# Criticality-aware fractional drift block

- ID: 2574
- Canonical URL: https://synthcore.org/idea/2574/criticality-aware-fractional-drift-block
- API JSON: https://synthcore.org/api/idea/2574.json
- API Markdown: https://synthcore.org/api/idea/2574.md
- Verification status: unverified
- Source: [arXiv:2608.19571](https://arxiv.org/abs/2608.19571)
- Category: architecture
- Solves: stability, scalability, accuracy
- ML areas: cnn, ssm, training-dynamics, initialization
- Math tags: pde, harmonic-analysis, dynamical-systems
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace an unconstrained multiscale residual block by the sum of a fractional diffusion branch and a drift or transport branch whose strength follows the PDE scaling law. At finer spatial scales, the drift coefficient is multiplied by R^{2s-1}; this suppresses unstable transport when s>1/2 while preserving equal-strength diffusion and drift at the critical value s=1/2.

## Mathematical statement

The paper studies the drift-diffusion equation $(-\Delta)^s u+b(x)\cdot\nabla u=f$ for $s\in[1/2,1)$, where $(-\Delta)^s$ has differential order $2s$, $b\cdot\nabla$ has order 1, and $f$ is a forcing term. Under the spatial rescaling $u_R(x)=u(x_0+Rx)$, the rescaled drift is $b_R(x)=R^{2s-1}b(x_0+Rx)$. Here R is the physical scale relative to a reference scale, x_0 is the center of the local patch, and b_R is the drift coefficient in rescaled coordinates. For s>1/2, R^{2s-1}    o0 as R    o0, so drift is lower order at fine scales; for s=1/2, the factor is exactly one and drift remains critical. The neural adaptation represents $(-\Delta)^s$ by a Fourier multiplier $|\xi|^{2s}$ or a spectral graph fractional power, and represents $b\cdot\nabla$ by a learned first-order directional convolution. The scaling factor is applied per resolution level rather than learned without constraint.

## Key formulas

- $$(-\Delta)^{s}u+b(x)\cdot\nabla u=f$$
- $$u_R(x)=u(x_0+Rx),\qquad b_R(x)=R^{2s-1}b(x_0+Rx)$$
- $$\mathcal{F}[(-\Delta)^s u](\xi)=|\xi|^{2s}\,\hat u(\xi)$$
- $$h_{\ell+1}=h_\ell+\eta\left[-\tau_\ell\,\mathcal{F}^{-1}\!\left(|\xi|^{2s}\mathcal{F}[h_\ell]\right)+R_\ell^{\,2s-1}\,\mathcal{D}_{\theta,\ell}(h_\ell)+\mathcal{F}_{\theta,\ell}(h_\ell)\right]$$

## Implementation notes

Integrate this into a 2-D CNN or Fourier neural operator residual block operating on feature maps h with shape B x C x H x W. Use one fixed or learnable fractional order s in [0.5,1), and define R_l as the pixel spacing or patch scale at level l relative to the finest-resolution reference, for example R_l=2^l. The diffusion branch computes a channelwise FFT, multiplies each frequency by |xi|^{2s}, applies the inverse FFT, and optionally uses a learned positive channel scale tau_l. The drift branch D_theta,l is a depthwise or grouped 3x3 convolution approximating a directional derivative. A stronger version predicts a two-channel vector field v_theta(h) and computes v_x times partial_x h plus v_y times partial_y h using centered finite differences. Multiply this branch by R_l^{2s-1}, while keeping the diffusion coefficient independent of R_l. Add a standard pointwise MLP or 1x1 convolution branch and an outer residual step size eta. Pseudocode: initialize h; compute D_s=IFFT(|xi|^{2s} FFT(h)); compute G=v_x*partial_x h+v_y*partial_y h; set h <- h+eta*(-tau*D_s+R^{2s-1}*G+MLP(h)). The exponent 2s-1 is taken directly from the paper; FFT frequencies, finite-difference derivatives, and R are computed deterministically. Estimate s using a grid over {0.5,0.6,0.75,0.9}, or learn it through s=0.5+0.49*sigmoid(a). First test on CIFAR-10 at 32x32 and 64x64 upsampled resolution using a six-block ResNet baseline with matched parameters and FLOPs. Compare unconstrained drift, scaling R^{2s-1}, and inverse scaling. Measure training-loss oscillation, gradient norm, accuracy at both resolutions, and degradation when transferring weights from 32x32 to 64x64. Success means lower gradient variance and better cross-resolution accuracy at equal compute, especially for s>0.5; the critical case should retain a non-negligible drift contribution rather than collapse it.

## Disclaimer

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