Once-Reinforced Random Walk

• Once-Reinforced Random Walk is a self-interacting process where each edge's weight permanently increases after its first traversal.
• It exhibits sharp phase transitions in recurrence, transience, and spatial growth depending on graph geometry and reinforcement parameters.
• Analytical methods like martingale techniques, capacity estimates, and variational formulas provide critical insights into its asymptotic behavior.

A once-reinforced random walk (ORRW) is a paradigmatic self-interacting stochastic process where the propensity to cross an edge increases permanently after the edge is traversed for the first time. This model interpolates between simple random walk and various reinforced or self-avoiding models and exhibits nontrivial phase transitions in recurrence, transience, range growth, and limiting shape, depending on the graph geometry and reinforcement parameter. ORRW has been systematically studied on the integers, trees, various lattices, general graphs, and oriented graphs. Recent advances employ multiscale capacity methods, martingale and change-of-measure techniques, and variational formulas to analyze its asymptotic properties.

1. Formal Definition and Basic Structure

The standard discrete-time ORRW on a locally finite undirected graph $G = (V, E)$ with reinforcement parameter $\delta > 0$ evolves as a nearest-neighbor process $(X_n)_{n \geq 0}$ with the following transition rule. Let $E_n$ be the set of edges crossed up to time $n$:

$$E_n = \{ \{u, v\} \in E : \exists k \leq n, \{X_{k-1}, X_k\} = \{u, v\} \}.$$

The probability of moving from $x$ to $y \sim x$ at time $n+1$ is:

$$\mathbb{P}[ X_{n+1}=y \mid \mathcal{F}_n ] = \frac{1 + \delta\,\mathbf{1}_{\{\{x, y\} \in E_n\}}}{\sum_{z \sim x} (1 + \delta\,\mathbf{1}_{\{\{x, z\} \in E_n\}})}.$$

Each edge's weight changes from $1$ to $1+\delta$ upon first traversal and remains at $1+\delta$ thereafter. This mechanism makes the process highly non-Markovian when viewed solely in terms of $(X_n)$, as transition probabilities depend on the history of edge usage.

In continuous time, the analogous process may define jump rates for each edge $e = \{i, j\}$ as $a$ if untraversed and $1$ if traversed at least once, emphasizing once-only reinforcement (Collevecchio et al., 4 Sep 2025).

2. Recurrence, Transience, and Phase Transitions

The recurrence or transience of ORRW is highly sensitive to both the underlying graph and the reinforcement parameter.

• On $\mathbb{Z}$ and $\mathbb{N}$: Any strictly positive reinforcement preserves recurrence due to electrical network arguments and martingale potentials. ORRW on the line or half-line remains recurrent for all $\delta > 0$ (Amir et al., 2015).
• On trees: A critical phenomenon arises on trees with polynomial growth. For a tree $T$, the branching-ruin number $br_r(T)$ quantifies the growth rate. The phase transition is sharp:

$\delta_c = br_r(T),$

so that the ORRW is recurrent if $\delta > br_r(T)$ and transient if $\delta < br_r(T)$. On regular (exponential growth) trees, $br_r(T) = 0$ and ORRW is always recurrent [$1710.00567$, $1604.07631$].

• On $\mathbb{Z}^d$-like trees: The critical line is explicit: for the canonical $\mathbb{Z}^d$-like trees, recurrence occurs for $a > \log_2 d$ and transience for $a < \log_2 d$ (Kious et al., 2016).
• On Euclidean lattices $\mathbb{Z}^d$: The question is subtle. For $d \geq 6$ and sufficiently small reinforcement, ORRW is provably transient and diffusive, converging to Brownian motion under diffusive scaling. The process is predicted to exhibit a recurrence/transience phase transition as reinforcement varies, but the precise critical parameter remains unknown for $d = 3, 4, 5$ [$2601.17972$]. On non-amenable graphs, small reinforcement ensures transience (Collevecchio et al., 4 Sep 2025).
• Adaptive versus non-adaptive environments: Adaptivity of reinforcement is essential for nontrivial behavior; nonadaptive reinforcement cannot break recurrence on recurrent graphs (Amir et al., 2015).

3. Range, Shape, and Large Deviation Results

The range and spatial profile of ORRW display rich and model-dependent fluctuations.

• Shape Theorem and Fluctuations on $\mathbb{Z} \times \Gamma$: For reinforced walks on cylinder graphs $\mathbb{Z} \times \Gamma$ (with $\Gamma$ finite), high reinforcement ($\delta \geq C|\Gamma|^{40}$) produces recurrence, with the set of visited sites growing linearly in both directions along $\mathbb{Z}$, with polynomially controlled fluctuations of width $n^{1/\delta^{1/8}}$ at time where $2|\Gamma| n$ sites have been visited (Kious et al., 2018).
• Range Growth in $\mathbb{Z}$: The law of the range admits explicit scaling limits; $\mathbb{E}[R_n]/\sqrt{n} \sim \sqrt{2/\pi} J_1(c)$ with $J_1(c)$ an explicit integral, and all moments have precise asymptotics derived from generating-function and Tauberian analysis (Pfaffelhuber et al., 2019).
• High dimensions ($d \geq 2$): For general graphs, probabilistic change-of-measure formulas yield large deviation bounds for the size of the range, with exponents depending explicitly on dimension and reinforcement (Collevecchio et al., 4 Sep 2025).
• No Deterministic Asymptotic Shape in Oriented Models: In oriented settings, there is provably no deterministic limiting shape for the visited set; the walk's boundary retains stochastic fluctuations at macroscopic scale (Collevecchio et al., 4 Sep 2025).

4. Empirical Measures, Cover Times, and Large Deviations

• Empirical Occupation: On finite graphs, the empirical occupation measure $L^n = \frac{1}{n} \sum_{i=0}^{n-1} \delta_{X_i}$ for ORRW satisfies a Large Deviation Principle (LDP) with a variational rate function $I_\delta$. For $\delta \leq 1$, the rate does not change with reinforcement, but for $\delta > 1$, rare-event paths are exponentially favored toward repetitively traversing known edges, resulting in slower exploration and a strictly lower rate function. This behavior induces a non-differentiability at the critical point $\delta_c = 1$ (Huang et al., 2022).
• Edge Cover Time: The edge-cover time $C_E$ (the time to visit every edge at least once) for ORRW on finite graphs has exponential tail decay:

$$\mathbb{P}(C_E > n) \sim \exp(-\alpha_c^1(\delta) n),$$

where $\alpha_c^1(\delta)$ is characterized by a variational formula involving the relative entropy of invariant kernels and is strictly decreasing in $\delta$. For star and triangle graphs, as $\delta \to 0$ the exponent diverges, while for other graphs with shielding structure, the $\delta \to 0$ limit is finite (Huang et al., 8 May 2025).

5. Analytical Techniques and Key Proof Ingredients

The analysis of ORRW relies on a confluence of tools:

• Electrical Network Theory: Martingale potentials, resistance arguments, and the use of Thomson's principle facilitate recurrence and transience proofs on lines and trees (Kious et al., 2018, Amir et al., 2015).
• Martingale and Regeneration Structures: Martingale constructions (e.g., for position plus cumulative drift in horizontally reinforced cylinders) and regeneration times (in transient tree settings) provide tight control over excursions and fluctuation bounds (Kious et al., 2018, Zhang, 2012, Collevecchio et al., 2017).
• Capacity and Multiscale Induction: In high dimensions, capacity estimates and spatial independence via the “demon” device enforce "nowhere heaviness" of the path and allow concatenation arguments essential for diffusive scaling and coupling to Brownian motion (Elboim et al., 25 Jan 2026).
• Change-of-Measure Formulas: Recent work develops Radon–Nikodym derivatives that express ORRW probabilities relative to simple random walk, enabling quantitative large deviation bounds and local time representations obeying tree-like, “polymer” formulas (Collevecchio et al., 4 Sep 2025).
• Variational and Weak-Convergence Methods: LDPs for empirical measures and occupation statistics are proven by lifting to Markovian processes on oriented edge spaces and using convex analysis and entropy techniques (Huang et al., 2022, Huang et al., 8 May 2025).
• Generating Functions and Tauberian Theory: In one dimension, closed-form generating functions for the range, coupled with Tauberian theorems, yield explicit asymptotics for the distribution and moments of the range (Pfaffelhuber et al., 2019).

6. Generalizations, Limitations, and Open Problems

Significant extensions and limitations include:

• Generalized Reinforcement and Non-Markovian Self-Interaction: Many results extend to generalized ORRW models with variable, possibly biased reinforcement or to walks in random and dynamically evolving environments, provided once-only monotonicity and quasi-independence hold [$1710.00567$].
• Comparison with Linearly Reinforced Random Walks (LRRW): The one-time nature of ORRW reinforcement makes it more tractable than linear reinforcement, where transience/recurrence criteria, especially on trees of higher degree, remain less understood and large deviation principles may fail or have different tail behaviors (Zhang, 2012).
• Phase Transition in High-Dimensional Lattices: Precise determination of critical parameters for recurrence/transience on $\mathbb{Z}^d$ for $3 \leq d \leq 5$ is open, as is mathematical understanding of the “compact” phase for large reinforcement in high dimensions (Elboim et al., 25 Jan 2026).
• Limiting Shape and Range Growth: For oriented or non-Euclidean variants, deterministic limiting shapes fail to emerge, and the exact fluctuation exponents in general graphs are not known (Collevecchio et al., 4 Sep 2025, Kious et al., 2018).
• Open Technical Questions: These include explicit computation of the rate functions in large deviation results, positive recurrence in recurrent regimes, monotonicity properties of the speed as a function of reinforcement, and extension of capacity methods to lower dimensions.

ORRW thus serves as a robust model for the analysis of self-interacting, history-dependent random processes, illuminating the interplay between geometry, reinforcement, and stochastic dynamics across a spectrum of graph theoretic settings.