I agree 100%, the general statement is that a circuit should give the same response, no matter the way it is simulated. My sentence was not so clear however, let me try to explain myself better:
The general point to assess is whether the Noise Figure of a TX mixer depends or not on the power level of its IF input. My intuition says that is should *not* depend on it, if we can say that the IF power level is within the linear operation region of the IF port. Moreover, we are interested in the noise contribution of the Mixer - i.e. the NF and its relatives - only when this contribution is relevant, that is when both input signal and noise power are small, so that even a small contribution by the 'amplifying' (actually 'mixing') device can be of importance. So I would even say that for increasing IF input signal the output SNR will be dominated by the input SNR, with the addittive contribution of the Mixer becoming progressively negligible.
NF = 10*log(SNRi/SNRo) = 10*log[Pi/Po*No/Ni] = 10*log[1/G*(G*Ni+Nd)/Ni] = 10*log[1+Nd/GNi]
..From which I think one could see that if we increase Si by keeping SNRi constant, we have to increase Ni as well, which makes Nd - the contribution of the Mixer in this example - less important. At least, for addittive noise.
That's for the intuitive part.
For the simulation part, I would have run a PSS/PNoise sim with the LO as the *only* tone, and the IF port set at a DC level. Maybe, sweep the DC both common and differential mode to see whether some non-idealities pop up. This, based on my 'theory', but also on the fact that I did not know better, in terms of simulator capabilities.
Reading your link this morning, gave me the impression that HB could handle widely spaced tones (like 2MHz and 2GHz) in a *clever* way. I never used HB as such inside Virtuoso. Only PSS (and HB inside it sometimes). If I had to do this in PSS, I would have set the fundamental to 2MHz and 1000 harmonics, to capture the 2GHz tone. Actually, many more harmonics because I want to see the harms of 2GHz as well.
I tried it in the past, wanting to compare ENV with PSS. Yes, results agree, but PSS this way is not very efficient :)
From the Appendix you linked to I got the impression that HB is cleverer and can handle this math problem more efficiently, so I suggested to bennys just to try and see whether indeed an higher - but still below saturation - IF power can really affect NF. After this online deployment of mine, I am even tempted to say that NF will be *better* (i.e. lower) for higher IF power :)