Abstract
We consider the estimation of high-dimensional covariance and correlation matrices under slow-decaying temporal dependence. For generalized thresholding estimators, convergence rates are obtained and properties of sparsistency and sign-consistency are established. The impact of temporal dependence on convergence rates is also investigated. An intuitive cross-validation method is proposed for the thresholding parameter selection, which shows good performance in simulations. Convergence rates are also obtained for banding method if the covariance or correlation matrix is bandable. The considered temporal dependence has longer memory than those in the current literature and has particular implications in analyzing resting-state fMRI data for brain connectivity studies.
Introduction
In the era of high-dimensional data, estimating the relationships between a large number of variables ($p$) from a limited number of observations ($n$) is a fundamental challenge. Traditional statistical methods for estimating covariance and precision matrices typically rely on the assumption of Independent and Identically Distributed (i.i.d.) observations. However, in fields such as finance and neuroscience, data points collected over time often exhibit significant temporal dependence. While much existing research focuses on "short-memory" dependence that decays rapidly, real-world signals like resting-state fMRI (rfMRI) scans of the brain often display "long-memory" dependence. This paper extends high-dimensional matrix estimation techniques—specifically generalized thresholding and CLIME—to accommodate such long-range temporal "drag."
Methodology
The researchers investigate several estimation techniques under the lens of temporal dependence:
- Generalized Thresholding: Applied to covariance and correlation matrices to eliminate noise and retain significant relationships.
- CLIME & SPICE: Methods for estimating the precision matrix (the inverse of the covariance matrix), which identifies direct connections between variables by accounting for the influence of others.
Innovation in Dependence Measurement
A key contribution of this work is a new measure of temporal dependence based on the Frobenius and spectral norms of autocorrelation matrices. Unlike traditional "mixing conditions," which are often difficult to verify in practice, this measure allows for a precise quantification of how temporal memory slows the convergence of high-dimensional estimates.
Gap-Block Cross-Validation
Standard cross-validation techniques often fail with time-series data because they assume observations can be shuffled. To address this, the authors propose "Gap-Block Cross-Validation," which involves splitting the data into blocks separated by temporal "gaps." This ensures that the training and validation sets remain sufficiently uncorrelated, respecting the underlying time-series structure.
Key Insights and Results
The study provides both theoretical and empirical evidence regarding the impact of long memory on statistical estimation.
Empirical Validation of Long Memory
Analysis of rfMRI data from the WU-Minn Human Connectome Project (comprising 1,190 images of 907 functional nodes) confirms that brain connectivity signals follow a Polynomial-Dominated Decay (PDD) model. The observed decay rate ($\alpha$) was approximately 0.30, significantly violating standard short-memory assumptions.
Theoretical Convergence and Sparsistency
The paper establishes rigorous convergence rates for sub-Gaussian, generalized sub-exponential, and polynomial-type tail distributions. Notably, the authors prove that methods like CLIME can still achieve minimax optimal convergence rates even with long-memory dependence, provided the dependence is not excessively extreme. Furthermore, they demonstrate "sparsistency"—the ability to correctly identify zero entries in a matrix with high probability—even under these challenging temporal conditions.
Brain Mapping Findings
In the rfMRI analysis, the proposed methods revealed:
- 1.47% of brain node pairs are marginally connected (simple correlation).
- 2.71% are directly connected (partial correlation via CLIME).
- The identification of "hubs" in the right inferior parietal cortex, which are known components of the brain's default mode network.
Discussion and Future Work
While this research provides a robust framework for high-dimensional estimation under long-memory dependence, several areas for future exploration remain. The current model primarily focuses on autocorrelation within individual variables rather than complex cross-correlations between different variables across time lags. Additionally, while Gap-Block Cross-Validation performed well in simulations, a formal theoretical justification for this specific method is still needed. Finally, the assumption of stationarity may need to be relaxed for longer fMRI sessions where brain states may shift dynamically.
Summary updated