I am using PSS+PNOISE to estimate output jitter of a limiting amplifier as follows:-

1) Drive the input with a single tone at Fin
2) Run PSS with shooting method, beat frequency Fin, select number of output harmonics N such that Fin*N > circuit's 0dB gain
3) Run PNOISE with noisetype=jitter, maximum sideband=N, reference sideband=0, frequency sweep from 1Hz to Fin/2
4) Plot Jee with integration limits 1Hz to Fin/2

So far I have been happy with this procedure - the results for Jee etc are reasonable and agree with a few TRAN+NOISE sims run as spot checks. It also agrees with recommendations elsewhere on the forum (e.g. www.designers-guide.org/Forum/YaBB.pl?num=1092399689) and my somewhat incomplete understanding of exactly how the simulator works its magic.

Recently, other members of my design group have questioned this approach. What they propose is mostly similar, but change PNOISE as follows:-

3a) Run PNOISE with noisetype=jitter, Sideband array of indices, additional indices=0, reference sideband=0, frequency sweep from 1Hz to 'very high' (well beyond circuit 0dB gain)
4a) Plot Jee with integration limits 1Hz to 'very high' (same as above)

Jitter numbers for both methods are very close (less than 1% mismatch, although not identical). However, when I run a few stripped-down test cases the actual output noise (using noisetype=sources) shows bigger deltas, up to 10% or so for the particular examples I have run so far.

My questions are therefore:-

a) Which of these methods is 'correct'?
(Maybe they are equivalent? Maybe they are both wrong?)

b) Can someone explain (or recommend a good reference to explain) the difference, if any, so I can improve my understanding?

Many thanks!

Pete

Hi vp1953,

Yes. Using the second method with more than a single harmonic and noise integration bandwidth > Fin/2 results in significantly higher jitter reported by PNOISE.

As I understand it, this is due to aliasing.

(Note: This was the result that sparked the debate. Several designers used the original method, but with integration bandwidth >> Fin/2 and got the "wrong" result. Therefore the suggestion was to use a single harmonic. I prefer the original method, but am hard pressed to explain why to myself, let alone others...)

Thanks,

Pete

Actually TRAN+NOISE matches quite well too (I run for a few hundred cycles, export the zero crossings into a spreadsheet and calculate Jee by curve fitting a Gaussian CDF). I get an answer within about 5% of PSS+PNOISE (either of the methods in my original post).

Thanks,

Pete

Hi Frank,

Thanks for the advice - I have looked at sc-filters.pdf and believe the PNOISE described there to calculate sampled noise is identical to what I called method 1 in my original post.

I agree that the two methods can only be equivalent when there is little or no frequency translation of noise. Since (I think) that is the situation I am simulating, my provisional conclusion is that in this case the two methods ARE equivalent. However I wish I could provide a rigorous argument to support (or refute!) it rather than relying on a small quantity of numerical data.

I will try to experiment with more/less non-linearity, more cascaded stages etc to see if I can isolate any systematic differences between the two methods.

Thanks,

Pete