Continuous-Time Random Walks Explained

Updated 26 February 2026

CTRWs are stochastic processes characterized by random jumps and waiting times governed by specific probability laws, underpinning models of anomalous transport.
They provide a framework for deriving fractional partial differential equations that capture non-Markovian dynamics and scaling limits in diverse applications.
Extensions of CTRWs, including reset mechanisms, coupled jumps, and network models, offer versatile tools for simulating and analyzing complex dynamical systems.

A continuous-time random walk (CTRW) is a stochastic process in which a particle undergoes random spatial displacements ("jumps") at random epochs ("waiting times"), with both sequences typically drawn from specified probability laws. CTRWs generalize classical random walk and renewal processes and serve as a fundamental modeling tool for anomalous transport, non-Markovian dynamics, and non-Gaussian diffusive phenomena in physics, biology, finance, and network science. The framework encompasses both uncoupled and coupled space-time increments, correlated processes, extensions to networks, and scaling limits leading to fractional partial differential equations. Below, core structural aspects and main results are organized by major research themes and technical developments.

1. Fundamental Structure: Definition, Master Equations, and Classes

A CTRW is specified by sequences of spatial jumps $\{\xi_k\}$ (i.i.d. or Markov) and strictly positive waiting times $\{\tau_k\}$ ; in the uncoupled case, their joint law factorizes as $\lambda(\xi)\psi(\tau)$ . Define renewals $t_0=0$ , $t_n = \sum_{i=1}^n \tau_i$ , and the counting process $N(t) = \max\{ n: t_n \leq t \}$ . The position process is $X(t) = \sum_{i=1}^{N(t)} \xi_i$ ; its law $p(x,t)$ satisfies the integro-differential (Montroll–Weiss) master equation,

$p(x,t) = \Psi(t)\delta(x) + \int_0^t d\tau \,\psi(\tau) \int_{-\infty}^\infty d\xi\, \lambda(\xi)p(x-\xi,t-\tau),$

where $\Psi(t)$ is the survival probability. In Laplace–Fourier space, the propagator is

$\widehat{\widetilde{p}}(k,s) = \frac{1-\widetilde\psi(s)}{s[1-\widehat\lambda(k)\widetilde\psi(s)]}$

(Colantoni et al., 10 Jul 2025).

Important classes:

Uncoupled CTRW: spatial and temporal increments are independent (0802.3769).
Coupled CTRW: space-time pairs $(\xi_k, \tau_k)$ may be dependent through a joint law $K(d\xi, d\tau)$ , allowing for forward- or backward-coupling (Barczyk et al., 2012, Meerschaert et al., 2012).
Correlated-jump CTRW: jumps are correlated via an underlying Markov chain, producing fractional Pearson diffusions as scaling limits (Leonenko et al., 2017).

2. Scaling Limits and Fractional Partial Differential Equations

Under appropriate scaling, CTRWs converge to non-Markovian processes. Let space and time be rescaled via $x \sim a_n$ , $t \sim b_n$ . If waiting times have infinite mean ( $\psi(\tau) \sim \tau^{-1-\beta}, 0<\beta<1$ ), the inverse subordinator $E(t)$ appears: $X^n(t) = S^n(N^n(t)) \xrightarrow{J_1} A(E(t)),$ where $A$ is, for example, Brownian motion or a Lévy process (Burr, 2011, Busani, 2015).

The one-dimensional density $p(x,t)$ satisfies the fractional Fokker–Planck equation: $D_t^\beta p(x, t) = \mathcal{L}_x p(x, t) + \text{boundary terms},$ where $D_t^\beta$ is the Caputo or Riemann–Liouville fractional derivative, and $\mathcal{L}_x$ is the generator of $A$ (Busani, 2015, Leonenko et al., 2017). For multi-time distributions, a hierarchy of finite-dimensional fractional PDEs (FD-FFPEs) arises, uniquely characterizing the process beyond its marginals (Busani, 2015, Meerschaert et al., 2012).

3. Memory, Correlations, and Generalizations

Correlated jumps and fractional Pearson limits

Correlated-jump CTRWs, with increments generated via urn-scheme Markov chains (e.g., Bernoulli–Laplace, Wright–Fisher), produce limiting diffusions with non-independent increments (Ornstein-Uhlenbeck, Jacobi, CIR diffusions), subordinated by $\beta$ -stable inverse subordinators to yield fractional Pearson diffusions (fPDs). The Fokker–Planck equation becomes: $D_t^\beta p_\beta(x, t; y) = -\partial_x[\mu(x)p_\beta] + \frac{1}{2}\partial_x^2[\sigma^2(x)p_\beta]$ with drift $\mu(x)$ and diffusion $\sigma^2(x)$ given by the limiting Pearson process (Leonenko et al., 2017).

Coupled and age-dependent extensions

Forward- and backward-coupling—the choice of whether the jump is associated with the preceding or following waiting time—yields distinct scaling limits ("overshooting" or "undershooting" subordination), analyzable via marked point process representations (Barczyk et al., 2012). Semi-Markov embeddings allow the full finite-dimensional distribution to be constructed via transition kernels for the augmented Markov process $(X(t), V(t))$ , where $V(t)$ is the "age" since last renewal (Meerschaert et al., 2012, Gill et al., 2016).

Correlated waiting times

Long-range dependence in the waiting-time sequence, introduced via "block-repetition" mechanisms, produces slow, power-law decay of nonlinear autocorrelations (e.g., volatility clustering) even for thin-tailed marginal $\psi(\tau)$ . In such models,

$C(\tau) \sim \tau^{-(\rho-2)}, \qquad \rho \in (2,4)$

with $\rho$ the exponent of the run-length distribution (Klamut et al., 2019).

Continuous-path interpolations

For modeling continuous particle tracks, CTRW trajectories can be interpolated to yield "continuous-path CTRWs" (CPCTRWs). Their scaling limits and convergence properties differ, often requiring Skorokhod $M_1$ topology for functional limit theorems (Zebrowski et al., 2013). The mapping from cadlag "staircase" CTRW paths to continuous interpolants can have nontrivial effects on limiting distributions.

4. Algorithmic Approaches and Simulation

Numerical solutions

Master equations for discrete-time/space approximations of bivariate Markov processes $(X(t),V(t))$ enable direct simulation with arbitrary initial conditions. Efficient algorithms are possible via discrete lattice schemes, with convergence to the Skorokhod limit proven under natural conditions (Gill et al., 2016). Monte Carlo simulation for CTRWs, and for functionals such as stochastic integrals and quadratic variation, is straightforward due to the pure-jump structure (0802.3769).

Path-ensemble statistics on networks

CTRW statistics on arbitrary networks (with state-dependent waiting-time distributions and transition probabilities) are computable via path-ensemble recursion relations. Exact relations connect moments of first-passage time and path length, and efficient algorithms (e.g., PathMAN) can resolve high-order statistics, highlighting the significance of memory and non-exponential waiting times (Manhart et al., 2015).

5. CTRWs with Resetting, Network and Graph Extensions

Stochastic resetting

A CTRW under arbitrary stochastic resetting to the origin results in a new CTRW structure with identical jump-size distribution but a modified counting process. The Volterra integral equation for the PDF $p(x,t)$ becomes: $p(x,t) = R(t)\rho(x, t) + \int_0^t r(\tau) p(x, t-\tau)\,d\tau,$ where $R(t)$ is the survival probability for no reset, $r(\tau)$ the reset interval density, and $\rho(x,t)$ the non-reset CTRW propagator (Colantoni et al., 10 Jul 2025). Existence of a stationary state requires the mean reset time to be finite, and the so-called "zero-law" stipulates that the probability to jump dominates resetting at large times.

Time-varying graphs and dynamic environments

CTRWs on time-varying graphs (dynamic networks) exhibit nontrivial stationary behavior, depending on the interplay between walker and network dynamics. Analytically tractable regimes include fast-walker, slow-walker, and degree-coupled dynamics, with closed-form results for stationary distributions under time-ergodic and T-connected graph processes. For degree-proportional rates, the stationary distribution is uniform regardless of dynamic variations (Figueiredo et al., 2011).

6. Large Deviations and Non-Gaussian Fluctuations

The large deviation rate function for a CTRW with jumps and waiting times characterized by respective rate functions $\mathcal{R}(u)$ and $\mathcal{I}(\xi)$ obeys

$\mathcal{X}(z) = \inf_{\xi>0} \left\{ \frac{\mathcal{R}(z\xi) + \mathcal{I}(\xi)}{\xi} \right\},$

yielding sub-quadratic, quadratic, or nontrivial tail forms depending on the short-time behavior of the waiting-time law. For example, for exponential waiting times, $\mathcal{X}(x) \sim |x| \sqrt{\ln |x|}$ ; for Pareto Type I, the leading behavior is quadratic (Pacheco-Pozo et al., 2020). These forms highlight the influence of short-time waiting-time statistics on rare-event trajectories.

7. Langevin Formulation and Noise Structure

Subdiffusive CTRWs, characterized by power-law waiting times, can be expressed via a Langevin equation in physical time: $\dot{Y}(t) = \sqrt{2\sigma}\,\bar{\xi}(t),$ where $\bar{\xi}(t)$ is a non-Gaussian noise which reproduces all moments of the original subordinated process $Y(t) = X(S(t))$ (Cairoli et al., 2015). While the mean-square displacement and two-point function coincide with that of scaled Brownian motion (SBM), only higher-order correlations distinguish the two. Extensions to arbitrary force fields, waiting-time distributions, and the inclusion of external driving are possible within this framework, providing a unified approach to anomalous dynamics and non-Gaussian fluctuations.