The Sheaf Laplacian: Spectral Theory for Sheaves
Published:
Constructing the Sheaf Laplacian
Intuition First: The standard graph Laplacian penalises adjacent nodes for being different (it is the gradient flow of \(\sum (x_u - x_v)^2\)). The sheaf Laplacian instead penalises adjacent nodes for being inconsistent after transformation (the flow of \(\sum \lVert \mathcal{F}_{v \trianglelefteq e} x_v - \mathcal{F}_{u \trianglelefteq e} x_u \rVert^2\)). With identity maps, these are the same. With learned maps, “consistent” can mean “opposite in a structured way”, which is exactly what heterophilic graphs need.
Given a cellular sheaf \(\mathcal{F}\) on a graph \(G\) with coboundary \(\delta\), the sheaf Laplacian is
This is positive semi-definite, since \(x^{\top} L_{\mathcal{F}} x = \lVert \delta x \rVert^2 \ge 0\) for all \(x\).
Node-wise it reads
which says: measure the disagreement at each incident edge, pull it back into \(v\)’s stalk, and sum.
Block structure: \(L_{\mathcal{F}}\) is a block matrix indexed by nodes. For a graph with \(n\) nodes each having stalk \(\mathbb{R}^d\), \(L_{\mathcal{F}} \in \mathbb{R}^{nd \times nd}\). The blocks are:
Note \((L_{\mathcal{F}})_{uv} = (L_{\mathcal{F}})_{vu}^{\top}\), so the matrix is symmetric as it must be.
Connection to the Standard Graph Laplacian
The graph Laplacian is exactly the special case in which every stalk is \(\mathbb{R}\) and every restriction map is the identity:
so \(L_{\mathcal{F}} = D - A = L\), the ordinary \(n \times n\) graph Laplacian. If we keep the identity maps but widen the stalks to \(\mathbb{R}^d\), we instead get
the Kronecker product that applies the same scalar Laplacian independently in each of the \(d\) coordinates. Widening the stalks alone buys nothing; it is the non-trivial restriction maps that “twist” the off-diagonal blocks and change how features from different nodes interact.
The Sheaf Dirichlet Energy
The quadratic form
measures the total sheaf disagreement over the graph, how much the restriction maps disagree across all edges when applied to the signal \(x\).
- \(E_{\mathcal{F}}(x) = 0 \iff x \in H^0(G;\mathcal{F})\) (perfect consistency)
- \(E_{\mathcal{F}}(x)\) large \(\iff x\) has large disagreement at many edges
(The same quantity is usually written with the normalised Laplacian, \(E_{\mathcal{F}}(x) = x^{\top}\Delta_{\mathcal{F}} x\), which inserts a \(D_v^{-1/2}\) factor at each endpoint; the characterisation of the zero set is unchanged.)
Spectral view: eigenvectors of \(L_{\mathcal{F}}\) with small eigenvalues correspond to signals with low sheaf Dirichlet energy, near-consistent signals. Diffusion with \(L_{\mathcal{F}}\) drives signals toward its kernel.
Sheaf Diffusion
The heat equation on the sheaf is
whose Euler discretisation with unit step is
This is sheaf diffusion, the generalisation of graph heat diffusion. At each step, each node’s features are updated using the sheaf-weighted contributions of its neighbours.
In practice one uses the normalised sheaf Laplacian, whose spectrum is bounded:
and the update \(X \leftarrow (I - \Delta_{\mathcal{F}}) X\), which is the sheaf analogue of the normalised GCN propagation \(\hat{A} = D^{-1/2} A D^{-1/2}\).
Worked Example: 2-Node Sheaf Laplacian
Setup: two nodes \(u, v\) connected by one edge \(e\), stalks \(\mathbb{R}^2\), restriction maps
Diagonal blocks: \((L_{\mathcal{F}})_{uu} = I^{\top} I = I\) and \((L_{\mathcal{F}})_{vv} = R^{\top} R = I\) (both maps are orthogonal).
Off-diagonal blocks:
Full \(4 \times 4\) sheaf Laplacian (ordering the coordinates \(u_1, u_2, v_1, v_2\)):
Global sections (kernel): \(\delta x = 0\) means \(R x_v - x_u = 0\), i.e. \(x_u = R x_v\). Since \(R\) is invertible, any \(x_v \in \mathbb{R}^2\) extends to a section, so
The kernel is not trivial, and it could not be. A single edge is a tree, so there are no cycles around which transport could fail to close, and for an \(O(d)\)-bundle over a connected graph one has \(\dim H^0 \le d\) with equality exactly when transport is path-independent (Bodnar et al., 2022, Lemma 6). To get a trivial harmonic space you need a cycle whose holonomy has no fixed vector, as in the triangle example of the previous post.
Spectral Properties
Kernel: \(L_{\mathcal{F}} x = 0 \iff \delta x = 0 \iff x\) is a global section. So
and the number of zero eigenvalues equals the dimension of the space of global sections.
For the trivial sheaf with \(\mathbb{R}^d\) stalks on a graph with \(c\) connected components, \(\dim \ker L_{\mathcal{F}} = c\,d\): one \(d\)-dimensional constant per component. With \(d = 1\) this is the familiar “one constant per component” of the graph Laplacian.
For a non-trivial sheaf the kernel can be smaller, and, this is the whole point, it can consist of signals that are not constant. For an \(O(d)\)-bundle the ceiling is \(\dim H^0 \le d\), reached exactly when the transport around every cycle is the identity.
Spectral gap: the smallest non-zero eigenvalue of \(L_{\mathcal{F}}\) determines how fast sheaf diffusion converges to the harmonic space. For \(O(d)\)-bundles this gap is controlled by how far the transport maps are from being path-independent.
Normalised Sheaf Laplacian Spectrum
The eigenvalues of \(\Delta_{\mathcal{F}}\) lie in \([0, 2]\):
- \(0\): global sections (consistent signals)
- close to \(0\): nearly consistent signals
- close to \(2\): maximally inconsistent signals
This is the same range as the standard normalised Laplacian. The difference: with non-trivial restriction maps, “consistency” is defined relative to the sheaf structure, not raw feature equality.
Summary
| Quantity | Formula | Interpretation |
|---|---|---|
| Coboundary | \((\delta x)_e = \mathcal{F}_{v \trianglelefteq e} x_v - \mathcal{F}_{u \trianglelefteq e} x_u\) | Sheaf disagreement at edge \(e\) |
| Sheaf Laplacian | \(L_{\mathcal{F}} = \delta^{\top}\delta\) | Total disagreement operator |
| Dirichlet energy | \(E_{\mathcal{F}}(x) = x^{\top} L_{\mathcal{F}} x\) | Total inconsistency of signal \(x\) |
| Harmonic space | \(\ker L_{\mathcal{F}} = H^0(G;\mathcal{F})\) | Global sections (consistent signals) |
| Diffusion step | \(X \leftarrow (I - \Delta_{\mathcal{F}}) X\) | Reduces inconsistency; generalises GCN |
The sheaf Laplacian is the central object for sheaf-based graph learning. It generalises the standard graph Laplacian by incorporating edge-level structure, making it possible to define diffusion that respects per-edge feature transformations rather than forcing raw feature equality.
References
- Hansen, J., & Ghrist, R. (2019). Toward a Spectral Theory of Cellular Sheaves. Journal of Applied and Computational Topology (defines \(L_{\mathcal{F}} = \delta^{\top}\delta\), its harmonic space, and the spectral bounds used above).
- Hansen, J., & Gebhart, T. (2020). Sheaf Neural Networks. NeurIPS 2020 GRL+ Workshop (first sheaf convolutional layer built on the sheaf Laplacian; shows the trivial sheaf recovers GCN).
- Bodnar, C., Di Giovanni, F., Chamberlain, B. P., Liò, P., & Bronstein, M. M. (2022). Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNNs. NeurIPS 2022 (analyses the harmonic space of \(\Delta_{\mathcal{F}}\), including the bound \(\dim H^0 \le d\) for \(O(d)\)-bundles and its link to path-independent transport).
- Ebli, S., Defferrard, M., & Spreemann, G. (2020). Simplicial Neural Networks. NeurIPS 2020 TDA & Beyond Workshop (related Hodge Laplacian approach on simplicial complexes, providing topological context).