Learning, Qimai Li et al. (2018). This study rigorously characterizes the graph convolution operation as Laplacian smoothing, mathematically framing the oversmoothing problem that neural sheaf diffusion is designed to solve.
Paper: Measuring and Relieving the Over-smoothing Problem for Graph Neural Networks from the Topological View, Deli Chen et al. (2019). It provides a topological perspective and quantitative distance metrics (MAD/MADGap) for analyzing how network structure induces oversmoothing in deep GNNs.
Paper: Simple and Deep Graph Convolutional Networks, Ming Chen et al. (2020). This work analyzes the depth constraints of graph convolutions and proposes residual architectures to overcome oversmoothing, serving as a primary deep learning baseline.
Paper: Geometric Deep Learning: Going beyond Euclidean data, Michael M. Bronstein et al. (2016). It formalizes the geometric deep learning paradigm that unifies spectral and spatial graph convolutions, which sheaf diffusion extends via algebraic topology.
Paper: Semi-Supervised Classification with Graph Convolutional Networks, Thomas N. Kipf et al. (2017). This foundational text introduces the standard Graph Convolutional Network and its implicit diffusion operator, whose trivial geometry neural sheaf diffusion generalizes.

Paper: Graph-Coupled Oscillator Networks, T. Konstantin Rusch et al. (2022). This paper extends dynamic continuous-time approaches to deep graph learning by formulating GNNs as systems of coupled oscillators to prevent oversmoothing and gradient collapse.
Paper: Finding Global Homophily in Graph Neural Networks When Meeting Heterophily, Xiang Li et al. (2022). It offers an alternative global aggregation framework for tackling severe heterophily by finding distant homophilous nodes across the network in linear time.
Paper: A Generalization of ViT/MLP-Mixer to Graphs, Xiaoxin He et al. (2023). This work tackles over-squashing and long-range interactions on complex graphs by scaling patch-based Mixer architectures with linear computational complexity.
Paper: Benchmarking Graph Neural Networks, Vijay Prakash Dwivedi et al. (2023). It establishes a standardized medium-scale benchmarking suite to evaluate advanced anisotropic and topological message-passing frameworks under controlled parameter budgets.

Cover for Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs

Abstract

Cellular sheaves equip graphs with a “geometrical” structure by assigning vector spaces and linear maps to nodes and edges. Graph Neural Networks (GNNs) implicitly assume a graph with a trivial underlying sheaf. This choice is reflected in the structure of the graph Laplacian operator, the properties of the associated diffusion equation, and the characteristics of the convolutional models that discretise this equation. In this paper, we use cellular sheaf theory to show that the underlying geometry of the graph is deeply linked with the performance of GNNs in heterophilic settings and their oversmoothing behaviour. By considering a hierarchy of increasingly general sheaves, we study how the ability of the sheaf diffusion process to achieve linear separation of the classes in the infinite time limit expands. At the same time, we prove that when the sheaf is non-trivial, discretised parametric diffusion processes have greater control than GNNs over their asymptotic behaviour. On the practical side, we study how sheaves can be learned from data. The resulting sheaf diffusion models have many desirable properties that address the limitations of classical graph diffusion equations (and corresponding GNN models) and obtain competitive results in heterophilic settings. Overall, our work provides new connections between GNNs and algebraic topology and would be of interest to both fields.

Topics

cellular diffusion equations diffusion probabilistic models geometry graph neural networks topological structure numerical solution convolutional codes bipartite graphs

Keywords

Neural Sheaf Diffusion cellular sheaf theory heterophilic graphs sheaf Laplacians Sheaf Convolutional Network oversmoothing in GNNs sheaf learning restriction maps algebraic topology

1 Introduction
2 Background
3 The Expressive Power of Sheaf Diffusion
3.1 Harmonic Space of Sheaf Laplacians
3.2 The Linear Separation Power of Sheaf Diffusion
4 Expressive Power of Sheaf Convolutions
5 Neural Sheaf Diffusion and Sheaf Learning
6 Experiments
7 Related Work, Discussion, and Conclusion
Acknowledgments and Disclosure of Funding
References
Checklist

Knowls

Knowl 1 — Neural Sheaf Diffusion and Discrete Residual Layer
model/method
Neural Sheaf Diffusion (NSD) models diffusion on an undirected graph $G=(V, E)$ with $|V|=n$ nodes, where each node stalk is assigned a vector space $\mathbb{R}^d$ and restriction maps evolve over time based on node representations.

In continuous time, for a feature matrix $\mathbf{X}(t) \in \mathbb{R}^{nd \times f}$ representing $f$ feature channels across all node stalks, the continuous diffusion PDE is: $\dot{\mathbf{X}}(t) = -\sigma\left(\Delta_{\mathcal{F}(t)} (\mathbf{I}_n \otimes \mathbf{W}_1) \mathbf{X}(t) \mathbf{W}_2\right)$ where $\Delta_{\mathcal{F}(t)}$ is the normalized sheaf Laplacian of a time-evolving cellular sheaf $(G, \mathcal{F}(t))$ computed from the data via $(G, \mathcal{F}(t)) = g(G, \mathbf{X}(t); \theta)$ , $\mathbf{W}_1 \in \mathbb{R}^{d \times d}$ mixes stalk dimensions, $\mathbf{W}_2 \in \mathbb{R}^{f \times f}$ transforms feature channels, $\mathbf{I}_n$ is the $n \times n$ identity matrix, $\otimes$ denotes the Kronecker product, and $\sigma$ is a non-linear activation function.

The time-discretized residual layer parameterized at step $t$ is: $\mathbf{X}_{t+1} = \mathbf{X}_t - \sigma\left(\Delta_{\mathcal{F}(t)} (\mathbf{I}_n \otimes \mathbf{W}_1^t) \mathbf{X}_t \mathbf{W}_2^t\right)$ where $\mathbf{W}_1^t \in \mathbb{R}^{d \times d}$ and $\mathbf{W}_2^t \in \mathbb{R}^{f_t \times f_{t+1}}$ are layer-specific trainable weight matrices. Raw input node features are first projected by an MLP and reshaped to $\mathbf{X}_0 \in \mathbb{R}^{nd \times f}$ , and a final linear layer performs node-level classification on the output representation.
Knowl 2 — Sheaf Learning Parametrisations for Neural Sheaf Diffusion
model/method
In Neural Sheaf Diffusion, restriction maps of the sheaf are learned end-to-end from local node representations. For each directed edge pair $(v, u) \in E$ belonging to edge $e = (v, u)$ , a parametric function $\Phi$ predicts the $d \times d$ restriction matrix $\mathbf{F}_{v \trianglelefteq e} = \Phi(\mathbf{x}_v, \mathbf{x}_u)$ , where $\mathbf{x}_v, \mathbf{x}_u \in \mathbb{R}^{df}$ stack all $f$ channels of node stalks. The function is parameterized as $\Phi(\mathbf{x}_v, \mathbf{x}_u) = \sigma(\mathbf{V} [\mathbf{x}_v \,\|\, \mathbf{x}_u])$ followed by output reshaping, where $[\cdot \,\|\, \cdot]$ denotes concatenation, $\mathbf{V}$ is a learnable matrix, and $\sigma$ is an activation function.

Three structural variants of $\Phi$ are used:
1. Diagonal Sheaves (Diag-NSD): $\mathbf{F}_{v \trianglelefteq e}$ is constrained to be a diagonal matrix, $\mathbf{F}_{v \trianglelefteq e} = \text{diag}(\mathbf{w}_{v \trianglelefteq e})$ with $\mathbf{w}_{v \trianglelefteq e} \in \mathbb{R}^d$ . This produces a sheaf Laplacian with diagonal blocks, requiring fewer parameters and faster sparse matrix multiplications.
2. Orthogonal Sheaves / Discrete Bundles (O(d)-NSD): $\mathbf{F}_{v \trianglelefteq e} \in \text{O}(d)$ , where orthogonal matrices are parameterized via compositions of Householder reflections: $\mathbf{H}(\mathbf{u}) = \mathbf{I}_d - 2 \frac{\mathbf{u} \mathbf{u}^\top}{\|\mathbf{u}\|_2^2}$ Orthogonality prevents overfitting, couples stalk dimensions, and ensures degree normalization blocks satisfy $D_{vv} = d_v \mathbf{I}_d$ .
3. General Sheaves (Gen-NSD): $\mathbf{F}_{v \trianglelefteq e} \in \mathbb{R}^{d \times d}$ is an unconstrained matrix, providing maximum geometric flexibility at the cost of computing $D^{-1/2}$ via singular value decomposition (SVD).
Knowl 3 — Cellular Sheaf, Sheaf Laplacian, and Sheaf Dirichlet Energy
definition
Let $G = (V, E)$ be an undirected graph. A cellular sheaf $(G, \mathcal{F})$ assigns a vector space (stalk) $\mathcal{F}(v)$ to each node $v \in V$ , a vector space $\mathcal{F}(e)$ to each edge $e \in E$ , and a linear restriction map $\mathcal{F}_{v \trianglelefteq e}: \mathcal{F}(v) \to \mathcal{F}(e)$ for each incident node-edge pair $v \trianglelefteq e$ .

When all node stalks have fixed dimension $d$ , the space of $0$ -cochains is $C^0(G; \mathcal{F}) := \bigoplus_{v \in V} \mathcal{F}(v) \cong \mathbb{R}^{nd}$ , and a $0$ -cochain $\mathbf{x} \in C^0(G; \mathcal{F})$ is the concatenation of stalk vectors $\mathbf{x}_v \in \mathcal{F}(v)$ . The sheaf Laplacian $L_{\mathcal{F}}: C^0(G; \mathcal{F}) \to C^0(G; \mathcal{F})$ is an $nd \times nd$ positive semi-definite block matrix defined node-wise by: $(L_{\mathcal{F}} \mathbf{x})_v := \sum_{u: (v, u) \in E} \mathcal{F}_{v \trianglelefteq e}^\top (\mathcal{F}_{v \trianglelefteq e} \mathbf{x}_v - \mathcal{F}_{u \trianglelefteq e} \mathbf{x}_u)$ Its diagonal block at node $v$ is $L_{\mathcal{F} vv} = \sum_{v \trianglelefteq e} \mathcal{F}_{v \trianglelefteq e}^\top \mathcal{F}_{v \trianglelefteq e}$ , and the off-diagonal block for edge $(v, u) = e$ is $L_{\mathcal{F} vu} = -\mathcal{F}_{v \trianglelefteq e}^\top \mathcal{F}_{u \trianglelefteq e}$ . Let $D$ be the block-diagonal matrix of $L_{\mathcal{F}}$ . The normalized sheaf Laplacian is $\Delta_{\mathcal{F}} := D^{-1/2} L_{\mathcal{F}} D^{-1/2}$ .

The space of global sections $H^0(G; \mathcal{F}) := \{\mathbf{x} \in C^0(G; \mathcal{F}) : \mathcal{F}_{v \trianglelefteq e} \mathbf{x}_v = \mathcal{F}_{u \trianglelefteq e} \mathbf{x}_u, \forall (v, u) = e \in E\}$ is isomorphic to the harmonic space $\ker(L_{\mathcal{F}})$ . For a signal $\mathbf{x} \in \mathbb{R}^{nd}$ (or multi-channel $\mathbf{X} \in \mathbb{R}^{nd \times f}$ ), the sheaf Dirichlet energy is: $\mathcal{E}_{\mathcal{F}}(\mathbf{x}) := \mathbf{x}^\top \Delta_{\mathcal{F}} \mathbf{x} = \frac{1}{2} \sum_{(v, u) = e \in E} \|\mathcal{F}_{v \trianglelefteq e} D_{vv}^{-1/2} \mathbf{x}_v - \mathcal{F}_{u \trianglelefteq e} D_{uu}^{-1/2} \mathbf{x}_u\|_2^2$ and $\mathcal{E}_{\mathcal{F}}(\mathbf{X}) := \text{trace}(\mathbf{X}^\top \Delta_{\mathcal{F}} \mathbf{X})$ , with $\mathcal{E}_{\mathcal{F}}(\mathbf{X}) = 0 \iff \mathbf{X} \in \ker(\Delta_{\mathcal{F}})$ .
Knowl 4 — Expressive Power and Multi-Class Linear Separation via Stalk Dimension Hierarchy
theoretical result
A hypothesis class of sheaves with $d$ -dimensional stalks $\mathcal{H}^d$ has linear separation power over a family of graphs $\mathcal{G}$ if, for any labeled graph $G = (V, E) \in \mathcal{G}$ , there exists a sheaf $(G, \mathcal{F}) \in \mathcal{H}^d$ such that continuous sheaf diffusion $\dot{\mathbf{X}}(t) = -\Delta_{\mathcal{F}} \mathbf{X}(t)$ linearly separates the classes of $G$ as $t \to \infty$ for almost all initial conditions $\mathbf{X}(0) \in \mathbb{R}^{nd \times f}$ .

The separation capacity depends directly on stalk dimension $d$ :
1. Fundamental 1D Multi-Class Limitation: If $G$ is a connected graph with $C \ge 3$ classes, the hypothesis class $\mathcal{H}^1 := \{(G, \mathcal{F}) : \det(\mathcal{F}_{v \trianglelefteq e}) \neq 0\}$ of 1-dimensional invertible sheaves cannot linearly separate the classes of $G$ for any initial conditions $\mathbf{X}(0) \in \mathbb{R}^{n \times f}$ , because $\dim(\ker(\Delta_{\mathcal{F}})) \le 1$ regardless of the channel count $f$ .
2. Diagonal Invertible Sheaves ( $\mathcal{H}^d_{\text{diag}}$ ): For diagonal invertible sheaves $\mathcal{H}^d_{\text{diag}} := \{(G, \mathcal{F}) : \mathcal{F}_{v \trianglelefteq e} \text{ is diagonal}, \det(\mathcal{F}_{v \trianglelefteq e}) \neq 0\}$ , if stalk dimension $d \ge C$ , then $\mathcal{H}^d_{\text{diag}}$ has linear separation power over all connected graphs with $C \ge 3$ classes.
3. Orthogonal Sheaves / Discrete Bundles ( $\mathcal{H}^d_{\text{orth}}$ ): For orthogonal sheaves $\mathcal{H}^d_{\text{orth}} := \{(G, \mathcal{F}) : \mathcal{F}_{v \trianglelefteq e} \in \text{O}(d)\}$ , for stalk dimensions $d \in \{2, 4\}$ , $\mathcal{H}^d_{\text{orth}}$ has linear separation power over all connected graphs with $C \le 2d$ classes.
Knowl 5 — Linear Separation and Transport Polarization in Binary Heterophilic Graphs
theoretical result
Consider continuous sheaf diffusion $\dot{\mathbf{X}}(t) = -\Delta_{\mathcal{F}} \mathbf{X}(t)$ on a connected graph $G = (V, E)$ partitioned into two classes $A, B \subset V$ :
1. Symmetric Invertible Sheaves ( $\mathcal{H}^1_{\text{sym}}$ ): For sheaves where $\mathcal{F}_{v \trianglelefteq e} = \mathcal{F}_{u \trianglelefteq e} \in \mathbb{R} \setminus \{0\}$ (which includes standard and strictly positive-weighted graph Laplacians):
  - If every node in $A$ has at least one neighbor in $A$ , $\mathcal{H}^1_{\text{sym}}$ has linear separation power over $G$ .
  - If $G = (A, B, E)$ is a connected bipartite graph with $|A| = |B|$ , $\mathcal{H}^1_{\text{sym}}$ cannot linearly separate the two classes for any initial condition $\mathbf{X}(0) \in \mathbb{R}^{n \times f}$ in the infinite time limit.
2. Non-Symmetric Invertible Sheaves ( $\mathcal{H}^1$ ): Allowing asymmetric restriction maps $\mathcal{F}_{v \trianglelefteq e} = -\alpha_e$ for $v \in A$ and $\mathcal{F}_{u \trianglelefteq e} = \alpha_e$ for $u \in B$ with $\alpha_e > 0$ yields a discrete $\text{O}(1)$ -bundle with edge transport $\mathcal{F}_{v \trianglelefteq e}^\top \mathcal{F}_{u \trianglelefteq e} = -1$ across inter-class edges and $+1$ across intra-class edges. Sheaf diffusion under this sheaf linearly separates $A$ and $B$ for almost all initial conditions, giving $\mathcal{H}^1$ linear separation power over all binary connected graphs by polarising class representations into opposite signs.
Knowl 6 — Dirichlet Energy Contraction and Kernel Escape in Sheaf Convolutional Networks
theoretical result
Let $\mathbf{Y} = \sigma\left((\mathbf{I}_{nd} - \Delta_{\mathcal{F}})(\mathbf{I}_n \otimes \mathbf{W}_1)\mathbf{X}\mathbf{W}_2\right) \in \mathbb{R}^{nd \times f_2}$ be a Sheaf Convolutional Network (SCN) layer where $\mathbf{X} \in \mathbb{R}^{nd \times f_1}$ , $\mathbf{W}_1 \in \mathbb{R}^{d \times d}$ , $\mathbf{W}_2 \in \mathbb{R}^{f_1 \times f_2}$ , and $\sigma$ is ReLU or LeakyReLU. Define $\lambda^* := \max_{i: \lambda_i^{\mathcal{F}} > 0} (\lambda_i^{\mathcal{F}} - 1)^2 \le 1$ , where $\lambda_i^{\mathcal{F}}$ are eigenvalues of the normalized sheaf Laplacian $\Delta_{\mathcal{F}}$ .
1. Dirichlet Energy Contraction: For positive 1D sheaves $(G, \mathcal{F}) \in \mathcal{H}^1_+ := \{(G, \mathcal{F}) : \mathcal{F}_{v \trianglelefteq e} \mathcal{F}_{u \trianglelefteq e} > 0\}$ (including standard and positive-weighted graph Laplacians), and for symmetric orthogonal bundles $(G, \mathcal{F}) \in \mathcal{H}^d_{\text{orth,sym}} := \mathcal{H}^d_{\text{orth}} \cap \mathcal{H}^d_{\text{sym}}$ , the sheaf Dirichlet energy contracts according to: $\mathcal{E}_{\mathcal{F}}(\mathbf{Y}) \le \lambda^* \|\mathbf{W}_1\|_2^2 \|\mathbf{W}_2^\top\|_2^2 \mathcal{E}_{\mathcal{F}}(\mathbf{X})$ When $\lambda^* \|\mathbf{W}_1\|_2^2 \|\mathbf{W}_2^\top\|_2^2 < 1$ , representations exponentially collapse into $\ker(\Delta_{\mathcal{F}})$ , causing oversmoothing.
2. Kernel Escape: For any connected graph $G$ and any $\varepsilon > 0$ , there exists a non-symmetric sheaf $(G, \mathcal{F}) \notin \mathcal{H}^d_{\text{sym}}$ , a weight matrix $\mathbf{W}_1$ with $\|\mathbf{W}_1\|_2 < \varepsilon$ , and a feature vector $\mathbf{x}$ such that: $\mathcal{E}_{\mathcal{F}}((\mathbf{I}_n \otimes \mathbf{W}_1)\mathbf{x}) > \mathcal{E}_{\mathcal{F}}(\mathbf{x})$ Thus, SCNs with non-symmetric sheaves can increase Dirichlet energy and escape the Laplacian kernel even when using low-norm weights.
Knowl 7 — Spectral Gap and Cheeger-type Bounds for Discrete Vector Bundles
theoretical result
Let $\mathcal{F}$ be a discrete $\text{O}(d)$ bundle (where restriction maps $\mathcal{F}_{v \trianglelefteq e} \in \text{O}(d)$ are orthogonal matrices) over a connected undirected graph $G = (V, E)$ with $n$ nodes, diameter $\text{diam}(G)$ , and maximum degree $d_{\max}$ . Transport along a path $\gamma_{v \to u} = (v, v_1, \dots, v_\ell, u)$ is defined as: $\mathbf{P}_{\gamma_{v \to u}} := (\mathcal{F}_{u \trianglelefteq (v_\ell, u)}^\top \mathcal{F}_{v_\ell \trianglelefteq (v_\ell, u)}) \dots (\mathcal{F}_{v_1 \trianglelefteq (v, v_1)}^\top \mathcal{F}_{v \trianglelefteq (v, v_1)}): \mathcal{F}(v) \to \mathcal{F}(u)$ The smallest eigenvalue $\lambda_0^{\mathcal{F}}$ of $\Delta_{\mathcal{F}}$ and the space of global sections $H^0(G; \mathcal{F})$ satisfy:
1. Spectral Gap Upper Bound: If $r := \max_{\gamma_{v \to u}, \gamma'_{v \to u}} \|\mathbf{P}_{\gamma_{v \to u}} - \mathbf{P}_{\gamma'_{v \to u}}\|$ , then $\lambda_0^{\mathcal{F}} \le r^2 / 2$ . When transport is path-independent ( $r = 0$ ), $\lambda_0^{\mathcal{F}} = 0$ .
2. Cycle Consistency: If $\mathbf{x} \in H^0(G; \mathcal{F})$ , then for any cycle $\gamma_{v \to v}$ based at node $v$ , $\mathbf{x}_v \in \ker(\mathbf{P}_{\gamma_{v \to v}} - \mathbf{I}_d)$ .
3. Cheeger-type Lower Bound: If $\|(\mathbf{P}_{\gamma_{v \to v}} - \mathbf{I}_d)\mathbf{x}_v\| \ge \epsilon \|\mathbf{x}_v\|$ for all cycles $\gamma_{v \to v}$ , then: $\lambda_0^{\mathcal{F}} \ge \frac{\epsilon^2}{2 \text{diam}(G) n d_{\max}}$
4. Dimension of Harmonic Space: $\dim(H^0(G; \mathcal{F})) \le d$ , and $\dim(H^0(G; \mathcal{F})) = d$ if and only if transport across $G$ is path-independent.
Knowl 8 — Universal Approximation of Cellular Sheaves via Local MLPs
theoretical result
Let $G = (V, E)$ be a finite undirected graph with node feature matrix $\mathbf{X}$ assigning vector $\mathbf{x}_v$ to each node $v \in V$ . If all ordered incident node-pair feature vectors are distinct, meaning $(\mathbf{x}_v, \mathbf{x}_u) \neq (\mathbf{x}_w, \mathbf{x}_z)$ for all distinct directed edges $(v, u) \neq (w, z) \in E$ , and $\Phi: \mathbb{R}^{2df} \to \mathbb{R}^{d \times d}$ is a multilayer perceptron (MLP) with sufficient capacity, then $\Phi$ can learn any arbitrary cellular sheaf $(G, \mathcal{F})$ over $G$ by assigning $\mathcal{F}_{v \trianglelefteq e} = \Phi(\mathbf{x}_v, \mathbf{x}_u)$ for every incident pair $v \trianglelefteq e = (v, u)$ .
Knowl 9 — Computational Complexity of Neural Sheaf Diffusion
theoretical result
Let $G = (V, E)$ be an undirected graph with $n$ nodes and $m$ edges. Let $c = d \times f$ denote the total node representation size, where $d$ is the stalk dimension and $f$ is the number of feature channels.
1. A standard Graph Convolutional Network (GCN) layer on representations of dimension $c$ has computational complexity $\mathcal{O}(n c^2 + m c)$ .
2. A Neural Sheaf Diffusion layer with diagonal restriction maps (Diag-NSD) has computational complexity $\mathcal{O}(n c^2 + m d c)$ .
3. A Neural Sheaf Diffusion layer with orthogonal ( $\text{O}(d)$ -NSD) or general (Gen-NSD) restriction maps has computational complexity $\mathcal{O}(n(c^2 + d^3) + m(c d^2 + d^3))$ .
For typical stalk dimensions $1 \le d \le 5$ , the computational overhead of NSD compared to standard GCN is a small constant factor.

Knowl 10 — Node Classification Performance on Heterophilic and Homophilic Benchmarks

data/table

Neural Sheaf Diffusion models (Diag-NSD, $\text{O}(d)$ -NSD, and Gen-NSD) were evaluated across nine node classification benchmarks with edge homophily levels $h$ ranging from 0.11 (highly heterophilic) to 0.81 (highly homophilic). Evaluation was performed over 10 fixed random splits (48%/32%/20% train/val/test per class), reporting mean classification accuracy (%) and standard deviation.

Dataset	Texas	Wisconsin	Film	Squirrel	Chameleon	Cornell	Citeseer	Pubmed	Cora
Hom. level ( $h$ )	0.11	0.21	0.22	0.22	0.23	0.30	0.74	0.80	0.81
#Nodes	183	251	7,600	5,201	2,277	183	3,327	18,717	2,708
#Edges	295	466	26,752	198,493	31,421	280	4,676	44,327	5,278
#Classes	5	5	5	5	5	5	7	3	6
Diag-NSD	85.676.95	88.632.75	37.791.01	54.781.81	68.681.73	86.497.35	77.141.85	89.420.43	87.141.06
O( $d$ )-NSD	85.955.51	89.414.74	37.811.15	56.341.32	68.041.58	84.864.71	76.701.57	89.490.40	86.901.13
Gen-NSD	82.975.13	89.213.84	37.801.22	53.171.31	67.931.58	85.686.51	76.321.65	89.330.35	87.301.15
GGCN	84.864.55	86.863.29	37.541.56	55.171.58	71.141.84	85.686.63	77.141.45	89.150.37	87.951.05
H2GCN	84.867.23	87.654.98	35.701.00	36.481.86	60.112.15	82.705.28	77.111.57	89.490.38	87.871.20
GPRGNN	78.384.36	82.944.21	34.631.22	31.611.24	46.581.71	80.278.11	77.131.67	87.540.38	87.951.18
FAGCN	82.436.89	82.947.95	34.871.25	42.590.79	55.223.19	79.199.79	N/A	N/A	N/A
MixHop	77.847.73	75.884.90	32.222.34	43.801.48	60.502.53	73.516.34	76.261.33	85.310.61	87.610.85
GCNII	77.573.83	80.393.40	37.441.30	38.471.58	63.863.04	77.863.79	77.331.48	90.150.43	88.371.25
Geom-GCN	66.762.72	64.513.66	31.591.15	38.150.92	60.002.81	60.543.67	78.021.15	89.950.47	85.351.57
PairNorm	60.274.34	48.436.14	27.401.24	50.442.04	62.742.82	58.923.15	73.591.47	87.530.44	85.791.01
GraphSAGE	82.436.14	81.185.56	34.230.99	41.610.74	58.731.68	75.955.01	76.041.30	88.450.50	86.901.04
GCN	55.145.16	51.763.06	27.321.10	53.432.01	64.822.24	60.545.30	76.501.36	88.420.50	86.981.27
GAT	52.166.63	49.414.09	27.440.89	40.721.55	60.262.50	61.895.05	76.551.23	87.301.10	86.330.48
MLP	80.814.75	85.293.31	36.530.70	28.771.56	46.212.99	81.896.40	74.021.90	87.160.37	75.692.00

NSD models rank first on 5 out of 6 heterophilic benchmarks ( $h < 0.3$ ) and rank in the top three on 8 of the 9 benchmarks overall, while remaining within 1% of top performance on homophilic graphs ( $h \ge 0.74$ ). Among the NSD variants, $\text{O}(d)$ -NSD performs best overall due to orthogonal constraints preventing overfitting while allowing non-trivial multi-dimensional feature mixing.

Coverage note — No substantial contributed material was omitted from the paper.

References

1.Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Hrayr Harutyunyan, Nazanin Alipourfard, Kristina Lerman, Greg Ver Steeg, and Aram Galstyan. Mixhop: Higher-order graph convo- lutional architectures via sparsified neighborhood mixing. In The Thirty-sixth International Conference on Machine Learning (ICML), 2019. URL http://proceedings.mlr.press/ v97/abu-el-haija19a/abu-el-haija19a.pdf.
2.Afonso S Bandeira, Amit Singer, and Daniel A Spielman. A Cheeger inequality for the graph connection laplacian. SIAM Journal on Matrix Analysis and Applications, 34(4):1611–1630, 2013.
3.Federico Barbero, Cristian Bodnar, Haitz Sáez de Ocáriz Borde, Michael Bronstein, Petar Veličković, and Pietro Liò. Sheaf neural networks with connection laplacians. In ICML 2022 Workshop on Topology, Algebra, and Geometry in Machine Learning, 2022.
4.Federico Barbero, Cristian Bodnar, Haitz Sáez de Ocáriz Borde, and Pietro Lio. Sheaf attention networks. In NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations, 2022.
5.Lukas Biewald. Experiment tracking with weights and biases, 2020. URL https://www. wandb.com/. Software available from wandb.com.
6.Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. ISBN 0387310738.
7.Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. In AAAI. AAAI Press, 2021.
8.Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yu Guang Wang, Pietro Liò, Guido Montúfar, and Michael Bronstein. Weisfeiler and Lehman Go Cellular: CW Networks. In NeurIPS, 2021.
9.Cristian Bodnar, Fabrizio Frasca, Yuguang Wang, Nina Otter, Guido F Montufar, Pietro Lió, and Michael Bronstein. Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks. In ICML, 2021.
10.Glen E Bredon. Sheaf theory, volume 170. Springer Science & Business Media, 2012.
11.Marc Brockschmidt. Gnn-film: Graph neural networks with feature-wise linear modulation. In International Conference on Machine Learning, pages 1144–1152. PMLR, 2020.
12.Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. In ICLR, 2014.
13.Chen Cai and Yusu Wang. A note on over-smoothing for graph neural networks. arXiv:2006.13318, 2020.
14.Benjamin Paul Chamberlain, James Rowbottom, Davide Eynard, Francesco Di Giovanni, Dong Xiaowen, and Michael M Bronstein. Beltrami flow and neural diffusion on graphs. In NeurIPS, 2021.
15.Benjamin Paul Chamberlain, James Rowbottom, Maria Goronova, Stefan Webb, Emanuele Rossi, and Michael M Bronstein. Grand: Graph neural diffusion. In ICML, 2021.
16.Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1725–1735. PMLR, 13–18 Jul 2020. URL https://proceedings. mlr.press/v119/chen20v.html.
17.Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=n6jl7fLxrP.
18.Justin Michael Curry. Sheaves, cosheaves and applications. University of Pennsylvania, 2014.
19.Pim de Haan, Taco Cohen, and Max Welling. Natural graph networks. In NeurIPS, 2020.
20.Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NIPS, 2016.
21.Lun Du, Xiaozhou Shi, Qiang Fu, Xiaojun Ma, Hengyu Liu, Shi Han, and Dongmei Zhang. Gbk-gnn: Gated bi-kernel graph neural networks for modeling both homophily and heterophily. In Proceedings of the ACM Web Conference 2022, pages 1550–1558, 2022.
22.Andrew Dudzik and Petar Veličković. Graph neural networks are dynamic programmers. arXiv preprint arXiv:2203.15544, 2022.
23.Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. arXiv:2003.00982, 2020.
24.Tingran Gao, Jacek Brodzki, and Sayan Mukherjee. The geometry of synchronization problems and learning group actions. Discrete & Computational Geometry, 65(1):150–211, 2021.
25.Robert Ghrist and Hans Riess. Cellular sheaves of lattices and the tarski laplacian. arXiv:2007.04099, 2020.
26.Robert W Ghrist. Elementary applied topology, volume 1. Createspace Seattle, 2014.
27.Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICML, 2017.
28.Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by backpropagation through structure. In ICNN, 1996.
29.Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In IJCNN, 2005.
30.Pim De Haan, Maurice Weiler, Taco Cohen, and Max Welling. Gauge equivariant mesh CNNs: Anisotropic convolutions on geometric graphs. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=Jnspzp-oIZE.
31.William L Hamilton, Rex Ying, and Jure Leskovec. Representation learning on graphs: Methods and applications. IEEE Data Engineering Bulletin, 2017.
32.Jakob Hansen and Thomas Gebhart. Sheaf neural networks. In NeurIPS 2020 Workshop on Topological Data Analysis and Beyond, 2020.
33.Jakob Hansen and Robert Ghrist. Learning sheaf laplacians from smooth signals. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5446–5450. IEEE, 2019.
34.Jakob Hansen and Robert Ghrist. Toward a spectral theory of cellular sheaves. Journal of Applied and Computational Topology, 3(4):315–358, 2019.
35.Jakob Hansen and Robert Ghrist. Opinion dynamics on discourse sheaves. SIAM Journal on Applied Mathematics, 81(5):2033–2060, 2021.
36.Kurt Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks, 4(2):251–257, 1991.
37.Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359–366, 1989.
38.Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.
39.Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017.
40.Hyun Deok Lee. On some matrix inequalities. Korean Journal of Mathematics, 16(4):565–571, 2008.
41.Vijay Lingam, Rahul Ragesh, Arun Iyer, and Sundararajan Sellamanickam. Simple truncated svd based model for node classification on heterophilic graphs. arXiv preprint arXiv:2106.12807, 2021.
42.Sitao Luan, Chenqing Hua, Qincheng Lu, Jiaqi Zhu, Mingde Zhao, Shuyuan Zhang, Xiao-Wen Chang, and Doina Precup. Is heterophily a real nightmare for graph neural networks to do node classification? arXiv preprint arXiv:2109.05641, 2021.
43.Saunders Mac Lane. Categories for the working mathematician, volume 5. Springer Science & Business Media, 2013.
44.Saunders MacLane and Ieke Moerdijk. Sheaves in geometry and logic: A first introduction to topos theory. Springer Science & Business Media, 2012.
45.Zakaria Mhammedi, Andrew Hellicar, Ashfaqur Rahman, and James Bailey. Efficient orthogonal parametrisation of recurrent neural networks using householder reflections. In International Conference on Machine Learning, pages 2401–2409. PMLR, 2017.
46.Christopher Morris, Martin Ritzert, Matthias Fey, William L Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and Leman go neural: Higher-order graph neural networks. In AAAI, 2019.
47.Galileo Namata, Ben London, Lise Getoor, Bert Huang, and U Edu. Query-driven active surveying for collective classification. In 10th International Workshop on Mining and Learning with Graphs, volume 8, page 1, 2012.
48.H Nt and T Maehara. Revisiting graph neural networks: all we have is low pass filters. arXiv preprint arXiv:1812.08434v4, 2019.
49.Anton Obukhov. Efficient householder transformation in pytorch, 2021. URL www.github. com/toshas/torch-householder. Version: 1.0.1, DOI: 10.5281/zenodo.5068733.
50.K Oono and T Suzuki. Graph neural networks exponentially lose expressive power for node classification. In International Conference on Learning Representations, 2020.
51.Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. arXiv:1905.10947, 2019.
52.John Palowitch, Anton Tsitsulin, Brandon Mayer, and Bryan Perozzi. Graphworld: Fake graphs bring real insights for gnns. arXiv preprint arXiv:2203.00112, 2022.
53.Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. arXiv preprint arXiv:2002.05287, 2020.
54.Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.
55.David Pfau, Irina Higgins, Alex Botev, and Sébastien Racanière. Disentangling by subspace diffusion. Advances in Neural Information Processing Systems, 33:17403–17415, 2020.
56.Daniel Rosiak. Sheaf theory through examples. MIT Press, 2022.
57.Benedek Rozemberczki, Carl Allen, and Rik Sarkar. Multi-scale attributed node embedding. Journal of Complex Networks, 9(2):cnab014, 2021.
58.Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Trans. Neural Networks, 20(1):61–80, 2008.
59.Richard D Schafer. An introduction to nonassociative algebras. Courier Dover Publications, 2017.
60.Luis Scoccola and Jose A Perea. Approximate and discrete euclidean vector bundles. arXiv preprint arXiv:2104.07563, 2021.
61.Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi- Rad. Collective classification in network data. AI magazine, 29(3):93–93, 2008.
62.Allen Dudley Shepard. A cellular description of the derived category of a stratified space. PhD thesis, Brown University, 1985.
63.Amit Singer and H-T Wu. Vector diffusion maps and the connection laplacian. Communications on pure and applied mathematics, 65(8):1067–1144, 2012.
64.Alessandro Sperduti. Encoding labeled graphs by labeling RAAM. In NIPS, 1994.
65.Julian Suk, Lorenzo Giusti, Tamir Hemo, Miguel Lopez, Konstantinos Barmpas, and Cristian Bodnar. Surfing on the neural sheaf. In NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations, 2022.
66.Jie Tang, Jimeng Sun, Chi Wang, and Zi Yang. Social influence analysis in large-scale networks. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 807–816, 2009.
67.Loring W Tu. Manifolds. In An Introduction to Manifolds, pages 47–83. Springer, 2011.
68.Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In ICLR, 2018.
69.Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr, Christopher Fifty, Tao Yu, and Kilian Q Weinberger. Simplifying graph convolutional networks. In ICML, 2019.
70.Louis-Pascal Xhonneux, Meng Qu, and Jian Tang. Continuous graph neural networks. In ICML, 2020.
71.Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In ICLR, 2019.
72.Yujun Yan, Milad Hashemi, Kevin Swersky, Yaoqing Yang, and Danai Koutra. Two sides of the same coin: Heterophily and oversmoothing in graph convolutional neural networks. arXiv:2102.06462, 2021.
73.Fouad El Zein and Jawad Snoussi. Local systems and constructible sheaves. In Arrangements, local systems and singularities, pages 111–153. Springer, 2009.
74.Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum? id=rkecl1rtwB.
75.Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances in Neural Information Processing Systems, 33, 2020.

Citation

MLA

Bodnar, C., et al. “Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs”. Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 18527–41, https://proceedings.neurips.cc/paper_files/paper/2022/file/75c45fca2aa416ada062b26cc4fb7641-Paper-Conference.pdf.

APA

Bodnar, C., Di Giovanni, F., Chamberlain, B., Lió, P., & Bronstein, M. (2022). Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs. Advances in Neural Information Processing Systems, 35, 18527–18541. https://proceedings.neurips.cc/paper_files/paper/2022/file/75c45fca2aa416ada062b26cc4fb7641-Paper-Conference.pdf

Chicago

Bodnar, C., F. Di Giovanni, B. Chamberlain, P. Lió, and M. Bronstein. 2022. “Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs”. Advances in Neural Information Processing Systems 35: 18527–41. https://proceedings.neurips.cc/paper_files/paper/2022/file/75c45fca2aa416ada062b26cc4fb7641-Paper-Conference.pdf.

Harvard

Bodnar, C. et al. (2022) “Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs”, Advances in Neural Information Processing Systems. Curran Associates, Inc., pp. 18527–18541. Available at: https://proceedings.neurips.cc/paper_files/paper/2022/file/75c45fca2aa416ada062b26cc4fb7641-Paper-Conference.pdf.

Vancouver

1. Bodnar C, Di Giovanni F, Chamberlain B, Lió P, Bronstein M (2022) Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs. In: Advances in Neural Information Processing Systems. Curran Associates, Inc., pp 18527–18541

BibTeX

@inproceedings{bodnar2022neural,
  title = {Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs},
  author = {Bodnar, Cristian and Di Giovanni, Francesco and Chamberlain, Benjamin and Lió, Pietro and Bronstein, Michael},
  year = {2022},
  booktitle = {Advances in Neural Information Processing Systems},
  publisher = {Curran Associates, Inc.},
  volume = {35},
  pages = {18527-18541},
  url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/75c45fca2aa416ada062b26cc4fb7641-Paper-Conference.pdf}
}

Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors

Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Neural Sheaf Diffusion and Discrete Residual Layer

Knowl 2 — Sheaf Learning Parametrisations for Neural Sheaf Diffusion

Knowl 3 — Cellular Sheaf, Sheaf Laplacian, and Sheaf Dirichlet Energy

Knowl 4 — Expressive Power and Multi-Class Linear Separation via Stalk Dimension Hierarchy

Knowl 5 — Linear Separation and Transport Polarization in Binary Heterophilic Graphs

Knowl 6 — Dirichlet Energy Contraction and Kernel Escape in Sheaf Convolutional Networks

Knowl 7 — Spectral Gap and Cheeger-type Bounds for Discrete Vector Bundles

Knowl 8 — Universal Approximation of Cellular Sheaves via Local MLPs