Linearly Edge-Reinforced Random Walks

• Linearly Edge-Reinforced Random Walks are self-interacting processes where the likelihood of traversing an edge increases linearly with its past usage, highlighting inherent memory effects.
• Their probabilistic representation as mixtures of reversible Markov chains and RWRE enables rigorous classification of recurrence and transience across various graph structures.
• Detailed phase transitions on lattices, trees, and directed graphs reveal critical thresholds and scaling limits, offering insights into network behavior and optimization.

A linearly edge-reinforced random walk (LERRW) is a self-interacting process on a graph in which the probability that an edge is traversed at a given step is increased linearly with the number of previous traversals of that edge. This class of processes, introduced in a canonical form by Coppersmith and Diaconis (1986), is notable for its connection to mixtures of reversible Markov chains and random walks in random environments (RWRE), its manifestation of phase transitions between recurrence and transience, and its connections to models such as the vertex-reinforced jump process and supersymmetric sigma models. Formal definitions, rigorous classification of recurrence/transience, explicit probabilistic representations, and core asymptotic results are available across a range of settings, including lattice graphs, trees, and random environments.

1. Definition and Core Properties

Given a connected, locally finite (possibly infinite) undirected graph $G = (V, E)$, with positive initial edge weights $a_e > 0$, the linearly edge-reinforced random walk $(X_n)_{n \geq 0}$ is specified by the update mechanism: $$w_n(e) = a_e + \#\{1 \leq k \leq n : \{X_{k-1}, X_k\} = e\}$$ At each step $n$, if $X_n = x$, the probability to move to adjacent $y$ is

$$\mathrm{P}(X_{n+1} = y \mid X_n = x, w_n(\cdot)) = \frac{w_n(\{x, y\})}{\sum_{z \sim x} w_n(\{x, z\})}$$

After traversing edge $\{x, y\}$, its weight increases by 1. This non-Markovian dynamics is partially exchangeable: the law of any path depends on the empirical counts of edge traversals, not their temporal order.

A crucial probabilistic structure is that in recurrent regimes, the path law is a mixture of reversible Markov chains—specifically, a time-homogeneous Markov chain with conductances $W: E \to (0, \infty)$ drawn from an explicit mixing measure $\mu$ (Angel et al., 2012, Sabot et al., 2016, Sabot et al., 2011).

2. Probability Representations: Mixture of Markov Chains and RWRE

LERRW paths on finite (or recurrent) graphs are distributed as Markov chains in a random environment. Explicitly:

• Sample edge weights $W_e$ independently (e.g., $W_e \sim \text{Gamma}(a_e, 1)$ for continuous-time or VRJP representations (Sabot et al., 2011)).
• Conditioned on $W$, the process is a reversible Markov chain with transitions $P_W(x, y) = W(\{x, y\})/\sum_z W(\{x, z\})$.
• The overall law is a mixture over the space of positive edge weights, using a "magic formula" for the stationary (mixing) measure in finite settings (Angel et al., 2012, Bacallado et al., 2021).

In one-dimensional and tree settings, the process further reduces to a random walk in a random environment (RWRE) with i.i.d. Beta or Dirichlet transition probabilities, enabling electrical network and regeneration-time tools (Michel, 2023, Takei, 2020, Sabot et al., 2016).

3. Phase Transitions: Recurrence, Transience, and Critical Regimes

Bounded-Degree Graphs and Phase Transition

• Recurrence for small initial weights: On bounded-degree graphs, there exists $a_0(\Delta) > 0$ such that if all $a_e \leq a_0$, the LERRW is a.s. recurrent (Angel et al., 2012, Michel, 2023). The proof uses exponential decay estimates of the mixing environment away from the starting vertex.
• Transience for large initial weights: On non-amenable graphs (positive Cheeger constant), for large enough $a_e \geq a_1(\Delta, h)$, the LERRW is a.s. transient (Angel et al., 2012). The transience exploits a stabilization argument for edge traversal frequencies, percolation of "good" vertices, and a restriction to subgraphs with anchored expansion.

Lattices and Trees

• Phase transition on $\mathbb{Z}^d$, $d \geq 3$: There is a critical value $a_c(d)$ s.t. for $a_e > a_c(d)$, the walk is transient, while for $a_e < a_c(d)$, recurrence is expected (proved for strips, conjectured in $\mathbb{Z}^2$) (Disertori et al., 2014). This transition was resolved using connections to VRJP and supersymmetric sigma-models.
• Phase transition on trees: On infinite trees, the recurrence/transience criterion is expressed in terms of the branching number $\mathrm{br}(T)$ and critical reinforcement $\Delta_0$ solving $E(\Delta_0) = 1/\mathrm{br}(T)$, where $E(\Delta) = \min_{0 \leq t \leq 1} \mathbb{E}[A^t]$ for the environment ratio $A$ (Michel, 2023).

Half-Line and Critical Exponents

For LERRW on $\mathbb{Z}_+$ with edge weights $x^\alpha \vee 1$ and reinforcement $\Delta$ (Takei, 2020, Hu et al., 22 Jan 2026):

• The walk is recurrent if and only if $\alpha \leq 1$.
• In the recurrent case, the growth rate is:

$$\limsup_{n \to \infty} \frac{X_n}{[K(\alpha, \Delta) \log n]^{1/(1-\alpha)}}=1,\quad K(\alpha, \Delta)>0$$

For $\alpha=1$, further phase transitions in scaling exponents are observed.

4. Asymptotic Behavior and Scaling Results

Large Deviations on Trees

• On regular trees in the transient regime, the upper-tail deviation $\mathbb{P}(h(X_n) \geq (T+\varepsilon)n)$ decays exponentially in $n$, while the lower tail is only polynomial for the linearly reinforced model, differing from once-reinforced walks where both tails decay exponentially (Zhang, 2012).

Invariance Principles and Scaling Limits

• On critical Galton–Watson trees with suitable initial edge weights, the LERRW scales to a diffusion in a random Gaussian snake potential on the scaling limit of the tree. In the recurrent regime ($\alpha \leq 1$), the displacement satisfies detailed scaling exponents, with stretched-exponential or subdiffusive behavior depending on parameters (Andriopoulos et al., 2021).
• In random environments (Dirichlet), the speed and phase transition for positive/zero-velocity regimes are characterized in explicit formulas tied to the offspring distribution and Dirichlet parameters. Conditions for positive speed are sharp, with explicit dependence on the existence of "pipes" or "dead ends" and moment restrictions (Qian et al., 2024).

5. Generalizations, Variants, and Connections

Directed and Higher-Order Reinforced Walks

• The *-edge-reinforced random walk ($*$-ERRW) generalizes the reversible LERRW to non-reversible/Yaglom-reversible mixtures, directed graphs, and higher-order Markov chains. The mixing law extends the magic formula to this broader setting, and partial exchangeability ensures a representation as a mixture over Yaglom-reversible Markov chains (Bacallado et al., 2021).

Multiple Walkers and Interaction

• When multiple reinforcing walkers operate on the same evolving weight environment, behavior remains "all-or-nothing": on $\mathbb{Z}$, either all walkers are recurrent or all have finite range; the edge-proportion process retains martingale properties in certain simplified networks (Michel, 2023).

Biased LERRWs

• Introducing multiplicative or additive bias in transition probabilities induces phase transitions between recurrence and transience at the critical value of the bias parameter. For example, in the additive-bias model on $\mathbb{Z}$, recurrence holds iff $\lambda \leq 1$, with speed positive only if $\lambda > 3$ (Michel, 2023). The RWRE representation remains effective for sharp criteria.

Emergent Path Structures

• In network optimization contexts, path-reinforced random walks with traversed-path-dependent rewards (a generalization of linearly reinforced walks) can lead to emergent shortest-path usages. For strictly decreasing reward functions of path length, only minimum-length paths retain positive limiting probability, and traversal probabilities of longer paths decay algebraically in the number of episodes (Figueiredo et al., 2016).

6. Analytical Techniques and Representations

Sigma-Model Connections and Field-Theoretic Methods

• For LERRW and VRJP on lattice settings, the path measure and fluctuation properties derive from representations involving the supersymmetric hyperbolic sigma model. Ward identities, multiscale induction, and explicit density formulæ for occupation fields are crucial for proving boundedness of fluctuations and deducing transience in higher dimensions (Disertori et al., 2014, Sabot et al., 2011).

Martingale and Urn Techniques

• The edge-weight proportion vector forms a martingale, implying almost sure convergence of edge-usage frequencies. Reduction to Pólya urns for single-vertex or simplified scenarios yields explicit Beta-limit laws and de Finetti-type mixtures (Michel, 2023).

RWRE and Electrical Network Methods

• Embedding into RWRE with Beta/Dirichlet distributed environments enables use of effective resistance, hitting-time estimates, and environmental invariance principles. Electrical network analogies are fundamental in classifying recurrence and in scaling limit proofs (Takei, 2020, Andriopoulos et al., 2021, Sabot et al., 2016).

7. Open Problems and Research Directions

• Fine scaling of LERRW in the transient regime in higher dimensions or on non-tree graphs remains incompletely characterized.
• Extension of the RWRE representation and associated analytical tools to cyclic graphs or non-linear reinforcement mechanisms is only partially understood.
• Relationships to continuous-time counterparts (vertex-reinforced jump processes), supersymmetric sigma models, and random Schrödinger operators offer further structure for both probabilistic and physical inquiry.
• The impact of network heterogeneities, reinforcement tuning, and higher-order memory effects remain rich for further study, especially in random or evolving graph settings.

---

Key references: (Angel et al., 2012, Disertori et al., 2014, Takei, 2020, Andriopoulos et al., 2021, Zhang, 2012, Sabot et al., 2016, Sabot et al., 2011, Figueiredo et al., 2016, Bacallado et al., 2021, Michel, 2023, Qian et al., 2024, Michel, 2023, Hu et al., 22 Jan 2026).