The magnitude of the reference and the signal are, in fact, identical in the frequency domain for every frequency bin. Just take IMABS() of the complex FFT of both on Excel. It's only after the dot product is taken of the reference and the signal that the magnitude drops for a phase shift. This, of course, holds for the INV FFT.
SPICE seems to give the same results. In the FFT multiply two noisy signals that are identical except for phase angles by the same reference. As might be expected this will knock down the noise peaks in both signals by a few dB, at least more than the signal peaks, and make you feel like you are doing some kind of reference filtering.
But if you divide the small phase angle signal by the large phase angle signal you get a dip at the noise frequency in the quotient This might indicate the signal with the larger phase angle has more noise. This would be expected if the "reference frequency filter" needs to have 0 phase angle for best results.
SPICE doesn't give you the real and imaginary parts so you can't see everything they are doing. That's why I'm qualifying everything so much.
Phase sensitive rectification, of course, will also give different results for different phase angles. To get the amplitude you must know the phase angle.
That may be the problem.
Thanks. Couldn't a bank be used in the time domain (PSR) when you don't know the phase angle? It might require a high SNR level.
Bret Cahill