Hi Ken,

First of all, thank you very much for making all these articles available. You are covering lots of important subjects and everything is very well explained; your site is really a great resource.

I just read your article about PLL phase noise and jitter modeling (PLLnoise+jitter.pdf) and I am feeling a little uncomfortable about the method used for simulating jitter. I see the following possible problems:

It is very difficult to directly compare the models with the real circuits. As the procedure for extracting the jitter values is rather complex, it is easy to make errors without noticing it.

Even after simulating over large time intervals and treating the results with Matlab (which some people might not have available), the resulting curves still look rather "noisy" and this uncertainty might not be acceptable for everyone.

I would prefer an approach where the noise in the voltage-domain models is not expressed as time-domain jitter but as small-signal spectral noise, just like you did it in the phase-domain models. In my opinion, this would have the following advantages:

It is very easy to extract the spectral noise from a noise or pnoise analysis and to compare noise simulation results for the model and the real circuit.

Unlike in the phase domain, one could replace some models in the PLL with the real circuits and simulate them together with the models for the other blocks. This would offer another possibility to verify the accuracy of the models.

Also unlike in the phase domain, effects of frequency conversion would be automatically taken into account.

You would get very nice and smooth curves for the phase noise without having to use external programs. If you are really interested in jitter, this would still be no problem. After all, jitter is simply the integral of the strobed noise at the threshold crossing, divided by the slope of the signal at this point.

I realize that this approach has some problems of its own. The biggest problem is probably that pss convergence is required. This means that all models must be written without hidden states. There may also be problems with large divide ratios. Fractional-N synthesizers are a special problem, some creative solution would be necessary here (maybe a qpss analysis could help?).

I wonder if you have already got some experience with this kind of approach and if there is something I have overlooked which makes it really impractical. If not, I believe the advantages mentioned above would make it worthwhile to explore how far one can get in this direction. What do you think?

Best regards,

Frank

Well, to be honest, my hope was really that you (or someone else) would volunteer to do this work... (or, even better, already had done something along these lines).