Model-based end-to-end learning for a self-homodyne coherent system

Abstract

We investigate the statistical properties of the inherent intensity fluctuation in a low-cost and low-complexity self-homodyne coherent system employing an amplified spontaneous emission (ASE) source. The noise distribution model of the considered system is established, which is shown to be highly consistent with the experimental results for a 10 GBd 256-ary quadrature amplitude modulation (QAM) signal transmission over a 10 m duplex fiber. With the help of the proposed noise model, we then design advanced mappers and demappers. The optimized system alleviates the need for ASE bandwidth and is evaluated by applying forward error correction codes. Furthermore, we demonstrate an information rate increase of 6.67% with respect to 64-QAM.