Categorithm

Unifying Sheaf and Equivariant Neural Networks

Two structural priors on graph neural networks may be one construction seen from two sides. A cellular-sheaf supplies shape, heterogeneous relations and a nontrivial notion of “agreeing across an edge”, whereas equivariance supplies symmetry, features that transform correctly under . The question is whether a single message-passing rule is equivariant by construction because its restriction maps are intertwiners.

The two priors

A sheaf neural network attaches a stalk, a vector space, to each node and edge of a graph, together with restriction maps that say how a node’s data is compared along an edge. Diffusion runs by the sheaf Laplacian. An -equivariant network instead attaches representations of the rotation group to each node and constrains every map to be an intertwiner, so the whole computation commutes with rigid motions.

Can the restriction maps be intertwiners?

Let each stalk carry a representation of (or ), and demand the sheaf’s restriction maps be equivariant, so that they are intertwiners between the stalk representations of the two incident cells. The sheaf Laplacian is then assembled from equivariant blocks, so its diffusion is equivariant by construction, with no extra symmetrisation. Sheaf networks would be the “shape” axis and equivariant networks the “symmetry” axis of a single object, an equivariant cellular sheaf.

What would have to be checked

Is the equivariant sheaf Laplacian still a Laplacian?

The good properties of the sheaf Laplacian, positive semidefiniteness and a harmonic space that reads off global consistency, rely on the restriction maps being linear over a fixed field. If they are instead intertwiners between distinct irreps, does the coboundary still compose to a self-adjoint operator, and does its kernel still mean “globally consistent and symmetric”? And does the construction subsume both parents as two degenerate limits, plain sheaf networks at trivial representations and plain equivariant GNNs at a discrete sheaf?

If it works, geometric molecular models built for rigid-motion alignment and relational models built on sheaves become instances of one design.

Recent work realises much of this unification. Maruyama’s order-equivariant neural networks put an equivariant bundle over a face poset, where the poset supplies locality and a group action supplies symmetry, and this single class generalises both ordinary graph message passing and sheaf networks. It characterises all linear order-equivariant maps, and extended to category-equivariant networks it reads equivariance as naturality. The same programme proves the first universal approximation theorem for sheaf networks, so the hard constraint turns out to be free in approximation power, and what it buys instead is per-edge transport.

The sharper checks above are then answered concretely for electronic structure. Harish’s equivariant cellular sheaf takes the restriction maps to be -steerable two-center kernels read off bond geometry, exactly the intertwiners asked for, and shows that the positive-semidefinite-shifted single-particle molecular Hamiltonian in a localised atomic-orbital basis is the Laplacian of that sheaf. The construction stays - and permutation-equivariant, strictly generalises -equivariant message passing and CW networks, and reads chemistry off the cohomology, since counts non-bonding orbitals while the Hodge -Laplacian captures ring and delocalisation structure. The equivariant sheaf Laplacian is therefore a genuine Laplacian whose kernel carries meaning, and what remains open is the correlated rung, an equivariant sheaf for the two-electron marginal rather than the single-particle Hamiltonian.

References

  • K. Harish, Equivariant Cellular Sheaves for Molecular Electronic Structure: Bridging Sheaf Cohomology and E(3)-Equivariant Hamiltonian Learning (arXiv:2608.23571).
  • Maruyama, order-equivariant neural networks (arXiv:2607.03798); category-equivariant neural networks (arXiv:2511.18417).
  • J. Hansen, T. Gebhart, Sheaf Neural Networks (arXiv:2012.06333); C. Bodnar et al., Neural Sheaf Diffusion (arXiv:2202.04579).
  • F. Barbero et al., Sheaf Neural Networks with Connection Laplacians (arXiv:2206.08702).