This type of artifact isn't unheard-of, but it's usually not hard to spot because the divots in the PN trace don't look much like a real property of the DUT. At least with direct-to-ADC digital analyzers, the effect of spectral collapse on the noise plot seems to be time variant in many instances. If you wait a while or repeat the measurement the artifact will often resolve itself... but of course it might come back later. It's good that the NIST folks are trying to take some of the empiricism out of the picture, as I've never been content with my own understanding of where some of these effects really come from.
The spectral-collapse phenomenon is likely causing some headaches in some R&D work I'm doing with multichannel downconverters. The more hardware that you add to the signal paths between the input splitter and ADCs, the more opportunities there are for phase drift, various sources of EMI, and differential delays to corrupt the cross spectrum. As the paper points out, these artifacts can potentially corrupt multiple FFT segments and be mistaken for a genuine aspect of the measurement, if they share the same noise slope and are stable enough. The implications of crosstalk are pretty scary in the latter regard. The longer paper points out that these artifacts are independent of the Syx estimator used to represent the complex cross spectrum average as a positive real that can be displayed on a log scale. That's true in the long run, since the I component will dominate the plot regardless of estimator after the instrument noise falls out, but it's also true that plots made with the vector magnitude estimator sqrt(<avg I>^2 + <avg Q>^2) look a lot better when you watch them evolve in real time. Using an estimator based on I alone such as Rubiola's recommended max(<avg I>, FLT_MIN) gives you a 3 dB SNR advantage during the time while the measurement is converging, when the average I and Q magnitudes are similar, but it also exhibits many more transient "divots" during that part of the measurement since it depends on only one averaged variable rather than two. Real-only estimators look bad enough in practice -- and provide such a small advantage at the end of the day -- that I chose to use the magnitude estimator in the TimePod/3120A driver. I suspect that most other commercial signal analyzers do this as well, for the same reason. Ideally the instrument would allow you to choose the estimator... but look at all of the stuff the user would have to read in order to understand how to use the feature. :) -- john Miles Design LLC > -----Original Message----- > From: [email protected] [mailto:[email protected]] On > Behalf Of [email protected] > Sent: Tuesday, May 27, 2014 3:32 AM > To: [email protected] > Subject: [time-nuts] Over optimistic PN measurements > > There has been some controversy over the Phase noise of CMOS logic devices. > Perhaps the apparently anomalously low PN measures are due to the use of > cross > correlation in the phase noise measurement equipment and the occurrence of > phenomena detailed in the recent NIST papers: > > http://tf.boulder.nist.gov/general/pdf/2697.pdf > > http://tf.boulder.nist.gov/general/pdf/2698.pdf > > Such collapse of the cross-spectral function may also be present in the PN plots > shown in the datasheets for some OCXOs. > > Bruce _______________________________________________ time-nuts mailing list -- [email protected] To unsubscribe, go to https://www.febo.com/cgi-bin/mailman/listinfo/time-nuts and follow the instructions there.
