Alessio Borgi

Alessio Borgi

PhD in Graph Neural Networks and Generative AI 👨🏻‍💻🤖

Sheaf Neural Networks: A Complete Research Guide

9 minute read

Published:

What this book covers: Standard GNNs average a node with its neighbours, which quietly assumes every node measures the world in the same units. Sheaf Neural Networks drop that assumption: each node gets its own vector space, each edge gets a learned linear map between them, and comparison happens only after transport. One operator — the sheaf Laplacian — then covers heterophily, directional structure, and a principled account of oversmoothing.
Two nodes v and u, each with a vector-space stalk, connected through a shared edge stalk by restriction maps and their transposes
The whole object in one picture: nodes v and u carry stalks F(v) and F(u), the edge carries F(e), and the restriction maps F(v◁e) and F(u◁e) transport vectors into the shared edge space. Φ is the learned function that produces those maps from node features (Bodnar et al., 2022).

The big idea

Do not force neighbouring nodes to be equal. Learn how they should be related, through a linear map attached to each edge.

Why it matters

That single change gives one language for heterophily, signed and directional relations, and diffusion richer than a plain graph Laplacian allows.

Where the book is

The paper chapters are live now. The foundations and theory chapters are still being written, so this overview carries the maths you need to read them.

The problem, stated precisely

A GCN layer propagates with \(\hat{A} = \tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}\), which is an averaging operator. Averaging has a fixed point, and repeated averaging converges to it: features collapse towards a single degree-scaled direction. That is oversmoothing, and it is not a bug in the implementation — it is what averaging does.

It also encodes an assumption. Adding \(h_u\) to \(h_v\) only means something if the two vectors are expressed in the same basis. On a heterophilic graph, where an edge signals difference rather than similarity, that assumption is actively wrong.

Intuition first — the weather-station analogy. Picture a network of weather stations. Each measures temperature, but in its own units: some Celsius, some Fahrenheit, some a proprietary scale. Knowing that two adjacent stations read \(22\) and \(71\) tells you nothing until you know the conversion between them. That conversion is the restriction map. A global section is an assignment of readings — one per station — on which every adjacent pair agrees after conversion. The sheaf Laplacian is the penalty for disagreement measured in those converted units. Sheaf neural networks learn the conversions from data. That is the entire conceptual leap.
A restriction map as a unit conversion between two weather stationsNode u reads 22 degrees Celsius and node v reads 71.6 degrees Fahrenheit. Both restriction maps carry those readings into a shared edge space, where both equal 295.15 kelvin, so the two stations agree. A restriction map is a unit conversion node u 22 °C node v 71.6 °F 295.15 K shared edge space transport from u transport from v The readings look nothing alike — 22 against 71.6 — yet they are the same temperature. Agreement is checked in the shared space, never in the local units.
Two stations disagree numerically and agree physically. A sheaf makes that distinction first-class: the restriction maps are the conversions, and the sheaf Laplacian measures disagreement only after they have been applied. Sheaf GNNs learn the conversions.

The core object

A cellular sheaf \(\mathcal{F}\) on a graph \(G = (V, E)\) assigns:

  • a node stalk \(\mathcal{F}(v) \cong \mathbb{R}^{d}\) to each \(v \in V\);
  • an edge stalk \(\mathcal{F}(e) \cong \mathbb{R}^{d}\) to each \(e \in E\);
  • a restriction map \(\mathcal{F}_{v \trianglelefteq e} : \mathcal{F}(v) \to \mathcal{F}(e)\) for each incident pair.

The coboundary \(\delta_0\) measures disagreement across an edge, after transport:

\[ (\delta_0 x)_e \;=\; \mathcal{F}_{v \trianglelefteq e}\, x_v \;-\; \mathcal{F}_{u \trianglelefteq e}\, x_u , \qquad e = (u,v). \]

The sheaf Laplacian is \(\Delta_{\mathcal{F}} = \delta_0^{\top}\delta_0\), a block matrix with

\[ (\Delta_{\mathcal{F}})_{vv} = \sum_{e \ni v} \mathcal{F}_{v \trianglelefteq e}^{\top}\mathcal{F}_{v \trianglelefteq e}, \qquad (\Delta_{\mathcal{F}})_{uv} = -\,\mathcal{F}_{u \trianglelefteq e}^{\top}\mathcal{F}_{v \trianglelefteq e}. \]

Because it is built as \(\delta_0^{\top}\delta_0\), it is symmetric and positive semi-definite, and its kernel is exactly the space of global sections. Set every restriction map to the identity and it collapses to the familiar case:

\[ \Delta_{\mathcal{F}} \;=\; L \otimes I_d , \qquad L = D - A . \]

So the graph Laplacian is the sheaf Laplacian of the trivial sheaf. Everything a GCN does, a sheaf GNN can do by choosing identity maps — and it has \(d \times d\) more room per edge when identity is the wrong choice.

The mental shift: a sheaf GNN never asks whether neighbours are similar. It asks how one neighbour should be transported into another's frame before comparison at all. Keep that sentence and the rest of the book follows.

The pipeline, end to end

Definitions settle nothing until you watch a signal move through them. Here is one full step, from raw node features to a diffusion update.

The sheaf neural network pipeline in five stagesRaw scalar node features are lifted to two-dimensional stalks; each stalk is transported into the shared edge space by its restriction map; the difference of the transported vectors is the coboundary on that edge; the coboundary is pulled back to the nodes to give the diffusion update. From raw features to one diffusion step ① raw features ② sheaf lifting ③ transport ④ coboundary ⑤ diffusion u v x⁠ᵤ = 1 x⁠ᵥ = 2 one scalar per node 1 0 0 2 Xᵤ Xᵥ each node → a stalk in ℝ² d = 2 here 1 0 −2 0 Fᵤ Xᵤ Fᵥ Xᵥ both land in ℱ(e) Fᵥ rotates by 90° −3 0 (δ₀X)ₑ disagreement measured after transport 3,0 0,3 pulled back to nodes X ← X − Δ⁠₣ X the same signal, carried through every stage
One diffusion step on a two-node graph with stalk dimension d = 2. The restriction map at u is the identity and at v is a 90° rotation, so the two nodes genuinely disagree once both are expressed in the edge's frame.

Every number above is checkable. With \(\mathcal{F}_{u \trianglelefteq e} = I\) and \(\mathcal{F}_{v \trianglelefteq e} = R(90^\circ)\):

\[ \begin{aligned} \mathcal{F}_{u \trianglelefteq e} X_u &= \begin{pmatrix} 1 \\ 0 \end{pmatrix}, &\qquad \mathcal{F}_{v \trianglelefteq e} X_v &= \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}\!\begin{pmatrix} 0 \\ 2 \end{pmatrix} = \begin{pmatrix} -2 \\ 0 \end{pmatrix}, \\[4pt] (\delta_0 X)_e &= \begin{pmatrix} -3 \\ 0 \end{pmatrix}, &\qquad (\Delta_{\mathcal{F}} X)_u &= \begin{pmatrix} 3 \\ 0 \end{pmatrix}, \quad (\Delta_{\mathcal{F}} X)_v = \begin{pmatrix} 0 \\ 3 \end{pmatrix}. \end{aligned} \]

The resulting \(\Delta_{\mathcal{F}}\) has eigenvalues \(\{0, 0, 2, 2\}\), so its kernel is two-dimensional: on this graph there is a whole plane of assignments the sheaf considers globally consistent. A graph Laplacian on two connected nodes has a one-dimensional kernel — the constants. That gap is the extra room sheaf diffusion has to work in.

What changes, concretely

 Standard GCNSheaf GNN
Operatorgraph Laplacian \(L\)sheaf Laplacian \(\Delta_{\mathcal{F}}\)
Edge weightone scalara \(d \times d\) linear map
Propagation\(H \leftarrow \hat{A}H\)\(H \leftarrow (I - \Delta_{\mathcal{F}})H\)
Depth limitcollapses to a constant directionconverges to \(\ker(\Delta_{\mathcal{F}})\), which can separate classes
Heterophilydestructive averagingsigned or rotating maps make disagreement meaningful

The fourth row is the theoretical heart. Oversmoothing is not avoided by making the operator weaker; it is avoided by making its null space richer. A graph Laplacian has an essentially one-dimensional-per-component null space. A sheaf Laplacian’s null space is the space of global sections, and with the right restriction maps that space is large enough to hold a separating assignment.

The whole story in one paragraph

A sheaf equips every node and edge with a vector space and every incidence with a linear map; the sheaf Laplacian then measures inconsistency after transporting signals through those maps. Accept that one formulation and a set of separately-studied GNN problems start to look like one problem: heterophily, sign structure, gauge symmetry, directional flow, and the null-space account of oversmoothing.

What is live now

The paper chapters are published; the foundations and theory chapters are still in draft. If you are arriving new, read them in this order:

Suggested reading order

  1. PolyNSD — replaces the fixed diffusion step with a Chebyshev polynomial in \(\Delta_{\mathcal{F}}\). The clearest entry point to what the operator actually does.
  2. DNSD — why sheaf diffusion stalls at depth, and what it costs to fix. Read this second: it is the sharpest critique of the framework from inside it.
  3. HetSheaf — sheaves on heterogeneous graphs, where node and edge types condition the restriction maps.
  4. SheafPool — graph-level readout that respects the sheaf's basis ambiguity.
  5. BrainDyn — sheaves inside a neural ODE, applied to brain dynamics. The most different thing in the book.

Still to come: the foundations run (what a sheaf is, cellular sheaves on graphs, cohomology, the Laplacian spectrum, connection Laplacians), the theory run (oversmoothing, heterophily, oversquashing, expressiveness, Hodge decomposition), and the extensions run (cosheaves, simplicial sheaves, multi-relational and temporal sheaves).

Key papers at a glance

Hansen & Gebhart (2020) — Sheaf Neural Networks. The first sheaf GNN. Restriction maps are fixed by hand rather than learned, which makes it the cleanest illustration of what the operator does on its own.
Bodnar et al. (2022) — Neural Sheaf Diffusion. NeurIPS 2022. Learns the restriction maps from node features, and proves the null-space result that makes sheaves a genuine answer to heterophily and oversmoothing rather than a heuristic.
Barbero et al. (2022) — Sheaf Attention Networks. NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations. Combines restriction maps with attention-weighted aggregation.
Borgi, Silvestri & Liò (2025) — Polynomial Neural Sheaf Diffusion. A degree-\(K\) Chebyshev polynomial in the normalised sheaf Laplacian, with diagonal restriction maps sufficing.
Bourgerie, Girdzijauskas & Fodor (2026) — Deep Neural Sheaf Diffusion. Diagnoses why the Laplacian's signal vanishes with depth and replaces it with a sheaf adjacency operator.
One honest caveat. Sheaf GNNs are not a free win. The extra machinery costs parameters and compute, the theory that motivates them applies to the linear diffusion rather than the trained network, and on several small heterophilic benchmarks a plain MPNN still wins. The DNSD chapter works through a concrete case where the published tables and the appendix tell different stories.

References

You May Also Enjoy

The Standard Library Modules Worth Knowing Before You pip install

8 minute read

Published:

Python ships with a counter, a queue, a memoiser, a CLI parser, a test framework’s smaller cousin, and a JSON-safe date library already installed — most of what people reach for a package to solve is one import away.

Files and Context Managers: Why with Is Not Optional

9 minute read

Published:

A file object that falls out of scope without with does get closed eventually — but ‘eventually’ means whenever the garbage collector gets around to it, which is not a promise any program handling more than a handful of files can live with.

Errors and Exceptions: EAFP, the Hierarchy, and Reading a Traceback

9 minute read

Published:

Exceptions in Python are not exceptional. They are a normal control-flow mechanism that the language leans on so heavily that the idiomatic style is to try the operation and handle the failure, rather than check first — which is faster, shorter, and free of race conditions.