# Stable Modal State-Space Bottleneck

- ID: 1565
- Canonical URL: https://synthcore.org/idea/1565/stable-modal-state-space-bottleneck
- API JSON: https://synthcore.org/api/idea/1565.json
- API Markdown: https://synthcore.org/api/idea/1565.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.28926](https://arxiv.org/abs/2607.28926)
- Category: architecture
- Solves: stability, accuracy, generalization
- ML areas: ssm, rnn, world-model, training-dynamics
- Math tags: dynamical-systems, control-theory, spectral-theory, linear-algebra, pde
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Insert a continuous spatial trunk and a pole-constrained modal state-space branch into a spatiotemporal predictor. The model represents a field as a sum of learned spatial modes and exponentially evolving modal coordinates, so long-horizon behavior is controlled by explicit poles rather than by an unconstrained recurrent transition matrix.

## Mathematical statement

Let $\mathbf{s}$ be a spatial coordinate, $t_k=k\Delta t$ a discrete time, $r$ the number of modes, $\phi_i(\mathbf{s})$ the $i$th trunk output, and $q_i(t)$ its modal coordinate. Reconstruct the field as $\hat{u}(\mathbf{s},t)=\sum_{i=1}^{r}\phi_i(\mathbf{s})q_i(t)$. Enforce orthonormality under spatial measure $d\mathbf{s}$ using $\int_\Omega\phi_i(\mathbf{s})\phi_j(\mathbf{s})d\mathbf{s}=\delta_{ij}$. Each damped mode has continuous pole $\mu_i=-\sigma_i+j\omega_{d,i}$, with $\sigma_i\geq 0$, natural frequency $\omega_{n,i}=\sqrt{\sigma_i^2+\omega_{d,i}^2}$, and damping ratio $\zeta_i=\sigma_i/\omega_{n,i}$. For discrete prediction, use $\alpha_i=\exp(\mu_i\Delta t)$ and $q_i[k+1]=\alpha_i q_i[k]$; stability is guaranteed when $|\alpha_i|=\exp(-\sigma_i\Delta t)<1$, equivalently $\sigma_i>0$. The projection-consistency condition is $q_i(t)\approx\int_\Omega\phi_i(\mathbf{s})u(\mathbf{s},t)d\mathbf{s}$. These relations transfer the paper's reconstruction, orthogonality, pole-residue, and temporal projection mechanisms into a forecasting model.

## Key formulas

- $$\hat{u}(\mathbf{s},t)=\sum_{i=1}^{r}\phi_i(\mathbf{s})q_i(t),\qquad \int_{\Omega}\phi_i(\mathbf{s})\phi_j(\mathbf{s})\,d\mathbf{s}=\delta_{ij}.$$
- $$\mu_i=-\sigma_i+j\omega_{d,i},\qquad \omega_{n,i}=\sqrt{\sigma_i^2+\omega_{d,i}^2},\qquad \zeta_i=\frac{\sigma_i}{\omega_{n,i}},\qquad \alpha_i=e^{\mu_i\Delta t}.$$
- $$q_i[k+1]=\alpha_iq_i[k],\qquad |\alpha_i|=e^{-\sigma_i\Delta t}<1\ \Longleftrightarrow\ \sigma_i>0.$$
- $$\mathcal{L}=\frac{1}{N}\sum_{n,k}\|u_n(\cdot,t_k)-\hat{u}_n(\cdot,t_k)\|^2+\lambda_o\|\Phi^\top W\Phi-I_r\|_F^2+\lambda_p\frac{1}{NrT}\sum_{n,i,k}\left|q_{n,i}[k]-\sum_{x\in\mathcal{G}}w_x\phi_i(x)u_n(x,t_k)\right|^2.$$

## Implementation notes

1. Exact integration point: use this module as the latent bottleneck of a mesh-based video predictor, neural PDE surrogate, or graph/SSM forecaster. An MLP trunk receives coordinates $\mathbf{s}$ and outputs $r$ values $\phi_i(\mathbf{s})$. An encoder maps the first observed field frames to complex modal amplitudes $q_i[0]$ and, optionally, slowly varying forcing terms. Parameterize damping as $\sigma_i=\operatorname{softplus}(a_i)+\epsilon$ and frequency as $\omega_{d,i}=\operatorname{softplus}(b_i)$, guaranteeing stable free decay. 2. Pseudocode: evaluate $\Phi_{xi}=\phi_i(x)$ on the training grid; normalize with the quadrature mass matrix $W$; add $\|\Phi^TW\Phi-I\|_F^2$; encode $q[0]$; for each future step compute $q_i[k+1]=\exp((-\sigma_i+j\omega_{d,i})\Delta t)q_i[k]$; decode $\hat u[:,k]=\Phi\operatorname{Re}(q[:,k])$; backpropagate reconstruction plus projection losses. For driven systems, add $q_i[k+1]=\alpha_iq_i[k]+\beta_i^Tf[k]$. 3. Computed from the paper: modal reconstruction, orthonormality, pole parameterization, and projection consistency. Estimated empirically: the number of modes $r$, quadrature weights $W$, encoder initialization, loss weights, and whether the modal basis remains useful under nonlinear or forced data. 4. First cheap experiment: train on snapshots of a 1-D damped Euler-Bernoulli beam or a synthetic sum of damped sinusoids, comparing a GRU/standard latent SSM against this module at equal parameter count. Sweep the unconstrained discrete spectral radius and verify that the modal model has no free-decay divergence while baselines diverge when $\rho(A)>1$. The quantitative prediction is that learned decay envelopes satisfy $\log|q_i[k]|\approx-\sigma_i k\Delta t+c_i$ with fitted slope error below 20%, and that the rollout remains bounded for every learned mode because $|\alpha_i|<1$. Also measure $\|\Phi^TW\Phi-I\|_F$ and MAC; the intended signature is orthogonality error below $10^{-2}$ and substantially slower long-horizon error growth than the unconstrained SSM.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: inconclusive
- Verdict: Built a reusable PyTorch StableModalBottleneck with weighted-QR spatial mode normalization, softplus-positive damping, complex exponential poles, projection, and stable rollout. The mechanism manifested: measured decay slopes matched -sigma*dt to 1.2e-16, the sigma=0 boundary and unstable negative-damping cases matched theory, and weighted orthogonality error was 2.85e-7 in the module smoke test. The modal forecast did not beat the direct spatial AR baseline in this noisy toy setting, so the stability mechanism is supported but an accuracy win was not observed.

### Mechanism check

- Verdict: Built a reusable PyTorch StableModalBottleneck with weighted-QR spatial mode normalization, softplus-positive damping, complex exponential poles, projection, and stable rollout. The mechanism manifested: measured decay slopes matched -sigma*dt to 1.2e-16, the sigma=0 boundary and unstable negative-damping cases matched theory, and weighted orthogonality error was 2.85e-7 in the module smoke test. The modal forecast did not beat the direct spatial AR baseline in this noisy toy setting, so the stability mechanism is supported but an accuracy win was not observed.
- Confidence: 8/10
- Limitations: Only a synthetic 1-D damped modal field was tested; no learned end-to-end trunk training, forcing inputs, nonlinear dynamics, irregular quadrature, GRU baseline, parameter/FLOP matching, or long-horizon real PDE/video benchmark was evaluated. The MVP uses real initial modal coordinates and reconstructs the real part of complex free decay.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.77805
- Paired wins: 5/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/455/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/455/bench_stage2.py)
- [experiment.py](https://synthcore.org/code/455/experiment.py)
- [report.md](https://synthcore.org/code/455/report.md)
- [report_bench_2026-08-31T123745.md](https://synthcore.org/code/455/report_bench_2026-08-31T123745.md)
- [results.json](https://synthcore.org/code/455/results.json)
- [smoke.py](https://synthcore.org/code/455/smoke.py)
- [stable_modal.py](https://synthcore.org/code/455/stable_modal.py)
- [Download all files as ZIP](https://synthcore.org/download/455)

## Disclaimer

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