# Erdelyi-Kober Log-Scale Mixer

- ID: 2044
- Canonical URL: https://synthcore.org/idea/2044/erdelyi-kober-log-scale-mixer
- API JSON: https://synthcore.org/api/idea/2044.json
- API Markdown: https://synthcore.org/api/idea/2044.md
- Verification status: unverified
- Source: [arXiv:2608.09401](https://arxiv.org/abs/2608.09401)
- Category: architecture
- Solves: accuracy, stability, scalability
- ML areas: cnn, transformer, attention
- Math tags: harmonic-analysis, functional-analysis, measure-theory
- Ratings: usefulness 6/10; difficulty 4/10; novelty 8/10

## Idea description

Replace generic cross-scale mixing with a fixed-shape or lightly parameterized Erdelyi-Kober fractional convolution over logarithmic scale. The fractional order controls how strongly nearby scales are emphasized, while the exponential tail parameter controls the receptive field over distant scales, providing an interpretable alternative to dense cross-scale attention.

## Mathematical statement

For order \(\alpha>0\), parameter \(\beta>0\), and \(\gamma\in\mathbb{R}\), the paper defines the Erdelyi-Kober operator by \((I_{\beta,\gamma}^{\alpha}f)(x)=\frac{\beta x^{-\beta(\alpha+\gamma)}}{\Gamma(\alpha)}\int_0^x (x^\beta-t^\beta)^{\alpha-1}t^{\beta\gamma+\beta-1}f(t)\,dt\). Here \(x\) is the positive output-scale coordinate, \(t\) is the input-scale coordinate, \(f\) is a feature function, and \(\Gamma\) is the Gamma function. Setting \(x=e^s\) and \(t=xe^{-u}=e^{s-u}\), with \(u\geq0\), yields \((I_{\beta,\gamma}^{\alpha}f)(e^s)=\int_0^\infty k_{\alpha,\beta,\gamma}(u)f(e^{s-u})\,du\), where \(k_{\alpha,\beta,\gamma}(u)=\frac{\beta}{\Gamma(\alpha)}(1-e^{-\beta u})^{\alpha-1}e^{-\beta(\gamma+1)u}\). For \(\gamma>-1\), the kernel has finite mass \(M=\Gamma(\gamma+1)/\Gamma(\alpha+\gamma+1)\). Normalize by \(\bar{k}=k/M\) so constant features across log-scale are preserved. The factor \((1-e^{-\beta u})^{\alpha-1}\) controls near-zero behavior, while \(e^{-\beta(\gamma+1)u}\) controls the distant-scale tail.

## Key formulas

- $$(I_{\beta,\gamma}^{\alpha}f)(x)=\frac{\beta x^{-\beta(\alpha+\gamma)}}{\Gamma(\alpha)}\int_0^x(x^\beta-t^\beta)^{\alpha-1}t^{\beta\gamma+\beta-1}f(t)\,dt$$
- $$(I_{\beta,\gamma}^{\alpha}f)(e^s)=\int_0^\infty\frac{\beta}{\Gamma(\alpha)}(1-e^{-\beta u})^{\alpha-1}e^{-\beta(\gamma+1)u}f(e^{s-u})\,du$$
- $$M_{\alpha,\beta,\gamma}=\int_0^\infty k_{\alpha,\beta,\gamma}(u)\,du=\frac{\Gamma(\gamma+1)}{\Gamma(\alpha+\gamma+1)},\qquad \bar{k}(u)=\frac{k_{\alpha,\beta,\gamma}(u)}{M_{\alpha,\beta,\gamma}}$$
- $$Y_i=\sum_{j=0}^{\min(i,L)}w_jF_{i-j},\qquad w_j=\frac{\bar{k}(j\Delta s)\Delta s}{\sum_{\ell=0}^{L}\bar{k}(\ell\Delta s)\Delta s}$$

## Implementation notes

Integrate the mixer at the cross-scale feature tensor, rather than inside the channel MLP. For a CNN, construct a feature pyramid with levels indexed by uniformly spaced log-scales \(s_i=s_0+i\Delta s\), project each level to the same channel width, and apply the operation independently per channel or after a learned \(1\times1\) projection. For a transformer, group tokens by resolution or patch scale and apply the same operation before the next attention block. Use this procedure: (1) choose a kernel length \(L\), spacing \(\Delta s\), and initial parameters \(\alpha=1\), \(\beta=1\), \(\gamma=0\); (2) compute \(u_j=j\Delta s\) and the unnormalized kernel weights \(q_j=\frac{\beta}{\Gamma(\alpha)}(1-e^{-\beta u_j})^{\alpha-1}e^{-\beta(\gamma+1)u_j}\); (3) avoid the singular endpoint when \(\alpha<1\) by using midpoint quadrature, replacing \(u_0\) with \(\Delta s/2\), or by analytically integrating the first bin; (4) normalize \(w_j=q_j/\sum_{\ell=0}^Lq_\ell\); (5) compute the causal grouped convolution \(Y_i=\sum_{j=0}^{\min(i,L)}w_jF_{i-j}\); and (6) return \(Z_i=F_i+\sigma(g)Y_i\), where \(g\) is a learned scalar or per-channel gate. Parameterize \(\alpha=\operatorname{softplus}(a)+\epsilon\), \(\beta=\operatorname{softplus}(b)+\epsilon\), and \(\gamma=-1+\operatorname{softplus}(c)\), ensuring integrability. The exact mathematical quantities are the analytic kernel and its mass; empirical choices are \(L\), \(\Delta s\), channel projections, and learned parameters. First test on CIFAR-100 with a small ResNet-FPN or ConvNeXt-Tiny using three resolution levels. Compare against nearest-neighbor plus \(1\times1\) fusion and a learned depthwise cross-scale convolution at matched FLOPs. Measure top-1 accuracy, robustness under input resizing, gradient variance, and wall-clock cost. Success is higher accuracy or resize robustness at equal FLOPs, with no loss spikes or unstable gradients.

## Disclaimer

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