Nonlinear Channel Equalization Using Gaussian Processes Regression in IMDD Fiber Link

Gaussian processes regression (GPR)-aided nonlinear channel equalizer (CE) is experimentally demonstrated in a multi-level intensity modulation and direct detection fiber link. In this scheme, the GPR model is used to estimate the transmitted symbols or the corresponding nonlinear distortions after pre-processing. The experimental results show that GPR-aided nonlinear CE has better nonlinear tolerance than conventional linear and nonlinear filter-based CE. It is also shown that the GPR model in the nonlinear channel equalization process can be understood as an optimized single-layer neural network model with infinite width. Finally, we reveal the relationship between the key coefficients in GPR model and parameters in fiber link through both experiment and simulation.