No abstract

No abstract

Optical communication systems often include nonlinear components, such as Mach-Zehnder modulators (MZMs). The components' nonlinearity may have a deleterious effect on the system performance, especially when the memory effect is included. One of the most common methods to mitigate the resultant distortion effects is digital predistortion (DPD) based on the Volterra polynomial model. Typically, the DPD coefficients are extracted using the least squares approach. However, naive implementation of the Volterra polynomial model usually introduces significant complexity due to the large number of model coefficients. Here, we propose the use of orthogonal polynomial basis functions for efficient DPD implementation. The orthogonal basis enables the estimation of each coefficient separately, which provides a significant computational gain. Furthermore, the amount of dominant terms in the orthogonal basis representation is significantly lower, which leads to dramatic relaxation in the DPD implementation. In addition to the analytical modeling approach, a nonparametric method is developed by applying eigenvalue decomposition on the correlation matrix, which is required for the case of dependent random variables, or for the case of unknown probability distribution functions. MZM-based lab experiments and simulations were performed, which indicated a potential saving of 50%-80% in the amount of useful dimensions.

High-speed optical communication systems may suffer from a combination of impairments such as memory effect and nonlinear behavior of the optoelectronic components. Nonlinear digital pre-distortion (DPD) is one of the well-known technique to alleviate these effects. As typical implementation of Volterra-based DPD is considered complex and consumes high power, more efficient orthogonal-based Volterra series representation has been proposed. Previous works offered ways to perform efficient grading of the most dominant dimensions based on the combination of the dimensions variances and the signal projection. Here, it is shown that normalization of the data dynamic range further improves this method and decreases significantly the number of required dimensions. Using normalization combined with the previous methods, maximizes the DPD performance by means of error vector magnitude (EVM) and bit error rate (BER), while minimizing the DPD complexity in the terms of required series dimensions. Extensive simulation and lab measurements indicate a potential saving of up to 87% in the number of dimensions with a negligible performance penalty.

No abstract