Stationarity of Manifold Time Series

Junhao Zhu, Dehan Kong, Z. Q. Zhang, Zhenhua LinJournal of the American Statistical Association 2024 · arXiv

TL;DR

This paper introduces the first formal definitions of stationarity for time series data that lives on curved geometric spaces (manifolds). They also develop novel statistical tests to detect this stationarity, which are crucial for analyzing complex data in fields like biology and finance.

What Problem Does This Paper Address?

Many real-world datasets are not simple vectors in Euclidean space but rather lie on curved surfaces, or "manifolds." Examples include cell-type proportions (on a sphere), or correlation matrices (on a manifold of positive-definite matrices). Analyzing these "manifold time series" requires understanding their temporal behavior, specifically their "stationarity." Stationarity implies that the statistical properties of the time series (like its mean and variance) don't change over time. However, existing definitions and tests for stationarity are designed for Euclidean data and don't account for the complexities of curved spaces. This paper bridges this gap by defining and testing stationarity for manifold time series.

Methods

The authors propose new definitions for first-order and second-order stationarity tailored for manifold time series. They then develop statistical tests for these properties.

  • First-order stationarity test: This test uses a Cumulative Sum (CUSUM) of residuals in the tangent space at the sample intrinsic mean. A key innovation is accounting for the manifold's curvature, which complicates the standard CUSUM approach. They use Gaussian multiplier bootstrap to estimate the distribution of the test statistic.
  • Second-order stationarity test: This test is based on local spectral density. It adapts techniques from functional time series analysis to the manifold setting. The test essentially checks if the spectral density of the time series remains constant over time.

Datasets

The paper uses a single-cell RNA sequencing dataset from Schiebinger et al. (2019) to demonstrate the practical application of their methods.

Key Insights and Ideas

  • Curvature Matters: The curvature of the manifold significantly impacts how stationarity is defined and tested. Standard Euclidean methods are insufficient.
  • Formalizing Stationarity on Manifolds: The paper provides the first rigorous definitions of stationarity for these complex data types, enabling more precise analysis.
  • Bootstrap for Manifold Data: The use of Gaussian multiplier bootstrap is extended to handle the complexities of manifold data, providing a robust way to estimate test statistic distributions.
  • Biological Relevance: The methods are shown to be applicable to real biological data, providing insights into cellular development dynamics.

Key Results

  • First-order stationarity test: Numerical simulations show that the proposed method controls Type-I error rates well, outperforming methods that ignore curvature or treat manifold data as Euclidean. The test also demonstrates good power to detect non-stationarity.
  • Second-order stationarity test: Simulations show that the test has good Type-I error control, though it can be slightly inflated for small sample sizes due to limitations of the adapted method. The test also exhibits increasing power as the deviation from stationarity grows.
  • Real Data Application: Applying the first-order stationarity test to single-cell RNA sequencing data revealed significant temporal changes in cell-type proportions, indicating that the system was not in dynamic equilibrium. The second-order stationarity test provided numerical support for an assumption of constant uncertainty in the biological model.

Limitations and Future Work

  • Block Size Dependency: The spectral density-based test for second-order stationarity depends on the choice of block size, which currently lacks a universally accepted benchmark. This is a common issue in time series analysis and requires further improvement.
  • Generalizing to Metric Spaces: Extending the framework to general metric spaces is a challenging open question.
  • Detecting Breakpoints: The authors suggest future work on extending their framework to detect structural breaks (change points) in manifold time series, particularly in neuroscience applications.
  • Wasserstein Space: Generalizing the framework to Wasserstein spaces, which can be viewed as infinite-dimensional Hilbert manifolds, is another interesting avenue for future research.

Summary updated

Abstract

In modern interdisciplinary research, manifold time series data have been garnering more attention. A critical question in analyzing such data is ``stationarity'', which reflects the underlying dynamic behavior and is crucial across various fields like cell biology, neuroscience and empirical finance. Yet, there has been an absence of a formal definition of stationarity that is tailored to manifold time series. This work bridges this gap by proposing the first definitions of first-order and second-order stationarity for manifold time series. Additionally, we develop novel statistical procedures to test the stationarity of manifold time series and study their asymptotic properties. Our methods account for the curved nature of manifolds, leading to a more intricate analysis than that in Euclidean space. The effectiveness of our methods is evaluated through numerical simulations and their practical merits are demonstrated through analyzing a cell-type proportion time series dataset from a paper recently published in Cell. The first-order stationarity test result aligns with the biological findings of this paper, while the second-order stationarity test provides numerical support for a critical assumption made therein.
BibTeX
@article{zhu2024stationarity,
  author = {Junhao Zhu and Dehan Kong and Z. Q. Zhang and Zhenhua Lin},
  title = {{Stationarity of Manifold Time Series}},
  journal = {Journal of the American Statistical Association},
  year = {2024},
}
Figure