Efficient Multidimensional Polynomial Filtering for Nonlinear Digital Predistortion

In all of these cases, one-dimensional coefficient swaths of the kernel are used, taking advantage of the ability to perform fast one-dimensional convolutions. Using multidimensional filters, however, has advantages. Filtering in multiple dimensions more easily addresses asymmetric nonlinearities that sometimes occur in practice. To generalize the division of the Volterra kernel to use multidimensional subkernels, we have created the cube coefficient subspace (CCS) architecture [5], which builds a polynomial signal processor with small nonlinear filters of arbitrary dimension and polynomial order. This allows us to adapt our system in more dimensions with fewer components, which leads to lower computational complexity for a given modeling error performance criterion.