# S3-Holonomy Message Passing

- ID: 75
- Canonical URL: https://synthcore.org/idea/75/s3-holonomy-message-passing
- API JSON: https://synthcore.org/api/idea/75.json
- API Markdown: https://synthcore.org/api/idea/75.md
- Verification status: unverified
- Source: [arXiv:2608.22814](https://arxiv.org/abs/2608.22814)
- Category: architecture
- Solves: accuracy, generalization, stability
- ML areas: graph-nn, embedding, regularization
- Math tags: graph-theory, topology, algebra, representation-theory
- Ratings: usefulness 5/10; difficulty 5/10; novelty 6/10

## Idea description

Build a graph neural network on the dual graph of a triangulated surface whose messages are transported by \(\mathfrak{S}_3\) permutation matrices associated with adjacent-face color transports. This removes dependence on arbitrary local color-label choices and gives the network an explicit representation of noncontractible topology through holonomy around cycles.

## Mathematical statement

For a triangulation \(G\), let \(f_0,f_1,\ldots,f_k\) be a sequence of adjacent triangular faces and let \(V(f_i)\) denote the three vertices of face \(f_i\). Starting from an initial face coloring \(c:V(f_0)\to\{1,2,3\}\), define \(\sigma_{G,W_0,f_0}=\mathrm{id}\), and recursively choose \(\sigma_{G,W_i,f_0}\in\mathfrak{S}_3\) so that the transported coloring agrees with the previous transport on the vertices shared by \(f_{i-1}\) and \(f_i\). For a closed walk \(W\), the resulting permutation is the monodromy. The paper states \(\sigma_{G,W^{-1},f}=\sigma_{G,W,f}^{-1}\), homotopic closed walks have equal monodromy, and every contractible closed walk has identity monodromy. Represent each directed dual edge \(e\) by a permutation matrix \(P_e\in\mathbb{R}^{3\times3}\), with \(P_{e^{-1}}=P_e^{-1}\). For a cycle \(C=(e_1,\ldots,e_m)\), its holonomy is \(P_C=P_{e_m}\cdots P_{e_1}\). A feature with a three-color channel transforms using the natural representation \(P_e\), so transported message passing is equivariant to local relabelings.

## Key formulas

- $$\sigma_{G,W_0,f}=\mathrm{id},\qquad \sigma_{G,W_i,f}\big|_{V(f_{i-1})\cap V(f_i)}=\sigma_{G,W_{i-1},f}\big|_{V(f_{i-1})\cap V(f_i)}.$$
- $$\sigma_{G,W^{-1},f}=\bigl(\sigma_{G,W,f}\bigr)^{-1},\qquad W_1^*\simeq W_2^*\ \Longrightarrow\ \sigma_{G,W_1,f}=\sigma_{G,W_2,f},\qquad W\text{ contractible}\ \Longrightarrow\ \sigma_{G,W,f}=\mathrm{id}.$$
- $$P_C=P_{e_m}P_{e_{m-1}}\cdots P_{e_1},\qquad P_{e^{-1}}=P_e^{-1},\qquad m_f^{(\ell)}=\sum_{g\in N(f)}\phi_\ell\!\left((P_{g\to f}\otimes I_d)h_g^{(\ell)}\right).$$
- $$\mathcal{L}_{\mathrm{cycle}}=\sum_{C\in\mathcal{C}_{\mathrm{contractible}}}\left\|P_C-I_3\right\|_F^2+\lambda\sum_{C\in\mathcal{C}_{\mathrm{known}}}\left\|P_C-P_C^{\mathrm{target}}\right\|_F^2.$

## Implementation notes

Integrate this into a face-level or vertex-level GNN for triangulated meshes, surface graphs, or synthetic graphs embedded on planar, toroidal, and higher-genus surfaces. Construct the dual graph: each triangular face is a node and each shared primal edge is a directed dual edge. Assign each face an ordering of its three vertices and compute the local three-coloring. For an edge \(f\to g\), identify the two shared vertices and choose the unique permutation \(T_{f\to g}\in S_3\) mapping their colors in face \(f\) to their colors in face \(g\); store its permutation matrix \(P_{f\to g}\). Set the reverse matrix to \(P_{g\to f}=P_{f\to g}^{-1}\). At layer \(\ell\), use `msg_f=0; for g in neighbors(f): msg_f += MLP_l((P[g,f] kron I_d) @ h_g); h_f=Update_l(h_f,msg_f)`. If features have only scalar channels, concatenate transported color-channel features with ordinary invariant features. Sample random global and local \(S_3\) relabelings during training. Explicitly multiply matrices around short dual cycles and add \(\mathcal{L}_{cycle}\); contractible cycles must target \(I_3\), while noncontractible cycle targets can be retained as topology descriptors rather than forced to identity. If colorings are unavailable, predict one of six permutations per directed edge with a categorical softmax, tie reverse-edge logits through inverse permutation, and train with the cycle loss. Start with a two-layer GNN on synthetic triangulations whose labels are invariant to color names. Compare a GCN or GAT, the same network with raw color IDs, and the holonomy model at equal parameter count. Evaluate accuracy after random relabeling, cross-triangulation transfer, calibration, and training variance. Success means accuracy remains stable under relabeling and improves on toroidal or higher-genus graphs; removing transport or replacing all \(P_e\) by identity should remove the gain.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.