Subspace methods break down below a critical SNR (typically 0-10 dB). Parametric AR methods are more robust at low SNR but assume the model order is known incorrectly.
– Schmidt (1986): MUSIC separates the observation space into signal subspace and noise subspace using eigenvalue decomposition of the autocorrelation matrix. modern spectral estimation theory and application pdf
| Method | Resolution | SNR Threshold | Computational Cost | Best For | | :--- | :--- | :--- | :--- | :--- | | | Poor (1/N) | High | Low | Real-time, high SNR, long data | | AR (Burg) | High | Moderate | Medium | Short data records, speech, EEG | | MUSIC | Super (theoretical) | Needs moderate-high SNR | High (SVD) | DOA estimation, harmonic retrieval | | ESPIRIT | Super | Needs moderate-high SNR | Medium-High | Radar array processing | Subspace methods break down below a critical SNR
: Improves on the periodogram by averaging spectra from overlapping segments to reduce variance. | Method | Resolution | SNR Threshold |