A Stationarity-and-Coupling Criterion for Training-Free Time-Lagged Spectral Embeddings of Multivariate Time Series

In three linesStudy of a training-free descriptor for multivariate time series using time-lagged correlation matrices truncated at the Marchenko-Pastur edge. Authors propose a stationarity-and-coupling criterion to predict applicability: succeeds on Sleep-EDF (88.5%), BCI-IV-2a, MIT-BIH, ESC-50 but fails on non-stationary or power-discriminated data, validating the predictive criterion.

Summary generated by Claude — human-verified