Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Probabilistic Amplitude Shaping for Nonlinear Fiber Channels

Mohammad Taha Askari, Lutz Lampe, Amirhossein Ghazisaeidi

TL;DR

The paper addresses NLIN in coherent fiber channels by noting that traditional PAS ignores temporal dependencies in symbol sequences. It introduces NPAS, a PAS-compatible autoregressive neural network that learns the joint distribution over unsigned amplitudes, trained end-to-end with a differentiable AM perturbative fiber model and an adjusted BCE loss to maximize AIR. NPAS achieves performance comparable to NPS with greater stability at large blocklengths and shows tangible gains over sequence selection (e.g., >0.5 dB SNR and ~0.1 bits/2D AIR on a 205 km link with 5×64-QAM at 50 GBd). The work demonstrates that joint-distribution learning can be integrated within PAS without sacrificing compatibility or practicality, though multi-span extensions remain for future work.

Abstract

We introduce neural probabilistic amplitude shaping, a joint-distribution learning framework for coherent fiber systems. The proposed scheme provides a 0.5 dB signal-to-noise ratio gain over sequence selection for dual-polarized 64-QAM transmission across a single-span 205 km link.

Neural Probabilistic Amplitude Shaping for Nonlinear Fiber Channels

TL;DR

The paper addresses NLIN in coherent fiber channels by noting that traditional PAS ignores temporal dependencies in symbol sequences. It introduces NPAS, a PAS-compatible autoregressive neural network that learns the joint distribution over unsigned amplitudes, trained end-to-end with a differentiable AM perturbative fiber model and an adjusted BCE loss to maximize AIR. NPAS achieves performance comparable to NPS with greater stability at large blocklengths and shows tangible gains over sequence selection (e.g., >0.5 dB SNR and ~0.1 bits/2D AIR on a 205 km link with 5×64-QAM at 50 GBd). The work demonstrates that joint-distribution learning can be integrated within PAS without sacrificing compatibility or practicality, though multi-span extensions remain for future work.

Abstract

We introduce neural probabilistic amplitude shaping, a joint-distribution learning framework for coherent fiber systems. The proposed scheme provides a 0.5 dB signal-to-noise ratio gain over sequence selection for dual-polarized 64-QAM transmission across a single-span 205 km link.
Paper Structure (5 sections, 1 equation, 2 figures)

This paper contains 5 sections, 1 equation, 2 figures.

Table of Contents

  1. Introduction
  2. NPAS Architecture and Training
  3. Numerical Results
  4. Conclusion
  5. Acknowledgements

Figures (2)

  • Figure 1: NPAS end-to-end training pipeline.
  • Figure 2: Performance comparison of NPAS with conventional PAS using ESS and with NPS. (a) AIR vs. blocklength $L$; (b) effective SNR vs. launch power; and (c) AIR vs. launch power.