# Log-Fractional Scale-Derivative Layer

- ID: 749
- Canonical URL: https://synthcore.org/idea/749/log-fractional-scale-derivative-layer
- API JSON: https://synthcore.org/api/idea/749.json
- API Markdown: https://synthcore.org/api/idea/749.md
- Verification status: unverified
- Source: [arXiv:2607.11462](https://arxiv.org/abs/2607.11462)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: cnn, architecture, training
- Math tags: harmonic-analysis, functional-analysis, numerical-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add a feature transformation that approximates the derivative of a fractional diffusion operator with respect to its order. Instead of only smoothing features with one fractional order, the layer exposes whether a feature changes rapidly across spatial scales, which can help with textures, edges, and multiscale patterns.

## Mathematical statement

The paper defines the fractional p-Laplacian by $(-\Delta_p)^t u(x)=C_{d,t,p}\,\mathrm{P.V.}\int_{\mathbb{R}^{d}}\frac{|u(x)-u(y)|^{p-2}(u(x)-u(y))}{|x-y|^{d+tp}}\,dy$, where $u$ is a feature function, $t\in(0,1)$ is the fractional order, $p\in(1,\infty)$, $d$ is the spatial dimension, and $C_{d,t,p}>0$ is a normalization constant. Its order derivative is $(-\Delta_p)^{s+\log}u=\left.\frac{d}{dt}(-\Delta_p)^t u\right|_{t=s}$. In the linear case $p=2$, differentiating the kernel gives $\mathcal{K}^{s+\log}(r)=C_{d,s}(B_{d,s}-2\log r)r^{-d-2s}$, where $r=|x-y|$ and $B_{d,s}=\partial_s\log C_{d,s,2}$. Under the standard Fourier normalization, $(-\Delta)^s$ has multiplier $|\xi|^{2s}$ and its order derivative has multiplier $2|\xi|^{2s}\log|\xi|$. The logarithmic factor is the transferable asset: it weights spatial frequencies according to scale sensitivity and changes sign around a characteristic scale.

## Key formulas

- $$(-\Delta_p)^t u(x):=C_{d,t,p}\,\mathrm{P.V.}\int_{\mathbb{R}^{d}}\frac{|u(x)-u(y)|^{p-2}(u(x)-u(y))}{|x-y|^{d+tp}}\,dy$$
- $$(-\Delta_p)^{s+\log}u(x):=\left.\frac{d}{dt}(-\Delta_p)^t u(x)\right|_{t=s}$$
- $$\mathcal{K}^{s+\log}(r)=C_{d,s}\frac{B_{d,s}-2\log r}{r^{d+2s}},\qquad B_{d,s}=\partial_s\log C_{d,s,2}$$
- $$\widehat{\mathcal{L}_{s+\log}u}(\xi)=2|\xi|^{2s}\log|\xi|\,\widehat{u}(\xi)$$

## Implementation notes

Integrate the layer after a convolutional feature map $X\in\mathbb{R}^{B\times C\times H\times W}$ and before the next normalization or residual block. Start with the linear $p=2$ version because it avoids the numerical difficulty of a principal-value nonlinear integral. Construct 2-D FFT frequency grids $\xi_x,\xi_y$, let $r_\xi=\sqrt{\xi_x^2+\xi_y^2}$, and define $m_s(\xi)=2(r_\xi+\epsilon)^{2s}\log(r_\xi+\epsilon)$ with the DC multiplier explicitly set to zero. Compute $Z=\mathrm{IFFT2}(m_s\odot\mathrm{FFT2}(X))$ independently for every batch item and channel, then return $X+\alpha Z$ or concatenate $[X,Z]$ followed by a learned $1\times1$ convolution. Use fixed $s\in\{0.25,0.5,0.75\}$ initially; an optional learned parameter is $s=0.1+0.8\,\sigma(\rho)$ per channel. The exact spectral multiplier comes from the paper's order derivative; FFT discretization, boundary padding, $\epsilon$, and the residual gate $\alpha$ are engineering choices. As a spatial cross-check, use offsets with weights $w_\delta=C_{d,s}(B_{d,s}-2\log\|\delta\|)\|\delta\|^{-d-2s}$ and apply $\sum_\delta w_\delta(X-\mathrm{shift}_\delta X)$. First test on CIFAR-10 with a small ResNet-18, comparing the baseline, a fractional layer with multiplier $|\xi|^{2s}$, and the proposed log-fractional layer at matched channels and training FLOPs. Measure clean accuracy, CIFAR-10-C accuracy, loss descent, gradient norms, and wall-clock cost. A useful success signal is at least 0.5 percentage points better corruption accuracy or faster convergence at equal final clean accuracy with less than 15% extra training time.

## Disclaimer

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