Abstract
To capture the slowly time-varying spectral content of real-world time-series, a common paradigm is to partition the data into approximately stationary intervals and perform inference in the time-frequency domain. However, this approach lacks a corresponding nonstationary time-domain generative model for the entire data and thus, time-domain inference occurs in each interval separately. This results in distortion/discontinuity around interval boundaries and can consequently lead to erroneous inferences based on any quantities derived from the posterior, such as the phase. To address these shortcomings, we propose the Piecewise Locally Stationary Oscillation (PLSO) model for decomposing time-series data with slowly time-varying spectra into several oscillatory, piecewise-stationary processes. PLSO, as a nonstationary time-domain generative model, enables inference on the entire time-series without boundary effects and simultaneously provides a characterization of its time-varying spectral properties. We also propose a novel two-stage inference algorithm that combines Kalman theory and an accelerated proximal gradient algorithm. We demonstrate these points through experiments on simulated data and real neural data from the rat and the human brain.
Introduction
Analyzing signals that evolve over time—such as the rhythmic firing of neurons or shifting brain rhythms—presents a fundamental challenge in data science. Traditional approaches often face a trade-off: assume the signal is "stationary" (unchanging) to leverage powerful mathematical tools, or partition the data into short windows, which can obscure long-term dependencies. The Short-Time Fourier Transform (STFT) is a common windowed approach, but it treats intervals as independent islands, leading to "jumps" or discontinuities at window boundaries. These artifacts are particularly problematic when inferring phase information, which is critical in fields like neuroscience for understanding the timing of events relative to oscillatory cycles.
The PLSO Framework
The Piecewise Locally Stationary Oscillation (PLSO) model addresses these shortcomings by providing a generative framework designed to decompose nonstationary time-series into smooth, oscillatory components while maintaining mathematical continuity across time intervals.
Unlike previous methods that only constrain parameter smoothness, PLSO enforces stochastic continuity on the signal itself. It utilizes a state-space representation where each component is modeled as a "stochastic harmonic oscillator." This ensures that even as the power of an oscillation shifts between windows, the underlying state remains continuous, effectively eliminating boundary artifacts.
Two-Stage Inference
To handle the computational complexity of nonstationary signals, the authors developed a two-stage inference algorithm:
- Spectral Tracking: Employs the Whittle likelihood (a frequency-domain approximation) and an Accelerated Proximal Gradient (APG) algorithm to estimate how the power of each frequency band changes over time.
- Signal Reconstruction: Once the spectral dynamics are known, the model uses Kalman smoothing and Forward Filtering Backward Sampling (FFBS) to reconstruct the actual time-domain waveforms and their precise phases.
Experimental Validation
The researchers validated PLSO across three domains: simulated data, rat hippocampal Local Field Potentials (LFP), and human EEG recordings.
In simulations, PLSO effectively eliminated boundary artifacts. As illustrated in Figure 1, PLSO's time-domain reconstruction closely follows the ground truth without the discontinuities characteristic of windowed methods.
In neural data analysis, PLSO's ability to preserve phase information proved vital. When analyzing rat hippocampal theta waves (5–10 Hz), PLSO successfully captured spike-phase coupling—the tendency of neurons to fire at specific phases of an oscillation. Unlike standard methods, PLSO provided statistically significant results because its credible intervals were not degraded by boundary noise.
Furthermore, in human EEG data recorded during anesthesia, the model's regularization naturally suppressed movement artifacts. These sudden bursts of noise typically create "vertical streaks" in standard spectrograms, but PLSO provided a clearer view of the shifting brain rhythms as patients transitioned through levels of consciousness.
Conclusion
PLSO offers a computationally efficient and mathematically rigorous alternative to traditional windowing and Gaussian Process-based methods. By providing a generative framework in the time domain, it enables the analysis of "messy" real-world oscillations without the distortions inherent in previous techniques. Future work may focus on automating model selection and incorporating adaptive window lengths to capture signals with varying temporal dynamics.
Paper Citation: Song, A. H., Ba, D., & Brown, E. N. (2021). PLSO: A generative framework for decomposing nonstationary time-series into piecewise stationary oscillatory components. Proceedings of the 37th Conference on Uncertainty in Artificial Intelligence (UAI).
Summary updated