[time-nuts] New algorithm, better ADEV

[Magnus Danielson]

Tom,

On 01/18/2015 02:44 PM, Tom Van Baak wrote:
> Li Ang,
>
> Nice data. Thanks for sharing.
>
> In a "self test" like this, try using a phase plot ('p') or TDEV plot ('t'). As you see, the X72 ADEV or MDEV plot is sort of boring, showing mostly a -1 slope to forever.

In the phase (and frequency) plot, I like to use the linear residue plot 
(toggle with 'r') to see the residues. Removing the offset and linear 
slope of the data, you can usually get a better view on what's going on 
in the phase.

Swapping around between wrapped phase 'w', phase 'p' and frequency 'f' 
is may somemtimes help to understand what's going wrong. Sometimes the 
wrong frequency data makes TimeLab unable to unwrap the phase correctly, 
but that is much more uncommon now.

Swapping between TDEV 't' and MDEV 'm' helps in similar way as swapping 
between phase and frequency.

To check if there is a drift limit at the upper end, using the Hadamard 
deviation 'h' helps.

It would help immensely to have an FFT plot of the data, since it 
provides a much better plot on systematics. Then again, the FFT would 
probably also need the phase and frequency scalings for similar reasons 
as the time and noise stability tests.

> A phase plot gives you an idea of the instrument rms and peak-to-peak noise. It also reveals experimental setup or run errors. For example, you appear to have a problem (unusual phase jump) near second #86760.

Some disruptions, which is easy to see in phase/frequency plot, will 
completely spoil your ADEV and friends.

> If you have a clean data set, a TDEV plot gives you a view of the time resolution of the instrument. Because it removes the -1 slope of MDEV it more clearly shows resolution as a function of time. When it rises at larger tau, then you have evidence of environmental sensitivity in your instrument.

I too strongly advice people to use the TDEV plot as a complementary 
view, because the ADEV/MDEV may only show a subtle wobble, but the TDEV 
shows a much clearer illustration of the phase stability effect. It 
naturally all depends on what "problem" you have, but it should be a 
good habbit to have these different views as they high-light different 
problems better.

> Now, the next step is using two different, but very stable, references that have a known frequency offset. The problem with using the same REF/DUT is that you do not get exposure across the entire 100 ns period of the 10 MHz frequency. An alternative is to try the experiment a few times, with say, a 10, 20, and 50 ns cable delay in one of the paths.

Indeed. The systematics of a counter/setup varies over the period of the 
reference. Some of this is due to cross-talk between channels, some of 
them is due to cross-talk between reference and channel, some of it is 
systematic noise in creating a higher frequency coarse counter from the 
reference and then some of it is non-linearities in the interpolator 
within the period of the coarse counter period.
Sweeping over this using frequency offset will average any systematic 
offsets out.

De-correlating the instances of start and stop with cable delay will 
remove the cross-talk between the channels. Usually a 10 ns cable 
suffice to get cross-talk between channels almost completely canceled.

Counter design and it's many subtleties is so much fun. :)

Cheers,
Magnus