# Reachability-Tube Monitor for Hidden States

- ID: 2581
- Canonical URL: https://synthcore.org/idea/2581/reachability-tube-monitor-for-hidden-states
- API JSON: https://synthcore.org/api/idea/2581.json
- API Markdown: https://synthcore.org/api/idea/2581.md
- Verification status: unverified
- Source: [arXiv:2608.19678](https://arxiv.org/abs/2608.19678)
- Category: dynamics
- Solves: stability, generalization
- ML areas: rnn, ssm, world-model, training-dynamics
- Math tags: control-theory, dynamical-systems, linear-algebra, numerical-analysis
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Attach a low-dimensional reachable-set monitor to an RNN or state-space model and propagate the set of hidden states allowed by bounded inputs, parameter uncertainty, and process noise. Penalize or reset hidden states that leave the predicted tube, turning the paper's instantaneous set-membership fault test into a robust neural-state validity test.

## Mathematical statement

Represent the neural hidden dynamics as $h_{t+1}=f_\theta(h_t,x_t)+w_t$, where $h_t\in\mathbb{R}^n$ is the hidden state, $x_t$ is the input, and $w_t\in\mathcal W$ is bounded disturbance. Use a reduced state $z_t=Ph_t\in\mathbb{R}^2$, where $P\in\mathbb{R}^{2\times n}$ is a fixed projection learned from hidden-state trajectories. Around the current nominal state, approximate the reduced dynamics by $z_{t+1}\approx A_tz_t+B_tx_t+d_t+\epsilon_t$, where $A_t$ and $B_t$ are local Jacobians, $d_t$ is an offset, and $\epsilon_t\in\mathcal E_t$ bounds linearization error. Propagate an interval or zonotope reachable set $\mathcal Z_t$ using Minkowski sums. The distance of the observed projected state $Ph_t$ from $\mathcal Z_t$ is an instantaneous validity residual. A hinge penalty or projection correction suppresses excursions. For locally linear dynamics, boundedness is expected when the effective spectral radius satisfies $\rho(A_t)<1$.

## Key formulas

- $$h_{t+1}=f_\theta(h_t,x_t)+w_t,\qquad w_t\in\mathcal W,$$
- $$z_t=Ph_t,\qquad z_{t+1}\approx A_tz_t+B_tx_t+d_t+\epsilon_t,$$
- $$\mathcal Z_{t+1}=A_t\mathcal Z_t\oplus B_t\mathcal X_t\oplus\{d_t\}\oplus\mathcal E_t,$$
- $$\mathcal L_{\mathrm{tube}}=\frac{1}{T}\sum_{t=1}^{T}\max\left(0,\operatorname{dist}(Ph_t,\mathcal Z_t)-\delta\right)^2.$$

## Implementation notes

1. Integration point: use a GRU, vanilla RNN, or linear state-space model with hidden state $h_t$; retain the ordinary task loss and add the reachable-tube penalty. Compute a two-dimensional projection $P$ by PCA on hidden trajectories from the training set, and optionally compare it with two random orthogonal rows. Maintain an axis-aligned interval $\mathcal Z_t=[\ell_t,u_t]$ in reduced coordinates. 2. Pseudocode: initialize $\mathcal Z_0$ around $Ph_0$; execute the recurrent update; obtain local Jacobians $A_t=\partial(Pf_\theta)/\partial z$ and $B_t=\partial(Pf_\theta)/\partial x$ using autodiff; propagate interval bounds with interval arithmetic, adding the known input box $\mathcal X_t$ and error box $\mathcal E_t$; calculate $r_t=\operatorname{dist}(Ph_t,\mathcal Z_t)$; add $\max(0,r_t-\delta)^2$ to the loss. A correction variant projects $Ph_t$ to the nearest point in $\mathcal Z_t$ before the next update. 3. The paper-derived operations are reduced-order reachable-set propagation and set-membership testing. Estimate $P$, Jacobians, disturbance bounds, and linearization-error bounds empirically; choose $\mathcal E_t$ to cover 99% of training one-step residuals, then inflate it by a factor $\kappa$. 4. First experiment: train a 64-unit RNN on sequential MNIST and noisy sine-wave prediction, comparing the baseline, tube penalty, and projection correction under clean and perturbed inputs while sweeping recurrent spectral radius. The quantitative prediction is that violation frequency remains below 1% at the calibrated noise level and rises sharply when the effective local spectral radius crosses $\rho(A_t)=1$. The first tube violation should precede loss explosion or long-horizon error growth by at least one prediction horizon.

## Disclaimer

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