Learning a Time-Frequency Predistortion for Optical Coherent Digital Sub-Carrier Multiplexing

The most significant channel impairments in optical coherent frequency division multiplex-systems are inter-(sub-)carrier distortions, i.e., cross-phase modulation and fourwave mixing. Maximizing the throughput in such a system requires mitigation strategies to compensate for them. One common approach is to apply a computationally complex nonlinearity compensation over all interfering channels, which is not feasible at today’s baudrates. In this work, we instead propose a learned symbolwise and cross-carrier predistortion for coherent optical digital sub-carrier systems that is embedded in an autoencodersetup and trained end-to-end. For a four sub-carrier-system we were able to achieve a spectral efficiency gain of ca. 0.12 bit/s/Hz compared to a per-carrier digital back-propagation, where ca. 65% stem from the cross-carrier ability of the trainable architectural template. The convolutional processing in frequency domain allows to extrapolate the application and results to even higher numbers of sub-carriers. Examining the results of the autoencoder shows that a specific ring constellation in combination with a predistortion that uses all available memory was learned.