Experimental approach

We first generated simulated time-series data of length $N_{t}=128$ from an auto-regressive process of order $P=2$ with a peaked spectrum. We then computed the DFT, calculated the magnitude-squared of the bins, keeping the $N=65$ unique bins as our "raw data" ${\bf x}$. We then created the $65\times 3$ matrix ${\bf A}$ that computed the first three auto-correlation lags (0 through 2), then computed the feature (5.1).