Edge-Reinforced Random Walk (ERRW)

• ERRW is a self-interacting random walk where each edge’s weight increases with traversals, resulting in non-Markovian behavior and dynamic transition probabilities.
• Mixture representations connect ERRW to models like VRJP and random Schrödinger operators, enabling rigorous analysis via the magic formula and operator methods.
• The walk’s behavior varies with reinforcement strength and graph structure, exhibiting regimes of recurrence, transience, and localization in super-linear settings.

Edge-reinforced random walk (ERRW) is a self-interacting random walk on a graph in which transition probabilities depend on past edge traversals: an edge that has been crossed more often becomes more likely to be crossed again. In the standard linearly reinforced model on a locally finite, connected graph $G=(V,E)$ with initial positive edge weights $(a_e)_{e\in E}$, the process is non-Markovian in the vertex variable because the transition kernel evolves with the trajectory. Over the last decade, ERRW has been linked to the Vertex-Reinforced Jump Process (VRJP), to random walks in random environments, to the supersymmetric hyperbolic sigma model, and to random Schrödinger operators; these representations have made possible sharp results on recurrence, transience, phase transitions, scaling limits, and statistical inference (Sabot et al., 2011, Disertori et al., 2014).

1. Definition and reinforcement mechanism

For the standard ERRW, if $X_n$ denotes the walker position at step $n$, then

$$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$

where

$Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$

Thus, each traversal increments the weight of the traversed edge, and the next-step law is proportional to current incident edge weights (Disertori et al., 2014).

This is the linearly edge-reinforced random walk. A related formulation used in one-dimensional and tree settings writes the edge weight after $\ell$ traversals as

$f(\ell,x)=f(0,x)+\ell \Delta,$

with reinforcement parameter $\Delta\ge 0$; on the half-line, for the edge $\{x,x+1\}$, one specific family of initial weights is $(a_e)_{e\in E}$0 (Takei, 2020). In the general infinite-graph literature, the model is often described as losing the Markov property because transition probabilities evolve over time, while still admitting representation as a mixture of Markov chains on suitable state spaces (Michel, 2023).

The central qualitative feature is path dependence. ERRW favors previously used edges, but the effect of that preference depends strongly on the reinforcement regime and on graph geometry. The literature distinguishes at least three asymptotic regimes that should not be conflated: linearly reinforced ERRW, super-linearly reinforced ERRW, and directed or non-reversible generalizations. Their long-term behavior differs sharply, from recurrence to transience to localization on a single attracting edge (Cotar et al., 2015).

2. Mixture representations, the magic formula, and operator methods

A decisive development was the representation of ERRW in terms of VRJP with random conductances. On any locally finite graph, ERRW is equal in law to the discrete-time process associated to VRJP in random conductances $(a_e)_{e\in E}$1, independently for each edge (Sabot et al., 2011). For VRJP, if $(a_e)_{e\in E}$2, the rate to jump to a neighbor $(a_e)_{e\in E}$3 at time $(a_e)_{e\in E}$4 is proportional to

$(a_e)_{e\in E}$5

After a time change, the corresponding rates can be written in the form $(a_e)_{e\in E}$6 (Sabot et al., 2011).

This representation converts ERRW into a random walk in a random reversible environment. Conditionally on the mixing field $(a_e)_{e\in E}$7, the effective conductances are

$(a_e)_{e\in E}$8

and the corresponding Markov chain is reversible with respect to these conductances (Disertori et al., 2014). On finite graphs, the associated mixing measure is the classical “magic formula,” a density involving edge weights, vertex weights, and a spanning-tree factor $(a_e)_{e\in E}$9; via the VRJP connection and a new exponential family, the normalizing constant of this formula can be computed directly, answering a question raised by Diaconis (Sabot et al., 2015).

The same structure has an operator-theoretic form. With $X_n$0 off the diagonal and $X_n$1 on the diagonal, the random Schrödinger operator is

$X_n$2

with Green function $X_n$3 (Sabot et al., 2015). The field defining the VRJP mixing measure satisfies

$X_n$4

which ties the ERRW/VRJP environment to a random potential problem (Sabot et al., 2015). On infinite graphs, a 1-dependent random potential $X_n$5 and a martingale-limit field $X_n$6 enter the mixture representation; in the transient case, $X_n$7 is a positive generalized eigenfunction with eigenvalue $X_n$8 for $X_n$9 (Sabot et al., 2015).

A second decisive bridge is to the supersymmetric hyperbolic sigma model. The limiting measure of the centered occupation field of VRJP can be interpreted as a supersymmetric hyperbolic sigma model, and through the ERRW–VRJP equivalence this imports methods from mathematical physics, including Ward identities and effective-resistance estimates, into reinforced-walk analysis (Sabot et al., 2011).

3. Recurrence, transience, and phase transition

The most prominent structural result is the existence of a nontrivial phase transition on $n$0 for $n$1. There exists $n$2 such that if all $n$3, the ERRW is transient almost surely; this proves transience for small reinforcement, establishes a phase transition between recurrent and transient behavior, and resolves the open problem posed by Diaconis in 1986 (Disertori et al., 2014). The proof adapts the quasi-diffusive analysis of the supersymmetric hyperbolic model, using the ERRW-as-VRJP-with-Gamma-conductances representation and Ward identities.

The key estimates are field-theoretic and electrical. One obtains bounds on fluctuations of $n$4, for instance

$n$5

for $n$6, together with Ward identities such as

$n$7

where $n$8 is an effective resistance in the corresponding conductance network (Disertori et al., 2014). Transience is then tied to the resistance formula

$n$9

At the opposite end, recurrence for strong reinforcement had already been established. For any bound $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$0 on degree, there exists $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$1 such that the linearly edge-reinforced random walk is recurrent for $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$2 on graphs of bounded degree $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$3 (Michel, 2023). On $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$4 with $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$5, the monotonicity theorem shows that increasing initial weights makes ERRW more transient, so the recurrence/transience transition is unique (Poudevigne, 2019).

Dimension $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$6 is exceptional. ERRW on $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$7 with constant weights is recurrent for all $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$8, and together with the $$P(X_{n+1}=j\,|\,\mathcal F_n) = \mathbf 1_{\{j\sim X_n\}} \frac{Z_n(\{X_n,j\})}{\sum_{k\sim X_n} Z_n(\{X_n,k\})},$$9 weak-reinforcement transience result this gives a full answer to the old question of Diaconis (Sabot et al., 2015). In $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$0, weak reinforcement also admits a functional central limit theorem: for sufficiently large constant weights, the diffusively rescaled walk converges to Brownian motion with non-degenerate isotropic diffusion matrix (Sabot et al., 2015).

A common misconception is that reinforcement necessarily implies localization. The ERRW literature shows the contrary: linearly reinforced ERRW can be recurrent, transient, or diffusive depending on dimension and initial weights, whereas strong localization onto one edge is characteristic of super-linear reinforcement rather than of the standard linear model (Cotar et al., 2015).

4. One-dimensional, tree, and random-tree regimes

On $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$1, the standard linearly edge-reinforced random walk is always recurrent irrespective of the initial edge weights (Michel, 2023). On the half-line $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$2, with edge $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$3 given initial weight $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$4 and linear increment $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$5, the walk is recurrent if and only if $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$6 (Takei, 2020). In the recurrent regime with $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$7 and $Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$8,

$Z_n(e)=a_e+\text{(number of traversals of %%%%4%%%% up to time %%%%5%%%%)}.$9

which is a law-of-the-iterated-logarithm-type statement showing that positive reinforcement drastically slows the walk relative to the unreinforced case (Takei, 2020). In the critical case $\ell$0, there is a phase transition in speed at $\ell$1 (Takei, 2020).

On infinite trees, recurrence and transience are characterized in terms of the branching number and a reinforcement parameter. For a tree $\ell$2 with branching number $\ell$3, one introduces

$\ell$4

and the critical parameter $\ell$5 is determined by $\ell$6. If $\ell$7, ERRW on $\ell$8 is transient; if $\ell$9, it is recurrent (Michel, 2023). For Galton–Watson trees with mean offspring $f(\ell,x)=f(0,x)+\ell \Delta,$0, the same formula applies with $f(\ell,x)=f(0,x)+\ell \Delta,$1 in place of $f(\ell,x)=f(0,x)+\ell \Delta,$2 (Michel, 2023).

On critical Galton–Watson trees in the recurrent regime $f(\ell,x)=f(0,x)+\ell \Delta,$3, an invariance principle holds: suitably rescaled linearly edge-reinforced random walks converge to a diffusion on the $f(\ell,x)=f(0,x)+\ell \Delta,$4-stable tree, with resistance metric and speed measure expressed through a tree-indexed Gaussian field (Andriopoulos et al., 2021). In the transient regime on these random trees, there is still no positive speed: the discrete ERRW never has positive speed, even when the initial edge weights are strongly biased away from the root (Andriopoulos et al., 2021). This suggests that random fractal tree geometry can dominate directional bias.

5. Non-reversible, interacting, and strongly reinforced variants

A major non-reversible extension is the $f(\ell,x)=f(0,x)+\ell \Delta,$5-ERRW, defined on directed graphs endowed with an involution $f(\ell,x)=f(0,x)+\ell \Delta,$6 on vertices and hence on edges. Its reinforced weights are

$f(\ell,x)=f(0,x)+\ell \Delta,$7

and the process generalizes both the classical ERRW and random walk in Dirichlet environment (Bacallado et al., 2021). Under the divergence condition

$f(\ell,x)=f(0,x)+\ell \Delta,$8

the $f(\ell,x)=f(0,x)+\ell \Delta,$9-ERRW is partially exchangeable and hence a random walk in a random environment. Its mixing law is explicit and extends the classical magic formula from mixtures of reversible Markov chains to mixtures of Yaglom reversible Markov chains (Bacallado et al., 2021).

Interaction between finitely many walkers does not, in the available results, fundamentally alter the macroscopic dichotomy seen for a single walker. On a three-node segment with two walkers and linear reinforcement, the left-edge weight proportion is a bounded martingale at certain stopping times, hence converges almost surely to a random limit (Gantert et al., 2023). On $\Delta\ge 0$0 with arbitrary finite $\Delta\ge 0$1 and very general reinforcement, either all walkers are recurrent or all walkers have finite range; no mixed behavior occurs (Gantert et al., 2023).

The strongly reinforced regime behaves differently from the linear one. For super-linear reinforcement on arbitrary infinite connected graphs of bounded degree, if the reinforcement weight function $\Delta\ge 0$2 is reciprocally summable, then the walk traverses a random attracting edge at all large times, settling a conjecture of Sellke (Cotar et al., 2015). This is genuine localization: eventually the walk moves back and forth across a single random edge forever.

These variants clarify that “ERRW” is not a single asymptotic universality class. Linearity, reversibility, and the number of interacting walkers each matter at the level of limiting behavior.

6. Statistical and information-theoretic viewpoints

Recent work treats ERRW not only as a probabilistic object but also as a statistical model. On finite connected graphs with positive initial edge weights, one can ask whether the initial weights $\Delta\ge 0$3 are identifiable from observed trajectories. Using the magic formula and explicit moment identities, a generalized method of moments estimator has been proposed for estimating $\Delta\ge 0$4 from multiple independent sample trajectories (Qinghua et al., 8 Mar 2025). The analysis is non-asymptotic and exploits a hyperbolic Gaussian representation of the random environment.

A central negative result is that a single trajectory, even infinitely long, is never sufficient for parameter identification (Qinghua et al., 8 Mar 2025). By contrast, with multiple i.i.d. trajectories one obtains explicit sample-complexity bounds. The estimation method uses moment equations built from random variables such as

$\Delta\ge 0$5

where $\Delta\ge 0$6 is the environment-induced transition probability across edge $\Delta\ge 0$7 (Qinghua et al., 8 Mar 2025).

The environment laws themselves admit an information-theoretic analysis. For finite graphs, the magic-formula family forms a regular exponential family in the initial weights, enabling explicit formulas for the entropy rate and for Kullback–Leibler divergences between ERRW models (Qinghua et al., 21 May 2026). The entropy rate has an annealed representation,

$\Delta\ge 0$8

and the KL divergence between two environment laws $\Delta\ge 0$9 takes the form

$\{x,x+1\}$0

with

$\{x,x+1\}$1

The trajectory-level KL divergence converges to the environment-level KL divergence, and the gap is described by an explicit posterior-gap identity (Qinghua et al., 21 May 2026).

These developments place ERRW within contemporary statistical theory for dependent data. A plausible implication is that the random-environment representation is not merely an analytical convenience: it is also the correct parameterization for inference, testing, and information measures.