✓✓ Beats tuned baseline 2026

Small-gain-certified surrogate controller

Implementation & benchmark of arXiv:2607.04362 — Approximate Feedback Linearization for a Nonlinear Hyperbolic PDE Class -- Part II: Neural Operator

Usefulness7/10

Difficulty5/10

Novelty7/10

Training Stability Accuracy Generalization World model Ssm Rl Control theory Dynamical systems Functional analysis Approximation theory

Source paper: Approximate Feedback Linearization for a Nonlinear Hyperbolic PDE Class -- Part II: Neural Operator arXiv:2607.04362 ⓘ · analyzed Aug 29, 2026

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 Not peer-reviewed.

Idea description

Train a neural controller as a uniformly accurate surrogate of a trusted but expensive controller, and use a measured small-gain condition to decide whether the surrogate is safe for closed-loop deployment. The approximation tolerance becomes an interpretable residual-state budget instead of an opaque validation metric.

Formulas

$$\left|\mathcal{U}_{N}(\mathbf{f}_{N},u)-\widehat{\mathcal{U}}_{N}(\mathbf{f}_{N},u)\right|\leq\varepsilon,\qquad \forall(\mathbf f_N,u)\in\mathcal F_N^{\star}(B_f,L_f)\times\mathcal V(R,L_u),$$

$$\mathrm{Lip}_{K}(r_0)<1,\qquad r_0<\bar r:=\min(r^*,\rho_f),\qquad r^*:=\mathrm{Lip}_{K}^{-1}(1),$$

$$\|u(t)\|_{\infty}\ \text{eventually enters a residual ball of radius}\ \gamma(\varepsilon),\qquad \gamma(\varepsilon)=O(\varepsilon).$$

$$\widehat{\mathcal U}_{\mathrm{deploy}}(f,u)=\begin{cases}\widehat{\mathcal U}_N(f,u),&\widehat{\mathrm{Lip}}_K(r_0)<1\ \text{and}\ \|u\|_\infty\le r_0,\\\mathcal U_N(f,u)\ \text{or a conservative fallback},&\text{otherwise.}\end{cases}$$

Mathematical statement

Theorem 2 guarantees a neural-operator surrogate $\widehat{\mathcal U}_N$ with uniform error $|\mathcal U_N(\mathbf f_N,u)-\widehat{\mathcal U}_N(\mathbf f_N,u)|\leq\varepsilon$ over the compact plant/state input class. Theorem 3 assumes a local feedback Lipschitz constant $\mathrm{Lip}_K(r_0)<1$, where $r_0$ is the allowed state radius, and requires $r_0<\bar r=\min(r^*,\rho_f)$, with $r^*=\mathrm{Lip}_K^{-1}(1)$ and $\rho_f$ the plant's admissible basin radius. Under these assumptions, the closed loop is practically stable and its residual class-$\mathcal K$ bound $\gamma(\varepsilon)$ scales linearly with the uniform surrogate error $\varepsilon$. Here $K$ denotes the closed-loop feedback map, $\mathrm{Lip}_K(r_0)$ is its local state Lipschitz constant, and $\gamma$ maps controller approximation error to the final state residual. The engineering principle is to minimize both approximation error and local feedback gain, then validate the inequality before deployment.

Implementation notes

Use a high-fidelity controller, optimizer, or solver-derived policy as the teacher and a neural operator as the deployable student. The integration point is the action computation from a plant descriptor $f$ and sampled state $u$; no kernel PDE or teacher solve should be needed during normal inference. Generate training examples by sampling plant coefficients and closed-loop states inside a declared radius $r_0$, evaluate the teacher action, and minimize supervised imitation error. At each validation checkpoint, estimate $\hat\varepsilon$ as the maximum absolute action error over dense random samples, adversarially perturbed inputs, and rollout-generated states. Estimate the local feedback gain with finite differences: for states $u_i$ satisfying $\|u_i\|_\infty\le r_0$, draw normalized perturbations $v_i$, compute $g_i=|\hat U(f_i,u_i+\delta v_i)-\hat U(f_i,u_i)|/(\delta\|v_i\|_\infty)$, and set $\widehat{\mathrm{Lip}}_K(r_0)=\max_i g_i$. Add $\alpha\max(0,g_i-\ell)^2$ to the loss for a target gain $\ell<1$. The paper supplies the small-gain criterion and predicts a residual radius linear in $\varepsilon$; the finite-sample error and gain are empirical estimates. Start with a 1D hyperbolic PDE discretized into 32--128 cells and compare the teacher, an unconstrained surrogate, and the gain-regularized surrogate. Plot final state residual versus measured $\hat\varepsilon$, and measure inference latency and unstable-rollout rate. Success means substantially lower inference cost, no unstable rollouts inside the validated region, and an approximately linear residual-error relationship.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Not reproduced ✗

Stage 2 · Practical benchmark result: Significant win ✓

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

How to read this: The implementation beat the baseline, but the claimed mechanism did not reproduce at benchmark scale. The gain may come from another effect; treat the cause as unresolved.

Stage 1 — Mechanism check agent confidence 8/10

Built a scalar small-gain-certified surrogate-controller MVP with teacher imitation, finite-difference gain estimation, gain regularization, rollouts, and reproducible sweeps. The mechanism manifested: strict |q|>1 cases grew, |q|<1 cases contracted, the marginal |q|=1 case was not asymptotically stable, and residuals matched epsilon/(1-|q|) to 1.9e-14 relative error across gains. The neural surrogate achieved estimated gain below 1 and zero unstable rollouts, but gain regularization did not improve the rollout metric in this run.

Agent confidence: 8/10

Baseline: Unconstrained MLP: max action error 0.0602, estimated local gain 0.7559, 0/20 unstable rollouts, mean final state 0.00211, latency 0.549 ms
Idea: Gain-regularized MLP: max action error 0.0483, estimated local gain 0.7612, 0/20 unstable rollouts, mean final state 0.00877, latency 0.558 ms; residual sweep matched predicted linear amplification exactly

Limitations: This is a scalar discrete-time toy analogue, not the paper's 1D hyperbolic PDE or neural operator. The empirical gain is finite-sample and finite-difference estimated rather than a formal uniform certificate; no adversarial validation, teacher-solver cost comparison, fallback controller, or statistically repeated seeds were tested.

How to run: python3 small_gain_experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2607.04362) Sep 2, 2026

Beats the tuned baseline by 19.7% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0.0007
Idea mean: 0.0005
Effect (Δ): -0.0001 (−19.7%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±5.5%
Mechanism: Not confirmed ✗
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

The built-in dynamics target is one-step/future-state regression rather than a full teacher-controller deployment loop. No teacher solve cost, explicit rollout instability rate, formal uniform certificate, adversarial validation, or direct NN-scale residual-versus-surrogate-error sweep was tested.

How to run: python3 stage2_small_gain_bench.py

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

Stage 2 — Benchmark Aug 30, 2026 latest Worked ✓

Agent confidence: 8/10

Implemented a gain-regularized rnn_small surrogate on the structurally matched dynamics track. Across 8 paired seeds, mean MSE improved from 0.0006559636 to 0.0005264665, with paired delta -0.0001294971 and permutation p=0.0081; this is a significant win under the bench decision rule. The analytic small-gain residual check passed to 1.39e-15 relative error, while NN-scale residual-versus-error linearity was not directly measured.

Baseline: rnn_small dynamics; swept best lr=0.003, epochs=16, weight_decay=0.0001; mean MSE 0.0006559635658049956
Idea: Same rnn_small and budget with finite-difference gain penalty; best alpha=0.8, target=0.85, delta=0.01; mean MSE 0.0005264664760034066
Paired delta: -0.00012949708980158903
Permutation p: 0.0081
Relative improvement: 19.74% lower mean MSE

Benchmark result