Alessio Borgi

Alessio Borgi

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

Sheaf Cohomology: Sections, Cochains, and What H⁰ and H¹ Mean

7 minute read

Published:

TL;DR: The sheaf cochain complex 0 → C⁰ →^{δ₀} C¹ → 0 gives two cohomology groups: H⁰ = ker(δ₀) (global sections — consistent signals) and H¹ = C¹/im(δ₀) (obstruction — edge disagreements that cannot be explained by any node assignment). For graph learning: H⁰ is the attractor of sheaf diffusion; dim(H⁰) controls the long-range information retained; H¹ captures topological obstructions to global consistency.

The Cochain Complex

Given a cellular sheaf F on a graph G, the cochain complex is:

0 → C⁰(G, F) →^{δ₀} C¹(G, F) → 0

where:

  • C⁰ = ∏_{v∈V} F(v) ≅ ℝ^{Nd} — node-level data
  • C¹ = ∏_{e∈E} F(e) ≅ ℝ^{Ed} — edge-level data
  • δ₀ : C⁰ → C¹ is the coboundary operator: (δ₀x)e = F{v▷e}x_v − F_{u▷e}x_u

For a graph, this is a 2-term complex (there are no 2-cells). The cohomology groups are:

H⁰(G, F) = ker(δ₀) (zeroth cohomology = global sections) H¹(G, F) = C¹(G, F) / im(δ₀) (first cohomology = obstruction to consistency)

H⁰: Global Sections

H⁰(G, F) = ker(δ₀) is the vector space of global sections — node assignments x = (x_v) such that for every edge e = (u,v):

F_{u▷e} x_u = F_{v▷e} x_v

Dimension of H⁰: For a connected graph with trivial (identity) sheaf, dim(H⁰) = d — one d-dimensional constant function per component. For a sheaf with orthogonal maps and trivial holonomy, dim(H⁰) = d. For a sheaf with maps that have non-trivial kernel interactions, dim(H⁰) can be larger.

Euler characteristic: For a connected graph G:

χ(G, F) = dim H⁰ − dim H¹ = d(|V| − |E|) = d · (1 − |E| + |V| − 1) = d · χ(G)
where χ(G) =V−Eis the graph Euler characteristic (= 1 for trees, = 1−g for graphs with g independent cycles).

H¹: Obstruction to Global Consistency

H¹(G, F) = C¹(G, F)/im(δ₀) measures how far C¹ is from being “explained” by C⁰.

An element of C¹ is an assignment y = (y_e) of vectors to edges. y is in im(δ₀) if and only if there exists a node assignment x such that y_e = F_{v▷e}x_v − F_{u▷e}x_u — i.e., y is a “disagreement signal” that can be attributed to a global node assignment.

y ∈ H¹ means: y is an edge-level signal that cannot be explained by any node assignment. This is a topological obstruction — it exists because of cycles in the graph where the holonomy of the restriction maps is non-trivial.

Dimension of H¹:

dim H¹ = Ed − (Nd − dim H⁰) using rank-nullity on δ₀ = Ed − Nd + dim H⁰
For a connected graph: dim H¹ = d·E− d·V+ dim H⁰ = d·(E−V+1) + (dim H⁰ − d).
For trees:E=V−1, so dim H¹ = dim H⁰ − d. If the sheaf has no “extra” global sections (dim H⁰ = d), then dim H¹ = 0 — trees always have trivial H¹.

For graphs with cycles: dim H¹ ≥ d·(number of independent cycles).

Graph learning interpretation of H¹: A non-zero H¹ means the sheaf has "frustrated cycles" — closed paths where the composition of restriction maps is not the identity. In physics language, this is holonomy (the gauge field has non-zero curvature around loops). In practice, large H¹ means the graph has richer structure that cannot be encoded in node assignments alone — this is information that sheaf diffusion processes differently than GCN.

The Hodge Decomposition

For a cellular sheaf, the space of 1-cochains C¹ decomposes as:

C¹ = im(δ₀) ⊕ ker(δ₀ᵀ) ⊕ H¹(G, F)

Wait — for a 2-term complex there is no further differential. The Hodge decomposition is:

C¹ = im(δ₀) ⊕ ker(δ₀ᵀ)

where:

  • im(δ₀): exact 1-cochains — edge disagreements attributable to node assignments
  • ker(δ₀ᵀ): co-closed 1-cochains — edge signals that “don’t accumulate” at nodes

H¹ = ker(δ₀ᵀ) / (im(δ₀) ∩ ker(δ₀ᵀ)) = ker(δ₀ᵀ) when the complex has trivial overlap. In the Hodge sense, harmonic 1-cochains (in ker(δ₀ᵀ) and “orthogonal to” im(δ₀)) represent H¹.

For sheaves, the harmonic space is ker(Δ₁) where Δ₁ = δ₀δ₀ᵀ is the down-Laplacian on edges. A 1-cochain y is harmonic if δ₀ᵀ y = 0 and δ₀ is not defined here since C² = 0. So harmonic 1-cochains = ker(δ₀ᵀ).

Betti Numbers and Graph Topology

The Betti numbers of the sheaf are:

  • β₀ = dim H⁰ — number of “independent global sections”
  • β₁ = dim H¹ — dimension of the obstruction space

For the constant sheaf (all maps = identity, d=1):

  • β₀ = number of connected components of G
  • β₁ = number of independent cycles of G (first Betti number)

Sheaf cohomology generalises ordinary graph cohomology: the constant sheaf recovers the classical topological invariants.

Computing H⁰ in Practice

In a sheaf GNN, the space of global sections ker(δ₀) is the long-time attractor of the sheaf diffusion equation. Computing it exactly requires computing the null space of Δ_F — an (Nd)×(Nd) matrix — which is too expensive at scale.

In practice, sheaf GNNs approximate the projection onto ker(Δ_F) by:

  1. Running K steps of diffusion X ← (I − αΔ_F^{norm})X
  2. Adding skip connections to preserve information outside ker(Δ_F)
  3. Using the output of each layer (not just the final step) as features

The skip connections are the crucial ingredient: without them, only ker(Δ_F) information survives at large K.

Example: Triangle Graph with Signed Sheaf

Consider a triangle graph G: nodes {1,2,3}, edges {e₁₂, e₂₃, e₁₃}, stalk dimension d=1.

Signed sheaf: all restriction maps are ±1 scalars. Assign +1 to all except F_{3▷e₁₃} = −1.

The coboundary:

  • (δ₀x)_{e₁₂} = x₂ − x₁
  • (δ₀x)_{e₂₃} = x₃ − x₂
  • (δ₀x)_{e₁₃} = −x₃ − x₁

Global sections: x₂=x₁, x₃=x₂, −x₃=x₁ → x₁=x₂=x₃=−x₁ → x₁=0. So H⁰ = {0} — no nontrivial global sections. The sheaf is frustrated: there is no consistent assignment.

dim H¹ =E·d −V·d + dim H⁰ = 3−3+0 = 0. But wait — using χ: χ(G) = 3−3 = 0, so dim H⁰ − dim H¹ = 0 → dim H¹ = dim H⁰ = 0. The sheaf is cohomologically trivial, even though it has no global sections.

This example shows the subtlety: a sheaf can be frustrated (no global sections) while still having trivial H¹. The frustration doesn’t create H¹; rather, it is captured by H⁰ vanishing.

Why Cohomology Matters for GNNs

The dimension of H⁰ directly controls what information sheaf diffusion retains at large depth:

  • Large dim(H⁰): the model retains a rich subspace, enabling complex long-range representations
  • Small dim(H⁰) (e.g., 0): diffusion is contractive and discards most information

Learning restriction maps (as in NSD) means learning the dimension of H⁰ implicitly — the model adapts the global section space to the task. This is a fundamentally different approach from choosing a fixed aggregation kernel.

References

  • Bodnar, C., Giovanni, F. D., Chamberlain, B. P., Liò, P., & Bronstein, M. M. (2022). Neural Sheaf Diffusion. NeurIPS 2022 (uses the H⁰ null space structure to explain why sheaf diffusion avoids oversmoothing).
  • Ghrist, R. (2014). Elementary Applied Topology. Createspace (ch. 5–6 cover sheaf cohomology with graph examples and the Euler characteristic formula).
  • Robinson, M. (2014). Topological Signal Processing. Springer 2014 (ch. 3–4 develop the cochain complex and Hodge decomposition for cellular sheaves, with signal processing applications).

You May Also Enjoy

Open Problems and Future Directions in Sheaf Neural Networks

5 minute read

Published:

Sheaf Neural Networks are a young field with many open research problems. This post surveys the most important unsolved questions: scalability to million-node graphs, stochastic sheaves for uncertainty quantification, sheaf autoencoders for generative models, connections to persistent homology, and unification with other topological deep learning frameworks.

Sheaf GNNs for Molecular Property Prediction

7 minute read

Published:

Molecular graphs are naturally heterophilic — atoms of different types bonded together — and have rich geometric structure (bond angles, torsion angles). Sheaf GNNs can encode this structure via restriction maps that represent the relational geometry between bonded atoms. This post explores how sheaf theory applies to molecular property prediction.

Sheaf GNNs for Heterophilic Node Classification: Benchmarks and Results

6 minute read

Published:

A systematic comparison of sheaf GNNs against baselines on standard heterophilic node classification benchmarks. Which datasets benefit most from sheaf structure? How does stalk dimension d affect performance? When do sheaves win — and when don’t they?

Sheaves Meet Attention: Transformer-Inspired Sheaf Models

6 minute read

Published:

Graph Transformers use global attention to capture long-range dependencies that local message passing misses. Sheaf Transformers combine sheaf-based local aggregation (with learnable relational geometry) with global attention (for long-range dependencies). This post surveys the design space of sheaf-augmented attention models.