Abstract
We derive and study a significance test for determining whether a panel of functional time series is separable. In the context of this paper, separability means that the covariance structure factors into the product of two functions, one depending only on time and the other depending only on the coordinates of the panel. Separability is a property that can dramatically improve computational efficiency by substantially reducing model complexity. It is especially useful for functional data, as it implies that the functional principal components are the same for each member of the panel. However, such an assumption must be verified before proceeding with further inference. Our approach is based on functional norm differences and provides a test with well‐controlled size and high power. We establish our procedure quite generally, allowing one to test separability of autocovariances as well. In addition to an asymptotic justification, our methodology is validated by a simulation study. It is applied to functional panels of particulate pollution and stock market data.
Introduction
In Functional Data Analysis (FDA), researchers frequently encounter "panels"—collections of curves observed over time, such as daily pollution profiles across multiple cities or intraday stock prices for a group of companies. A primary challenge in analyzing these panels is the "curse of dimensionality." Modeling the full covariance structure, which describes how every point on every curve relates to every other point, is computationally intensive and often impractical.
To address this, researchers often assume separability. This assumption posits that the global covariance structure can be factored into two independent components: one representing the "spatial" relationship between different entities in the panel and another representing the "temporal" dynamics along the individual curves. While separability allows for significant model simplification, assuming it without verification can lead to biased results. This paper introduces a rigorous significance test to determine if a panel of functional time series is truly separable.
Addressing Temporal Dependence
A critical limitation of previous separability tests is the assumption that observations are independent and identically distributed (i.i.d.). In reality, functional data often exhibits "memory," where today's observations are correlated with yesterday's. Ignoring this temporal dependence leads to "over-rejection," where a test might incorrectly claim data is non-separable simply because it mistook temporal correlation for a lack of separability.
This paper provides a novel framework that remains accurate even in the presence of such dependence. By utilizing a Central Limit Theorem designed for weakly dependent sequences ($L^p-m$-approximable), the authors ensure the test is robust for real-world time series data.
Methodology
The proposed test is based on the functional norm difference. It compares a "free" estimate of the data's covariance to a "constrained" estimate that assumes separability. Key components of the methodology include:
- Partial Trace Operators: These are used to isolate the spatial and temporal components from the global covariance structure.
- Dimension Reduction: The infinite-dimensional curves are projected onto a finite number of Functional Principal Components (FPCs). The test is robust to the number of components chosen, provided they explain the majority (85-95%) of the variance.
- Asymptotic Theory: Using the Delta method, the authors derive the test's distribution as a weighted sum of chi-squared variables.
- Long-Run Covariance: A "Bartlett-type" estimator is employed to account for dependencies over time, ensuring the test maintains a precise 5% nominal error rate in simulations where older tests failed.
Empirical Applications
The methodology was applied to two distinct real-world datasets with contrasting results:
Environmental Monitoring: Nitrogen Dioxide ($NO_2$)
The authors examined daily maximum one-hour $NO_2$ concentrations from nine locations on the US East Coast. The test strongly rejected the null hypothesis of separability ($P < 0.0001$). This indicates that the daily "rhythm" of pollution is fundamentally different across locations (e.g., Boston vs. New York City). For environmental scientists, this means a one-size-fits-all temporal model is insufficient for this region.
Finance: Blue-Chip Stock Returns
The test was also applied to minute-by-minute intraday returns for 10 US blue-chip companies over 1,377 trading days. Interestingly, the test failed to reject separability ($P \approx 0.22$). This suggests that for these major companies, the structure of intraday price movements is sufficiently similar to allow for more efficient, separable models in risk assessment and portfolio management.
Conclusion
Verifying the assumption of separability is a vital step in functional data modeling. By providing a test that is robust to temporal dependence and scale-invariant, this research offers a practical tool for simplifying complex datasets without sacrificing statistical integrity. Future work may extend these methods to non-stationary series or higher-dimensional tensor structures.
Summary updated