Which Bits Should Ride the Sacrificed Ring? Perceptual Bit Mapping for Nonlinear Fiber

Long optical fibers have a hard throughput ceiling. Turning up the laser power raises the signal-to-noise ratio only up to a point, because glass responds to intense light by bending the light's phase (the Kerr effect). Beyond an optimal power, this self-inflicted distortion grows faster than the signal. The best transmit alphabets for this regime look strange. When engineers optimized amplitude-and-phase rings of points for a fiber with this power-dependent phase noise, they found that the best design pushes the outermost points to very high power. Those points are deliberately sacrificed: they absorb most of the phase damage so that the inner rings stay clean, and the average symbol error rate drops (Design of APSK Constellations for Coherent Optical Channels with Nonlinear Phase Noise). The consequence is that, in the nonlinear regime, some bit positions in each transmitted symbol are far less reliable than others. Which positions are unreliable depends on launch power, and the ordering can be the reverse of what holds on an ordinary noisy channel.

Compressed images and video are equally uneven. A few bits (low-frequency coefficients, the most significant bit-planes) decide whether a picture is recognizable, and many others only add fine texture. Vision research has also shown that counting bit errors, or even pixel squared error, is a poor stand-in for what people see. The proposal is to bring these two facts together. We will design the transmit alphabet and the assignment of bits to symbol positions so that they minimize the expected perceived damage to transmitted images, using the power-dependent reliability profile of the nonlinear fiber. The bits a viewer cares about should never ride the sacrificed ring. We expect this to matter most past the optimal launch power, where alphabet designs tuned for average error and protection schemes tuned for linear channels both make the wrong choice.

Core Claim

On a nonlinear fiber, the reliability ranking of bit positions within a symbol changes with launch power. At high power it inverts relative to the additive-Gaussian-noise case that conventional unequal-protection designs assume. We therefore expect a mapping chosen against a perceptual objective, with power-dependent reliabilities, to deliver visibly better images at the same net rate and launch power than two alternatives: an alphabet tuned for average error, and an importance-ordered mapping inherited from linear channels. The objective is

$$ D(\mathcal{C}, \pi, P) = \sum_{k} w_k \, p_k(\mathcal{C}, \pi, P), $$

where $\mathcal{C}$ is the constellation geometry, $\pi$ assigns source bit classes $k$ to labeling positions, $P$ is launch power, $p_k$ is the error probability that position assignment $\pi$ gives class $k$ on this fiber, and $w_k$ is the measured perceptual damage caused by one error in class $k$. The novel ingredient is $p_k$'s dependence on $P$ under Kerr distortion. The perceptual weights $w_k$ are the second, testable ingredient.

Relation to Prior Work

Three fiber papers supply the channel side and leave the application side untouched. The APSK design study optimizes ring constellations and bit labelings under a memoryless nonlinear-phase-noise model with two-stage detection, reporting average error only. We reuse its channel model and its Nelder–Mead search, but we replace its objective. Demodulation and Detection Schemes for a Memoryless Optical WDM Channel proves that Euclidean-distance detection hits an error floor at high power on nonlinear channels, while matched-demodulation receivers do not. That result constrains us: we must evaluate under strong receivers so that gains are not an artifact of a weak detector. Meta-DSP improves nonlinear equalization across symbol rates and reports Q-factor. We use it as an optional receiver front-end to test whether our gains survive good equalization.

The end-to-end autoencoder of Deep Learning Autoencoders for Reducing PAPR in Coherent Optical Systems is the closest learned joint design. It trains transmitter and receiver over a simulated coherent optical link with a squared-error reconstruction loss plus a peak-power penalty, on symbols rather than images. We differ in what is optimized (perceptual damage of real image payloads) and in where the leverage comes from (power-dependent bit reliability, not peak power). Its finding that learned gains vanish below 25 dB optical SNR also bounds the regime in which we should expect any transmitter-side gain.

On the perception side, Perceptual Fidelity Aware Mean Squared Error shows that Gaussian prefiltering before squared error raises rank correlation with human scores on TID2008 from 0.7718 to 0.9162 while staying cheap and convex. That makes it a practical source of $w_k$. BiRQA is a lightweight full-reference metric (SROCC 0.988 on LIVE) that is hardened against adversarial inputs, but it admits weakness on structural distortions. UNIQUE supplies pairwise human-preference training and the observation that rater disagreement peaks at mid quality. None of these metrics has been checked against the burst-like, coefficient-level damage that channel errors produce. We therefore treat metric validity as something we must establish, not assume.

Research Program

We will build the pipeline in three stages.

The first stage defines the payload. Images from TID2013 and KADID-10K reference sets are coded into block-transform coefficient bit-planes. A small header is always heavily protected, and resynchronization markers keep each error local. With this setup the experiment measures mapping effects rather than decoder crashes. We group bits into roughly eight classes by frequency band and bit-plane.

The second stage builds the channel. We will use the memoryless nonlinear-phase-noise model from the APSK study and a split-step simulation of the nonlinear Schrödinger equation with dispersion and several wavelength channels. The split-step simulation addresses that study's own stated limitation of ignoring dispersion memory. From both, we extract $p_k$ as a function of launch power for 16-point QAM and for optimized 16-APSK under every candidate labeling.

The third stage measures the perceptual weights $w_k$ by injecting single-class errors at controlled rates and scoring the results. Before trusting any metric, we will run a small paired-comparison human study on fiber-error-distorted images, scored with UNIQUE-style pairwise fidelity, and select among PSNR, PAMSE, SSIM and BiRQA by agreement with those judgments. We then minimize $D$ jointly over ring radii, phase offsets and the class-to-position map, for each launch power on a grid that spans the linear optimum and the regime past it.

Baselines

The naive reference is square 16-QAM with Gray labeling, with bits assigned in stream order and no importance awareness. This is what a bit-pipe transceiver does.

The strongest existing channel-side method is the optimized 16-APSK with two-stage detection and best average-error labeling from the APSK study. It reports a 3.2 dB power-margin gain over 16-QAM at symbol error probability $10^{-2}$, and a 0.4 dB gain from offset labeling over Gray for 8-APSK. We will reimplement it on the identical channel and payload, giving its search the same Nelder–Mead iteration budget that ours uses.

The strongest unequal-protection competitor is the same APSK geometry with bits assigned by importance according to the reliability ordering computed under additive Gaussian noise. This is the classical hierarchical-modulation recipe. It isolates whether power-aware reliability matters.

A coding-side competitor applies stronger error-correcting code rates to important classes over a plain alphabet, matched in net rate.

Two ablations isolate the key ingredients. One replaces perceptual weights with squared-error weights. The other replaces split-step-measured $p_k$ with memoryless-model $p_k$.

For the perceptual outcome itself, we found no reported result on Lacuna for image quality delivered over a nonlinear fiber link, so every number on that axis will be new. Beating the baselines means a statistically significant improvement in human preference and in the validated metric at equal net rate and launch power, past the nonlinear optimum. That improvement must also come without a loss at or below the optimum.

Why not just…?

Why not just rely on separation and strong forward error correction? When the coded error rate is truly zero, our method has nothing to add. We will show that region explicitly, as the low-power part of every curve. The case for this work rests on operating near or past the error-correction threshold, for example to maximize net rate or in low-latency links with light coding. There, the sacrificed ring concentrates residual errors in specific positions. The file's autoencoder results, with error floors at small training noise and no gain below 25 dB optical SNR, show that error-free operation cannot be assumed near the limit. Still, the honest summary is that this is a near-threshold technique, not a universal one.

Why not just use classical unequal error protection by coding or hierarchical modulation? This is the real threat. If the reliability ordering never inverts across powers, linear-channel hierarchical mapping should match us. The coded-protection baseline may then win outright at moderate power. The power-sweep comparison against both baselines settles this. If no inversion appears in split-step simulations with dispersion, we will report that the memoryless-model inversion is an artifact and the claim fails.

Why not just weight by squared error, or run SSIM or LPIPS at the receiver? Squared-error weights may rank bit classes almost identically to perceptual ones. The weight ablation tests this directly, and if it ties, the contribution reduces to nonlinear-aware mapping. Receiver-side metrics like SSIM do not change what is transmitted. They are also non-convex and, per the PAMSE paper, ten times slower than squared error. We use them only to estimate weights offline.

Why not just lower the launch power into the linear regime? That forfeits margin. The APSK result quantifies 3.2 dB available at high power, so we compare at the power that maximizes each scheme's own quality.

Decisive Checks

Three results decide the outcome. First, the split-step simulation must show a launch power above which the reliability ranking of bit positions differs from the Gaussian-noise ranking. Without that, the core claim is false. Second, past that power, the perceptual mapping must beat power-blind hierarchical mapping and rate-matched coded protection in human paired preference. Third, the advantage must persist under a strong matched-demodulation receiver and with Meta-DSP-style equalization in front. If it disappears there, the effect belongs to weak receivers rather than to the fiber.