[time-nuts] Second FTS4060 shows Drift, is it me? Good Links

Tom Van Baak tvb at leapsecond.com
Tue Feb 7 21:51:12 EST 2006


Let me clarify this a bit more.

> It's my understanding the a properly working Cs standard probably has it's
> frequency slightly off from perfect, but does NOT have any drift in
> frequency.

Agreed. Over a month there should be no obvious
frequency drift in a Cs standard. However, from minute
to minute, or even day to day, you *will* see frequency
changes. But unlike Qz or Rb, there should be no
consistent, long-term trend in frequency with Cs.

> So a Time Interval plot, like those in the links above, should be just a
> straight line, i.e. it should not have a parabolic shape.

Correct; a time interval or phase plot will be straight
(with some slope) but usually never flat (zero slope).
In the real world the line will also have plenty of bumps
due to noise. Over time (days, weeks) a thin line will

> The slope of the TI plot is the first derivative of the equation of the
> points and should have included two terms.  The second derivative gives

Right, the first derivative of phase is frequency. When
you plot Cs frequency you should see a flat line (zero
slope) but usually the flat line is not centered at zero;
i.e., there is a constant, non-zero frequency offset. If
you plot a year of data you should see a very narrow
flat line. If you plot a couple of days or weeks the noise
level may leave some doubt about how flat the line it.

> aging rate which should be zero for a Cs standard. When I recompute based

Right, the second derivative of frequency is drift. Within
the limits of noise and length of the data set, this should
be a flat line (zero slope) centered at zero (no drift).

> the Excell equation coefficients the answer is off, so I think that Excell
> is giving rough values.  But in any case the second derivative is just a
> single number whose units would be ns/day/day.

True, ns/day/day works, although in general the units
are time to the power minus 1, which are more often
scaled and displayed as "per day". You can scale
your ns/day/day by 86400/1e9 to get "per day".

I can help you with Excel - I use it here for some of
my analysis and have not found errors with it. If you
are manually computing statistics with large numbers
it helps to first normalize all values about the mean
to avoid loss of precision in floating point calculations.

Remember also that single numbers (such as the
second derivative) are not constants; they are a
measured quantity with an uncertainty that can be
calculated (as in standard deviation).

> Don't have plots of s/n 1013 vs. s/n 1227.  I just have one SR620 counter.
> Is there a common way to check the TI counter, maybe based on a coax

These sorts timing experiments work well most of the
time. Your apparent Cs drift still baffles me and I'd like
to get to the bottom of it. Often one has cross checks
(multiple counters, or multiple references, such as your
use of both GPS and Loran-C) to eliminate or at least
to isolate errors in equipment. Send me email offline
and, short of shipping the 4060's up here, I'll see what
I can do to sort it out.


More information about the time-nuts mailing list