[ntp:questions] NTP phase lock loop inputs and outputs?

Unruh unruh-spam at physics.ubc.ca
Sun Jun 1 18:58:14 UTC 2008

"David L. Mills" <mills at udel.edu> writes:


>Judah and I have come around the issue of white/flicker frequency noise 

What does this sentence mean? Does it mean you have come around to the same
viewpoint, or that you are still arguing about it. 

>with cold rocks (grotty computer oscillators). Determining accurate 

Exactly why you would call them "cold rocks" and then use them as though
they meant something to anyone else I do not know. 

>Allan deviation plots is something of a black art and at hazard to 
>insiduous resonances. I found with careful technique, as documented in 
>previous papers and my book, that phase noise and frequency noise 
>characteristics of cold rocks are generally separable, mainly because 
>the flicker noise is so bad. Therefore, the intersection of the phase 

I assume by flicker noise you mean 1/f noise.

>characteristic and frequency characteristic is generally well defined, 
>which I call the Allan intercept. This is why the NTP discipline is 
>mainly responsive to phase noise below the intercept and frequency noise 

Except that the Allan intercept depends on a whole bunch of things,
including the noise in the measurement process ( network delays and noise
for example). There is no single Allan intercept. It depends on the system.
The Allan intercept can be minutes or days depending even on exactly the
same computer.

Also as Levine says and my measuements confirm, on the timescale of a few hours, it is the daily temp
variations, not flicker noise or any other semi Markovian process (ie
uncorrelated noise in the frequency domain) that dominates. Computers tend
to get used during the day, and thus heat up during the day, and not at
night. One would like an adaptive process which tries to measure the
current Allan or equivalent (ie is the frequency noise due to the white
measurement noise or is it more correlated) rather than relying on some
"average" intercept. 

I would note that chrony has an algorithm to try to do just that-- by
estimating whetehr or not a least squares linear fit to the phase errors is
a reasonable statistical estimate or whether the measurements indicate that
this model is failing. Ie, it tries to dynamically adapt to the actual
noise type.



>Bruce wrote:

>> The only thing that you can conclude from the referenced paper is that 
>> for very short measurement intervals, order of a second, the noise 
>> process is white phase.  Obviously a simple average is optimal to 
>> determine the time from a group of such closely spaced measurements.
>> The paper then describes an algorithm to adapt to the decidedly 
>> non-stationary measurement noise and local oscillator noise processes as 
>> the averaging time is increased.  Flicker and random walk of frequency 
>> are mentioned, as well as white frequency.  However he attributes this 
>> to the crystal oscillator, which I believe is a mistake.  Crystal 
>> oscillators do not exhibit white frequency noise over longer averaging 
>> times.  If they did, we probably would not need Rubidium oscillators.
>> All of these multiple noise processes operating simultaneously and 
>> completely uncorrelated with each other are what make the precision time 
>> transfer and frequency control art so interesting and allow so many 
>> different ideas for how to best set the clock for the particular 
>> application.
>> Bruce
>> Unruh wrote:
>>> Bruce <bmpenrod at comcast.net> writes:
>>>> My understanding is that least squares is optimal only when the 
>>>> residuals are white.  For measurements of atomic frequency standards 
>>> Yes, and I clearly made that assumption. And the assumption is generally
>>> true for short enough time intervals. If the netwrok delays are really
>>> really short or if the time over which you are determining the 
>>> frequency is
>>> very long, then that is a bad assumption.
>>>> such as Rubidium or Cesium, the noise process is dominated by white 
>>>> frequency noise, and in this case linear regression yields the 
>>>> optimal estimate of frequency.  A little snooping around on the NIST 
>>>> web site will provide the relevant backup info.
>>>> For so many other precision timing and frequency applications, the 
>>>> noise processes are decidedly un-white.  David Allan developed his 
>>>> famous two-sample variance to handle these divergent, non-stationary 
>>>> noise processes.  For instance, quartz oscillators are dominated by 
>>>> flicker frequency noise for averaging times greater than about 100 
>>>> milliseconds, and eventually turn to random walk at longer averaging 
>>>> times.  Selective Availability of GPS (when it was in effect) was a 
>>>> white phase noise process that modulated the time transfer for 
>>>> un-keyed users.  The statistics of network time transfer via ntp are 
>>>> undoubtedly divergent, but I have not seen any data that showed it to 
>>>> be white frequency noise dominant.
>>> All the data suggests that it is white and the explicit assumption of 
>>> that
>>> Levine paper is that it is white.
>>>> So, it is not clear that linear regression is optimal for estimating 
>>>> the frequency via ntp, unless someone has determined the statistics 
>>>> to be white frequency.  I personally have not performed the 
>>>> measurements to make that determination, but it would not surprise me 
>>>> if Judah Levine has.
>>> And he explicitly assumes that the network delay noise is white. His 
>>> whole
>>> procedure is to make a large ( eg 25-50) of ntp type measurements at one
>>> time (withing seconds) then wait a long time ( 1/4 of a day) and doing it
>>> again. He estimates teh frequency by averaging the measurements at any 
>>> one
>>> time, and then using that average phase error to determine the frequency.
>>> The number of measurements at one instant is determined by requiring that
>>> the frequency error due to the white noise ( decreases as sqrt(n)) is 
>>> equal
>>> to the other errors (flicker noise, etc) or equals the pre determined 
>>> error
>>> level wanted.
>>>> Bruce
>>>> Unruh wrote:
>>>>> "David L. Mills" <mills at udel.edu> writes:
>>>>>> Bill,
>>>>>> Read it again. Judah takes multiple samples to reduce the phase 
>>>>>> noise, not to improve the frequency estimation.
>>>>> Dave: The frequency estimate is done by subtracting two phase
>>>>> determinations. Thus the phase noise enters the frequency 
>>>>> determination. By
>>>>> reducing the phase noise you reduce the frequency noise as well. I 
>>>>> think
>>>>> you need to read it again, but us just telling the other to read 
>>>>> properly
>>>>> will not help.
>>>>> The frequency estimate is obtained in NTP and in his procedure by 
>>>>> making
>>>>> phase measurements. f_i= (y_i-y_{i-1})/T
>>>>> If y_i=z_i+e_i where z_i is the "true" time and e_i is a gaussian 
>>>>> random
>>>>> variable, then delta f_i= sqrt( <e_i^2>+<e_{i-1}^2>)/T
>>>>> By reducing <e_i^2> you reduce delta f_i. And as you point out, you can
>>>>> reduce <e_i^2> by making a bunch of measurements. Those measurements 
>>>>> can be
>>>>> all done at the end points or spread over the time interval T. The 
>>>>> latter
>>>>> is not quite as effective in reducing delta f_i since many of the
>>>>> measurements do not have as long a "lever arm" as if they were all 
>>>>> at the
>>>>> endpoints, and that is why uniform sampling is about sqrt(3) worse than
>>>>> clustering at the end points. But in either case, the more 
>>>>> measurements you
>>>>> make the more you reduce the uncertainty in the frequency estimate.
>>>>> Anyway, at this point everyone else has enough information to make 
>>>>> up their
>>>>> own mind.
>>>>>> Dave
>>>>>> Unruh wrote:
>>>>>>> You must have read a different paper than that one. I found it 
>>>>>>> (through our
>>>>>>> library) and it says that if you have n measurements in a time 
>>>>>>> period T,
>>>>>>> the best strategy is to take n/2 measurements at the beginning of 
>>>>>>> the time
>>>>>>> and n/2 at the end to minimize the effect of the white noise phase 
>>>>>>> error on the
>>>>>>> frequency estimate. That is perfectly true, and gives an error 
>>>>>>> which goes
>>>>>>> as sqrt(4/n)delta/T rather than sqrt(12/n)(delta/T) for equally 
>>>>>>> spaced
>>>>>>> measurements (assuming large n) T is the total time interval and 
>>>>>>> delta is the std dev of each phase measurement . But it certainly 
>>>>>>> does NOT say that if you have n
>>>>>>> measurements, just use the first and last one to estimate the slope.
>>>>>>> If you have n measurements, the best estimate of the slope is to 
>>>>>>> do a least
>>>>>>> squares fit. If they are equally spaced, the center third do not 
>>>>>>> help much
>>>>>>> (nor do they hinder), but a least squares fit is always the best 
>>>>>>> thing to
>>>>>>> do.
>>>>>>> "David L. Mills" <mills at udel.edu> writes:
>>>>>>>> Bill,
>>>>>>>> NIST doesn't agree with you. Only the first and last are truly 
>>>>>>>> significant. Reference: Levine, J. Time synchronization over the 
>>>>>>>> Internet using an adaptive frequency locked loop. IEEE Trans. 
>>>>>>>> UFFC, 46(4), 888-896, 1999.
>>>>>>>> Dave
>>>>>>>> Unruh wrote:
>>>>>>>>> "David L. Mills" <mills at udel.edu> writes:
>>>>>>>>>> Bill,
>>>>>>>>>> Ahem. The first point I made was that least-squares doesn't 
>>>>>>>>>> help the frequency estimate. The next point you made is that 
>>>>>>>>>> least-squares improves the phase estimate. The last point you 
>>>>>>>>>> made is that phase noise 
>>>>>>>>> No. The point I tried to make was the least squares improved the 
>>>>>>>>> FREQUENCY estimate by sqrt(n/6) for large n, where n is the 
>>>>>>>>> number of points (assumed
>>>>>>>>> equally spaced) at which the phase is measured. I am sorry that 
>>>>>>>>> the way I
>>>>>>>>> phrased it could have been misunderstood.
>>>>>>>>> The phase is ALSO improved proportional to sqrt(n)
>>>>>>>>> . This assumes uncorrelated phase errors dominate the error budget.
>>>>>>>>>> is not important. Our points have been made and further 
>>>>>>>>>> discussion would be boring.
>>>>>>>>> Except you misunderstood my point. It may still be boring to you.
>>>>>>>>>> Dave
>>>>>>>>>> Unruh wrote:
>>>>>>>>>>> "David L. Mills" <mills at udel.edu> writes:
>>>>>>>>>>>> Bill,
>>>>>>>>>>>> If you need only the frequency, least-squares doesn't help a 
>>>>>>>>>>>> lot; all you need are the first and last points during the 
>>>>>>>>>>>> measurement interval. 
>>>>>>>>>>> Well, no. If you have random phase noise, a least squares fit 
>>>>>>>>>>> will improve
>>>>>>>>>>> the above estimate by roughly sqrt(n/4) where n is the number 
>>>>>>>>>>> of points.
>>>>>>>>>>> That can be significant. It is certainly true that the end 
>>>>>>>>>>> points have the
>>>>>>>>>>> most weight ( which is why the factor of 1/4). Ie, if you have 
>>>>>>>>>>> 64 points,
>>>>>>>>>>> you are better by about a factor of 4 which is not insignificant.
>>>>>>>>>>>> The NIST LOCKCLOCK and nptd FLL disciplines compute the 
>>>>>>>>>>>> frequency directly and exponentially average successive 
>>>>>>>>>>>> intervals. The NTP discipline is in fact a hybrid PLL/FLL 
>>>>>>>>>>>> where the PLL dominates below the Allan intercept and FLL 
>>>>>>>>>>>> above it and also when started without a frequency file. The 
>>>>>>>>>>>> trick is to separate the phase component from the frequency 
>>>>>>>>>>>> component, which requires some delicate computations. This 
>>>>>>>>>>>> allows the frequency to be accurately computed as above, yet 
>>>>>>>>>>>> allows a phase correction during the measurement interval.
>>>>>>>>>>> He of course is not interested in phase corrections.
>>>>>>>>>>>> Dave
>>>>>>>>>>>> Unruh wrote:
>>>>>>>>>>>>> David Woolley <david at ex.djwhome.demon.co.uk.invalid> writes:
>>>>>>>>>>>>>> Unruh wrote:
>>>>>>>>>>>>>>> I do not understand this. You seem to be measuring the 
>>>>>>>>>>>>>>> offsets, not the
>>>>>>>>>>>>>>> frequencies. The offset is irrelevant. What you want to do 
>>>>>>>>>>>>>>> is to measure
>>>>>>>>>>>>>> Measuring phase error to control frequency is pretty much 
>>>>>>>>>>>>>> THE standard way of doing it in modern electronics.  It's 
>>>>>>>>>>>>>> called a phase locked loop 
>>>>>>>>>>>>> Sure. In the case of ntp you want to have zero phase error. 
>>>>>>>>>>>>> ntp reduces the
>>>>>>>>>>>>> phase error slowly by changing the frequency. This has the 
>>>>>>>>>>>>> advantage that
>>>>>>>>>>>>> the frequency error also gets reduced (slowly). He wants to 
>>>>>>>>>>>>> reduce the
>>>>>>>>>>>>> frequency error only. He does not give a damn about the 
>>>>>>>>>>>>> phase error
>>>>>>>>>>>>> apparently. Thus you do NOT want to reduce the frequecy 
>>>>>>>>>>>>> error by attacking
>>>>>>>>>>>>> the phase error. That is a slow way of doing it. You want to 
>>>>>>>>>>>>> estimate the
>>>>>>>>>>>>> frequency error directly. Now in his case he is doing so by 
>>>>>>>>>>>>> measuring the
>>>>>>>>>>>>> phase, so you need at least two phase measurements to 
>>>>>>>>>>>>> estimate the
>>>>>>>>>>>>> frequency error. But you do NOT want to reduce the frequency 
>>>>>>>>>>>>> error by
>>>>>>>>>>>>> reducing the phase error-- far too slow.
>>>>>>>>>>>>> One way of reducing the frequency error is to use the ntp 
>>>>>>>>>>>>> procedure but
>>>>>>>>>>>>> applied to the frequency. But you must feed in an estimate 
>>>>>>>>>>>>> of the frequecy
>>>>>>>>>>>>> error. Anothr way is the chrony technique. -- collect phase 
>>>>>>>>>>>>> points, do a
>>>>>>>>>>>>> least squares fit to find the frequency, and then use that 
>>>>>>>>>>>>> information to
>>>>>>>>>>>>> drive the frequecy to zero. To reuse past data, also correct 
>>>>>>>>>>>>> the prior
>>>>>>>>>>>>> phase measurements by the change in frequency.
>>>>>>>>>>>>> (t_{i-j}-=(t_{i}-t_{i-j}) df
>>>>>>>>>>>>>> (PLL) and it is getting difficult to find any piece of 
>>>>>>>>>>>>>> electrnics that doesn't include one these days.  E.g. the 
>>>>>>>>>>>>>> typical digitally tuned radio 
>>>>>>>>>>>>> A PLL is a dirt simply thing to impliment electronically. A 
>>>>>>>>>>>>> few resistors
>>>>>>>>>>>>> and capacitors. It however is a very simply Markovian 
>>>>>>>>>>>>> process. There is far
>>>>>>>>>>>>> more information in the data than that, and digititally it 
>>>>>>>>>>>>> is easy to
>>>>>>>>>>>>> impliment far more complex feedback loops than that.
>>>>>>>>>>>>>> or TV has a crystal oscillator, which is divided down to 
>>>>>>>>>>>>>> the channel spacing or a sub-multiple, and a configurable 
>>>>>>>>>>>>>> divider on the local oscillator divides that down to the 
>>>>>>>>>>>>>> same frequency.  The resulting two signals are then phase 
>>>>>>>>>>>>>> locked, by measuring the phase error on each cycle, low 
>>>>>>>>>>>>>> pass filtering it, and using it to control the local 
>>>>>>>>>>>>>> oscillator frequency, resulting in their matching in 
>>>>>>>>>>>>>> frequency, and having some constant phase error.
>>>>>>>>>>>>>>> the offset twice, and ask if the difference is constant or 
>>>>>>>>>>>>>>> not. Ie, th
>>>>>>>>>>>>>>> eoffset does not correspond to being off by 5Hz. 
>>>>>>>>>>>>>> ntpd only uses this method on a cold start, to get the 
>>>>>>>>>>>>>> initial coarse calibration.  Typical electronic 
>>>>>>>>>>>>>> implementations don't use it at all, but either do a 
>>>>>>>>>>>>>> frequency sweep or simply open up the low pass filter, to 
>>>>>>>>>>>>>> get initial lock.
>>>>>>>>>>>>> And? You are claiming that that is efficient or easy? I 
>>>>>>>>>>>>> would claim the
>>>>>>>>>>>>> latter. And his requirements are NOT ntp's requirements. He 
>>>>>>>>>>>>> does not care
>>>>>>>>>>>>> about the phase errors. He is onlyconcerned about the 
>>>>>>>>>>>>> frequency errors.
>>>>>>>>>>>>> driving the frequency errors to zero by driving the phase 
>>>>>>>>>>>>> errors to zero is
>>>>>>>>>>>>> not a very efficient technique-- unless of course you want 
>>>>>>>>>>>>> the phase errors
>>>>>>>>>>>>> to be zero( as ntp does, and he does not).

More information about the questions mailing list