Correlated Gaussian Random Walks (CGRW)

• Correlated Gaussian Random Walks (CGRW) are stochastic processes with Gaussian increments whose distributions depend on previous steps via a strategy function.
• They generalize simple Gaussian walks by incorporating correlations, using models like AR(1) and sign-based variations to capture behaviors such as ballistic spread and bimodal displacements.
• Simulation studies and scaling analyses validate CGRW’s applicability to fields from animal dispersal to cosmology, while open challenges remain in non-linear and high-dimensional extensions.

A Correlated Gaussian Random Walk (CGRW) is a discrete-time stochastic process in which the increments are Gaussian, with the additional feature that the distribution of each increment depends on previous steps, typically via a "memory" or "strategy" function. CGRWs generalize the classical simple (uncorrelated) Gaussian random walk by introducing step-to-step correlations, allowing the process to model directional persistence or more complex dependencies. These models are central in the quantitative study of animal movement, transport phenomena with memory effects, anomalous diffusion, and as components in the construction of Gaussian fields with tunable long-range dependence.

1. Mathematical Definition and Core Properties

Let $S_n = \sum_{k=1}^n x_k$ with each $x_k \in \mathbb{R}^d$ the $k$th step. In the uncorrelated case (Simple Gaussian Random Walk, SGRW), the increments $x_1, ..., x_n$ are i.i.d. $\sim N(0, \sigma^2 I_d)$. The hallmark of the CGRW is that $x_k$, given the previous increment $x_{k-1}$, is distributed as

$$P(x_k \mid x_{k-1}) = (2\pi \sigma^2)^{-d/2} \exp\left[ -\frac{\|x_k - \mu(x_{k-1})\|^2}{2\sigma^2} \right]$$

where $\mu: \mathbb{R}^d \to \mathbb{R}^d$ is the strategy function and $\sigma > 0$ controls the scale of the residual noise. The entire trajectory density is then

$$P(x_1, ..., x_n) = P(x_1) \prod_{k=2}^n P(x_k | x_{k-1}),$$

with the initial memory $\mu(x_0) = \mu_0$ (often chosen symmetrically).

A prominent example is the AR(1) linear memory: $\mu(x) = \rho x$ with $|\rho|<1$, which yields: $x_k | x_{k-1} \sim N(\rho x_{k-1}, \sigma^2 I_d)$ and

$\operatorname{Corr}(x_k, x_{k-1}) = \rho.$

The marginal law for displacement $S_n$ is again Gaussian with $\mathbf{E}[S_n]=0$ and

$$\operatorname{Var}(S_n) = \sigma^2\left[n + 2 \sum_{m=1}^{n-1} (n-m) \rho^m\right].$$

In limiting cases, for $\rho=0$ uncorrelated walks are recovered, and for $\rho \to 1^-$ (maximal persistence) the variance scales as $\sim \sigma^2 n^2$, indicating ballistic spread (Bagarti, 2011).

2. Nonlinear One-Dimensional CGRW Models: Memory, Bimodality, and Persistence

Distinct qualitative behaviors arise for nonlinear choices of $\mu$. Two classic one-dimensional cases are:

• Sign-only CGRW: $\mu(x) = \operatorname{sgn}(x) \in \{-1, +1\}$. For strong memory ($\sigma \ll 1$), the process exhibits persistent motion in one direction, with the probability density for $S_n$ becoming a bimodal sum of Gaussians centered at $\pm n$, reflecting two pure persistent trajectories, each broadened by Gaussian fluctuations:

$$p_{\rm CGRW}(S_n) \approx \frac{1}{2\sqrt{2\pi n}\sigma} \left\{ e^{-\frac{(S_n-n)^2}{2n\sigma^2}} + e^{-\frac{(S_n+n)^2}{2n\sigma^2}} \right\}.$$

For weak memory ($\sigma \gg 1$), the model reduces to an uncorrelated walk with Gaussian density centered at 0. Directional persistence is quantified by the probability of $n$ consecutive steps of the same sign, $$p_n = [\tfrac12(1 + \operatorname{erf}(1/\sqrt{2}\sigma))]^n$$; this decays much more slowly than $2^{-n}$ for small $\sigma$, indicating marked persistence (Bagarti, 2011).

• Threshold-and-sign CGRW: $\mu(x) = \operatorname{sgn}(x)\, p$ for $|x| \le x_0$, $\mu(x) = 0$ otherwise, with $p \gg \sigma \gg x_0$. This model captures "cluster-and-jump" patterns, with clusters of small, persistent steps punctuated by uncorrelated large excursions. In this regime, directional persistence disappears ($p_n = 2^{-n}$), because large steps rapidly decorrelate the trajectory.

These models illustrate that correlation structure alone does not guarantee persistence; the detailed form of $\mu$ and the interplay with $\sigma$ control the qualitative movement regimes (Bagarti, 2011).

3. Scaling, Diffusion Limits, and Fractional Dynamics

CGRW processes are central to constructing limit objects with non-Markovian or anomalous scaling, notably in continuous-time or in long-memory regimes:

• Correlated Continuous-Time Random Walks (CTRW): For increments $Y_n = \sum_{j=0}^\infty c_j Z_{n-j}$ (with $Z_n$ i.i.d., $c_j \sim j^{H-1-1/\alpha}$), and i.i.d. exponential waiting times $J_n$ (Poisson clock), the process $X(t) = S(N(t))$ converges, under scaling, to fractional Brownian motion when $Z_n$ are Gaussian and $H>1/2$:

$$\frac{X(t)}{\sigma\,(t/\mu)^H} \xrightarrow{t\to\infty} W_H(1)$$

where $W_H$ is standard fBM of Hurst $H$ (0809.1612).

• Governing Equations: The self-similar limits are governed by time-dependent diffusion equations:

$$\partial_t p(x,t) = D_H(t)\, \partial_{xx} p(x,t), \quad D_H(t) = 2H \sigma^2 \mu^{-2H} t^{2H-1}$$

For heavy-tailed waiting times or jumps, space-time fractional PDEs emerge, encoding non-local memory (0809.1612).

• Operator-Scaling Gaussian Fields: Aggregation of multidimensional persistence-walks with random, possibly dependent, persistence parameters, yields in the scaling limit a rich family of operator-scaling fields, including both fractional Brownian sheets and new anisotropic structures, contingent on the dependence of the persistence parameters and the scaling regime ("critical speed") (Shen et al., 2017).

4. Quantitative and Numerical Results: Simulation Insights

Monte Carlo simulations confirm the theoretical predictions for CGRW models:

• In the sign-only CGRW with strong memory, the histogram of total displacement $S_n$ is strongly bimodal, agreeing with the analytic form; the root-mean-square displacement grows linearly with $n$, evidencing superdiffusion.
• In the threshold-and-sign model, histograms exhibit oscillatory structure with peaks near multiples of $p$ ("clusters"), which is lost as $\sigma$ increases.
• Persistence probabilities decay as predicted in asymmetric memory settings, and collapses to $2^{-n}$ when memory is reset by large steps (Bagarti, 2011).

These simulations validate the analytic forms and clarify the transition between diffusive, persistent, and mixed movement regimes.

5. Applications in Animal Dispersal Modeling

CGRW models provide a unified framework for modeling movement ecology:

• By varying $\mu$ and $\sigma$, the model covers a spectrum from random-walk-like diffusion (weak or absent memory) to persistent (ballistic) movement and cluster-and-jump patterns.
• The strategy function $\mu$ represents the organism's behavioral tendency, and $1/\sigma$ captures memory fidelity.
• Matching such models to tracking data allows inference of behavioral parameters and memory scales, elucidating the mechanistic origins of observed movement patterns (Bagarti, 2011, Michelot et al., 2018).

Extensions to continuous-time settings (e.g., the continuous-time correlated random walk, CTCRW) further enable integration with irregularly sampled telemetry data and regime-switching models for state-dependent movement (Michelot et al., 2018).

6. Connections to Excursion Set Theory and Random Environments

CGRWs appear in diverse contexts beyond movement ecology:

• Excursion Set Theory: In cosmological structure formation, the smoothed density field $\delta(S)$ as a function of variance $S$ follows a CGRW. Filter-induced correlations render the process non-Markovian. The first-crossing distribution (e.g., for halo formation thresholds) can be computed via path-integral and integral-equation approaches, with correlation corrections formulated in terms of the deviation kernel $\Delta(S,S')$ (Farahi et al., 2013).
• Persistence in Correlated Environments: In random walks in correlated Gaussian environments, step correlations govern persistence probabilities and critical exponents. For $H \in [1/2,1)$ correlations regular-varying with index $2H-2$, the survival probability decays like $(\log N)^{-(1-H)/H}$, with implications for branching processes with correlated reproduction (Aurzada et al., 2016).

These connections show that CGRW-type dependencies play foundational roles in fields ranging from statistical physics to astrophysics to probability theory.

7. Extensions and Open Problems

While CGRW frameworks are analytically tractable in several regimes, challenges remain:

• Exact expressions for non-linear or high-dimensional strategy functions $\mu$ are often unavailable; analysis is restricted to asymptotic or simulation-based investigation (Bagarti, 2011).
• Determination of slowly-varying terms in persistence exponents in correlated environments is incomplete (Aurzada et al., 2016).
• Functional limit theorems for branching processes under correlated environments require further study, especially under heavy-tailed, non-Gaussian, or ergodic settings (Aurzada et al., 2016, 0809.1612).
• The extension of these models to state-switching or regime-dependent frameworks is active, leveraging computational methods such as MCMC and Kalman filtering for inference in high-noise or irregularly-observed settings (Michelot et al., 2018).

CGRW models remain a versatile and foundational tool in modern stochastic analysis, providing the structural basis for modeling persistence, memory, and complex dependence in random walk processes across natural and applied sciences.