Discrete Gauss–Bonnet Graph Attention
Implementation & benchmark of arXiv:2608.06405 — Elements of finite geometry I
Source paper: Elements of finite geometry I arXiv:2608.06405 ⓘ · analyzed Aug 31, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Compute each graph node's discrete curvature from the numbers of simplices in its neighbor-induced unit sphere, then inject this scalar into message-passing or attention logits. Add an optional topology-aware feature channel so that nodes with identical degree but different local clique structure receive different representations.
Formulas
Mathematical statement
For a finite graph or clique complex G, let f_k(G) be the number of k-dimensional simplices, where a k-simplex has k+1 vertices, and let the Euler characteristic be chi(G)=sum_{k>=0}(-1)^k f_k(G). For a vertex v, its unit sphere S(v) is the subgraph induced by the neighbors of v; f_k(S(v)) counts its k-dimensional simplices. If the maximal simplex dimension of G is q, the paper defines K(v)=1-sum_{k=0}^{q-1}((-1)^k f_k(S(v))/(k+1)). The discrete Gauss–Bonnet theorem states chi(G)=sum_{v in V}K(v). The factor 1/(k+1) distributes each k-simplex equally among its vertices, while the alternating signs reproduce the Euler valuation. The neural transfer uses K(v) as a fixed node feature or as an additive attention bias, preserving the exact local-to-global topological signal.
Implementation notes
Integrate this into a GAT or transformer-style graph message-passing layer at the attention-logit stage. For every input graph, construct its clique complex up to a small maximum dimension q-1, typically q=3 or q=4. Count vertices, edges, triangles, and optionally tetrahedra in each node's neighbor-induced subgraph. Compute K[v] using the paper's formula, then normalize it per graph as K_norm[v]=(K[v]-mean(K))/(std(K)+1e-6). Pseudocode: simplices=enumerate_cliques(G,max_dim=q-1); for v: sphere=induced_subgraph(neighbors(v)); K[v]=1; for k in range(q): K[v]-=(-1)**k*count_k_simplices(sphere)/(k+1); for edge(v,u): logit[v,u]=dot(Q[v],Kproj[u])/sqrt(d)+beta*(K_norm[v]+K_norm[u])/2; alpha[v]=softmax(logit[v,:]); hnew[v]=sum_u alpha[v,u]*V[u]. Use the exact curvature values as a frozen preprocessing feature rather than learning them. For large graphs, cap clique enumeration and estimate higher-dimensional counts by bounded random clique sampling, while retaining exact vertex and edge counts. Verify the preprocessing identity numerically with abs(sum(K)-chi)<1e-6, where chi is computed from the same capped simplex complex. First compare a two-layer curvature-biased GAT against a vanilla GAT and GCN on MUTAG, PROTEINS, and synthetic graphs whose labels depend on triangle density or Euler characteristic. Sweep beta in {0, 0.1, 0.3, 1}. Success means higher validation accuracy or better sample efficiency at equal parameter count and FLOPs, particularly on splits where degree statistics are matched but clique topology differs; also monitor attention entropy for instability.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.