Generally speaking a standpoint of 10GHz and Greater, Air Filled, metal walled resonant cavities are linear devices. But however as you very well know, RF connectors have an associated 3rd Intercept Point (generally quite high, dbm wise). This nonlinearity arises from a schottky type junction formed at the contacts due to the metal conduction band discontinuities. Therefor I propose a mixer cavity based on this principle. Lets take a simple TE101 rectangular waveguide cavity. Build the cavity using two different metals that produce a schottky type junction. Make the Split at the middle of the "b-dimension" walls where you very well know that is where the highest current density near resonance is located. Now we need to inject the two LO/IF signals. Since the resonant cavity bandwidth is quite low ~3-5% these two signals need to be almost equal. Couple to the cavity as you prefer, coaxial probe, Iris whatever. Now the nonlinearity in the wall excited the normal intermodulation products m*if+n*lo m,n=+/- 0,1,2...
.
The lower frequency excitations are now imediately extincted as the cavity can not support them, but the lo+if term can be excited as a higher order cavity mode, TE202 or others. It is your job the efficiently extract this essentally doubled frequency for the overmoded cavity.
In this case you have effectively designed a frequency mixer with output 2*If, 2*Lo and IF+LO, some filtering or a choke may be required to remove the fundementals however.
To be used as a down converter, this scheme is useless as it provides no lo/if isolation and therefor the lo sidebands will totally swamp out whatever if you may have generated in the form of an IF.
I hope this is what you are looking for. My experience is the waveguide type cavities. The DR and MS types may as well have promise in producing mixers with 50dB conversion loss as well.
