# Phase-Margin Graph Propagation

- ID: 574
- Canonical URL: https://synthcore.org/idea/574/phase-margin-graph-propagation
- API JSON: https://synthcore.org/api/idea/574.json
- API Markdown: https://synthcore.org/api/idea/574.md
- Verification status: unverified
- Source: [arXiv:2607.08260](https://arxiv.org/abs/2607.08260)
- Category: dynamics
- Solves: stability, accuracy
- ML areas: graph-nn, training, regularization
- Math tags: graph-theory, dynamical-systems, linear-algebra, control-theory
- Ratings: usefulness 5/10; difficulty 5/10; novelty 7/10

## Idea description

Replace fixed graph-convolution weights with edge couplings that depend on learned node amplitudes and relative phases, following the power-grid stability construction. Add trainable positive diagonal margins that dominate aggregate phase-weighted incident coupling, then use the resulting operator in a residual or recurrent GNN layer. This creates an operating-point-aware propagation rule intended to reduce oversmoothing, exploding iterates, and sensitivity to graph degree or edge loading.

## Mathematical statement

The paper's lossless network model uses symmetric susceptances B_{ij}=B_{ji}\geq 0, node amplitudes v_i, and node phases     heta_i. The edge sensitivity used in its stability matrix is w_{ij}=B_{ij}v_i^{\circ}v_j^{\circ}/\cos(    heta_i^{\circ}-    heta_j^{\circ}), where the superscript \circ denotes the operating point and the denominator should be positive and bounded away from zero. Equation (60) gives the diagonal correction T_{ii}=1/k_i^q-2\sum_{j
eq i}w_{ij}, where k_i^q>0 is an effective droop gain. We transfer this as a conservative graph-layer margin: define learned positive amplitudes v_i=\operatorname{softplus}(a_i)+\epsilon and bounded phases     heta_i=\pi    anh(\phi_i), compute w_{ij}, and parameterize k_i^q so that T_{ii}\geq    au>0. This uses the paper's exact operating-point correction as a local normalization and regularizer; T alone is not asserted to be the paper's complete stability certificate.

## Key formulas

- $$p_i=\sum_{j=1}^{n}v_i v_j B_{ij}\sin(\theta_i-\theta_j)+p_i^{\rm load},\qquad q_i=-\sum_{j=1}^{n}v_i v_j B_{ij}\cos(\theta_i-\theta_j)+q_i^{\rm load}.$$
- $$T_{ii}=\frac{1}{k_i^q}-2\sum_{j\neq i}\frac{B_{ij}v_i^{\circ}v_j^{\circ}}{\cos(\theta_i^{\circ}-\theta_j^{\circ})}=\frac{1}{k_i^q}-2\sum_{j=1}^{n}\frac{B_{ij}v_i^{\circ}v_j^{\circ}}{\cos(\theta_i^{\circ}-\theta_j^{\circ})}.$$
- $$w_{ij}=\frac{B_{ij}v_i v_j}{\max(\cos(\theta_i-\theta_j),\epsilon_c)},\qquad k_i^q=\left(2\sum_{j:(i,j)\in E}w_{ij}+\operatorname{softplus}(r_i)+\tau\right)^{-1},\qquad T_{ii}=\operatorname{softplus}(r_i)+\tau>0.$$
- $$h_i^{(\ell+1)}=h_i^{(\ell)}+\eta\,\sigma\!\left(W_0h_i^{(\ell)}+\sum_{j:(i,j)\in E}\frac{w_{ij}}{\sqrt{(\sum_m w_{im}+\epsilon)(\sum_m w_{jm}+\epsilon)}}W_1h_j^{(\ell)}\right).$$

## Implementation notes

(1) Integration point: use this construction in a graph-convolution or recurrent message-passing layer. Give each node two auxiliary scalar parameters a_i and phi_i, either learned globally per node or predicted from its current hidden state by a small linear head. Set v_i=softplus(a_i)+1e-4 and theta_i=pi*tanh(phi_i). Preserve the physical graph sparsity pattern and do not densify the adjacency. (2) Pseudocode: compute c_ij=cos(theta_i-theta_j); clamp it with c_safe=max(c_ij,epsilon_c), using epsilon_c=0.1 initially; compute w_ij=B_ij*v_i*v_j/c_safe; compute d_i=sum_{j:(i,j) in E}w_ij; set k_i=1/(2*d_i+softplus(r_i)+tau); set T_i=1/k_i-2*d_i; normalize messages with w_ij/sqrt((d_i+eps)(d_j+eps)); and apply h_next=h+eta*activation(W0*h+message). Add L_margin=lambda*mean(relu(tau-T_i)^2) if k_i is not hard-parameterized. (3) Computed from the paper: the sine/cosine operating-point dependence, edge weight w_ij, and diagonal correction T_i. Estimated empirically: amplitudes and phases if predicted from hidden states, together with eta, tau, epsilon_c, and lambda. For diagnostics, estimate the smallest eigenvalue of the symmetrized propagation Jacobian on small graphs. (4) First experiment: train a 6-layer GCN and a 10-step recurrent GNN on Cora, ogbn-arxiv subgraphs, and synthetic graphs with strong degree heterogeneity. Compare standard symmetric-normalized adjacency, a GCN with learned scalar edge gates, and this phase-margin layer at matched parameter count and FLOPs. Measure validation accuracy, gradient-norm variance, hidden-state norm across recurrent steps, and the largest stable learning rate. Success means fewer exploding or vanishing runs, stable propagation for more steps, or improved accuracy at equal depth; a useful target is at least twice the stable propagation depth without accuracy degradation.

## Disclaimer

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