# Reciprocal-Lattice Gauge-Covariant Bloch Network

- ID: 2287
- Canonical URL: https://synthcore.org/idea/2287/reciprocal-lattice-gauge-covariant-bloch-network
- API JSON: https://synthcore.org/api/idea/2287.json
- API Markdown: https://synthcore.org/api/idea/2287.md
- Verification status: beats_baseline
- Source: [arXiv:2608.14348](https://arxiv.org/abs/2608.14348)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: embedding, mlp, regularization
- Math tags: pde, harmonic-analysis, symmetry
- Ratings: usefulness 7/10; difficulty 4/10; novelty 8/10

## Idea description

Build a Bloch-conditioned neural model whose periodic-factor representation transforms covariantly when the supplied Bloch wavenumber is shifted by a reciprocal lattice vector. Either canonicalize q to the first Brillouin zone or augment training with mathematically paired examples whose outputs differ by the exact phase gauge. This prevents the network from learning inconsistent predictions for physically identical Bloch modes.

## Mathematical statement

For a one-dimensional cell of length a, the physical field satisfies E''(x)+k^2E(x)=0 with Bloch boundary conditions E(a)=zE(0), E'(a)=zE'(0), where z=exp(iqa). Writing E(x)=\widetilde{E}_q(x)exp(iqx) gives the periodic-factor operator L(q)\widetilde{E}=\widetilde{E}''+2iq\widetilde{E}'+(k^2-q^2)\widetilde{E}=0 and \widetilde{E}_q(x+a)=\widetilde{E}_q(x). Let \kappa=2\pi/a. Since exp(i(q+m\kappa)x)=exp(iqx)exp(im\kappa x), the same physical field has the gauge transformation \widetilde{E}_{q+m\kappa}(x)=\widetilde{E}_q(x)exp(-im\kappa x). The homogeneous-cell aliases are \widetilde{E}_m(x)=exp(im\kappa x), q_{m,\pm}=\pm k-m\kappa. The implementation exploits this exact reciprocal-lattice covariance rather than treating q as an ordinary scalar feature.

## Key formulas

- $$E(x)=\widetilde{E}_q(x)e^{iqx},\qquad \widetilde{E}_q(x+a)=\widetilde{E}_q(x),$$
- $$L(q)\widetilde{E}=\widetilde{E}''+2iq\widetilde{E}'+(k^2-q^2)\widetilde{E}=0,$$
- $$\widetilde{E}_{q+m\kappa}(x)=\widetilde{E}_q(x)e^{-im\kappa x},\qquad \kappa=\frac{2\pi}{a},\quad m\in\mathbb{Z},$$
- $$\widehat{F}(q+m\kappa)[x]=e^{-im\kappa x}\widehat{F}(q)[x]$$

## Implementation notes

Integrate this at the input/output interface of a Bloch-conditioned MLP or Fourier neural operator that maps a periodic material profile and Bloch wavenumber q to a complex periodic field, band frequency, or mode coefficient tensor. Represent complex fields as two real channels sampled at cell locations x_j. First implement the canonicalized baseline: compute m=round(q/\kappa), q_0=q-m\kappa, feed q_0 to the network, predict u_0(x_j), and return u(x_j)=u_0(x_j)exp(iq_0x_j)exp(im\kappa x_j). If the target is a periodic factor, transform it as exp(-im\kappa x_j)u_0(x_j) according to the gauge convention. For the stronger version, retain raw q and add a covariance loss. Use m in {-2,-1,1,2}, generate q_shift=q+m\kappa, and transform the reference prediction with y_shift=exp(-im\kappa x)y(q). Optimize task_loss(y(q),target(q))+lambda_cov mean_j|y(q_shift)-exp(-im\kappa x_j)y(q)|^2. The quantities \kappa and the phase factor are exact; prediction errors and lambda_cov are empirical. Start with a 1D periodic coefficient dataset and a small Fourier neural operator, comparing raw-q conditioning, Brillouin-zone clipping, and gauge covariance. Measure error at unseen reciprocal shifts and consistency of reconstructed physical fields. Success means near-zero post-gauge discrepancy and lower error on aliased queries without increasing model size.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built a reciprocal-lattice gauge-aware Bloch MLP wrapper and numerical verification suite. For shifts m=-3..3, phase reconstruction error was 3.43e-16 and operator covariance residual was 6.60e-15, matching machine-precision predictions; the naive untransformed alias mismatch was exactly 1.4142 as predicted. In the toy MLP, mean nonzero-shift RMSE fell from 0.8810 for raw-q conditioning to 0.02344 with canonicalization plus the exact gauge phase, demonstrating the promised mechanism.

### Mechanism check

- Verdict: Built a reciprocal-lattice gauge-aware Bloch MLP wrapper and numerical verification suite. For shifts m=-3..3, phase reconstruction error was 3.43e-16 and operator covariance residual was 6.60e-15, matching machine-precision predictions; the naive untransformed alias mismatch was exactly 1.4142 as predicted. In the toy MLP, mean nonzero-shift RMSE fell from 0.8810 for raw-q conditioning to 0.02344 with canonicalization plus the exact gauge phase, demonstrating the promised mechanism.
- Confidence: 9/10
- Limitations: This is a small synthetic 1D experiment with a hand-designed periodic target and an MLP, not a full material-profile dataset or Fourier neural operator. It tests canonicalization and exact gauge wrapping, but not the stronger covariance-loss training variant, physical band-structure accuracy, FLOPs, or large-scale generalization.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/910/bench_report.json)
- [bloch_track.py](https://synthcore.org/code/910/bloch_track.py)
- [experiment.py](https://synthcore.org/code/910/experiment.py)
- [report.md](https://synthcore.org/code/910/report.md)
- [report_bench_2026-09-01T144646.md](https://synthcore.org/code/910/report_bench_2026-09-01T144646.md)
- [results.json](https://synthcore.org/code/910/results.json)
- [stage2_bench.py](https://synthcore.org/code/910/stage2_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/910)

## Disclaimer

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