# Gauge-Equivariant Complex Message Passing

- ID: 105
- Canonical URL: https://synthcore.org/idea/105/gauge-equivariant-complex-message-passing
- API JSON: https://synthcore.org/api/idea/105.json
- API Markdown: https://synthcore.org/api/idea/105.md
- Verification status: unverified
- Source: [arXiv:2608.23655](https://arxiv.org/abs/2608.23655)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: graph-nn, embedding
- Math tags: graph-theory, representation-theory, linear-algebra
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Replace real nonnegative graph-edge weights in a directed GNN with unit-modulus complex gains and complex node states. Enforce node-wise U(1) gauge equivariance, so arbitrary phase choices at individual nodes cannot change predictions while gauge-invariant cycle phases remain available to encode directed relational structure.

## Mathematical statement

For a directed graph D with arc set A, each arc (u,v) receives a gain φ(u,v)∈T={z∈C:|z|=1}. The gain adjacency is A(Φ)_{ij}=φ(v_i,v_j) on arcs and zero otherwise. Under a node switching function σ:V→T, the switched gain is φ^σ(u,v)=σ(u)^{-1}φ(u,v)σ(v). A walk gain is φ(W)=∏_s φ(v_s,v_{s+1}), and cycle gains are invariant under switching because node phases telescope. Let node features transform as h_u^σ=σ(u)^{-1}h_u. Then the message m_u=∑_(u,v) φ(u,v)W h_v transforms as m_u^σ=σ(u)^{-1}m_u, provided W is complex-linear and shared across nodes. A nonlinear update must therefore be phase-equivariant, for example f(z)=z g(|z|), where g is a real MLP.

## Key formulas

- $$A(\Phi)_{ij}=\begin{cases}\varphi(v_i,v_j),&(v_i,v_j)\in\mathcal A,\\0,&\text{otherwise},\end{cases}\qquad \varphi(u,v)\in\mathbb T,\ |\varphi(u,v)|=1.$$
- $$\varphi^{\sigma}(u,v)=\sigma(u)^{-1}\varphi(u,v)\sigma(v),\qquad \varphi(W)=\prod_{s=0}^{k-1}\varphi(v_{i_s},v_{i_{s+1}}).$$
- $$m_u=\sum_{(u,v)\in\mathcal A}\varphi(u,v)W h_v,\qquad h_u^{\sigma}=\sigma(u)^{-1}h_u\ \Longrightarrow\ m_u^{\sigma}=\sigma(u)^{-1}m_u.$$
- $$f(z)=z\,g(|z|),\qquad f(\sigma^{-1}z)=\sigma^{-1}f(z)\quad (|\sigma|=1).$$

## Implementation notes

1. Integration point: implement a directed message-passing layer for a graph batch. Store each node state as a complex tensor with shape [batch,nodes,channels], and assign every directed edge either a fixed gain φ_uv=e^{iθ_uv} derived from edge features or a learnable phase θ_uv. Do not use independent real and imaginary nonlinearities, since they break gauge equivariance. 2. Pseudocode: initialize complex states h; for each layer, compute q_v=complex_linear(h_v), messages m_u=sum over outgoing arcs (u,v) of exp(iθ_uv)q_v, normalize by degree, set z_u=W_self h_u+m_u, and update h_u=z_u*MLP_abs(|z_u|). Optionally concatenate invariant features |h_u| and sampled cycle gains. For a gauge test, sample random node phases σ_u, replace θ_uv by θ_uv−angle(σ_u)+angle(σ_v), and replace h_u by exp(−i angle(σ_u))h_u; outputs should transform by the same node phase. 3. The paper supplies the gain algebra and telescoping invariance; estimate phases from data or learn them, and use ordinary backpropagation through sin and cos. 4. First experiment: compare a 3-layer real directed GCN/GAT, a complex layer without gauge augmentation, and this equivariant layer on Cora or Citeseer with randomly oriented edges and a synthetic task where labels depend on directed cycle phase products. Measure accuracy, robustness to random node-wise phase reparameterization, and oversmoothing via pairwise feature cosine similarity. Success is unchanged predictions under gauge transformations and improved accuracy on cycle-dependent tasks at comparable parameter count.

## Disclaimer

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