Tdnoise option in pnoise analysis can be used to analyze circuits with cyclotationary noise sources. However, when it is used to characterize oscillators (see the following scenario, figures are in the attached pdf file).

Fig.1 shows a CMOS cross-coupled pair (noiseless) negative-gm biased by an ideal tail current source (a noiseless resistor is added to model gds, and an additional verilog-A gm is used to inject current noise into the negative-gm). The only noise source is the gm in parallel with the tail current.

A large differential sinusoidal voltage is applied to the two terminals of the negative-gm, and a verilog-A module measures the differential current between the negative-gm. Tdnoise analysis gives results in Fig.2. In Fig.2, the noise PSD of the sampled noise across the entire period shows that the differential-mode noise nulls when the two transistors in the negative-gm are both in saturation region (when currents from the two terminals of the negative-gm are in transition). This is understandable since when the tail current is the only noise source, there is no differential-mode current if the negative-gm behaves like a fully-differential circuit.

Fig.3 shows an LC oscillator using the negative-gm shown in Fig.1. When tdnoise is again used by the oscillator option in PSS analysis is checked, the differential current output noise changes its proprty. In this case (see Fig.4), the differential current noise of the negative-gm peaks when the negative-gm is in saturation region (currents in transition).

If the tdnoise analysis is equivalent to adding an ideal sampler after the designated output, why these two test cases give so much different results? By definition, the noise sources in an oscillator are not really cyclostationary because there is no real reference sources setting up the periodicity in an autonomous circuit. Does this imply that tdnoise is not realy valid in simulating an oscillator?

Does anybody know what makes this difference?