# Ellipsoidal Robust Lyapunov Training

- ID: 2072
- Canonical URL: https://synthcore.org/idea/2072/ellipsoidal-robust-lyapunov-training
- API JSON: https://synthcore.org/api/idea/2072.json
- API Markdown: https://synthcore.org/api/idea/2072.md
- Verification status: unverified
- Source: [arXiv:2608.10158](https://arxiv.org/abs/2608.10158)
- Category: dynamics
- Solves: stability, generalization
- ML areas: mlp, optimizer, training-dynamics, world-model
- Math tags: control-theory, optimization, convex-analysis, dynamical-systems
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Train a neural state-feedback controller together with a positive Lyapunov critic so that the closed-loop system decreases a Lyapunov function for every plant matrix inside the data-consistent uncertainty ellipsoid. Replace the paper's exact SOS constraints by differentiable sampled constraints or inner maximization over uncertain plant parameters, yielding a controller that is explicitly robust to measurement noise and system-identification error.

## Mathematical statement

The paper defines a nominal model [A_star B_star] and a data-consistent plant set C by the matrix ellipsoid $\mathcal{C}:=\{[A~B]:([A~B]-[A_\star~B_\star])^{\top}\mathbf{Q}^{-1}([A~B]-[A_\star~B_\star])\preceq\mathbf{A}^{-1}\}$, where $\mathbf{Q}\succ0$ and $\mathbf{A}\succ0$ determine the uncertainty shape. Its Lemma 3 states that the data-derived set contains this plant set when $\mathcal{R}\preceq\mathbf{Q}^{-1}$ and $[A_\star~B_\star]^{\top}\mathcal{R}[A_\star~B_\star]-Q_3\succeq\cdots$, so robust inequalities over the data set imply inequalities for all plants that could have generated the observations. The stabilization inequality is represented in the extracted text by $\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}\preceq\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}^{\top}[A_\star~B_\star]^{\top}\mathcal{R}[A_\star~B_\star]\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}-\mathcal{Q}(x)$, where $Z(x)$ and $W(x)$ are polynomial feature matrices, $u(x)$ is the feedback law, $\mathcal{R}$ is a robustness multiplier, and $\mathcal{Q}(x)\succeq0$ is the desired dissipation/decrease term. For neural adaptation, use a scalar Lyapunov function $V_\phi(x)>0$, closed-loop dynamics $f_{A,B,\theta}(x)=Ax+B\pi_\theta(x)$, and enforce $V_\phi(f_{A,B,\theta}(x))-V_\phi(x)\le-\alpha\|x\|^2$ for uncertain $(A,B)\in\mathcal C$.

## Key formulas

- $$\mathcal{C}:=\left\{[A~B]:\Big([A~B]-[A_\star~B_\star]\Big)^{\top}\mathbf{Q}^{-1}\Big([A~B]-[A_\star~B_\star]\Big)\preceq\mathbf{A}^{-1}\right\}.$$
- $$\mathcal{R}\preceq\mathbf{Q}^{-1},\qquad [A_\star~B_\star]^{\top}\mathcal{R}[A_\star~B_\star]-Q_3\succeq\cdots.$$
- $$\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}\preceq\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}^{\top}[A_\star~B_\star]^{\top}\mathcal{R}[A_\star~B_\star]\begin{bmatrix}Z(x)\\W(x)u(x)\end{bmatrix}-\mathcal{Q}(x).$$
- $$\mathcal{L}_{\mathrm{rob}}(\theta,\phi)=\frac{1}{NK}\sum_{i=1}^{N}\sum_{k=1}^{K}\operatorname{softplus}\!\left(V_\phi\!\left(A_kx_i+B_k\pi_\theta(x_i)\right)-V_\phi(x_i)+\alpha\|x_i\|^2\right),\quad (A_k,B_k)\sim\mathcal C.$$

## Implementation notes

(1) Integration point: apply this during training of an MLP controller $u=\pi_\theta(x)$ for a learned linear or locally linear world model. Add a Lyapunov MLP $V_\phi(x)$ with scalar output; parameterize it as $V_\phi(x)=\|g_\phi(x)\|^2+\varepsilon\|x\|^2$ to guarantee positivity. Estimate the nominal matrices $A_\star,B_\star$ from transition data and construct positive-definite $\mathbf Q$ and $\mathbf A$ from the residual covariance and a regularization floor. The uncertainty set is the ellipsoid in the paper's formula. (2) Pseudocode: initialize $\theta,\phi$; at each batch sample states $x_i$ and compute $u_i=\pi_\theta(x_i)$. Generate $K$ uncertain models $(A_k,B_k)$ by sampling a matrix direction $E$ with $\|\mathbf Q^{-1/2}E\|_F\le1$ and setting $[A_k~B_k]=[A_\star~B_\star]+\mathbf Q^{1/2}E\mathbf A^{-1/2}$, or maximize the Lyapunov violation over $E$ with 3--5 projected ascent steps. Compute $r_{ik}=V_\phi(A_kx_i+B_k u_i)-V_\phi(x_i)+\alpha\|x_i\|^2$ and backpropagate $\mathcal L_{task}+\lambda_{rob}\operatorname{softplus}(r_{ik})$. Include a positivity penalty or normalization $V_\phi(0)=0$. (3) Computed from the paper's mathematics: the ellipsoidal uncertainty geometry, the robust plant samples, and the decrease inequality. Estimated empirically: $A_\star,B_\star$, residual matrices $\mathbf Q,\mathbf A$, the Lyapunov network, and the multiplier $\lambda_{rob}$. For a small polynomial system, optionally replace the sampled loss by an SOS certificate: fit polynomial approximations of $\pi_\theta$ and $V_\phi$, then use an SDP solver to verify $V(f(x))-V(x)+\alpha\|x\|^2\le0$ on a bounded domain. (4) First experiment: use a 2D unstable linear system with randomly perturbed $A,B$, train a two-layer controller on one-step prediction or quadratic regulation, and compare vanilla policy training, nominal-model Lyapunov training, and ellipsoidal robust training. Evaluate rollout failure rate under 100 unseen plants, maximum state norm, task cost, and the fraction of sampled states violating the decrease inequality. Success is a lower catastrophic-rollout rate and fewer Lyapunov violations at matched control cost; a strong signal would be stable rollouts for uncertainty radii 2x larger than the nominal baseline tolerates.

## Disclaimer

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