TCXO (shopping for replacement)

Cell phone interference, you name it.

Cheers

We make time-domain instruments. My customers, and my competition, specify jitter, not phase noise, of pulse outputs. The hardest one to get right is jitter between an external, asynchronous trigger and a delayed output. I wouldn't know how to measure that in the frequency domain.

The phase noise of our XO time base is just one of many things that can contribute to ultimate delay jitter. It only dominates at long delays, milliseconds and longer.

There are several phase-noise-to-jitter calculators around, including PhaseNoise102.exe, authored by one of our own. I think you are saying that they only work if the phase noise sources are uncorrelated; in other words, they don't really work.

I couldn't say in general without more research, but any procedure relying solely on the 1-D power spectrum is doomed.

Cheers

Phil Hobbs

There's a good reason for that: it doesn't exist in the frequency domain, u nless you have complete phase vs frequency information. A logic signal's sh arp edges involve correlations between many frequencies. It's not possible to estimate the jitter from various phase noise contributions without knowi ng the slope of the logic transitions; a given voltage perturbation causes a timing error proportional to the ratio of its slope to that of the logic input.

Cheers

Phil Hobbs

It doesn't look like just a wobbly table but a well tuned (low friction) wobbly table.

That may be a bit pessimistic. When you integrate a chunk of a Fourier plot to obtain jitter, you use the magnitude, so phase cancellation is not an issue. This is a time-tested, industry-standard thing to do, and something that every instrument that measures phase noise is expected to support. Obviously you want to be careful when there are visible spurs between the integration limits, but the software normally removes those first, possibly treating them separately in order to classify the jitter into random and deterministic components. There are all sorts of subclasses in those categories as well.

To the extent a delta function looks like a burst of noise to a jitter integration routine, that's because it *is* a burst of what might as well be noise. If you're trying to answer the question, "What does this sound like?", you can do that by integrating a piece of a digital or analog sweep regardless of where it came from. If you need to go beyond that, you may indeed need to go back to the time domain, but chances are reasonably good that you don't.

-- john, KE5FX

(Actually I should be more careful here... there are cases where the magnitude or real-only output of a cross-spectrum analyzer can underestimate noise near the thermal floor due to the Johnson noise of the resistor(s) in the channel splitter. But that's a corner case of a corner case, and still something of an open topic. It's been happening forever but people only started to look into it recently.)

-- john

Indeed you should. I'm more of an RF guy myself, but I've learned to be ver y wary of schemes that import unexamined assumptions on a large scale.

In a world where only uncorrelated Gaussian white-noise sources are importa nt, there's really some excuse for thinking that way. It works fine for RF data communications, which is where I got my start in engineering as well.

ly.)

Time-domain logic stuff is a different situation. If you're measuring femto second jitter on a 10-MHz square wave, you care very much about the amplitu des and phases of the harmonics of that 10 MHz. Their amplitudes may go as

In the ideal situation, as you make the bandwidth wider the noise amplitude goes as sqrt(B) whereas the transition width goes as 1/B, so you continue to win lower jitter with bandwidth.

The net is that RF habits have to be reexamined when nonsinusoidal waveform s are in view.

Cheers

Phil Hobbs

The gear that measures jitter behaves like a frequency counter: it has a comparator threshold that we set to about half the pulse height. We may have a 1 or 2 ns rise time, but we measure its delay and jitter to a picosecond.

A frequency domain measurement of time delays makes no sense. If we did that, our customers would think we were way beyond the usual lunatic fringe.

You are making less and less sense.

I think John Miles has stated the obvious. Integrating the Fourier spectrum is an industry standard procedure.

However, there may be more than one distribution. So time domain measurements are very important.

You need both.

All the high frequency VCXO's I have looked at, without exception, use the integration method to spec RMS jitter.

Here are two examples:

Connor-Winfield VPLD54TEM

Jitter:

(BW=10 Hz to 20 MHz) : 5ps RMS (BW=12 kHz to 80 MHz) : 1ps RMS

http://pdf.datasheet.directory/datasheets-1/connor-winfield/VPLD54TEM-

Abracon VCXO

Phase jitter RMS (12kHz to 20MHz offset) 1.0 - 1.8 ps See Note #2

Note #2: The rms jitter integrated over 12kHz to 20MHz Bandwidth is dependent on the carrier and whether or not the final frequency is achieved without engaging the Fractional Mode

Presumably, the jitter below 12kHz depends on the bandwidth and gain of the PLL loop the oscillator is locked to.

It takes a set of closely tuned high Q resonators. Metronomes are neither.

Jeroen Belleman

Any oscillator can be injection locked. Examples abound. Here is one:

"A Study of Injection Pulling and Locking in Oscillators"

A similar phenomenon is tidal locking, such as between the earth and the moon.

Early in life, I worked in hard disk drive manufacturing. Divisions that made the platters noticed their yields went down when the women on the assembly line synchronized their periods. The voc's or pheromones released were sufficient to wreck the subtle iron-epoxy formula used to coat the disks.

No high Q stuff there.

Of course. Given enough power, you can make anything waggle at the imposed frequency. Weren't we discussing more subtle effects?

That *is* a high Q system.

That's silly, at first sight. A more credible argument would be that they botched the job because they didn't feel well. Do you have conclusive evidence for that claim?

Jeroen Belleman

It isn't the industry that's being silly, it's you. A barefoot OCXO in a qu iet environment can be plausibly assumed to be dominated by fundamental noi se sources and mechanical instability in the resonator, so the phase correl ations are small and integrating the power spectrum works OK.

JL is building time domain stuff that's chock full of fast edges and FPGAs and SMPSes and so forth, so the random phase approximation is far from safe .

Cheers

Phil Hobbs

It doen't take power. As JL noticed, two oscillators in the same room can lock together. I have seen the same thing.

This phenmenon has been studied for many years. I wish I could remember the name of the guy who did one of the best analysis. I think it starts with a "P".

How do you measure the Q?

The yield was determined when the paint was sprayed on the platters. That was entirely automatic and the girls had nothing to do with the process. They only moved the components through the process and assembled the final product. They did not do anything that had any effect on the yield.

This was a recurring theme in the disk industry. They had years of data to show it was real.

Did you read John Miles post of Sept 19? I repeat it here for your enjoyment.

------------------------------------------------ From: "John Miles, KE5FX" > I couldn't say in general without more research, but any procedure

That may be a bit pessimistic. When you integrate a chunk of a Fourier plot to obtain jitter, you use the magnitude, so phase cancellation is not an issue. This is a time-tested, industry-standard thing to do, and something that every instrument that measures phase noise is expected to support. Obviously you want to be careful when there are visible spurs between the integration limits, but the software normally removes those first, possibly treating them separately in order to classify the jitter into random and deterministic components. There are all sorts of subclasses in those categories as well.

To the extent a delta function looks like a burst of noise to a jitter integration routine, that's because it *is* a burst of what might as well be noise. If you're trying to answer the question, "What does this sound like?", you can do that by integrating a piece of a digital or analog sweep regardless of where it came from. If you need to go beyond that, you may indeed need to go back to the time domain, but chances are reasonably good that you don't.

-- john, KE5FX

---------------------------------------------------- I posted two examples of high frequency oscillators that integrate the Fourier spectrum over a defined frequency range. The spectrum is measured in a production environment. Do you think that is free of RFI and EMI?

If you look at the actual spectrum of many oscillators on the market or published in papers, you can see spikes at the line frequency and harmonics. Most likely from ripple on the power supply. They show that better filtering and isolation is needed.

The Wenzel series of low phase noise oscillators was plagued for years by sensitivity to noise and ripple on the power supply. He only recently came up with versions that reduce the sensitivity.

As I have stated, time domain and frequency domain measurements are both useful. I believe they are both needed.

The only reason that we use the phase-noise-to-jitter conversion math is to qualify candidate oscillators when the OCXO manufacturer doesn't give us jitter vs time data. The final selection is based on measured jitter.

Cheap surface-mount VCXOs and TCXOs vary wildly in jitter performance, with no useful correlation to their data sheets. We just hunt down the good ones. We use those in our lowest-end boxes, where an OCXO won't fit or is optional.

For short delays, instrument jitter is dominated by things other than the XO timebase. Like switching power supplies on the same board.