Phase noise numbers are only meaningful together with how they were measured. The three common methods have very different floors and failure modes.

1. Direct spectrum analyzer method
The quickest method: look at the carrier and read the noise at an offset.
L(f) [dBc/Hz] = P_noise(f) [dBm in RBW] - P_carrier [dBm] - 10 log10(RBW)
Example: carrier 0 dBm, noise reading at 10 kHz offset -80 dBm in a 1 kHz RBW gives -80 - 0 - 30 = -110 dBc/Hz.
Two corrections matter when doing it by hand: the noise bandwidth of the RBW filter is not exactly the nominal RBW, and averaging a log-detected noise signal reads low by about 2.5 dB. Use the analyzer's noise marker or phase noise personality, which applies both.
Limitations:
- The analyzer's own LO phase noise is the floor. Check the datasheet curve at your offset; you need the DUT at least 10 dB above it.
- It measures total noise, AM plus PM. For most oscillators PM dominates close in, but far out they can be comparable.
- A drifting source smears the close-in skirt. Free-running VCOs often cannot be measured below a few kHz offset this way.

2. Phase detector (quadrature) method
The DUT and a low-noise reference drive a mixer used as a phase detector. A slow PLL holds them in quadrature (90 deg), where the mixer output is proportional to the phase difference: v = K_phi * delta_phi. The baseband output goes to an FFT analyzer, and after calibrating K_phi you read S_phi(f) and L(f) = S_phi(f)/2 directly.
Advantages: the mixer suppresses AM noise, and the floor is set by the reference and the baseband noise, not by an analyzer LO.
The catch: you measure the sum of both sources. If the reference is equal to the DUT, the result is 3 dB worse than the DUT. With the reference 6 dB better the error is about 1 dB, with 10 dB about 0.4 dB, and with 20 dB about 0.04 dB. Also, the PLL suppresses noise inside its loop bandwidth, so results below the loop bandwidth need correction or a narrower loop.

3. Cross-correlation
Two independent phase-detector channels, each with its own reference, measure the same DUT. The DUT noise is common to both channels; each reference and channel noise is not. Averaging the cross-spectrum over N correlations reduces the uncorrelated part by about 5 log10(N) dB: 10 dB for 100 correlations, 20 dB for 10,000. This is how commercial analyzers measure sources that are better than any single reference available.
Watch out for:
- Measurement time: each extra 10 dB costs 100 times more averaging, painful at low offsets.
- Correlated contamination: shared supplies, a shared reference clock or crosstalk between channels does not average out and sets a hard floor.
- Cross-spectrum artefacts: with insufficient averaging the estimate can show dips or collapse below the true value at some offsets. Trust the result only where it has stopped changing with more averaging.

4. From L(f) to jitter
The rms phase deviation is phi_rms = sqrt(2 * integral of 10^(L(f)/10) df) radians, and rms jitter = phi_rms / (2 pi f0).
Example: a flat -120 dBc/Hz from 10 kHz to 10 MHz gives phi_rms = 4.47 mrad (0.256 deg); at f0 = 1 GHz that is 0.71 ps rms. Always state the integration limits with a jitter number, since they change it a lot.

5. Bench checklist
- Spurs are not noise. Report them separately in dBc, not dBc/Hz.
- Power supply ripple and ground loops appear as spurs at 50/60 Hz and harmonics, and switching-regulator spurs at tens to hundreds of kHz. Try a battery or linear supply before blaming the oscillator.
- Use enough input level for the instrument floor but stay below its compression point.
- Measure the instrument floor with its own reference or a known clean source on the same day.

If two labs disagree by 5 dB on the same oscillator, the first question to ask is which method and which reference each of them used.