# Zeta-Corrected Singular Integral Layer

- ID: 808
- Canonical URL: https://synthcore.org/idea/808/zeta-corrected-singular-integral-layer
- API JSON: https://synthcore.org/api/idea/808.json
- API Markdown: https://synthcore.org/api/idea/808.md
- Verification status: mechanism_failed
- Source: [arXiv:2607.12390](https://arxiv.org/abs/2607.12390)
- Category: architecture
- Solves: accuracy, scalability, sample-efficiency
- ML areas: graph-nn, world-model, inference-speedup
- Math tags: numerical-analysis, harmonic-analysis, special-functions, approximation-theory
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Construct a periodic neural integral layer whose fixed singular kernel behaves like |y|^{-s} near the origin, but whose samples on the uniform grid are replaced on a small symmetric stencil by SinCoTrap correction weights. The correction cancels low-order Taylor errors caused by sampling the singularity, while all nonlocal grid points remain unchanged. Increasing the correction order from p=0 to p=1 or p=2 should reduce discretization error without increasing global grid resolution.

## Mathematical statement

The paper considers I[φ·σ]=∫_V φ(x)σ(x) dx with σ(x)=|x−x_0|^{-s}, where d is the spatial dimension, 0<s<d, V is a periodic unit cell, and φ is smooth away from the isolated singularity. Its corrected trapezoidal rule changes only a fixed number of weights near the singular grid point and achieves error O(h^{2p+2+d−s}), where h is grid spacing and p is correction order. The correction coefficients are generated from the analytically continued generalized lattice zeta function Z_{2α}(s)=Σ_{n∈Z^d\{0}} n^{2α}/|n|^s, initially defined for Re(s)>d+2|α|_1, with multi-index α=(α_1,...,α_d) and n^{2α}=∏_k n_k^{2α_k}. For an implementable approximation, use a symmetric stencil S and solve the even-moment system Σ_{r∈S} c_r r^{2α}=b_α(d,s) for |α|_1≤p, where b_α(d,s) is obtained from analytically continued zeta values; a practical initialization is b_α=AC[Z_{2α}(s)]/(2α)!, followed by calibration against a high-resolution reference. Symmetry makes odd moments vanish automatically and keeps the stored coefficient set small.

## Key formulas

- $$I[\varphi\cdot\sigma]=\int_V\varphi(\boldsymbol{x})\sigma(\boldsymbol{x})\,\mathrm{d}\boldsymbol{x},\qquad \sigma(\boldsymbol{x})=\frac{1}{|\boldsymbol{x}-\boldsymbol{x}_0|^s},\qquad 0<s<d.$$
- $$Z_{2\boldsymbol{\alpha}}(s):=\sum_{\boldsymbol{n}\in\mathbb{Z}^d\setminus\{\boldsymbol{0}\}}\frac{\boldsymbol{n}^{2\boldsymbol{\alpha}}}{|\boldsymbol{n}|^s},\qquad \operatorname{Re}(s)>d+2|\boldsymbol{\alpha}|_1.$$
- $$\|I-\mathcal{Q}_h^{(p)}\|=O\!\left(h^{2p+2+d-s}\right).$$
- $$\mathcal{Q}_h^{(p)}f(\boldsymbol{x}_i)=h^d\!\left[\sum_{\boldsymbol{j}\notin S}|h\boldsymbol{j}|^{-s}f(\boldsymbol{x}_i-h\boldsymbol{j})+h^{-s}\sum_{\boldsymbol{r}\in S}w_{\boldsymbol{r}}^{(p)}f(\boldsymbol{x}_i-h\boldsymbol{r})\right],\qquad \sum_{\boldsymbol{r}\in S}c_{\boldsymbol{r}}\boldsymbol{r}^{2\boldsymbol{\alpha}}=b_{\boldsymbol{\alpha}}(d,s),\quad |\boldsymbol{\alpha}|_1\le p.$$

## Implementation notes

(1) Exact integration point: add this as a periodic neural integral or convolution layer operating on a feature tensor F∈R^{B×C×n^d}. Fix a singular component K_sing(y)=a|y|^{-s}, optionally followed by a learned channel-mixing matrix A∈R^{C×C}; let a smooth learned branch model the nonsingular remainder. Apply the correction only to displacement samples in a fixed stencil S around zero during both training and inference. (2) Pseudocode: choose grid spacing h=1/n, dimension d, exponent s, correction order p, and symmetric stencil S={r∈Z^d:||r||_∞≤R}. Offline, evaluate or approximate AC[Z_{2α}(s)] for all |α|_1≤p, form b_α=AC[Z_{2α}(s)]/(2α)!, and solve the constrained least-squares system M c=b with M_{α,r}=r^{2α}; impose c_r=c_{−r} and use a minimum-norm solution. At runtime, for every grid site i, compute G_i=h^d[Σ_{j∉S}|hj|^{-s}F_{i−j}+h^{−s}Σ_{r∈S}w_rF_{i−r}] using periodic indexing, then output Y_i=φ(A G_i+W F_i). Set w_r to the ordinary sampled value |r|^{-s} plus correction c_r, with a finite learned or fixed value at r=0. (3) Mathematics versus empirical estimation: d,s,p, the moment equations, predicted order 2p+2+d−s, and zeta-based coefficients come from the paper. Estimate analytically continued zeta values numerically with Ewald splitting or fit coefficients against a very fine-grid reference, and monitor the moment residual ||Mc−b||_2. Train only the singular amplitude a and channel matrix A initially; freeze correction weights for the first ablation. (4) First cheap experiment: use a 2D periodic Burgers or advection-diffusion surrogate with inputs and targets on 32×32, 64×64, and 128×128 grids. Compare a naive periodic neural convolution against corrected layers with p=0,1,2, keeping width, optimizer, parameter count, and wall-clock budget fixed. Generate smooth random Fourier fields and compute reference singular convolutions on a 512×512 grid. The expected signal is a log-log discretization slope near 2p+2+d−s and lower validation error at equal resolution; success means p=1 or p=2 reaches the baseline accuracy with fewer grid points while adding only O(|S|) local work per site.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a readable periodic 2-D ZetaSingularIntegral layer in zeta_layer.py and a calibrated singular quadrature experiment in experiment.py. The even-moment systems solved to near machine precision, but the predicted slopes 3.5, 5.5, and 7.5 were not confirmed; observed smooth-field slopes were approximately 0.01, 0.05, and -3.09 for p=0,1,2, with nonmonotonic errors. Thus there is no reliable evidence of the claimed discretization-order win.

### Mechanism check

- Verdict: Built a readable periodic 2-D ZetaSingularIntegral layer in zeta_layer.py and a calibrated singular quadrature experiment in experiment.py. The even-moment systems solved to near machine precision, but the predicted slopes 3.5, 5.5, and 7.5 were not confirmed; observed smooth-field slopes were approximately 0.01, 0.05, and -3.09 for p=0,1,2, with nonmonotonic errors. Thus there is no reliable evidence of the claimed discretization-order win.
- Confidence: 9/10
- Limitations: No analytically continued generalized zeta/Ewald implementation, no Burgers or learned neural training experiment, no wall-clock or FLOP comparison, and no independent 512x512 reference. Calibration used exact polar cell moments rather than paper-derived analytically continued zeta coefficients, so this tests a local moment-calibration proxy rather than a fully paper-faithful zeta implementation.

## Artifacts

- [experiment.py](https://synthcore.org/code/329/experiment.py)
- [report.md](https://synthcore.org/code/329/report.md)
- [results.json](https://synthcore.org/code/329/results.json)
- [results_final.json](https://synthcore.org/code/329/results_final.json)
- [results_fixed.json](https://synthcore.org/code/329/results_fixed.json)
- [results_fixed2.json](https://synthcore.org/code/329/results_fixed2.json)
- [zeta_layer.py](https://synthcore.org/code/329/zeta_layer.py)
- [Download all files as ZIP](https://synthcore.org/download/329)

## Disclaimer

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