Abstract: It has recently been observed that wide-sense self-similar processes have a rich linear structure analogous to that of wide-sense stationary processes. In this paper, a reproducing kernel Hilbert space (RKHS) approach is used to characterize this structure. The RKHS associated with a self-similar process on a variety of simple index sets has a straightforward description, provided that the scale-spectrum of the process can be factored. This RKHS description makes use of the Mellin transform and linear self-similar systems in much the same way that Laplace transforms and linear time-invariant systems are used to study stationary processes. The RKHS results are applied to solve linear problems including projection, polynomial signal detection and polynomial amplitude estimation, for general wide-sense self-similar processes. These solutions are applied specifically to fractional Brownian motion (fBm). Minimum variance unbiased estimators are given for the amplitudes of polynomial trends in fBm, and two new innovations representations for fBm are presented.
Introduction
In signal processing, wide-sense stationary (WSS) processes are the standard because their statistical properties are invariant to time shifts. This invariance allows for a powerful toolbox including Fourier transforms, linear time-invariant (LTI) systems, and spectral density analysis. However, many natural and engineered systems—ranging from network traffic to financial markets—exhibit self-similarity. In these processes, the statistics are invariant to scaling rather than shifting.
While self-similar processes are ubiquitous, they lack the elegant linear analysis framework available to stationary processes. This paper bridges that gap by developing a theory for self-similar processes that mirrors the classic stationary theory, replacing time-shifts with time-scales.
Methodology
The authors establish a unified framework for self-similar processes using three primary mathematical pillars:
- Lamperti’s Transformation: This transformation acts as the fundamental link, mapping a self-similar process to a corresponding stationary process. It effectively converts scaling operations in the self-similar domain into time-shifts in the stationary domain.
- The Mellin Transform: Just as the Laplace or Fourier transform is used to analyze frequency in stationary systems, the Mellin transform is used here to define a "scale-spectrum." This provides a scale-domain equivalent to power spectral density.
- RKHS Isomorphism: The authors prove that the Reproducing Kernel Hilbert Space (RKHS) of a self-similar process is isomorphic to the RKHS of its stationary counterpart. This allows linear problems (like estimation and detection) to be solved in whichever domain is most convenient.
By combining these tools, the authors define "Self-Similar Systems"—the scale-invariant analogues of LTI systems. This enables "spectral factorization" in the scale domain, which is essential for whitening noise or designing optimal filters.
Key Insights
- Symmetry Shift: The core realization is that scaling is a fundamental symmetry. By shifting perspective from the Laplace transform (associated with translation) to the Mellin transform (associated with dilation), the mathematics of self-similarity becomes as tractable as stationary theory.
- Polynomial Eigenfunctions: In stationary theory, complex exponentials ($e^{st}$) are the eigenfunctions of LTI systems. In self-similar theory, polynomials ($t^a$) take this role. This makes the detection and estimation of polynomial trends (e.g., linear drift) mathematically straightforward.
- Mellin-Causal Factorization: The authors demonstrate that the scale-spectrum can be factored to create causal whitening filters, leading to "innovations representations" where complex processes like fractional Brownian motion (fBm) are generated from simple white noise.
Key Results
The framework yields several practical applications for signal processing:
- Optimal Detrending: The authors derived the Minimum Variance Unbiased Estimator (MVUE) for polynomial trends in fBm. Interestingly, for fBm with a Hurst parameter ($H$) between 0.5 and 1, they found that a simple estimate using only the final data point is nearly as efficient as the mathematically optimal filter.
- New fBm Representations: The paper introduces two new ways to represent fBm as the output of a self-similar system driven by white noise, providing valuable tools for the simulation and synthesis of these processes.
Limitations and Future Work
While the framework is robust, it is currently limited to wide-sense self-similarity, focusing on first and second moments (mean and covariance). It does not yet account for higher-order statistics or processes with infinite variance, such as Alpha-stable distributions. Additionally, extending this to "jointly self-similar" processes (multi-variable signals) requires complex matrix spectral factorization, which remains a significant challenge for future research.
Summary updated