Gain and noise figure: why the datasheet number is not the acceptance number

Your specification says noise figure not exceeding 5.5 dB, and it looked like a watertight line. At acceptance the supplier produces 5.2; the laboratory you appointed measures 5.9. Neither of them is cheating: they measured between different points, with different methods, and no document says which of the two numbers prevails. That line settles nothing: the amplifier stays in the line or goes back depending on who wrote the better contract.

On 14 May 2025 the ITU-T approved the sixth edition of ITU-T G.661, Definitions and test methods for the relevant generic parameters of optical amplifier devices and subsystems: in force, freely downloadable, the first revision since the 2007 edition. Eighteen years, for a device that has meanwhile found its way into every metro backbone and every amplified span handed over with an acceptance record. The problem with your acceptance test sits in three notes nobody reads.

Gain is measured between two points nobody can touch.

G.661 defines gain as the increase in signal optical power from the input reference point R′ to the output reference point S′. Then, in note 2 to clause 3.2.1.1, it adds the sentence that breaks the comparison: “This definition is different from the corresponding IEC definition in that it includes the effects of any connectors directly associated with the OA device at both the input and output.” Clause 3.2 says the same thing the other way round: inside the values of gain and noise figure sit the loss of both pairs of connectors directly associated with the device and the corresponding uncertainty. The connector pairs are not outside the measurement. They are part of the declared figure.

The ITU-T measures with the connectors in; the IEC, which writes the test methods, counts them differently. ITU-T G.665 (11/2025, in force) devotes a whole appendix to the divergence and closes it with a working rule: “Therefore, ITU-T span losses will always differ from the IEC test method losses by the loss of one joint or connector, as appropriate.” Always, not occasionally. A pair of connectors is counted in tenths of a dB: the same order of magnitude as the difference the two parties are arguing about.

And those points R′ and S′, where the number is supposed to be read? ITU-T G.662 (07/2005, in force) calls them virtual R and S points, because the amplifier is neither a regenerator nor a terminal. G.665 puts it more bluntly: “Often, it is a point inside of the fibre which must be cut to explicitly access the interface.” The reference plane of the contractual figure is not a bulkhead you can unscrew.

For noise figure, the reference test method has not been chosen yet.

Clause 7 of G.661 carries a table: for each group of parameters, the applicable IEC test methods. For noise it lists five: optical spectrum analyser (IEC 61290-3-1), electrical spectrum analyser (IEC 61290-3-2) and three multichannel methods in the IEC 61290-10 series. Five routes to a number that appears in the specification exactly once.

Which one prevails? None, for now. Note 2 to clause 7: “The comparative evaluation of the test methods given in the IEC basic specifications is currently under development.” When it becomes available, the note adds, the chosen reference test methods and the possible alternative test methods will be indicated. That is written in a 2025 edition. Until that evaluation arrives, there is no designated reference test method for the noise figure of an optical amplifier: if the specification does not pick one, nobody does.

And those methods, note 1 warns just above, answer to IEC definitions: you can measure with the right instrument, in the right way, and still be measuring a different quantity from the one written into the contract.

The test conditions are half the number.

A value without its conditions, for G.661, is not a parameter; clause 3.2 makes that binding: “The temperature and state of polarization should be kept constant or controlled or be measured and reported together with the measured parameter.” On noise figure it repeats that the operating conditions it is specified at must be stated, and it separates nominal conditions, the ones the manufacturer suggests for normal use, from limiting conditions, with every user-adjustable parameter at its maximum rating.

Then there is multichannel operation, which is to say almost everything. Channel gain, says clause 3.2.1.53, “is dependent on the input power level of all signals”. A gain measured with eight channels lit and a gain measured with forty are not the same gain: an acceptance record that does not state the channel allocation — count, frequencies, powers — cannot be verified by anyone.

Then reflectance: in three separate clauses G.661 repeats that noise figure is the parameter most sensitive to reflectance at the amplifier ports. On site that means a dirty connector end face at the input moves the very number under dispute.

With distributed amplification, the value can be negative.

On Raman systems the trap changes shape. G.661 fixes the convention — by convention the noise figure is a positive number — but for distributed amplifiers it introduces the effective noise figure and warns, in clause 3.2.1.45: “It is thus possible for the effective noise figure of distributed amplification to be negative, expressed in dB.” Anyone rejecting a negative value because it looks impossible is rejecting a correct measurement of a different quantity. Gain works the same way: G.665 separates on-off gain, how much the power grows when the pump is switched on, from net gain, which is the on-off gain minus the loss between the reference points. Twelve dB of on-off gain over a span that loses fifteen makes a negative net gain.

What to write into the specification.

Which definition: ITU-T G.661, connectors included and reference points R′ and S′, or the IEC one. One line, before the price.
Which test method, cited by part number, and the same method for acceptance and for any counter-verification.
The operating conditions: temperature, channel allocation (count, frequencies, input powers), state of polarisation controlled or reported in the record.
With distributed amplification: state whether the required value is on-off or net gain, noise figure or effective noise figure.
Reflectance at the ports and end-face cleanliness recorded before the measurement, not after the dispute.
The uncertainty of the measurement chain, with the calibration certificate attached: without a declared uncertainty, 5.2 and 5.9 are the same number or two different numbers depending on who is speaking.

What we do not know.

We have not read the test methods in the IEC 61290 series: they are paid documents, and we know their numbers and titles only from table 1 of G.661; we do not know what they prescribe, nor which editions are current. We do not know when the comparative evaluation of the methods will arrive: the recommendation gives no date. Compliance with ITU-T Recommendations remains voluntary — “Compliance with this Recommendation is voluntary” — and binds only where the contract calls it up. A warning about editions: G.661 cites G.665 in its 2005 version, but the edition in force is the one approved on 29 November 2025, six months later. The appendix on connectors is still there: we read it there, in the text in force, not in the one cited.

The point.

The point is not whether the amplifier is any good: it is that the acceptance record, as it is written today, does not let you say. When we write or review a specification, definition, method and conditions become contractual lines and a check performed at acceptance, with the measurement report handed over together with the work. And when the amplifiers number in the dozens — the backbone of a carrier, two interconnected data centres, the network of a public body — gain, noise figure and reflectance stop being scattered PDFs and become one map of the network, where an AI sees which amplifier is drifting away from its acceptance value before a channel goes out of range. That AI runs inside the client’s perimeter — on-premise or in a dedicated cloud with data centres in Italy — together with CSIDIA, the other company in the group. Our own crews, no subcontracting.

About to sign off an acceptance record on an amplified span, or having one disputed? Talk to one of our engineers: method and conditions are written before the measurement, not after.