△ Mechanism confirmed, baseline not beaten 2026

OSL-QIB Contractive State Observer

Implementation & benchmark of arXiv:2608.30977 — Adaptive Observer of Nonlinear One-Sided Lipschitz Systems Using Estimated State Regressors With Finite Excitation

Usefulness7/10

Difficulty6/10

Novelty8/10

Dynamics Stability Accuracy Ssm Rnn World model Training dynamics Control theory Dynamical systems Optimization Linear algebra

Source paper: Adaptive Observer of Nonlinear One-Sided Lipschitz Systems Using Estimated State Regressors With Finite Excitation arXiv:2608.30977 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Add an observer correction to a recurrent or state-space neural model and constrain its local dynamics so latent-state errors contract according to a quadratic Lyapunov certificate. The design tolerates nonlinear residuals that are not globally Lipschitz, provided their one-sided growth and quadratic inner-bound constants satisfy a computable matrix inequality.

Formulas

$$\dot{\widehat{x}}=A\widehat{x}+\varphi(\widehat{x},u)+L(y-C\widehat{x}),\qquad e=\widehat{x}-x,$$

$$e^{\mathsf T}[\varphi(\widehat{x},u)-\varphi(x,u)]\leq\rho\|e\|^2,\qquad \|\varphi(\widehat{x},u)-\varphi(x,u)\|^2\leq\alpha\|e\|^2,$$

$$(A-LC)^{\mathsf T}P+P(A-LC)+(2\rho+\alpha/\epsilon)P+\epsilon P^2\preceq-\kappa I,\qquad P\succ0,$$

$$V=e^{\mathsf T}Pe\quad\Longrightarrow\quad \dot V\leq-\kappa\|e\|^2,\qquad V(t)\lesssim V(0)\exp\left(-\frac{\kappa t}{\lambda_{\max}(P)}\right).$$

Mathematical statement

Let the learned latent dynamics be $\dot{x}=Ax+\varphi(x,u)$, and define the observer $\dot{\widehat{x}}=A\widehat{x}+\varphi(\widehat{x},u)+L(y-C\widehat{x})$, where $C$ selects observed features and $L$ is an observer gain. For error $e=\widehat{x}-x$, let $\Delta\varphi=\varphi(\widehat{x},u)-\varphi(x,u)$. Assume the one-sided Lipschitz bound $e^{\mathsf T}\Delta\varphi\leq\rho\|e\|^2$ and the quadratic inner-bound condition $\|\Delta\varphi\|^2\leq\alpha\|e\|^2$, with constants $\rho$ and $\alpha$. A sufficient quadratic Lyapunov condition for $V=e^{\mathsf T}Pe$, with $P\succ0$, is $(A-LC)^{\mathsf T}P+P(A-LC)+(2\rho+\alpha/\epsilon)P+\epsilon P^2\preceq-\kappa I$, where $\epsilon>0$ is a Young-inequality multiplier and $\kappa>0$ is a decay margin. In the disturbance-free case this gives exponential error decay.

Implementation notes

1. Integration point: wrap a neural ODE, continuous-time state-space model, or recurrent hidden-state update with an observer correction. Maintain a predicted state $\widehat{x}$ and observed embedding $y$, and inject $L(y-C\widehat{x})$ before discretization. 2. Pseudocode: estimate local pairs $(e_j,\Delta\varphi_j)$ from teacher-forced trajectories; set conservative bounds $\rho=\max_j e_j^{\mathsf T}\Delta\varphi_j/(\|e_j\|^2+\delta)$ and $\alpha=\max_j\|\Delta\varphi_j\|^2/(\|e_j\|^2+\delta)$; solve an offline semidefinite program for $P\succ0$ and $L$, maximizing $\kappa$ subject to the displayed inequality; then update $\widehat{x}_{k+1}=\widehat{x}_k+h[A\widehat{x}_k+\varphi_\psi(\widehat{x}_k,u_k)+L(y_k-C\widehat{x}_k)]$. Reject or reduce neural updates whenever the largest eigenvalue of the certificate matrix exceeds $-\kappa$. 3. The OSL/QIB assumptions and Lyapunov inequality are the paper mechanism; $\rho$ and $\alpha$ are estimated from minibatches and should be inflated for safety. 4. First cheap experiment: train a latent neural ODE on partially observed damped Duffing trajectories, comparing an unconstrained observer, a fixed-gain observer, and the certified observer under observation noise. Sweep integration step $h$, and record $V_k=e_k^{\mathsf T}Pe_k$. The predicted signature is exponential decay with slope near or above $-\kappa/\lambda_{\max}(P)$ without disturbances, followed by a disturbance-dependent floor under noise. Scaling the neural nonlinearity until the certificate eigenvalue crosses zero should produce hidden-state divergence near that predicted boundary.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Inconclusive

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a scalar OSL-QIB contractive observer MVP with exact bounds for phi(x)=gamma*tanh(x), Lyapunov certificate evaluation, parameter sweeps, and noisy tracking comparison. The sampled OSL/QIB maxima stayed below theoretical rho=1.7 and alpha=2.89; the certificate sign boundary was observed at gamma=0.8 versus predicted 0.75 with grid spacing 0.1, and the gain required for kappa=0.8 was observed at 1.8, exactly matching prediction. The certificate upper bound held at every rate-sweep point, while the observer reduced noisy MSE from 1.298e-3 to 2.416e-5, approximately 54x.

Agent confidence: 8/10

Baseline: No observer correction: noisy tracking MSE 0.0012980 ± 0.0000000
Idea: Certified observer: noisy tracking MSE 0.00002416 ± 0.00001099; certificate boundary predicted 0.75 and observed grid transition 0.8; gain threshold predicted and observed at 1.8

Limitations: This is a scalar analytic toy rather than a trained neural ODE or Duffing model, and it uses hand-specified OSL/QIB constants instead of minibatch estimates or an SDP. The exact nonlinear stability boundary was analytically identified but not independently localized with a dense empirical sweep; disturbance-floor behavior was tested only with fixed observation noise and one input trajectory.

How to run: python3 observer_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.30977) Sep 2, 2026

Beats the tuned baseline by -0 metric points (p=0.0081, wins 8 / 0 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0.0001
Effect (Δ): -0 (n/a; negative = idea better)
Wins: 8 / 0 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: Inconclusive
Baseline tuning: swept over 0 configs

Limitations:

No explicit observation noise was present in the built-in dynamics track. The MVP uses a fixed diagonal correction and P=I rather than a learned full gain and SDP certificate; contraction was measured locally on trained models.

How to run: python3 bench_observer.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 2, 2026 latest Worked ✓

Agent confidence: 8/10

The observer improved the matched GRU recurrent system on the controlled-pendulum dynamics track. Mean test MSE decreased from 0.000101881 to 0.000052282, all 8 paired seeds improved, and the permutation p-value was 0.0081. The trained-model contraction signature was also confirmed.

Baseline: MSE 0.000101881 ± 0.000069692
Delta: -0.0000495987
Idea: MSE 0.000052282 ± 0.000043368
P value: 0.0081
Paired wins: 8/8

Benchmark result