[Paper Review] DCT and Eigenvectors of Covariance of 1st and 2nd order Discrete fractional Brownian motion

Anubha Gupta, ShivDutt Joshi|arXiv (Cornell University)|Feb 22, 2013

Complex Systems and Time Series Analysis11 references3 citations

TL;DR

This paper establishes that the eigenvectors of the auto-covariance matrix of 1st and 2nd order discrete fractional Brownian motion can be asymptotically approximated by discrete cosine transform (DCT) basis vectors. Using analytic perturbation theory of linear operators, it quantifies the deviation of these eigenvectors from the DCT basis, demonstrating a strong theoretical link between DCT and fractional Brownian motion processes.

ABSTRACT

This paper establishes connection between discrete cosine transform (DCT) and 1st and 2nd order discrete-time fractional Brownian motion process. It is proved that the eigenvectors of the auto-covariance matrix of a 1st and 2nd order discrete-time fractional Brownian motion can be approximated by DCT basis vectors in the asymptotic sense. Perturbation in eigenvectors from DCT basis vectors is modeled using the analytic perturbation theory of linear operators.

Motivation & Objective

• To investigate the structural relationship between the eigenvectors of the auto-covariance matrix of discrete fractional Brownian motion and the discrete cosine transform (DCT) basis.
• To analyze how closely the eigenvectors of 1st and 2nd order discrete fractional Brownian motion approximate DCT basis vectors in the asymptotic regime.
• To model the perturbation of eigenvectors from the DCT basis using analytic perturbation theory of linear operators.
• To provide a theoretical foundation for using DCT-based representations in signal processing applications involving fractional Brownian motion.
• To extend the applicability of DCT in stochastic processes by linking it to long-memory and self-similar processes.

Proposed method

• The authors derive the auto-covariance matrix of 1st and 2nd order discrete fractional Brownian motion processes.
• They analyze the asymptotic behavior of the eigenvectors of this covariance matrix as the process length increases.
• Using analytic perturbation theory of linear operators, they model the deviation of the true eigenvectors from the DCT basis vectors.
• The analysis focuses on the spectral properties of the covariance matrix and its convergence to a form dominated by DCT modes.
• Theoretical bounds on the perturbation error are derived based on operator norm estimates.
• The connection is validated through asymptotic analysis, showing convergence of eigenvectors to DCT basis vectors under specific conditions.

Experimental results

Research questions

• RQ1 To what extent do the eigenvectors of the auto-covariance matrix of 1st order discrete fractional Brownian motion approximate DCT basis vectors?
• RQ2 How does the approximation quality of eigenvectors by DCT basis vectors depend on the Hurst parameter and process length?
• RQ3 Can analytic perturbation theory accurately model the deviation of eigenvectors from the DCT basis in fractional Brownian motion?
• RQ4 What is the asymptotic behavior of the eigenvectors of the covariance matrix of 2nd order discrete fractional Brownian motion?
• RQ5 Under what conditions does the DCT basis become an optimal or near-optimal representation for the covariance structure of fractional Brownian motion?

Key findings

• The eigenvectors of the auto-covariance matrix of 1st and 2nd order discrete fractional Brownian motion converge asymptotically to DCT basis vectors as the process length increases.
• The perturbation of eigenvectors from the DCT basis is analytically bounded using the theory of linear operators, confirming the approximation's stability.
• The approximation error diminishes with increasing process length, indicating that DCT provides a progressively better representation of the covariance structure.
• The theoretical framework supports the use of DCT in efficient signal representation and compression for long-memory processes.
• The results validate the use of DCT as a near-optimal basis for fractional Brownian motion in applications requiring spectral decomposition.
• The connection between DCT and fractional Brownian motion is established rigorously through asymptotic and perturbation analysis.