Introduction
Fractional Gaussian noise (fGn) and fractional Brownian motion (fBm) are essential stochastic models used to describe processes with long-range dependence (LRD). These models are characterized by the Hurst exponent ($H$), which quantifies the intensity of the signal's "memory." In practical applications—ranging from network traffic analysis to economic forecasting—these continuous signals must be sampled and quantized into digital formats. However, the quantization process introduces errors. This paper investigates whether these errors inherit the complex correlations of the original signal or if they can be treated as simple white noise, a distinction critical for the accurate estimation of $H$.
Theoretical Framework
The study employs high-resolution quantization theory to analyze discrete-time fGn and fBm signals modeled as ARFIMA (Auto-Regressive Fractionally Integrated Moving Average) processes. By representing quantization error through a sigma-delta modulator framework and utilizing characteristic functions, the authors provide formal proofs (Theorems 1 and 2) regarding the error's distribution. The analysis covers the full spectrum of the Hurst exponent: anti-persistent ($H < 0.5$), standard Brownian motion ($H = 0.5$), and persistent/long-memory ($H > 0.5$) behaviors.
Results and Analysis
Asymptotic Uniformity and Independence
The mathematical proofs demonstrate that as quantizer resolution increases, the error distribution converges to a uniform distribution between $[-1/2, 1/2]$. Crucially, the error is shown to be asymptotically uncorrelated with itself over time and uncorrelated with the signal itself. This "white noise" simplification holds regardless of the complexity or long-range dependencies inherent in the underlying fGn or fBm process.
Principal Component Analysis (PCA)
To validate these findings, the authors performed PCA on simulated signals. As shown in the eigenvalue distribution, the fBm signal exhibits a distinct power-law decay in its eigenvalues, reflecting its high correlation. In contrast, the quantization error produces a flat eigenvalue spectrum, which is the mathematical signature of uncorrelated white noise.
Spectral Flatness
Further validation is provided through Power Spectral Density (PSD) analysis. The PSD of the quantization error remains flat across the normalized frequency range. This spectral flatness confirms that the digitization process does not "leak" the long-range dependencies of the original signal into the noise term, provided the resolution is sufficiently high.
Conclusion
This research establishes that under high-resolution conditions, quantization error in LRD signals behaves as additive white noise. While this simplifies the modeling of digital stochastic processes, the authors note that this assumption may fail in low-bit environments (e.g., 1-bit or 2-bit quantization) where the error could become correlated with the signal. This work provides a theoretical foundation for refining Hurst exponent estimators to account for digitization noise in real-world datasets.
Summary updated