Analytic wavelets for multivariate time series analysis

Sophie Achard, Irène Gannaz, Marianne Clausel, François Roueff2017 · DOI

TL;DR

This paper introduces a novel construction for "analytic wavelets," which are complex wavelets that can better capture the phase information in multivariate time series data. This advancement is crucial for analyzing complex phenomena like long-memory processes in fields such as finance and network traffic, where phase dynamics are often overlooked by traditional methods.

What Problem Does This Paper Address?

Many real-world phenomena, like stock market fluctuations or internet traffic patterns, exhibit "long-memory" properties, meaning past events have a persistent influence on future behavior. Analyzing these multivariate time series (where multiple related signals are observed simultaneously) presents a challenge: not only do we need to understand the long-memory characteristics, but also how the different time series are coupled. Traditional wavelet analysis, which uses real-valued wavelets, often struggles with phase information, particularly at low frequencies. This can lead to a loss of crucial insights into the interdependence between time series. The paper aims to bridge this gap by developing and demonstrating the utility of analytic wavelets, which are designed to explicitly capture phase information.

Methods

The core contribution of this paper lies in the construction and theoretical validation of "approximately analytic wavelets" using a common-factor design. This approach builds upon existing work on Hilbert transform pairs of wavelets but provides a rigorous proof of existence and an explicit formula for the wavelet coefficients.

Key methodological aspects include:

  • Common-Factor Design: The construction of the analytic wavelets relies on a shared "common factor" that allows for the simultaneous design of filters for both the real and imaginary parts of the wavelet.
  • Approximation of Analyticity: Since perfect analytic wavelets with compact support are impossible, the authors focus on achieving an "approximate" analyticity. This is achieved by ensuring the wavelet's Fourier transform is nearly zero for negative frequencies.
  • Explicit Formula for Wavelet Coefficients: A significant theoretical contribution is the derivation of an explicit formula for the coefficients of the approximate analytic wavelets. This formula allows for the quantification of the quality of the analyticity approximation.
  • Proof of Existence: The paper provides a mathematical proof demonstrating that these approximately analytic wavelets can indeed be constructed for any valid parameters.
  • Application to Multivariate Long-Memory Processes: The authors then apply these analytic wavelets to analyze simulated bivariate ARFIMA(0,d,0) processes, which are a standard model for multivariate long-memory data.

Datasets

The paper does not use real-world datasets for evaluation. Instead, it relies on simulated bivariate ARFIMA(0,d,0) processes with varying long-memory parameters (d) to demonstrate the effectiveness of the proposed analytic wavelets. These simulations are designed to highlight scenarios where phase information is critical.

Key Insights and Ideas

  • Phase Matters in Long-Memory Processes: The paper strongly emphasizes that the phase of the spectral density at low frequencies is crucial for understanding the coupling between multivariate long-memory time series. Traditional real wavelets can obscure this information.
  • Analytic Wavelets Unlock Phase Information: By explicitly constructing and using analytic wavelets, researchers can recover and utilize phase information that is otherwise lost. This is demonstrated through simulations where real wavelets fail to capture important correlations.
  • Theoretical Foundation for Approximate Analyticity: The paper provides a much-needed theoretical underpinning for the construction of approximate analytic wavelets, moving beyond empirical design to rigorous mathematical validation.
  • Quantifiable Approximation Quality: The explicit formulas derived allow researchers to assess how "analytic" their wavelets are, providing control over the approximation quality based on design parameters.

Key Results

The paper demonstrates the following key results:

  • Existence of Analytic Wavelets: The authors prove that approximately analytic wavelets, constructed using the common-factor design, exist for any valid parameters.
  • Explicit Construction: An explicit formula for the wavelet coefficients is provided, enabling practical implementation and analysis of the approximation quality.
  • Improved Phase Recovery in Simulations: Simulations of multivariate long-memory processes show that analytic wavelets successfully recover the correlation coefficient (p) even in cases where the phase is critical (e.g., when d₁ - d₂ is close to 1 modulo 2). In these scenarios, real wavelets lead to vanishing correlations, making it impossible to identify p.
  • Distinguishing Real and Imaginary Parts: The analysis of wavelet coefficients from analytic wavelets reveals that both their real and imaginary parts contribute to identifying the modulus and phase of the spectral density at zero frequency, offering a more complete picture than real wavelets alone.

Limitations and Future Work

  • Numerical Instability: The authors acknowledge observing a numerical instability when varying the parameters L and M of the analytic wavelets. This issue requires further investigation in future work.
  • Theoretical Quality of Approximation: While the paper provides an explicit formula to quantify the approximation, further theoretical results on the quality of this approximation are needed. This is mentioned as a direction for future research, potentially in a separate paper.
  • Statistical Properties of Estimators: The paper suggests that deriving statistical properties of estimators based on these analytic wavelets is a crucial next step for practical applications in parameter estimation for multivariate long-memory models.
  • Algorithm Guarantee: The paper mentions a constructive proof leading to a numerical solution but notes that an algorithm is proposed "without guarantee of success," implying that further refinement or alternative algorithms might be necessary.

Summary updated

Abstract

International audience
BibTeX
@misc{achard2017analytic,
  author = {Sophie Achard and Irène Gannaz and Marianne Clausel and François Roueff},
  title = {{Analytic wavelets for multivariate time series analysis}},
  year = {2017},
  doi = {10.1117/12.2272928},
}
Figure