Sparse Tensor Network System Identification for Nonlinear Circuit Macromodeling

Abstract

Volterra-series nonlinear system identification has recently been transformed into a tensor decomposition and opti-mization problem. Such tensor network reformulation with small tensor cores overcomes the curse of dimensionality in deploying high order Volterra kernels, and allows the fast simulation of arbitrarily nonlinear systems. This paper further advances the tensor core identification by adding sparsity: pruning unimportant weights and retraining the remaining weights, while retaining most accuracy. The resultant sparse macromodels are hardware friendly and can facilitate low power signal processing.