Abstract
This paper explores the mathematical mapping between a random walk in a disordered environment and a random walk with memory. By utilizing Bayesian inference to update transition probabilities based on a particle's history, the authors demonstrate that quenched disorder is functionally equivalent to a dynamical retrieval of memory. This framework provides a natural explanation for complex physical phenomena such as aging, sub-diffusion, and the peculiar localization properties of random walkers in disordered media.
Introduction: The Challenge of Disordered Media
In statistical physics, studying diffusion in disordered media—such as a particle navigating a porous rock or a signal traversing a noisy network—traditionally requires a "double averaging" process. Researchers must first average over the stochastic paths a particle takes in a specific, fixed environment (quenched dynamics) and then average those results over all possible realizations of the disorder.
This approach is conceptually cumbersome and computationally expensive. Furthermore, it often leads to "aging" effects, where the system's behavior depends on its history, breaking standard time-translational invariance. This paper proposes a way to bypass pre-defined disorder by using "annealed" dynamics, where the environment is effectively generated on-the-fly by the particle's own movement history.
Methodology: Mapping Disorder to Memory
The authors establish an equivalence between a 1D Random Walk in Random Environment (RWRE) and a Random Walk with Memory (RWM) through a Bayesian mapping.
Bayesian Inference and Polya Urns
Instead of assigning a fixed, unknown probability $p_i$ to each site $i$, the observer updates their knowledge of the site based on the walker's behavior. If a walker has visited site $i$ a total of $n$ times and moved to the right $k$ times, the "effective" transition probability is derived using Bayes' rule:
$$p^a_{n,k} = \frac{k+1}{n+2}$$
This formula is identical to a Polya Urn process, where the history of previous "successes" (rightward jumps) increases the likelihood of future successes. This suggests that the walker doesn't "feel" a pre-existing landscape but rather "builds" one based on its history.
Generalizing the Dynamics
To explore different regimes of disorder, the authors introduce a parameter $a$ into the transition rule:
$$p^a_{n,k} = \frac{k+1}{n+2} + a \sin\left(2\pi \frac{k+1}{n+2}\right)$$
This generalized model allows the system to be tuned between localization (trapping), standard diffusion, and the "marginal" case of Sinai diffusion.
Key Results and Phase Transitions
The RWM model successfully reproduces the statistics of random walks in random environments without requiring the disorder to be specified upfront. The process self-selects a specific realization of disorder consistent with its own history—a phenomenon termed "dynamical retrieval."
The system exhibits distinct phases based on the parameter $a$:
- $a > 0$ (Standard Diffusion): The walker eventually ignores its memory, and the transition probabilities converge toward $p \to 1/2$, leading to normal diffusive behavior.
- $a < 0$ (Localization): The walker becomes trapped on specific sites as the retrieved disorder distribution develops sharp peaks at the boundaries ($p \to 0$ or $p \to 1$).
- $a = 0$ (Sinai Diffusion): This is the critical marginal case where the retrieved disorder distribution is continuous and uniform. The model correctly recovers the famous Sinai scaling law, where displacement grows incredibly slowly as $(\ln t)^2$.
Discussion and Future Directions
This memory-based framework offers a more intuitive explanation for "aging" in disordered systems. Aging is not an anomaly but a natural consequence of transition probabilities evolving as the walker gains more "experience" at each site.
Beyond theoretical physics, the authors suggest that this approach could have applications in optimization. Instead of fixing a complex disorder and searching for an optimal state (as in Simulated Annealing), one could allow the dynamics to "grow" a consistent disorder realization. While the current work is primarily 1D, extending this mapping to higher dimensions and non-uniform initial distributions remains a promising area for future research.
Summary updated