# Invariant-Guided Error-Compensating Rollouts

- ID: 2607
- Canonical URL: https://synthcore.org/idea/2607/invariant-guided-error-compensating-rollouts
- API JSON: https://synthcore.org/api/idea/2607.json
- API Markdown: https://synthcore.org/api/idea/2607.md
- Verification status: unverified
- Source: [arXiv:2608.20139](https://arxiv.org/abs/2608.20139)
- Category: dynamics
- Solves: stability, accuracy, speedup
- ML areas: world-model, rl, training-dynamics
- Math tags: dynamical-systems, numerical-analysis, control-theory
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Train a small controller to choose the next integration step size in a learned dynamical model using only deviations of conserved or slowly varying quantities. Unlike standard local adaptive solvers, optimize the complete rollout objective, allowing a later coarse step to compensate for an earlier discretization error. The controller can reduce the number of model evaluations while preserving long-horizon behavior.

## Mathematical statement

The paper uses the second-order Suzuki-Trotter approximation for a Hamiltonian split into two noncommuting parts: $H=H_{1}+H_{2}$ and $U(\delta t)=e^{-iH_{1}\delta t/2}e^{-iH_{2}\delta t}e^{-iH_{1}\delta t/2}$. Here $H$ is the full generator, $H_{1},H_{2}$ are subgenerators, $\delta t$ is the selected step size, and $U(\delta t)$ is the approximate one-step evolution operator. Noncommutativity causes discretization error, which appears as drift in conserved quantities. The paper defines the normalized energy diagnostic $\mathcal{E}_{n}=\langle\psi(t_{n})|H|\psi(t_{n})\rangle/L$, where $\psi(t_n)$ is the state after step $n$, $t_n$ is simulated time, $H$ is the conserved generator, and $L$ is a normalization such as system size. For a neural rollout, replace the quantum state by latent or physical state $z_n$, replace $H$ by a known or learned scalar invariant $C(z)$, and use $r_n=-|C(z_n)-C(z_0)|/s_C$ plus an action-cost penalty. The transferable property is global error orchestration: choose $\delta t_n$ from trajectory-level return rather than minimizing the one-step error $\|z_{n+1}-\Phi_{\delta t_n}(z_n)\|$. If two approximately conservative subflows $f_1,f_2$ are available, implement the adapted splitting update $z_{n+1}=\Phi^{(1)}_{\delta t_n/2}(\Phi^{(2)}_{\delta t_n}(\Phi^{(1)}_{\delta t_n/2}(z_n)))$.

## Key formulas

- $$H=H_{1}+H_{2},\qquad U(\delta t)=e^{-iH_{1}\delta t/2}e^{-iH_{2}\delta t}e^{-iH_{1}\delta t/2}.$$
- $$\mathcal{E}_{n}=\langle\psi(t_{n})|H|\psi(t_{n})\rangle/L.$$
- $$z_{n+1}=\Phi^{(1)}_{\delta t_n/2}\!\left(\Phi^{(2)}_{\delta t_n}\!\left(\Phi^{(1)}_{\delta t_n/2}(z_n)\right)\right),\qquad \delta t_n=\pi_\theta(o_n),$$
- $$r_n=-\frac{|C(z_{n+1})-C(z_0)|}{s_C}-\lambda\frac{\delta t_{\min}}{\delta t_n},\qquad J(\theta)=\mathbb{E}\left[\sum_{n=0}^{N-1}\gamma^n r_n-\beta\frac{|C(z_N)-C(z_0)|}{s_C}\right].$$

## Implementation notes

Integrate this at the rollout or inference loop of a neural world model, Hamiltonian neural network, or learned simulator rather than inside the network's parameter optimizer. The model must expose either two composable subflows $\Phi^{(1)}$ and $\Phi^{(2)}$, as in the paper's split evolution, or two vector-field components that can be advanced sequentially. Build a lightweight policy $\pi_\theta$ that receives $o_n=[t_n,\delta t_{n-1},(C(z_n)-C(z_0))/s_C,\operatorname{Var}(C(z_{n-k:n}))/s_C^2,\|z_n-z_{n-1}\|]$ and outputs a bounded action: $\delta t_n=\delta t_{\min}+(\delta t_{\max}-\delta t_{\min})\operatorname{sigmoid}(a_n)$. Pseudocode: initialize $z=z_0$, $t=0$; for $n=0,...,N-1$, compute $a=\pi_\theta(o)$, set $\delta t$ from the bounded action, apply subflow 1 for $\delta t/2$, subflow 2 for $\delta t$, then subflow 1 for $\delta t/2$, compute $C(z)$, assign $r_n=-|C(z)-C(z_0)|/s_C-\lambda\delta t_{\min}/\delta t$, update the policy with PPO or a critic, and continue until the time horizon. The mathematical quantities are the symmetric split composition, invariant-drift reward, and terminal long-horizon penalty; estimate $s_C$ from the training rollout distribution and estimate the variance feature using a rolling window. Train on small systems or short trajectories where exact dense rollouts are available, but do not give the policy the exact target state. First test on a 2D harmonic oscillator or spring-mass world model with a fixed compute budget, comparing fixed-step leapfrog or RK4, standard local adaptive stepping, and the learned controller. Measure invariant drift versus number of function evaluations, long-horizon state error, and rollout stability. Success is a lower energy-drift curve at equal evaluations or at least 25% fewer evaluations at equal terminal error; additionally test transfer to unseen initial conditions and to a larger system.

## Disclaimer

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