[time-nuts] Re: Is this the right way to compare the short term accuracy of two frequency counters?

[Tom Van Baak]

Erik,

See some plots of testing a 53132A counter in time interval mode:

http://leapsecond.com/pages/53132/

Yes, using stdev (rms) is ok. You can also use ADEV and expect to see a 
line with slope of -1; the noise of the counter is the value at tau 0 
[*]. You can also use TDEV and it should be a relatively flat line; this 
also reflects the noise floor of the counter. If your stdev varies a lot 
depending on sample size, or if your ADEV plot isn't a very straight 
line with slope exactly -1, there's a problem.

To determine the noise floor of a counter it's best to use different 
sources, independent in frequency and also independent in phase. That 
rules out dividers or multipliers. The phase needs to "sweep" across a 
wide range. What you want to avoid are "sweet spots" where there is some 
constant relationship between the counter's internal reference and the 
external input(s) that are being measured. A sweep will help identify 
any subtle frequency or phase pulling effects. You can see this 
dramatically in the 53132A plots.

So in frequency mode make sure it's two different oscillators. In time 
interval mode make sure it's 3 independent sources; all drifting in 
phase wrt to each other. You're welcome to use only one source too but 
that will likely give an overly optimistic report of counter resolution.

An analogy is if you built a 5 digit 20 volt DIY voltmeter and only 
tested it at 0.0000 volts, or maybe only at 12.345 volts. Sure, you can 
take hundreds of readings and computer the stdev. But a better way is to 
generate a precise triangle wave from 0 to 20 volts in steps of 20 uV 
and then look at the readings to see how accurate they are across the 
entire range.

Most frequency counters have artifacts that cause some level of 
non-linearity depending on where the input(s) land wrt to the reference. 
It could be minor, but it's something that contributes to the accuracy 
and noise floor, especially with interpolating or "enhanced resolution" 
counter. It's ok. You just want to know how good or bad it is. It can 
even be a diagnostic to find PCB layout or f/w bugs.

The 10 MHz effect that Bob mentions is:

http://leapsecond.com/pages/53132/53132-reduced-resolution.gif

/tvb


On 1/24/2022 6:47 AM, Bob kb8tq wrote:
> Hi
>
> First off, yes, standard deviation is a pretty good way to look at what a counter
> is doing. Reducing the answer to time ( = picoseconds ) is usually the easy way
> to look at the data.
>
> Next up, counters have a *lot* of things that impact what they do. The slew rate
> of the input signal is a big one on most counters ( = square waves and sine
> waves will give you different answers).  It is best to do your testing with the
> type of signal you are most likely to use.
>
> Some counters do very odd things when the input and reference are tightly coupled.
> An SR620 counting it’s reference output is one good example of this. Best practice
> is to run two independent sources. One supplies the reference, a second one
> generates the test signal.
>
> HP counters (and likely some others) have interesting “dead spots” at 10 MHz, at
> 10MHz / N and 10 MHz * N. to get the best performance numbers, test with a signal
> that is not in one of these regions.
>
> It is worth looking at the ADEV / phase noise of your test signal. You can indeed
> get into trouble there…..
>
> Fun !!!
>
> Bob
>
>> On Jan 24, 2022, at 8:43 AM, Erik Kaashoek <erik at kaashoek.com> wrote:
>>
>> For some project I'm trying to establish the short term accuracy of a
>> frequency counter versus the gate time.
>> As using the Allan Deviation for this type of measurement did lead to
>> extensive discussion over the validity of using ADEV for measuring the
>> short term performance of a counter I tried to find a different, but still
>> relevant way to establish the performance.
>> To exclude as much as possible external and long term factors I'm using a
>> single fairly stable OCXO (short term error below 1e-10) to output 10 MHz.
>> This 10 MHz goes into an SI5351 as reference for its PLL and the SI5351
>> outputs two frequencies from the same VCO, one at 10 MHz into input A of
>> the counter and one at 10.00003319 MHz into input B of the counter. The
>> counter is setup to measure the ratio of A/B and to display the STDDEV of
>> the ratio over n=100. The STDEV of counter B is calculated as the square
>> root of  ( (the sum of the squares of the difference between the measured
>> ratio and the average ratio ) divided by the number of measurements )
>> I'm aware the SI5351 uses a fractional divider but I hope the impact is
>> below the measurement accuracy required.
>> Doing this test with two counter gave these results:
>>
>> Counter A
>> Gate time :  STDDEV
>> 1 s : 1.0-10
>> 0.1 s : 1.0e-9
>> 0.02 s : 6.5e-9
>>
>> Counter B
>> Gate time :  STDDEV
>> 1 s : 1.3--9
>> 0.1 s : 1.5e-8
>> 0.02 s : 1.4e-7
>>
>> The results have been verified by performing multiple measurements. Counter
>> A and B are both fractional counters that use interpolation.
>>
>> The manual of the Agilent 53132A specifies the worst case RMS  error of a
>> frequency measurement for different gate times and an input frequency of
>> 10MHz as:
>>
>> Agilent 53132A
>> Gate time : Max RMS error (estimated)
>> 1 s :  2e-10
>> 0.1 s : 2e-9
>> 0.02 s :  5e-8
>>
>> Assuming the RMS error and the STDDEV are the same the steps with gate time
>> change of the Agilent and Counter A seem to be comparable but Counter B
>> behaves a bit different for 0.02 s gate time.
>>
>> This leads me to the following questions:
>> Is measuring the STDDEV of the ratio of two input frequencies derived from
>> the same timebase a valid way to assess the short term measurement accuracy
>> of a frequency counter?
>> If not, how should this be done?
>> If yes, do the numbers I'v listed above make sense?
>> _______________________________________________
>> time-nuts mailing list -- time-nuts at lists.febo.com -- To unsubscribe send an email to time-nuts-leave at lists.febo.com
>> To unsubscribe, go to and follow the instructions there.
> _______________________________________________
> time-nuts mailing list -- time-nuts at lists.febo.com -- To unsubscribe send an email to time-nuts-leave at lists.febo.com
> To unsubscribe, go to and follow the instructions there.