PLSO: A generative framework for decomposing nonstationary time-series into piecewise stationary oscillatory components

Andrew H. Song, Demba E. Ba, Emery N. Brown

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 often introduces boundary artifacts and lacks a formal generative model. We introduce the Piecewise Locally Stationary Oscillation (PLSO) framework, which models nonstationary time-series as a superposition of latent oscillatory components. By combining a frequency-domain Whittle likelihood with a time-domain state-space model, PLSO enables smooth, continuous decomposition and robust phase estimation.

Introduction

Most real-world signals—such as brain waves (EEG) or seismic data—are nonstationary, meaning their frequency content evolves over time. The standard analytical approach involves partitioning the data into small windows and assuming stationarity within each, a method exemplified by the Short-Time Fourier Transform (STFT). However, this "windowing" creates significant challenges: independent processing of windows often leads to "jumps" or distortions at boundaries. These artifacts are particularly problematic in fields like neuroscience, where precise timing and phase information (the exact positioning of peaks and troughs) are critical for understanding neural dynamics. Furthermore, many existing methods lack an underlying generative model, hindering formal statistical inference and uncertainty quantification.

The PLSO Framework

The authors propose the Piecewise Locally Stationary Oscillation (PLSO) model, which treats a time-series as a superposition of $J$ independent oscillatory components. Unlike traditional windowed methods, PLSO is built on a state-space foundation that ensures continuity across the entire signal.

The graphical model of PLSO illustrates how latent oscillatory components ($z$) and their time-varying variances ($\sigma^2$) relate to the observed data ($y$) across multiple intervals ($M$).

Key Technical Features

  • Stochastic Harmonic Oscillators: Each component is modeled in the time domain as a stochastic oscillator, allowing it to capture a range of frequencies rather than just a single pure tone.
  • Markovian Dynamics: To ensure smooth evolution, the model imposes a relationship between the power (variance) of components in adjacent windows, effectively pooling information across the time-series.
  • Two-Stage Inference: The framework utilizes an efficient two-stage algorithm. First, it uses the Whittle likelihood (a frequency-domain approximation) and an Accelerated Proximal Gradient algorithm to estimate spectral properties. Second, it employs Kalman filtering and smoothing to reconstruct continuous time-domain trajectories and calculate phase.

Experimental Results

The researchers evaluated PLSO using simulated data, rat hippocampal Local Field Potentials (LFP), and human EEG recorded during anesthesia.

Accuracy and Continuity

In simulations involving amplitude-modulated oscillations, PLSO achieved significantly lower Mean Squared Error (MSE) than regularized STFT. Crucially, its "jump" metric—a measure of boundary discontinuity—was nearly 200 times lower than STFT-based methods. As shown in the time-domain reconstruction, PLSO maintains smoothness across window boundaries where traditional methods fail.

A comparison of signal reconstruction shows how STFT (blue) suffers from discontinuities at window boundaries (gray vertical lines), while PLSO (red) remains smooth and closely follows the ground truth (black).

Neuroscience and Clinical Applications

On rat LFP data, PLSO accurately recovered the preferred firing phase of neurons, matching results from traditional bandpass filtering while providing a more robust statistical framework. In the human EEG dataset, PLSO successfully tracked the smooth spectral transitions of "alpha" and "slow" wave signatures during propofol-induced anesthesia. While independent window methods produced noisy, flickering power estimates, PLSO's regularization yielded a clean, interpretable spectrogram.

Conclusion

PLSO addresses the fundamental tension between local stationarity and global continuity in nonstationary signal analysis. By providing a generative model with continuous phase estimation and credible intervals, it offers a superior alternative to traditional windowing techniques. Future work will focus on automatic identification of the number of components and extending the model to capture cross-frequency interactions.

Summary updated

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-f...
BibTeX
@inproceedings{song2021plso,
  author = {Andrew H. Song and Demba E. Ba and Emery N. Brown},
  title = {{PLSO: A generative framework for decomposing nonstationary time-series into piecewise stationary oscillatory components}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=dPOovDTjoAz},
}
Figure