# Continuous Ergodic Projection for Recurrent States

- ID: 2281
- Canonical URL: https://synthcore.org/idea/2281/continuous-ergodic-projection-for-recurrent-states
- API JSON: https://synthcore.org/api/idea/2281.json
- API Markdown: https://synthcore.org/api/idea/2281.md
- Verification status: unverified
- Source: [arXiv:2608.14175](https://arxiv.org/abs/2608.14175)
- Category: dynamics
- Solves: stability, generalization, accuracy
- ML areas: rnn, ssm, world-model, regularization
- Math tags: dynamical-systems, functional-analysis, measure-theory, probability
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Given a learned recurrent dynamics map, estimate a state-dependent invariant measure from each trajectory and use integration against that measure as a projection onto long-term invariant features. Penalize discontinuities of this projection between nearby states and assign zero mass to trajectories whose feature norms escape, producing a principled distinction between convergent attractors and divergent rollouts.

## Mathematical statement

Let $T_s:X\to X$ be the learned semigroup action indexed by $s\in S$, let $U_s f=f\circ T_s$ be its Koopman operator on continuous observables $f\in C_0(X)$, and let $E(x)$ be the pointwise ergodic map: $E(x)=\mu_x$ when the orbit closure of $x$ supports a unique invariant probability measure and $E(x)=0$ for escaping orbits. The invariant projection of an observable is $Pf(x)=\int_X f(y)\,dE(x)(y)$. The paper identifies weak-star continuity of $E$ together with vanishing at infinity as the condition corresponding to mean ergodicity, meaning that temporal averages converge to $Pf$. In a neural implementation, $E(x)$ is represented by normalized trajectory weights when the rollout remains bounded and by a subprobability mass $m(x)\in[0,1]$ that tends to zero for escaping trajectories; continuity is enforced through Wasserstein or kernel discrepancies between $E(x)$ and $E(x')$ for nearby states.

## Key formulas

- $$E(x)=\begin{cases}\mu_x,&\overline{\{T_sx:s\in S\}}\text{ is compact and uniquely ergodic},\\0,&\text{the orbit escapes to infinity},\end{cases}\qquad Pf(x)=\int_X f(y)\,dE(x)(y).$$
- $$A_Nf(x)=\frac{1}{N}\sum_{k=0}^{N-1}f(T^kx)\ \longrightarrow\ Pf(x),\qquad \|A_Nf-Pf\|_{\infty}\to0$$
- $$\mathcal L_{\mathrm{cont}}=\mathbb E_{x,x':\,d(x,x')<\delta}\left[\frac{W_1(\widehat E(x),\widehat E(x'))}{d(x,x')+\varepsilon}\right],\qquad \mathcal L_{\mathrm{escape}}=\mathbb E_x[(\widehat m(x)-\exp(-\alpha R(x)))^2]$$

## Implementation notes

Use an RNN, state-space model, or learned world model with latent update $z_{t+1}=T_\theta(z_t,u_t)$. Draw two nearby initial latent states $z_0,z'_0$, roll each for $N$ steps, and construct an empirical subprobability measure $\widehat E(z_0)=\widehat m(z_0)\sum_{t=0}^{N-1}w_t\delta_{z_t}$, with $w_t\ge0$ and $\sum_t w_t=1$. A bounded rollout should have mass near one; estimate an escape score $R(z_0)=N^{-1}\sum_t\|z_t\|^2$ or terminal growth rate, and train a mass head toward $\widehat m\approx\exp(-\alpha R)$ after latent-scale normalization. Compute projected features $P_\theta f(z_0)=\widehat m\sum_tw_tf_\phi(z_t)$ and add task loss, $\lambda_c\mathcal L_{\mathrm{cont}}$, and $\lambda_e\mathcal L_{\mathrm{escape}}$. Pseudocode: roll out both states; calculate weights, mass, and projected features; evaluate task loss; calculate a differentiable kernel approximation to $W_1$; backpropagate the total loss. The paper supplies the invariant-projection and continuity mechanism; finite-horizon weights, escape scores, and mass targets are empirical approximations. First test a GRU on Lorenz-63 or a damped oscillator, comparing ordinary teacher forcing with this regularizer. Measure projection residual $\|A_Nf-P_\theta f\|$ and the continuity ratio $W_1(\widehat E(z),\widehat E(z'))/(\|z-z'\|+\varepsilon)$. The prediction is that bounded trajectories have decreasing residual, approximately $O(N^{-1/2})$ under mixing, while escaping trajectories have decreasing mass. On a known uniquely ergodic rotation, the learned projection should match the analytic invariant mean within 10-20%, and the averaging residual should decrease monotonically with horizon.

## Disclaimer

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