Revisiting Efficient Multi-Step Nonlinearity Compensation with Machine Learning: An Experimental Demonstration

Nonlinearity compensation in fiber optical communication systems has been for a long time considered a key enabler for going beyond the "capacity crunch". One of the guiding principles for the design of practical nonlinearity compensation schemes appears to be that fewer steps are better and more efficient. In this paper, we challenge this assumption and show how to carefully design multi-step approaches that can lead to better performance-complexity trade-offs than their few-step counterparts. We consider the recently proposed learned digital backpropagation (LDBP) approach, where the linear steps in the split-step method are re-interpreted as general linear functions, similar to the weight matrices in a deep neural network. Our main contribution lies in an experimental demonstration of this approach for a 25 Gbaud single-channel optical transmission system. It is shown how LDBP can be integrated into a coherent receiver DSP chain and successfully trained in the presence of various hardware impairments. Our results show that LDBP with limited complexity can achieve better performance than standard DBP by using very short, but jointly optimized, finite-impulse response filters in each step. This paper also provides an overview of recently proposed extensions of LDBP and we comment on potentially interesting avenues for future work.