One of the issues I've been facing was that the range of the eigenvalues
was quite large, and I couldn't easily see a natural threshold to set to
separate the signal and noise subspaces.
This visualisation provides a histogram for this scenario, to assist in
choosing the parameter (which is different in different configurations).
The original assumption that we can use eigh because the autocorrelation
matrix is Hermitian is not true. Although the true autocorrelation
matrix is Hermitian, this is not necessarily the case for the estimate we
use.
