State-space multitaper time-frequency analysis.

Area of Science:

• Signal Processing
• Statistical Inference
• Time Series Analysis

Background:

• Time series data are increasingly prevalent across diverse scientific fields.
• Existing time-varying spectral methods lack a statistical inference framework for entire series.
• Nonstationary time series analysis requires robust methods for accurate characterization.

Purpose of the Study:

• To introduce state-space multitaper (SS-MT) methods for statistical inference in time-varying spectral analysis.
• To provide a framework for analyzing nonstationary time series with enhanced accuracy and reproducibility.
• To enable statistical comparisons of spectral properties across arbitrary time series segments.

Main Methods:

• Modeling nonstationary time series as sequential Gaussian processes.
• Utilizing a frequency-domain random-walk model to link spectral representations across intervals.
• Employing parallel 1D complex Kalman filters and expectation-maximization for parameter estimation.

Main Results:

• SS-MT methods provide enhanced spectral resolution and significant noise reduction (up to 10 dB) compared to standard multitaper (MT).
• The framework allows for robust statistical comparisons of spectral properties between different time series segments.
• SS-MT successfully extracts time-domain estimates of signal components.

Conclusions:

• SS-MT offers a broadly applicable, empirical Bayes' framework for statistical inference in nonstationary time series.
• This paradigm ensures accurate and reproducible findings in complex time series analyses.
• SS-MT is effective for analyzing diverse data, including simulated series and EEG recordings.