So I started fully characterizing the beat detection path.
As a part of the characterization works, I measured the spectra of the RFPD noise as well.
The noise is totally dominated by that of the RFPD (i.e. not by an RF amplifier).
I am going to check the noise curve by comparing with a LISO model (or a simple analytical model) in order to make sure the noise is reasonable.
The red curve represents the dark noise of the RFPD, which is amplified by a low noise amp, ZFL-1000LN.
The blue curve is a noise of only ZFL-1000LN with a 50 Ohm terminator at its input.
The last curve is noise of the network analyzer AG4395A itself.
It is clear that the noise is dominated by that of RFPD. It has a broad hill around 100MHz and a spike at 16MHz.
Gain of ZFL-1000LN = 25.5 dB (measured)
Applied voltage to ZFL-1000LN = +15.0 V
Bias voltage on PD = -150 V