[time-nuts] timelab

[Magnus Danielson]

On 07/05/2012 07:14 AM, Don Latham wrote:
> So if the 5370 determines the data rate, then in calculating adev or
> other such products, we're assuming the time differences are ergodic
> because we're not getting them sequentially, but rather selecting the
> delay intervals at long inetrvals between them?

It's trying to determine the tau0. It's actually determining the T 
factor, the time between measurement results. The difference T-tau0 will 
be your deadtime and it will bias your measurements. This bias function 
was first shown by Dr. Allan in his Feb 1966 paper. Since we have 
modernized the analysis somewhat. Please see

http://en.wikipedia.org/wiki/Allan_variance#Bias_functions
and in particular the B2 bias function.

There is two ways to approach it for propper values:

1) Set up the measurement for no (apparent) deadtime. I use a PPS (or 
other rate) trigger out of something like the TADD-2 either for start or 
for measurement arming.

2) Compensate for it using the bias function.

If you ever "skip over" measurements, toss the measurement result or at 
least stop. The series ends when a "skip" occurs and a new series 
begins. You can combine the results, but it is tricky, error prone and 
you loose data on long taus which is what is hard to get much of anyway.

Hence, just averaging the sample arrival time without critically analyze 
the pauses can lead you into incorrect values as things will be biased 
in all kinds of ways. Inferring the time-base on arrival time should be 
avoided whenever possible.

Cheers,
Magnus