I think I posted this on the old TheRadioBoard forums, but I forgot about it, and recently again had to learn the following lesson: when simulating self-oscillating mixers in LTspice or similar software, you need a really small timestep to properly observe the IF output. A too-large timestep will lead to a strangely noisy output and apparent absence of the IF signal at the IF output.
As an example, I simulated a self-oscillating mixer oscillating at 7 MHz and fed with a 5 MHz RF input signal (AF-modulated with a 10 kHz signal). In order to see the IF signal at the IF output (a resonant LC tank), I needed a timestep of 100 picoseconds, meaning 1000 steps for 1 microsecond. Anything larger led to an excessively "noisy" output, where no trace or only a very distorted version of the IF output could be seen. The apparent "noise" and false signals created with a too-large timestep are due to temporal aliasing, meaning that LTspice does not have enough simulated data points to capture the small details of the waveforms.
Another difficulty with a self-oscillating mixer is that the output at the IF port (the LC tank tuned to the IF frequency) has a very large LO component. To see the IF output, you have to zoom in to the top-edge of the output signal at the IF port, ignoring the large LO component, and focusing only on the small variations in the top-edge of the signal. These small variations "riding" on top of the output signal occur at the frequency of the IF (in this example, 2 MHz) and are visible as small "grass-like" peaks and valleys on the top edge. Again, if the timestep is too large, these small variations will get highly distorted due to inadequate sampling, and the IF signal will not be visible and will be drowned in artificial "noise" caused by the inaccurate simulation results.
To visually confirm if the timestep is adequate, zoom in closely to the narrowest peaks in the output window, and confirm that even the narrowest peaks appear smooth and curved. If they appear sharp and jagged, drawn with only a few straight-line segments joined together at sharp angles, then this indicates that the timestep is too large and should be reduced by a factor of ten, repeatedly until the curves appear completely smooth.
It's also important to turn off all waveform compression in LTspice, in order to force LTspice to actually use the specified timestep and to do a full simulation every timestep.
Here's an example of a correctly-simulated self-oscillating mixer output. Note that the first 6000 uS of the simulation data have been discarded (to omit the oscillator's start-up behavior and only caputure the final steady-state behavior), so where the graph shows 0 uS it is actually already 6000 uS into the simulation.
This is the circuit. For this example, ignore the Q3/Q4/Q5 IF filters and amplifier. In this example, the IF output is taken at the first IF tank, at the "hot" end of the L4 inductor.
The output at the IF port ("hot" end of L4 inductor) shows a strong LO component:
If we zoom in, we see a 7 MHz LO signal (7 peaks in one microsecond):
So where is the IF signal? We have to zoom out again, and zoom in on the top edge of the signal.
You can barely see that there is a small-amplitude, lower-frequency signal appearing on the top edge. Let's zoom in further:
The "grass-like" signal "riding on top" of the larger LO signal increases and decreases in amplitude with a period of 100 uS -- corresponding exactly to the 10 kHz AF modulation of the RF input signal. So we are indeed seeing the frequency-converted RF signal, now present as a very small IF signal "riding on top of" the huge LO signal that is leaking through to the IF output.
For reference, here is a similar plot, but with the RF input signal displayed below in blue: a 10-kHz modulated 5 MHz signal. Note that the measured RF signal (taken from the left side of R14) is not quite a pure 5 MHz signal, because it has been "contaminated" by the LO signal bleeding out of the RF port. But the point is that the modulation contained in the original 5 MHz RF input signal has been faithfully reproduced in the lower-frequency, 2 MHz IF signal that is "riding on top of" the LO signal, available at the IF output port.
Zooming in further, we can see that the small peaks and valleys of the IF signal occur two times in one microsecond -- indicating a frequency of 2 MHz, which is indeed our IF frequency.
Zooming in even further, let's confirm that our timestep is small enough to capture the correct, smoothly-curved waveform shapes.
These peaks look kind of sharp! Are they really sharp, or are they curved? Let's zoom in further:
And zooming in further still to one single peak:
Here we can finally see that the curve is indeed smooth and that our timestep is small enough. If the timestep were too large, this peak would appear rough and sharp, and when zooming out, the "grass-like" signal riding on top of the LO would be extremely noisy and the input 10 kHz modulation would no longer be visible in the IF output -- in other words, the small IF output signal would have disappeared due to simulation inaccuracies caused by a too-large timestep.
So, simulating self-oscillating mixers is a bit of tedious work in LTspice, but it can be done.
I hope to eventually make some comparisons of conversion gain between the self-oscillating mixer and a separate LO-mixer combination.