Intermittent process analysis with scattering moments

Joan Bruna, Stéphane Mallat, Emmanuel Bacry, Jean–François MuzyThe Annals of Statistics 2015 · DOI · arXiv

Abstract

Scattering moments provide nonparametric models of random processes with stationary increments. They are expected values of random variables computed with a nonexpansive operator, obtained by iteratively applying wavelet transforms and modulus nonlinearities, which preserves the variance. First- and second-order scattering moments are shown to characterize intermittency and self-similarity properties of multiscale processes. Scattering moments of Poisson processes, fractional Brownian motions, Lévy processes and multifractal random walks are shown to have characteristic decay. The Generalized Method of Simulated Moments is applied to scattering moments to estimate data generating models. Numerical applications are shown on financial time-series and on energy dissipation of turbulent flows.

Introduction

In complex systems such as fluid dynamics and quantitative finance, signals often exhibit "intermittency"—periods of relative quiescence punctuated by sudden, high-amplitude bursts of activity. Characterizing these non-Gaussian behaviors is critical for modeling and risk assessment. Traditionally, researchers have relied on high-order polynomial moments (e.g., skewness, kurtosis) to describe these fluctuations. However, these statistics are notoriously unstable; they suffer from high variance and require vast amounts of data for reliable estimation.

This paper introduces scattering moments as a robust, multi-scale alternative. By leveraging the scattering transform—a cascade of wavelet transforms and modulus nonlinearities—the authors provide a stable, nonparametric framework to quantify intermittency and self-similarity without the instabilities inherent in high-power polynomial statistics.

Methodology: The Scattering Transform

The core of the approach is a nonexpansive operator that preserves variance while extracting multi-scale information. Unlike traditional moments that use powers ($x^q$), the scattering transform applies a sequence of wavelet transforms followed by a modulus operator ($| \cdot |$). This iterative process creates a representation that is stable to small deformations and noise.

A sample path of an intermittent process, characterized by sudden bursts of activity that are difficult to capture with standard statistics.

The authors define two primary descriptors:

  1. First-order scattering moments: These capture the general scaling properties of the process, such as the Hurst exponent in fractional Brownian motion.
  2. Second-order scattering moments: These measure the variability of "bursts" across different scales. This allows for a precise quantification of intermittency.

To bridge the gap between theoretical models and empirical data, the Generalized Method of Simulated Moments (GMM) is employed. By comparing the scattering moments of observed data against simulated models (e.g., Multifractal Random Walks), researchers can estimate model parameters with significantly lower variance than standard wavelet-based methods.

Characterizing Intermittency and Self-Similarity

A key theoretical contribution is the identification of characteristic decay patterns in scattering moments. For Gaussian processes, such as fractional Brownian motion, normalized second-order scattering moments decay with a specific "Gaussian slope" of $-1/2$. Any deviation from this slope serves as a mathematical "smoking gun" for non-Gaussian intermittency.

Second-order scattering moments for various processes. The characteristic decay and alignment across scales provide a signature for intermittency and self-similarity.

Furthermore, for truly self-similar (fractal) processes, normalized scattering moments become a univariate function of the scale ratio. This property provides a rigorous test for stationarity across scales, allowing researchers to determine if a process maintains its structural properties across different time horizons.

Empirical Applications

The framework was validated on several high-impact datasets:

  • Turbulence: Analysis of energy dissipation in a gaseous Helium jet revealed that standard models like Multifractal Random Walks (MRW) and Lévy processes fail to fully capture the nuances of real turbulence. The study found that self-similarity is violated in the inertial range, likely due to the breakdown of the Taylor hypothesis.
  • Finance: In contrast to turbulence, financial time series (Euro-Bund futures and S&P 100 index) showed strong evidence of self-similarity. Interestingly, the S&P 100 was found to be significantly more intermittent than the Euro-Bund, reflecting the higher frequency of "bursty" price movements in equity markets.

Conclusion

Scattering moments offer a powerful new lens for analyzing multiscale intermittent processes. By replacing unstable high-order polynomials with stable, nonexpansive operators, this framework enables more accurate parameter estimation and model selection. While currently focused on 1D time series, the potential for extending these rigorous mathematical characterizations to 2D textures and higher-dimensional data remains a promising avenue for future research.

Reference: Bruna, J., Mallat, S., Bacry, E., & Muzy, J. F. (2015). Intermittent process analysis with scattering moments. The Annals of Statistics, 43(1), 323-351.

Summary updated

Abstract

Scattering moments provide nonparametric models of random processes with stationary increments. They are expected values of random variables computed with a nonexpansive operator, obtained by iteratively applying wavelet transforms and modulus nonlinearities, which preserves the variance. First- and second-order scattering moments are shown to characterize intermittency and self-similarity properties of multiscale processes. Scattering moments of Poisson processes, fractional Brownian motions, Lévy processes and multifractal random walks are shown to have characteristic decay. The Generalized Method of Simulated Moments is applied to scattering moments to estimate data generating models. Numerical applications are shown on financial time-series and on energy dissipation of turbulent flows.
BibTeX
@article{bruna2015intermittent,
  author = {Joan Bruna and Stéphane Mallat and Emmanuel Bacry and Jean–François Muzy},
  title = {{Intermittent process analysis with scattering moments}},
  journal = {The Annals of Statistics},
  year = {2015},
  doi = {10.1214/14-aos1276},
}
Figure