You raise here an interesting
question. It is true that manufacturers of active
devices don't provide the tolearnces. These tolerances
would be the deviations of the device intrinsic
parameters like intrinsic capacitances, inductances,
equivalent noise sources etc, which are very hard
to model and which are process-dependant.
Basically the S-parmaeters of an active RF device
are being extracted from the non-linear model provided
by the manufacturer. A deviation from the S-paramaters
e.g. lot to lot or temperature dependant will be
dervided by a deviation in the device internal parmeters.
For a given device, a given X% deviation in S21
(gain), is caused by a given Y% deviation in several
intrinsic paramaters of the non-linear model. You
can ask the device's manufacturer what is for example
the expected deviation in S21 from lot to lot, and
according to that to change the parameters of the
model, which mostly affect the device's gain, by
this to get an idea about the estimateded tolerances
of the device's parametrs.
Hope this helps!