LUT-based end-to-end optical communication system for constellation shaping

This paper presents a look-up table (LUT)-based end-to-end (E2E) optical communication system for constellation shaping. The LUT encoder enables efficient initialization and enhances the convergence performance of the E2E learning system. We explore geometric constellation shaping, probabilistic constellation shaping, as well as joint geometric and probabilistic constellation shaping (GeoPCS), with GeoPCS achieving the highest mutual information gain of 0.2104 bit/sym over standard QAM. To address the gradient backpropagation blocking problem, the paper introduces a conditional generative adversarial network as a surrogate channel to facilitate the backpropagation of gradients from the decoder to the encoder, allowing for efficient optimization of the system. The proposed system's performance is validated through the improved generalized mutual information, demonstrating its potential for constellation optimization in real optical communication systems.