✗ Failed on benchmark 2026

MPDI-Certified Neural Observer

Implementation & benchmark of arXiv:2607.19926 — Integrating Deep Learning and Contraction Theory for Robust Nonlinear State Estimation via Unsupervised Scientific Machine Learning

Usefulness8/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Generalization Rnn Ssm Optimizer Training dynamics Control theory Dynamical systems Differential geometry Optimization Linear algebra

Source paper: Integrating Deep Learning and Contraction Theory for Robust Nonlinear State Estimation via Unsupervised Scientific Machine Learning arXiv:2607.19926 ⓘ · analyzed Aug 30, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained recurrent or neural-ODE state update with a copy of the known or learned plant dynamics plus a neural output-error correction, and train both the correction and a contraction metric using a pointwise matrix inequality penalty. The resulting observer should forget initialization exponentially and should amplify measurement noise by a quantitatively bounded factor rather than exhibiting unconstrained recurrent error growth.

Formulas

$$\dot{\hat x}=f(\hat x)+K_\theta\bigl(\hat x,y-h(\hat x)\bigr),\qquad A_\theta(\hat x,y)=\frac{\partial}{\partial\hat x}\left[f(\hat x)+K_\theta\bigl(\hat x,y-h(\hat x)\bigr)\right].$$

$$M_\phi(z)=L_\phi(z)L_\phi(z)^\top+\epsilon I,\qquad \mathcal{C}(z,y)=\dot M_\phi+A_\theta^\top M_\phi+M_\phi A_\theta+2\lambda M_\phi\preceq0,$$

$$\dot M_\phi(z)=\sum_{i=1}^{n}\frac{\partial M_\phi(z)}{\partial z_i}F_{\theta,i}(z,y),\qquad \mathcal{L}_{\mathrm{contr}}=\mathbb{E}_{(z,y)}\left[\max\left(0,\lambda_{\max}\left(M_\phi^{-1/2}\mathcal{C}M_\phi^{-1/2}\right)\right)^2\right].$$

$$\|e(t)\|\le\sqrt{\frac{M}{m}}e^{-\lambda t}\|e(0)\|+\sqrt{\frac{M}{m}}\frac{\bar b}{\lambda}\sup_{0\le s\le t}\|\nu(s)\|,\qquad mI\preceq M_\phi\preceq MI,$$

Mathematical statement

For a plant $\dot x=f(x)$, $y=h(x)+\nu$, use the observer $\dot{\hat x}=f(\hat x)+K_\theta(\hat x,y-h(\hat x))$, where $K_\theta$ is a neural correction, $\theta$ are its parameters, and $\nu$ is measurement noise. With measurement held fixed, define the observer vector field $F_\theta(\hat x,y)=f(\hat x)+K_\theta(\hat x,y-h(\hat x))$ and its state Jacobian $A_\theta(\hat x,y)=\partial F_\theta/\partial\hat x$. Learn a symmetric positive-definite metric $M_\phi(z)=L_\phi(z)L_\phi(z)^\top+\epsilon I$, where $L_\phi$ is lower triangular, $\phi$ are metric parameters, and $\epsilon>0$. The contraction MPDI is $\mathcal{C}(z,y)=\dot M_\phi+A_\theta^\top M_\phi+M_\phi A_\theta+2\lambda M_\phi\preceq0$, where $z=\hat x$, $\lambda>0$ is the target contraction rate, and $\dot M_\phi=\sum_i(\partial M_\phi/\partial z_i)F_{\theta,i}$. Penalize the largest positive eigenvalue of the normalized residual. If $mI\preceq M_\phi\preceq MI$, contraction gives a distance bound $\|e(t)\|\le\sqrt{M/m}\,e^{-\lambda t}\|e(0)\|$ for noiseless observations. Treating noise as an input with correction gain $B_\theta=\partial K_\theta/\partial(y-h(\hat x))$, the corresponding exponential input-to-state bound has the form $\|e(t)\|\lesssim\sqrt{M/m}e^{-\lambda t}\|e(0)\|+\sqrt{M/m}\,\|B_\theta\|\lambda^{-1}\sup_{s\le t}\|\nu(s)\|$.

Implementation notes

1. Integration point: use this as the recurrent transition for a neural state-space model or as the observer in a neural ODE. Given a measured sequence $y_k$, integrate $\hat x$ with RK4 or Euler using the plant drift plus the correction network $K_\theta(\hat x_k,y_k-h(\hat x_k))$. If $f$ is unknown, replace it with a second MLP $f_\psi(\hat x)$, but first test the method with a known synthetic drift so the contraction mechanism is isolated. Parameterize the metric as a Cholesky network $M=L L^\top+10^{-4}I$, and optionally clamp the diagonal of $L$ to enforce numerical eigenvalue bounds. 2. Pseudocode: sample collocation states $z_j$ and measurements $y_j$; compute $F_j=f(z_j)+K_\theta(z_j,y_j-h(z_j))$; use automatic differentiation to obtain $A_j=\partial F_j/\partial z_j$ and $\dot M_j=\sum_i(\partial_iM_j)F_{j,i}$; form $C_j=\dot M_j+A_j^\top M_j+M_jA_j+2\lambda M_j$; optimize $L=L_{\mathrm{prediction}}+\alpha\mathcal{L}_{\mathrm{contr}}+\beta L_{\mathrm{metric}}+\gamma\mathbb{E}\|K_\theta\|^2$, where $L_{\mathrm{metric}}$ penalizes metric eigenvalues outside $[m,M]$. 3. Computed versus estimated: the MPDI, Jacobians, and metric derivatives are computed by autodiff; global bounds $m,M$, the worst residual, and the noise gain $\bar b=\sup\|M^{1/2}B_\theta\|$ are estimated by dense validation grids, random sampling, and interval enlargement around the training domain. 4. First cheap experiment: train on a two-dimensional Van der Pol oscillator with 4,000 uniformly sampled collocation points and compare an unconstrained GRU/neural observer against the contraction-regularized observer at equal parameter count. Sweep the target rate $\lambda$ and integration step $\Delta t$. The predicted signatures are exponential error slopes no slower than approximately $-\lambda$ in the metric norm, a sharp discrete-time failure boundary near $\Delta t\lambda\approx2$ for forward Euler in the locally linearized error dynamics, and steady-state noise error scaling approximately as $\bar b\sigma_\nu/\lambda$ for Gaussian measurement noise standard deviation $\sigma_\nu$. Require the measured decay rate to be within 20% of the certified lower bound and the noise-error slope versus $1/\lambda$ to be approximately linear.

Verification

Failed on benchmark

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: No measurable effect

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 and ran a reproducible scalar known-plant MPDI observer MVP with autodiff Jacobian verification, contraction-rate sweep, Euler stability-bound sweep, inverse-lambda disturbance-gain sweep, and an unstable unconstrained baseline. All five mechanism checks passed: the MPDI threshold occurred exactly at lambda=a+q, measured decay rates were within 0.21% of theory, the Euler boundary appeared at h(a+q)=2, and disturbance steady-state error matched b*sigma/lambda within 0.005%. The mechanism is numerically manifested, but this is not evidence for a learned neural observer advantage.

Agent confidence: 8/10

Baseline: Unconstrained correction q=-0.8: final absolute error 10.8285 and growth rate 0.2978
Idea: Certified correction q=1.0: final absolute error 3.825e-6 and decay rate 1.5592; MPDI formula exact, contraction threshold exact, Euler boundary confirmed, inverse-lambda gain ratios [1.00005, 1.00000, 1.00000, 1.00000]

Limitations: Only a scalar linear plant with constant metric M=1 was tested; no neural K_theta, learned metric M_phi, Van der Pol dynamics, sequence training, noisy measurement-learning comparison, GRU baseline, multidimensional geometry, or computational-cost analysis was evaluated. The baseline comparison is illustrative rather than a matched trained-model benchmark.

How to run: python3 mpdi_observer_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.19926) Aug 30, 2026

No significant difference vs. the tuned baseline (−-0.0%, p=1, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0028
Idea mean: 0.0028
Effect (Δ): 0 (−-0.0%; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Not confirmed ✗
Practical effect: No effect
Baseline tuning: swept over 3 configs

Limitations:

The implementation uses a Euclidean input-output Jacobian surrogate rather than a full learned state-dependent Cholesky metric and continuous-time observer integration. Only the built-in 400-sample dynamics benchmark, 10 epochs, and three shared learning-rate settings were tested; no Van der Pol sequence, noisy-observation rollout, global metric bounds, or computational-cost analysis was run.

How to run: python3 stage2_bench.py

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

Stage 2 — Benchmark Aug 30, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented an MPDI-inspired contraction-penalty intervention on the structurally matched dynamics track using the shared rnn_small architecture. The scalar autodiff MPDI identity passed, and the trained model's measured Jacobian signature was below the selected surrogate threshold, but test MSE was exactly identical to baseline, with paired delta 0 and permutation p=1.0. Thus no neural-training win was observed.

Baseline: dynamics/rnn_small; best lr=0.01, weight_decay=0.0; 8-seed test MSE mean 0.0027758953 ± 0.0017084856.
Idea: contraction-penalty rnn_small; best lr=0.01, weight_decay=0.0; 8-seed test MSE mean 0.0027758953 ± 0.0017084856; paired delta 0.0, p=1.0.

Benchmark result