Phase Noise Head to Head: Siglent SSA3032X vs HP 8566B vs SignalHound SA124B

[hendorog]

Currently I have a cool problem - too many Spectrum Analysers!
I have had a SignalHound SA124B for quite a while, I've just acquired an HP 8566B in really nice condition, and member tautech has lent me his Siglent SSA3032X for a comparison.

I am interested in low phase noise equipment, and this is not often covered in any detail in modern equipment. So I thought it was a good chance to compare these three instruments and see what they've got.

Before this gets going, important to point out that I'm not an RF engineer and am not pretending to be one, so advice, education and suggestions are welcome.

Firstly to measure the phase noise contribution of an SA, the signal generator needs to have lower noise than the LO of the SA itself.

I am using a Marconi 2041 for this test, which 10kHz - 2.7GHz and is very low noise. Typical performance is below 140dBc/Hz at 1GHz@10kHz offset. I haven't the capability to confirm this yet, so am operating under the rash assumption that my example is somewhere close to this spec.
https://www.avionteq.com/Document/IFR-Aeroflex-Marconi-2042-Specification-Sheet.pdf

Some comments on the procedure:
Setup is just RF out of the Marconi and into the SA's one by one. The source is set to 0dBm, which meant that dBm/Hz is close enough to dBc/Hz for it to not matter.
The 10MHz references are not coupled together as I didn't want to influence the PN. Because of this, in some cases there will be small errors due to me not checking the frequency is centered exactly.

Method
The SignalHound has a built in Phase Noise test mode, so I used that.

The HP 8566B has the very nice KE5FX PN utility. That makes life simple so I used that with an HP 82357A USB-GPIB adapter.

The Siglent required a bit more work to create a phase noise trace. I started the trace at an offset of 1kHz from the carrier, and ended it at 1MHz offset. I set the RBW to 10Hz to get keep the sweep time under control, and used the Math function to deduct 10dBm to correct for the RBW not being zero - this normalised the trace to 1Hz. Then I set markers at 1kHz, 10kHz, 100kHz and 500kHz. Using noise markers would get the marker values, but the trace itself would not have been normalised to 1Hz.

I tried using a log mode sweep, but I couldn't get it to work on narrow spans. Turning it on forced the SA into sweep mode which was much slower than FFT mode, and limited the RBW I could use. It did appear to work correctly on wider spans though.

The CSV data from the SSA was exported and graphed in Octave to get a Log-Log plot. After doing that I realised it would be good idea to graph all three on the same chart, to make the comparisons much better.

I will post what I have so far anyway, first test was 2.7GHz, which is the max of the 2041. Feedback welcome.

[Performa01]

Nice comparison - and we can draw the conclusion that all reasonable budget instruments are quite similar in this regard, whereas the HP 8566B outperforms them by a clear margin - probably even more pronounced at higher frequencies like 2.7GHz as in your example.

For good comparability, you should use identical start frequencies for the phase plots though.

I happen to have a screenshot for the SH SA44B with external reference, which is required for close-in phase noise measurements (when looking at the HP phase noise plot below 1kHz, it appears it would benefit from an external high quality reference clock as well). Please note that the measurement starts at 1Hz carrier distance and 1kHz is at the center of the x-axis.

The results for 2.7GHz are unsurprisingly very similar to the SA124B, yet a little worse for the SA44B, weirdly enough because of the external reference, which brings a huge improvement below a couple hundred Hertz, but slightly degrades phase noise at higher carrier distances.

The phase noise is plotted for three different center frequencies: 5, 250 and 2700MHz. Below some 150MHz, the Signal Hound shows funny jumps in the phase noise plot, which limits its usefulness for this kind of measurements at lower frequencies.


MG3633A_PNP_5_250_2700MHz

[hendorog]

Yes I overlooked the start offset of the HP being different to the others until it came time to post, also I realised late in the piece that the Y range doesn't match. I will remedy that.

Thanks for the SA44 traces. From what I can see it looks very similar to the SA124.
One interesting point is how the SH PN gets much better as the frequency drops. OTOH while the HP is good at 2.7GHz, it gets no better at 20MHz - at offsets > 1kHz

[Performa01]

Oh - that's interesting!

I also have a screenshot for 25MHz (close enough!) and there the HP 8566B doesn't look superior anymore. Quite the contrary, in the range of 1-70kHz carrier distance it looks significantly worse...


MG3633A_PNP_25MHz

[hendorog]

Hehe, yes I was surprised as well by that too. It is the winner at low frequencies.

I am still a little skeptical though as I noticed that the PN is influenced by switching the spur detection off vs spur detection on.  So I want to check that it really can measure something at that level.

Attached is a measurement at 20MHz starting at 10Hz offset - references not linked still.

[hendorog]

OK, I have managed to combine it all onto one chart, using Octave. Here it is at 20MHz, 1GHz and 2.7GHz.

At 1GHz and above is where the HP is starts to show it's class. The Siglent also comes back into it and is ahead of the SignalHound at 2.7GHz.

This is getting interesting, I'm going to use a doubler to see what it looks like above that.

Edit: will do that tomorrow.

[Performa01]

Okay, conclusion so far: the HP 8566B beats the SSA3032X by a clear margin, no matter what center frequency. Compared with the SH this is only true at and above 1GHz, whereas at lower frequencies the SH is the clear winner and even at 1GHz there is basically just the span from 20-200kHz where the HP is some 8dB better. Apart from that, the SH generally struggles to get below some -125dBc, even at 1MHz distance from the carrier.

Interesting results in any case. To be honest, I would have thought that the SSA performs a little better than that. Even at 2.7GHz it is hardly any better than the SH - a couple dB lower phase noise below 20kHz carrier distance, but then also a couple dB more at and above 100kHz.

This raises the question whether the test method is appropriate to get the most out of it. After all, the SSA appears to be the only contender that does not have a dedicated phase noise test utility. But then again, your latest results have been obtained by using the same test method for all three analyzers, right?

[tautech]

Just jumping in to keep up with the play and get further developments in my post replies.
Hendorog and I have chatted about this a bit and I was party to his initial checks a couple of weekends back.
I guess he'll explore some suspicions he has WRT the componentry Siglent used and how SSA's could be improved.
Staying tuned. 

[hendorog]

This raises the question whether the test method is appropriate to get the most out of it. After all, the SSA appears to be the only contender that does not have a dedicated phase noise test utility. But then again, your latest results have been obtained by using the same test method for all three analyzers, right?

Good question. No the test method is not the same for all three, I have taken advantage of automated tools to save time.

The SSA measurement is outlined above - I set the span from min offset to max offset and reduced the RBW to 10Hz, and made a 10dB adjustment to allow for that. The detector was set to Avg.

The SH does multiple sweeps which are consolidated to build up its PN display, and so does the KE5FX utility for the HP.

Assuming the PN utility from KE5FX is open source, the most efficient way is probably to add the SSA and SH to that to make them all the same.

[KE5FX]

The SH does multiple sweeps which are consolidated to build up its PN display, and so does the KE5FX utility for the HP.

Assuming the PN utility from KE5FX is open source, the most efficient way is probably to add the SSA and SH to that to make them all the same.

It's not "open source" in the trademarked sense of having a viral or semi-viral license and a Git repository with an official maintainer, but it's "open source" in the sense that you get a copy of the development directory with the .cpp sources and makefile when you run the installer, whether you want them or not.   Patches to support new instruments are always welcome. 

There is already some code in specan.cpp to talk to the Signal Hound API, added to support the original SA44 in the SSM utility.  PN uses separate routines, but if you did want to add SAxxx support to it, the specan.cpp code would be a good starting point. 
 
Sounds like your 8566B has a late serial number, probably over 27xx, correct?  Those results are the cleanest I've seen.  Normally the 8568B would be the best way to go if you wanted the best PN performance in a heavyweight benchtop analyzer, but that particular 8566B seems almost as good near 10 kHz.

[G0HZU]

I think that the first LO of the HP8566B will only be at 3021.4MHz when tuned to 2700MHz because this is in band 2 on this analyser. Maybe retest it at the upper end of band 1? eg try something like 2400MHz where the first LO will be up around 6GHz.

Your phase noise plots look really lumpy for the HP8566B. It shouldn't look lumpy like this and the noise levels in the 1kHz to 7kHz offset range look to be a bit too good even with the lumps. Edit: I just fired my old HP8566B up to check it and got -101dBc/Hz at a 5kHz offset at 20MHz and -98dBc/Hz at a 5kHz offset at 1GHz.

The HP8566B was the only classic 1980s lab analyser we never bought or rented at work and this was mainly because other lab analysers had lower phase noise and better dynamic range in the ranges we were looking at in the early 1990s. By the time we needed the 22GHz range it offered the HP8563E was available and the PSA arrived a few years later.

A better benchmark for a lab analyser that covers 0-2.9GHz would be the HP8560E series from the 1990s. The close in phase noise for the HP8560E was quite good.

[KE5FX]

I think that the first LO of the HP8566B will only be at 3021.4MHz when tuned to 2700MHz because this is in band 2 on this analyser. Maybe retest it at the upper end of band 1? eg try something like 2400MHz where the first LO will be up around 6GHz.

Your phase noise plots look really lumpy for the HP8566B. It shouldn't look lumpy like this and the noise levels in the 1kHz to 7kHz offset range look to be a bit too good even with the lumps. Edit: I just fired my old HP8566B up to check it and got -101dBc/Hz at a 5kHz offset at 20MHz and -98dBc/Hz at a 5kHz offset at 1GHz.

The HP8566B was the only classic 1980s lab analyser we never bought or rented at work and this was mainly because other lab analysers had lower phase noise and better dynamic range in the ranges we were looking at in the early 1990s. By the time we needed the 22GHz range it offered the HP8563E was available and the PSA arrived a few years later.

A better benchmark for a lab analyser that covers 0-2.9GHz would be the HP8560E series from the 1990s. The close in phase noise for the HP8560E was quite good.

The lumpiness isn't too surprising if it really is performing that well.  If you have lower-than-usual PN in a given example, you're more likely to see evidence of small internal spur clusters that are far below the specification, and that would normally be buried in the noise.

There were some noticeable improvements over time.  An 8566 from the late 90s will tend to be several dB quieter near 10 kHz than an older model from the early 1980s.   The noise level also doesn't have a straightforward relationship to the carrier frequency due to the way the M/N synthesizer works. 

None of these analyzers is quiet enough by itself to measure the noise from a decent signal source directly.  The newer Signal Hound models would be good candidates, though.

[hendorog]

The SH does multiple sweeps which are consolidated to build up its PN display, and so does the KE5FX utility for the HP.

Assuming the PN utility from KE5FX is open source, the most efficient way is probably to add the SSA and SH to that to make them all the same.

It's not "open source" in the trademarked sense of having a viral or semi-viral license and a Git repository with an official maintainer, but it's "open source" in the sense that you get a copy of the development directory with the .cpp sources and makefile when you run the installer, whether you want them or not.   Patches to support new instruments are always welcome. 

There is already some code in specan.cpp to talk to the Signal Hound API, added to support the original SA44 in the SSM utility.  PN uses separate routines, but if you did want to add SAxxx support to it, the specan.cpp code would be a good starting point. 
 
Sounds like your 8566B has a late serial number, probably over 27xx, correct?  Those results are the cleanest I've seen.  Normally the 8568B would be the best way to go if you wanted the best PN performance in a heavyweight benchtop analyzer, but that particular 8566B seems almost as good near 10 kHz.

Thanks, yes that was the definition of open source I meant, i.e. just access to the source, not the full legal definition I think I've used the SA44 SSM util on my SA124, will have a poke around with the code and see what I can do.

The serial is 3026xxx so yes it is pretty new - I think is 30 years after 1970? 1960. (Edited)

I used 20dB clipping and a 0dB signal. I left the 10dB attenuation setting alone - i.e. 10dB attenuation enabled.

[hendorog]

I think that the first LO of the HP8566B will only be at 3021.4MHz when tuned to 2700MHz because this is in band 2 on this analyser. Maybe retest it at the upper end of band 1? eg try something like 2400MHz where the first LO will be up around 6GHz.

Your phase noise plots look really lumpy for the HP8566B. It shouldn't look lumpy like this and the noise levels in the 1kHz to 7kHz offset range look to be a bit too good even with the lumps. Edit: I just fired my old HP8566B up to check it and got -101dBc/Hz at a 5kHz offset at 20MHz and -98dBc/Hz at a 5kHz offset at 1GHz.

The HP8566B was the only classic 1980s lab analyser we never bought or rented at work and this was mainly because other lab analysers had lower phase noise and better dynamic range in the ranges we were looking at in the early 1990s. By the time we needed the 22GHz range it offered the HP8563E was available and the PSA arrived a few years later.

A better benchmark for a lab analyser that covers 0-2.9GHz would be the HP8560E series from the 1990s. The close in phase noise for the HP8560E was quite good.

I have a friend with a 3048A setup and an HP8563E. He doesn't live close by, but I could be travelling his way this weekend...

I will do a run at 2400MHz and post that.

[B1nary]

As an owner of the Siglent 3032, I found this very interesting. I'll be following along. Good job!

[hendorog]

Here are the traces from the HP comparing 2.4GHz - top end of band 1-  with 2.7GHz - band 2

2.7 GHz does have lower noise around 10kHz as predicted by G0HZU.

[hendorog]

Here is a measurement of my SignalHound 12.4 GHz Tracking Generator.

This unit has very ordinary phase noise and so is well within the capabilities of the three instruments.

They all line up quite well for most of the range.

The SSA only has a few data points at low offsets because I am just grabbing all of the data in one sweep. The other two combine multiple sweeps to improve resolution.
I think I'm hitting the noise floor of the SSA at 1MHz offset, since this is a lower power source I could set 0dB attenuation.

[G0HZU]

There were some noticeable improvements over time.  An 8566 from the late 90s will tend to be several dB quieter near 10 kHz than an older model from the early 1980s.

OK thanks. I've only ever used a couple of 8566 analysers so I don't have much experience of how much the performance was changed over the years. The one I have here is quite old as it is an A that has had the factory conversion to a B. I assume this just upgraded the controller (and a few minor components) and kept the RF bits as they were. This was probably done after 1989 as the CRT shows a 1989 date for the B firmware. Mine has phase noise that is slightly better than the typical phase noise plots in HP's specs for close in phase noise for the HP8566B and it generally seems to be about 5dB cleaner than the 'typical' HP spec at a 100kHz offset.

Mine seems about 5dB better at 100kHz offset at 2.4GHz than henderog's results indicate. But my close in phase noise at 10kHz offset at 2.4GHz is about -96dBc/Hz when tested with my 2024 sig gen. However, it's probably more like -97dBc/Hz because the 2024 sig gen will be adding a small amount of noise here. But this is still much worse than Henderog's results at this offset.

The phase noise response on mine is really smooth with no lumpy bits. I'm used to seeing smooth phase noise on all HP/Agilent/Keysight lab analysers at work. I've used loads of the HP856xE analysers at work. We have owned/rented loads of them over the years and they are really consistent (between analysers) for phase noise on band 1. They are much cleaner at offsets <20kHz than my HP8566B for example.

[rf-loop]

I set the RBW to 10Hz to get keep the sweep time under control, and used the Math function to deduct 10dBm to correct for the RBW not being zero - this normalised the trace to 1Hz. Then I set markers at 1kHz, 10kHz, 100kHz and 500kHz. Using noise markers would get the marker values, but the trace itself would not have been normalised to 1Hz.

Perhaps nitpick but..

In your 1. image there is red marked marker values and they match with visible trace. Is this trace corrected for noise dBm/Hz

I am not now near SSA for test but I believe you have of course noted difference between gaussian shape filters produced level and if we talk about example random noise dBm/Hz levels what need around 2.5dB correction.
When some use 10Hz gaussian type RBW filter for random noise and deduct 10dB it is after then "normalized" for 1Hz gaussian filter level. But if we talk about random noise dBm/Hz it is defined for rectangle filter, just as 1Hz wide slice from noise. For this, it need typically around 2.5dB correction.
If I remember right (my old memory is not perfect) Siglent SSA noise markers function  do this correction.

This is quite well explained in (destroyed HP)Agilent-Keysight AN-150, starting page 53.

[hendorog]

I set the RBW to 10Hz to get keep the sweep time under control, and used the Math function to deduct 10dBm to correct for the RBW not being zero - this normalised the trace to 1Hz. Then I set markers at 1kHz, 10kHz, 100kHz and 500kHz. Using noise markers would get the marker values, but the trace itself would not have been normalised to 1Hz.

Perhaps nitpick but..

In your 1. image there is red marked marker values and they match with visible trace. Is this trace corrected for noise dBm/Hz

I am not now near SSA for test but I believe you have of course noted difference between gaussian shape filters produced level and if we talk about example random noise dBm/Hz levels what need around 2.5dB correction.
When some use 10Hz gaussian type RBW filter for random noise and deduct 10dB it is after then "normalized" for 1Hz gaussian filter level. But if we talk about random noise dBm/Hz it is defined for rectangle filter, just as 1Hz wide slice from noise. For this, it need typically around 2.5dB correction.
If I remember right (my old memory is not perfect) Siglent SSA noise markers function  do this correction.

This is quite well explained in (destroyed HP)Agilent-Keysight AN-150, starting page 53.

Aha, thanks. No I haven't made that correction for filter shape.

I did notice that the SSA trace was a dB or two different from the others, and different from the dBm/Hz markers on the SSA but I didn't connect the dots. That will explain it.

[hendorog]

I measured the SSA 3032X Tracking Generator at a range of frequencies. It is quite good, certainly much cleaner than the SignalHound tracking generator I already measured.

The second plot shows the SignalHound TG vs the SSA 3032X TG. There is a small difference in frequencies, but clearly the SSA is far better here.

This was measured by downconverting the SSA Track gen output to 5MHz using an ADL5801 mixer and the 2041 sig gen. Then directly measuring at 5MHz using the SA124B, which is nice and clean there. The 3.2GHz measurement was done at 500MHz as the 2041 maxed out at 2.7GHz.

While doing this, I also noticed that the SSA LO range is from 3.93GHz (-45dBm) to 7.13GHz (-60dBm) - by observing the leakage at the TG port.

[G0HZU]

The cleanest regular analyser I have here for 100kHz offsets is my fairly ancient Advantest TR4172. It manages -128dBc/Hz at 100kHz offset across HF/VHF and up to 1GHz. The only spectrum analyser I have that beats the SignalHound at close offsets down at 20MHz is my 8GHz Tek RSA3408A. It beats it by maybe 15-20dB at offsets below 10kHz.

[hendorog]

The cleanest regular analyser I have here for 100kHz offsets is my fairly ancient Advantest TR4172. It manages -128dBc/Hz at 100kHz offset across HF/VHF and up to 1GHz. The only spectrum analyser I have that beats the SignalHound at close offsets down at 20MHz is my 8GHz Tek RSA3408A. It beats it by maybe 15-20dB at offsets below 10kHz.

Very nice!

[hendorog]

As Tautech mentioned earlier, we were wondering why the SSA3032X PN is at the level it is at.
There seemed to be the potential for it to be better, given the high quality parts used, and at first glance most of the PLL's seem capable of lower phase noise performance.

Here are the PLL chips which were found in the teardowns:

1st LO: Hittite HMC703 + Z-Comm VCO + Doubler (3.9-7.1 GHz)
2nd LO: Hittite HMC835 (3.1 GHz)
3rd LO: Hittite HMC835 Divide by 4 (800 MHz)
Tracking generator LO: HMC835 (3.9 GHz)
ADC PLL: ADF4001

So what is limiting the phase noise to around -100 dBc/Hz at 10kHz?

I checked the datasheets for the phase noise specs, and this is what I take from it:

ADC clock:
The ADF4001 stuck out to me, as  the datasheet shows it can only achieve -100 dBc/Hz at any offset at 200MHz - even though is is supposed to be 'Ultra Low Phase Noise'.
I think this is because it is an old part with a datasheet revision of 2003. The ADF4002 is a bit newer and has better performance but still not brilliant.

However, this PLL clocks the ADC, and I think it is tuning a 40Mhz crystal. This is a factor of 8 5 below the datasheet frequency, so theoretically 18 14dBc/Hz lower noise.
Also, I don't know how much effect ADC clock noise will have on the phase noise of the SA itself? Any comments?
And of course, the ADC clock does not influence the PN of the tracking generator, which is also a bit worse than -100dBc/Hz close in.

TG:
The TG takes the 1st LO and downconverts it using an HMC835 which must be running at about 3.9GHz. The HMC835 datasheet shows that it might be just a little bit better than -100dBc/Hz at 3.9GHz.

I thought the TG would be a good way to determine the PN of the LO. But it turned out that according to the datasheet the HMC835 hits about -100dBc/Hz @ 10kHz at it's operating frequency of 3.9GHz.
Since this is close to what I'm seeing then I still don't know what the limit is.

1st LO:
This is implemented by an HMC703 and a 2-4 GHz VCO, which is then doubled.
My first thought was that this would not be an issue. According to the datasheet the HMC703 can get close to -110dBc/Hz at 8GHz and around 10kHz offset.
(It only runs at ~3.6GHz in the SSA so would be ~6dB better than this, but since it is doubled in the SSA there is a 6dB penalty - so the 8GHz figure is close enough)

The caveat with this however, is that the HMC703 can only reach its best PN in 'HiK' mode, which uses a higher charge pump current. According to the datasheet this requires an active loop filter. The teardown shows a passive loop filter, so I assume that this mode was not able to be used.
Therefore the 1st LO may be only around -100 or -101dBc/Hz at the top end of it's sweep at 7.1 GHz. It may be about -106dBc/Hz at the bottom end of the sweep - 3.9GHz.

2nd LO:
The 2nd LO is another HMC835 at about 3.1 GHz, and so should be below -105dBc/Hz according to the datasheet.
So any improvements in the 1st LO past this level would not improve the SA function.

3rd LO:
This is divided twice from the 2nd LO and therefore should be 12dB better than the 2nd LO.

The phase noise of the SSA itself is fairly static until about 2GHz. Then it starts to degrade as the frequency is increased.
So I've concluded that either the ADC clock, or the 2nd LO, or both must be the limiting factor until 2GHz as they don't sweep and so their PN will not change with frequency. After that the 1st LO becomes the limiting factor as it is sweeping.

I was initially thinking that there was a single main driver of the PN, but now I don't think this is the case. I think that there are two, perhaps more, parts which limit the PN to about the same level.

I'm interested to hear any thoughts.
How do you determine when the ADC clock PLL quality becomes critical to the PN performance of a SA?
Anyone able to shed any light on the HiK mode in the HMC703 and any ideas why an active filter design would not have been used to take advantage of it?

Datasheets:
ADF4001 - ADC PLL
http://www.analog.com/media/en/technical-documentation/data-sheets/ADF4001.pdf
ADF4002 - replacement for above
http://www.analog.com/media/en/technical-documentation/data-sheets/ADF4002.pdf

HMC835 - 2nd LO and TG
http://www.analog.com/media/en/technical-documentation/data-sheets/hmc835.pdf
HMC703 - 1st LO
http://www.analog.com/media/en/technical-documentation/data-sheets/hmc703.pdf

[G0HZU]

The PD Freq for the 200MHz example in the ADF4001 datasheet is only 200kHz. It should be possible to run at a much higher PD Freq and I'd expect the SSA to run it at 10MHz to get much lower phase noise inside the loop bandwidth? But that is just a guess.

[G0HZU]

I stuck in the numbers to ADI SIMPLL for the Zcomm VCO used for LO1 and the 703PLL chip and also a typical 10MHz reference and I got the PN response below for LO1:

This shows the noise contribution of all the major components in the first LO and the (hand drawn) green trace shows the 6dB degradation in phase noise due to the external doubler. I had to add this in by hand and it is the top 'Total' trace in ADI SIMPLL lifted by 6dB.

I chose a loop BW of 80kHz and an active loop filter using a low noise opamp and a phase detect frequency of 10MHz . But it may run the PD at a higher frequency than this. I'm just guessing but the results seem to agree with your plots at the lower end of the analyser range.

[G0HZU]

Lowering the loop BW a bit and choosing a lower noise opamp (in the PLL filter) gets it closer? It would help if we knew the PD frequency and maybe a few other things. But the plot below shows a ballpark phase noise response using the same major parts as the LO1 design in the Siglent SSA.

[hendorog]

Yes that is looking very close. Simulation matching reality!

[hendorog]

The cleanest regular analyser I have here for 100kHz offsets is my fairly ancient Advantest TR4172. It manages -128dBc/Hz at 100kHz offset across HF/VHF and up to 1GHz. The only spectrum analyser I have that beats the SignalHound at close offsets down at 20MHz is my 8GHz Tek RSA3408A. It beats it by maybe 15-20dB at offsets below 10kHz.

This might be of interest - I have been experimenting with 'reducing' the phase noise floor of the SignalHound.

The one which appears to work is NFE, basically just measuring the noise floor and subtracting it from the measurement - as described here: http://literature.cdn.keysight.com/litweb/pdf/5966-4008E.pdf

I have applied this to phase noise by measuring a very clean signal and then subtracting that trace from subsequent measurements. It provides about a 10dB improvement in the PN floor.

The process is basically:

Record baseline sweep of clean signal with averaging set to 1000. The Sig gen was set to Noise mode 1 which is its lowest PN mode.
Convert to dBc/Hz => dB - 17.5 for RBW of 100Hz
Convert to mW => 10 ^ (mW/10)
Call this baseline

Record actual measurement with averaging set to 1000 - I switched the noise mode of the Sig Gen to 'Normal' which is not as clean to provide the measurement.
Convert to dBc/Hz => dB - 17.5 for RBW of 100Hz
Convert to mW => 10 ^ (mW/10)
Call this mn

Calculate corrected value in mW => abs(m1 - baseline)
Convert to dBc/Hz => 10 * log(mW)
Call this mn_adjusted

Chart frequency offset vs mn_adjusted

Here is a google spreadsheet template which does this. It is view only, but you should be able to create a copy of this for yourself (File menu, Make a Copy) and then plug your own numbers into the Baseline and Measurement sheets. 

https://docs.google.com/spreadsheets/d/18VlzMaz6_csRroIl1bk0MpA9q1dbmOahmxpGaENg_uw/edit?usp=sharing


This image shows the results I got measuring the 2041 in Normal noise mode with and without this:
https://docs.google.com/spreadsheets/d/e/2PACX-1vRdEo1EnRjM5oINIitp9IARpjr7TZMN9ZJw90Clzg3Rpl_2rZZi9MDotEedda1psExLc_pL910KjQDz/pubchart?oid=680692226&format=interactive