I don't know why it took me this long to realize this, but I just realized that the zener resistor value plays a large role in determining the error "attenuation" factor of this circuit.
"Rz" in the above circuit was given as 1k. This is what turns 1mV of op amp voltage error into 1uA of zener current error.
If you instead used 2k for Rz (and for Rf), 1mV of op amp voltage error becomes 0.5uA of zener current error.
So, if you have some additional voltage headroom on your op amp, you can increase this value a bit to get some additional error attenuation.
This isn't so important with the LM399, because it already has a low output impedance (of about 1R). However, for other zeners (like the 2DW232) which have higher output impedance, they won't have as good attenuation in this circuit. Here, using higher value resistors for Rz and Rf would help improve their attenuation.
For example, if you had a 15V Vcc for your op amp, and it can safely swing the output up to 12V, you could use 5k for Rz and Rf for additional error attenuation.
In addition to that:
The error attenuation is maximized near a gain of 1 and loses significance as gain increases.
This is due to the "1 +" term in the gain equation for non-inverting opamp: gain = 1 + Rf/Rg.
The key parameters of opamps - that have an effect here - are input referred, that means this parameters have to be multiplied by the gain.
Comparison for different Rz @ Iz = 1mA, Vz = 6.95V with gain(Rz) = 1 + (Rz x Iz) / Vz and dIz = (gain x dVin) / Rz, with given dVin = 1mV (opamp tempco, drift, noise ...)
1k => gain = 1.14 => dIz = 1.1µA
2k => gain = 1.29 (+12% to 1k) => dIz = 0.64µA (-78% to 1k)
4k => gain = 1.58 (+22% to 2k) => dIz = 0.39µA (-63% to 2k)
8k => gain = 2.15 (+37% to 4k) => dIz = 0.27µA (-46% to 4k)
The conclusion remains:
Rz should be set
as high as possible (for given max output voltage of opamp)
to have the
least significant influence on Iz from errors in Vout
from opamp over temperature/time.
Errors in Vout are resulting from opamp due to tempco, drift, noise, ...
The next question arises:
How do errors from
feedback network Rf/Rg due to tempco-match, drift-match, noise (gain-error) contribute to the choice for Rz?
It is easy recognizeable that there is a trade-off/target conflict for this...
And the final question:
Where is the
optimum for Rz in therms of influence for all changes in Vout over temperature/time?