# Transverse Fourier Collision Control

- ID: 2756
- Canonical URL: https://synthcore.org/idea/2756/transverse-fourier-collision-control
- API JSON: https://synthcore.org/api/idea/2756.json
- API Markdown: https://synthcore.org/api/idea/2756.md
- Verification status: unverified
- Source: [arXiv:2608.27544](https://arxiv.org/abs/2608.27544)
- Category: architecture
- Solves: stability, accuracy, scalability
- ML areas: mlp, cnn, regularization, training
- Math tags: harmonic-analysis, geometry, combinatorics, linear-algebra
- Ratings: usefulness 5/10; difficulty 6/10; novelty 7/10

## Idea description

Construct a Fourier layer whose active frequencies lie on several nonparallel polygonal patches or thin annular sectors, and cap repeated difference vectors generated by pairs of patches. The bounded-multiplicity geometry limits how many input frequency pairs can contribute to the same output frequency, potentially reducing spectral aliasing and gradient variance in nonlinear Fourier mixing.

## Mathematical statement

For a convex polygon P, let S_i and S_j be two nonparallel sides, and let Gamma_{i,R}^{(A)} and Gamma_{j,R}^{(A)} be lattice-frequency subsets lying near the corresponding A-trimmed neighborhoods of the dilated sides R S_i and R S_j. The paper proves the separation estimate |xi-zeta| >= c_P A - 2 for every pair (xi,zeta) in these sets, where c_P > 0 depends only on P and A is the tangential distance from the side endpoints. It also proves that, for every difference vector k in Z^2, the number of pairs satisfying zeta-xi=k is bounded by a constant depending only on P. For a neural spectral product, define m(k) as the number of retained frequency pairs producing difference k. The adaptation either enforces m(k) <= M by hard pair selection or penalizes excess collisions with sum_k [m(k)-M]_+^2. The mathematical asset is bounded multiplicity: each output difference frequency receives only a controlled number of cross-patch interactions.

## Key formulas

- $$|\xi-\zeta|\geqslant c_{P}A-2$$
- $$\#\Bigl\{(\xi,\zeta)\in\Gamma_{i,R}^{(A)}\times\Gamma_{j,R}^{(A)}:\zeta-\xi=k\Bigr\}\lesssim_{P}1$$
- $$m(k):=\#\{(i,j,\xi,\zeta):\xi\in\Gamma_i,\zeta\in\Gamma_j,\zeta-\xi=k\},\qquad \mathcal L_{\mathrm{coll}}=\sum_{k\in\mathcal K}[m(k)-M]_+^2$$
- $$\widehat{(fg)}(k)=\sum_{\xi+\zeta=k}\widehat f(\xi)\widehat g(\zeta),\qquad \widehat h(k)=\sum_{\zeta-\xi=k}W_{\xi,\zeta}\widehat f(\xi)\widehat g(\zeta)$$

## Implementation notes

(1) Integration point: modify the spectral mixing operation in a 2D Fourier neural operator, Fourier MLP, or spectral CNN. Represent each feature map by an H by W FFT, retain frequencies in a band near radius R, and divide the retained lattice frequencies into 3 to 8 patches associated with nonparallel sides of a fixed convex polygon such as a square or hexagon. Use separate complex weights for cross-patch products and then apply an inverse FFT followed by the usual residual block. (2) Pseudocode: precompute Gamma_i as the integer frequencies within radial distance h of R S_i and at least A tangential units from the side endpoints. For each pair xi in Gamma_i and zeta in Gamma_j with i not equal to j, compute k equal to zeta minus xi and insert the pair into a dictionary indexed by k. Either retain only the first M pairs per key or attach a learnable gate and optimize the collision penalty. During the forward pass, compute FFTs F and G, then set H[k] equal to the sum of W[xi,zeta] F[xi] G[zeta] over retained pairs with zeta minus xi equal to k. Add IFFT(H) to the residual stream. (3) Compute exactly from the geometry: patch membership, difference keys, multiplicities, and the separation test |xi-zeta| >= c_P A - 2. Estimate empirically the useful cap M, the value of c_P for the chosen polygon, spectral energy retention, and conditioning. (4) First experiment: train a 6-layer Fourier neural operator on a 64 by 64 periodic Burgers or Navier-Stokes dataset. Compare dense spectral multiplication, random frequency partitions, and polygonal transverse patches at equal parameter count and FFT cost. Measure validation error, loss at equal FLOPs, gradient-norm variance, high-frequency energy, and divergent-run frequency. A successful result would be lower validation error or substantially fewer unstable runs at the same retained-mode budget.

## Disclaimer

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