# Fractional-to-Local RG Residual Block

- ID: 2320
- Canonical URL: https://synthcore.org/idea/2320/fractional-to-local-rg-residual-block
- API JSON: https://synthcore.org/api/idea/2320.json
- API Markdown: https://synthcore.org/api/idea/2320.md
- Verification status: unverified
- Source: [arXiv:2608.15120](https://arxiv.org/abs/2608.15120)
- Category: architecture
- Solves: stability, accuracy, scalability
- ML areas: ssm, graph-nn, transformer, training-dynamics
- Math tags: renormalization, statistical-mechanics, dynamical-systems, spectral-theory, pde
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace a single local message-passing or convolution operator by a spectrally controlled mixture of fractional and ordinary diffusion. The exponent σ is learned or scheduled, while a crossover gate forces the model to change parameterization near the renormalization-group threshold σ*=2, allowing long-range propagation when useful without retaining an unnecessarily nonlocal operator at short scales.

## Mathematical statement

The paper's long-range field theory uses the quadratic operator $(-\Delta)^{\sigma/2}$ and the cubic coupling's canonical distance from marginality $\epsilon'=3\sigma-d$. In $d=6-\epsilon$ and with $\delta=2-\sigma$, this becomes $\epsilon'=\epsilon-3\delta$; the perturbative long-range regime is $0<\delta<\epsilon/3$, and the proposed crossover is $\sigma_*=2$. For a neural feature vector $x\in\mathbb{R}^n$ on a graph or grid with positive semidefinite Laplacian $L=U\operatorname{diag}(\lambda_i)U^\top$, define $L^{\sigma/2}=U\operatorname{diag}(\lambda_i^{\sigma/2})U^\top$. A residual diffusion step is $x^+=x-hL^{\sigma/2}x$; its linear stability condition is $0<h<2/\lambda_{\max}(L^{\sigma/2})$. The transferable mechanism is continuous interpolation from fractional propagation for $\sigma<2$ to local Laplacian propagation at $\sigma=2$, rather than extrapolating the long-range branch beyond its validity.

## Key formulas

- $$\mathcal{S}[\phi]=\int d^d x\left[\frac{1}{2}\phi_a\left(\mu^\sigma r+(-\Delta)^{\sigma/2}\right)\phi_a+\frac{g\mu^{\epsilon'/2}}{3!}d_{abc}\phi_a\phi_b\phi_c\right],\qquad \epsilon'=3\sigma-d.$$
- $$d=6-\epsilon,\qquad \delta=2-\sigma,\qquad \epsilon'=\epsilon-3\delta,\qquad 0<\delta<\epsilon/3,\qquad \sigma_*=2.$$
- $$L^{\sigma/2}=U\operatorname{diag}(\lambda_i^{\sigma/2})U^\top,\qquad x_{k+1}=x_k-hL^{\sigma/2}x_k,\qquad 0<h<\frac{2}{\lambda_{\max}(L^{\sigma/2})}.$$
- $$x_{\ell+1}=x_\ell+\alpha_\ell F_\theta(x_\ell)+\beta_\ell\left[-h_\ell L^{\sigma_\ell/2}x_\ell\right],\qquad \sigma_\ell=2-\delta_\ell.$$

## Implementation notes

1. Integration point: use this block in a graph neural network, neural operator, vision state-space model, or transformer token mixer. Construct the graph or grid Laplacian L once; for large n, estimate L^{\sigma/2}x with Lanczos/Krylov methods or approximate it by a Chebyshev polynomial. Parameterize \sigma=\sigma_{\min}+(2-\sigma_{\min})\operatorname{sigmoid}(a), so the default branch remains below the predicted crossover. Mix the fractional branch with an ordinary local branch using q=\operatorname{sigmoid}(b(\sigma-2)) or, for a strictly continuous model, q=\operatorname{sigmoid}(b(\sigma-2)/    au). 2. Pseudocode: compute y_frac=frac_laplacian(x,sigma); compute y_local=Lx; set y=(1-q)y_frac+q y_local; update x=x+alpha*MLP(x)-beta*h*y. Estimate lambda_max(L^{sigma/2}) by 10-step power iteration and clip h to 0.9*2/lambda_max. Add a weak penalty gamma*(sigma-2)^2 only if unconstrained training collapses to an extreme exponent. 3. Computed from the paper: the fractional kinetic operator, canonical dimension epsilon'=epsilon-3delta, long-range validity condition, and crossover sigma*=2. Estimated empirically: sigma, gate q, spectral radius, and task loss. 4. First cheap experiment: train a 6-layer graph network on Cora or a 2-D reaction-diffusion next-step prediction task, comparing GCN, fixed sigma=1.5, fixed sigma=2, and the learnable crossover block. Sweep h across the predicted linear boundary h_c=2/lambda_max(L^{sigma/2}); measure whether exploding activations begin near h_c and whether the learned sigma or gate changes rapidly within |sigma-2|<0.1. The quantitative architectural prediction is continuous behavior for sigma<2 followed by a slope change in validation loss, learned spectral response, or gate occupancy at sigma approximately 2; the measured transition should lie within 20% of sigma=2. For a purely linear block, the norm decay should satisfy mode-wise |x_k^{(i)}|=|1-h lambda_i^{sigma/2}|^k|x_0^{(i)}| below h_c.

## Disclaimer

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