[time-nuts] ADEV vs MDEV

[Magnus Danielson]

Pete Rawson wrote:
> Efforts are underway to develop a low cost DMTD apparatus with
> demonstrated stability measurements of 1E-13 in 1s. It seems that
> existing TI counters can reach this goal in 10s. (using MDEV estimate
> or 100+s. using ADEV estimate). The question is; does the MDEV tool
> provide an appropriate measure of stability in this time range, or is
> the ADEV estimate a more correct answer? 

Your question is incorrect. ADEV and MDEV is two different animals of 
measurements. They are kind of similar, but MDEV can distinguish f² and 
f³ from each other. The trivial non-overlapping ADEV estimator is not 
using the samples even nearly as efficient as MDEV, so you need to use 
the overlapping ADEV estimator to compare numbers, but you must look at 
the details of how they react to different noise-types to understand 
what happens. So which gives the more correct answer depends on which 
question you are asking.

> The TI performance I'm referring to is the 20-25 ps, single shot TI,
> typical for theHP5370A/B, the SR620 or the CNT81/91. I have data
> from my CNT81showing MDEV < 1E-13 in 10s. and I believe the
> other counters behave similarly.

Care should be taken when comparing numbers, in fact those single-number 
values can be confusing as they combined a systematic error with that of 
a noise which depends among other things on the slew-rate of the input 
signal.

The CNT90/91 have different modes of measurements, and the predictor 
data is only relevant when the frequency/period measures is taken, not 
when the time-stamp data-record is being used. The HP53131/53132 has 
similar issues, but again type of data-record used is important to recall.

> I would appreciate any comments or observations on this topic.
> My motivation is to discover the simplest scheme for making 
> stability measurements at this performance level; this is NOT
> even close to the state-of-the-art, but can still be useful.

The post-processing can use various variants of estimators, which is 
relative cheap form of upgrade. The total variance and Theo variance 
will provide you very good measures for the same sequence of data. The 
filter approach provides another interesting approach. I have not 
implemented the filter-approach variants just yet, but intend to do so.

Cheers,
Magnus