On the accuracy of spectral density estimation