# Horizon-Adaptive Neural Tube Rollouts

- ID: 2765
- Canonical URL: https://synthcore.org/idea/2765/horizon-adaptive-neural-tube-rollouts
- API JSON: https://synthcore.org/api/idea/2765.json
- API Markdown: https://synthcore.org/api/idea/2765.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.27659](https://arxiv.org/abs/2608.27659)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: rnn, ssm, world-model, rl, training-dynamics
- Math tags: control-theory, dynamical-systems, linear-algebra, optimization, numerical-analysis
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Attach a robust, horizon-dependent uncertainty tube to a recurrent neural state-space model or learned policy. Instead of training only the nominal rollout, propagate state-estimation, model, and disturbance uncertainty through local Jacobians and impose a loss that keeps the tube inside task constraints. The method should be especially useful when short-horizon predictions are accurate but small Jacobian gains cause long-horizon divergence.

## Mathematical statement

The paper uses a time-varying model $\mathbf{x}_{k+1}=A_k\mathbf{x}_k+B_k\mathbf{u}_k+\mathbf{w}_k$, where $\mathbf{x}_k$ is the state, $\mathbf{u}_k$ the control, $A_k$ and $B_k$ vary with phase, and the bounded disturbance satisfies $|\mathbf{w}_k|\leq\bar{\mathbf{w}}$ componentwise. Transfer this construction to a neural transition $z_{k+1}=f_\theta(z_k,u_k)$ with nominal state $\hat z_k$ and radius vector $r_k\geq0$. If $J_k=\partial f_\theta/\partial z$ evaluated at $(\hat z_k,u_k)$ is the local state Jacobian and $d_k$ bounds model, input, and process uncertainty, then an axis-aligned robust tube obeys $r_{k+1}=|J_k|r_k+d_k$. For a feedback policy $u_k=\pi_\phi(z_k)$, use the closed-loop Jacobian $J_k^{\mathrm{cl}}=f_z+f_u\pi_z$. A constraint $g(z)\leq0$ is robustly enforced with $g(\hat z_k)+\|\nabla g(\hat z_k)\|_1r_k\leq0$. In a stationary regime, $\rho(J^{\mathrm{cl}})<1$ predicts bounded asymptotic tubes; $\rho(J^{\mathrm{cl}})>1$ predicts exponential growth.

## Key formulas

- $$\mathbf{x}_{k+1}=A_k\mathbf{x}_k+B_k\mathbf{u}_k+\mathbf{w}_k,\qquad |\mathbf{w}_k|\leq\bar{\mathbf{w}}.$$
- $$r_{k+1}=|J_k|r_k+d_k,\qquad J_k=\left.\frac{\partial f_\theta(z,u)}{\partial z}\right|_{(\hat z_k,u_k)}.$$
- $$r_H=\left(\prod_{j=0}^{H-1}|J_j|\right)r_0+\sum_{i=0}^{H-1}\left(\prod_{j=i+1}^{H-1}|J_j|\right)d_i.$$
- $$g(\hat z_k)+\|\nabla g(\hat z_k)\|_1r_k\leq0,\qquad \rho(J^{\mathrm{cl}})<1\ \text{predicts bounded asymptotic tubes}.$$

## Implementation notes

1) Integration point: use a GRU, residual RNN, or neural state-space model with transition $z_{k+1}=f_\theta(z_k,u_k)$. During training, maintain the nominal hidden state $\hat z_k$ and an elementwise radius vector $r_k$; for a policy, include the policy derivative in the closed-loop Jacobian. 2) Pseudocode: initialize $\hat z=E(o_0)$ and $r=r_0$; at each step compute $u=\pi_\phi(\hat z)$, update $\hat z\leftarrow f_\theta(\hat z,u)$, obtain $J$ by automatic differentiation, update $r\leftarrow |J|r+d$, and add $\lambda_r\|r\|_1$ plus penalties for violated robust constraints $g(\hat z)+|\nabla g|r>0$ to the ordinary prediction loss. Stop gradients through $r$ initially if memory is limiting. 3) Compute the tube recursion and Jacobians exactly; estimate $d_k$ from held-out transition-residual quantiles, and estimate $r_0$ from repeated encoder evaluations or initial-state uncertainty. Use a high residual quantile such as 99%. 4) First cheap experiment: train a small GRU world model on CartPole or a nonlinear pendulum with injected observation and transition noise, comparing ordinary multi-step training with tube-regularized training at equal parameter count. Sweep rollout horizon and a scalar Jacobian-gain multiplier. The quantitative prediction is a sharp change near the measured spectral-radius boundary $\rho(J^{\mathrm{cl}})=1$: below it, median $r_k$ should plateau; above it, $\log r_k$ should grow with slope approximately $\log\rho(J^{\mathrm{cl}})$. Constraint violations should decrease when robust backoff inequalities hold.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: no_effect
- Verdict: Built an exact axis-aligned neural tube recursion with autograd Jacobians, robust constraint backoff, and a matched small neural state-space comparison. The toy verification confirmed all three mechanism predictions: the stability boundary was observed at gamma=1.25 (=1/lambda), unstable growth slopes matched log(gamma·lambda) within 0.02%, and steady-state radius scaled linearly with disturbance magnitude to numerical precision. The tube regularizer did not improve this particular neural task: rollout MSE was essentially unchanged and constraints were never violated by either model, so the mechanism is validated but a practical win was not demonstrated.

### Mechanism check

- Verdict: Built an exact axis-aligned neural tube recursion with autograd Jacobians, robust constraint backoff, and a matched small neural state-space comparison. The toy verification confirmed all three mechanism predictions: the stability boundary was observed at gamma=1.25 (=1/lambda), unstable growth slopes matched log(gamma·lambda) within 0.02%, and steady-state radius scaled linearly with disturbance magnitude to numerical precision. The tube regularizer did not improve this particular neural task: rollout MSE was essentially unchanged and constraints were never violated by either model, so the mechanism is validated but a practical win was not demonstrated.
- Confidence: 9/10
- Limitations: The neural comparison used a tiny synthetic stable plant, fixed uncertainty radii rather than learned held-out 99th-percentile residuals, one-step tube regularization rather than full horizon-adaptive training, and no genuinely active constraints. GPU execution, CartPole, policy closed-loop Jacobians, and larger or harder long-horizon tasks were not tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.58005
- Paired wins: 5/8
- Benchmark verdict: no measurable effect

## Artifacts

- [bench_report.json](https://synthcore.org/code/1092/bench_report.json)
- [bench_tube.py](https://synthcore.org/code/1092/bench_tube.py)
- [report.md](https://synthcore.org/code/1092/report.md)
- [report_bench_2026-09-02T002452.md](https://synthcore.org/code/1092/report_bench_2026-09-02T002452.md)
- [results.json](https://synthcore.org/code/1092/results.json)
- [tube_experiment.py](https://synthcore.org/code/1092/tube_experiment.py)
- [Download all files as ZIP](https://synthcore.org/download/1092)

## Disclaimer

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