An F-Type Normal-Reference Calibration for High-Dimensional Covariance Testing
Abstract
An F-type normal-reference calibration is developed for testing equality of two high-dimensional covariance matrices with unknown means. The test statistic is a ratio whose numerator measures the Frobenius covariance discrepancy and whose denominator estimates its null expectation. Their difference is the complete Li-Chen statistic. Under suitable moment and trace conditions, we establish a Gaussian-reference approximation. After standardization, the statistic and its reference share the normal or nonnormal limits determined by a convergent normalized spectrum. The reference is a ratio of independent chi-square-type mixtures. Applying Welch-Satterthwaite matching to its two components gives an ordinary F approximation, with degrees of freedom estimated from independent four-observation blocks. The effect of estimating the population means is controlled by an exact expansion, and local-power and consistency results are established. The ordinary F approximation is asymptotically valid under a diffuse spectrum, whereas a covariance spike can produce a nonvanishing approximation error. Simulation studies examine empirical size and power under normal and nonnormal sampling. Averaging the trace estimates over fixed partitions improves their stability, but does not remove size inflation in the strongly correlated settings.
Related
- High-Dimensional Two-Sample Covariance Testing with Null-Preserving Transformations
- Calibration of robust covariance estimation under heavy tails
- Matrix-Aware Proper Scoring Rules and Significance Testing for Correlation and Covariance Forecasts in Python
- A Three-Way Testing Framework for Quantifying Epistemic Calibration Uncertainty in SBI
- High-Dimensional Regularization of the Spatial Sign Covariance Matrix for Robust Shape Estimation
BibTeX
@misc{wang2026ftype,
title = {{An F-Type Normal-Reference Calibration for High-Dimensional Covariance Testing}},
author = {Pengfei Wang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.30497},
eprint = {2610.30497},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.30497},
}
RIS
TY - UNPB
TI - An F-Type Normal-Reference Calibration for High-Dimensional Covariance Testing
AU - Pengfei Wang
PY - 2026
DA - 2026/10/21
PB - arXiv
DO - 10.48550/arxiv.2610.30497
UR - https://arxiv.org/abs/2610.30497
AB - An F-type normal-reference calibration is developed for testing equality of two high-dimensional covariance matrices with unknown means. The test statistic is a ratio whose numerator measures the Frobenius covariance discrepancy and whose denominator estimates its null expectation. Their difference is the complete Li-Chen statistic. Under suitable moment and trace conditions, we establish a Gaussian-reference approximation. After standardization, the statistic and its reference share the normal or nonnormal limits determined by a convergent normalized spectrum. The reference is a ratio of independent chi-square-type mixtures. Applying Welch-Satterthwaite matching to its two components gives an ordinary F approximation, with degrees of freedom estimated from independent four-observation blocks. The effect of estimating the population means is controlled by an exact expansion, and local-power and consistency results are established. The ordinary F approximation is asymptotically valid under a diffuse spectrum, whereas a covariance spike can produce a nonvanishing approximation error. Simulation studies examine empirical size and power under normal and nonnormal sampling. Averaging the trace estimates over fixed partitions improves their stability, but does not remove size inflation in the strongly correlated settings.
ER -