ISS Backstepping Latent Regulator
Implementation & benchmark of arXiv:2608.17262 — Nonadaptive Learning in Robust Nonlinear Output Regulation
Source paper: Nonadaptive Learning in Robust Nonlinear Output Regulation arXiv:2608.17262 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace unconstrained latent or neural-ODE dynamics with a strict-feedback cascade whose virtual controls are generated recursively by nonadaptive backstepping. Add a fixed internal-model oscillator when the desired output contains known-frequency periodic components, so the network tracks persistent targets without learning an unstable long-memory representation. The controller is designed to tolerate bounded neural-model mismatch and disturbances through an input-to-state stability margin.
Formulas
Mathematical statement
Assume the latent error has strict-feedback form $\dot e_i=e_{i+1}+d_i(e,t)$ for $i=1,\ldots,r-1$ and $\dot e_r=u+d_r(e,t)$, where $e_i\in\mathbb{R}^{d}$, $u\in\mathbb{R}^{d}$, and the unknown mismatch terms satisfy $\|d_i(e,t)\|\leq \bar d_i+\ell_i\|e\|$. Define virtual controls recursively by $\alpha_1=-k_1e_1$ and $\alpha_i=-k_ie_i+\dot\alpha_{i-1}$ for $i>1$, and errors $z_i=e_i-\alpha_{i-1}$ with $\alpha_0=0$. A practical implementation replaces unavailable derivatives $\dot\alpha_i$ by input-driven filters $\tau_i\dot v_i=-v_i+\alpha_i$, using $v_i$ as the virtual control. With $V=\frac12\sum_{i=1}^{r}\|z_i\|^2$, the design aims to enforce $\dot V\leq-\lambda V+c\|d\|^2$, where $\lambda>0$ is obtained by choosing gains above the coupling and filter-error bounds. For a periodic reference generated by $\dot w=Sw$, with $S$ having simple purely imaginary eigenvalues, augment the latent state with $w$ and use $u=u_{\mathrm{fb}}+C_uw$.
Implementation notes
1) Integration point: use a neural ODE, recurrent state-space model, or deep equilibrium model with a designated latent chain $h_1,\ldots,h_r$. Interpret the prediction residual and target residual as $e_1$, and expose higher latent derivatives as the remaining strict-feedback states. Add an oscillator state $w$ only for periodic references; initialize $S$ from known frequencies or estimate frequencies offline, but keep controller gains fixed during training. 2) Pseudocode: compute $e_1=h_1-h_{1,\mathrm{target}}$; set $\alpha_1=-k_1e_1$; for $i=2,\ldots,r-1$, update $\tau_i\dot v_i=-v_i+\alpha_i$ and set $\alpha_i=-k_ie_i+\dot v_{i-1}$; apply final forcing $u=-k_re_r+\dot v_{r-1}+C_uw$; integrate the network and oscillator $\dot w=Sw$. Monitor $V=\frac12\sum_i\|e_i-\alpha_{i-1}\|^2$. 3) Compute residuals, filter states, $V$, and empirical mismatch slopes $\ell_i$ using Jacobian-vector products on minibatches; estimate the aggregate coupling bound $\Gamma$ as a high quantile over held-out trajectories. The paper supplies the ISS/Lyapunov mechanism and gain-separation requirement, while the Jacobian bounds are estimated empirically. 4) First cheap experiment: train a two-layer neural ODE on a forced Duffing oscillator and compare vanilla latent dynamics, spectral normalization, and this regulator under frequency-shifted forcing. Sweep equal gains $k_i=k$ and measure divergence. The quantitative prediction is a stability transition near $k_c=\Gamma$: above it, $V(t)$ should decay approximately as $V(0)e^{-2(k-\Gamma)t}$; below it, trajectories should show persistent growth or large oscillations. With bounded disturbance magnitude $\|d\|$, the asymptotic residual should scale like $\|d\|/\sqrt{k-\Gamma}$.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.