[time-nuts] Update on Rb Performance

[WarrenS]

John said in part;

> ignore the  last two (ADEV) plot points as there isn't enough data for
> them to be very meaningful.
> you need a lot more than 10 days data to draw any real conclusions;

IMHO, ADEV is not the right tool or even a very useful tool for evaluation
the long term performance of these Rb Osc.
Whenever possible, best to record the Raw data so other tools,
such as available in TimeLab can be used to better compare the different 
Rbs.

I have to wonder how many others notice ADEV's severe limitation
in its able to reliable predict even 1 day in the future from 10 days
worth of past data.
Where as, ADEV is a great tool for measuring 1 sec noise,
It is a poor tool to use for predicting future 1 day errors.
This is especially true when the majority of the errors are due to
measurable systematic things such as ageing and tempCo,
as is the case with these Rb oscillators.

This seems like a case of using the wrong tool for the job.
ADEV's usefulness and power is its ability to predict future
errors from past "random" Noise.
For this, it helps to have at least a 10 to one and preferable a 100
to one ratio of raw data to tau.
Because of the randomness of "Noise", in order to get good consistent
ADEV results, The larger the ratio of recorded data to tau the better the
answers.
So to get a good consistent 1 sec ADEV answer,
best to have 100 seconds or more of raw data.

To get good 1 day ADEV answers from "noisy" data,
would need to have 100 or so days worth of raw data.
One of the ironic things of trying to use ADEV plots to predict long term
future errors, is that during the 100 days of recording raw data,
the systematic things that are causing the errors will have likely changed
enough that the raw long data run may still be near useless to predict
future errors over even the next day or two using ADEV.

The ratio of time that raw data needs to be collected compared to
future predicable time, can be greatly improved by using any number
of more appropriate tools.
By using something rather than ADEV, taking 10 days worth of the right data,
one can then better predict the next 10 or even 100 days in the future when
the major error sources are systematic rather than Random.

from  <http://en.wikipedia.org/wiki/Allan_variance>
"The Allan variance is intended to estimate stability due to noise processes
and not that of systematic errors or imperfections such as frequency drift
or temperature effects"

ws

********************
----- Original Message ----- 
From: "John Ackermann N8UR"
Subject: Re: [time-nuts] Update on Rb Performance

Hi Warren -- for these tests, I wasn't capturing raw data, just using the
tables and graphs that come out of the TSC box.

John

*************
On Feb 18, 2012, at 11:57 PM, "ws at Yahoo"

> John
>
> If you have the raw phase data, can you post a plot of what the well
> filtered freq offset looks like over that 10 day period?
>
> I've have found a properly filtered high resolution freq vs. time plot
> provides a lot more useful information than the couple of data numbers of
> a ADEV plot for evaluating long term performance of an Osc and helps
> separate all the many different possible causes of  poor ADEV numbers.
>
> This is because then one can see the shape and magnitude of the Freq
> drift, therefore being able to see if the freq drift has a short term
> cycle due to temperature or if it is linear due to ageing or 2nd order due
> to still stabilizing or if it contains freq jumps due to 1/f flicker, or a
> single large jump due ...etc,  etc.
>
> To be of any long term use, the freq data must be filtered over a long
> enough time period, such as a 1 hr running averaged, so the plot is more
> than just the 1 sec noise shown on most freq plots.
...
> ws
> ****************
>
> John Ackermann N8UR jra at febo.com
>
> This isn't the real long-term stability test I'm planning to do, but I
> did let the measurement continue on the last unit I was testing (an
> Efratrom FRS-type) out to 10+ days, which should give fairly reasonable
> data out to 100K seconds.  An ADEV plot is attached.  I would ignore the
> last two plot points as there isn't enough data for them to be very
> meaningful.
>
> Bottom line is that Efratom specs the FRS units at <1e-10/day, and this
> one seems to do more than an order of magnitude better.  But also looks
> like you need a lot more than 10 days data to draw any real conclusions;
> you can look at this plot and think that the ADEV is maybe heading back
> down after a peak near 1e-11.
>
> John