[time-nuts] Training period for a Rb clock using GPS

Hal Murray hmurray at megapathdsl.net
Thu Jun 3 18:46:04 UTC 2010

> What is the relationship between the training time and the appropriate value
> of the time constant (currently set at 18 hours)? The time constant isn't
> the size of a moving average window is it?

My assumption is that the filter is a something like a simple exponential 
decay.  The error will decay (roughly) like 1/e per time-constant.  There 
could easily be scale factors due to quirks in documentation or 

For a given time constant, how long you need to wait depends on both how far 
off you start and how good you need to get.

> Ok, great. So if we can train for h hours we should set the time constant
> somewhere between h/10 and h/5. It would be safer to pick something closer
> to h/10 since when the clock powers up it might "start" in the wrong place
> so a smaller value helps the clock move quickly into the right area, but h/5
> will act as a better buffer against hanging bridges. Is my reasoning
> correct? 

I think you are close, but you also have to consider if the unit will be 
warmed up and stable quickly enough.  If the warm up time is close to the 
training time, it's not likely to work very well.

I think this type of problem is best solved by a few experiments.  Do you 
know how accurate you need it to be?  Can you get close enough?

Another consideration is temperature shift.  Are you training the unit in the 
same environment as you will be using it?

These are my opinions, not necessarily my employer's.  I hate spam.

More information about the time-nuts mailing list