Abstract
In this paper, we propose a new inference procedure for understanding non-stationary processes, under the framework of evolutionary spectra developed by Priestley. Among various frameworks of modeling non-stationary processes, the distinguishing feature of the evolutionary spectra is its focus on the physical meaning of frequency. The classical estimate of the evolutionary spectral density is based on a double-window technique consisting of a short-time Fourier transform and a smoothing. However, smoothing is known to suffer from the so-called bias leakage problem. By incorporating Thomson's multitaper method that was originally designed for stationary processes, we propose an improved estimate of the evolutionary spectral density, and analyze its bias/variance/resolution tradeoff. As an application of the new estimate, we further propose a non-parametric rank-based stationarity test, and provide various experimental studies.
Introduction
In classical signal processing, the assumption of stationarity—where statistical properties remain constant over time—is often violated by real-world signals such as EEG readings, seismic waves, and economic indices. Priestley’s "evolutionary spectra" framework offers a solution by allowing frequency components to evolve over time while maintaining their physical interpretation. However, traditional estimation methods using a "double-window" technique (short-time Fourier transform followed by smoothing) are prone to bias leakage, where spectral information from different time-frequency regions distorts the estimate. Furthermore, standard stationarity tests like the Priestley-Subba Rao (PSR) test often fail when data deviates from Gaussian distributions.
Methodology
This research bridges the gap between stationary and non-stationary analysis by applying the Multitaper Method (MTM) to evolutionary spectra.
Multitaper Estimation
Instead of a single window, MTM utilizes a set of orthogonal Slepian sequences (tapers). Averaging results across these tapers significantly reduces variance and mitigates bias leakage. The relationship between the number of tapers ($K$) and the window length ($N$) is crucial for balancing resolution and stability.
Rank-Based Stationarity Test (RS Test)
Building on the improved estimator, the authors introduce a non-parametric stationarity test based on the Friedman test. By analyzing the ranks of spectral estimates rather than raw values, the RS test achieves high robustness against outliers and non-normal data distributions, providing a conservative alternative to traditional ANOVA-based methods.
Key Insights: The Non-Stationarity Penalty
A central finding of this work is the "non-stationarity penalty." In stationary analysis, increasing the window length ($N$) typically improves estimation accuracy. However, for non-stationary signals, there is a distinct tipping point.
As shown in Figure 1, if $N$ becomes too large, the signal's underlying properties change within the window, causing the Mean Squared Error (MSE) to increase (blue curve). This behavior quantifies the "approximately stationary" segment length and provides a theoretical guide for window selection in dynamic environments.
Results and Applications
Performance and Accuracy
Theoretical analysis shows the multitaper estimate's MSE scales as $O(\frac{\log N}{K} + W^2 + \frac{B_g}{B_X})$, where $W$ is resolution and $B_X$ is the process's characteristic width. In simulations, the RS test maintained an empirical size (false positive rate) close to the nominal 5% level, whereas the traditional PSR test over-rejected stationarity in up to 78% of non-Gaussian cases.
Real-World Data
- ECG Analysis: The method successfully characterized biological non-stationarity in RR interval time series from the MIT-BIH database.
- Economic Indices: Testing Real Exchange Rates (RER) revealed that while the UK and Japan have relatively stationary rates, China’s exchange rate is significantly non-stationary, challenging certain assumptions in "Purchasing Power Parity" theory.
Conclusion
By integrating Thomson’s multitaper method into the evolutionary spectra framework, this work provides a more accurate and robust toolkit for non-stationary signal analysis. While parameter selection for $N$ and $K$ remains an area for heuristic refinement, the proposed RS test offers a significant advancement in reliability for non-Gaussian datasets.
Summary updated