1. Introduction — why physics still governs

Every decision in an optical network — the reach of a wavelength, the modulation order a transponder selects, the margin a planner allocates, the alarm threshold a receiver trips — ultimately resolves to physics. The DSP can do extraordinary things with the complex optical field it receives. It cannot invent photons the fiber threw away, undo a polarization rotation that happened faster than its tracking bandwidth, or unwind a Kerr interaction that has already shifted phase between two channels. Fiber physics sets the ceiling. Every other engineering choice sits below it.

This article is an engineer's working view of that ceiling. It covers the phenomena that actually limit deployed coherent systems — attenuation, chromatic dispersion, PMD, Kerr-effect nonlinearity, ASE noise accumulation in amplifier chains — and the metrics (OSNR, GOSNR, required-OSNR-at-BER-threshold) that translate physics into engineering decisions. Where numbers appear, they are drawn from ITU-T fiber characterization data, vendor-published loss coefficients for deployed fiber types, and OIF-published symbol rates and channel widths for current-generation coherent interfaces.

The calibration target is: a newly-graduated optical engineer should finish this article with a complete mental model of how a link closes; a fifteen-year system architect should find at least one correlation or number that sharpens a planning assumption they have been using from memory. Every section is designed to answer three questions in sequence — what is the physical mechanism, what does it do in this system, and what happens if it is absent or fails.

2. The spectrum — where light goes and why

Silica-based single-mode fiber is transparent enough to carry information across oceans, but only within a few specific wavelength windows. The boundaries of those windows are set by the same four physical mechanisms that determine which parts of the radio spectrum propagate through the atmosphere: intrinsic material loss, resonance absorption by impurities, Rayleigh scattering that scales as λ⁻⁴, and infrared absorption by silica itself at wavelengths above about 1600 nm. The result is a characteristic loss curve with a minimum near 1550 nm, a deep valley at 1310 nm, and a water-absorption peak near 1383 nm that modern fibers either suppress (low-water-peak variants) or exploit less aggressively (legacy fiber).

The ITU-T standardises the usable bands against this loss spectrum. The most-deployed fiber type today, ITU-T G.652.D standard single-mode fiber, is characterised for use from 1260 nm to 1625 nm, with the E-band water peak suppressed. G.654.E fiber — the ultra-low-loss, large-effective-area variant used on new submarine and long-haul terrestrial builds — targets 1530 – 1625 nm with attenuation as low as 0.15 – 0.17 dB/km and an effective area in excess of 110 µm². G.655 non-zero dispersion-shifted fiber was designed for the C-band era of dispersion-managed 10 Gb/s transport and is now a legacy constraint for coherent upgrades rather than a deployment target.

The band names in Figure 1 are not marketing labels — they determine which amplifier technology is available. The C-band (1530 – 1565 nm) is the historical home of erbium-doped fiber amplifiers because the erbium energy levels produce efficient gain in exactly that window. The L-band (1565 – 1625 nm) also supports erbium gain but with a longer gain fiber and somewhat lower power efficiency. Together, C and L give approximately 10 THz of amplified spectrum — the substrate on which DWDM at scale is built. Everything outside those two bands is either shorter-reach IM-DD territory (1310 nm) or requires amplification technology other than EDFA, which is why the S-band and U-band have historically carried pump, supervisory, and maintenance traffic rather than revenue-bearing wavelengths.

A working engineer should internalise two spectrum facts. First, the total usable C+L spectrum is finite — roughly 9.6 THz under typical guard-band conditions, corresponding to about 96 channels at 100 GHz or 128 channels at 75 GHz spacing. Second, the attenuation differences between bands are real and directly enter span-loss calculations: a 1590 nm L-band wavelength on standard G.652 fiber sees approximately 0.02 dB/km more loss than a 1550 nm C-band wavelength, which over 80 km compounds to 1.6 dB of extra span loss — a margin an OSNR-limited design may not have.

3. Attenuation — the first-order loss physics

Attenuation is the coefficient that converts kilometres of fiber into dB of signal loss. It is the first parameter an engineer plugs into a link budget, and it is the one most often taken from a nominal value rather than the actual characterisation of the deployed cable. The difference matters: a 0.02 dB/km error across an 80 km span is 1.6 dB — enough to push a margin-limited 16QAM channel below its required OSNR.

The physics has three contributors. Rayleigh scattering is the dominant mechanism below the infrared absorption tail; it arises from microscopic refractive-index fluctuations frozen into the glass during fiber draw and scales as λ⁻⁴. Material absorption — primarily from trace hydroxyl (OH⁻) ions — creates the 1383 nm water peak that makes E-band operation impractical on legacy fiber. Infrared absorption by the silica matrix itself becomes significant above 1600 nm and sets the long-wavelength cutoff. Waveguide imperfections (microbending, macrobending, splice losses) contribute additional loss that is not wavelength-selective but accumulates along the route.

Table 1 summarises the attenuation coefficients typically used for deployed fiber types. These values come from vendor-published reference tables and ITU-T fiber characterisation data. For link budget work, the key insight is the roughly 0.08 dB/km spread between the best G.654-class fiber and legacy dispersion-shifted fiber — a factor that over 1000 km of cable plant corresponds to 80 dB of total span loss difference, or approximately four additional amplifier spans' worth of ASE accumulation.

The span-loss calculation that drops out of this table is straightforward in form but consequential in practice:

/* Total span loss for a single amplifier span */
L_span = α × D + L_splices + L_connectors + L_margin

where:
  α           fiber attenuation coefficient (dB/km)
  D           span length (km)
  L_splices   ~0.05 – 0.10 dB per fusion splice
  L_connectors ~0.20 – 0.50 dB per connector pair
  L_margin    fiber degradation reserve, 2 – 3 dB typical

Worked example — 80 km G.652.D span:
  L_span = 0.20 × 80 + 0.05×8 + 0.5×2 + 2.5
       = 16.0 + 0.4 + 1.0 + 2.5 = 19.9 dB

Same 80 km on G.654.E (α = 0.167):
  L_span = 0.167 × 80 + 3.9 = 17.3 dB
       → 2.6 dB improvement from fiber alone

That 2.6 dB is not a theoretical benefit. It is the gain an amplifier chain does not need to provide, which means an ASE contribution that is never added. Over an 8-span long-haul link, choosing G.654.E over G.652.D recovers about 2.6 dB of per-span OSNR budget — enough to change the modulation format that closes on a given route from 16QAM back to 32QAM, or to extend the reach of a 64QAM channel by roughly a factor of two. This is the cleanest, most linear trade in optical networking: pay more for better fiber, get more capacity out of the same route. The reason it is not universal is that the fiber is installed once and the cost is amortised over decades — upgrades target transponders, amplifiers, and DSPs instead.

3.1 Wavelength dependence and band tilt

Attenuation is not flat across C+L. A typical G.652.D fiber has about 0.19 dB/km at 1550 nm, rising to 0.21 – 0.22 dB/km at the L-band edge near 1620 nm. In a fully loaded C+L system with 96+ channels per band, this wavelength-dependent loss manifests as a tilt across the spectrum after each span. Amplifiers flatten it imperfectly — EDFA gain flatness is typically 1 – 2 dB peak-to-peak after a gain flattening filter — and stimulated Raman scattering (SRS) inside the fiber transfers power from shorter to longer wavelengths, adding another 1 – 3 dB of tilt across the full C+L window. Every line system deploys dynamic gain equalisers (DGEs) or wavelength-selective switches with DGE capability to keep the tilt inside the per-channel OSNR budget.

For design, the practical rule is that wavelengths near the band edges carry 1 – 2 dB less OSNR margin than band-centre wavelengths in a fully loaded system. Channel plans that put high-rate, margin-limited services on edge wavelengths are fighting physics. The more robust approach is to keep the hardest modulation formats in the middle of the band and accept that edge slots will carry lower-rate, higher-margin traffic.

Sanjay Yadav

Optical Communications & Network Automation Expert | Author of 3 Books for Optical Engineers | Founder, MapYourTech

Optical networking engineer with nearly two decades of experience across DWDM, OTN, coherent optics, submarine systems, and cloud infrastructure. Founder of MapYourTech. Read full bio →

Follow on LinkedIn

Related Articles on MapYourTech

Free · Jul 11, 2026

Connector Types and Their Loss Budgets: SC, LC, MPO

Automation · Jul 11, 2026

Cooled vs Uncooled Lasers: The Pluggable Power Trade-off

Automation · Jul 11, 2026

Automation Blast Radius: Scoping What a Bad Intent Can Touch

Analysis · Jul 11, 2026

ASE Channel Loading: Running a Line System Full From Day One