The true Zeners work. Yes, they are softer. But it is not important if it is 100 mV more or less.
Maybe it would work without the diodes but I wanted less unknowns / easier changes
when debugging the bias system.
One of the pictures is the noise of Zener diodes. One can easily see the strong rise
of the noise vs. rise of Vz. Between 2V7 and 6V are worlds, in the same series.
Diodes are tested with a 1k or 2k wire resistor from 10 NiCd cells.
The NiCds add no visible noise.
The steep rise with 1/f**3 on the left side of the plots goes on the measuring
amplifier resp. its input coupling capacitor. 6 years ago, that was 100 uF foil,
not enough by far. The short & 60 Ohms are on the amplifier side of the cap
and show what the amplifier can do really with a low impedance DUT.
A 4700 uF wet slug tantal healed that. The input stage of the preamp consists of
20 ADA4898 op amps (10 pairs) in par and averaged into an inverter.
Noise in the flat part is abt. 220 pV/rt Hz . The 0 dB line is 1 nV/rt HZ.
I now have an amplifier based on 16 CPH3910 FETs that is nearly in the same
league but with less noise current and a reasonable input cap. I must find
the time time for the write-up. The smaller noise current also opens the
way to cross correlation with the Agilent 89441A. The effect of the noise current
is common mode in the DUT and would not average away.
Among LEDs, blue ones are bad. King of LEDs is HP/Agilent/Avago
HLMP-6000 if you want low noise. Otherwise, it's quite a dim bulb.
There is also a LT3042 for comparison.
For completeness, there are also some regulators.
The LT3042 is standing out.
The LM329 is from Digikey, marked NS, years after TI bought NS.
It does not look like a Zener, more like a bad bandgap.