The usual explanation is that light bounces along the inside of a glass strand by total internal reflection. That is adequate for thick multimode fiber and misleading for the single mode fiber carrying essentially all long distance traffic, where the core is a few wavelengths across and a bouncing ray describes nothing physical.
What limits a real link is not whether light stays in the glass. It is loss, dispersion, and nonlinearity, and each sets a different ceiling.
Why light stays in the core
A fiber is a core surrounded by cladding of slightly lower refractive index, typically by less than one percent. Light crossing into a lower index bends away from the normal, and beyond a critical angle it does not cross at all, so rays entering within a certain cone stay trapped.
The size of that cone is the numerical aperture, fixed by the index difference. A larger difference accepts more light and tolerates sloppier alignment, which is why multimode fiber is cheaper to terminate, and it makes the fiber worse at everything else.
Modal dispersion and the end of ray optics
In a wide core light follows many distinct paths. A ray down the axis covers less distance than one bouncing steeply along the boundary, so they arrive at different times, and a pulse launched as a sharp edge emerges smeared. Beyond some distance and bit rate the smearing runs adjacent pulses together.
This is modal dispersion, and it limits step index multimode fiber to short reach regardless of available power.
Graded index fiber varies the refractive index smoothly across the core rather than in a step. Light travels faster where the index is lower, so a ray wandering toward the cladding covers its longer path at higher speed, and with a roughly parabolic profile the transit times nearly equalise. It is an elegant fix and still a fix, because multiple paths continue to exist.
Single mode removes the paths
Single mode fiber removes the paths instead. Shrink the core to around eight or nine micrometres, comparable to the wavelength carried, and only one guided mode propagates. At that scale the ray picture stops meaning anything and the field must be treated as a waveguide solution. Whether a fiber is single moded depends on core radius, index difference, and wavelength together, combined into one normalised parameter that must stay below approximately 2.405.
The same fiber that is single moded at 1550 nm supports several modes at a short enough wavelength.
Where the light goes
Attenuation in silica is not flat across wavelength, and the shape of that curve decided which wavelengths the industry standardised on. At short wavelengths the dominant loss is Rayleigh scattering from density variations frozen into the glass as it cooled, and it falls steeply, scaling with the inverse fourth power of wavelength.
At long wavelengths loss rises again as the silica lattice begins absorbing in the infrared. Between them is a minimum. Older fiber also carried a pronounced absorption peak near 1383 nm from residual hydroxyl contamination, and removing it opened a previously unusable band.
The result is two operating windows. Around 1310 nm loss is roughly 0.32 to 0.35 decibels per kilometre. Around 1550 nm it reaches the minimum at roughly 0.17 to 0.2. That difference sounds modest until multiplied by distance, where it becomes the difference between spending half a power budget and spending nearly all of it.
The two windows compared
| Window | Loss per km | Dispersion | Amplification |
|---|---|---|---|
| 1310 nm | 0.32 to 0.35 dB | Near zero | Not available |
| 1550 nm | 0.17 to 0.2 dB | About 17 ps/nm/km | Erbium doped fiber |
Two windows, one conflict
Standard single mode fiber has a second property that does not align with the first. Chromatic dispersion passes through zero near 1310 nm and rises to roughly seventeen picoseconds per nanometre per kilometre at 1550 nm. The wavelength with the lowest loss is therefore not the wavelength with the lowest dispersion, and a designer chooses which limit to fight.
No transmitter is perfectly monochromatic and modulation broadens its spectrum further, so chromatic dispersion spreads those components apart in time and the pulse widens exactly as under modal dispersion, from a different cause. For decades this drove elaborate countermeasures, including fiber engineered to move the dispersion zero and spools of deliberately opposite dispersion inserted to cancel the accumulated effect.
Polarisation mode dispersion adds a subtler limit, because a real fiber is never perfectly circular or free of stress, so the two polarisation states travel at fractionally different speeds, and that difference varies randomly along the fiber and drifts with temperature.
When more power stops helping
Silica is very nearly linear, but not entirely. At the intensities present in a small core carrying amplified signals the refractive index itself varies slightly with intensity. Self phase modulation shifts a pulse's own spectrum as its intensity rises and falls, cross phase modulation lets one wavelength disturb its neighbours through the same mechanism, and four wave mixing generates new frequencies from combinations of existing ones that arrive as interference no filter can remove.
The structural consequence is that optical links have an optimum launch power rather than a maximum. Below it performance is limited by noise, above it by nonlinearity, and raising power past the optimum makes the link worse. That is not intuitive for anyone used to electrical systems where more signal is generally better.
Amplification, and then digital correction
The decisive advantage of the 1550 nm window was not loss alone. Fiber doped with erbium and pumped by a laser amplifies a passing signal directly in the optical domain across a window covering roughly 1530 to 1565 nm, without converting anything to electrical form. Before that, distance meant regeneration, which meant detecting and retransmitting every channel at every site.
Optical amplification handles all wavelengths at once and is indifferent to modulation format, so capacity can be raised later by changing only the terminal equipment, and dense wavelength division multiplexing became viable because the amplifier did not care how many channels passed through it.
Coherent detection changed the rules
Coherent detection also freed the modulation format. Once amplitude and phase are both available, systems can carry many bits per symbol using quadrature amplitude modulation, trading the signal to noise ratio a link requires against the spectral efficiency it achieves, chosen according to the reach actually needed rather than fixed at design time.
Contemporary high rate systems then added coherent detection. Rather than measuring received intensity, the receiver mixes the incoming signal against a local laser and recovers amplitude and phase on both polarisations, then processes the result digitally. Once the full field exists in the digital domain, chromatic dispersion becomes a linear distortion that a filter can undo and polarisation effects can be tracked continuously.
This inverted decades of practice, because optical dispersion compensation is now usually omitted, since leaving the dispersion in place and correcting it electrically works better and costs less.
The failures that actually happen
Everything above concerns the fiber. In operational networks the losses that cause problems are at the ends of it. A fusion splice contributes a small fraction of a decibel, a mated connector pair contributes more, and its loss depends on alignment, end face geometry, and above all cleanliness.
A single particle sits directly in a beam a few micrometres across and can add several decibels or reflect power back toward the transmitter. Reflections destabilise laser sources and circulate power through amplified systems in ways the design did not anticipate, which is why angled end faces exist and why unused ports are capped.
The physics of the glass is largely solved. The physics of a dirty ferrule is what takes links down.
Note: light in silica travels at roughly two thirds of its vacuum speed, close to five microseconds per kilometre. No optical engineering changes that, which is why intercontinental latency is a matter of route length rather than equipment.