# Cactus-Graph Phase Budgeting

- ID: 2194
- Canonical URL: https://synthcore.org/idea/2194/cactus-graph-phase-budgeting
- API JSON: https://synthcore.org/api/idea/2194.json
- API Markdown: https://synthcore.org/api/idea/2194.md
- Verification status: unverified
- Source: [arXiv:2608.12591](https://arxiv.org/abs/2608.12591)
- Category: architecture
- Solves: stability, scalability, accuracy
- ML areas: architecture, training-dynamics, regularization, mlp, transformer
- Math tags: dynamical-systems, control-theory, spectral-theory, linear-algebra, graph-theory
- Ratings: usefulness 6/10; difficulty 7/10; novelty 8/10

## Idea description

Build neural computation graphs with explicitly phase-budgeted serial and parallel branches, treating serial compositions as SRG products and parallel residual branches as SRG sums. Allocate phase centers theta_i so that every loop or branch aggregate stays away from -1, enabling stability-aware architecture search and constructive control of branch gains.

## Mathematical statement

For matrix sets A_i, the paper gives the sum enclosure SRG_theta(sum_i A_i) contained in sum_i SRG_theta(A_i) when at least N-1 sets have the theta-chord property. It gives the product enclosure SRG_theta(product_i A_i) contained in product_i SRG_theta_i(A_i) when at least N-1 sets have the corresponding arc property. The neural analogue uses edge Jacobians J_i, serial path products P_q=product_{k in q}J_k, and a parallel aggregate S=sum_q P_q. Define each path phase budget theta_q=sum_{k in q}theta_k, construct an approximate enclosure S_hat by multiplying point sets along paths and adding them across paths, and require -1 not in S_hat. A differentiable surrogate is d=min_{z in S_hat}|1+z|, the distance of the enclosure from the critical feedback point.

## Key formulas

- $$\mathrm{SRG}_{\theta}\left(\sum_{i=1}^{N}A_i\right)\subseteq\sum_{i=1}^{N}\mathrm{SRG}_{\theta}(\mathcal{A}_i),$$
- $$\mathrm{SRG}_{\theta}\left(\prod_{i=1}^{N}A_i\right)\subseteq\prod_{i=1}^{N}\mathrm{SRG}_{\theta_i}(\mathcal{A}_i),\qquad \theta=\sum_{i=1}^{N}\theta_i,$$
- $$-1\notin\sum_q\prod_{k\in\mathcal{V}_q}\mathrm{SRG}_{\theta_k}(\mathcal{A}_k)\Longrightarrow I+\sum_q\prod_{k\in\mathcal{V}_q}J_k\text{ is nonsingular},$$
- $$\mathcal{L}_{\mathrm{cactus}}=\lambda_d\left[\max\left(0,d_{\min}-\min_{z\in\widehat{\mathcal{S}}}|1+z|\right)\right]^2+\lambda_\theta\sum_k(\theta_k-\theta_k^{\mathrm{target}})^2.$$

## Implementation notes

Implement a graph module containing serial chains and parallel skip branches, such as a residual-in-residual MLP or a transformer layer with multiple attention paths. Associate every edge with a scalar phase parameter theta_k and a positive gain parameter; these can be learned or used only as certificate variables. During a periodic certification step, estimate each edge Jacobian on a minibatch using Jacobian-vector products, approximate its SRG by a polar grid of complex points, multiply point sets along serial paths, and add the resulting sets across parallel paths. Compute d=min_z|1+z|. If exact point-set operations are expensive, use interval boxes in log-gain and phase coordinates. Add the hinge penalty or use feedback gain control: multiply branch output by c_q, with c_q updated to c_q min(1,d/d_min), whenever the margin is too small. The chord and arc composition rules and -1 avoidance condition are imported from the paper; finite Jacobian probes and set approximations are estimated. First test on CIFAR-10 using a 12-layer residual CNN or 24-layer residual MLP, comparing unrestricted parallel branches against phase-budgeted branches at equal parameter count. Track training loss, gradient norms, Jacobian conditioning, and d versus depth. The falsifiable prediction is a transition when the aggregate enclosure touches -1: the unbudgeted model should show oscillatory optimization and a sharp increase in ||(I+sum_q P_q)^(-1)||, whereas gain control should keep d above d_min and prevent the transition.

## Disclaimer

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