I have noticed that the spectrum given by equation (16) in "Simulating switched-capacitor filters..." differs quite a bit from the "unsampled" spectrum that Spectre computes. This is true even when using a very large number of sidebands, and is most prominent when mT/RC is not all that large (e.g. 5 or so, see attached example in which I simulated 1000 sidebands). The *integral* of the spectrum seems to be always correct though.

I suspect that the derivation of (16) implicitely assumes that the two noise components Sh and Sc are uncorrelated in frequency. Is that really true? Mathematically, one could derive the spectrum by looking at the Fourier transform of the time average autocorrelation function. This seems to be quite a messy exercise though. Has anyone seen a derivation through that route?

I have already ruled out the possibility that the discrepancy I am seeing is due to the "whiteness" assumption in Ss. It is failry easy to show that Ss is white for just about any mT/RC greater than 3.

Thanks,
Boris