[time-nuts] Sub Pico Second Phase logger

[Bruce Griffiths]

Joseph M Gwinn wrote:
> Bruce,
>
>
> time-nuts-bounces at febo.com wrote on 12/16/2008 09:05:55 PM:
>
>   
>> Joseph M Gwinn wrote:
>>     
>>> Bruce,
>>>
>>>
>>> time-nuts-bounces at febo.com wrote on 12/15/2008 05:31:27 PM:
>>>
>>>
>>>       
>>>> Joseph M Gwinn wrote:
>>>>
>>>>         
>>>>> Bruce,
>>>>>
>>>>>
>>>>> time-nuts-bounces at febo.com wrote on 12/15/2008 04:34:34 PM:
>>>>>
>>>>>
>>>>>           
>>> [snip]
>>>       
>>>> I'll look into doing this [MDEV and ADEV].
>>>> Real time filtering and decimation may be impractical, in the short 
>>>>         
> term
>   
>>>> at least, as most signal processing libraries only process 16 bit 
>>>> samples.
>>>>         
>>>> Most real time spectrum analysis programs are similarly afflicted in
>>>> that they only process 16 bit samples.
>>>>
>>>>         
>>> I don't see why we would need realtime filtering.  Data reduction can 
>>>       
> be 
>   
>>> offline, so we ought to be able to use 32-bit or 64-bit arithmetic.
>>>
>>> Given that we will inspect Allan Deviation data in a log-log plot, one 
>>>       
> can 
>   
>>> save much processing time by spacing the tau values to be computed 
>>> uniformly in log tau.  I've played with this in Mathematica, and it 
>>>       
> does 
>   
>>> work and yields a large speedup factor.  It should also help with 
>>>       
> Plotter 
>   
>>> and Win2K limits.  One trick is to ensure that one computes each tau 
>>>       
> value 
>   
>>> at most once.  This check is needed because with close spacing, the 
>>>       
> round 
>   
>>> function will yield the same tau values multiple times for small 
>>>       
> values of 
>   
>>> tau.
>>>
>>> Joe
>>>
>>>
>>>       
>> Joe
>>
>> Real time processing certainly isn't required to characterise the 
>>     
> performance.
>   
>> However some may be tempted to do this, it's probably possible with a 
>>     
> sufficently fast machine.
>
> If we are looking for thermal effects, with a characteristic timescale of 
> tens of minutes to hours, the concept of realtime can be very generous.
>
>
>   
>> I was just highlighting a problem with some available signal processing
>> libraries which may have been developed before sound cards with
>> resolutions of more than 16 bits became available.
>> Some (perhaps most) real time spectrum display software also has this
>> problem (eg baudline, Virtins etc).
>>     
>
> I would assume that there are newer libraries now, and libraries available 
> as source code can be updated and recompiled.
>
> 20 Log[ 2^16 ]= 96 dB.  This isn't awful, and we will get the entire 
> 16-bit range if the ADC is 24 bits (with ENOB of 19-20 bits) and we scale 
> and round the samples properly.
>
> As I think about it, the 16-bit limit must be for embedded signal 
> processing code, and math libraries intended for use on ordinary computers 
> will be at least 32 bit or 64-bit float, so it should not be difficult to 
> come by the necessary code.
>
>
>   
>> It isn't necessary to use a pair of mixers and an offset source to
>> characterise the sound card, driving both sound card inputs from the
>> same audio source should suffice.
>>     
>
> Yes.  One input at a time, with the other input shorted, so we can also 
> see the crosstalk.
>
>
>   
>> The audio source need not have low ultra low distortion (the IF output
>> signals in a dual mixer system won't have ultra low distortion) or very
>> high frequency stability (the IF output signals in a dual mixer system
>> won't necessarily have particularly high frequency stability).
>>     
>
> But ... but ... but ... I thought Time Nuts used only atomic frequency 
> refs, and crystals only if oven stabilized.
>
>   
If one mixes down a 10MHz source to 100Hz the fractional frequency
instability (of the beat frequency) is magnified by a factor of 1E5 over
that of the 10MHz source.
This assumes that the offset source has significantly lower instability
than the source under test.
In the special case when the offset source and the test source are phase
locked the offset frequency will have much greater stability.

>  
>   
>> A standard RC audio oscillator with distortion lower than 1% or so
>> should suffice.
>> At least the resultant frequency fluctuations should thoroughly exercise
>> the phase extraction algorithms.
>>
>> Another option would be to low pass filter the output of a divider.
>> Using a sound card to generate the test signal is also possible but it
>> can potentially introduce extraneous noise and other artifacts such as
>> phase truncation spurs.
>>     
>
> If one chooses the test frequencies correctly, one can eliminate the 
> spurs.  The trick is to choose frequencies that lead to DDS tuning words 
> that have zeroes in the accumulator bits that are truncated (that is, do 
> not make it into the sin/cos lookup table).
>
>
>   
This just adds another layer of complexity for little immediate gain.
Making the algorithms robust against small drifts in beat frequency is
more useful in the general case (when 2 different test sources are being
compared) than just assuming that the the beat frequency is very stable
and fixed.
> Step one of planning an experiment is to decide on the objectives.  The 
> large scale objective is to determine which sound cards are suitable for a 
> number of time-related tasks, so we should enumerate and describe these 
> tasks. 
>
> Task 1.  The immediate task is to receive and digitize the sinewave output 
> from a mixer, the sinewave being the offset frequency coming out of a DMTD 
> apparatus. Offset frequencies will range from 10 Hz to 1 KHz, will be 
> known with great precision from the design of the apparatus, and need not 
> be measured.  This sinewave is high amplitude (at least one volt rms, 
> matched to the needs of the soundcard) and very high SNR.  This will be 
> done in two channels in parallel.  The signals are at the same frequency 
> but differ in phase.  The intent is to extract the phases of these two 
> sinewaves, the difference in phase being the ultimate output.  The phase 
> of a signal will be extracted by least-squares fitting of a sine function 
> to the measured data.
>
> And so on.  We need to list the tasks, and to use this task list to inform 
> the experiment design.
>
>
>   
The immediate task is actually to evaluate sound cards for their
suitability, preferably without the added cost and complexity of a DDS
LO and mixer.
Once this evaluation is done, using a mixer and a DDS based LO to
generate a beat frequency is the next step.
Eliminating the mixer and DDS allows a greater number of participants at
this stage than would otherwise be the case.

10Hz resolution whilst avoiding phase truncation spurs is impractical
with a DDS chip by itself.
Depending on the DDS and its clock frequency, the frequency spacing of
phase truncation spur free outputs may be as large as several kHz.
A few divide and mix stages will be required to achieve a spur free
resolution of 10Hz.
A DDS chip with higher resolution phase outputs after truncation such as
the AD99XX series are better in this respect than the earlier AD98XX series.

To broaden participation we need to broaden the scope of the project to
include dual mixer system with statistically independent test sources as
well as the more specialised case where the 2 input frequencies differ
only in phase.

1) Evaluate sound cards for suitablility.
Initially use simple less stable sources and follow up with more stable
test sources for the more promising cards.
Need to measure crosstalk, temporal instability of interchannel phase
shift, system noise etc.

2) Develop robust algorithms for phase extraction.
Use the data produced by the less stable sources and that produced by
the more stable sources

3) Repeat testing using a dual mixer system complete with offset LO.
Test frequencies identical to evaluate system noise floor.

4) Repeat testing using a dual mixer system complete with offset LO.
Test frequencies differ to help the effect of residual crosstalk and
other artifacts.

5) Split the project into 2 branches:
A) where mixer inputs differ only by a phase shift to be measured.
Useful for measuring effect on ADEV of various components and their
phase shift tempcos etc.

B) Where the mixer input test sources are statistically independent.
Useful for measuring pairwise source ADEV etc.

Bruce