Low-Complexity Frequency-Dependent Linearizers Based on Parallel Bias-Modulus and Bias-ReLU Operations

TL;DR

This paper presents low-complexity, frequency-dependent linearizers inspired by CNNs that effectively suppress nonlinear distortion in analog-to-digital interfaces, outperforming traditional methods with simpler implementation.

Contribution

Introduces novel frequency-dependent linearizers based on parallel bias-modulus and bias-ReLU operations, with a matrix inversion design that avoids iterative optimization.

Findings

Achieves 20-30 dB SNDR improvement

Lower implementation complexity than Hammerstein linearizers

Effective for wideband multi-tone signals and noise

Abstract

This paper introduces low-complexity frequency-dependent (memory) linearizers designed to suppress nonlinear distortion in analog-to-digital interfaces. Two different linearizers are considered, based on nonlinearity models which correspond to sampling before and after the nonlinearity operations, respectively. The proposed linearizers are inspired by convolutional neural networks but have an order-of-magnitude lower implementation complexity compared to existing neural-network-based linearizer schemes. The proposed linearizers can also outperform the traditional parallel Hammerstein (as well as Wiener) linearizers even when the nonlinearities have been generated through a Hammerstein model. Further, a design procedure is proposed in which the linearizer parameters are obtained through matrix inversion. This eliminates the need for costly and time-consuming iterative nonconvex optimization which is traditionally associated with neural network training. The design effectively handles a wide range of wideband multi-tone signals and filtered white noise. Examples demonstrate significant signal-to-noise-and-distortion ratio (SNDR) improvements of some $20$--$30$ dB, as well as a lower implementation complexity than the Hammerstein linearizers.