Shadowing-Constrained Latent Rollouts
Implementation & benchmark of arXiv:2607.17325 — Shadowing property and transitivity of a set-valued map and its inverse limit
Source paper: Shadowing property and transitivity of a set-valued map and its inverse limit arXiv:2607.17325 ⓘ · analyzed Aug 30, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Replace a deterministic latent transition with a set-valued relation consisting of all next states within a learned tolerance of the predicted transition, and train the model so noisy or approximate latent rollouts are shadowed by valid exact trajectories. Use forward and inverse-limit consistency losses to make the same robustness property visible in finite sequence windows.
Formulas
Mathematical statement
Let $(X,d)$ be a compact latent metric space and let $F:X\to 2^X$ be a set-valued transition relation. A finite sequence $(x_0,\ldots,x_T)$ is a $\delta$-pseudo-orbit if $d(x_{t+1},F(x_t))\leq\delta$ for every $t$, where $d(y,F(x))=\inf_{z\in F(x)}d(y,z)$. The map has the $(\delta,\varepsilon)$ shadowing property if every such pseudo-orbit is uniformly close to an exact orbit $(z_t)$ satisfying $z_{t+1}\in F(z_t)$ and $d(x_t,z_t)\leq\varepsilon$ for all $t$. The paper proves, under surjectivity and upper semicontinuity, that shadowing of $F$ is equivalent to shadowing of the shift on the generalized inverse-limit space $\varprojlim F=\{(x_0,x_1,\ldots):x_t\in F(x_{t+1})\}$. In the neural adaptation, $F_\theta(x)=\{g_\theta(x)+r:r\in R_\theta(x)\}$ is a learned uncertainty set, and the practical certificate is bounded projection distance from a noisy rollout to a valid relation-consistent trajectory.
Implementation notes
1. Exact integration point: use a recurrent state-space model or world model with latent state $h_t$. Let the network output a mean transition $g_\theta(h_t,u_t)$ and positive diagonal radii $\rho_\theta(h_t,u_t)$; define $F_\theta(h_t,u_t)$ as an axis-aligned ellipsoid or box centered at $g_\theta$ with those radii. Train on encoded latent transitions, adding the relation loss in the formula and a reverse-window loss obtained by fitting the inverse relation, so every observed state is compatible with at least one predecessor.
2. Pseudocode: encode a length-$T$ sequence to $h_{0:T}$; compute $c_t=g_\theta(h_t,u_t)$ and radii $\rho_t$; compute normalized violation $v_t=\max(0,\|h_{t+1}-c_t\|_{M_t}-\rho_t)$; minimize task loss plus $\lambda\sum_t v_t^2$. For robustness testing, perturb each state by Gaussian noise or inject transition defects of size $\delta$, then optimize a projected trajectory $z_{0:T}$ using the second displayed objective. Report the maximum correction $\max_t d(h_t,z_t)$.
3. Computed versus estimated: relation distances and the projection objective are directly computed. The paper supplies the qualitative equivalence between forward-map and inverse-limit shadowing, but not a numerical $\delta(\varepsilon)$ bound; estimate the empirical shadowing curve by sweeping defect magnitude $\delta$ and measuring the smallest correction radius $\varepsilon$ on held-out sequences. Enforce approximate surjectivity by retaining a nonzero predecessor density and rejecting latent states with no compatible predecessor.
4. First cheap experiment: train a GRU or diagonal SSM on sequential MNIST, permuted-MNIST, or a synthetic nonlinear oscillator, comparing a deterministic baseline against the set-valued model with and without $\mathcal L_{\mathrm{rel}}$. At test time inject independent transition perturbations of size $\delta$ for horizons 10, 50, and 100. The quantitative prediction is a shadowing regime: below a critical defect scale $\delta_c(\varepsilon,T)$, projection correction remains bounded and grows weakly with horizon, whereas above it correction increases sharply or optimization fails. Forward rollout and inverse-limit projection should show the same transition within 20%; this agreement is the primary test, not merely final task accuracy.
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.