[time-nuts] Interesting paper: Don't GPSD' your Rb...

[Magnus Danielson]

On 05/06/2012 09:01 PM, Bill Hawkins wrote:
> Here is a different tactic for disciplining Rb from GPS/TXCO -
>
> Consider the relatively (relative to a second) long stability of
> an Rb oscillator and the not-so-good stability of GPS. Perhaps
> using 1 PPS for a sampling period for stabilizing Rb is way too
> short. Maybe 1000 seconds is better. That's way too long for an
> analog integrator to do, so a microcomputer is required.

Using analog integrator for 1 PPS stuff is hardish and your performance 
will most probably suffer. You want your integrator to be digital as it 
is clearly a better memory over time.

However, this is not to be confused with the time-constant, which should 
be much higher than 1000 s to use the stability of the rubidium where 
the GPS noise is worse. See PRS-10 manual for good illustration.

> Count the Rb and GPS 1 PPS signals for 1000 counts of Rb 1 PPS.
> You'll need to interpolate between 1 PPS GPS intervals to the
> level of accuracy required, so maybe we count cycles of 10 MHz
> from the GPS, using as many registers (cascaded integer counters)
> as required for 1 E10 (or more) counts (2540BE400 hex). At the
> end of 1000 seconds, use the difference between the lowest counter
> reading and 0xEB400 (times an appropriate gain) to tweak the value
> for the DAC doing the fine correction to the control voltage or
> current. Use the upper counters for a sanity check on the reading.

Why wait with the updates of the DAC? By incrementally average for each 
second I think you get a smoother transitions.

You sure want time-constants in excess of 1000 s, but you can achieve 
that by using 1 s updates. Again, read the PRS-10 manual for a fairly 
good description.

> As may be evident, I have never tried to discipline an Rb, but I
> am well aware of the effect of sample time on the control of a
> long time constant loop. True, the effect is strongest when dead
> time dominates the time constant, but that is not the case here.
> Still, I think there is value in using a long sampling time for
> the control action.

I fail to see the benefit, and I have many times learned the hard way to 
see the downside of too low update rate.

> Comments accepted with enthusiasm.

:)

Cheers,
Magnus