# Identifiability-Gated Latent Dynamics

- ID: 1638
- Canonical URL: https://synthcore.org/idea/1638/identifiability-gated-latent-dynamics
- API JSON: https://synthcore.org/api/idea/1638.json
- API Markdown: https://synthcore.org/api/idea/1638.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.00466](https://arxiv.org/abs/2608.00466)
- Category: dynamics
- Solves: stability, accuracy, sample-efficiency
- ML areas: world-model, ssm, rnn, training-dynamics, regularization
- Math tags: dynamical-systems, control-theory, statistics, linear-algebra, optimization
- Ratings: usefulness 8/10; difficulty 6/10; novelty 7/10

## Idea description

Augment a latent neural state-space model with an observable-coordinate residual that is first learned flexibly and then projected onto a constrained library of interpretable coupling terms. Train or collect data only after checking that the trajectory sufficiently excites the candidate terms; this prevents a latent model from fitting arbitrary hidden-state effects that are unidentifiable from the observations.

## Mathematical statement

Let $z_o(t)\in\mathbb{R}^{d_o}$ be observed coordinates, $z_h(t)\in\mathbb{R}^{d_h}$ hidden coordinates, $u(t)$ known inputs, and $p$ physical parameters. The measured dynamics are decomposed as $\dot z_o=F_o^{\mathrm{obs}}(z_o,u,p)+\Delta(z_o,z_h,u,p)$. Represent the coupling as $\Delta\approx B\Phi(z_o,u)$, where $\Phi\in\mathbb{R}^{K}$ is a fixed candidate-library vector and $B\in\mathbb{R}^{d_o\times K}$ is learned. Identifiability of the projected coefficients requires a persistently rich feature matrix $\Phi_T=[\Phi(t_1)^\top;\ldots;\Phi(t_N)^\top]$; a directly testable sufficient condition is $\lambda_{\min}(N^{-1}\Phi_T^\top\Phi_T)>\gamma>0$. If this eigenvalue is near zero, different hidden couplings produce the same observed trajectory and no estimator can reliably distinguish them. Use a neural latent state $h$ for reconstructing hidden effects, but apply the library constraint only to the observable derivative residual.

## Key formulas

- $$\dot{\bm z}=\bm F(\bm z,\bm u(t),\bm p),\qquad \bm z=\begin{bmatrix}\bm z_{\mathrm o}\\\bm z_{\mathrm h}\end{bmatrix}$$
- $$\dot{\bm z}_{\mathrm o}=\bm F_{\mathrm o}^{\mathrm{obs}}(\bm z_{\mathrm o},\bm u,\bm p)+\Delta(\bm z_{\mathrm o},\bm z_{\mathrm h},\bm u,\bm p)$$
- $$\Delta(\bm z_{\mathrm o},\bm z_{\mathrm h},\bm u,\bm p)\approx B\Phi(\bm z_{\mathrm o},\bm u),\qquad \lambda_{\min}\!\left({1\over N}\Phi_T^{\top}\Phi_T\right)\geq\gamma>0$$
- $$\hat B=\arg\min_B\sum_{n=1}^{N}\left\|\dot z_{o,n}-F_o^{\mathrm{obs}}(z_{o,n},u_n)-B\Phi(z_{o,n},u_n)\right\|_2^2+\lambda\|B\|_1$$

## Implementation notes

Use a latent neural ODE or recurrent state-space model. The encoder maps a recent observation window $y_{t-L:t}$ to a latent hidden state $h_t$; the decoder predicts $y_t$, while the transition contains two branches: a known or learned measured-coordinate base $F_o^{obs}(y_t,u_t)$ and a coupling branch $B\Phi(y_t,u_t)$. A separate small MLP may predict latent-state evolution $\dot h=G_\theta(h,y,u)$, but its contribution to the observed derivative must pass through the constrained coupling branch. Pseudocode: (1) encode $h_t$; (2) compute $r_t=\dot y_t-F_o^{obs}(y_t,u_t)$ using finite differences or an automatic-differentiation observation model; (3) compute $\phi_t=\Phi(y_t,u_t)$; (4) update the neural model on prediction loss plus $\|r_t-B\phi_t\|^2$; (5) every $M$ epochs solve the ridge or elastic-net regression for $B$ and zero coefficients whose confidence intervals include zero; (6) reject a symbolic update when $\lambda_{\min}(N^{-1}\Phi^\top\Phi)<\gamma$. Quantities inherited from the paper are the measured/hidden partition and residual decomposition. Estimate derivatives, feature covariance, standard errors, and $\gamma$ empirically using minibatch or held-out trajectories. The first cheap experiment is a Duffing-like partially observed oscillator or CartPole with one hidden velocity, comparing a vanilla latent neural ODE against this model on trajectory prediction. Sweep excitation: constant input, random input, and a frequency/speed sweep. The quantitative prediction is that constant-input data will yield a near-zero minimum Gram eigenvalue and unstable coefficient estimates across seeds, whereas the sweep will increase $\lambda_{\min}$ and sharply reduce coefficient variance; prediction error and coefficient variance should improve primarily once $\lambda_{\min}>\gamma$, not smoothly with training time.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a reproducible identifiability-gated observable residual estimator with library phi=[1,x,u,xu], Gram-eigenvalue rejection, and a flexible polynomial baseline. The mechanism manifested: constant input produced lambda_min≈0 and 0% acceptance; weak-excitation scaling followed the predicted lambda_min proportional to amplitude^2 with observed exponent 1.994 versus predicted 2.0; the gate crossed between amplitudes 0.1 and 0.3, with interpolated threshold 0.215. Accepted coefficient uncertainty also followed the predicted 1/sqrt(N lambda_min) behavior, with normalized values 0.53–0.59, and the gated model had lower test MSE than the flexible baseline at N=500 (0.00642 vs 0.00689).

### Mechanism check

- Verdict: Built a reproducible identifiability-gated observable residual estimator with library phi=[1,x,u,xu], Gram-eigenvalue rejection, and a flexible polynomial baseline. The mechanism manifested: constant input produced lambda_min≈0 and 0% acceptance; weak-excitation scaling followed the predicted lambda_min proportional to amplitude^2 with observed exponent 1.994 versus predicted 2.0; the gate crossed between amplitudes 0.1 and 0.3, with interpolated threshold 0.215. Accepted coefficient uncertainty also followed the predicted 1/sqrt(N lambda_min) behavior, with normalized values 0.53–0.59, and the gated model had lower test MSE than the flexible baseline at N=500 (0.00642 vs 0.00689).
- Confidence: 9/10
- Limitations: This is a controlled linear residual toy study rather than a neural latent ODE or recurrent state-space model, and it uses synthetic measured coordinates with directly observed residuals rather than finite-difference trajectory derivatives. No real physical dataset, hidden-state encoder/decoder, elastic-net confidence intervals, or long-horizon latent prediction was tested; the flexible baseline is a fixed degree-5 ridge model.

### Practical benchmark

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

## Artifacts

- [bench_report.json](https://synthcore.org/code/483/bench_report.json)
- [bench_stage2.py](https://synthcore.org/code/483/bench_stage2.py)
- [identifiability_experiment.py](https://synthcore.org/code/483/identifiability_experiment.py)
- [report.md](https://synthcore.org/code/483/report.md)
- [report_bench_2026-08-31T152243.md](https://synthcore.org/code/483/report_bench_2026-08-31T152243.md)
- [results.json](https://synthcore.org/code/483/results.json)
- [Download all files as ZIP](https://synthcore.org/download/483)

## Disclaimer

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