[time-nuts] Z3805A cooling requirements?
magnus at rubidium.dyndns.org
Tue Dec 25 02:12:51 UTC 2012
On 12/24/2012 06:47 PM, Tom Van Baak wrote:
> Hi Mark,
>> I wouldn't place much emphasis on the Adev data for Tau's of less than 80 seconds.
> Actually, just the opposite; the ADEV at short tau is very close to correct.
>> I've collected some ADEV data as well but don't entirely trust it yet.
> Right, it's the ADEV for longer tau that is completely misleading. Let me explain.
> Realize that Allan deviation numbers are statistics; essentially they predict
> how constant the future frequency might be, based on a sampling of measured
> frequency in the past. For a statistic like this to be relevant you want to
> have at least 3, but more likely tens to hundreds of past measurements in
> order to have confidence in the prediction.
Here I want to point out what this is the statistics off, and it is the
statistics noise, normalized to white frequency noise. It is however not
statistics of systematic effects.
> Plotting many Allan deviation statistics, each with a different sampling
> interval, on a log-log plot gives even more information; the slope of the
> line reveals noise types.
Which is the original intent of the ADEV/MDEV curves. MDEV is being
preferred as it helps to distinguish two noise forms that ADEV failed to
handle. For far-out noises, ADEV does just was well with less processing
> Now, there is no problem observing transient phenomenon like temperature
> changes (or phase jumps or frequency jumps or loose cables or pets
> jumping onto the bench). They show up dramatically in phase or frequency
> strip plots. You can see how quickly the effect occurs. You can measure
> the magnitude of the effect. You can measure how long it takes to recover.
> This is all useful: you get numbers like tempco or thermal Q. But using
> standard deviation or RMS or Allan deviation, or any other *statistic*
> on this data is not the right thing to do -- because you have only a sample of one.
Consider if you wrap the time-sequence to re-occur at the same period.
If this is the signal you have, then it is valid. If you include more
"un-eventfull" time and wrap that, then this wrinkle has less part of
the overall time, and thus is averaged out. Assume you keep extending
with un-eventfull time to double each time you end up averaging the
particular wrinkle out of the plot, but still only approach an
approximation as a single wrinkle only occur once.
Those, the ADEV tool-set isn't going to give you very meaningful
interpretation of that wrinkle.
> On the other hand, if you encounter tens or hundreds of these
> transients over hours or days or months, then it is perfectly
> valid to use statistics like standard or Allan deviation to
> describe the probability of the transient occurring; the
> magnitude of the effect, etc. Now you have enough events to
> offer a future prediction based on many samples in the past.
Here I don't agree. Re-occuring "wrinkles" is systematic effects, and
the impact of systematic effects is different to those of noise forms.
A sine modulation of frequency and the way we can estimate it's impact
on future time is quite different from that of inherent noise sources.
Also, it doesn't scale to the white frequency noise.
Similarly, other systematic effects should be separated out of the data
before noise analysis.
> Does this make sense? In your case "removing air flow" is only one event.
Indeed. In my experience, forced air as such does not need to be the
culprit, it just optimize the coupling between ambient temperature
changes and the oscillator. Varying forced air rate also counts as
inducing temperature gradients.
> I know it's easy to make ADEV/MDEV plots using Plotter or Timelab
> but that doesn't mean it's appropriate in every case. When your
> data has an accidental data glitch or an intentional transient,
> it's best not to use statistics to describe that one event.
In fact, looking at SP 1065 for instance, cleaning your data of such
events is assumed normal procedure.
More information about the time-nuts