Measuring phase noise on a budget: spectrum analyzer method vs. dedicated test set, and where each breaks down

Phase noise can make or break a synthesizer design, but not everyone has a dedicated phase noise analyzer on the bench. Here's my practical take on what a spectrum analyzer can and can't do.

The SA method: carrier at center, narrow RBW, measure noise power in a 1 Hz-normalized bandwidth at the offset of interest, subtract carrier power. Simple on paper. In practice:

1. The SA's own LO phase noise sets the floor. If your DUT is within ~10 dB of the SA's own noise, you're measuring the analyzer, not the DUT. Check the SA datasheet first — this kills more measurements than any other mistake.
2. AM noise leaks in. The SA can't separate AM from PM noise. If your source has significant AM noise (many PLLs do at close-in offsets), your "phase noise" number is pessimistic. A quick sanity check: AM-detector measurement or compare against a known-clean reference.
3. Close-in offsets (<1 kHz) are where SAs struggle most — 1/f noise of the SA's own LO dominates.

When to rent or borrow a dedicated test set: characterizing OCXOs, low-noise DROs, or any time you need close-in data you can actually trust. Cross-correlation instruments buy you 10–20 dB of extra sensitivity — worth it for a week-long rental when a design decision hangs on the number.

Practical tips that helped me:
- Always measure with the DUT at the same drive level you'll use in the system; PN is level-dependent.
- Warm up both instruments for at least 30 minutes.
- Log the SA's own noise floor measurement alongside every DUT plot — future-you will thank present-you.

What's the lowest phase noise you've credibly measured with just a spectrum analyzer, and how did you convince yourself the number was real?

A practical summary, RF_Bench_Notes — point 1 (measuring the analyser instead of the DUT) is the most common phase-noise mistake we see, so here are the numbers that go with it.

1. How close to the floor is too close? The analyser and DUT noise add in power. If the DUT is 10 dB above the analyser floor, the reading is high by 10·log(1 + 0.1) ≈ 0.41 dB. At 3 dB above it is high by about 1.76 dB. So "10 dB margin" really means "about 0.4 dB of error", not zero.
2. Normalisation. L(f) = noise power in 1 Hz − carrier power. If you use a plain marker rather than the analyser's noise-marker function, correct for the RBW filter's noise bandwidth (around 1.06–1.2 × RBW depending on the filter shape) and for log averaging, which under-reads random noise by about 2.5 dB. Modern noise markers apply both corrections, but older instruments and manual methods do not.
3. Drift. A free-running VCO drifts during the sweep, which smears close-in offsets. A phase-detector test set locks to the DUT and tracks it, which is another reason close-in data from a spectrum analyser is hard to trust.
4. Cross-correlation gain grows only as about 5·log10(N): 100 correlations give roughly 10 dB of improvement and 10,000 give roughly 20 dB. That is why deep measurements can take minutes per trace.

A related editorial piece is at https://www.rfforum.com/t/56 if readers want the three methods side by side.

Question for readers: how do you handle discrete spurs in the plot? Do you report them separately in dBc, or let them integrate into your RMS jitter number?

— rfforum 编辑部 / Editorial Team