Frequency estimation accuracy of ROCKET

We assess the frequency estimation accuracy of the recently introduced reduced rank autoregressive linear predictor called reduced order correlation kernel estimation technique (ROCKET). We compare the frequency estimation performance of ROCKET to both the conventional full rank autoregressive (FR-AR) method and the theoretical limit imposed by the Cramer-Rao bound (CRB). The analysis includes estimation accuracy as a function of signal-to-noise ratio (SNR), data length, and subspace rank. Simulations reveal that ROCKET can approach the CRB for a much greater range of SNR levels and for shorter data sequences than FR-AR. Perhaps more importantly, ROCKET's performance is shown to be very robust to subspace rank selection. This means that a priori knowledge of the upperbound of the number of frequencies present is not crucial to this reduced rank algorithm. Finally, it is shown that a small frequency estimation bias appears when the subspace rank is well below the signal rank.