Evaluating Gaussianity of heterogeneous fractional Brownian motion

Technical Summary: Evaluating Gaussianity of Heterogeneous Fractional Brownian Motion

Problem Statement
Heterogeneous diffusion processes, characterized by time-varying dynamics and fluctuations in diffusivity, are prevalent in diverse fields ranging from protein motion in living cells to financial markets. While fractional Brownian motion (FBM) is a standard model for anomalous diffusion with long-range temporal correlations, it is inherently a Gaussian process. However, real-world systems often exhibit "heterogeneous FBM," where the diffusivity ($D$) or the Hurst exponent ($H$) fluctuates over time. A critical challenge in analyzing these systems is distinguishing between homogeneous FBM and heterogeneous variants. While deviations from Gaussianity are a hallmark of heterogeneity, a rigorous characterization of Gaussianity in heterogeneous FBM—specifically Switching Fractional Brownian Motion (SFBM)—has been lacking. Existing goodness-of-fit tests (e.g., Shapiro-Wilk, Kolmogorov-Smirnov) often yield complex statistics that are difficult to derive analytically for specific models. This paper addresses the need for a tractable metric to quantify non-Gaussianity in SFBM, where diffusivity fluctuates while temporal correlations are preserved.

Methodology
The authors investigate SFBM, a model where the generalized diffusion coefficient $D(t)$ is a stochastic process switching between two states ($D_+$ and $D_-$), while the underlying FBM retains its Hurst exponent $H$. The study focuses on two switching scenarios:

1. Markovian Switching: Both states have exponential dwell time distributions.
2. Non-Markovian Switching: At least one state follows a power-law dwell time distribution (specifically Pareto), leading to diverging first moments.

To characterize non-Gaussianity, the authors employ kurtosis ($K(t)$), defined as the ratio of the fourth moment to the square of the second moment (mean square displacement). They derive an exact analytical expression for the kurtosis of this heterogeneous process by:

• Expressing the SFBM process $X(t)$ as a Riemann-Liouville fractional integral with a fluctuating diffusivity.
• Utilizing the conditional expectation over Gaussian noise and applying Isserlis' theorem to evaluate the fourth moment.
• Deriving a closed-form series solution for the kurtosis in the Markovian case and performing numerical integration for the general case.

The analytical results are validated against extensive numerical simulations (100,000 trajectories) using Monte Carlo methods. Additionally, the authors compare the efficacy of kurtosis against the Hellinger distance, a metric that measures the divergence between the process's probability density function (PDF) and a reference Gaussian distribution.

Key Contributions and Results

• Exact Kurtosis Derivation: The paper provides an exact analytical expression for the kurtosis of SFBM with Markovian switching (Eq. 26), expressed as a convergent series involving the Pochhammer symbol. This allows for precise calculation without relying solely on simulation.
• Convergence to Gaussianity in Markovian Systems: For SFBM with exponential (Markovian) dwell times, the study demonstrates that while the process is non-Gaussian at short times, the kurtosis converges to 3 (the Gaussian value) at long times ($t \gg \tau_{\pm}$). This indicates that memoryless switching leads to statistical homogenization over time.
• Persistent Non-Gaussianity in Non-Markovian Systems: In contrast, systems with power-law dwell times (PL-Exp or PL-PL scenarios) exhibit persistent non-Gaussian behavior. The kurtosis remains significantly above 3 (e.g., converging to ~3.75 for $H=0.3$ and ~3.60 for $H=0.7$) even at long timescales. This highlights that heavy-tailed distributions prevent the system from becoming Gaussian.
• Initial Condition Dependence: The authors show that for non-Markovian switching, the short-time evolution of kurtosis is sensitive to initial conditions (the probability of starting in a specific diffusivity state), whereas Markovian systems rapidly lose this sensitivity.
• Comparison with Hellinger Distance: The study finds that the Hellinger distance yields results consistent with kurtosis but does not provide additional discriminatory power regarding non-Gaussianity in these specific models. The simplicity of kurtosis makes it a preferable metric.

Significance and Claims
The paper claims that kurtosis is a robust, simple, and effective metric for identifying non-Gaussian behavior in heterogeneous anomalous diffusion processes. The authors emphasize that:

1. Detection Capability: Kurtosis can effectively distinguish between Markovian heterogeneous processes (which eventually appear Gaussian) and non-Markovian processes (which remain non-Gaussian).
2. Practical Utility: Unlike complex statistical tests, kurtosis is easily calculable and interpretable, making it accessible for researchers in biophysics, finance, and other fields dealing with complex systems.
3. Experimental Implications: The convergence of Markovian SFBM to Gaussianity at long times suggests that experimental data must have sufficient temporal resolution to capture state-switching dynamics; otherwise, underlying heterogeneities may be masked.
4. Universality: The authors suggest that kurtosis may serve as a universal tool for characterizing non-Gaussianity in broader classes of heterogeneous models beyond SFBM, though this requires further research.

The study concludes that while the Hellinger distance and other PDF-based metrics are valid, kurtosis remains the primary choice for its balance of simplicity and effectiveness in quantifying the non-Gaussian nature of heterogeneous fractional Brownian motion.