Aaron,

  First, my guess is that you are using Shooting Newton with the
moderate error preset. If so the issue is the following:
- At moderate signal levels, the distortion is above the numerical
 noise floor and the slope is 3dB/dB
- At lower signal levels, the distortion is below the numerical noise
 floor. The numerical noise floor is set by reltol, possibly
 reltol* fundamental. In this region the distortion has a slope of
 roughly 1dB/dB just like the fundamental.

The issue is not simulator accuracy. The issue is that your simulation
setup is not appropriate for the accuracy you are asking the simulator
to provide.

Some solutions:
1) Use harmonic balance
2) Use a tighter error preset, for example, conservative
   - This change will increase simulation and degrade convergence
      in saturation
3) Use higerorder=yes,
   - This change will significantly increase simulation and may
      degrade convergence in saturation

Shooting Newton is based on transient analysis and transient analysis
does not have infinite accuracy and that improving the accuracy of
transient analysis costs time.

In general, if you are interested in the results for low levels, try
harmonic balance.

                                                        Best Regards,

                                                          Sheldon

Aaron,

  Some circuits can be analyzed using their quiescent, dc, operating
point, some can be analyzed using their periodic operating point, and
some circuits do not have a steady-state. Shooting Newton and
Harmonic Balance are algorithms for calculating the periodic steady-
state of circuits. Shooting Newton calculates the periodic steady-state
response in the time domain. Harmonic Balance calculates the periodic
steady-state response in the frequency domain.

Shooting Newton uses transient analysis to calculate the periodic
steady-state.

Harmonic Balance calculates the frequency domain response
directly.

Shooting Newton and Harmonic Balance complement each other.

Shooting Newton is useful when time domain waveform changes
quickly, for example, switched capacitor circuits, and when the
circuit is dependent on an instantaneous event, for example,
Colpitt's Oscillators. Shooting Newton is also efficient for single
tone periodic steady-state analysis.

Harmonic Balance is useful when time domain waveforms do not
change quickly or the circuit is dependent on instantaneous events.
Harmonic Balance is also useful when the circuit characteristics can
be more effectively described in the frequency domain. Most RF
applications fall into this category. Harmonic Balance is more
efficient than Shooting Newton for multiple tone periodic steady-
state analysis.

Transient analysis is prior to the steady-state calculation is used to
improve convergence. For Shooting Newton, circuits often have
start-up transients and these need to settle (somewhat) before
starting the periodic analysis. For Harmonic Balance, some circuits
have fast changing waveforms, for example, dividers and running
transient analysis first can help the HB analysis converge.

It might be useful to think about it as there a set of tools available
to you and transient analysis is the hammer. You can do anything
with it, but it is not also the most efficient tool. DC/small signal
and PSS/small signal are not as generally applicable, however,
they are very good at the things they can do.

Hope this helps!

                                                       Best Regards,

                                                         Sheldon

I have seen this issue with BSIM CMOS models that have not been properly extracted. Unfortunately, I am not experienced with BSIM model extract, so I really do not know what the differences are.

Shooting methods is an iterative time domain solver for finding the steady-state periodic time domain solution to circuit problem. Periodic is defined as the be first point and last point are the same value (I'm pretty sure this is the only restriction).

Flexbalance is a harmonic balance solver that essentially solves the circuit in the frequency domain. All sources are defined as sinusoids (harmonic series for squarish signals), all linear components are defined by their respective frequency response, and intrinsic nonlinear functions are described by their time domain I-V characteristic. Flexbalance deals with nonlinearities using an IFFT to go from frequency domain to time domain (to apply signals to the intrinsic nonlinearity) and FFT to go back to the frequency domain. The solver iteratively goes back and forth until the harmonics at the interface stop changing (balance).

QPSS is even different in the sense that this solver can decouple large signal responses from so called moderate signal responses. QPSS flexbalance really speeds up mixer and amplifier IM2 and IM3 measurements. For mixers you make the LO the large signal and for amplifier one of the two-tone input signals.

My position is that I always trust transient first. If there is a problem with transient then there is either a model or a simulation setup problem. Then I progress to the other techniques as needed after establishing a transient baseline.