[time-nuts] Recommended parameters for Timelab ADEV analysis

[Magnus Danielson]

Adrian,

On 15/11/12 03:59, Adrian wrote:
> Hi Edgardo,
>
> however, the 53132A is considerably more sensitive in frequency mode as
> we have recently discovered.
> I saw an astoundingly low ADEV noise floor of some 6E-13 at t=10 sec as
> opposed to only 2.5E-11 in TI mode.
> I don't have a 53132A, only a 53131A that I don't use anymore for ADEV,
> so I can't be more specific.

No. The filtering that the 53132A does on frequency data does not make 
it more sensitive in the ADEV context, it pre-filters the data to 
improve frequency reading, yes, but that pre filtering cause bias in the 
ADEV measurement and when you compensate for that bias you are back 
where you started. We have been over this many times in the history of 
the list, there is papers explaining it, so let's not again spread this 
misconception. The 53132A is still just a 150 ps resolution counter, 
which forms the 1/f limit slope on the ADEV plot (1/f^2 power slope).

Doing a ADEV with the 53132A frequency mode gives you the stability 
measure of that measurement mode, true, the trouble is that the 
filtering will also applied to the source, so we do not get closer to 
the source while we see an lower value in the plot and fools ourselves 
that we got closer. The only thing we have achieved is a lower value, 
but the relative distance between source and counter noise limit remains 
the same.

The improvement you may do is to combat trigger jitter, so squaring the 
signal up could get you a bit closer to the counters abilities.

If you want to break the trigger jitter and resolution noise limit of 
the pure counter, besides squaring up you got to look into mixer 
enhancements, and you end up doing the Dual Mixer Time Difference (DMTD) 
game, but filterings such as that in the 53132A isn't it.

This filtering, which also applies to the later Pendulum counters such 
as CNT-90, isn't a bad thing when you want to provide a higher frequency 
resolution while still maintaining relatively high reporting rate. It's 
actually a very good method. It's just that you don't get "pure" ADEV 
that way, and we already have a systematic method of doing something 
similar called the modified Allan Deviation (MDEV) which actually builds 
on such filtering, but applied in a more systematic way. The MDEV has 
known different behaviour to the noises compared to ADEV, and this fact 
is used to separate noise-variants better. Dr. Allan even is eager to 
point out that MDEV is actually fixing what is broken with ADEV, and we 
should be using MDEV. I tend to agree.

This type of prefiltering of 53132A frequency readings will not improve 
MDEV measures either.

So, that is the wrong tool for improving our measures.

The only thing is that it for most of the quality stuff we measure won't 
do much harm, since it only dominates the part of the curve where we are 
we are usually counter limited, so we just got another shape of that, 
but it adds nothing and we gain nothing. Hence, it's not helping us one 
bit. Only subtle benefit would be that auto-scaling could work a little 
better, but I doubt it's a strong enough reason.

Please enlight me if I missed something important.

Cheers,
Magnus