Random Walk in Changing Environments

• Random Walk in Changing Environments (RWCE) is a stochastic process defined on time-varying graphs where edge weights or transition rules evolve either exogenously or based on the walker’s history.
• Studies on RWCE establish rigorous recurrence, transience, and mixing criteria using electrical network methods, distinguishing adaptive and nonadaptive dynamics on structures like ℤ² and trees.
• RWCE models find applications in domains such as self-interacting random walks, distributed algorithms, and genetic ancestry, providing insights into long-term behavior in dynamic settings.

A Random Walk in Changing Environment (RWCE) is a generalized stochastic process in which the substrate of the random walk—most commonly graph edge-weights or transition rules—evolves over time, either exogenously or as a function of the walk's trajectory. This broad class encompasses self-interacting walks (e.g., reinforced random walk), dynamically evolving conductance models, random walks in dynamic random environments, and processes on evolving graphs. The study of RWCE includes fundamental questions of recurrence, transience, limit theorems, and mixing properties, with rigorous criteria now available for several prominent subclasses.

1. Formal Definitions and Principal Classes

Let $G = (V,E)$ be a locally finite, connected graph. An RWCE is defined via a sequence of (possibly random) edge-weight functions $\{C_t : E \to [0, \infty) \}_{t \geq 0}$ (conductances), or equivalently time-dependent weighted graphs $G_t = (V, E, C_t)$ (Amir et al., 2015, Park et al., 2024). The law of the walk $\{X_t\}$ is specified by

$$\mathbb{P}[ X_{t+1}=y \mid \mathcal{F}_t ] = \frac{C_t(X_t, y) \mathbf{1}_{ \{ y \sim X_t \} } }{\sum_{z \sim X_t} C_t(X_t, z)},$$

where $\mathcal{F}_t$ denotes the history up to time $t$.

Key subclasses:

• Adaptive RWCE: The evolution of $G_{t+1}$ may depend on the walk's past $X_0, \dots, X_t$ (e.g., once-reinforced RW, bridge-burning RW).
• Non-adaptive RWCE: The sequence $\{G_t\}$ is determined independently of the walk (Amir et al., 2015).
• Monotone RWCE: Edge weights are monotonic in $t$; monotone-increasing (reinforcement) and monotone-decreasing (depletion) variants both arise.
• Bounded RWCE: For all $t$ and $e$, $$0 < C_{\min}(e) \leq C_t(e) \leq C_{\max}(e) < \infty$$ (Park et al., 2024).

The properness condition requires all nonzero edge weights to remain finite at all times.

2. Recurrence and Transience Criteria on Trees and $\mathbb{N}$

In the absence of cycles (e.g., on $\mathbb{N}$ or a tree), recurrence/transience can be characterized by classical electrical network criteria applied to the initial or limiting (supremal/infimal) edge-weight configurations and the monotonicity of the evolution (Amir et al., 2015, Park et al., 2024). The main theorems are as follows:

• Monotone increasing, bounded-above RWCE is recurrent if and only if the limiting conductance profile is recurrent: $\sum_{i=0}^\infty 1/w_\infty(i) = \infty$.
• Monotone increasing, bounded-below RWCE is transient if the initial conductance profile is transient: $\sum_{i=0}^\infty 1/w_0(i) < \infty$.
• Monotone decreasing, bounded-below RWCE is transient if the lower bound graph is transient.
• Monotone decreasing, bounded-ratio RWCE is recurrent if the initial graph is recurrent (Amir et al., 2015).

On trees, analogous theorems hold, with effective resistances along unique paths controlling behavior (Amir et al., 2015).

For general locally finite connected graphs, Park and Ray proved that if the time-total variation of resistances is summable across all edges,

$$\sum_{t=0}^{\infty} \sum_{e \in E} |R_t(e) - R_{t+1}(e)| < \infty,$$

then the RWCE inherits the recurrence or transience of the initial weighted graph $(G, C_0)$ (Park et al., 2024). This encompasses slowly changing RWCEs and recovers all the above as corollaries under bounded monotonicity.

3. RWCEs on Graphs with Cycles and Adaptive Effects

When $G$ contains cycles (e.g., $\mathbb{Z}^2$), the above recurrence/transience criteria can fail, especially for adaptive RWCEs. For instance, one can construct a monotone adaptive walk (MAW) on $\mathbb{Z}^2$, where horizontal edge weights are increased dependent on the walk's history, yielding a process which is transient despite all fixed-weighted configurations being recurrent (Amir et al., 2015). This demonstrates the potential for adaptive mechanisms to fundamentally change walk behavior beyond what is possible in static or nonadaptive environments.

However, it is conjectured that this transition to transience on recurrent graphs like $\mathbb{Z}^2$ cannot occur in the nonadaptive, monotone, bounded setting; such RWCE would always remain recurrent—a dichotomy between adaptive and nonadaptive environmental changes.

4. Link to Self-interacting and Excited Walks

Many classical self-interacting random walks can be cast as RWCEs:

• Once-reinforced random walk: Edge conductances jump from 1 to $c > 1$ upon first traversal.
• Bridge-burning walk: Conductance of an edge drops to 0 upon use, removing the edge.
• True self-avoiding walk with bond repulsion: Conductances evolve as $w_t(e) = \exp(-c\cdot k_t(e))$ based on visit counts.

In excited or deterministic environments (e.g., deterministic cookies per site), growth remains sublinear and large deviations exhibit exponential tails. The walk's diffusion properties are tightly linked to the environment's updating rule and initial data (Matic et al., 2014).

5. Dynamic Random Environment: Mixing, Limit Theorems, and Structure

If the environment itself evolves stochastically—either independently (random walk among random walks, random walks on evolving graphs), via hidden Markov processes, or by interacting particle systems—the analysis must blend classical random walk theory with the temporal mixing properties of the environment.

• In edge-Markovian evolving graphs, mixing times of the RWCE are controlled by both the stationary edge density (relative to the connectivity threshold) and the rate of environmental evolution. Too rapid a change or too sparse instantaneous connections can preclude mixing, while appropriate density and slow changes recover static-graph mixing rates (Cai et al., 2020).
• In one-dimensional dynamic random environments, annealed and quenched limit theorems, ballisticity criteria, and large deviation principles can be established—often through multiscale renormalization, regeneration times, and space-time decoupling arguments (Hilário et al., 2014).
• For general Markovian or weakly mixing environments, existence, uniqueness, and mutual absolute continuity of stationary measures for the "environment seen from the walker" have been proved under various quantitative and qualitative mixing conditions, including slow and polynomial mixing (Bethuelsen et al., 2016, Redig et al., 2012).
• Regeneration structures yield SLLN, CLT, and large deviation bounds in slowly mixing or slowly changing environments, provided key space-time decoupling assumptions can be verified (Birkner et al., 2015, Hilário et al., 2014).

6. Applications and Extensions

RWCE models appear throughout probability theory and applications, including:

• Branching random walks with barriers in random environments: Survival/extinction thresholds and precise decay rates are governed by associated RWCEs (Lv et al., 2022).
• Population and genetic ancestry models with local regulation: The spatial embedding of lineages corresponds to RWCEs in time-reversed dynamic environments (Birkner et al., 2015).
• Distributed algorithms in dynamic networks: Protocol efficiency is directly impacted by the mixing properties of random walks on edge-Markovian evolving graphs (Cai et al., 2020).
• Interacting particle systems: Random walks coupled to (e.g.) contact processes or exclusion processes showcase how environmental spectral gap and ergodicity drive walker behavior and absorption times (Masi et al., 2013).

Further foundational directions concern the extension of martingale and electrical-network approaches to less restrictive variation conditions, as well as deeper exploration of decoupling under local rather than global metrics (Park et al., 2024).

7. Summary Table: Key Recurrence/Transience Criteria

Setting	RWCE Property	Recurrence/Transience Character
$\mathbb{N}$, tree	Monotone increasing, bounded	Recurrence iff limit conductance recurrent (Amir et al., 2015)
$\mathbb{N}$, tree	Monotone increasing, bounded below	Transience iff initial conductance transient (Amir et al., 2015)
General graph	Bounded, $\ell^1$-finite resistance increments	Same as initial graph (Park et al., 2024)
$\mathbb{Z}^2$	Adaptive monotone, bounded	Transience possible (MAW) (Amir et al., 2015)
$\mathbb{Z}^2$	Nonadaptive monotone, bounded	Conjectured always recurrent (Amir et al., 2015)

This structure underpins the rigorous understanding of random walks in time-dependent environments, offering a pathway to both proof and computation of precise stochastic properties in dynamically structured spaces.