# Dispersion-Matched Unitary Fourier State Space

- ID: 313
- Canonical URL: https://synthcore.org/idea/313/dispersion-matched-unitary-fourier-state-space
- API JSON: https://synthcore.org/api/idea/313.json
- API Markdown: https://synthcore.org/api/idea/313.md
- Verification status: mechanism_works
- Source: [arXiv:2607.00749](https://arxiv.org/abs/2607.00749)
- Category: dynamics
- Solves: stability, accuracy, sample-efficiency
- ML areas: ssm, cnn, initialization, training-dynamics
- Math tags: spectral-theory, dynamical-systems, pde, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace unconstrained per-frequency recurrent dynamics in a Fourier neural operator or spectral state-space model with oscillators initialized from the plasma dispersion relation \(\omega_k=\sqrt{\underline{b}^{2}+|k|^{2}}\). Each Fourier mode first undergoes a norm-preserving rotation at its prescribed frequency, while a small learned residual and optional nonnegative damping account for task-specific dynamics. This should reduce phase drift and exploding or vanishing activations when modeling long-horizon wave-like or oscillatory data.

## Mathematical statement

The paper derives the transverse eigenfrequency from the Fourier-mode eigenvalue equation \(-\lambda^{2}\mathbbm{E}_{3}=-(\underline{b}^{2}+|k|^{2})\mathbbm{E}_{3}\), where \(k\in\mathbb{R}^{d}\) is the spatial Fourier wavevector, \(|k|\) is its Euclidean norm, \(\underline{b}>0\) is the background coupling or plasma frequency, \(\mathbbm{E}_{3}\) is the transverse electric-field Fourier coefficient, and \(\lambda\) is the temporal oscillation frequency. Hence the paper obtains \(\lambda_{\perp k}=\sqrt{\underline{b}^{2}+|k|^{2}}\). The corresponding generator has eigenvalues \(\pm i\lambda_{\perp k}\), so its ideal evolution is a rotation and preserves the squared amplitude of each mode. Implement this with the real rotation matrix \(R(\theta)=\begin{bmatrix}\cos\theta&-\sin\theta\\\sin\theta&\cos\theta\end{bmatrix}\), optionally multiplied by \(e^{-\gamma_k\Delta t}\) with learned or fixed \(\gamma_k\geq0\). The learned network should predict only residual forcing or corrections to this analytically stable base evolution.

## Key formulas

- $$-\lambda^{2}\,\mathbbm{E}_{3}=-(\underline{b}^{2}+|k|^{2})\,\mathbbm{E}_{3}.$$
- $$\lambda_{\perp k}=\sqrt{\underline{b}^{2}+|k|^{2}}.$$
- $$R_k(\Delta t)=\begin{bmatrix}\cos(\lambda_{\perp k}\Delta t)&-\sin(\lambda_{\perp k}\Delta t)\\\sin(\lambda_{\perp k}\Delta t)&\cos(\lambda_{\perp k}\Delta t)\end{bmatrix},\qquad R_k(\Delta t)^{\mathsf T}R_k(\Delta t)=I.$$
- $$z_{k,t+1}=e^{-\gamma_k\Delta t}R_k(\Delta t)z_{k,t}+\Delta t\,F_{\theta}(z_t,x_t)_k.$$

## Implementation notes

Integrate this into a Fourier neural operator, spectral convolution layer, or sequence state-space model that already transforms features into spatial or temporal Fourier modes. For an input tensor x of shape [batch, time, channels, spatial dimensions], compute a real or complex FFT over the spatial dimensions, retain a fixed set of modes k, and represent each selected mode with a two-dimensional real state z_k containing cosine and sine components. Precompute omega_k=sqrt(b^2+sum_j k_j^2), using b as either a fixed hyperparameter or a positive trainable scalar b=softplus(beta). At every recurrent or propagation step, form theta=omega_k*dt and update z_k_base=R_k(theta) z_k, where R_k is the 2-by-2 rotation in the formula. Optionally use gamma_k=softplus(gamma_raw[k]) and multiply the rotation by exp(-gamma_k*dt). Add a learned forcing network F_theta from the current spectral state and input; use z_next=z_base+dt*F_theta. To preserve the intended inductive bias, initialize the forcing projection near zero and initialize gamma near zero, so the model starts as a stable oscillator. Estimate no eigenvalues from data: omega_k is computed analytically from the mode index and b. Compare against an otherwise identical Fourier layer with unconstrained learned complex multipliers or a standard diagonal SSM. First run on a small 1D or 2D wave-equation forecasting benchmark and on long-context synthetic sums of sinusoids, using a 4-layer model and horizons 10 times the training window. Measure validation error versus forecast horizon, activation norm, gradient norm, and training loss at equal parameter count and FLOPs. Success means lower long-horizon phase error, bounded activation norms, and equal-or-faster loss descent. Use ablations for fixed dispersion, learned dispersion, and unconstrained frequencies.

## Verification

- Status: mechanism_works
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a compact dispersion-matched oscillator forecaster with the prescribed analytic frequencies and an unconstrained learned 2x2 transition baseline. The numerical checks confirmed RᵀR=I, amplitude preservation, and the dispersion equation to approximately 1e-15. On data generated by the same dispersion law, the unitary model had zero clean forecast error while the noisy, data-limited baseline reached 0.7365 mean state MSE over the long horizon; this is a real signal, but it is substantially an oracle-matched synthetic result rather than evidence of broad neural-network superiority.

### Mechanism check

- Verdict: Built a compact dispersion-matched oscillator forecaster with the prescribed analytic frequencies and an unconstrained learned 2x2 transition baseline. The numerical checks confirmed RᵀR=I, amplitude preservation, and the dispersion equation to approximately 1e-15. On data generated by the same dispersion law, the unitary model had zero clean forecast error while the noisy, data-limited baseline reached 0.7365 mean state MSE over the long horizon; this is a real signal, but it is substantially an oracle-matched synthetic result rather than evidence of broad neural-network superiority.
- Confidence: 6/10
- Limitations: This is a NumPy state-transition MVP, not a trained Fourier neural operator or SSM with learned residual forcing/damping. The synthetic target exactly follows the proposed known dispersion relation, so the unitary method receives oracle structural information; no real wave benchmark, learned-versus-fixed b ablation, residual network, FLOP comparison, or multi-seed statistical study was tested.

## Artifacts

- [experiment.py](https://synthcore.org/code/88/experiment.py)
- [report.md](https://synthcore.org/code/88/report.md)
- [results.json](https://synthcore.org/code/88/results.json)
- [Download all files as ZIP](https://synthcore.org/download/88)

## Disclaimer

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