# Fixed/Anti-Fixed Signed Message Passing

- ID: 1685
- Canonical URL: https://synthcore.org/idea/1685/fixed-anti-fixed-signed-message-passing
- API JSON: https://synthcore.org/api/idea/1685.json
- API Markdown: https://synthcore.org/api/idea/1685.md
- Verification status: unverified
- Source: [arXiv:2608.01716](https://arxiv.org/abs/2608.01716)
- Category: architecture
- Solves: accuracy, generalization, stability
- ML areas: graph-nn, embedding
- Math tags: geometry, linear-algebra, representation-theory
- Ratings: usefulness 6/10; difficulty 4/10; novelty 5/10

## Idea description

Replace a conventional signed-graph message-passing layer with two coupled feature channels: a fixed channel invariant under switching and an anti-fixed channel that changes sign under switching. Unsigned aggregation updates invariant features, while signed aggregation updates anti-invariant features, implementing the paper's sphere-plus-involution representation at the hidden-state level.

## Mathematical statement

The paper represents signed-graph colours by unit vectors on a sphere equipped with an orthogonal involution J satisfying J^2=I. The fixed subspace is E_+={x:Jx=x}, and the anti-fixed subspace is E_-={x:Jx=-x}; each representation decomposes as x=u+v with u in E_+ and v in E_-. Applying the involution to one endpoint realizes switching: J(u+v)=u-v. The paper's Lovász-type geometric condition uses an adjacent-pair inner product 1/(1-q), where q is the continuous colouring parameter. We implement J(u,v)=(u,-v), with fixed and anti-fixed hidden channels, and use signed edge aggregation to realize the endpoint-switching rule.

## Key formulas

- $$\overline{\vartheta}(G):=\vartheta(\overline{G})$$
- $$\langle z_i,z_j\rangle=\frac{1}{1-q}\quad\text{for prescribed adjacent pairs}$$
- $$J^2=I,\qquad E_+=\{x:Jx=x\},\qquad E_- =\{x:Jx=-x\},\qquad J(u+v)=u-v$$
- $$h_i^{\ell+1}=\phi\!\left(W_0h_i^\ell+\sum_{j\in N(i)}W_+h_j^\ell+\sum_{j\in N(i)}\sigma_{ij}W_-h_j^\ell\right)$$

## Implementation notes

Integrate this as a drop-in layer in a signed GNN before node classification or graph pooling. Store every node state as h_i=(u_i,v_i), where u_i in R^d is fixed and v_i in R^d is anti-fixed. For each node, compute the unsigned aggregate a_i=sum_{j in N(i)} h_j and signed aggregate b_i=sum_{j in N(i)} sigma_ij h_j, with sigma_ij in {+1,-1}. Update u_i with an MLP receiving [u_i,a_i^u,b_i^u] and update v_i with a separate MLP receiving [v_i,a_i^v,b_i^v]. Optionally normalize (u_i,v_i) to unit norm to match the spherical representation. For a switching gauge s_i in {+1,-1}, transform edge signs as sigma'_ij=s_i sigma_ij s_j, leave u_i unchanged, and transform v_i to s_i v_i. Check equivariance by measuring ||u'_i-u_i|| and ||v'_i-s_i v_i||. Pool only invariant quantities, such as sum_i u_i and sum_i(v_i squared), or use a permutation-invariant readout after an invariant projection. Add a graph-output consistency loss between the original graph and randomly switched copies. The involution and switching law are mathematically fixed; channel widths, MLPs, normalization, and nonlinearities are empirical choices. First test on signed stochastic block models and a small real signed-node-classification dataset against a two-layer signed GCN. Success means unchanged graph predictions under random switching, improved test accuracy on unseen gauges, and stable gradients with only a modest FLOP increase.

## Disclaimer

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