# Dissipative Neural State-Space Identification

- ID: 3038
- Canonical URL: https://synthcore.org/idea/3038/dissipative-neural-state-space-identification
- API JSON: https://synthcore.org/api/idea/3038.json
- API Markdown: https://synthcore.org/api/idea/3038.md
- Verification status: mechanism_failed
- Source: [arXiv:2609.02071](https://arxiv.org/abs/2609.02071)
- Category: regularization
- Solves: stability, accuracy, generalization
- ML areas: ssm, rnn, world-model, regularization, training-dynamics
- Math tags: control-theory, dynamical-systems, optimization, convex-analysis
- Ratings: usefulness 8/10; difficulty 6/10; novelty 6/10

## Idea description

Attach a learned nonnegative storage function to a neural state-space model and penalize violations of a strict dissipativity inequality during rollout training. The resulting telescoping inequality limits cumulative output deviation and provides a monitor for whether long-horizon simulations are entering a stable turnpike regime.

## Mathematical statement

Let ℓₖ be the output prediction stage loss, ℓ* a reference optimal stage loss, Vφ(x) ≥ 0 a storage function, and d(yₖ,yₖ*) an output distance to a reference trajectory. Strict dissipativity requires ℓₖ − ℓ* ≥ Vφ(xₖ₊₁) − Vφ(xₖ) + α(d(yₖ,yₖ*)), where α(r) > 0 for r > 0. Summing over time telescopes the storage terms and bounds cumulative deviation by accumulated excess loss plus Vφ(x₀) − Vφ(x_N). Parameterize Vφ(x) = ||gφ(x)||² + δ||x||², with δ > 0, and use α(r) = ρr². Penalize the squared positive part of the inequality residual. The storage decrease and positive output penalty are the mechanisms that prevent persistent trajectory deviation.

## Key formulas

- $$ℓ_k−ℓ^*≥V_φ(x_{k+1})−V_φ(x_k)+α(d(y_k,y_k^*))$$
- $$Σ_{k=0}^{N-1}α(d(y_k,y_k^*))≤Σ_{k=0}^{N-1}(ℓ_k−ℓ^*)+V_φ(x_0)−V_φ(x_N)$$
- $$L_diss=(1/(BN))Σ_{b=1}^{B}Σ_{k=0}^{N-1}[V_φ(x_{k+1})−V_φ(x_k)+ρd_k²−ℓ_k+ℓ_hat^*]_+²$$
- $$V_φ(x)=||g_φ(x)||²+δ||x||²,\quad δ>0,\quad α(r)=ρr²,\quad ρ>0$$

## Implementation notes

Add a small storage network gφ to the neural state-space model. During each rollout, compute xₖ₊₁, yₖ, the ordinary prediction loss ℓₖ, and Vφ(xₖ). Add λdiss times Ldiss to the prediction objective. Use Vφ(x) = ||gφ(x)||² + δ||x||² with a small fixed δ to ensure nonnegativity. Set dₖ to the squared output error relative to the measured target, or use a separately optimized free-initial-state reference trajectory when comparing fixed and free solutions. Estimate ℓ_hat* as a stop-gradient exponential moving minimum of per-step losses; alternatively maintain an auxiliary free-initial-state rollout and use its average loss. Backpropagate through fθ and Vφ, but periodically freeze θ and evaluate dissipativity residuals on held-out sequences. The paper supplies the storage-function telescoping argument; the storage network and constants are learned or estimated. First experiment: compare an unconstrained MLP state-space model with the regularized model on noisy Duffing and Lorenz-63 data for one-step and 100-step prediction. Plot residuals, cumulative output error, and fixed/free initialization gaps versus horizon. The prediction is that average residual becomes nonpositive, cumulative squared deviation saturates or grows sublinearly, and the fixed/free gap remains bounded. Persistent positive residuals or linear cumulative growth falsify the proposed certificate.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a compact Duffing neural state-space identification MVP with a nonnegative learned storage function, squared positive-part dissipativity penalty, 100-step rollout evaluation, and a direct telescoping sanity check. The math check passed numerically: maximum telescoping error was 7.1e-15 and the constructed certified bound had slack of about -20. In the fixed-seed mini-experiment, the dissipative model had worse MSE than the unconstrained baseline (0.1619 vs 0.0341 at 100 steps; 0.001406 vs 0.001326 at 10 steps), while 59.8% of residuals remained positive and mean residual was 0.01445, so the promised stability/accuracy effect was not observed.

### Mechanism check

- Verdict: Built a compact Duffing neural state-space identification MVP with a nonnegative learned storage function, squared positive-part dissipativity penalty, 100-step rollout evaluation, and a direct telescoping sanity check. The math check passed numerically: maximum telescoping error was 7.1e-15 and the constructed certified bound had slack of about -20. In the fixed-seed mini-experiment, the dissipative model had worse MSE than the unconstrained baseline (0.1619 vs 0.0341 at 100 steps; 0.001406 vs 0.001326 at 10 steps), while 59.8% of residuals remained positive and mean residual was 0.01445, so the promised stability/accuracy effect was not observed.
- Confidence: 9/10
- Limitations: Only a small damped Duffing toy system and one fixed random seed were tested; Lorenz-63, noisy-data sweeps, fixed/free initialization gaps, FLOPs, plots, hyperparameter searches, and longer training were not evaluated. The baseline residual fields are intentionally NaN because no storage certificate is defined for the baseline.

## Artifacts

- [experiment.py](https://synthcore.org/code/1213/experiment.py)
- [report.md](https://synthcore.org/code/1213/report.md)
- [results.json](https://synthcore.org/code/1213/results.json)
- [Download all files as ZIP](https://synthcore.org/download/1213)

## Disclaimer

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