{
 "artifacts": null,
 "category": "regularization",
 "description": "Use the random-attractor construction as a training and inference diagnostic: initialize latent trajectories far in the past with different states but the same recent noise sequence, then measure whether they contract toward the same current set. This detects whether a stochastic recurrent model has a bounded, reproducible random attractor or instead exhibits discretization-induced divergence and spurious long-term modes.",
 "formulas_latex": [
  "$$A_{n,h}(\\omega)=\\Phi_h^n(\\theta_{-n}\\omega,B),$$",
  "$$\\Phi_h(\\omega,A_h(\\omega))=A_h(\\theta_1\\omega),\\qquad \\operatorname{dist}\\!\\left(A_{n,h}(\\omega),A_h(\\omega)\\right)\\xrightarrow[n\\to\\infty]{}0,$$",
  "$$D_{n,h}(\\omega)=\\max_{i,j}\\left\\|\\Phi_h^n(\\theta_{-n}\\omega,z_i)-\\Phi_h^n(\\theta_{-n}\\omega,z_j)\\right\\|\\approx D_{0,h}e^{\\lambda_hnh}\\quad(\\lambda_h\u003c0).$$"
 ],
 "id": 2811,
 "implementation": "Apply this monitor to a stochastic RNN, state-space world model, or diffusion latent transition model during validation and optionally as a regularizer. The same noise realization must be replayed across different initial states; otherwise ordinary noise variability is confused with pullback attraction. Sample a noise block $\\omega_{-n:-1}$ and $K$ initial states $z_{-n}^{(i)}$ from a fixed bounded ball. Run every state through the identical sequence of stochastic updates and compute the pairwise diameter, or an RMS approximation, after every step. Repeat for several noise blocks and record $D_{n,h}$. If a stable attractor is desired, add $L_{\\mathrm{attr}}=\\sum_{r=r_0}^{R}[\\log D_{r+1,h}-\\log D_{r,h}-\\widehat\\lambda_hh]_+^2$, where $[x]_+=\\max(x,0)$. Trajectories and diameters are directly computed; the attractor is approximated by the terminal point cloud $A_{R,h}$; the decay exponent is fitted by regressing $\\log D_{n,h}$ against $nh$. First experiment: use a two-dimensional noisy RNN on a synthetic damped oscillator with 32 initial states and 128-step histories, comparing Euler-Maruyama at several step sizes with a semi-implicit update. The prediction is exponential pullback contraction whose fitted slope matches the independently measured $\\widehat\\lambda_h$. Near the stability boundary, the contraction time should grow approximately as $1/|\\lambda_h|$; a coarse solver with a spurious positive exponent should show persistent or increasing diameter.",
 "math_summary": "Let $\\Phi_h(\\omega,z)$ be the stochastic neural one-step map, $\\theta_{-n}\\omega$ denote the noise sequence shifted $n$ steps into the past, and $B$ be a bounded set of initial latent states. The pullback image is $A_{n,h}(\\omega)=\\Phi_h^n(\\theta_{-n}\\omega,B)$. A random attractor is a family of compact sets $A_h(\\omega)$ satisfying invariance $\\Phi_h(\\omega,A_h(\\omega))=A_h(\\theta_1\\omega)$ and pullback attraction $\\operatorname{dist}(A_{n,h}(\\omega),A_h(\\omega))\\to0$. A negative top Lyapunov exponent predicts exponential contraction of the image diameter, $D_{n,h}\\approx D_{0,h}e^{\\lambda_hnh}$, until the attractor's intrinsic diameter is reached.",
 "math_tags": [
  "dynamical-systems",
  "stochastic-processes",
  "numerical-analysis",
  "control-theory"
 ],
 "ml_areas": [
  "rnn",
  "ssm",
  "world-model",
  "training-dynamics",
  "regularization"
 ],
 "paper": {
  "arxiv_id": "2608.29149",
  "arxiv_url": "https://arxiv.org/abs/2608.29149",
  "summary_what_math_gives_to_ml": "The paper provides a transferable numerical-dynamics mechanism: a discretization of a stochastic continuous-time system can preserve both its random attractor and the sign of the almost-sure Lyapunov exponent of a distinguished invariant solution. This is stronger than ordinary numerical convergence because it preserves qualitative stability classification, not merely finite-time trajectories. The most promising neural-network use is a structure-aware discretization and stability monitor for stochastic neural ODEs, recurrent state-space models, or diffusion-like latent dynamics, with the step size restricted until the discrete Lyapunov classification agrees with the continuous-time estimate.",
  "title": "Random attractors and almost-sure stability under discretization of a stochastic autoparametric system",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 4,
  "novelty": 8,
  "usefulness": 7
 },
 "solves": [
  "stability",
  "accuracy",
  "generalization"
 ],
 "title": "Pullback random-attractor monitor",
 "url": "https://synthcore.org/idea/2811/pullback-random-attractor-monitor",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)"
 }
}