PCA based Hurst exponent estimator for fBm signals under disturbances

Li Li, Jianming Hu, Yudong Chen, Yi ZhangIEEE TSP 2009 · OpenReview

Abstract

In this paper, the validity of the PCA eigenspectrum-based Hurst exponent estimator for single fractional Brownian motion (fBm) signals is proved. Furthermore, the application of this estimator to fBm signals corrupted by external disturbances is discussed. Theoretical analysis and experimental results demonstrate that the estimator remains effective for: 1) mixed fBm signals with varying Hurst exponents, 2) fBm signals corrupted with additive Gaussian white noise (provided the signal-to-noise ratio is sufficient), and 3) fBm signals corrupted with additive deterministic sine/cosine signals. However, estimation accuracy is shown to be sensitive to the SNR in the first two scenarios.

Introduction

Fractional Brownian motion (fBm) is a stochastic process widely utilized to model complex systems exhibiting "long-term memory" or self-similarity. This behavior is characterized by the Hurst exponent ($H$), a parameter critical for understanding the dynamics of everything from financial markets to biological signals.

One of the most effective ways to estimate $H$ is through Principal Component Analysis (PCA). This method relies on the relationship between the Hurst exponent and the rate at which the eigenvalues ($\lambda_n$) of the signal's autocorrelation function decay. Specifically, for an fBm signal, the eigenvalues are expected to follow a power-law decay: $\lambda_n \sim n^{-2H-1}$. This paper provides a rigorous theoretical foundation for this decay rate and explores the estimator's robustness in real-world conditions where signals are rarely pure.

Theoretical Foundation: Repairing the Proof

The validity of the PCA-based estimator hinges on the Karhunen-Loève (K-L) expansion, which represents a stochastic process as an infinite linear combination of orthogonal functions. In previous literature, attempts to prove the $n^{-2H-1}$ decay rate encountered mathematical hurdles. Specifically, some proofs utilized series expansions where the basis functions were linearly independent but not strictly orthogonal.

In signal processing, treating the coefficients of a non-orthogonal expansion as if they were eigenvalues of a K-L expansion is a significant oversight. This paper addresses this by employing orthogonal projections. By projecting non-orthogonal basis functions (a Riesz basis) onto a set of orthonormal basis functions, the author establishes a direct mathematical link between the expansion coefficients and the true eigenvalues. Using asymptotic analysis of the zeros of Bessel functions ($J_{-H}$ and $J_{1-H}$), the paper confirms that the expected power-law decay holds true despite the non-orthogonality of the initial basis.

Robustness Under Disturbances

While the theoretical proof validates the estimator for pure fBm signals, practical applications often involve "corrupted" data. The paper extends the analysis to three common types of disturbances:

1. Mixed fBm Signals

When multiple fBm processes with different Hurst exponents are superimposed, the PCA estimator can still provide meaningful results. However, the resulting estimate is influenced by the relative strengths (SNR) of the constituent signals, often reflecting the dominant component's characteristics.

2. Additive Gaussian White Noise

In the presence of white noise, the eigenvalue spectrum is shifted. While the power-law decay of the fBm component remains visible in the lower-order eigenvalues, the noise floor eventually dominates the higher-order eigenvalues. The estimator remains valid as long as the SNR is not too small, allowing the $n^{-2H-1}$ trend to be extracted before the noise floor is reached.

3. Deterministic Sine/Cosine Signals

Deterministic disturbances, such as periodic oscillations, often appear as spikes or specific deviations in the eigenspectrum. The study shows that the PCA-based approach is remarkably resilient to these additive deterministic signals, maintaining its ability to estimate the underlying $H$ value of the stochastic component.

Key Insights

  • The Role of Riesz Bases: The paper demonstrates that even if a basis is not orthonormal, it can be mapped to an orthonormal one using a linear bounded bijective operator while preserving essential asymptotic properties.
  • Bridging Methodologies: The work successfully links the "direct attack" on integral kernels with series expansion approaches, providing a unified view of fBm eigenvalue asymptotics.
  • Fundamental Limits: The paper notes that for very smooth kernels, the decay rate $O(n^{-1})$ represents a fundamental limit, suggesting that estimator behavior may shift as $H$ approaches 0.

Conclusion

By providing a rigorous mathematical "remedy" for the eigenvalue decay proof, this paper solidifies the theoretical standing of PCA-based Hurst exponent estimators. The confirmation of the $\lambda_n \approx n^{-2H-1}$ relationship, combined with the demonstrated robustness against noise and deterministic disturbances, ensures that this method remains a powerful tool for analyzing long-memory processes in diverse scientific fields.

Summary updated

Abstract

In this paper, the validity of PCA eigenspectrum based Hurst exponent estimator proposed in[J. B. Gao, Y. Cao, and J.-M. Lee, ldquoPrincipal Component Analysis of 1/ <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">f</i> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">alpha</sup> noise,rdquo <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Phys.</i> <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Lett.</i> <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</i> , vol. 314, no. 5-6, pp. 392-400, 2003] for single fBm signal is proved. Moreover, how to apply this estimator for fBm signals corrupted with some other signals are discussed. Theoretical analysis and experiments show that it can also be used for 1) mixed fBm signals with different Hurst exponents, 2) fBm signals corrupted with additive Gaussian white noise when the signal-to-noise ratio (SNR) is not too small, and 3) fBm signals corrupted with additive deterministic sine/cosine signals. However, the estimation accuracy depends on the SNR value for the first two situations.
BibTeX
@article{li2009pca,
  author = {Li Li and Jianming Hu and Yudong Chen and Yi Zhang},
  title = {{PCA based Hurst exponent estimator for fBm signals under disturbances}},
  journal = {IEEE Trans. Signal Process.},
  year = {2009},
  url = {https://openreview.net/forum?id=kt7AQitRCF},
}
Figure