Hi Simon, you can play around with different averaging ideas with e.g. allantools (or some other adev-library). The script for the attached image is here: https://gist.github.com/aewallin/a10966ac50846e264ca3ee97f2e0d832 (the script also plots time-series, PSDs, and histograms, which might illustrate things)
For an SDR-based phase-meter the key point is not that the individual sampled I/Q measurements would be superbly good (at femtosecond level) but that there are lots and lots of samples sampled quickly with some reasonable noise. If you just compute ADEV/MDEV from the raw high-bandwidth data it will average down, but not as far as we want. For white phase-noise downsampling (and reducing the bandwidth) by two decades reduces the 1-sample ADEV by one decade. The 'process gain' for an SDR thus comes from bandwidth-reduction by e.g. 6-decades and gives an ADEV-reduction of 3-decades. This is shown in the image below where at tau=1s the raw simulated data is at 2e-12, and after 6-decades of downsampling we are at 2e-15. In practice the challenge is to find an SDR that truly has white phase-noise also at low (1Hz) offset from the carrier. Usually there is an 1/f shape towards zero frequency. The averaging trick stops working at some point, and with the mentioned N210 radio we have not seen 1s ADEVs below 1e-14 (IIRC). The 2e-12/tau level for 1MS/s should be roughly right for an N210 fed with two 10MHz signals (again, IIRC). hope this helps, Anders [image: SDR_phasenoise_Figure_2.png] On Sat, May 2, 2020 at 4:22 PM Simon Lewis <[email protected]> wrote: > Hi all, > > I have a fairly newbie question on averaging of phase data, prior to ADEV. > > In a paper by Sherman and Jordens (https://arxiv.org/pdf/1605.03505.pdf) > on > oscillator metrology using SDR (Ettus N210), for long term measurements (1 > tau up) the group estimates an average phase to reduce the data volume. > They use both a novel lower-bound variance estimate, and what I assume was > a rectangular average of N samples per second (original data rate was > 1Msamples/s). Both apparently provided similar results. > > From what I've read/tried to understand, pre-averaging phase isn't > practically a good idea, considering the ADEV natively does this, and you > end up with a lower estimate than reality. They state that time interval > counters effectively do this for frequency measurements. > > I know that one can do this with caveats, and in essence this is what MDEV > does (I believe?), but is it not more 'real' to just downsample the phase > data? That is, drop N-1 samples per second (not decimating which would > filter the higher order components), and then ADEV the downsampled data. > I'm sure I'm missing some understanding here! > > Thanks, > Simon > _______________________________________________ > time-nuts mailing list -- [email protected] > To unsubscribe, go to > http://lists.febo.com/mailman/listinfo/time-nuts_lists.febo.com > and follow the instructions there. >
_______________________________________________ time-nuts mailing list -- [email protected] To unsubscribe, go to http://lists.febo.com/mailman/listinfo/time-nuts_lists.febo.com and follow the instructions there.
