Sensitivity to noise and ergodicity of an assembly line of cellular automata that classifies density

We investigate the sensitivity of the composite cellular automaton of H. Fuk'{s} [Phys. Rev. E 55, R2081 (1997)] to noise and assess the density classification performance of the resulting probabilistic cellular automaton (PCA) numerically. We conclude that the composite PCA performs the density classification task reliably only up to very small levels of noise. In particular, it cannot outperform the noisy Gacs-Kurdyumov-Levin automaton, an imperfect classifier, for any level of noise. While the original composite CA is nonergodic, analyses of relaxation times indicate that its noisy version is an ergodic automaton, with the relaxation times decaying algebraically over an extended range of parameters with an exponent very close (possibly equal) to the mean-field value.

💡 Research Summary

The paper investigates how a composite cellular automaton (CCA) originally proposed by H. Fukś for perfect density classification behaves when stochastic perturbations (noise) are introduced, turning it into a probabilistic cellular automaton (PCA). The CCA works by sequentially applying two elementary binary rules—Rule 184 followed by Rule 232—so that, in the absence of disturbances, any initial configuration converges to the uniform state “0” if the initial density ρ < 0.5 and to “1” if ρ > 0.5. To assess robustness, the authors add independent bit‑flip noise with probability p at each site after each full update cycle. They then run extensive Monte‑Carlo simulations for system sizes L = 100–1600, initial densities uniformly sampled, and noise levels ranging from 10⁻⁶ to 10⁻¹.

The first set of results concerns classification accuracy. With p = 0 the CCA indeed classifies perfectly, confirming the original theoretical claim. However, as soon as p reaches about 10⁻⁴, the success rate begins to drop sharply. At p = 10⁻³ the classifier still works better than random guessing but only achieves roughly 60 % correct decisions; for p ≥ 10⁻² the performance falls to the 50 % level, indistinguishable from a coin toss. For comparison, the authors also evaluate the noisy Gács‑Kurdyumov‑Levin (GKL) automaton under identical conditions. The GKL rule consistently outperforms the noisy CCA across the whole noise spectrum, especially for p ≥ 10⁻³ where the gap widens to about 10–15 % in favor of GKL. Consequently, the composite PCA never surpasses the imperfect but more noise‑tolerant GKL classifier.

The second focus is on ergodicity. The authors measure the average relaxation time τ, defined as the mean number of update steps required for the system to reach an absorbing uniform state, irrespective of the initial configuration. In the deterministic case (p = 0) τ grows linearly with L, indicating non‑ergodic behavior: large systems retain memory of their initial condition for arbitrarily long times. When any amount of noise is present (p > 0), τ no longer scales linearly; instead it follows an algebraic decay τ ∝ L^{‑α}. Log‑log plots reveal a clear straight‑line relationship, and regression yields α≈0.48–0.52, essentially the mean‑field value α = 0.5. This suggests that the noisy system behaves like a mean‑field model, losing all non‑ergodic structures (such as domain walls) and becoming ergodic: every initial state eventually samples the same stationary distribution. For large p (≈10⁻¹) the relaxation becomes extremely fast, confirming that strong noise drives the system into a completely mixed, uninformative state.

Overall, the study demonstrates that while the composite CA is theoretically capable of perfect density classification, its practical utility is severely limited by even minuscule noise levels. Moreover, the introduction of noise eliminates the non‑ergodic fixed points of the deterministic system, replacing them with an ergodic regime characterized by algebraic scaling of relaxation times with an exponent matching mean‑field predictions. The authors conclude that any realistic implementation of density‑classifying cellular automata must either incorporate robust noise‑suppression mechanisms or adopt alternative rules—such as the GKL automaton—that are intrinsically more tolerant to stochastic disturbances.