Step-Reinforced Random Walk

• Step-Reinforced Random Walk is a discrete-time non-Markovian process that combines fresh independent steps with uniformly repeated past increments.
• It employs an urn representation to track reinforced step counts and reveals a critical phase transition at the reinforcement parameter a = 1/2 across different dimensions.
• The model offers insights into recurrence versus transience behavior, linking classical random walks with advanced reinforcement mechanisms like elephant random walks.

A step-reinforced random walk (SRRW) is a discrete-time non-Markovian random walk in which the next increment is obtained either by drawing a fresh step from a fixed law or by repeating a uniformly chosen past increment. In the standard formulation on $\mathbb Z^d$, the process remembers and reinforces past increments rather than past positions. This class includes the elephant random walk as a special case, and recent work establishes a sharp recurrence/transience phase diagram whose critical reinforcement value is $a_c=\frac12$: in dimensions $d=1,2$ there is a low-dimensional transition at one-half, whereas in dimensions $d\ge 3$ the walk is transient for every reinforcement parameter in $[0,1)$ under suitable moment assumptions (Qin, 2024).

1. Definition and reinforcement mechanism

Let $(X_n)_{n\ge1}$ be the increment sequence and

$S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$

The first increment is sampled from a fixed step distribution $\mu$ on $\mathbb Z^d$,

$X_1\sim \mu.$

For each $a_c=\frac12$0, independently of the past, one tosses a Bernoulli$a_c=\frac12$1 variable $a_c=\frac12$2. With probability $a_c=\frac12$3, the walk repeats one of its previous steps, chosen uniformly: $a_c=\frac12$4 With probability $a_c=\frac12$5, the walk makes a fresh independent step: $a_c=\frac12$6 This is the standard step-reinforcement mechanism (Qin, 2024).

The dependence is global in time: the law of $a_c=\frac12$7 depends on the entire increment history through the empirical multiset of previously realized steps. The model is therefore non-Markovian even when the underlying fresh-step law $a_c=\frac12$8 is simple. At the same time, the reinforcement rule remains structurally tractable because the copied step is sampled uniformly from prior increments.

A useful representation is given by the step counts $a_c=\frac12$9, which evolve like an infinite-color Pólya urn. In that representation the walk can be written as

$d=1,2$0

This urn viewpoint isolates the reinforcement in the evolution of step frequencies and is central in the modern analysis of SRRWs (Qin, 2024).

2. Assumptions, nondegeneracy, and basic notions

The recurrence/transience theory is formulated for centered, non-degenerate increment laws with suitable moments. The centeredness assumption is

$d=1,2$1

since nonzero drift trivially gives transience. The basic finite-variance condition is

$d=1,2$2

and several results require the stronger $d=1,2$3 moment condition

$d=1,2$4

In the critical two-dimensional case $d=1,2$5, a fourth-moment assumption is imposed: $d=1,2$6 These are the main regularity hypotheses in the phase-transition theorem (Qin, 2024).

The walk is assumed to be genuinely $d=1,2$7-dimensional. Under finite second moments, this is equivalent to

$d=1,2$8

This excludes lower-dimensional support hidden inside $d=1,2$9, which would otherwise reduce the effective dimension of the recurrence/transience problem (Qin, 2024).

The same analysis uses the following definitions. Recurrence means

$d\ge 3$0

and transience means

$d\ge 3$1

These are the natural analogues of the classical dichotomy for ordinary random walks, but applied to a process whose increments are history-dependent rather than independent (Qin, 2024).

3. Phase transition at one-half

The critical reinforcement value is

$d\ge 3$2

This threshold separates diffusive or recurrent behavior from superdiffusive or transient behavior in low dimensions (Qin, 2024).

The phase diagram can be summarized as follows.

Dimension	Reinforcement regime	Behavior
$d\ge 3$3	$d\ge 3$4	recurrent
$d\ge 3$5	$d\ge 3$6	transient
$d\ge 3$7	$d\ge 3$8	recurrent
$d\ge 3$9	$[0,1)$0	transient
$[0,1)$1	$[0,1)$2 with fourth moment	transient
$[0,1)$3	all $[0,1)$4	transient

In dimension $[0,1)$5, if the SRRW is genuinely one-dimensional and $[0,1)$6 has a $[0,1)$7 moment, then the walk is recurrent for $[0,1)$8 and transient for $[0,1)$9. In dimension $(X_n)_{n\ge1}$0, if the SRRW is genuinely two-dimensional and $(X_n)_{n\ge1}$1 has a $(X_n)_{n\ge1}$2 moment, then the walk is recurrent for $(X_n)_{n\ge1}$3 and transient for $(X_n)_{n\ge1}$4. In dimensions $(X_n)_{n\ge1}$5, if the SRRW is genuinely $(X_n)_{n\ge1}$6-dimensional and $(X_n)_{n\ge1}$7 has a $(X_n)_{n\ge1}$8 moment, then the SRRW is transient for every $(X_n)_{n\ge1}$9 (Qin, 2024).

The critical two-dimensional case is borderline but still transient under the stronger fourth-moment condition. More precisely,

$S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$0

and

$S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$1

Thus the walk escapes to infinity, but with borderline growth rather than the clearer superdiffusive scaling seen above threshold (Qin, 2024).

For $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$2, only the second-moment assumption is needed to obtain a superdiffusive limit: $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$3 exists almost surely, and the limit is non-atomic: $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$4 This implies transience throughout the superdiffusive regime (Qin, 2024).

The same 2024 analysis explicitly resolves a conjecture of Bertoin: in dimensions $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$5, recurrence holds in the subcritical regime under mild moment assumptions, while higher dimensions are always transient. A common misconception is that reinforcement necessarily makes the walk more recurrent; the low-dimensional results show that this is only true below the threshold $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$6, whereas above that threshold reinforcement instead drives transience, and in $S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$7 it never restores recurrence (Qin, 2024).

4. Proof architecture

The proof strategy combines an urn representation, Lyapunov-function arguments adapted to non-Markovian dynamics, and a quasi-martingale analysis at criticality (Qin, 2024).

The first component is the infinite-color Pólya urn or recursive-tree representation. Writing

$S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$8

the analysis studies weighted fluctuations

$S_0=0,\qquad S_n=\sum_{k=1}^n X_k.$9

Decay estimates for $\mu$0, including almost sure convergence rates and $\mu$1 bounds, control the conditional moments of future increments. This converts the memory of the walk into a tractable empirical-process problem on reinforced step counts (Qin, 2024).

The second component adapts Lamperti’s method to a non-Markovian setting. Different Lyapunov functions are used in different dimensions: $\mu$2

$\mu$3

and, in dimension $\mu$4, a decaying function such as

$\mu$5

for large $\mu$6. These functions are chosen so that the expected one-step drift becomes negative, or otherwise suitable for a transience argument, after expansion in small increments and use of moment bounds together with urn estimates (Qin, 2024).

The third component addresses the critical two-dimensional case $\mu$7. There the argument uses the quasi-martingale

$\mu$8

and proves $\mu$9 almost surely. This yields

$\mathbb Z^d$0

which identifies the borderline growth rate and rules out recurrence at criticality (Qin, 2024).

5. Relation to elephant walks, counterbalancing, and fluctuation theory

The SRRW framework is broader than the original elephant random walk. The elephant model arises as a special case, but the general step distribution $\mathbb Z^d$1 need not be Rademacher, need not be bounded, and the state space can be genuinely multidimensional. In the superdiffusive regime $\mathbb Z^d$2, a general SRRW admits the deterministic normalization

$\mathbb Z^d$3

for which $\mathbb Z^d$4 is a martingale, and the second-order fluctuations satisfy

$\mathbb Z^d$5

for a non-degenerate random limit $\mathbb Z^d$6 (Bertenghi, 2021).

Subcritical positive and negative reinforcement also admit functional Gaussian limits. For centered $\mathbb Z^d$7 and $\mathbb Z^d$8, the triplet consisting of the ordinary walk and its positive and negative reinforced versions converges jointly to Gaussian processes that admit stochastic-integral representations. In the positive case there is a phase transition at $\mathbb Z^d$9, while in the negative case the walk remains diffusive for all $X_1\sim \mu.$0 under finite second moments (Bertenghi et al., 2021).

A particularly important contrast is provided by the counterbalanced or negatively step-reinforced random walk, where reinforcement copies a uniformly chosen past step with the opposite sign. If $X_1\sim \mu.$1,

$X_1\sim \mu.$2

and if $X_1\sim \mu.$3,

$X_1\sim \mu.$4

In this negative-reinforcement setting there is no phase transition in $X_1\sim \mu.$5: the walk is always diffusive at the $X_1\sim \mu.$6 scale when $X_1\sim \mu.$7 (Bertoin, 2020).

Pathwise limit theory has also been sharpened. Strong laws of large numbers and strong invariance principles have been proved for positively and negatively step-reinforced random walks via martingale difference sequences and truncation arguments (Hu et al., 2023). Under a finite third moment, Berry–Esseen bounds quantify the rates of convergence to normality for both positive and negative reinforcement: for the positive model in the Gaussian regime $X_1\sim \mu.$8, and for the negative model for all $X_1\sim \mu.$9 (Hu, 3 Apr 2025).

6. Generalizations, geometry, and continuous-time analogues

The classical SRRW on $a_c=\frac12$00 has been extended in several directions. One recent generalization places the walk on groups and allows arbitrary transformations of copied steps. In that framework, when $a_c=\frac12$01, $a_c=\frac12$02 is genuinely $a_c=\frac12$03-dimensional, and $a_c=\frac12$04, one has the Euclidean upper bound

$a_c=\frac12$05

for every $a_c=\frac12$06, which implies transience in all dimensions $a_c=\frac12$07. On countable groups the decay of transition probabilities is controlled by the isoperimetric profile of the Cayley graph, and this framework includes counterbalanced steps, unbalanced step-reinforced walks, echoed-step random walks, and the elephant random walk (Peres et al., 8 Apr 2026).

On finite groups, step reinforcement interacts with mixing rather than recurrence. If the classical walk is irreducible and aperiodic, the SRRW converges exponentially fast to the uniform distribution. On odd cycles there is a phase transition at $a_c=\frac12$08: for $a_c=\frac12$09 the mixing time remains of order $a_c=\frac12$10, at $a_c=\frac12$11 it is at least of order $a_c=\frac12$12, and for $a_c=\frac12$13,

$a_c=\frac12$14

so reinforcement can accelerate mixing (Peres et al., 8 Apr 2026).

Continuous-time analogues replace the discrete reinforced increment sequence by reinforced Lévy noise. For a Lévy process with Blumenthal–Getoor index $a_c=\frac12$15, the critical threshold is

$a_c=\frac12$16

and the admissible regime is $a_c=\frac12$17. In that regime, the step-reinforced random walks obtained from discrete-time skeletons converge in finite-dimensional distributions to a noise reinforced Lévy process built from the Lévy–Itô decomposition and Yule–Simon multiplicities (Bertoin, 2018).

The memory rule itself can also be generalized. In step-reinforced random walks with regularly varying memory, the past index is not sampled uniformly but with probabilities proportional to a regularly varying sequence $a_c=\frac12$18 of index $a_c=\frac12$19. Under finite second moments this produces a new critical point

$a_c=\frac12$20

with diffusive behavior below $a_c=\frac12$21, superdiffusive behavior above it, and a critical regime in which either almost sure convergence or Gaussian weak convergence can occur depending on the boundedness of a sequence associated with $a_c=\frac12$22 (Majumdar et al., 9 May 2025).

Taken together, these developments show that step reinforcement is not a single asymptotic mechanism but a family of long-range dependence structures whose behavior depends jointly on dimension, group geometry, sign structure, moment assumptions, and the rule by which the past is sampled. The one-half threshold on $a_c=\frac12$23 is the canonical result for the classical uniformly sampled model, but the broader literature demonstrates that this threshold can persist, disappear, or be replaced when the reinforcement scheme itself is modified (Qin, 2024).