Correlated Classical Random Walk Analysis

• Correlated classical random walks are models where successive steps depend on a persistence parameter (q) that induces persistence for q > 1/2 and anti-persistence for q < 1/2.
• Exact finite-time laws and scaling transformations reveal distinct displacement and return statistics that differ from those of uncorrelated random walks.
• Operator and graph formulations connect these models to spectral properties and offer insights into parallels with coined quantum walks.

A correlated classical random walk is a random-walk model in which successive displacements are not independent. In the standard one-dimensional discrete-time formulation, each step has fixed length $\delta x$, occurs after fixed time $\delta t$, and has sign $\nu_n\in\{-1,+1\}$, but the probability that two consecutive steps have the same sign is $q\neq 1/2$. This one-step memory produces persistence for $q>1/2$, anti-persistence for $q<1/2$, and the ordinary uncorrelated walk at $q=1/2$. The topic has developed along several axes: exact finite-time combinatorics for displacement and return laws, scaling transformations across sampling intervals, inhomogeneous ensembles, first-passage and record statistics, graph- and operator-theoretic formulations, and systematic comparison with coined quantum walks (1207.1240, Kiumi et al., 2022, Hantzko et al., 31 Jul 2025).

1. Foundational definitions and parameterizations

In the discrete-time correlated walk, the position is built from step increments

$X_n-X_{n-1}=\nu_n,$

with physical coordinates

$x_k=k\,\delta x,\qquad t_n=n\,\delta t.$

The persistence parameter is

$q=\Pr(\nu_{n+1}=\nu_n).$

The standard interpretation is immediate: $\delta t$0 gives a persistent walk, $\delta t$1 an anti-persistent walk, and $\delta t$2 the ordinary uncorrelated random walk. In this formulation the process is a Markov chain, because the next step depends only on the immediately preceding step, not on the full history. A common misconception is that “correlated” here necessarily means non-Markovian; for the basic single-walker model that is not the case (1207.1240).

A useful exact diagnostic is the step-sign autocorrelation

$\delta t$3

so the discrete-time correlations decay exponentially in the lag $\delta t$4. In the inhomogeneous-ensemble literature this same correlation structure is expressed as a velocity autocorrelation

$\delta t$5

where $\delta t$6 sets the crossover from short-time ballistic behavior to long-time diffusive behavior (1207.1240, Stadler et al., 2012).

An equivalent one-step-memory parameterization specifies the next move conditional on the previous direction. If the previous step was to the left, then

$\delta t$7

If the previous step was to the right, then

$\delta t$8

This is encoded by the transition matrix

$\delta t$9

with $\nu_n\in\{-1,+1\}$0. Two special cases are emphasized: if $\nu_n\in\{-1,+1\}$1, the walker tends to continue in the same direction with probability $\nu_n\in\{-1,+1\}$2; if $\nu_n\in\{-1,+1\}$3, the walk becomes uncorrelated with the past; and if $\nu_n\in\{-1,+1\}$4, it reduces to the usual symmetric random walk (Kiumi et al., 2022).

A further notational variant, used to compare directly with coined quantum walks, promotes the current direction to an internal state in $\nu_n\in\{-1,+1\}$5. The correlation parameter is then $\nu_n\in\{-1,+1\}$6: if the walker previously moved in a given direction, it keeps that direction with probability $\nu_n\in\{-1,+1\}$7 and flips with probability $\nu_n\in\{-1,+1\}$8. The velocity update is governed by the doubly stochastic matrix

$\nu_n\in\{-1,+1\}$9

This formulation makes explicit that the model is a persistent two-state velocity process followed by a conditional spatial shift (Hantzko et al., 31 Jul 2025).

2. Exact finite-time laws

For the one-dimensional lattice walk with equal initial left/right probability, the exact displacement distribution in lattice units is

$q\neq 1/2$0

subject to the conditions that $q\neq 1/2$1 and $q\neq 1/2$2 have the same parity and that $q\neq 1/2$3 for $q\neq 1/2$4. With physical units,

$q\neq 1/2$5

and interpolation is used off the lattice. The central point is that $q\neq 1/2$6 changes the full displacement law, not merely the variance (1207.1240).

The mean squared displacement also has a closed form: $q\neq 1/2$7 and in physical units

$q\neq 1/2$8

At long times the walk is diffusive, but the diffusion constant is $q\neq 1/2$9-dependent (1207.1240).

A separate exact observable is the return probability on $q>1/2$0. If $q>1/2$1 denotes the sum of all paths with $q>1/2$2 left steps and $q>1/2$3 right steps, then for an initial state $q>1/2$4 with $q>1/2$5,

$q>1/2$6

Returning to the origin requires $q>1/2$7, so the return probability is nonzero only at even times. The path-counting method yields a closed form in terms of Legendre polynomials. With

$q>1/2$8

the paper gives $q>1/2$9 explicitly as a Legendre-polynomial combination in $q<1/2$0 and $q<1/2$1. When $q<1/2$2, the return probability does not depend on the initial state. In the ordinary-random-walk limit $q<1/2$3, one obtains

$q<1/2$4

and for the symmetric case $q<1/2$5,

$q<1/2$6

The generating function is elementary, involving a square root rather than elliptic integrals (Kiumi et al., 2022).

These exact laws show that nearest-neighbor correlation is not a perturbative refinement of the simple random walk but a distinct combinatorial object. The displacement law, return law, and second moment all retain closed-form structure, yet differ qualitatively from the binomial/Gaussian profile of the uncorrelated case.

3. Scaling, coarse graining, and heterogeneous ensembles

A central scaling result states that a correlated discrete-time random walk with parameters

$q<1/2$7

can be mapped to a walk with finer sampling interval

$q<1/2$8

while preserving long-time displacement statistics. Matching the step autocorrelation under time rescaling gives

$q<1/2$9

with $q=1/2$0 for $q=1/2$1 and $q=1/2$2 for $q=1/2$3. Matching the asymptotic mean squared displacement then requires

$q=1/2$4

The full transformation is therefore

$q=1/2$5

A common misunderstanding is that autocorrelation matching alone is sufficient; the construction explicitly shows that rescaling $q=1/2$6 is also necessary. The equivalence is a long-time one, not a step-by-step microscopic identity (1207.1240).

The same logic extends to inhomogeneous ensembles in which each walker $q=1/2$7 has its own persistence $q=1/2$8, and possibly its own step length $q=1/2$9. For a distribution $X_n-X_{n-1}=\nu_n,$0,

$X_n-X_{n-1}=\nu_n,$1

and

$X_n-X_{n-1}=\nu_n,$2

Although each constituent walker remains a simple Markovian correlated random walker, the ensemble can display markedly different statistics: the averaged displacement density can become leptokurtic, and the mean squared displacement can increase approximately like a fractional power law with lag time. This is an explicit instance in which complex ensemble-level behavior does not require a single non-Markovian microscopic walker (Stadler et al., 2012).

For the special case of a uniform distribution $X_n-X_{n-1}=\nu_n,$3 on $X_n-X_{n-1}=\nu_n,$4, the averaged step autocorrelation becomes

$X_n-X_{n-1}=\nu_n,$5

where $X_n-X_{n-1}=\nu_n,$6 for even $X_n-X_{n-1}=\nu_n,$7 and $X_n-X_{n-1}=\nu_n,$8 for odd $X_n-X_{n-1}=\nu_n,$9. The corresponding exact ensemble MSD is

$x_k=k\,\delta x,\qquad t_n=n\,\delta t.$0

with $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$1 for even $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$2 and $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$3 for odd $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$4. The reported growth is approximately a fractional power law with an effective exponent around $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$5 over a broad range of lag times, but the paper stresses that this is a broad crossover generated by a superposition of correlation times, not a true asymptotic power law from a scale-free process (Stadler et al., 2012).

The ensemble framework is also applied to temporally fluctuating parameters $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$6, provided they remain approximately constant over windows $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$7. This suggests a bridge between heterogeneous populations and nonstationary single-particle trajectories without changing the effective lag-time statistics (Stadler et al., 2012).

4. First-passage structure, survival, and record statistics

In a space-continuous one-dimensional variant, the walk is defined by

$x_k=k\,\delta x,\qquad t_n=n\,\delta t.$8

where $x_k=k\,\delta x,\qquad t_n=n\,\delta t.$9 are i.i.d. with continuous density $q=\Pr(\nu_{n+1}=\nu_n).$0, while the signs satisfy

$q=\Pr(\nu_{n+1}=\nu_n).$1

The persistence parameter is again $q=\Pr(\nu_{n+1}=\nu_n).$2. Under the continuity assumption on $q=\Pr(\nu_{n+1}=\nu_n).$3, the survival probability

$q=\Pr(\nu_{n+1}=\nu_n).$4

is independent of the jump distribution for any finite $q=\Pr(\nu_{n+1}=\nu_n).$5, a universality traced to the Sparre-Andersen theorem. The exact formula is

$q=\Pr(\nu_{n+1}=\nu_n).$6

with generating function

$q=\Pr(\nu_{n+1}=\nu_n).$7

For fixed $q=\Pr(\nu_{n+1}=\nu_n).$8,

$q=\Pr(\nu_{n+1}=\nu_n).$9

In the scaling regime $\delta t$00, $\delta t$01, $\delta t$02, the model converges to the run-and-tumble particle, and the survival probability becomes

$\delta t$03

The same framework yields the exact distribution of the time $\delta t$04 at which the maximum is reached,

$\delta t$05

and the large-$\delta t$06 arcsine law

$\delta t$07

For record statistics, the number of records $\delta t$08 has the same asymptotic scaling form as for an uncorrelated random walk with effective step number

$\delta t$09

This effective-time renormalization is one of the sharpest universal statements available for correlated classical walks with sign persistence (Lacroix-A-Chez-Toine et al., 2020).

A different correlated-step model produces a qualitatively different regime. There the increments are $\delta t$10, but

$\delta t$11

The exact variance is

$\delta t$12

so $\delta t$13, indicating super-diffusion. The step-step correlations do not decay: $\delta t$14 Moreover,

$\delta t$15

where the limiting speed has density

$\delta t$16

This walk is transient, record counts grow linearly rather than diffusively, and the limiting record distribution

$\delta t$17

is nonzero for fixed $\delta t$18. The contrast with the diffusive sign-persistence model is conceptually important: “correlated random walk” is not a single universality class, but a family of memory mechanisms with distinct asymptotic consequences (Kearney, 2019).

5. Spectral, graph, and finite-domain formulations

Correlated classical random walks admit natural operator formulations on graphs. One construction starts from the Grover matrix $\delta t$19 of the Grover walk on a connected graph $\delta t$20 and defines a classical walk on the arc set $\delta t$21 by

$\delta t$22

The resulting process is a classical Markov chain on arcs, not vertices, so the next transition depends on the incoming direction. Its transition matrix can be written as $\delta t$23, and the characteristic polynomial $\delta t$24 is obtained באמצעות a determinant formula for the generalized weighted zeta function. Explicit spectral formulas are then derived for connected $\delta t$25-regular graphs and connected semiregular bipartite graphs. For $\delta t$26, the nontrivial correlation disappears: each nonzero transition probability becomes $\delta t$27, the walk becomes essentially the simple random walk on arcs, and the Grover matrix is a Hadamard matrix (Komatsu et al., 2020).

A related “Walk/Zeta Correspondence” treats CRW on discrete tori $\delta t$28 with an internal coin state. The walk operator is matrix-valued,

$\delta t$29

where the coin matrix $\delta t$30 has nonnegative entries and each column sums to $\delta t$31. Ordinary random walks arise as the degenerate case in which all columns are identical, so the internal state no longer matters. The zeta function is defined by

$\delta t$32

and its logarithm generates traces of powers of $\delta t$33. Fourier analysis reduces the determinant to a product over momentum modes,

$\delta t$34

yielding explicit zeta formulas for three- and four-state CRW on the one-dimensional torus and four-state CRW on the two-dimensional torus. This places correlated classical walks in a spectral framework parallel to that used for quantum walks (Konno et al., 2021).

On a finite path $\delta t$35, the CRW is formulated on the two-component space

$\delta t$36

with time evolution operator $\delta t$37, where $\delta t$38 is a site-dependent coin and $\delta t$39 is a boundary-preserving shift. In the isospectral-coin case, a Jacobi matrix $\delta t$40 captures the essential spectral data. Most eigenvalues of $\delta t$41 are obtained from the quadratic relation

$\delta t$42

where $\delta t$43 is an eigenvalue of $\delta t$44, and there is always an additional eigenvalue

$\delta t$45

The long-time limiting distribution exists and is the stationary distribution of an associated birth-and-death chain (Ide et al., 2023).

In a continuous-space finite-domain setting, the correlated random walk system is written for right- and left-moving densities $\delta t$46 and $\delta t$47: $\delta t$48 with absorbing boundary conditions. In variables $\delta t$49 and $\delta t$50, this becomes

$\delta t$51

and hence the telegraph equation

$\delta t$52

The operator has compact resolvent, the semigroup is eventually compact, and all solutions decay exponentially with optimal exponent equal to the dominant eigenvalue $\delta t$53. The spectrum is characterized by the transcendental equations

$\delta t$54

and

$\delta t$55

This formulation makes precise the persistent-motion interpretation common in biological modeling and chromatography (Menacho et al., 17 Jan 2025).

6. Comparison principles and major extensions

The correlated classical random walk is the natural classical analogue of a one-dimensional coined quantum walk when both are formulated on $\delta t$56, with an internal direction update followed by a conditional shift. The asymptotic distinction is sharp. For the classical walk with $\delta t$57, the normalized position converges weakly to a centered Gaussian with variance

$\delta t$58

so the generic classical walk remains diffusive with

$\delta t$59

By contrast, non-trivial coined quantum walks have non-Gaussian limiting laws on compact support and asymptotic variance $\delta t$60. Only the extremal classical cases align with boundary quantum cases: $\delta t$61 gives quadratic growth, while $\delta t$62 yields an alternating walk matching the Pauli-$\delta t$63 boundary behavior. This comparison clarifies that classical correlation changes the diffusion constant, whereas quantum coherence changes the qualitative transport regime (Hantzko et al., 31 Jul 2025).

For systems of two correlated classical random walks on a lattice, one can decompose the pair into a walk of common movements, a walk of counter movements, and a random time change. If $\delta t$64 and $\delta t$65 are the two walks and $\delta t$66 counts the common-direction steps, then

$\delta t$67

Under symmetric marginal step laws, $\delta t$68 and $\delta t$69 are independent simple random walks, so the entire dependence structure is carried by the clock $\delta t$70. The three processes $\delta t$71, $\delta t$72, and $\delta t$73 are mutually independent if and only if Condition C1 holds. This decomposition isolates correlation into a time-change mechanism rather than into nontrivial marginal step laws (Chen et al., 2018).

A continuous-time extension replaces deterministic jump times by renewal times while retaining correlated spatial increments. In correlated CTRWs, the jumps are generated by a Markov chain arising from urn-scheme models, while the waiting times are independent and lie in the domain of attraction of a positively skewed $\delta t$74-stable law. The scaling limits are not Lévy processes but fractional Pearson diffusions: $\delta t$75 This extends the notion of correlated classical random walk from nearest-neighbor persistence to decoupled CTRWs with correlated jumps and heavy-tailed waiting times (Leonenko et al., 2017).

Another extension connects correlated random walks to integrable discrete dynamics. In the symmetric case $\delta t$76, eliminating one internal component from

$\delta t$77

yields the discrete correlated diffusion equation

$\delta t$78

Through a generalized discrete Cole–Hopf transformation and ultradiscretization, this produces a variant of the ultradiscrete Burgers equation with a cellular-automaton interpretation as a traffic-flow model (Fukuda et al., 2021).

Taken together, these developments show that the correlated classical random walk is not a single model but a coherent research domain organized around a common principle: successive displacements are linked by internal state, persistence, or time-change structure. Depending on how that linkage is implemented, the resulting process may remain diffusive, exhibit effective coarse-grained anomalous scaling, converge to run-and-tumble dynamics, acquire exact spectral descriptions on graphs, or serve as a classical foil for quantum transport.