from the [time-nuts] mailing list:

Magnus Danielson magnus@rubidium.dyndns.org via febo.com
4:46 PM (3 hours ago)
to Discussion

Fellow time-nuts,

In a private discussion, I got somewhat inspired, so I wrote this relating to what jitter is. It's the pre-breakfast, won't get out of bed version. Since it was enjoyed by a fellow time-nut, I share it with a little larger audience.

Jitter as such is the part of the phase deviations being “fast” while wander is part of the phase deviations being “slow”. Jitter and Wander is measured in seconds, degrees or Unit Interval depending on the application. Jitter is reported in RMS or peak-to-peak values depending on where it is being used, but since jitter has a underlying Gaussian distribution, the longer you measure the higher peak-to-peak value gets and you don't get very smarter. For gaussian distribition the RMS value is much better. However… jitter can have other additive components than gaussian jitter, so over the last 10-15 years or so it has become increasingly popular to use software tools to separate Random Jitter (with Gaussian distribtion) from Deterministic Jitter (which has stable peak-to-peak values, such as an added sine-modulation). Then this can be broken down further to include intersymbolic interference, crosstalk etc. This is the market that Wavecrest was pursuing with their counters, a market which has the needs. Today Tektronix, Agilent and LeCroy have this processing as a software option into their scopes.

Jitter for a signal is measured with a PLL clock recovery, and the output of the mixer is tapped and then filtering is done to measure different standard values. Then a RMS value of the aggregate is reported. By using this method, a cheap compliance testing can be done, and the filter frequencies is standardised. They also relate to the sinus jitter tolerance curves, which also relate to the MTIE tolerance curves. These tolerance curves work in two ways, the output tolerance must be below the curve while the input tolerance must be above, when they are the output will always be tolerated by the input.

For telecom use, aggregate numbers for jitter and wander is reported in UI after filtering. A UI is a scaled unit of time, where 1 UI is the time of the shortest symboltime (or symbol time quanta). For simple compliance getting the aggregate readings suffice, but for detailed analysis the data is processed according to TDEV and MTIE numbers. TDEV addresses the random properties of the signal noise, while MTIE covers the systematic properties of the signal noise. The neat thing is that MTIE numbers in their UI scale provides very good hints about jitter compensation buffer size. It correlate nicely. Also, the corner frequencies correlate well with the bandwidth of the jitter damping PLL being in parallel with the jitter compensation buffer. With the plots in the standards you can read out all the key parameter for your design!

The separation between jitter and wander now becomes a little easier to explain. They relate to different sources, so in telecom the separation is being said to occur at 10 Hz, but to be honest, this is a separation only meaning full in the context of PDH, SONET and SDH systems. It's not meaningful in the context of other signals or systems, not by any form of automagic anyway. Looking at the aggregate signals and how they best are measured and characterized the separation makes sense. Jitter uses the sinusoidal tolerance curves while wander used TDEV and MTIE.

I like to compare jitter and wander to wow and flutter. Wow is the low frequency modulation that typically occur when the record slips a little in angle compared to the record player disc. This causes a low frequency modulation causing a “Wooow” sound on voices. This is clearly the wander of the record player industry. Flutter comes as a result of the drive mechanism, steps in the motors, gears and lack of damping through rubber band and heavy turntable disc.

Actually jitter (shaking) and wander (walking around) is just colloquial terms just was wow and flutter, but they all relate to particular phase modulation aspects.

So the question is, how does one relate ADEV or TDEV or phase or phase
noise to a 1PPS jitter number? Or, is it safe to say for T&F use, that
TDEV at tau 1 s is a more useful measure of 1PPS “jitter” anyway?

Now, if your jitter (and wander) is dominated by gaussian jitter, then yes, your ADEV and TDEV correlate at 1 s to the jitter (do notice the sqrt(3) between these measures). The reason is that ADEV as such is scaled to match the white noise readings.

If your jitter has sufficiently strong components of non-gaussian character, you will need to measure them separately and join them after proper scaling.

Consider that you have a bit error rate (BER) requirement of 10-12. This is very common number, so it provides the basis of the rule of thumb. You get a bit error when you sample the wrong bit value, so in a simplified model you would get a bit error when you sample 1 UI away from your average (it's actually half, but it is compensated by the fact that previous or next bit may be half value) and then the deterministic components peak-to-peak value and the gaussian peak-to-peak value may not together be beyond one. As I said, the gaussian distribution does not have a stable peak-to-peak, but it has a peak-to-peak matching a statistic area under the curve, and using that we can use the RMS to peak-to-peak scale of 14 (as rule of thumb number) to scale the RMS value to a peak-to-peak. This is why a typical jitter requirement is 0,07 UI. That reads out as BER of 10-12 for gaussian jitter sources.

Thus combining them require knowledge about your application.

Anyway, all this is the things we do in the telecom side of time measurements, but that is now creeping into high speed digital design, which has the same basic properties now.

So jitter can be many things and many values, all depending on what you do.

Cheers, Magnus