[time-nuts] Notes on tight-PLL performance versus TSC 5120A

WarrenS warrensjmail-one at yahoo.com
Thu Jun 3 13:35:07 UTC 2010


Bruce posted a bunch of stuff

But
It has finally got to the real point

Bruce finally did admit that by doing high speed oversampling I would get my 
desired integration.
He just doesn't want me to do it that way, He wants we to do it in S/W

Bruce there are so may trades offs to any design that one can make, probable 
not real helpful to tell one what his should be.
I did it in hardware case that was the BEST CHOICE FOR ME, I did not have to 
do ANY  Software that way.
and the hardware was not a limiting factor.

Yes I know how to make tradeoffs in software and hardware and now that you 
have admitted on you last post, that it is all this has really has been 
about, we can move on.  You want to make yours in S/W or VtoF, no problem. 
go for it.
Software is not always the best answers. I have found that for me If I have 
to open up any software, I can add about 5 x to the project's time.
sometimes it is worth it, but for a one unit thing, when it is not needed 
not likely for me.

I'm happy with my Hardware results for now, thank you.
When I need to do something that a $1 PIC can not do in Hardware, I'll 
consider going the S/W way.

ONCE AGAIN for all the bystanders, the reason I did it in H/W by 
oversampling is because that way I DID NOT NEED ANY NEW S/W.
To me that is a bid deal.
and if anyone  missed it, reread Bruce's email, It says Oversampling works, 
and IF you want to get down to the details of which works better, then 
oversampling works better than software at normal sampling rates, because 
there is less compromise with the oversampling than with the S/W solution 
when working with low taus.
The S/W just tries to approximate integration, and Does not do as good of 
job as H/W Oversampling OR VtoF.  But FtoV causes too much compromises in 
other areas to be a good solution for me.

The real truth seems to be that Bruce has finally figured out what I was 
talking about when it comes to Oversampling and so has had to changed his 
argument mid stream on why I should not do it.   But that's OK

ws

***************

WarrenS wrote:
> Bruce posted:
>
>> It would be much easier if Warren limited his commentary to the
>> actual results and omitted the wild speculation
> OK, It works with everything that it has been tried on and gives the
> same answers as the TSC5120A, including the Osc shake test.
> Now your turn to try and find something that it does not work on.  (&
> see end  Item)
>
>
>> Phase is the integral of frequency so phase differences sampled at
>> intervals of say T are equivalent to frequencies averaged over time T
>> and sampled at the end of the sample interval. Thus sampling the time
>> average frequency every T seconds is equivalent to sampling the phase
>> difference every T seconds.
> Equivalent Information Yes, but not equal to at all.
> Please tell me you think you can now process both of those data set
> with the same algorithm and get the same results?
> Then we can all have something to laugh at.
>
Another somewhat misleading statement since the frequency averages can
be calculated from the phase differences.
The algorithm is exactly the same as integrating the frequency error is
the same as calculating the phase change accumulated over the averaging
time.

>
>
>>> Warrens implementation improves on the original NIST implementation
>>> by oversampling.
>>>
>> Actually it degrades the simplicity and accuracy of the NIST
>> implementation by replacing the integration inherent when using the
>> counter and VFC with an approximation to the required frequency
>> integral. Fortunately the accuracy can largely be recovered by using
>> the appropriate signal processing algorithms.
>>
> Think so?  Lets see how well the VFC does at 1 ms and $10.00.
> I have NO trouble doing 1 sec integration, with errors that are far
> less than theirs.
> (& see end  item)
>
Yet again you seem to miss the point, cost is not relevant to the
discussion of the correct signal processing technique.


>>>>> Why Warren omits this crucial step when all it requires is a little
>>>>> digital signal processing as all the required information is
>>>>> available
>>>>> from the sampled EFC voltage remains a mystery.
> Hay, The tester logs raw data, if you want to make a S/W filter to run
> the Raw data thru go for it,
> BUT before you waste your time, you should try and find at least one
> case that the current system does not work on.
> (& see end item)
>
Yet again you demonstrate a profound lack of understanding of higher
order numerical integration techniques.

>
>>>
>> Its not that the method cant be easily fixed so that it produces
>> accurate ADEV measures for an extremely wide range of sources with
>> divergent phase noise spectra, its the extreme reluctance to do the
>> signal processing correctly (its not that this even incurs extra
>> hardware costs) that is perplexing.
>
> See above and  Send me the easily fixed S/W.  I'm ready to try it.
> The first test is to make sure it does not brake what is already working,
> then I can try it on anything that the existing S/W does not work on.
> Oh yet, that is going to be a bit of a problem, there is no known
> device the existing software does not work on.
> You'll need to send that along with the Software.
How about hydrogen masers and cryogenic sapphire resonators?

>
> this is the END ITEM
> If you want to save a whole lot of time and not do extra S/W etc,
> I'll just make the oversample to tau0 ratio larger. That will fix any
> integration and phase noise and errors that you can come up with.
> If you do not understand why that is the case, then you still have no
> understanding of what I'm doing.
>

> Lets forget all the other BS for now, and just concentrate on the
> single statement.
> "I'll  increase the oversample to tau0 ratio, that will fix it"
That isnt always even possible or even a cost effective option.
Another problem with the tight PLL method is that the PLL bandwidth is
limited by the variable bandwidth of the EFC circuit (a few kHz for a
10811A).
Thus accurate operation down to Tau =1ms may be somewhat problematic.
> If you think it is false, you really do need to go back to
> oversampling school.
> If you admit it is true, then we really do not have much else to talk
> about, because it fixes all your present concerns.
> It is now as simple as that.  To discuss anything else is a total
> waste of time.
>
Which is your way of saying that you don't understand the alternative
more accurate methods and won't consider them.
> ws
>
> ***********************
>
Bruce
> ----- Original Message ----- From: "Bruce Griffiths"
> <bruce.griffiths at xtra.co.nz>
> To: "Discussion of precise time and frequency measurement"
> <time-nuts at febo.com>
> Sent: Thursday, June 03, 2010 12:27 AM
> Subject: Re: [time-nuts] Notes on tight-PLL performance versus TSC 5120A
>
>
>> Steve Rooke wrote:
>>> On 3 June 2010 15:46, Bruce Griffiths<bruce.griffiths at xtra.co.nz>
>>> wrote:
>>>
>>>> WarrenS wrote:
>>>>
>>>>> As Bruce says "It remains a mystery"  to him why this works.
>>>>>
>>>> It doesnt, it only appears to in a very restricted set of
>>>> circumstances.
>>>>
>>> Bruce, I don't understand you, when presented with visual evidence
>>> that this method works you still deny it.
>>>
>>>
>> What visual evidence??
>> There is no proof that the technique works well in every case.
>> Only that for the range of Tau tested and for the particular source
>> pair used that it appears to.
>>
>>>>> Not one of my best skills, but I'll try to explain it once again.
>>>>> Now that they see it works, maybe someone else will be able to put
>>>>> this
>>>>> into words that Bruce will be able to finally understand.
>>>>>
>>>>> The only requirement needed for the Frequency data log to be give
>>>>> correct
>>>>> ADEV readings, is to get good, Averaged, integrated, Frequency
>>>>> data, with no
>>>>> dead time, and no aliasing, over the tau0 time period.
>>>>> Each Tau0 Frequency sample is ideally completely independent from all
>>>>> others. If it can do one right then it can get them ALL right.
>>>>> In a single tau0 sample there is NO SUCH THING as a certain type
>>>>> of long
>>>>> term noise, Just the average freq over that single time period.
>>>>>
>>>>>
>>>>>
>>>> Misleading as usual, your knowledge of statistics is woefully
>>>> inadequate
>>>> leading to incorrect conclusions.
>>>>
>>> Well, what are are the "woefully inadequate" conclusions then? Please
>>> give us your full reasoning.
>>>
>>>
>> A simple example is that for a small number of samples a stability
>> metric like the ordinary (unfiltered) phase variance standard
>> deviation may appear to be stable, whereas with a sufficiently large
>> number of samples the instability of the metric itself becomes
>> evident whenever divergent noise processes like flicker phase noise,
>> random walk frequency noise are present.
>>
>> /"Each Tau0 Frequency sample is ideally completely independent from all
>> others."
>> /
>>
>> The above statement is incorrect as the finite bandwidth necessarily
>> imparts a correlation between samples they can only be strictly
>> independent if the bandwidth is infinite.
>>
>> /"In a single tau0 sample there is NO SUCH THING as a certain type of
>> long
>> term noise, Just the average freq over that single time period."
>> /
>>
>> The above statement imparts no useful information.
>> It would be much easier and less bandwidth wasted if the circuit
>> schematics and useful documentation on the  algorithms employed were
>> available.
>> Extracting any useful information seems somewhat akin to pulling teeth.
>>
>>>> The crucial integration/averaging to get good tau0 data, that Bruce
>>>> can not
>>>> see for some unknown reason, is done
>>>>
>>>> Only in your imagination.
>>>>
>>> One would assume that this method only works when Warren does it as
>>> his "imagination" is required for it to work, but wait, John Miles has
>>> managed to get point for point identical data against a TSC, how can
>>> that be Bruce? Please give answers, not insults.
>>>
>>>
>> Read the following paper:
>>
>> http://hal.archives-ouvertes.fr/docs/00/37/63/05/PDF/alaa_p1_v4a.pdf
>>
>> which shows the relationship between AVAR etc, filters and the
>> ordinary phase variance.
>> The paper also outlines the techniques that should be used with the
>> sampled frequency difference data from a tight PLL.
>>
>>
>>>>> with an analog filter  set to about the Tau0 Freq and by
>>>>> oversampling at
>>>>> about about a 10 to one ratio, and averaging the oversampled
>>>>> frequency
>>>>> readings down to tau0.
>>>>>
>>>> That doesn't work as it has the wrong transfer function.
>>>>
>>> Again, it it does not work, how come the evidence shows that it does,
>>> how do you explain that Bruce?
>>>
>> The evidence doesn't show this at all.
>> It merely indicates that for the devices tested that the phase noise
>> spectral components in the region where the filter responses of the
>> ADEV and WDEV differ (its not ADEV so it shouldnt be labelled as
>> such) dont appear to be significant for the 2 sources compared and
>> the tau range over which the testing was done.
>> Extrapolation of such results to predict that the technique will
>> produce such agreement with other devices with differing phase noise
>> characteristics is unrelaible.
>>
>> You are confusing producing the same numbers in specific cases with
>> the ability to do so in general.
>> There is no guarantee that such agreement will occur with a given
>> pair of sources.
>> Such agreement in general isnt possible as the equivalent phase noise
>> filters have different frequency responses.
>>
>> Stability measures like AVAR can be shown to be the equal to the
>> ordinary variance of the phase difference at the output of a very
>> specific phase noise filter.
>> WDEV has a phase noise filter with a different frequency response so
>> that it doesnt actually measure ADEV.
>>
>>>>> It is not perfect, but plenty close enough for the plot to match the
>>>>> output of the TSC 5120A over the whole tau range.
>>>>> There are a few other subtle details on how to insure that
>>>>> aliasing and
>>>>> over filtering do not become a problem, but first things first,
>>>>> one needs to understand how the integration is being done.
>>>>>
>>>> Sloppy and misleading "explanation" as usual.
>>>>
>>> You sound like a school teacher marking a pupil's work. Perhaps not
>>> everyone is as eloquent as Shakespeare with their English, there is no
>>> need to resort to this form of denigration. I find your explanations
>>> on things very cryptic and hard to follow but I don'r resort to this
>>> sort of abuse.
>>>
>>>
>> It would be much easier if Warren limited his commentary to the
>> actual results and omitted the wild speculation (and the metaphysics).
>>
>>>>> The integration secret  (which is no secret to anyone but Bruce)
>>>>> is to
>>>>> analog filter, Oversample, then average the Frequency data at a
>>>>> rate much
>>>>> faster than the tau0 data rate.
>>>>>
>>>> Which again is misleading as you specify neither the averaging
>>>> method nor
>>>> the analog filter.
>>>>
>>> Has been explained by John who wrote the method and is available for
>>> you to review.
>>>
>>>
>> I've seen it and its somewhat shy of the optimum signal processing
>> technique.
>>>>> That alone should be enough information for any knowledgeable
>>>>> designer to
>>>>> understand.
>>>>>
>>>> Its not and you should know that it isnt.
>>>> You draw conclusions that are neither supported by measurement nor
>>>> theory.
>>>>
>>> So the visual evidence before your very eyes which clearly shows that
>>> this works is not sufficient for you to understand that this
>>> measurement works.
>>>
>> Read the theory outlined in the paper and maybe you''ll begin to
>> understand my objections to statements that the technique measures ADEV.
>>>
>>>>> ws
>>>>>
>>>>> ps)
>>>>> Do note, I'm working with Frequency here and not phase, that may
>>>>> be what
>>>>> is confusing some.
>>>>>
>>>> When will you understand that phase differences and differences of
>>>> average
>>>> frequency (unit weight to frequency measures over the sampling
>>>> interval zero
>>>> weight outside) are equivalent.
>>>>
>>> Bruce, you do know that this is the NIST tight-loop PLL method which
>>> produces frequency measurements and not the loose-loop PLL method
>>> which produces phase difference data I hope.
>>>
>>>
>> Of course I am aware of that.
>>
>> Phase is the integral of frequency so phase differences sampled at
>> intervals of say T are equivalent to frequencies averaged over time T
>> and sampled at the end of the sample interval. Thus sampling the time
>> average frequency every T seconds is equivalent to sampling the phase
>> difference every T seconds.
>>
>>>
>>>>> The problem with that page is that you show the original NIST
>>>>> implementation which actually produces valid ADEV measures whereas
>>>>> Warren's implementation omits the crucial integration/averaging (his
>>>>> figurative handwaving antics don't change this) and hence actually
>>>>> has a
>>>>> different phase noise frequency response than that of the filter
>>>>> implied
>>>>> by the definition of AVAR.
>>>>>
>>> Warrens implementation improves on the original NIST implementation by
>>> oversampling.
>>>
>> Actually it degrades the simplicity and accuracy of the NIST
>> implementation by replacing the integration inherent when using the
>> counter and VFC with an approximation to the required frequency
>> integral. Fortunately the accuracy can largely be recovered by using
>> the appropriate signal processing algorithms.
>>
>>>
>>>>> Why Warren omits this crucial step when all it requires is a little
>>>>> digital signal processing as all the required information is
>>>>> available
>>>>> from the sampled EFC voltage remains a mystery.
>>>>>
>>> I'm intrigued Bruce, please explain to us in detail what you are
>>> actually getting at here?
>>>
>>>
>> Read the paper on stability variances and filters.
>>>>> The method as implemented by Warren produces a frequency stability
>>>>> metric which may be useful for comparing the stability of some
>>>>> sources,
>>>>> however it does not measure ADEV.
>>>>>
>>> The needle is stuck again, Bruce, look at the results, as rose by any
>>> other name would smell as sweet.
>>>
>>>
>> Poetry is irrelevant, the fact that the equivalent filters for AVAR
>> and WVAR differ should be of concern.
>>>>> Under a restricted set of circumstances such as when white phase
>>>>> noise
>>>>> or drift dominate the measures so calculated my be close to the
>>>>> measured
>>>>> ADEV obtained by a method wth the correct response to the various
>>>>> phase
>>>>> noise frequency components, however this doesnt mean that the
>>>>> measures
>>>>> are actually ADEV measures it merely means that the phase noise
>>>>> frequency components in the region where the frequency response of
>>>>> the 2
>>>>> methods differ significantly, are not significant.
>>>>>
>>> You keep coming up with imaginary ways that you think this method
>>> would fail to produce the right answer but you've not produced a
>>> source that can be tested in the REAL World. You talked about Warrens
>>> imagination earlier, well I'm calling this on you now. Lets have some
>>> concrete example that shows this method is not usable or shut up.
>>> Warren has put his money where his mouth is, now it's your turn.
>>>
>>>
>> Its obvious if you understand the theory.
>> Otherwise an infinite set of tests with an infinite number of sources
>> with an infinite variety of phase noise spectra are required to show
>> the technique works in all cases.
>>
>>
>>> Bruce, I really do admire your knowledge and intelligence generally
>>> but sometimes you really need to take a step back and smell the
>>> coffee.
>>>
>>>
>> Its not that the method cant be easily fixed so that it produces
>> accurate ADEV measures for an extremely wide range of sources with
>> divergent phase noise spectra, its the extreme reluctance to do the
>> signal processing correctly (its not that this even incurs extra
>> hardware costs) that is perplexing.
>>
>>> My apologies to the list for openly expressing my feelings on this.
>>>
>> Bruce
>>
>>
>> 





More information about the Time-nuts_lists.febo.com mailing list