[A covariance matrix correction method for effective sample size mismatch in event-related potential]

Covariance matrix estimation is a fundamental problem in event-related potential (ERP) signal processing. Conventional methods systematically overestimate the effective sample size due to temporal autocorrelation and nonstationarity in ERP data. This overestimation degrades the accuracy of covariance matrix estimation, thereby compromising the performance and stability of downstream tasks such as blind source separation and brain decoding. To address this problem, we propose an autocorrelation-corrected Ledoit-Wolf method (AC-LW), which estimates a variance inflation factor from pooled channel-wise lag-one autocorrelations and applies a bounded correction to the Ledoit-Wolf estimator. In controlled tests with strong temporal autocorrelation and nonstationary residuals, AC-LW reduced the squared Frobenius error by 79.47% compared with empirical covariance estimation and by 17.40% compared with the best-performing baseline, the Bartz-Müller method. In typical downstream task evaluations involving ERP components such as N400 and P300, the improved covariance estimation quality brought significant classification performance gains to spatial filter-based single-trial decoding. Overall, AC-LW provides a lightweight and robust correction for electroencephalography spatial covariance matrix estimation.