# Two-Channel Fractal Renormalization Network

- ID: 1723
- Canonical URL: https://synthcore.org/idea/1723/two-channel-fractal-renormalization-network
- API JSON: https://synthcore.org/api/idea/1723.json
- API Markdown: https://synthcore.org/api/idea/1723.md
- Verification status: beats_baseline
- Source: [arXiv:2608.02741](https://arxiv.org/abs/2608.02741)
- Category: architecture
- Solves: accuracy, stability, scalability
- ML areas: graph-nn, transformer, training-dynamics, regularization
- Math tags: renormalization, statistical-mechanics, dynamical-systems, bifurcations, combinatorics
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Replace ordinary depth-wise feature propagation by a ternary hierarchical block that recursively aggregates three child representations while maintaining separate neutral and defect channels. The block is initialized from the Sierpinski six-vertex recursion, then optionally learns a bounded correction. The neutral channel preserves the paper's cubic mixing law, while the defect channel provides a controlled route for long-range and nonlocal interactions.

## Mathematical statement

For six-vertex weights x and y = z = lambda, let Z_n be the charge-free partition function and Q_n the constrained partition function with opposite charges at two corner vertices. The exact Sierpinski recursion contains Z_(n+1) = lambda^3 Z_n^3 + 2 x^3 Q_n^3, together with a corresponding recursion Q_(n+1) = H(Z_n,Q_n;x,lambda) from the paper. Define the normalized defect ratio r_n = Q_n / Z_n. Its renormalization evolution is r_(n+1) = F(r_n;x,lambda), where F is obtained by dividing H by the Z recursion. A fixed point r_* changes stability when the absolute value of F'(r_*) equals one. For vector neural features, the analogous condition is that the spectral radius of the local Jacobian crosses one. The neural block uses neutral features a_n and defect features b_n, with elementwise cubic neutral mixing and a learned bounded defect map.

## Key formulas

- $$Z_{n+1}=\lambda^3 Z_n^3+2x^3 Q_n^3,\qquad y=z=\lambda.$$
- $$r_n=\frac{Q_n}{Z_n},\qquad r_{n+1}=F(r_n;x,\lambda),\qquad \left|\frac{\partial F}{\partial r}(r_*;x,\lambda)\right|=1.$$
- $$a_{n+1}=\lambda^3 a_n^{\odot 3}+2x^3 b_n^{\odot 3},\qquad b_{n+1}=G_\theta(a_n,b_n),$$
- $$\rho\left(J_{(a,b)}\right)<1\ \text{implies local contraction, while}\ \rho\left(J_{(a,b)}\right)>1\ \text{implies perturbation amplification}.$$

## Implementation notes

Use this block as a replacement for a 3-to-1 pooling or message-passing stage in a hierarchical image, graph, or token model. Partition the input into three child groups recursively. Each child emits two vectors: a neutral content vector a and a defect or disagreement vector b. First align the three child vectors with shared linear projections. Then compute the neutral parent using the paper-inspired rule a_parent = lambda^3 times the elementwise product of the three a vectors plus 2 times x^3 times the elementwise product of the three b vectors. Compute b_parent with a small gated MLP G_theta that receives all six child vectors. Apply RMS normalization after the cubic operation and use softplus parameterizations for x and lambda. At each level, estimate the largest local Jacobian eigenvalue with two power-iteration steps or Hutchinson-vector products. Add a penalty max(0, rho_hat - 1 + epsilon)^2 when contractive propagation is desired; alternatively target rho_hat near one for maximal signal persistence. The paper supplies the ternary recursion, the two boundary-state interpretation, and the fixed-point stability criterion. The learned defect map, Jacobian estimate, and task loss are empirical additions. First test a six-level model on CIFAR-10 using a small hierarchical CNN, comparing an ordinary pooling baseline and the fractal block at matched parameter count. Sweep q = x/lambda and measure loss, gradient amplification, and maximum stable depth. The predicted signature is a sharp change when the dominant Jacobian eigenvalue crosses one: below the boundary, perturbations decay approximately as rho^n; above it, perturbations grow or remain persistent. Near the deconfinement-like setting, active routing paths should scale approximately as L^1.3 with hierarchy depth L, rather than as the full number of sites.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: helps
- Verdict: Built the exact Sierpinski ratio recursion check and a runnable PyTorch two-channel ternary block with cubic neutral mixing and bounded defect routing. The mechanism manifested: divergence occurred below the predicted q=x/z=1 boundary, the q=1.2 marginal scaling gave mean n(R*-R)=2.097 versus predicted 2.4, and q=0.8 showed a double-log growth increment of 0.977 versus predicted 1.0. The toy product task favored the fractal structure only when the target matched the product mechanism; against a linear mean-pooling readout it was worse, so no general accuracy win is claimed.

### Mechanism check

- Verdict: Built the exact Sierpinski ratio recursion check and a runnable PyTorch two-channel ternary block with cubic neutral mixing and bounded defect routing. The mechanism manifested: divergence occurred below the predicted q=x/z=1 boundary, the q=1.2 marginal scaling gave mean n(R*-R)=2.097 versus predicted 2.4, and q=0.8 showed a double-log growth increment of 0.977 versus predicted 1.0. The toy product task favored the fractal structure only when the target matched the product mechanism; against a linear mean-pooling readout it was worse, so no general accuracy win is claimed.
- Confidence: 8/10
- Limitations: Did not run the proposed CIFAR-10 six-level CNN, matched-parameter training comparison, Jacobian regularization during learning, active-path scaling, or a full paper-scale partition-function implementation. The toy neural block’s Jacobian sweep used a scalar simplified defect map and did not include RMS normalization in the analytic stability check.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0152
- Paired wins: 7/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/637/bench_report.json)
- [bench_run.py](https://synthcore.org/code/637/bench_run.py)
- [experiment.py](https://synthcore.org/code/637/experiment.py)
- [fractal_block.py](https://synthcore.org/code/637/fractal_block.py)
- [report.md](https://synthcore.org/code/637/report.md)
- [report_bench_2026-08-31T203843.md](https://synthcore.org/code/637/report_bench_2026-08-31T203843.md)
- [results.json](https://synthcore.org/code/637/results.json)
- [smoke_test.py](https://synthcore.org/code/637/smoke_test.py)
- [Download all files as ZIP](https://synthcore.org/download/637)

## Disclaimer

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