# Newton-polygon anisotropic spectral regularizer

- ID: 2173
- Canonical URL: https://synthcore.org/idea/2173/newton-polygon-anisotropic-spectral-regularizer
- API JSON: https://synthcore.org/api/idea/2173.json
- API Markdown: https://synthcore.org/api/idea/2173.md
- Verification status: unverified
- Source: [arXiv:2608.12250](https://arxiv.org/abs/2608.12250)
- Category: regularization
- Solves: accuracy, stability, sample-efficiency
- ML areas: transformer, ssm, regularization, world-model
- Math tags: functional-analysis, harmonic-analysis, geometry
- Ratings: usefulness 5/10; difficulty 4/10; novelty 6/10

## Idea description

Regularize a spatiotemporal neural model with spectral penalties corresponding to several temporal-spatial scaling laws rather than using a single isotropic smoothness penalty. The model can remain spatially detailed while suppressing temporal oscillations, or learn the opposite preference when the data demand it.

## Mathematical statement

The paper defines the elementary anisotropic weight function $w_{o_y}(\tau,\xi)=\langle\tau\rangle^y+\langle\xi\rangle$ for $y>0$, where $\tau$ is temporal frequency, $\xi$ is spatial frequency, and $\langle z\rangle=(1+|z|^2)^{1/2}$. It also defines the monomial weight $w_{(s,r)}(\tau,\xi)=\langle\tau\rangle^s\langle\xi\rangle^r$, where $s$ is temporal order and $r$ is spatial order. A Newton polygon is the convex geometric object whose vertices $(r_k,s_k)$ encode the competing orders. We use one weight for each selected vertex or edge and penalize the weighted Fourier energy of a feature field $u$. The coefficients $\alpha_k$ are nonnegative mixture weights, optionally learned with a softmax; the mathematical weights are fixed by the selected polygon while $\alpha_k$ and the global coefficient $\lambda$ are fitted empirically.

## Key formulas

- $$w_{o_y}(\tau,\xi)=\langle\tau\rangle^{y}+\langle\xi\rangle,\qquad y\in(0,\infty),$$
- $$w_{(s,r)}(\tau,\xi):=\langle\tau\rangle^{s}\langle\xi\rangle^{r},$$
- $$\mathcal{R}_{\mathrm{NP}}(u)=\sum_{k=1}^{K}\alpha_k\sum_{\tau,\xi}\left(\langle\tau\rangle^{s_k}+\langle\xi\rangle^{r_k}\right)^2\left|\widehat{u}(\tau,\xi)\right|^2,\qquad \alpha_k\geq 0,\quad\sum_{k=1}^{K}\alpha_k=1,$$
- $$\mathcal{L}=\mathcal{L}_{\mathrm{task}}+\lambda\mathcal{R}_{\mathrm{NP}}(u),\qquad (r_k,s_k)\in\mathcal{N}_{V},$$

## Implementation notes

Integrate the regularizer after a spatiotemporal feature block in a video transformer, ConvLSTM, SSM, or neural operator. Let $u\in\mathbb{R}^{B\times T\times H\times W\times C}$ be the feature tensor. Subtract its mean over $(T,H,W)$, apply an orthonormal three-dimensional FFT over time and the two spatial axes, and call the coefficient $\widehat{u}[b,c,\tau,\xi_x,\xi_y]$. Define $|\xi|=(\xi_x^2+\xi_y^2)^{1/2}$ and $\langle\tau\rangle=(1+\tau^2)^{1/2}$, with analogous notation for $\langle\xi\rangle$. Choose a small candidate set such as $(r_k,s_k)\in\{(1,1/2),(1,1),(2,1),(1,2)\}$. For every candidate compute $w_k=\langle\tau\rangle^{s_k}+\langle\xi\rangle^{r_k}$ and accumulate $E_k=\sum_{b,c,\tau,\xi}w_k^2|\widehat{u}|^2$. Divide each $E_k$ by $E_0=\sum|\widehat{u}|^2+10^{-8}$ to avoid scale dependence. Use $\mathcal{R}_{\mathrm{NP}}=\sum_k\alpha_kE_k/E_0$ and add $\lambda\mathcal{R}_{\mathrm{NP}}$ to the task loss. Begin with uniform $\alpha_k$; then parameterize $\alpha_k=\exp(a_k)/\sum_j\exp(a_j)$ and learn $a_k$, optionally adding a small entropy penalty. Compute FFT frequencies and weights exactly; estimate only $\alpha_k$ and $\lambda$ by validation. The first experiment should use 12-frame Moving-MNIST and a small ConvLSTM or video ViT. Compare against no regularization and isotropic Fourier L2 regularization at equal training FLOPs. Track validation MSE, temporal flicker, spatial edge sharpness, and performance when test videos use a different frame rate. Success means lower validation error and reduced temporal jitter without loss of spatial detail.

## Disclaimer

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