How to Measure OSNR: Interpolation, Integration and In-Band Methods

Scope

This article gives the procedure for measuring the optical signal-to-noise ratio (OSNR) of a channel in an amplified link with a grating-based optical spectrum analyzer (OSA). It covers the reference-bandwidth convention, the out-of-band interpolation method of IEC 61280-2-9 and the corrections it needs, the integration required for channels wider than the reference bandwidth, and the in-band methods used when the noise between channels no longer represents the noise inside them. OSA architecture and its specifications are covered in Optical Spectrum Analyzer Fundamentals; how OSNR accumulates along a chain of amplifiers is covered in the noise figure and amplified spontaneous emission entries, and the choice of amplifier in SOA vs EDFA vs Raman.

Definition and reference bandwidth

OSNR is the ratio of the channel's signal power to the noise power in a fixed reference bandwidth $B_\text{ref}$, normally 0.1 nm:

$$\text{OSNR} \;=\; 10\log_{10}\!\left(\frac{P_\text{sig}}{N_\text{ref}}\right),$$

where $N_\text{ref}$ is the ASE power in $B_\text{ref}$ at the channel's wavelength, including both polarizations. At 1550 nm, 0.1 nm corresponds to 12.48 GHz. The signal power is the whole power of the channel, however wide it is; the noise is always referred to 0.1 nm. OSNR is therefore not the signal-to-noise ratio inside the channel, and a reported value without its reference bandwidth cannot be compared with another.

Equipment

The OSA's dynamic range at a small offset from a strong carrier, set by stray light in the monochromator, limits the highest OSNR that can be measured; an OSA whose stray-light floor sits 40 dB below the carrier at the measurement offset cannot report 45 dB.

Procedure: out-of-band interpolation

This is the method of IEC 61280-2-9 and the default on most OSAs. It assumes the ASE spectrum is smooth, so that the noise under the channel can be estimated from the noise beside it.

1. Set the OSA

Set the span to include the channel and its neighbors, and the RBW to 0.1 nm or less. The RBW must be narrow enough that the filter's skirt, at the offset where the noise will be read, has fallen below the noise floor; for channels on a 50 GHz grid, 0.4 nm apart, that forces 0.1 nm or finer. Use enough averaging or video bandwidth reduction that the noise trace is smooth to a few tenths of a dB.

2. Measure the signal power

For a channel narrower than the RBW (a 10 Gb/s intensity-modulated channel, for example), read the peak. For a channel wider than the RBW, integrate the spectral density across the channel's full width, since the peak reading captures only a slice. A 64 GBd coherent channel is about 0.51 nm wide at 1550 nm, five times the reference bandwidth. Most OSAs offer this as a "channel integration" or "WDM analysis" mode with a set integration width; set it to the channel's occupied bandwidth or its grid slot.

3. Measure the noise beside the channel

Read the noise level on each side of the channel at an offset where the channel's own spectrum has fallen well below the floor, usually half the channel spacing. Interpolate linearly between the two readings to the channel's center wavelength.

4. Correct the noise to the reference bandwidth

The noise reading is in the OSA's actual noise-equivalent bandwidth, not in 0.1 nm. The correction is

$$N_\text{ref} \;=\; N_\text{meas} + 10\log_{10}\!\left(\frac{B_\text{ref}}{B_\text{eff}}\right),$$

where $B_\text{eff}$ is the effective noise bandwidth of the RBW filter. It differs from the nominal RBW setting because the filter shape is not rectangular; the instrument's calibration gives it. A nominal 0.1 nm setting with a 0.07 nm effective noise bandwidth reads noise 1.55 dB low, and an uncorrected OSNR is too high by the same amount.

5. Subtract the noise from the signal

The integrated channel power in step 2 includes the noise within the integration width. For an integration width $B_\text{ch}$,

$$P_\text{sig} \;=\; P_\text{tot} - N_\text{ref}\,\frac{B_\text{ch}}{B_\text{ref}}$$

in linear units. Worked example for a 75 GHz slot at 1550 nm with a total integrated power of −10.0 dBm:

The correction is negligible at high OSNR and approaches a dB near the operating points of coherent receivers, where it matters most.

6. Record the conditions

Report OSNR with its reference bandwidth, the RBW and effective bandwidth used, the integration width, and the interpolation offset. Without these, a later measurement of the same link cannot be compared.

When interpolation fails

The interpolation method rests on an assumption that modern networks break in three ways.

Filtered noise. A channel that has passed through ROADMs or other wavelength-selective filters after the last amplifier has had the noise between channels cut away, while the noise inside the passband remains. The interpolated floor is then too low and the reported OSNR too high, by as much as the filter's rejection.

No gap between channels. Coherent channels at 60 GBd and above on a 75 GHz grid, and flexible-grid channels packed edge to edge, leave no wavelength at which the noise floor can be seen between them. The OSA reads the skirts of the neighboring channels instead.

Spectral shaping. Channels shaped with a raised-cosine filter have a flat top and steep edges; there is no spectral region inside the channel's own slot that is free of signal.

In each case the error has one sign: the method overestimates OSNR. A link that measures comfortably by interpolation but runs at an unexpectedly high pre-FEC error rate is the usual symptom.

In-band methods

In-band methods estimate the noise inside the channel itself.

Channel-off. Turn the channel off at its transmitter and measure the noise at its wavelength directly. The method is exact if the amplifiers' operating point does not change when the channel is removed, which holds for a lightly loaded amplifier in constant-gain mode and fails for one in constant-output-power mode, where the remaining channels and the ASE rise to fill the gap. It also takes the channel out of service.

Polarization nulling. For a single-polarization signal, a polarization controller and polarizer ahead of the OSA are adjusted to extinguish the signal, leaving half of the unpolarized ASE; the noise in the channel is twice that reading. The method fails for polarization-multiplexed coherent signals, which fill both polarizations, and is disturbed by polarization-mode dispersion that depolarizes a signal across its spectrum.

Reference-spectrum subtraction. The channel's spectrum is recorded at the transmitter, where the noise is negligible, and scaled to match the measured spectrum at the point where signal dominates; the difference between the two is the noise. It works for polarization-multiplexed signals and filtered channels, at the cost of access to the transmitter and the assumption that the signal's spectral shape has not changed along the way.

Receiver estimates. A coherent receiver's DSP reports an estimate derived from the constellation, which includes nonlinear noise and transceiver noise as well as ASE (see What the Coherent DSP Actually Does). It is the quantity that predicts the error rate, and it is not the same quantity as the ASE-only OSNR an OSA measures; the two should not be compared without saying which is which.

Verification

Check the OSA's reading against a known case before trusting it on a link. A transmitter connected directly to the OSA should report an OSNR limited only by its own noise and the OSA's dynamic range, typically well above 35 dB. Adding a known ASE source through a coupler at a measured power should move the reported OSNR by the amount calculated from that power. For the interpolation method, the noise readings on the two sides of a channel should agree within about a dB on a flat ASE spectrum; a larger difference means the floor is tilted or one side is reading a neighbor's skirt.

Common failure modes

Nominal RBW used as the noise bandwidth. The OSNR is high by the ratio of nominal to effective bandwidth, often more than a dB.

Peak reading on a wide channel. The signal power is underestimated by roughly the ratio of channel width to RBW, and OSNR is low by as much as 7 dB for a 0.51 nm channel read in 0.1 nm.

Noise read on a neighbor's skirt. The noise is overestimated and OSNR underestimated; move the offset or narrow the RBW.

Interpolation after a filter. OSNR is overestimated, as described above; use an in-band method.

Reference bandwidth unstated. An OSNR in 12.5 GHz, in 0.1 nm, and in the signal bandwidth are three different numbers; the conversion is a bandwidth ratio in dB.

References: IEC 61280-2-9, Fibre optic communication subsystem test procedures, Part 2-9: Digital systems, Optical signal-to-noise ratio measurement for dense wavelength-division multiplexed systems; ITU-T Recommendation G.697, Optical monitoring for dense wavelength division multiplexing systems; D. Derickson (ed.), Fiber Optic Test and Measurement (Prentice Hall, 1998), chapters on optical spectrum analysis and optical amplifier testing.