The effect of the number of noise eigenvectors used and quantisation errors on the performance of the MUSIC algorithm

The authors show that quantisation errors reduce the resolution of the MUSIC algorithm, and that the loss in performance is significant if less than 14 bit digitization is used. However, in practice other system errors will make the ADC requirements less stringent and 12 bit ADCs (which will provide 10 to 11 bit quantization in a practical system) should give adequate performance, since quantization errors will still not be the main cause of loss of performance. It is also shown that not all the noise subspace eigenvectors need be used to produce the MUSIC spectrum, and that using less than all the noise subspace eigenvectors will improve performance and robustness, while also reducing computation. Use of just one noise subspace eigenvector is not recommended in practice, since no averaging down of spurious cusps in the spectrum occurs, leading to false results