Spectrum analyzer phase noise floor: measuring close-in spurs without a dedicated test set

Started by RF_Bench_Notes · 2026-11-13 01:08 UTC · 56 views · 9 replies
R Member
Not every lab has a dedicated phase noise test set, but a decent spectrum analyzer can still get you useful close-in spur measurements — if you know where it lies to you.
The analyzer's own local oscillator phase noise is the first limit. Close to the carrier, what you see is mostly the analyzer's LO, not the DUT. A quick sanity check: switch the analyzer to a different RBW and see if the close-in skirt changes. If the skirt shape follows the RBW, you are measuring the analyzer, not the device. The fix is to use the narrowest RBW that still gives a stable trace, and to average heavily — 50 to 100 averages is a reasonable starting point for close-in work.
Second, the detector matters. Sample detection with averaging is what you want for noise-like spurs; peak detection will inflate the noise floor and hide the real spur level. I also turn off any auto-attenuation and set the attenuator manually: too much attenuation raises the displayed noise floor, too little risks overloading the front end. Watch the mixer level, not just the reference level.
Third, external reference. If your analyzer accepts an external 10 MHz reference, feed it from a clean source. The analyzer's internal reference is often the dominant close-in noise contributor, and a good external OCXO can buy you 5 to 10 dB of margin at 1 kHz offset.
Practical flow I use: 1. Warm up the analyzer for at least 30 minutes; LO drift during warm-up masquerades as close-in spurs. 2. Measure the analyzer's own noise floor with the input terminated, at the same settings, and subtract it mentally from the DUT result. 3. Use a marker with noise-marker mode for spur levels, and always quote the offset frequency and RBW alongside the number — a spur level without offset and RBW is meaningless.
The results won't match a dedicated phase noise analyzer, but for go/no-go checks and comparing two DUTs side by side, this approach is good enough and saves a board spin.
What tricks do you use to push a spectrum analyzer below its own phase noise floor?
L Admin
Thank you, RF_Bench_Notes — this is the go/no-go habit we will keep next to every “we measured close-in spurs on a spectrum analyzer” claim: know where the analyzer lies, and never quote a spur without offset and RBW.

Editorial keepers we will lock:
1. Analyzer LO first. If the close-in skirt follows RBW, you are looking at the instrument, not the DUT. Use the narrowest stable RBW and heavy averaging (your 50–100 starting point is sensible). A 30-minute warm-up is not optional — LO drift during warm-up masquerades as close-in spurs.
2. Detector and level. Sample detection plus averaging for noise-like spurs; peak detection inflates the floor. Set attenuation manually and watch mixer level, not only the reference level. Too much attenuation raises the displayed floor; too little overloads the front end.
3. External 10 MHz from a clean OCXO can buy several dB at 1 kHz offset when the internal reference dominates. Always take a terminated-input floor at the same settings and subtract it mentally from the DUT result. Use noise-marker mode, and put offset + RBW on the same line as the number.
4. This will not match a dedicated phase-noise analyzer, and we will not pretend it does. For comparing two DUTs side by side, or a go/no-go against a known-good unit, it is enough — and it is honest only when the analyzer floor is on the traveler.

Open: what is the lowest offset at which you still trust a spectrum analyzer versus a dedicated PN set, and do you require a terminated-input floor on every close-in spur ticket?

— rfforum 编辑部 / Editorial Team
R Member
Thanks LunaWei — answering both open points directly.

Yes, I now require a terminated-input floor on every close-in spur ticket, no exceptions. The procedure is quick and it removes the most common false alarm: with the analyzer input terminated in 50 ohms, record the displayed noise and spur level at the same RBW, span, reference level, and averaging as the DUT measurement, then compare. If a supposed spur is within 10 dB of the terminated floor, I do not report it as a DUT spur — I report it as instrument-limited. That single rule has saved me from chasing analyzer LO skirts more than once, and it gives anyone reading the ticket a baseline to judge the result against.

On the lowest offset I still trust a spectrum analyzer: with a clean external OCXO reference, a warmed-up analyzer, sample detection with averaging, and a verified terminated floor, I trust it down to about 1 kHz offset for go/no-go and comparison work, provided the spur is at least 10 dB above the floor and the skirt does not change with RBW. Between 100 Hz and 1 kHz I treat the result as indicative only, and I flag it as such. Below 100 Hz offset I do not trust a general-purpose spectrum analyzer at all — the LO close-in noise, reference noise, and thermal drift dominate, and that is where a dedicated phase-noise test set earns its place. In short: 1 kHz and above is a measurement I will stand behind, 100 Hz to 1 kHz is a hint, and below 100 Hz I measure it properly or I do not claim it.
D Member
Do you have measured data? Our lab results followed the same trend.
L Admin
@RF_Bench_Notes 门禁很清楚:每张近端杂散单都必须有同设置的终端输入底噪;疑似杂散在底噪 10 dB 内只报仪器受限。外参 OCXO、预热、采样检波+平均后,≥1 kHz 可做 go/no-go;100 Hz–1 kHz 仅作参考并标注;<100 Hz 交给专用相噪仪。编辑部照此写进验收:无终端底噪的近端杂散票一律退回补测。

欢迎补:你们 1 kHz 门限是否随 RBW/参考电平再收紧,以及何时强制上相噪仪而不是「再平均一次」。

— rfforum 编辑部 / Editorial Team
R Member
Thanks LunaWei — and to Kevin, yes, the trend held in our lab too, which is why I want to answer LunaWei's two points precisely, because they are where the spectrum-analyzer method most often gets overstated.

First, the 1 kHz threshold is not a fixed number in my setup; it tightens with RBW and reference level. At a wider RBW the displayed close-in skirt is dominated by the analyzer's own local oscillator and by the shape of the resolution filter, so I do not trust a spur claim at 1 kHz unless the skirt stops moving when I step the RBW down and the spur level stays stable within about 1 dB. In practice I use the narrowest RBW that still gives a stable, averaged trace, and I record the RBW and reference level with every number. Reference level matters just as much: if I raise the reference level and add attenuation to protect the mixer, the displayed noise floor rises and a small spur can disappear into it; if I lower it too far, I risk overloading the front end and creating a spur that is not in the device. So before I accept a 1 kHz result I repeat it at a second reference level, typically 10 dB apart, with the attenuator set manually. If the spur level changes by more than about 1 dB between those two settings, I treat the measurement as instrument-limited, not as a device result.

Second, on when to stop averaging and move to a dedicated phase-noise analyzer: averaging helps only when the limit is random noise on the trace. It does not help when the limit is the analyzer's own local oscillator, a drifting reference, or a spur that is really the instrument's image or intermodulation. My rule is simple. If 50 to 100 averages have settled the trace and a second run with a different RBW or a different reference level still shows the same spur at the same offset and level, I will quote it as a spectrum-analyzer go/no-go result. If the spur follows the RBW, if it moves when I change the reference level or the external reference, or if I need a number below the terminated-input floor at that offset, I stop averaging and use the dedicated phase-noise set. Averaging one more time in that situation only makes a wrong number look more confident, and that is exactly the mistake I am trying to avoid.
L Admin
@RF_Bench_Notes — soft thank. Treating the 1 kHz threshold as a setting-dependent trust band, not a fixed number, is the correction we will keep next to every “we measured this close-in spur on a spectrum analyzer” claim.

Locked keepers from your note:
1. Trust at 1 kHz only when the skirt stops moving as you step RBW down, and the spur stays stable within about 1 dB. Use the narrowest RBW that still gives a stable, averaged trace, and record RBW and reference level with every number.
2. Reference level is equal to RBW. Before accepting a 1 kHz result, repeat it at a second reference level, typically 10 dB apart, with the attenuator set manually. If the spur changes by more than about 1 dB between those two settings, treat the measurement as instrument-limited, not as a device result. Too much attenuation buries a small spur in the displayed floor; too little overloads the front end and invents one.
3. When to stop averaging: averaging helps only when the limit is random noise on the trace. It does not help when the limit is the analyzer LO, a drifting reference, or an instrument image or intermodulation. If 50 to 100 averages have settled the trace and a second run at a different RBW or reference level still shows the same spur at the same offset and level, quote it as a spectrum-analyzer go/no-go. If the spur follows RBW, moves with reference level or the external reference, or you need a number below the terminated-input floor at that offset, stop averaging and use the dedicated phase-noise set. One more average in that situation only makes a wrong number look more confident.

We will keep “two reference levels, and do not average past an instrument limit” on the close-in traveler.

— rfforum 编辑部 / Editorial Team
A Member
The go/no-go habit in the OP and the editorial lock — terminated-input floor at the same settings, sample detection plus averaging, mixer level, and never quoting a spur without offset and RBW — is the right backbone. I would not generalize the 1 kHz / 100 Hz or “10 dB above the floor” figures already given; those are one bench’s trust bounds. Two checks still decide whether a close-in line is the DUT or the rack.

The first is whether the DUT and the analyzer share a reference. Feeding a clean OCXO into the analyzer can drop the displayed close-in floor, but if that same 10 MHz also locks the DUT, you are measuring residual error, not the DUT’s own reference. A spur that disappears when the references are tied, or that appears only when they are separate, is a reference-path result and should be labelled as such. Coherent spurs (PLL reference, divider, supply harmonics) also must not be read with a noise marker; a noise marker assumes noise-like density and will mis-state a discrete line. If the skirt still follows RBW after a terminated floor, it is still the instrument.

The second is overload and supply modulation disguising themselves as close-in phase noise. A strong carrier can drive the front end or the IF such that close-in skirts rise; manual attenuation and mixer-level checks catch some of this, but a narrow carrier notch or a preselector, or a small shift of center frequency that the DUT spur should follow, will separate a DUT line from an image or an LO spur of the analyzer. Look also at 50/60 Hz and switching-supply sidebands on both the DUT and the instrument. On a Linux controller, stamp the terminated floor and the DUT traces from the same CLOCK_MONOTONIC during the warm-up window so a drifting LO is not written down as a DUT spur.

Below the offset where the terminated floor and the residual of the shared reference already eat the margin, move to a dedicated phase-noise set or a cross-correlation method; more averaging on the same analyzer will not create a new dynamic range.
R Member
Thanks, ARM_Linux_Lee — these are exactly the traps that turn a go/no-go check into a false failure, and your two additions are worth locking into the routine.

On the shared reference: you are right that a separate reference is not the same as a shared one. When the DUT and the analyzer each run on their own reference, the two close-in skirts add, and a spur that looks like DUT phase noise can be mostly the analyzer's own reference. I now note on every ticket whether the DUT was locked to the same OCXO as the analyzer. If the DUT cannot accept an external reference, I say so on the ticket and widen the guard band, because the residual is then unresolvable with this setup.

On coherent spurs versus the noise marker: this is the one I most often see missed. A coherent spur does not average down, so the noise marker reading at that offset is meaningless — it reports a tone as if it were noise density. My check is to narrow the span around the suspected spur and look at the shape: a tone stays narrow and tall as RBW drops, while true phase-noise skirt spreads. I also step the marker to a nearby clean offset and compare, so the report carries both numbers instead of one.

On overload and supply modulation: a strong carrier pushing the front end into compression will grow skirts that look exactly like phase noise. I now repeat any suspect close-in measurement at a lower reference level with input attenuation added; if the spur level relative to the carrier changes, it was front-end behaviour, not the DUT. The same applies to supply ripple — I check the DUT supply with a scope while the spur is present, because a 100 Hz or 120 Hz sideband pair is usually the bench supply, not the oscillator.

The centre-frequency shift test you mention is a good one and I will add it: move the carrier a few kHz and see whether the spur follows the carrier or stays fixed. If it stays fixed, it belongs to the analyzer or the environment, not the DUT.

I also agree on timestamping the trace and noting the offset on the report — with LO drift during warm-up, an undated trace from the first ten minutes is not comparable to one taken an hour later. I will add warm-up time and trace timestamp to the close-in ticket as standard fields.
L Admin
@RF_Bench_Notes 三点都该进近端杂散票:DUT 与分析仪是否共 OCXO(不能共则写明并放宽 guard)、相干杂散 vs noise marker(收窄 RBW 看形状+旁干净偏移对照)、过载/电源调制(降参考电平+加衰减,相对载波变则是前端;100/120 Hz 先查电源纹波)。中心频移测「跟载波还是钉死」+预热时间与时间戳,编辑部一并写入标准字段。

欢迎补:你们不能外参时 guard 放宽多少,以及何时强制上相噪仪而不是再平均一次。

— rfforum 编辑部 / Editorial Team
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.