A fiber ships with an insertion loss printed on its test card. Six months later the same fiber is measured again and the number is worse. Nobody dropped it. Nobody repaired it. The console has not moved and the fiber has been in the same room the whole time. What changed is the interface, and the reason it changed is that alignment is not a property of a component — it is a property of a relationship, one that gets re-negotiated every time a connector is unscrewed and seated again.
That distinction explains most of the surprises in this subject. A fiber’s core diameter is fixed when the preform is drawn, and nothing done to it later will alter it. The connector’s alignment, by contrast, is re-decided at every mating, by two eccentrics that are free to rotate relative to each other, across a set of tolerances that all stack in the same direction when you are unlucky. This article is about where those tolerances come from, what each one costs, and why the same micrometer of error is a rounding error on one fiber and a ruined link on another.
In telecom a decibel is information. In power delivery it is heat
Start with the unit itself, because the same number means two different things depending on what is flowing through the fiber.
On a telecom link carrying milliwatts, 0.5 dB of alignment loss means a slightly weaker signal and a receiver that has to work a little harder. Nothing heats up, because there is nothing to heat. On a power delivery fiber, that decibel is a fraction of the energy that did not make it down the fiber — and energy that does not leave through the tip has to go somewhere. It goes into the interface.
The arithmetic is unforgiving and it scales linearly with power. On a 30 W delivery, 0.5 dB is 3.3 W dissipated at the joint. On the 120 W class used for holmium work, the same 0.5 dB is over 13 W. Put that in a kilowatt-scale fiber laser and 0.5 dB means roughly 100 W dumped into the splice point at once — enough to make the glass glow, soften or fail outright. High-power engineering literature states this plainly: at milliwatt levels a 0.5 dB loss is just a weaker signal, and at kilowatt levels it is a thermal event at a single point.
This is the reason alignment deserves an article of its own in a medical context rather than being treated as an installation detail. It is not a question of how much light arrives. It is a question of where the light that does not arrive ends up, and the answer is always the same place: at the interface, as heat, a few millimeters from the connector body that is supposed to be carrying that heat away.
Alignment loss is three different geometries, not one
“Poor alignment” gets used as a single phrase, which hides the fact that three distinct mechanical errors produce three distinct optical penalties, with different sensitivities and different remedies.
| Error | What is misaligned | How loss scales | In high-power medical delivery |
| Lateral offset | The two core centers sit off-axis from each other | Quadratic in offset relative to the mode field, for a small-core fiber | The dominant term, but heavily discounted by large cores |
| Angular tilt | The two fiber axes meet at an angle | Rises with the square of the angle | Set by cleave and polish angle; small angles matter |
| Longitudinal gap | The two end faces do not touch | Fixed Fresnel penalty per glass-to-air interface | Deliberately accepted — the air gap |
The third one deserves an immediate clarification, because it is the one that gets misread as a defect. In a small-core telecom joint, a gap between the faces is a failure: it adds loss and it turns the joint into a reflector. In a high-power medical termination, the gap is a specified design feature, chosen so that nothing absorbing sits in front of the fiber facet. It is not a tolerance that slipped. Our own bare fiber is offered in both flat-polished and air-gap versions, which is the clearest way to say it: the gap is a variant you order, not a manufacturing accident you tolerate.
So when the total insertion loss of a medical interface is quoted, part of it is a deliberate design cost and part of it is genuine misalignment. Only the second part is what this article is about, and separating them is the first step in knowing whether a number is a problem.
Where lateral offset actually comes from: four tolerances and a sum
Lateral offset is the largest term in most real joints, and it is worth knowing that it is not produced in one place. It is the sum of four separate tolerances, each contributed by a different part of the assembly:
- Ferrule bore eccentricity — the hole drilled through the ferrule is not perfectly concentric with the ferrule’s outer diameter.
- Fiber core concentricity — inside the glass itself, the core is not perfectly centered in the cladding. This one is set at the draw and is a property of the fiber, not the connector.
- Ferrule outer-diameter tolerance — the ferrule that must fit into a sleeve is itself manufactured to a tolerance, and that tolerance is an alignment error budget item.
- Sleeve fit — the alignment sleeve that brings the two ferrules coaxial either holds them tightly or it does not.
Four tolerances, all in the same few-micrometers range, all capable of pointing the same way. This is the reason the industry’s concentricity specifications are quoted in micrometers — a high-power SMA cable specification gives under 6 µm — and the reason a well-made joint is described not as “well aligned” but as “within budget.” When a joint underperforms, the diagnosis is the sum, and the four contributors are rarely equally responsible.
The fourth item on that list has a behavior pattern worth knowing about, because it produces a signature that is easy to misread. A sleeve that has worn or developed a crack no longer holds the ferrules on axis, and the resulting loss is not stable — it changes each time the connector is re-seated. That pattern points at the mating hardware rather than at the connector, and it is one of the few alignment problems in this field that announces itself clearly by being inconsistent.
Why a 400 µm core forgives what a 9 µm core cannot
Now the part that makes the medical case unusual, and it comes down to a comparison of two formulas that describe the same physical event.
For a small-core single-mode fiber, the loss from a lateral offset d is approximated by the Gaussian-beam relation L ≈ 4.34 (d/w)², where w is the mode field radius — half the mode field diameter. For a standard single-mode fiber at 1310 nm, where the mode field diameter is about 9.2 µm and therefore w ≈ 4.6 µm, the numbers work out like this:
| Lateral offset | Loss (small-core single-mode) |
| 0.5 µm | 0.05 dB |
| 1.0 µm | 0.21 dB |
| 1.5 µm | 0.46 dB |
| 2.0 µm | 0.82 dB |
| 3.0 µm | about 1.8 dB — over 30 % of the light misses the core |
| 5.0 µm | over 5 dB — over 70 % lost |
One micrometer. A twentieth of the width of a human hair, an error no machinist would call an error, costs a fifth of a decibel on a 9 µm core. This is why telecom connector geometry is specified to fractions of a micron and why that world needed keys, domed polishes and interferometric grading.
A large-core medical fiber plays a different game. When the core is 400 µm across, a lateral offset does not move light into a cladding mode; it simply removes a sliver from the overlap of two large circles. Estimating that overlap geometrically — the shared area of two circles of equal radius offset by d, as a fraction of the full area — puts the same micrometer of error at a very different cost:
| Lateral offset | Loss (400 µm core, geometric overlap) |
| 1 µm | about 0.014 dB |
| 50 µm | about 0.75 dB |
| 100 µm | about 1.6 dB |
The comparison is the point of the whole article. The same 1 µm of eccentricity is roughly 0.21 dB on a 9 µm core and roughly 0.014 dB on a 400 µm core — about fifteen times cheaper, and it stays cheap out to offsets that would have destroyed a single-mode link twice over. A 50 µm offset, which is an unthinkable assembly error in telecom, costs less than a decibel here.
This is the quantitative foundation under the connector choices described elsewhere on the site. It is why a 200 to 1000 µm bare fiber can be terminated with an unkeyed SMA-905 and still be a sound engineering decision: the alignment penalty that makes an unkeyed connector unacceptable for single-mode fiber is diluted by roughly an order of magnitude at medical core diameters. The connector did not get better. The fiber got tolerant.
The unkeyed penalty, quantified
Which does not mean the penalty is zero, and it is worth being precise about what an unkeyed connector actually costs, because the cost is not what most people assume.
An unkeyed interface allows the ferrule to rotate freely. Since the ferrule bore is eccentric and the fiber core is eccentric within the cladding, the total core offset at any given mating is a vector sum of two eccentricities whose relative orientation depends on how the connector happened to stop when it was tightened. Seat it once and the two eccentricities may partially cancel. Seat it again at a different angle and they add. Nothing about the joint changed; the draw changed.
The consequence is not a large insertion loss. It is an insertion loss that is not repeatable — a joint that measures 0.15 dB one day and 0.4 dB the next with the same parts and the same console. In a telecom link that variation would be intolerable and would be blamed on the connector. In a medical delivery fiber, where a few tenths of a decibel is tolerable in absolute terms, the variation itself becomes the real issue, because it is invisible: nobody notices a 0.2 dB change, and nobody measures it, so the interface drifts quietly while the operator compensates by turning the power up or working more slowly.
That is the mechanism behind the opening scenario. A fiber does not fail. It degrades along a curve, in small increments, at the one point in the optical path that is disturbed routinely. And because the increments are individually unremarkable, the accumulation is what finally shows up.
What actually moves the loss upward in service
Alignment errors are set at assembly, but the loss is written in service, by a short list of mechanisms that all push in the same direction.
Contamination is the most common field cause and the most underrated. A single particle sitting over the core can add anything from a fraction of a decibel to several decibels, and — this is the part that matters — when the connector is mated, that particle is pressed into the glass. The loss it causes can become permanent, and the damage it leaves behind becomes an absorption site for the next session. The interface does not recover.
Sleeve and mating-hardware wear produces the instability described above: loss that changes on every re-seat. It is diagnosed by its inconsistency rather than by its magnitude.
End-face geometry failure — an incorrect radius of curvature, a displaced apex, a fiber that protrudes or sits undercut relative to the ferrule — recreates a gap where there should not be one, which adds the fixed Fresnel penalty on top of the alignment loss. In an interface that has deliberately accepted one gap, an unintended second one is pure loss with none of the thermal benefit.
Repeated mating accumulates all of the above, and in an unkeyed interface it adds rotational abrasion that a keyed connector would not see.
Thermal cycling moves the problem into the metal. A ferrule and its sleeve expand by different amounts at different temperatures, and a joint that was concentric on the bench is not necessarily concentric at operating temperature. Materials chosen for heat dissipation in high-power medical connectors are chosen for exactly this reason — the thermal path is the primary function, and dimensional stability under that thermal load is the constraint that comes with it.