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

[Magnus Danielson]

On 05/06/2012 08:47 AM, Poul-Henning Kamp wrote:
> In message<20120506021212.EC21A800073 at ip-64-139-1-69.sjc.megapath.net>, Hal Murray writes:
>>
>> phk at phk.freebsd.dk said:
>>> If you cannot apply the negative sawtooth, you will get better results by
>>> disciplining almost any random quartz xtal, ovenized or not to the GPS,
>>> divide it down to PPS and then discipline the PRS10 to that.
>>
>> I don't understand that.  What am I missing?
>
> You are missing that the average of the 1PPS pulse only can be trusted
> to be zero over a timescale of many hours.
>
> This is an error-source distinct from GPS reception, caused by the
> picking a preexisting flank nearest to the epoch, with no attempt
> to keep the average of the resulting error zero.
>
> Imagine that the GPS receivers clock happens to run on perfect
> frequency for a while:  That means that the flank used to generate
> the PPS will have a fixed location relative to the epoch, for instance
> always 12 nanoseconds early or late.
>
> I belive that some GPS receivers have deliberately de-tuned Xtals
> for this very reason, but unfortunately that is only a partial
> fix, as the problem is a modulus-issue, so not only is perfect
> frequency bad, but "perfect +/- n Hz" is equally bad.

Since GPSes typically have a TCXO, it is not as stable in relationship 
to temperature as you would wish. Whenever the TCXO frequency is near an 
integer multiple of the PPS, then it will select the same number of 
clock cycles over a long run, and then averaging effects of alternating 
between two nearby clock-cycles are gone.

It would be possible to do error accumulation in the GPS that would 
steer the PPS to on average be more accurate. It should not be that many 
lines of code to achieve it.

Oh, if we only had hooks so we could insert some code into the GPS 
receivers...

>
> The hanging bridges Tom has plots of on leapsecond.com, arises when
> the frequency of the GPS xtal changes.
>
> At one point I tried putting a 1W resister close to the xtal and
> feed it with a very slow sine-wave to see if "jittering" it would
> get me an average of zero of shorter timespans.  My experiment
> was inconclusive, but the idea is not unsound.

Indeed. But you would not be "jittering" it, you would "wander" it. It 
would be a good experiment. Another approach would be steer the XTAL 
using the statistics and then control the heater.

Locking it up to .5 Hz (i.e. 50% of the transitions is short and the 
others long) would be possible, but I don't think you gain much in 
lowering the quantization floor, so I rather think a slow triangle 
modulation (that you also remove from the TI measurements) is preferred.

Cheers,
Magnus